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;
164 #define ME_CACHE_SIZE 1024
170 
171  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)
172 
175 }SnowContext;
176 
177 /* Tables */
178 extern const uint8_t * const ff_obmc_tab[4];
179 extern uint8_t ff_qexp[QROOT];
181 
182 /* C bits used by mmx/sse2/altivec */
183 
184 static av_always_inline void snow_interleave_line_header(int * i, int width, IDWTELEM * low, IDWTELEM * high){
185  (*i) = (width) - 2;
186 
187  if (width & 1){
188  low[(*i)+1] = low[((*i)+1)>>1];
189  (*i)--;
190  }
191 }
192 
193 static av_always_inline void snow_interleave_line_footer(int * i, IDWTELEM * low, IDWTELEM * high){
194  for (; (*i)>=0; (*i)-=2){
195  low[(*i)+1] = high[(*i)>>1];
196  low[*i] = low[(*i)>>1];
197  }
198 }
199 
200 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){
201  for(; i<w; i++){
202  dst[i] = src[i] - ((mul * (ref[i] + ref[i + 1]) + add) >> shift);
203  }
204 
205  if((width^lift_high)&1){
206  dst[w] = src[w] - ((mul * 2 * ref[w] + add) >> shift);
207  }
208 }
209 
211  for(; i<w; i++){
212  dst[i] = src[i] + ((ref[i] + ref[(i+1)]+W_BO + 4 * src[i]) >> W_BS);
213  }
214 
215  if(width&1){
216  dst[w] = src[w] + ((2 * ref[w] + W_BO + 4 * src[w]) >> W_BS);
217  }
218 }
219 
220 /* common code */
221 
229 void ff_snow_pred_block(SnowContext *s, uint8_t *dst, uint8_t *tmp, int stride,
230  int sx, int sy, int b_w, int b_h, BlockNode *block,
231  int plane_index, int w, int h);
232 /* common inline functions */
233 //XXX doublecheck all of them should stay inlined
234 
235 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){
236  const int w= s->b_width << s->block_max_depth;
237  const int rem_depth= s->block_max_depth - level;
238  const int index= (x + y*w) << rem_depth;
239  const int block_w= 1<<rem_depth;
241  int i,j;
242 
243  block.color[0]= l;
244  block.color[1]= cb;
245  block.color[2]= cr;
246  block.mx= mx;
247  block.my= my;
248  block.ref= ref;
249  block.type= type;
250  block.level= level;
251 
252  for(j=0; j<block_w; j++){
253  for(i=0; i<block_w; i++){
254  s->block[index + i + j*w]= block;
255  }
256  }
257 }
258 
259 static inline void pred_mv(SnowContext *s, int *mx, int *my, int ref,
260  const BlockNode *left, const BlockNode *top, const BlockNode *tr){
261  if(s->ref_frames == 1){
262  *mx = mid_pred(left->mx, top->mx, tr->mx);
263  *my = mid_pred(left->my, top->my, tr->my);
264  }else{
265  const int *scale = ff_scale_mv_ref[ref];
266  *mx = mid_pred((left->mx * scale[left->ref] + 128) >>8,
267  (top ->mx * scale[top ->ref] + 128) >>8,
268  (tr ->mx * scale[tr ->ref] + 128) >>8);
269  *my = mid_pred((left->my * scale[left->ref] + 128) >>8,
270  (top ->my * scale[top ->ref] + 128) >>8,
271  (tr ->my * scale[tr ->ref] + 128) >>8);
272  }
273 }
274 
276  if((a->type&BLOCK_INTRA) && (b->type&BLOCK_INTRA)){
277  return !((a->color[0] - b->color[0]) | (a->color[1] - b->color[1]) | (a->color[2] - b->color[2]));
278  }else{
279  return !((a->mx - b->mx) | (a->my - b->my) | (a->ref - b->ref) | ((a->type ^ b->type)&BLOCK_INTRA));
280  }
281 }
282 
283 //FIXME name cleanup (b_w, block_w, b_width stuff)
284 //XXX should we really inline it?
285 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){
286  const int b_width = s->b_width << s->block_max_depth;
287  const int b_height= s->b_height << s->block_max_depth;
288  const int b_stride= b_width;
289  BlockNode *lt= &s->block[b_x + b_y*b_stride];
290  BlockNode *rt= lt+1;
291  BlockNode *lb= lt+b_stride;
292  BlockNode *rb= lb+1;
293  uint8_t *block[4];
294  int tmp_step= src_stride >= 7*MB_SIZE ? MB_SIZE : MB_SIZE*src_stride;
295  uint8_t *tmp = s->scratchbuf;
296  uint8_t *ptmp;
297  int x,y;
298 
299  if(b_x<0){
300  lt= rt;
301  lb= rb;
302  }else if(b_x + 1 >= b_width){
303  rt= lt;
304  rb= lb;
305  }
306  if(b_y<0){
307  lt= lb;
308  rt= rb;
309  }else if(b_y + 1 >= b_height){
310  lb= lt;
311  rb= rt;
312  }
313 
314  if(src_x<0){ //FIXME merge with prev & always round internal width up to *16
315  obmc -= src_x;
316  b_w += src_x;
317  if(!sliced && !offset_dst)
318  dst -= src_x;
319  src_x=0;
320  }else if(src_x + b_w > w){
321  b_w = w - src_x;
322  }
323  if(src_y<0){
324  obmc -= src_y*obmc_stride;
325  b_h += src_y;
326  if(!sliced && !offset_dst)
327  dst -= src_y*dst_stride;
328  src_y=0;
329  }else if(src_y + b_h> h){
330  b_h = h - src_y;
331  }
332 
333  if(b_w<=0 || b_h<=0) return;
334 
335  av_assert2(src_stride > 2*MB_SIZE + 5);
336 
337  if(!sliced && offset_dst)
338  dst += src_x + src_y*dst_stride;
339  dst8+= src_x + src_y*src_stride;
340 // src += src_x + src_y*src_stride;
341 
342  ptmp= tmp + 3*tmp_step;
343  block[0]= ptmp;
344  ptmp+=tmp_step;
345  ff_snow_pred_block(s, block[0], tmp, src_stride, src_x, src_y, b_w, b_h, lt, plane_index, w, h);
346 
347  if(same_block(lt, rt)){
348  block[1]= block[0];
349  }else{
350  block[1]= ptmp;
351  ptmp+=tmp_step;
352  ff_snow_pred_block(s, block[1], tmp, src_stride, src_x, src_y, b_w, b_h, rt, plane_index, w, h);
353  }
354 
355  if(same_block(lt, lb)){
356  block[2]= block[0];
357  }else if(same_block(rt, lb)){
358  block[2]= block[1];
359  }else{
360  block[2]= ptmp;
361  ptmp+=tmp_step;
362  ff_snow_pred_block(s, block[2], tmp, src_stride, src_x, src_y, b_w, b_h, lb, plane_index, w, h);
363  }
364 
365  if(same_block(lt, rb) ){
366  block[3]= block[0];
367  }else if(same_block(rt, rb)){
368  block[3]= block[1];
369  }else if(same_block(lb, rb)){
370  block[3]= block[2];
371  }else{
372  block[3]= ptmp;
373  ff_snow_pred_block(s, block[3], tmp, src_stride, src_x, src_y, b_w, b_h, rb, plane_index, w, h);
374  }
375  if(sliced){
376  s->dwt.inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
377  }else{
378  for(y=0; y<b_h; y++){
379  //FIXME ugly misuse of obmc_stride
380  const uint8_t *obmc1= obmc + y*obmc_stride;
381  const uint8_t *obmc2= obmc1+ (obmc_stride>>1);
382  const uint8_t *obmc3= obmc1+ obmc_stride*(obmc_stride>>1);
383  const uint8_t *obmc4= obmc3+ (obmc_stride>>1);
384  for(x=0; x<b_w; x++){
385  int v= obmc1[x] * block[3][x + y*src_stride]
386  +obmc2[x] * block[2][x + y*src_stride]
387  +obmc3[x] * block[1][x + y*src_stride]
388  +obmc4[x] * block[0][x + y*src_stride];
389 
390  v <<= 8 - LOG2_OBMC_MAX;
391  if(FRAC_BITS != 8){
392  v >>= 8 - FRAC_BITS;
393  }
394  if(add){
395  v += dst[x + y*dst_stride];
396  v = (v + (1<<(FRAC_BITS-1))) >> FRAC_BITS;
397  if(v&(~255)) v= ~(v>>31);
398  dst8[x + y*src_stride] = v;
399  }else{
400  dst[x + y*dst_stride] -= v;
401  }
402  }
403  }
404  }
405 }
406 
407 static av_always_inline void predict_slice(SnowContext *s, IDWTELEM *buf, int plane_index, int add, int mb_y){
408  Plane *p= &s->plane[plane_index];
409  const int mb_w= s->b_width << s->block_max_depth;
410  const int mb_h= s->b_height << s->block_max_depth;
411  int x, y, mb_x;
412  int block_size = MB_SIZE >> s->block_max_depth;
413  int block_w = plane_index ? block_size>>s->chroma_h_shift : block_size;
414  int block_h = plane_index ? block_size>>s->chroma_v_shift : block_size;
415  const uint8_t *obmc = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
416  const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
417  int ref_stride= s->current_picture.linesize[plane_index];
418  uint8_t *dst8= s->current_picture.data[plane_index];
419  int w= p->width;
420  int h= p->height;
421  av_assert2(s->chroma_h_shift == s->chroma_v_shift); // obmc params assume squares
422  if(s->keyframe || (s->avctx->debug&512)){
423  if(mb_y==mb_h)
424  return;
425 
426  if(add){
427  for(y=block_h*mb_y; y<FFMIN(h,block_h*(mb_y+1)); y++){
428  for(x=0; x<w; x++){
429  int v= buf[x + y*w] + (128<<FRAC_BITS) + (1<<(FRAC_BITS-1));
430  v >>= FRAC_BITS;
431  if(v&(~255)) v= ~(v>>31);
432  dst8[x + y*ref_stride]= v;
433  }
434  }
435  }else{
436  for(y=block_h*mb_y; y<FFMIN(h,block_h*(mb_y+1)); y++){
437  for(x=0; x<w; x++){
438  buf[x + y*w]-= 128<<FRAC_BITS;
439  }
440  }
441  }
442 
443  return;
444  }
445 
446  for(mb_x=0; mb_x<=mb_w; mb_x++){
447  add_yblock(s, 0, NULL, buf, dst8, obmc,
448  block_w*mb_x - block_w/2,
449  block_h*mb_y - block_h/2,
450  block_w, block_h,
451  w, h,
452  w, ref_stride, obmc_stride,
453  mb_x - 1, mb_y - 1,
454  add, 1, plane_index);
455  }
456 }
457 
458 static av_always_inline void predict_plane(SnowContext *s, IDWTELEM *buf, int plane_index, int add){
459  const int mb_h= s->b_height << s->block_max_depth;
460  int mb_y;
461  for(mb_y=0; mb_y<=mb_h; mb_y++)
462  predict_slice(s, buf, plane_index, add, mb_y);
463 }
464 
465 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){
466  const int w= s->b_width << s->block_max_depth;
467  const int rem_depth= s->block_max_depth - level;
468  const int index= (x + y*w) << rem_depth;
469  const int block_w= 1<<rem_depth;
470  const int block_h= 1<<rem_depth; //FIXME "w!=h"
472  int i,j;
473 
474  block.color[0]= l;
475  block.color[1]= cb;
476  block.color[2]= cr;
477  block.mx= mx;
478  block.my= my;
479  block.ref= ref;
480  block.type= type;
481  block.level= level;
482 
483  for(j=0; j<block_h; j++){
484  for(i=0; i<block_w; i++){
485  s->block[index + i + j*w]= block;
486  }
487  }
488 }
489 
490 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){
491  SnowContext *s = c->avctx->priv_data;
492  const int offset[3]= {
493  y*c-> stride + x,
494  ((y*c->uvstride + x)>>s->chroma_h_shift),
495  ((y*c->uvstride + x)>>s->chroma_h_shift),
496  };
497  int i;
498  for(i=0; i<3; i++){
499  c->src[0][i]= src [i];
500  c->ref[0][i]= ref [i] + offset[i];
501  }
502  av_assert2(!ref_index);
503 }
504 
505 
506 /* bitstream functions */
507 
508 extern const int8_t ff_quant3bA[256];
509 
510 #define QEXPSHIFT (7-FRAC_BITS+8) //FIXME try to change this to 0
511 
512 static inline void put_symbol(RangeCoder *c, uint8_t *state, int v, int is_signed){
513  int i;
514 
515  if(v){
516  const int a= FFABS(v);
517  const int e= av_log2(a);
518  const int el= FFMIN(e, 10);
519  put_rac(c, state+0, 0);
520 
521  for(i=0; i<el; i++){
522  put_rac(c, state+1+i, 1); //1..10
523  }
524  for(; i<e; i++){
525  put_rac(c, state+1+9, 1); //1..10
526  }
527  put_rac(c, state+1+FFMIN(i,9), 0);
528 
529  for(i=e-1; i>=el; i--){
530  put_rac(c, state+22+9, (a>>i)&1); //22..31
531  }
532  for(; i>=0; i--){
533  put_rac(c, state+22+i, (a>>i)&1); //22..31
534  }
535 
536  if(is_signed)
537  put_rac(c, state+11 + el, v < 0); //11..21
538  }else{
539  put_rac(c, state+0, 1);
540  }
541 }
542 
543 static inline int get_symbol(RangeCoder *c, uint8_t *state, int is_signed){
544  if(get_rac(c, state+0))
545  return 0;
546  else{
547  int i, e, a;
548  e= 0;
549  while(get_rac(c, state+1 + FFMIN(e,9))){ //1..10
550  e++;
551  }
552 
553  a= 1;
554  for(i=e-1; i>=0; i--){
555  a += a + get_rac(c, state+22 + FFMIN(i,9)); //22..31
556  }
557 
558  e= -(is_signed && get_rac(c, state+11 + FFMIN(e,10))); //11..21
559  return (a^e)-e;
560  }
561 }
562 
563 static inline void put_symbol2(RangeCoder *c, uint8_t *state, int v, int log2){
564  int i;
565  int r= log2>=0 ? 1<<log2 : 1;
566 
567  av_assert2(v>=0);
568  av_assert2(log2>=-4);
569 
570  while(v >= r){
571  put_rac(c, state+4+log2, 1);
572  v -= r;
573  log2++;
574  if(log2>0) r+=r;
575  }
576  put_rac(c, state+4+log2, 0);
577 
578  for(i=log2-1; i>=0; i--){
579  put_rac(c, state+31-i, (v>>i)&1);
580  }
581 }
582 
583 static inline int get_symbol2(RangeCoder *c, uint8_t *state, int log2){
584  int i;
585  int r= log2>=0 ? 1<<log2 : 1;
586  int v=0;
587 
588  av_assert2(log2>=-4);
589 
590  while(log2<28 && get_rac(c, state+4+log2)){
591  v+= r;
592  log2++;
593  if(log2>0) r+=r;
594  }
595 
596  for(i=log2-1; i>=0; i--){
597  v+= get_rac(c, state+31-i)<<i;
598  }
599 
600  return v;
601 }
602 
603 static inline void unpack_coeffs(SnowContext *s, SubBand *b, SubBand * parent, int orientation){
604  const int w= b->width;
605  const int h= b->height;
606  int x,y;
607 
608  int run, runs;
609  x_and_coeff *xc= b->x_coeff;
610  x_and_coeff *prev_xc= NULL;
611  x_and_coeff *prev2_xc= xc;
612  x_and_coeff *parent_xc= parent ? parent->x_coeff : NULL;
613  x_and_coeff *prev_parent_xc= parent_xc;
614 
615  runs= get_symbol2(&s->c, b->state[30], 0);
616  if(runs-- > 0) run= get_symbol2(&s->c, b->state[1], 3);
617  else run= INT_MAX;
618 
619  for(y=0; y<h; y++){
620  int v=0;
621  int lt=0, t=0, rt=0;
622 
623  if(y && prev_xc->x == 0){
624  rt= prev_xc->coeff;
625  }
626  for(x=0; x<w; x++){
627  int p=0;
628  const int l= v;
629 
630  lt= t; t= rt;
631 
632  if(y){
633  if(prev_xc->x <= x)
634  prev_xc++;
635  if(prev_xc->x == x + 1)
636  rt= prev_xc->coeff;
637  else
638  rt=0;
639  }
640  if(parent_xc){
641  if(x>>1 > parent_xc->x){
642  parent_xc++;
643  }
644  if(x>>1 == parent_xc->x){
645  p= parent_xc->coeff;
646  }
647  }
648  if(/*ll|*/l|lt|t|rt|p){
649  int context= av_log2(/*FFABS(ll) + */3*(l>>1) + (lt>>1) + (t&~1) + (rt>>1) + (p>>1));
650 
651  v=get_rac(&s->c, &b->state[0][context]);
652  if(v){
653  v= 2*(get_symbol2(&s->c, b->state[context + 2], context-4) + 1);
654  v+=get_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l&0xFF] + 3*ff_quant3bA[t&0xFF]]);
655 
656  xc->x=x;
657  (xc++)->coeff= v;
658  }
659  }else{
660  if(!run){
661  if(runs-- > 0) run= get_symbol2(&s->c, b->state[1], 3);
662  else run= INT_MAX;
663  v= 2*(get_symbol2(&s->c, b->state[0 + 2], 0-4) + 1);
664  v+=get_rac(&s->c, &b->state[0][16 + 1 + 3]);
665 
666  xc->x=x;
667  (xc++)->coeff= v;
668  }else{
669  int max_run;
670  run--;
671  v=0;
672  av_assert2(run >= 0);
673  if(y) max_run= FFMIN(run, prev_xc->x - x - 2);
674  else max_run= FFMIN(run, w-x-1);
675  if(parent_xc)
676  max_run= FFMIN(max_run, 2*parent_xc->x - x - 1);
677  av_assert2(max_run >= 0 && max_run <= run);
678 
679  x+= max_run;
680  run-= max_run;
681  }
682  }
683  }
684  (xc++)->x= w+1; //end marker
685  prev_xc= prev2_xc;
686  prev2_xc= xc;
687 
688  if(parent_xc){
689  if(y&1){
690  while(parent_xc->x != parent->width+1)
691  parent_xc++;
692  parent_xc++;
693  prev_parent_xc= parent_xc;
694  }else{
695  parent_xc= prev_parent_xc;
696  }
697  }
698  }
699 
700  (xc++)->x= w+1; //end marker
701 }
702 
703 #endif /* AVCODEC_SNOW_H */