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 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 
22 #ifndef AVCODEC_SNOW_H
23 #define AVCODEC_SNOW_H
24 
25 #include "dsputil.h"
26 #include "dwt.h"
27 
28 #include "rangecoder.h"
29 #include "mathops.h"
30 #include "mpegvideo.h"
31 
32 #define MID_STATE 128
33 
34 #define MAX_PLANES 4
35 #define QSHIFT 5
36 #define QROOT (1<<QSHIFT)
37 #define LOSSLESS_QLOG -128
38 #define FRAC_BITS 4
39 #define MAX_REF_FRAMES 8
40 
41 #define LOG2_OBMC_MAX 8
42 #define OBMC_MAX (1<<(LOG2_OBMC_MAX))
43 typedef struct BlockNode{
44  int16_t mx;
45  int16_t my;
49 //#define TYPE_SPLIT 1
50 #define BLOCK_INTRA 1
51 #define BLOCK_OPT 2
52 //#define TYPE_NOCOLOR 4
53  uint8_t level; //FIXME merge into type?
54 }BlockNode;
55 
56 static const BlockNode null_block= { //FIXME add border maybe
57  .color= {128,128,128},
58  .mx= 0,
59  .my= 0,
60  .ref= 0,
61  .type= 0,
62  .level= 0,
63 };
64 
65 #define LOG2_MB_SIZE 4
66 #define MB_SIZE (1<<LOG2_MB_SIZE)
67 #define ENCODER_EXTRA_BITS 4
68 #define HTAPS_MAX 8
69 
70 typedef struct x_and_coeff{
71  int16_t x;
72  uint16_t coeff;
73 } x_and_coeff;
74 
75 typedef struct SubBand{
76  int level;
77  int stride;
78  int width;
79  int height;
80  int qlog;
87  struct SubBand *parent;
88  uint8_t state[/*7*2*/ 7 + 512][32];
89 }SubBand;
90 
91 typedef struct Plane{
92  int width;
93  int height;
95 
96  int htaps;
97  int8_t hcoeff[HTAPS_MAX/2];
98  int diag_mc;
99  int fast_mc;
100 
104 }Plane;
105 
106 typedef struct SnowContext{
107  AVClass *class;
119 // uint8_t q_context[16];
121  uint8_t block_state[128 + 32*128];
122  int keyframe;
124  int version;
133  int16_t (*ref_mvs[MAX_REF_FRAMES])[2];
144  int qlog;
146  int lambda;
147  int lambda2;
148  int pass1_rc;
149  int mv_scale;
151  int qbias;
153 #define QBIAS_SHIFT 3
154  int b_width;
155  int b_height;
160 #define ME_CACHE_SIZE 1024
165 
166  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)
167 
170 }SnowContext;
171 
172 /* Tables */
173 extern const uint8_t * const ff_obmc_tab[4];
174 extern uint8_t ff_qexp[QROOT];
176 
177 /* C bits used by mmx/sse2/altivec */
178 
179 static av_always_inline void snow_interleave_line_header(int * i, int width, IDWTELEM * low, IDWTELEM * high){
180  (*i) = (width) - 2;
181 
182  if (width & 1){
183  low[(*i)+1] = low[((*i)+1)>>1];
184  (*i)--;
185  }
186 }
187 
188 static av_always_inline void snow_interleave_line_footer(int * i, IDWTELEM * low, IDWTELEM * high){
189  for (; (*i)>=0; (*i)-=2){
190  low[(*i)+1] = high[(*i)>>1];
191  low[*i] = low[(*i)>>1];
192  }
193 }
194 
195 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){
196  for(; i<w; i++){
197  dst[i] = src[i] - ((mul * (ref[i] + ref[i + 1]) + add) >> shift);
198  }
199 
200  if((width^lift_high)&1){
201  dst[w] = src[w] - ((mul * 2 * ref[w] + add) >> shift);
202  }
203 }
204 
205 static av_always_inline void snow_horizontal_compose_liftS_lead_out(int i, IDWTELEM * dst, IDWTELEM * src, IDWTELEM * ref, int width, int w){
206  for(; i<w; i++){
207  dst[i] = src[i] + ((ref[i] + ref[(i+1)]+W_BO + 4 * src[i]) >> W_BS);
208  }
209 
210  if(width&1){
211  dst[w] = src[w] + ((2 * ref[w] + W_BO + 4 * src[w]) >> W_BS);
212  }
213 }
214 
215 /* common code */
216 
224 void ff_snow_pred_block(SnowContext *s, uint8_t *dst, uint8_t *tmp, int stride,
225  int sx, int sy, int b_w, int b_h, BlockNode *block,
226  int plane_index, int w, int h);
227 /* common inline functions */
228 //XXX doublecheck all of them should stay inlined
229 
230 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){
231  const int w= s->b_width << s->block_max_depth;
232  const int rem_depth= s->block_max_depth - level;
233  const int index= (x + y*w) << rem_depth;
234  const int block_w= 1<<rem_depth;
236  int i,j;
237 
238  block.color[0]= l;
239  block.color[1]= cb;
240  block.color[2]= cr;
241  block.mx= mx;
242  block.my= my;
243  block.ref= ref;
244  block.type= type;
245  block.level= level;
246 
247  for(j=0; j<block_w; j++){
248  for(i=0; i<block_w; i++){
249  s->block[index + i + j*w]= block;
250  }
251  }
252 }
253 
254 static inline void pred_mv(SnowContext *s, int *mx, int *my, int ref,
255  const BlockNode *left, const BlockNode *top, const BlockNode *tr){
256  if(s->ref_frames == 1){
257  *mx = mid_pred(left->mx, top->mx, tr->mx);
258  *my = mid_pred(left->my, top->my, tr->my);
259  }else{
260  const int *scale = ff_scale_mv_ref[ref];
261  *mx = mid_pred((left->mx * scale[left->ref] + 128) >>8,
262  (top ->mx * scale[top ->ref] + 128) >>8,
263  (tr ->mx * scale[tr ->ref] + 128) >>8);
264  *my = mid_pred((left->my * scale[left->ref] + 128) >>8,
265  (top ->my * scale[top ->ref] + 128) >>8,
266  (tr ->my * scale[tr ->ref] + 128) >>8);
267  }
268 }
269 
271  if((a->type&BLOCK_INTRA) && (b->type&BLOCK_INTRA)){
272  return !((a->color[0] - b->color[0]) | (a->color[1] - b->color[1]) | (a->color[2] - b->color[2]));
273  }else{
274  return !((a->mx - b->mx) | (a->my - b->my) | (a->ref - b->ref) | ((a->type ^ b->type)&BLOCK_INTRA));
275  }
276 }
277 
278 //FIXME name cleanup (b_w, block_w, b_width stuff)
279 //XXX should we really inline it?
280 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){
281  const int b_width = s->b_width << s->block_max_depth;
282  const int b_height= s->b_height << s->block_max_depth;
283  const int b_stride= b_width;
284  BlockNode *lt= &s->block[b_x + b_y*b_stride];
285  BlockNode *rt= lt+1;
286  BlockNode *lb= lt+b_stride;
287  BlockNode *rb= lb+1;
288  uint8_t *block[4];
289  int tmp_step= src_stride >= 7*MB_SIZE ? MB_SIZE : MB_SIZE*src_stride;
290  uint8_t *tmp = s->scratchbuf;
291  uint8_t *ptmp;
292  int x,y;
293 
294  if(b_x<0){
295  lt= rt;
296  lb= rb;
297  }else if(b_x + 1 >= b_width){
298  rt= lt;
299  rb= lb;
300  }
301  if(b_y<0){
302  lt= lb;
303  rt= rb;
304  }else if(b_y + 1 >= b_height){
305  lb= lt;
306  rb= rt;
307  }
308 
309  if(src_x<0){ //FIXME merge with prev & always round internal width up to *16
310  obmc -= src_x;
311  b_w += src_x;
312  if(!sliced && !offset_dst)
313  dst -= src_x;
314  src_x=0;
315  }else if(src_x + b_w > w){
316  b_w = w - src_x;
317  }
318  if(src_y<0){
319  obmc -= src_y*obmc_stride;
320  b_h += src_y;
321  if(!sliced && !offset_dst)
322  dst -= src_y*dst_stride;
323  src_y=0;
324  }else if(src_y + b_h> h){
325  b_h = h - src_y;
326  }
327 
328  if(b_w<=0 || b_h<=0) return;
329 
330  assert(src_stride > 2*MB_SIZE + 5);
331 
332  if(!sliced && offset_dst)
333  dst += src_x + src_y*dst_stride;
334  dst8+= src_x + src_y*src_stride;
335 // src += src_x + src_y*src_stride;
336 
337  ptmp= tmp + 3*tmp_step;
338  block[0]= ptmp;
339  ptmp+=tmp_step;
340  ff_snow_pred_block(s, block[0], tmp, src_stride, src_x, src_y, b_w, b_h, lt, plane_index, w, h);
341 
342  if(same_block(lt, rt)){
343  block[1]= block[0];
344  }else{
345  block[1]= ptmp;
346  ptmp+=tmp_step;
347  ff_snow_pred_block(s, block[1], tmp, src_stride, src_x, src_y, b_w, b_h, rt, plane_index, w, h);
348  }
349 
350  if(same_block(lt, lb)){
351  block[2]= block[0];
352  }else if(same_block(rt, lb)){
353  block[2]= block[1];
354  }else{
355  block[2]= ptmp;
356  ptmp+=tmp_step;
357  ff_snow_pred_block(s, block[2], tmp, src_stride, src_x, src_y, b_w, b_h, lb, plane_index, w, h);
358  }
359 
360  if(same_block(lt, rb) ){
361  block[3]= block[0];
362  }else if(same_block(rt, rb)){
363  block[3]= block[1];
364  }else if(same_block(lb, rb)){
365  block[3]= block[2];
366  }else{
367  block[3]= ptmp;
368  ff_snow_pred_block(s, block[3], tmp, src_stride, src_x, src_y, b_w, b_h, rb, plane_index, w, h);
369  }
370  if(sliced){
371  s->dwt.inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
372  }else{
373  for(y=0; y<b_h; y++){
374  //FIXME ugly misuse of obmc_stride
375  const uint8_t *obmc1= obmc + y*obmc_stride;
376  const uint8_t *obmc2= obmc1+ (obmc_stride>>1);
377  const uint8_t *obmc3= obmc1+ obmc_stride*(obmc_stride>>1);
378  const uint8_t *obmc4= obmc3+ (obmc_stride>>1);
379  for(x=0; x<b_w; x++){
380  int v= obmc1[x] * block[3][x + y*src_stride]
381  +obmc2[x] * block[2][x + y*src_stride]
382  +obmc3[x] * block[1][x + y*src_stride]
383  +obmc4[x] * block[0][x + y*src_stride];
384 
385  v <<= 8 - LOG2_OBMC_MAX;
386  if(FRAC_BITS != 8){
387  v >>= 8 - FRAC_BITS;
388  }
389  if(add){
390  v += dst[x + y*dst_stride];
391  v = (v + (1<<(FRAC_BITS-1))) >> FRAC_BITS;
392  if(v&(~255)) v= ~(v>>31);
393  dst8[x + y*src_stride] = v;
394  }else{
395  dst[x + y*dst_stride] -= v;
396  }
397  }
398  }
399  }
400 }
401 
402 static av_always_inline void predict_slice(SnowContext *s, IDWTELEM *buf, int plane_index, int add, int mb_y){
403  Plane *p= &s->plane[plane_index];
404  const int mb_w= s->b_width << s->block_max_depth;
405  const int mb_h= s->b_height << s->block_max_depth;
406  int x, y, mb_x;
407  int block_size = MB_SIZE >> s->block_max_depth;
408  int block_w = plane_index ? block_size/2 : block_size;
409  const uint8_t *obmc = plane_index ? ff_obmc_tab[s->block_max_depth+1] : ff_obmc_tab[s->block_max_depth];
410  const int obmc_stride= plane_index ? block_size : 2*block_size;
411  int ref_stride= s->current_picture.linesize[plane_index];
412  uint8_t *dst8= s->current_picture.data[plane_index];
413  int w= p->width;
414  int h= p->height;
415 
416  if(s->keyframe || (s->avctx->debug&512)){
417  if(mb_y==mb_h)
418  return;
419 
420  if(add){
421  for(y=block_w*mb_y; y<FFMIN(h,block_w*(mb_y+1)); y++){
422  for(x=0; x<w; x++){
423  int v= buf[x + y*w] + (128<<FRAC_BITS) + (1<<(FRAC_BITS-1));
424  v >>= FRAC_BITS;
425  if(v&(~255)) v= ~(v>>31);
426  dst8[x + y*ref_stride]= v;
427  }
428  }
429  }else{
430  for(y=block_w*mb_y; y<FFMIN(h,block_w*(mb_y+1)); y++){
431  for(x=0; x<w; x++){
432  buf[x + y*w]-= 128<<FRAC_BITS;
433  }
434  }
435  }
436 
437  return;
438  }
439 
440  for(mb_x=0; mb_x<=mb_w; mb_x++){
441  add_yblock(s, 0, NULL, buf, dst8, obmc,
442  block_w*mb_x - block_w/2,
443  block_w*mb_y - block_w/2,
444  block_w, block_w,
445  w, h,
446  w, ref_stride, obmc_stride,
447  mb_x - 1, mb_y - 1,
448  add, 1, plane_index);
449  }
450 }
451 
452 static av_always_inline void predict_plane(SnowContext *s, IDWTELEM *buf, int plane_index, int add){
453  const int mb_h= s->b_height << s->block_max_depth;
454  int mb_y;
455  for(mb_y=0; mb_y<=mb_h; mb_y++)
456  predict_slice(s, buf, plane_index, add, mb_y);
457 }
458 
459 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){
460  const int w= s->b_width << s->block_max_depth;
461  const int rem_depth= s->block_max_depth - level;
462  const int index= (x + y*w) << rem_depth;
463  const int block_w= 1<<rem_depth;
465  int i,j;
466 
467  block.color[0]= l;
468  block.color[1]= cb;
469  block.color[2]= cr;
470  block.mx= mx;
471  block.my= my;
472  block.ref= ref;
473  block.type= type;
474  block.level= level;
475 
476  for(j=0; j<block_w; j++){
477  for(i=0; i<block_w; i++){
478  s->block[index + i + j*w]= block;
479  }
480  }
481 }
482 
483 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){
484  const int offset[3]= {
485  y*c-> stride + x,
486  ((y*c->uvstride + x)>>1),
487  ((y*c->uvstride + x)>>1),
488  };
489  int i;
490  for(i=0; i<3; i++){
491  c->src[0][i]= src [i];
492  c->ref[0][i]= ref [i] + offset[i];
493  }
494  assert(!ref_index);
495 }
496 
497 
498 /* bitstream functions */
499 
500 extern const int8_t ff_quant3bA[256];
501 
502 #define QEXPSHIFT (7-FRAC_BITS+8) //FIXME try to change this to 0
503 
504 static inline void put_symbol(RangeCoder *c, uint8_t *state, int v, int is_signed){
505  int i;
506 
507  if(v){
508  const int a= FFABS(v);
509  const int e= av_log2(a);
510  const int el= FFMIN(e, 10);
511  put_rac(c, state+0, 0);
512 
513  for(i=0; i<el; i++){
514  put_rac(c, state+1+i, 1); //1..10
515  }
516  for(; i<e; i++){
517  put_rac(c, state+1+9, 1); //1..10
518  }
519  put_rac(c, state+1+FFMIN(i,9), 0);
520 
521  for(i=e-1; i>=el; i--){
522  put_rac(c, state+22+9, (a>>i)&1); //22..31
523  }
524  for(; i>=0; i--){
525  put_rac(c, state+22+i, (a>>i)&1); //22..31
526  }
527 
528  if(is_signed)
529  put_rac(c, state+11 + el, v < 0); //11..21
530  }else{
531  put_rac(c, state+0, 1);
532  }
533 }
534 
535 static inline int get_symbol(RangeCoder *c, uint8_t *state, int is_signed){
536  if(get_rac(c, state+0))
537  return 0;
538  else{
539  int i, e, a;
540  e= 0;
541  while(get_rac(c, state+1 + FFMIN(e,9))){ //1..10
542  e++;
543  }
544 
545  a= 1;
546  for(i=e-1; i>=0; i--){
547  a += a + get_rac(c, state+22 + FFMIN(i,9)); //22..31
548  }
549 
550  e= -(is_signed && get_rac(c, state+11 + FFMIN(e,10))); //11..21
551  return (a^e)-e;
552  }
553 }
554 
555 static inline void put_symbol2(RangeCoder *c, uint8_t *state, int v, int log2){
556  int i;
557  int r= log2>=0 ? 1<<log2 : 1;
558 
559  assert(v>=0);
560  assert(log2>=-4);
561 
562  while(v >= r){
563  put_rac(c, state+4+log2, 1);
564  v -= r;
565  log2++;
566  if(log2>0) r+=r;
567  }
568  put_rac(c, state+4+log2, 0);
569 
570  for(i=log2-1; i>=0; i--){
571  put_rac(c, state+31-i, (v>>i)&1);
572  }
573 }
574 
575 static inline int get_symbol2(RangeCoder *c, uint8_t *state, int log2){
576  int i;
577  int r= log2>=0 ? 1<<log2 : 1;
578  int v=0;
579 
580  assert(log2>=-4);
581 
582  while(get_rac(c, state+4+log2)){
583  v+= r;
584  log2++;
585  if(log2>0) r+=r;
586  }
587 
588  for(i=log2-1; i>=0; i--){
589  v+= get_rac(c, state+31-i)<<i;
590  }
591 
592  return v;
593 }
594 
595 static inline void unpack_coeffs(SnowContext *s, SubBand *b, SubBand * parent, int orientation){
596  const int w= b->width;
597  const int h= b->height;
598  int x,y;
599 
600  int run, runs;
601  x_and_coeff *xc= b->x_coeff;
602  x_and_coeff *prev_xc= NULL;
603  x_and_coeff *prev2_xc= xc;
604  x_and_coeff *parent_xc= parent ? parent->x_coeff : NULL;
605  x_and_coeff *prev_parent_xc= parent_xc;
606 
607  runs= get_symbol2(&s->c, b->state[30], 0);
608  if(runs-- > 0) run= get_symbol2(&s->c, b->state[1], 3);
609  else run= INT_MAX;
610 
611  for(y=0; y<h; y++){
612  int v=0;
613  int lt=0, t=0, rt=0;
614 
615  if(y && prev_xc->x == 0){
616  rt= prev_xc->coeff;
617  }
618  for(x=0; x<w; x++){
619  int p=0;
620  const int l= v;
621 
622  lt= t; t= rt;
623 
624  if(y){
625  if(prev_xc->x <= x)
626  prev_xc++;
627  if(prev_xc->x == x + 1)
628  rt= prev_xc->coeff;
629  else
630  rt=0;
631  }
632  if(parent_xc){
633  if(x>>1 > parent_xc->x){
634  parent_xc++;
635  }
636  if(x>>1 == parent_xc->x){
637  p= parent_xc->coeff;
638  }
639  }
640  if(/*ll|*/l|lt|t|rt|p){
641  int context= av_log2(/*FFABS(ll) + */3*(l>>1) + (lt>>1) + (t&~1) + (rt>>1) + (p>>1));
642 
643  v=get_rac(&s->c, &b->state[0][context]);
644  if(v){
645  v= 2*(get_symbol2(&s->c, b->state[context + 2], context-4) + 1);
646  v+=get_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l&0xFF] + 3*ff_quant3bA[t&0xFF]]);
647 
648  xc->x=x;
649  (xc++)->coeff= v;
650  }
651  }else{
652  if(!run){
653  if(runs-- > 0) run= get_symbol2(&s->c, b->state[1], 3);
654  else run= INT_MAX;
655  v= 2*(get_symbol2(&s->c, b->state[0 + 2], 0-4) + 1);
656  v+=get_rac(&s->c, &b->state[0][16 + 1 + 3]);
657 
658  xc->x=x;
659  (xc++)->coeff= v;
660  }else{
661  int max_run;
662  run--;
663  v=0;
664 
665  if(y) max_run= FFMIN(run, prev_xc->x - x - 2);
666  else max_run= FFMIN(run, w-x-1);
667  if(parent_xc)
668  max_run= FFMIN(max_run, 2*parent_xc->x - x - 1);
669  x+= max_run;
670  run-= max_run;
671  }
672  }
673  }
674  (xc++)->x= w+1; //end marker
675  prev_xc= prev2_xc;
676  prev2_xc= xc;
677 
678  if(parent_xc){
679  if(y&1){
680  while(parent_xc->x != parent->width+1)
681  parent_xc++;
682  parent_xc++;
683  prev_parent_xc= parent_xc;
684  }else{
685  parent_xc= prev_parent_xc;
686  }
687  }
688  }
689 
690  (xc++)->x= w+1; //end marker
691 }
692 
693 #endif /* AVCODEC_SNOW_H */