FFmpeg
 All Data Structures Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
snow.h
Go to the documentation of this file.
1 /*
2  * Copyright (C) 2004 Michael Niedermayer <michaelni@gmx.at>
3  * Copyright (C) 2006 Robert Edele <yartrebo@earthlink.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 #ifndef AVCODEC_SNOW_H
23 #define AVCODEC_SNOW_H
24 
25 #include "dsputil.h"
26 #include "snow_dwt.h"
27 
28 #include "rangecoder.h"
29 #include "mathops.h"
30 #include "mpegvideo.h"
31 #include "h264qpel.h"
32 
33 #define MID_STATE 128
34 
35 #define MAX_PLANES 4
36 #define QSHIFT 5
37 #define QROOT (1<<QSHIFT)
38 #define LOSSLESS_QLOG -128
39 #define FRAC_BITS 4
40 #define MAX_REF_FRAMES 8
41 
42 #define LOG2_OBMC_MAX 8
43 #define OBMC_MAX (1<<(LOG2_OBMC_MAX))
44 typedef struct BlockNode{
45  int16_t mx;
46  int16_t my;
50 //#define TYPE_SPLIT 1
51 #define BLOCK_INTRA 1
52 #define BLOCK_OPT 2
53 //#define TYPE_NOCOLOR 4
54  uint8_t level; //FIXME merge into type?
55 }BlockNode;
56 
57 static const BlockNode null_block= { //FIXME add border maybe
58  .color= {128,128,128},
59  .mx= 0,
60  .my= 0,
61  .ref= 0,
62  .type= 0,
63  .level= 0,
64 };
65 
66 #define LOG2_MB_SIZE 4
67 #define MB_SIZE (1<<LOG2_MB_SIZE)
68 #define ENCODER_EXTRA_BITS 4
69 #define HTAPS_MAX 8
70 
71 typedef struct x_and_coeff{
72  int16_t x;
73  uint16_t coeff;
74 } x_and_coeff;
75 
76 typedef struct SubBand{
77  int level;
78  int stride;
79  int width;
80  int height;
81  int qlog; ///< log(qscale)/log[2^(1/6)]
83  IDWTELEM *ibuf;
86  int stride_line; ///< Stride measured in lines, not pixels.
88  struct SubBand *parent;
89  uint8_t state[/*7*2*/ 7 + 512][32];
90 }SubBand;
91 
92 typedef struct Plane{
93  int width;
94  int height;
96 
97  int htaps;
98  int8_t hcoeff[HTAPS_MAX/2];
99  int diag_mc;
100  int fast_mc;
101 
105 }Plane;
106 
107 typedef struct SnowContext{
108  AVClass *class;
116  AVFrame input_picture; ///< new_picture with the internal linesizes
121 // uint8_t q_context[16];
123  uint8_t block_state[128 + 32*128];
124  int keyframe;
126  int version;
135  int16_t (*ref_mvs[MAX_REF_FRAMES])[2];
146  int qlog;
148  int lambda;
149  int lambda2;
150  int pass1_rc;
151  int mv_scale;
153  int qbias;
155 #define QBIAS_SHIFT 3
156  int b_width;
157  int b_height;
162 #define ME_CACHE_SIZE 1024
168 
169  MpegEncContext m; // needed for motion estimation, should not be used for anything else, the idea is to eventually make the motion estimation independent of MpegEncContext, so this will be removed then (FIXME/XXX)
170 
173 }SnowContext;
174 
175 /* Tables */
176 extern const uint8_t * const ff_obmc_tab[4];
177 extern uint8_t ff_qexp[QROOT];
179 
180 /* C bits used by mmx/sse2/altivec */
181 
182 static av_always_inline void snow_interleave_line_header(int * i, int width, IDWTELEM * low, IDWTELEM * high){
183  (*i) = (width) - 2;
184 
185  if (width & 1){
186  low[(*i)+1] = low[((*i)+1)>>1];
187  (*i)--;
188  }
189 }
190 
191 static av_always_inline void snow_interleave_line_footer(int * i, IDWTELEM * low, IDWTELEM * high){
192  for (; (*i)>=0; (*i)-=2){
193  low[(*i)+1] = high[(*i)>>1];
194  low[*i] = low[(*i)>>1];
195  }
196 }
197 
198 static av_always_inline void snow_horizontal_compose_lift_lead_out(int i, IDWTELEM * dst, IDWTELEM * src, IDWTELEM * ref, int width, int w, int lift_high, int mul, int add, int shift){
199  for(; i<w; i++){
200  dst[i] = src[i] - ((mul * (ref[i] + ref[i + 1]) + add) >> shift);
201  }
202 
203  if((width^lift_high)&1){
204  dst[w] = src[w] - ((mul * 2 * ref[w] + add) >> shift);
205  }
206 }
207 
209  for(; i<w; i++){
210  dst[i] = src[i] + ((ref[i] + ref[(i+1)]+W_BO + 4 * src[i]) >> W_BS);
211  }
212 
213  if(width&1){
214  dst[w] = src[w] + ((2 * ref[w] + W_BO + 4 * src[w]) >> W_BS);
215  }
216 }
217 
218 /* common code */
219 
228  int sx, int sy, int b_w, int b_h, BlockNode *block,
229  int plane_index, int w, int h);
230 /* common inline functions */
231 //XXX doublecheck all of them should stay inlined
232 
233 static inline void snow_set_blocks(SnowContext *s, int level, int x, int y, int l, int cb, int cr, int mx, int my, int ref, int type){
234  const int w= s->b_width << s->block_max_depth;
235  const int rem_depth= s->block_max_depth - level;
236  const int index= (x + y*w) << rem_depth;
237  const int block_w= 1<<rem_depth;
239  int i,j;
240 
241  block.color[0]= l;
242  block.color[1]= cb;
243  block.color[2]= cr;
244  block.mx= mx;
245  block.my= my;
246  block.ref= ref;
247  block.type= type;
248  block.level= level;
249 
250  for(j=0; j<block_w; j++){
251  for(i=0; i<block_w; i++){
252  s->block[index + i + j*w]= block;
253  }
254  }
255 }
256 
257 static inline void pred_mv(SnowContext *s, int *mx, int *my, int ref,
258  const BlockNode *left, const BlockNode *top, const BlockNode *tr){
259  if(s->ref_frames == 1){
260  *mx = mid_pred(left->mx, top->mx, tr->mx);
261  *my = mid_pred(left->my, top->my, tr->my);
262  }else{
263  const int *scale = ff_scale_mv_ref[ref];
264  *mx = mid_pred((left->mx * scale[left->ref] + 128) >>8,
265  (top ->mx * scale[top ->ref] + 128) >>8,
266  (tr ->mx * scale[tr ->ref] + 128) >>8);
267  *my = mid_pred((left->my * scale[left->ref] + 128) >>8,
268  (top ->my * scale[top ->ref] + 128) >>8,
269  (tr ->my * scale[tr ->ref] + 128) >>8);
270  }
271 }
272 
274  if((a->type&BLOCK_INTRA) && (b->type&BLOCK_INTRA)){
275  return !((a->color[0] - b->color[0]) | (a->color[1] - b->color[1]) | (a->color[2] - b->color[2]));
276  }else{
277  return !((a->mx - b->mx) | (a->my - b->my) | (a->ref - b->ref) | ((a->type ^ b->type)&BLOCK_INTRA));
278  }
279 }
280 
281 //FIXME name cleanup (b_w, block_w, b_width stuff)
282 //XXX should we really inline it?
283 static av_always_inline void add_yblock(SnowContext *s, int sliced, slice_buffer *sb, IDWTELEM *dst, uint8_t *dst8, const uint8_t *obmc, int src_x, int src_y, int b_w, int b_h, int w, int h, int dst_stride, int src_stride, int obmc_stride, int b_x, int b_y, int add, int offset_dst, int plane_index){
284  const int b_width = s->b_width << s->block_max_depth;
285  const int b_height= s->b_height << s->block_max_depth;
286  const int b_stride= b_width;
287  BlockNode *lt= &s->block[b_x + b_y*b_stride];
288  BlockNode *rt= lt+1;
289  BlockNode *lb= lt+b_stride;
290  BlockNode *rb= lb+1;
291  uint8_t *block[4];
292  int tmp_step= src_stride >= 7*MB_SIZE ? MB_SIZE : MB_SIZE*src_stride;
293  uint8_t *tmp = s->scratchbuf;
294  uint8_t *ptmp;
295  int x,y;
296 
297  if(b_x<0){
298  lt= rt;
299  lb= rb;
300  }else if(b_x + 1 >= b_width){
301  rt= lt;
302  rb= lb;
303  }
304  if(b_y<0){
305  lt= lb;
306  rt= rb;
307  }else if(b_y + 1 >= b_height){
308  lb= lt;
309  rb= rt;
310  }
311 
312  if(src_x<0){ //FIXME merge with prev & always round internal width up to *16
313  obmc -= src_x;
314  b_w += src_x;
315  if(!sliced && !offset_dst)
316  dst -= src_x;
317  src_x=0;
318  }else if(src_x + b_w > w){
319  b_w = w - src_x;
320  }
321  if(src_y<0){
322  obmc -= src_y*obmc_stride;
323  b_h += src_y;
324  if(!sliced && !offset_dst)
325  dst -= src_y*dst_stride;
326  src_y=0;
327  }else if(src_y + b_h> h){
328  b_h = h - src_y;
329  }
330 
331  if(b_w<=0 || b_h<=0) return;
332 
333  av_assert2(src_stride > 2*MB_SIZE + 5);
334 
335  if(!sliced && offset_dst)
336  dst += src_x + src_y*dst_stride;
337  dst8+= src_x + src_y*src_stride;
338 // src += src_x + src_y*src_stride;
339 
340  ptmp= tmp + 3*tmp_step;
341  block[0]= ptmp;
342  ptmp+=tmp_step;
343  ff_snow_pred_block(s, block[0], tmp, src_stride, src_x, src_y, b_w, b_h, lt, plane_index, w, h);
344 
345  if(same_block(lt, rt)){
346  block[1]= block[0];
347  }else{
348  block[1]= ptmp;
349  ptmp+=tmp_step;
350  ff_snow_pred_block(s, block[1], tmp, src_stride, src_x, src_y, b_w, b_h, rt, plane_index, w, h);
351  }
352 
353  if(same_block(lt, lb)){
354  block[2]= block[0];
355  }else if(same_block(rt, lb)){
356  block[2]= block[1];
357  }else{
358  block[2]= ptmp;
359  ptmp+=tmp_step;
360  ff_snow_pred_block(s, block[2], tmp, src_stride, src_x, src_y, b_w, b_h, lb, plane_index, w, h);
361  }
362 
363  if(same_block(lt, rb) ){
364  block[3]= block[0];
365  }else if(same_block(rt, rb)){
366  block[3]= block[1];
367  }else if(same_block(lb, rb)){
368  block[3]= block[2];
369  }else{
370  block[3]= ptmp;
371  ff_snow_pred_block(s, block[3], tmp, src_stride, src_x, src_y, b_w, b_h, rb, plane_index, w, h);
372  }
373  if(sliced){
374  s->dwt.inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
375  }else{
376  for(y=0; y<b_h; y++){
377  //FIXME ugly misuse of obmc_stride
378  const uint8_t *obmc1= obmc + y*obmc_stride;
379  const uint8_t *obmc2= obmc1+ (obmc_stride>>1);
380  const uint8_t *obmc3= obmc1+ obmc_stride*(obmc_stride>>1);
381  const uint8_t *obmc4= obmc3+ (obmc_stride>>1);
382  for(x=0; x<b_w; x++){
383  int v= obmc1[x] * block[3][x + y*src_stride]
384  +obmc2[x] * block[2][x + y*src_stride]
385  +obmc3[x] * block[1][x + y*src_stride]
386  +obmc4[x] * block[0][x + y*src_stride];
387 
388  v <<= 8 - LOG2_OBMC_MAX;
389  if(FRAC_BITS != 8){
390  v >>= 8 - FRAC_BITS;
391  }
392  if(add){
393  v += dst[x + y*dst_stride];
394  v = (v + (1<<(FRAC_BITS-1))) >> FRAC_BITS;
395  if(v&(~255)) v= ~(v>>31);
396  dst8[x + y*src_stride] = v;
397  }else{
398  dst[x + y*dst_stride] -= v;
399  }
400  }
401  }
402  }
403 }
404 
405 static av_always_inline void predict_slice(SnowContext *s, IDWTELEM *buf, int plane_index, int add, int mb_y){
406  Plane *p= &s->plane[plane_index];
407  const int mb_w= s->b_width << s->block_max_depth;
408  const int mb_h= s->b_height << s->block_max_depth;
409  int x, y, mb_x;
410  int block_size = MB_SIZE >> s->block_max_depth;
411  int block_w = plane_index ? block_size>>s->chroma_h_shift : block_size;
412  int block_h = plane_index ? block_size>>s->chroma_v_shift : block_size;
413  const uint8_t *obmc = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
414  const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
415  int ref_stride= s->current_picture.linesize[plane_index];
416  uint8_t *dst8= s->current_picture.data[plane_index];
417  int w= p->width;
418  int h= p->height;
419  av_assert2(s->chroma_h_shift == s->chroma_v_shift); // obmc params assume squares
420  if(s->keyframe || (s->avctx->debug&512)){
421  if(mb_y==mb_h)
422  return;
423 
424  if(add){
425  for(y=block_h*mb_y; y<FFMIN(h,block_h*(mb_y+1)); y++){
426  for(x=0; x<w; x++){
427  int v= buf[x + y*w] + (128<<FRAC_BITS) + (1<<(FRAC_BITS-1));
428  v >>= FRAC_BITS;
429  if(v&(~255)) v= ~(v>>31);
430  dst8[x + y*ref_stride]= v;
431  }
432  }
433  }else{
434  for(y=block_h*mb_y; y<FFMIN(h,block_h*(mb_y+1)); y++){
435  for(x=0; x<w; x++){
436  buf[x + y*w]-= 128<<FRAC_BITS;
437  }
438  }
439  }
440 
441  return;
442  }
443 
444  for(mb_x=0; mb_x<=mb_w; mb_x++){
445  add_yblock(s, 0, NULL, buf, dst8, obmc,
446  block_w*mb_x - block_w/2,
447  block_h*mb_y - block_h/2,
448  block_w, block_h,
449  w, h,
450  w, ref_stride, obmc_stride,
451  mb_x - 1, mb_y - 1,
452  add, 1, plane_index);
453  }
454 }
455 
456 static av_always_inline void predict_plane(SnowContext *s, IDWTELEM *buf, int plane_index, int add){
457  const int mb_h= s->b_height << s->block_max_depth;
458  int mb_y;
459  for(mb_y=0; mb_y<=mb_h; mb_y++)
460  predict_slice(s, buf, plane_index, add, mb_y);
461 }
462 
463 static inline void set_blocks(SnowContext *s, int level, int x, int y, int l, int cb, int cr, int mx, int my, int ref, int type){
464  const int w= s->b_width << s->block_max_depth;
465  const int rem_depth= s->block_max_depth - level;
466  const int index= (x + y*w) << rem_depth;
467  const int block_w= 1<<rem_depth;
468  const int block_h= 1<<rem_depth; //FIXME "w!=h"
470  int i,j;
471 
472  block.color[0]= l;
473  block.color[1]= cb;
474  block.color[2]= cr;
475  block.mx= mx;
476  block.my= my;
477  block.ref= ref;
478  block.type= type;
479  block.level= level;
480 
481  for(j=0; j<block_h; j++){
482  for(i=0; i<block_w; i++){
483  s->block[index + i + j*w]= block;
484  }
485  }
486 }
487 
488 static inline void init_ref(MotionEstContext *c, uint8_t *src[3], uint8_t *ref[3], uint8_t *ref2[3], int x, int y, int ref_index){
489  SnowContext *s = c->avctx->priv_data;
490  const int offset[3]= {
491  y*c-> stride + x,
492  ((y*c->uvstride + x)>>s->chroma_h_shift),
493  ((y*c->uvstride + x)>>s->chroma_h_shift),
494  };
495  int i;
496  for(i=0; i<3; i++){
497  c->src[0][i]= src [i];
498  c->ref[0][i]= ref [i] + offset[i];
499  }
500  av_assert2(!ref_index);
501 }
502 
503 
504 /* bitstream functions */
505 
506 extern const int8_t ff_quant3bA[256];
507 
508 #define QEXPSHIFT (7-FRAC_BITS+8) //FIXME try to change this to 0
509 
510 static inline void put_symbol(RangeCoder *c, uint8_t *state, int v, int is_signed){
511  int i;
512 
513  if(v){
514  const int a= FFABS(v);
515  const int e= av_log2(a);
516  const int el= FFMIN(e, 10);
517  put_rac(c, state+0, 0);
518 
519  for(i=0; i<el; i++){
520  put_rac(c, state+1+i, 1); //1..10
521  }
522  for(; i<e; i++){
523  put_rac(c, state+1+9, 1); //1..10
524  }
525  put_rac(c, state+1+FFMIN(i,9), 0);
526 
527  for(i=e-1; i>=el; i--){
528  put_rac(c, state+22+9, (a>>i)&1); //22..31
529  }
530  for(; i>=0; i--){
531  put_rac(c, state+22+i, (a>>i)&1); //22..31
532  }
533 
534  if(is_signed)
535  put_rac(c, state+11 + el, v < 0); //11..21
536  }else{
537  put_rac(c, state+0, 1);
538  }
539 }
540 
541 static inline int get_symbol(RangeCoder *c, uint8_t *state, int is_signed){
542  if(get_rac(c, state+0))
543  return 0;
544  else{
545  int i, e, a;
546  e= 0;
547  while(get_rac(c, state+1 + FFMIN(e,9))){ //1..10
548  e++;
549  }
550 
551  a= 1;
552  for(i=e-1; i>=0; i--){
553  a += a + get_rac(c, state+22 + FFMIN(i,9)); //22..31
554  }
555 
556  e= -(is_signed && get_rac(c, state+11 + FFMIN(e,10))); //11..21
557  return (a^e)-e;
558  }
559 }
560 
561 static inline void put_symbol2(RangeCoder *c, uint8_t *state, int v, int log2){
562  int i;
563  int r= log2>=0 ? 1<<log2 : 1;
564 
565  av_assert2(v>=0);
566  av_assert2(log2>=-4);
567 
568  while(v >= r){
569  put_rac(c, state+4+log2, 1);
570  v -= r;
571  log2++;
572  if(log2>0) r+=r;
573  }
574  put_rac(c, state+4+log2, 0);
575 
576  for(i=log2-1; i>=0; i--){
577  put_rac(c, state+31-i, (v>>i)&1);
578  }
579 }
580 
581 static inline int get_symbol2(RangeCoder *c, uint8_t *state, int log2){
582  int i;
583  int r= log2>=0 ? 1<<log2 : 1;
584  int v=0;
585 
586  av_assert2(log2>=-4);
587 
588  while(log2<28 && get_rac(c, state+4+log2)){
589  v+= r;
590  log2++;
591  if(log2>0) r+=r;
592  }
593 
594  for(i=log2-1; i>=0; i--){
595  v+= get_rac(c, state+31-i)<<i;
596  }
597 
598  return v;
599 }
600 
601 static inline void unpack_coeffs(SnowContext *s, SubBand *b, SubBand * parent, int orientation){
602  const int w= b->width;
603  const int h= b->height;
604  int x,y;
605 
606  int run, runs;
607  x_and_coeff *xc= b->x_coeff;
608  x_and_coeff *prev_xc= NULL;
609  x_and_coeff *prev2_xc= xc;
610  x_and_coeff *parent_xc= parent ? parent->x_coeff : NULL;
611  x_and_coeff *prev_parent_xc= parent_xc;
612 
613  runs= get_symbol2(&s->c, b->state[30], 0);
614  if(runs-- > 0) run= get_symbol2(&s->c, b->state[1], 3);
615  else run= INT_MAX;
616 
617  for(y=0; y<h; y++){
618  int v=0;
619  int lt=0, t=0, rt=0;
620 
621  if(y && prev_xc->x == 0){
622  rt= prev_xc->coeff;
623  }
624  for(x=0; x<w; x++){
625  int p=0;
626  const int l= v;
627 
628  lt= t; t= rt;
629 
630  if(y){
631  if(prev_xc->x <= x)
632  prev_xc++;
633  if(prev_xc->x == x + 1)
634  rt= prev_xc->coeff;
635  else
636  rt=0;
637  }
638  if(parent_xc){
639  if(x>>1 > parent_xc->x){
640  parent_xc++;
641  }
642  if(x>>1 == parent_xc->x){
643  p= parent_xc->coeff;
644  }
645  }
646  if(/*ll|*/l|lt|t|rt|p){
647  int context= av_log2(/*FFABS(ll) + */3*(l>>1) + (lt>>1) + (t&~1) + (rt>>1) + (p>>1));
648 
649  v=get_rac(&s->c, &b->state[0][context]);
650  if(v){
651  v= 2*(get_symbol2(&s->c, b->state[context + 2], context-4) + 1);
652  v+=get_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l&0xFF] + 3*ff_quant3bA[t&0xFF]]);
653 
654  xc->x=x;
655  (xc++)->coeff= v;
656  }
657  }else{
658  if(!run){
659  if(runs-- > 0) run= get_symbol2(&s->c, b->state[1], 3);
660  else run= INT_MAX;
661  v= 2*(get_symbol2(&s->c, b->state[0 + 2], 0-4) + 1);
662  v+=get_rac(&s->c, &b->state[0][16 + 1 + 3]);
663 
664  xc->x=x;
665  (xc++)->coeff= v;
666  }else{
667  int max_run;
668  run--;
669  v=0;
670  av_assert2(run >= 0);
671  if(y) max_run= FFMIN(run, prev_xc->x - x - 2);
672  else max_run= FFMIN(run, w-x-1);
673  if(parent_xc)
674  max_run= FFMIN(max_run, 2*parent_xc->x - x - 1);
675  av_assert2(max_run >= 0 && max_run <= run);
676 
677  x+= max_run;
678  run-= max_run;
679  }
680  }
681  }
682  (xc++)->x= w+1; //end marker
683  prev_xc= prev2_xc;
684  prev2_xc= xc;
685 
686  if(parent_xc){
687  if(y&1){
688  while(parent_xc->x != parent->width+1)
689  parent_xc++;
690  parent_xc++;
691  prev_parent_xc= parent_xc;
692  }else{
693  parent_xc= prev_parent_xc;
694  }
695  }
696  }
697 
698  (xc++)->x= w+1; //end marker
699 }
700 
701 #endif /* AVCODEC_SNOW_H */