FFmpeg
 All Data Structures Namespaces 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 "hpeldsp.h"
27 #include "snow_dwt.h"
28 
29 #include "rangecoder.h"
30 #include "mathops.h"
31 #include "mpegvideo.h"
32 #include "h264qpel.h"
33 
34 #define MID_STATE 128
35 
36 #define MAX_PLANES 4
37 #define QSHIFT 5
38 #define QROOT (1<<QSHIFT)
39 #define LOSSLESS_QLOG -128
40 #define FRAC_BITS 4
41 #define MAX_REF_FRAMES 8
42 
43 #define LOG2_OBMC_MAX 8
44 #define OBMC_MAX (1<<(LOG2_OBMC_MAX))
45 typedef struct BlockNode{
46  int16_t mx;
47  int16_t my;
51 //#define TYPE_SPLIT 1
52 #define BLOCK_INTRA 1
53 #define BLOCK_OPT 2
54 //#define TYPE_NOCOLOR 4
55  uint8_t level; //FIXME merge into type?
56 }BlockNode;
57 
58 static const BlockNode null_block= { //FIXME add border maybe
59  .color= {128,128,128},
60  .mx= 0,
61  .my= 0,
62  .ref= 0,
63  .type= 0,
64  .level= 0,
65 };
66 
67 #define LOG2_MB_SIZE 4
68 #define MB_SIZE (1<<LOG2_MB_SIZE)
69 #define ENCODER_EXTRA_BITS 4
70 #define HTAPS_MAX 8
71 
72 typedef struct x_and_coeff{
73  int16_t x;
74  uint16_t coeff;
75 } x_and_coeff;
76 
77 typedef struct SubBand{
78  int level;
79  int stride;
80  int width;
81  int height;
82  int qlog; ///< log(qscale)/log[2^(1/6)]
84  IDWTELEM *ibuf;
87  int stride_line; ///< Stride measured in lines, not pixels.
89  struct SubBand *parent;
90  uint8_t state[/*7*2*/ 7 + 512][32];
91 }SubBand;
92 
93 typedef struct Plane{
94  int width;
95  int height;
97 
98  int htaps;
99  int8_t hcoeff[HTAPS_MAX/2];
100  int diag_mc;
101  int fast_mc;
102 
106 }Plane;
107 
108 typedef struct SnowContext{
109  AVClass *class;
118  AVFrame *input_picture; ///< new_picture with the internal linesizes
123 // uint8_t q_context[16];
125  uint8_t block_state[128 + 32*128];
126  int keyframe;
128  int version;
137  int16_t (*ref_mvs[MAX_REF_FRAMES])[2];
148  int qlog;
150  int lambda;
151  int lambda2;
152  int pass1_rc;
153  int mv_scale;
155  int qbias;
157 #define QBIAS_SHIFT 3
158  int b_width;
159  int b_height;
165 #define ME_CACHE_SIZE 1024
171 
172  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)
173 
176 }SnowContext;
177 
178 /* Tables */
179 extern const uint8_t * const ff_obmc_tab[4];
180 extern uint8_t ff_qexp[QROOT];
182 
183 /* C bits used by mmx/sse2/altivec */
184 
185 static av_always_inline void snow_interleave_line_header(int * i, int width, IDWTELEM * low, IDWTELEM * high){
186  (*i) = (width) - 2;
187 
188  if (width & 1){
189  low[(*i)+1] = low[((*i)+1)>>1];
190  (*i)--;
191  }
192 }
193 
194 static av_always_inline void snow_interleave_line_footer(int * i, IDWTELEM * low, IDWTELEM * high){
195  for (; (*i)>=0; (*i)-=2){
196  low[(*i)+1] = high[(*i)>>1];
197  low[*i] = low[(*i)>>1];
198  }
199 }
200 
201 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){
202  for(; i<w; i++){
203  dst[i] = src[i] - ((mul * (ref[i] + ref[i + 1]) + add) >> shift);
204  }
205 
206  if((width^lift_high)&1){
207  dst[w] = src[w] - ((mul * 2 * ref[w] + add) >> shift);
208  }
209 }
210 
212  for(; i<w; i++){
213  dst[i] = src[i] + ((ref[i] + ref[(i+1)]+W_BO + 4 * src[i]) >> W_BS);
214  }
215 
216  if(width&1){
217  dst[w] = src[w] + ((2 * ref[w] + W_BO + 4 * src[w]) >> W_BS);
218  }
219 }
220 
221 /* common code */
222 
230 void ff_snow_pred_block(SnowContext *s, uint8_t *dst, uint8_t *tmp, ptrdiff_t stride,
231  int sx, int sy, int b_w, int b_h, BlockNode *block,
232  int plane_index, int w, int h);
233 /* common inline functions */
234 //XXX doublecheck all of them should stay inlined
235 
236 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){
237  const int w= s->b_width << s->block_max_depth;
238  const int rem_depth= s->block_max_depth - level;
239  const int index= (x + y*w) << rem_depth;
240  const int block_w= 1<<rem_depth;
242  int i,j;
243 
244  block.color[0]= l;
245  block.color[1]= cb;
246  block.color[2]= cr;
247  block.mx= mx;
248  block.my= my;
249  block.ref= ref;
250  block.type= type;
251  block.level= level;
252 
253  for(j=0; j<block_w; j++){
254  for(i=0; i<block_w; i++){
255  s->block[index + i + j*w]= block;
256  }
257  }
258 }
259 
260 static inline void pred_mv(SnowContext *s, int *mx, int *my, int ref,
261  const BlockNode *left, const BlockNode *top, const BlockNode *tr){
262  if(s->ref_frames == 1){
263  *mx = mid_pred(left->mx, top->mx, tr->mx);
264  *my = mid_pred(left->my, top->my, tr->my);
265  }else{
266  const int *scale = ff_scale_mv_ref[ref];
267  *mx = mid_pred((left->mx * scale[left->ref] + 128) >>8,
268  (top ->mx * scale[top ->ref] + 128) >>8,
269  (tr ->mx * scale[tr ->ref] + 128) >>8);
270  *my = mid_pred((left->my * scale[left->ref] + 128) >>8,
271  (top ->my * scale[top ->ref] + 128) >>8,
272  (tr ->my * scale[tr ->ref] + 128) >>8);
273  }
274 }
275 
277  if((a->type&BLOCK_INTRA) && (b->type&BLOCK_INTRA)){
278  return !((a->color[0] - b->color[0]) | (a->color[1] - b->color[1]) | (a->color[2] - b->color[2]));
279  }else{
280  return !((a->mx - b->mx) | (a->my - b->my) | (a->ref - b->ref) | ((a->type ^ b->type)&BLOCK_INTRA));
281  }
282 }
283 
284 //FIXME name cleanup (b_w, block_w, b_width stuff)
285 //XXX should we really inline it?
286 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){
287  const int b_width = s->b_width << s->block_max_depth;
288  const int b_height= s->b_height << s->block_max_depth;
289  const int b_stride= b_width;
290  BlockNode *lt= &s->block[b_x + b_y*b_stride];
291  BlockNode *rt= lt+1;
292  BlockNode *lb= lt+b_stride;
293  BlockNode *rb= lb+1;
294  uint8_t *block[4];
295  int tmp_step= src_stride >= 7*MB_SIZE ? MB_SIZE : MB_SIZE*src_stride;
296  uint8_t *tmp = s->scratchbuf;
297  uint8_t *ptmp;
298  int x,y;
299 
300  if(b_x<0){
301  lt= rt;
302  lb= rb;
303  }else if(b_x + 1 >= b_width){
304  rt= lt;
305  rb= lb;
306  }
307  if(b_y<0){
308  lt= lb;
309  rt= rb;
310  }else if(b_y + 1 >= b_height){
311  lb= lt;
312  rb= rt;
313  }
314 
315  if(src_x<0){ //FIXME merge with prev & always round internal width up to *16
316  obmc -= src_x;
317  b_w += src_x;
318  if(!sliced && !offset_dst)
319  dst -= src_x;
320  src_x=0;
321  }else if(src_x + b_w > w){
322  b_w = w - src_x;
323  }
324  if(src_y<0){
325  obmc -= src_y*obmc_stride;
326  b_h += src_y;
327  if(!sliced && !offset_dst)
328  dst -= src_y*dst_stride;
329  src_y=0;
330  }else if(src_y + b_h> h){
331  b_h = h - src_y;
332  }
333 
334  if(b_w<=0 || b_h<=0) return;
335 
336  av_assert2(src_stride > 2*MB_SIZE + 5);
337 
338  if(!sliced && offset_dst)
339  dst += src_x + src_y*dst_stride;
340  dst8+= src_x + src_y*src_stride;
341 // src += src_x + src_y*src_stride;
342 
343  ptmp= tmp + 3*tmp_step;
344  block[0]= ptmp;
345  ptmp+=tmp_step;
346  ff_snow_pred_block(s, block[0], tmp, src_stride, src_x, src_y, b_w, b_h, lt, plane_index, w, h);
347 
348  if(same_block(lt, rt)){
349  block[1]= block[0];
350  }else{
351  block[1]= ptmp;
352  ptmp+=tmp_step;
353  ff_snow_pred_block(s, block[1], tmp, src_stride, src_x, src_y, b_w, b_h, rt, plane_index, w, h);
354  }
355 
356  if(same_block(lt, lb)){
357  block[2]= block[0];
358  }else if(same_block(rt, lb)){
359  block[2]= block[1];
360  }else{
361  block[2]= ptmp;
362  ptmp+=tmp_step;
363  ff_snow_pred_block(s, block[2], tmp, src_stride, src_x, src_y, b_w, b_h, lb, plane_index, w, h);
364  }
365 
366  if(same_block(lt, rb) ){
367  block[3]= block[0];
368  }else if(same_block(rt, rb)){
369  block[3]= block[1];
370  }else if(same_block(lb, rb)){
371  block[3]= block[2];
372  }else{
373  block[3]= ptmp;
374  ff_snow_pred_block(s, block[3], tmp, src_stride, src_x, src_y, b_w, b_h, rb, plane_index, w, h);
375  }
376  if(sliced){
377  s->dwt.inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
378  }else{
379  for(y=0; y<b_h; y++){
380  //FIXME ugly misuse of obmc_stride
381  const uint8_t *obmc1= obmc + y*obmc_stride;
382  const uint8_t *obmc2= obmc1+ (obmc_stride>>1);
383  const uint8_t *obmc3= obmc1+ obmc_stride*(obmc_stride>>1);
384  const uint8_t *obmc4= obmc3+ (obmc_stride>>1);
385  for(x=0; x<b_w; x++){
386  int v= obmc1[x] * block[3][x + y*src_stride]
387  +obmc2[x] * block[2][x + y*src_stride]
388  +obmc3[x] * block[1][x + y*src_stride]
389  +obmc4[x] * block[0][x + y*src_stride];
390 
391  v <<= 8 - LOG2_OBMC_MAX;
392  if(FRAC_BITS != 8){
393  v >>= 8 - FRAC_BITS;
394  }
395  if(add){
396  v += dst[x + y*dst_stride];
397  v = (v + (1<<(FRAC_BITS-1))) >> FRAC_BITS;
398  if(v&(~255)) v= ~(v>>31);
399  dst8[x + y*src_stride] = v;
400  }else{
401  dst[x + y*dst_stride] -= v;
402  }
403  }
404  }
405  }
406 }
407 
408 static av_always_inline void predict_slice(SnowContext *s, IDWTELEM *buf, int plane_index, int add, int mb_y){
409  Plane *p= &s->plane[plane_index];
410  const int mb_w= s->b_width << s->block_max_depth;
411  const int mb_h= s->b_height << s->block_max_depth;
412  int x, y, mb_x;
413  int block_size = MB_SIZE >> s->block_max_depth;
414  int block_w = plane_index ? block_size>>s->chroma_h_shift : block_size;
415  int block_h = plane_index ? block_size>>s->chroma_v_shift : block_size;
416  const uint8_t *obmc = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
417  const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
418  int ref_stride= s->current_picture->linesize[plane_index];
419  uint8_t *dst8= s->current_picture->data[plane_index];
420  int w= p->width;
421  int h= p->height;
422  av_assert2(s->chroma_h_shift == s->chroma_v_shift); // obmc params assume squares
423  if(s->keyframe || (s->avctx->debug&512)){
424  if(mb_y==mb_h)
425  return;
426 
427  if(add){
428  for(y=block_h*mb_y; y<FFMIN(h,block_h*(mb_y+1)); y++){
429  for(x=0; x<w; x++){
430  int v= buf[x + y*w] + (128<<FRAC_BITS) + (1<<(FRAC_BITS-1));
431  v >>= FRAC_BITS;
432  if(v&(~255)) v= ~(v>>31);
433  dst8[x + y*ref_stride]= v;
434  }
435  }
436  }else{
437  for(y=block_h*mb_y; y<FFMIN(h,block_h*(mb_y+1)); y++){
438  for(x=0; x<w; x++){
439  buf[x + y*w]-= 128<<FRAC_BITS;
440  }
441  }
442  }
443 
444  return;
445  }
446 
447  for(mb_x=0; mb_x<=mb_w; mb_x++){
448  add_yblock(s, 0, NULL, buf, dst8, obmc,
449  block_w*mb_x - block_w/2,
450  block_h*mb_y - block_h/2,
451  block_w, block_h,
452  w, h,
453  w, ref_stride, obmc_stride,
454  mb_x - 1, mb_y - 1,
455  add, 1, plane_index);
456  }
457 }
458 
459 static av_always_inline void predict_plane(SnowContext *s, IDWTELEM *buf, int plane_index, int add){
460  const int mb_h= s->b_height << s->block_max_depth;
461  int mb_y;
462  for(mb_y=0; mb_y<=mb_h; mb_y++)
463  predict_slice(s, buf, plane_index, add, mb_y);
464 }
465 
466 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){
467  const int w= s->b_width << s->block_max_depth;
468  const int rem_depth= s->block_max_depth - level;
469  const int index= (x + y*w) << rem_depth;
470  const int block_w= 1<<rem_depth;
471  const int block_h= 1<<rem_depth; //FIXME "w!=h"
473  int i,j;
474 
475  block.color[0]= l;
476  block.color[1]= cb;
477  block.color[2]= cr;
478  block.mx= mx;
479  block.my= my;
480  block.ref= ref;
481  block.type= type;
482  block.level= level;
483 
484  for(j=0; j<block_h; j++){
485  for(i=0; i<block_w; i++){
486  s->block[index + i + j*w]= block;
487  }
488  }
489 }
490 
491 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){
492  SnowContext *s = c->avctx->priv_data;
493  const int offset[3]= {
494  y*c-> stride + x,
495  ((y*c->uvstride + x)>>s->chroma_h_shift),
496  ((y*c->uvstride + x)>>s->chroma_h_shift),
497  };
498  int i;
499  for(i=0; i<3; i++){
500  c->src[0][i]= src [i];
501  c->ref[0][i]= ref [i] + offset[i];
502  }
503  av_assert2(!ref_index);
504 }
505 
506 
507 /* bitstream functions */
508 
509 extern const int8_t ff_quant3bA[256];
510 
511 #define QEXPSHIFT (7-FRAC_BITS+8) //FIXME try to change this to 0
512 
513 static inline void put_symbol(RangeCoder *c, uint8_t *state, int v, int is_signed){
514  int i;
515 
516  if(v){
517  const int a= FFABS(v);
518  const int e= av_log2(a);
519  const int el= FFMIN(e, 10);
520  put_rac(c, state+0, 0);
521 
522  for(i=0; i<el; i++){
523  put_rac(c, state+1+i, 1); //1..10
524  }
525  for(; i<e; i++){
526  put_rac(c, state+1+9, 1); //1..10
527  }
528  put_rac(c, state+1+FFMIN(i,9), 0);
529 
530  for(i=e-1; i>=el; i--){
531  put_rac(c, state+22+9, (a>>i)&1); //22..31
532  }
533  for(; i>=0; i--){
534  put_rac(c, state+22+i, (a>>i)&1); //22..31
535  }
536 
537  if(is_signed)
538  put_rac(c, state+11 + el, v < 0); //11..21
539  }else{
540  put_rac(c, state+0, 1);
541  }
542 }
543 
544 static inline int get_symbol(RangeCoder *c, uint8_t *state, int is_signed){
545  if(get_rac(c, state+0))
546  return 0;
547  else{
548  int i, e, a;
549  e= 0;
550  while(get_rac(c, state+1 + FFMIN(e,9))){ //1..10
551  e++;
552  }
553 
554  a= 1;
555  for(i=e-1; i>=0; i--){
556  a += a + get_rac(c, state+22 + FFMIN(i,9)); //22..31
557  }
558 
559  e= -(is_signed && get_rac(c, state+11 + FFMIN(e,10))); //11..21
560  return (a^e)-e;
561  }
562 }
563 
564 static inline void put_symbol2(RangeCoder *c, uint8_t *state, int v, int log2){
565  int i;
566  int r= log2>=0 ? 1<<log2 : 1;
567 
568  av_assert2(v>=0);
569  av_assert2(log2>=-4);
570 
571  while(v >= r){
572  put_rac(c, state+4+log2, 1);
573  v -= r;
574  log2++;
575  if(log2>0) r+=r;
576  }
577  put_rac(c, state+4+log2, 0);
578 
579  for(i=log2-1; i>=0; i--){
580  put_rac(c, state+31-i, (v>>i)&1);
581  }
582 }
583 
584 static inline int get_symbol2(RangeCoder *c, uint8_t *state, int log2){
585  int i;
586  int r= log2>=0 ? 1<<log2 : 1;
587  int v=0;
588 
589  av_assert2(log2>=-4);
590 
591  while(log2<28 && get_rac(c, state+4+log2)){
592  v+= r;
593  log2++;
594  if(log2>0) r+=r;
595  }
596 
597  for(i=log2-1; i>=0; i--){
598  v+= get_rac(c, state+31-i)<<i;
599  }
600 
601  return v;
602 }
603 
604 static inline void unpack_coeffs(SnowContext *s, SubBand *b, SubBand * parent, int orientation){
605  const int w= b->width;
606  const int h= b->height;
607  int x,y;
608 
609  int run, runs;
610  x_and_coeff *xc= b->x_coeff;
611  x_and_coeff *prev_xc= NULL;
612  x_and_coeff *prev2_xc= xc;
613  x_and_coeff *parent_xc= parent ? parent->x_coeff : NULL;
614  x_and_coeff *prev_parent_xc= parent_xc;
615 
616  runs= get_symbol2(&s->c, b->state[30], 0);
617  if(runs-- > 0) run= get_symbol2(&s->c, b->state[1], 3);
618  else run= INT_MAX;
619 
620  for(y=0; y<h; y++){
621  int v=0;
622  int lt=0, t=0, rt=0;
623 
624  if(y && prev_xc->x == 0){
625  rt= prev_xc->coeff;
626  }
627  for(x=0; x<w; x++){
628  int p=0;
629  const int l= v;
630 
631  lt= t; t= rt;
632 
633  if(y){
634  if(prev_xc->x <= x)
635  prev_xc++;
636  if(prev_xc->x == x + 1)
637  rt= prev_xc->coeff;
638  else
639  rt=0;
640  }
641  if(parent_xc){
642  if(x>>1 > parent_xc->x){
643  parent_xc++;
644  }
645  if(x>>1 == parent_xc->x){
646  p= parent_xc->coeff;
647  }
648  }
649  if(/*ll|*/l|lt|t|rt|p){
650  int context= av_log2(/*FFABS(ll) + */3*(l>>1) + (lt>>1) + (t&~1) + (rt>>1) + (p>>1));
651 
652  v=get_rac(&s->c, &b->state[0][context]);
653  if(v){
654  v= 2*(get_symbol2(&s->c, b->state[context + 2], context-4) + 1);
655  v+=get_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l&0xFF] + 3*ff_quant3bA[t&0xFF]]);
656 
657  xc->x=x;
658  (xc++)->coeff= v;
659  }
660  }else{
661  if(!run){
662  if(runs-- > 0) run= get_symbol2(&s->c, b->state[1], 3);
663  else run= INT_MAX;
664  v= 2*(get_symbol2(&s->c, b->state[0 + 2], 0-4) + 1);
665  v+=get_rac(&s->c, &b->state[0][16 + 1 + 3]);
666 
667  xc->x=x;
668  (xc++)->coeff= v;
669  }else{
670  int max_run;
671  run--;
672  v=0;
673  av_assert2(run >= 0);
674  if(y) max_run= FFMIN(run, prev_xc->x - x - 2);
675  else max_run= FFMIN(run, w-x-1);
676  if(parent_xc)
677  max_run= FFMIN(max_run, 2*parent_xc->x - x - 1);
678  av_assert2(max_run >= 0 && max_run <= run);
679 
680  x+= max_run;
681  run-= max_run;
682  }
683  }
684  }
685  (xc++)->x= w+1; //end marker
686  prev_xc= prev2_xc;
687  prev2_xc= xc;
688 
689  if(parent_xc){
690  if(y&1){
691  while(parent_xc->x != parent->width+1)
692  parent_xc++;
693  parent_xc++;
694  prev_parent_xc= parent_xc;
695  }else{
696  parent_xc= prev_parent_xc;
697  }
698  }
699  }
700 
701  (xc++)->x= w+1; //end marker
702 }
703 
704 #endif /* AVCODEC_SNOW_H */