Merge commit 'aa45b90804ab21175b8c116bd8e5eb4b4e85fbcb' into release/0.8
[ffmpeg:ffmpeg.git] / libavcodec / svq1enc.c
1 /*
2  * SVQ1 Encoder
3  * Copyright (C) 2004 Mike Melanson <melanson@pcisys.net>
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21
22 /**
23  * @file
24  * Sorenson Vector Quantizer #1 (SVQ1) video codec.
25  * For more information of the SVQ1 algorithm, visit:
26  *   http://www.pcisys.net/~melanson/codecs/
27  */
28
29
30 #include "avcodec.h"
31 #include "dsputil.h"
32 #include "mpegvideo.h"
33 #include "h263.h"
34 #include "internal.h"
35
36 #include "svq1.h"
37 #include "svq1enc_cb.h"
38
39 #undef NDEBUG
40 #include <assert.h>
41
42
43 typedef struct SVQ1Context {
44     MpegEncContext m; // needed for motion estimation, should not be used for anything else, the idea is to make the motion estimation eventually independent of MpegEncContext, so this will be removed then (FIXME/XXX)
45     AVCodecContext *avctx;
46     DSPContext dsp;
47     AVFrame picture;
48     AVFrame current_picture;
49     AVFrame last_picture;
50     PutBitContext pb;
51     GetBitContext gb;
52
53     PutBitContext reorder_pb[6]; //why ooh why this sick breadth first order, everything is slower and more complex
54
55     int frame_width;
56     int frame_height;
57
58     /* Y plane block dimensions */
59     int y_block_width;
60     int y_block_height;
61
62     /* U & V plane (C planes) block dimensions */
63     int c_block_width;
64     int c_block_height;
65
66     uint16_t *mb_type;
67     uint32_t *dummy;
68     int16_t (*motion_val8[3])[2];
69     int16_t (*motion_val16[3])[2];
70
71     int64_t rd_total;
72
73     uint8_t *scratchbuf;
74 } SVQ1Context;
75
76 static void svq1_write_header(SVQ1Context *s, int frame_type)
77 {
78     int i;
79
80     /* frame code */
81     put_bits(&s->pb, 22, 0x20);
82
83     /* temporal reference (sure hope this is a "don't care") */
84     put_bits(&s->pb, 8, 0x00);
85
86     /* frame type */
87     put_bits(&s->pb, 2, frame_type - 1);
88
89     if (frame_type == AV_PICTURE_TYPE_I) {
90
91         /* no checksum since frame code is 0x20 */
92
93         /* no embedded string either */
94
95         /* output 5 unknown bits (2 + 2 + 1) */
96         put_bits(&s->pb, 5, 2); /* 2 needed by quicktime decoder */
97
98         i= ff_match_2uint16(ff_svq1_frame_size_table, FF_ARRAY_ELEMS(ff_svq1_frame_size_table), s->frame_width, s->frame_height);
99         put_bits(&s->pb, 3, i);
100
101         if (i == 7)
102         {
103                 put_bits(&s->pb, 12, s->frame_width);
104                 put_bits(&s->pb, 12, s->frame_height);
105         }
106     }
107
108     /* no checksum or extra data (next 2 bits get 0) */
109     put_bits(&s->pb, 2, 0);
110 }
111
112
113 #define QUALITY_THRESHOLD 100
114 #define THRESHOLD_MULTIPLIER 0.6
115
116 #if HAVE_ALTIVEC
117 #undef vector
118 #endif
119
120 static int encode_block(SVQ1Context *s, uint8_t *src, uint8_t *ref, uint8_t *decoded, int stride, int level, int threshold, int lambda, int intra){
121     int count, y, x, i, j, split, best_mean, best_score, best_count;
122     int best_vector[6];
123     int block_sum[7]= {0, 0, 0, 0, 0, 0};
124     int w= 2<<((level+2)>>1);
125     int h= 2<<((level+1)>>1);
126     int size=w*h;
127     int16_t block[7][256];
128     const int8_t *codebook_sum, *codebook;
129     const uint16_t (*mean_vlc)[2];
130     const uint8_t (*multistage_vlc)[2];
131
132     best_score=0;
133     //FIXME optimize, this doenst need to be done multiple times
134     if(intra){
135         codebook_sum= svq1_intra_codebook_sum[level];
136         codebook= ff_svq1_intra_codebooks[level];
137         mean_vlc= ff_svq1_intra_mean_vlc;
138         multistage_vlc= ff_svq1_intra_multistage_vlc[level];
139         for(y=0; y<h; y++){
140             for(x=0; x<w; x++){
141                 int v= src[x + y*stride];
142                 block[0][x + w*y]= v;
143                 best_score += v*v;
144                 block_sum[0] += v;
145             }
146         }
147     }else{
148         codebook_sum= svq1_inter_codebook_sum[level];
149         codebook= ff_svq1_inter_codebooks[level];
150         mean_vlc= ff_svq1_inter_mean_vlc + 256;
151         multistage_vlc= ff_svq1_inter_multistage_vlc[level];
152         for(y=0; y<h; y++){
153             for(x=0; x<w; x++){
154                 int v= src[x + y*stride] - ref[x + y*stride];
155                 block[0][x + w*y]= v;
156                 best_score += v*v;
157                 block_sum[0] += v;
158             }
159         }
160     }
161
162     best_count=0;
163     best_score -= ((block_sum[0]*block_sum[0])>>(level+3));
164     best_mean= (block_sum[0] + (size>>1)) >> (level+3);
165
166     if(level<4){
167         for(count=1; count<7; count++){
168             int best_vector_score= INT_MAX;
169             int best_vector_sum=-999, best_vector_mean=-999;
170             const int stage= count-1;
171             const int8_t *vector;
172
173             for(i=0; i<16; i++){
174                 int sum= codebook_sum[stage*16 + i];
175                 int sqr, diff, score;
176
177                 vector = codebook + stage*size*16 + i*size;
178                 sqr = s->dsp.ssd_int8_vs_int16(vector, block[stage], size);
179                 diff= block_sum[stage] - sum;
180                 score= sqr - ((diff*(int64_t)diff)>>(level+3)); //FIXME 64bit slooow
181                 if(score < best_vector_score){
182                     int mean= (diff + (size>>1)) >> (level+3);
183                     assert(mean >-300 && mean<300);
184                     mean= av_clip(mean, intra?0:-256, 255);
185                     best_vector_score= score;
186                     best_vector[stage]= i;
187                     best_vector_sum= sum;
188                     best_vector_mean= mean;
189                 }
190             }
191             assert(best_vector_mean != -999);
192             vector= codebook + stage*size*16 + best_vector[stage]*size;
193             for(j=0; j<size; j++){
194                 block[stage+1][j] = block[stage][j] - vector[j];
195             }
196             block_sum[stage+1]= block_sum[stage] - best_vector_sum;
197             best_vector_score +=
198                 lambda*(+ 1 + 4*count
199                         + multistage_vlc[1+count][1]
200                         + mean_vlc[best_vector_mean][1]);
201
202             if(best_vector_score < best_score){
203                 best_score= best_vector_score;
204                 best_count= count;
205                 best_mean= best_vector_mean;
206             }
207         }
208     }
209
210     split=0;
211     if(best_score > threshold && level){
212         int score=0;
213         int offset= (level&1) ? stride*h/2 : w/2;
214         PutBitContext backup[6];
215
216         for(i=level-1; i>=0; i--){
217             backup[i]= s->reorder_pb[i];
218         }
219         score += encode_block(s, src         , ref         , decoded         , stride, level-1, threshold>>1, lambda, intra);
220         score += encode_block(s, src + offset, ref + offset, decoded + offset, stride, level-1, threshold>>1, lambda, intra);
221         score += lambda;
222
223         if(score < best_score){
224             best_score= score;
225             split=1;
226         }else{
227             for(i=level-1; i>=0; i--){
228                 s->reorder_pb[i]= backup[i];
229             }
230         }
231     }
232     if (level > 0)
233         put_bits(&s->reorder_pb[level], 1, split);
234
235     if(!split){
236         assert((best_mean >= 0 && best_mean<256) || !intra);
237         assert(best_mean >= -256 && best_mean<256);
238         assert(best_count >=0 && best_count<7);
239         assert(level<4 || best_count==0);
240
241         /* output the encoding */
242         put_bits(&s->reorder_pb[level],
243             multistage_vlc[1 + best_count][1],
244             multistage_vlc[1 + best_count][0]);
245         put_bits(&s->reorder_pb[level], mean_vlc[best_mean][1],
246             mean_vlc[best_mean][0]);
247
248         for (i = 0; i < best_count; i++){
249             assert(best_vector[i]>=0 && best_vector[i]<16);
250             put_bits(&s->reorder_pb[level], 4, best_vector[i]);
251         }
252
253         for(y=0; y<h; y++){
254             for(x=0; x<w; x++){
255                 decoded[x + y*stride]= src[x + y*stride] - block[best_count][x + w*y] + best_mean;
256             }
257         }
258     }
259
260     return best_score;
261 }
262
263
264 static int svq1_encode_plane(SVQ1Context *s, int plane, unsigned char *src_plane, unsigned char *ref_plane, unsigned char *decoded_plane,
265     int width, int height, int src_stride, int stride)
266 {
267     int x, y;
268     int i;
269     int block_width, block_height;
270     int level;
271     int threshold[6];
272     uint8_t *src = s->scratchbuf + stride * 16;
273     const int lambda= (s->picture.quality*s->picture.quality) >> (2*FF_LAMBDA_SHIFT);
274
275     /* figure out the acceptable level thresholds in advance */
276     threshold[5] = QUALITY_THRESHOLD;
277     for (level = 4; level >= 0; level--)
278         threshold[level] = threshold[level + 1] * THRESHOLD_MULTIPLIER;
279
280     block_width = (width + 15) / 16;
281     block_height = (height + 15) / 16;
282
283     if(s->picture.pict_type == AV_PICTURE_TYPE_P){
284         s->m.avctx= s->avctx;
285         s->m.current_picture_ptr= &s->m.current_picture;
286         s->m.last_picture_ptr   = &s->m.last_picture;
287         s->m.last_picture.data[0]= ref_plane;
288         s->m.linesize=
289         s->m.last_picture.linesize[0]=
290         s->m.new_picture.linesize[0]=
291         s->m.current_picture.linesize[0]= stride;
292         s->m.width= width;
293         s->m.height= height;
294         s->m.mb_width= block_width;
295         s->m.mb_height= block_height;
296         s->m.mb_stride= s->m.mb_width+1;
297         s->m.b8_stride= 2*s->m.mb_width+1;
298         s->m.f_code=1;
299         s->m.pict_type= s->picture.pict_type;
300         s->m.me_method= s->avctx->me_method;
301         s->m.me.scene_change_score=0;
302         s->m.flags= s->avctx->flags;
303 //        s->m.out_format = FMT_H263;
304 //        s->m.unrestricted_mv= 1;
305
306         s->m.lambda= s->picture.quality;
307         s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
308         s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
309
310         if(!s->motion_val8[plane]){
311             s->motion_val8 [plane]= av_mallocz((s->m.b8_stride*block_height*2 + 2)*2*sizeof(int16_t));
312             s->motion_val16[plane]= av_mallocz((s->m.mb_stride*(block_height + 2) + 1)*2*sizeof(int16_t));
313         }
314
315         s->m.mb_type= s->mb_type;
316
317         //dummies, to avoid segfaults
318         s->m.current_picture.mb_mean=   (uint8_t *)s->dummy;
319         s->m.current_picture.mb_var=    (uint16_t*)s->dummy;
320         s->m.current_picture.mc_mb_var= (uint16_t*)s->dummy;
321         s->m.current_picture.mb_type= s->dummy;
322
323         s->m.current_picture.motion_val[0]= s->motion_val8[plane] + 2;
324         s->m.p_mv_table= s->motion_val16[plane] + s->m.mb_stride + 1;
325         s->m.dsp= s->dsp; //move
326         ff_init_me(&s->m);
327
328         s->m.me.dia_size= s->avctx->dia_size;
329         s->m.first_slice_line=1;
330         for (y = 0; y < block_height; y++) {
331             s->m.new_picture.data[0]= src - y*16*stride; //ugly
332             s->m.mb_y= y;
333
334             for(i=0; i<16 && i + 16*y<height; i++){
335                 memcpy(&src[i*stride], &src_plane[(i+16*y)*src_stride], width);
336                 for(x=width; x<16*block_width; x++)
337                     src[i*stride+x]= src[i*stride+x-1];
338             }
339             for(; i<16 && i + 16*y<16*block_height; i++)
340                 memcpy(&src[i*stride], &src[(i-1)*stride], 16*block_width);
341
342             for (x = 0; x < block_width; x++) {
343                 s->m.mb_x= x;
344                 ff_init_block_index(&s->m);
345                 ff_update_block_index(&s->m);
346
347                 ff_estimate_p_frame_motion(&s->m, x, y);
348             }
349             s->m.first_slice_line=0;
350         }
351
352         ff_fix_long_p_mvs(&s->m);
353         ff_fix_long_mvs(&s->m, NULL, 0, s->m.p_mv_table, s->m.f_code, CANDIDATE_MB_TYPE_INTER, 0);
354     }
355
356     s->m.first_slice_line=1;
357     for (y = 0; y < block_height; y++) {
358         for(i=0; i<16 && i + 16*y<height; i++){
359             memcpy(&src[i*stride], &src_plane[(i+16*y)*src_stride], width);
360             for(x=width; x<16*block_width; x++)
361                 src[i*stride+x]= src[i*stride+x-1];
362         }
363         for(; i<16 && i + 16*y<16*block_height; i++)
364             memcpy(&src[i*stride], &src[(i-1)*stride], 16*block_width);
365
366         s->m.mb_y= y;
367         for (x = 0; x < block_width; x++) {
368             uint8_t reorder_buffer[3][6][7*32];
369             int count[3][6];
370             int offset = y * 16 * stride + x * 16;
371             uint8_t *decoded= decoded_plane + offset;
372             uint8_t *ref= ref_plane + offset;
373             int score[4]={0,0,0,0}, best;
374             uint8_t *temp = s->scratchbuf;
375
376             if(s->pb.buf_end - s->pb.buf - (put_bits_count(&s->pb)>>3) < 3000){ //FIXME check size
377                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
378                 return -1;
379             }
380
381             s->m.mb_x= x;
382             ff_init_block_index(&s->m);
383             ff_update_block_index(&s->m);
384
385             if(s->picture.pict_type == AV_PICTURE_TYPE_I || (s->m.mb_type[x + y*s->m.mb_stride]&CANDIDATE_MB_TYPE_INTRA)){
386                 for(i=0; i<6; i++){
387                     init_put_bits(&s->reorder_pb[i], reorder_buffer[0][i], 7*32);
388                 }
389                 if(s->picture.pict_type == AV_PICTURE_TYPE_P){
390                     const uint8_t *vlc= ff_svq1_block_type_vlc[SVQ1_BLOCK_INTRA];
391                     put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
392                     score[0]= vlc[1]*lambda;
393                 }
394                 score[0]+= encode_block(s, src+16*x, NULL, temp, stride, 5, 64, lambda, 1);
395                 for(i=0; i<6; i++){
396                     count[0][i]= put_bits_count(&s->reorder_pb[i]);
397                     flush_put_bits(&s->reorder_pb[i]);
398                 }
399             }else
400                 score[0]= INT_MAX;
401
402             best=0;
403
404             if(s->picture.pict_type == AV_PICTURE_TYPE_P){
405                 const uint8_t *vlc= ff_svq1_block_type_vlc[SVQ1_BLOCK_INTER];
406                 int mx, my, pred_x, pred_y, dxy;
407                 int16_t *motion_ptr;
408
409                 motion_ptr= ff_h263_pred_motion(&s->m, 0, 0, &pred_x, &pred_y);
410                 if(s->m.mb_type[x + y*s->m.mb_stride]&CANDIDATE_MB_TYPE_INTER){
411                     for(i=0; i<6; i++)
412                         init_put_bits(&s->reorder_pb[i], reorder_buffer[1][i], 7*32);
413
414                     put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
415
416                     s->m.pb= s->reorder_pb[5];
417                     mx= motion_ptr[0];
418                     my= motion_ptr[1];
419                     assert(mx>=-32 && mx<=31);
420                     assert(my>=-32 && my<=31);
421                     assert(pred_x>=-32 && pred_x<=31);
422                     assert(pred_y>=-32 && pred_y<=31);
423                     ff_h263_encode_motion(&s->m, mx - pred_x, 1);
424                     ff_h263_encode_motion(&s->m, my - pred_y, 1);
425                     s->reorder_pb[5]= s->m.pb;
426                     score[1] += lambda*put_bits_count(&s->reorder_pb[5]);
427
428                     dxy= (mx&1) + 2*(my&1);
429
430                     s->dsp.put_pixels_tab[0][dxy](temp+16, ref + (mx>>1) + stride*(my>>1), stride, 16);
431
432                     score[1]+= encode_block(s, src+16*x, temp+16, decoded, stride, 5, 64, lambda, 0);
433                     best= score[1] <= score[0];
434
435                     vlc= ff_svq1_block_type_vlc[SVQ1_BLOCK_SKIP];
436                     score[2]= s->dsp.sse[0](NULL, src+16*x, ref, stride, 16);
437                     score[2]+= vlc[1]*lambda;
438                     if(score[2] < score[best] && mx==0 && my==0){
439                         best=2;
440                         s->dsp.put_pixels_tab[0][0](decoded, ref, stride, 16);
441                         for(i=0; i<6; i++){
442                             count[2][i]=0;
443                         }
444                         put_bits(&s->pb, vlc[1], vlc[0]);
445                     }
446                 }
447
448                 if(best==1){
449                     for(i=0; i<6; i++){
450                         count[1][i]= put_bits_count(&s->reorder_pb[i]);
451                         flush_put_bits(&s->reorder_pb[i]);
452                     }
453                 }else{
454                     motion_ptr[0                 ] = motion_ptr[1                 ]=
455                     motion_ptr[2                 ] = motion_ptr[3                 ]=
456                     motion_ptr[0+2*s->m.b8_stride] = motion_ptr[1+2*s->m.b8_stride]=
457                     motion_ptr[2+2*s->m.b8_stride] = motion_ptr[3+2*s->m.b8_stride]=0;
458                 }
459             }
460
461             s->rd_total += score[best];
462
463             for(i=5; i>=0; i--){
464                 ff_copy_bits(&s->pb, reorder_buffer[best][i], count[best][i]);
465             }
466             if(best==0){
467                 s->dsp.put_pixels_tab[0][0](decoded, temp, stride, 16);
468             }
469         }
470         s->m.first_slice_line=0;
471     }
472     return 0;
473 }
474
475 static av_cold int svq1_encode_init(AVCodecContext *avctx)
476 {
477     SVQ1Context * const s = avctx->priv_data;
478
479     dsputil_init(&s->dsp, avctx);
480     avctx->coded_frame= (AVFrame*)&s->picture;
481
482     s->frame_width = avctx->width;
483     s->frame_height = avctx->height;
484
485     s->y_block_width = (s->frame_width + 15) / 16;
486     s->y_block_height = (s->frame_height + 15) / 16;
487
488     s->c_block_width = (s->frame_width / 4 + 15) / 16;
489     s->c_block_height = (s->frame_height / 4 + 15) / 16;
490
491     s->avctx= avctx;
492     s->m.avctx= avctx;
493     s->m.me.temp      =
494     s->m.me.scratchpad= av_mallocz((avctx->width+64)*2*16*2*sizeof(uint8_t));
495     s->m.me.map       = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
496     s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
497     s->mb_type        = av_mallocz((s->y_block_width+1)*s->y_block_height*sizeof(int16_t));
498     s->dummy          = av_mallocz((s->y_block_width+1)*s->y_block_height*sizeof(int32_t));
499     ff_h263_encode_init(&s->m); //mv_penalty
500
501     return 0;
502 }
503
504 static int svq1_encode_frame(AVCodecContext *avctx, unsigned char *buf,
505     int buf_size, void *data)
506 {
507     SVQ1Context * const s = avctx->priv_data;
508     AVFrame *pict = data;
509     AVFrame * const p= (AVFrame*)&s->picture;
510     AVFrame temp;
511     int i;
512
513     if(avctx->pix_fmt != PIX_FMT_YUV410P){
514         av_log(avctx, AV_LOG_ERROR, "unsupported pixel format\n");
515         return -1;
516     }
517
518     if(!s->current_picture.data[0]){
519         avctx->get_buffer(avctx, &s->current_picture);
520         avctx->get_buffer(avctx, &s->last_picture);
521         s->scratchbuf = av_malloc(s->current_picture.linesize[0] * 16 * 2);
522     }
523
524     temp= s->current_picture;
525     s->current_picture= s->last_picture;
526     s->last_picture= temp;
527
528     init_put_bits(&s->pb, buf, buf_size);
529
530     *p = *pict;
531     p->pict_type = avctx->gop_size && avctx->frame_number % avctx->gop_size ? AV_PICTURE_TYPE_P : AV_PICTURE_TYPE_I;
532     p->key_frame = p->pict_type == AV_PICTURE_TYPE_I;
533
534     svq1_write_header(s, p->pict_type);
535     for(i=0; i<3; i++){
536         if(svq1_encode_plane(s, i,
537             s->picture.data[i], s->last_picture.data[i], s->current_picture.data[i],
538             s->frame_width / (i?4:1), s->frame_height / (i?4:1),
539             s->picture.linesize[i], s->current_picture.linesize[i]) < 0)
540                 return -1;
541     }
542
543 //    align_put_bits(&s->pb);
544     while(put_bits_count(&s->pb) & 31)
545         put_bits(&s->pb, 1, 0);
546
547     flush_put_bits(&s->pb);
548
549     return put_bits_count(&s->pb) / 8;
550 }
551
552 static av_cold int svq1_encode_end(AVCodecContext *avctx)
553 {
554     SVQ1Context * const s = avctx->priv_data;
555     int i;
556
557     av_log(avctx, AV_LOG_DEBUG, "RD: %f\n", s->rd_total/(double)(avctx->width*avctx->height*avctx->frame_number));
558
559     av_freep(&s->m.me.scratchpad);
560     av_freep(&s->m.me.map);
561     av_freep(&s->m.me.score_map);
562     av_freep(&s->mb_type);
563     av_freep(&s->dummy);
564     av_freep(&s->scratchbuf);
565
566     for(i=0; i<3; i++){
567         av_freep(&s->motion_val8[i]);
568         av_freep(&s->motion_val16[i]);
569     }
570
571     return 0;
572 }
573
574
575 AVCodec ff_svq1_encoder = {
576     "svq1",
577     AVMEDIA_TYPE_VIDEO,
578     CODEC_ID_SVQ1,
579     sizeof(SVQ1Context),
580     svq1_encode_init,
581     svq1_encode_frame,
582     svq1_encode_end,
583     .pix_fmts= (const enum PixelFormat[]){PIX_FMT_YUV410P, PIX_FMT_NONE},
584     .long_name= NULL_IF_CONFIG_SMALL("Sorenson Vector Quantizer 1 / Sorenson Video 1 / SVQ1"),
585 };