h263/p encoder: add 'obmc' private option.
[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 Libav.
9  *
10  * Libav is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU Lesser General Public
12  * License as published by the Free Software Foundation; either
13  * version 2.1 of the License, or (at your option) any later version.
14  *
15  * Libav is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18  * Lesser General Public License for more details.
19  *
20  * You should have received a copy of the GNU Lesser General Public
21  * License along with Libav; if not, write to the Free Software
22  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23  */
24
25 /**
26  * @file
27  * The simplest mpeg encoder (well, it was the simplest!).
28  */
29
30 #include "libavutil/intmath.h"
31 #include "libavutil/mathematics.h"
32 #include "libavutil/opt.h"
33 #include "avcodec.h"
34 #include "dsputil.h"
35 #include "mpegvideo.h"
36 #include "mpegvideo_common.h"
37 #include "h263.h"
38 #include "mjpegenc.h"
39 #include "msmpeg4.h"
40 #include "faandct.h"
41 #include "thread.h"
42 #include "aandcttab.h"
43 #include "flv.h"
44 #include "mpeg4video.h"
45 #include "internal.h"
46 #include <limits.h>
47
48 //#undef NDEBUG
49 //#include <assert.h>
50
51 static int encode_picture(MpegEncContext *s, int picture_number);
52 static int dct_quantize_refine(MpegEncContext *s, DCTELEM *block, int16_t *weight, DCTELEM *orig, int n, int qscale);
53 static int sse_mb(MpegEncContext *s);
54 static void denoise_dct_c(MpegEncContext *s, DCTELEM *block);
55 static int dct_quantize_trellis_c(MpegEncContext *s, DCTELEM *block, int n, int qscale, int *overflow);
56
57 /* enable all paranoid tests for rounding, overflows, etc... */
58 //#define PARANOID
59
60 //#define DEBUG
61
62 static uint8_t default_mv_penalty[MAX_FCODE+1][MAX_MV*2+1];
63 static uint8_t default_fcode_tab[MAX_MV*2+1];
64
65 void ff_convert_matrix(DSPContext *dsp, int (*qmat)[64], uint16_t (*qmat16)[2][64],
66                            const uint16_t *quant_matrix, int bias, int qmin, int qmax, int intra)
67 {
68     int qscale;
69     int shift=0;
70
71     for(qscale=qmin; qscale<=qmax; qscale++){
72         int i;
73         if (dsp->fdct == ff_jpeg_fdct_islow_8 ||
74             dsp->fdct == ff_jpeg_fdct_islow_10
75 #ifdef FAAN_POSTSCALE
76             || dsp->fdct == ff_faandct
77 #endif
78             ) {
79             for(i=0;i<64;i++) {
80                 const int j= dsp->idct_permutation[i];
81                 /* 16 <= qscale * quant_matrix[i] <= 7905 */
82                 /* 19952             <= ff_aanscales[i] * qscale * quant_matrix[i]               <= 249205026 */
83                 /* (1 << 36) / 19952 >= (1 << 36) / (ff_aanscales[i] * qscale * quant_matrix[i]) >= (1 << 36) / 249205026 */
84                 /* 3444240           >= (1 << 36) / (ff_aanscales[i] * qscale * quant_matrix[i]) >= 275 */
85
86                 qmat[qscale][i] = (int)((UINT64_C(1) << QMAT_SHIFT) /
87                                 (qscale * quant_matrix[j]));
88             }
89         } else if (dsp->fdct == fdct_ifast
90 #ifndef FAAN_POSTSCALE
91                    || dsp->fdct == ff_faandct
92 #endif
93                    ) {
94             for(i=0;i<64;i++) {
95                 const int j= dsp->idct_permutation[i];
96                 /* 16 <= qscale * quant_matrix[i] <= 7905 */
97                 /* 19952             <= ff_aanscales[i] * qscale * quant_matrix[i]               <= 249205026 */
98                 /* (1 << 36) / 19952 >= (1 << 36) / (ff_aanscales[i] * qscale * quant_matrix[i]) >= (1<<36)/249205026 */
99                 /* 3444240           >= (1 << 36) / (ff_aanscales[i] * qscale * quant_matrix[i]) >= 275 */
100
101                 qmat[qscale][i] = (int)((UINT64_C(1) << (QMAT_SHIFT + 14)) /
102                                 (ff_aanscales[i] * qscale * quant_matrix[j]));
103             }
104         } else {
105             for(i=0;i<64;i++) {
106                 const int j= dsp->idct_permutation[i];
107                 /* We can safely suppose that 16 <= quant_matrix[i] <= 255
108                    So 16           <= qscale * quant_matrix[i]             <= 7905
109                    so (1<<19) / 16 >= (1<<19) / (qscale * quant_matrix[i]) >= (1<<19) / 7905
110                    so 32768        >= (1<<19) / (qscale * quant_matrix[i]) >= 67
111                 */
112                 qmat[qscale][i] = (int)((UINT64_C(1) << QMAT_SHIFT) / (qscale * quant_matrix[j]));
113 //                qmat  [qscale][i] = (1 << QMAT_SHIFT_MMX) / (qscale * quant_matrix[i]);
114                 qmat16[qscale][0][i] = (1 << QMAT_SHIFT_MMX) / (qscale * quant_matrix[j]);
115
116                 if(qmat16[qscale][0][i]==0 || qmat16[qscale][0][i]==128*256) qmat16[qscale][0][i]=128*256-1;
117                 qmat16[qscale][1][i]= ROUNDED_DIV(bias<<(16-QUANT_BIAS_SHIFT), qmat16[qscale][0][i]);
118             }
119         }
120
121         for(i=intra; i<64; i++){
122             int64_t max= 8191;
123             if (dsp->fdct == fdct_ifast
124 #ifndef FAAN_POSTSCALE
125                    || dsp->fdct == ff_faandct
126 #endif
127                    ) {
128                 max = (8191LL*ff_aanscales[i]) >> 14;
129             }
130             while(((max * qmat[qscale][i]) >> shift) > INT_MAX){
131                 shift++;
132             }
133         }
134     }
135     if(shift){
136         av_log(NULL, AV_LOG_INFO, "Warning, QMAT_SHIFT is larger than %d, overflows possible\n", QMAT_SHIFT - shift);
137     }
138 }
139
140 static inline void update_qscale(MpegEncContext *s){
141     s->qscale= (s->lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
142     s->qscale= av_clip(s->qscale, s->avctx->qmin, s->avctx->qmax);
143
144     s->lambda2= (s->lambda*s->lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
145 }
146
147 void ff_write_quant_matrix(PutBitContext *pb, uint16_t *matrix){
148     int i;
149
150     if(matrix){
151         put_bits(pb, 1, 1);
152         for(i=0;i<64;i++) {
153             put_bits(pb, 8, matrix[ ff_zigzag_direct[i] ]);
154         }
155     }else
156         put_bits(pb, 1, 0);
157 }
158
159 /**
160  * init s->current_picture.qscale_table from s->lambda_table
161  */
162 void ff_init_qscale_tab(MpegEncContext *s){
163     int8_t * const qscale_table = s->current_picture.f.qscale_table;
164     int i;
165
166     for(i=0; i<s->mb_num; i++){
167         unsigned int lam= s->lambda_table[ s->mb_index2xy[i] ];
168         int qp= (lam*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
169         qscale_table[ s->mb_index2xy[i] ]= av_clip(qp, s->avctx->qmin, s->avctx->qmax);
170     }
171 }
172
173 static void copy_picture_attributes(MpegEncContext *s, AVFrame *dst, AVFrame *src){
174     int i;
175
176     dst->pict_type              = src->pict_type;
177     dst->quality                = src->quality;
178     dst->coded_picture_number   = src->coded_picture_number;
179     dst->display_picture_number = src->display_picture_number;
180 //    dst->reference              = src->reference;
181     dst->pts                    = src->pts;
182     dst->interlaced_frame       = src->interlaced_frame;
183     dst->top_field_first        = src->top_field_first;
184
185     if(s->avctx->me_threshold){
186         if(!src->motion_val[0])
187             av_log(s->avctx, AV_LOG_ERROR, "AVFrame.motion_val not set!\n");
188         if(!src->mb_type)
189             av_log(s->avctx, AV_LOG_ERROR, "AVFrame.mb_type not set!\n");
190         if(!src->ref_index[0])
191             av_log(s->avctx, AV_LOG_ERROR, "AVFrame.ref_index not set!\n");
192         if(src->motion_subsample_log2 != dst->motion_subsample_log2)
193             av_log(s->avctx, AV_LOG_ERROR, "AVFrame.motion_subsample_log2 doesn't match! (%d!=%d)\n",
194             src->motion_subsample_log2, dst->motion_subsample_log2);
195
196         memcpy(dst->mb_type, src->mb_type, s->mb_stride * s->mb_height * sizeof(dst->mb_type[0]));
197
198         for(i=0; i<2; i++){
199             int stride= ((16*s->mb_width )>>src->motion_subsample_log2) + 1;
200             int height= ((16*s->mb_height)>>src->motion_subsample_log2);
201
202             if(src->motion_val[i] && src->motion_val[i] != dst->motion_val[i]){
203                 memcpy(dst->motion_val[i], src->motion_val[i], 2*stride*height*sizeof(int16_t));
204             }
205             if(src->ref_index[i] && src->ref_index[i] != dst->ref_index[i]){
206                 memcpy(dst->ref_index[i], src->ref_index[i], s->mb_stride*4*s->mb_height*sizeof(int8_t));
207             }
208         }
209     }
210 }
211
212 static void update_duplicate_context_after_me(MpegEncContext *dst, MpegEncContext *src){
213 #define COPY(a) dst->a= src->a
214     COPY(pict_type);
215     COPY(current_picture);
216     COPY(f_code);
217     COPY(b_code);
218     COPY(qscale);
219     COPY(lambda);
220     COPY(lambda2);
221     COPY(picture_in_gop_number);
222     COPY(gop_picture_number);
223     COPY(frame_pred_frame_dct); //FIXME don't set in encode_header
224     COPY(progressive_frame); //FIXME don't set in encode_header
225     COPY(partitioned_frame); //FIXME don't set in encode_header
226 #undef COPY
227 }
228
229 /**
230  * sets the given MpegEncContext to defaults for encoding.
231  * the changed fields will not depend upon the prior state of the MpegEncContext.
232  */
233 static void MPV_encode_defaults(MpegEncContext *s){
234     int i;
235     MPV_common_defaults(s);
236
237     for(i=-16; i<16; i++){
238         default_fcode_tab[i + MAX_MV]= 1;
239     }
240     s->me.mv_penalty= default_mv_penalty;
241     s->fcode_tab= default_fcode_tab;
242 }
243
244 /* init video encoder */
245 av_cold int MPV_encode_init(AVCodecContext *avctx)
246 {
247     MpegEncContext *s = avctx->priv_data;
248     int i;
249     int chroma_h_shift, chroma_v_shift;
250
251     MPV_encode_defaults(s);
252
253     switch (avctx->codec_id) {
254     case CODEC_ID_MPEG2VIDEO:
255         if(avctx->pix_fmt != PIX_FMT_YUV420P && avctx->pix_fmt != PIX_FMT_YUV422P){
256             av_log(avctx, AV_LOG_ERROR, "only YUV420 and YUV422 are supported\n");
257             return -1;
258         }
259         break;
260     case CODEC_ID_LJPEG:
261         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 &&
262            ((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)){
263             av_log(avctx, AV_LOG_ERROR, "colorspace not supported in LJPEG\n");
264             return -1;
265         }
266         break;
267     case CODEC_ID_MJPEG:
268         if(avctx->pix_fmt != PIX_FMT_YUVJ420P && avctx->pix_fmt != PIX_FMT_YUVJ422P &&
269            ((avctx->pix_fmt != PIX_FMT_YUV420P && avctx->pix_fmt != PIX_FMT_YUV422P) || avctx->strict_std_compliance>FF_COMPLIANCE_UNOFFICIAL)){
270             av_log(avctx, AV_LOG_ERROR, "colorspace not supported in jpeg\n");
271             return -1;
272         }
273         break;
274     default:
275         if(avctx->pix_fmt != PIX_FMT_YUV420P){
276             av_log(avctx, AV_LOG_ERROR, "only YUV420 is supported\n");
277             return -1;
278         }
279     }
280
281     switch (avctx->pix_fmt) {
282     case PIX_FMT_YUVJ422P:
283     case PIX_FMT_YUV422P:
284         s->chroma_format = CHROMA_422;
285         break;
286     case PIX_FMT_YUVJ420P:
287     case PIX_FMT_YUV420P:
288     default:
289         s->chroma_format = CHROMA_420;
290         break;
291     }
292
293     s->bit_rate = avctx->bit_rate;
294     s->width = avctx->width;
295     s->height = avctx->height;
296     if(avctx->gop_size > 600 && avctx->strict_std_compliance>FF_COMPLIANCE_EXPERIMENTAL){
297         av_log(avctx, AV_LOG_ERROR, "Warning keyframe interval too large! reducing it ...\n");
298         avctx->gop_size=600;
299     }
300     s->gop_size = avctx->gop_size;
301     s->avctx = avctx;
302     s->flags= avctx->flags;
303     s->flags2= avctx->flags2;
304     s->max_b_frames= avctx->max_b_frames;
305     s->codec_id= avctx->codec->id;
306     s->luma_elim_threshold  = avctx->luma_elim_threshold;
307     s->chroma_elim_threshold= avctx->chroma_elim_threshold;
308     s->strict_std_compliance= avctx->strict_std_compliance;
309 #if FF_API_MPEGVIDEO_GLOBAL_OPTS
310     if (avctx->flags & CODEC_FLAG_PART)
311         s->data_partitioning = 1;
312 #endif
313     s->quarter_sample= (avctx->flags & CODEC_FLAG_QPEL)!=0;
314     s->mpeg_quant= avctx->mpeg_quant;
315     s->rtp_mode= !!avctx->rtp_payload_size;
316     s->intra_dc_precision= avctx->intra_dc_precision;
317     s->user_specified_pts = AV_NOPTS_VALUE;
318
319     if (s->gop_size <= 1) {
320         s->intra_only = 1;
321         s->gop_size = 12;
322     } else {
323         s->intra_only = 0;
324     }
325
326     s->me_method = avctx->me_method;
327
328     /* Fixed QSCALE */
329     s->fixed_qscale = !!(avctx->flags & CODEC_FLAG_QSCALE);
330
331     s->adaptive_quant= (   s->avctx->lumi_masking
332                         || s->avctx->dark_masking
333                         || s->avctx->temporal_cplx_masking
334                         || s->avctx->spatial_cplx_masking
335                         || s->avctx->p_masking
336                         || s->avctx->border_masking
337                         || (s->flags&CODEC_FLAG_QP_RD))
338                        && !s->fixed_qscale;
339
340     s->loop_filter= !!(s->flags & CODEC_FLAG_LOOP_FILTER);
341 #if FF_API_MPEGVIDEO_GLOBAL_OPTS
342     s->alternate_scan= !!(s->flags & CODEC_FLAG_ALT_SCAN);
343     s->intra_vlc_format= !!(s->flags2 & CODEC_FLAG2_INTRA_VLC);
344     s->q_scale_type= !!(s->flags2 & CODEC_FLAG2_NON_LINEAR_QUANT);
345     s->obmc= !!(s->flags & CODEC_FLAG_OBMC);
346 #endif
347
348     if(avctx->rc_max_rate && !avctx->rc_buffer_size){
349         av_log(avctx, AV_LOG_ERROR, "a vbv buffer size is needed, for encoding with a maximum bitrate\n");
350         return -1;
351     }
352
353     if(avctx->rc_min_rate && avctx->rc_max_rate != avctx->rc_min_rate){
354         av_log(avctx, AV_LOG_INFO, "Warning min_rate > 0 but min_rate != max_rate isn't recommended!\n");
355     }
356
357     if(avctx->rc_min_rate && avctx->rc_min_rate > avctx->bit_rate){
358         av_log(avctx, AV_LOG_ERROR, "bitrate below min bitrate\n");
359         return -1;
360     }
361
362     if(avctx->rc_max_rate && avctx->rc_max_rate < avctx->bit_rate){
363         av_log(avctx, AV_LOG_INFO, "bitrate above max bitrate\n");
364         return -1;
365     }
366
367     if(avctx->rc_max_rate && avctx->rc_max_rate == avctx->bit_rate && avctx->rc_max_rate != avctx->rc_min_rate){
368         av_log(avctx, AV_LOG_INFO, "impossible bitrate constraints, this will fail\n");
369     }
370
371     if(avctx->rc_buffer_size && avctx->bit_rate*(int64_t)avctx->time_base.num > avctx->rc_buffer_size * (int64_t)avctx->time_base.den){
372         av_log(avctx, AV_LOG_ERROR, "VBV buffer too small for bitrate\n");
373         return -1;
374     }
375
376     if(!s->fixed_qscale && avctx->bit_rate*av_q2d(avctx->time_base) > avctx->bit_rate_tolerance){
377         av_log(avctx, AV_LOG_ERROR, "bitrate tolerance too small for bitrate\n");
378         return -1;
379     }
380
381     if(   s->avctx->rc_max_rate && s->avctx->rc_min_rate == s->avctx->rc_max_rate
382        && (s->codec_id == CODEC_ID_MPEG1VIDEO || s->codec_id == CODEC_ID_MPEG2VIDEO)
383        && 90000LL * (avctx->rc_buffer_size-1) > s->avctx->rc_max_rate*0xFFFFLL){
384
385         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");
386     }
387
388     if((s->flags & CODEC_FLAG_4MV) && s->codec_id != CODEC_ID_MPEG4
389        && s->codec_id != CODEC_ID_H263 && s->codec_id != CODEC_ID_H263P && s->codec_id != CODEC_ID_FLV1){
390         av_log(avctx, AV_LOG_ERROR, "4MV not supported by codec\n");
391         return -1;
392     }
393
394     if(s->obmc && s->avctx->mb_decision != FF_MB_DECISION_SIMPLE){
395         av_log(avctx, AV_LOG_ERROR, "OBMC is only supported with simple mb decision\n");
396         return -1;
397     }
398
399 #if FF_API_MPEGVIDEO_GLOBAL_OPTS
400     if(s->obmc && s->codec_id != CODEC_ID_H263 && s->codec_id != CODEC_ID_H263P){
401         av_log(avctx, AV_LOG_ERROR, "OBMC is only supported with H263(+)\n");
402         return -1;
403     }
404 #endif
405
406     if(s->quarter_sample && s->codec_id != CODEC_ID_MPEG4){
407         av_log(avctx, AV_LOG_ERROR, "qpel not supported by codec\n");
408         return -1;
409     }
410
411 #if FF_API_MPEGVIDEO_GLOBAL_OPTS
412     if(s->data_partitioning && s->codec_id != CODEC_ID_MPEG4){
413         av_log(avctx, AV_LOG_ERROR, "data partitioning not supported by codec\n");
414         return -1;
415     }
416 #endif
417
418     if(s->max_b_frames && s->codec_id != CODEC_ID_MPEG4 && s->codec_id != CODEC_ID_MPEG1VIDEO && s->codec_id != CODEC_ID_MPEG2VIDEO){
419         av_log(avctx, AV_LOG_ERROR, "b frames not supported by codec\n");
420         return -1;
421     }
422
423     if ((s->codec_id == CODEC_ID_MPEG4 || s->codec_id == CODEC_ID_H263 ||
424          s->codec_id == CODEC_ID_H263P) &&
425         (avctx->sample_aspect_ratio.num > 255 || avctx->sample_aspect_ratio.den > 255)) {
426         av_log(avctx, AV_LOG_ERROR, "Invalid pixel aspect ratio %i/%i, limit is 255/255\n",
427                avctx->sample_aspect_ratio.num, avctx->sample_aspect_ratio.den);
428         return -1;
429     }
430
431     if((s->flags & (CODEC_FLAG_INTERLACED_DCT|CODEC_FLAG_INTERLACED_ME|CODEC_FLAG_ALT_SCAN))
432        && s->codec_id != CODEC_ID_MPEG4 && s->codec_id != CODEC_ID_MPEG2VIDEO){
433         av_log(avctx, AV_LOG_ERROR, "interlacing not supported by codec\n");
434         return -1;
435     }
436
437     if(s->mpeg_quant && s->codec_id != CODEC_ID_MPEG4){ //FIXME mpeg2 uses that too
438         av_log(avctx, AV_LOG_ERROR, "mpeg2 style quantization not supported by codec\n");
439         return -1;
440     }
441
442     if((s->flags & CODEC_FLAG_CBP_RD) && !avctx->trellis){
443         av_log(avctx, AV_LOG_ERROR, "CBP RD needs trellis quant\n");
444         return -1;
445     }
446
447     if((s->flags & CODEC_FLAG_QP_RD) && s->avctx->mb_decision != FF_MB_DECISION_RD){
448         av_log(avctx, AV_LOG_ERROR, "QP RD needs mbd=2\n");
449         return -1;
450     }
451
452     if(s->avctx->scenechange_threshold < 1000000000 && (s->flags & CODEC_FLAG_CLOSED_GOP)){
453         av_log(avctx, AV_LOG_ERROR, "closed gop with scene change detection are not supported yet, set threshold to 1000000000\n");
454         return -1;
455     }
456
457     if((s->flags2 & CODEC_FLAG2_INTRA_VLC) && s->codec_id != CODEC_ID_MPEG2VIDEO){
458         av_log(avctx, AV_LOG_ERROR, "intra vlc table not supported by codec\n");
459         return -1;
460     }
461
462     if(s->flags & CODEC_FLAG_LOW_DELAY){
463         if (s->codec_id != CODEC_ID_MPEG2VIDEO){
464             av_log(avctx, AV_LOG_ERROR, "low delay forcing is only available for mpeg2\n");
465             return -1;
466         }
467         if (s->max_b_frames != 0){
468             av_log(avctx, AV_LOG_ERROR, "b frames cannot be used with low delay\n");
469             return -1;
470         }
471     }
472
473     if(s->q_scale_type == 1){
474 #if FF_API_MPEGVIDEO_GLOBAL_OPTS
475         if(s->codec_id != CODEC_ID_MPEG2VIDEO){
476             av_log(avctx, AV_LOG_ERROR, "non linear quant is only available for mpeg2\n");
477             return -1;
478         }
479 #endif
480         if(avctx->qmax > 12){
481             av_log(avctx, AV_LOG_ERROR, "non linear quant only supports qmax <= 12 currently\n");
482             return -1;
483         }
484     }
485
486     if(s->avctx->thread_count > 1 && s->codec_id != CODEC_ID_MPEG4
487        && s->codec_id != CODEC_ID_MPEG1VIDEO && s->codec_id != CODEC_ID_MPEG2VIDEO
488        && (s->codec_id != CODEC_ID_H263P || !(s->flags & CODEC_FLAG_H263P_SLICE_STRUCT))){
489         av_log(avctx, AV_LOG_ERROR, "multi threaded encoding not supported by codec\n");
490         return -1;
491     }
492
493     if(s->avctx->thread_count < 1){
494         av_log(avctx, AV_LOG_ERROR, "automatic thread number detection not supported by codec, patch welcome\n");
495         return -1;
496     }
497
498     if(s->avctx->thread_count > 1)
499         s->rtp_mode= 1;
500
501     if(!avctx->time_base.den || !avctx->time_base.num){
502         av_log(avctx, AV_LOG_ERROR, "framerate not set\n");
503         return -1;
504     }
505
506     i= (INT_MAX/2+128)>>8;
507     if(avctx->me_threshold >= i){
508         av_log(avctx, AV_LOG_ERROR, "me_threshold too large, max is %d\n", i - 1);
509         return -1;
510     }
511     if(avctx->mb_threshold >= i){
512         av_log(avctx, AV_LOG_ERROR, "mb_threshold too large, max is %d\n", i - 1);
513         return -1;
514     }
515
516     if(avctx->b_frame_strategy && (avctx->flags&CODEC_FLAG_PASS2)){
517         av_log(avctx, AV_LOG_INFO, "notice: b_frame_strategy only affects the first pass\n");
518         avctx->b_frame_strategy = 0;
519     }
520
521     i= av_gcd(avctx->time_base.den, avctx->time_base.num);
522     if(i > 1){
523         av_log(avctx, AV_LOG_INFO, "removing common factors from framerate\n");
524         avctx->time_base.den /= i;
525         avctx->time_base.num /= i;
526 //        return -1;
527     }
528
529     if(s->mpeg_quant || s->codec_id==CODEC_ID_MPEG1VIDEO || s->codec_id==CODEC_ID_MPEG2VIDEO || s->codec_id==CODEC_ID_MJPEG){
530         s->intra_quant_bias= 3<<(QUANT_BIAS_SHIFT-3); //(a + x*3/8)/x
531         s->inter_quant_bias= 0;
532     }else{
533         s->intra_quant_bias=0;
534         s->inter_quant_bias=-(1<<(QUANT_BIAS_SHIFT-2)); //(a - x/4)/x
535     }
536
537     if(avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)
538         s->intra_quant_bias= avctx->intra_quant_bias;
539     if(avctx->inter_quant_bias != FF_DEFAULT_QUANT_BIAS)
540         s->inter_quant_bias= avctx->inter_quant_bias;
541
542     avcodec_get_chroma_sub_sample(avctx->pix_fmt, &chroma_h_shift, &chroma_v_shift);
543
544     if(avctx->codec_id == CODEC_ID_MPEG4 && s->avctx->time_base.den > (1<<16)-1){
545         av_log(avctx, AV_LOG_ERROR, "timebase %d/%d not supported by MPEG 4 standard, "
546                "the maximum admitted value for the timebase denominator is %d\n",
547                s->avctx->time_base.num, s->avctx->time_base.den, (1<<16)-1);
548         return -1;
549     }
550     s->time_increment_bits = av_log2(s->avctx->time_base.den - 1) + 1;
551
552     switch(avctx->codec->id) {
553     case CODEC_ID_MPEG1VIDEO:
554         s->out_format = FMT_MPEG1;
555         s->low_delay= !!(s->flags & CODEC_FLAG_LOW_DELAY);
556         avctx->delay= s->low_delay ? 0 : (s->max_b_frames + 1);
557         break;
558     case CODEC_ID_MPEG2VIDEO:
559         s->out_format = FMT_MPEG1;
560         s->low_delay= !!(s->flags & CODEC_FLAG_LOW_DELAY);
561         avctx->delay= s->low_delay ? 0 : (s->max_b_frames + 1);
562         s->rtp_mode= 1;
563         break;
564     case CODEC_ID_LJPEG:
565     case CODEC_ID_MJPEG:
566         s->out_format = FMT_MJPEG;
567         s->intra_only = 1; /* force intra only for jpeg */
568         if(avctx->codec->id == CODEC_ID_LJPEG && avctx->pix_fmt == PIX_FMT_BGRA){
569             s->mjpeg_vsample[0] = s->mjpeg_hsample[0] =
570             s->mjpeg_vsample[1] = s->mjpeg_hsample[1] =
571             s->mjpeg_vsample[2] = s->mjpeg_hsample[2] = 1;
572         }else{
573             s->mjpeg_vsample[0] = 2;
574             s->mjpeg_vsample[1] = 2>>chroma_v_shift;
575             s->mjpeg_vsample[2] = 2>>chroma_v_shift;
576             s->mjpeg_hsample[0] = 2;
577             s->mjpeg_hsample[1] = 2>>chroma_h_shift;
578             s->mjpeg_hsample[2] = 2>>chroma_h_shift;
579         }
580         if (!(CONFIG_MJPEG_ENCODER || CONFIG_LJPEG_ENCODER)
581             || ff_mjpeg_encode_init(s) < 0)
582             return -1;
583         avctx->delay=0;
584         s->low_delay=1;
585         break;
586     case CODEC_ID_H261:
587         if (!CONFIG_H261_ENCODER)  return -1;
588         if (ff_h261_get_picture_format(s->width, s->height) < 0) {
589             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);
590             return -1;
591         }
592         s->out_format = FMT_H261;
593         avctx->delay=0;
594         s->low_delay=1;
595         break;
596     case CODEC_ID_H263:
597         if (!CONFIG_H263_ENCODER)  return -1;
598         if (ff_match_2uint16(h263_format, FF_ARRAY_ELEMS(h263_format), s->width, s->height) == 8) {
599             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);
600             return -1;
601         }
602         s->out_format = FMT_H263;
603         avctx->delay=0;
604         s->low_delay=1;
605         break;
606     case CODEC_ID_H263P:
607         s->out_format = FMT_H263;
608         s->h263_plus = 1;
609         /* Fx */
610 #if FF_API_MPEGVIDEO_GLOBAL_OPTS
611         if (avctx->flags & CODEC_FLAG_H263P_UMV)
612             s->umvplus = 1;
613         if (avctx->flags & CODEC_FLAG_H263P_AIV)
614             s->alt_inter_vlc = 1;
615 #endif
616         s->h263_aic= (avctx->flags & CODEC_FLAG_AC_PRED) ? 1:0;
617         s->modified_quant= s->h263_aic;
618         s->loop_filter= (avctx->flags & CODEC_FLAG_LOOP_FILTER) ? 1:0;
619         s->unrestricted_mv= s->obmc || s->loop_filter || s->umvplus;
620         s->h263_slice_structured= (s->flags & CODEC_FLAG_H263P_SLICE_STRUCT) ? 1:0;
621
622         /* /Fx */
623         /* These are just to be sure */
624         avctx->delay=0;
625         s->low_delay=1;
626         break;
627     case CODEC_ID_FLV1:
628         s->out_format = FMT_H263;
629         s->h263_flv = 2; /* format = 1; 11-bit codes */
630         s->unrestricted_mv = 1;
631         s->rtp_mode=0; /* don't allow GOB */
632         avctx->delay=0;
633         s->low_delay=1;
634         break;
635     case CODEC_ID_RV10:
636         s->out_format = FMT_H263;
637         avctx->delay=0;
638         s->low_delay=1;
639         break;
640     case CODEC_ID_RV20:
641         s->out_format = FMT_H263;
642         avctx->delay=0;
643         s->low_delay=1;
644         s->modified_quant=1;
645         s->h263_aic=1;
646         s->h263_plus=1;
647         s->loop_filter=1;
648         s->unrestricted_mv= 0;
649         break;
650     case CODEC_ID_MPEG4:
651         s->out_format = FMT_H263;
652         s->h263_pred = 1;
653         s->unrestricted_mv = 1;
654         s->low_delay= s->max_b_frames ? 0 : 1;
655         avctx->delay= s->low_delay ? 0 : (s->max_b_frames + 1);
656         break;
657     case CODEC_ID_MSMPEG4V2:
658         s->out_format = FMT_H263;
659         s->h263_pred = 1;
660         s->unrestricted_mv = 1;
661         s->msmpeg4_version= 2;
662         avctx->delay=0;
663         s->low_delay=1;
664         break;
665     case CODEC_ID_MSMPEG4V3:
666         s->out_format = FMT_H263;
667         s->h263_pred = 1;
668         s->unrestricted_mv = 1;
669         s->msmpeg4_version= 3;
670         s->flipflop_rounding=1;
671         avctx->delay=0;
672         s->low_delay=1;
673         break;
674     case CODEC_ID_WMV1:
675         s->out_format = FMT_H263;
676         s->h263_pred = 1;
677         s->unrestricted_mv = 1;
678         s->msmpeg4_version= 4;
679         s->flipflop_rounding=1;
680         avctx->delay=0;
681         s->low_delay=1;
682         break;
683     case CODEC_ID_WMV2:
684         s->out_format = FMT_H263;
685         s->h263_pred = 1;
686         s->unrestricted_mv = 1;
687         s->msmpeg4_version= 5;
688         s->flipflop_rounding=1;
689         avctx->delay=0;
690         s->low_delay=1;
691         break;
692     default:
693         return -1;
694     }
695
696     avctx->has_b_frames= !s->low_delay;
697
698     s->encoding = 1;
699
700     s->progressive_frame=
701     s->progressive_sequence= !(avctx->flags & (CODEC_FLAG_INTERLACED_DCT|CODEC_FLAG_INTERLACED_ME|CODEC_FLAG_ALT_SCAN));
702
703     /* init */
704     if (MPV_common_init(s) < 0)
705         return -1;
706
707     if(!s->dct_quantize)
708         s->dct_quantize = dct_quantize_c;
709     if(!s->denoise_dct)
710         s->denoise_dct = denoise_dct_c;
711     s->fast_dct_quantize = s->dct_quantize;
712     if(avctx->trellis)
713         s->dct_quantize = dct_quantize_trellis_c;
714
715     if((CONFIG_H263P_ENCODER || CONFIG_RV20_ENCODER) && s->modified_quant)
716         s->chroma_qscale_table= ff_h263_chroma_qscale_table;
717
718     s->quant_precision=5;
719
720     ff_set_cmp(&s->dsp, s->dsp.ildct_cmp, s->avctx->ildct_cmp);
721     ff_set_cmp(&s->dsp, s->dsp.frame_skip_cmp, s->avctx->frame_skip_cmp);
722
723     if (CONFIG_H261_ENCODER && s->out_format == FMT_H261)
724         ff_h261_encode_init(s);
725     if (CONFIG_H263_ENCODER && s->out_format == FMT_H263)
726         h263_encode_init(s);
727     if (CONFIG_MSMPEG4_ENCODER && s->msmpeg4_version)
728         ff_msmpeg4_encode_init(s);
729     if ((CONFIG_MPEG1VIDEO_ENCODER || CONFIG_MPEG2VIDEO_ENCODER)
730         && s->out_format == FMT_MPEG1)
731         ff_mpeg1_encode_init(s);
732
733     /* init q matrix */
734     for(i=0;i<64;i++) {
735         int j= s->dsp.idct_permutation[i];
736         if(CONFIG_MPEG4_ENCODER && s->codec_id==CODEC_ID_MPEG4 && s->mpeg_quant){
737             s->intra_matrix[j] = ff_mpeg4_default_intra_matrix[i];
738             s->inter_matrix[j] = ff_mpeg4_default_non_intra_matrix[i];
739         }else if(s->out_format == FMT_H263 || s->out_format == FMT_H261){
740             s->intra_matrix[j] =
741             s->inter_matrix[j] = ff_mpeg1_default_non_intra_matrix[i];
742         }else
743         { /* mpeg1/2 */
744             s->intra_matrix[j] = ff_mpeg1_default_intra_matrix[i];
745             s->inter_matrix[j] = ff_mpeg1_default_non_intra_matrix[i];
746         }
747         if(s->avctx->intra_matrix)
748             s->intra_matrix[j] = s->avctx->intra_matrix[i];
749         if(s->avctx->inter_matrix)
750             s->inter_matrix[j] = s->avctx->inter_matrix[i];
751     }
752
753     /* precompute matrix */
754     /* for mjpeg, we do include qscale in the matrix */
755     if (s->out_format != FMT_MJPEG) {
756         ff_convert_matrix(&s->dsp, s->q_intra_matrix, s->q_intra_matrix16,
757                        s->intra_matrix, s->intra_quant_bias, avctx->qmin, 31, 1);
758         ff_convert_matrix(&s->dsp, s->q_inter_matrix, s->q_inter_matrix16,
759                        s->inter_matrix, s->inter_quant_bias, avctx->qmin, 31, 0);
760     }
761
762     if(ff_rate_control_init(s) < 0)
763         return -1;
764
765     return 0;
766 }
767
768 av_cold int MPV_encode_end(AVCodecContext *avctx)
769 {
770     MpegEncContext *s = avctx->priv_data;
771
772     ff_rate_control_uninit(s);
773
774     MPV_common_end(s);
775     if ((CONFIG_MJPEG_ENCODER || CONFIG_LJPEG_ENCODER) && s->out_format == FMT_MJPEG)
776         ff_mjpeg_encode_close(s);
777
778     av_freep(&avctx->extradata);
779
780     return 0;
781 }
782
783 static int get_sae(uint8_t *src, int ref, int stride){
784     int x,y;
785     int acc=0;
786
787     for(y=0; y<16; y++){
788         for(x=0; x<16; x++){
789             acc+= FFABS(src[x+y*stride] - ref);
790         }
791     }
792
793     return acc;
794 }
795
796 static int get_intra_count(MpegEncContext *s, uint8_t *src, uint8_t *ref, int stride){
797     int x, y, w, h;
798     int acc=0;
799
800     w= s->width &~15;
801     h= s->height&~15;
802
803     for(y=0; y<h; y+=16){
804         for(x=0; x<w; x+=16){
805             int offset= x + y*stride;
806             int sad = s->dsp.sad[0](NULL, src + offset, ref + offset, stride, 16);
807             int mean= (s->dsp.pix_sum(src + offset, stride) + 128)>>8;
808             int sae = get_sae(src + offset, mean, stride);
809
810             acc+= sae + 500 < sad;
811         }
812     }
813     return acc;
814 }
815
816
817 static int load_input_picture(MpegEncContext *s, AVFrame *pic_arg){
818     AVFrame *pic=NULL;
819     int64_t pts;
820     int i;
821     const int encoding_delay= s->max_b_frames;
822     int direct=1;
823
824     if(pic_arg){
825         pts= pic_arg->pts;
826         pic_arg->display_picture_number= s->input_picture_number++;
827
828         if(pts != AV_NOPTS_VALUE){
829             if(s->user_specified_pts != AV_NOPTS_VALUE){
830                 int64_t time= pts;
831                 int64_t last= s->user_specified_pts;
832
833                 if(time <= last){
834                     av_log(s->avctx, AV_LOG_ERROR, "Error, Invalid timestamp=%"PRId64", last=%"PRId64"\n", pts, s->user_specified_pts);
835                     return -1;
836                 }
837             }
838             s->user_specified_pts= pts;
839         }else{
840             if(s->user_specified_pts != AV_NOPTS_VALUE){
841                 s->user_specified_pts=
842                 pts= s->user_specified_pts + 1;
843                 av_log(s->avctx, AV_LOG_INFO, "Warning: AVFrame.pts=? trying to guess (%"PRId64")\n", pts);
844             }else{
845                 pts= pic_arg->display_picture_number;
846             }
847         }
848     }
849
850   if(pic_arg){
851     if(encoding_delay && !(s->flags&CODEC_FLAG_INPUT_PRESERVED)) direct=0;
852     if(pic_arg->linesize[0] != s->linesize) direct=0;
853     if(pic_arg->linesize[1] != s->uvlinesize) direct=0;
854     if(pic_arg->linesize[2] != s->uvlinesize) direct=0;
855
856 //    av_log(AV_LOG_DEBUG, "%d %d %d %d\n",pic_arg->linesize[0], pic_arg->linesize[1], s->linesize, s->uvlinesize);
857
858     if(direct){
859         i= ff_find_unused_picture(s, 1);
860
861         pic= (AVFrame*)&s->picture[i];
862         pic->reference= 3;
863
864         for(i=0; i<4; i++){
865             pic->data[i]= pic_arg->data[i];
866             pic->linesize[i]= pic_arg->linesize[i];
867         }
868         if(ff_alloc_picture(s, (Picture*)pic, 1) < 0){
869             return -1;
870         }
871     }else{
872         i= ff_find_unused_picture(s, 0);
873
874         pic= (AVFrame*)&s->picture[i];
875         pic->reference= 3;
876
877         if(ff_alloc_picture(s, (Picture*)pic, 0) < 0){
878             return -1;
879         }
880
881         if(   pic->data[0] + INPLACE_OFFSET == pic_arg->data[0]
882            && pic->data[1] + INPLACE_OFFSET == pic_arg->data[1]
883            && pic->data[2] + INPLACE_OFFSET == pic_arg->data[2]){
884        // empty
885         }else{
886             int h_chroma_shift, v_chroma_shift;
887             avcodec_get_chroma_sub_sample(s->avctx->pix_fmt, &h_chroma_shift, &v_chroma_shift);
888
889             for(i=0; i<3; i++){
890                 int src_stride= pic_arg->linesize[i];
891                 int dst_stride= i ? s->uvlinesize : s->linesize;
892                 int h_shift= i ? h_chroma_shift : 0;
893                 int v_shift= i ? v_chroma_shift : 0;
894                 int w= s->width >>h_shift;
895                 int h= s->height>>v_shift;
896                 uint8_t *src= pic_arg->data[i];
897                 uint8_t *dst= pic->data[i];
898
899                 if(!s->avctx->rc_buffer_size)
900                     dst +=INPLACE_OFFSET;
901
902                 if(src_stride==dst_stride)
903                     memcpy(dst, src, src_stride*h);
904                 else{
905                     while(h--){
906                         memcpy(dst, src, w);
907                         dst += dst_stride;
908                         src += src_stride;
909                     }
910                 }
911             }
912         }
913     }
914     copy_picture_attributes(s, pic, pic_arg);
915     pic->pts= pts; //we set this here to avoid modifiying pic_arg
916   }
917
918     /* shift buffer entries */
919     for(i=1; i<MAX_PICTURE_COUNT /*s->encoding_delay+1*/; i++)
920         s->input_picture[i-1]= s->input_picture[i];
921
922     s->input_picture[encoding_delay]= (Picture*)pic;
923
924     return 0;
925 }
926
927 static int skip_check(MpegEncContext *s, Picture *p, Picture *ref){
928     int x, y, plane;
929     int score=0;
930     int64_t score64=0;
931
932     for(plane=0; plane<3; plane++){
933         const int stride = p->f.linesize[plane];
934         const int bw= plane ? 1 : 2;
935         for(y=0; y<s->mb_height*bw; y++){
936             for(x=0; x<s->mb_width*bw; x++){
937                 int off = p->f.type == FF_BUFFER_TYPE_SHARED ? 0: 16;
938                 int v   = s->dsp.frame_skip_cmp[1](s, p->f.data[plane] + 8*(x + y*stride)+off, ref->f.data[plane] + 8*(x + y*stride), stride, 8);
939
940                 switch(s->avctx->frame_skip_exp){
941                     case 0: score= FFMAX(score, v); break;
942                     case 1: score+= FFABS(v);break;
943                     case 2: score+= v*v;break;
944                     case 3: score64+= FFABS(v*v*(int64_t)v);break;
945                     case 4: score64+= v*v*(int64_t)(v*v);break;
946                 }
947             }
948         }
949     }
950
951     if(score) score64= score;
952
953     if(score64 < s->avctx->frame_skip_threshold)
954         return 1;
955     if(score64 < ((s->avctx->frame_skip_factor * (int64_t)s->lambda)>>8))
956         return 1;
957     return 0;
958 }
959
960 static int estimate_best_b_count(MpegEncContext *s){
961     AVCodec *codec= avcodec_find_encoder(s->avctx->codec_id);
962     AVCodecContext *c = avcodec_alloc_context3(NULL);
963     AVFrame input[FF_MAX_B_FRAMES+2];
964     const int scale= s->avctx->brd_scale;
965     int i, j, out_size, p_lambda, b_lambda, lambda2;
966     int outbuf_size= s->width * s->height; //FIXME
967     uint8_t *outbuf= av_malloc(outbuf_size);
968     int64_t best_rd= INT64_MAX;
969     int best_b_count= -1;
970
971     assert(scale>=0 && scale <=3);
972
973 //    emms_c();
974     p_lambda= s->last_lambda_for[AV_PICTURE_TYPE_P]; //s->next_picture_ptr->quality;
975     b_lambda= s->last_lambda_for[AV_PICTURE_TYPE_B]; //p_lambda *FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset;
976     if(!b_lambda) b_lambda= p_lambda; //FIXME we should do this somewhere else
977     lambda2= (b_lambda*b_lambda + (1<<FF_LAMBDA_SHIFT)/2 ) >> FF_LAMBDA_SHIFT;
978
979     c->width = s->width >> scale;
980     c->height= s->height>> scale;
981     c->flags= CODEC_FLAG_QSCALE | CODEC_FLAG_PSNR | CODEC_FLAG_INPUT_PRESERVED /*| CODEC_FLAG_EMU_EDGE*/;
982     c->flags|= s->avctx->flags & CODEC_FLAG_QPEL;
983     c->mb_decision= s->avctx->mb_decision;
984     c->me_cmp= s->avctx->me_cmp;
985     c->mb_cmp= s->avctx->mb_cmp;
986     c->me_sub_cmp= s->avctx->me_sub_cmp;
987     c->pix_fmt = PIX_FMT_YUV420P;
988     c->time_base= s->avctx->time_base;
989     c->max_b_frames= s->max_b_frames;
990
991     if (avcodec_open2(c, codec, NULL) < 0)
992         return -1;
993
994     for(i=0; i<s->max_b_frames+2; i++){
995         int ysize= c->width*c->height;
996         int csize= (c->width/2)*(c->height/2);
997         Picture pre_input, *pre_input_ptr= i ? s->input_picture[i-1] : s->next_picture_ptr;
998
999         avcodec_get_frame_defaults(&input[i]);
1000         input[i].data[0]= av_malloc(ysize + 2*csize);
1001         input[i].data[1]= input[i].data[0] + ysize;
1002         input[i].data[2]= input[i].data[1] + csize;
1003         input[i].linesize[0]= c->width;
1004         input[i].linesize[1]=
1005         input[i].linesize[2]= c->width/2;
1006
1007         if(pre_input_ptr && (!i || s->input_picture[i-1])) {
1008             pre_input= *pre_input_ptr;
1009
1010             if (pre_input.f.type != FF_BUFFER_TYPE_SHARED && i) {
1011                 pre_input.f.data[0] += INPLACE_OFFSET;
1012                 pre_input.f.data[1] += INPLACE_OFFSET;
1013                 pre_input.f.data[2] += INPLACE_OFFSET;
1014             }
1015
1016             s->dsp.shrink[scale](input[i].data[0], input[i].linesize[0], pre_input.f.data[0], pre_input.f.linesize[0], c->width,      c->height);
1017             s->dsp.shrink[scale](input[i].data[1], input[i].linesize[1], pre_input.f.data[1], pre_input.f.linesize[1], c->width >> 1, c->height >> 1);
1018             s->dsp.shrink[scale](input[i].data[2], input[i].linesize[2], pre_input.f.data[2], pre_input.f.linesize[2], c->width >> 1, c->height >> 1);
1019         }
1020     }
1021
1022     for(j=0; j<s->max_b_frames+1; j++){
1023         int64_t rd=0;
1024
1025         if(!s->input_picture[j])
1026             break;
1027
1028         c->error[0]= c->error[1]= c->error[2]= 0;
1029
1030         input[0].pict_type= AV_PICTURE_TYPE_I;
1031         input[0].quality= 1 * FF_QP2LAMBDA;
1032         out_size = avcodec_encode_video(c, outbuf, outbuf_size, &input[0]);
1033 //        rd += (out_size * lambda2) >> FF_LAMBDA_SHIFT;
1034
1035         for(i=0; i<s->max_b_frames+1; i++){
1036             int is_p= i % (j+1) == j || i==s->max_b_frames;
1037
1038             input[i+1].pict_type= is_p ? AV_PICTURE_TYPE_P : AV_PICTURE_TYPE_B;
1039             input[i+1].quality= is_p ? p_lambda : b_lambda;
1040             out_size = avcodec_encode_video(c, outbuf, outbuf_size, &input[i+1]);
1041             rd += (out_size * lambda2) >> (FF_LAMBDA_SHIFT - 3);
1042         }
1043
1044         /* get the delayed frames */
1045         while(out_size){
1046             out_size = avcodec_encode_video(c, outbuf, outbuf_size, NULL);
1047             rd += (out_size * lambda2) >> (FF_LAMBDA_SHIFT - 3);
1048         }
1049
1050         rd += c->error[0] + c->error[1] + c->error[2];
1051
1052         if(rd < best_rd){
1053             best_rd= rd;
1054             best_b_count= j;
1055         }
1056     }
1057
1058     av_freep(&outbuf);
1059     avcodec_close(c);
1060     av_freep(&c);
1061
1062     for(i=0; i<s->max_b_frames+2; i++){
1063         av_freep(&input[i].data[0]);
1064     }
1065
1066     return best_b_count;
1067 }
1068
1069 static int select_input_picture(MpegEncContext *s){
1070     int i;
1071
1072     for(i=1; i<MAX_PICTURE_COUNT; i++)
1073         s->reordered_input_picture[i-1]= s->reordered_input_picture[i];
1074     s->reordered_input_picture[MAX_PICTURE_COUNT-1]= NULL;
1075
1076     /* set next picture type & ordering */
1077     if(s->reordered_input_picture[0]==NULL && s->input_picture[0]){
1078         if(/*s->picture_in_gop_number >= s->gop_size ||*/ s->next_picture_ptr==NULL || s->intra_only){
1079             s->reordered_input_picture[0]= s->input_picture[0];
1080             s->reordered_input_picture[0]->f.pict_type = AV_PICTURE_TYPE_I;
1081             s->reordered_input_picture[0]->f.coded_picture_number = s->coded_picture_number++;
1082         }else{
1083             int b_frames;
1084
1085             if(s->avctx->frame_skip_threshold || s->avctx->frame_skip_factor){
1086                 if(s->picture_in_gop_number < s->gop_size && skip_check(s, s->input_picture[0], s->next_picture_ptr)){
1087                 //FIXME check that te gop check above is +-1 correct
1088 //av_log(NULL, AV_LOG_DEBUG, "skip %p %"PRId64"\n", s->input_picture[0]->f.data[0], s->input_picture[0]->pts);
1089
1090                     if (s->input_picture[0]->f.type == FF_BUFFER_TYPE_SHARED) {
1091                         for(i=0; i<4; i++)
1092                             s->input_picture[0]->f.data[i] = NULL;
1093                         s->input_picture[0]->f.type = 0;
1094                     }else{
1095                         assert(   s->input_picture[0]->type==FF_BUFFER_TYPE_USER
1096                                || s->input_picture[0]->type==FF_BUFFER_TYPE_INTERNAL);
1097
1098                         s->avctx->release_buffer(s->avctx, (AVFrame*)s->input_picture[0]);
1099                     }
1100
1101                     emms_c();
1102                     ff_vbv_update(s, 0);
1103
1104                     goto no_output_pic;
1105                 }
1106             }
1107
1108             if(s->flags&CODEC_FLAG_PASS2){
1109                 for(i=0; i<s->max_b_frames+1; i++){
1110                     int pict_num = s->input_picture[0]->f.display_picture_number + i;
1111
1112                     if(pict_num >= s->rc_context.num_entries)
1113                         break;
1114                     if(!s->input_picture[i]){
1115                         s->rc_context.entry[pict_num-1].new_pict_type = AV_PICTURE_TYPE_P;
1116                         break;
1117                     }
1118
1119                     s->input_picture[i]->f.pict_type =
1120                         s->rc_context.entry[pict_num].new_pict_type;
1121                 }
1122             }
1123
1124             if(s->avctx->b_frame_strategy==0){
1125                 b_frames= s->max_b_frames;
1126                 while(b_frames && !s->input_picture[b_frames]) b_frames--;
1127             }else if(s->avctx->b_frame_strategy==1){
1128                 for(i=1; i<s->max_b_frames+1; i++){
1129                     if(s->input_picture[i] && s->input_picture[i]->b_frame_score==0){
1130                         s->input_picture[i]->b_frame_score=
1131                             get_intra_count(s, s->input_picture[i  ]->f.data[0],
1132                                                s->input_picture[i-1]->f.data[0], s->linesize) + 1;
1133                     }
1134                 }
1135                 for(i=0; i<s->max_b_frames+1; i++){
1136                     if(s->input_picture[i]==NULL || s->input_picture[i]->b_frame_score - 1 > s->mb_num/s->avctx->b_sensitivity) break;
1137                 }
1138
1139                 b_frames= FFMAX(0, i-1);
1140
1141                 /* reset scores */
1142                 for(i=0; i<b_frames+1; i++){
1143                     s->input_picture[i]->b_frame_score=0;
1144                 }
1145             }else if(s->avctx->b_frame_strategy==2){
1146                 b_frames= estimate_best_b_count(s);
1147             }else{
1148                 av_log(s->avctx, AV_LOG_ERROR, "illegal b frame strategy\n");
1149                 b_frames=0;
1150             }
1151
1152             emms_c();
1153 //static int b_count=0;
1154 //b_count+= b_frames;
1155 //av_log(s->avctx, AV_LOG_DEBUG, "b_frames: %d\n", b_count);
1156
1157             for(i= b_frames - 1; i>=0; i--){
1158                 int type = s->input_picture[i]->f.pict_type;
1159                 if(type && type != AV_PICTURE_TYPE_B)
1160                     b_frames= i;
1161             }
1162             if (s->input_picture[b_frames]->f.pict_type == AV_PICTURE_TYPE_B && b_frames == s->max_b_frames){
1163                 av_log(s->avctx, AV_LOG_ERROR, "warning, too many b frames in a row\n");
1164             }
1165
1166             if(s->picture_in_gop_number + b_frames >= s->gop_size){
1167               if((s->flags2 & CODEC_FLAG2_STRICT_GOP) && s->gop_size > s->picture_in_gop_number){
1168                     b_frames= s->gop_size - s->picture_in_gop_number - 1;
1169               }else{
1170                 if(s->flags & CODEC_FLAG_CLOSED_GOP)
1171                     b_frames=0;
1172                 s->input_picture[b_frames]->f.pict_type = AV_PICTURE_TYPE_I;
1173               }
1174             }
1175
1176             if(   (s->flags & CODEC_FLAG_CLOSED_GOP)
1177                && b_frames
1178                && s->input_picture[b_frames]->f.pict_type== AV_PICTURE_TYPE_I)
1179                 b_frames--;
1180
1181             s->reordered_input_picture[0]= s->input_picture[b_frames];
1182             if (s->reordered_input_picture[0]->f.pict_type != AV_PICTURE_TYPE_I)
1183                 s->reordered_input_picture[0]->f.pict_type = AV_PICTURE_TYPE_P;
1184             s->reordered_input_picture[0]->f.coded_picture_number = s->coded_picture_number++;
1185             for(i=0; i<b_frames; i++){
1186                 s->reordered_input_picture[i + 1] = s->input_picture[i];
1187                 s->reordered_input_picture[i + 1]->f.pict_type = AV_PICTURE_TYPE_B;
1188                 s->reordered_input_picture[i + 1]->f.coded_picture_number = s->coded_picture_number++;
1189             }
1190         }
1191     }
1192 no_output_pic:
1193     if(s->reordered_input_picture[0]){
1194         s->reordered_input_picture[0]->f.reference = s->reordered_input_picture[0]->f.pict_type!=AV_PICTURE_TYPE_B ? 3 : 0;
1195
1196         ff_copy_picture(&s->new_picture, s->reordered_input_picture[0]);
1197
1198         if (s->reordered_input_picture[0]->f.type == FF_BUFFER_TYPE_SHARED || s->avctx->rc_buffer_size) {
1199             // input is a shared pix, so we can't modifiy it -> alloc a new one & ensure that the shared one is reuseable
1200
1201             int i= ff_find_unused_picture(s, 0);
1202             Picture *pic= &s->picture[i];
1203
1204             pic->f.reference = s->reordered_input_picture[0]->f.reference;
1205             if(ff_alloc_picture(s, pic, 0) < 0){
1206                 return -1;
1207             }
1208
1209             /* mark us unused / free shared pic */
1210             if (s->reordered_input_picture[0]->f.type == FF_BUFFER_TYPE_INTERNAL)
1211                 s->avctx->release_buffer(s->avctx, (AVFrame*)s->reordered_input_picture[0]);
1212             for(i=0; i<4; i++)
1213                 s->reordered_input_picture[0]->f.data[i] = NULL;
1214             s->reordered_input_picture[0]->f.type = 0;
1215
1216             copy_picture_attributes(s, (AVFrame*)pic, (AVFrame*)s->reordered_input_picture[0]);
1217
1218             s->current_picture_ptr= pic;
1219         }else{
1220             // input is not a shared pix -> reuse buffer for current_pix
1221
1222             assert(   s->reordered_input_picture[0]->type==FF_BUFFER_TYPE_USER
1223                    || s->reordered_input_picture[0]->type==FF_BUFFER_TYPE_INTERNAL);
1224
1225             s->current_picture_ptr= s->reordered_input_picture[0];
1226             for(i=0; i<4; i++){
1227                 s->new_picture.f.data[i] += INPLACE_OFFSET;
1228             }
1229         }
1230         ff_copy_picture(&s->current_picture, s->current_picture_ptr);
1231
1232         s->picture_number = s->new_picture.f.display_picture_number;
1233 //printf("dpn:%d\n", s->picture_number);
1234     }else{
1235        memset(&s->new_picture, 0, sizeof(Picture));
1236     }
1237     return 0;
1238 }
1239
1240 int MPV_encode_picture(AVCodecContext *avctx,
1241                        unsigned char *buf, int buf_size, void *data)
1242 {
1243     MpegEncContext *s = avctx->priv_data;
1244     AVFrame *pic_arg = data;
1245     int i, stuffing_count, context_count = avctx->thread_count;
1246
1247     for(i=0; i<context_count; i++){
1248         int start_y= s->thread_context[i]->start_mb_y;
1249         int   end_y= s->thread_context[i]->  end_mb_y;
1250         int h= s->mb_height;
1251         uint8_t *start= buf + (size_t)(((int64_t) buf_size)*start_y/h);
1252         uint8_t *end  = buf + (size_t)(((int64_t) buf_size)*  end_y/h);
1253
1254         init_put_bits(&s->thread_context[i]->pb, start, end - start);
1255     }
1256
1257     s->picture_in_gop_number++;
1258
1259     if(load_input_picture(s, pic_arg) < 0)
1260         return -1;
1261
1262     if(select_input_picture(s) < 0){
1263         return -1;
1264     }
1265
1266     /* output? */
1267     if (s->new_picture.f.data[0]) {
1268         s->pict_type = s->new_picture.f.pict_type;
1269 //emms_c();
1270 //printf("qs:%f %f %d\n", s->new_picture.quality, s->current_picture.quality, s->qscale);
1271         MPV_frame_start(s, avctx);
1272 vbv_retry:
1273         if (encode_picture(s, s->picture_number) < 0)
1274             return -1;
1275
1276         avctx->header_bits = s->header_bits;
1277         avctx->mv_bits     = s->mv_bits;
1278         avctx->misc_bits   = s->misc_bits;
1279         avctx->i_tex_bits  = s->i_tex_bits;
1280         avctx->p_tex_bits  = s->p_tex_bits;
1281         avctx->i_count     = s->i_count;
1282         avctx->p_count     = s->mb_num - s->i_count - s->skip_count; //FIXME f/b_count in avctx
1283         avctx->skip_count  = s->skip_count;
1284
1285         MPV_frame_end(s);
1286
1287         if (CONFIG_MJPEG_ENCODER && s->out_format == FMT_MJPEG)
1288             ff_mjpeg_encode_picture_trailer(s);
1289
1290         if(avctx->rc_buffer_size){
1291             RateControlContext *rcc= &s->rc_context;
1292             int max_size= rcc->buffer_index * avctx->rc_max_available_vbv_use;
1293
1294             if(put_bits_count(&s->pb) > max_size && s->lambda < s->avctx->lmax){
1295                 s->next_lambda= FFMAX(s->lambda+1, s->lambda*(s->qscale+1) / s->qscale);
1296                 if(s->adaptive_quant){
1297                     int i;
1298                     for(i=0; i<s->mb_height*s->mb_stride; i++)
1299                         s->lambda_table[i]= FFMAX(s->lambda_table[i]+1, s->lambda_table[i]*(s->qscale+1) / s->qscale);
1300                 }
1301                 s->mb_skipped = 0;        //done in MPV_frame_start()
1302                 if(s->pict_type==AV_PICTURE_TYPE_P){ //done in encode_picture() so we must undo it
1303                     if(s->flipflop_rounding || s->codec_id == CODEC_ID_H263P || s->codec_id == CODEC_ID_MPEG4)
1304                         s->no_rounding ^= 1;
1305                 }
1306                 if(s->pict_type!=AV_PICTURE_TYPE_B){
1307                     s->time_base= s->last_time_base;
1308                     s->last_non_b_time= s->time - s->pp_time;
1309                 }
1310 //                av_log(NULL, AV_LOG_ERROR, "R:%d ", s->next_lambda);
1311                 for(i=0; i<context_count; i++){
1312                     PutBitContext *pb= &s->thread_context[i]->pb;
1313                     init_put_bits(pb, pb->buf, pb->buf_end - pb->buf);
1314                 }
1315                 goto vbv_retry;
1316             }
1317
1318             assert(s->avctx->rc_max_rate);
1319         }
1320
1321         if(s->flags&CODEC_FLAG_PASS1)
1322             ff_write_pass1_stats(s);
1323
1324         for(i=0; i<4; i++){
1325             s->current_picture_ptr->f.error[i]  = s->current_picture.f.error[i];
1326             avctx->error[i]                        += s->current_picture_ptr->f.error[i];
1327         }
1328
1329         if(s->flags&CODEC_FLAG_PASS1)
1330             assert(avctx->header_bits + avctx->mv_bits + avctx->misc_bits + avctx->i_tex_bits + avctx->p_tex_bits == put_bits_count(&s->pb));
1331         flush_put_bits(&s->pb);
1332         s->frame_bits  = put_bits_count(&s->pb);
1333
1334         stuffing_count= ff_vbv_update(s, s->frame_bits);
1335         if(stuffing_count){
1336             if(s->pb.buf_end - s->pb.buf - (put_bits_count(&s->pb)>>3) < stuffing_count + 50){
1337                 av_log(s->avctx, AV_LOG_ERROR, "stuffing too large\n");
1338                 return -1;
1339             }
1340
1341             switch(s->codec_id){
1342             case CODEC_ID_MPEG1VIDEO:
1343             case CODEC_ID_MPEG2VIDEO:
1344                 while(stuffing_count--){
1345                     put_bits(&s->pb, 8, 0);
1346                 }
1347             break;
1348             case CODEC_ID_MPEG4:
1349                 put_bits(&s->pb, 16, 0);
1350                 put_bits(&s->pb, 16, 0x1C3);
1351                 stuffing_count -= 4;
1352                 while(stuffing_count--){
1353                     put_bits(&s->pb, 8, 0xFF);
1354                 }
1355             break;
1356             default:
1357                 av_log(s->avctx, AV_LOG_ERROR, "vbv buffer overflow\n");
1358             }
1359             flush_put_bits(&s->pb);
1360             s->frame_bits  = put_bits_count(&s->pb);
1361         }
1362
1363         /* update mpeg1/2 vbv_delay for CBR */
1364         if(s->avctx->rc_max_rate && s->avctx->rc_min_rate == s->avctx->rc_max_rate && s->out_format == FMT_MPEG1
1365            && 90000LL * (avctx->rc_buffer_size-1) <= s->avctx->rc_max_rate*0xFFFFLL){
1366             int vbv_delay, min_delay;
1367             double inbits = s->avctx->rc_max_rate*av_q2d(s->avctx->time_base);
1368             int    minbits= s->frame_bits - 8*(s->vbv_delay_ptr - s->pb.buf - 1);
1369             double bits   = s->rc_context.buffer_index + minbits - inbits;
1370
1371             if(bits<0)
1372                 av_log(s->avctx, AV_LOG_ERROR, "Internal error, negative bits\n");
1373
1374             assert(s->repeat_first_field==0);
1375
1376             vbv_delay=     bits * 90000                               / s->avctx->rc_max_rate;
1377             min_delay= (minbits * 90000LL + s->avctx->rc_max_rate - 1)/ s->avctx->rc_max_rate;
1378
1379             vbv_delay= FFMAX(vbv_delay, min_delay);
1380
1381             assert(vbv_delay < 0xFFFF);
1382
1383             s->vbv_delay_ptr[0] &= 0xF8;
1384             s->vbv_delay_ptr[0] |= vbv_delay>>13;
1385             s->vbv_delay_ptr[1]  = vbv_delay>>5;
1386             s->vbv_delay_ptr[2] &= 0x07;
1387             s->vbv_delay_ptr[2] |= vbv_delay<<3;
1388             avctx->vbv_delay = vbv_delay*300;
1389         }
1390         s->total_bits += s->frame_bits;
1391         avctx->frame_bits  = s->frame_bits;
1392     }else{
1393         assert((put_bits_ptr(&s->pb) == s->pb.buf));
1394         s->frame_bits=0;
1395     }
1396     assert((s->frame_bits&7)==0);
1397
1398     return s->frame_bits/8;
1399 }
1400
1401 static inline void dct_single_coeff_elimination(MpegEncContext *s, int n, int threshold)
1402 {
1403     static const char tab[64]=
1404         {3,2,2,1,1,1,1,1,
1405          1,1,1,1,1,1,1,1,
1406          1,1,1,1,1,1,1,1,
1407          0,0,0,0,0,0,0,0,
1408          0,0,0,0,0,0,0,0,
1409          0,0,0,0,0,0,0,0,
1410          0,0,0,0,0,0,0,0,
1411          0,0,0,0,0,0,0,0};
1412     int score=0;
1413     int run=0;
1414     int i;
1415     DCTELEM *block= s->block[n];
1416     const int last_index= s->block_last_index[n];
1417     int skip_dc;
1418
1419     if(threshold<0){
1420         skip_dc=0;
1421         threshold= -threshold;
1422     }else
1423         skip_dc=1;
1424
1425     /* Are all we could set to zero already zero? */
1426     if(last_index<=skip_dc - 1) return;
1427
1428     for(i=0; i<=last_index; i++){
1429         const int j = s->intra_scantable.permutated[i];
1430         const int level = FFABS(block[j]);
1431         if(level==1){
1432             if(skip_dc && i==0) continue;
1433             score+= tab[run];
1434             run=0;
1435         }else if(level>1){
1436             return;
1437         }else{
1438             run++;
1439         }
1440     }
1441     if(score >= threshold) return;
1442     for(i=skip_dc; i<=last_index; i++){
1443         const int j = s->intra_scantable.permutated[i];
1444         block[j]=0;
1445     }
1446     if(block[0]) s->block_last_index[n]= 0;
1447     else         s->block_last_index[n]= -1;
1448 }
1449
1450 static inline void clip_coeffs(MpegEncContext *s, DCTELEM *block, int last_index)
1451 {
1452     int i;
1453     const int maxlevel= s->max_qcoeff;
1454     const int minlevel= s->min_qcoeff;
1455     int overflow=0;
1456
1457     if(s->mb_intra){
1458         i=1; //skip clipping of intra dc
1459     }else
1460         i=0;
1461
1462     for(;i<=last_index; i++){
1463         const int j= s->intra_scantable.permutated[i];
1464         int level = block[j];
1465
1466         if     (level>maxlevel){
1467             level=maxlevel;
1468             overflow++;
1469         }else if(level<minlevel){
1470             level=minlevel;
1471             overflow++;
1472         }
1473
1474         block[j]= level;
1475     }
1476
1477     if(overflow && s->avctx->mb_decision == FF_MB_DECISION_SIMPLE)
1478         av_log(s->avctx, AV_LOG_INFO, "warning, clipping %d dct coefficients to %d..%d\n", overflow, minlevel, maxlevel);
1479 }
1480
1481 static void get_visual_weight(int16_t *weight, uint8_t *ptr, int stride){
1482     int x, y;
1483 //FIXME optimize
1484     for(y=0; y<8; y++){
1485         for(x=0; x<8; x++){
1486             int x2, y2;
1487             int sum=0;
1488             int sqr=0;
1489             int count=0;
1490
1491             for(y2= FFMAX(y-1, 0); y2 < FFMIN(8, y+2); y2++){
1492                 for(x2= FFMAX(x-1, 0); x2 < FFMIN(8, x+2); x2++){
1493                     int v= ptr[x2 + y2*stride];
1494                     sum += v;
1495                     sqr += v*v;
1496                     count++;
1497                 }
1498             }
1499             weight[x + 8*y]= (36*ff_sqrt(count*sqr - sum*sum)) / count;
1500         }
1501     }
1502 }
1503
1504 static av_always_inline void encode_mb_internal(MpegEncContext *s, int motion_x, int motion_y, int mb_block_height, int mb_block_count)
1505 {
1506     int16_t weight[8][64];
1507     DCTELEM orig[8][64];
1508     const int mb_x= s->mb_x;
1509     const int mb_y= s->mb_y;
1510     int i;
1511     int skip_dct[8];
1512     int dct_offset   = s->linesize*8; //default for progressive frames
1513     uint8_t *ptr_y, *ptr_cb, *ptr_cr;
1514     int wrap_y, wrap_c;
1515
1516     for(i=0; i<mb_block_count; i++) skip_dct[i]=s->skipdct;
1517
1518     if(s->adaptive_quant){
1519         const int last_qp= s->qscale;
1520         const int mb_xy= mb_x + mb_y*s->mb_stride;
1521
1522         s->lambda= s->lambda_table[mb_xy];
1523         update_qscale(s);
1524
1525         if(!(s->flags&CODEC_FLAG_QP_RD)){
1526             s->qscale = s->current_picture_ptr->f.qscale_table[mb_xy];
1527             s->dquant= s->qscale - last_qp;
1528
1529             if(s->out_format==FMT_H263){
1530                 s->dquant= av_clip(s->dquant, -2, 2);
1531
1532                 if(s->codec_id==CODEC_ID_MPEG4){
1533                     if(!s->mb_intra){
1534                         if(s->pict_type == AV_PICTURE_TYPE_B){
1535                             if(s->dquant&1 || s->mv_dir&MV_DIRECT)
1536                                 s->dquant= 0;
1537                         }
1538                         if(s->mv_type==MV_TYPE_8X8)
1539                             s->dquant=0;
1540                     }
1541                 }
1542             }
1543         }
1544         ff_set_qscale(s, last_qp + s->dquant);
1545     }else if(s->flags&CODEC_FLAG_QP_RD)
1546         ff_set_qscale(s, s->qscale + s->dquant);
1547
1548     wrap_y = s->linesize;
1549     wrap_c = s->uvlinesize;
1550     ptr_y  = s->new_picture.f.data[0] + (mb_y * 16 * wrap_y) + mb_x * 16;
1551     ptr_cb = s->new_picture.f.data[1] + (mb_y * mb_block_height * wrap_c) + mb_x * 8;
1552     ptr_cr = s->new_picture.f.data[2] + (mb_y * mb_block_height * wrap_c) + mb_x * 8;
1553
1554     if(mb_x*16+16 > s->width || mb_y*16+16 > s->height){
1555         uint8_t *ebuf= s->edge_emu_buffer + 32;
1556         s->dsp.emulated_edge_mc(ebuf            , ptr_y , wrap_y,16,16,mb_x*16,mb_y*16, s->width   , s->height);
1557         ptr_y= ebuf;
1558         s->dsp.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);
1559         ptr_cb= ebuf+18*wrap_y;
1560         s->dsp.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);
1561         ptr_cr= ebuf+18*wrap_y+8;
1562     }
1563
1564     if (s->mb_intra) {
1565         if(s->flags&CODEC_FLAG_INTERLACED_DCT){
1566             int progressive_score, interlaced_score;
1567
1568             s->interlaced_dct=0;
1569             progressive_score= s->dsp.ildct_cmp[4](s, ptr_y           , NULL, wrap_y, 8)
1570                               +s->dsp.ildct_cmp[4](s, ptr_y + wrap_y*8, NULL, wrap_y, 8) - 400;
1571
1572             if(progressive_score > 0){
1573                 interlaced_score = s->dsp.ildct_cmp[4](s, ptr_y           , NULL, wrap_y*2, 8)
1574                                   +s->dsp.ildct_cmp[4](s, ptr_y + wrap_y  , NULL, wrap_y*2, 8);
1575                 if(progressive_score > interlaced_score){
1576                     s->interlaced_dct=1;
1577
1578                     dct_offset= wrap_y;
1579                     wrap_y<<=1;
1580                     if (s->chroma_format == CHROMA_422)
1581                         wrap_c<<=1;
1582                 }
1583             }
1584         }
1585
1586         s->dsp.get_pixels(s->block[0], ptr_y                 , wrap_y);
1587         s->dsp.get_pixels(s->block[1], ptr_y              + 8, wrap_y);
1588         s->dsp.get_pixels(s->block[2], ptr_y + dct_offset    , wrap_y);
1589         s->dsp.get_pixels(s->block[3], ptr_y + dct_offset + 8, wrap_y);
1590
1591         if(s->flags&CODEC_FLAG_GRAY){
1592             skip_dct[4]= 1;
1593             skip_dct[5]= 1;
1594         }else{
1595             s->dsp.get_pixels(s->block[4], ptr_cb, wrap_c);
1596             s->dsp.get_pixels(s->block[5], ptr_cr, wrap_c);
1597             if(!s->chroma_y_shift){ /* 422 */
1598                 s->dsp.get_pixels(s->block[6], ptr_cb + (dct_offset>>1), wrap_c);
1599                 s->dsp.get_pixels(s->block[7], ptr_cr + (dct_offset>>1), wrap_c);
1600             }
1601         }
1602     }else{
1603         op_pixels_func (*op_pix)[4];
1604         qpel_mc_func (*op_qpix)[16];
1605         uint8_t *dest_y, *dest_cb, *dest_cr;
1606
1607         dest_y  = s->dest[0];
1608         dest_cb = s->dest[1];
1609         dest_cr = s->dest[2];
1610
1611         if ((!s->no_rounding) || s->pict_type==AV_PICTURE_TYPE_B){
1612             op_pix = s->dsp.put_pixels_tab;
1613             op_qpix= s->dsp.put_qpel_pixels_tab;
1614         }else{
1615             op_pix = s->dsp.put_no_rnd_pixels_tab;
1616             op_qpix= s->dsp.put_no_rnd_qpel_pixels_tab;
1617         }
1618
1619         if (s->mv_dir & MV_DIR_FORWARD) {
1620             MPV_motion(s, dest_y, dest_cb, dest_cr, 0, s->last_picture.f.data, op_pix, op_qpix);
1621             op_pix = s->dsp.avg_pixels_tab;
1622             op_qpix= s->dsp.avg_qpel_pixels_tab;
1623         }
1624         if (s->mv_dir & MV_DIR_BACKWARD) {
1625             MPV_motion(s, dest_y, dest_cb, dest_cr, 1, s->next_picture.f.data, op_pix, op_qpix);
1626         }
1627
1628         if(s->flags&CODEC_FLAG_INTERLACED_DCT){
1629             int progressive_score, interlaced_score;
1630
1631             s->interlaced_dct=0;
1632             progressive_score= s->dsp.ildct_cmp[0](s, dest_y           , ptr_y           , wrap_y, 8)
1633                               +s->dsp.ildct_cmp[0](s, dest_y + wrap_y*8, ptr_y + wrap_y*8, wrap_y, 8) - 400;
1634
1635             if(s->avctx->ildct_cmp == FF_CMP_VSSE) progressive_score -= 400;
1636
1637             if(progressive_score>0){
1638                 interlaced_score = s->dsp.ildct_cmp[0](s, dest_y           , ptr_y           , wrap_y*2, 8)
1639                                   +s->dsp.ildct_cmp[0](s, dest_y + wrap_y  , ptr_y + wrap_y  , wrap_y*2, 8);
1640
1641                 if(progressive_score > interlaced_score){
1642                     s->interlaced_dct=1;
1643
1644                     dct_offset= wrap_y;
1645                     wrap_y<<=1;
1646                     if (s->chroma_format == CHROMA_422)
1647                         wrap_c<<=1;
1648                 }
1649             }
1650         }
1651
1652         s->dsp.diff_pixels(s->block[0], ptr_y                 , dest_y                 , wrap_y);
1653         s->dsp.diff_pixels(s->block[1], ptr_y              + 8, dest_y              + 8, wrap_y);
1654         s->dsp.diff_pixels(s->block[2], ptr_y + dct_offset    , dest_y + dct_offset    , wrap_y);
1655         s->dsp.diff_pixels(s->block[3], ptr_y + dct_offset + 8, dest_y + dct_offset + 8, wrap_y);
1656
1657         if(s->flags&CODEC_FLAG_GRAY){
1658             skip_dct[4]= 1;
1659             skip_dct[5]= 1;
1660         }else{
1661             s->dsp.diff_pixels(s->block[4], ptr_cb, dest_cb, wrap_c);
1662             s->dsp.diff_pixels(s->block[5], ptr_cr, dest_cr, wrap_c);
1663             if(!s->chroma_y_shift){ /* 422 */
1664                 s->dsp.diff_pixels(s->block[6], ptr_cb + (dct_offset>>1), dest_cb + (dct_offset>>1), wrap_c);
1665                 s->dsp.diff_pixels(s->block[7], ptr_cr + (dct_offset>>1), dest_cr + (dct_offset>>1), wrap_c);
1666             }
1667         }
1668         /* pre quantization */
1669         if(s->current_picture.mc_mb_var[s->mb_stride*mb_y+ mb_x]<2*s->qscale*s->qscale){
1670             //FIXME optimize
1671             if(s->dsp.sad[1](NULL, ptr_y               , dest_y               , wrap_y, 8) < 20*s->qscale) skip_dct[0]= 1;
1672             if(s->dsp.sad[1](NULL, ptr_y            + 8, dest_y            + 8, wrap_y, 8) < 20*s->qscale) skip_dct[1]= 1;
1673             if(s->dsp.sad[1](NULL, ptr_y +dct_offset   , dest_y +dct_offset   , wrap_y, 8) < 20*s->qscale) skip_dct[2]= 1;
1674             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;
1675             if(s->dsp.sad[1](NULL, ptr_cb              , dest_cb              , wrap_c, 8) < 20*s->qscale) skip_dct[4]= 1;
1676             if(s->dsp.sad[1](NULL, ptr_cr              , dest_cr              , wrap_c, 8) < 20*s->qscale) skip_dct[5]= 1;
1677             if(!s->chroma_y_shift){ /* 422 */
1678                 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;
1679                 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;
1680             }
1681         }
1682     }
1683
1684     if(s->avctx->quantizer_noise_shaping){
1685         if(!skip_dct[0]) get_visual_weight(weight[0], ptr_y                 , wrap_y);
1686         if(!skip_dct[1]) get_visual_weight(weight[1], ptr_y              + 8, wrap_y);
1687         if(!skip_dct[2]) get_visual_weight(weight[2], ptr_y + dct_offset    , wrap_y);
1688         if(!skip_dct[3]) get_visual_weight(weight[3], ptr_y + dct_offset + 8, wrap_y);
1689         if(!skip_dct[4]) get_visual_weight(weight[4], ptr_cb                , wrap_c);
1690         if(!skip_dct[5]) get_visual_weight(weight[5], ptr_cr                , wrap_c);
1691         if(!s->chroma_y_shift){ /* 422 */
1692             if(!skip_dct[6]) get_visual_weight(weight[6], ptr_cb + (dct_offset>>1), wrap_c);
1693             if(!skip_dct[7]) get_visual_weight(weight[7], ptr_cr + (dct_offset>>1), wrap_c);
1694         }
1695         memcpy(orig[0], s->block[0], sizeof(DCTELEM)*64*mb_block_count);
1696     }
1697
1698     /* DCT & quantize */
1699     assert(s->out_format!=FMT_MJPEG || s->qscale==8);
1700     {
1701         for(i=0;i<mb_block_count;i++) {
1702             if(!skip_dct[i]){
1703                 int overflow;
1704                 s->block_last_index[i] = s->dct_quantize(s, s->block[i], i, s->qscale, &overflow);
1705             // FIXME we could decide to change to quantizer instead of clipping
1706             // JS: I don't think that would be a good idea it could lower quality instead
1707             //     of improve it. Just INTRADC clipping deserves changes in quantizer
1708                 if (overflow) clip_coeffs(s, s->block[i], s->block_last_index[i]);
1709             }else
1710                 s->block_last_index[i]= -1;
1711         }
1712         if(s->avctx->quantizer_noise_shaping){
1713             for(i=0;i<mb_block_count;i++) {
1714                 if(!skip_dct[i]){
1715                     s->block_last_index[i] = dct_quantize_refine(s, s->block[i], weight[i], orig[i], i, s->qscale);
1716                 }
1717             }
1718         }
1719
1720         if(s->luma_elim_threshold && !s->mb_intra)
1721             for(i=0; i<4; i++)
1722                 dct_single_coeff_elimination(s, i, s->luma_elim_threshold);
1723         if(s->chroma_elim_threshold && !s->mb_intra)
1724             for(i=4; i<mb_block_count; i++)
1725                 dct_single_coeff_elimination(s, i, s->chroma_elim_threshold);
1726
1727         if(s->flags & CODEC_FLAG_CBP_RD){
1728             for(i=0;i<mb_block_count;i++) {
1729                 if(s->block_last_index[i] == -1)
1730                     s->coded_score[i]= INT_MAX/256;
1731             }
1732         }
1733     }
1734
1735     if((s->flags&CODEC_FLAG_GRAY) && s->mb_intra){
1736         s->block_last_index[4]=
1737         s->block_last_index[5]= 0;
1738         s->block[4][0]=
1739         s->block[5][0]= (1024 + s->c_dc_scale/2)/ s->c_dc_scale;
1740     }
1741
1742     //non c quantize code returns incorrect block_last_index FIXME
1743     if(s->alternate_scan && s->dct_quantize != dct_quantize_c){
1744         for(i=0; i<mb_block_count; i++){
1745             int j;
1746             if(s->block_last_index[i]>0){
1747                 for(j=63; j>0; j--){
1748                     if(s->block[i][ s->intra_scantable.permutated[j] ]) break;
1749                 }
1750                 s->block_last_index[i]= j;
1751             }
1752         }
1753     }
1754
1755     /* huffman encode */
1756     switch(s->codec_id){ //FIXME funct ptr could be slightly faster
1757     case CODEC_ID_MPEG1VIDEO:
1758     case CODEC_ID_MPEG2VIDEO:
1759         if (CONFIG_MPEG1VIDEO_ENCODER || CONFIG_MPEG2VIDEO_ENCODER)
1760             mpeg1_encode_mb(s, s->block, motion_x, motion_y);
1761         break;
1762     case CODEC_ID_MPEG4:
1763         if (CONFIG_MPEG4_ENCODER)
1764             mpeg4_encode_mb(s, s->block, motion_x, motion_y);
1765         break;
1766     case CODEC_ID_MSMPEG4V2:
1767     case CODEC_ID_MSMPEG4V3:
1768     case CODEC_ID_WMV1:
1769         if (CONFIG_MSMPEG4_ENCODER)
1770             msmpeg4_encode_mb(s, s->block, motion_x, motion_y);
1771         break;
1772     case CODEC_ID_WMV2:
1773         if (CONFIG_WMV2_ENCODER)
1774             ff_wmv2_encode_mb(s, s->block, motion_x, motion_y);
1775         break;
1776     case CODEC_ID_H261:
1777         if (CONFIG_H261_ENCODER)
1778             ff_h261_encode_mb(s, s->block, motion_x, motion_y);
1779         break;
1780     case CODEC_ID_H263:
1781     case CODEC_ID_H263P:
1782     case CODEC_ID_FLV1:
1783     case CODEC_ID_RV10:
1784     case CODEC_ID_RV20:
1785         if (CONFIG_H263_ENCODER)
1786             h263_encode_mb(s, s->block, motion_x, motion_y);
1787         break;
1788     case CODEC_ID_MJPEG:
1789         if (CONFIG_MJPEG_ENCODER)
1790             ff_mjpeg_encode_mb(s, s->block);
1791         break;
1792     default:
1793         assert(0);
1794     }
1795 }
1796
1797 static av_always_inline void encode_mb(MpegEncContext *s, int motion_x, int motion_y)
1798 {
1799     if (s->chroma_format == CHROMA_420) encode_mb_internal(s, motion_x, motion_y,  8, 6);
1800     else                                encode_mb_internal(s, motion_x, motion_y, 16, 8);
1801 }
1802
1803 static inline void copy_context_before_encode(MpegEncContext *d, MpegEncContext *s, int type){
1804     int i;
1805
1806     memcpy(d->last_mv, s->last_mv, 2*2*2*sizeof(int)); //FIXME is memcpy faster than a loop?
1807
1808     /* mpeg1 */
1809     d->mb_skip_run= s->mb_skip_run;
1810     for(i=0; i<3; i++)
1811         d->last_dc[i]= s->last_dc[i];
1812
1813     /* statistics */
1814     d->mv_bits= s->mv_bits;
1815     d->i_tex_bits= s->i_tex_bits;
1816     d->p_tex_bits= s->p_tex_bits;
1817     d->i_count= s->i_count;
1818     d->f_count= s->f_count;
1819     d->b_count= s->b_count;
1820     d->skip_count= s->skip_count;
1821     d->misc_bits= s->misc_bits;
1822     d->last_bits= 0;
1823
1824     d->mb_skipped= 0;
1825     d->qscale= s->qscale;
1826     d->dquant= s->dquant;
1827
1828     d->esc3_level_length= s->esc3_level_length;
1829 }
1830
1831 static inline void copy_context_after_encode(MpegEncContext *d, MpegEncContext *s, int type){
1832     int i;
1833
1834     memcpy(d->mv, s->mv, 2*4*2*sizeof(int));
1835     memcpy(d->last_mv, s->last_mv, 2*2*2*sizeof(int)); //FIXME is memcpy faster than a loop?
1836
1837     /* mpeg1 */
1838     d->mb_skip_run= s->mb_skip_run;
1839     for(i=0; i<3; i++)
1840         d->last_dc[i]= s->last_dc[i];
1841
1842     /* statistics */
1843     d->mv_bits= s->mv_bits;
1844     d->i_tex_bits= s->i_tex_bits;
1845     d->p_tex_bits= s->p_tex_bits;
1846     d->i_count= s->i_count;
1847     d->f_count= s->f_count;
1848     d->b_count= s->b_count;
1849     d->skip_count= s->skip_count;
1850     d->misc_bits= s->misc_bits;
1851
1852     d->mb_intra= s->mb_intra;
1853     d->mb_skipped= s->mb_skipped;
1854     d->mv_type= s->mv_type;
1855     d->mv_dir= s->mv_dir;
1856     d->pb= s->pb;
1857     if(s->data_partitioning){
1858         d->pb2= s->pb2;
1859         d->tex_pb= s->tex_pb;
1860     }
1861     d->block= s->block;
1862     for(i=0; i<8; i++)
1863         d->block_last_index[i]= s->block_last_index[i];
1864     d->interlaced_dct= s->interlaced_dct;
1865     d->qscale= s->qscale;
1866
1867     d->esc3_level_length= s->esc3_level_length;
1868 }
1869
1870 static inline void encode_mb_hq(MpegEncContext *s, MpegEncContext *backup, MpegEncContext *best, int type,
1871                            PutBitContext pb[2], PutBitContext pb2[2], PutBitContext tex_pb[2],
1872                            int *dmin, int *next_block, int motion_x, int motion_y)
1873 {
1874     int score;
1875     uint8_t *dest_backup[3];
1876
1877     copy_context_before_encode(s, backup, type);
1878
1879     s->block= s->blocks[*next_block];
1880     s->pb= pb[*next_block];
1881     if(s->data_partitioning){
1882         s->pb2   = pb2   [*next_block];
1883         s->tex_pb= tex_pb[*next_block];
1884     }
1885
1886     if(*next_block){
1887         memcpy(dest_backup, s->dest, sizeof(s->dest));
1888         s->dest[0] = s->rd_scratchpad;
1889         s->dest[1] = s->rd_scratchpad + 16*s->linesize;
1890         s->dest[2] = s->rd_scratchpad + 16*s->linesize + 8;
1891         assert(s->linesize >= 32); //FIXME
1892     }
1893
1894     encode_mb(s, motion_x, motion_y);
1895
1896     score= put_bits_count(&s->pb);
1897     if(s->data_partitioning){
1898         score+= put_bits_count(&s->pb2);
1899         score+= put_bits_count(&s->tex_pb);
1900     }
1901
1902     if(s->avctx->mb_decision == FF_MB_DECISION_RD){
1903         MPV_decode_mb(s, s->block);
1904
1905         score *= s->lambda2;
1906         score += sse_mb(s) << FF_LAMBDA_SHIFT;
1907     }
1908
1909     if(*next_block){
1910         memcpy(s->dest, dest_backup, sizeof(s->dest));
1911     }
1912
1913     if(score<*dmin){
1914         *dmin= score;
1915         *next_block^=1;
1916
1917         copy_context_after_encode(best, s, type);
1918     }
1919 }
1920
1921 static int sse(MpegEncContext *s, uint8_t *src1, uint8_t *src2, int w, int h, int stride){
1922     uint32_t *sq = ff_squareTbl + 256;
1923     int acc=0;
1924     int x,y;
1925
1926     if(w==16 && h==16)
1927         return s->dsp.sse[0](NULL, src1, src2, stride, 16);
1928     else if(w==8 && h==8)
1929         return s->dsp.sse[1](NULL, src1, src2, stride, 8);
1930
1931     for(y=0; y<h; y++){
1932         for(x=0; x<w; x++){
1933             acc+= sq[src1[x + y*stride] - src2[x + y*stride]];
1934         }
1935     }
1936
1937     assert(acc>=0);
1938
1939     return acc;
1940 }
1941
1942 static int sse_mb(MpegEncContext *s){
1943     int w= 16;
1944     int h= 16;
1945
1946     if(s->mb_x*16 + 16 > s->width ) w= s->width - s->mb_x*16;
1947     if(s->mb_y*16 + 16 > s->height) h= s->height- s->mb_y*16;
1948
1949     if(w==16 && h==16)
1950       if(s->avctx->mb_cmp == FF_CMP_NSSE){
1951         return  s->dsp.nsse[0](s, s->new_picture.f.data[0] + s->mb_x*16 + s->mb_y*s->linesize*16, s->dest[0], s->linesize, 16)
1952                +s->dsp.nsse[1](s, s->new_picture.f.data[1] + s->mb_x*8  + s->mb_y*s->uvlinesize*8,s->dest[1], s->uvlinesize, 8)
1953                +s->dsp.nsse[1](s, s->new_picture.f.data[2] + s->mb_x*8  + s->mb_y*s->uvlinesize*8,s->dest[2], s->uvlinesize, 8);
1954       }else{
1955         return  s->dsp.sse[0](NULL, s->new_picture.f.data[0] + s->mb_x*16 + s->mb_y*s->linesize*16, s->dest[0], s->linesize, 16)
1956                +s->dsp.sse[1](NULL, s->new_picture.f.data[1] + s->mb_x*8  + s->mb_y*s->uvlinesize*8,s->dest[1], s->uvlinesize, 8)
1957                +s->dsp.sse[1](NULL, s->new_picture.f.data[2] + s->mb_x*8  + s->mb_y*s->uvlinesize*8,s->dest[2], s->uvlinesize, 8);
1958       }
1959     else
1960         return  sse(s, s->new_picture.f.data[0] + s->mb_x*16 + s->mb_y*s->linesize*16, s->dest[0], w, h, s->linesize)
1961                +sse(s, s->new_picture.f.data[1] + s->mb_x*8  + s->mb_y*s->uvlinesize*8,s->dest[1], w>>1, h>>1, s->uvlinesize)
1962                +sse(s, s->new_picture.f.data[2] + s->mb_x*8  + s->mb_y*s->uvlinesize*8,s->dest[2], w>>1, h>>1, s->uvlinesize);
1963 }
1964
1965 static int pre_estimate_motion_thread(AVCodecContext *c, void *arg){
1966     MpegEncContext *s= *(void**)arg;
1967
1968
1969     s->me.pre_pass=1;
1970     s->me.dia_size= s->avctx->pre_dia_size;
1971     s->first_slice_line=1;
1972     for(s->mb_y= s->end_mb_y-1; s->mb_y >= s->start_mb_y; s->mb_y--) {
1973         for(s->mb_x=s->mb_width-1; s->mb_x >=0 ;s->mb_x--) {
1974             ff_pre_estimate_p_frame_motion(s, s->mb_x, s->mb_y);
1975         }
1976         s->first_slice_line=0;
1977     }
1978
1979     s->me.pre_pass=0;
1980
1981     return 0;
1982 }
1983
1984 static int estimate_motion_thread(AVCodecContext *c, void *arg){
1985     MpegEncContext *s= *(void**)arg;
1986
1987     ff_check_alignment();
1988
1989     s->me.dia_size= s->avctx->dia_size;
1990     s->first_slice_line=1;
1991     for(s->mb_y= s->start_mb_y; s->mb_y < s->end_mb_y; s->mb_y++) {
1992         s->mb_x=0; //for block init below
1993         ff_init_block_index(s);
1994         for(s->mb_x=0; s->mb_x < s->mb_width; s->mb_x++) {
1995             s->block_index[0]+=2;
1996             s->block_index[1]+=2;
1997             s->block_index[2]+=2;
1998             s->block_index[3]+=2;
1999
2000             /* compute motion vector & mb_type and store in context */
2001             if(s->pict_type==AV_PICTURE_TYPE_B)
2002                 ff_estimate_b_frame_motion(s, s->mb_x, s->mb_y);
2003             else
2004                 ff_estimate_p_frame_motion(s, s->mb_x, s->mb_y);
2005         }
2006         s->first_slice_line=0;
2007     }
2008     return 0;
2009 }
2010
2011 static int mb_var_thread(AVCodecContext *c, void *arg){
2012     MpegEncContext *s= *(void**)arg;
2013     int mb_x, mb_y;
2014
2015     ff_check_alignment();
2016
2017     for(mb_y=s->start_mb_y; mb_y < s->end_mb_y; mb_y++) {
2018         for(mb_x=0; mb_x < s->mb_width; mb_x++) {
2019             int xx = mb_x * 16;
2020             int yy = mb_y * 16;
2021             uint8_t *pix = s->new_picture.f.data[0] + (yy * s->linesize) + xx;
2022             int varc;
2023             int sum = s->dsp.pix_sum(pix, s->linesize);
2024
2025             varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
2026
2027             s->current_picture.mb_var [s->mb_stride * mb_y + mb_x] = varc;
2028             s->current_picture.mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
2029             s->me.mb_var_sum_temp    += varc;
2030         }
2031     }
2032     return 0;
2033 }
2034
2035 static void write_slice_end(MpegEncContext *s){
2036     if(CONFIG_MPEG4_ENCODER && s->codec_id==CODEC_ID_MPEG4){
2037         if(s->partitioned_frame){
2038             ff_mpeg4_merge_partitions(s);
2039         }
2040
2041         ff_mpeg4_stuffing(&s->pb);
2042     }else if(CONFIG_MJPEG_ENCODER && s->out_format == FMT_MJPEG){
2043         ff_mjpeg_encode_stuffing(&s->pb);
2044     }
2045
2046     align_put_bits(&s->pb);
2047     flush_put_bits(&s->pb);
2048
2049     if((s->flags&CODEC_FLAG_PASS1) && !s->partitioned_frame)
2050         s->misc_bits+= get_bits_diff(s);
2051 }
2052
2053 static int encode_thread(AVCodecContext *c, void *arg){
2054     MpegEncContext *s= *(void**)arg;
2055     int mb_x, mb_y, pdif = 0;
2056     int chr_h= 16>>s->chroma_y_shift;
2057     int i, j;
2058     MpegEncContext best_s, backup_s;
2059     uint8_t bit_buf[2][MAX_MB_BYTES];
2060     uint8_t bit_buf2[2][MAX_MB_BYTES];
2061     uint8_t bit_buf_tex[2][MAX_MB_BYTES];
2062     PutBitContext pb[2], pb2[2], tex_pb[2];
2063 //printf("%d->%d\n", s->resync_mb_y, s->end_mb_y);
2064
2065     ff_check_alignment();
2066
2067     for(i=0; i<2; i++){
2068         init_put_bits(&pb    [i], bit_buf    [i], MAX_MB_BYTES);
2069         init_put_bits(&pb2   [i], bit_buf2   [i], MAX_MB_BYTES);
2070         init_put_bits(&tex_pb[i], bit_buf_tex[i], MAX_MB_BYTES);
2071     }
2072
2073     s->last_bits= put_bits_count(&s->pb);
2074     s->mv_bits=0;
2075     s->misc_bits=0;
2076     s->i_tex_bits=0;
2077     s->p_tex_bits=0;
2078     s->i_count=0;
2079     s->f_count=0;
2080     s->b_count=0;
2081     s->skip_count=0;
2082
2083     for(i=0; i<3; i++){
2084         /* init last dc values */
2085         /* note: quant matrix value (8) is implied here */
2086         s->last_dc[i] = 128 << s->intra_dc_precision;
2087
2088         s->current_picture.f.error[i] = 0;
2089     }
2090     s->mb_skip_run = 0;
2091     memset(s->last_mv, 0, sizeof(s->last_mv));
2092
2093     s->last_mv_dir = 0;
2094
2095     switch(s->codec_id){
2096     case CODEC_ID_H263:
2097     case CODEC_ID_H263P:
2098     case CODEC_ID_FLV1:
2099         if (CONFIG_H263_ENCODER)
2100             s->gob_index = ff_h263_get_gob_height(s);
2101         break;
2102     case CODEC_ID_MPEG4:
2103         if(CONFIG_MPEG4_ENCODER && s->partitioned_frame)
2104             ff_mpeg4_init_partitions(s);
2105         break;
2106     }
2107
2108     s->resync_mb_x=0;
2109     s->resync_mb_y=0;
2110     s->first_slice_line = 1;
2111     s->ptr_lastgob = s->pb.buf;
2112     for(mb_y= s->start_mb_y; mb_y < s->end_mb_y; mb_y++) {
2113 //    printf("row %d at %X\n", s->mb_y, (int)s);
2114         s->mb_x=0;
2115         s->mb_y= mb_y;
2116
2117         ff_set_qscale(s, s->qscale);
2118         ff_init_block_index(s);
2119
2120         for(mb_x=0; mb_x < s->mb_width; mb_x++) {
2121             int xy= mb_y*s->mb_stride + mb_x; // removed const, H261 needs to adjust this
2122             int mb_type= s->mb_type[xy];
2123 //            int d;
2124             int dmin= INT_MAX;
2125             int dir;
2126
2127             if(s->pb.buf_end - s->pb.buf - (put_bits_count(&s->pb)>>3) < MAX_MB_BYTES){
2128                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
2129                 return -1;
2130             }
2131             if(s->data_partitioning){
2132                 if(   s->pb2   .buf_end - s->pb2   .buf - (put_bits_count(&s->    pb2)>>3) < MAX_MB_BYTES
2133                    || s->tex_pb.buf_end - s->tex_pb.buf - (put_bits_count(&s->tex_pb )>>3) < MAX_MB_BYTES){
2134                     av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
2135                     return -1;
2136                 }
2137             }
2138
2139             s->mb_x = mb_x;
2140             s->mb_y = mb_y;  // moved into loop, can get changed by H.261
2141             ff_update_block_index(s);
2142
2143             if(CONFIG_H261_ENCODER && s->codec_id == CODEC_ID_H261){
2144                 ff_h261_reorder_mb_index(s);
2145                 xy= s->mb_y*s->mb_stride + s->mb_x;
2146                 mb_type= s->mb_type[xy];
2147             }
2148
2149             /* write gob / video packet header  */
2150             if(s->rtp_mode){
2151                 int current_packet_size, is_gob_start;
2152
2153                 current_packet_size= ((put_bits_count(&s->pb)+7)>>3) - (s->ptr_lastgob - s->pb.buf);
2154
2155                 is_gob_start= s->avctx->rtp_payload_size && current_packet_size >= s->avctx->rtp_payload_size && mb_y + mb_x>0;
2156
2157                 if(s->start_mb_y == mb_y && mb_y > 0 && mb_x==0) is_gob_start=1;
2158
2159                 switch(s->codec_id){
2160                 case CODEC_ID_H263:
2161                 case CODEC_ID_H263P:
2162                     if(!s->h263_slice_structured)
2163                         if(s->mb_x || s->mb_y%s->gob_index) is_gob_start=0;
2164                     break;
2165                 case CODEC_ID_MPEG2VIDEO:
2166                     if(s->mb_x==0 && s->mb_y!=0) is_gob_start=1;
2167                 case CODEC_ID_MPEG1VIDEO:
2168                     if(s->mb_skip_run) is_gob_start=0;
2169                     break;
2170                 }
2171
2172                 if(is_gob_start){
2173                     if(s->start_mb_y != mb_y || mb_x!=0){
2174                         write_slice_end(s);
2175
2176                         if(CONFIG_MPEG4_ENCODER && s->codec_id==CODEC_ID_MPEG4 && s->partitioned_frame){
2177                             ff_mpeg4_init_partitions(s);
2178                         }
2179                     }
2180
2181                     assert((put_bits_count(&s->pb)&7) == 0);
2182                     current_packet_size= put_bits_ptr(&s->pb) - s->ptr_lastgob;
2183
2184                     if(s->avctx->error_rate && s->resync_mb_x + s->resync_mb_y > 0){
2185                         int r= put_bits_count(&s->pb)/8 + s->picture_number + 16 + s->mb_x + s->mb_y;
2186                         int d= 100 / s->avctx->error_rate;
2187                         if(r % d == 0){
2188                             current_packet_size=0;
2189                             s->pb.buf_ptr= s->ptr_lastgob;
2190                             assert(put_bits_ptr(&s->pb) == s->ptr_lastgob);
2191                         }
2192                     }
2193
2194                     if (s->avctx->rtp_callback){
2195                         int number_mb = (mb_y - s->resync_mb_y)*s->mb_width + mb_x - s->resync_mb_x;
2196                         s->avctx->rtp_callback(s->avctx, s->ptr_lastgob, current_packet_size, number_mb);
2197                     }
2198
2199                     switch(s->codec_id){
2200                     case CODEC_ID_MPEG4:
2201                         if (CONFIG_MPEG4_ENCODER) {
2202                             ff_mpeg4_encode_video_packet_header(s);
2203                             ff_mpeg4_clean_buffers(s);
2204                         }
2205                     break;
2206                     case CODEC_ID_MPEG1VIDEO:
2207                     case CODEC_ID_MPEG2VIDEO:
2208                         if (CONFIG_MPEG1VIDEO_ENCODER || CONFIG_MPEG2VIDEO_ENCODER) {
2209                             ff_mpeg1_encode_slice_header(s);
2210                             ff_mpeg1_clean_buffers(s);
2211                         }
2212                     break;
2213                     case CODEC_ID_H263:
2214                     case CODEC_ID_H263P:
2215                         if (CONFIG_H263_ENCODER)
2216                             h263_encode_gob_header(s, mb_y);
2217                     break;
2218                     }
2219
2220                     if(s->flags&CODEC_FLAG_PASS1){
2221                         int bits= put_bits_count(&s->pb);
2222                         s->misc_bits+= bits - s->last_bits;
2223                         s->last_bits= bits;
2224                     }
2225
2226                     s->ptr_lastgob += current_packet_size;
2227                     s->first_slice_line=1;
2228                     s->resync_mb_x=mb_x;
2229                     s->resync_mb_y=mb_y;
2230                 }
2231             }
2232
2233             if(  (s->resync_mb_x   == s->mb_x)
2234                && s->resync_mb_y+1 == s->mb_y){
2235                 s->first_slice_line=0;
2236             }
2237
2238             s->mb_skipped=0;
2239             s->dquant=0; //only for QP_RD
2240
2241             if(mb_type & (mb_type-1) || (s->flags & CODEC_FLAG_QP_RD)){ // more than 1 MB type possible or CODEC_FLAG_QP_RD
2242                 int next_block=0;
2243                 int pb_bits_count, pb2_bits_count, tex_pb_bits_count;
2244
2245                 copy_context_before_encode(&backup_s, s, -1);
2246                 backup_s.pb= s->pb;
2247                 best_s.data_partitioning= s->data_partitioning;
2248                 best_s.partitioned_frame= s->partitioned_frame;
2249                 if(s->data_partitioning){
2250                     backup_s.pb2= s->pb2;
2251                     backup_s.tex_pb= s->tex_pb;
2252                 }
2253
2254                 if(mb_type&CANDIDATE_MB_TYPE_INTER){
2255                     s->mv_dir = MV_DIR_FORWARD;
2256                     s->mv_type = MV_TYPE_16X16;
2257                     s->mb_intra= 0;
2258                     s->mv[0][0][0] = s->p_mv_table[xy][0];
2259                     s->mv[0][0][1] = s->p_mv_table[xy][1];
2260                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_INTER, pb, pb2, tex_pb,
2261                                  &dmin, &next_block, s->mv[0][0][0], s->mv[0][0][1]);
2262                 }
2263                 if(mb_type&CANDIDATE_MB_TYPE_INTER_I){
2264                     s->mv_dir = MV_DIR_FORWARD;
2265                     s->mv_type = MV_TYPE_FIELD;
2266                     s->mb_intra= 0;
2267                     for(i=0; i<2; i++){
2268                         j= s->field_select[0][i] = s->p_field_select_table[i][xy];
2269                         s->mv[0][i][0] = s->p_field_mv_table[i][j][xy][0];
2270                         s->mv[0][i][1] = s->p_field_mv_table[i][j][xy][1];
2271                     }
2272                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_INTER_I, pb, pb2, tex_pb,
2273                                  &dmin, &next_block, 0, 0);
2274                 }
2275                 if(mb_type&CANDIDATE_MB_TYPE_SKIPPED){
2276                     s->mv_dir = MV_DIR_FORWARD;
2277                     s->mv_type = MV_TYPE_16X16;
2278                     s->mb_intra= 0;
2279                     s->mv[0][0][0] = 0;
2280                     s->mv[0][0][1] = 0;
2281                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_SKIPPED, pb, pb2, tex_pb,
2282                                  &dmin, &next_block, s->mv[0][0][0], s->mv[0][0][1]);
2283                 }
2284                 if(mb_type&CANDIDATE_MB_TYPE_INTER4V){
2285                     s->mv_dir = MV_DIR_FORWARD;
2286                     s->mv_type = MV_TYPE_8X8;
2287                     s->mb_intra= 0;
2288                     for(i=0; i<4; i++){
2289                         s->mv[0][i][0] = s->current_picture.f.motion_val[0][s->block_index[i]][0];
2290                         s->mv[0][i][1] = s->current_picture.f.motion_val[0][s->block_index[i]][1];
2291                     }
2292                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_INTER4V, pb, pb2, tex_pb,
2293                                  &dmin, &next_block, 0, 0);
2294                 }
2295                 if(mb_type&CANDIDATE_MB_TYPE_FORWARD){
2296                     s->mv_dir = MV_DIR_FORWARD;
2297                     s->mv_type = MV_TYPE_16X16;
2298                     s->mb_intra= 0;
2299                     s->mv[0][0][0] = s->b_forw_mv_table[xy][0];
2300                     s->mv[0][0][1] = s->b_forw_mv_table[xy][1];
2301                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_FORWARD, pb, pb2, tex_pb,
2302                                  &dmin, &next_block, s->mv[0][0][0], s->mv[0][0][1]);
2303                 }
2304                 if(mb_type&CANDIDATE_MB_TYPE_BACKWARD){
2305                     s->mv_dir = MV_DIR_BACKWARD;
2306                     s->mv_type = MV_TYPE_16X16;
2307                     s->mb_intra= 0;
2308                     s->mv[1][0][0] = s->b_back_mv_table[xy][0];
2309                     s->mv[1][0][1] = s->b_back_mv_table[xy][1];
2310                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_BACKWARD, pb, pb2, tex_pb,
2311                                  &dmin, &next_block, s->mv[1][0][0], s->mv[1][0][1]);
2312                 }
2313                 if(mb_type&CANDIDATE_MB_TYPE_BIDIR){
2314                     s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD;
2315                     s->mv_type = MV_TYPE_16X16;
2316                     s->mb_intra= 0;
2317                     s->mv[0][0][0] = s->b_bidir_forw_mv_table[xy][0];
2318                     s->mv[0][0][1] = s->b_bidir_forw_mv_table[xy][1];
2319                     s->mv[1][0][0] = s->b_bidir_back_mv_table[xy][0];
2320                     s->mv[1][0][1] = s->b_bidir_back_mv_table[xy][1];
2321                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_BIDIR, pb, pb2, tex_pb,
2322                                  &dmin, &next_block, 0, 0);
2323                 }
2324                 if(mb_type&CANDIDATE_MB_TYPE_FORWARD_I){
2325                     s->mv_dir = MV_DIR_FORWARD;
2326                     s->mv_type = MV_TYPE_FIELD;
2327                     s->mb_intra= 0;
2328                     for(i=0; i<2; i++){
2329                         j= s->field_select[0][i] = s->b_field_select_table[0][i][xy];
2330                         s->mv[0][i][0] = s->b_field_mv_table[0][i][j][xy][0];
2331                         s->mv[0][i][1] = s->b_field_mv_table[0][i][j][xy][1];
2332                     }
2333                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_FORWARD_I, pb, pb2, tex_pb,
2334                                  &dmin, &next_block, 0, 0);
2335                 }
2336                 if(mb_type&CANDIDATE_MB_TYPE_BACKWARD_I){
2337                     s->mv_dir = MV_DIR_BACKWARD;
2338                     s->mv_type = MV_TYPE_FIELD;
2339                     s->mb_intra= 0;
2340                     for(i=0; i<2; i++){
2341                         j= s->field_select[1][i] = s->b_field_select_table[1][i][xy];
2342                         s->mv[1][i][0] = s->b_field_mv_table[1][i][j][xy][0];
2343                         s->mv[1][i][1] = s->b_field_mv_table[1][i][j][xy][1];
2344                     }
2345                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_BACKWARD_I, pb, pb2, tex_pb,
2346                                  &dmin, &next_block, 0, 0);
2347                 }
2348                 if(mb_type&CANDIDATE_MB_TYPE_BIDIR_I){
2349                     s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD;
2350                     s->mv_type = MV_TYPE_FIELD;
2351                     s->mb_intra= 0;
2352                     for(dir=0; dir<2; dir++){
2353                         for(i=0; i<2; i++){
2354                             j= s->field_select[dir][i] = s->b_field_select_table[dir][i][xy];
2355                             s->mv[dir][i][0] = s->b_field_mv_table[dir][i][j][xy][0];
2356                             s->mv[dir][i][1] = s->b_field_mv_table[dir][i][j][xy][1];
2357                         }
2358                     }
2359                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_BIDIR_I, pb, pb2, tex_pb,
2360                                  &dmin, &next_block, 0, 0);
2361                 }
2362                 if(mb_type&CANDIDATE_MB_TYPE_INTRA){
2363                     s->mv_dir = 0;
2364                     s->mv_type = MV_TYPE_16X16;
2365                     s->mb_intra= 1;
2366                     s->mv[0][0][0] = 0;
2367                     s->mv[0][0][1] = 0;
2368                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_INTRA, pb, pb2, tex_pb,
2369                                  &dmin, &next_block, 0, 0);
2370                     if(s->h263_pred || s->h263_aic){
2371                         if(best_s.mb_intra)
2372                             s->mbintra_table[mb_x + mb_y*s->mb_stride]=1;
2373                         else
2374                             ff_clean_intra_table_entries(s); //old mode?
2375                     }
2376                 }
2377
2378                 if((s->flags & CODEC_FLAG_QP_RD) && dmin < INT_MAX){
2379                     if(best_s.mv_type==MV_TYPE_16X16){ //FIXME move 4mv after QPRD
2380                         const int last_qp= backup_s.qscale;
2381                         int qpi, qp, dc[6];
2382                         DCTELEM ac[6][16];
2383                         const int mvdir= (best_s.mv_dir&MV_DIR_BACKWARD) ? 1 : 0;
2384                         static const int dquant_tab[4]={-1,1,-2,2};
2385
2386                         assert(backup_s.dquant == 0);
2387
2388                         //FIXME intra
2389                         s->mv_dir= best_s.mv_dir;
2390                         s->mv_type = MV_TYPE_16X16;
2391                         s->mb_intra= best_s.mb_intra;
2392                         s->mv[0][0][0] = best_s.mv[0][0][0];
2393                         s->mv[0][0][1] = best_s.mv[0][0][1];
2394                         s->mv[1][0][0] = best_s.mv[1][0][0];
2395                         s->mv[1][0][1] = best_s.mv[1][0][1];
2396
2397                         qpi = s->pict_type == AV_PICTURE_TYPE_B ? 2 : 0;
2398                         for(; qpi<4; qpi++){
2399                             int dquant= dquant_tab[qpi];
2400                             qp= last_qp + dquant;
2401                             if(qp < s->avctx->qmin || qp > s->avctx->qmax)
2402                                 continue;
2403                             backup_s.dquant= dquant;
2404                             if(s->mb_intra && s->dc_val[0]){
2405                                 for(i=0; i<6; i++){
2406                                     dc[i]= s->dc_val[0][ s->block_index[i] ];
2407                                     memcpy(ac[i], s->ac_val[0][s->block_index[i]], sizeof(DCTELEM)*16);
2408                                 }
2409                             }
2410
2411                             encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_INTER /* wrong but unused */, pb, pb2, tex_pb,
2412                                          &dmin, &next_block, s->mv[mvdir][0][0], s->mv[mvdir][0][1]);
2413                             if(best_s.qscale != qp){
2414                                 if(s->mb_intra && s->dc_val[0]){
2415                                     for(i=0; i<6; i++){
2416                                         s->dc_val[0][ s->block_index[i] ]= dc[i];
2417                                         memcpy(s->ac_val[0][s->block_index[i]], ac[i], sizeof(DCTELEM)*16);
2418                                     }
2419                                 }
2420                             }
2421                         }
2422                     }
2423                 }
2424                 if(CONFIG_MPEG4_ENCODER && mb_type&CANDIDATE_MB_TYPE_DIRECT){
2425                     int mx= s->b_direct_mv_table[xy][0];
2426                     int my= s->b_direct_mv_table[xy][1];
2427
2428                     backup_s.dquant = 0;
2429                     s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD | MV_DIRECT;
2430                     s->mb_intra= 0;
2431                     ff_mpeg4_set_direct_mv(s, mx, my);
2432                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_DIRECT, pb, pb2, tex_pb,
2433                                  &dmin, &next_block, mx, my);
2434                 }
2435                 if(CONFIG_MPEG4_ENCODER && mb_type&CANDIDATE_MB_TYPE_DIRECT0){
2436                     backup_s.dquant = 0;
2437                     s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD | MV_DIRECT;
2438                     s->mb_intra= 0;
2439                     ff_mpeg4_set_direct_mv(s, 0, 0);
2440                     encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_DIRECT, pb, pb2, tex_pb,
2441                                  &dmin, &next_block, 0, 0);
2442                 }
2443                 if(!best_s.mb_intra && s->flags2&CODEC_FLAG2_SKIP_RD){
2444                     int coded=0;
2445                     for(i=0; i<6; i++)
2446                         coded |= s->block_last_index[i];
2447                     if(coded){
2448                         int mx,my;
2449                         memcpy(s->mv, best_s.mv, sizeof(s->mv));
2450                         if(CONFIG_MPEG4_ENCODER && best_s.mv_dir & MV_DIRECT){
2451                             mx=my=0; //FIXME find the one we actually used
2452                             ff_mpeg4_set_direct_mv(s, mx, my);
2453                         }else if(best_s.mv_dir&MV_DIR_BACKWARD){
2454                             mx= s->mv[1][0][0];
2455                             my= s->mv[1][0][1];
2456                         }else{
2457                             mx= s->mv[0][0][0];
2458                             my= s->mv[0][0][1];
2459                         }
2460
2461                         s->mv_dir= best_s.mv_dir;
2462                         s->mv_type = best_s.mv_type;
2463                         s->mb_intra= 0;
2464 /*                        s->mv[0][0][0] = best_s.mv[0][0][0];
2465                         s->mv[0][0][1] = best_s.mv[0][0][1];
2466                         s->mv[1][0][0] = best_s.mv[1][0][0];
2467                         s->mv[1][0][1] = best_s.mv[1][0][1];*/
2468                         backup_s.dquant= 0;
2469                         s->skipdct=1;
2470                         encode_mb_hq(s, &backup_s, &best_s, CANDIDATE_MB_TYPE_INTER /* wrong but unused */, pb, pb2, tex_pb,
2471                                         &dmin, &next_block, mx, my);
2472                         s->skipdct=0;
2473                     }
2474                 }
2475
2476                 s->current_picture.f.qscale_table[xy] = best_s.qscale;
2477
2478                 copy_context_after_encode(s, &best_s, -1);
2479
2480                 pb_bits_count= put_bits_count(&s->pb);
2481                 flush_put_bits(&s->pb);
2482                 ff_copy_bits(&backup_s.pb, bit_buf[next_block^1], pb_bits_count);
2483                 s->pb= backup_s.pb;
2484
2485                 if(s->data_partitioning){
2486                     pb2_bits_count= put_bits_count(&s->pb2);
2487                     flush_put_bits(&s->pb2);
2488                     ff_copy_bits(&backup_s.pb2, bit_buf2[next_block^1], pb2_bits_count);
2489                     s->pb2= backup_s.pb2;
2490
2491                     tex_pb_bits_count= put_bits_count(&s->tex_pb);
2492                     flush_put_bits(&s->tex_pb);
2493                     ff_copy_bits(&backup_s.tex_pb, bit_buf_tex[next_block^1], tex_pb_bits_count);
2494                     s->tex_pb= backup_s.tex_pb;
2495                 }
2496                 s->last_bits= put_bits_count(&s->pb);
2497
2498                 if (CONFIG_H263_ENCODER &&
2499                     s->out_format == FMT_H263 && s->pict_type!=AV_PICTURE_TYPE_B)
2500                     ff_h263_update_motion_val(s);
2501
2502                 if(next_block==0){ //FIXME 16 vs linesize16
2503                     s->dsp.put_pixels_tab[0][0](s->dest[0], s->rd_scratchpad                     , s->linesize  ,16);
2504                     s->dsp.put_pixels_tab[1][0](s->dest[1], s->rd_scratchpad + 16*s->linesize    , s->uvlinesize, 8);
2505                     s->dsp.put_pixels_tab[1][0](s->dest[2], s->rd_scratchpad + 16*s->linesize + 8, s->uvlinesize, 8);
2506                 }
2507
2508                 if(s->avctx->mb_decision == FF_MB_DECISION_BITS)
2509                     MPV_decode_mb(s, s->block);
2510             } else {
2511                 int motion_x = 0, motion_y = 0;
2512                 s->mv_type=MV_TYPE_16X16;
2513                 // only one MB-Type possible
2514
2515                 switch(mb_type){
2516                 case CANDIDATE_MB_TYPE_INTRA:
2517                     s->mv_dir = 0;
2518                     s->mb_intra= 1;
2519                     motion_x= s->mv[0][0][0] = 0;
2520                     motion_y= s->mv[0][0][1] = 0;
2521                     break;
2522                 case CANDIDATE_MB_TYPE_INTER:
2523                     s->mv_dir = MV_DIR_FORWARD;
2524                     s->mb_intra= 0;
2525                     motion_x= s->mv[0][0][0] = s->p_mv_table[xy][0];
2526                     motion_y= s->mv[0][0][1] = s->p_mv_table[xy][1];
2527                     break;
2528                 case CANDIDATE_MB_TYPE_INTER_I:
2529                     s->mv_dir = MV_DIR_FORWARD;
2530                     s->mv_type = MV_TYPE_FIELD;
2531                     s->mb_intra= 0;
2532                     for(i=0; i<2; i++){
2533                         j= s->field_select[0][i] = s->p_field_select_table[i][xy];
2534                         s->mv[0][i][0] = s->p_field_mv_table[i][j][xy][0];
2535                         s->mv[0][i][1] = s->p_field_mv_table[i][j][xy][1];
2536                     }
2537                     break;
2538                 case CANDIDATE_MB_TYPE_INTER4V:
2539                     s->mv_dir = MV_DIR_FORWARD;
2540                     s->mv_type = MV_TYPE_8X8;
2541                     s->mb_intra= 0;
2542                     for(i=0; i<4; i++){
2543                         s->mv[0][i][0] = s->current_picture.f.motion_val[0][s->block_index[i]][0];
2544                         s->mv[0][i][1] = s->current_picture.f.motion_val[0][s->block_index[i]][1];
2545                     }
2546                     break;
2547                 case CANDIDATE_MB_TYPE_DIRECT:
2548                     if (CONFIG_MPEG4_ENCODER) {
2549                         s->mv_dir = MV_DIR_FORWARD|MV_DIR_BACKWARD|MV_DIRECT;
2550                         s->mb_intra= 0;
2551                         motion_x=s->b_direct_mv_table[xy][0];
2552                         motion_y=s->b_direct_mv_table[xy][1];
2553                         ff_mpeg4_set_direct_mv(s, motion_x, motion_y);
2554                     }
2555                     break;
2556                 case CANDIDATE_MB_TYPE_DIRECT0:
2557                     if (CONFIG_MPEG4_ENCODER) {
2558                         s->mv_dir = MV_DIR_FORWARD|MV_DIR_BACKWARD|MV_DIRECT;
2559                         s->mb_intra= 0;
2560                         ff_mpeg4_set_direct_mv(s, 0, 0);
2561                     }
2562                     break;
2563                 case CANDIDATE_MB_TYPE_BIDIR:
2564                     s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD;
2565                     s->mb_intra= 0;
2566                     s->mv[0][0][0] = s->b_bidir_forw_mv_table[xy][0];
2567                     s->mv[0][0][1] = s->b_bidir_forw_mv_table[xy][1];
2568                     s->mv[1][0][0] = s->b_bidir_back_mv_table[xy][0];
2569                     s->mv[1][0][1] = s->b_bidir_back_mv_table[xy][1];
2570                     break;
2571                 case CANDIDATE_MB_TYPE_BACKWARD:
2572                     s->mv_dir = MV_DIR_BACKWARD;
2573                     s->mb_intra= 0;
2574                     motion_x= s->mv[1][0][0] = s->b_back_mv_table[xy][0];
2575                     motion_y= s->mv[1][0][1] = s->b_back_mv_table[xy][1];
2576                     break;
2577                 case CANDIDATE_MB_TYPE_FORWARD:
2578                     s->mv_dir = MV_DIR_FORWARD;
2579                     s->mb_intra= 0;
2580                     motion_x= s->mv[0][0][0] = s->b_forw_mv_table[xy][0];
2581                     motion_y= s->mv[0][0][1] = s->b_forw_mv_table[xy][1];
2582 //                    printf(" %d %d ", motion_x, motion_y);
2583                     break;
2584                 case CANDIDATE_MB_TYPE_FORWARD_I:
2585                     s->mv_dir = MV_DIR_FORWARD;
2586                     s->mv_type = MV_TYPE_FIELD;
2587                     s->mb_intra= 0;
2588                     for(i=0; i<2; i++){
2589                         j= s->field_select[0][i] = s->b_field_select_table[0][i][xy];
2590                         s->mv[0][i][0] = s->b_field_mv_table[0][i][j][xy][0];
2591                         s->mv[0][i][1] = s->b_field_mv_table[0][i][j][xy][1];
2592                     }
2593                     break;
2594                 case CANDIDATE_MB_TYPE_BACKWARD_I:
2595                     s->mv_dir = MV_DIR_BACKWARD;
2596                     s->mv_type = MV_TYPE_FIELD;
2597                     s->mb_intra= 0;
2598                     for(i=0; i<2; i++){
2599                         j= s->field_select[1][i] = s->b_field_select_table[1][i][xy];
2600                         s->mv[1][i][0] = s->b_field_mv_table[1][i][j][xy][0];
2601                         s->mv[1][i][1] = s->b_field_mv_table[1][i][j][xy][1];
2602                     }
2603                     break;
2604                 case CANDIDATE_MB_TYPE_BIDIR_I:
2605                     s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD;
2606                     s->mv_type = MV_TYPE_FIELD;
2607                     s->mb_intra= 0;
2608                     for(dir=0; dir<2; dir++){
2609                         for(i=0; i<2; i++){
2610                             j= s->field_select[dir][i] = s->b_field_select_table[dir][i][xy];
2611                             s->mv[dir][i][0] = s->b_field_mv_table[dir][i][j][xy][0];
2612                             s->mv[dir][i][1] = s->b_field_mv_table[dir][i][j][xy][1];
2613                         }
2614                     }
2615                     break;
2616                 default:
2617                     av_log(s->avctx, AV_LOG_ERROR, "illegal MB type\n");
2618                 }
2619
2620                 encode_mb(s, motion_x, motion_y);
2621
2622                 // RAL: Update last macroblock type
2623                 s->last_mv_dir = s->mv_dir;
2624
2625                 if (CONFIG_H263_ENCODER &&
2626                     s->out_format == FMT_H263 && s->pict_type!=AV_PICTURE_TYPE_B)
2627                     ff_h263_update_motion_val(s);
2628
2629                 MPV_decode_mb(s, s->block);
2630             }
2631
2632             /* clean the MV table in IPS frames for direct mode in B frames */
2633             if(s->mb_intra /* && I,P,S_TYPE */){
2634                 s->p_mv_table[xy][0]=0;
2635                 s->p_mv_table[xy][1]=0;
2636             }
2637
2638             if(s->flags&CODEC_FLAG_PSNR){
2639                 int w= 16;
2640                 int h= 16;
2641
2642                 if(s->mb_x*16 + 16 > s->width ) w= s->width - s->mb_x*16;
2643                 if(s->mb_y*16 + 16 > s->height) h= s->height- s->mb_y*16;
2644
2645                 s->current_picture.f.error[0] += sse(
2646                     s, s->new_picture.f.data[0] + s->mb_x*16 + s->mb_y*s->linesize*16,
2647                     s->dest[0], w, h, s->linesize);
2648                 s->current_picture.f.error[1] += sse(
2649                     s, s->new_picture.f.data[1] + s->mb_x*8  + s->mb_y*s->uvlinesize*chr_h,
2650                     s->dest[1], w>>1, h>>s->chroma_y_shift, s->uvlinesize);
2651                 s->current_picture.f.error[2] += sse(
2652                     s, s->new_picture.f.data[2] + s->mb_x*8  + s->mb_y*s->uvlinesize*chr_h,
2653                     s->dest[2], w>>1, h>>s->chroma_y_shift, s->uvlinesize);
2654             }
2655             if(s->loop_filter){
2656                 if(CONFIG_H263_ENCODER && s->out_format == FMT_H263)
2657                     ff_h263_loop_filter(s);
2658             }
2659 //printf("MB %d %d bits\n", s->mb_x+s->mb_y*s->mb_stride, put_bits_count(&s->pb));
2660         }
2661     }
2662
2663     //not beautiful here but we must write it before flushing so it has to be here
2664     if (CONFIG_MSMPEG4_ENCODER && s->msmpeg4_version && s->msmpeg4_version<4 && s->pict_type == AV_PICTURE_TYPE_I)
2665         msmpeg4_encode_ext_header(s);
2666
2667     write_slice_end(s);
2668
2669     /* Send the last GOB if RTP */
2670     if (s->avctx->rtp_callback) {
2671         int number_mb = (mb_y - s->resync_mb_y)*s->mb_width - s->resync_mb_x;
2672         pdif = put_bits_ptr(&s->pb) - s->ptr_lastgob;
2673         /* Call the RTP callback to send the last GOB */
2674         emms_c();
2675         s->avctx->rtp_callback(s->avctx, s->ptr_lastgob, pdif, number_mb);
2676     }
2677
2678     return 0;
2679 }
2680
2681 #define MERGE(field) dst->field += src->field; src->field=0
2682 static void merge_context_after_me(MpegEncContext *dst, MpegEncContext *src){
2683     MERGE(me.scene_change_score);
2684     MERGE(me.mc_mb_var_sum_temp);
2685     MERGE(me.mb_var_sum_temp);
2686 }
2687
2688 static void merge_context_after_encode(MpegEncContext *dst, MpegEncContext *src){
2689     int i;
2690
2691     MERGE(dct_count[0]); //note, the other dct vars are not part of the context
2692     MERGE(dct_count[1]);
2693     MERGE(mv_bits);
2694     MERGE(i_tex_bits);
2695     MERGE(p_tex_bits);
2696     MERGE(i_count);
2697     MERGE(f_count);
2698     MERGE(b_count);
2699     MERGE(skip_count);
2700     MERGE(misc_bits);
2701     MERGE(error_count);
2702     MERGE(padding_bug_score);
2703     MERGE(current_picture.f.error[0]);
2704     MERGE(current_picture.f.error[1]);
2705     MERGE(current_picture.f.error[2]);
2706
2707     if(dst->avctx->noise_reduction){
2708         for(i=0; i<64; i++){
2709             MERGE(dct_error_sum[0][i]);
2710             MERGE(dct_error_sum[1][i]);
2711         }
2712     }
2713
2714     assert(put_bits_count(&src->pb) % 8 ==0);
2715     assert(put_bits_count(&dst->pb) % 8 ==0);
2716     ff_copy_bits(&dst->pb, src->pb.buf, put_bits_count(&src->pb));
2717     flush_put_bits(&dst->pb);
2718 }
2719
2720 static int estimate_qp(MpegEncContext *s, int dry_run){
2721     if (s->next_lambda){
2722         s->current_picture_ptr->f.quality =
2723         s->current_picture.f.quality = s->next_lambda;
2724         if(!dry_run) s->next_lambda= 0;
2725     } else if (!s->fixed_qscale) {
2726         s->current_picture_ptr->f.quality =
2727         s->current_picture.f.quality = ff_rate_estimate_qscale(s, dry_run);
2728         if (s->current_picture.f.quality < 0)
2729             return -1;
2730     }
2731
2732     if(s->adaptive_quant){
2733         switch(s->codec_id){
2734         case CODEC_ID_MPEG4:
2735             if (CONFIG_MPEG4_ENCODER)
2736                 ff_clean_mpeg4_qscales(s);
2737             break;
2738         case CODEC_ID_H263:
2739         case CODEC_ID_H263P:
2740         case CODEC_ID_FLV1:
2741             if (CONFIG_H263_ENCODER)
2742                 ff_clean_h263_qscales(s);
2743             break;
2744         default:
2745             ff_init_qscale_tab(s);
2746         }
2747
2748         s->lambda= s->lambda_table[0];
2749         //FIXME broken
2750     }else
2751         s->lambda = s->current_picture.f.quality;
2752 //printf("%d %d\n", s->avctx->global_quality, s->current_picture.quality);
2753     update_qscale(s);
2754     return 0;
2755 }
2756
2757 /* must be called before writing the header */
2758 static void set_frame_distances(MpegEncContext * s){
2759     assert(s->current_picture_ptr->pts != AV_NOPTS_VALUE);
2760     s->time = s->current_picture_ptr->f.pts * s->avctx->time_base.num;
2761
2762     if(s->pict_type==AV_PICTURE_TYPE_B){
2763         s->pb_time= s->pp_time - (s->last_non_b_time - s->time);
2764         assert(s->pb_time > 0 && s->pb_time < s->pp_time);
2765     }else{
2766         s->pp_time= s->time - s->last_non_b_time;
2767         s->last_non_b_time= s->time;
2768         assert(s->picture_number==0 || s->pp_time > 0);
2769     }
2770 }
2771
2772 static int encode_picture(MpegEncContext *s, int picture_number)
2773 {
2774     int i;
2775     int bits;
2776     int context_count = s->avctx->thread_count;
2777
2778     s->picture_number = picture_number;
2779
2780     /* Reset the average MB variance */
2781     s->me.mb_var_sum_temp    =
2782     s->me.mc_mb_var_sum_temp = 0;
2783
2784     /* we need to initialize some time vars before we can encode b-frames */
2785     // RAL: Condition added for MPEG1VIDEO
2786     if (s->codec_id == CODEC_ID_MPEG1VIDEO || s->codec_id == CODEC_ID_MPEG2VIDEO || (s->h263_pred && !s->msmpeg4_version))
2787         set_frame_distances(s);
2788     if(CONFIG_MPEG4_ENCODER && s->codec_id == CODEC_ID_MPEG4)
2789         ff_set_mpeg4_time(s);
2790
2791     s->me.scene_change_score=0;
2792
2793 //    s->lambda= s->current_picture_ptr->quality; //FIXME qscale / ... stuff for ME rate distortion
2794
2795     if(s->pict_type==AV_PICTURE_TYPE_I){
2796         if(s->msmpeg4_version >= 3) s->no_rounding=1;
2797         else                        s->no_rounding=0;
2798     }else if(s->pict_type!=AV_PICTURE_TYPE_B){
2799         if(s->flipflop_rounding || s->codec_id == CODEC_ID_H263P || s->codec_id == CODEC_ID_MPEG4)
2800             s->no_rounding ^= 1;
2801     }
2802
2803     if(s->flags & CODEC_FLAG_PASS2){
2804         if (estimate_qp(s,1) < 0)
2805             return -1;
2806         ff_get_2pass_fcode(s);
2807     }else if(!(s->flags & CODEC_FLAG_QSCALE)){
2808         if(s->pict_type==AV_PICTURE_TYPE_B)
2809             s->lambda= s->last_lambda_for[s->pict_type];
2810         else
2811             s->lambda= s->last_lambda_for[s->last_non_b_pict_type];
2812         update_qscale(s);
2813     }
2814
2815     s->mb_intra=0; //for the rate distortion & bit compare functions
2816     for(i=1; i<context_count; i++){
2817         ff_update_duplicate_context(s->thread_context[i], s);
2818     }
2819
2820     if(ff_init_me(s)<0)
2821         return -1;
2822
2823     /* Estimate motion for every MB */
2824     if(s->pict_type != AV_PICTURE_TYPE_I){
2825         s->lambda = (s->lambda * s->avctx->me_penalty_compensation + 128)>>8;
2826         s->lambda2= (s->lambda2* (int64_t)s->avctx->me_penalty_compensation + 128)>>8;
2827         if(s->pict_type != AV_PICTURE_TYPE_B && s->avctx->me_threshold==0){
2828             if((s->avctx->pre_me && s->last_non_b_pict_type==AV_PICTURE_TYPE_I) || s->avctx->pre_me==2){
2829                 s->avctx->execute(s->avctx, pre_estimate_motion_thread, &s->thread_context[0], NULL, context_count, sizeof(void*));
2830             }
2831         }
2832
2833         s->avctx->execute(s->avctx, estimate_motion_thread, &s->thread_context[0], NULL, context_count, sizeof(void*));
2834     }else /* if(s->pict_type == AV_PICTURE_TYPE_I) */{
2835         /* I-Frame */
2836         for(i=0; i<s->mb_stride*s->mb_height; i++)
2837             s->mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
2838
2839         if(!s->fixed_qscale){
2840             /* finding spatial complexity for I-frame rate control */
2841             s->avctx->execute(s->avctx, mb_var_thread, &s->thread_context[0], NULL, context_count, sizeof(void*));
2842         }
2843     }
2844     for(i=1; i<context_count; i++){
2845         merge_context_after_me(s, s->thread_context[i]);
2846     }
2847     s->current_picture.mc_mb_var_sum= s->current_picture_ptr->mc_mb_var_sum= s->me.mc_mb_var_sum_temp;
2848     s->current_picture.   mb_var_sum= s->current_picture_ptr->   mb_var_sum= s->me.   mb_var_sum_temp;
2849     emms_c();
2850
2851     if(s->me.scene_change_score > s->avctx->scenechange_threshold && s->pict_type == AV_PICTURE_TYPE_P){
2852         s->pict_type= AV_PICTURE_TYPE_I;
2853         for(i=0; i<s->mb_stride*s->mb_height; i++)
2854             s->mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
2855 //printf("Scene change detected, encoding as I Frame %d %d\n", s->current_picture.mb_var_sum, s->current_picture.mc_mb_var_sum);
2856     }
2857
2858     if(!s->umvplus){
2859         if(s->pict_type==AV_PICTURE_TYPE_P || s->pict_type==AV_PICTURE_TYPE_S) {
2860             s->f_code= ff_get_best_fcode(s, s->p_mv_table, CANDIDATE_MB_TYPE_INTER);
2861
2862             if(s->flags & CODEC_FLAG_INTERLACED_ME){
2863                 int a,b;
2864                 a= ff_get_best_fcode(s, s->p_field_mv_table[0][0], CANDIDATE_MB_TYPE_INTER_I); //FIXME field_select
2865                 b= ff_get_best_fcode(s, s->p_field_mv_table[1][1], CANDIDATE_MB_TYPE_INTER_I);
2866                 s->f_code= FFMAX3(s->f_code, a, b);
2867             }
2868
2869             ff_fix_long_p_mvs(s);
2870             ff_fix_long_mvs(s, NULL, 0, s->p_mv_table, s->f_code, CANDIDATE_MB_TYPE_INTER, 0);
2871             if(s->flags & CODEC_FLAG_INTERLACED_ME){
2872                 int j;
2873                 for(i=0; i<2; i++){
2874                     for(j=0; j<2; j++)
2875                         ff_fix_long_mvs(s, s->p_field_select_table[i], j,
2876                                         s->p_field_mv_table[i][j], s->f_code, CANDIDATE_MB_TYPE_INTER_I, 0);
2877                 }
2878             }
2879         }
2880
2881         if(s->pict_type==AV_PICTURE_TYPE_B){
2882             int a, b;
2883
2884             a = ff_get_best_fcode(s, s->b_forw_mv_table, CANDIDATE_MB_TYPE_FORWARD);
2885             b = ff_get_best_fcode(s, s->b_bidir_forw_mv_table, CANDIDATE_MB_TYPE_BIDIR);
2886             s->f_code = FFMAX(a, b);
2887
2888             a = ff_get_best_fcode(s, s->b_back_mv_table, CANDIDATE_MB_TYPE_BACKWARD);
2889             b = ff_get_best_fcode(s, s->b_bidir_back_mv_table, CANDIDATE_MB_TYPE_BIDIR);
2890             s->b_code = FFMAX(a, b);
2891
2892             ff_fix_long_mvs(s, NULL, 0, s->b_forw_mv_table, s->f_code, CANDIDATE_MB_TYPE_FORWARD, 1);
2893             ff_fix_long_mvs(s, NULL, 0, s->b_back_mv_table, s->b_code, CANDIDATE_MB_TYPE_BACKWARD, 1);
2894             ff_fix_long_mvs(s, NULL, 0, s->b_bidir_forw_mv_table, s->f_code, CANDIDATE_MB_TYPE_BIDIR, 1);
2895             ff_fix_long_mvs(s, NULL, 0, s->b_bidir_back_mv_table, s->b_code, CANDIDATE_MB_TYPE_BIDIR, 1);
2896             if(s->flags & CODEC_FLAG_INTERLACED_ME){
2897                 int dir, j;
2898                 for(dir=0; dir<2; dir++){
2899                     for(i=0; i<2; i++){
2900                         for(j=0; j<2; j++){
2901                             int type= dir ? (CANDIDATE_MB_TYPE_BACKWARD_I|CANDIDATE_MB_TYPE_BIDIR_I)
2902                                           : (CANDIDATE_MB_TYPE_FORWARD_I |CANDIDATE_MB_TYPE_BIDIR_I);
2903                             ff_fix_long_mvs(s, s->b_field_select_table[dir][i], j,
2904                                             s->b_field_mv_table[dir][i][j], dir ? s->b_code : s->f_code, type, 1);
2905                         }
2906                     }
2907                 }
2908             }
2909         }
2910     }
2911
2912     if (estimate_qp(s, 0) < 0)
2913         return -1;
2914
2915     if(s->qscale < 3 && s->max_qcoeff<=128 && s->pict_type==AV_PICTURE_TYPE_I && !(s->flags & CODEC_FLAG_QSCALE))
2916         s->qscale= 3; //reduce clipping problems
2917
2918     if (s->out_format == FMT_MJPEG) {
2919         /* for mjpeg, we do include qscale in the matrix */
2920         for(i=1;i<64;i++){
2921             int j= s->dsp.idct_permutation[i];
2922
2923             s->intra_matrix[j] = av_clip_uint8((ff_mpeg1_default_intra_matrix[i] * s->qscale) >> 3);
2924         }
2925         s->y_dc_scale_table=
2926         s->c_dc_scale_table= ff_mpeg2_dc_scale_table[s->intra_dc_precision];
2927         s->intra_matrix[0] = ff_mpeg2_dc_scale_table[s->intra_dc_precision][8];
2928         ff_convert_matrix(&s->dsp, s->q_intra_matrix, s->q_intra_matrix16,
2929                        s->intra_matrix, s->intra_quant_bias, 8, 8, 1);
2930         s->qscale= 8;
2931     }
2932
2933     //FIXME var duplication
2934     s->current_picture_ptr->f.key_frame =
2935     s->current_picture.f.key_frame = s->pict_type == AV_PICTURE_TYPE_I; //FIXME pic_ptr
2936     s->current_picture_ptr->f.pict_type =
2937     s->current_picture.f.pict_type = s->pict_type;
2938
2939     if (s->current_picture.f.key_frame)
2940         s->picture_in_gop_number=0;
2941
2942     s->last_bits= put_bits_count(&s->pb);
2943     switch(s->out_format) {
2944     case FMT_MJPEG:
2945         if (CONFIG_MJPEG_ENCODER)
2946             ff_mjpeg_encode_picture_header(s);
2947         break;
2948     case FMT_H261:
2949         if (CONFIG_H261_ENCODER)
2950             ff_h261_encode_picture_header(s, picture_number);
2951         break;
2952     case FMT_H263:
2953         if (CONFIG_WMV2_ENCODER && s->codec_id == CODEC_ID_WMV2)
2954             ff_wmv2_encode_picture_header(s, picture_number);
2955         else if (CONFIG_MSMPEG4_ENCODER && s->msmpeg4_version)
2956             msmpeg4_encode_picture_header(s, picture_number);
2957         else if (CONFIG_MPEG4_ENCODER && s->h263_pred)
2958             mpeg4_encode_picture_header(s, picture_number);
2959         else if (CONFIG_RV10_ENCODER && s->codec_id == CODEC_ID_RV10)
2960             rv10_encode_picture_header(s, picture_number);
2961         else if (CONFIG_RV20_ENCODER && s->codec_id == CODEC_ID_RV20)
2962             rv20_encode_picture_header(s, picture_number);
2963         else if (CONFIG_FLV_ENCODER && s->codec_id == CODEC_ID_FLV1)
2964             ff_flv_encode_picture_header(s, picture_number);
2965         else if (CONFIG_H263_ENCODER)
2966             h263_encode_picture_header(s, picture_number);
2967         break;
2968     case FMT_MPEG1:
2969         if (CONFIG_MPEG1VIDEO_ENCODER || CONFIG_MPEG2VIDEO_ENCODER)
2970             mpeg1_encode_picture_header(s, picture_number);
2971         break;
2972     case FMT_H264:
2973         break;
2974     default:
2975         assert(0);
2976     }
2977     bits= put_bits_count(&s->pb);
2978     s->header_bits= bits - s->last_bits;
2979
2980     for(i=1; i<context_count; i++){
2981         update_duplicate_context_after_me(s->thread_context[i], s);
2982     }
2983     s->avctx->execute(s->avctx, encode_thread, &s->thread_context[0], NULL, context_count, sizeof(void*));
2984     for(i=1; i<context_count; i++){
2985         merge_context_after_encode(s, s->thread_context[i]);
2986     }
2987     emms_c();
2988     return 0;
2989 }
2990
2991 static void denoise_dct_c(MpegEncContext *s, DCTELEM *block){
2992     const int intra= s->mb_intra;
2993     int i;
2994
2995     s->dct_count[intra]++;
2996
2997     for(i=0; i<64; i++){
2998         int level= block[i];
2999
3000         if(level){
3001             if(level>0){
3002                 s->dct_error_sum[intra][i] += level;
3003                 level -= s->dct_offset[intra][i];
3004                 if(level<0) level=0;
3005             }else{
3006                 s->dct_error_sum[intra][i] -= level;
3007                 level += s->dct_offset[intra][i];
3008                 if(level>0) level=0;
3009             }
3010             block[i]= level;
3011         }
3012     }
3013 }
3014
3015 static int dct_quantize_trellis_c(MpegEncContext *s,
3016                                   DCTELEM *block, int n,
3017                                   int qscale, int *overflow){
3018     const int *qmat;
3019     const uint8_t *scantable= s->intra_scantable.scantable;
3020     const uint8_t *perm_scantable= s->intra_scantable.permutated;
3021     int max=0;
3022     unsigned int threshold1, threshold2;
3023     int bias=0;
3024     int run_tab[65];
3025     int level_tab[65];
3026     int score_tab[65];
3027     int survivor[65];
3028     int survivor_count;
3029     int last_run=0;
3030     int last_level=0;
3031     int last_score= 0;
3032     int last_i;
3033     int coeff[2][64];
3034     int coeff_count[64];
3035     int qmul, qadd, start_i, last_non_zero, i, dc;
3036     const int esc_length= s->ac_esc_length;
3037     uint8_t * length;
3038     uint8_t * last_length;
3039     const int lambda= s->lambda2 >> (FF_LAMBDA_SHIFT - 6);
3040
3041     s->dsp.fdct (block);
3042
3043     if(s->dct_error_sum)
3044         s->denoise_dct(s, block);
3045     qmul= qscale*16;
3046     qadd= ((qscale-1)|1)*8;
3047
3048     if (s->mb_intra) {
3049         int q;
3050         if (!s->h263_aic) {
3051             if (n < 4)
3052                 q = s->y_dc_scale;
3053             else
3054                 q = s->c_dc_scale;
3055             q = q << 3;
3056         } else{
3057             /* For AIC we skip quant/dequant of INTRADC */
3058             q = 1 << 3;
3059             qadd=0;
3060         }
3061
3062         /* note: block[0] is assumed to be positive */
3063         block[0] = (block[0] + (q >> 1)) / q;
3064         start_i = 1;
3065         last_non_zero = 0;
3066         qmat = s->q_intra_matrix[qscale];
3067         if(s->mpeg_quant || s->out_format == FMT_MPEG1)
3068             bias= 1<<(QMAT_SHIFT-1);
3069         length     = s->intra_ac_vlc_length;
3070         last_length= s->intra_ac_vlc_last_length;
3071     } else {
3072         start_i = 0;
3073         last_non_zero = -1;
3074         qmat = s->q_inter_matrix[qscale];
3075         length     = s->inter_ac_vlc_length;
3076         last_length= s->inter_ac_vlc_last_length;
3077     }
3078     last_i= start_i;
3079
3080     threshold1= (1<<QMAT_SHIFT) - bias - 1;
3081     threshold2= (threshold1<<1);
3082
3083     for(i=63; i>=start_i; i--) {
3084         const int j = scantable[i];
3085         int level = block[j] * qmat[j];
3086
3087         if(((unsigned)(level+threshold1))>threshold2){
3088             last_non_zero = i;
3089             break;
3090         }
3091     }
3092
3093     for(i=start_i; i<=last_non_zero; i++) {
3094         const int j = scantable[i];
3095         int level = block[j] * qmat[j];
3096
3097 //        if(   bias+level >= (1<<(QMAT_SHIFT - 3))
3098 //           || bias-level >= (1<<(QMAT_SHIFT - 3))){
3099         if(((unsigned)(level+threshold1))>threshold2){
3100             if(level>0){
3101                 level= (bias + level)>>QMAT_SHIFT;
3102                 coeff[0][i]= level;
3103                 coeff[1][i]= level-1;
3104 //                coeff[2][k]= level-2;
3105             }else{
3106                 level= (bias - level)>>QMAT_SHIFT;
3107                 coeff[0][i]= -level;
3108                 coeff[1][i]= -level+1;
3109 //                coeff[2][k]= -level+2;
3110             }
3111             coeff_count[i]= FFMIN(level, 2);
3112             assert(coeff_count[i]);
3113             max |=level;
3114         }else{
3115             coeff[0][i]= (level>>31)|1;
3116             coeff_count[i]= 1;
3117         }
3118     }
3119
3120     *overflow= s->max_qcoeff < max; //overflow might have happened
3121
3122     if(last_non_zero < start_i){
3123         memset(block + start_i, 0, (64-start_i)*sizeof(DCTELEM));
3124         return last_non_zero;
3125     }
3126
3127     score_tab[start_i]= 0;
3128     survivor[0]= start_i;
3129     survivor_count= 1;
3130
3131     for(i=start_i; i<=last_non_zero; i++){
3132         int level_index, j, zero_distortion;
3133         int dct_coeff= FFABS(block[ scantable[i] ]);
3134         int best_score=256*256*256*120;
3135
3136         if (   s->dsp.fdct == fdct_ifast
3137 #ifndef FAAN_POSTSCALE
3138             || s->dsp.fdct == ff_faandct
3139 #endif
3140            )
3141             dct_coeff= (dct_coeff*ff_inv_aanscales[ scantable[i] ]) >> 12;
3142         zero_distortion= dct_coeff*dct_coeff;
3143
3144         for(level_index=0; level_index < coeff_count[i]; level_index++){
3145             int distortion;
3146             int level= coeff[level_index][i];
3147             const int alevel= FFABS(level);
3148             int unquant_coeff;
3149
3150             assert(level);
3151
3152             if(s->out_format == FMT_H263){
3153                 unquant_coeff= alevel*qmul + qadd;
3154             }else{ //MPEG1
3155                 j= s->dsp.idct_permutation[ scantable[i] ]; //FIXME optimize
3156                 if(s->mb_intra){
3157                         unquant_coeff = (int)(  alevel  * qscale * s->intra_matrix[j]) >> 3;
3158                         unquant_coeff =   (unquant_coeff - 1) | 1;
3159                 }else{
3160                         unquant_coeff = (((  alevel  << 1) + 1) * qscale * ((int) s->inter_matrix[j])) >> 4;
3161                         unquant_coeff =   (unquant_coeff - 1) | 1;
3162                 }
3163                 unquant_coeff<<= 3;
3164             }
3165
3166             distortion= (unquant_coeff - dct_coeff) * (unquant_coeff - dct_coeff) - zero_distortion;
3167             level+=64;
3168             if((level&(~127)) == 0){
3169                 for(j=survivor_count-1; j>=0; j--){
3170                     int run= i - survivor[j];
3171                     int score= distortion + length[UNI_AC_ENC_INDEX(run, level)]*lambda;
3172                     score += score_tab[i-run];
3173
3174                     if(score < best_score){
3175                         best_score= score;
3176                         run_tab[i+1]= run;
3177                         level_tab[i+1]= level-64;
3178                     }
3179                 }
3180
3181                 if(s->out_format == FMT_H263){
3182                     for(j=survivor_count-1; j>=0; j--){
3183                         int run= i - survivor[j];
3184                         int score= distortion + last_length[UNI_AC_ENC_INDEX(run, level)]*lambda;
3185                         score += score_tab[i-run];
3186                         if(score < last_score){
3187                             last_score= score;
3188                             last_run= run;
3189                             last_level= level-64;
3190                             last_i= i+1;
3191                         }
3192                     }
3193                 }
3194             }else{
3195                 distortion += esc_length*lambda;
3196                 for(j=survivor_count-1; j>=0; j--){
3197                     int run= i - survivor[j];
3198                     int score= distortion + score_tab[i-run];
3199
3200                     if(score < best_score){
3201                         best_score= score;
3202                         run_tab[i+1]= run;
3203                         level_tab[i+1]= level-64;
3204                     }
3205                 }
3206
3207                 if(s->out_format == FMT_H263){
3208                   for(j=survivor_count-1; j>=0; j--){
3209                         int run= i - survivor[j];
3210                         int score= distortion + score_tab[i-run];
3211                         if(score < last_score){
3212                             last_score= score;
3213                             last_run= run;
3214                             last_level= level-64;
3215                             last_i= i+1;
3216                         }
3217                     }
3218                 }
3219             }
3220         }
3221
3222         score_tab[i+1]= best_score;
3223
3224         //Note: there is a vlc code in mpeg4 which is 1 bit shorter then another one with a shorter run and the same level
3225         if(last_non_zero <= 27){
3226             for(; survivor_count; survivor_count--){
3227                 if(score_tab[ survivor[survivor_count-1] ] <= best_score)
3228                     break;
3229             }
3230         }else{
3231             for(; survivor_count; survivor_count--){
3232                 if(score_tab[ survivor[survivor_count-1] ] <= best_score + lambda)
3233                     break;
3234             }
3235         }
3236
3237         survivor[ survivor_count++ ]= i+1;
3238     }
3239
3240     if(s->out_format != FMT_H263){
3241         last_score= 256*256*256*120;
3242         for(i= survivor[0]; i<=last_non_zero + 1; i++){
3243             int score= score_tab[i];
3244             if(i) score += lambda*2; //FIXME exacter?
3245
3246             if(score < last_score){
3247                 last_score= score;
3248                 last_i= i;
3249                 last_level= level_tab[i];
3250                 last_run= run_tab[i];
3251             }
3252         }
3253     }
3254
3255     s->coded_score[n] = last_score;
3256
3257     dc= FFABS(block[0]);
3258     last_non_zero= last_i - 1;
3259     memset(block + start_i, 0, (64-start_i)*sizeof(DCTELEM));
3260
3261     if(last_non_zero < start_i)
3262         return last_non_zero;
3263
3264     if(last_non_zero == 0 && start_i == 0){
3265         int best_level= 0;
3266         int best_score= dc * dc;
3267
3268         for(i=0; i<coeff_count[0]; i++){
3269             int level= coeff[i][0];
3270             int alevel= FFABS(level);
3271             int unquant_coeff, score, distortion;
3272
3273             if(s->out_format == FMT_H263){
3274                     unquant_coeff= (alevel*qmul + qadd)>>3;
3275             }else{ //MPEG1
3276                     unquant_coeff = (((  alevel  << 1) + 1) * qscale * ((int) s->inter_matrix[0])) >> 4;
3277                     unquant_coeff =   (unquant_coeff - 1) | 1;
3278             }
3279             unquant_coeff = (unquant_coeff + 4) >> 3;
3280             unquant_coeff<<= 3 + 3;
3281
3282             distortion= (unquant_coeff - dc) * (unquant_coeff - dc);
3283             level+=64;
3284             if((level&(~127)) == 0) score= distortion + last_length[UNI_AC_ENC_INDEX(0, level)]*lambda;
3285             else                    score= distortion + esc_length*lambda;
3286
3287             if(score < best_score){
3288                 best_score= score;
3289                 best_level= level - 64;
3290             }
3291         }
3292         block[0]= best_level;
3293         s->coded_score[n] = best_score - dc*dc;
3294         if(best_level == 0) return -1;
3295         else                return last_non_zero;
3296     }
3297
3298     i= last_i;
3299     assert(last_level);
3300
3301     block[ perm_scantable[last_non_zero] ]= last_level;
3302     i -= last_run + 1;
3303
3304     for(; i>start_i; i -= run_tab[i] + 1){
3305         block[ perm_scantable[i-1] ]= level_tab[i];
3306     }
3307
3308     return last_non_zero;
3309 }
3310
3311 //#define REFINE_STATS 1
3312 static int16_t basis[64][64];
3313
3314 static void build_basis(uint8_t *perm){
3315     int i, j, x, y;
3316     emms_c();
3317     for(i=0; i<8; i++){
3318         for(j=0; j<8; j++){
3319             for(y=0; y<8; y++){
3320                 for(x=0; x<8; x++){
3321                     double s= 0.25*(1<<BASIS_SHIFT);
3322                     int index= 8*i + j;
3323                     int perm_index= perm[index];
3324                     if(i==0) s*= sqrt(0.5);
3325                     if(j==0) s*= sqrt(0.5);
3326                     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)));
3327                 }
3328             }
3329         }
3330     }
3331 }
3332
3333 static int dct_quantize_refine(MpegEncContext *s, //FIXME breaks denoise?
3334                         DCTELEM *block, int16_t *weight, DCTELEM *orig,
3335                         int n, int qscale){
3336     int16_t rem[64];
3337     LOCAL_ALIGNED_16(DCTELEM, d1, [64]);
3338     const uint8_t *scantable= s->intra_scantable.scantable;
3339     const uint8_t *perm_scantable= s->intra_scantable.permutated;
3340 //    unsigned int threshold1, threshold2;
3341 //    int bias=0;
3342     int run_tab[65];
3343     int prev_run=0;
3344     int prev_level=0;
3345     int qmul, qadd, start_i, last_non_zero, i, dc;
3346     uint8_t * length;
3347     uint8_t * last_length;
3348     int lambda;
3349     int rle_index, run, q = 1, sum; //q is only used when s->mb_intra is true
3350 #ifdef REFINE_STATS
3351 static int count=0;
3352 static int after_last=0;
3353 static int to_zero=0;
3354 static int from_zero=0;
3355 static int raise=0;
3356 static int lower=0;
3357 static int messed_sign=0;
3358 #endif
3359
3360     if(basis[0][0] == 0)
3361         build_basis(s->dsp.idct_permutation);
3362
3363     qmul= qscale*2;
3364     qadd= (qscale-1)|1;
3365     if (s->mb_intra) {
3366         if (!s->h263_aic) {
3367             if (n < 4)
3368                 q = s->y_dc_scale;
3369             else
3370                 q = s->c_dc_scale;
3371         } else{
3372             /* For AIC we skip quant/dequant of INTRADC */
3373             q = 1;
3374             qadd=0;
3375         }
3376         q <<= RECON_SHIFT-3;
3377         /* note: block[0] is assumed to be positive */
3378         dc= block[0]*q;
3379 //        block[0] = (block[0] + (q >> 1)) / q;
3380         start_i = 1;
3381 //        if(s->mpeg_quant || s->out_format == FMT_MPEG1)
3382 //            bias= 1<<(QMAT_SHIFT-1);
3383         length     = s->intra_ac_vlc_length;
3384         last_length= s->intra_ac_vlc_last_length;
3385     } else {
3386         dc= 0;
3387         start_i = 0;
3388         length     = s->inter_ac_vlc_length;
3389         last_length= s->inter_ac_vlc_last_length;
3390     }
3391     last_non_zero = s->block_last_index[n];
3392
3393 #ifdef REFINE_STATS
3394 {START_TIMER
3395 #endif
3396     dc += (1<<(RECON_SHIFT-1));
3397     for(i=0; i<64; i++){
3398         rem[i]= dc - (orig[i]<<RECON_SHIFT); //FIXME  use orig dirrectly instead of copying to rem[]
3399     }
3400 #ifdef REFINE_STATS
3401 STOP_TIMER("memset rem[]")}
3402 #endif
3403     sum=0;
3404     for(i=0; i<64; i++){
3405         int one= 36;
3406         int qns=4;
3407         int w;
3408
3409         w= FFABS(weight[i]) + qns*one;
3410         w= 15 + (48*qns*one + w/2)/w; // 16 .. 63
3411
3412         weight[i] = w;
3413 //        w=weight[i] = (63*qns + (w/2)) / w;
3414
3415         assert(w>0);
3416         assert(w<(1<<6));
3417         sum += w*w;
3418     }
3419     lambda= sum*(uint64_t)s->lambda2 >> (FF_LAMBDA_SHIFT - 6 + 6 + 6 + 6);
3420 #ifdef REFINE_STATS
3421 {START_TIMER
3422 #endif
3423     run=0;
3424     rle_index=0;
3425     for(i=start_i; i<=last_non_zero; i++){
3426         int j= perm_scantable[i];
3427         const int level= block[j];
3428         int coeff;
3429
3430         if(level){
3431             if(level<0) coeff= qmul*level - qadd;
3432             else        coeff= qmul*level + qadd;
3433             run_tab[rle_index++]=run;
3434             run=0;
3435
3436             s->dsp.add_8x8basis(rem, basis[j], coeff);
3437         }else{
3438             run++;
3439         }
3440     }
3441 #ifdef REFINE_STATS
3442 if(last_non_zero>0){
3443 STOP_TIMER("init rem[]")
3444 }
3445 }
3446
3447 {START_TIMER
3448 #endif
3449     for(;;){
3450         int best_score=s->dsp.try_8x8basis(rem, weight, basis[0], 0);
3451         int best_coeff=0;
3452         int best_change=0;
3453         int run2, best_unquant_change=0, analyze_gradient;
3454 #ifdef REFINE_STATS
3455 {START_TIMER
3456 #endif
3457         analyze_gradient = last_non_zero > 2 || s->avctx->quantizer_noise_shaping >= 3;
3458
3459         if(analyze_gradient){
3460 #ifdef REFINE_STATS
3461 {START_TIMER
3462 #endif
3463             for(i=0; i<64; i++){
3464                 int w= weight[i];
3465
3466                 d1[i] = (rem[i]*w*w + (1<<(RECON_SHIFT+12-1)))>>(RECON_SHIFT+12);
3467             }
3468 #ifdef REFINE_STATS
3469 STOP_TIMER("rem*w*w")}
3470 {START_TIMER
3471 #endif
3472             s->dsp.fdct(d1);
3473 #ifdef REFINE_STATS
3474 STOP_TIMER("dct")}
3475 #endif
3476         }
3477
3478         if(start_i){
3479             const int level= block[0];
3480             int change, old_coeff;
3481
3482             assert(s->mb_intra);
3483
3484             old_coeff= q*level;
3485
3486             for(change=-1; change<=1; change+=2){
3487                 int new_level= level + change;
3488                 int score, new_coeff;
3489
3490                 new_coeff= q*new_level;
3491                 if(new_coeff >= 2048 || new_coeff < 0)
3492                     continue;
3493
3494                 score= s->dsp.try_8x8basis(rem, weight, basis[0], new_coeff - old_coeff);
3495                 if(score<best_score){
3496                     best_score= score;
3497                     best_coeff= 0;
3498                     best_change= change;
3499                     best_unquant_change= new_coeff - old_coeff;
3500                 }
3501             }
3502         }
3503
3504         run=0;
3505         rle_index=0;
3506         run2= run_tab[rle_index++];
3507         prev_level=0;
3508         prev_run=0;
3509
3510         for(i=start_i; i<64; i++){
3511             int j= perm_scantable[i];
3512             const int level= block[j];
3513             int change, old_coeff;
3514
3515             if(s->avctx->quantizer_noise_shaping < 3 && i > last_non_zero + 1)
3516                 break;
3517
3518             if(level){
3519                 if(level<0) old_coeff= qmul*level - qadd;
3520                 else        old_coeff= qmul*level + qadd;
3521                 run2= run_tab[rle_index++]; //FIXME ! maybe after last
3522             }else{
3523                 old_coeff=0;
3524                 run2--;
3525                 assert(run2>=0 || i >= last_non_zero );
3526             }
3527
3528             for(change=-1; change<=1; change+=2){
3529                 int new_level= level + change;
3530                 int score, new_coeff, unquant_change;
3531
3532                 score=0;
3533                 if(s->avctx->quantizer_noise_shaping < 2 && FFABS(new_level) > FFABS(level))
3534                    continue;
3535
3536                 if(new_level){
3537                     if(new_level<0) new_coeff= qmul*new_level - qadd;
3538                     else            new_coeff= qmul*new_level + qadd;
3539                     if(new_coeff >= 2048 || new_coeff <= -2048)
3540                         continue;
3541                     //FIXME check for overflow
3542
3543                     if(level){
3544                         if(level < 63 && level > -63){
3545                             if(i < last_non_zero)
3546                                 score +=   length[UNI_AC_ENC_INDEX(run, new_level+64)]
3547                                          - length[UNI_AC_ENC_INDEX(run, level+64)];
3548                             else
3549                                 score +=   last_length[UNI_AC_ENC_INDEX(run, new_level+64)]
3550                                          - last_length[UNI_AC_ENC_INDEX(run, level+64)];
3551                         }
3552                     }else{
3553                         assert(FFABS(new_level)==1);
3554
3555                         if(analyze_gradient){
3556                             int g= d1[ scantable[i] ];
3557                             if(g && (g^new_level) >= 0)
3558                                 continue;
3559                         }
3560
3561                         if(i < last_non_zero){
3562                             int next_i= i + run2 + 1;
3563                             int next_level= block[ perm_scantable[next_i] ] + 64;
3564
3565                             if(next_level&(~127))
3566                                 next_level= 0;
3567
3568                             if(next_i < last_non_zero)
3569                                 score +=   length[UNI_AC_ENC_INDEX(run, 65)]
3570                                          + length[UNI_AC_ENC_INDEX(run2, next_level)]
3571                                          - length[UNI_AC_ENC_INDEX(run + run2 + 1, next_level)];
3572                             else
3573                                 score +=  length[UNI_AC_ENC_INDEX(run, 65)]
3574                                         + last_length[UNI_AC_ENC_INDEX(run2, next_level)]
3575                                         - last_length[UNI_AC_ENC_INDEX(run + run2 + 1, next_level)];
3576                         }else{
3577                             score += last_length[UNI_AC_ENC_INDEX(run, 65)];
3578                             if(prev_level){
3579                                 score +=  length[UNI_AC_ENC_INDEX(prev_run, prev_level)]
3580                                         - last_length[UNI_AC_ENC_INDEX(prev_run, prev_level)];
3581                             }
3582                         }
3583                     }
3584                 }else{
3585                     new_coeff=0;
3586                     assert(FFABS(level)==1);
3587
3588                     if(i < last_non_zero){
3589                         int next_i= i + run2 + 1;
3590                         int next_level= block[ perm_scantable[next_i] ] + 64;
3591
3592                         if(next_level&(~127))
3593                             next_level= 0;
3594
3595                         if(next_i < last_non_zero)
3596                             score +=   length[UNI_AC_ENC_INDEX(run + run2 + 1, next_level)]
3597                                      - 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 + run2 + 1, next_level)]
3601                                      - last_length[UNI_AC_ENC_INDEX(run2, next_level)]
3602                                      - length[UNI_AC_ENC_INDEX(run, 65)];
3603                     }else{
3604                         score += -last_length[UNI_AC_ENC_INDEX(run, 65)];
3605                         if(prev_level){
3606                             score +=  last_length[UNI_AC_ENC_INDEX(prev_run, prev_level)]
3607                                     - length[UNI_AC_ENC_INDEX(prev_run, prev_level)];
3608                         }
3609                     }
3610                 }
3611
3612                 score *= lambda;
3613
3614                 unquant_change= new_coeff - old_coeff;
3615                 assert((score < 100*lambda && score > -100*lambda) || lambda==0);
3616
3617                 score+= s->dsp.try_8x8basis(rem, weight, basis[j], unquant_change);
3618                 if(score<best_score){
3619                     best_score= score;
3620                     best_coeff= i;
3621                     best_change= change;
3622                     best_unquant_change= unquant_change;
3623                 }
3624             }
3625             if(level){
3626                 prev_level= level + 64;
3627                 if(prev_level&(~127))
3628                     prev_level= 0;
3629                 prev_run= run;
3630                 run=0;
3631             }else{
3632                 run++;
3633             }
3634         }
3635 #ifdef REFINE_STATS
3636 STOP_TIMER("iterative step")}
3637 #endif
3638
3639         if(best_change){
3640             int j= perm_scantable[ best_coeff ];
3641
3642             block[j] += best_change;
3643
3644             if(best_coeff > last_non_zero){
3645                 last_non_zero= best_coeff;
3646                 assert(block[j]);
3647 #ifdef REFINE_STATS
3648 after_last++;
3649 #endif
3650             }else{
3651 #ifdef REFINE_STATS
3652 if(block[j]){
3653     if(block[j] - best_change){
3654         if(FFABS(block[j]) > FFABS(block[j] - best_change)){
3655             raise++;
3656         }else{
3657             lower++;
3658         }
3659     }else{
3660         from_zero++;
3661     }
3662 }else{
3663     to_zero++;
3664 }
3665 #endif
3666                 for(; last_non_zero>=start_i; last_non_zero--){
3667                     if(block[perm_scantable[last_non_zero]])
3668                         break;
3669                 }
3670             }
3671 #ifdef REFINE_STATS
3672 count++;
3673 if(256*256*256*64 % count == 0){
3674     printf("after_last:%d to_zero:%d from_zero:%d raise:%d lower:%d sign:%d xyp:%d/%d/%d\n", after_last, to_zero, from_zero, raise, lower, messed_sign, s->mb_x, s->mb_y, s->picture_number);
3675 }
3676 #endif
3677             run=0;
3678             rle_index=0;
3679             for(i=start_i; i<=last_non_zero; i++){
3680                 int j= perm_scantable[i];
3681                 const int level= block[j];
3682
3683                  if(level){
3684                      run_tab[rle_index++]=run;
3685                      run=0;
3686                  }else{
3687                      run++;
3688                  }
3689             }
3690
3691             s->dsp.add_8x8basis(rem, basis[j], best_unquant_change);
3692         }else{
3693             break;
3694         }
3695     }
3696 #ifdef REFINE_STATS
3697 if(last_non_zero>0){
3698 STOP_TIMER("iterative search")
3699 }
3700 }
3701 #endif
3702
3703     return last_non_zero;
3704 }
3705
3706 int dct_quantize_c(MpegEncContext *s,
3707                         DCTELEM *block, int n,
3708                         int qscale, int *overflow)
3709 {
3710     int i, j, level, last_non_zero, q, start_i;
3711     const int *qmat;
3712     const uint8_t *scantable= s->intra_scantable.scantable;
3713     int bias;
3714     int max=0;
3715     unsigned int threshold1, threshold2;
3716
3717     s->dsp.fdct (block);
3718
3719     if(s->dct_error_sum)
3720         s->denoise_dct(s, block);
3721
3722     if (s->mb_intra) {
3723         if (!s->h263_aic) {
3724             if (n < 4)
3725                 q = s->y_dc_scale;
3726             else
3727                 q = s->c_dc_scale;
3728             q = q << 3;
3729         } else
3730             /* For AIC we skip quant/dequant of INTRADC */
3731             q = 1 << 3;
3732
3733         /* note: block[0] is assumed to be positive */
3734         block[0] = (block[0] + (q >> 1)) / q;
3735         start_i = 1;
3736         last_non_zero = 0;
3737         qmat = s->q_intra_matrix[qscale];
3738         bias= s->intra_quant_bias<<(QMAT_SHIFT - QUANT_BIAS_SHIFT);
3739     } else {
3740         start_i = 0;
3741         last_non_zero = -1;
3742         qmat = s->q_inter_matrix[qscale];
3743         bias= s->inter_quant_bias<<(QMAT_SHIFT - QUANT_BIAS_SHIFT);
3744     }
3745     threshold1= (1<<QMAT_SHIFT) - bias - 1;
3746     threshold2= (threshold1<<1);
3747     for(i=63;i>=start_i;i--) {
3748         j = scantable[i];
3749         level = block[j] * qmat[j];
3750
3751         if(((unsigned)(level+threshold1))>threshold2){
3752             last_non_zero = i;
3753             break;
3754         }else{
3755             block[j]=0;
3756         }
3757     }
3758     for(i=start_i; i<=last_non_zero; i++) {
3759         j = scantable[i];
3760         level = block[j] * qmat[j];
3761
3762 //        if(   bias+level >= (1<<QMAT_SHIFT)
3763 //           || bias-level >= (1<<QMAT_SHIFT)){
3764         if(((unsigned)(level+threshold1))>threshold2){
3765             if(level>0){
3766                 level= (bias + level)>>QMAT_SHIFT;
3767                 block[j]= level;
3768             }else{
3769                 level= (bias - level)>>QMAT_SHIFT;
3770                 block[j]= -level;
3771             }
3772             max |=level;
3773         }else{
3774             block[j]=0;
3775         }
3776     }
3777     *overflow= s->max_qcoeff < max; //overflow might have happened
3778
3779     /* we need this permutation so that we correct the IDCT, we only permute the !=0 elements */
3780     if (s->dsp.idct_permutation_type != FF_NO_IDCT_PERM)
3781         ff_block_permute(block, s->dsp.idct_permutation, scantable, last_non_zero);
3782
3783     return last_non_zero;
3784 }
3785
3786 #define OFFSET(x) offsetof(MpegEncContext, x)
3787 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
3788 static const AVOption h263_options[] = {
3789     { "obmc",         "use overlapped block motion compensation.", OFFSET(obmc), FF_OPT_TYPE_INT, { 0 }, 0, 1, VE },
3790     { NULL },
3791 };
3792
3793 static const AVClass h263_class = {
3794     .class_name = "H.263 encoder",
3795     .item_name  = av_default_item_name,
3796     .option     = h263_options,
3797     .version    = LIBAVUTIL_VERSION_INT,
3798 };
3799
3800 AVCodec ff_h263_encoder = {
3801     .name           = "h263",
3802     .type           = AVMEDIA_TYPE_VIDEO,
3803     .id             = CODEC_ID_H263,
3804     .priv_data_size = sizeof(MpegEncContext),
3805     .init           = MPV_encode_init,
3806     .encode         = MPV_encode_picture,
3807     .close          = MPV_encode_end,
3808     .pix_fmts= (const enum PixelFormat[]){PIX_FMT_YUV420P, PIX_FMT_NONE},
3809     .long_name= NULL_IF_CONFIG_SMALL("H.263 / H.263-1996"),
3810     .priv_class     = &h263_class,
3811 };
3812
3813 static const AVOption h263p_options[] = {
3814     { "umv",        "Use unlimited motion vectors.",    OFFSET(umvplus), FF_OPT_TYPE_INT, { 0 }, 0, 1, VE },
3815     { "aiv",        "Use alternative inter VLC.",       OFFSET(alt_inter_vlc), FF_OPT_TYPE_INT, { 0 }, 0, 1, VE },
3816     { "obmc",       "use overlapped block motion compensation.", OFFSET(obmc), FF_OPT_TYPE_INT, { 0 }, 0, 1, VE },
3817     { NULL },
3818 };
3819 static const AVClass h263p_class = {
3820     .class_name = "H.263p encoder",
3821     .item_name  = av_default_item_name,
3822     .option     = h263p_options,
3823     .version    = LIBAVUTIL_VERSION_INT,
3824 };
3825
3826 AVCodec ff_h263p_encoder = {
3827     .name           = "h263p",
3828     .type           = AVMEDIA_TYPE_VIDEO,
3829     .id             = CODEC_ID_H263P,
3830     .priv_data_size = sizeof(MpegEncContext),
3831     .init           = MPV_encode_init,
3832     .encode         = MPV_encode_picture,
3833     .close          = MPV_encode_end,
3834     .capabilities = CODEC_CAP_SLICE_THREADS,
3835     .pix_fmts= (const enum PixelFormat[]){PIX_FMT_YUV420P, PIX_FMT_NONE},
3836     .long_name= NULL_IF_CONFIG_SMALL("H.263+ / H.263-1998 / H.263 version 2"),
3837     .priv_class     = &h263p_class,
3838 };
3839
3840 AVCodec ff_msmpeg4v2_encoder = {
3841     .name           = "msmpeg4v2",
3842     .type           = AVMEDIA_TYPE_VIDEO,
3843     .id             = CODEC_ID_MSMPEG4V2,
3844     .priv_data_size = sizeof(MpegEncContext),
3845     .init           = MPV_encode_init,
3846     .encode         = MPV_encode_picture,
3847     .close          = MPV_encode_end,
3848     .pix_fmts= (const enum PixelFormat[]){PIX_FMT_YUV420P, PIX_FMT_NONE},
3849     .long_name= NULL_IF_CONFIG_SMALL("MPEG-4 part 2 Microsoft variant version 2"),
3850 };
3851
3852 AVCodec ff_msmpeg4v3_encoder = {
3853     .name           = "msmpeg4",
3854     .type           = AVMEDIA_TYPE_VIDEO,
3855     .id             = CODEC_ID_MSMPEG4V3,
3856     .priv_data_size = sizeof(MpegEncContext),
3857     .init           = MPV_encode_init,
3858     .encode         = MPV_encode_picture,
3859     .close          = MPV_encode_end,
3860     .pix_fmts= (const enum PixelFormat[]){PIX_FMT_YUV420P, PIX_FMT_NONE},
3861     .long_name= NULL_IF_CONFIG_SMALL("MPEG-4 part 2 Microsoft variant version 3"),
3862 };
3863
3864 AVCodec ff_wmv1_encoder = {
3865     .name           = "wmv1",
3866     .type           = AVMEDIA_TYPE_VIDEO,
3867     .id             = CODEC_ID_WMV1,
3868     .priv_data_size = sizeof(MpegEncContext),
3869     .init           = MPV_encode_init,
3870     .encode         = MPV_encode_picture,
3871     .close          = MPV_encode_end,
3872     .pix_fmts= (const enum PixelFormat[]){PIX_FMT_YUV420P, PIX_FMT_NONE},
3873     .long_name= NULL_IF_CONFIG_SMALL("Windows Media Video 7"),
3874 };