]> git.sesse.net Git - ffmpeg/blob - libavcodec/svq1enc.c
Add a function to match a 2 element vector of uint16_t and use it in h263 and svq1
[ffmpeg] / 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 libavcodec/svq1enc.c
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
35 #include "svq1.h"
36 #include "svq1enc_cb.h"
37
38 #undef NDEBUG
39 #include <assert.h>
40
41
42 typedef struct SVQ1Context {
43     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)
44     AVCodecContext *avctx;
45     DSPContext dsp;
46     AVFrame picture;
47     AVFrame current_picture;
48     AVFrame last_picture;
49     PutBitContext pb;
50     GetBitContext gb;
51
52     PutBitContext reorder_pb[6]; //why ooh why this sick breadth first order, everything is slower and more complex
53
54     int frame_width;
55     int frame_height;
56
57     /* Y plane block dimensions */
58     int y_block_width;
59     int y_block_height;
60
61     /* U & V plane (C planes) block dimensions */
62     int c_block_width;
63     int c_block_height;
64
65     uint16_t *mb_type;
66     uint32_t *dummy;
67     int16_t (*motion_val8[3])[2];
68     int16_t (*motion_val16[3])[2];
69
70     int64_t rd_total;
71
72     uint8_t *scratchbuf;
73 } SVQ1Context;
74
75 static void svq1_write_header(SVQ1Context *s, int frame_type)
76 {
77     int i;
78
79     /* frame code */
80     put_bits(&s->pb, 22, 0x20);
81
82     /* temporal reference (sure hope this is a "don't care") */
83     put_bits(&s->pb, 8, 0x00);
84
85     /* frame type */
86     put_bits(&s->pb, 2, frame_type - 1);
87
88     if (frame_type == FF_I_TYPE) {
89
90         /* no checksum since frame code is 0x20 */
91
92         /* no embedded string either */
93
94         /* output 5 unknown bits (2 + 2 + 1) */
95         put_bits(&s->pb, 5, 2); /* 2 needed by quicktime decoder */
96
97         i= ff_match_2uint16(ff_svq1_frame_size_table, FF_ARRAY_ELEMS(ff_svq1_frame_size_table), s->frame_width, s->frame_height);
98         put_bits(&s->pb, 3, i);
99
100         if (i == 7)
101         {
102                 put_bits(&s->pb, 12, s->frame_width);
103                 put_bits(&s->pb, 12, s->frame_height);
104         }
105     }
106
107     /* no checksum or extra data (next 2 bits get 0) */
108     put_bits(&s->pb, 2, 0);
109 }
110
111
112 #define QUALITY_THRESHOLD 100
113 #define THRESHOLD_MULTIPLIER 0.6
114
115 #if HAVE_ALTIVEC
116 #undef vector
117 #endif
118
119 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){
120     int count, y, x, i, j, split, best_mean, best_score, best_count;
121     int best_vector[6];
122     int block_sum[7]= {0, 0, 0, 0, 0, 0};
123     int w= 2<<((level+2)>>1);
124     int h= 2<<((level+1)>>1);
125     int size=w*h;
126     int16_t block[7][256];
127     const int8_t *codebook_sum, *codebook;
128     const uint16_t (*mean_vlc)[2];
129     const uint8_t (*multistage_vlc)[2];
130
131     best_score=0;
132     //FIXME optimize, this doenst need to be done multiple times
133     if(intra){
134         codebook_sum= svq1_intra_codebook_sum[level];
135         codebook= ff_svq1_intra_codebooks[level];
136         mean_vlc= ff_svq1_intra_mean_vlc;
137         multistage_vlc= ff_svq1_intra_multistage_vlc[level];
138         for(y=0; y<h; y++){
139             for(x=0; x<w; x++){
140                 int v= src[x + y*stride];
141                 block[0][x + w*y]= v;
142                 best_score += v*v;
143                 block_sum[0] += v;
144             }
145         }
146     }else{
147         codebook_sum= svq1_inter_codebook_sum[level];
148         codebook= ff_svq1_inter_codebooks[level];
149         mean_vlc= ff_svq1_inter_mean_vlc + 256;
150         multistage_vlc= ff_svq1_inter_multistage_vlc[level];
151         for(y=0; y<h; y++){
152             for(x=0; x<w; x++){
153                 int v= src[x + y*stride] - ref[x + y*stride];
154                 block[0][x + w*y]= v;
155                 best_score += v*v;
156                 block_sum[0] += v;
157             }
158         }
159     }
160
161     best_count=0;
162     best_score -= ((block_sum[0]*block_sum[0])>>(level+3));
163     best_mean= (block_sum[0] + (size>>1)) >> (level+3);
164
165     if(level<4){
166         for(count=1; count<7; count++){
167             int best_vector_score= INT_MAX;
168             int best_vector_sum=-999, best_vector_mean=-999;
169             const int stage= count-1;
170             const int8_t *vector;
171
172             for(i=0; i<16; i++){
173                 int sum= codebook_sum[stage*16 + i];
174                 int sqr, diff, score;
175
176                 vector = codebook + stage*size*16 + i*size;
177                 sqr = s->dsp.ssd_int8_vs_int16(vector, block[stage], size);
178                 diff= block_sum[stage] - sum;
179                 score= sqr - ((diff*(int64_t)diff)>>(level+3)); //FIXME 64bit slooow
180                 if(score < best_vector_score){
181                     int mean= (diff + (size>>1)) >> (level+3);
182                     assert(mean >-300 && mean<300);
183                     mean= av_clip(mean, intra?0:-256, 255);
184                     best_vector_score= score;
185                     best_vector[stage]= i;
186                     best_vector_sum= sum;
187                     best_vector_mean= mean;
188                 }
189             }
190             assert(best_vector_mean != -999);
191             vector= codebook + stage*size*16 + best_vector[stage]*size;
192             for(j=0; j<size; j++){
193                 block[stage+1][j] = block[stage][j] - vector[j];
194             }
195             block_sum[stage+1]= block_sum[stage] - best_vector_sum;
196             best_vector_score +=
197                 lambda*(+ 1 + 4*count
198                         + multistage_vlc[1+count][1]
199                         + mean_vlc[best_vector_mean][1]);
200
201             if(best_vector_score < best_score){
202                 best_score= best_vector_score;
203                 best_count= count;
204                 best_mean= best_vector_mean;
205             }
206         }
207     }
208
209     split=0;
210     if(best_score > threshold && level){
211         int score=0;
212         int offset= (level&1) ? stride*h/2 : w/2;
213         PutBitContext backup[6];
214
215         for(i=level-1; i>=0; i--){
216             backup[i]= s->reorder_pb[i];
217         }
218         score += encode_block(s, src         , ref         , decoded         , stride, level-1, threshold>>1, lambda, intra);
219         score += encode_block(s, src + offset, ref + offset, decoded + offset, stride, level-1, threshold>>1, lambda, intra);
220         score += lambda;
221
222         if(score < best_score){
223             best_score= score;
224             split=1;
225         }else{
226             for(i=level-1; i>=0; i--){
227                 s->reorder_pb[i]= backup[i];
228             }
229         }
230     }
231     if (level > 0)
232         put_bits(&s->reorder_pb[level], 1, split);
233
234     if(!split){
235         assert((best_mean >= 0 && best_mean<256) || !intra);
236         assert(best_mean >= -256 && best_mean<256);
237         assert(best_count >=0 && best_count<7);
238         assert(level<4 || best_count==0);
239
240         /* output the encoding */
241         put_bits(&s->reorder_pb[level],
242             multistage_vlc[1 + best_count][1],
243             multistage_vlc[1 + best_count][0]);
244         put_bits(&s->reorder_pb[level], mean_vlc[best_mean][1],
245             mean_vlc[best_mean][0]);
246
247         for (i = 0; i < best_count; i++){
248             assert(best_vector[i]>=0 && best_vector[i]<16);
249             put_bits(&s->reorder_pb[level], 4, best_vector[i]);
250         }
251
252         for(y=0; y<h; y++){
253             for(x=0; x<w; x++){
254                 decoded[x + y*stride]= src[x + y*stride] - block[best_count][x + w*y] + best_mean;
255             }
256         }
257     }
258
259     return best_score;
260 }
261
262
263 static int svq1_encode_plane(SVQ1Context *s, int plane, unsigned char *src_plane, unsigned char *ref_plane, unsigned char *decoded_plane,
264     int width, int height, int src_stride, int stride)
265 {
266     int x, y;
267     int i;
268     int block_width, block_height;
269     int level;
270     int threshold[6];
271     const int lambda= (s->picture.quality*s->picture.quality) >> (2*FF_LAMBDA_SHIFT);
272
273     /* figure out the acceptable level thresholds in advance */
274     threshold[5] = QUALITY_THRESHOLD;
275     for (level = 4; level >= 0; level--)
276         threshold[level] = threshold[level + 1] * THRESHOLD_MULTIPLIER;
277
278     block_width = (width + 15) / 16;
279     block_height = (height + 15) / 16;
280
281     if(s->picture.pict_type == FF_P_TYPE){
282         s->m.avctx= s->avctx;
283         s->m.current_picture_ptr= &s->m.current_picture;
284         s->m.last_picture_ptr   = &s->m.last_picture;
285         s->m.last_picture.data[0]= ref_plane;
286         s->m.linesize=
287         s->m.last_picture.linesize[0]=
288         s->m.new_picture.linesize[0]=
289         s->m.current_picture.linesize[0]= stride;
290         s->m.width= width;
291         s->m.height= height;
292         s->m.mb_width= block_width;
293         s->m.mb_height= block_height;
294         s->m.mb_stride= s->m.mb_width+1;
295         s->m.b8_stride= 2*s->m.mb_width+1;
296         s->m.f_code=1;
297         s->m.pict_type= s->picture.pict_type;
298         s->m.me_method= s->avctx->me_method;
299         s->m.me.scene_change_score=0;
300         s->m.flags= s->avctx->flags;
301 //        s->m.out_format = FMT_H263;
302 //        s->m.unrestricted_mv= 1;
303
304         s->m.lambda= s->picture.quality;
305         s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
306         s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
307
308         if(!s->motion_val8[plane]){
309             s->motion_val8 [plane]= av_mallocz((s->m.b8_stride*block_height*2 + 2)*2*sizeof(int16_t));
310             s->motion_val16[plane]= av_mallocz((s->m.mb_stride*(block_height + 2) + 1)*2*sizeof(int16_t));
311         }
312
313         s->m.mb_type= s->mb_type;
314
315         //dummies, to avoid segfaults
316         s->m.current_picture.mb_mean=   (uint8_t *)s->dummy;
317         s->m.current_picture.mb_var=    (uint16_t*)s->dummy;
318         s->m.current_picture.mc_mb_var= (uint16_t*)s->dummy;
319         s->m.current_picture.mb_type= s->dummy;
320
321         s->m.current_picture.motion_val[0]= s->motion_val8[plane] + 2;
322         s->m.p_mv_table= s->motion_val16[plane] + s->m.mb_stride + 1;
323         s->m.dsp= s->dsp; //move
324         ff_init_me(&s->m);
325
326         s->m.me.dia_size= s->avctx->dia_size;
327         s->m.first_slice_line=1;
328         for (y = 0; y < block_height; y++) {
329             uint8_t src[stride*16];
330
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         uint8_t src[stride*16];
359
360         for(i=0; i<16 && i + 16*y<height; i++){
361             memcpy(&src[i*stride], &src_plane[(i+16*y)*src_stride], width);
362             for(x=width; x<16*block_width; x++)
363                 src[i*stride+x]= src[i*stride+x-1];
364         }
365         for(; i<16 && i + 16*y<16*block_height; i++)
366             memcpy(&src[i*stride], &src[(i-1)*stride], 16*block_width);
367
368         s->m.mb_y= y;
369         for (x = 0; x < block_width; x++) {
370             uint8_t reorder_buffer[3][6][7*32];
371             int count[3][6];
372             int offset = y * 16 * stride + x * 16;
373             uint8_t *decoded= decoded_plane + offset;
374             uint8_t *ref= ref_plane + offset;
375             int score[4]={0,0,0,0}, best;
376             uint8_t *temp = s->scratchbuf;
377
378             if(s->pb.buf_end - s->pb.buf - (put_bits_count(&s->pb)>>3) < 3000){ //FIXME check size
379                 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
380                 return -1;
381             }
382
383             s->m.mb_x= x;
384             ff_init_block_index(&s->m);
385             ff_update_block_index(&s->m);
386
387             if(s->picture.pict_type == FF_I_TYPE || (s->m.mb_type[x + y*s->m.mb_stride]&CANDIDATE_MB_TYPE_INTRA)){
388                 for(i=0; i<6; i++){
389                     init_put_bits(&s->reorder_pb[i], reorder_buffer[0][i], 7*32);
390                 }
391                 if(s->picture.pict_type == FF_P_TYPE){
392                     const uint8_t *vlc= ff_svq1_block_type_vlc[SVQ1_BLOCK_INTRA];
393                     put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
394                     score[0]= vlc[1]*lambda;
395                 }
396                 score[0]+= encode_block(s, src+16*x, NULL, temp, stride, 5, 64, lambda, 1);
397                 for(i=0; i<6; i++){
398                     count[0][i]= put_bits_count(&s->reorder_pb[i]);
399                     flush_put_bits(&s->reorder_pb[i]);
400                 }
401             }else
402                 score[0]= INT_MAX;
403
404             best=0;
405
406             if(s->picture.pict_type == FF_P_TYPE){
407                 const uint8_t *vlc= ff_svq1_block_type_vlc[SVQ1_BLOCK_INTER];
408                 int mx, my, pred_x, pred_y, dxy;
409                 int16_t *motion_ptr;
410
411                 motion_ptr= h263_pred_motion(&s->m, 0, 0, &pred_x, &pred_y);
412                 if(s->m.mb_type[x + y*s->m.mb_stride]&CANDIDATE_MB_TYPE_INTER){
413                     for(i=0; i<6; i++)
414                         init_put_bits(&s->reorder_pb[i], reorder_buffer[1][i], 7*32);
415
416                     put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
417
418                     s->m.pb= s->reorder_pb[5];
419                     mx= motion_ptr[0];
420                     my= motion_ptr[1];
421                     assert(mx>=-32 && mx<=31);
422                     assert(my>=-32 && my<=31);
423                     assert(pred_x>=-32 && pred_x<=31);
424                     assert(pred_y>=-32 && pred_y<=31);
425                     ff_h263_encode_motion(&s->m, mx - pred_x, 1);
426                     ff_h263_encode_motion(&s->m, my - pred_y, 1);
427                     s->reorder_pb[5]= s->m.pb;
428                     score[1] += lambda*put_bits_count(&s->reorder_pb[5]);
429
430                     dxy= (mx&1) + 2*(my&1);
431
432                     s->dsp.put_pixels_tab[0][dxy](temp+16, ref + (mx>>1) + stride*(my>>1), stride, 16);
433
434                     score[1]+= encode_block(s, src+16*x, temp+16, decoded, stride, 5, 64, lambda, 0);
435                     best= score[1] <= score[0];
436
437                     vlc= ff_svq1_block_type_vlc[SVQ1_BLOCK_SKIP];
438                     score[2]= s->dsp.sse[0](NULL, src+16*x, ref, stride, 16);
439                     score[2]+= vlc[1]*lambda;
440                     if(score[2] < score[best] && mx==0 && my==0){
441                         best=2;
442                         s->dsp.put_pixels_tab[0][0](decoded, ref, stride, 16);
443                         for(i=0; i<6; i++){
444                             count[2][i]=0;
445                         }
446                         put_bits(&s->pb, vlc[1], vlc[0]);
447                     }
448                 }
449
450                 if(best==1){
451                     for(i=0; i<6; i++){
452                         count[1][i]= put_bits_count(&s->reorder_pb[i]);
453                         flush_put_bits(&s->reorder_pb[i]);
454                     }
455                 }else{
456                     motion_ptr[0                 ] = motion_ptr[1                 ]=
457                     motion_ptr[2                 ] = motion_ptr[3                 ]=
458                     motion_ptr[0+2*s->m.b8_stride] = motion_ptr[1+2*s->m.b8_stride]=
459                     motion_ptr[2+2*s->m.b8_stride] = motion_ptr[3+2*s->m.b8_stride]=0;
460                 }
461             }
462
463             s->rd_total += score[best];
464
465             for(i=5; i>=0; i--){
466                 ff_copy_bits(&s->pb, reorder_buffer[best][i], count[best][i]);
467             }
468             if(best==0){
469                 s->dsp.put_pixels_tab[0][0](decoded, temp, stride, 16);
470             }
471         }
472         s->m.first_slice_line=0;
473     }
474     return 0;
475 }
476
477 static av_cold int svq1_encode_init(AVCodecContext *avctx)
478 {
479     SVQ1Context * const s = avctx->priv_data;
480
481     dsputil_init(&s->dsp, avctx);
482     avctx->coded_frame= (AVFrame*)&s->picture;
483
484     s->frame_width = avctx->width;
485     s->frame_height = avctx->height;
486
487     s->y_block_width = (s->frame_width + 15) / 16;
488     s->y_block_height = (s->frame_height + 15) / 16;
489
490     s->c_block_width = (s->frame_width / 4 + 15) / 16;
491     s->c_block_height = (s->frame_height / 4 + 15) / 16;
492
493     s->avctx= avctx;
494     s->m.avctx= avctx;
495     s->m.me.temp      =
496     s->m.me.scratchpad= av_mallocz((avctx->width+64)*2*16*2*sizeof(uint8_t));
497     s->m.me.map       = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
498     s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
499     s->mb_type        = av_mallocz((s->y_block_width+1)*s->y_block_height*sizeof(int16_t));
500     s->dummy          = av_mallocz((s->y_block_width+1)*s->y_block_height*sizeof(int32_t));
501     h263_encode_init(&s->m); //mv_penalty
502
503     return 0;
504 }
505
506 static int svq1_encode_frame(AVCodecContext *avctx, unsigned char *buf,
507     int buf_size, void *data)
508 {
509     SVQ1Context * const s = avctx->priv_data;
510     AVFrame *pict = data;
511     AVFrame * const p= (AVFrame*)&s->picture;
512     AVFrame temp;
513     int i;
514
515     if(avctx->pix_fmt != PIX_FMT_YUV410P){
516         av_log(avctx, AV_LOG_ERROR, "unsupported pixel format\n");
517         return -1;
518     }
519
520     if(!s->current_picture.data[0]){
521         avctx->get_buffer(avctx, &s->current_picture);
522         avctx->get_buffer(avctx, &s->last_picture);
523         s->scratchbuf = av_malloc(s->current_picture.linesize[0] * 16);
524     }
525
526     temp= s->current_picture;
527     s->current_picture= s->last_picture;
528     s->last_picture= temp;
529
530     init_put_bits(&s->pb, buf, buf_size);
531
532     *p = *pict;
533     p->pict_type = avctx->gop_size && avctx->frame_number % avctx->gop_size ? FF_P_TYPE : FF_I_TYPE;
534     p->key_frame = p->pict_type == FF_I_TYPE;
535
536     svq1_write_header(s, p->pict_type);
537     for(i=0; i<3; i++){
538         if(svq1_encode_plane(s, i,
539             s->picture.data[i], s->last_picture.data[i], s->current_picture.data[i],
540             s->frame_width / (i?4:1), s->frame_height / (i?4:1),
541             s->picture.linesize[i], s->current_picture.linesize[i]) < 0)
542                 return -1;
543     }
544
545 //    align_put_bits(&s->pb);
546     while(put_bits_count(&s->pb) & 31)
547         put_bits(&s->pb, 1, 0);
548
549     flush_put_bits(&s->pb);
550
551     return put_bits_count(&s->pb) / 8;
552 }
553
554 static av_cold int svq1_encode_end(AVCodecContext *avctx)
555 {
556     SVQ1Context * const s = avctx->priv_data;
557     int i;
558
559     av_log(avctx, AV_LOG_DEBUG, "RD: %f\n", s->rd_total/(double)(avctx->width*avctx->height*avctx->frame_number));
560
561     av_freep(&s->m.me.scratchpad);
562     av_freep(&s->m.me.map);
563     av_freep(&s->m.me.score_map);
564     av_freep(&s->mb_type);
565     av_freep(&s->dummy);
566     av_freep(&s->scratchbuf);
567
568     for(i=0; i<3; i++){
569         av_freep(&s->motion_val8[i]);
570         av_freep(&s->motion_val16[i]);
571     }
572
573     return 0;
574 }
575
576
577 AVCodec svq1_encoder = {
578     "svq1",
579     CODEC_TYPE_VIDEO,
580     CODEC_ID_SVQ1,
581     sizeof(SVQ1Context),
582     svq1_encode_init,
583     svq1_encode_frame,
584     svq1_encode_end,
585     .pix_fmts= (const enum PixelFormat[]){PIX_FMT_YUV410P, PIX_FMT_NONE},
586     .long_name= NULL_IF_CONFIG_SMALL("Sorenson Vector Quantizer 1 / Sorenson Video 1 / SVQ1"),
587 };