ef830affd7d695e46026e04dfd22c5053e3352d7
[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  * 4MV & hq & B-frame encoding stuff by Michael Niedermayer <michaelni@gmx.at>
7  *
8  * This file is part of FFmpeg.
9  *
10  * FFmpeg is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU Lesser General Public
12  * License as published by the Free Software Foundation; either
13  * version 2.1 of the License, or (at your option) any later version.
14  *
15  * FFmpeg is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18  * Lesser General Public License for more details.
19  *
20  * You should have received a copy of the GNU Lesser General Public
21  * License along with FFmpeg; if not, write to the Free Software
22  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23  */
24
25 /**
26  * @file
27  * The simplest mpeg encoder (well, it was the simplest!).
28  */
29
30 #include "libavutil/intmath.h"
31 #include "avcodec.h"
32 #include "dsputil.h"
33 #include "mpegvideo.h"
34 #include "mpegvideo_common.h"
35 #include "h263.h"
36 #include "mjpegenc.h"
37 #include "msmpeg4.h"
38 #include "faandct.h"
39 #include "aandcttab.h"
40 #include "flv.h"
41 #include "mpeg4video.h"
42 #include "internal.h"
43 #include <limits.h>
44
45 //#undef NDEBUG
46 //#include <assert.h>
47
48 static int encode_picture(MpegEncContext *s, int picture_number);
49 static int dct_quantize_refine(MpegEncContext *s, DCTELEM *block, int16_t *weight, DCTELEM *orig, int n, int qscale);
50 static int sse_mb(MpegEncContext *s);
51 static void denoise_dct_c(MpegEncContext *s, DCTELEM *block);
52 static int dct_quantize_trellis_c(MpegEncContext *s, DCTELEM *block, int n, int qscale, int *overflow);
53
54 /* enable all paranoid tests for rounding, overflows, etc... */
55 //#define PARANOID
56
57 //#define DEBUG
58
59 static uint8_t default_mv_penalty[MAX_FCODE+1][MAX_MV*2+1];
60 static uint8_t default_fcode_tab[MAX_MV*2+1];
61
62 void ff_convert_matrix(DSPContext *dsp, int (*qmat)[64], uint16_t (*qmat16)[2][64],
63                            const uint16_t *quant_matrix, int bias, int qmin, int qmax, int intra)
64 {
65     int qscale;
66     int shift=0;
67
68     for(qscale=qmin; qscale<=qmax; qscale++){
69         int i;
70         if (dsp->fdct == ff_jpeg_fdct_islow
71 #ifdef FAAN_POSTSCALE
72             || dsp->fdct == ff_faandct
73 #endif
74             ) {
75             for(i=0;i<64;i++) {
76                 const int j= dsp->idct_permutation[i];
77                 /* 16 <= qscale * quant_matrix[i] <= 7905 */
78                 /* 19952             <= ff_aanscales[i] * qscale * quant_matrix[i]               <= 249205026 */
79                 /* (1 << 36) / 19952 >= (1 << 36) / (ff_aanscales[i] * qscale * quant_matrix[i]) >= (1 << 36) / 249205026 */
80                 /* 3444240           >= (1 << 36) / (ff_aanscales[i] * qscale * quant_matrix[i]) >= 275 */
81
82                 qmat[qscale][i] = (int)((UINT64_C(1) << QMAT_SHIFT) /
83                                 (qscale * quant_matrix[j]));
84             }
85         } else if (dsp->fdct == fdct_ifast
86 #ifndef FAAN_POSTSCALE
87                    || dsp->fdct == ff_faandct
88 #endif
89                    ) {
90             for(i=0;i<64;i++) {
91                 const int j= dsp->idct_permutation[i];
92                 /* 16 <= qscale * quant_matrix[i] <= 7905 */
93                 /* 19952             <= ff_aanscales[i] * qscale * quant_matrix[i]               <= 249205026 */
94                 /* (1 << 36) / 19952 >= (1 << 36) / (ff_aanscales[i] * qscale * quant_matrix[i]) >= (1<<36)/249205026 */
95                 /* 3444240           >= (1 << 36) / (ff_aanscales[i] * qscale * quant_matrix[i]) >= 275 */
96
97                 qmat[qscale][i] = (int)((UINT64_C(1) << (QMAT_SHIFT + 14)) /
98                                 (ff_aanscales[i] * qscale * quant_matrix[j]));
99             }
100         } else {
101             for(i=0;i<64;i++) {
102                 const int j= dsp->idct_permutation[i];
103                 /* We can safely suppose that 16 <= quant_matrix[i] <= 255
104                    So 16           <= qscale * quant_matrix[i]             <= 7905
105                    so (1<<19) / 16 >= (1<<19) / (qscale * quant_matrix[i]) >= (1<<19) / 7905
106                    so 32768        >= (1<<19) / (qscale * quant_matrix[i]) >= 67
107                 */
108                 qmat[qscale][i] = (int)((UINT64_C(1) << QMAT_SHIFT) / (qscale * quant_matrix[j]));
109 //                qmat  [qscale][i] = (1 << QMAT_SHIFT_MMX) / (qscale * quant_matrix[i]);
110                 qmat16[qscale][0][i] = (1 << QMAT_SHIFT_MMX) / (qscale * quant_matrix[j]);
111
112                 if(qmat16[qscale][0][i]==0 || qmat16[qscale][0][i]==128*256) qmat16[qscale][0][i]=128*256-1;
113                 qmat16[qscale][1][i]= ROUNDED_DIV(bias<<(16-QUANT_BIAS_SHIFT), qmat16[qscale][0][i]);
114             }
115         }
116
117         for(i=intra; i<64; i++){
118             int64_t max= 8191;
119             if (dsp->fdct == fdct_ifast
120 #ifndef FAAN_POSTSCALE
121                    || dsp->fdct == ff_faandct
122 #endif
123                    ) {
124                 max = (8191LL*ff_aanscales[i]) >> 14;
125             }
126             while(((max * qmat[qscale][i]) >> shift) > INT_MAX){
127                 shift++;
128             }
129         }
130     }
131     if(shift){
132         av_log(NULL, AV_LOG_INFO, "Warning, QMAT_SHIFT is larger than %d, overflows possible\n", QMAT_SHIFT - shift);
133     }
134 }
135
136 static inline void update_qscale(MpegEncContext *s){
137     s->qscale= (s->lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
138     s->qscale= av_clip(s->qscale, s->avctx->qmin, s->avctx->qmax);
139
140     s->lambda2= (s->lambda*s->lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
141 }
142
143 void ff_write_quant_matrix(PutBitContext *pb, uint16_t *matrix){
144     int i;
145
146     if(matrix){
147         put_bits(pb, 1, 1);
148         for(i=0;i<64;i++) {
149             put_bits(pb, 8, matrix[ ff_zigzag_direct[i] ]);
150         }
151     }else
152         put_bits(pb, 1, 0);
153 }
154
155 /**
156  * init s->current_picture.qscale_table from s->lambda_table
157  */
158 void ff_init_qscale_tab(MpegEncContext *s){
159     int8_t * const qscale_table= s->current_picture.qscale_table;
160     int i;
161
162     for(i=0; i<s->mb_num; i++){
163         unsigned int lam= s->lambda_table[ s->mb_index2xy[i] ];
164         int qp= (lam*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
165         qscale_table[ s->mb_index2xy[i] ]= av_clip(qp, s->avctx->qmin, s->avctx->qmax);
166     }
167 }
168
169 static void copy_picture_attributes(MpegEncContext *s, AVFrame *dst, AVFrame *src){
170     int i;
171
172     dst->pict_type              = src->pict_type;
173     dst->quality                = src->quality;
174     dst->coded_picture_number   = src->coded_picture_number;
175     dst->display_picture_number = src->display_picture_number;
176 //    dst->reference              = src->reference;
177     dst->pts                    = src->pts;
178     dst->interlaced_frame       = src->interlaced_frame;
179     dst->top_field_first        = src->top_field_first;
180
181     if(s->avctx->me_threshold){
182         if(!src->motion_val[0])
183             av_log(s->avctx, AV_LOG_ERROR, "AVFrame.motion_val not set!\n");
184         if(!src->mb_type)
185             av_log(s->avctx, AV_LOG_ERROR, "AVFrame.mb_type not set!\n");
186         if(!src->ref_index[0])
187             av_log(s->avctx, AV_LOG_ERROR, "AVFrame.ref_index not set!\n");
188         if(src->motion_subsample_log2 != dst->motion_subsample_log2)
189             av_log(s->avctx, AV_LOG_ERROR, "AVFrame.motion_subsample_log2 doesn't match! (%d!=%d)\n",
190             src->motion_subsample_log2, dst->motion_subsample_log2);
191
192         memcpy(dst->mb_type, src->mb_type, s->mb_stride * s->mb_height * sizeof(dst->mb_type[0]));
193
194         for(i=0; i<2; i++){
195             int stride= ((16*s->mb_width )>>src->motion_subsample_log2) + 1;
196             int height= ((16*s->mb_height)>>src->motion_subsample_log2);
197
198             if(src->motion_val[i] && src->motion_val[i] != dst->motion_val[i]){
199                 memcpy(dst->motion_val[i], src->motion_val[i], 2*stride*height*sizeof(int16_t));
200             }
201             if(src->ref_index[i] && src->ref_index[i] != dst->ref_index[i]){
202                 memcpy(dst->ref_index[i], src->ref_index[i], s->mb_stride*4*s->mb_height*sizeof(int8_t));
203             }
204         }
205     }
206 }
207
208 static void update_duplicate_context_after_me(MpegEncContext *dst, MpegEncContext *src){
209 #define COPY(a) dst->a= src->a
210     COPY(pict_type);
211     COPY(current_picture);
212     COPY(f_code);
213     COPY(b_code);
214     COPY(qscale);
215     COPY(lambda);
216     COPY(lambda2);
217     COPY(picture_in_gop_number);
218     COPY(gop_picture_number);
219     COPY(frame_pred_frame_dct); //FIXME don't set in encode_header
220     COPY(progressive_frame); //FIXME don't set in encode_header
221     COPY(partitioned_frame); //FIXME don't set in encode_header
222 #undef COPY
223 }
224
225 /**
226  * sets the given MpegEncContext to defaults for encoding.
227  * the changed fields will not depend upon the prior state of the MpegEncContext.
228  */
229 static void MPV_encode_defaults(MpegEncContext *s){
230     int i;
231     MPV_common_defaults(s);
232
233     for(i=-16; i<16; i++){
234         default_fcode_tab[i + MAX_MV]= 1;
235     }
236     s->me.mv_penalty= default_mv_penalty;
237     s->fcode_tab= default_fcode_tab;
238 }
239
240 /* init video encoder */
241 av_cold int MPV_encode_init(AVCodecContext *avctx)
242 {
243     MpegEncContext *s = avctx->priv_data;
244     int i;
245     int chroma_h_shift, chroma_v_shift;
246
247     MPV_encode_defaults(s);
248
249     switch (avctx->codec_id) {
250     case CODEC_ID_MPEG2VIDEO:
251         if(avctx->pix_fmt != PIX_FMT_YUV420P && avctx->pix_fmt != PIX_FMT_YUV422P){
252             av_log(avctx, AV_LOG_ERROR, "only YUV420 and YUV422 are supported\n");
253             return -1;
254         }
255         break;
256     case CODEC_ID_LJPEG:
257         if(avctx->pix_fmt != PIX_FMT_YUVJ420P && avctx->pix_fmt != PIX_FMT_YUVJ422P && avctx->pix_fmt != PIX_FMT_YUVJ444P && avctx->pix_fmt != PIX_FMT_BGRA &&
258            ((avctx->pix_fmt != PIX_FMT_YUV420P && avctx->pix_fmt != PIX_FMT_YUV422P && avctx->pix_fmt != PIX_FMT_YUV444P) || avctx->strict_std_compliance>FF_COMPLIANCE_UNOFFICIAL)){
259             av_log(avctx, AV_LOG_ERROR, "colorspace not supported in LJPEG\n");
260             return -1;
261         }
262         break;
263     case CODEC_ID_MJPEG:
264         if(avctx->pix_fmt != PIX_FMT_YUVJ420P && avctx->pix_fmt != PIX_FMT_YUVJ422P &&
265            ((avctx->pix_fmt != PIX_FMT_YUV420P && avctx->pix_fmt != PIX_FMT_YUV422P) || avctx->strict_std_compliance>FF_COMPLIANCE_UNOFFICIAL)){
266             av_log(avctx, AV_LOG_ERROR, "colorspace not supported in jpeg\n");
267             return -1;
268         }
269         break;
270     default:
271         if(avctx->pix_fmt != PIX_FMT_YUV420P){
272             av_log(avctx, AV_LOG_ERROR, "only YUV420 is supported\n");
273             return -1;
274         }
275     }
276
277     switch (avctx->pix_fmt) {
278     case PIX_FMT_YUVJ422P:
279     case PIX_FMT_YUV422P:
280         s->chroma_format = CHROMA_422;
281         break;
282     case PIX_FMT_YUVJ420P:
283     case PIX_FMT_YUV420P:
284     default:
285         s->chroma_format = CHROMA_420;
286         break;
287     }
288
289     s->bit_rate = avctx->bit_rate;
290     s->width = avctx->width;
291     s->height = avctx->height;
292     if(avctx->gop_size > 600 && avctx->strict_std_compliance>FF_COMPLIANCE_EXPERIMENTAL){
293         av_log(avctx, AV_LOG_ERROR, "Warning keyframe interval too large! reducing it ...\n");
294         avctx->gop_size=600;
295     }
296     s->gop_size = avctx->gop_size;
297     s->avctx = avctx;
298     s->flags= avctx->flags;
299     s->flags2= avctx->flags2;
300     s->max_b_frames= avctx->max_b_frames;
301     s->codec_id= avctx->codec->id;
302     s->luma_elim_threshold  = avctx->luma_elim_threshold;
303     s->chroma_elim_threshold= avctx->chroma_elim_threshold;
304     s->strict_std_compliance= avctx->strict_std_compliance;
305     s->data_partitioning= avctx->flags & CODEC_FLAG_PART;
306     s->quarter_sample= (avctx->flags & CODEC_FLAG_QPEL)!=0;
307     s->mpeg_quant= avctx->mpeg_quant;
308     s->rtp_mode= !!avctx->rtp_payload_size;
309     s->intra_dc_precision= avctx->intra_dc_precision;
310     s->user_specified_pts = AV_NOPTS_VALUE;
311
312     if (s->gop_size <= 1) {
313         s->intra_only = 1;
314         s->gop_size = 12;
315     } else {
316         s->intra_only = 0;
317     }
318
319     s->me_method = avctx->me_method;
320
321     /* Fixed QSCALE */
322     s->fixed_qscale = !!(avctx->flags & CODEC_FLAG_QSCALE);
323
324     s->adaptive_quant= (   s->avctx->lumi_masking
325                         || s->avctx->dark_masking
326                         || s->avctx->temporal_cplx_masking
327                         || s->avctx->spatial_cplx_masking
328                         || s->avctx->p_masking
329                         || s->avctx->border_masking
330                         || (s->flags&CODEC_FLAG_QP_RD))
331                        && !s->fixed_qscale;
332
333     s->obmc= !!(s->flags & CODEC_FLAG_OBMC);
334     s->loop_filter= !!(s->flags & CODEC_FLAG_LOOP_FILTER);
335     s->alternate_scan= !!(s->flags & CODEC_FLAG_ALT_SCAN);
336     s->intra_vlc_format= !!(s->flags2 & CODEC_FLAG2_INTRA_VLC);
337     s->q_scale_type= !!(s->flags2 & CODEC_FLAG2_NON_LINEAR_QUANT);
338
339     if(avctx->rc_max_rate && !avctx->rc_buffer_size){
340         av_log(avctx, AV_LOG_ERROR, "a vbv buffer size is needed, for encoding with a maximum bitrate\n");
341         return -1;
342     }
343
344     if(avctx->rc_min_rate && avctx->rc_max_rate != avctx->rc_min_rate){
345         av_log(avctx, AV_LOG_INFO, "Warning min_rate > 0 but min_rate != max_rate isn't recommended!\n");
346     }
347
348     if(avctx->rc_min_rate && avctx->rc_min_rate > avctx->bit_rate){
349         av_log(avctx, AV_LOG_ERROR, "bitrate below min bitrate\n");
350         return -1;
351     }
352
353     if(avctx->rc_max_rate && avctx->rc_max_rate < avctx->bit_rate){
354         av_log(avctx, AV_LOG_INFO, "bitrate above max bitrate\n");
355         return -1;
356     }
357
358     if(avctx->rc_max_rate && avctx->rc_max_rate == avctx->bit_rate && avctx->rc_max_rate != avctx->rc_min_rate){
359         av_log(avctx, AV_LOG_INFO, "impossible bitrate constraints, this will fail\n");
360     }
361
362     if(avctx->rc_buffer_size && avctx->bit_rate*(int64_t)avctx->time_base.num > avctx->rc_buffer_size * (int64_t)avctx->time_base.den){
363         av_log(avctx, AV_LOG_ERROR, "VBV buffer too small for bitrate\n");
364         return -1;
365     }
366
367     if(!s->fixed_qscale && avctx->bit_rate*av_q2d(avctx->time_base) > avctx->bit_rate_tolerance){
368         av_log(avctx, AV_LOG_ERROR, "bitrate tolerance too small for bitrate\n");
369         return -1;
370     }
371
372     if(   s->avctx->rc_max_rate && s->avctx->rc_min_rate == s->avctx->rc_max_rate
373        && (s->codec_id == CODEC_ID_MPEG1VIDEO || s->codec_id == CODEC_ID_MPEG2VIDEO)
374        && 90000LL * (avctx->rc_buffer_size-1) > s->avctx->rc_max_rate*0xFFFFLL){
375
376         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");
377     }
378
379     if((s->flags & CODEC_FLAG_4MV) && s->codec_id != CODEC_ID_MPEG4
380        && s->codec_id != CODEC_ID_H263 && s->codec_id != CODEC_ID_H263P && s->codec_id != CODEC_ID_FLV1){
381         av_log(avctx, AV_LOG_ERROR, "4MV not supported by codec\n");
382         return -1;
383     }
384
385     if(s->obmc && s->avctx->mb_decision != FF_MB_DECISION_SIMPLE){
386         av_log(avctx, AV_LOG_ERROR, "OBMC is only supported with simple mb decision\n");
387         return -1;
388     }
389
390     if(s->obmc && s->codec_id != CODEC_ID_H263 && s->codec_id != CODEC_ID_H263P){
391         av_log(avctx, AV_LOG_ERROR, "OBMC is only supported with H263(+)\n");
392         return -1;
393     }
394
395     if(s->quarter_sample && s->codec_id != CODEC_ID_MPEG4){
396         av_log(avctx, AV_LOG_ERROR, "qpel not supported by codec\n");
397         return -1;
398     }
399
400     if(s->data_partitioning && s->codec_id != CODEC_ID_MPEG4){
401         av_log(avctx, AV_LOG_ERROR, "data partitioning not supported by codec\n");
402         return -1;
403     }
404
405     if(s->max_b_frames && s->codec_id != CODEC_ID_MPEG4 && s->codec_id != CODEC_ID_MPEG1VIDEO && s->codec_id != CODEC_ID_MPEG2VIDEO){
406         av_log(avctx, AV_LOG_ERROR, "b frames not supported by codec\n");
407         return -1;
408     }
409
410     if ((s->codec_id == CODEC_ID_MPEG4 || s->codec_id == CODEC_ID_H263 ||
411          s->codec_id == CODEC_ID_H263P) &&
412         (avctx->sample_aspect_ratio.num > 255 || avctx->sample_aspect_ratio.den > 255)) {
413         av_log(avctx, AV_LOG_ERROR, "Invalid pixel aspect ratio %i/%i, limit is 255/255\n",
414                avctx->sample_aspect_ratio.num, avctx->sample_aspect_ratio.den);
415         return -1;
416     }
417
418     if((s->flags & (CODEC_FLAG_INTERLACED_DCT|CODEC_FLAG_INTERLACED_ME|CODEC_FLAG_ALT_SCAN))
419        && s->codec_id != CODEC_ID_MPEG4 && s->codec_id != CODEC_ID_MPEG2VIDEO){
420         av_log(avctx, AV_LOG_ERROR, "interlacing not supported by codec\n");
421         return -1;
422     }
423
424     if(s->mpeg_quant && s->codec_id != CODEC_ID_MPEG4){ //FIXME mpeg2 uses that too
425         av_log(avctx, AV_LOG_ERROR, "mpeg2 style quantization not supported by codec\n");
426         return -1;
427     }
428
429     if((s->flags & CODEC_FLAG_CBP_RD) && !avctx->trellis){
430         av_log(avctx, AV_LOG_ERROR, "CBP RD needs trellis quant\n");
431         return -1;
432     }
433
434     if((s->flags & CODEC_FLAG_QP_RD) && s->avctx->mb_decision != FF_MB_DECISION_RD){
435         av_log(avctx, AV_LOG_ERROR, "QP RD needs mbd=2\n");
436         return -1;
437     }
438
439     if(s->avctx->scenechange_threshold < 1000000000 && (s->flags & CODEC_FLAG_CLOSED_GOP)){
440         av_log(avctx, AV_LOG_ERROR, "closed gop with scene change detection are not supported yet, set threshold to 1000000000\n");
441         return -1;
442     }
443
444     if((s->flags2 & CODEC_FLAG2_INTRA_VLC) && s->codec_id != CODEC_ID_MPEG2VIDEO){
445         av_log(avctx, AV_LOG_ERROR, "intra vlc table not supported by codec\n");
446         return -1;
447     }
448
449     if(s->flags & CODEC_FLAG_LOW_DELAY){
450         if (s->codec_id != CODEC_ID_MPEG2VIDEO){
451             av_log(avctx, AV_LOG_ERROR, "low delay forcing is only available for mpeg2\n");
452             return -1;
453         }
454         if (s->max_b_frames != 0){
455             av_log(avctx, AV_LOG_ERROR, "b frames cannot be used with low delay\n");
456             return -1;
457         }
458     }
459
460     if(s->q_scale_type == 1){
461         if(s->codec_id != CODEC_ID_MPEG2VIDEO){
462             av_log(avctx, AV_LOG_ERROR, "non linear quant is only available for mpeg2\n");
463             return -1;
464         }
465         if(avctx->qmax > 12){
466             av_log(avctx, AV_LOG_ERROR, "non linear quant only supports qmax <= 12 currently\n");
467             return -1;
468         }
469     }
470
471     if(s->avctx->thread_count > 1 && s->codec_id != CODEC_ID_MPEG4
472        && s->codec_id != CODEC_ID_MPEG1VIDEO && s->codec_id != CODEC_ID_MPEG2VIDEO
473        && (s->codec_id != CODEC_ID_H263P || !(s->flags & CODEC_FLAG_H263P_SLICE_STRUCT))){
474         av_log(avctx, AV_LOG_ERROR, "multi threaded encoding not supported by codec\n");
475         return -1;
476     }
477
478     if(s->avctx->thread_count < 1){
479         av_log(avctx, AV_LOG_ERROR, "automatic thread number detection not supported by codec, patch welcome\n");
480         return -1;
481     }
482
483     if(s->avctx->thread_count > 1)
484         s->rtp_mode= 1;
485
486     if(!avctx->time_base.den || !avctx->time_base.num){
487         av_log(avctx, AV_LOG_ERROR, "framerate not set\n");
488         return -1;
489     }
490
491     i= (INT_MAX/2+128)>>8;
492     if(avctx->me_threshold >= i){
493         av_log(avctx, AV_LOG_ERROR, "me_threshold too large, max is %d\n", i - 1);
494         return -1;
495     }
496     if(avctx->mb_threshold >= i){
497         av_log(avctx, AV_LOG_ERROR, "mb_threshold too large, max is %d\n", i - 1);
498         return -1;
499     }
500
501     if(avctx->b_frame_strategy && (avctx->flags&CODEC_FLAG_PASS2)){
502         av_log(avctx, AV_LOG_INFO, "notice: b_frame_strategy only affects the first pass\n");
503         avctx->b_frame_strategy = 0;
504     }
505
506     i= av_gcd(avctx->time_base.den, avctx->time_base.num);
507     if(i > 1){
508         av_log(avctx, AV_LOG_INFO, "removing common factors from framerate\n");
509         avctx->time_base.den /= i;
510         avctx->time_base.num /= i;
511 //        return -1;
512     }
513
514     if(s->mpeg_quant || s->codec_id==CODEC_ID_MPEG1VIDEO || s->codec_id==CODEC_ID_MPEG2VIDEO || s->codec_id==CODEC_ID_MJPEG){
515         s->intra_quant_bias= 3<<(QUANT_BIAS_SHIFT-3); //(a + x*3/8)/x
516         s->inter_quant_bias= 0;
517     }else{
518         s->intra_quant_bias=0;
519         s->inter_quant_bias=-(1<<(QUANT_BIAS_SHIFT-2)); //(a - x/4)/x
520     }
521
522     if(avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)
523         s->intra_quant_bias= avctx->intra_quant_bias;
524     if(avctx->inter_quant_bias != FF_DEFAULT_QUANT_BIAS)
525         s->inter_quant_bias= avctx->inter_quant_bias;
526
527     avcodec_get_chroma_sub_sample(avctx->pix_fmt, &chroma_h_shift, &chroma_v_shift);
528
529     if(avctx->codec_id == CODEC_ID_MPEG4 && s->avctx->time_base.den > (1<<16)-1){
530         av_log(avctx, AV_LOG_ERROR, "timebase %d/%d not supported by MPEG 4 standard, "
531                "the maximum admitted value for the timebase denominator is %d\n",
532                s->avctx->time_base.num, s->avctx->time_base.den, (1<<16)-1);
533         return -1;
534     }
535     s->time_increment_bits = av_log2(s->avctx->time_base.den - 1) + 1;
536
537     switch(avctx->codec->id) {
538     case CODEC_ID_MPEG1VIDEO:
539         s->out_format = FMT_MPEG1;
540         s->low_delay= !!(s->flags & CODEC_FLAG_LOW_DELAY);
541         avctx->delay= s->low_delay ? 0 : (s->max_b_frames + 1);
542         break;
543     case CODEC_ID_MPEG2VIDEO:
544         s->out_format = FMT_MPEG1;
545         s->low_delay= !!(s->flags & CODEC_FLAG_LOW_DELAY);
546         avctx->delay= s->low_delay ? 0 : (s->max_b_frames + 1);
547         s->rtp_mode= 1;
548         break;
549     case CODEC_ID_LJPEG:
550     case CODEC_ID_MJPEG:
551         s->out_format = FMT_MJPEG;
552         s->intra_only = 1; /* force intra only for jpeg */
553         if(avctx->codec->id == CODEC_ID_LJPEG && avctx->pix_fmt == PIX_FMT_BGRA){
554             s->mjpeg_vsample[0] = s->mjpeg_hsample[0] =
555             s->mjpeg_vsample[1] = s->mjpeg_hsample[1] =
556             s->mjpeg_vsample[2] = s->mjpeg_hsample[2] = 1;
557         }else{
558             s->mjpeg_vsample[0] = 2;
559             s->mjpeg_vsample[1] = 2>>chroma_v_shift;
560             s->mjpeg_vsample[2] = 2>>chroma_v_shift;
561             s->mjpeg_hsample[0] = 2;
562             s->mjpeg_hsample[1] = 2>>chroma_h_shift;
563             s->mjpeg_hsample[2] = 2>>chroma_h_shift;
564         }
565         if (!(CONFIG_MJPEG_ENCODER || CONFIG_LJPEG_ENCODER)
566             || ff_mjpeg_encode_init(s) < 0)
567             return -1;
568         avctx->delay=0;
569         s->low_delay=1;
570         break;
571     case CODEC_ID_H261:
572         if (!CONFIG_H261_ENCODER)  return -1;
573         if (ff_h261_get_picture_format(s->width, s->height) < 0) {
574             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);
575             return -1;
576         }
577         s->out_format = FMT_H261;
578         avctx->delay=0;
579         s->low_delay=1;
580         break;
581     case CODEC_ID_H263:
582         if (!CONFIG_H263_ENCODER)  return -1;
583         if (ff_match_2uint16(h263_format, FF_ARRAY_ELEMS(h263_format), s->width, s->height) == 8) {
584             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);
585             return -1;
586         }
587         s->out_format = FMT_H263;
588         s->obmc= (avctx->flags & CODEC_FLAG_OBMC) ? 1:0;
589         avctx->delay=0;
590         s->low_delay=1;
591         break;
592     case CODEC_ID_H263P:
593         s->out_format = FMT_H263;
594         s->h263_plus = 1;
595         /* Fx */
596         s->umvplus = (avctx->flags & CODEC_FLAG_H263P_UMV) ? 1:0;
597         s->h263_aic= (avctx->flags & CODEC_FLAG_AC_PRED) ? 1:0;
598         s->modified_quant= s->h263_aic;
599         s->alt_inter_vlc= (avctx->flags & CODEC_FLAG_H263P_AIV) ? 1:0;
600         s->obmc= (avctx->flags & CODEC_FLAG_OBMC) ? 1:0;
601         s->loop_filter= (avctx->flags & CODEC_FLAG_LOOP_FILTER) ? 1:0;
602         s->unrestricted_mv= s->obmc || s->loop_filter || s->umvplus;
603         s->h263_slice_structured= (s->flags & CODEC_FLAG_H263P_SLICE_STRUCT) ? 1:0;
604
605         /* /Fx */
606         /* These are just to be sure */
607         avctx->delay=0;
608         s->low_delay=1;
609         break;
610     case CODEC_ID_FLV1:
611         s->out_format = FMT_H263;
612         s->h263_flv = 2; /* format = 1; 11-bit codes */
613         s->unrestricted_mv = 1;
614         s->rtp_mode=0; /* don't allow GOB */
615         avctx->delay=0;
616         s->low_delay=1;
617         break;
618     case CODEC_ID_RV10:
619         s->out_format = FMT_H263;
620         avctx->delay=0;
621         s->low_delay=1;
622         break;
623     case CODEC_ID_RV20:
624         s->out_format = FMT_H263;
625         avctx->delay=0;
626         s->low_delay=1;
627         s->modified_quant=1;
628         s->h263_aic=1;
629         s->h263_plus=1;
630         s->loop_filter=1;
631         s->unrestricted_mv= 0;
632         break;
633     case CODEC_ID_MPEG4:
634         s->out_format = FMT_H263;
635         s->h263_pred = 1;
636         s->unrestricted_mv = 1;
637         s->low_delay= s->max_b_frames ? 0 : 1;
638         avctx->delay= s->low_delay ? 0 : (s->max_b_frames + 1);
639         break;
640     case CODEC_ID_MSMPEG4V1:
641         s->out_format = FMT_H263;
642         s->h263_msmpeg4 = 1;
643         s->h263_pred = 1;
644         s->unrestricted_mv = 1;
645         s->msmpeg4_version= 1;
646         avctx->delay=0;
647         s->low_delay=1;
648         break;
649     case CODEC_ID_MSMPEG4V2:
650         s->out_format = FMT_H263;
651         s->h263_msmpeg4 = 1;
652         s->h263_pred = 1;
653         s->unrestricted_mv = 1;
654         s->msmpeg4_version= 2;
655         avctx->delay=0;
656         s->low_delay=1;
657         break;
658     case CODEC_ID_MSMPEG4V3:
659         s->out_format = FMT_H263;
660         s->h263_msmpeg4 = 1;
661         s->h263_pred = 1;
662         s->unrestricted_mv = 1;
663         s->msmpeg4_version= 3;
664         s->flipflop_rounding=1;
665         avctx->delay=0;
666         s->low_delay=1;
667         break;
668     case CODEC_ID_WMV1:
669         s->out_format = FMT_H263;
670         s->h263_msmpeg4 = 1;
671         s->h263_pred = 1;
672         s->unrestricted_mv = 1;
673         s->msmpeg4_version= 4;
674         s->flipflop_rounding=1;
675         avctx->delay=0;
676         s->low_delay=1;
677         break;
678     case CODEC_ID_WMV2:
679         s->out_format = FMT_H263;
680         s->h263_msmpeg4 = 1;
681         s->h263_pred = 1;
682         s->unrestricted_mv = 1;
683         s->msmpeg4_version= 5;
684         s->flipflop_rounding=1;
685         avctx->delay=0;
686         s->low_delay=1;
687         break;
688     default:
689         return -1;
690     }
691
692     avctx->has_b_frames= !s->low_delay;
693
694     s->encoding = 1;
695
696     s->progressive_frame=
697     s->progressive_sequence= !(avctx->flags & (CODEC_FLAG_INTERLACED_DCT|CODEC_FLAG_INTERLACED_ME|CODEC_FLAG_ALT_SCAN));
698
699     /* init */
700     if (MPV_common_init(s) < 0)
701         return -1;
702
703     if(!s->dct_quantize)
704         s->dct_quantize = dct_quantize_c;
705     if(!s->denoise_dct)
706         s->denoise_dct = denoise_dct_c;
707     s->fast_dct_quantize = s->dct_quantize;
708     if(avctx->trellis)
709         s->dct_quantize = dct_quantize_trellis_c;
710
711     if((CONFIG_H263P_ENCODER || CONFIG_RV20_ENCODER) && s->modified_quant)
712         s->chroma_qscale_table= ff_h263_chroma_qscale_table;
713
714     s->quant_precision=5;
715
716     ff_set_cmp(&s->dsp, s->dsp.ildct_cmp, s->avctx->ildct_cmp);
717     ff_set_cmp(&s->dsp, s->dsp.frame_skip_cmp, s->avctx->frame_skip_cmp);
718
719     if (CONFIG_H261_ENCODER && s->out_format == FMT_H261)
720         ff_h261_encode_init(s);
721     if (CONFIG_H263_ENCODER && s->out_format == FMT_H263)
722         h263_encode_init(s);
723     if (CONFIG_MSMPEG4_ENCODER && s->msmpeg4_version)
724         ff_msmpeg4_encode_init(s);
725     if ((CONFIG_MPEG1VIDEO_ENCODER || CONFIG_MPEG2VIDEO_ENCODER)
726         && s->out_format == FMT_MPEG1)
727         ff_mpeg1_encode_init(s);
728
729     /* init q matrix */
730     for(i=0;i<64;i++) {
731         int j= s->dsp.idct_permutation[i];
732         if(CONFIG_MPEG4_ENCODER && s->codec_id==CODEC_ID_MPEG4 && s->mpeg_quant){
733             s->intra_matrix[j] = ff_mpeg4_default_intra_matrix[i];
734             s->inter_matrix[j] = ff_mpeg4_default_non_intra_matrix[i];
735         }else if(s->out_format == FMT_H263 || s->out_format == FMT_H261){
736             s->intra_matrix[j] =
737             s->inter_matrix[j] = ff_mpeg1_default_non_intra_matrix[i];
738         }else
739         { /* mpeg1/2 */
740             s->intra_matrix[j] = ff_mpeg1_default_intra_matrix[i];
741             s->inter_matrix[j] = ff_mpeg1_default_non_intra_matrix[i];
742         }
743         if(s->avctx->intra_matrix)
744             s->intra_matrix[j] = s->avctx->intra_matrix[i];
745         if(s->avctx->inter_matrix)
746             s->inter_matrix[j] = s->avctx->inter_matrix[i];
747     }
748
749     /* precompute matrix */
750     /* for mjpeg, we do include qscale in the matrix */
751     if (s->out_format != FMT_MJPEG) {
752         ff_convert_matrix(&s->dsp, s->q_intra_matrix, s->q_intra_matrix16,
753                        s->intra_matrix, s->intra_quant_bias, avctx->qmin, 31, 1);
754         ff_convert_matrix(&s->dsp, s->q_inter_matrix, s->q_inter_matrix16,
755                        s->inter_matrix, s->inter_quant_bias, avctx->qmin, 31, 0);
756     }
757
758     if(ff_rate_control_init(s) < 0)
759         return -1;
760
761     return 0;
762 }
763
764 av_cold int MPV_encode_end(AVCodecContext *avctx)
765 {
766     MpegEncContext *s = avctx->priv_data;
767
768     ff_rate_control_uninit(s);
769
770     MPV_common_end(s);
771     if ((CONFIG_MJPEG_ENCODER || CONFIG_LJPEG_ENCODER) && s->out_format == FMT_MJPEG)
772         ff_mjpeg_encode_close(s);
773
774     av_freep(&avctx->extradata);
775
776     return 0;
777 }
778
779 static int get_sae(uint8_t *src, int ref, int stride){
780     int x,y;
781     int acc=0;
782
783     for(y=0; y<16; y++){
784         for(x=0; x<16; x++){
785             acc+= FFABS(src[x+y*stride] - ref);
786         }
787     }
788
789     return acc;
790 }
791
792 static int get_intra_count(MpegEncContext *s, uint8_t *src, uint8_t *ref, int stride){
793     int x, y, w, h;
794     int acc=0;
795
796     w= s->width &~15;
797     h= s->height&~15;
798
799     for(y=0; y<h; y+=16){
800         for(x=0; x<w; x+=16){
801             int offset= x + y*stride;
802             int sad = s->dsp.sad[0](NULL, src + offset, ref + offset, stride, 16);
803             int mean= (s->dsp.pix_sum(src + offset, stride) + 128)>>8;
804             int sae = get_sae(src + offset, mean, stride);
805
806             acc+= sae + 500 < sad;
807         }
808     }
809     return acc;
810 }
811
812
813 static int load_input_picture(MpegEncContext *s, AVFrame *pic_arg){
814     AVFrame *pic=NULL;
815     int64_t pts;
816     int i;
817     const int encoding_delay= s->max_b_frames;
818     int direct=1;
819
820     if(pic_arg){
821         pts= pic_arg->pts;
822         pic_arg->display_picture_number= s->input_picture_number++;
823
824         if(pts != AV_NOPTS_VALUE){
825             if(s->user_specified_pts != AV_NOPTS_VALUE){
826                 int64_t time= pts;
827                 int64_t last= s->user_specified_pts;
828
829                 if(time <= last){
830                     av_log(s->avctx, AV_LOG_ERROR, "Error, Invalid timestamp=%"PRId64", last=%"PRId64"\n", pts, s->user_specified_pts);
831                     return -1;
832                 }
833             }
834             s->user_specified_pts= pts;
835         }else{
836             if(s->user_specified_pts != AV_NOPTS_VALUE){
837                 s->user_specified_pts=
838                 pts= s->user_specified_pts + 1;
839                 av_log(s->avctx, AV_LOG_INFO, "Warning: AVFrame.pts=? trying to guess (%"PRId64")\n", pts);
840             }else{
841                 pts= pic_arg->display_picture_number;
842             }
843         }
844     }
845
846   if(pic_arg){
847     if(encoding_delay && !(s->flags&CODEC_FLAG_INPUT_PRESERVED)) direct=0;
848     if(pic_arg->linesize[0] != s->linesize) direct=0;
849     if(pic_arg->linesize[1] != s->uvlinesize) direct=0;
850     if(pic_arg->linesize[2] != s->uvlinesize) direct=0;
851
852 //    av_log(AV_LOG_DEBUG, "%d %d %d %d\n",pic_arg->linesize[0], pic_arg->linesize[1], s->linesize, s->uvlinesize);
853
854     if(direct){
855         i= ff_find_unused_picture(s, 1);
856
857         pic= (AVFrame*)&s->picture[i];
858         pic->reference= 3;
859
860         for(i=0; i<4; i++){
861             pic->data[i]= pic_arg->data[i];
862             pic->linesize[i]= pic_arg->linesize[i];
863         }
864         if(ff_alloc_picture(s, (Picture*)pic, 1) < 0){
865             return -1;
866         }
867     }else{
868         i= ff_find_unused_picture(s, 0);
869
870         pic= (AVFrame*)&s->picture[i];
871         pic->reference= 3;
872
873         if(ff_alloc_picture(s, (Picture*)pic, 0) < 0){
874             return -1;
875         }
876
877         if(   pic->data[0] + INPLACE_OFFSET == pic_arg->data[0]
878            && pic->data[1] + INPLACE_OFFSET == pic_arg->data[1]
879            && pic->data[2] + INPLACE_OFFSET == pic_arg->data[2]){
880        // empty
881         }else{
882             int h_chroma_shift, v_chroma_shift;
883             avcodec_get_chroma_sub_sample(s->avctx->pix_fmt, &h_chroma_shift, &v_chroma_shift);
884
885             for(i=0; i<3; i++){
886                 int src_stride= pic_arg->linesize[i];
887                 int dst_stride= i ? s->uvlinesize : s->linesize;
888                 int h_shift= i ? h_chroma_shift : 0;
889                 int v_shift= i ? v_chroma_shift : 0;
890                 int w= s->width >>h_shift;
891                 int h= s->height>>v_shift;
892                 uint8_t *src= pic_arg->data[i];
893                 uint8_t *dst= pic->data[i];
894
895                 if(!s->avctx->rc_buffer_size)
896                     dst +=INPLACE_OFFSET;
897
898                 if(src_stride==dst_stride)
899                     memcpy(dst, src, src_stride*h);
900                 else{
901                     while(h--){
902                         memcpy(dst, src, w);
903                         dst += dst_stride;
904                         src += src_stride;
905                     }
906                 }
907             }
908         }
909     }
910     copy_picture_attributes(s, pic, pic_arg);
911     pic->pts= pts; //we set this here to avoid modifiying pic_arg
912   }
913
914     /* shift buffer entries */
915     for(i=1; i<MAX_PICTURE_COUNT /*s->encoding_delay+1*/; i++)
916         s->input_picture[i-1]= s->input_picture[i];
917
918     s->input_picture[encoding_delay]= (Picture*)pic;
919
920     return 0;
921 }
922
923 static int skip_check(MpegEncContext *s, Picture *p, Picture *ref){
924     int x, y, plane;
925     int score=0;
926     int64_t score64=0;
927
928     for(plane=0; plane<3; plane++){
929         const int stride= p->linesize[plane];
930         const int bw= plane ? 1 : 2;
931         for(y=0; y<s->mb_height*bw; y++){
932             for(x=0; x<s->mb_width*bw; x++){
933                 int off= p->type == FF_BUFFER_TYPE_SHARED ? 0: 16;
934                 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);
935
936                 switch(s->avctx->frame_skip_exp){
937                     case 0: score= FFMAX(score, v); break;
938                     case 1: score+= FFABS(v);break;
939                     case 2: score+= v*v;break;
940                     case 3: score64+= FFABS(v*v*(int64_t)v);break;
941                     case 4: score64+= v*v*(int64_t)(v*v);break;
942                 }
943             }
944         }
945     }
946
947     if(score) score64= score;
948
949     if(score64 < s->avctx->frame_skip_threshold)
950         return 1;
951     if(score64 < ((s->avctx->frame_skip_factor * (int64_t)s->lambda)>>8))
952         return 1;
953     return 0;
954 }
955
956 static int estimate_best_b_count(MpegEncContext *s){
957     AVCodec *codec= avcodec_find_encoder(s->avctx->codec_id);
958     AVCodecContext *c= avcodec_alloc_context();
959     AVFrame input[FF_MAX_B_FRAMES+2];
960     const int scale= s->avctx->brd_scale;
961     int i, j, out_size, p_lambda, b_lambda, lambda2;
962     int outbuf_size= s->width * s->height; //FIXME
963     uint8_t *outbuf= av_malloc(outbuf_size);
964     int64_t best_rd= INT64_MAX;
965     int best_b_count= -1;
966
967     assert(scale>=0 && scale <=3);
968
969 //    emms_c();
970     p_lambda= s->last_lambda_for[FF_P_TYPE]; //s->next_picture_ptr->quality;
971     b_lambda= s->last_lambda_for[FF_B_TYPE]; //p_lambda *FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset;
972     if(!b_lambda) b_lambda= p_lambda; //FIXME we should do this somewhere else
973     lambda2= (b_lambda*b_lambda + (1<<FF_LAMBDA_SHIFT)/2 ) >> FF_LAMBDA_SHIFT;
974
975     c->width = s->width >> scale;
976     c->height= s->height>> scale;
977     c->flags= CODEC_FLAG_QSCALE | CODEC_FLAG_PSNR | CODEC_FLAG_INPUT_PRESERVED /*| CODEC_FLAG_EMU_EDGE*/;
978     c->flags|= s->avctx->flags & CODEC_FLAG_QPEL;
979     c->mb_decision= s->avctx->mb_decision;
980     c->me_cmp= s->avctx->me_cmp;
981     c->mb_cmp= s->avctx->mb_cmp;
982     c->me_sub_cmp= s->avctx->me_sub_cmp;
983     c->pix_fmt = PIX_FMT_YUV420P;
984     c->time_base= s->avctx->time_base;
985     c->max_b_frames= s->max_b_frames;
986
987     if (avcodec_open(c, codec) < 0)
988         return -1;
989
990     for(i=0; i<s->max_b_frames+2; i++){
991         int ysize= c->width*c->height;
992         int csize= (c->width/2)*(c->height/2);
993         Picture pre_input, *pre_input_ptr= i ? s->input_picture[i-1] : s->next_picture_ptr;
994
995         avcodec_get_frame_defaults(&input[i]);
996         input[i].data[0]= av_malloc(ysize + 2*csize);
997         input[i].data[1]= input[i].data[0] + ysize;
998         input[i].data[2]= input[i].data[1] + csize;
999         input[i].linesize[0]= c->width;
1000         input[i].linesize[1]=
1001         input[i].linesize[2]= c->width/2;
1002
1003         if(pre_input_ptr && (!i || s->input_picture[i-1])) {
1004             pre_input= *pre_input_ptr;
1005
1006             if(pre_input.type != FF_BUFFER_TYPE_SHARED && i) {
1007                 pre_input.data[0]+=INPLACE_OFFSET;
1008                 pre_input.data[1]+=INPLACE_OFFSET;
1009                 pre_input.data[2]+=INPLACE_OFFSET;
1010             }
1011
1012             s->dsp.shrink[scale](input[i].data[0], input[i].linesize[0], pre_input.data[0], pre_input.linesize[0], c->width, c->height);
1013             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);
1014             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);
1015         }
1016     }
1017
1018     for(j=0; j<s->max_b_frames+1; j++){
1019         int64_t rd=0;
1020
1021         if(!s->input_picture[j])
1022             break;
1023
1024         c->error[0]= c->error[1]= c->error[2]= 0;
1025
1026         input[0].pict_type= FF_I_TYPE;
1027         input[0].quality= 1 * FF_QP2LAMBDA;
1028         out_size = avcodec_encode_video(c, outbuf, outbuf_size, &input[0]);
1029 //        rd += (out_size * lambda2) >> FF_LAMBDA_SHIFT;
1030
1031         for(i=0; i<s->max_b_frames+1; i++){
1032             int is_p= i % (j+1) == j || i==s->max_b_frames;
1033
1034             input[i+1].pict_type= is_p ? FF_P_TYPE : FF_B_TYPE;
1035             input[i+1].quality= is_p ? p_lambda : b_lambda;
1036             out_size = avcodec_encode_video(c, outbuf, outbuf_size, &input[i+1]);
1037             rd += (out_size * lambda2) >> (FF_LAMBDA_SHIFT - 3);
1038         }
1039
1040         /* get the delayed frames */
1041         while(out_size){
1042             out_size = avcodec_encode_video(c, outbuf, outbuf_size, NULL);
1043             rd += (out_size * lambda2) >> (FF_LAMBDA_SHIFT - 3);
1044         }
1045
1046         rd += c->error[0] + c->error[1] + c->error[2];
1047
1048         if(rd < best_rd){
1049             best_rd= rd;
1050             best_b_count= j;
1051         }
1052     }
1053
1054     av_freep(&outbuf);
1055     avcodec_close(c);
1056     av_freep(&c);
1057
1058     for(i=0; i<s->max_b_frames+2; i++){
1059         av_freep(&input[i].data[0]);
1060     }
1061
1062     return best_b_count;
1063 }
1064
1065 static int select_input_picture(MpegEncContext *s){
1066     int i;
1067
1068     for(i=1; i<MAX_PICTURE_COUNT; i++)
1069         s->reordered_input_picture[i-1]= s->reordered_input_picture[i];
1070     s->reordered_input_picture[MAX_PICTURE_COUNT-1]= NULL;
1071
1072     /* set next picture type & ordering */
1073     if(s->reordered_input_picture[0]==NULL && s->input_picture[0]){
1074         if(/*s->picture_in_gop_number >= s->gop_size ||*/ s->next_picture_ptr==NULL || s->intra_only){
1075             s->reordered_input_picture[0]= s->input_picture[0];
1076             s->reordered_input_picture[0]->pict_type= FF_I_TYPE;
1077             s->reordered_input_picture[0]->coded_picture_number= s->coded_picture_number++;
1078         }else{
1079             int b_frames;
1080
1081             if(s->avctx->frame_skip_threshold || s->avctx->frame_skip_factor){
1082                 if(s->picture_in_gop_number < s->gop_size && skip_check(s, s->input_picture[0], s->next_picture_ptr)){
1083                 //FIXME check that te gop check above is +-1 correct
1084 //av_log(NULL, AV_LOG_DEBUG, "skip %p %"PRId64"\n", s->input_picture[0]->data[0], s->input_picture[0]->pts);
1085
1086                     if(s->input_picture[0]->type == FF_BUFFER_TYPE_SHARED){
1087                         for(i=0; i<4; i++)
1088                             s->input_picture[0]->data[i]= NULL;
1089                         s->input_picture[0]->type= 0;
1090                     }else{
1091                         assert(   s->input_picture[0]->type==FF_BUFFER_TYPE_USER
1092                                || s->input_picture[0]->type==FF_BUFFER_TYPE_INTERNAL);
1093
1094                         s->avctx->release_buffer(s->avctx, (AVFrame*)s->input_picture[0]);
1095                     }
1096
1097                     emms_c();
1098                     ff_vbv_update(s, 0);
1099
1100                     goto no_output_pic;
1101                 }
1102             }
1103
1104             if(s->flags&CODEC_FLAG_PASS2){
1105                 for(i=0; i<s->max_b_frames+1; i++){
1106                     int pict_num= s->input_picture[0]->display_picture_number + i;
1107
1108                     if(pict_num >= s->rc_context.num_entries)
1109                         break;
1110                     if(!s->input_picture[i]){
1111                         s->rc_context.entry[pict_num-1].new_pict_type = FF_P_TYPE;
1112                         break;
1113                     }
1114
1115                     s->input_picture[i]->pict_type=
1116                         s->rc_context.entry[pict_num].new_pict_type;
1117                 }
1118             }
1119
1120             if(s->avctx->b_frame_strategy==0){
1121                 b_frames= s->max_b_frames;
1122                 while(b_frames && !s->input_picture[b_frames]) b_frames--;
1123             }else if(s->avctx->b_frame_strategy==1){
1124                 for(i=1; i<s->max_b_frames+1; i++){
1125                     if(s->input_picture[i] && s->input_picture[i]->b_frame_score==0){
1126                         s->input_picture[i]->b_frame_score=
1127                             get_intra_count(s, s->input_picture[i  ]->data[0],
1128                                                s->input_picture[i-1]->data[0], s->linesize) + 1;
1129                     }
1130                 }
1131                 for(i=0; i<s->max_b_frames+1; i++){
1132                     if(s->input_picture[i]==NULL || s->input_picture[i]->b_frame_score - 1 > s->mb_num/s->avctx->b_sensitivity) break;
1133                 }
1134
1135                 b_frames= FFMAX(0, i-1);
1136
1137                 /* reset scores */
1138                 for(i=0; i<b_frames+1; i++){
1139                     s->input_picture[i]->b_frame_score=0;
1140                 }
1141             }else if(s->avctx->b_frame_strategy==2){
1142                 b_frames= estimate_best_b_count(s);
1143             }else{
1144                 av_log(s->avctx, AV_LOG_ERROR, "illegal b frame strategy\n");
1145                 b_frames=0;
1146             }
1147
1148             emms_c();
1149 //static int b_count=0;
1150 //b_count+= b_frames;
1151 //av_log(s->avctx, AV_LOG_DEBUG, "b_frames: %d\n", b_count);
1152
1153             for(i= b_frames - 1; i>=0; i--){
1154                 int type= s->input_picture[i]->pict_type;
1155                 if(type && type != FF_B_TYPE)
1156                     b_frames= i;
1157             }
1158             if(s->input_picture[b_frames]->pict_type == FF_B_TYPE && b_frames == s->max_b_frames){
1159                 av_log(s->avctx, AV_LOG_ERROR, "warning, too many b frames in a row\n");
1160             }
1161
1162             if(s->picture_in_gop_number + b_frames >= s->gop_size){
1163               if((s->flags2 & CODEC_FLAG2_STRICT_GOP) && s->gop_size > s->picture_in_gop_number){
1164                     b_frames= s->gop_size - s->picture_in_gop_number - 1;
1165               }else{
1166                 if(s->flags & CODEC_FLAG_CLOSED_GOP)
1167                     b_frames=0;
1168                 s->input_picture[b_frames]->pict_type= FF_I_TYPE;
1169               }
1170             }
1171
1172             if(   (s->flags & CODEC_FLAG_CLOSED_GOP)
1173                && b_frames
1174                && s->input_picture[b_frames]->pict_type== FF_I_TYPE)
1175                 b_frames--;
1176
1177             s->reordered_input_picture[0]= s->input_picture[b_frames];
1178             if(s->reordered_input_picture[0]->pict_type != FF_I_TYPE)
1179                 s->reordered_input_picture[0]->pict_type= FF_P_TYPE;
1180             s->reordered_input_picture[0]->coded_picture_number= s->coded_picture_number++;
1181             for(i=0; i<b_frames; i++){
1182                 s->reordered_input_picture[i+1]= s->input_picture[i];
1183                 s->reordered_input_picture[i+1]->pict_type= FF_B_TYPE;
1184                 s->reordered_input_picture[i+1]->coded_picture_number= s->coded_picture_number++;
1185             }
1186         }
1187     }
1188 no_output_pic:
1189     if(s->reordered_input_picture[0]){
1190         s->reordered_input_picture[0]->reference= s->reordered_input_picture[0]->pict_type!=FF_B_TYPE ? 3 : 0;
1191
1192         ff_copy_picture(&s->new_picture, s->reordered_input_picture[0]);
1193
1194         if(s->reordered_input_picture[0]->type == FF_BUFFER_TYPE_SHARED || s->avctx->rc_buffer_size){
1195             // input is a shared pix, so we can't modifiy it -> alloc a new one & ensure that the shared one is reuseable
1196
1197             int i= ff_find_unused_picture(s, 0);
1198             Picture *pic= &s->picture[i];
1199
1200             pic->reference              = s->reordered_input_picture[0]->reference;
1201             if(ff_alloc_picture(s, pic, 0) < 0){
1202                 return -1;
1203             }
1204
1205             /* mark us unused / free shared pic */
1206             if(s->reordered_input_picture[0]->type == FF_BUFFER_TYPE_INTERNAL)
1207                 s->avctx->release_buffer(s->avctx, (AVFrame*)s->reordered_input_picture[0]);
1208             for(i=0; i<4; i++)
1209                 s->reordered_input_picture[0]->data[i]= NULL;
1210             s->reordered_input_picture[0]->type= 0;
1211
1212             copy_picture_attributes(s, (AVFrame*)pic, (AVFrame*)s->reordered_input_picture[0]);
1213
1214             s->current_picture_ptr= pic;
1215         }else{
1216             // input is not a shared pix -> reuse buffer for current_pix
1217
1218             assert(   s->reordered_input_picture[0]->type==FF_BUFFER_TYPE_USER
1219                    || s->reordered_input_picture[0]->type==FF_BUFFER_TYPE_INTERNAL);
1220
1221             s->current_picture_ptr= s->reordered_input_picture[0];
1222             for(i=0; i<4; i++){
1223                 s->new_picture.data[i]+= INPLACE_OFFSET;
1224             }
1225         }
1226         ff_copy_picture(&s->current_picture, s->current_picture_ptr);
1227
1228         s->picture_number= s->new_picture.display_picture_number;
1229 //printf("dpn:%d\n", s->picture_number);
1230     }else{
1231        memset(&s->new_picture, 0, sizeof(Picture));
1232     }
1233     return 0;
1234 }
1235
1236 int MPV_encode_picture(AVCodecContext *avctx,
1237                        unsigned char *buf, int buf_size, void *data)
1238 {
1239     MpegEncContext *s = avctx->priv_data;
1240     AVFrame *pic_arg = data;
1241     int i, stuffing_count;
1242
1243     for(i=0; i<avctx->thread_count; i++){
1244         int start_y= s->thread_context[i]->start_mb_y;
1245         int   end_y= s->thread_context[i]->  end_mb_y;
1246         int h= s->mb_height;
1247         uint8_t *start= buf + (size_t)(((int64_t) buf_size)*start_y/h);
1248         uint8_t *end  = buf + (size_t)(((int64_t) buf_size)*  end_y/h);
1249
1250         init_put_bits(&s->thread_context[i]->pb, start, end - start);
1251     }
1252
1253     s->picture_in_gop_number++;
1254
1255     if(load_input_picture(s, pic_arg) < 0)
1256         return -1;
1257
1258     if(select_input_picture(s) < 0){
1259         return -1;
1260     }
1261
1262     /* output? */
1263     if(s->new_picture.data[0]){
1264         s->pict_type= s->new_picture.pict_type;
1265 //emms_c();
1266 //printf("qs:%f %f %d\n", s->new_picture.quality, s->current_picture.quality, s->qscale);
1267         MPV_frame_start(s, avctx);
1268 vbv_retry:
1269         if (encode_picture(s, s->picture_number) < 0)
1270             return -1;
1271
1272         avctx->header_bits = s->header_bits;
1273         avctx->mv_bits     = s->mv_bits;
1274         avctx->misc_bits   = s->misc_bits;
1275         avctx->i_tex_bits  = s->i_tex_bits;
1276         avctx->p_tex_bits  = s->p_tex_bits;
1277         avctx->i_count     = s->i_count;
1278         avctx->p_count     = s->mb_num - s->i_count - s->skip_count; //FIXME f/b_count in avctx
1279         avctx->skip_count  = s->skip_count;
1280
1281         MPV_frame_end(s);
1282
1283         if (CONFIG_MJPEG_ENCODER && s->out_format == FMT_MJPEG)
1284             ff_mjpeg_encode_picture_trailer(s);
1285
1286         if(avctx->rc_buffer_size){
1287             RateControlContext *rcc= &s->rc_context;
1288             int max_size= rcc->buffer_index * avctx->rc_max_available_vbv_use;
1289
1290             if(put_bits_count(&s->pb) > max_size && s->lambda < s->avctx->lmax){
1291                 s->next_lambda= FFMAX(s->lambda+1, s->lambda*(s->qscale+1) / s->qscale);
1292                 if(s->adaptive_quant){
1293                     int i;
1294                     for(i=0; i<s->mb_height*s->mb_stride; i++)
1295                         s->lambda_table[i]= FFMAX(s->lambda_table[i]+1, s->lambda_table[i]*(s->qscale+1) / s->qscale);
1296                 }
1297                 s->mb_skipped = 0;        //done in MPV_frame_start()
1298                 if(s->pict_type==FF_P_TYPE){ //done in encode_picture() so we must undo it
1299                     if(s->flipflop_rounding || s->codec_id == CODEC_ID_H263P || s->codec_id == CODEC_ID_MPEG4)
1300                         s->no_rounding ^= 1;
1301                 }
1302                 if(s->pict_type!=FF_B_TYPE){
1303                     s->time_base= s->last_time_base;
1304                     s->last_non_b_time= s->time - s->pp_time;
1305                 }
1306 //                av_log(NULL, AV_LOG_ERROR, "R:%d ", s->next_lambda);
1307                 for(i=0; i<avctx->thread_count; i++){
1308                     PutBitContext *pb= &s->thread_context[i]->pb;
1309                     init_put_bits(pb, pb->buf, pb->buf_end - pb->buf);
1310                 }
1311                 goto vbv_retry;
1312             }
1313
1314             assert(s->avctx->rc_max_rate);
1315         }
1316
1317         if(s->flags&CODEC_FLAG_PASS1)
1318             ff_write_pass1_stats(s);
1319
1320         for(i=0; i<4; i++){
1321             s->current_picture_ptr->error[i]= s->current_picture.error[i];
1322             avctx->error[i] += s->current_picture_ptr->error[i];
1323         }
1324
1325         if(s->flags&CODEC_FLAG_PASS1)
1326             assert(avctx->header_bits + avctx->mv_bits + avctx->misc_bits + avctx->i_tex_bits + avctx->p_tex_bits == put_bits_count(&s->pb));
1327         flush_put_bits(&s->pb);
1328         s->frame_bits  = put_bits_count(&s->pb);
1329
1330         stuffing_count= ff_vbv_update(s, s->frame_bits);
1331         if(stuffing_count){
1332             if(s->pb.buf_end - s->pb.buf - (put_bits_count(&s->pb)>>3) < stuffing_count + 50){
1333                 av_log(s->avctx, AV_LOG_ERROR, "stuffing too large\n");
1334                 return -1;
1335             }
1336
1337             switch(s->codec_id){
1338             case CODEC_ID_MPEG1VIDEO:
1339             case CODEC_ID_MPEG2VIDEO:
1340                 while(stuffing_count--){
1341                     put_bits(&s->pb, 8, 0);
1342                 }
1343             break;
1344             case CODEC_ID_MPEG4:
1345                 put_bits(&s->pb, 16, 0);
1346                 put_bits(&s->pb, 16, 0x1C3);
1347                 stuffing_count -= 4;
1348                 while(stuffing_count--){
1349                     put_bits(&s->pb, 8, 0xFF);
1350                 }
1351             break;
1352             default:
1353                 av_log(s->avctx, AV_LOG_ERROR, "vbv buffer overflow\n");
1354             }
1355             flush_put_bits(&s->pb);
1356             s->frame_bits  = put_bits_count(&s->pb);
1357         }
1358
1359         /* update mpeg1/2 vbv_delay for CBR */
1360         if(s->avctx->rc_max_rate && s->avctx->rc_min_rate == s->avctx->rc_max_rate && s->out_format == FMT_MPEG1
1361            && 90000LL * (avctx->rc_buffer_size-1) <= s->avctx->rc_max_rate*0xFFFFLL){
1362             int vbv_delay, min_delay;
1363             double inbits = s->avctx->rc_max_rate*av_q2d(s->avctx->time_base);
1364             int    minbits= s->frame_bits - 8*(s->vbv_delay_ptr - s->pb.buf - 1);
1365             double bits   = s->rc_context.buffer_index + minbits - inbits;
1366
1367             if(bits<0)
1368                 av_log(s->avctx, AV_LOG_ERROR, "Internal error, negative bits\n");
1369
1370             assert(s->repeat_first_field==0);
1371
1372             vbv_delay=     bits * 90000                               / s->avctx->rc_max_rate;
1373             min_delay= (minbits * 90000LL + s->avctx->rc_max_rate - 1)/ s->avctx->rc_max_rate;
1374
1375             vbv_delay= FFMAX(vbv_delay, min_delay);
1376
1377             assert(vbv_delay < 0xFFFF);
1378
1379             s->vbv_delay_ptr[0] &= 0xF8;
1380             s->vbv_delay_ptr[0] |= vbv_delay>>13;
1381             s->vbv_delay_ptr[1]  = vbv_delay>>5;
1382             s->vbv_delay_ptr[2] &= 0x07;
1383             s->vbv_delay_ptr[2] |= vbv_delay<<3;
1384         }
1385         s->total_bits += s->frame_bits;
1386         avctx->frame_bits  = s->frame_bits;
1387     }else{
1388         assert((put_bits_ptr(&s->pb) == s->pb.buf));
1389         s->frame_bits=0;
1390     }
1391     assert((s->frame_bits&7)==0);
1392
1393     return s->frame_bits/8;
1394 }
1395
1396 static inline void dct_single_coeff_elimination(MpegEncContext *s, int n, int threshold)
1397 {
1398     static const char tab[64]=
1399         {3,2,2,1,1,1,1,1,
1400          1,1,1,1,1,1,1,1,
1401          1,1,1,1,1,1,1,1,
1402          0,0,0,0,0,0,0,0,
1403          0,0,0,0,0,0,0,0,
1404          0,0,0,0,0,0,0,0,
1405          0,0,0,0,0,0,0,0,
1406          0,0,0,0,0,0,0,0};
1407     int score=0;
1408     int run=0;
1409     int i;
1410     DCTELEM *block= s->block[n];
1411     const int last_index= s->block_last_index[n];
1412     int skip_dc;
1413
1414     if(threshold<0){
1415         skip_dc=0;
1416         threshold= -threshold;
1417     }else
1418         skip_dc=1;
1419
1420     /* Are all we could set to zero already zero? */
1421     if(last_index<=skip_dc - 1) return;
1422
1423     for(i=0; i<=last_index; i++){
1424         const int j = s->intra_scantable.permutated[i];
1425         const int level = FFABS(block[j]);
1426         if(level==1){
1427             if(skip_dc && i==0) continue;
1428             score+= tab[run];
1429             run=0;
1430         }else if(level>1){
1431             return;
1432         }else{
1433             run++;
1434         }
1435     }
1436     if(score >= threshold) return;
1437     for(i=skip_dc; i<=last_index; i++){
1438         const int j = s->intra_scantable.permutated[i];
1439         block[j]=0;
1440     }
1441     if(block[0]) s->block_last_index[n]= 0;
1442     else         s->block_last_index[n]= -1;
1443 }
1444
1445 static inline void clip_coeffs(MpegEncContext *s, DCTELEM *block, int last_index)
1446 {
1447     int i;
1448     const int maxlevel= s->max_qcoeff;
1449     const int minlevel= s->min_qcoeff;
1450     int overflow=0;
1451
1452     if(s->mb_intra){
1453         i=1; //skip clipping of intra dc
1454     }else
1455         i=0;
1456
1457     for(;i<=last_index; i++){
1458         const int j= s->intra_scantable.permutated[i];
1459         int level = block[j];
1460
1461         if     (level>maxlevel){
1462             level=maxlevel;
1463             overflow++;
1464         }else if(level<minlevel){
1465             level=minlevel;
1466             overflow++;
1467         }
1468
1469         block[j]= level;
1470     }
1471
1472     if(overflow && s->avctx->mb_decision == FF_MB_DECISION_SIMPLE)
1473         av_log(s->avctx, AV_LOG_INFO, "warning, clipping %d dct coefficients to %d..%d\n", overflow, minlevel, maxlevel);
1474 }
1475
1476 static void get_visual_weight(int16_t *weight, uint8_t *ptr, int stride){
1477     int x, y;
1478 //FIXME optimize
1479     for(y=0; y<8; y++){
1480         for(x=0; x<8; x++){
1481             int x2, y2;
1482             int sum=0;
1483             int sqr=0;
1484             int count=0;
1485
1486             for(y2= FFMAX(y-1, 0); y2 < FFMIN(8, y+2); y2++){
1487                 for(x2= FFMAX(x-1, 0); x2 < FFMIN(8, x+2); x2++){
1488                     int v= ptr[x2 + y2*stride];
1489                     sum += v;
1490                     sqr += v*v;
1491                     count++;
1492                 }
1493             }
1494             weight[x + 8*y]= (36*ff_sqrt(count*sqr - sum*sum)) / count;
1495         }
1496     }
1497 }
1498
1499 static av_always_inline void encode_mb_internal(MpegEncContext *s, int motion_x, int motion_y, int mb_block_height, int mb_block_count)
1500 {
1501     int16_t weight[8][64];
1502     DCTELEM orig[8][64];
1503     const int mb_x= s->mb_x;
1504     const int mb_y= s->mb_y;
1505     int i;
1506     int skip_dct[8];
1507     int dct_offset   = s->linesize*8; //default for progressive frames
1508     uint8_t *ptr_y, *ptr_cb, *ptr_cr;
1509     int wrap_y, wrap_c;
1510
1511     for(i=0; i<mb_block_count; i++) skip_dct[i]=s->skipdct;
1512
1513     if(s->adaptive_quant){
1514         const int last_qp= s->qscale;
1515         const int mb_xy= mb_x + mb_y*s->mb_stride;
1516
1517         s->lambda= s->lambda_table[mb_xy];
1518         update_qscale(s);
1519
1520         if(!(s->flags&CODEC_FLAG_QP_RD)){
1521             s->qscale= s->current_picture_ptr->qscale_table[mb_xy];
1522             s->dquant= s->qscale - last_qp;
1523
1524             if(s->out_format==FMT_H263){
1525                 s->dquant= av_clip(s->dquant, -2, 2);
1526
1527                 if(s->codec_id==CODEC_ID_MPEG4){
1528                     if(!s->mb_intra){
1529                         if(s->pict_type == FF_B_TYPE){
1530                             if(s->dquant&1 || s->mv_dir&MV_DIRECT)
1531                                 s->dquant= 0;
1532                         }
1533                         if(s->mv_type==MV_TYPE_8X8)
1534                             s->dquant=0;
1535                     }
1536                 }
1537             }
1538         }
1539         ff_set_qscale(s, last_qp + s->dquant);
1540     }else if(s->flags&CODEC_FLAG_QP_RD)
1541         ff_set_qscale(s, s->qscale + s->dquant);
1542
1543     wrap_y = s->linesize;
1544     wrap_c = s->uvlinesize;
1545     ptr_y = s->new_picture.data[0] + (mb_y * 16 * wrap_y) + mb_x * 16;
1546     ptr_cb = s->new_picture.data[1] + (mb_y * mb_block_height * wrap_c) + mb_x * 8;
1547     ptr_cr = s->new_picture.data[2] + (mb_y * mb_block_height * wrap_c) + mb_x * 8;
1548
1549     if(mb_x*16+16 > s->width || mb_y*16+16 > s->height){
1550         uint8_t *ebuf= s->edge_emu_buffer + 32;
1551         ff_emulated_edge_mc(ebuf            , ptr_y , wrap_y,16,16,mb_x*16,mb_y*16, s->width   , s->height);
1552         ptr_y= ebuf;
1553         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);
1554         ptr_cb= ebuf+18*wrap_y;
1555         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);
1556         ptr_cr= ebuf+18*wrap_y+8;
1557     }
1558
1559     if (s->mb_intra) {
1560         if(s->flags&CODEC_FLAG_INTERLACED_DCT){
1561             int progressive_score, interlaced_score;
1562
1563             s->interlaced_dct=0;
1564             progressive_score= s->dsp.ildct_cmp[4](s, ptr_y           , NULL, wrap_y, 8)
1565                               +s->dsp.ildct_cmp[4](s, ptr_y + wrap_y*8, NULL, wrap_y, 8) - 400;
1566
1567             if(progressive_score > 0){
1568                 interlaced_score = s->dsp.ildct_cmp[4](s, ptr_y           , NULL, wrap_y*2, 8)
1569                                   +s->dsp.ildct_cmp[4](s, ptr_y + wrap_y  , NULL, wrap_y*2, 8);
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.get_pixels(s->block[0], ptr_y                 , wrap_y);
1582         s->dsp.get_pixels(s->block[1], ptr_y              + 8, wrap_y);
1583         s->dsp.get_pixels(s->block[2], ptr_y + dct_offset    , wrap_y);
1584         s->dsp.get_pixels(s->block[3], ptr_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.get_pixels(s->block[4], ptr_cb, wrap_c);
1591             s->dsp.get_pixels(s->block[5], ptr_cr, wrap_c);
1592             if(!s->chroma_y_shift){ /* 422 */
1593                 s->dsp.get_pixels(s->block[6], ptr_cb + (dct_offset>>1), wrap_c);
1594                 s->dsp.get_pixels(s->block[7], ptr_cr + (dct_offset>>1), wrap_c);
1595             }
1596         }
1597     }else{
1598         op_pixels_func (*op_pix)[4];
1599         qpel_mc_func (*op_qpix)[16];
1600         uint8_t *dest_y, *dest_cb, *dest_cr;
1601
1602         dest_y  = s->dest[0];
1603         dest_cb = s->dest[1];
1604         dest_cr = s->dest[2];
1605
1606         if ((!s->no_rounding) || s->pict_type==FF_B_TYPE){
1607             op_pix = s->dsp.put_pixels_tab;
1608             op_qpix= s->dsp.put_qpel_pixels_tab;
1609         }else{
1610             op_pix = s->dsp.put_no_rnd_pixels_tab;
1611             op_qpix= s->dsp.put_no_rnd_qpel_pixels_tab;
1612         }
1613
1614         if (s->mv_dir & MV_DIR_FORWARD) {
1615             MPV_motion(s, dest_y, dest_cb, dest_cr, 0, s->last_picture.data, op_pix, op_qpix);
1616             op_pix = s->dsp.avg_pixels_tab;
1617             op_qpix= s->dsp.avg_qpel_pixels_tab;
1618         }
1619         if (s->mv_dir & MV_DIR_BACKWARD) {
1620             MPV_motion(s, dest_y, dest_cb, dest_cr, 1, s->next_picture.data, op_pix, op_qpix);
1621         }
1622
1623         if(s->flags&CODEC_FLAG_INTERLACED_DCT){
1624             int progressive_score, interlaced_score;
1625
1626             s->interlaced_dct=0;
1627             progressive_score= s->dsp.ildct_cmp[0](s, dest_y           , ptr_y           , wrap_y, 8)
1628                               +s->dsp.ildct_cmp[0](s, dest_y + wrap_y*8, ptr_y + wrap_y*8, wrap_y, 8) - 400;
1629
1630             if(s->avctx->ildct_cmp == FF_CMP_VSSE) progressive_score -= 400;
1631
1632             if(progressive_score>0){
1633                 interlaced_score = s->dsp.ildct_cmp[0](s, dest_y           , ptr_y           , wrap_y*2, 8)
1634                                   +s->dsp.ildct_cmp[0](s, dest_y + wrap_y  , ptr_y + wrap_y  , wrap_y*2, 8);
1635
1636                 if(progressive_score > interlaced_score){
1637                     s->interlaced_dct=1;
1638
1639                     dct_offset= wrap_y;
1640                     wrap_y<<=1;
1641                     if (s->chroma_format == CHROMA_422)
1642                         wrap_c<<=1;
1643                 }
1644             }
1645         }
1646
1647         s->dsp.diff_pixels(s->block[0], ptr_y                 , dest_y                 , wrap_y);
1648         s->dsp.diff_pixels(s->block[1], ptr_y              + 8, dest_y              + 8, wrap_y);
1649         s->dsp.diff_pixels(s->block[2], ptr_y + dct_offset    , dest_y + dct_offset    , wrap_y);
1650         s->dsp.diff_pixels(s->block[3], ptr_y + dct_offset + 8, dest_y + dct_offset + 8, wrap_y);
1651
1652         if(s->flags&CODEC_FLAG_GRAY){
1653             skip_dct[4]= 1;
1654             skip_dct[5]= 1;
1655         }else{
1656             s->dsp.diff_pixels(s->block[4], ptr_cb, dest_cb, wrap_c);
1657             s->dsp.diff_pixels(s->block[5], ptr_cr, dest_cr, wrap_c);
1658             if(!s->chroma_y_shift){ /* 422 */
1659                 s->dsp.diff_pixels(s->block[6], ptr_cb + (dct_offset>>1), dest_cb + (dct_offset>>1), wrap_c);
1660                 s->dsp.diff_pixels(s->block[7], ptr_cr + (dct_offset>>1), dest_cr + (dct_offset>>1), wrap_c);
1661             }
1662         }
1663         /* pre quantization */
1664         if(s->current_picture.mc_mb_var[s->mb_stride*mb_y+ mb_x]<2*s->qscale*s->qscale){
1665             //FIXME optimize
1666             if(s->dsp.sad[1](NULL, ptr_y               , dest_y               , wrap_y, 8) < 20*s->qscale) skip_dct[0]= 1;
1667             if(s->dsp.sad[1](NULL, ptr_y            + 8, dest_y            + 8, wrap_y, 8) < 20*s->qscale) skip_dct[1]= 1;
1668             if(s->dsp.sad[1](NULL, ptr_y +dct_offset   , dest_y +dct_offset   , wrap_y, 8) < 20*s->qscale) skip_dct[2]= 1;
1669             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;
1670             if(s->dsp.sad[1](NULL, ptr_cb              , dest_cb              , wrap_c, 8) < 20*s->qscale) skip_dct[4]= 1;
1671             if(s->dsp.sad[1](NULL, ptr_cr              , dest_cr              , wrap_c, 8) < 20*s->qscale) skip_dct[5]= 1;
1672             if(!s->chroma_y_shift){ /* 422 */
1673                 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;
1674                 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;
1675             }
1676         }
1677     }
1678
1679     if(s->avctx->quantizer_noise_shaping){
1680         if(!skip_dct[0]) get_visual_weight(weight[0], ptr_y                 , wrap_y);
1681         if(!skip_dct[1]) get_visual_weight(weight[1], ptr_y              + 8, wrap_y);
1682         if(!skip_dct[2]) get_visual_weight(weight[2], ptr_y + dct_offset    , wrap_y);
1683         if(!skip_dct[3]) get_visual_weight(weight[3], ptr_y + dct_offset + 8, wrap_y);
1684         if(!skip_dct[4]) get_visual_weight(weight[4], ptr_cb                , wrap_c);
1685         if(!skip_dct[5]) get_visual_weight(weight[5], ptr_cr                , wrap_c);
1686         if(!s->chroma_y_shift){ /* 422 */
1687             if(!skip_dct[6]) get_visual_weight(weight[6], ptr_cb + (dct_offset>>1), wrap_c);
1688             if(!skip_dct[7]) get_visual_weight(weight[7], ptr_cr + (dct_offset>>1), wrap_c);
1689         }
1690         memcpy(orig[0], s->block[0], sizeof(DCTELEM)*64*mb_block_count);
1691     }
1692
1693     /* DCT & quantize */
1694     assert(s->out_format!=FMT_MJPEG || s->qscale==8);
1695     {
1696         for(i=0;i<mb_block_count;i++) {
1697             if(!skip_dct[i]){
1698                 int overflow;
1699                 s->block_last_index[i] = s->dct_quantize(s, s->block[i], i, s->qscale, &overflow);
1700             // FIXME we could decide to change to quantizer instead of clipping
1701             // JS: I don't think that would be a good idea it could lower quality instead
1702             //     of improve it. Just INTRADC clipping deserves changes in quantizer
1703                 if (overflow) clip_coeffs(s, s->block[i], s->block_last_index[i]);
1704             }else
1705                 s->block_last_index[i]= -1;
1706         }
1707         if(s->avctx->quantizer_noise_shaping){
1708             for(i=0;i<mb_block_count;i++) {
1709                 if(!skip_dct[i]){
1710                     s->block_last_index[i] = dct_quantize_refine(s, s->block[i], weight[i], orig[i], i, s->qscale);
1711                 }
1712             }
1713         }
1714
1715         if(s->luma_elim_threshold && !s->mb_intra)
1716             for(i=0; i<4; i++)
1717                 dct_single_coeff_elimination(s, i, s->luma_elim_threshold);
1718         if(s->chroma_elim_threshold && !s->mb_intra)
1719             for(i=4; i<mb_block_count; i++)
1720                 dct_single_coeff_elimination(s, i, s->chroma_elim_threshold);
1721
1722         if(s->flags & CODEC_FLAG_CBP_RD){
1723             for(i=0;i<mb_block_count;i++) {
1724                 if(s->block_last_index[i] == -1)
1725                     s->coded_score[i]= INT_MAX/256;
1726             }
1727         }
1728     }
1729
1730     if((s->flags&CODEC_FLAG_GRAY) && s->mb_intra){
1731         s->block_last_index[4]=
1732         s->block_last_index[5]= 0;
1733         s->block[4][0]=
1734         s->block[5][0]= (1024 + s->c_dc_scale/2)/ s->c_dc_scale;
1735     }
1736
1737     //non c quantize code returns incorrect block_last_index FIXME
1738     if(s->alternate_scan && s->dct_quantize != dct_quantize_c){
1739         for(i=0; i<mb_block_count; i++){
1740             int j;
1741             if(s->block_last_index[i]>0){
1742                 for(j=63; j>0; j--){
1743                     if(s->block[i][ s->intra_scantable.permutated[j] ]) break;
1744                 }
1745                 s->block_last_index[i]= j;
1746             }
1747         }
1748     }
1749
1750     /* huffman encode */
1751     switch(s->codec_id){ //FIXME funct ptr could be slightly faster
1752     case CODEC_ID_MPEG1VIDEO:
1753     case CODEC_ID_MPEG2VIDEO:
1754         if (CONFIG_MPEG1VIDEO_ENCODER || CONFIG_MPEG2VIDEO_ENCODER)
1755             mpeg1_encode_mb(s, s->block, motion_x, motion_y);
1756         break;
1757     case CODEC_ID_MPEG4:
1758         if (CONFIG_MPEG4_ENCODER)
1759             mpeg4_encode_mb(s, s->block, motion_x, motion_y);
1760         break;
1761     case CODEC_ID_MSMPEG4V2:
1762     case CODEC_ID_MSMPEG4V3:
1763     case CODEC_ID_WMV1:
1764         if (CONFIG_MSMPEG4_ENCODER)
1765             msmpeg4_encode_mb(s, s->block, motion_x, motion_y);
1766         break;
1767     case CODEC_ID_WMV2:
1768         if (CONFIG_WMV2_ENCODER)
1769             ff_wmv2_encode_mb(s, s->block, motion_x, motion_y);
1770         break;
1771     case CODEC_ID_H261:
1772         if (CONFIG_H261_ENCODER)
1773             ff_h261_encode_mb(s, s->block, motion_x, motion_y);
1774         break;
1775     case CODEC_ID_H263:
1776     case CODEC_ID_H263P:
1777     case CODEC_ID_FLV1:
1778     case CODEC_ID_RV10:
1779     case CODEC_ID_RV20:
1780         if (CONFIG_H263_ENCODER)
1781             h263_encode_mb(s, s->block, motion_x, motion_y);
1782         break;
1783     case CODEC_ID_MJPEG:
1784         if (CONFIG_MJPEG_ENCODER)
1785             ff_mjpeg_encode_mb(s, s->block);
1786         break;
1787     default:
1788         assert(0);
1789     }
1790 }
1791
1792 static av_always_inline void encode_mb(MpegEncContext *s, int motion_x, int motion_y)
1793 {
1794     if (s->chroma_format == CHROMA_420) encode_mb_internal(s, motion_x, motion_y,  8, 6);
1795     else                                encode_mb_internal(s, motion_x, motion_y, 16, 8);
1796 }
1797
1798 static inline void copy_context_before_encode(MpegEncContext *d, MpegEncContext *s, int type){
1799     int i;
1800
1801     memcpy(d->last_mv, s->last_mv, 2*2*2*sizeof(int)); //FIXME is memcpy faster then a loop?
1802
1803     /* mpeg1 */
1804     d->mb_skip_run= s->mb_skip_run;
1805     for(i=0; i<3; i++)
1806         d->last_dc[i]= s->last_dc[i];
1807
1808     /* statistics */
1809     d->mv_bits= s->mv_bits;
1810     d->i_tex_bits= s->i_tex_bits;
1811     d->p_tex_bits= s->p_tex_bits;
1812     d->i_count= s->i_count;
1813     d->f_count= s->f_count;
1814     d->b_count= s->b_count;
1815     d->skip_count= s->skip_count;
1816     d->misc_bits= s->misc_bits;
1817     d->last_bits= 0;
1818
1819     d->mb_skipped= 0;
1820     d->qscale= s->qscale;
1821     d->dquant= s->dquant;
1822
1823     d->esc3_level_length= s->esc3_level_length;
1824 }
1825
1826 static inline void copy_context_after_encode(MpegEncContext *d, MpegEncContext *s, int type){
1827     int i;
1828
1829     memcpy(d->mv, s->mv, 2*4*2*sizeof(int));
1830     memcpy(d->last_mv, s->last_mv, 2*2*2*sizeof(int)); //FIXME is memcpy faster then a loop?
1831
1832     /* mpeg1 */
1833     d->mb_skip_run= s->mb_skip_run;
1834     for(i=0; i<3; i++)
1835         d->last_dc[i]= s->last_dc[i];
1836
1837     /* statistics */
1838     d->mv_bits= s->mv_bits;
1839     d->i_tex_bits= s->i_tex_bits;
1840     d->p_tex_bits= s->p_tex_bits;
1841     d->i_count= s->i_count;
1842     d->f_count= s->f_count;
1843     d->b_count= s->b_count;
1844     d->skip_count= s->skip_count;
1845     d->misc_bits= s->misc_bits;
1846
1847     d->mb_intra= s->mb_intra;
1848     d->mb_skipped= s->mb_skipped;
1849     d->mv_type= s->mv_type;
1850     d->mv_dir= s->mv_dir;
1851     d->pb= s->pb;
1852     if(s->data_partitioning){
1853         d->pb2= s->pb2;
1854         d->tex_pb= s->tex_pb;
1855     }
1856     d->block= s->block;
1857     for(i=0; i<8; i++)
1858         d->block_last_index[i]= s->block_last_index[i];
1859     d->interlaced_dct= s->interlaced_dct;
1860     d->qscale= s->qscale;
1861
1862     d->esc3_level_length= s->esc3_level_length;
1863 }
1864
1865 static inline void encode_mb_hq(MpegEncContext *s, MpegEncContext *backup, MpegEncContext *best, int type,
1866                            PutBitContext pb[2], PutBitContext pb2[2], PutBitContext tex_pb[2],
1867                            int *dmin, int *next_block, int motion_x, int motion_y)
1868 {
1869     int score;
1870     uint8_t *dest_backup[3];
1871
1872     copy_context_before_encode(s, backup, type);
1873
1874     s->block= s->blocks[*next_block];
1875     s->pb= pb[*next_block];
1876     if(s->data_partitioning){
1877         s->pb2   = pb2   [*next_block];
1878         s->tex_pb= tex_pb[*next_block];
1879     }
1880
1881     if(*next_block){
1882         memcpy(dest_backup, s->dest, sizeof(s->dest));
1883         s->dest[0] = s->rd_scratchpad;
1884         s->dest[1] = s->rd_scratchpad + 16*s->linesize;
1885         s->dest[2] = s->rd_scratchpad + 16*s->linesize + 8;
1886         assert(s->linesize >= 32); //FIXME
1887     }
1888
1889     encode_mb(s, motion_x, motion_y);
1890
1891     score= put_bits_count(&s->pb);
1892     if(s->data_partitioning){
1893         score+= put_bits_count(&s->pb2);
1894         score+= put_bits_count(&s->tex_pb);
1895     }
1896
1897     if(s->avctx->mb_decision == FF_MB_DECISION_RD){
1898         MPV_decode_mb(s, s->block);
1899
1900         score *= s->lambda2;
1901         score += sse_mb(s) << FF_LAMBDA_SHIFT;
1902     }
1903
1904     if(*next_block){
1905         memcpy(s->dest, dest_backup, sizeof(s->dest));
1906     }
1907
1908     if(score<*dmin){
1909         *dmin= score;
1910         *next_block^=1;
1911
1912         copy_context_after_encode(best, s, type);
1913     }
1914 }
1915
1916 static int sse(MpegEncContext *s, uint8_t *src1, uint8_t *src2, int w, int h, int stride){
1917     uint32_t *sq = ff_squareTbl + 256;
1918     int acc=0;
1919     int x,y;
1920
1921     if(w==16 && h==16)
1922         return s->dsp.sse[0](NULL, src1, src2, stride, 16);
1923     else if(w==8 && h==8)
1924         return s->dsp.sse[1](NULL, src1, src2, stride, 8);
1925
1926     for(y=0; y<h; y++){
1927         for(x=0; x<w; x++){
1928             acc+= sq[src1[x + y*stride] - src2[x + y*stride]];
1929         }
1930     }
1931
1932     assert(acc>=0);
1933
1934     return acc;
1935 }
1936
1937 static int sse_mb(MpegEncContext *s){
1938     int w= 16;
1939     int h= 16;
1940
1941     if(s->mb_x*16 + 16 > s->width ) w= s->width - s->mb_x*16;
1942     if(s->mb_y*16 + 16 > s->height) h= s->height- s->mb_y*16;
1943
1944     if(w==16 && h==16)
1945       if(s->avctx->mb_cmp == FF_CMP_NSSE){
1946         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)
1947                +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)
1948                +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);
1949       }else{
1950         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)
1951                +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)
1952                +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);
1953       }
1954     else
1955         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)
1956                +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)
1957                +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);
1958 }
1959
1960 static int pre_estimate_motion_thread(AVCodecContext *c, void *arg){
1961     MpegEncContext *s= *(void**)arg;
1962
1963
1964     s->me.pre_pass=1;
1965     s->me.dia_size= s->avctx->pre_dia_size;
1966     s->first_slice_line=1;
1967     for(s->mb_y= s->end_mb_y-1; s->mb_y >= s->start_mb_y; s->mb_y--) {
1968         for(s->mb_x=s->mb_width-1; s->mb_x >=0 ;s->mb_x--) {
1969             ff_pre_estimate_p_frame_motion(s, s->mb_x, s->mb_y);
1970         }
1971         s->first_slice_line=0;
1972     }
1973
1974     s->me.pre_pass=0;
1975
1976     return 0;
1977 }
1978
1979 static int estimate_motion_thread(AVCodecContext *c, void *arg){
1980     MpegEncContext *s= *(void**)arg;
1981
1982     ff_check_alignment();
1983
1984     s->me.dia_size= s->avctx->dia_size;
1985     s->first_slice_line=1;
1986     for(s->mb_y= s->start_mb_y; s->mb_y < s->end_mb_y; s->mb_y++) {
1987         s->mb_x=0; //for block init below
1988         ff_init_block_index(s);
1989         for(s->mb_x=0; s->mb_x < s->mb_width; s->mb_x++) {
1990             s->block_index[0]+=2;
1991             s->block_index[1]+=2;
1992             s->block_index[2]+=2;
1993             s->block_index[3]+=2;
1994
1995             /* compute motion vector & mb_type and store in context */
1996             if(s->pict_type==FF_B_TYPE)
1997                 ff_estimate_b_frame_motion(s, s->mb_x, s->mb_y);
1998             else
1999                 ff_estimate_p_frame_motion(s, s->mb_x, s->mb_y);
2000         }
2001         s->first_slice_line=0;
2002     }
2003     return 0;
2004 }
2005
2006 static int mb_var_thread(AVCodecContext *c, void *arg){
2007     MpegEncContext *s= *(void**)arg;
2008     int mb_x, mb_y;
2009
2010     ff_check_alignment();
2011
2012     for(mb_y=s->start_mb_y; mb_y < s->end_mb_y; mb_y++) {
2013         for(mb_x=0; mb_x < s->mb_width; mb_x++) {
2014             int xx = mb_x * 16;
2015             int yy = mb_y * 16;
2016             uint8_t *pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
2017             int varc;
2018             int sum = s->dsp.pix_sum(pix, s->linesize);
2019
2020             varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
2021
2022             s->current_picture.mb_var [s->mb_stride * mb_y + mb_x] = varc;
2023             s->current_picture.mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
2024             s->me.mb_var_sum_temp    += varc;
2025         }
2026     }
2027     return 0;
2028 }
2029
2030 static void write_slice_end(MpegEncContext *s){
2031     if(CONFIG_MPEG4_ENCODER && s->codec_id==CODEC_ID_MPEG4){
2032         if(s->partitioned_frame){
2033             ff_mpeg4_merge_partitions(s);
2034         }
2035
2036         ff_mpeg4_stuffing(&s->pb);
2037     }else if(CONFIG_MJPEG_ENCODER && s->out_format == FMT_MJPEG){
2038         ff_mjpeg_encode_stuffing(&s->pb);
2039     }
2040
2041     align_put_bits(&s->pb);
2042     flush_put_bits(&s->pb);
2043
2044     if((s->flags&CODEC_FLAG_PASS1) && !s->partitioned_frame)
2045         s->misc_bits+= get_bits_diff(s);
2046 }
2047
2048 static int encode_thread(AVCodecContext *c, void *arg){
2049     MpegEncContext *s= *(void**)arg;
2050     int mb_x, mb_y, pdif = 0;
2051     int chr_h= 16>>s->chroma_y_shift;
2052     int i, j;
2053     MpegEncContext best_s, backup_s;
2054     uint8_t bit_buf[2][MAX_MB_BYTES];
2055     uint8_t bit_buf2[2][MAX_MB_BYTES];
2056     uint8_t bit_buf_tex[2][MAX_MB_BYTES];
2057     PutBitContext pb[2], pb2[2], tex_pb[2];
2058 //printf("%d->%d\n", s->resync_mb_y, s->end_mb_y);
2059
2060     ff_check_alignment();
2061
2062     for(i=0; i<2; i++){
2063         init_put_bits(&pb    [i], bit_buf    [i], MAX_MB_BYTES);
2064         init_put_bits(&pb2   [i], bit_buf2   [i], MAX_MB_BYTES);
2065         init_put_bits(&tex_pb[i], bit_buf_tex[i], MAX_MB_BYTES);
2066     }
2067
2068     s->last_bits= put_bits_count(&s->pb);
2069     s->mv_bits=0;
2070     s->misc_bits=0;
2071     s->i_tex_bits=0;
2072     s->p_tex_bits=0;
2073     s->i_count=0;
2074     s->f_count=0;
2075     s->b_count=0;
2076     s->skip_count=0;
2077
2078     for(i=0; i<3; i++){
2079         /* init last dc values */
2080         /* note: quant matrix value (8) is implied here */
2081         s->last_dc[i] = 128 << s->intra_dc_precision;
2082
2083         s->current_picture.error[i] = 0;
2084     }
2085     s->mb_skip_run = 0;
2086     memset(s->last_mv, 0, sizeof(s->last_mv));
2087
2088     s->last_mv_dir = 0;
2089
2090     switch(s->codec_id){
2091     case CODEC_ID_H263:
2092     case CODEC_ID_H263P:
2093     case CODEC_ID_FLV1:
2094         if (CONFIG_H263_ENCODER)
2095             s->gob_index = ff_h263_get_gob_height(s);
2096         break;
2097     case CODEC_ID_MPEG4:
2098         if(CONFIG_MPEG4_ENCODER && s->partitioned_frame)
2099             ff_mpeg4_init_partitions(s);
2100         break;
2101     }
2102
2103     s->resync_mb_x=0;
2104     s->resync_mb_y=0;
2105     s->first_slice_line = 1;
2106     s->ptr_lastgob = s->pb.buf;
2107     for(mb_y= s->start_mb_y; mb_y < s->end_mb_y; mb_y++) {
2108 //    printf("row %d at %X\n", s->mb_y, (int)s);
2109         s->mb_x=0;
2110         s->mb_y= mb_y;
2111
2112         ff_set_qscale(s, s->qscale);
2113         ff_init_block_index(s);
2114
2115         for(mb_x=0; mb_x < s->mb_width; mb_x++) {
2116             int xy= mb_y*s->mb_stride + mb_x; // removed const, H261 needs to adjust this
2117             int mb_type= s->mb_type[xy];
2118 //            int d;
2119             int dmin= INT_MAX;
2120             int dir;
2121
2122             if(s->pb.buf_end - s->pb.buf - (put_bits_count(&s->pb)>>3) < MAX_MB_BYTES){
2123                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
2124                 return -1;
2125             }
2126             if(s->data_partitioning){
2127                 if(   s->pb2   .buf_end - s->pb2   .buf - (put_bits_count(&s->    pb2)>>3) < MAX_MB_BYTES
2128                    || s->tex_pb.buf_end - s->tex_pb.buf - (put_bits_count(&s->tex_pb )>>3) < MAX_MB_BYTES){
2129                     av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
2130                     return -1;
2131                 }
2132             }
2133
2134             s->mb_x = mb_x;
2135             s->mb_y = mb_y;  // moved into loop, can get changed by H.261
2136             ff_update_block_index(s);
2137
2138             if(CONFIG_H261_ENCODER && s->codec_id == CODEC_ID_H261){
2139                 ff_h261_reorder_mb_index(s);
2140                 xy= s->mb_y*s->mb_stride + s->mb_x;
2141                 mb_type= s->mb_type[xy];
2142             }
2143
2144             /* write gob / video packet header  */
2145             if(s->rtp_mode){
2146                 int current_packet_size, is_gob_start;
2147
2148                 current_packet_size= ((put_bits_count(&s->pb)+7)>>3) - (s->ptr_lastgob - s->pb.buf);
2149
2150                 is_gob_start= s->avctx->rtp_payload_size && current_packet_size >= s->avctx->rtp_payload_size && mb_y + mb_x>0;
2151
2152                 if(s->start_mb_y == mb_y && mb_y > 0 && mb_x==0) is_gob_start=1;
2153
2154                 switch(s->codec_id){
2155                 case CODEC_ID_H263:
2156                 case CODEC_ID_H263P:
2157                     if(!s->h263_slice_structured)
2158                         if(s->mb_x || s->mb_y%s->gob_index) is_gob_start=0;
2159                     break;
2160                 case CODEC_ID_MPEG2VIDEO:
2161                     if(s->mb_x==0 && s->mb_y!=0) is_gob_start=1;
2162                 case CODEC_ID_MPEG1VIDEO:
2163                     if(s->mb_skip_run) is_gob_start=0;
2164                     break;
2165                 }
2166
2167                 if(is_gob_start){
2168                     if(s->start_mb_y != mb_y || mb_x!=0){
2169                         write_slice_end(s);
2170
2171                         if(CONFIG_MPEG4_ENCODER && s->codec_id==CODEC_ID_MPEG4 && s->partitioned_frame){
2172                             ff_mpeg4_init_partitions(s);
2173                         }
2174                     }
2175
2176                     assert((put_bits_count(&s->pb)&7) == 0);
2177                     current_packet_size= put_bits_ptr(&s->pb) - s->ptr_lastgob;
2178
2179                     if(s->avctx->error_rate && s->resync_mb_x + s->resync_mb_y > 0){
2180                         int r= put_bits_count(&s->pb)/8 + s->picture_number + 16 + s->mb_x + s->mb_y;
2181                         int d= 100 / s->avctx->error_rate;
2182                         if(r % d == 0){
2183                             current_packet_size=0;
2184 #ifndef ALT_BITSTREAM_WRITER
2185                             s->pb.buf_ptr= s->ptr_lastgob;
2186 #endif
2187                             assert(put_bits_ptr(&s->pb) == s->ptr_lastgob);
2188                         }
2189                     }
2190
2191                     if (s->avctx->rtp_callback){
2192                         int number_mb = (mb_y - s->resync_mb_y)*s->mb_width + mb_x - s->resync_mb_x;
2193                         s->avctx->rtp_callback(s->avctx, s->ptr_lastgob, current_packet_size, number_mb);
2194                     }
2195
2196                     switch(s->codec_id){
2197                     case CODEC_ID_MPEG4:
2198                         if (CONFIG_MPEG4_ENCODER) {
2199                             ff_mpeg4_encode_video_packet_header(s);
2200                             ff_mpeg4_clean_buffers(s);
2201                         }
2202                     break;
2203                     case CODEC_ID_MPEG1VIDEO:
2204                     case CODEC_ID_MPEG2VIDEO:
2205                         if (CONFIG_MPEG1VIDEO_ENCODER || CONFIG_MPEG2VIDEO_ENCODER) {
2206                             ff_mpeg1_encode_slice_header(s);
2207                             ff_mpeg1_clean_buffers(s);
2208                         }
2209                     break;
2210                     case CODEC_ID_H263:
2211                     case CODEC_ID_H263P:
2212                         if (CONFIG_H263_ENCODER)
2213                             h263_encode_gob_header(s, mb_y);
2214                     break;
2215                     }
2216
2217                     if(s->flags&CODEC_FLAG_PASS1){
2218                         int bits= put_bits_count(&s->pb);
2219                         s->misc_bits+= bits - s->last_bits;
2220                         s->last_bits= bits;
2221                     }
2222
2223                     s->ptr_lastgob += current_packet_size;
2224                     s->first_slice_line=1;
2225                     s->resync_mb_x=mb_x;
2226                     s->resync_mb_y=mb_y;
2227                 }
2228             }
2229
2230             if(  (s->resync_mb_x   == s->mb_x)
2231                && s->resync_mb_y+1 == s->mb_y){
2232                 s->first_slice_line=0;
2233             }
2234
2235             s->mb_skipped=0;
2236             s->dquant=0; //only for QP_RD
2237
2238             if(mb_type & (mb_type-1) || (s->flags & CODEC_FLAG_QP_RD)){ // more than 1 MB type possible or CODEC_FLAG_QP_RD
2239                 int next_block=0;
2240                 int pb_bits_count, pb2_bits_count, tex_pb_bits_count;
2241
2242                 copy_context_before_encode(&backup_s, s, -1);
2243                 backup_s.pb= s->pb;
2244                 best_s.data_partitioning= s->data_partitioning;
2245                 best_s.partitioned_frame= s->partitioned_frame;
2246                 if(s->data_partitioning){
2247                     backup_s.pb2= s->pb2;
2248                     backup_s.tex_pb= s->tex_pb;
2249                 }
2250
2251                 if(mb_type&CANDIDATE_MB_TYPE_INTER){
2252                     s->mv_dir = MV_DIR_FORWARD;
2253                     s->mv_type = MV_TYPE_16X16;
2254                     s->mb_intra= 0;
2255                     s->mv[0][0][0] = s->p_mv_table[xy][0];
2256                     s->mv[0][0][1] = s->p_mv_table[xy][1];
2257                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_INTER, pb, pb2, tex_pb,
2258                                  &dmin, &next_block, s->mv[0][0][0], s->mv[0][0][1]);
2259                 }
2260                 if(mb_type&CANDIDATE_MB_TYPE_INTER_I){
2261                     s->mv_dir = MV_DIR_FORWARD;
2262                     s->mv_type = MV_TYPE_FIELD;
2263                     s->mb_intra= 0;
2264                     for(i=0; i<2; i++){
2265                         j= s->field_select[0][i] = s->p_field_select_table[i][xy];
2266                         s->mv[0][i][0] = s->p_field_mv_table[i][j][xy][0];
2267                         s->mv[0][i][1] = s->p_field_mv_table[i][j][xy][1];
2268                     }
2269                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_INTER_I, pb, pb2, tex_pb,
2270                                  &dmin, &next_block, 0, 0);
2271                 }
2272                 if(mb_type&CANDIDATE_MB_TYPE_SKIPPED){
2273                     s->mv_dir = MV_DIR_FORWARD;
2274                     s->mv_type = MV_TYPE_16X16;
2275                     s->mb_intra= 0;
2276                     s->mv[0][0][0] = 0;
2277                     s->mv[0][0][1] = 0;
2278                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_SKIPPED, pb, pb2, tex_pb,
2279                                  &dmin, &next_block, s->mv[0][0][0], s->mv[0][0][1]);
2280                 }
2281                 if(mb_type&CANDIDATE_MB_TYPE_INTER4V){
2282                     s->mv_dir = MV_DIR_FORWARD;
2283                     s->mv_type = MV_TYPE_8X8;
2284                     s->mb_intra= 0;
2285                     for(i=0; i<4; i++){
2286                         s->mv[0][i][0] = s->current_picture.motion_val[0][s->block_index[i]][0];
2287                         s->mv[0][i][1] = s->current_picture.motion_val[0][s->block_index[i]][1];
2288                     }
2289                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_INTER4V, pb, pb2, tex_pb,
2290                                  &dmin, &next_block, 0, 0);
2291                 }
2292                 if(mb_type&CANDIDATE_MB_TYPE_FORWARD){
2293                     s->mv_dir = MV_DIR_FORWARD;
2294                     s->mv_type = MV_TYPE_16X16;
2295                     s->mb_intra= 0;
2296                     s->mv[0][0][0] = s->b_forw_mv_table[xy][0];
2297                     s->mv[0][0][1] = s->b_forw_mv_table[xy][1];
2298                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_FORWARD, pb, pb2, tex_pb,
2299                                  &dmin, &next_block, s->mv[0][0][0], s->mv[0][0][1]);
2300                 }
2301                 if(mb_type&CANDIDATE_MB_TYPE_BACKWARD){
2302                     s->mv_dir = MV_DIR_BACKWARD;
2303                     s->mv_type = MV_TYPE_16X16;
2304                     s->mb_intra= 0;
2305                     s->mv[1][0][0] = s->b_back_mv_table[xy][0];
2306                     s->mv[1][0][1] = s->b_back_mv_table[xy][1];
2307                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_BACKWARD, pb, pb2, tex_pb,
2308                                  &dmin, &next_block, s->mv[1][0][0], s->mv[1][0][1]);
2309                 }
2310                 if(mb_type&CANDIDATE_MB_TYPE_BIDIR){
2311                     s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD;
2312                     s->mv_type = MV_TYPE_16X16;
2313                     s->mb_intra= 0;
2314                     s->mv[0][0][0] = s->b_bidir_forw_mv_table[xy][0];
2315                     s->mv[0][0][1] = s->b_bidir_forw_mv_table[xy][1];
2316                     s->mv[1][0][0] = s->b_bidir_back_mv_table[xy][0];
2317                     s->mv[1][0][1] = s->b_bidir_back_mv_table[xy][1];
2318                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_BIDIR, pb, pb2, tex_pb,
2319                                  &dmin, &next_block, 0, 0);
2320                 }
2321                 if(mb_type&CANDIDATE_MB_TYPE_FORWARD_I){
2322                     s->mv_dir = MV_DIR_FORWARD;
2323                     s->mv_type = MV_TYPE_FIELD;
2324                     s->mb_intra= 0;
2325                     for(i=0; i<2; i++){
2326                         j= s->field_select[0][i] = s->b_field_select_table[0][i][xy];
2327                         s->mv[0][i][0] = s->b_field_mv_table[0][i][j][xy][0];
2328                         s->mv[0][i][1] = s->b_field_mv_table[0][i][j][xy][1];
2329                     }
2330                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_FORWARD_I, pb, pb2, tex_pb,
2331                                  &dmin, &next_block, 0, 0);
2332                 }
2333                 if(mb_type&CANDIDATE_MB_TYPE_BACKWARD_I){
2334                     s->mv_dir = MV_DIR_BACKWARD;
2335                     s->mv_type = MV_TYPE_FIELD;
2336                     s->mb_intra= 0;
2337                     for(i=0; i<2; i++){
2338                         j= s->field_select[1][i] = s->b_field_select_table[1][i][xy];
2339                         s->mv[1][i][0] = s->b_field_mv_table[1][i][j][xy][0];
2340                         s->mv[1][i][1] = s->b_field_mv_table[1][i][j][xy][1];
2341                     }
2342                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_BACKWARD_I, pb, pb2, tex_pb,
2343                                  &dmin, &next_block, 0, 0);
2344                 }
2345                 if(mb_type&CANDIDATE_MB_TYPE_BIDIR_I){
2346                     s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD;
2347                     s->mv_type = MV_TYPE_FIELD;
2348                     s->mb_intra= 0;
2349                     for(dir=0; dir<2; dir++){
2350                         for(i=0; i<2; i++){
2351                             j= s->field_select[dir][i] = s->b_field_select_table[dir][i][xy];
2352                             s->mv[dir][i][0] = s->b_field_mv_table[dir][i][j][xy][0];
2353                             s->mv[dir][i][1] = s->b_field_mv_table[dir][i][j][xy][1];
2354                         }
2355                     }
2356                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_BIDIR_I, pb, pb2, tex_pb,
2357                                  &dmin, &next_block, 0, 0);
2358                 }
2359                 if(mb_type&CANDIDATE_MB_TYPE_INTRA){
2360                     s->mv_dir = 0;
2361                     s->mv_type = MV_TYPE_16X16;
2362                     s->mb_intra= 1;
2363                     s->mv[0][0][0] = 0;
2364                     s->mv[0][0][1] = 0;
2365                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_INTRA, pb, pb2, tex_pb,
2366                                  &dmin, &next_block, 0, 0);
2367                     if(s->h263_pred || s->h263_aic){
2368                         if(best_s.mb_intra)
2369                             s->mbintra_table[mb_x + mb_y*s->mb_stride]=1;
2370                         else
2371                             ff_clean_intra_table_entries(s); //old mode?
2372                     }
2373                 }
2374
2375                 if((s->flags & CODEC_FLAG_QP_RD) && dmin < INT_MAX){
2376                     if(best_s.mv_type==MV_TYPE_16X16){ //FIXME move 4mv after QPRD
2377                         const int last_qp= backup_s.qscale;
2378                         int qpi, qp, dc[6];
2379                         DCTELEM ac[6][16];
2380                         const int mvdir= (best_s.mv_dir&MV_DIR_BACKWARD) ? 1 : 0;
2381                         static const int dquant_tab[4]={-1,1,-2,2};
2382
2383                         assert(backup_s.dquant == 0);
2384
2385                         //FIXME intra
2386                         s->mv_dir= best_s.mv_dir;
2387                         s->mv_type = MV_TYPE_16X16;
2388                         s->mb_intra= best_s.mb_intra;
2389                         s->mv[0][0][0] = best_s.mv[0][0][0];
2390                         s->mv[0][0][1] = best_s.mv[0][0][1];
2391                         s->mv[1][0][0] = best_s.mv[1][0][0];
2392                         s->mv[1][0][1] = best_s.mv[1][0][1];
2393
2394                         qpi = s->pict_type == FF_B_TYPE ? 2 : 0;
2395                         for(; qpi<4; qpi++){
2396                             int dquant= dquant_tab[qpi];
2397                             qp= last_qp + dquant;
2398                             if(qp < s->avctx->qmin || qp > s->avctx->qmax)
2399                                 continue;
2400                             backup_s.dquant= dquant;
2401                             if(s->mb_intra && s->dc_val[0]){
2402                                 for(i=0; i<6; i++){
2403                                     dc[i]= s->dc_val[0][ s->block_index[i] ];
2404                                     memcpy(ac[i], s->ac_val[0][s->block_index[i]], sizeof(DCTELEM)*16);
2405                                 }
2406                             }
2407
2408                             encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_INTER /* wrong but unused */, pb, pb2, tex_pb,
2409                                          &dmin, &next_block, s->mv[mvdir][0][0], s->mv[mvdir][0][1]);
2410                             if(best_s.qscale != qp){
2411                                 if(s->mb_intra && s->dc_val[0]){
2412                                     for(i=0; i<6; i++){
2413                                         s->dc_val[0][ s->block_index[i] ]= dc[i];
2414                                         memcpy(s->ac_val[0][s->block_index[i]], ac[i], sizeof(DCTELEM)*16);
2415                                     }
2416                                 }
2417                             }
2418                         }
2419                     }
2420                 }
2421                 if(CONFIG_MPEG4_ENCODER && mb_type&CANDIDATE_MB_TYPE_DIRECT){
2422                     int mx= s->b_direct_mv_table[xy][0];
2423                     int my= s->b_direct_mv_table[xy][1];
2424
2425                     backup_s.dquant = 0;
2426                     s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD | MV_DIRECT;
2427                     s->mb_intra= 0;
2428                     ff_mpeg4_set_direct_mv(s, mx, my);
2429                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_DIRECT, pb, pb2, tex_pb,
2430                                  &dmin, &next_block, mx, my);
2431                 }
2432                 if(CONFIG_MPEG4_ENCODER && mb_type&CANDIDATE_MB_TYPE_DIRECT0){
2433                     backup_s.dquant = 0;
2434                     s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD | MV_DIRECT;
2435                     s->mb_intra= 0;
2436                     ff_mpeg4_set_direct_mv(s, 0, 0);
2437                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_DIRECT, pb, pb2, tex_pb,
2438                                  &dmin, &next_block, 0, 0);
2439                 }
2440                 if(!best_s.mb_intra && s->flags2&CODEC_FLAG2_SKIP_RD){
2441                     int coded=0;
2442                     for(i=0; i<6; i++)
2443                         coded |= s->block_last_index[i];
2444                     if(coded){
2445                         int mx,my;
2446                         memcpy(s->mv, best_s.mv, sizeof(s->mv));
2447                         if(CONFIG_MPEG4_ENCODER && best_s.mv_dir & MV_DIRECT){
2448                             mx=my=0; //FIXME find the one we actually used
2449                             ff_mpeg4_set_direct_mv(s, mx, my);
2450                         }else if(best_s.mv_dir&MV_DIR_BACKWARD){
2451                             mx= s->mv[1][0][0];
2452                             my= s->mv[1][0][1];
2453                         }else{
2454                             mx= s->mv[0][0][0];
2455                             my= s->mv[0][0][1];
2456                         }
2457
2458                         s->mv_dir= best_s.mv_dir;
2459                         s->mv_type = best_s.mv_type;
2460                         s->mb_intra= 0;
2461 /*                        s->mv[0][0][0] = best_s.mv[0][0][0];
2462                         s->mv[0][0][1] = best_s.mv[0][0][1];
2463                         s->mv[1][0][0] = best_s.mv[1][0][0];
2464                         s->mv[1][0][1] = best_s.mv[1][0][1];*/
2465                         backup_s.dquant= 0;
2466                         s->skipdct=1;
2467                         encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_INTER /* wrong but unused */, pb, pb2, tex_pb,
2468                                         &dmin, &next_block, mx, my);
2469                         s->skipdct=0;
2470                     }
2471                 }
2472
2473                 s->current_picture.qscale_table[xy]= best_s.qscale;
2474
2475                 copy_context_after_encode(s, &best_s, -1);
2476
2477                 pb_bits_count= put_bits_count(&s->pb);
2478                 flush_put_bits(&s->pb);
2479                 ff_copy_bits(&backup_s.pb, bit_buf[next_block^1], pb_bits_count);
2480                 s->pb= backup_s.pb;
2481
2482                 if(s->data_partitioning){
2483                     pb2_bits_count= put_bits_count(&s->pb2);
2484                     flush_put_bits(&s->pb2);
2485                     ff_copy_bits(&backup_s.pb2, bit_buf2[next_block^1], pb2_bits_count);
2486                     s->pb2= backup_s.pb2;
2487
2488                     tex_pb_bits_count= put_bits_count(&s->tex_pb);
2489                     flush_put_bits(&s->tex_pb);
2490                     ff_copy_bits(&backup_s.tex_pb, bit_buf_tex[next_block^1], tex_pb_bits_count);
2491                     s->tex_pb= backup_s.tex_pb;
2492                 }
2493                 s->last_bits= put_bits_count(&s->pb);
2494
2495                 if (CONFIG_H263_ENCODER &&
2496                     s->out_format == FMT_H263 && s->pict_type!=FF_B_TYPE)
2497                     ff_h263_update_motion_val(s);
2498
2499                 if(next_block==0){ //FIXME 16 vs linesize16
2500                     s->dsp.put_pixels_tab[0][0](s->dest[0], s->rd_scratchpad                     , s->linesize  ,16);
2501                     s->dsp.put_pixels_tab[1][0](s->dest[1], s->rd_scratchpad + 16*s->linesize    , s->uvlinesize, 8);
2502                     s->dsp.put_pixels_tab[1][0](s->dest[2], s->rd_scratchpad + 16*s->linesize + 8, s->uvlinesize, 8);
2503                 }
2504
2505                 if(s->avctx->mb_decision == FF_MB_DECISION_BITS)
2506                     MPV_decode_mb(s, s->block);
2507             } else {
2508                 int motion_x = 0, motion_y = 0;
2509                 s->mv_type=MV_TYPE_16X16;
2510                 // only one MB-Type possible
2511
2512                 switch(mb_type){
2513                 case CANDIDATE_MB_TYPE_INTRA:
2514                     s->mv_dir = 0;
2515                     s->mb_intra= 1;
2516                     motion_x= s->mv[0][0][0] = 0;
2517                     motion_y= s->mv[0][0][1] = 0;
2518                     break;
2519                 case CANDIDATE_MB_TYPE_INTER:
2520                     s->mv_dir = MV_DIR_FORWARD;
2521                     s->mb_intra= 0;
2522                     motion_x= s->mv[0][0][0] = s->p_mv_table[xy][0];
2523                     motion_y= s->mv[0][0][1] = s->p_mv_table[xy][1];
2524                     break;
2525                 case CANDIDATE_MB_TYPE_INTER_I:
2526                     s->mv_dir = MV_DIR_FORWARD;
2527                     s->mv_type = MV_TYPE_FIELD;
2528                     s->mb_intra= 0;
2529                     for(i=0; i<2; i++){
2530                         j= s->field_select[0][i] = s->p_field_select_table[i][xy];
2531                         s->mv[0][i][0] = s->p_field_mv_table[i][j][xy][0];
2532                         s->mv[0][i][1] = s->p_field_mv_table[i][j][xy][1];
2533                     }
2534                     break;
2535                 case CANDIDATE_MB_TYPE_INTER4V:
2536                     s->mv_dir = MV_DIR_FORWARD;
2537                     s->mv_type = MV_TYPE_8X8;
2538                     s->mb_intra= 0;
2539                     for(i=0; i<4; i++){
2540                         s->mv[0][i][0] = s->current_picture.motion_val[0][s->block_index[i]][0];
2541                         s->mv[0][i][1] = s->current_picture.motion_val[0][s->block_index[i]][1];
2542                     }
2543                     break;
2544                 case CANDIDATE_MB_TYPE_DIRECT:
2545                     if (CONFIG_MPEG4_ENCODER) {
2546                         s->mv_dir = MV_DIR_FORWARD|MV_DIR_BACKWARD|MV_DIRECT;
2547                         s->mb_intra= 0;
2548                         motion_x=s->b_direct_mv_table[xy][0];
2549                         motion_y=s->b_direct_mv_table[xy][1];
2550                         ff_mpeg4_set_direct_mv(s, motion_x, motion_y);
2551                     }
2552                     break;
2553                 case CANDIDATE_MB_TYPE_DIRECT0:
2554                     if (CONFIG_MPEG4_ENCODER) {
2555                         s->mv_dir = MV_DIR_FORWARD|MV_DIR_BACKWARD|MV_DIRECT;
2556                         s->mb_intra= 0;
2557                         ff_mpeg4_set_direct_mv(s, 0, 0);
2558                     }
2559                     break;
2560                 case CANDIDATE_MB_TYPE_BIDIR:
2561                     s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD;
2562                     s->mb_intra= 0;
2563                     s->mv[0][0][0] = s->b_bidir_forw_mv_table[xy][0];
2564                     s->mv[0][0][1] = s->b_bidir_forw_mv_table[xy][1];
2565                     s->mv[1][0][0] = s->b_bidir_back_mv_table[xy][0];
2566                     s->mv[1][0][1] = s->b_bidir_back_mv_table[xy][1];
2567                     break;
2568                 case CANDIDATE_MB_TYPE_BACKWARD:
2569                     s->mv_dir = MV_DIR_BACKWARD;
2570                     s->mb_intra= 0;
2571                     motion_x= s->mv[1][0][0] = s->b_back_mv_table[xy][0];
2572                     motion_y= s->mv[1][0][1] = s->b_back_mv_table[xy][1];
2573                     break;
2574                 case CANDIDATE_MB_TYPE_FORWARD:
2575                     s->mv_dir = MV_DIR_FORWARD;
2576                     s->mb_intra= 0;
2577                     motion_x= s->mv[0][0][0] = s->b_forw_mv_table[xy][0];
2578                     motion_y= s->mv[0][0][1] = s->b_forw_mv_table[xy][1];
2579 //                    printf(" %d %d ", motion_x, motion_y);
2580                     break;
2581                 case CANDIDATE_MB_TYPE_FORWARD_I:
2582                     s->mv_dir = MV_DIR_FORWARD;
2583                     s->mv_type = MV_TYPE_FIELD;
2584                     s->mb_intra= 0;
2585                     for(i=0; i<2; i++){
2586                         j= s->field_select[0][i] = s->b_field_select_table[0][i][xy];
2587                         s->mv[0][i][0] = s->b_field_mv_table[0][i][j][xy][0];
2588                         s->mv[0][i][1] = s->b_field_mv_table[0][i][j][xy][1];
2589                     }
2590                     break;
2591                 case CANDIDATE_MB_TYPE_BACKWARD_I:
2592                     s->mv_dir = MV_DIR_BACKWARD;
2593                     s->mv_type = MV_TYPE_FIELD;
2594                     s->mb_intra= 0;
2595                     for(i=0; i<2; i++){
2596                         j= s->field_select[1][i] = s->b_field_select_table[1][i][xy];
2597                         s->mv[1][i][0] = s->b_field_mv_table[1][i][j][xy][0];
2598                         s->mv[1][i][1] = s->b_field_mv_table[1][i][j][xy][1];
2599                     }
2600                     break;
2601                 case CANDIDATE_MB_TYPE_BIDIR_I:
2602                     s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD;
2603                     s->mv_type = MV_TYPE_FIELD;
2604                     s->mb_intra= 0;
2605                     for(dir=0; dir<2; dir++){
2606                         for(i=0; i<2; i++){
2607                             j= s->field_select[dir][i] = s->b_field_select_table[dir][i][xy];
2608                             s->mv[dir][i][0] = s->b_field_mv_table[dir][i][j][xy][0];
2609                             s->mv[dir][i][1] = s->b_field_mv_table[dir][i][j][xy][1];
2610                         }
2611                     }
2612                     break;
2613                 default:
2614                     av_log(s->avctx, AV_LOG_ERROR, "illegal MB type\n");
2615                 }
2616
2617                 encode_mb(s, motion_x, motion_y);
2618
2619                 // RAL: Update last macroblock type
2620                 s->last_mv_dir = s->mv_dir;
2621
2622                 if (CONFIG_H263_ENCODER &&
2623                     s->out_format == FMT_H263 && s->pict_type!=FF_B_TYPE)
2624                     ff_h263_update_motion_val(s);
2625
2626                 MPV_decode_mb(s, s->block);
2627             }
2628
2629             /* clean the MV table in IPS frames for direct mode in B frames */
2630             if(s->mb_intra /* && I,P,S_TYPE */){
2631                 s->p_mv_table[xy][0]=0;
2632                 s->p_mv_table[xy][1]=0;
2633             }
2634
2635             if(s->flags&CODEC_FLAG_PSNR){
2636                 int w= 16;
2637                 int h= 16;
2638
2639                 if(s->mb_x*16 + 16 > s->width ) w= s->width - s->mb_x*16;
2640                 if(s->mb_y*16 + 16 > s->height) h= s->height- s->mb_y*16;
2641
2642                 s->current_picture.error[0] += sse(
2643                     s, s->new_picture.data[0] + s->mb_x*16 + s->mb_y*s->linesize*16,
2644                     s->dest[0], w, h, s->linesize);
2645                 s->current_picture.error[1] += sse(
2646                     s, s->new_picture.data[1] + s->mb_x*8  + s->mb_y*s->uvlinesize*chr_h,
2647                     s->dest[1], w>>1, h>>s->chroma_y_shift, s->uvlinesize);
2648                 s->current_picture.error[2] += sse(
2649                     s, s->new_picture.data[2] + s->mb_x*8  + s->mb_y*s->uvlinesize*chr_h,
2650                     s->dest[2], w>>1, h>>s->chroma_y_shift, s->uvlinesize);
2651             }
2652             if(s->loop_filter){
2653                 if(CONFIG_H263_ENCODER && s->out_format == FMT_H263)
2654                     ff_h263_loop_filter(s);
2655             }
2656 //printf("MB %d %d bits\n", s->mb_x+s->mb_y*s->mb_stride, put_bits_count(&s->pb));
2657         }
2658     }
2659
2660     //not beautiful here but we must write it before flushing so it has to be here
2661     if (CONFIG_MSMPEG4_ENCODER && s->msmpeg4_version && s->msmpeg4_version<4 && s->pict_type == FF_I_TYPE)
2662         msmpeg4_encode_ext_header(s);
2663
2664     write_slice_end(s);
2665
2666     /* Send the last GOB if RTP */
2667     if (s->avctx->rtp_callback) {
2668         int number_mb = (mb_y - s->resync_mb_y)*s->mb_width - s->resync_mb_x;
2669         pdif = put_bits_ptr(&s->pb) - s->ptr_lastgob;
2670         /* Call the RTP callback to send the last GOB */
2671         emms_c();
2672         s->avctx->rtp_callback(s->avctx, s->ptr_lastgob, pdif, number_mb);
2673     }
2674
2675     return 0;
2676 }
2677
2678 #define MERGE(field) dst->field += src->field; src->field=0
2679 static void merge_context_after_me(MpegEncContext *dst, MpegEncContext *src){
2680     MERGE(me.scene_change_score);
2681     MERGE(me.mc_mb_var_sum_temp);
2682     MERGE(me.mb_var_sum_temp);
2683 }
2684
2685 static void merge_context_after_encode(MpegEncContext *dst, MpegEncContext *src){
2686     int i;
2687
2688     MERGE(dct_count[0]); //note, the other dct vars are not part of the context
2689     MERGE(dct_count[1]);
2690     MERGE(mv_bits);
2691     MERGE(i_tex_bits);
2692     MERGE(p_tex_bits);
2693     MERGE(i_count);
2694     MERGE(f_count);
2695     MERGE(b_count);
2696     MERGE(skip_count);
2697     MERGE(misc_bits);
2698     MERGE(error_count);
2699     MERGE(padding_bug_score);
2700     MERGE(current_picture.error[0]);
2701     MERGE(current_picture.error[1]);
2702     MERGE(current_picture.error[2]);
2703
2704     if(dst->avctx->noise_reduction){
2705         for(i=0; i<64; i++){
2706             MERGE(dct_error_sum[0][i]);
2707             MERGE(dct_error_sum[1][i]);
2708         }
2709     }
2710
2711     assert(put_bits_count(&src->pb) % 8 ==0);
2712     assert(put_bits_count(&dst->pb) % 8 ==0);
2713     ff_copy_bits(&dst->pb, src->pb.buf, put_bits_count(&src->pb));
2714     flush_put_bits(&dst->pb);
2715 }
2716
2717 static int estimate_qp(MpegEncContext *s, int dry_run){
2718     if (s->next_lambda){
2719         s->current_picture_ptr->quality=
2720         s->current_picture.quality = s->next_lambda;
2721         if(!dry_run) s->next_lambda= 0;
2722     } else if (!s->fixed_qscale) {
2723         s->current_picture_ptr->quality=
2724         s->current_picture.quality = ff_rate_estimate_qscale(s, dry_run);
2725         if (s->current_picture.quality < 0)
2726             return -1;
2727     }
2728
2729     if(s->adaptive_quant){
2730         switch(s->codec_id){
2731         case CODEC_ID_MPEG4:
2732             if (CONFIG_MPEG4_ENCODER)
2733                 ff_clean_mpeg4_qscales(s);
2734             break;
2735         case CODEC_ID_H263:
2736         case CODEC_ID_H263P:
2737         case CODEC_ID_FLV1:
2738             if (CONFIG_H263_ENCODER)
2739                 ff_clean_h263_qscales(s);
2740             break;
2741         default:
2742             ff_init_qscale_tab(s);
2743         }
2744
2745         s->lambda= s->lambda_table[0];
2746         //FIXME broken
2747     }else
2748         s->lambda= s->current_picture.quality;
2749 //printf("%d %d\n", s->avctx->global_quality, s->current_picture.quality);
2750     update_qscale(s);
2751     return 0;
2752 }
2753
2754 /* must be called before writing the header */
2755 static void set_frame_distances(MpegEncContext * s){
2756     assert(s->current_picture_ptr->pts != AV_NOPTS_VALUE);
2757     s->time= s->current_picture_ptr->pts*s->avctx->time_base.num;
2758
2759     if(s->pict_type==FF_B_TYPE){
2760         s->pb_time= s->pp_time - (s->last_non_b_time - s->time);
2761         assert(s->pb_time > 0 && s->pb_time < s->pp_time);
2762     }else{
2763         s->pp_time= s->time - s->last_non_b_time;
2764         s->last_non_b_time= s->time;
2765         assert(s->picture_number==0 || s->pp_time > 0);
2766     }
2767 }
2768
2769 static int encode_picture(MpegEncContext *s, int picture_number)
2770 {
2771     int i;
2772     int bits;
2773
2774     s->picture_number = picture_number;
2775
2776     /* Reset the average MB variance */
2777     s->me.mb_var_sum_temp    =
2778     s->me.mc_mb_var_sum_temp = 0;
2779
2780     /* we need to initialize some time vars before we can encode b-frames */
2781     // RAL: Condition added for MPEG1VIDEO
2782     if (s->codec_id == CODEC_ID_MPEG1VIDEO || s->codec_id == CODEC_ID_MPEG2VIDEO || (s->h263_pred && !s->h263_msmpeg4))
2783         set_frame_distances(s);
2784     if(CONFIG_MPEG4_ENCODER && s->codec_id == CODEC_ID_MPEG4)
2785         ff_set_mpeg4_time(s);
2786
2787     s->me.scene_change_score=0;
2788
2789 //    s->lambda= s->current_picture_ptr->quality; //FIXME qscale / ... stuff for ME rate distortion
2790
2791     if(s->pict_type==FF_I_TYPE){
2792         if(s->msmpeg4_version >= 3) s->no_rounding=1;
2793         else                        s->no_rounding=0;
2794     }else if(s->pict_type!=FF_B_TYPE){
2795         if(s->flipflop_rounding || s->codec_id == CODEC_ID_H263P || s->codec_id == CODEC_ID_MPEG4)
2796             s->no_rounding ^= 1;
2797     }
2798
2799     if(s->flags & CODEC_FLAG_PASS2){
2800         if (estimate_qp(s,1) < 0)
2801             return -1;
2802         ff_get_2pass_fcode(s);
2803     }else if(!(s->flags & CODEC_FLAG_QSCALE)){
2804         if(s->pict_type==FF_B_TYPE)
2805             s->lambda= s->last_lambda_for[s->pict_type];
2806         else
2807             s->lambda= s->last_lambda_for[s->last_non_b_pict_type];
2808         update_qscale(s);
2809     }
2810
2811     s->mb_intra=0; //for the rate distortion & bit compare functions
2812     for(i=1; i<s->avctx->thread_count; i++){
2813         ff_update_duplicate_context(s->thread_context[i], s);
2814     }
2815
2816     if(ff_init_me(s)<0)
2817         return -1;
2818
2819     /* Estimate motion for every MB */
2820     if(s->pict_type != FF_I_TYPE){
2821         s->lambda = (s->lambda * s->avctx->me_penalty_compensation + 128)>>8;
2822         s->lambda2= (s->lambda2* (int64_t)s->avctx->me_penalty_compensation + 128)>>8;
2823         if(s->pict_type != FF_B_TYPE && s->avctx->me_threshold==0){
2824             if((s->avctx->pre_me && s->last_non_b_pict_type==FF_I_TYPE) || s->avctx->pre_me==2){
2825                 s->avctx->execute(s->avctx, pre_estimate_motion_thread, &s->thread_context[0], NULL, s->avctx->thread_count, sizeof(void*));
2826             }
2827         }
2828
2829         s->avctx->execute(s->avctx, estimate_motion_thread, &s->thread_context[0], NULL, s->avctx->thread_count, sizeof(void*));
2830     }else /* if(s->pict_type == FF_I_TYPE) */{
2831         /* I-Frame */
2832         for(i=0; i<s->mb_stride*s->mb_height; i++)
2833             s->mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
2834
2835         if(!s->fixed_qscale){
2836             /* finding spatial complexity for I-frame rate control */
2837             s->avctx->execute(s->avctx, mb_var_thread, &s->thread_context[0], NULL, s->avctx->thread_count, sizeof(void*));
2838         }
2839     }
2840     for(i=1; i<s->avctx->thread_count; i++){
2841         merge_context_after_me(s, s->thread_context[i]);
2842     }
2843     s->current_picture.mc_mb_var_sum= s->current_picture_ptr->mc_mb_var_sum= s->me.mc_mb_var_sum_temp;
2844     s->current_picture.   mb_var_sum= s->current_picture_ptr->   mb_var_sum= s->me.   mb_var_sum_temp;
2845     emms_c();
2846
2847     if(s->me.scene_change_score > s->avctx->scenechange_threshold && s->pict_type == FF_P_TYPE){
2848         s->pict_type= FF_I_TYPE;
2849         for(i=0; i<s->mb_stride*s->mb_height; i++)
2850             s->mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
2851 //printf("Scene change detected, encoding as I Frame %d %d\n", s->current_picture.mb_var_sum, s->current_picture.mc_mb_var_sum);
2852     }
2853
2854     if(!s->umvplus){
2855         if(s->pict_type==FF_P_TYPE || s->pict_type==FF_S_TYPE) {
2856             s->f_code= ff_get_best_fcode(s, s->p_mv_table, CANDIDATE_MB_TYPE_INTER);
2857
2858             if(s->flags & CODEC_FLAG_INTERLACED_ME){
2859                 int a,b;
2860                 a= ff_get_best_fcode(s, s->p_field_mv_table[0][0], CANDIDATE_MB_TYPE_INTER_I); //FIXME field_select
2861                 b= ff_get_best_fcode(s, s->p_field_mv_table[1][1], CANDIDATE_MB_TYPE_INTER_I);
2862                 s->f_code= FFMAX3(s->f_code, a, b);
2863             }
2864
2865             ff_fix_long_p_mvs(s);
2866             ff_fix_long_mvs(s, NULL, 0, s->p_mv_table, s->f_code, CANDIDATE_MB_TYPE_INTER, 0);
2867             if(s->flags & CODEC_FLAG_INTERLACED_ME){
2868                 int j;
2869                 for(i=0; i<2; i++){
2870                     for(j=0; j<2; j++)
2871                         ff_fix_long_mvs(s, s->p_field_select_table[i], j,
2872                                         s->p_field_mv_table[i][j], s->f_code, CANDIDATE_MB_TYPE_INTER_I, 0);
2873                 }
2874             }
2875         }
2876
2877         if(s->pict_type==FF_B_TYPE){
2878             int a, b;
2879
2880             a = ff_get_best_fcode(s, s->b_forw_mv_table, CANDIDATE_MB_TYPE_FORWARD);
2881             b = ff_get_best_fcode(s, s->b_bidir_forw_mv_table, CANDIDATE_MB_TYPE_BIDIR);
2882             s->f_code = FFMAX(a, b);
2883
2884             a = ff_get_best_fcode(s, s->b_back_mv_table, CANDIDATE_MB_TYPE_BACKWARD);
2885             b = ff_get_best_fcode(s, s->b_bidir_back_mv_table, CANDIDATE_MB_TYPE_BIDIR);
2886             s->b_code = FFMAX(a, b);
2887
2888             ff_fix_long_mvs(s, NULL, 0, s->b_forw_mv_table, s->f_code, CANDIDATE_MB_TYPE_FORWARD, 1);
2889             ff_fix_long_mvs(s, NULL, 0, s->b_back_mv_table, s->b_code, CANDIDATE_MB_TYPE_BACKWARD, 1);
2890             ff_fix_long_mvs(s, NULL, 0, s->b_bidir_forw_mv_table, s->f_code, CANDIDATE_MB_TYPE_BIDIR, 1);
2891             ff_fix_long_mvs(s, NULL, 0, s->b_bidir_back_mv_table, s->b_code, CANDIDATE_MB_TYPE_BIDIR, 1);
2892             if(s->flags & CODEC_FLAG_INTERLACED_ME){
2893                 int dir, j;
2894                 for(dir=0; dir<2; dir++){
2895                     for(i=0; i<2; i++){
2896                         for(j=0; j<2; j++){
2897                             int type= dir ? (CANDIDATE_MB_TYPE_BACKWARD_I|CANDIDATE_MB_TYPE_BIDIR_I)
2898                                           : (CANDIDATE_MB_TYPE_FORWARD_I |CANDIDATE_MB_TYPE_BIDIR_I);
2899                             ff_fix_long_mvs(s, s->b_field_select_table[dir][i], j,
2900                                             s->b_field_mv_table[dir][i][j], dir ? s->b_code : s->f_code, type, 1);
2901                         }
2902                     }
2903                 }
2904             }
2905         }
2906     }
2907
2908     if (estimate_qp(s, 0) < 0)
2909         return -1;
2910
2911     if(s->qscale < 3 && s->max_qcoeff<=128 && s->pict_type==FF_I_TYPE && !(s->flags & CODEC_FLAG_QSCALE))
2912         s->qscale= 3; //reduce clipping problems
2913
2914     if (s->out_format == FMT_MJPEG) {
2915         /* for mjpeg, we do include qscale in the matrix */
2916         for(i=1;i<64;i++){
2917             int j= s->dsp.idct_permutation[i];
2918
2919             s->intra_matrix[j] = av_clip_uint8((ff_mpeg1_default_intra_matrix[i] * s->qscale) >> 3);
2920         }
2921         s->y_dc_scale_table=
2922         s->c_dc_scale_table= ff_mpeg2_dc_scale_table[s->intra_dc_precision];
2923         s->intra_matrix[0] = ff_mpeg2_dc_scale_table[s->intra_dc_precision][8];
2924         ff_convert_matrix(&s->dsp, s->q_intra_matrix, s->q_intra_matrix16,
2925                        s->intra_matrix, s->intra_quant_bias, 8, 8, 1);
2926         s->qscale= 8;
2927     }
2928
2929     //FIXME var duplication
2930     s->current_picture_ptr->key_frame=
2931     s->current_picture.key_frame= s->pict_type == FF_I_TYPE; //FIXME pic_ptr
2932     s->current_picture_ptr->pict_type=
2933     s->current_picture.pict_type= s->pict_type;
2934
2935     if(s->current_picture.key_frame)
2936         s->picture_in_gop_number=0;
2937
2938     s->last_bits= put_bits_count(&s->pb);
2939     switch(s->out_format) {
2940     case FMT_MJPEG:
2941         if (CONFIG_MJPEG_ENCODER)
2942             ff_mjpeg_encode_picture_header(s);
2943         break;
2944     case FMT_H261:
2945         if (CONFIG_H261_ENCODER)
2946             ff_h261_encode_picture_header(s, picture_number);
2947         break;
2948     case FMT_H263:
2949         if (CONFIG_WMV2_ENCODER && s->codec_id == CODEC_ID_WMV2)
2950             ff_wmv2_encode_picture_header(s, picture_number);
2951         else if (CONFIG_MSMPEG4_ENCODER && s->h263_msmpeg4)
2952             msmpeg4_encode_picture_header(s, picture_number);
2953         else if (CONFIG_MPEG4_ENCODER && s->h263_pred)
2954             mpeg4_encode_picture_header(s, picture_number);
2955         else if (CONFIG_RV10_ENCODER && s->codec_id == CODEC_ID_RV10)
2956             rv10_encode_picture_header(s, picture_number);
2957         else if (CONFIG_RV20_ENCODER && s->codec_id == CODEC_ID_RV20)
2958             rv20_encode_picture_header(s, picture_number);
2959         else if (CONFIG_FLV_ENCODER && s->codec_id == CODEC_ID_FLV1)
2960             ff_flv_encode_picture_header(s, picture_number);
2961         else if (CONFIG_H263_ENCODER)
2962             h263_encode_picture_header(s, picture_number);
2963         break;
2964     case FMT_MPEG1:
2965         if (CONFIG_MPEG1VIDEO_ENCODER || CONFIG_MPEG2VIDEO_ENCODER)
2966             mpeg1_encode_picture_header(s, picture_number);
2967         break;
2968     case FMT_H264:
2969         break;
2970     default:
2971         assert(0);
2972     }
2973     bits= put_bits_count(&s->pb);
2974     s->header_bits= bits - s->last_bits;
2975
2976     for(i=1; i<s->avctx->thread_count; i++){
2977         update_duplicate_context_after_me(s->thread_context[i], s);
2978     }
2979     s->avctx->execute(s->avctx, encode_thread, &s->thread_context[0], NULL, s->avctx->thread_count, sizeof(void*));
2980     for(i=1; i<s->avctx->thread_count; i++){
2981         merge_context_after_encode(s, s->thread_context[i]);
2982     }
2983     emms_c();
2984     return 0;
2985 }
2986
2987 static void denoise_dct_c(MpegEncContext *s, DCTELEM *block){
2988     const int intra= s->mb_intra;
2989     int i;
2990
2991     s->dct_count[intra]++;
2992
2993     for(i=0; i<64; i++){
2994         int level= block[i];
2995
2996         if(level){
2997             if(level>0){
2998                 s->dct_error_sum[intra][i] += level;
2999                 level -= s->dct_offset[intra][i];
3000                 if(level<0) level=0;
3001             }else{
3002                 s->dct_error_sum[intra][i] -= level;
3003                 level += s->dct_offset[intra][i];
3004                 if(level>0) level=0;
3005             }
3006             block[i]= level;
3007         }
3008     }
3009 }
3010
3011 static int dct_quantize_trellis_c(MpegEncContext *s,
3012                                   DCTELEM *block, int n,
3013                                   int qscale, int *overflow){
3014     const int *qmat;
3015     const uint8_t *scantable= s->intra_scantable.scantable;
3016     const uint8_t *perm_scantable= s->intra_scantable.permutated;
3017     int max=0;
3018     unsigned int threshold1, threshold2;
3019     int bias=0;
3020     int run_tab[65];
3021     int level_tab[65];
3022     int score_tab[65];
3023     int survivor[65];
3024     int survivor_count;
3025     int last_run=0;
3026     int last_level=0;
3027     int last_score= 0;
3028     int last_i;
3029     int coeff[2][64];
3030     int coeff_count[64];
3031     int qmul, qadd, start_i, last_non_zero, i, dc;
3032     const int esc_length= s->ac_esc_length;
3033     uint8_t * length;
3034     uint8_t * last_length;
3035     const int lambda= s->lambda2 >> (FF_LAMBDA_SHIFT - 6);
3036
3037     s->dsp.fdct (block);
3038
3039     if(s->dct_error_sum)
3040         s->denoise_dct(s, block);
3041     qmul= qscale*16;
3042     qadd= ((qscale-1)|1)*8;
3043
3044     if (s->mb_intra) {
3045         int q;
3046         if (!s->h263_aic) {
3047             if (n < 4)
3048                 q = s->y_dc_scale;
3049             else
3050                 q = s->c_dc_scale;
3051             q = q << 3;
3052         } else{
3053             /* For AIC we skip quant/dequant of INTRADC */
3054             q = 1 << 3;
3055             qadd=0;
3056         }
3057
3058         /* note: block[0] is assumed to be positive */
3059         block[0] = (block[0] + (q >> 1)) / q;
3060         start_i = 1;
3061         last_non_zero = 0;
3062         qmat = s->q_intra_matrix[qscale];
3063         if(s->mpeg_quant || s->out_format == FMT_MPEG1)
3064             bias= 1<<(QMAT_SHIFT-1);
3065         length     = s->intra_ac_vlc_length;
3066         last_length= s->intra_ac_vlc_last_length;
3067     } else {
3068         start_i = 0;
3069         last_non_zero = -1;
3070         qmat = s->q_inter_matrix[qscale];
3071         length     = s->inter_ac_vlc_length;
3072         last_length= s->inter_ac_vlc_last_length;
3073     }
3074     last_i= start_i;
3075
3076     threshold1= (1<<QMAT_SHIFT) - bias - 1;
3077     threshold2= (threshold1<<1);
3078
3079     for(i=63; i>=start_i; i--) {
3080         const int j = scantable[i];
3081         int level = block[j] * qmat[j];
3082
3083         if(((unsigned)(level+threshold1))>threshold2){
3084             last_non_zero = i;
3085             break;
3086         }
3087     }
3088
3089     for(i=start_i; i<=last_non_zero; i++) {
3090         const int j = scantable[i];
3091         int level = block[j] * qmat[j];
3092
3093 //        if(   bias+level >= (1<<(QMAT_SHIFT - 3))
3094 //           || bias-level >= (1<<(QMAT_SHIFT - 3))){
3095         if(((unsigned)(level+threshold1))>threshold2){
3096             if(level>0){
3097                 level= (bias + level)>>QMAT_SHIFT;
3098                 coeff[0][i]= level;
3099                 coeff[1][i]= level-1;
3100 //                coeff[2][k]= level-2;
3101             }else{
3102                 level= (bias - level)>>QMAT_SHIFT;
3103                 coeff[0][i]= -level;
3104                 coeff[1][i]= -level+1;
3105 //                coeff[2][k]= -level+2;
3106             }
3107             coeff_count[i]= FFMIN(level, 2);
3108             assert(coeff_count[i]);
3109             max |=level;
3110         }else{
3111             coeff[0][i]= (level>>31)|1;
3112             coeff_count[i]= 1;
3113         }
3114     }
3115
3116     *overflow= s->max_qcoeff < max; //overflow might have happened
3117
3118     if(last_non_zero < start_i){
3119         memset(block + start_i, 0, (64-start_i)*sizeof(DCTELEM));
3120         return last_non_zero;
3121     }
3122
3123     score_tab[start_i]= 0;
3124     survivor[0]= start_i;
3125     survivor_count= 1;
3126
3127     for(i=start_i; i<=last_non_zero; i++){
3128         int level_index, j, zero_distortion;
3129         int dct_coeff= FFABS(block[ scantable[i] ]);
3130         int best_score=256*256*256*120;
3131
3132         if (   s->dsp.fdct == fdct_ifast
3133 #ifndef FAAN_POSTSCALE
3134             || s->dsp.fdct == ff_faandct
3135 #endif
3136            )
3137             dct_coeff= (dct_coeff*ff_inv_aanscales[ scantable[i] ]) >> 12;
3138         zero_distortion= dct_coeff*dct_coeff;
3139
3140         for(level_index=0; level_index < coeff_count[i]; level_index++){
3141             int distortion;
3142             int level= coeff[level_index][i];
3143             const int alevel= FFABS(level);
3144             int unquant_coeff;
3145
3146             assert(level);
3147
3148             if(s->out_format == FMT_H263){
3149                 unquant_coeff= alevel*qmul + qadd;
3150             }else{ //MPEG1
3151                 j= s->dsp.idct_permutation[ scantable[i] ]; //FIXME optimize
3152                 if(s->mb_intra){
3153                         unquant_coeff = (int)(  alevel  * qscale * s->intra_matrix[j]) >> 3;
3154                         unquant_coeff =   (unquant_coeff - 1) | 1;
3155                 }else{
3156                         unquant_coeff = (((  alevel  << 1) + 1) * qscale * ((int) s->inter_matrix[j])) >> 4;
3157                         unquant_coeff =   (unquant_coeff - 1) | 1;
3158                 }
3159                 unquant_coeff<<= 3;
3160             }
3161
3162             distortion= (unquant_coeff - dct_coeff) * (unquant_coeff - dct_coeff) - zero_distortion;
3163             level+=64;
3164             if((level&(~127)) == 0){
3165                 for(j=survivor_count-1; j>=0; j--){
3166                     int run= i - survivor[j];
3167                     int score= distortion + length[UNI_AC_ENC_INDEX(run, level)]*lambda;
3168                     score += score_tab[i-run];
3169
3170                     if(score < best_score){
3171                         best_score= score;
3172                         run_tab[i+1]= run;
3173                         level_tab[i+1]= level-64;
3174                     }
3175                 }
3176
3177                 if(s->out_format == FMT_H263){
3178                     for(j=survivor_count-1; j>=0; j--){
3179                         int run= i - survivor[j];
3180                         int score= distortion + last_length[UNI_AC_ENC_INDEX(run, level)]*lambda;
3181                         score += score_tab[i-run];
3182                         if(score < last_score){
3183                             last_score= score;
3184                             last_run= run;
3185                             last_level= level-64;
3186                             last_i= i+1;
3187                         }
3188                     }
3189                 }
3190             }else{
3191                 distortion += esc_length*lambda;
3192                 for(j=survivor_count-1; j>=0; j--){
3193                     int run= i - survivor[j];
3194                     int score= distortion + score_tab[i-run];
3195
3196                     if(score < best_score){
3197                         best_score= score;
3198                         run_tab[i+1]= run;
3199                         level_tab[i+1]= level-64;
3200                     }
3201                 }
3202
3203                 if(s->out_format == FMT_H263){
3204                   for(j=survivor_count-1; j>=0; j--){
3205                         int run= i - survivor[j];
3206                         int score= distortion + score_tab[i-run];
3207                         if(score < last_score){
3208                             last_score= score;
3209                             last_run= run;
3210                             last_level= level-64;
3211                             last_i= i+1;
3212                         }
3213                     }
3214                 }
3215             }
3216         }
3217
3218         score_tab[i+1]= best_score;
3219
3220         //Note: there is a vlc code in mpeg4 which is 1 bit shorter then another one with a shorter run and the same level
3221         if(last_non_zero <= 27){
3222             for(; survivor_count; survivor_count--){
3223                 if(score_tab[ survivor[survivor_count-1] ] <= best_score)
3224                     break;
3225             }
3226         }else{
3227             for(; survivor_count; survivor_count--){
3228                 if(score_tab[ survivor[survivor_count-1] ] <= best_score + lambda)
3229                     break;
3230             }
3231         }
3232
3233         survivor[ survivor_count++ ]= i+1;
3234     }
3235
3236     if(s->out_format != FMT_H263){
3237         last_score= 256*256*256*120;
3238         for(i= survivor[0]; i<=last_non_zero + 1; i++){
3239             int score= score_tab[i];
3240             if(i) score += lambda*2; //FIXME exacter?
3241
3242             if(score < last_score){
3243                 last_score= score;
3244                 last_i= i;
3245                 last_level= level_tab[i];
3246                 last_run= run_tab[i];
3247             }
3248         }
3249     }
3250
3251     s->coded_score[n] = last_score;
3252
3253     dc= FFABS(block[0]);
3254     last_non_zero= last_i - 1;
3255     memset(block + start_i, 0, (64-start_i)*sizeof(DCTELEM));
3256
3257     if(last_non_zero < start_i)
3258         return last_non_zero;
3259
3260     if(last_non_zero == 0 && start_i == 0){
3261         int best_level= 0;
3262         int best_score= dc * dc;
3263
3264         for(i=0; i<coeff_count[0]; i++){
3265             int level= coeff[i][0];
3266             int alevel= FFABS(level);
3267             int unquant_coeff, score, distortion;
3268
3269             if(s->out_format == FMT_H263){
3270                     unquant_coeff= (alevel*qmul + qadd)>>3;
3271             }else{ //MPEG1
3272                     unquant_coeff = (((  alevel  << 1) + 1) * qscale * ((int) s->inter_matrix[0])) >> 4;
3273                     unquant_coeff =   (unquant_coeff - 1) | 1;
3274             }
3275             unquant_coeff = (unquant_coeff + 4) >> 3;
3276             unquant_coeff<<= 3 + 3;
3277
3278             distortion= (unquant_coeff - dc) * (unquant_coeff - dc);
3279             level+=64;
3280             if((level&(~127)) == 0) score= distortion + last_length[UNI_AC_ENC_INDEX(0, level)]*lambda;
3281             else                    score= distortion + esc_length*lambda;
3282
3283             if(score < best_score){
3284                 best_score= score;
3285                 best_level= level - 64;
3286             }
3287         }
3288         block[0]= best_level;
3289         s->coded_score[n] = best_score - dc*dc;
3290         if(best_level == 0) return -1;
3291         else                return last_non_zero;
3292     }
3293
3294     i= last_i;
3295     assert(last_level);
3296
3297     block[ perm_scantable[last_non_zero] ]= last_level;
3298     i -= last_run + 1;
3299
3300     for(; i>start_i; i -= run_tab[i] + 1){
3301         block[ perm_scantable[i-1] ]= level_tab[i];
3302     }
3303
3304     return last_non_zero;
3305 }
3306
3307 //#define REFINE_STATS 1
3308 static int16_t basis[64][64];
3309
3310 static void build_basis(uint8_t *perm){
3311     int i, j, x, y;
3312     emms_c();
3313     for(i=0; i<8; i++){
3314         for(j=0; j<8; j++){
3315             for(y=0; y<8; y++){
3316                 for(x=0; x<8; x++){
3317                     double s= 0.25*(1<<BASIS_SHIFT);
3318                     int index= 8*i + j;
3319                     int perm_index= perm[index];
3320                     if(i==0) s*= sqrt(0.5);
3321                     if(j==0) s*= sqrt(0.5);
3322                     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)));
3323                 }
3324             }
3325         }
3326     }
3327 }
3328
3329 static int dct_quantize_refine(MpegEncContext *s, //FIXME breaks denoise?
3330                         DCTELEM *block, int16_t *weight, DCTELEM *orig,
3331                         int n, int qscale){
3332     int16_t rem[64];
3333     LOCAL_ALIGNED_16(DCTELEM, d1, [64]);
3334     const uint8_t *scantable= s->intra_scantable.scantable;
3335     const uint8_t *perm_scantable= s->intra_scantable.permutated;
3336 //    unsigned int threshold1, threshold2;
3337 //    int bias=0;
3338     int run_tab[65];
3339     int prev_run=0;
3340     int prev_level=0;
3341     int qmul, qadd, start_i, last_non_zero, i, dc;
3342     uint8_t * length;
3343     uint8_t * last_length;
3344     int lambda;
3345     int rle_index, run, q = 1, sum; //q is only used when s->mb_intra is true
3346 #ifdef REFINE_STATS
3347 static int count=0;
3348 static int after_last=0;
3349 static int to_zero=0;
3350 static int from_zero=0;
3351 static int raise=0;
3352 static int lower=0;
3353 static int messed_sign=0;
3354 #endif
3355
3356     if(basis[0][0] == 0)
3357         build_basis(s->dsp.idct_permutation);
3358
3359     qmul= qscale*2;
3360     qadd= (qscale-1)|1;
3361     if (s->mb_intra) {
3362         if (!s->h263_aic) {
3363             if (n < 4)
3364                 q = s->y_dc_scale;
3365             else
3366                 q = s->c_dc_scale;
3367         } else{
3368             /* For AIC we skip quant/dequant of INTRADC */
3369             q = 1;
3370             qadd=0;
3371         }
3372         q <<= RECON_SHIFT-3;
3373         /* note: block[0] is assumed to be positive */
3374         dc= block[0]*q;
3375 //        block[0] = (block[0] + (q >> 1)) / q;
3376         start_i = 1;
3377 //        if(s->mpeg_quant || s->out_format == FMT_MPEG1)
3378 //            bias= 1<<(QMAT_SHIFT-1);
3379         length     = s->intra_ac_vlc_length;
3380         last_length= s->intra_ac_vlc_last_length;
3381     } else {
3382         dc= 0;
3383         start_i = 0;
3384         length     = s->inter_ac_vlc_length;
3385         last_length= s->inter_ac_vlc_last_length;
3386     }
3387     last_non_zero = s->block_last_index[n];
3388
3389 #ifdef REFINE_STATS
3390 {START_TIMER
3391 #endif
3392     dc += (1<<(RECON_SHIFT-1));
3393     for(i=0; i<64; i++){
3394         rem[i]= dc - (orig[i]<<RECON_SHIFT); //FIXME  use orig dirrectly instead of copying to rem[]
3395     }
3396 #ifdef REFINE_STATS
3397 STOP_TIMER("memset rem[]")}
3398 #endif
3399     sum=0;
3400     for(i=0; i<64; i++){
3401         int one= 36;
3402         int qns=4;
3403         int w;
3404
3405         w= FFABS(weight[i]) + qns*one;
3406         w= 15 + (48*qns*one + w/2)/w; // 16 .. 63
3407
3408         weight[i] = w;
3409 //        w=weight[i] = (63*qns + (w/2)) / w;
3410
3411         assert(w>0);
3412         assert(w<(1<<6));
3413         sum += w*w;
3414     }
3415     lambda= sum*(uint64_t)s->lambda2 >> (FF_LAMBDA_SHIFT - 6 + 6 + 6 + 6);
3416 #ifdef REFINE_STATS
3417 {START_TIMER
3418 #endif
3419     run=0;
3420     rle_index=0;
3421     for(i=start_i; i<=last_non_zero; i++){
3422         int j= perm_scantable[i];
3423         const int level= block[j];
3424         int coeff;
3425
3426         if(level){
3427             if(level<0) coeff= qmul*level - qadd;
3428             else        coeff= qmul*level + qadd;
3429             run_tab[rle_index++]=run;
3430             run=0;
3431
3432             s->dsp.add_8x8basis(rem, basis[j], coeff);
3433         }else{
3434             run++;
3435         }
3436     }
3437 #ifdef REFINE_STATS
3438 if(last_non_zero>0){
3439 STOP_TIMER("init rem[]")
3440 }
3441 }
3442
3443 {START_TIMER
3444 #endif
3445     for(;;){
3446         int best_score=s->dsp.try_8x8basis(rem, weight, basis[0], 0);
3447         int best_coeff=0;
3448         int best_change=0;
3449         int run2, best_unquant_change=0, analyze_gradient;
3450 #ifdef REFINE_STATS
3451 {START_TIMER
3452 #endif
3453         analyze_gradient = last_non_zero > 2 || s->avctx->quantizer_noise_shaping >= 3;
3454
3455         if(analyze_gradient){
3456 #ifdef REFINE_STATS
3457 {START_TIMER
3458 #endif
3459             for(i=0; i<64; i++){
3460                 int w= weight[i];
3461
3462                 d1[i] = (rem[i]*w*w + (1<<(RECON_SHIFT+12-1)))>>(RECON_SHIFT+12);
3463             }
3464 #ifdef REFINE_STATS
3465 STOP_TIMER("rem*w*w")}
3466 {START_TIMER
3467 #endif
3468             s->dsp.fdct(d1);
3469 #ifdef REFINE_STATS
3470 STOP_TIMER("dct")}
3471 #endif
3472         }
3473
3474         if(start_i){
3475             const int level= block[0];
3476             int change, old_coeff;
3477
3478             assert(s->mb_intra);
3479
3480             old_coeff= q*level;
3481
3482             for(change=-1; change<=1; change+=2){
3483                 int new_level= level + change;
3484                 int score, new_coeff;
3485
3486                 new_coeff= q*new_level;
3487                 if(new_coeff >= 2048 || new_coeff < 0)
3488                     continue;
3489
3490                 score= s->dsp.try_8x8basis(rem, weight, basis[0], new_coeff - old_coeff);
3491                 if(score<best_score){
3492                     best_score= score;
3493                     best_coeff= 0;
3494                     best_change= change;
3495                     best_unquant_change= new_coeff - old_coeff;
3496                 }
3497             }
3498         }
3499
3500         run=0;
3501         rle_index=0;
3502         run2= run_tab[rle_index++];
3503         prev_level=0;
3504         prev_run=0;
3505
3506         for(i=start_i; i<64; i++){
3507             int j= perm_scantable[i];
3508             const int level= block[j];
3509             int change, old_coeff;
3510
3511             if(s->avctx->quantizer_noise_shaping < 3 && i > last_non_zero + 1)
3512                 break;
3513
3514             if(level){
3515                 if(level<0) old_coeff= qmul*level - qadd;
3516                 else        old_coeff= qmul*level + qadd;
3517                 run2= run_tab[rle_index++]; //FIXME ! maybe after last
3518             }else{
3519                 old_coeff=0;
3520                 run2--;
3521                 assert(run2>=0 || i >= last_non_zero );
3522             }
3523
3524             for(change=-1; change<=1; change+=2){
3525                 int new_level= level + change;
3526                 int score, new_coeff, unquant_change;
3527
3528                 score=0;
3529                 if(s->avctx->quantizer_noise_shaping < 2 && FFABS(new_level) > FFABS(level))
3530                    continue;
3531
3532                 if(new_level){
3533                     if(new_level<0) new_coeff= qmul*new_level - qadd;
3534                     else            new_coeff= qmul*new_level + qadd;
3535                     if(new_coeff >= 2048 || new_coeff <= -2048)
3536                         continue;
3537                     //FIXME check for overflow
3538
3539                     if(level){
3540                         if(level < 63 && level > -63){
3541                             if(i < last_non_zero)
3542                                 score +=   length[UNI_AC_ENC_INDEX(run, new_level+64)]
3543                                          - length[UNI_AC_ENC_INDEX(run, level+64)];
3544                             else
3545                                 score +=   last_length[UNI_AC_ENC_INDEX(run, new_level+64)]
3546                                          - last_length[UNI_AC_ENC_INDEX(run, level+64)];
3547                         }
3548                     }else{
3549                         assert(FFABS(new_level)==1);
3550
3551                         if(analyze_gradient){
3552                             int g= d1[ scantable[i] ];
3553                             if(g && (g^new_level) >= 0)
3554                                 continue;
3555                         }
3556
3557                         if(i < last_non_zero){
3558                             int next_i= i + run2 + 1;
3559                             int next_level= block[ perm_scantable[next_i] ] + 64;
3560
3561                             if(next_level&(~127))
3562                                 next_level= 0;
3563
3564                             if(next_i < last_non_zero)
3565                                 score +=   length[UNI_AC_ENC_INDEX(run, 65)]
3566                                          + length[UNI_AC_ENC_INDEX(run2, next_level)]
3567                                          - length[UNI_AC_ENC_INDEX(run + run2 + 1, next_level)];
3568                             else
3569                                 score +=  length[UNI_AC_ENC_INDEX(run, 65)]
3570                                         + last_length[UNI_AC_ENC_INDEX(run2, next_level)]
3571                                         - last_length[UNI_AC_ENC_INDEX(run + run2 + 1, next_level)];
3572                         }else{
3573                             score += last_length[UNI_AC_ENC_INDEX(run, 65)];
3574                             if(prev_level){
3575                                 score +=  length[UNI_AC_ENC_INDEX(prev_run, prev_level)]
3576                                         - last_length[UNI_AC_ENC_INDEX(prev_run, prev_level)];
3577                             }
3578                         }
3579                     }
3580                 }else{
3581                     new_coeff=0;
3582                     assert(FFABS(level)==1);
3583
3584                     if(i < last_non_zero){
3585                         int next_i= i + run2 + 1;
3586                         int next_level= block[ perm_scantable[next_i] ] + 64;
3587
3588                         if(next_level&(~127))
3589                             next_level= 0;
3590
3591                         if(next_i < last_non_zero)
3592                             score +=   length[UNI_AC_ENC_INDEX(run + run2 + 1, next_level)]
3593                                      - length[UNI_AC_ENC_INDEX(run2, next_level)]
3594                                      - length[UNI_AC_ENC_INDEX(run, 65)];
3595                         else
3596                             score +=   last_length[UNI_AC_ENC_INDEX(run + run2 + 1, next_level)]
3597                                      - last_length[UNI_AC_ENC_INDEX(run2, next_level)]
3598                                      - length[UNI_AC_ENC_INDEX(run, 65)];
3599                     }else{
3600                         score += -last_length[UNI_AC_ENC_INDEX(run, 65)];
3601                         if(prev_level){
3602                             score +=  last_length[UNI_AC_ENC_INDEX(prev_run, prev_level)]
3603                                     - length[UNI_AC_ENC_INDEX(prev_run, prev_level)];
3604                         }
3605                     }
3606                 }
3607
3608                 score *= lambda;
3609
3610                 unquant_change= new_coeff - old_coeff;
3611                 assert((score < 100*lambda && score > -100*lambda) || lambda==0);
3612
3613                 score+= s->dsp.try_8x8basis(rem, weight, basis[j], unquant_change);
3614                 if(score<best_score){
3615                     best_score= score;
3616                     best_coeff= i;
3617                     best_change= change;
3618                     best_unquant_change= unquant_change;
3619                 }
3620             }
3621             if(level){
3622                 prev_level= level + 64;
3623                 if(prev_level&