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 
33 #define FF_MPV_OFFSET(x) (offsetof(MpegEncContext, x) + offsetof(SnowContext, m))
34 #include "mpegvideo.h"
35 #include "h264qpel.h"
36 
37 #define MID_STATE 128
38 
39 #define MAX_PLANES 4
40 #define QSHIFT 5
41 #define QROOT (1<<QSHIFT)
42 #define LOSSLESS_QLOG -128
43 #define FRAC_BITS 4
44 #define MAX_REF_FRAMES 8
45 
46 #define LOG2_OBMC_MAX 8
47 #define OBMC_MAX (1<<(LOG2_OBMC_MAX))
48 typedef struct BlockNode{
49  int16_t mx;
50  int16_t my;
54 //#define TYPE_SPLIT 1
55 #define BLOCK_INTRA 1
56 #define BLOCK_OPT 2
57 //#define TYPE_NOCOLOR 4
58  uint8_t level; //FIXME merge into type?
59 }BlockNode;
60 
61 static const BlockNode null_block= { //FIXME add border maybe
62  .color= {128,128,128},
63  .mx= 0,
64  .my= 0,
65  .ref= 0,
66  .type= 0,
67  .level= 0,
68 };
69 
70 #define LOG2_MB_SIZE 4
71 #define MB_SIZE (1<<LOG2_MB_SIZE)
72 #define ENCODER_EXTRA_BITS 4
73 #define HTAPS_MAX 8
74 
75 typedef struct x_and_coeff{
76  int16_t x;
77  uint16_t coeff;
78 } x_and_coeff;
79 
80 typedef struct SubBand{
81  int level;
82  int stride;
83  int width;
84  int height;
85  int qlog; ///< log(qscale)/log[2^(1/6)]
87  IDWTELEM *ibuf;
90  int stride_line; ///< Stride measured in lines, not pixels.
92  struct SubBand *parent;
93  uint8_t state[/*7*2*/ 7 + 512][32];
94 }SubBand;
95 
96 typedef struct Plane{
97  int width;
98  int height;
100 
101  int htaps;
102  int8_t hcoeff[HTAPS_MAX/2];
103  int diag_mc;
104  int fast_mc;
105 
109 }Plane;
110 
111 typedef struct SnowContext{
112  AVClass *class;
123  AVFrame *input_picture; ///< new_picture with the internal linesizes
128 // uint8_t q_context[16];
130  uint8_t block_state[128 + 32*128];
131  int keyframe;
133  int version;
142  int16_t (*ref_mvs[MAX_REF_FRAMES])[2];
153  int qlog;
155  int lambda;
156  int lambda2;
157  int pass1_rc;
158  int mv_scale;
160  int qbias;
162 #define QBIAS_SHIFT 3
163  int b_width;
164  int b_height;
170 #define ME_CACHE_SIZE 1024
176 
177  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)
178 
181 }SnowContext;
182 
183 /* Tables */
184 extern const uint8_t * const ff_obmc_tab[4];
185 extern uint8_t ff_qexp[QROOT];
187 
188 /* C bits used by mmx/sse2/altivec */
189 
190 static av_always_inline void snow_interleave_line_header(int * i, int width, IDWTELEM * low, IDWTELEM * high){
191  (*i) = (width) - 2;
192 
193  if (width & 1){
194  low[(*i)+1] = low[((*i)+1)>>1];
195  (*i)--;
196  }
197 }
198 
199 static av_always_inline void snow_interleave_line_footer(int * i, IDWTELEM * low, IDWTELEM * high){
200  for (; (*i)>=0; (*i)-=2){
201  low[(*i)+1] = high[(*i)>>1];
202  low[*i] = low[(*i)>>1];
203  }
204 }
205 
206 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){
207  for(; i<w; i++){
208  dst[i] = src[i] - ((mul * (ref[i] + ref[i + 1]) + add) >> shift);
209  }
210 
211  if((width^lift_high)&1){
212  dst[w] = src[w] - ((mul * 2 * ref[w] + add) >> shift);
213  }
214 }
215 
217  for(; i<w; i++){
218  dst[i] = src[i] + ((ref[i] + ref[(i+1)]+W_BO + 4 * src[i]) >> W_BS);
219  }
220 
221  if(width&1){
222  dst[w] = src[w] + ((2 * ref[w] + W_BO + 4 * src[w]) >> W_BS);
223  }
224 }
225 
226 /* common code */
227 
235 void ff_snow_pred_block(SnowContext *s, uint8_t *dst, uint8_t *tmp, ptrdiff_t stride,
236  int sx, int sy, int b_w, int b_h, const BlockNode *block,
237  int plane_index, int w, int h);
239 /* common inline functions */
240 //XXX doublecheck all of them should stay inlined
241 
242 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){
243  const int w= s->b_width << s->block_max_depth;
244  const int rem_depth= s->block_max_depth - level;
245  const int index= (x + y*w) << rem_depth;
246  const int block_w= 1<<rem_depth;
248  int i,j;
249 
250  block.color[0]= l;
251  block.color[1]= cb;
252  block.color[2]= cr;
253  block.mx= mx;
254  block.my= my;
255  block.ref= ref;
256  block.type= type;
257  block.level= level;
258 
259  for(j=0; j<block_w; j++){
260  for(i=0; i<block_w; i++){
261  s->block[index + i + j*w]= block;
262  }
263  }
264 }
265 
266 static inline void pred_mv(SnowContext *s, int *mx, int *my, int ref,
267  const BlockNode *left, const BlockNode *top, const BlockNode *tr){
268  if(s->ref_frames == 1){
269  *mx = mid_pred(left->mx, top->mx, tr->mx);
270  *my = mid_pred(left->my, top->my, tr->my);
271  }else{
272  const int *scale = ff_scale_mv_ref[ref];
273  *mx = mid_pred((left->mx * scale[left->ref] + 128) >>8,
274  (top ->mx * scale[top ->ref] + 128) >>8,
275  (tr ->mx * scale[tr ->ref] + 128) >>8);
276  *my = mid_pred((left->my * scale[left->ref] + 128) >>8,
277  (top ->my * scale[top ->ref] + 128) >>8,
278  (tr ->my * scale[tr ->ref] + 128) >>8);
279  }
280 }
281 
283  if((a->type&BLOCK_INTRA) && (b->type&BLOCK_INTRA)){
284  return !((a->color[0] - b->color[0]) | (a->color[1] - b->color[1]) | (a->color[2] - b->color[2]));
285  }else{
286  return !((a->mx - b->mx) | (a->my - b->my) | (a->ref - b->ref) | ((a->type ^ b->type)&BLOCK_INTRA));
287  }
288 }
289 
290 //FIXME name cleanup (b_w, block_w, b_width stuff)
291 //XXX should we really inline it?
292 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){
293  const int b_width = s->b_width << s->block_max_depth;
294  const int b_height= s->b_height << s->block_max_depth;
295  const int b_stride= b_width;
296  BlockNode *lt= &s->block[b_x + b_y*b_stride];
297  BlockNode *rt= lt+1;
298  BlockNode *lb= lt+b_stride;
299  BlockNode *rb= lb+1;
300  uint8_t *block[4];
301  int tmp_step= src_stride >= 7*MB_SIZE ? MB_SIZE : MB_SIZE*src_stride;
302  uint8_t *tmp = s->scratchbuf;
303  uint8_t *ptmp;
304  int x,y;
305 
306  if(b_x<0){
307  lt= rt;
308  lb= rb;
309  }else if(b_x + 1 >= b_width){
310  rt= lt;
311  rb= lb;
312  }
313  if(b_y<0){
314  lt= lb;
315  rt= rb;
316  }else if(b_y + 1 >= b_height){
317  lb= lt;
318  rb= rt;
319  }
320 
321  if(src_x<0){ //FIXME merge with prev & always round internal width up to *16
322  obmc -= src_x;
323  b_w += src_x;
324  if(!sliced && !offset_dst)
325  dst -= src_x;
326  src_x=0;
327  }
328  if(src_x + b_w > w){
329  b_w = w - src_x;
330  }
331  if(src_y<0){
332  obmc -= src_y*obmc_stride;
333  b_h += src_y;
334  if(!sliced && !offset_dst)
335  dst -= src_y*dst_stride;
336  src_y=0;
337  }
338  if(src_y + b_h> h){
339  b_h = h - src_y;
340  }
341 
342  if(b_w<=0 || b_h<=0) return;
343 
344  av_assert2(src_stride > 2*MB_SIZE + 5);
345 
346  if(!sliced && offset_dst)
347  dst += src_x + src_y*dst_stride;
348  dst8+= src_x + src_y*src_stride;
349 // src += src_x + src_y*src_stride;
350 
351  ptmp= tmp + 3*tmp_step;
352  block[0]= ptmp;
353  ptmp+=tmp_step;
354  ff_snow_pred_block(s, block[0], tmp, src_stride, src_x, src_y, b_w, b_h, lt, plane_index, w, h);
355 
356  if(same_block(lt, rt)){
357  block[1]= block[0];
358  }else{
359  block[1]= ptmp;
360  ptmp+=tmp_step;
361  ff_snow_pred_block(s, block[1], tmp, src_stride, src_x, src_y, b_w, b_h, rt, plane_index, w, h);
362  }
363 
364  if(same_block(lt, lb)){
365  block[2]= block[0];
366  }else if(same_block(rt, lb)){
367  block[2]= block[1];
368  }else{
369  block[2]= ptmp;
370  ptmp+=tmp_step;
371  ff_snow_pred_block(s, block[2], tmp, src_stride, src_x, src_y, b_w, b_h, lb, plane_index, w, h);
372  }
373 
374  if(same_block(lt, rb) ){
375  block[3]= block[0];
376  }else if(same_block(rt, rb)){
377  block[3]= block[1];
378  }else if(same_block(lb, rb)){
379  block[3]= block[2];
380  }else{
381  block[3]= ptmp;
382  ff_snow_pred_block(s, block[3], tmp, src_stride, src_x, src_y, b_w, b_h, rb, plane_index, w, h);
383  }
384  if(sliced){
385  s->dwt.inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
386  }else{
387  for(y=0; y<b_h; y++){
388  //FIXME ugly misuse of obmc_stride
389  const uint8_t *obmc1= obmc + y*obmc_stride;
390  const uint8_t *obmc2= obmc1+ (obmc_stride>>1);
391  const uint8_t *obmc3= obmc1+ obmc_stride*(obmc_stride>>1);
392  const uint8_t *obmc4= obmc3+ (obmc_stride>>1);
393  for(x=0; x<b_w; x++){
394  int v= obmc1[x] * block[3][x + y*src_stride]
395  +obmc2[x] * block[2][x + y*src_stride]
396  +obmc3[x] * block[1][x + y*src_stride]
397  +obmc4[x] * block[0][x + y*src_stride];
398 
399  v <<= 8 - LOG2_OBMC_MAX;
400  if(FRAC_BITS != 8){
401  v >>= 8 - FRAC_BITS;
402  }
403  if(add){
404  v += dst[x + y*dst_stride];
405  v = (v + (1<<(FRAC_BITS-1))) >> FRAC_BITS;
406  if(v&(~255)) v= ~(v>>31);
407  dst8[x + y*src_stride] = v;
408  }else{
409  dst[x + y*dst_stride] -= v;
410  }
411  }
412  }
413  }
414 }
415 
416 static av_always_inline void predict_slice(SnowContext *s, IDWTELEM *buf, int plane_index, int add, int mb_y){
417  Plane *p= &s->plane[plane_index];
418  const int mb_w= s->b_width << s->block_max_depth;
419  const int mb_h= s->b_height << s->block_max_depth;
420  int x, y, mb_x;
421  int block_size = MB_SIZE >> s->block_max_depth;
422  int block_w = plane_index ? block_size>>s->chroma_h_shift : block_size;
423  int block_h = plane_index ? block_size>>s->chroma_v_shift : block_size;
424  const uint8_t *obmc = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
425  const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
426  int ref_stride= s->current_picture->linesize[plane_index];
427  uint8_t *dst8= s->current_picture->data[plane_index];
428  int w= p->width;
429  int h= p->height;
430  av_assert2(s->chroma_h_shift == s->chroma_v_shift); // obmc params assume squares
431  if(s->keyframe || (s->avctx->debug&512)){
432  if(mb_y==mb_h)
433  return;
434 
435  if(add){
436  for(y=block_h*mb_y; y<FFMIN(h,block_h*(mb_y+1)); y++){
437  for(x=0; x<w; x++){
438  int v= buf[x + y*w] + (128<<FRAC_BITS) + (1<<(FRAC_BITS-1));
439  v >>= FRAC_BITS;
440  if(v&(~255)) v= ~(v>>31);
441  dst8[x + y*ref_stride]= v;
442  }
443  }
444  }else{
445  for(y=block_h*mb_y; y<FFMIN(h,block_h*(mb_y+1)); y++){
446  for(x=0; x<w; x++){
447  buf[x + y*w]-= 128<<FRAC_BITS;
448  }
449  }
450  }
451 
452  return;
453  }
454 
455  for(mb_x=0; mb_x<=mb_w; mb_x++){
456  add_yblock(s, 0, NULL, buf, dst8, obmc,
457  block_w*mb_x - block_w/2,
458  block_h*mb_y - block_h/2,
459  block_w, block_h,
460  w, h,
461  w, ref_stride, obmc_stride,
462  mb_x - 1, mb_y - 1,
463  add, 1, plane_index);
464  }
465 }
466 
467 static av_always_inline void predict_plane(SnowContext *s, IDWTELEM *buf, int plane_index, int add){
468  const int mb_h= s->b_height << s->block_max_depth;
469  int mb_y;
470  for(mb_y=0; mb_y<=mb_h; mb_y++)
471  predict_slice(s, buf, plane_index, add, mb_y);
472 }
473 
474 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){
475  const int w= s->b_width << s->block_max_depth;
476  const int rem_depth= s->block_max_depth - level;
477  const int index= (x + y*w) << rem_depth;
478  const int block_w= 1<<rem_depth;
479  const int block_h= 1<<rem_depth; //FIXME "w!=h"
481  int i,j;
482 
483  block.color[0]= l;
484  block.color[1]= cb;
485  block.color[2]= cr;
486  block.mx= mx;
487  block.my= my;
488  block.ref= ref;
489  block.type= type;
490  block.level= level;
491 
492  for(j=0; j<block_h; j++){
493  for(i=0; i<block_w; i++){
494  s->block[index + i + j*w]= block;
495  }
496  }
497 }
498 
499 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){
500  SnowContext *s = c->avctx->priv_data;
501  const int offset[3]= {
502  y*c-> stride + x,
503  ((y*c->uvstride + x)>>s->chroma_h_shift),
504  ((y*c->uvstride + x)>>s->chroma_h_shift),
505  };
506  int i;
507  for(i=0; i<3; i++){
508  c->src[0][i]= src [i];
509  c->ref[0][i]= ref [i] + offset[i];
510  }
511  av_assert2(!ref_index);
512 }
513 
514 
515 /* bitstream functions */
516 
517 extern const int8_t ff_quant3bA[256];
518 
519 #define QEXPSHIFT (7-FRAC_BITS+8) //FIXME try to change this to 0
520 
521 static inline void put_symbol(RangeCoder *c, uint8_t *state, int v, int is_signed){
522  int i;
523 
524  if(v){
525  const int a= FFABS(v);
526  const int e= av_log2(a);
527  const int el= FFMIN(e, 10);
528  put_rac(c, state+0, 0);
529 
530  for(i=0; i<el; i++){
531  put_rac(c, state+1+i, 1); //1..10
532  }
533  for(; i<e; i++){
534  put_rac(c, state+1+9, 1); //1..10
535  }
536  put_rac(c, state+1+FFMIN(i,9), 0);
537 
538  for(i=e-1; i>=el; i--){
539  put_rac(c, state+22+9, (a>>i)&1); //22..31
540  }
541  for(; i>=0; i--){
542  put_rac(c, state+22+i, (a>>i)&1); //22..31
543  }
544 
545  if(is_signed)
546  put_rac(c, state+11 + el, v < 0); //11..21
547  }else{
548  put_rac(c, state+0, 1);
549  }
550 }
551 
552 static inline int get_symbol(RangeCoder *c, uint8_t *state, int is_signed){
553  if(get_rac(c, state+0))
554  return 0;
555  else{
556  int i, e, a;
557  e= 0;
558  while(get_rac(c, state+1 + FFMIN(e,9))){ //1..10
559  e++;
560  }
561 
562  a= 1;
563  for(i=e-1; i>=0; i--){
564  a += a + get_rac(c, state+22 + FFMIN(i,9)); //22..31
565  }
566 
567  e= -(is_signed && get_rac(c, state+11 + FFMIN(e,10))); //11..21
568  return (a^e)-e;
569  }
570 }
571 
572 static inline void put_symbol2(RangeCoder *c, uint8_t *state, int v, int log2){
573  int i;
574  int r= log2>=0 ? 1<<log2 : 1;
575 
576  av_assert2(v>=0);
577  av_assert2(log2>=-4);
578 
579  while(v >= r){
580  put_rac(c, state+4+log2, 1);
581  v -= r;
582  log2++;
583  if(log2>0) r+=r;
584  }
585  put_rac(c, state+4+log2, 0);
586 
587  for(i=log2-1; i>=0; i--){
588  put_rac(c, state+31-i, (v>>i)&1);
589  }
590 }
591 
592 static inline int get_symbol2(RangeCoder *c, uint8_t *state, int log2){
593  int i;
594  int r= log2>=0 ? 1<<log2 : 1;
595  int v=0;
596 
597  av_assert2(log2>=-4);
598 
599  while(log2<28 && get_rac(c, state+4+log2)){
600  v+= r;
601  log2++;
602  if(log2>0) r+=r;
603  }
604 
605  for(i=log2-1; i>=0; i--){
606  v+= get_rac(c, state+31-i)<<i;
607  }
608 
609  return v;
610 }
611 
612 static inline void unpack_coeffs(SnowContext *s, SubBand *b, SubBand * parent, int orientation){
613  const int w= b->width;
614  const int h= b->height;
615  int x,y;
616 
617  int run, runs;
618  x_and_coeff *xc= b->x_coeff;
619  x_and_coeff *prev_xc= NULL;
620  x_and_coeff *prev2_xc= xc;
621  x_and_coeff *parent_xc= parent ? parent->x_coeff : NULL;
622  x_and_coeff *prev_parent_xc= parent_xc;
623 
624  runs= get_symbol2(&s->c, b->state[30], 0);
625  if(runs-- > 0) run= get_symbol2(&s->c, b->state[1], 3);
626  else run= INT_MAX;
627 
628  for(y=0; y<h; y++){
629  int v=0;
630  int lt=0, t=0, rt=0;
631 
632  if(y && prev_xc->x == 0){
633  rt= prev_xc->coeff;
634  }
635  for(x=0; x<w; x++){
636  int p=0;
637  const int l= v;
638 
639  lt= t; t= rt;
640 
641  if(y){
642  if(prev_xc->x <= x)
643  prev_xc++;
644  if(prev_xc->x == x + 1)
645  rt= prev_xc->coeff;
646  else
647  rt=0;
648  }
649  if(parent_xc){
650  if(x>>1 > parent_xc->x){
651  parent_xc++;
652  }
653  if(x>>1 == parent_xc->x){
654  p= parent_xc->coeff;
655  }
656  }
657  if(/*ll|*/l|lt|t|rt|p){
658  int context= av_log2(/*FFABS(ll) + */3*(l>>1) + (lt>>1) + (t&~1) + (rt>>1) + (p>>1));
659 
660  v=get_rac(&s->c, &b->state[0][context]);
661  if(v){
662  v= 2*(get_symbol2(&s->c, b->state[context + 2], context-4) + 1);
663  v+=get_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l&0xFF] + 3*ff_quant3bA[t&0xFF]]);
664  if ((uint16_t)v != v) {
665  av_log(s->avctx, AV_LOG_ERROR, "Coefficient damaged\n");
666  v = 1;
667  }
668  xc->x=x;
669  (xc++)->coeff= v;
670  }
671  }else{
672  if(!run){
673  if(runs-- > 0) run= get_symbol2(&s->c, b->state[1], 3);
674  else run= INT_MAX;
675  v= 2*(get_symbol2(&s->c, b->state[0 + 2], 0-4) + 1);
676  v+=get_rac(&s->c, &b->state[0][16 + 1 + 3]);
677  if ((uint16_t)v != v) {
678  av_log(s->avctx, AV_LOG_ERROR, "Coefficient damaged\n");
679  v = 1;
680  }
681 
682  xc->x=x;
683  (xc++)->coeff= v;
684  }else{
685  int max_run;
686  run--;
687  v=0;
688  av_assert2(run >= 0);
689  if(y) max_run= FFMIN(run, prev_xc->x - x - 2);
690  else max_run= FFMIN(run, w-x-1);
691  if(parent_xc)
692  max_run= FFMIN(max_run, 2*parent_xc->x - x - 1);
693  av_assert2(max_run >= 0 && max_run <= run);
694 
695  x+= max_run;
696  run-= max_run;
697  }
698  }
699  }
700  (xc++)->x= w+1; //end marker
701  prev_xc= prev2_xc;
702  prev2_xc= xc;
703 
704  if(parent_xc){
705  if(y&1){
706  while(parent_xc->x != parent->width+1)
707  parent_xc++;
708  parent_xc++;
709  prev_parent_xc= parent_xc;
710  }else{
711  parent_xc= prev_parent_xc;
712  }
713  }
714  }
715 
716  (xc++)->x= w+1; //end marker
717 }
718 
719 #endif /* AVCODEC_SNOW_H */