Description: split mpegvideo.c
[ffmpeg.git] / libavcodec / mpegvideo_enc.c
1 /*
2  * The simplest mpeg encoder (well, it was the simplest!)
3  * Copyright (c) 2000,2001 Fabrice Bellard.
4  * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
5  *
6  * This file is part of FFmpeg.
7  *
8  * FFmpeg is free software; you can redistribute it and/or
9  * modify it under the terms of the GNU Lesser General Public
10  * License as published by the Free Software Foundation; either
11  * version 2.1 of the License, or (at your option) any later version.
12  *
13  * FFmpeg is distributed in the hope that it will be useful,
14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
16  * Lesser General Public License for more details.
17  *
18  * You should have received a copy of the GNU Lesser General Public
19  * License along with FFmpeg; if not, write to the Free Software
20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21  *
22  * 4MV & hq & b-frame encoding stuff by Michael Niedermayer <michaelni@gmx.at>
23  */
24
25 /**
26  * @file mpegvideo_enc.c
27  * The simplest mpeg encoder (well, it was the simplest!).
28  */
29
30 #include "avcodec.h"
31 #include "dsputil.h"
32 #include "mpegvideo.h"
33 #include "mpegvideo_common.h"
34 #include "mjpegenc.h"
35 #include "msmpeg4.h"
36 #include "faandct.h"
37 #include <limits.h>
38
39 //#undef NDEBUG
40 //#include <assert.h>
41
42 static int encode_picture(MpegEncContext *s, int picture_number);
43 static int dct_quantize_refine(MpegEncContext *s, DCTELEM *block, int16_t *weight, DCTELEM *orig, int n, int qscale);
44 static int sse_mb(MpegEncContext *s);
45
46 /* enable all paranoid tests for rounding, overflows, etc... */
47 //#define PARANOID
48
49 //#define DEBUG
50
51 static const uint16_t aanscales[64] = {
52     /* precomputed values scaled up by 14 bits */
53     16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
54     22725, 31521, 29692, 26722, 22725, 17855, 12299,  6270,
55     21407, 29692, 27969, 25172, 21407, 16819, 11585,  5906,
56     19266, 26722, 25172, 22654, 19266, 15137, 10426,  5315,
57     16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
58     12873, 17855, 16819, 15137, 12873, 10114,  6967,  3552,
59     8867 , 12299, 11585, 10426,  8867,  6967,  4799,  2446,
60     4520 ,  6270,  5906,  5315,  4520,  3552,  2446,  1247
61 };
62
63 static uint8_t default_mv_penalty[MAX_FCODE+1][MAX_MV*2+1];
64 static uint8_t default_fcode_tab[MAX_MV*2+1];
65
66 enum PixelFormat ff_yuv420p_list[2]= {PIX_FMT_YUV420P, -1};
67
68 static void convert_matrix(DSPContext *dsp, int (*qmat)[64], uint16_t (*qmat16)[2][64],
69                            const uint16_t *quant_matrix, int bias, int qmin, int qmax, int intra)
70 {
71     int qscale;
72     int shift=0;
73
74     for(qscale=qmin; qscale<=qmax; qscale++){
75         int i;
76         if (dsp->fdct == ff_jpeg_fdct_islow
77 #ifdef FAAN_POSTSCALE
78             || dsp->fdct == ff_faandct
79 #endif
80             ) {
81             for(i=0;i<64;i++) {
82                 const int j= dsp->idct_permutation[i];
83                 /* 16 <= qscale * quant_matrix[i] <= 7905 */
84                 /* 19952         <= aanscales[i] * qscale * quant_matrix[i]           <= 249205026 */
85                 /* (1<<36)/19952 >= (1<<36)/(aanscales[i] * qscale * quant_matrix[i]) >= (1<<36)/249205026 */
86                 /* 3444240       >= (1<<36)/(aanscales[i] * qscale * quant_matrix[i]) >= 275 */
87
88                 qmat[qscale][i] = (int)((UINT64_C(1) << QMAT_SHIFT) /
89                                 (qscale * quant_matrix[j]));
90             }
91         } else if (dsp->fdct == fdct_ifast
92 #ifndef FAAN_POSTSCALE
93                    || dsp->fdct == ff_faandct
94 #endif
95                    ) {
96             for(i=0;i<64;i++) {
97                 const int j= dsp->idct_permutation[i];
98                 /* 16 <= qscale * quant_matrix[i] <= 7905 */
99                 /* 19952         <= aanscales[i] * qscale * quant_matrix[i]           <= 249205026 */
100                 /* (1<<36)/19952 >= (1<<36)/(aanscales[i] * qscale * quant_matrix[i]) >= (1<<36)/249205026 */
101                 /* 3444240       >= (1<<36)/(aanscales[i] * qscale * quant_matrix[i]) >= 275 */
102
103                 qmat[qscale][i] = (int)((UINT64_C(1) << (QMAT_SHIFT + 14)) /
104                                 (aanscales[i] * qscale * quant_matrix[j]));
105             }
106         } else {
107             for(i=0;i<64;i++) {
108                 const int j= dsp->idct_permutation[i];
109                 /* We can safely suppose that 16 <= quant_matrix[i] <= 255
110                    So 16           <= qscale * quant_matrix[i]             <= 7905
111                    so (1<<19) / 16 >= (1<<19) / (qscale * quant_matrix[i]) >= (1<<19) / 7905
112                    so 32768        >= (1<<19) / (qscale * quant_matrix[i]) >= 67
113                 */
114                 qmat[qscale][i] = (int)((UINT64_C(1) << QMAT_SHIFT) / (qscale * quant_matrix[j]));
115 //                qmat  [qscale][i] = (1 << QMAT_SHIFT_MMX) / (qscale * quant_matrix[i]);
116                 qmat16[qscale][0][i] = (1 << QMAT_SHIFT_MMX) / (qscale * quant_matrix[j]);
117
118                 if(qmat16[qscale][0][i]==0 || qmat16[qscale][0][i]==128*256) qmat16[qscale][0][i]=128*256-1;
119                 qmat16[qscale][1][i]= ROUNDED_DIV(bias<<(16-QUANT_BIAS_SHIFT), qmat16[qscale][0][i]);
120             }
121         }
122
123         for(i=intra; i<64; i++){
124             int64_t max= 8191;
125             if (dsp->fdct == fdct_ifast
126 #ifndef FAAN_POSTSCALE
127                    || dsp->fdct == ff_faandct
128 #endif
129                    ) {
130                 max= (8191LL*aanscales[i]) >> 14;
131             }
132             while(((max * qmat[qscale][i]) >> shift) > INT_MAX){
133                 shift++;
134             }
135         }
136     }
137     if(shift){
138         av_log(NULL, AV_LOG_INFO, "Warning, QMAT_SHIFT is larger than %d, overflows possible\n", QMAT_SHIFT - shift);
139     }
140 }
141
142 static inline void update_qscale(MpegEncContext *s){
143     s->qscale= (s->lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
144     s->qscale= av_clip(s->qscale, s->avctx->qmin, s->avctx->qmax);
145
146     s->lambda2= (s->lambda*s->lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
147 }
148
149 void ff_write_quant_matrix(PutBitContext *pb, uint16_t *matrix){
150     int i;
151
152     if(matrix){
153         put_bits(pb, 1, 1);
154         for(i=0;i<64;i++) {
155             put_bits(pb, 8, matrix[ ff_zigzag_direct[i] ]);
156         }
157     }else
158         put_bits(pb, 1, 0);
159 }
160
161 static void copy_picture_attributes(MpegEncContext *s, AVFrame *dst, AVFrame *src){
162     int i;
163
164     dst->pict_type              = src->pict_type;
165     dst->quality                = src->quality;
166     dst->coded_picture_number   = src->coded_picture_number;
167     dst->display_picture_number = src->display_picture_number;
168 //    dst->reference              = src->reference;
169     dst->pts                    = src->pts;
170     dst->interlaced_frame       = src->interlaced_frame;
171     dst->top_field_first        = src->top_field_first;
172
173     if(s->avctx->me_threshold){
174         if(!src->motion_val[0])
175             av_log(s->avctx, AV_LOG_ERROR, "AVFrame.motion_val not set!\n");
176         if(!src->mb_type)
177             av_log(s->avctx, AV_LOG_ERROR, "AVFrame.mb_type not set!\n");
178         if(!src->ref_index[0])
179             av_log(s->avctx, AV_LOG_ERROR, "AVFrame.ref_index not set!\n");
180         if(src->motion_subsample_log2 != dst->motion_subsample_log2)
181             av_log(s->avctx, AV_LOG_ERROR, "AVFrame.motion_subsample_log2 doesn't match! (%d!=%d)\n",
182             src->motion_subsample_log2, dst->motion_subsample_log2);
183
184         memcpy(dst->mb_type, src->mb_type, s->mb_stride * s->mb_height * sizeof(dst->mb_type[0]));
185
186         for(i=0; i<2; i++){
187             int stride= ((16*s->mb_width )>>src->motion_subsample_log2) + 1;
188             int height= ((16*s->mb_height)>>src->motion_subsample_log2);
189
190             if(src->motion_val[i] && src->motion_val[i] != dst->motion_val[i]){
191                 memcpy(dst->motion_val[i], src->motion_val[i], 2*stride*height*sizeof(int16_t));
192             }
193             if(src->ref_index[i] && src->ref_index[i] != dst->ref_index[i]){
194                 memcpy(dst->ref_index[i], src->ref_index[i], s->b8_stride*2*s->mb_height*sizeof(int8_t));
195             }
196         }
197     }
198 }
199
200 static void update_duplicate_context_after_me(MpegEncContext *dst, MpegEncContext *src){
201 #define COPY(a) dst->a= src->a
202     COPY(pict_type);
203     COPY(current_picture);
204     COPY(f_code);
205     COPY(b_code);
206     COPY(qscale);
207     COPY(lambda);
208     COPY(lambda2);
209     COPY(picture_in_gop_number);
210     COPY(gop_picture_number);
211     COPY(frame_pred_frame_dct); //FIXME don't set in encode_header
212     COPY(progressive_frame); //FIXME don't set in encode_header
213     COPY(partitioned_frame); //FIXME don't set in encode_header
214 #undef COPY
215 }
216
217 /**
218  * sets the given MpegEncContext to defaults for encoding.
219  * the changed fields will not depend upon the prior state of the MpegEncContext.
220  */
221 static void MPV_encode_defaults(MpegEncContext *s){
222     int i;
223     MPV_common_defaults(s);
224
225     for(i=-16; i<16; i++){
226         default_fcode_tab[i + MAX_MV]= 1;
227     }
228     s->me.mv_penalty= default_mv_penalty;
229     s->fcode_tab= default_fcode_tab;
230 }
231
232 /* init video encoder */
233 int MPV_encode_init(AVCodecContext *avctx)
234 {
235     MpegEncContext *s = avctx->priv_data;
236     int i;
237     int chroma_h_shift, chroma_v_shift;
238
239     MPV_encode_defaults(s);
240
241     switch (avctx->codec_id) {
242     case CODEC_ID_MPEG2VIDEO:
243         if(avctx->pix_fmt != PIX_FMT_YUV420P && avctx->pix_fmt != PIX_FMT_YUV422P){
244             av_log(avctx, AV_LOG_ERROR, "only YUV420 and YUV422 are supported\n");
245             return -1;
246         }
247         break;
248     case CODEC_ID_LJPEG:
249     case CODEC_ID_MJPEG:
250         if(avctx->pix_fmt != PIX_FMT_YUVJ420P && avctx->pix_fmt != PIX_FMT_YUVJ422P &&
251            ((avctx->pix_fmt != PIX_FMT_YUV420P && avctx->pix_fmt != PIX_FMT_YUV422P) || avctx->strict_std_compliance>FF_COMPLIANCE_INOFFICIAL)){
252             av_log(avctx, AV_LOG_ERROR, "colorspace not supported in jpeg\n");
253             return -1;
254         }
255         break;
256     default:
257         if(avctx->pix_fmt != PIX_FMT_YUV420P){
258             av_log(avctx, AV_LOG_ERROR, "only YUV420 is supported\n");
259             return -1;
260         }
261     }
262
263     switch (avctx->pix_fmt) {
264     case PIX_FMT_YUVJ422P:
265     case PIX_FMT_YUV422P:
266         s->chroma_format = CHROMA_422;
267         break;
268     case PIX_FMT_YUVJ420P:
269     case PIX_FMT_YUV420P:
270     default:
271         s->chroma_format = CHROMA_420;
272         break;
273     }
274
275     s->bit_rate = avctx->bit_rate;
276     s->width = avctx->width;
277     s->height = avctx->height;
278     if(avctx->gop_size > 600 && avctx->strict_std_compliance>FF_COMPLIANCE_EXPERIMENTAL){
279         av_log(avctx, AV_LOG_ERROR, "Warning keyframe interval too large! reducing it ...\n");
280         avctx->gop_size=600;
281     }
282     s->gop_size = avctx->gop_size;
283     s->avctx = avctx;
284     s->flags= avctx->flags;
285     s->flags2= avctx->flags2;
286     s->max_b_frames= avctx->max_b_frames;
287     s->codec_id= avctx->codec->id;
288     s->luma_elim_threshold  = avctx->luma_elim_threshold;
289     s->chroma_elim_threshold= avctx->chroma_elim_threshold;
290     s->strict_std_compliance= avctx->strict_std_compliance;
291     s->data_partitioning= avctx->flags & CODEC_FLAG_PART;
292     s->quarter_sample= (avctx->flags & CODEC_FLAG_QPEL)!=0;
293     s->mpeg_quant= avctx->mpeg_quant;
294     s->rtp_mode= !!avctx->rtp_payload_size;
295     s->intra_dc_precision= avctx->intra_dc_precision;
296     s->user_specified_pts = AV_NOPTS_VALUE;
297
298     if (s->gop_size <= 1) {
299         s->intra_only = 1;
300         s->gop_size = 12;
301     } else {
302         s->intra_only = 0;
303     }
304
305     s->me_method = avctx->me_method;
306
307     /* Fixed QSCALE */
308     s->fixed_qscale = !!(avctx->flags & CODEC_FLAG_QSCALE);
309
310     s->adaptive_quant= (   s->avctx->lumi_masking
311                         || s->avctx->dark_masking
312                         || s->avctx->temporal_cplx_masking
313                         || s->avctx->spatial_cplx_masking
314                         || s->avctx->p_masking
315                         || s->avctx->border_masking
316                         || (s->flags&CODEC_FLAG_QP_RD))
317                        && !s->fixed_qscale;
318
319     s->obmc= !!(s->flags & CODEC_FLAG_OBMC);
320     s->loop_filter= !!(s->flags & CODEC_FLAG_LOOP_FILTER);
321     s->alternate_scan= !!(s->flags & CODEC_FLAG_ALT_SCAN);
322     s->intra_vlc_format= !!(s->flags2 & CODEC_FLAG2_INTRA_VLC);
323     s->q_scale_type= !!(s->flags2 & CODEC_FLAG2_NON_LINEAR_QUANT);
324
325     if(avctx->rc_max_rate && !avctx->rc_buffer_size){
326         av_log(avctx, AV_LOG_ERROR, "a vbv buffer size is needed, for encoding with a maximum bitrate\n");
327         return -1;
328     }
329
330     if(avctx->rc_min_rate && avctx->rc_max_rate != avctx->rc_min_rate){
331         av_log(avctx, AV_LOG_INFO, "Warning min_rate > 0 but min_rate != max_rate isn't recommended!\n");
332     }
333
334     if(avctx->rc_min_rate && avctx->rc_min_rate > avctx->bit_rate){
335         av_log(avctx, AV_LOG_ERROR, "bitrate below min bitrate\n");
336         return -1;
337     }
338
339     if(avctx->rc_max_rate && avctx->rc_max_rate < avctx->bit_rate){
340         av_log(avctx, AV_LOG_INFO, "bitrate above max bitrate\n");
341         return -1;
342     }
343
344     if(avctx->rc_buffer_size && avctx->bit_rate*av_q2d(avctx->time_base) > avctx->rc_buffer_size){
345         av_log(avctx, AV_LOG_ERROR, "VBV buffer too small for bitrate\n");
346         return -1;
347     }
348
349     if(avctx->bit_rate*av_q2d(avctx->time_base) > avctx->bit_rate_tolerance){
350         av_log(avctx, AV_LOG_ERROR, "bitrate tolerance too small for bitrate\n");
351         return -1;
352     }
353
354     if(   s->avctx->rc_max_rate && s->avctx->rc_min_rate == s->avctx->rc_max_rate
355        && (s->codec_id == CODEC_ID_MPEG1VIDEO || s->codec_id == CODEC_ID_MPEG2VIDEO)
356        && 90000LL * (avctx->rc_buffer_size-1) > s->avctx->rc_max_rate*0xFFFFLL){
357
358         av_log(avctx, AV_LOG_INFO, "Warning vbv_delay will be set to 0xFFFF (=VBR) as the specified vbv buffer is too large for the given bitrate!\n");
359     }
360
361     if((s->flags & CODEC_FLAG_4MV) && s->codec_id != CODEC_ID_MPEG4
362        && s->codec_id != CODEC_ID_H263 && s->codec_id != CODEC_ID_H263P && s->codec_id != CODEC_ID_FLV1){
363         av_log(avctx, AV_LOG_ERROR, "4MV not supported by codec\n");
364         return -1;
365     }
366
367     if(s->obmc && s->avctx->mb_decision != FF_MB_DECISION_SIMPLE){
368         av_log(avctx, AV_LOG_ERROR, "OBMC is only supported with simple mb decision\n");
369         return -1;
370     }
371
372     if(s->obmc && s->codec_id != CODEC_ID_H263 && s->codec_id != CODEC_ID_H263P){
373         av_log(avctx, AV_LOG_ERROR, "OBMC is only supported with H263(+)\n");
374         return -1;
375     }
376
377     if(s->quarter_sample && s->codec_id != CODEC_ID_MPEG4){
378         av_log(avctx, AV_LOG_ERROR, "qpel not supported by codec\n");
379         return -1;
380     }
381
382     if(s->data_partitioning && s->codec_id != CODEC_ID_MPEG4){
383         av_log(avctx, AV_LOG_ERROR, "data partitioning not supported by codec\n");
384         return -1;
385     }
386
387     if(s->max_b_frames && s->codec_id != CODEC_ID_MPEG4 && s->codec_id != CODEC_ID_MPEG1VIDEO && s->codec_id != CODEC_ID_MPEG2VIDEO){
388         av_log(avctx, AV_LOG_ERROR, "b frames not supported by codec\n");
389         return -1;
390     }
391
392     if((s->flags & (CODEC_FLAG_INTERLACED_DCT|CODEC_FLAG_INTERLACED_ME|CODEC_FLAG_ALT_SCAN))
393        && s->codec_id != CODEC_ID_MPEG4 && s->codec_id != CODEC_ID_MPEG2VIDEO){
394         av_log(avctx, AV_LOG_ERROR, "interlacing not supported by codec\n");
395         return -1;
396     }
397
398     if(s->mpeg_quant && s->codec_id != CODEC_ID_MPEG4){ //FIXME mpeg2 uses that too
399         av_log(avctx, AV_LOG_ERROR, "mpeg2 style quantization not supported by codec\n");
400         return -1;
401     }
402
403     if((s->flags & CODEC_FLAG_CBP_RD) && !(s->flags & CODEC_FLAG_TRELLIS_QUANT)){
404         av_log(avctx, AV_LOG_ERROR, "CBP RD needs trellis quant\n");
405         return -1;
406     }
407
408     if((s->flags & CODEC_FLAG_QP_RD) && s->avctx->mb_decision != FF_MB_DECISION_RD){
409         av_log(avctx, AV_LOG_ERROR, "QP RD needs mbd=2\n");
410         return -1;
411     }
412
413     if(s->avctx->scenechange_threshold < 1000000000 && (s->flags & CODEC_FLAG_CLOSED_GOP)){
414         av_log(avctx, AV_LOG_ERROR, "closed gop with scene change detection arent supported yet, set threshold to 1000000000\n");
415         return -1;
416     }
417
418     if((s->flags2 & CODEC_FLAG2_INTRA_VLC) && s->codec_id != CODEC_ID_MPEG2VIDEO){
419         av_log(avctx, AV_LOG_ERROR, "intra vlc table not supported by codec\n");
420         return -1;
421     }
422
423     if(s->flags & CODEC_FLAG_LOW_DELAY){
424         if (s->codec_id != CODEC_ID_MPEG2VIDEO && s->codec_id != CODEC_ID_MPEG1VIDEO){
425             av_log(avctx, AV_LOG_ERROR, "low delay forcing is only available for mpeg1/2\n");
426             return -1;
427         }
428         if (s->max_b_frames != 0){
429             av_log(avctx, AV_LOG_ERROR, "b frames cannot be used with low delay\n");
430             return -1;
431         }
432     }
433
434     if(s->q_scale_type == 1){
435         if(s->codec_id != CODEC_ID_MPEG2VIDEO){
436             av_log(avctx, AV_LOG_ERROR, "non linear quant is only available for mpeg2\n");
437             return -1;
438         }
439         if(avctx->qmax > 12){
440             av_log(avctx, AV_LOG_ERROR, "non linear quant only supports qmax <= 12 currently\n");
441             return -1;
442         }
443     }
444
445     if(s->avctx->thread_count > 1 && s->codec_id != CODEC_ID_MPEG4
446        && s->codec_id != CODEC_ID_MPEG1VIDEO && s->codec_id != CODEC_ID_MPEG2VIDEO
447        && (s->codec_id != CODEC_ID_H263P || !(s->flags & CODEC_FLAG_H263P_SLICE_STRUCT))){
448         av_log(avctx, AV_LOG_ERROR, "multi threaded encoding not supported by codec\n");
449         return -1;
450     }
451
452     if(s->avctx->thread_count > 1)
453         s->rtp_mode= 1;
454
455     if(!avctx->time_base.den || !avctx->time_base.num){
456         av_log(avctx, AV_LOG_ERROR, "framerate not set\n");
457         return -1;
458     }
459
460     i= (INT_MAX/2+128)>>8;
461     if(avctx->me_threshold >= i){
462         av_log(avctx, AV_LOG_ERROR, "me_threshold too large, max is %d\n", i - 1);
463         return -1;
464     }
465     if(avctx->mb_threshold >= i){
466         av_log(avctx, AV_LOG_ERROR, "mb_threshold too large, max is %d\n", i - 1);
467         return -1;
468     }
469
470     if(avctx->b_frame_strategy && (avctx->flags&CODEC_FLAG_PASS2)){
471         av_log(avctx, AV_LOG_INFO, "notice: b_frame_strategy only affects the first pass\n");
472         avctx->b_frame_strategy = 0;
473     }
474
475     i= ff_gcd(avctx->time_base.den, avctx->time_base.num);
476     if(i > 1){
477         av_log(avctx, AV_LOG_INFO, "removing common factors from framerate\n");
478         avctx->time_base.den /= i;
479         avctx->time_base.num /= i;
480 //        return -1;
481     }
482
483     if(s->codec_id==CODEC_ID_MJPEG){
484         s->intra_quant_bias= 1<<(QUANT_BIAS_SHIFT-1); //(a + x/2)/x
485         s->inter_quant_bias= 0;
486     }else if(s->mpeg_quant || s->codec_id==CODEC_ID_MPEG1VIDEO || s->codec_id==CODEC_ID_MPEG2VIDEO){
487         s->intra_quant_bias= 3<<(QUANT_BIAS_SHIFT-3); //(a + x*3/8)/x
488         s->inter_quant_bias= 0;
489     }else{
490         s->intra_quant_bias=0;
491         s->inter_quant_bias=-(1<<(QUANT_BIAS_SHIFT-2)); //(a - x/4)/x
492     }
493
494     if(avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)
495         s->intra_quant_bias= avctx->intra_quant_bias;
496     if(avctx->inter_quant_bias != FF_DEFAULT_QUANT_BIAS)
497         s->inter_quant_bias= avctx->inter_quant_bias;
498
499     avcodec_get_chroma_sub_sample(avctx->pix_fmt, &chroma_h_shift, &chroma_v_shift);
500
501     if(avctx->codec_id == CODEC_ID_MPEG4 && s->avctx->time_base.den > (1<<16)-1){
502         av_log(avctx, AV_LOG_ERROR, "timebase not supported by mpeg 4 standard\n");
503         return -1;
504     }
505     s->time_increment_bits = av_log2(s->avctx->time_base.den - 1) + 1;
506
507     switch(avctx->codec->id) {
508     case CODEC_ID_MPEG1VIDEO:
509         s->out_format = FMT_MPEG1;
510         s->low_delay= !!(s->flags & CODEC_FLAG_LOW_DELAY);
511         avctx->delay= s->low_delay ? 0 : (s->max_b_frames + 1);
512         break;
513     case CODEC_ID_MPEG2VIDEO:
514         s->out_format = FMT_MPEG1;
515         s->low_delay= !!(s->flags & CODEC_FLAG_LOW_DELAY);
516         avctx->delay= s->low_delay ? 0 : (s->max_b_frames + 1);
517         s->rtp_mode= 1;
518         break;
519     case CODEC_ID_LJPEG:
520     case CODEC_ID_MJPEG:
521         s->out_format = FMT_MJPEG;
522         s->intra_only = 1; /* force intra only for jpeg */
523         s->mjpeg_vsample[0] = 2;
524         s->mjpeg_vsample[1] = 2>>chroma_v_shift;
525         s->mjpeg_vsample[2] = 2>>chroma_v_shift;
526         s->mjpeg_hsample[0] = 2;
527         s->mjpeg_hsample[1] = 2>>chroma_h_shift;
528         s->mjpeg_hsample[2] = 2>>chroma_h_shift;
529         if (!(ENABLE_MJPEG_ENCODER || ENABLE_LJPEG_ENCODER)
530             || ff_mjpeg_encode_init(s) < 0)
531             return -1;
532         avctx->delay=0;
533         s->low_delay=1;
534         break;
535     case CODEC_ID_H261:
536         if (!ENABLE_H261_ENCODER)  return -1;
537         if (ff_h261_get_picture_format(s->width, s->height) < 0) {
538             av_log(avctx, AV_LOG_ERROR, "The specified picture size of %dx%d is not valid for the H.261 codec.\nValid sizes are 176x144, 352x288\n", s->width, s->height);
539             return -1;
540         }
541         s->out_format = FMT_H261;
542         avctx->delay=0;
543         s->low_delay=1;
544         break;
545     case CODEC_ID_H263:
546         if (h263_get_picture_format(s->width, s->height) == 7) {
547             av_log(avctx, AV_LOG_INFO, "The specified picture size of %dx%d is not valid for the H.263 codec.\nValid sizes are 128x96, 176x144, 352x288, 704x576, and 1408x1152. Try H.263+.\n", s->width, s->height);
548             return -1;
549         }
550         s->out_format = FMT_H263;
551         s->obmc= (avctx->flags & CODEC_FLAG_OBMC) ? 1:0;
552         avctx->delay=0;
553         s->low_delay=1;
554         break;
555     case CODEC_ID_H263P:
556         s->out_format = FMT_H263;
557         s->h263_plus = 1;
558         /* Fx */
559         s->umvplus = (avctx->flags & CODEC_FLAG_H263P_UMV) ? 1:0;
560         s->h263_aic= (avctx->flags & CODEC_FLAG_AC_PRED) ? 1:0;
561         s->modified_quant= s->h263_aic;
562         s->alt_inter_vlc= (avctx->flags & CODEC_FLAG_H263P_AIV) ? 1:0;
563         s->obmc= (avctx->flags & CODEC_FLAG_OBMC) ? 1:0;
564         s->loop_filter= (avctx->flags & CODEC_FLAG_LOOP_FILTER) ? 1:0;
565         s->unrestricted_mv= s->obmc || s->loop_filter || s->umvplus;
566         s->h263_slice_structured= (s->flags & CODEC_FLAG_H263P_SLICE_STRUCT) ? 1:0;
567
568         /* /Fx */
569         /* These are just to be sure */
570         avctx->delay=0;
571         s->low_delay=1;
572         break;
573     case CODEC_ID_FLV1:
574         s->out_format = FMT_H263;
575         s->h263_flv = 2; /* format = 1; 11-bit codes */
576         s->unrestricted_mv = 1;
577         s->rtp_mode=0; /* don't allow GOB */
578         avctx->delay=0;
579         s->low_delay=1;
580         break;
581     case CODEC_ID_RV10:
582         s->out_format = FMT_H263;
583         avctx->delay=0;
584         s->low_delay=1;
585         break;
586     case CODEC_ID_RV20:
587         s->out_format = FMT_H263;
588         avctx->delay=0;
589         s->low_delay=1;
590         s->modified_quant=1;
591         s->h263_aic=1;
592         s->h263_plus=1;
593         s->loop_filter=1;
594         s->unrestricted_mv= s->obmc || s->loop_filter || s->umvplus;
595         break;
596     case CODEC_ID_MPEG4:
597         s->out_format = FMT_H263;
598         s->h263_pred = 1;
599         s->unrestricted_mv = 1;
600         s->low_delay= s->max_b_frames ? 0 : 1;
601         avctx->delay= s->low_delay ? 0 : (s->max_b_frames + 1);
602         break;
603     case CODEC_ID_MSMPEG4V1:
604         s->out_format = FMT_H263;
605         s->h263_msmpeg4 = 1;
606         s->h263_pred = 1;
607         s->unrestricted_mv = 1;
608         s->msmpeg4_version= 1;
609         avctx->delay=0;
610         s->low_delay=1;
611         break;
612     case CODEC_ID_MSMPEG4V2:
613         s->out_format = FMT_H263;
614         s->h263_msmpeg4 = 1;
615         s->h263_pred = 1;
616         s->unrestricted_mv = 1;
617         s->msmpeg4_version= 2;
618         avctx->delay=0;
619         s->low_delay=1;
620         break;
621     case CODEC_ID_MSMPEG4V3:
622         s->out_format = FMT_H263;
623         s->h263_msmpeg4 = 1;
624         s->h263_pred = 1;
625         s->unrestricted_mv = 1;
626         s->msmpeg4_version= 3;
627         s->flipflop_rounding=1;
628         avctx->delay=0;
629         s->low_delay=1;
630         break;
631     case CODEC_ID_WMV1:
632         s->out_format = FMT_H263;
633         s->h263_msmpeg4 = 1;
634         s->h263_pred = 1;
635         s->unrestricted_mv = 1;
636         s->msmpeg4_version= 4;
637         s->flipflop_rounding=1;
638         avctx->delay=0;
639         s->low_delay=1;
640         break;
641     case CODEC_ID_WMV2:
642         s->out_format = FMT_H263;
643         s->h263_msmpeg4 = 1;
644         s->h263_pred = 1;
645         s->unrestricted_mv = 1;
646         s->msmpeg4_version= 5;
647         s->flipflop_rounding=1;
648         avctx->delay=0;
649         s->low_delay=1;
650         break;
651     default:
652         return -1;
653     }
654
655     avctx->has_b_frames= !s->low_delay;
656
657     s->encoding = 1;
658
659     /* init */
660     if (MPV_common_init(s) < 0)
661         return -1;
662
663     if(s->modified_quant)
664         s->chroma_qscale_table= ff_h263_chroma_qscale_table;
665     s->progressive_frame=
666     s->progressive_sequence= !(avctx->flags & (CODEC_FLAG_INTERLACED_DCT|CODEC_FLAG_INTERLACED_ME|CODEC_FLAG_ALT_SCAN));
667     s->quant_precision=5;
668
669     ff_set_cmp(&s->dsp, s->dsp.ildct_cmp, s->avctx->ildct_cmp);
670     ff_set_cmp(&s->dsp, s->dsp.frame_skip_cmp, s->avctx->frame_skip_cmp);
671
672     if (ENABLE_H261_ENCODER && s->out_format == FMT_H261)
673         ff_h261_encode_init(s);
674     if (s->out_format == FMT_H263)
675         h263_encode_init(s);
676     if (ENABLE_MSMPEG4_ENCODER && s->msmpeg4_version)
677         ff_msmpeg4_encode_init(s);
678     if (s->out_format == FMT_MPEG1)
679         ff_mpeg1_encode_init(s);
680
681     /* init q matrix */
682     for(i=0;i<64;i++) {
683         int j= s->dsp.idct_permutation[i];
684         if(s->codec_id==CODEC_ID_MPEG4 && s->mpeg_quant){
685             s->intra_matrix[j] = ff_mpeg4_default_intra_matrix[i];
686             s->inter_matrix[j] = ff_mpeg4_default_non_intra_matrix[i];
687         }else if(s->out_format == FMT_H263 || s->out_format == FMT_H261){
688             s->intra_matrix[j] =
689             s->inter_matrix[j] = ff_mpeg1_default_non_intra_matrix[i];
690         }else
691         { /* mpeg1/2 */
692             s->intra_matrix[j] = ff_mpeg1_default_intra_matrix[i];
693             s->inter_matrix[j] = ff_mpeg1_default_non_intra_matrix[i];
694         }
695         if(s->avctx->intra_matrix)
696             s->intra_matrix[j] = s->avctx->intra_matrix[i];
697         if(s->avctx->inter_matrix)
698             s->inter_matrix[j] = s->avctx->inter_matrix[i];
699     }
700
701     /* precompute matrix */
702     /* for mjpeg, we do include qscale in the matrix */
703     if (s->out_format != FMT_MJPEG) {
704         convert_matrix(&s->dsp, s->q_intra_matrix, s->q_intra_matrix16,
705                        s->intra_matrix, s->intra_quant_bias, avctx->qmin, 31, 1);
706         convert_matrix(&s->dsp, s->q_inter_matrix, s->q_inter_matrix16,
707                        s->inter_matrix, s->inter_quant_bias, avctx->qmin, 31, 0);
708     }
709
710     if(ff_rate_control_init(s) < 0)
711         return -1;
712
713     return 0;
714 }
715
716 int MPV_encode_end(AVCodecContext *avctx)
717 {
718     MpegEncContext *s = avctx->priv_data;
719
720     ff_rate_control_uninit(s);
721
722     MPV_common_end(s);
723     if ((ENABLE_MJPEG_ENCODER || ENABLE_LJPEG_ENCODER) && s->out_format == FMT_MJPEG)
724         ff_mjpeg_encode_close(s);
725
726     av_freep(&avctx->extradata);
727
728     return 0;
729 }
730
731 static int get_sae(uint8_t *src, int ref, int stride){
732     int x,y;
733     int acc=0;
734
735     for(y=0; y<16; y++){
736         for(x=0; x<16; x++){
737             acc+= FFABS(src[x+y*stride] - ref);
738         }
739     }
740
741     return acc;
742 }
743
744 static int get_intra_count(MpegEncContext *s, uint8_t *src, uint8_t *ref, int stride){
745     int x, y, w, h;
746     int acc=0;
747
748     w= s->width &~15;
749     h= s->height&~15;
750
751     for(y=0; y<h; y+=16){
752         for(x=0; x<w; x+=16){
753             int offset= x + y*stride;
754             int sad = s->dsp.sad[0](NULL, src + offset, ref + offset, stride, 16);
755             int mean= (s->dsp.pix_sum(src + offset, stride) + 128)>>8;
756             int sae = get_sae(src + offset, mean, stride);
757
758             acc+= sae + 500 < sad;
759         }
760     }
761     return acc;
762 }
763
764
765 static int load_input_picture(MpegEncContext *s, AVFrame *pic_arg){
766     AVFrame *pic=NULL;
767     int64_t pts;
768     int i;
769     const int encoding_delay= s->max_b_frames;
770     int direct=1;
771
772     if(pic_arg){
773         pts= pic_arg->pts;
774         pic_arg->display_picture_number= s->input_picture_number++;
775
776         if(pts != AV_NOPTS_VALUE){
777             if(s->user_specified_pts != AV_NOPTS_VALUE){
778                 int64_t time= pts;
779                 int64_t last= s->user_specified_pts;
780
781                 if(time <= last){
782                     av_log(s->avctx, AV_LOG_ERROR, "Error, Invalid timestamp=%"PRId64", last=%"PRId64"\n", pts, s->user_specified_pts);
783                     return -1;
784                 }
785             }
786             s->user_specified_pts= pts;
787         }else{
788             if(s->user_specified_pts != AV_NOPTS_VALUE){
789                 s->user_specified_pts=
790                 pts= s->user_specified_pts + 1;
791                 av_log(s->avctx, AV_LOG_INFO, "Warning: AVFrame.pts=? trying to guess (%"PRId64")\n", pts);
792             }else{
793                 pts= pic_arg->display_picture_number;
794             }
795         }
796     }
797
798   if(pic_arg){
799     if(encoding_delay && !(s->flags&CODEC_FLAG_INPUT_PRESERVED)) direct=0;
800     if(pic_arg->linesize[0] != s->linesize) direct=0;
801     if(pic_arg->linesize[1] != s->uvlinesize) direct=0;
802     if(pic_arg->linesize[2] != s->uvlinesize) direct=0;
803
804 //    av_log(AV_LOG_DEBUG, "%d %d %d %d\n",pic_arg->linesize[0], pic_arg->linesize[1], s->linesize, s->uvlinesize);
805
806     if(direct){
807         i= ff_find_unused_picture(s, 1);
808
809         pic= (AVFrame*)&s->picture[i];
810         pic->reference= 3;
811
812         for(i=0; i<4; i++){
813             pic->data[i]= pic_arg->data[i];
814             pic->linesize[i]= pic_arg->linesize[i];
815         }
816         alloc_picture(s, (Picture*)pic, 1);
817     }else{
818         i= ff_find_unused_picture(s, 0);
819
820         pic= (AVFrame*)&s->picture[i];
821         pic->reference= 3;
822
823         alloc_picture(s, (Picture*)pic, 0);
824
825         if(   pic->data[0] + INPLACE_OFFSET == pic_arg->data[0]
826            && pic->data[1] + INPLACE_OFFSET == pic_arg->data[1]
827            && pic->data[2] + INPLACE_OFFSET == pic_arg->data[2]){
828        // empty
829         }else{
830             int h_chroma_shift, v_chroma_shift;
831             avcodec_get_chroma_sub_sample(s->avctx->pix_fmt, &h_chroma_shift, &v_chroma_shift);
832
833             for(i=0; i<3; i++){
834                 int src_stride= pic_arg->linesize[i];
835                 int dst_stride= i ? s->uvlinesize : s->linesize;
836                 int h_shift= i ? h_chroma_shift : 0;
837                 int v_shift= i ? v_chroma_shift : 0;
838                 int w= s->width >>h_shift;
839                 int h= s->height>>v_shift;
840                 uint8_t *src= pic_arg->data[i];
841                 uint8_t *dst= pic->data[i];
842
843                 if(!s->avctx->rc_buffer_size)
844                     dst +=INPLACE_OFFSET;
845
846                 if(src_stride==dst_stride)
847                     memcpy(dst, src, src_stride*h);
848                 else{
849                     while(h--){
850                         memcpy(dst, src, w);
851                         dst += dst_stride;
852                         src += src_stride;
853                     }
854                 }
855             }
856         }
857     }
858     copy_picture_attributes(s, pic, pic_arg);
859     pic->pts= pts; //we set this here to avoid modifiying pic_arg
860   }
861
862     /* shift buffer entries */
863     for(i=1; i<MAX_PICTURE_COUNT /*s->encoding_delay+1*/; i++)
864         s->input_picture[i-1]= s->input_picture[i];
865
866     s->input_picture[encoding_delay]= (Picture*)pic;
867
868     return 0;
869 }
870
871 static int skip_check(MpegEncContext *s, Picture *p, Picture *ref){
872     int x, y, plane;
873     int score=0;
874     int64_t score64=0;
875
876     for(plane=0; plane<3; plane++){
877         const int stride= p->linesize[plane];
878         const int bw= plane ? 1 : 2;
879         for(y=0; y<s->mb_height*bw; y++){
880             for(x=0; x<s->mb_width*bw; x++){
881                 int off= p->type == FF_BUFFER_TYPE_SHARED ? 0: 16;
882                 int v= s->dsp.frame_skip_cmp[1](s, p->data[plane] + 8*(x + y*stride)+off, ref->data[plane] + 8*(x + y*stride), stride, 8);
883
884                 switch(s->avctx->frame_skip_exp){
885                     case 0: score= FFMAX(score, v); break;
886                     case 1: score+= FFABS(v);break;
887                     case 2: score+= v*v;break;
888                     case 3: score64+= FFABS(v*v*(int64_t)v);break;
889                     case 4: score64+= v*v*(int64_t)(v*v);break;
890                 }
891             }
892         }
893     }
894
895     if(score) score64= score;
896
897     if(score64 < s->avctx->frame_skip_threshold)
898         return 1;
899     if(score64 < ((s->avctx->frame_skip_factor * (int64_t)s->lambda)>>8))
900         return 1;
901     return 0;
902 }
903
904 static int estimate_best_b_count(MpegEncContext *s){
905     AVCodec *codec= avcodec_find_encoder(s->avctx->codec_id);
906     AVCodecContext *c= avcodec_alloc_context();
907     AVFrame input[FF_MAX_B_FRAMES+2];
908     const int scale= s->avctx->brd_scale;
909     int i, j, out_size, p_lambda, b_lambda, lambda2;
910     int outbuf_size= s->width * s->height; //FIXME
911     uint8_t *outbuf= av_malloc(outbuf_size);
912     int64_t best_rd= INT64_MAX;
913     int best_b_count= -1;
914
915     assert(scale>=0 && scale <=3);
916
917 //    emms_c();
918     p_lambda= s->last_lambda_for[P_TYPE]; //s->next_picture_ptr->quality;
919     b_lambda= s->last_lambda_for[B_TYPE]; //p_lambda *FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset;
920     if(!b_lambda) b_lambda= p_lambda; //FIXME we should do this somewhere else
921     lambda2= (b_lambda*b_lambda + (1<<FF_LAMBDA_SHIFT)/2 ) >> FF_LAMBDA_SHIFT;
922
923     c->width = s->width >> scale;
924     c->height= s->height>> scale;
925     c->flags= CODEC_FLAG_QSCALE | CODEC_FLAG_PSNR | CODEC_FLAG_INPUT_PRESERVED /*| CODEC_FLAG_EMU_EDGE*/;
926     c->flags|= s->avctx->flags & CODEC_FLAG_QPEL;
927     c->mb_decision= s->avctx->mb_decision;
928     c->me_cmp= s->avctx->me_cmp;
929     c->mb_cmp= s->avctx->mb_cmp;
930     c->me_sub_cmp= s->avctx->me_sub_cmp;
931     c->pix_fmt = PIX_FMT_YUV420P;
932     c->time_base= s->avctx->time_base;
933     c->max_b_frames= s->max_b_frames;
934
935     if (avcodec_open(c, codec) < 0)
936         return -1;
937
938     for(i=0; i<s->max_b_frames+2; i++){
939         int ysize= c->width*c->height;
940         int csize= (c->width/2)*(c->height/2);
941         Picture pre_input, *pre_input_ptr= i ? s->input_picture[i-1] : s->next_picture_ptr;
942
943         avcodec_get_frame_defaults(&input[i]);
944         input[i].data[0]= av_malloc(ysize + 2*csize);
945         input[i].data[1]= input[i].data[0] + ysize;
946         input[i].data[2]= input[i].data[1] + csize;
947         input[i].linesize[0]= c->width;
948         input[i].linesize[1]=
949         input[i].linesize[2]= c->width/2;
950
951         if(pre_input_ptr && (!i || s->input_picture[i-1])) {
952             pre_input= *pre_input_ptr;
953
954             if(pre_input.type != FF_BUFFER_TYPE_SHARED && i) {
955                 pre_input.data[0]+=INPLACE_OFFSET;
956                 pre_input.data[1]+=INPLACE_OFFSET;
957                 pre_input.data[2]+=INPLACE_OFFSET;
958             }
959
960             s->dsp.shrink[scale](input[i].data[0], input[i].linesize[0], pre_input.data[0], pre_input.linesize[0], c->width, c->height);
961             s->dsp.shrink[scale](input[i].data[1], input[i].linesize[1], pre_input.data[1], pre_input.linesize[1], c->width>>1, c->height>>1);
962             s->dsp.shrink[scale](input[i].data[2], input[i].linesize[2], pre_input.data[2], pre_input.linesize[2], c->width>>1, c->height>>1);
963         }
964     }
965
966     for(j=0; j<s->max_b_frames+1; j++){
967         int64_t rd=0;
968
969         if(!s->input_picture[j])
970             break;
971
972         c->error[0]= c->error[1]= c->error[2]= 0;
973
974         input[0].pict_type= I_TYPE;
975         input[0].quality= 1 * FF_QP2LAMBDA;
976         out_size = avcodec_encode_video(c, outbuf, outbuf_size, &input[0]);
977 //        rd += (out_size * lambda2) >> FF_LAMBDA_SHIFT;
978
979         for(i=0; i<s->max_b_frames+1; i++){
980             int is_p= i % (j+1) == j || i==s->max_b_frames;
981
982             input[i+1].pict_type= is_p ? P_TYPE : B_TYPE;
983             input[i+1].quality= is_p ? p_lambda : b_lambda;
984             out_size = avcodec_encode_video(c, outbuf, outbuf_size, &input[i+1]);
985             rd += (out_size * lambda2) >> (FF_LAMBDA_SHIFT - 3);
986         }
987
988         /* get the delayed frames */
989         while(out_size){
990             out_size = avcodec_encode_video(c, outbuf, outbuf_size, NULL);
991             rd += (out_size * lambda2) >> (FF_LAMBDA_SHIFT - 3);
992         }
993
994         rd += c->error[0] + c->error[1] + c->error[2];
995
996         if(rd < best_rd){
997             best_rd= rd;
998             best_b_count= j;
999         }
1000     }
1001
1002     av_freep(&outbuf);
1003     avcodec_close(c);
1004     av_freep(&c);
1005
1006     for(i=0; i<s->max_b_frames+2; i++){
1007         av_freep(&input[i].data[0]);
1008     }
1009
1010     return best_b_count;
1011 }
1012
1013 static void select_input_picture(MpegEncContext *s){
1014     int i;
1015
1016     for(i=1; i<MAX_PICTURE_COUNT; i++)
1017         s->reordered_input_picture[i-1]= s->reordered_input_picture[i];
1018     s->reordered_input_picture[MAX_PICTURE_COUNT-1]= NULL;
1019
1020     /* set next picture type & ordering */
1021     if(s->reordered_input_picture[0]==NULL && s->input_picture[0]){
1022         if(/*s->picture_in_gop_number >= s->gop_size ||*/ s->next_picture_ptr==NULL || s->intra_only){
1023             s->reordered_input_picture[0]= s->input_picture[0];
1024             s->reordered_input_picture[0]->pict_type= I_TYPE;
1025             s->reordered_input_picture[0]->coded_picture_number= s->coded_picture_number++;
1026         }else{
1027             int b_frames;
1028
1029             if(s->avctx->frame_skip_threshold || s->avctx->frame_skip_factor){
1030                 if(s->picture_in_gop_number < s->gop_size && skip_check(s, s->input_picture[0], s->next_picture_ptr)){
1031                 //FIXME check that te gop check above is +-1 correct
1032 //av_log(NULL, AV_LOG_DEBUG, "skip %p %"PRId64"\n", s->input_picture[0]->data[0], s->input_picture[0]->pts);
1033
1034                     if(s->input_picture[0]->type == FF_BUFFER_TYPE_SHARED){
1035                         for(i=0; i<4; i++)
1036                             s->input_picture[0]->data[i]= NULL;
1037                         s->input_picture[0]->type= 0;
1038                     }else{
1039                         assert(   s->input_picture[0]->type==FF_BUFFER_TYPE_USER
1040                                || s->input_picture[0]->type==FF_BUFFER_TYPE_INTERNAL);
1041
1042                         s->avctx->release_buffer(s->avctx, (AVFrame*)s->input_picture[0]);
1043                     }
1044
1045                     emms_c();
1046                     ff_vbv_update(s, 0);
1047
1048                     goto no_output_pic;
1049                 }
1050             }
1051
1052             if(s->flags&CODEC_FLAG_PASS2){
1053                 for(i=0; i<s->max_b_frames+1; i++){
1054                     int pict_num= s->input_picture[0]->display_picture_number + i;
1055
1056                     if(pict_num >= s->rc_context.num_entries)
1057                         break;
1058                     if(!s->input_picture[i]){
1059                         s->rc_context.entry[pict_num-1].new_pict_type = P_TYPE;
1060                         break;
1061                     }
1062
1063                     s->input_picture[i]->pict_type=
1064                         s->rc_context.entry[pict_num].new_pict_type;
1065                 }
1066             }
1067
1068             if(s->avctx->b_frame_strategy==0){
1069                 b_frames= s->max_b_frames;
1070                 while(b_frames && !s->input_picture[b_frames]) b_frames--;
1071             }else if(s->avctx->b_frame_strategy==1){
1072                 for(i=1; i<s->max_b_frames+1; i++){
1073                     if(s->input_picture[i] && s->input_picture[i]->b_frame_score==0){
1074                         s->input_picture[i]->b_frame_score=
1075                             get_intra_count(s, s->input_picture[i  ]->data[0],
1076                                                s->input_picture[i-1]->data[0], s->linesize) + 1;
1077                     }
1078                 }
1079                 for(i=0; i<s->max_b_frames+1; i++){
1080                     if(s->input_picture[i]==NULL || s->input_picture[i]->b_frame_score - 1 > s->mb_num/s->avctx->b_sensitivity) break;
1081                 }
1082
1083                 b_frames= FFMAX(0, i-1);
1084
1085                 /* reset scores */
1086                 for(i=0; i<b_frames+1; i++){
1087                     s->input_picture[i]->b_frame_score=0;
1088                 }
1089             }else if(s->avctx->b_frame_strategy==2){
1090                 b_frames= estimate_best_b_count(s);
1091             }else{
1092                 av_log(s->avctx, AV_LOG_ERROR, "illegal b frame strategy\n");
1093                 b_frames=0;
1094             }
1095
1096             emms_c();
1097 //static int b_count=0;
1098 //b_count+= b_frames;
1099 //av_log(s->avctx, AV_LOG_DEBUG, "b_frames: %d\n", b_count);
1100
1101             for(i= b_frames - 1; i>=0; i--){
1102                 int type= s->input_picture[i]->pict_type;
1103                 if(type && type != B_TYPE)
1104                     b_frames= i;
1105             }
1106             if(s->input_picture[b_frames]->pict_type == B_TYPE && b_frames == s->max_b_frames){
1107                 av_log(s->avctx, AV_LOG_ERROR, "warning, too many b frames in a row\n");
1108             }
1109
1110             if(s->picture_in_gop_number + b_frames >= s->gop_size){
1111               if((s->flags2 & CODEC_FLAG2_STRICT_GOP) && s->gop_size > s->picture_in_gop_number){
1112                     b_frames= s->gop_size - s->picture_in_gop_number - 1;
1113               }else{
1114                 if(s->flags & CODEC_FLAG_CLOSED_GOP)
1115                     b_frames=0;
1116                 s->input_picture[b_frames]->pict_type= I_TYPE;
1117               }
1118             }
1119
1120             if(   (s->flags & CODEC_FLAG_CLOSED_GOP)
1121                && b_frames
1122                && s->input_picture[b_frames]->pict_type== I_TYPE)
1123                 b_frames--;
1124
1125             s->reordered_input_picture[0]= s->input_picture[b_frames];
1126             if(s->reordered_input_picture[0]->pict_type != I_TYPE)
1127                 s->reordered_input_picture[0]->pict_type= P_TYPE;
1128             s->reordered_input_picture[0]->coded_picture_number= s->coded_picture_number++;
1129             for(i=0; i<b_frames; i++){
1130                 s->reordered_input_picture[i+1]= s->input_picture[i];
1131                 s->reordered_input_picture[i+1]->pict_type= B_TYPE;
1132                 s->reordered_input_picture[i+1]->coded_picture_number= s->coded_picture_number++;
1133             }
1134         }
1135     }
1136 no_output_pic:
1137     if(s->reordered_input_picture[0]){
1138         s->reordered_input_picture[0]->reference= s->reordered_input_picture[0]->pict_type!=B_TYPE ? 3 : 0;
1139
1140         copy_picture(&s->new_picture, s->reordered_input_picture[0]);
1141
1142         if(s->reordered_input_picture[0]->type == FF_BUFFER_TYPE_SHARED || s->avctx->rc_buffer_size){
1143             // input is a shared pix, so we can't modifiy it -> alloc a new one & ensure that the shared one is reuseable
1144
1145             int i= ff_find_unused_picture(s, 0);
1146             Picture *pic= &s->picture[i];
1147
1148             pic->reference              = s->reordered_input_picture[0]->reference;
1149             alloc_picture(s, pic, 0);
1150
1151             /* mark us unused / free shared pic */
1152             if(s->reordered_input_picture[0]->type == FF_BUFFER_TYPE_INTERNAL)
1153                 s->avctx->release_buffer(s->avctx, (AVFrame*)s->reordered_input_picture[0]);
1154             for(i=0; i<4; i++)
1155                 s->reordered_input_picture[0]->data[i]= NULL;
1156             s->reordered_input_picture[0]->type= 0;
1157
1158             copy_picture_attributes(s, (AVFrame*)pic, (AVFrame*)s->reordered_input_picture[0]);
1159
1160             s->current_picture_ptr= pic;
1161         }else{
1162             // input is not a shared pix -> reuse buffer for current_pix
1163
1164             assert(   s->reordered_input_picture[0]->type==FF_BUFFER_TYPE_USER
1165                    || s->reordered_input_picture[0]->type==FF_BUFFER_TYPE_INTERNAL);
1166
1167             s->current_picture_ptr= s->reordered_input_picture[0];
1168             for(i=0; i<4; i++){
1169                 s->new_picture.data[i]+= INPLACE_OFFSET;
1170             }
1171         }
1172         copy_picture(&s->current_picture, s->current_picture_ptr);
1173
1174         s->picture_number= s->new_picture.display_picture_number;
1175 //printf("dpn:%d\n", s->picture_number);
1176     }else{
1177        memset(&s->new_picture, 0, sizeof(Picture));
1178     }
1179 }
1180
1181 int MPV_encode_picture(AVCodecContext *avctx,
1182                        unsigned char *buf, int buf_size, void *data)
1183 {
1184     MpegEncContext *s = avctx->priv_data;
1185     AVFrame *pic_arg = data;
1186     int i, stuffing_count;
1187
1188     for(i=0; i<avctx->thread_count; i++){
1189         int start_y= s->thread_context[i]->start_mb_y;
1190         int   end_y= s->thread_context[i]->  end_mb_y;
1191         int h= s->mb_height;
1192         uint8_t *start= buf + (size_t)(((int64_t) buf_size)*start_y/h);
1193         uint8_t *end  = buf + (size_t)(((int64_t) buf_size)*  end_y/h);
1194
1195         init_put_bits(&s->thread_context[i]->pb, start, end - start);
1196     }
1197
1198     s->picture_in_gop_number++;
1199
1200     if(load_input_picture(s, pic_arg) < 0)
1201         return -1;
1202
1203     select_input_picture(s);
1204
1205     /* output? */
1206     if(s->new_picture.data[0]){
1207         s->pict_type= s->new_picture.pict_type;
1208 //emms_c();
1209 //printf("qs:%f %f %d\n", s->new_picture.quality, s->current_picture.quality, s->qscale);
1210         MPV_frame_start(s, avctx);
1211 vbv_retry:
1212         if (encode_picture(s, s->picture_number) < 0)
1213             return -1;
1214
1215         avctx->real_pict_num  = s->picture_number;
1216         avctx->header_bits = s->header_bits;
1217         avctx->mv_bits     = s->mv_bits;
1218         avctx->misc_bits   = s->misc_bits;
1219         avctx->i_tex_bits  = s->i_tex_bits;
1220         avctx->p_tex_bits  = s->p_tex_bits;
1221         avctx->i_count     = s->i_count;
1222         avctx->p_count     = s->mb_num - s->i_count - s->skip_count; //FIXME f/b_count in avctx
1223         avctx->skip_count  = s->skip_count;
1224
1225         MPV_frame_end(s);
1226
1227         if (ENABLE_MJPEG_ENCODER && s->out_format == FMT_MJPEG)
1228             ff_mjpeg_encode_picture_trailer(s);
1229
1230         if(avctx->rc_buffer_size){
1231             RateControlContext *rcc= &s->rc_context;
1232             int max_size= rcc->buffer_index/3;
1233
1234             if(put_bits_count(&s->pb) > max_size && s->lambda < s->avctx->lmax){
1235                 s->next_lambda= FFMAX(s->lambda+1, s->lambda*(s->qscale+1) / s->qscale);
1236                 if(s->adaptive_quant){
1237                     int i;
1238                     for(i=0; i<s->mb_height*s->mb_stride; i++)
1239                         s->lambda_table[i]= FFMAX(s->lambda_table[i]+1, s->lambda_table[i]*(s->qscale+1) / s->qscale);
1240                 }
1241                 s->mb_skipped = 0;        //done in MPV_frame_start()
1242                 if(s->pict_type==P_TYPE){ //done in encode_picture() so we must undo it
1243                     if(s->flipflop_rounding || s->codec_id == CODEC_ID_H263P || s->codec_id == CODEC_ID_MPEG4)
1244                         s->no_rounding ^= 1;
1245                 }
1246                 if(s->pict_type!=B_TYPE){
1247                     s->time_base= s->last_time_base;
1248                     s->last_non_b_time= s->time - s->pp_time;
1249                 }
1250 //                av_log(NULL, AV_LOG_ERROR, "R:%d ", s->next_lambda);
1251                 for(i=0; i<avctx->thread_count; i++){
1252                     PutBitContext *pb= &s->thread_context[i]->pb;
1253                     init_put_bits(pb, pb->buf, pb->buf_end - pb->buf);
1254                 }
1255                 goto vbv_retry;
1256             }
1257
1258             assert(s->avctx->rc_max_rate);
1259         }
1260
1261         if(s->flags&CODEC_FLAG_PASS1)
1262             ff_write_pass1_stats(s);
1263
1264         for(i=0; i<4; i++){
1265             s->current_picture_ptr->error[i]= s->current_picture.error[i];
1266             avctx->error[i] += s->current_picture_ptr->error[i];
1267         }
1268
1269         if(s->flags&CODEC_FLAG_PASS1)
1270             assert(avctx->header_bits + avctx->mv_bits + avctx->misc_bits + avctx->i_tex_bits + avctx->p_tex_bits == put_bits_count(&s->pb));
1271         flush_put_bits(&s->pb);
1272         s->frame_bits  = put_bits_count(&s->pb);
1273
1274         stuffing_count= ff_vbv_update(s, s->frame_bits);
1275         if(stuffing_count){
1276             if(s->pb.buf_end - s->pb.buf - (put_bits_count(&s->pb)>>3) < stuffing_count + 50){
1277                 av_log(s->avctx, AV_LOG_ERROR, "stuffing too large\n");
1278                 return -1;
1279             }
1280
1281             switch(s->codec_id){
1282             case CODEC_ID_MPEG1VIDEO:
1283             case CODEC_ID_MPEG2VIDEO:
1284                 while(stuffing_count--){
1285                     put_bits(&s->pb, 8, 0);
1286                 }
1287             break;
1288             case CODEC_ID_MPEG4:
1289                 put_bits(&s->pb, 16, 0);
1290                 put_bits(&s->pb, 16, 0x1C3);
1291                 stuffing_count -= 4;
1292                 while(stuffing_count--){
1293                     put_bits(&s->pb, 8, 0xFF);
1294                 }
1295             break;
1296             default:
1297                 av_log(s->avctx, AV_LOG_ERROR, "vbv buffer overflow\n");
1298             }
1299             flush_put_bits(&s->pb);
1300             s->frame_bits  = put_bits_count(&s->pb);
1301         }
1302
1303         /* update mpeg1/2 vbv_delay for CBR */
1304         if(s->avctx->rc_max_rate && s->avctx->rc_min_rate == s->avctx->rc_max_rate && s->out_format == FMT_MPEG1
1305            && 90000LL * (avctx->rc_buffer_size-1) <= s->avctx->rc_max_rate*0xFFFFLL){
1306             int vbv_delay;
1307
1308             assert(s->repeat_first_field==0);
1309
1310             vbv_delay= lrintf(90000 * s->rc_context.buffer_index / s->avctx->rc_max_rate);
1311             assert(vbv_delay < 0xFFFF);
1312
1313             s->vbv_delay_ptr[0] &= 0xF8;
1314             s->vbv_delay_ptr[0] |= vbv_delay>>13;
1315             s->vbv_delay_ptr[1]  = vbv_delay>>5;
1316             s->vbv_delay_ptr[2] &= 0x07;
1317             s->vbv_delay_ptr[2] |= vbv_delay<<3;
1318         }
1319         s->total_bits += s->frame_bits;
1320         avctx->frame_bits  = s->frame_bits;
1321     }else{
1322         assert((pbBufPtr(&s->pb) == s->pb.buf));
1323         s->frame_bits=0;
1324     }
1325     assert((s->frame_bits&7)==0);
1326
1327     return s->frame_bits/8;
1328 }
1329
1330 static inline void dct_single_coeff_elimination(MpegEncContext *s, int n, int threshold)
1331 {
1332     static const char tab[64]=
1333         {3,2,2,1,1,1,1,1,
1334          1,1,1,1,1,1,1,1,
1335          1,1,1,1,1,1,1,1,
1336          0,0,0,0,0,0,0,0,
1337          0,0,0,0,0,0,0,0,
1338          0,0,0,0,0,0,0,0,
1339          0,0,0,0,0,0,0,0,
1340          0,0,0,0,0,0,0,0};
1341     int score=0;
1342     int run=0;
1343     int i;
1344     DCTELEM *block= s->block[n];
1345     const int last_index= s->block_last_index[n];
1346     int skip_dc;
1347
1348     if(threshold<0){
1349         skip_dc=0;
1350         threshold= -threshold;
1351     }else
1352         skip_dc=1;
1353
1354     /* are all which we could set to zero are allready zero? */
1355     if(last_index<=skip_dc - 1) return;
1356
1357     for(i=0; i<=last_index; i++){
1358         const int j = s->intra_scantable.permutated[i];
1359         const int level = FFABS(block[j]);
1360         if(level==1){
1361             if(skip_dc && i==0) continue;
1362             score+= tab[run];
1363             run=0;
1364         }else if(level>1){
1365             return;
1366         }else{
1367             run++;
1368         }
1369     }
1370     if(score >= threshold) return;
1371     for(i=skip_dc; i<=last_index; i++){
1372         const int j = s->intra_scantable.permutated[i];
1373         block[j]=0;
1374     }
1375     if(block[0]) s->block_last_index[n]= 0;
1376     else         s->block_last_index[n]= -1;
1377 }
1378
1379 static inline void clip_coeffs(MpegEncContext *s, DCTELEM *block, int last_index)
1380 {
1381     int i;
1382     const int maxlevel= s->max_qcoeff;
1383     const int minlevel= s->min_qcoeff;
1384     int overflow=0;
1385
1386     if(s->mb_intra){
1387         i=1; //skip clipping of intra dc
1388     }else
1389         i=0;
1390
1391     for(;i<=last_index; i++){
1392         const int j= s->intra_scantable.permutated[i];
1393         int level = block[j];
1394
1395         if     (level>maxlevel){
1396             level=maxlevel;
1397             overflow++;
1398         }else if(level<minlevel){
1399             level=minlevel;
1400             overflow++;
1401         }
1402
1403         block[j]= level;
1404     }
1405
1406     if(overflow && s->avctx->mb_decision == FF_MB_DECISION_SIMPLE)
1407         av_log(s->avctx, AV_LOG_INFO, "warning, clipping %d dct coefficients to %d..%d\n", overflow, minlevel, maxlevel);
1408 }
1409
1410 static void get_vissual_weight(int16_t *weight, uint8_t *ptr, int stride){
1411     int x, y;
1412 //FIXME optimize
1413     for(y=0; y<8; y++){
1414         for(x=0; x<8; x++){
1415             int x2, y2;
1416             int sum=0;
1417             int sqr=0;
1418             int count=0;
1419
1420             for(y2= FFMAX(y-1, 0); y2 < FFMIN(8, y+2); y2++){
1421                 for(x2= FFMAX(x-1, 0); x2 < FFMIN(8, x+2); x2++){
1422                     int v= ptr[x2 + y2*stride];
1423                     sum += v;
1424                     sqr += v*v;
1425                     count++;
1426                 }
1427             }
1428             weight[x + 8*y]= (36*ff_sqrt(count*sqr - sum*sum)) / count;
1429         }
1430     }
1431 }
1432
1433 static av_always_inline void encode_mb_internal(MpegEncContext *s, int motion_x, int motion_y, int mb_block_height, int mb_block_count)
1434 {
1435     int16_t weight[8][64];
1436     DCTELEM orig[8][64];
1437     const int mb_x= s->mb_x;
1438     const int mb_y= s->mb_y;
1439     int i;
1440     int skip_dct[8];
1441     int dct_offset   = s->linesize*8; //default for progressive frames
1442     uint8_t *ptr_y, *ptr_cb, *ptr_cr;
1443     int wrap_y, wrap_c;
1444
1445     for(i=0; i<mb_block_count; i++) skip_dct[i]=s->skipdct;
1446
1447     if(s->adaptive_quant){
1448         const int last_qp= s->qscale;
1449         const int mb_xy= mb_x + mb_y*s->mb_stride;
1450
1451         s->lambda= s->lambda_table[mb_xy];
1452         update_qscale(s);
1453
1454         if(!(s->flags&CODEC_FLAG_QP_RD)){
1455             s->qscale= s->current_picture_ptr->qscale_table[mb_xy];
1456             s->dquant= s->qscale - last_qp;
1457
1458             if(s->out_format==FMT_H263){
1459                 s->dquant= av_clip(s->dquant, -2, 2);
1460
1461                 if(s->codec_id==CODEC_ID_MPEG4){
1462                     if(!s->mb_intra){
1463                         if(s->pict_type == B_TYPE){
1464                             if(s->dquant&1 || s->mv_dir&MV_DIRECT)
1465                                 s->dquant= 0;
1466                         }
1467                         if(s->mv_type==MV_TYPE_8X8)
1468                             s->dquant=0;
1469                     }
1470                 }
1471             }
1472         }
1473         ff_set_qscale(s, last_qp + s->dquant);
1474     }else if(s->flags&CODEC_FLAG_QP_RD)
1475         ff_set_qscale(s, s->qscale + s->dquant);
1476
1477     wrap_y = s->linesize;
1478     wrap_c = s->uvlinesize;
1479     ptr_y = s->new_picture.data[0] + (mb_y * 16 * wrap_y) + mb_x * 16;
1480     ptr_cb = s->new_picture.data[1] + (mb_y * mb_block_height * wrap_c) + mb_x * 8;
1481     ptr_cr = s->new_picture.data[2] + (mb_y * mb_block_height * wrap_c) + mb_x * 8;
1482
1483     if(mb_x*16+16 > s->width || mb_y*16+16 > s->height){
1484         uint8_t *ebuf= s->edge_emu_buffer + 32;
1485         ff_emulated_edge_mc(ebuf            , ptr_y , wrap_y,16,16,mb_x*16,mb_y*16, s->width   , s->height);
1486         ptr_y= ebuf;
1487         ff_emulated_edge_mc(ebuf+18*wrap_y  , ptr_cb, wrap_c, 8, mb_block_height, mb_x*8, mb_y*8, s->width>>1, s->height>>1);
1488         ptr_cb= ebuf+18*wrap_y;
1489         ff_emulated_edge_mc(ebuf+18*wrap_y+8, ptr_cr, wrap_c, 8, mb_block_height, mb_x*8, mb_y*8, s->width>>1, s->height>>1);
1490         ptr_cr= ebuf+18*wrap_y+8;
1491     }
1492
1493     if (s->mb_intra) {
1494         if(s->flags&CODEC_FLAG_INTERLACED_DCT){
1495             int progressive_score, interlaced_score;
1496
1497             s->interlaced_dct=0;
1498             progressive_score= s->dsp.ildct_cmp[4](s, ptr_y           , NULL, wrap_y, 8)
1499                               +s->dsp.ildct_cmp[4](s, ptr_y + wrap_y*8, NULL, wrap_y, 8) - 400;
1500
1501             if(progressive_score > 0){
1502                 interlaced_score = s->dsp.ildct_cmp[4](s, ptr_y           , NULL, wrap_y*2, 8)
1503                                   +s->dsp.ildct_cmp[4](s, ptr_y + wrap_y  , NULL, wrap_y*2, 8);
1504                 if(progressive_score > interlaced_score){
1505                     s->interlaced_dct=1;
1506
1507                     dct_offset= wrap_y;
1508                     wrap_y<<=1;
1509                     if (s->chroma_format == CHROMA_422)
1510                         wrap_c<<=1;
1511                 }
1512             }
1513         }
1514
1515         s->dsp.get_pixels(s->block[0], ptr_y                 , wrap_y);
1516         s->dsp.get_pixels(s->block[1], ptr_y              + 8, wrap_y);
1517         s->dsp.get_pixels(s->block[2], ptr_y + dct_offset    , wrap_y);
1518         s->dsp.get_pixels(s->block[3], ptr_y + dct_offset + 8, wrap_y);
1519
1520         if(s->flags&CODEC_FLAG_GRAY){
1521             skip_dct[4]= 1;
1522             skip_dct[5]= 1;
1523         }else{
1524             s->dsp.get_pixels(s->block[4], ptr_cb, wrap_c);
1525             s->dsp.get_pixels(s->block[5], ptr_cr, wrap_c);
1526             if(!s->chroma_y_shift){ /* 422 */
1527                 s->dsp.get_pixels(s->block[6], ptr_cb + (dct_offset>>1), wrap_c);
1528                 s->dsp.get_pixels(s->block[7], ptr_cr + (dct_offset>>1), wrap_c);
1529             }
1530         }
1531     }else{
1532         op_pixels_func (*op_pix)[4];
1533         qpel_mc_func (*op_qpix)[16];
1534         uint8_t *dest_y, *dest_cb, *dest_cr;
1535
1536         dest_y  = s->dest[0];
1537         dest_cb = s->dest[1];
1538         dest_cr = s->dest[2];
1539
1540         if ((!s->no_rounding) || s->pict_type==B_TYPE){
1541             op_pix = s->dsp.put_pixels_tab;
1542             op_qpix= s->dsp.put_qpel_pixels_tab;
1543         }else{
1544             op_pix = s->dsp.put_no_rnd_pixels_tab;
1545             op_qpix= s->dsp.put_no_rnd_qpel_pixels_tab;
1546         }
1547
1548         if (s->mv_dir & MV_DIR_FORWARD) {
1549             MPV_motion(s, dest_y, dest_cb, dest_cr, 0, s->last_picture.data, op_pix, op_qpix);
1550             op_pix = s->dsp.avg_pixels_tab;
1551             op_qpix= s->dsp.avg_qpel_pixels_tab;
1552         }
1553         if (s->mv_dir & MV_DIR_BACKWARD) {
1554             MPV_motion(s, dest_y, dest_cb, dest_cr, 1, s->next_picture.data, op_pix, op_qpix);
1555         }
1556
1557         if(s->flags&CODEC_FLAG_INTERLACED_DCT){
1558             int progressive_score, interlaced_score;
1559
1560             s->interlaced_dct=0;
1561             progressive_score= s->dsp.ildct_cmp[0](s, dest_y           , ptr_y           , wrap_y, 8)
1562                               +s->dsp.ildct_cmp[0](s, dest_y + wrap_y*8, ptr_y + wrap_y*8, wrap_y, 8) - 400;
1563
1564             if(s->avctx->ildct_cmp == FF_CMP_VSSE) progressive_score -= 400;
1565
1566             if(progressive_score>0){
1567                 interlaced_score = s->dsp.ildct_cmp[0](s, dest_y           , ptr_y           , wrap_y*2, 8)
1568                                   +s->dsp.ildct_cmp[0](s, dest_y + wrap_y  , ptr_y + wrap_y  , wrap_y*2, 8);
1569
1570                 if(progressive_score > interlaced_score){
1571                     s->interlaced_dct=1;
1572
1573                     dct_offset= wrap_y;
1574                     wrap_y<<=1;
1575                     if (s->chroma_format == CHROMA_422)
1576                         wrap_c<<=1;
1577                 }
1578             }
1579         }
1580
1581         s->dsp.diff_pixels(s->block[0], ptr_y                 , dest_y                 , wrap_y);
1582         s->dsp.diff_pixels(s->block[1], ptr_y              + 8, dest_y              + 8, wrap_y);
1583         s->dsp.diff_pixels(s->block[2], ptr_y + dct_offset    , dest_y + dct_offset    , wrap_y);
1584         s->dsp.diff_pixels(s->block[3], ptr_y + dct_offset + 8, dest_y + dct_offset + 8, wrap_y);
1585
1586         if(s->flags&CODEC_FLAG_GRAY){
1587             skip_dct[4]= 1;
1588             skip_dct[5]= 1;
1589         }else{
1590             s->dsp.diff_pixels(s->block[4], ptr_cb, dest_cb, wrap_c);
1591             s->dsp.diff_pixels(s->block[5], ptr_cr, dest_cr, wrap_c);
1592             if(!s->chroma_y_shift){ /* 422 */
1593                 s->dsp.diff_pixels(s->block[6], ptr_cb + (dct_offset>>1), dest_cb + (dct_offset>>1), wrap_c);
1594                 s->dsp.diff_pixels(s->block[7], ptr_cr + (dct_offset>>1), dest_cr + (dct_offset>>1), wrap_c);
1595             }
1596         }
1597         /* pre quantization */
1598         if(s->current_picture.mc_mb_var[s->mb_stride*mb_y+ mb_x]<2*s->qscale*s->qscale){
1599             //FIXME optimize
1600             if(s->dsp.sad[1](NULL, ptr_y               , dest_y               , wrap_y, 8) < 20*s->qscale) skip_dct[0]= 1;
1601             if(s->dsp.sad[1](NULL, ptr_y            + 8, dest_y            + 8, wrap_y, 8) < 20*s->qscale) skip_dct[1]= 1;
1602             if(s->dsp.sad[1](NULL, ptr_y +dct_offset   , dest_y +dct_offset   , wrap_y, 8) < 20*s->qscale) skip_dct[2]= 1;
1603             if(s->dsp.sad[1](NULL, ptr_y +dct_offset+ 8, dest_y +dct_offset+ 8, wrap_y, 8) < 20*s->qscale) skip_dct[3]= 1;
1604             if(s->dsp.sad[1](NULL, ptr_cb              , dest_cb              , wrap_c, 8) < 20*s->qscale) skip_dct[4]= 1;
1605             if(s->dsp.sad[1](NULL, ptr_cr              , dest_cr              , wrap_c, 8) < 20*s->qscale) skip_dct[5]= 1;
1606             if(!s->chroma_y_shift){ /* 422 */
1607                 if(s->dsp.sad[1](NULL, ptr_cb +(dct_offset>>1), dest_cb +(dct_offset>>1), wrap_c, 8) < 20*s->qscale) skip_dct[6]= 1;
1608                 if(s->dsp.sad[1](NULL, ptr_cr +(dct_offset>>1), dest_cr +(dct_offset>>1), wrap_c, 8) < 20*s->qscale) skip_dct[7]= 1;
1609             }
1610         }
1611     }
1612
1613     if(s->avctx->quantizer_noise_shaping){
1614         if(!skip_dct[0]) get_vissual_weight(weight[0], ptr_y                 , wrap_y);
1615         if(!skip_dct[1]) get_vissual_weight(weight[1], ptr_y              + 8, wrap_y);
1616         if(!skip_dct[2]) get_vissual_weight(weight[2], ptr_y + dct_offset    , wrap_y);
1617         if(!skip_dct[3]) get_vissual_weight(weight[3], ptr_y + dct_offset + 8, wrap_y);
1618         if(!skip_dct[4]) get_vissual_weight(weight[4], ptr_cb                , wrap_c);
1619         if(!skip_dct[5]) get_vissual_weight(weight[5], ptr_cr                , wrap_c);
1620         if(!s->chroma_y_shift){ /* 422 */
1621             if(!skip_dct[6]) get_vissual_weight(weight[6], ptr_cb + (dct_offset>>1), wrap_c);
1622             if(!skip_dct[7]) get_vissual_weight(weight[7], ptr_cr + (dct_offset>>1), wrap_c);
1623         }
1624         memcpy(orig[0], s->block[0], sizeof(DCTELEM)*64*mb_block_count);
1625     }
1626
1627     /* DCT & quantize */
1628     assert(s->out_format!=FMT_MJPEG || s->qscale==8);
1629     {
1630         for(i=0;i<mb_block_count;i++) {
1631             if(!skip_dct[i]){
1632                 int overflow;
1633                 s->block_last_index[i] = s->dct_quantize(s, s->block[i], i, s->qscale, &overflow);
1634             // FIXME we could decide to change to quantizer instead of clipping
1635             // JS: I don't think that would be a good idea it could lower quality instead
1636             //     of improve it. Just INTRADC clipping deserves changes in quantizer
1637                 if (overflow) clip_coeffs(s, s->block[i], s->block_last_index[i]);
1638             }else
1639                 s->block_last_index[i]= -1;
1640         }
1641         if(s->avctx->quantizer_noise_shaping){
1642             for(i=0;i<mb_block_count;i++) {
1643                 if(!skip_dct[i]){
1644                     s->block_last_index[i] = dct_quantize_refine(s, s->block[i], weight[i], orig[i], i, s->qscale);
1645                 }
1646             }
1647         }
1648
1649         if(s->luma_elim_threshold && !s->mb_intra)
1650             for(i=0; i<4; i++)
1651                 dct_single_coeff_elimination(s, i, s->luma_elim_threshold);
1652         if(s->chroma_elim_threshold && !s->mb_intra)
1653             for(i=4; i<mb_block_count; i++)
1654                 dct_single_coeff_elimination(s, i, s->chroma_elim_threshold);
1655
1656         if(s->flags & CODEC_FLAG_CBP_RD){
1657             for(i=0;i<mb_block_count;i++) {
1658                 if(s->block_last_index[i] == -1)
1659                     s->coded_score[i]= INT_MAX/256;
1660             }
1661         }
1662     }
1663
1664     if((s->flags&CODEC_FLAG_GRAY) && s->mb_intra){
1665         s->block_last_index[4]=
1666         s->block_last_index[5]= 0;
1667         s->block[4][0]=
1668         s->block[5][0]= (1024 + s->c_dc_scale/2)/ s->c_dc_scale;
1669     }
1670
1671     //non c quantize code returns incorrect block_last_index FIXME
1672     if(s->alternate_scan && s->dct_quantize != dct_quantize_c){
1673         for(i=0; i<mb_block_count; i++){
1674             int j;
1675             if(s->block_last_index[i]>0){
1676                 for(j=63; j>0; j--){
1677                     if(s->block[i][ s->intra_scantable.permutated[j] ]) break;
1678                 }
1679                 s->block_last_index[i]= j;
1680             }
1681         }
1682     }
1683
1684     /* huffman encode */
1685     switch(s->codec_id){ //FIXME funct ptr could be slightly faster
1686     case CODEC_ID_MPEG1VIDEO:
1687     case CODEC_ID_MPEG2VIDEO:
1688         mpeg1_encode_mb(s, s->block, motion_x, motion_y); break;
1689     case CODEC_ID_MPEG4:
1690         mpeg4_encode_mb(s, s->block, motion_x, motion_y); break;
1691     case CODEC_ID_MSMPEG4V2:
1692     case CODEC_ID_MSMPEG4V3:
1693     case CODEC_ID_WMV1:
1694         if (ENABLE_MSMPEG4_ENCODER)
1695             msmpeg4_encode_mb(s, s->block, motion_x, motion_y);
1696         break;
1697     case CODEC_ID_WMV2:
1698         if (ENABLE_WMV2_ENCODER)
1699             ff_wmv2_encode_mb(s, s->block, motion_x, motion_y);
1700         break;
1701     case CODEC_ID_H261:
1702         if (ENABLE_H261_ENCODER)
1703             ff_h261_encode_mb(s, s->block, motion_x, motion_y);
1704         break;
1705     case CODEC_ID_H263:
1706     case CODEC_ID_H263P:
1707     case CODEC_ID_FLV1:
1708     case CODEC_ID_RV10:
1709     case CODEC_ID_RV20:
1710         h263_encode_mb(s, s->block, motion_x, motion_y); break;
1711     case CODEC_ID_MJPEG:
1712         if (ENABLE_MJPEG_ENCODER)
1713             ff_mjpeg_encode_mb(s, s->block);
1714         break;
1715     default:
1716         assert(0);
1717     }
1718 }
1719
1720 static av_always_inline void encode_mb(MpegEncContext *s, int motion_x, int motion_y)
1721 {
1722     if (s->chroma_format == CHROMA_420) encode_mb_internal(s, motion_x, motion_y,  8, 6);
1723     else                                encode_mb_internal(s, motion_x, motion_y, 16, 8);
1724 }
1725
1726 void ff_copy_bits(PutBitContext *pb, uint8_t *src, int length)
1727 {
1728     const uint16_t *srcw= (uint16_t*)src;
1729     int words= length>>4;
1730     int bits= length&15;
1731     int i;
1732
1733     if(length==0) return;
1734
1735     if(words < 16){
1736         for(i=0; i<words; i++) put_bits(pb, 16, be2me_16(srcw[i]));
1737     }else if(put_bits_count(pb)&7){
1738         for(i=0; i<words; i++) put_bits(pb, 16, be2me_16(srcw[i]));
1739     }else{
1740         for(i=0; put_bits_count(pb)&31; i++)
1741             put_bits(pb, 8, src[i]);
1742         flush_put_bits(pb);
1743         memcpy(pbBufPtr(pb), src+i, 2*words-i);
1744         skip_put_bytes(pb, 2*words-i);
1745     }
1746
1747     put_bits(pb, bits, be2me_16(srcw[words])>>(16-bits));
1748 }
1749
1750 static inline void copy_context_before_encode(MpegEncContext *d, MpegEncContext *s, int type){
1751     int i;
1752
1753     memcpy(d->last_mv, s->last_mv, 2*2*2*sizeof(int)); //FIXME is memcpy faster then a loop?
1754
1755     /* mpeg1 */
1756     d->mb_skip_run= s->mb_skip_run;
1757     for(i=0; i<3; i++)
1758         d->last_dc[i]= s->last_dc[i];
1759
1760     /* statistics */
1761     d->mv_bits= s->mv_bits;
1762     d->i_tex_bits= s->i_tex_bits;
1763     d->p_tex_bits= s->p_tex_bits;
1764     d->i_count= s->i_count;
1765     d->f_count= s->f_count;
1766     d->b_count= s->b_count;
1767     d->skip_count= s->skip_count;
1768     d->misc_bits= s->misc_bits;
1769     d->last_bits= 0;
1770
1771     d->mb_skipped= 0;
1772     d->qscale= s->qscale;
1773     d->dquant= s->dquant;
1774 }
1775
1776 static inline void copy_context_after_encode(MpegEncContext *d, MpegEncContext *s, int type){
1777     int i;
1778
1779     memcpy(d->mv, s->mv, 2*4*2*sizeof(int));
1780     memcpy(d->last_mv, s->last_mv, 2*2*2*sizeof(int)); //FIXME is memcpy faster then a loop?
1781
1782     /* mpeg1 */
1783     d->mb_skip_run= s->mb_skip_run;
1784     for(i=0; i<3; i++)
1785         d->last_dc[i]= s->last_dc[i];
1786
1787     /* statistics */
1788     d->mv_bits= s->mv_bits;
1789     d->i_tex_bits= s->i_tex_bits;
1790     d->p_tex_bits= s->p_tex_bits;
1791     d->i_count= s->i_count;
1792     d->f_count= s->f_count;
1793     d->b_count= s->b_count;
1794     d->skip_count= s->skip_count;
1795     d->misc_bits= s->misc_bits;
1796
1797     d->mb_intra= s->mb_intra;
1798     d->mb_skipped= s->mb_skipped;
1799     d->mv_type= s->mv_type;
1800     d->mv_dir= s->mv_dir;
1801     d->pb= s->pb;
1802     if(s->data_partitioning){
1803         d->pb2= s->pb2;
1804         d->tex_pb= s->tex_pb;
1805     }
1806     d->block= s->block;
1807     for(i=0; i<8; i++)
1808         d->block_last_index[i]= s->block_last_index[i];
1809     d->interlaced_dct= s->interlaced_dct;
1810     d->qscale= s->qscale;
1811 }
1812
1813 static inline void encode_mb_hq(MpegEncContext *s, MpegEncContext *backup, MpegEncContext *best, int type,
1814                            PutBitContext pb[2], PutBitContext pb2[2], PutBitContext tex_pb[2],
1815                            int *dmin, int *next_block, int motion_x, int motion_y)
1816 {
1817     int score;
1818     uint8_t *dest_backup[3];
1819
1820     copy_context_before_encode(s, backup, type);
1821
1822     s->block= s->blocks[*next_block];
1823     s->pb= pb[*next_block];
1824     if(s->data_partitioning){
1825         s->pb2   = pb2   [*next_block];
1826         s->tex_pb= tex_pb[*next_block];
1827     }
1828
1829     if(*next_block){
1830         memcpy(dest_backup, s->dest, sizeof(s->dest));
1831         s->dest[0] = s->rd_scratchpad;
1832         s->dest[1] = s->rd_scratchpad + 16*s->linesize;
1833         s->dest[2] = s->rd_scratchpad + 16*s->linesize + 8;
1834         assert(s->linesize >= 32); //FIXME
1835     }
1836
1837     encode_mb(s, motion_x, motion_y);
1838
1839     score= put_bits_count(&s->pb);
1840     if(s->data_partitioning){
1841         score+= put_bits_count(&s->pb2);
1842         score+= put_bits_count(&s->tex_pb);
1843     }
1844
1845     if(s->avctx->mb_decision == FF_MB_DECISION_RD){
1846         MPV_decode_mb(s, s->block);
1847
1848         score *= s->lambda2;
1849         score += sse_mb(s) << FF_LAMBDA_SHIFT;
1850     }
1851
1852     if(*next_block){
1853         memcpy(s->dest, dest_backup, sizeof(s->dest));
1854     }
1855
1856     if(score<*dmin){
1857         *dmin= score;
1858         *next_block^=1;
1859
1860         copy_context_after_encode(best, s, type);
1861     }
1862 }
1863
1864 static int sse(MpegEncContext *s, uint8_t *src1, uint8_t *src2, int w, int h, int stride){
1865     uint32_t *sq = ff_squareTbl + 256;
1866     int acc=0;
1867     int x,y;
1868
1869     if(w==16 && h==16)
1870         return s->dsp.sse[0](NULL, src1, src2, stride, 16);
1871     else if(w==8 && h==8)
1872         return s->dsp.sse[1](NULL, src1, src2, stride, 8);
1873
1874     for(y=0; y<h; y++){
1875         for(x=0; x<w; x++){
1876             acc+= sq[src1[x + y*stride] - src2[x + y*stride]];
1877         }
1878     }
1879
1880     assert(acc>=0);
1881
1882     return acc;
1883 }
1884
1885 static int sse_mb(MpegEncContext *s){
1886     int w= 16;
1887     int h= 16;
1888
1889     if(s->mb_x*16 + 16 > s->width ) w= s->width - s->mb_x*16;
1890     if(s->mb_y*16 + 16 > s->height) h= s->height- s->mb_y*16;
1891
1892     if(w==16 && h==16)
1893       if(s->avctx->mb_cmp == FF_CMP_NSSE){
1894         return  s->dsp.nsse[0](s, s->new_picture.data[0] + s->mb_x*16 + s->mb_y*s->linesize*16, s->dest[0], s->linesize, 16)
1895                +s->dsp.nsse[1](s, s->new_picture.data[1] + s->mb_x*8  + s->mb_y*s->uvlinesize*8,s->dest[1], s->uvlinesize, 8)
1896                +s->dsp.nsse[1](s, s->new_picture.data[2] + s->mb_x*8  + s->mb_y*s->uvlinesize*8,s->dest[2], s->uvlinesize, 8);
1897       }else{
1898         return  s->dsp.sse[0](NULL, s->new_picture.data[0] + s->mb_x*16 + s->mb_y*s->linesize*16, s->dest[0], s->linesize, 16)
1899                +s->dsp.sse[1](NULL, s->new_picture.data[1] + s->mb_x*8  + s->mb_y*s->uvlinesize*8,s->dest[1], s->uvlinesize, 8)
1900                +s->dsp.sse[1](NULL, s->new_picture.data[2] + s->mb_x*8  + s->mb_y*s->uvlinesize*8,s->dest[2], s->uvlinesize, 8);
1901       }
1902     else
1903         return  sse(s, s->new_picture.data[0] + s->mb_x*16 + s->mb_y*s->linesize*16, s->dest[0], w, h, s->linesize)
1904                +sse(s, s->new_picture.data[1] + s->mb_x*8  + s->mb_y*s->uvlinesize*8,s->dest[1], w>>1, h>>1, s->uvlinesize)
1905                +sse(s, s->new_picture.data[2] + s->mb_x*8  + s->mb_y*s->uvlinesize*8,s->dest[2], w>>1, h>>1, s->uvlinesize);
1906 }
1907
1908 static int pre_estimate_motion_thread(AVCodecContext *c, void *arg){
1909     MpegEncContext *s= arg;
1910
1911
1912     s->me.pre_pass=1;
1913     s->me.dia_size= s->avctx->pre_dia_size;
1914     s->first_slice_line=1;
1915     for(s->mb_y= s->end_mb_y-1; s->mb_y >= s->start_mb_y; s->mb_y--) {
1916         for(s->mb_x=s->mb_width-1; s->mb_x >=0 ;s->mb_x--) {
1917             ff_pre_estimate_p_frame_motion(s, s->mb_x, s->mb_y);
1918         }
1919         s->first_slice_line=0;
1920     }
1921
1922     s->me.pre_pass=0;
1923
1924     return 0;
1925 }
1926
1927 static int estimate_motion_thread(AVCodecContext *c, void *arg){
1928     MpegEncContext *s= arg;
1929
1930     ff_check_alignment();
1931
1932     s->me.dia_size= s->avctx->dia_size;
1933     s->first_slice_line=1;
1934     for(s->mb_y= s->start_mb_y; s->mb_y < s->end_mb_y; s->mb_y++) {
1935         s->mb_x=0; //for block init below
1936         ff_init_block_index(s);
1937         for(s->mb_x=0; s->mb_x < s->mb_width; s->mb_x++) {
1938             s->block_index[0]+=2;
1939             s->block_index[1]+=2;
1940             s->block_index[2]+=2;
1941             s->block_index[3]+=2;
1942
1943             /* compute motion vector & mb_type and store in context */
1944             if(s->pict_type==B_TYPE)
1945                 ff_estimate_b_frame_motion(s, s->mb_x, s->mb_y);
1946             else
1947                 ff_estimate_p_frame_motion(s, s->mb_x, s->mb_y);
1948         }
1949         s->first_slice_line=0;
1950     }
1951     return 0;
1952 }
1953
1954 static int mb_var_thread(AVCodecContext *c, void *arg){
1955     MpegEncContext *s= arg;
1956     int mb_x, mb_y;
1957
1958     ff_check_alignment();
1959
1960     for(mb_y=s->start_mb_y; mb_y < s->end_mb_y; mb_y++) {
1961         for(mb_x=0; mb_x < s->mb_width; mb_x++) {
1962             int xx = mb_x * 16;
1963             int yy = mb_y * 16;
1964             uint8_t *pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
1965             int varc;
1966             int sum = s->dsp.pix_sum(pix, s->linesize);
1967
1968             varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
1969
1970             s->current_picture.mb_var [s->mb_stride * mb_y + mb_x] = varc;
1971             s->current_picture.mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1972             s->me.mb_var_sum_temp    += varc;
1973         }
1974     }
1975     return 0;
1976 }
1977
1978 static void write_slice_end(MpegEncContext *s){
1979     if(s->codec_id==CODEC_ID_MPEG4){
1980         if(s->partitioned_frame){
1981             ff_mpeg4_merge_partitions(s);
1982         }
1983
1984         ff_mpeg4_stuffing(&s->pb);
1985     }else if(ENABLE_MJPEG_ENCODER && s->out_format == FMT_MJPEG){
1986         ff_mjpeg_encode_stuffing(&s->pb);
1987     }
1988
1989     align_put_bits(&s->pb);
1990     flush_put_bits(&s->pb);
1991
1992     if((s->flags&CODEC_FLAG_PASS1) && !s->partitioned_frame)
1993         s->misc_bits+= get_bits_diff(s);
1994 }
1995
1996 static int encode_thread(AVCodecContext *c, void *arg){
1997     MpegEncContext *s= arg;
1998     int mb_x, mb_y, pdif = 0;
1999     int i, j;
2000     MpegEncContext best_s, backup_s;
2001     uint8_t bit_buf[2][MAX_MB_BYTES];
2002     uint8_t bit_buf2[2][MAX_MB_BYTES];
2003     uint8_t bit_buf_tex[2][MAX_MB_BYTES];
2004     PutBitContext pb[2], pb2[2], tex_pb[2];
2005 //printf("%d->%d\n", s->resync_mb_y, s->end_mb_y);
2006
2007     ff_check_alignment();
2008
2009     for(i=0; i<2; i++){
2010         init_put_bits(&pb    [i], bit_buf    [i], MAX_MB_BYTES);
2011         init_put_bits(&pb2   [i], bit_buf2   [i], MAX_MB_BYTES);
2012         init_put_bits(&tex_pb[i], bit_buf_tex[i], MAX_MB_BYTES);
2013     }
2014
2015     s->last_bits= put_bits_count(&s->pb);
2016     s->mv_bits=0;
2017     s->misc_bits=0;
2018     s->i_tex_bits=0;
2019     s->p_tex_bits=0;
2020     s->i_count=0;
2021     s->f_count=0;
2022     s->b_count=0;
2023     s->skip_count=0;
2024
2025     for(i=0; i<3; i++){
2026         /* init last dc values */
2027         /* note: quant matrix value (8) is implied here */
2028         s->last_dc[i] = 128 << s->intra_dc_precision;
2029
2030         s->current_picture.error[i] = 0;
2031     }
2032     s->mb_skip_run = 0;
2033     memset(s->last_mv, 0, sizeof(s->last_mv));
2034
2035     s->last_mv_dir = 0;
2036
2037     switch(s->codec_id){
2038     case CODEC_ID_H263:
2039     case CODEC_ID_H263P:
2040     case CODEC_ID_FLV1:
2041         s->gob_index = ff_h263_get_gob_height(s);
2042         break;
2043     case CODEC_ID_MPEG4:
2044         if(s->partitioned_frame)
2045             ff_mpeg4_init_partitions(s);
2046         break;
2047     }
2048
2049     s->resync_mb_x=0;
2050     s->resync_mb_y=0;
2051     s->first_slice_line = 1;
2052     s->ptr_lastgob = s->pb.buf;
2053     for(mb_y= s->start_mb_y; mb_y < s->end_mb_y; mb_y++) {
2054 //    printf("row %d at %X\n", s->mb_y, (int)s);
2055         s->mb_x=0;
2056         s->mb_y= mb_y;
2057
2058         ff_set_qscale(s, s->qscale);
2059         ff_init_block_index(s);
2060
2061         for(mb_x=0; mb_x < s->mb_width; mb_x++) {
2062             int xy= mb_y*s->mb_stride + mb_x; // removed const, H261 needs to adjust this
2063             int mb_type= s->mb_type[xy];
2064 //            int d;
2065             int dmin= INT_MAX;
2066             int dir;
2067
2068             if(s->pb.buf_end - s->pb.buf - (put_bits_count(&s->pb)>>3) < MAX_MB_BYTES){
2069                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
2070                 return -1;
2071             }
2072             if(s->data_partitioning){
2073                 if(   s->pb2   .buf_end - s->pb2   .buf - (put_bits_count(&s->    pb2)>>3) < MAX_MB_BYTES
2074                    || s->tex_pb.buf_end - s->tex_pb.buf - (put_bits_count(&s->tex_pb )>>3) < MAX_MB_BYTES){
2075                     av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
2076                     return -1;
2077                 }
2078             }
2079
2080             s->mb_x = mb_x;
2081             s->mb_y = mb_y;  // moved into loop, can get changed by H.261
2082             ff_update_block_index(s);
2083
2084             if(ENABLE_H261_ENCODER && s->codec_id == CODEC_ID_H261){
2085                 ff_h261_reorder_mb_index(s);
2086                 xy= s->mb_y*s->mb_stride + s->mb_x;
2087                 mb_type= s->mb_type[xy];
2088             }
2089
2090             /* write gob / video packet header  */
2091             if(s->rtp_mode){
2092                 int current_packet_size, is_gob_start;
2093
2094                 current_packet_size= ((put_bits_count(&s->pb)+7)>>3) - (s->ptr_lastgob - s->pb.buf);
2095
2096                 is_gob_start= s->avctx->rtp_payload_size && current_packet_size >= s->avctx->rtp_payload_size && mb_y + mb_x>0;
2097
2098                 if(s->start_mb_y == mb_y && mb_y > 0 && mb_x==0) is_gob_start=1;
2099
2100                 switch(s->codec_id){
2101                 case CODEC_ID_H263:
2102                 case CODEC_ID_H263P:
2103                     if(!s->h263_slice_structured)
2104                         if(s->mb_x || s->mb_y%s->gob_index) is_gob_start=0;
2105                     break;
2106                 case CODEC_ID_MPEG2VIDEO:
2107                     if(s->mb_x==0 && s->mb_y!=0) is_gob_start=1;
2108                 case CODEC_ID_MPEG1VIDEO:
2109                     if(s->mb_skip_run) is_gob_start=0;
2110                     break;
2111                 }
2112
2113                 if(is_gob_start){
2114                     if(s->start_mb_y != mb_y || mb_x!=0){
2115                         write_slice_end(s);
2116
2117                         if(s->codec_id==CODEC_ID_MPEG4 && s->partitioned_frame){
2118                             ff_mpeg4_init_partitions(s);
2119                         }
2120                     }
2121
2122                     assert((put_bits_count(&s->pb)&7) == 0);
2123                     current_packet_size= pbBufPtr(&s->pb) - s->ptr_lastgob;
2124
2125                     if(s->avctx->error_rate && s->resync_mb_x + s->resync_mb_y > 0){
2126                         int r= put_bits_count(&s->pb)/8 + s->picture_number + 16 + s->mb_x + s->mb_y;
2127                         int d= 100 / s->avctx->error_rate;
2128                         if(r % d == 0){
2129                             current_packet_size=0;
2130 #ifndef ALT_BITSTREAM_WRITER
2131                             s->pb.buf_ptr= s->ptr_lastgob;
2132 #endif
2133                             assert(pbBufPtr(&s->pb) == s->ptr_lastgob);
2134                         }
2135                     }
2136
2137                     if (s->avctx->rtp_callback){
2138                         int number_mb = (mb_y - s->resync_mb_y)*s->mb_width + mb_x - s->resync_mb_x;
2139                         s->avctx->rtp_callback(s->avctx, s->ptr_lastgob, current_packet_size, number_mb);
2140                     }
2141
2142                     switch(s->codec_id){
2143                     case CODEC_ID_MPEG4:
2144                         ff_mpeg4_encode_video_packet_header(s);
2145                         ff_mpeg4_clean_buffers(s);
2146                     break;
2147                     case CODEC_ID_MPEG1VIDEO:
2148                     case CODEC_ID_MPEG2VIDEO:
2149                         ff_mpeg1_encode_slice_header(s);
2150                         ff_mpeg1_clean_buffers(s);
2151                     break;
2152                     case CODEC_ID_H263:
2153                     case CODEC_ID_H263P:
2154                         h263_encode_gob_header(s, mb_y);
2155                     break;
2156                     }
2157
2158                     if(s->flags&CODEC_FLAG_PASS1){
2159                         int bits= put_bits_count(&s->pb);
2160                         s->misc_bits+= bits - s->last_bits;
2161                         s->last_bits= bits;
2162                     }
2163
2164                     s->ptr_lastgob += current_packet_size;
2165                     s->first_slice_line=1;
2166                     s->resync_mb_x=mb_x;
2167                     s->resync_mb_y=mb_y;
2168                 }
2169             }
2170
2171             if(  (s->resync_mb_x   == s->mb_x)
2172                && s->resync_mb_y+1 == s->mb_y){
2173                 s->first_slice_line=0;
2174             }
2175
2176             s->mb_skipped=0;
2177             s->dquant=0; //only for QP_RD
2178
2179             if(mb_type & (mb_type-1) || (s->flags & CODEC_FLAG_QP_RD)){ // more than 1 MB type possible or CODEC_FLAG_QP_RD
2180                 int next_block=0;
2181                 int pb_bits_count, pb2_bits_count, tex_pb_bits_count;
2182
2183                 copy_context_before_encode(&backup_s, s, -1);
2184                 backup_s.pb= s->pb;
2185                 best_s.data_partitioning= s->data_partitioning;
2186                 best_s.partitioned_frame= s->partitioned_frame;
2187                 if(s->data_partitioning){
2188                     backup_s.pb2= s->pb2;
2189                     backup_s.tex_pb= s->tex_pb;
2190                 }
2191
2192                 if(mb_type&CANDIDATE_MB_TYPE_INTER){
2193                     s->mv_dir = MV_DIR_FORWARD;
2194                     s->mv_type = MV_TYPE_16X16;
2195                     s->mb_intra= 0;
2196                     s->mv[0][0][0] = s->p_mv_table[xy][0];
2197                     s->mv[0][0][1] = s->p_mv_table[xy][1];
2198                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_INTER, pb, pb2, tex_pb,
2199                                  &dmin, &next_block, s->mv[0][0][0], s->mv[0][0][1]);
2200                 }
2201                 if(mb_type&CANDIDATE_MB_TYPE_INTER_I){
2202                     s->mv_dir = MV_DIR_FORWARD;
2203                     s->mv_type = MV_TYPE_FIELD;
2204                     s->mb_intra= 0;
2205                     for(i=0; i<2; i++){
2206                         j= s->field_select[0][i] = s->p_field_select_table[i][xy];
2207                         s->mv[0][i][0] = s->p_field_mv_table[i][j][xy][0];
2208                         s->mv[0][i][1] = s->p_field_mv_table[i][j][xy][1];
2209                     }
2210                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_INTER_I, pb, pb2, tex_pb,
2211                                  &dmin, &next_block, 0, 0);
2212                 }
2213                 if(mb_type&CANDIDATE_MB_TYPE_SKIPPED){
2214                     s->mv_dir = MV_DIR_FORWARD;
2215                     s->mv_type = MV_TYPE_16X16;
2216                     s->mb_intra= 0;
2217                     s->mv[0][0][0] = 0;
2218                     s->mv[0][0][1] = 0;
2219                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_SKIPPED, pb, pb2, tex_pb,
2220                                  &dmin, &next_block, s->mv[0][0][0], s->mv[0][0][1]);
2221                 }
2222                 if(mb_type&CANDIDATE_MB_TYPE_INTER4V){
2223                     s->mv_dir = MV_DIR_FORWARD;
2224                     s->mv_type = MV_TYPE_8X8;
2225                     s->mb_intra= 0;
2226                     for(i=0; i<4; i++){
2227                         s->mv[0][i][0] = s->current_picture.motion_val[0][s->block_index[i]][0];
2228                         s->mv[0][i][1] = s->current_picture.motion_val[0][s->block_index[i]][1];
2229                     }
2230                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_INTER4V, pb, pb2, tex_pb,
2231                                  &dmin, &next_block, 0, 0);
2232                 }
2233                 if(mb_type&CANDIDATE_MB_TYPE_FORWARD){
2234                     s->mv_dir = MV_DIR_FORWARD;
2235                     s->mv_type = MV_TYPE_16X16;
2236                     s->mb_intra= 0;
2237                     s->mv[0][0][0] = s->b_forw_mv_table[xy][0];
2238                     s->mv[0][0][1] = s->b_forw_mv_table[xy][1];
2239                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_FORWARD, pb, pb2, tex_pb,
2240                                  &dmin, &next_block, s->mv[0][0][0], s->mv[0][0][1]);
2241                 }
2242                 if(mb_type&CANDIDATE_MB_TYPE_BACKWARD){
2243                     s->mv_dir = MV_DIR_BACKWARD;
2244                     s->mv_type = MV_TYPE_16X16;
2245                     s->mb_intra= 0;
2246                     s->mv[1][0][0] = s->b_back_mv_table[xy][0];
2247                     s->mv[1][0][1] = s->b_back_mv_table[xy][1];
2248                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_BACKWARD, pb, pb2, tex_pb,
2249                                  &dmin, &next_block, s->mv[1][0][0], s->mv[1][0][1]);
2250                 }
2251                 if(mb_type&CANDIDATE_MB_TYPE_BIDIR){
2252                     s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD;
2253                     s->mv_type = MV_TYPE_16X16;
2254                     s->mb_intra= 0;
2255                     s->mv[0][0][0] = s->b_bidir_forw_mv_table[xy][0];
2256                     s->mv[0][0][1] = s->b_bidir_forw_mv_table[xy][1];
2257                     s->mv[1][0][0] = s->b_bidir_back_mv_table[xy][0];
2258                     s->mv[1][0][1] = s->b_bidir_back_mv_table[xy][1];
2259                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_BIDIR, pb, pb2, tex_pb,
2260                                  &dmin, &next_block, 0, 0);
2261                 }
2262                 if(mb_type&CANDIDATE_MB_TYPE_FORWARD_I){
2263                     s->mv_dir = MV_DIR_FORWARD;
2264                     s->mv_type = MV_TYPE_FIELD;
2265                     s->mb_intra= 0;
2266                     for(i=0; i<2; i++){
2267                         j= s->field_select[0][i] = s->b_field_select_table[0][i][xy];
2268                         s->mv[0][i][0] = s->b_field_mv_table[0][i][j][xy][0];
2269                         s->mv[0][i][1] = s->b_field_mv_table[0][i][j][xy][1];
2270                     }
2271                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_FORWARD_I, pb, pb2, tex_pb,
2272                                  &dmin, &next_block, 0, 0);
2273                 }
2274                 if(mb_type&CANDIDATE_MB_TYPE_BACKWARD_I){
2275                     s->mv_dir = MV_DIR_BACKWARD;
2276                     s->mv_type = MV_TYPE_FIELD;
2277                     s->mb_intra= 0;
2278                     for(i=0; i<2; i++){
2279                         j= s->field_select[1][i] = s->b_field_select_table[1][i][xy];
2280                         s->mv[1][i][0] = s->b_field_mv_table[1][i][j][xy][0];
2281                         s->mv[1][i][1] = s->b_field_mv_table[1][i][j][xy][1];
2282                     }
2283                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_BACKWARD_I, pb, pb2, tex_pb,
2284                                  &dmin, &next_block, 0, 0);
2285                 }
2286                 if(mb_type&CANDIDATE_MB_TYPE_BIDIR_I){
2287                     s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD;
2288                     s->mv_type = MV_TYPE_FIELD;
2289                     s->mb_intra= 0;
2290                     for(dir=0; dir<2; dir++){
2291                         for(i=0; i<2; i++){
2292                             j= s->field_select[dir][i] = s->b_field_select_table[dir][i][xy];
2293                             s->mv[dir][i][0] = s->b_field_mv_table[dir][i][j][xy][0];
2294                             s->mv[dir][i][1] = s->b_field_mv_table[dir][i][j][xy][1];
2295                         }
2296                     }
2297                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_BIDIR_I, pb, pb2, tex_pb,
2298                                  &dmin, &next_block, 0, 0);
2299                 }
2300                 if(mb_type&CANDIDATE_MB_TYPE_INTRA){
2301                     s->mv_dir = 0;
2302                     s->mv_type = MV_TYPE_16X16;
2303                     s->mb_intra= 1;
2304                     s->mv[0][0][0] = 0;
2305                     s->mv[0][0][1] = 0;
2306                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_INTRA, pb, pb2, tex_pb,
2307                                  &dmin, &next_block, 0, 0);
2308                     if(s->h263_pred || s->h263_aic){
2309                         if(best_s.mb_intra)
2310                             s->mbintra_table[mb_x + mb_y*s->mb_stride]=1;
2311                         else
2312                             ff_clean_intra_table_entries(s); //old mode?
2313                     }
2314                 }
2315
2316                 if((s->flags & CODEC_FLAG_QP_RD) && dmin < INT_MAX){
2317                     if(best_s.mv_type==MV_TYPE_16X16){ //FIXME move 4mv after QPRD
2318                         const int last_qp= backup_s.qscale;
2319                         int qpi, qp, dc[6];
2320                         DCTELEM ac[6][16];
2321                         const int mvdir= (best_s.mv_dir&MV_DIR_BACKWARD) ? 1 : 0;
2322                         static const int dquant_tab[4]={-1,1,-2,2};
2323
2324                         assert(backup_s.dquant == 0);
2325
2326                         //FIXME intra
2327                         s->mv_dir= best_s.mv_dir;
2328                         s->mv_type = MV_TYPE_16X16;
2329                         s->mb_intra= best_s.mb_intra;
2330                         s->mv[0][0][0] = best_s.mv[0][0][0];
2331                         s->mv[0][0][1] = best_s.mv[0][0][1];
2332                         s->mv[1][0][0] = best_s.mv[1][0][0];
2333                         s->mv[1][0][1] = best_s.mv[1][0][1];
2334
2335                         qpi = s->pict_type == B_TYPE ? 2 : 0;
2336                         for(; qpi<4; qpi++){
2337                             int dquant= dquant_tab[qpi];
2338                             qp= last_qp + dquant;
2339                             if(qp < s->avctx->qmin || qp > s->avctx->qmax)
2340                                 continue;
2341                             backup_s.dquant= dquant;
2342                             if(s->mb_intra && s->dc_val[0]){
2343                                 for(i=0; i<6; i++){
2344                                     dc[i]= s->dc_val[0][ s->block_index[i] ];
2345                                     memcpy(ac[i], s->ac_val[0][s->block_index[i]], sizeof(DCTELEM)*16);
2346                                 }
2347                             }
2348
2349                             encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_INTER /* wrong but unused */, pb, pb2, tex_pb,
2350                                          &dmin, &next_block, s->mv[mvdir][0][0], s->mv[mvdir][0][1]);
2351                             if(best_s.qscale != qp){
2352                                 if(s->mb_intra && s->dc_val[0]){
2353                                     for(i=0; i<6; i++){
2354                                         s->dc_val[0][ s->block_index[i] ]= dc[i];
2355                                         memcpy(s->ac_val[0][s->block_index[i]], ac[i], sizeof(DCTELEM)*16);
2356                                     }
2357                                 }
2358                             }
2359                         }
2360                     }
2361                 }
2362                 if(mb_type&CANDIDATE_MB_TYPE_DIRECT){
2363                     int mx= s->b_direct_mv_table[xy][0];
2364                     int my= s->b_direct_mv_table[xy][1];
2365
2366                     backup_s.dquant = 0;
2367                     s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD | MV_DIRECT;
2368                     s->mb_intra= 0;
2369                     ff_mpeg4_set_direct_mv(s, mx, my);
2370                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_DIRECT, pb, pb2, tex_pb,
2371                                  &dmin, &next_block, mx, my);
2372                 }
2373                 if(mb_type&CANDIDATE_MB_TYPE_DIRECT0){
2374                     backup_s.dquant = 0;
2375                     s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD | MV_DIRECT;
2376                     s->mb_intra= 0;
2377                     ff_mpeg4_set_direct_mv(s, 0, 0);
2378                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_DIRECT, pb, pb2, tex_pb,
2379                                  &dmin, &next_block, 0, 0);
2380                 }
2381                 if(!best_s.mb_intra && s->flags2&CODEC_FLAG2_SKIP_RD){
2382                     int coded=0;
2383                     for(i=0; i<6; i++)
2384                         coded |= s->block_last_index[i];
2385                     if(coded){
2386                         int mx,my;
2387                         memcpy(s->mv, best_s.mv, sizeof(s->mv));
2388                         if(best_s.mv_dir & MV_DIRECT){
2389                             mx=my=0; //FIXME find the one we actually used
2390                             ff_mpeg4_set_direct_mv(s, mx, my);
2391                         }else if(best_s.mv_dir&MV_DIR_BACKWARD){
2392                             mx= s->mv[1][0][0];
2393                             my= s->mv[1][0][1];
2394                         }else{
2395                             mx= s->mv[0][0][0];
2396                             my= s->mv[0][0][1];
2397                         }
2398
2399                         s->mv_dir= best_s.mv_dir;
2400                         s->mv_type = best_s.mv_type;
2401                         s->mb_intra= 0;
2402 /*                        s->mv[0][0][0] = best_s.mv[0][0][0];
2403                         s->mv[0][0][1] = best_s.mv[0][0][1];
2404                         s->mv[1][0][0] = best_s.mv[1][0][0];
2405                         s->mv[1][0][1] = best_s.mv[1][0][1];*/
2406                         backup_s.dquant= 0;
2407                         s->skipdct=1;
2408                         encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_INTER /* wrong but unused */, pb, pb2, tex_pb,
2409                                         &dmin, &next_block, mx, my);
2410                         s->skipdct=0;
2411                     }
2412                 }
2413
2414                 s->current_picture.qscale_table[xy]= best_s.qscale;
2415
2416                 copy_context_after_encode(s, &best_s, -1);
2417
2418                 pb_bits_count= put_bits_count(&s->pb);
2419                 flush_put_bits(&s->pb);
2420                 ff_copy_bits(&backup_s.pb, bit_buf[next_block^1], pb_bits_count);
2421                 s->pb= backup_s.pb;
2422
2423                 if(s->data_partitioning){
2424                     pb2_bits_count= put_bits_count(&s->pb2);
2425                     flush_put_bits(&s->pb2);
2426                     ff_copy_bits(&backup_s.pb2, bit_buf2[next_block^1], pb2_bits_count);
2427                     s->pb2= backup_s.pb2;
2428
2429                     tex_pb_bits_count= put_bits_count(&s->tex_pb);
2430                     flush_put_bits(&s->tex_pb);
2431                     ff_copy_bits(&backup_s.tex_pb, bit_buf_tex[next_block^1], tex_pb_bits_count);
2432                     s->tex_pb= backup_s.tex_pb;
2433                 }
2434                 s->last_bits= put_bits_count(&s->pb);
2435
2436                 if (s->out_format == FMT_H263 && s->pict_type!=B_TYPE)
2437                     ff_h263_update_motion_val(s);
2438
2439                 if(next_block==0){ //FIXME 16 vs linesize16
2440                     s->dsp.put_pixels_tab[0][0](s->dest[0], s->rd_scratchpad                     , s->linesize  ,16);
2441                     s->dsp.put_pixels_tab[1][0](s->dest[1], s->rd_scratchpad + 16*s->linesize    , s->uvlinesize, 8);
2442                     s->dsp.put_pixels_tab[1][0](s->dest[2], s->rd_scratchpad + 16*s->linesize + 8, s->uvlinesize, 8);
2443                 }
2444
2445                 if(s->avctx->mb_decision == FF_MB_DECISION_BITS)
2446                     MPV_decode_mb(s, s->block);
2447             } else {
2448                 int motion_x = 0, motion_y = 0;
2449                 s->mv_type=MV_TYPE_16X16;
2450                 // only one MB-Type possible
2451
2452                 switch(mb_type){
2453                 case CANDIDATE_MB_TYPE_INTRA:
2454                     s->mv_dir = 0;
2455                     s->mb_intra= 1;
2456                     motion_x= s->mv[0][0][0] = 0;
2457                     motion_y= s->mv[0][0][1] = 0;
2458                     break;
2459                 case CANDIDATE_MB_TYPE_INTER:
2460                     s->mv_dir = MV_DIR_FORWARD;
2461                     s->mb_intra= 0;
2462                     motion_x= s->mv[0][0][0] = s->p_mv_table[xy][0];
2463                     motion_y= s->mv[0][0][1] = s->p_mv_table[xy][1];
2464                     break;
2465                 case CANDIDATE_MB_TYPE_INTER_I:
2466                     s->mv_dir = MV_DIR_FORWARD;
2467                     s->mv_type = MV_TYPE_FIELD;
2468                     s->mb_intra= 0;
2469                     for(i=0; i<2; i++){
2470                         j= s->field_select[0][i] = s->p_field_select_table[i][xy];
2471                         s->mv[0][i][0] = s->p_field_mv_table[i][j][xy][0];
2472                         s->mv[0][i][1] = s->p_field_mv_table[i][j][xy][1];
2473                     }
2474                     break;
2475                 case CANDIDATE_MB_TYPE_INTER4V:
2476                     s->mv_dir = MV_DIR_FORWARD;
2477                     s->mv_type = MV_TYPE_8X8;
2478                     s->mb_intra= 0;
2479                     for(i=0; i<4; i++){
2480                         s->mv[0][i][0] = s->current_picture.motion_val[0][s->block_index[i]][0];
2481                         s->mv[0][i][1] = s->current_picture.motion_val[0][s->block_index[i]][1];
2482                     }
2483                     break;
2484                 case CANDIDATE_MB_TYPE_DIRECT:
2485                     s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD | MV_DIRECT;
2486                     s->mb_intra= 0;
2487                     motion_x=s->b_direct_mv_table[xy][0];
2488                     motion_y=s->b_direct_mv_table[xy][1];
2489                     ff_mpeg4_set_direct_mv(s, motion_x, motion_y);
2490                     break;
2491                 case CANDIDATE_MB_TYPE_DIRECT0:
2492                     s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD | MV_DIRECT;
2493                     s->mb_intra= 0;
2494                     ff_mpeg4_set_direct_mv(s, 0, 0);
2495                     break;
2496                 case CANDIDATE_MB_TYPE_BIDIR:
2497                     s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD;
2498                     s->mb_intra= 0;
2499                     s->mv[0][0][0] = s->b_bidir_forw_mv_table[xy][0];
2500                     s->mv[0][0][1] = s->b_bidir_forw_mv_table[xy][1];
2501                     s->mv[1][0][0] = s->b_bidir_back_mv_table[xy][0];
2502                     s->mv[1][0][1] = s->b_bidir_back_mv_table[xy][1];
2503                     break;
2504                 case CANDIDATE_MB_TYPE_BACKWARD:
2505                     s->mv_dir = MV_DIR_BACKWARD;
2506                     s->mb_intra= 0;
2507                     motion_x= s->mv[1][0][0] = s->b_back_mv_table[xy][0];
2508                     motion_y= s->mv[1][0][1] = s->b_back_mv_table[xy][1];
2509                     break;
2510                 case CANDIDATE_MB_TYPE_FORWARD:
2511                     s->mv_dir = MV_DIR_FORWARD;
2512                     s->mb_intra= 0;
2513                     motion_x= s->mv[0][0][0] = s->b_forw_mv_table[xy][0];
2514                     motion_y= s->mv[0][0][1] = s->b_forw_mv_table[xy][1];
2515 //                    printf(" %d %d ", motion_x, motion_y);
2516                     break;
2517                 case CANDIDATE_MB_TYPE_FORWARD_I:
2518                     s->mv_dir = MV_DIR_FORWARD;
2519                     s->mv_type = MV_TYPE_FIELD;
2520                     s->mb_intra= 0;
2521                     for(i=0; i<2; i++){
2522                         j= s->field_select[0][i] = s->b_field_select_table[0][i][xy];
2523                         s->mv[0][i][0] = s->b_field_mv_table[0][i][j][xy][0];
2524                         s->mv[0][i][1] = s->b_field_mv_table[0][i][j][xy][1];
2525                     }
2526                     break;
2527                 case CANDIDATE_MB_TYPE_BACKWARD_I:
2528                     s->mv_dir = MV_DIR_BACKWARD;
2529                     s->mv_type = MV_TYPE_FIELD;
2530                     s->mb_intra= 0;
2531                     for(i=0; i<2; i++){
2532                         j= s->field_select[1][i] = s->b_field_select_table[1][i][xy];
2533                         s->mv[1][i][0] = s->b_field_mv_table[1][i][j][xy][0];
2534                         s->mv[1][i][1] = s->b_field_mv_table[1][i][j][xy][1];
2535                     }
2536                     break;
2537                 case CANDIDATE_MB_TYPE_BIDIR_I:
2538                     s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD;
2539                     s->mv_type = MV_TYPE_FIELD;
2540                     s->mb_intra= 0;
2541                     for(dir=0; dir<2; dir++){
2542                         for(i=0; i<2; i++){
2543                             j= s->field_select[dir][i] = s->b_field_select_table[dir][i][xy];
2544                             s->mv[dir][i][0] = s->b_field_mv_table[dir][i][j][xy][0];
2545                             s->mv[dir][i][1] = s->b_field_mv_table[dir][i][j][xy][1];
2546                         }
2547                     }
2548                     break;
2549                 default:
2550                     av_log(s->avctx, AV_LOG_ERROR, "illegal MB type\n");
2551                 }
2552
2553                 encode_mb(s, motion_x, motion_y);
2554
2555                 // RAL: Update last macroblock type
2556                 s->last_mv_dir = s->mv_dir;
2557
2558                 if (s->out_format == FMT_H263 && s->pict_type!=B_TYPE)
2559                     ff_h263_update_motion_val(s);
2560
2561                 MPV_decode_mb(s, s->block);
2562             }
2563
2564             /* clean the MV table in IPS frames for direct mode in B frames */
2565             if(s->mb_intra /* && I,P,S_TYPE */){
2566                 s->p_mv_table[xy][0]=0;
2567                 s->p_mv_table[xy][1]=0;
2568             }
2569
2570             if(s->flags&CODEC_FLAG_PSNR){
2571                 int w= 16;
2572                 int h= 16;
2573
2574                 if(s->mb_x*16 + 16 > s->width ) w= s->width - s->mb_x*16;
2575                 if(s->mb_y*16 + 16 > s->height) h= s->height- s->mb_y*16;
2576
2577                 s->current_picture.error[0] += sse(
2578                     s, s->new_picture.data[0] + s->mb_x*16 + s->mb_y*s->linesize*16,
2579                     s->dest[0], w, h, s->linesize);
2580                 s->current_picture.error[1] += sse(
2581                     s, s->new_picture.data[1] + s->mb_x*8  + s->mb_y*s->uvlinesize*8,
2582                     s->dest[1], w>>1, h>>1, s->uvlinesize);
2583                 s->current_picture.error[2] += sse(
2584                     s, s->new_picture    .data[2] + s->mb_x*8  + s->mb_y*s->uvlinesize*8,
2585                     s->dest[2], w>>1, h>>1, s->uvlinesize);
2586             }
2587             if(s->loop_filter){
2588                 if(s->out_format == FMT_H263)
2589                     ff_h263_loop_filter(s);
2590             }
2591 //printf("MB %d %d bits\n", s->mb_x+s->mb_y*s->mb_stride, put_bits_count(&s->pb));
2592         }
2593     }
2594
2595     //not beautiful here but we must write it before flushing so it has to be here
2596     if (ENABLE_MSMPEG4_ENCODER && s->msmpeg4_version && s->msmpeg4_version<4 && s->pict_type == I_TYPE)
2597         msmpeg4_encode_ext_header(s);
2598
2599     write_slice_end(s);
2600
2601     /* Send the last GOB if RTP */
2602     if (s->avctx->rtp_callback) {
2603         int number_mb = (mb_y - s->resync_mb_y)*s->mb_width - s->resync_mb_x;
2604         pdif = pbBufPtr(&s->pb) - s->ptr_lastgob;
2605         /* Call the RTP callback to send the last GOB */
2606         emms_c();
2607         s->avctx->rtp_callback(s->avctx, s->ptr_lastgob, pdif, number_mb);
2608     }
2609
2610     return 0;
2611 }
2612
2613 #define MERGE(field) dst->field += src->field; src->field=0
2614 static void merge_context_after_me(MpegEncContext *dst, MpegEncContext *src){
2615     MERGE(me.scene_change_score);
2616     MERGE(me.mc_mb_var_sum_temp);
2617     MERGE(me.mb_var_sum_temp);
2618 }
2619
2620 static void merge_context_after_encode(MpegEncContext *dst, MpegEncContext *src){
2621     int i;
2622
2623     MERGE(dct_count[0]); //note, the other dct vars are not part of the context
2624     MERGE(dct_count[1]);
2625     MERGE(mv_bits);
2626     MERGE(i_tex_bits);
2627     MERGE(p_tex_bits);
2628     MERGE(i_count);
2629     MERGE(f_count);
2630     MERGE(b_count);
2631     MERGE(skip_count);
2632     MERGE(misc_bits);
2633     MERGE(error_count);
2634     MERGE(padding_bug_score);
2635     MERGE(current_picture.error[0]);
2636     MERGE(current_picture.error[1]);
2637     MERGE(current_picture.error[2]);
2638
2639     if(dst->avctx->noise_reduction){
2640         for(i=0; i<64; i++){
2641             MERGE(dct_error_sum[0][i]);
2642             MERGE(dct_error_sum[1][i]);
2643         }
2644     }
2645
2646     assert(put_bits_count(&src->pb) % 8 ==0);
2647     assert(put_bits_count(&dst->pb) % 8 ==0);
2648     ff_copy_bits(&dst->pb, src->pb.buf, put_bits_count(&src->pb));
2649     flush_put_bits(&dst->pb);
2650 }
2651
2652 static int estimate_qp(MpegEncContext *s, int dry_run){
2653     if (s->next_lambda){
2654         s->current_picture_ptr->quality=
2655         s->current_picture.quality = s->next_lambda;
2656         if(!dry_run) s->next_lambda= 0;
2657     } else if (!s->fixed_qscale) {
2658         s->current_picture_ptr->quality=
2659         s->current_picture.quality = ff_rate_estimate_qscale(s, dry_run);
2660         if (s->current_picture.quality < 0)
2661             return -1;
2662     }
2663
2664     if(s->adaptive_quant){
2665         switch(s->codec_id){
2666         case CODEC_ID_MPEG4:
2667             ff_clean_mpeg4_qscales(s);
2668             break;
2669         case CODEC_ID_H263:
2670         case CODEC_ID_H263P:
2671         case CODEC_ID_FLV1:
2672             ff_clean_h263_qscales(s);
2673             break;
2674         }
2675
2676         s->lambda= s->lambda_table[0];
2677         //FIXME broken
2678     }else
2679         s->lambda= s->current_picture.quality;
2680 //printf("%d %d\n", s->avctx->global_quality, s->current_picture.quality);
2681     update_qscale(s);
2682     return 0;
2683 }
2684
2685 static int encode_picture(MpegEncContext *s, int picture_number)
2686 {
2687     int i;
2688     int bits;
2689
2690     s->picture_number = picture_number;
2691
2692     /* Reset the average MB variance */
2693     s->me.mb_var_sum_temp    =
2694     s->me.mc_mb_var_sum_temp = 0;
2695
2696     /* we need to initialize some time vars before we can encode b-frames */
2697     // RAL: Condition added for MPEG1VIDEO
2698     if (s->codec_id == CODEC_ID_MPEG1VIDEO || s->codec_id == CODEC_ID_MPEG2VIDEO || (s->h263_pred && !s->h263_msmpeg4))
2699         ff_set_mpeg4_time(s, s->picture_number);  //FIXME rename and use has_b_frames or similar
2700
2701     s->me.scene_change_score=0;
2702
2703 //    s->lambda= s->current_picture_ptr->quality; //FIXME qscale / ... stuff for ME ratedistoration
2704
2705     if(s->pict_type==I_TYPE){
2706         if(s->msmpeg4_version >= 3) s->no_rounding=1;
2707         else                        s->no_rounding=0;
2708     }else if(s->pict_type!=B_TYPE){
2709         if(s->flipflop_rounding || s->codec_id == CODEC_ID_H263P || s->codec_id == CODEC_ID_MPEG4)
2710             s->no_rounding ^= 1;
2711     }
2712
2713     if(s->flags & CODEC_FLAG_PASS2){
2714         if (estimate_qp(s,1) < 0)
2715             return -1;
2716         ff_get_2pass_fcode(s);
2717     }else if(!(s->flags & CODEC_FLAG_QSCALE)){
2718         if(s->pict_type==B_TYPE)
2719             s->lambda= s->last_lambda_for[s->pict_type];
2720         else
2721             s->lambda= s->last_lambda_for[s->last_non_b_pict_type];
2722         update_qscale(s);
2723     }
2724
2725     s->mb_intra=0; //for the rate distortion & bit compare functions
2726     for(i=1; i<s->avctx->thread_count; i++){
2727         ff_update_duplicate_context(s->thread_context[i], s);
2728     }
2729
2730     ff_init_me(s);
2731
2732     /* Estimate motion for every MB */
2733     if(s->pict_type != I_TYPE){
2734         s->lambda = (s->lambda * s->avctx->me_penalty_compensation + 128)>>8;
2735         s->lambda2= (s->lambda2* (int64_t)s->avctx->me_penalty_compensation + 128)>>8;
2736         if(s->pict_type != B_TYPE && s->avctx->me_threshold==0){
2737             if((s->avctx->pre_me && s->last_non_b_pict_type==I_TYPE) || s->avctx->pre_me==2){
2738                 s->avctx->execute(s->avctx, pre_estimate_motion_thread, (void**)&(s->thread_context[0]), NULL, s->avctx->thread_count);
2739             }
2740         }
2741
2742         s->avctx->execute(s->avctx, estimate_motion_thread, (void**)&(s->thread_context[0]), NULL, s->avctx->thread_count);
2743     }else /* if(s->pict_type == I_TYPE) */{
2744         /* I-Frame */
2745         for(i=0; i<s->mb_stride*s->mb_height; i++)
2746             s->mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
2747
2748         if(!s->fixed_qscale){
2749             /* finding spatial complexity for I-frame rate control */
2750             s->avctx->execute(s->avctx, mb_var_thread, (void**)&(s->thread_context[0]), NULL, s->avctx->thread_count);
2751         }
2752     }
2753     for(i=1; i<s->avctx->thread_count; i++){
2754         merge_context_after_me(s, s->thread_context[i]);
2755     }
2756     s->current_picture.mc_mb_var_sum= s->current_picture_ptr->mc_mb_var_sum= s->me.mc_mb_var_sum_temp;
2757     s->current_picture.   mb_var_sum= s->current_picture_ptr->   mb_var_sum= s->me.   mb_var_sum_temp;
2758     emms_c();
2759
2760     if(s->me.scene_change_score > s->avctx->scenechange_threshold && s->pict_type == P_TYPE){
2761         s->pict_type= I_TYPE;
2762         for(i=0; i<s->mb_stride*s->mb_height; i++)
2763             s->mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
2764 //printf("Scene change detected, encoding as I Frame %d %d\n", s->current_picture.mb_var_sum, s->current_picture.mc_mb_var_sum);
2765     }
2766
2767     if(!s->umvplus){
2768         if(s->pict_type==P_TYPE || s->pict_type==S_TYPE) {
2769             s->f_code= ff_get_best_fcode(s, s->p_mv_table, CANDIDATE_MB_TYPE_INTER);
2770
2771             if(s->flags & CODEC_FLAG_INTERLACED_ME){
2772                 int a,b;
2773                 a= ff_get_best_fcode(s, s->p_field_mv_table[0][0], CANDIDATE_MB_TYPE_INTER_I); //FIXME field_select
2774                 b= ff_get_best_fcode(s, s->p_field_mv_table[1][1], CANDIDATE_MB_TYPE_INTER_I);
2775                 s->f_code= FFMAX(s->f_code, FFMAX(a,b));
2776             }
2777
2778             ff_fix_long_p_mvs(s);
2779             ff_fix_long_mvs(s, NULL, 0, s->p_mv_table, s->f_code, CANDIDATE_MB_TYPE_INTER, 0);
2780             if(s->flags & CODEC_FLAG_INTERLACED_ME){
2781                 int j;
2782                 for(i=0; i<2; i++){
2783                     for(j=0; j<2; j++)
2784                         ff_fix_long_mvs(s, s->p_field_select_table[i], j,
2785                                         s->p_field_mv_table[i][j], s->f_code, CANDIDATE_MB_TYPE_INTER_I, 0);
2786                 }
2787             }
2788         }
2789
2790         if(s->pict_type==B_TYPE){
2791             int a, b;
2792
2793             a = ff_get_best_fcode(s, s->b_forw_mv_table, CANDIDATE_MB_TYPE_FORWARD);
2794             b = ff_get_best_fcode(s, s->b_bidir_forw_mv_table, CANDIDATE_MB_TYPE_BIDIR);
2795             s->f_code = FFMAX(a, b);
2796
2797             a = ff_get_best_fcode(s, s->b_back_mv_table, CANDIDATE_MB_TYPE_BACKWARD);
2798             b = ff_get_best_fcode(s, s->b_bidir_back_mv_table, CANDIDATE_MB_TYPE_BIDIR);
2799             s->b_code = FFMAX(a, b);
2800
2801             ff_fix_long_mvs(s, NULL, 0, s->b_forw_mv_table, s->f_code, CANDIDATE_MB_TYPE_FORWARD, 1);
2802             ff_fix_long_mvs(s, NULL, 0, s->b_back_mv_table, s->b_code, CANDIDATE_MB_TYPE_BACKWARD, 1);
2803             ff_fix_long_mvs(s, NULL, 0, s->b_bidir_forw_mv_table, s->f_code, CANDIDATE_MB_TYPE_BIDIR, 1);
2804             ff_fix_long_mvs(s, NULL, 0, s->b_bidir_back_mv_table, s->b_code, CANDIDATE_MB_TYPE_BIDIR, 1);
2805             if(s->flags & CODEC_FLAG_INTERLACED_ME){
2806                 int dir, j;
2807                 for(dir=0; dir<2; dir++){
2808                     for(i=0; i<2; i++){
2809                         for(j=0; j<2; j++){
2810                             int type= dir ? (CANDIDATE_MB_TYPE_BACKWARD_I|CANDIDATE_MB_TYPE_BIDIR_I)
2811                                           : (CANDIDATE_MB_TYPE_FORWARD_I |CANDIDATE_MB_TYPE_BIDIR_I);
2812                             ff_fix_long_mvs(s, s->b_field_select_table[dir][i], j,
2813                                             s->b_field_mv_table[dir][i][j], dir ? s->b_code : s->f_code, type, 1);
2814                         }
2815                     }
2816                 }
2817             }
2818         }
2819     }
2820
2821     if (estimate_qp(s, 0) < 0)
2822         return -1;
2823
2824     if(s->qscale < 3 && s->max_qcoeff<=128 && s->pict_type==I_TYPE && !(s->flags & CODEC_FLAG_QSCALE))
2825         s->qscale= 3; //reduce clipping problems
2826
2827     if (s->out_format == FMT_MJPEG) {
2828         /* for mjpeg, we do include qscale in the matrix */
2829         s->intra_matrix[0] = ff_mpeg1_default_intra_matrix[0];
2830         for(i=1;i<64;i++){
2831             int j= s->dsp.idct_permutation[i];
2832
2833             s->intra_matrix[j] = av_clip_uint8((ff_mpeg1_default_intra_matrix[i] * s->qscale) >> 3);
2834         }
2835         convert_matrix(&s->dsp, s->q_intra_matrix, s->q_intra_matrix16,
2836                        s->intra_matrix, s->intra_quant_bias, 8, 8, 1);
2837         s->qscale= 8;
2838     }
2839
2840     //FIXME var duplication
2841     s->current_picture_ptr->key_frame=
2842     s->current_picture.key_frame= s->pict_type == I_TYPE; //FIXME pic_ptr
2843     s->current_picture_ptr->pict_type=
2844     s->current_picture.pict_type= s->pict_type;
2845
2846     if(s->current_picture.key_frame)
2847         s->picture_in_gop_number=0;
2848
2849     s->last_bits= put_bits_count(&s->pb);
2850     switch(s->out_format) {
2851     case FMT_MJPEG:
2852         if (ENABLE_MJPEG_ENCODER)
2853             ff_mjpeg_encode_picture_header(s);
2854         break;
2855     case FMT_H261:
2856         if (ENABLE_H261_ENCODER)
2857             ff_h261_encode_picture_header(s, picture_number);
2858         break;
2859     case FMT_H263:
2860         if (ENABLE_WMV2_ENCODER && s->codec_id == CODEC_ID_WMV2)
2861             ff_wmv2_encode_picture_header(s, picture_number);
2862         else if (ENABLE_MSMPEG4_ENCODER && s->h263_msmpeg4)
2863             msmpeg4_encode_picture_header(s, picture_number);
2864         else if (s->h263_pred)
2865             mpeg4_encode_picture_header(s, picture_number);
2866         else if (ENABLE_RV10_ENCODER && s->codec_id == CODEC_ID_RV10)
2867             rv10_encode_picture_header(s, picture_number);
2868         else if (ENABLE_RV20_ENCODER && s->codec_id == CODEC_ID_RV20)
2869             rv20_encode_picture_header(s, picture_number);
2870         else if (s->codec_id == CODEC_ID_FLV1)
2871             ff_flv_encode_picture_header(s, picture_number);
2872         else
2873             h263_encode_picture_header(s, picture_number);
2874         break;
2875     case FMT_MPEG1:
2876         mpeg1_encode_picture_header(s, picture_number);
2877         break;
2878     case FMT_H264:
2879         break;
2880     default:
2881         assert(0);
2882     }
2883     bits= put_bits_count(&s->pb);
2884     s->header_bits= bits - s->last_bits;
2885
2886     for(i=1; i<s->avctx->thread_count; i++){
2887         update_duplicate_context_after_me(s->thread_context[i], s);
2888     }
2889     s->avctx->execute(s->avctx, encode_thread, (void**)&(s->thread_context[0]), NULL, s->avctx->thread_count);
2890     for(i=1; i<s->avctx->thread_count; i++){
2891         merge_context_after_encode(s, s->thread_context[i]);
2892     }
2893     emms_c();
2894     return 0;
2895 }
2896
2897 void  denoise_dct_c(MpegEncContext *s, DCTELEM *block){
2898     const int intra= s->mb_intra;
2899     int i;
2900
2901     s->dct_count[intra]++;
2902
2903     for(i=0; i<64; i++){
2904         int level= block[i];
2905
2906         if(level){
2907             if(level>0){
2908                 s->dct_error_sum[intra][i] += level;
2909                 level -= s->dct_offset[intra][i];
2910                 if(level<0) level=0;
2911             }else{
2912                 s->dct_error_sum[intra][i] -= level;
2913                 level += s->dct_offset[intra][i];
2914                 if(level>0) level=0;
2915             }
2916             block[i]= level;
2917         }
2918     }
2919 }
2920
2921 int dct_quantize_trellis_c(MpegEncContext *s,
2922                         DCTELEM *block, int n,
2923                         int qscale, int *overflow){
2924     const int *qmat;
2925     const uint8_t *scantable= s->intra_scantable.scantable;
2926     const uint8_t *perm_scantable= s->intra_scantable.permutated;
2927     int max=0;
2928     unsigned int threshold1, threshold2;
2929     int bias=0;
2930     int run_tab[65];
2931     int level_tab[65];
2932     int score_tab[65];
2933     int survivor[65];
2934     int survivor_count;
2935     int last_run=0;
2936     int last_level=0;
2937     int last_score= 0;
2938     int last_i;
2939     int coeff[2][64];
2940     int coeff_count[64];
2941     int qmul, qadd, start_i, last_non_zero, i, dc;
2942     const int esc_length= s->ac_esc_length;
2943     uint8_t * length;
2944     uint8_t * last_length;
2945     const int lambda= s->lambda2 >> (FF_LAMBDA_SHIFT - 6);
2946
2947     s->dsp.fdct (block);
2948
2949     if(s->dct_error_sum)
2950         s->denoise_dct(s, block);
2951     qmul= qscale*16;
2952     qadd= ((qscale-1)|1)*8;
2953
2954     if (s->mb_intra) {
2955         int q;
2956         if (!s->h263_aic) {
2957             if (n < 4)
2958                 q = s->y_dc_scale;
2959             else
2960                 q = s->c_dc_scale;
2961             q = q << 3;
2962         } else{
2963             /* For AIC we skip quant/dequant of INTRADC */
2964             q = 1 << 3;
2965             qadd=0;
2966         }
2967
2968         /* note: block[0] is assumed to be positive */
2969         block[0] = (block[0] + (q >> 1)) / q;
2970         start_i = 1;
2971         last_non_zero = 0;
2972         qmat = s->q_intra_matrix[qscale];
2973         if(s->mpeg_quant || s->out_format == FMT_MPEG1)
2974             bias= 1<<(QMAT_SHIFT-1);
2975         length     = s->intra_ac_vlc_length;
2976         last_length= s->intra_ac_vlc_last_length;
2977     } else {
2978         start_i = 0;
2979         last_non_zero = -1;
2980         qmat = s->q_inter_matrix[qscale];
2981         length     = s->inter_ac_vlc_length;
2982         last_length= s->inter_ac_vlc_last_length;
2983     }
2984     last_i= start_i;
2985
2986     threshold1= (1<<QMAT_SHIFT) - bias - 1;
2987     threshold2= (threshold1<<1);
2988
2989     for(i=63; i>=start_i; i--) {
2990         const int j = scantable[i];
2991         int level = block[j] * qmat[j];
2992
2993         if(((unsigned)(level+threshold1))>threshold2){
2994             last_non_zero = i;
2995             break;
2996         }
2997     }
2998
2999     for(i=start_i; i<=last_non_zero; i++) {
3000         const int j = scantable[i];
3001         int level = block[j] * qmat[j];
3002
3003 //        if(   bias+level >= (1<<(QMAT_SHIFT - 3))
3004 //           || bias-level >= (1<<(QMAT_SHIFT - 3))){
3005         if(((unsigned)(level+threshold1))>threshold2){
3006             if(level>0){
3007                 level= (bias + level)>>QMAT_SHIFT;
3008                 coeff[0][i]= level;
3009                 coeff[1][i]= level-1;
3010 //                coeff[2][k]= level-2;
3011             }else{
3012                 level= (bias - level)>>QMAT_SHIFT;
3013                 coeff[0][i]= -level;
3014                 coeff[1][i]= -level+1;
3015 //                coeff[2][k]= -level+2;
3016             }
3017             coeff_count[i]= FFMIN(level, 2);
3018             assert(coeff_count[i]);
3019             max |=level;
3020         }else{
3021             coeff[0][i]= (level>>31)|1;
3022             coeff_count[i]= 1;
3023         }
3024     }
3025
3026     *overflow= s->max_qcoeff < max; //overflow might have happened
3027
3028     if(last_non_zero < start_i){
3029         memset(block + start_i, 0, (64-start_i)*sizeof(DCTELEM));
3030         return last_non_zero;
3031     }
3032
3033     score_tab[start_i]= 0;
3034     survivor[0]= start_i;
3035     survivor_count= 1;
3036
3037     for(i=start_i; i<=last_non_zero; i++){
3038         int level_index, j;
3039         const int dct_coeff= FFABS(block[ scantable[i] ]);
3040         const int zero_distoration= dct_coeff*dct_coeff;
3041         int best_score=256*256*256*120;
3042         for(level_index=0; level_index < coeff_count[i]; level_index++){
3043             int distoration;
3044             int level= coeff[level_index][i];
3045             const int alevel= FFABS(level);
3046             int unquant_coeff;
3047
3048             assert(level);
3049
3050             if(s->out_format == FMT_H263){
3051                 unquant_coeff= alevel*qmul + qadd;
3052             }else{ //MPEG1
3053                 j= s->dsp.idct_permutation[ scantable[i] ]; //FIXME optimize
3054                 if(s->mb_intra){
3055                         unquant_coeff = (int)(  alevel  * qscale * s->intra_matrix[j]) >> 3;
3056                         unquant_coeff =   (unquant_coeff - 1) | 1;
3057                 }else{
3058                         unquant_coeff = (((  alevel  << 1) + 1) * qscale * ((int) s->inter_matrix[j])) >> 4;
3059                         unquant_coeff =   (unquant_coeff - 1) | 1;
3060                 }
3061                 unquant_coeff<<= 3;
3062             }
3063
3064             distoration= (unquant_coeff - dct_coeff) * (unquant_coeff - dct_coeff) - zero_distoration;
3065             level+=64;
3066             if((level&(~127)) == 0){
3067                 for(j=survivor_count-1; j>=0; j--){
3068                     int run= i - survivor[j];
3069                     int score= distoration + length[UNI_AC_ENC_INDEX(run, level)]*lambda;
3070                     score += score_tab[i-run];
3071
3072                     if(score < best_score){
3073                         best_score= score;
3074                         run_tab[i+1]= run;
3075                         level_tab[i+1]= level-64;
3076                     }
3077                 }
3078
3079                 if(s->out_format == FMT_H263){
3080                     for(j=survivor_count-1; j>=0; j--){
3081                         int run= i - survivor[j];
3082                         int score= distoration + last_length[UNI_AC_ENC_INDEX(run, level)]*lambda;
3083                         score += score_tab[i-run];
3084                         if(score < last_score){
3085                             last_score= score;
3086                             last_run= run;
3087                             last_level= level-64;
3088                             last_i= i+1;
3089                         }
3090                     }
3091                 }
3092             }else{
3093                 distoration += esc_length*lambda;
3094                 for(j=survivor_count-1; j>=0; j--){
3095                     int run= i - survivor[j];
3096                     int score= distoration + score_tab[i-run];
3097
3098                     if(score < best_score){
3099                         best_score= score;
3100                         run_tab[i+1]= run;
3101                         level_tab[i+1]= level-64;
3102                     }
3103                 }
3104
3105                 if(s->out_format == FMT_H263){
3106                   for(j=survivor_count-1; j>=0; j--){
3107                         int run= i - survivor[j];
3108                         int score= distoration + score_tab[i-run];
3109                         if(score < last_score){
3110                             last_score= score;
3111                             last_run= run;
3112                             last_level= level-64;
3113                             last_i= i+1;
3114                         }
3115                     }
3116                 }
3117             }
3118         }
3119
3120         score_tab[i+1]= best_score;
3121
3122         //Note: there is a vlc code in mpeg4 which is 1 bit shorter then another one with a shorter run and the same level
3123         if(last_non_zero <= 27){
3124             for(; survivor_count; survivor_count--){
3125                 if(score_tab[ survivor[survivor_count-1] ] <= best_score)
3126                     break;
3127             }
3128         }else{
3129             for(; survivor_count; survivor_count--){
3130                 if(score_tab[ survivor[survivor_count-1] ] <= best_score + lambda)
3131                     break;
3132             }
3133         }
3134
3135         survivor[ survivor_count++ ]= i+1;
3136     }
3137
3138     if(s->out_format != FMT_H263){
3139         last_score= 256*256*256*120;
3140         for(i= survivor[0]; i<=last_non_zero + 1; i++){
3141             int score= score_tab[i];
3142             if(i) score += lambda*2; //FIXME exacter?
3143
3144             if(score < last_score){
3145                 last_score= score;
3146                 last_i= i;
3147                 last_level= level_tab[i];
3148                 last_run= run_tab[i];
3149             }
3150         }
3151     }
3152
3153     s->coded_score[n] = last_score;
3154
3155     dc= FFABS(block[0]);
3156     last_non_zero= last_i - 1;
3157     memset(block + start_i, 0, (64-start_i)*sizeof(DCTELEM));
3158
3159     if(last_non_zero < start_i)
3160         return last_non_zero;
3161
3162     if(last_non_zero == 0 && start_i == 0){
3163         int best_level= 0;
3164         int best_score= dc * dc;
3165
3166         for(i=0; i<coeff_count[0]; i++){
3167             int level= coeff[i][0];
3168             int alevel= FFABS(level);
3169             int unquant_coeff, score, distortion;
3170
3171             if(s->out_format == FMT_H263){
3172                     unquant_coeff= (alevel*qmul + qadd)>>3;
3173             }else{ //MPEG1
3174                     unquant_coeff = (((  alevel  << 1) + 1) * qscale * ((int) s->inter_matrix[0])) >> 4;
3175                     unquant_coeff =   (unquant_coeff - 1) | 1;
3176             }
3177             unquant_coeff = (unquant_coeff + 4) >> 3;
3178             unquant_coeff<<= 3 + 3;
3179
3180             distortion= (unquant_coeff - dc) * (unquant_coeff - dc);
3181             level+=64;
3182             if((level&(~127)) == 0) score= distortion + last_length[UNI_AC_ENC_INDEX(0, level)]*lambda;
3183             else                    score= distortion + esc_length*lambda;
3184
3185             if(score < best_score){
3186                 best_score= score;
3187                 best_level= level - 64;
3188             }
3189         }
3190         block[0]= best_level;
3191         s->coded_score[n] = best_score - dc*dc;
3192         if(best_level == 0) return -1;
3193         else                return last_non_zero;
3194     }
3195
3196     i= last_i;
3197     assert(last_level);
3198
3199     block[ perm_scantable[last_non_zero] ]= last_level;
3200     i -= last_run + 1;
3201
3202     for(; i>start_i; i -= run_tab[i] + 1){
3203         block[ perm_scantable[i-1] ]= level_tab[i];
3204     }
3205
3206     return last_non_zero;
3207 }
3208
3209 //#define REFINE_STATS 1
3210 static int16_t basis[64][64];
3211
3212 static void build_basis(uint8_t *perm){
3213     int i, j, x, y;
3214     emms_c();
3215     for(i=0; i<8; i++){
3216         for(j=0; j<8; j++){
3217             for(y=0; y<8; y++){
3218                 for(x=0; x<8; x++){
3219                     double s= 0.25*(1<<BASIS_SHIFT);
3220                     int index= 8*i + j;
3221                     int perm_index= perm[index];
3222                     if(i==0) s*= sqrt(0.5);
3223                     if(j==0) s*= sqrt(0.5);
3224                     basis[perm_index][8*x + y]= lrintf(s * cos((M_PI/8.0)*i*(x+0.5)) * cos((M_PI/8.0)*j*(y+0.5)));
3225                 }
3226             }
3227         }
3228     }
3229 }
3230
3231 static int dct_quantize_refine(MpegEncContext *s, //FIXME breaks denoise?
3232                         DCTELEM *block, int16_t *weight, DCTELEM *orig,
3233                         int n, int qscale){
3234     int16_t rem[64];
3235     DECLARE_ALIGNED_16(DCTELEM, d1[64]);
3236     const int *qmat;
3237     const uint8_t *scantable= s->intra_scantable.scantable;
3238     const uint8_t *perm_scantable= s->intra_scantable.permutated;
3239 //    unsigned int threshold1, threshold2;
3240 //    int bias=0;
3241     int run_tab[65];
3242     int prev_run=0;
3243     int prev_level=0;
3244     int qmul, qadd, start_i, last_non_zero, i, dc;
3245     uint8_t * length;
3246     uint8_t * last_length;
3247     int lambda;
3248     int rle_index, run, q = 1, sum; //q is only used when s->mb_intra is true
3249 #ifdef REFINE_STATS
3250 static int count=0;
3251 static int after_last=0;
3252 static int to_zero=0;
3253 static int from_zero=0;
3254 static int raise=0;
3255 static int lower=0;
3256 static int messed_sign=0;
3257 #endif
3258
3259     if(basis[0][0] == 0)
3260         build_basis(s->dsp.idct_permutation);
3261
3262     qmul= qscale*2;
3263     qadd= (qscale-1)|1;
3264     if (s->mb_intra) {
3265         if (!s->h263_aic) {
3266             if (n < 4)
3267                 q = s->y_dc_scale;
3268             else
3269                 q = s->c_dc_scale;
3270         } else{
3271             /* For AIC we skip quant/dequant of INTRADC */
3272             q = 1;
3273             qadd=0;
3274         }
3275         q <<= RECON_SHIFT-3;
3276         /* note: block[0] is assumed to be positive */
3277         dc= block[0]*q;
3278 //        block[0] = (block[0] + (q >> 1)) / q;
3279         start_i = 1;
3280         qmat = s->q_intra_matrix[qscale];
3281 //        if(s->mpeg_quant || s->out_format == FMT_MPEG1)
3282 //            bias= 1<<(QMAT_SHIFT-1);
3283         length     = s->intra_ac_vlc_length;
3284         last_length= s->intra_ac_vlc_last_length;
3285     } else {
3286         dc= 0;
3287         start_i = 0;
3288         qmat = s->q_inter_matrix[qscale];
3289         length     = s->inter_ac_vlc_length;
3290         last_length= s->inter_ac_vlc_last_length;
3291     }
3292     last_non_zero = s->block_last_index[n];
3293
3294 #ifdef REFINE_STATS
3295 {START_TIMER
3296 #endif
3297     dc += (1<<(RECON_SHIFT-1));
3298     for(i=0; i<64; i++){
3299         rem[i]= dc - (orig[i]<<RECON_SHIFT); //FIXME  use orig dirrectly instead of copying to rem[]
3300     }
3301 #ifdef REFINE_STATS
3302 STOP_TIMER("memset rem[]")}
3303 #endif
3304     sum=0;
3305     for(i=0; i<64; i++){
3306         int one= 36;
3307         int qns=4;
3308         int w;
3309
3310         w= FFABS(weight[i]) + qns*one;
3311         w= 15 + (48*qns*one + w/2)/w; // 16 .. 63
3312
3313         weight[i] = w;
3314 //        w=weight[i] = (63*qns + (w/2)) / w;
3315
3316         assert(w>0);
3317         assert(w<(1<<6));
3318         sum += w*w;
3319     }
3320     lambda= sum*(uint64_t)s->lambda2 >> (FF_LAMBDA_SHIFT - 6 + 6 + 6 + 6);
3321 #ifdef REFINE_STATS
3322 {START_TIMER
3323 #endif
3324     run=0;
3325     rle_index=0;
3326     for(i=start_i; i<=last_non_zero; i++){
3327         int j= perm_scantable[i];
3328         const int level= block[j];
3329         int coeff;
3330
3331         if(level){
3332             if(level<0) coeff= qmul*level - qadd;
3333             else        coeff= qmul*level + qadd;
3334             run_tab[rle_index++]=run;
3335             run=0;
3336
3337             s->dsp.add_8x8basis(rem, basis[j], coeff);
3338         }else{
3339             run++;
3340         }
3341     }
3342 #ifdef REFINE_STATS
3343 if(last_non_zero>0){
3344 STOP_TIMER("init rem[]")
3345 }
3346 }
3347
3348 {START_TIMER
3349 #endif
3350     for(;;){
3351         int best_score=s->dsp.try_8x8basis(rem, weight, basis[0], 0);
3352         int best_coeff=0;
3353         int best_change=0;
3354         int run2, best_unquant_change=0, analyze_gradient;
3355 #ifdef REFINE_STATS
3356 {START_TIMER
3357 #endif
3358         analyze_gradient = last_non_zero > 2 || s->avctx->quantizer_noise_shaping >= 3;
3359
3360         if(analyze_gradient){
3361 #ifdef REFINE_STATS
3362 {START_TIMER
3363 #endif
3364             for(i=0; i<64; i++){
3365                 int w= weight[i];
3366
3367                 d1[i] = (rem[i]*w*w + (1<<(RECON_SHIFT+12-1)))>>(RECON_SHIFT+12);
3368             }
3369 #ifdef REFINE_STATS
3370 STOP_TIMER("rem*w*w")}
3371 {START_TIMER
3372 #endif
3373             s->dsp.fdct(d1);
3374 #ifdef REFINE_STATS
3375 STOP_TIMER("dct")}
3376 #endif
3377         }
3378
3379         if(start_i){
3380             const int level= block[0];
3381             int change, old_coeff;
3382
3383             assert(s->mb_intra);
3384
3385             old_coeff= q*level;
3386
3387             for(change=-1; change<=1; change+=2){
3388                 int new_level= level + change;
3389                 int score, new_coeff;
3390
3391                 new_coeff= q*new_level;
3392