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 
26 
27 #include "hpeldsp.h"
28 #include "me_cmp.h"
29 #include "qpeldsp.h"
30 #include "snow_dwt.h"
31 
32 #include "rangecoder.h"
33 #include "mathops.h"
34 
35 #define FF_MPV_OFFSET(x) (offsetof(MpegEncContext, x) + offsetof(SnowContext, m))
36 #include "mpegvideo.h"
37 #include "h264qpel.h"
38 
39 #define MID_STATE 128
40 
41 #define MAX_PLANES 4
42 #define QSHIFT 5
43 #define QROOT (1<<QSHIFT)
44 #define LOSSLESS_QLOG -128
45 #define FRAC_BITS 4
46 #define MAX_REF_FRAMES 8
47 
48 #define LOG2_OBMC_MAX 8
49 #define OBMC_MAX (1<<(LOG2_OBMC_MAX))
50 typedef struct BlockNode{
51  int16_t mx;
52  int16_t my;
56 //#define TYPE_SPLIT 1
57 #define BLOCK_INTRA 1
58 #define BLOCK_OPT 2
59 //#define TYPE_NOCOLOR 4
60  uint8_t level; //FIXME merge into type?
61 }BlockNode;
62 
63 static const BlockNode null_block= { //FIXME add border maybe
64  .color= {128,128,128},
65  .mx= 0,
66  .my= 0,
67  .ref= 0,
68  .type= 0,
69  .level= 0,
70 };
71 
72 #define LOG2_MB_SIZE 4
73 #define MB_SIZE (1<<LOG2_MB_SIZE)
74 #define ENCODER_EXTRA_BITS 4
75 #define HTAPS_MAX 8
76 
77 typedef struct x_and_coeff{
78  int16_t x;
79  uint16_t coeff;
80 } x_and_coeff;
81 
82 typedef struct SubBand{
83  int level;
84  int stride;
85  int width;
86  int height;
87  int qlog; ///< log(qscale)/log[2^(1/6)]
89  IDWTELEM *ibuf;
92  int stride_line; ///< Stride measured in lines, not pixels.
94  struct SubBand *parent;
95  uint8_t state[/*7*2*/ 7 + 512][32];
96 }SubBand;
97 
98 typedef struct Plane{
99  int width;
100  int height;
102 
103  int htaps;
104  int8_t hcoeff[HTAPS_MAX/2];
105  int diag_mc;
106  int fast_mc;
107 
111 }Plane;
112 
113 typedef struct SnowContext{
114  AVClass *class;
125  AVFrame *input_picture; ///< new_picture with the internal linesizes
130 // uint8_t q_context[16];
132  uint8_t block_state[128 + 32*128];
133  int keyframe;
135  int version;
144  int16_t (*ref_mvs[MAX_REF_FRAMES])[2];
155  int qlog;
157  int lambda;
158  int lambda2;
159  int pass1_rc;
160  int mv_scale;
162  int qbias;
164 #define QBIAS_SHIFT 3
165  int b_width;
166  int b_height;
172 #define ME_CACHE_SIZE 1024
179 
180  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)
181 
184 
187 }SnowContext;
188 
189 /* Tables */
190 extern const uint8_t * const ff_obmc_tab[4];
191 extern uint8_t ff_qexp[QROOT];
193 
194 /* C bits used by mmx/sse2/altivec */
195 
196 static av_always_inline void snow_interleave_line_header(int * i, int width, IDWTELEM * low, IDWTELEM * high){
197  (*i) = (width) - 2;
198 
199  if (width & 1){
200  low[(*i)+1] = low[((*i)+1)>>1];
201  (*i)--;
202  }
203 }
204 
205 static av_always_inline void snow_interleave_line_footer(int * i, IDWTELEM * low, IDWTELEM * high){
206  for (; (*i)>=0; (*i)-=2){
207  low[(*i)+1] = high[(*i)>>1];
208  low[*i] = low[(*i)>>1];
209  }
210 }
211 
212 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){
213  for(; i<w; i++){
214  dst[i] = src[i] - ((mul * (ref[i] + ref[i + 1]) + add) >> shift);
215  }
216 
217  if((width^lift_high)&1){
218  dst[w] = src[w] - ((mul * 2 * ref[w] + add) >> shift);
219  }
220 }
221 
223  for(; i<w; i++){
224  dst[i] = src[i] + ((ref[i] + ref[(i+1)]+W_BO + 4 * src[i]) >> W_BS);
225  }
226 
227  if(width&1){
228  dst[w] = src[w] + ((2 * ref[w] + W_BO + 4 * src[w]) >> W_BS);
229  }
230 }
231 
232 /* common code */
233 
241 void ff_snow_pred_block(SnowContext *s, uint8_t *dst, uint8_t *tmp, ptrdiff_t stride,
242  int sx, int sy, int b_w, int b_h, const BlockNode *block,
243  int plane_index, int w, int h);
245 /* common inline functions */
246 //XXX doublecheck all of them should stay inlined
247 
248 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){
249  const int w= s->b_width << s->block_max_depth;
250  const int rem_depth= s->block_max_depth - level;
251  const int index= (x + y*w) << rem_depth;
252  const int block_w= 1<<rem_depth;
254  int i,j;
255 
256  block.color[0]= l;
257  block.color[1]= cb;
258  block.color[2]= cr;
259  block.mx= mx;
260  block.my= my;
261  block.ref= ref;
262  block.type= type;
263  block.level= level;
264 
265  for(j=0; j<block_w; j++){
266  for(i=0; i<block_w; i++){
267  s->block[index + i + j*w]= block;
268  }
269  }
270 }
271 
272 static inline void pred_mv(SnowContext *s, int *mx, int *my, int ref,
273  const BlockNode *left, const BlockNode *top, const BlockNode *tr){
274  if(s->ref_frames == 1){
275  *mx = mid_pred(left->mx, top->mx, tr->mx);
276  *my = mid_pred(left->my, top->my, tr->my);
277  }else{
278  const int *scale = ff_scale_mv_ref[ref];
279  *mx = mid_pred((left->mx * scale[left->ref] + 128) >>8,
280  (top ->mx * scale[top ->ref] + 128) >>8,
281  (tr ->mx * scale[tr ->ref] + 128) >>8);
282  *my = mid_pred((left->my * scale[left->ref] + 128) >>8,
283  (top ->my * scale[top ->ref] + 128) >>8,
284  (tr ->my * scale[tr ->ref] + 128) >>8);
285  }
286 }
287 
289  if((a->type&BLOCK_INTRA) && (b->type&BLOCK_INTRA)){
290  return !((a->color[0] - b->color[0]) | (a->color[1] - b->color[1]) | (a->color[2] - b->color[2]));
291  }else{
292  return !((a->mx - b->mx) | (a->my - b->my) | (a->ref - b->ref) | ((a->type ^ b->type)&BLOCK_INTRA));
293  }
294 }
295 
296 //FIXME name cleanup (b_w, block_w, b_width stuff)
297 //XXX should we really inline it?
298 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){
299  const int b_width = s->b_width << s->block_max_depth;
300  const int b_height= s->b_height << s->block_max_depth;
301  const int b_stride= b_width;
302  BlockNode *lt= &s->block[b_x + b_y*b_stride];
303  BlockNode *rt= lt+1;
304  BlockNode *lb= lt+b_stride;
305  BlockNode *rb= lb+1;
306  uint8_t *block[4];
307  int tmp_step= src_stride >= 7*MB_SIZE ? MB_SIZE : MB_SIZE*src_stride;
308  uint8_t *tmp = s->scratchbuf;
309  uint8_t *ptmp;
310  int x,y;
311 
312  if(b_x<0){
313  lt= rt;
314  lb= rb;
315  }else if(b_x + 1 >= b_width){
316  rt= lt;
317  rb= lb;
318  }
319  if(b_y<0){
320  lt= lb;
321  rt= rb;
322  }else if(b_y + 1 >= b_height){
323  lb= lt;
324  rb= rt;
325  }
326 
327  if(src_x<0){ //FIXME merge with prev & always round internal width up to *16
328  obmc -= src_x;
329  b_w += src_x;
330  if(!sliced && !offset_dst)
331  dst -= src_x;
332  src_x=0;
333  }
334  if(src_x + b_w > w){
335  b_w = w - src_x;
336  }
337  if(src_y<0){
338  obmc -= src_y*obmc_stride;
339  b_h += src_y;
340  if(!sliced && !offset_dst)
341  dst -= src_y*dst_stride;
342  src_y=0;
343  }
344  if(src_y + b_h> h){
345  b_h = h - src_y;
346  }
347 
348  if(b_w<=0 || b_h<=0) return;
349 
350  av_assert2(src_stride > 2*MB_SIZE + 5);
351 
352  if(!sliced && offset_dst)
353  dst += src_x + src_y*dst_stride;
354  dst8+= src_x + src_y*src_stride;
355 // src += src_x + src_y*src_stride;
356 
357  ptmp= tmp + 3*tmp_step;
358  block[0]= ptmp;
359  ptmp+=tmp_step;
360  ff_snow_pred_block(s, block[0], tmp, src_stride, src_x, src_y, b_w, b_h, lt, plane_index, w, h);
361 
362  if(same_block(lt, rt)){
363  block[1]= block[0];
364  }else{
365  block[1]= ptmp;
366  ptmp+=tmp_step;
367  ff_snow_pred_block(s, block[1], tmp, src_stride, src_x, src_y, b_w, b_h, rt, plane_index, w, h);
368  }
369 
370  if(same_block(lt, lb)){
371  block[2]= block[0];
372  }else if(same_block(rt, lb)){
373  block[2]= block[1];
374  }else{
375  block[2]= ptmp;
376  ptmp+=tmp_step;
377  ff_snow_pred_block(s, block[2], tmp, src_stride, src_x, src_y, b_w, b_h, lb, plane_index, w, h);
378  }
379 
380  if(same_block(lt, rb) ){
381  block[3]= block[0];
382  }else if(same_block(rt, rb)){
383  block[3]= block[1];
384  }else if(same_block(lb, rb)){
385  block[3]= block[2];
386  }else{
387  block[3]= ptmp;
388  ff_snow_pred_block(s, block[3], tmp, src_stride, src_x, src_y, b_w, b_h, rb, plane_index, w, h);
389  }
390  if(sliced){
391  s->dwt.inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
392  }else{
393  for(y=0; y<b_h; y++){
394  //FIXME ugly misuse of obmc_stride
395  const uint8_t *obmc1= obmc + y*obmc_stride;
396  const uint8_t *obmc2= obmc1+ (obmc_stride>>1);
397  const uint8_t *obmc3= obmc1+ obmc_stride*(obmc_stride>>1);
398  const uint8_t *obmc4= obmc3+ (obmc_stride>>1);
399  for(x=0; x<b_w; x++){
400  int v= obmc1[x] * block[3][x + y*src_stride]
401  +obmc2[x] * block[2][x + y*src_stride]
402  +obmc3[x] * block[1][x + y*src_stride]
403  +obmc4[x] * block[0][x + y*src_stride];
404 
405  v <<= 8 - LOG2_OBMC_MAX;
406  if(FRAC_BITS != 8){
407  v >>= 8 - FRAC_BITS;
408  }
409  if(add){
410  v += dst[x + y*dst_stride];
411  v = (v + (1<<(FRAC_BITS-1))) >> FRAC_BITS;
412  if(v&(~255)) v= ~(v>>31);
413  dst8[x + y*src_stride] = v;
414  }else{
415  dst[x + y*dst_stride] -= v;
416  }
417  }
418  }
419  }
420 }
421 
422 static av_always_inline void predict_slice(SnowContext *s, IDWTELEM *buf, int plane_index, int add, int mb_y){
423  Plane *p= &s->plane[plane_index];
424  const int mb_w= s->b_width << s->block_max_depth;
425  const int mb_h= s->b_height << s->block_max_depth;
426  int x, y, mb_x;
427  int block_size = MB_SIZE >> s->block_max_depth;
428  int block_w = plane_index ? block_size>>s->chroma_h_shift : block_size;
429  int block_h = plane_index ? block_size>>s->chroma_v_shift : block_size;
430  const uint8_t *obmc = plane_index ? ff_obmc_tab[s->block_max_depth+s->chroma_h_shift] : ff_obmc_tab[s->block_max_depth];
431  const int obmc_stride= plane_index ? (2*block_size)>>s->chroma_h_shift : 2*block_size;
432  int ref_stride= s->current_picture->linesize[plane_index];
433  uint8_t *dst8= s->current_picture->data[plane_index];
434  int w= p->width;
435  int h= p->height;
436  av_assert2(s->chroma_h_shift == s->chroma_v_shift); // obmc params assume squares
437  if(s->keyframe || (s->avctx->debug&512)){
438  if(mb_y==mb_h)
439  return;
440 
441  if(add){
442  for(y=block_h*mb_y; y<FFMIN(h,block_h*(mb_y+1)); y++){
443  for(x=0; x<w; x++){
444  int v= buf[x + y*w] + (128<<FRAC_BITS) + (1<<(FRAC_BITS-1));
445  v >>= FRAC_BITS;
446  if(v&(~255)) v= ~(v>>31);
447  dst8[x + y*ref_stride]= v;
448  }
449  }
450  }else{
451  for(y=block_h*mb_y; y<FFMIN(h,block_h*(mb_y+1)); y++){
452  for(x=0; x<w; x++){
453  buf[x + y*w]-= 128<<FRAC_BITS;
454  }
455  }
456  }
457 
458  return;
459  }
460 
461  for(mb_x=0; mb_x<=mb_w; mb_x++){
462  add_yblock(s, 0, NULL, buf, dst8, obmc,
463  block_w*mb_x - block_w/2,
464  block_h*mb_y - block_h/2,
465  block_w, block_h,
466  w, h,
467  w, ref_stride, obmc_stride,
468  mb_x - 1, mb_y - 1,
469  add, 1, plane_index);
470  }
471 }
472 
473 static av_always_inline void predict_plane(SnowContext *s, IDWTELEM *buf, int plane_index, int add){
474  const int mb_h= s->b_height << s->block_max_depth;
475  int mb_y;
476  for(mb_y=0; mb_y<=mb_h; mb_y++)
477  predict_slice(s, buf, plane_index, add, mb_y);
478 }
479 
480 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){
481  const int w= s->b_width << s->block_max_depth;
482  const int rem_depth= s->block_max_depth - level;
483  const int index= (x + y*w) << rem_depth;
484  const int block_w= 1<<rem_depth;
485  const int block_h= 1<<rem_depth; //FIXME "w!=h"
487  int i,j;
488 
489  block.color[0]= l;
490  block.color[1]= cb;
491  block.color[2]= cr;
492  block.mx= mx;
493  block.my= my;
494  block.ref= ref;
495  block.type= type;
496  block.level= level;
497 
498  for(j=0; j<block_h; j++){
499  for(i=0; i<block_w; i++){
500  s->block[index + i + j*w]= block;
501  }
502  }
503 }
504 
505 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){
506  SnowContext *s = c->avctx->priv_data;
507  const int offset[3]= {
508  y*c-> stride + x,
509  ((y*c->uvstride + x)>>s->chroma_h_shift),
510  ((y*c->uvstride + x)>>s->chroma_h_shift),
511  };
512  int i;
513  for(i=0; i<3; i++){
514  c->src[0][i]= src [i];
515  c->ref[0][i]= ref [i] + offset[i];
516  }
517  av_assert2(!ref_index);
518 }
519 
520 
521 /* bitstream functions */
522 
523 extern const int8_t ff_quant3bA[256];
524 
525 #define QEXPSHIFT (7-FRAC_BITS+8) //FIXME try to change this to 0
526 
527 static inline void put_symbol(RangeCoder *c, uint8_t *state, int v, int is_signed){
528  int i;
529 
530  if(v){
531  const int a= FFABS(v);
532  const int e= av_log2(a);
533  const int el= FFMIN(e, 10);
534  put_rac(c, state+0, 0);
535 
536  for(i=0; i<el; i++){
537  put_rac(c, state+1+i, 1); //1..10
538  }
539  for(; i<e; i++){
540  put_rac(c, state+1+9, 1); //1..10
541  }
542  put_rac(c, state+1+FFMIN(i,9), 0);
543 
544  for(i=e-1; i>=el; i--){
545  put_rac(c, state+22+9, (a>>i)&1); //22..31
546  }
547  for(; i>=0; i--){
548  put_rac(c, state+22+i, (a>>i)&1); //22..31
549  }
550 
551  if(is_signed)
552  put_rac(c, state+11 + el, v < 0); //11..21
553  }else{
554  put_rac(c, state+0, 1);
555  }
556 }
557 
558 static inline int get_symbol(RangeCoder *c, uint8_t *state, int is_signed){
559  if(get_rac(c, state+0))
560  return 0;
561  else{
562  int i, e, a;
563  e= 0;
564  while(get_rac(c, state+1 + FFMIN(e,9))){ //1..10
565  e++;
566  }
567 
568  a= 1;
569  for(i=e-1; i>=0; i--){
570  a += a + get_rac(c, state+22 + FFMIN(i,9)); //22..31
571  }
572 
573  e= -(is_signed && get_rac(c, state+11 + FFMIN(e,10))); //11..21
574  return (a^e)-e;
575  }
576 }
577 
578 static inline void put_symbol2(RangeCoder *c, uint8_t *state, int v, int log2){
579  int i;
580  int r= log2>=0 ? 1<<log2 : 1;
581 
582  av_assert2(v>=0);
583  av_assert2(log2>=-4);
584 
585  while(v >= r){
586  put_rac(c, state+4+log2, 1);
587  v -= r;
588  log2++;
589  if(log2>0) r+=r;
590  }
591  put_rac(c, state+4+log2, 0);
592 
593  for(i=log2-1; i>=0; i--){
594  put_rac(c, state+31-i, (v>>i)&1);
595  }
596 }
597 
598 static inline int get_symbol2(RangeCoder *c, uint8_t *state, int log2){
599  int i;
600  int r= log2>=0 ? 1<<log2 : 1;
601  int v=0;
602 
603  av_assert2(log2>=-4);
604 
605  while(log2<28 && get_rac(c, state+4+log2)){
606  v+= r;
607  log2++;
608  if(log2>0) r+=r;
609  }
610 
611  for(i=log2-1; i>=0; i--){
612  v+= get_rac(c, state+31-i)<<i;
613  }
614 
615  return v;
616 }
617 
618 static inline void unpack_coeffs(SnowContext *s, SubBand *b, SubBand * parent, int orientation){
619  const int w= b->width;
620  const int h= b->height;
621  int x,y;
622 
623  int run, runs;
624  x_and_coeff *xc= b->x_coeff;
625  x_and_coeff *prev_xc= NULL;
626  x_and_coeff *prev2_xc= xc;
627  x_and_coeff *parent_xc= parent ? parent->x_coeff : NULL;
628  x_and_coeff *prev_parent_xc= parent_xc;
629 
630  runs= get_symbol2(&s->c, b->state[30], 0);
631  if(runs-- > 0) run= get_symbol2(&s->c, b->state[1], 3);
632  else run= INT_MAX;
633 
634  for(y=0; y<h; y++){
635  int v=0;
636  int lt=0, t=0, rt=0;
637 
638  if(y && prev_xc->x == 0){
639  rt= prev_xc->coeff;
640  }
641  for(x=0; x<w; x++){
642  int p=0;
643  const int l= v;
644 
645  lt= t; t= rt;
646 
647  if(y){
648  if(prev_xc->x <= x)
649  prev_xc++;
650  if(prev_xc->x == x + 1)
651  rt= prev_xc->coeff;
652  else
653  rt=0;
654  }
655  if(parent_xc){
656  if(x>>1 > parent_xc->x){
657  parent_xc++;
658  }
659  if(x>>1 == parent_xc->x){
660  p= parent_xc->coeff;
661  }
662  }
663  if(/*ll|*/l|lt|t|rt|p){
664  int context= av_log2(/*FFABS(ll) + */3*(l>>1) + (lt>>1) + (t&~1) + (rt>>1) + (p>>1));
665 
666  v=get_rac(&s->c, &b->state[0][context]);
667  if(v){
668  v= 2*(get_symbol2(&s->c, b->state[context + 2], context-4) + 1);
669  v+=get_rac(&s->c, &b->state[0][16 + 1 + 3 + ff_quant3bA[l&0xFF] + 3*ff_quant3bA[t&0xFF]]);
670  if ((uint16_t)v != v) {
671  av_log(s->avctx, AV_LOG_ERROR, "Coefficient damaged\n");
672  v = 1;
673  }
674  xc->x=x;
675  (xc++)->coeff= v;
676  }
677  }else{
678  if(!run){
679  if(runs-- > 0) run= get_symbol2(&s->c, b->state[1], 3);
680  else run= INT_MAX;
681  v= 2*(get_symbol2(&s->c, b->state[0 + 2], 0-4) + 1);
682  v+=get_rac(&s->c, &b->state[0][16 + 1 + 3]);
683  if ((uint16_t)v != v) {
684  av_log(s->avctx, AV_LOG_ERROR, "Coefficient damaged\n");
685  v = 1;
686  }
687 
688  xc->x=x;
689  (xc++)->coeff= v;
690  }else{
691  int max_run;
692  run--;
693  v=0;
694  av_assert2(run >= 0);
695  if(y) max_run= FFMIN(run, prev_xc->x - x - 2);
696  else max_run= FFMIN(run, w-x-1);
697  if(parent_xc)
698  max_run= FFMIN(max_run, 2*parent_xc->x - x - 1);
699  av_assert2(max_run >= 0 && max_run <= run);
700 
701  x+= max_run;
702  run-= max_run;
703  }
704  }
705  }
706  (xc++)->x= w+1; //end marker
707  prev_xc= prev2_xc;
708  prev2_xc= xc;
709 
710  if(parent_xc){
711  if(y&1){
712  while(parent_xc->x != parent->width+1)
713  parent_xc++;
714  parent_xc++;
715  prev_parent_xc= parent_xc;
716  }else{
717  parent_xc= prev_parent_xc;
718  }
719  }
720  }
721 
722  (xc++)->x= w+1; //end marker
723 }
724 
725 #endif /* AVCODEC_SNOW_H */