svq1enc.c
Go to the documentation of this file.
1 /*
2  * SVQ1 Encoder
3  * Copyright (C) 2004 Mike Melanson <melanson@pcisys.net>
4  *
5  * This file is part of Libav.
6  *
7  * Libav 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  * Libav 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 Libav; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21 
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)
52 
53  PutBitContext reorder_pb[6]; //why ooh why this sick breadth first order, everything is slower and more complex
54 
57 
58  /* Y plane block dimensions */
61 
62  /* U & V plane (C planes) block dimensions */
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 
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 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){
117  int count, y, x, i, j, split, best_mean, best_score, best_count;
118  int best_vector[6];
119  int block_sum[7]= {0, 0, 0, 0, 0, 0};
120  int w= 2<<((level+2)>>1);
121  int h= 2<<((level+1)>>1);
122  int size=w*h;
123  int16_t block[7][256];
124  const int8_t *codebook_sum, *codebook;
125  const uint16_t (*mean_vlc)[2];
126  const uint8_t (*multistage_vlc)[2];
127 
128  best_score=0;
129  //FIXME optimize, this doenst need to be done multiple times
130  if(intra){
131  codebook_sum= svq1_intra_codebook_sum[level];
132  codebook= ff_svq1_intra_codebooks[level];
133  mean_vlc= ff_svq1_intra_mean_vlc;
134  multistage_vlc= ff_svq1_intra_multistage_vlc[level];
135  for(y=0; y<h; y++){
136  for(x=0; x<w; x++){
137  int v= src[x + y*stride];
138  block[0][x + w*y]= v;
139  best_score += v*v;
140  block_sum[0] += v;
141  }
142  }
143  }else{
144  codebook_sum= svq1_inter_codebook_sum[level];
145  codebook= ff_svq1_inter_codebooks[level];
146  mean_vlc= ff_svq1_inter_mean_vlc + 256;
147  multistage_vlc= ff_svq1_inter_multistage_vlc[level];
148  for(y=0; y<h; y++){
149  for(x=0; x<w; x++){
150  int v= src[x + y*stride] - ref[x + y*stride];
151  block[0][x + w*y]= v;
152  best_score += v*v;
153  block_sum[0] += v;
154  }
155  }
156  }
157 
158  best_count=0;
159  best_score -= (int)(((unsigned)block_sum[0]*block_sum[0])>>(level+3));
160  best_mean= (block_sum[0] + (size>>1)) >> (level+3);
161 
162  if(level<4){
163  for(count=1; count<7; count++){
164  int best_vector_score= INT_MAX;
165  int best_vector_sum=-999, best_vector_mean=-999;
166  const int stage= count-1;
167  const int8_t *vector;
168 
169  for(i=0; i<16; i++){
170  int sum= codebook_sum[stage*16 + i];
171  int sqr, diff, score;
172 
173  vector = codebook + stage*size*16 + i*size;
174  sqr = s->dsp.ssd_int8_vs_int16(vector, block[stage], size);
175  diff= block_sum[stage] - sum;
176  score= sqr - ((diff*(int64_t)diff)>>(level+3)); //FIXME 64bit slooow
177  if(score < best_vector_score){
178  int mean= (diff + (size>>1)) >> (level+3);
179  assert(mean >-300 && mean<300);
180  mean= av_clip(mean, intra?0:-256, 255);
181  best_vector_score= score;
182  best_vector[stage]= i;
183  best_vector_sum= sum;
184  best_vector_mean= mean;
185  }
186  }
187  assert(best_vector_mean != -999);
188  vector= codebook + stage*size*16 + best_vector[stage]*size;
189  for(j=0; j<size; j++){
190  block[stage+1][j] = block[stage][j] - vector[j];
191  }
192  block_sum[stage+1]= block_sum[stage] - best_vector_sum;
193  best_vector_score +=
194  lambda*(+ 1 + 4*count
195  + multistage_vlc[1+count][1]
196  + mean_vlc[best_vector_mean][1]);
197 
198  if(best_vector_score < best_score){
199  best_score= best_vector_score;
200  best_count= count;
201  best_mean= best_vector_mean;
202  }
203  }
204  }
205 
206  split=0;
207  if(best_score > threshold && level){
208  int score=0;
209  int offset= (level&1) ? stride*h/2 : w/2;
210  PutBitContext backup[6];
211 
212  for(i=level-1; i>=0; i--){
213  backup[i]= s->reorder_pb[i];
214  }
215  score += encode_block(s, src , ref , decoded , stride, level-1, threshold>>1, lambda, intra);
216  score += encode_block(s, src + offset, ref + offset, decoded + offset, stride, level-1, threshold>>1, lambda, intra);
217  score += lambda;
218 
219  if(score < best_score){
220  best_score= score;
221  split=1;
222  }else{
223  for(i=level-1; i>=0; i--){
224  s->reorder_pb[i]= backup[i];
225  }
226  }
227  }
228  if (level > 0)
229  put_bits(&s->reorder_pb[level], 1, split);
230 
231  if(!split){
232  assert((best_mean >= 0 && best_mean<256) || !intra);
233  assert(best_mean >= -256 && best_mean<256);
234  assert(best_count >=0 && best_count<7);
235  assert(level<4 || best_count==0);
236 
237  /* output the encoding */
238  put_bits(&s->reorder_pb[level],
239  multistage_vlc[1 + best_count][1],
240  multistage_vlc[1 + best_count][0]);
241  put_bits(&s->reorder_pb[level], mean_vlc[best_mean][1],
242  mean_vlc[best_mean][0]);
243 
244  for (i = 0; i < best_count; i++){
245  assert(best_vector[i]>=0 && best_vector[i]<16);
246  put_bits(&s->reorder_pb[level], 4, best_vector[i]);
247  }
248 
249  for(y=0; y<h; y++){
250  for(x=0; x<w; x++){
251  decoded[x + y*stride]= src[x + y*stride] - block[best_count][x + w*y] + best_mean;
252  }
253  }
254  }
255 
256  return best_score;
257 }
258 
259 
260 static int svq1_encode_plane(SVQ1Context *s, int plane, unsigned char *src_plane, unsigned char *ref_plane, unsigned char *decoded_plane,
261  int width, int height, int src_stride, int stride)
262 {
263  int x, y;
264  int i;
265  int block_width, block_height;
266  int level;
267  int threshold[6];
268  uint8_t *src = s->scratchbuf + stride * 16;
269  const int lambda= (s->picture.quality*s->picture.quality) >> (2*FF_LAMBDA_SHIFT);
270 
271  /* figure out the acceptable level thresholds in advance */
272  threshold[5] = QUALITY_THRESHOLD;
273  for (level = 4; level >= 0; level--)
274  threshold[level] = threshold[level + 1] * THRESHOLD_MULTIPLIER;
275 
276  block_width = (width + 15) / 16;
277  block_height = (height + 15) / 16;
278 
280  s->m.avctx= s->avctx;
282  s->m.last_picture_ptr = &s->m.last_picture;
283  s->m.last_picture.f.data[0] = ref_plane;
284  s->m.linesize=
285  s->m.last_picture.f.linesize[0] =
286  s->m.new_picture.f.linesize[0] =
288  s->m.width= width;
289  s->m.height= height;
290  s->m.mb_width= block_width;
291  s->m.mb_height= block_height;
292  s->m.mb_stride= s->m.mb_width+1;
293  s->m.b8_stride= 2*s->m.mb_width+1;
294  s->m.f_code=1;
295  s->m.pict_type= s->picture.pict_type;
296  s->m.me_method= s->avctx->me_method;
297  s->m.me.scene_change_score=0;
298  s->m.flags= s->avctx->flags;
299 // s->m.out_format = FMT_H263;
300 // s->m.unrestricted_mv= 1;
301 
302  s->m.lambda= s->picture.quality;
303  s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
305 
306  if(!s->motion_val8[plane]){
307  s->motion_val8 [plane]= av_mallocz((s->m.b8_stride*block_height*2 + 2)*2*sizeof(int16_t));
308  s->motion_val16[plane]= av_mallocz((s->m.mb_stride*(block_height + 2) + 1)*2*sizeof(int16_t));
309  }
310 
311  s->m.mb_type= s->mb_type;
312 
313  //dummies, to avoid segfaults
314  s->m.current_picture.mb_mean= (uint8_t *)s->dummy;
315  s->m.current_picture.mb_var= (uint16_t*)s->dummy;
316  s->m.current_picture.mc_mb_var= (uint16_t*)s->dummy;
317  s->m.current_picture.f.mb_type = s->dummy;
318 
319  s->m.current_picture.f.motion_val[0] = s->motion_val8[plane] + 2;
320  s->m.p_mv_table= s->motion_val16[plane] + s->m.mb_stride + 1;
321  s->m.dsp= s->dsp; //move
322  ff_init_me(&s->m);
323 
324  s->m.me.dia_size= s->avctx->dia_size;
325  s->m.first_slice_line=1;
326  for (y = 0; y < block_height; y++) {
327  s->m.new_picture.f.data[0] = src - y*16*stride; //ugly
328  s->m.mb_y= y;
329 
330  for(i=0; i<16 && i + 16*y<height; i++){
331  memcpy(&src[i*stride], &src_plane[(i+16*y)*src_stride], width);
332  for(x=width; x<16*block_width; x++)
333  src[i*stride+x]= src[i*stride+x-1];
334  }
335  for(; i<16 && i + 16*y<16*block_height; i++)
336  memcpy(&src[i*stride], &src[(i-1)*stride], 16*block_width);
337 
338  for (x = 0; x < block_width; x++) {
339  s->m.mb_x= x;
340  ff_init_block_index(&s->m);
342 
343  ff_estimate_p_frame_motion(&s->m, x, y);
344  }
345  s->m.first_slice_line=0;
346  }
347 
348  ff_fix_long_p_mvs(&s->m);
350  }
351 
352  s->m.first_slice_line=1;
353  for (y = 0; y < block_height; y++) {
354  for(i=0; i<16 && i + 16*y<height; i++){
355  memcpy(&src[i*stride], &src_plane[(i+16*y)*src_stride], width);
356  for(x=width; x<16*block_width; x++)
357  src[i*stride+x]= src[i*stride+x-1];
358  }
359  for(; i<16 && i + 16*y<16*block_height; i++)
360  memcpy(&src[i*stride], &src[(i-1)*stride], 16*block_width);
361 
362  s->m.mb_y= y;
363  for (x = 0; x < block_width; x++) {
364  uint8_t reorder_buffer[3][6][7*32];
365  int count[3][6];
366  int offset = y * 16 * stride + x * 16;
367  uint8_t *decoded= decoded_plane + offset;
368  uint8_t *ref= ref_plane + offset;
369  int score[4]={0,0,0,0}, best;
370  uint8_t *temp = s->scratchbuf;
371 
372  if(s->pb.buf_end - s->pb.buf - (put_bits_count(&s->pb)>>3) < 3000){ //FIXME check size
373  av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
374  return -1;
375  }
376 
377  s->m.mb_x= x;
378  ff_init_block_index(&s->m);
380 
382  for(i=0; i<6; i++){
383  init_put_bits(&s->reorder_pb[i], reorder_buffer[0][i], 7*32);
384  }
386  const uint8_t *vlc= ff_svq1_block_type_vlc[SVQ1_BLOCK_INTRA];
387  put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
388  score[0]= vlc[1]*lambda;
389  }
390  score[0]+= encode_block(s, src+16*x, NULL, temp, stride, 5, 64, lambda, 1);
391  for(i=0; i<6; i++){
392  count[0][i]= put_bits_count(&s->reorder_pb[i]);
393  flush_put_bits(&s->reorder_pb[i]);
394  }
395  }else
396  score[0]= INT_MAX;
397 
398  best=0;
399 
401  const uint8_t *vlc= ff_svq1_block_type_vlc[SVQ1_BLOCK_INTER];
402  int mx, my, pred_x, pred_y, dxy;
403  int16_t *motion_ptr;
404 
405  motion_ptr= ff_h263_pred_motion(&s->m, 0, 0, &pred_x, &pred_y);
406  if(s->m.mb_type[x + y*s->m.mb_stride]&CANDIDATE_MB_TYPE_INTER){
407  for(i=0; i<6; i++)
408  init_put_bits(&s->reorder_pb[i], reorder_buffer[1][i], 7*32);
409 
410  put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
411 
412  s->m.pb= s->reorder_pb[5];
413  mx= motion_ptr[0];
414  my= motion_ptr[1];
415  assert(mx>=-32 && mx<=31);
416  assert(my>=-32 && my<=31);
417  assert(pred_x>=-32 && pred_x<=31);
418  assert(pred_y>=-32 && pred_y<=31);
419  ff_h263_encode_motion(&s->m, mx - pred_x, 1);
420  ff_h263_encode_motion(&s->m, my - pred_y, 1);
421  s->reorder_pb[5]= s->m.pb;
422  score[1] += lambda*put_bits_count(&s->reorder_pb[5]);
423 
424  dxy= (mx&1) + 2*(my&1);
425 
426  s->dsp.put_pixels_tab[0][dxy](temp+16, ref + (mx>>1) + stride*(my>>1), stride, 16);
427 
428  score[1]+= encode_block(s, src+16*x, temp+16, decoded, stride, 5, 64, lambda, 0);
429  best= score[1] <= score[0];
430 
432  score[2]= s->dsp.sse[0](NULL, src+16*x, ref, stride, 16);
433  score[2]+= vlc[1]*lambda;
434  if(score[2] < score[best] && mx==0 && my==0){
435  best=2;
436  s->dsp.put_pixels_tab[0][0](decoded, ref, stride, 16);
437  for(i=0; i<6; i++){
438  count[2][i]=0;
439  }
440  put_bits(&s->pb, vlc[1], vlc[0]);
441  }
442  }
443 
444  if(best==1){
445  for(i=0; i<6; i++){
446  count[1][i]= put_bits_count(&s->reorder_pb[i]);
447  flush_put_bits(&s->reorder_pb[i]);
448  }
449  }else{
450  motion_ptr[0 ] = motion_ptr[1 ]=
451  motion_ptr[2 ] = motion_ptr[3 ]=
452  motion_ptr[0+2*s->m.b8_stride] = motion_ptr[1+2*s->m.b8_stride]=
453  motion_ptr[2+2*s->m.b8_stride] = motion_ptr[3+2*s->m.b8_stride]=0;
454  }
455  }
456 
457  s->rd_total += score[best];
458 
459  for(i=5; i>=0; i--){
460  avpriv_copy_bits(&s->pb, reorder_buffer[best][i], count[best][i]);
461  }
462  if(best==0){
463  s->dsp.put_pixels_tab[0][0](decoded, temp, stride, 16);
464  }
465  }
466  s->m.first_slice_line=0;
467  }
468  return 0;
469 }
470 
472 {
473  SVQ1Context * const s = avctx->priv_data;
474 
475  dsputil_init(&s->dsp, avctx);
476  avctx->coded_frame= (AVFrame*)&s->picture;
477 
478  s->frame_width = avctx->width;
479  s->frame_height = avctx->height;
480 
481  s->y_block_width = (s->frame_width + 15) / 16;
482  s->y_block_height = (s->frame_height + 15) / 16;
483 
484  s->c_block_width = (s->frame_width / 4 + 15) / 16;
485  s->c_block_height = (s->frame_height / 4 + 15) / 16;
486 
487  s->avctx= avctx;
488  s->m.avctx= avctx;
490  s->m.me.temp =
491  s->m.me.scratchpad= av_mallocz((avctx->width+64)*2*16*2*sizeof(uint8_t));
492  s->m.me.map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
493  s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
494  s->mb_type = av_mallocz((s->y_block_width+1)*s->y_block_height*sizeof(int16_t));
495  s->dummy = av_mallocz((s->y_block_width+1)*s->y_block_height*sizeof(int32_t));
496  ff_h263_encode_init(&s->m); //mv_penalty
497 
498  return 0;
499 }
500 
501 static int svq1_encode_frame(AVCodecContext *avctx, unsigned char *buf,
502  int buf_size, void *data)
503 {
504  SVQ1Context * const s = avctx->priv_data;
505  AVFrame *pict = data;
506  AVFrame * const p= (AVFrame*)&s->picture;
507  AVFrame temp;
508  int i;
509 
510  if(avctx->pix_fmt != PIX_FMT_YUV410P){
511  av_log(avctx, AV_LOG_ERROR, "unsupported pixel format\n");
512  return -1;
513  }
514 
515  if(!s->current_picture.data[0]){
516  ff_get_buffer(avctx, &s->current_picture);
517  ff_get_buffer(avctx, &s->last_picture);
518  s->scratchbuf = av_malloc(s->current_picture.linesize[0] * 16 * 2);
519  }
520 
521  temp= s->current_picture;
523  s->last_picture= temp;
524 
525  init_put_bits(&s->pb, buf, buf_size);
526 
527  *p = *pict;
528  p->pict_type = avctx->gop_size && avctx->frame_number % avctx->gop_size ? AV_PICTURE_TYPE_P : AV_PICTURE_TYPE_I;
530 
532  for(i=0; i<3; i++){
533  if(svq1_encode_plane(s, i,
534  s->picture.data[i], s->last_picture.data[i], s->current_picture.data[i],
535  s->frame_width / (i?4:1), s->frame_height / (i?4:1),
536  s->picture.linesize[i], s->current_picture.linesize[i]) < 0)
537  return -1;
538  }
539 
540 // avpriv_align_put_bits(&s->pb);
541  while(put_bits_count(&s->pb) & 31)
542  put_bits(&s->pb, 1, 0);
543 
544  flush_put_bits(&s->pb);
545 
546  return put_bits_count(&s->pb) / 8;
547 }
548 
550 {
551  SVQ1Context * const s = avctx->priv_data;
552  int i;
553 
554  av_log(avctx, AV_LOG_DEBUG, "RD: %f\n", s->rd_total/(double)(avctx->width*avctx->height*avctx->frame_number));
555 
556  av_freep(&s->m.me.scratchpad);
557  av_freep(&s->m.me.map);
558  av_freep(&s->m.me.score_map);
559  av_freep(&s->mb_type);
560  av_freep(&s->dummy);
561  av_freep(&s->scratchbuf);
562 
563  for(i=0; i<3; i++){
564  av_freep(&s->motion_val8[i]);
565  av_freep(&s->motion_val16[i]);
566  }
567 
568  return 0;
569 }
570 
571 
573  .name = "svq1",
574  .type = AVMEDIA_TYPE_VIDEO,
575  .id = CODEC_ID_SVQ1,
576  .priv_data_size = sizeof(SVQ1Context),
578  .encode = svq1_encode_frame,
580  .pix_fmts= (const enum PixelFormat[]){PIX_FMT_YUV410P, PIX_FMT_NONE},
581  .long_name= NULL_IF_CONFIG_SMALL("Sorenson Vector Quantizer 1 / Sorenson Video 1 / SVQ1"),
582 };