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 "hpeldsp.h"
26 #include "me_cmp.h"
27 #include "qpeldsp.h"
28 #include "snow_dwt.h"
29 
30 #include "rangecoder.h"
31 #include "mathops.h"
32 #include "mpegvideo.h"
33 #include "h264qpel.h"
34 
35 #define MID_STATE 128
36 
37 #define MAX_PLANES 4
38 #define QSHIFT 5
39 #define QROOT (1<<QSHIFT)
40 #define LOSSLESS_QLOG -128
41 #define FRAC_BITS 4
42 #define MAX_REF_FRAMES 8
43 
44 #define LOG2_OBMC_MAX 8
45 #define OBMC_MAX (1<<(LOG2_OBMC_MAX))
46 typedef struct BlockNode{
47  int16_t mx;
48  int16_t my;
52 //#define TYPE_SPLIT 1
53 #define BLOCK_INTRA 1
54 #define BLOCK_OPT 2
55 //#define TYPE_NOCOLOR 4
56  uint8_t level; //FIXME merge into type?
57 }BlockNode;
58 
59 static const BlockNode null_block= { //FIXME add border maybe
60  .color= {128,128,128},
61  .mx= 0,
62  .my= 0,
63  .ref= 0,
64  .type= 0,
65  .level= 0,
66 };
67 
68 #define LOG2_MB_SIZE 4
69 #define MB_SIZE (1<<LOG2_MB_SIZE)
70 #define ENCODER_EXTRA_BITS 4
71 #define HTAPS_MAX 8
72 
73 typedef struct x_and_coeff{
74  int16_t x;
75  uint16_t coeff;
76 } x_and_coeff;
77 
78 typedef struct SubBand{
79  int level;
80  int stride;
81  int width;
82  int height;
83  int qlog; ///< log(qscale)/log[2^(1/6)]
85  IDWTELEM *ibuf;
88  int stride_line; ///< Stride measured in lines, not pixels.
90  struct SubBand *parent;
91  uint8_t state[/*7*2*/ 7 + 512][32];
92 }SubBand;
93 
94 typedef struct Plane{
95  int width;
96  int height;
98 
99  int htaps;
100  int8_t hcoeff[HTAPS_MAX/2];
101  int diag_mc;
102  int fast_mc;
103 
107 }Plane;
108 
109 typedef struct SnowContext{
110  AVClass *class;
121  AVFrame *input_picture; ///< new_picture with the internal linesizes
126 // uint8_t q_context[16];
128  uint8_t block_state[128 + 32*128];
129  int keyframe;
131  int version;
140  int16_t (*ref_mvs[MAX_REF_FRAMES])[2];
151  int qlog;
153  int lambda;
154  int lambda2;
155  int pass1_rc;
156  int mv_scale;
158  int qbias;
160 #define QBIAS_SHIFT 3
161  int b_width;
162  int b_height;
168 #define ME_CACHE_SIZE 1024
174 
175  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)
176 
179 }SnowContext;
180 
181 /* Tables */
182 extern const uint8_t * const ff_obmc_tab[4];
183 extern uint8_t ff_qexp[QROOT];
185 
186 /* C bits used by mmx/sse2/altivec */
187 
188 static av_always_inline void snow_interleave_line_header(int * i, int width, IDWTELEM * low, IDWTELEM * high){
189  (*i) = (width) - 2;
190 
191  if (width & 1){
192  low[(*i)+1] = low[((*i)+1)>>1];
193  (*i)--;
194  }
195 }
196 
197 static av_always_inline void snow_interleave_line_footer(int * i, IDWTELEM * low, IDWTELEM * high){
198  for (; (*i)>=0; (*i)-=2){
199  low[(*i)+1] = high[(*i)>>1];
200  low[*i] = low[(*i)>>1];
201  }
202 }
203 
204 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){
205  for(; i<w; i++){
206  dst[i] = src[i] - ((mul * (ref[i] + ref[i + 1]) + add) >> shift);
207  }
208 
209  if((width^lift_high)&1){
210  dst[w] = src[w] - ((mul * 2 * ref[w] + add) >> shift);
211  }
212 }
213 
215  for(; i<w; i++){
216  dst[i] = src[i] + ((ref[i] + ref[(i+1)]+W_BO + 4 * src[i]) >> W_BS);
217  }
218 
219  if(width&1){
220  dst[w] = src[w] + ((2 * ref[w] + W_BO + 4 * src[w]) >> W_BS);
221  }
222 }
223 
224 /* common code */
225 
233 void ff_snow_pred_block(SnowContext *s, uint8_t *dst, uint8_t *tmp, ptrdiff_t stride,
234  int sx, int sy, int b_w, int b_h, BlockNode *block,
235  int plane_index, int w, int h);
237 /* common inline functions */
238 //XXX doublecheck all of them should stay inlined
239 
240 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){
241  const int w= s->b_width << s->block_max_depth;
242  const int rem_depth= s->block_max_depth - level;
243  const int index= (x + y*w) << rem_depth;
244  const int block_w= 1<<rem_depth;
246  int i,j;
247 
248  block.color[0]= l;
249  block.color[1]= cb;
250  block.color[2]= cr;
251  block.mx= mx;
252  block.my= my;
253  block.ref= ref;
254  block.type= type;
255  block.level= level;
256 
257  for(j=0; j<block_w; j++){
258  for(i=0; i<block_w; i++){
259  s->block[index + i + j*w]= block;
260  }
261  }
262 }
263 
264 static inline void pred_mv(SnowContext *s, int *mx, int *my, int ref,
265  const BlockNode *left, const BlockNode *top, const BlockNode *tr){
266  if(s->ref_frames == 1){
267  *mx = mid_pred(left->mx, top->mx, tr->mx);
268  *my = mid_pred(left->my, top->my, tr->my);
269  }else{
270  const int *scale = ff_scale_mv_ref[ref];
271  *mx = mid_pred((left->mx * scale[left->ref] + 128) >>8,
272  (top ->mx * scale[top ->ref] + 128) >>8,
273  (tr ->mx * scale[tr ->ref] + 128) >>8);
274  *my = mid_pred((left->my * scale[left->ref] + 128) >>8,
275  (top ->my * scale[top ->ref] + 128) >>8,
276  (tr ->my * scale[tr ->ref] + 128) >>8);
277  }
278 }
279 
281  if((a->type&BLOCK_INTRA) && (b->type&BLOCK_INTRA)){
282  return !((a->color[0] - b->color[0]) | (a->color[1] - b->color[1]) | (a->color[2] - b->color[2]));
283  }else{
284  return !((a->mx - b->mx) | (a->my - b->my) | (a->ref - b->ref) | ((a->type ^ b->type)&BLOCK_INTRA));
285  }
286 }
287 
288 //FIXME name cleanup (b_w, block_w, b_width stuff)
289 //XXX should we really inline it?
290 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){
291  const int b_width = s->b_width << s->block_max_depth;
292  const int b_height= s->b_height << s->block_max_depth;
293  const int b_stride= b_width;
294  BlockNode *lt= &s->block[b_x + b_y*b_stride];
295  BlockNode *rt= lt+1;
296  BlockNode *lb= lt+b_stride;
297  BlockNode *rb= lb+1;
298  uint8_t *block[4];
299  int tmp_step= src_stride >= 7*MB_SIZE ? MB_SIZE : MB_SIZE*src_stride;
300  uint8_t *tmp = s->scratchbuf;
301  uint8_t *ptmp;
302  int x,y;
303 
304  if(b_x<0){
305  lt= rt;
306  lb= rb;
307  }else if(b_x + 1 >= b_width){
308  rt= lt;
309  rb= lb;
310  }
311  if(b_y<0){
312  lt= lb;
313  rt= rb;
314  }else if(b_y + 1 >= b_height){
315  lb= lt;
316  rb= rt;
317  }
318 
319  if(src_x<0){ //FIXME merge with prev & always round internal width up to *16
320  obmc -= src_x;
321  b_w += src_x;
322  if(!sliced && !offset_dst)
323  dst -= src_x;
324  src_x=0;
325  }
326  if(src_x + b_w > w){
327  b_w = w - src_x;
328  }
329  if(src_y<0){
330  obmc -= src_y*obmc_stride;
331  b_h += src_y;
332  if(!sliced && !offset_dst)
333  dst -= src_y*dst_stride;
334  src_y=0;
335  }
336  if(src_y + b_h> h){
337  b_h = h - src_y;
338  }
339 
340  if(b_w<=0 || b_h<=0) return;
341 
342  av_assert2(src_stride > 2*MB_SIZE + 5);
343 
344  if(!sliced && offset_dst)
345  dst += src_x + src_y*dst_stride;
346  dst8+= src_x + src_y*src_stride;
347 // src += src_x + src_y*src_stride;
348 
349  ptmp= tmp + 3*tmp_step;
350  block[0]= ptmp;
351  ptmp+=tmp_step;
352  ff_snow_pred_block(s, block[0], tmp, src_stride, src_x, src_y, b_w, b_h, lt, plane_index, w, h);
353 
354  if(same_block(lt, rt)){
355  block[1]= block[0];
356  }else{
357  block[1]= ptmp;
358  ptmp+=tmp_step;
359  ff_snow_pred_block(s, block[1], tmp, src_stride, src_x, src_y, b_w, b_h, rt, plane_index, w, h);
360  }
361 
362  if(same_block(lt, lb)){
363  block[2]= block[0];
364  }else if(same_block(rt, lb)){
365  block[2]= block[1];
366  }else{
367  block[2]= ptmp;
368  ptmp+=tmp_step;
369  ff_snow_pred_block(s, block[2], tmp, src_stride, src_x, src_y, b_w, b_h, lb, plane_index, w, h);
370  }
371 
372  if(same_block(lt, rb) ){
373  block[3]= block[0];
374  }else if(same_block(rt, rb)){
375  block[3]= block[1];
376  }else if(same_block(lb, rb)){
377  block[3]= block[2];
378  }else{
379  block[3]= ptmp;
380  ff_snow_pred_block(s, block[3], tmp, src_stride, src_x, src_y, b_w, b_h, rb, plane_index, w, h);
381  }
382  if(sliced){
383  s->dwt.inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
384  }else{
385  for(y=0; y<b_h; y++){
386  //FIXME ugly misuse of obmc_stride
387  const uint8_t *obmc1= obmc + y*obmc_stride;
388  const uint8_t *obmc2= obmc1+ (obmc_stride>>1);
389  const uint8_t *obmc3= obmc1+ obmc_stride*(obmc_stride>>1);
390  const uint8_t *obmc4= obmc3+ (obmc_stride>>1);
391  for(x=0; x<b_w; x++){
392  int v= obmc1[x] * block[3][x + y*src_stride]
393  +obmc2[x] * block[2][x + y*src_stride]
394  +obmc3[x] * block[1][x + y*src_stride]
395  +obmc4[x] * block[0][x + y*src_stride];
396 
397  v <<= 8 - LOG2_OBMC_MAX;
398  if(FRAC_BITS != 8){
399  v >>= 8 - FRAC_BITS;
400  }
401  if(add){
402  v += dst[x + y*dst_stride];
403  v = (v + (1<<(FRAC_BITS-1))) >> FRAC_BITS;
404  if(v&(~255)) v= ~(v>>31);
405  dst8[x + y*src_stride] = v;
406  }else{
407  dst[x + y*dst_stride] -= v;
408  }
409  }
410  }
411  }
412 }
413 
414 static av_always_inline void predict_slice(SnowContext *s, IDWTELEM *buf, int plane_index, int add, int mb_y){
415  Plane *p= &s->plane[plane_index];
416  const int mb_w= s->b_width << s->block_max_depth;
417  const int mb_h= s->b_height << s->block_max_depth;
418  int x, y, mb_x;
419  int block_size = MB_SIZE >> s->block_max_depth;
420  int block_w = plane_index ? block_size>>s->chroma_h_shift : block_size;
421  int block_h = plane_index ? block_size>>s->chroma_v_shift : block_size;
422  const uint8_t *obmc = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
423  const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
424  int ref_stride= s->current_picture->linesize[plane_index];
425  uint8_t *dst8= s->current_picture->data[plane_index];
426  int w= p->width;
427  int h= p->height;
428  av_assert2(s->chroma_h_shift == s->chroma_v_shift); // obmc params assume squares
429  if(s->keyframe || (s->avctx->debug&512)){
430  if(mb_y==mb_h)
431  return;
432 
433  if(add){
434  for(y=block_h*mb_y; y<FFMIN(h,block_h*(mb_y+1)); y++){
435  for(x=0; x<w; x++){
436  int v= buf[x + y*w] + (128<<FRAC_BITS) + (1<<(FRAC_BITS-1));
437  v >>= FRAC_BITS;
438  if(v&(~255)) v= ~(v>>31);
439  dst8[x + y*ref_stride]= v;
440  }
441  }
442  }else{
443  for(y=block_h*mb_y; y<FFMIN(h,block_h*(mb_y+1)); y++){
444  for(x=0; x<w; x++){
445  buf[x + y*w]-= 128<<FRAC_BITS;
446  }
447  }
448  }
449 
450  return;
451  }
452 
453  for(mb_x=0; mb_x<=mb_w; mb_x++){
454  add_yblock(s, 0, NULL, buf, dst8, obmc,
455  block_w*mb_x - block_w/2,
456  block_h*mb_y - block_h/2,
457  block_w, block_h,
458  w, h,
459  w, ref_stride, obmc_stride,
460  mb_x - 1, mb_y - 1,
461  add, 1, plane_index);
462  }
463 }
464 
465 static av_always_inline void predict_plane(SnowContext *s, IDWTELEM *buf, int plane_index, int add){
466  const int mb_h= s->b_height << s->block_max_depth;
467  int mb_y;
468  for(mb_y=0; mb_y<=mb_h; mb_y++)
469  predict_slice(s, buf, plane_index, add, mb_y);
470 }
471 
472 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){
473  const int w= s->b_width << s->block_max_depth;
474  const int rem_depth= s->block_max_depth - level;
475  const int index= (x + y*w) << rem_depth;
476  const int block_w= 1<<rem_depth;
477  const int block_h= 1<<rem_depth; //FIXME "w!=h"
479  int i,j;
480 
481  block.color[0]= l;
482  block.color[1]= cb;
483  block.color[2]= cr;
484  block.mx= mx;
485  block.my= my;
486  block.ref= ref;
487  block.type= type;
488  block.level= level;
489 
490  for(j=0; j<block_h; j++){
491  for(i=0; i<block_w; i++){
492  s->block[index + i + j*w]= block;
493  }
494  }
495 }
496 
497 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){
498  SnowContext *s = c->avctx->priv_data;
499  const int offset[3]= {
500  y*c-> stride + x,
501  ((y*c->uvstride + x)>>s->chroma_h_shift),
502  ((y*c->uvstride + x)>>s->chroma_h_shift),
503  };
504  int i;
505  for(i=0; i<3; i++){
506  c->src[0][i]= src [i];
507  c->ref[0][i]= ref [i] + offset[i];
508  }
509  av_assert2(!ref_index);
510 }
511 
512 
513 /* bitstream functions */
514 
515 extern const int8_t ff_quant3bA[256];
516 
517 #define QEXPSHIFT (7-FRAC_BITS+8) //FIXME try to change this to 0
518 
519 static inline void put_symbol(RangeCoder *c, uint8_t *state, int v, int is_signed){
520  int i;
521 
522  if(v){
523  const int a= FFABS(v);
524  const int e= av_log2(a);
525  const int el= FFMIN(e, 10);
526  put_rac(c, state+0, 0);
527 
528  for(i=0; i<el; i++){
529  put_rac(c, state+1+i, 1); //1..10
530  }
531  for(; i<e; i++){
532  put_rac(c, state+1+9, 1); //1..10
533  }
534  put_rac(c, state+1+FFMIN(i,9), 0);
535 
536  for(i=e-1; i>=el; i--){
537  put_rac(c, state+22+9, (a>>i)&1); //22..31
538  }
539  for(; i>=0; i--){
540  put_rac(c, state+22+i, (a>>i)&1); //22..31
541  }
542 
543  if(is_signed)
544  put_rac(c, state+11 + el, v < 0); //11..21
545  }else{
546  put_rac(c, state+0, 1);
547  }
548 }
549 
550 static inline int get_symbol(RangeCoder *c, uint8_t *state, int is_signed){
551  if(get_rac(c, state+0))
552  return 0;
553  else{
554  int i, e, a;
555  e= 0;
556  while(get_rac(c, state+1 + FFMIN(e,9))){ //1..10
557  e++;
558  }
559 
560  a= 1;
561  for(i=e-1; i>=0; i--){
562  a += a + get_rac(c, state+22 + FFMIN(i,9)); //22..31
563  }
564 
565  e= -(is_signed && get_rac(c, state+11 + FFMIN(e,10))); //11..21
566  return (a^e)-e;
567  }
568 }
569 
570 static inline void put_symbol2(RangeCoder *c, uint8_t *state, int v, int log2){
571  int i;
572  int r= log2>=0 ? 1<<log2 : 1;
573 
574  av_assert2(v>=0);
575  av_assert2(log2>=-4);
576 
577  while(v >= r){
578  put_rac(c, state+4+log2, 1);
579  v -= r;
580  log2++;
581  if(log2>0) r+=r;
582  }
583  put_rac(c, state+4+log2, 0);
584 
585  for(i=log2-1; i>=0; i--){
586  put_rac(c, state+31-i, (v>>i)&1);
587  }
588 }
589 
590 static inline int get_symbol2(RangeCoder *c, uint8_t *state, int log2){
591  int i;
592  int r= log2>=0 ? 1<<log2 : 1;
593  int v=0;
594 
595  av_assert2(log2>=-4);
596 
597  while(log2<28 && get_rac(c, state+4+log2)){
598  v+= r;
599  log2++;
600  if(log2>0) r+=r;
601  }
602 
603  for(i=log2-1; i>=0; i--){
604  v+= get_rac(c, state+31-i)<<i;
605  }
606 
607  return v;
608 }
609 
610 static inline void unpack_coeffs(SnowContext *s, SubBand *b, SubBand * parent, int orientation){
611  const int w= b->width;
612  const int h= b->height;
613  int x,y;
614 
615  int run, runs;
616  x_and_coeff *xc= b->x_coeff;
617  x_and_coeff *prev_xc= NULL;
618  x_and_coeff *prev2_xc= xc;
619  x_and_coeff *parent_xc= parent ? parent->x_coeff : NULL;
620  x_and_coeff *prev_parent_xc= parent_xc;
621 
622  runs= get_symbol2(&s->c, b->state[30], 0);
623  if(runs-- > 0) run= get_symbol2(&s->c, b->state[1], 3);
624  else run= INT_MAX;
625 
626  for(y=0; y<h; y++){
627  int v=0;
628  int lt=0, t=0, rt=0;
629 
630  if(y && prev_xc->x == 0){
631  rt= prev_xc->coeff;
632  }
633  for(x=0; x<w; x++){
634  int p=0;
635  const int l= v;
636 
637  lt= t; t= rt;
638 
639  if(y){
640  if(prev_xc->x <= x)
641  prev_xc++;
642  if(prev_xc->x == x + 1)
643  rt= prev_xc->coeff;
644  else
645  rt=0;
646  }
647  if(parent_xc){
648  if(x>>1 > parent_xc->x){
649  parent_xc++;
650  }
651  if(x>>1 == parent_xc->x){
652  p= parent_xc->coeff;
653  }
654  }
655  if(/*ll|*/l|lt|t|rt|p){
656  int context= av_log2(/*FFABS(ll) + */3*(l>>1) + (lt>>1) + (t&~1) + (rt>>1) + (p>>1));
657 
658  v=get_rac(&s->c, &b->state[0][context]);
659  if(v){
660  v= 2*(get_symbol2(&s->c, b->state[context + 2], context-4) + 1);
661  v+=get_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l&0xFF] + 3*ff_quant3bA[t&0xFF]]);
662  if ((uint16_t)v != v) {
663  av_log(s->avctx, AV_LOG_ERROR, "Coefficient damaged\n");
664  v = 1;
665  }
666  xc->x=x;
667  (xc++)->coeff= v;
668  }
669  }else{
670  if(!run){
671  if(runs-- > 0) run= get_symbol2(&s->c, b->state[1], 3);
672  else run= INT_MAX;
673  v= 2*(get_symbol2(&s->c, b->state[0 + 2], 0-4) + 1);
674  v+=get_rac(&s->c, &b->state[0][16 + 1 + 3]);
675  if ((uint16_t)v != v) {
676  av_log(s->avctx, AV_LOG_ERROR, "Coefficient damaged\n");
677  v = 1;
678  }
679 
680  xc->x=x;
681  (xc++)->coeff= v;
682  }else{
683  int max_run;
684  run--;
685  v=0;
686  av_assert2(run >= 0);
687  if(y) max_run= FFMIN(run, prev_xc->x - x - 2);
688  else max_run= FFMIN(run, w-x-1);
689  if(parent_xc)
690  max_run= FFMIN(max_run, 2*parent_xc->x - x - 1);
691  av_assert2(max_run >= 0 && max_run <= run);
692 
693  x+= max_run;
694  run-= max_run;
695  }
696  }
697  }
698  (xc++)->x= w+1; //end marker
699  prev_xc= prev2_xc;
700  prev2_xc= xc;
701 
702  if(parent_xc){
703  if(y&1){
704  while(parent_xc->x != parent->width+1)
705  parent_xc++;
706  parent_xc++;
707  prev_parent_xc= parent_xc;
708  }else{
709  parent_xc= prev_parent_xc;
710  }
711  }
712  }
713 
714  (xc++)->x= w+1; //end marker
715 }
716 
717 #endif /* AVCODEC_SNOW_H */