FFmpeg
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
roqvideoenc.c
Go to the documentation of this file.
1 /*
2  * RoQ Video Encoder.
3  *
4  * Copyright (C) 2007 Vitor Sessak <vitor1001@gmail.com>
5  * Copyright (C) 2004-2007 Eric Lasota
6  * Based on RoQ specs (C) 2001 Tim Ferguson
7  *
8  * This file is part of FFmpeg.
9  *
10  * FFmpeg is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU Lesser General Public
12  * License as published by the Free Software Foundation; either
13  * version 2.1 of the License, or (at your option) any later version.
14  *
15  * FFmpeg is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18  * Lesser General Public License for more details.
19  *
20  * You should have received a copy of the GNU Lesser General Public
21  * License along with FFmpeg; if not, write to the Free Software
22  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23  */
24 
25 /**
26  * @file
27  * id RoQ encoder by Vitor. Based on the Switchblade3 library and the
28  * Switchblade3 FFmpeg glue by Eric Lasota.
29  */
30 
31 /*
32  * COSTS:
33  * Level 1:
34  * SKIP - 2 bits
35  * MOTION - 2 + 8 bits
36  * CODEBOOK - 2 + 8 bits
37  * SUBDIVIDE - 2 + combined subcel cost
38  *
39  * Level 2:
40  * SKIP - 2 bits
41  * MOTION - 2 + 8 bits
42  * CODEBOOK - 2 + 8 bits
43  * SUBDIVIDE - 2 + 4*8 bits
44  *
45  * Maximum cost: 138 bits per cel
46  *
47  * Proper evaluation requires LCD fraction comparison, which requires
48  * Squared Error (SE) loss * savings increase
49  *
50  * Maximum savings increase: 136 bits
51  * Maximum SE loss without overflow: 31580641
52  * Components in 8x8 supercel: 192
53  * Maximum SE precision per component: 164482
54  * >65025, so no truncation is needed (phew)
55  */
56 
57 #include <string.h>
58 
59 #include "libavutil/attributes.h"
60 #include "libavutil/opt.h"
61 #include "roqvideo.h"
62 #include "bytestream.h"
63 #include "elbg.h"
64 #include "internal.h"
65 #include "mathops.h"
66 
67 #define CHROMA_BIAS 1
68 
69 /**
70  * Maximum number of generated 4x4 codebooks. Can't be 256 to workaround a
71  * Quake 3 bug.
72  */
73 #define MAX_CBS_4x4 256
74 
75 #define MAX_CBS_2x2 256 ///< Maximum number of 2x2 codebooks.
76 
77 /* The cast is useful when multiplying it by INT_MAX */
78 #define ROQ_LAMBDA_SCALE ((uint64_t) FF_LAMBDA_SCALE)
79 
80 /* Macroblock support functions */
81 static void unpack_roq_cell(roq_cell *cell, uint8_t u[4*3])
82 {
83  memcpy(u , cell->y, 4);
84  memset(u+4, cell->u, 4);
85  memset(u+8, cell->v, 4);
86 }
87 
88 static void unpack_roq_qcell(uint8_t cb2[], roq_qcell *qcell, uint8_t u[4*4*3])
89 {
90  int i,cp;
91  static const int offsets[4] = {0, 2, 8, 10};
92 
93  for (cp=0; cp<3; cp++)
94  for (i=0; i<4; i++) {
95  u[4*4*cp + offsets[i] ] = cb2[qcell->idx[i]*2*2*3 + 4*cp ];
96  u[4*4*cp + offsets[i]+1] = cb2[qcell->idx[i]*2*2*3 + 4*cp+1];
97  u[4*4*cp + offsets[i]+4] = cb2[qcell->idx[i]*2*2*3 + 4*cp+2];
98  u[4*4*cp + offsets[i]+5] = cb2[qcell->idx[i]*2*2*3 + 4*cp+3];
99  }
100 }
101 
102 
103 static void enlarge_roq_mb4(uint8_t base[3*16], uint8_t u[3*64])
104 {
105  int x,y,cp;
106 
107  for(cp=0; cp<3; cp++)
108  for(y=0; y<8; y++)
109  for(x=0; x<8; x++)
110  *u++ = base[(y/2)*4 + (x/2) + 16*cp];
111 }
112 
113 static inline int square(int x)
114 {
115  return x*x;
116 }
117 
118 static inline int eval_sse(const uint8_t *a, const uint8_t *b, int count)
119 {
120  int diff=0;
121 
122  while(count--)
123  diff += square(*b++ - *a++);
124 
125  return diff;
126 }
127 
128 // FIXME Could use DSPContext.sse, but it is not so speed critical (used
129 // just for motion estimation).
130 static int block_sse(uint8_t * const *buf1, uint8_t * const *buf2, int x1, int y1,
131  int x2, int y2, const int *stride1, const int *stride2, int size)
132 {
133  int i, k;
134  int sse=0;
135 
136  for (k=0; k<3; k++) {
137  int bias = (k ? CHROMA_BIAS : 4);
138  for (i=0; i<size; i++)
139  sse += bias*eval_sse(buf1[k] + (y1+i)*stride1[k] + x1,
140  buf2[k] + (y2+i)*stride2[k] + x2, size);
141  }
142 
143  return sse;
144 }
145 
146 static int eval_motion_dist(RoqContext *enc, int x, int y, motion_vect vect,
147  int size)
148 {
149  int mx=vect.d[0];
150  int my=vect.d[1];
151 
152  if (mx < -7 || mx > 7)
153  return INT_MAX;
154 
155  if (my < -7 || my > 7)
156  return INT_MAX;
157 
158  mx += x;
159  my += y;
160 
161  if ((unsigned) mx > enc->width-size || (unsigned) my > enc->height-size)
162  return INT_MAX;
163 
164  return block_sse(enc->frame_to_enc->data, enc->last_frame->data, x, y,
165  mx, my,
167  size);
168 }
169 
170 /**
171  * @return distortion between two macroblocks
172  */
173 static inline int squared_diff_macroblock(uint8_t a[], uint8_t b[], int size)
174 {
175  int cp, sdiff=0;
176 
177  for(cp=0;cp<3;cp++) {
178  int bias = (cp ? CHROMA_BIAS : 4);
179  sdiff += bias*eval_sse(a, b, size*size);
180  a += size*size;
181  b += size*size;
182  }
183 
184  return sdiff;
185 }
186 
187 typedef struct
188 {
189  int eval_dist[4];
192 
193  int subCels[4];
195  int cbEntry;
197 
198 typedef struct
199 {
200  int eval_dist[4];
202 
203  SubcelEvaluation subCels[4];
204 
206  int cbEntry;
207 
208  int sourceX, sourceY;
209 } CelEvaluation;
210 
211 typedef struct
212 {
213  int numCB4;
214  int numCB2;
215  int usedCB2[MAX_CBS_2x2];
216  int usedCB4[MAX_CBS_4x4];
217  uint8_t unpacked_cb2[MAX_CBS_2x2*2*2*3];
218  uint8_t unpacked_cb4[MAX_CBS_4x4*4*4*3];
219  uint8_t unpacked_cb4_enlarged[MAX_CBS_4x4*8*8*3];
220 } RoqCodebooks;
221 
222 /**
223  * Temporary vars
224  */
225 typedef struct RoqTempData
226 {
228 
229  int f2i4[MAX_CBS_4x4];
230  int i2f4[MAX_CBS_4x4];
231  int f2i2[MAX_CBS_2x2];
232  int i2f2[MAX_CBS_2x2];
233 
235 
236  int numCB4;
237  int numCB2;
238 
240 
242  int used_option[4];
243 } RoqTempdata;
244 
245 /**
246  * Initialize cel evaluators and set their source coordinates
247  */
248 static int create_cel_evals(RoqContext *enc, RoqTempdata *tempData)
249 {
250  int n=0, x, y, i;
251 
252  tempData->cel_evals = av_malloc_array(enc->width*enc->height/64, sizeof(CelEvaluation));
253  if (!tempData->cel_evals)
254  return AVERROR(ENOMEM);
255 
256  /* Map to the ROQ quadtree order */
257  for (y=0; y<enc->height; y+=16)
258  for (x=0; x<enc->width; x+=16)
259  for(i=0; i<4; i++) {
260  tempData->cel_evals[n ].sourceX = x + (i&1)*8;
261  tempData->cel_evals[n++].sourceY = y + (i&2)*4;
262  }
263 
264  return 0;
265 }
266 
267 /**
268  * Get macroblocks from parts of the image
269  */
270 static void get_frame_mb(const AVFrame *frame, int x, int y, uint8_t mb[], int dim)
271 {
272  int i, j, cp;
273 
274  for (cp=0; cp<3; cp++) {
275  int stride = frame->linesize[cp];
276  for (i=0; i<dim; i++)
277  for (j=0; j<dim; j++)
278  *mb++ = frame->data[cp][(y+i)*stride + x + j];
279  }
280 }
281 
282 /**
283  * Find the codebook with the lowest distortion from an image
284  */
285 static int index_mb(uint8_t cluster[], uint8_t cb[], int numCB,
286  int *outIndex, int dim)
287 {
288  int i, lDiff = INT_MAX, pick=0;
289 
290  /* Diff against the others */
291  for (i=0; i<numCB; i++) {
292  int diff = squared_diff_macroblock(cluster, cb + i*dim*dim*3, dim);
293  if (diff < lDiff) {
294  lDiff = diff;
295  pick = i;
296  }
297  }
298 
299  *outIndex = pick;
300  return lDiff;
301 }
302 
303 #define EVAL_MOTION(MOTION) \
304  do { \
305  diff = eval_motion_dist(enc, j, i, MOTION, blocksize); \
306  \
307  if (diff < lowestdiff) { \
308  lowestdiff = diff; \
309  bestpick = MOTION; \
310  } \
311  } while(0)
312 
313 static void motion_search(RoqContext *enc, int blocksize)
314 {
315  static const motion_vect offsets[8] = {
316  {{ 0,-1}},
317  {{ 0, 1}},
318  {{-1, 0}},
319  {{ 1, 0}},
320  {{-1, 1}},
321  {{ 1,-1}},
322  {{-1,-1}},
323  {{ 1, 1}},
324  };
325 
326  int diff, lowestdiff, oldbest;
327  int off[3];
328  motion_vect bestpick = {{0,0}};
329  int i, j, k, offset;
330 
331  motion_vect *last_motion;
332  motion_vect *this_motion;
333  motion_vect vect, vect2;
334 
335  int max=(enc->width/blocksize)*enc->height/blocksize;
336 
337  if (blocksize == 4) {
338  last_motion = enc->last_motion4;
339  this_motion = enc->this_motion4;
340  } else {
341  last_motion = enc->last_motion8;
342  this_motion = enc->this_motion8;
343  }
344 
345  for (i=0; i<enc->height; i+=blocksize)
346  for (j=0; j<enc->width; j+=blocksize) {
347  lowestdiff = eval_motion_dist(enc, j, i, (motion_vect) {{0,0}},
348  blocksize);
349  bestpick.d[0] = 0;
350  bestpick.d[1] = 0;
351 
352  if (blocksize == 4)
353  EVAL_MOTION(enc->this_motion8[(i/8)*(enc->width/8) + j/8]);
354 
355  offset = (i/blocksize)*enc->width/blocksize + j/blocksize;
356  if (offset < max && offset >= 0)
357  EVAL_MOTION(last_motion[offset]);
358 
359  offset++;
360  if (offset < max && offset >= 0)
361  EVAL_MOTION(last_motion[offset]);
362 
363  offset = (i/blocksize + 1)*enc->width/blocksize + j/blocksize;
364  if (offset < max && offset >= 0)
365  EVAL_MOTION(last_motion[offset]);
366 
367  off[0]= (i/blocksize)*enc->width/blocksize + j/blocksize - 1;
368  off[1]= off[0] - enc->width/blocksize + 1;
369  off[2]= off[1] + 1;
370 
371  if (i) {
372 
373  for(k=0; k<2; k++)
374  vect.d[k]= mid_pred(this_motion[off[0]].d[k],
375  this_motion[off[1]].d[k],
376  this_motion[off[2]].d[k]);
377 
378  EVAL_MOTION(vect);
379  for(k=0; k<3; k++)
380  EVAL_MOTION(this_motion[off[k]]);
381  } else if(j)
382  EVAL_MOTION(this_motion[off[0]]);
383 
384  vect = bestpick;
385 
386  oldbest = -1;
387  while (oldbest != lowestdiff) {
388  oldbest = lowestdiff;
389  for (k=0; k<8; k++) {
390  vect2 = vect;
391  vect2.d[0] += offsets[k].d[0];
392  vect2.d[1] += offsets[k].d[1];
393  EVAL_MOTION(vect2);
394  }
395  vect = bestpick;
396  }
397  offset = (i/blocksize)*enc->width/blocksize + j/blocksize;
398  this_motion[offset] = bestpick;
399  }
400 }
401 
402 /**
403  * Get distortion for all options available to a subcel
404  */
405 static void gather_data_for_subcel(SubcelEvaluation *subcel, int x,
406  int y, RoqContext *enc, RoqTempdata *tempData)
407 {
408  uint8_t mb4[4*4*3];
409  uint8_t mb2[2*2*3];
410  int cluster_index;
411  int i, best_dist;
412 
413  static const int bitsUsed[4] = {2, 10, 10, 34};
414 
415  if (enc->framesSinceKeyframe >= 1) {
416  subcel->motion = enc->this_motion4[y*enc->width/16 + x/4];
417 
418  subcel->eval_dist[RoQ_ID_FCC] =
419  eval_motion_dist(enc, x, y,
420  enc->this_motion4[y*enc->width/16 + x/4], 4);
421  } else
422  subcel->eval_dist[RoQ_ID_FCC] = INT_MAX;
423 
424  if (enc->framesSinceKeyframe >= 2)
426  enc->current_frame->data, x,
427  y, x, y,
428  enc->frame_to_enc->linesize,
429  enc->current_frame->linesize,
430  4);
431  else
432  subcel->eval_dist[RoQ_ID_MOT] = INT_MAX;
433 
434  cluster_index = y*enc->width/16 + x/4;
435 
436  get_frame_mb(enc->frame_to_enc, x, y, mb4, 4);
437 
438  subcel->eval_dist[RoQ_ID_SLD] = index_mb(mb4,
439  tempData->codebooks.unpacked_cb4,
440  tempData->codebooks.numCB4,
441  &subcel->cbEntry, 4);
442 
443  subcel->eval_dist[RoQ_ID_CCC] = 0;
444 
445  for(i=0;i<4;i++) {
446  subcel->subCels[i] = tempData->closest_cb2[cluster_index*4+i];
447 
448  get_frame_mb(enc->frame_to_enc, x+2*(i&1),
449  y+(i&2), mb2, 2);
450 
451  subcel->eval_dist[RoQ_ID_CCC] +=
452  squared_diff_macroblock(tempData->codebooks.unpacked_cb2 + subcel->subCels[i]*2*2*3, mb2, 2);
453  }
454 
455  best_dist = INT_MAX;
456  for (i=0; i<4; i++)
457  if (ROQ_LAMBDA_SCALE*subcel->eval_dist[i] + enc->lambda*bitsUsed[i] <
458  best_dist) {
459  subcel->best_coding = i;
460  subcel->best_bit_use = bitsUsed[i];
461  best_dist = ROQ_LAMBDA_SCALE*subcel->eval_dist[i] +
462  enc->lambda*bitsUsed[i];
463  }
464 }
465 
466 /**
467  * Get distortion for all options available to a cel
468  */
470  RoqTempdata *tempData)
471 {
472  uint8_t mb8[8*8*3];
473  int index = cel->sourceY*enc->width/64 + cel->sourceX/8;
474  int i, j, best_dist, divide_bit_use;
475 
476  int bitsUsed[4] = {2, 10, 10, 0};
477 
478  if (enc->framesSinceKeyframe >= 1) {
479  cel->motion = enc->this_motion8[index];
480 
481  cel->eval_dist[RoQ_ID_FCC] =
482  eval_motion_dist(enc, cel->sourceX, cel->sourceY,
483  enc->this_motion8[index], 8);
484  } else
485  cel->eval_dist[RoQ_ID_FCC] = INT_MAX;
486 
487  if (enc->framesSinceKeyframe >= 2)
489  enc->current_frame->data,
490  cel->sourceX, cel->sourceY,
491  cel->sourceX, cel->sourceY,
492  enc->frame_to_enc->linesize,
493  enc->current_frame->linesize,8);
494  else
495  cel->eval_dist[RoQ_ID_MOT] = INT_MAX;
496 
497  get_frame_mb(enc->frame_to_enc, cel->sourceX, cel->sourceY, mb8, 8);
498 
499  cel->eval_dist[RoQ_ID_SLD] =
501  tempData->codebooks.numCB4, &cel->cbEntry, 8);
502 
503  gather_data_for_subcel(cel->subCels + 0, cel->sourceX+0, cel->sourceY+0, enc, tempData);
504  gather_data_for_subcel(cel->subCels + 1, cel->sourceX+4, cel->sourceY+0, enc, tempData);
505  gather_data_for_subcel(cel->subCels + 2, cel->sourceX+0, cel->sourceY+4, enc, tempData);
506  gather_data_for_subcel(cel->subCels + 3, cel->sourceX+4, cel->sourceY+4, enc, tempData);
507 
508  cel->eval_dist[RoQ_ID_CCC] = 0;
509  divide_bit_use = 0;
510  for (i=0; i<4; i++) {
511  cel->eval_dist[RoQ_ID_CCC] +=
512  cel->subCels[i].eval_dist[cel->subCels[i].best_coding];
513  divide_bit_use += cel->subCels[i].best_bit_use;
514  }
515 
516  best_dist = INT_MAX;
517  bitsUsed[3] = 2 + divide_bit_use;
518 
519  for (i=0; i<4; i++)
520  if (ROQ_LAMBDA_SCALE*cel->eval_dist[i] + enc->lambda*bitsUsed[i] <
521  best_dist) {
522  cel->best_coding = i;
523  best_dist = ROQ_LAMBDA_SCALE*cel->eval_dist[i] +
524  enc->lambda*bitsUsed[i];
525  }
526 
527  tempData->used_option[cel->best_coding]++;
528  tempData->mainChunkSize += bitsUsed[cel->best_coding];
529 
530  if (cel->best_coding == RoQ_ID_SLD)
531  tempData->codebooks.usedCB4[cel->cbEntry]++;
532 
533  if (cel->best_coding == RoQ_ID_CCC)
534  for (i=0; i<4; i++) {
535  if (cel->subCels[i].best_coding == RoQ_ID_SLD)
536  tempData->codebooks.usedCB4[cel->subCels[i].cbEntry]++;
537  else if (cel->subCels[i].best_coding == RoQ_ID_CCC)
538  for (j=0; j<4; j++)
539  tempData->codebooks.usedCB2[cel->subCels[i].subCels[j]]++;
540  }
541 }
542 
543 static void remap_codebooks(RoqContext *enc, RoqTempdata *tempData)
544 {
545  int i, j, idx=0;
546 
547  /* Make remaps for the final codebook usage */
548  for (i=0; i<(enc->quake3_compat ? MAX_CBS_4x4-1 : MAX_CBS_4x4); i++) {
549  if (tempData->codebooks.usedCB4[i]) {
550  tempData->i2f4[i] = idx;
551  tempData->f2i4[idx] = i;
552  for (j=0; j<4; j++)
553  tempData->codebooks.usedCB2[enc->cb4x4[i].idx[j]]++;
554  idx++;
555  }
556  }
557 
558  tempData->numCB4 = idx;
559 
560  idx = 0;
561  for (i=0; i<MAX_CBS_2x2; i++) {
562  if (tempData->codebooks.usedCB2[i]) {
563  tempData->i2f2[i] = idx;
564  tempData->f2i2[idx] = i;
565  idx++;
566  }
567  }
568  tempData->numCB2 = idx;
569 
570 }
571 
572 /**
573  * Write codebook chunk
574  */
575 static void write_codebooks(RoqContext *enc, RoqTempdata *tempData)
576 {
577  int i, j;
578  uint8_t **outp= &enc->out_buf;
579 
580  if (tempData->numCB2) {
581  bytestream_put_le16(outp, RoQ_QUAD_CODEBOOK);
582  bytestream_put_le32(outp, tempData->numCB2*6 + tempData->numCB4*4);
583  bytestream_put_byte(outp, tempData->numCB4);
584  bytestream_put_byte(outp, tempData->numCB2);
585 
586  for (i=0; i<tempData->numCB2; i++) {
587  bytestream_put_buffer(outp, enc->cb2x2[tempData->f2i2[i]].y, 4);
588  bytestream_put_byte(outp, enc->cb2x2[tempData->f2i2[i]].u);
589  bytestream_put_byte(outp, enc->cb2x2[tempData->f2i2[i]].v);
590  }
591 
592  for (i=0; i<tempData->numCB4; i++)
593  for (j=0; j<4; j++)
594  bytestream_put_byte(outp, tempData->i2f2[enc->cb4x4[tempData->f2i4[i]].idx[j]]);
595 
596  }
597 }
598 
599 static inline uint8_t motion_arg(motion_vect mot)
600 {
601  uint8_t ax = 8 - ((uint8_t) mot.d[0]);
602  uint8_t ay = 8 - ((uint8_t) mot.d[1]);
603  return ((ax&15)<<4) | (ay&15);
604 }
605 
606 typedef struct
607 {
610  uint8_t argumentSpool[64];
613 } CodingSpool;
614 
615 /* NOTE: Typecodes must be spooled AFTER arguments!! */
617 {
618  s->typeSpool |= (type & 3) << (14 - s->typeSpoolLength);
619  s->typeSpoolLength += 2;
620  if (s->typeSpoolLength == 16) {
621  bytestream_put_le16(s->pout, s->typeSpool);
623  s->args - s->argumentSpool);
624  s->typeSpoolLength = 0;
625  s->typeSpool = 0;
626  s->args = s->argumentSpool;
627  }
628 }
629 
630 static void reconstruct_and_encode_image(RoqContext *enc, RoqTempdata *tempData, int w, int h, int numBlocks)
631 {
632  int i, j, k;
633  int x, y;
634  int subX, subY;
635  int dist=0;
636 
637  roq_qcell *qcell;
638  CelEvaluation *eval;
639 
640  CodingSpool spool;
641 
642  spool.typeSpool=0;
643  spool.typeSpoolLength=0;
644  spool.args = spool.argumentSpool;
645  spool.pout = &enc->out_buf;
646 
647  if (tempData->used_option[RoQ_ID_CCC]%2)
648  tempData->mainChunkSize+=8; //FIXME
649 
650  /* Write the video chunk header */
651  bytestream_put_le16(&enc->out_buf, RoQ_QUAD_VQ);
652  bytestream_put_le32(&enc->out_buf, tempData->mainChunkSize/8);
653  bytestream_put_byte(&enc->out_buf, 0x0);
654  bytestream_put_byte(&enc->out_buf, 0x0);
655 
656  for (i=0; i<numBlocks; i++) {
657  eval = tempData->cel_evals + i;
658 
659  x = eval->sourceX;
660  y = eval->sourceY;
661  dist += eval->eval_dist[eval->best_coding];
662 
663  switch (eval->best_coding) {
664  case RoQ_ID_MOT:
665  write_typecode(&spool, RoQ_ID_MOT);
666  break;
667 
668  case RoQ_ID_FCC:
669  bytestream_put_byte(&spool.args, motion_arg(eval->motion));
670 
671  write_typecode(&spool, RoQ_ID_FCC);
672  ff_apply_motion_8x8(enc, x, y,
673  eval->motion.d[0], eval->motion.d[1]);
674  break;
675 
676  case RoQ_ID_SLD:
677  bytestream_put_byte(&spool.args, tempData->i2f4[eval->cbEntry]);
678  write_typecode(&spool, RoQ_ID_SLD);
679 
680  qcell = enc->cb4x4 + eval->cbEntry;
681  ff_apply_vector_4x4(enc, x , y , enc->cb2x2 + qcell->idx[0]);
682  ff_apply_vector_4x4(enc, x+4, y , enc->cb2x2 + qcell->idx[1]);
683  ff_apply_vector_4x4(enc, x , y+4, enc->cb2x2 + qcell->idx[2]);
684  ff_apply_vector_4x4(enc, x+4, y+4, enc->cb2x2 + qcell->idx[3]);
685  break;
686 
687  case RoQ_ID_CCC:
688  write_typecode(&spool, RoQ_ID_CCC);
689 
690  for (j=0; j<4; j++) {
691  subX = x + 4*(j&1);
692  subY = y + 2*(j&2);
693 
694  switch(eval->subCels[j].best_coding) {
695  case RoQ_ID_MOT:
696  break;
697 
698  case RoQ_ID_FCC:
699  bytestream_put_byte(&spool.args,
700  motion_arg(eval->subCels[j].motion));
701 
702  ff_apply_motion_4x4(enc, subX, subY,
703  eval->subCels[j].motion.d[0],
704  eval->subCels[j].motion.d[1]);
705  break;
706 
707  case RoQ_ID_SLD:
708  bytestream_put_byte(&spool.args,
709  tempData->i2f4[eval->subCels[j].cbEntry]);
710 
711  qcell = enc->cb4x4 + eval->subCels[j].cbEntry;
712 
713  ff_apply_vector_2x2(enc, subX , subY ,
714  enc->cb2x2 + qcell->idx[0]);
715  ff_apply_vector_2x2(enc, subX+2, subY ,
716  enc->cb2x2 + qcell->idx[1]);
717  ff_apply_vector_2x2(enc, subX , subY+2,
718  enc->cb2x2 + qcell->idx[2]);
719  ff_apply_vector_2x2(enc, subX+2, subY+2,
720  enc->cb2x2 + qcell->idx[3]);
721  break;
722 
723  case RoQ_ID_CCC:
724  for (k=0; k<4; k++) {
725  int cb_idx = eval->subCels[j].subCels[k];
726  bytestream_put_byte(&spool.args,
727  tempData->i2f2[cb_idx]);
728 
729  ff_apply_vector_2x2(enc, subX + 2*(k&1), subY + (k&2),
730  enc->cb2x2 + cb_idx);
731  }
732  break;
733  }
734  write_typecode(&spool, eval->subCels[j].best_coding);
735  }
736  break;
737  }
738  }
739 
740  /* Flush the remainder of the argument/type spool */
741  while (spool.typeSpoolLength)
742  write_typecode(&spool, 0x0);
743 
744 #if 0
745  uint8_t *fdata[3] = {enc->frame_to_enc->data[0],
746  enc->frame_to_enc->data[1],
747  enc->frame_to_enc->data[2]};
748  uint8_t *cdata[3] = {enc->current_frame->data[0],
749  enc->current_frame->data[1],
750  enc->current_frame->data[2]};
751  av_log(enc->avctx, AV_LOG_ERROR, "Expected distortion: %i Actual: %i\n",
752  dist,
753  block_sse(fdata, cdata, 0, 0, 0, 0,
754  enc->frame_to_enc->linesize,
755  enc->current_frame->linesize,
756  enc->width)); //WARNING: Square dimensions implied...
757 #endif
758 }
759 
760 
761 /**
762  * Create a single YUV cell from a 2x2 section of the image
763  */
764 static inline void frame_block_to_cell(uint8_t *block, uint8_t * const *data,
765  int top, int left, const int *stride)
766 {
767  int i, j, u=0, v=0;
768 
769  for (i=0; i<2; i++)
770  for (j=0; j<2; j++) {
771  int x = (top+i)*stride[0] + left + j;
772  *block++ = data[0][x];
773  x = (top+i)*stride[1] + left + j;
774  u += data[1][x];
775  v += data[2][x];
776  }
777 
778  *block++ = (u+2)/4;
779  *block++ = (v+2)/4;
780 }
781 
782 /**
783  * Create YUV clusters for the entire image
784  */
785 static void create_clusters(const AVFrame *frame, int w, int h, uint8_t *yuvClusters)
786 {
787  int i, j, k, l;
788 
789  for (i=0; i<h; i+=4)
790  for (j=0; j<w; j+=4) {
791  for (k=0; k < 2; k++)
792  for (l=0; l < 2; l++)
793  frame_block_to_cell(yuvClusters + (l + 2*k)*6, frame->data,
794  i+2*k, j+2*l, frame->linesize);
795  yuvClusters += 24;
796  }
797 }
798 
799 static int generate_codebook(RoqContext *enc, RoqTempdata *tempdata,
800  int *points, int inputCount, roq_cell *results,
801  int size, int cbsize)
802 {
803  int i, j, k, ret = 0;
804  int c_size = size*size/4;
805  int *buf;
806  int *codebook = av_malloc_array(6*c_size, cbsize*sizeof(int));
807  int *closest_cb;
808 
809  if (!codebook)
810  return AVERROR(ENOMEM);
811 
812  if (size == 4) {
813  closest_cb = av_malloc_array(6*c_size, inputCount*sizeof(int));
814  if (!closest_cb) {
815  ret = AVERROR(ENOMEM);
816  goto out;
817  }
818  } else
819  closest_cb = tempdata->closest_cb2;
820 
821  ret = avpriv_init_elbg(points, 6 * c_size, inputCount, codebook,
822  cbsize, 1, closest_cb, &enc->randctx);
823  if (ret < 0)
824  goto out;
825  ret = avpriv_do_elbg(points, 6 * c_size, inputCount, codebook,
826  cbsize, 1, closest_cb, &enc->randctx);
827  if (ret < 0)
828  goto out;
829 
830  buf = codebook;
831  for (i=0; i<cbsize; i++)
832  for (k=0; k<c_size; k++) {
833  for(j=0; j<4; j++)
834  results->y[j] = *buf++;
835 
836  results->u = (*buf++ + CHROMA_BIAS/2)/CHROMA_BIAS;
837  results->v = (*buf++ + CHROMA_BIAS/2)/CHROMA_BIAS;
838  results++;
839  }
840 out:
841  if (size == 4)
842  av_free(closest_cb);
843  av_free(codebook);
844  return ret;
845 }
846 
847 static int generate_new_codebooks(RoqContext *enc, RoqTempdata *tempData)
848 {
849  int i, j, ret = 0;
850  RoqCodebooks *codebooks = &tempData->codebooks;
851  int max = enc->width*enc->height/16;
852  uint8_t mb2[3*4];
853  roq_cell *results4 = av_malloc(sizeof(roq_cell)*MAX_CBS_4x4*4);
854  uint8_t *yuvClusters=av_malloc_array(max, sizeof(int)*6*4);
855  int *points = av_malloc_array(max, 6*4*sizeof(int));
856  int bias;
857 
858  if (!results4 || !yuvClusters || !points) {
859  ret = AVERROR(ENOMEM);
860  goto out;
861  }
862 
863  /* Subsample YUV data */
864  create_clusters(enc->frame_to_enc, enc->width, enc->height, yuvClusters);
865 
866  /* Cast to integer and apply chroma bias */
867  for (i=0; i<max*24; i++) {
868  bias = ((i%6)<4) ? 1 : CHROMA_BIAS;
869  points[i] = bias*yuvClusters[i];
870  }
871 
872  /* Create 4x4 codebooks */
873  if ((ret = generate_codebook(enc, tempData, points, max,
874  results4, 4, (enc->quake3_compat ? MAX_CBS_4x4-1 : MAX_CBS_4x4))) < 0)
875  goto out;
876 
877  codebooks->numCB4 = (enc->quake3_compat ? MAX_CBS_4x4-1 : MAX_CBS_4x4);
878 
879  tempData->closest_cb2 = av_malloc_array(max, 4*sizeof(int));
880  if (!tempData->closest_cb2) {
881  ret = AVERROR(ENOMEM);
882  goto out;
883  }
884 
885  /* Create 2x2 codebooks */
886  if ((ret = generate_codebook(enc, tempData, points, max * 4,
887  enc->cb2x2, 2, MAX_CBS_2x2)) < 0)
888  goto out;
889 
890  codebooks->numCB2 = MAX_CBS_2x2;
891 
892  /* Unpack 2x2 codebook clusters */
893  for (i=0; i<codebooks->numCB2; i++)
894  unpack_roq_cell(enc->cb2x2 + i, codebooks->unpacked_cb2 + i*2*2*3);
895 
896  /* Index all 4x4 entries to the 2x2 entries, unpack, and enlarge */
897  for (i=0; i<codebooks->numCB4; i++) {
898  for (j=0; j<4; j++) {
899  unpack_roq_cell(&results4[4*i + j], mb2);
900  index_mb(mb2, codebooks->unpacked_cb2, codebooks->numCB2,
901  &enc->cb4x4[i].idx[j], 2);
902  }
903  unpack_roq_qcell(codebooks->unpacked_cb2, enc->cb4x4 + i,
904  codebooks->unpacked_cb4 + i*4*4*3);
905  enlarge_roq_mb4(codebooks->unpacked_cb4 + i*4*4*3,
906  codebooks->unpacked_cb4_enlarged + i*8*8*3);
907  }
908 out:
909  av_free(yuvClusters);
910  av_free(points);
911  av_free(results4);
912  return ret;
913 }
914 
915 static int roq_encode_video(RoqContext *enc)
916 {
917  RoqTempdata *tempData = enc->tmpData;
918  int i, ret;
919 
920  memset(tempData, 0, sizeof(*tempData));
921 
922  ret = create_cel_evals(enc, tempData);
923  if (ret < 0)
924  return ret;
925 
926  ret = generate_new_codebooks(enc, tempData);
927  if (ret < 0)
928  return ret;
929 
930  if (enc->framesSinceKeyframe >= 1) {
931  motion_search(enc, 8);
932  motion_search(enc, 4);
933  }
934 
935  retry_encode:
936  for (i=0; i<enc->width*enc->height/64; i++)
937  gather_data_for_cel(tempData->cel_evals + i, enc, tempData);
938 
939  /* Quake 3 can't handle chunks bigger than 65535 bytes */
940  if (tempData->mainChunkSize/8 > 65535 && enc->quake3_compat) {
941  if (enc->lambda > 100000) {
942  av_log(enc->avctx, AV_LOG_ERROR, "Cannot encode video in Quake compatible form\n");
943  return AVERROR(EINVAL);
944  }
945  av_log(enc->avctx, AV_LOG_ERROR,
946  "Warning, generated a frame too big for Quake (%d > 65535), "
947  "now switching to a bigger qscale value.\n",
948  tempData->mainChunkSize/8);
949  enc->lambda *= 1.5;
950  tempData->mainChunkSize = 0;
951  memset(tempData->used_option, 0, sizeof(tempData->used_option));
952  memset(tempData->codebooks.usedCB4, 0,
953  sizeof(tempData->codebooks.usedCB4));
954  memset(tempData->codebooks.usedCB2, 0,
955  sizeof(tempData->codebooks.usedCB2));
956 
957  goto retry_encode;
958  }
959 
960  remap_codebooks(enc, tempData);
961 
962  write_codebooks(enc, tempData);
963 
964  reconstruct_and_encode_image(enc, tempData, enc->width, enc->height,
965  enc->width*enc->height/64);
966 
967  enc->avctx->coded_frame = enc->current_frame;
968 
969  /* Rotate frame history */
970  FFSWAP(AVFrame *, enc->current_frame, enc->last_frame);
973 
974  av_freep(&tempData->cel_evals);
975  av_freep(&tempData->closest_cb2);
976 
977  enc->framesSinceKeyframe++;
978 
979  return 0;
980 }
981 
983 {
984  RoqContext *enc = avctx->priv_data;
985 
987  av_frame_free(&enc->last_frame);
988 
989  av_freep(&enc->tmpData);
990  av_freep(&enc->this_motion4);
991  av_freep(&enc->last_motion4);
992  av_freep(&enc->this_motion8);
993  av_freep(&enc->last_motion8);
994 
995  return 0;
996 }
997 
999 {
1000  RoqContext *enc = avctx->priv_data;
1001 
1002  av_lfg_init(&enc->randctx, 1);
1003 
1004  enc->framesSinceKeyframe = 0;
1005  if ((avctx->width & 0xf) || (avctx->height & 0xf)) {
1006  av_log(avctx, AV_LOG_ERROR, "Dimensions must be divisible by 16\n");
1007  return AVERROR(EINVAL);
1008  }
1009 
1010  if (avctx->width > 65535 || avctx->height > 65535) {
1011  av_log(avctx, AV_LOG_ERROR, "Dimensions are max %d\n", enc->quake3_compat ? 32768 : 65535);
1012  return AVERROR(EINVAL);
1013  }
1014 
1015  if (((avctx->width)&(avctx->width-1))||((avctx->height)&(avctx->height-1)))
1016  av_log(avctx, AV_LOG_ERROR, "Warning: dimensions not power of two, this is not supported by quake\n");
1017 
1018  enc->width = avctx->width;
1019  enc->height = avctx->height;
1020 
1021  enc->framesSinceKeyframe = 0;
1022  enc->first_frame = 1;
1023 
1024  enc->last_frame = av_frame_alloc();
1025  enc->current_frame = av_frame_alloc();
1026  if (!enc->last_frame || !enc->current_frame) {
1027  roq_encode_end(avctx);
1028  return AVERROR(ENOMEM);
1029  }
1030 
1031  enc->tmpData = av_malloc(sizeof(RoqTempdata));
1032 
1033  enc->this_motion4 =
1034  av_mallocz_array((enc->width*enc->height/16), sizeof(motion_vect));
1035 
1036  enc->last_motion4 =
1037  av_malloc_array ((enc->width*enc->height/16), sizeof(motion_vect));
1038 
1039  enc->this_motion8 =
1040  av_mallocz_array((enc->width*enc->height/64), sizeof(motion_vect));
1041 
1042  enc->last_motion8 =
1043  av_malloc_array ((enc->width*enc->height/64), sizeof(motion_vect));
1044 
1045  if (!enc->tmpData || !enc->this_motion4 || !enc->last_motion4 ||
1046  !enc->this_motion8 || !enc->last_motion8) {
1047  roq_encode_end(avctx);
1048  return AVERROR(ENOMEM);
1049  }
1050 
1051  return 0;
1052 }
1053 
1055 {
1056  /* ROQ info chunk */
1057  bytestream_put_le16(&enc->out_buf, RoQ_INFO);
1058 
1059  /* Size: 8 bytes */
1060  bytestream_put_le32(&enc->out_buf, 8);
1061 
1062  /* Unused argument */
1063  bytestream_put_byte(&enc->out_buf, 0x00);
1064  bytestream_put_byte(&enc->out_buf, 0x00);
1065 
1066  /* Width */
1067  bytestream_put_le16(&enc->out_buf, enc->width);
1068 
1069  /* Height */
1070  bytestream_put_le16(&enc->out_buf, enc->height);
1071 
1072  /* Unused in Quake 3, mimics the output of the real encoder */
1073  bytestream_put_byte(&enc->out_buf, 0x08);
1074  bytestream_put_byte(&enc->out_buf, 0x00);
1075  bytestream_put_byte(&enc->out_buf, 0x04);
1076  bytestream_put_byte(&enc->out_buf, 0x00);
1077 }
1078 
1080  const AVFrame *frame, int *got_packet)
1081 {
1082  RoqContext *enc = avctx->priv_data;
1083  int size, ret;
1084 
1085  enc->avctx = avctx;
1086 
1087  enc->frame_to_enc = frame;
1088 
1089  if (frame->quality)
1090  enc->lambda = frame->quality - 1;
1091  else
1092  enc->lambda = 2*ROQ_LAMBDA_SCALE;
1093 
1094  /* 138 bits max per 8x8 block +
1095  * 256 codebooks*(6 bytes 2x2 + 4 bytes 4x4) + 8 bytes frame header */
1096  size = ((enc->width * enc->height / 64) * 138 + 7) / 8 + 256 * (6 + 4) + 8;
1097  if ((ret = ff_alloc_packet2(avctx, pkt, size)) < 0)
1098  return ret;
1099  enc->out_buf = pkt->data;
1100 
1101  /* Check for I frame */
1102  if (enc->framesSinceKeyframe == avctx->gop_size)
1103  enc->framesSinceKeyframe = 0;
1104 
1105  if (enc->first_frame) {
1106  /* Alloc memory for the reconstruction data (we must know the stride
1107  for that) */
1108  if ((ret = ff_get_buffer(avctx, enc->current_frame, 0)) < 0 ||
1109  (ret = ff_get_buffer(avctx, enc->last_frame, 0)) < 0)
1110  return ret;
1111 
1112  /* Before the first video frame, write a "video info" chunk */
1114 
1115  enc->first_frame = 0;
1116  }
1117 
1118  /* Encode the actual frame */
1119  ret = roq_encode_video(enc);
1120  if (ret < 0)
1121  return ret;
1122 
1123  pkt->size = enc->out_buf - pkt->data;
1124  if (enc->framesSinceKeyframe == 1)
1125  pkt->flags |= AV_PKT_FLAG_KEY;
1126  *got_packet = 1;
1127 
1128  return 0;
1129 }
1130 
1131 #define OFFSET(x) offsetof(RoqContext, x)
1132 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1133 static const AVOption options[] = {
1134  { "quake3_compat", "Whether to respect known limitations in Quake 3 decoder", OFFSET(quake3_compat), AV_OPT_TYPE_INT, { .i64 = 1 }, 0, 1, VE },
1135  { NULL },
1136 };
1137 
1138 static const AVClass roq_class = {
1139  .class_name = "RoQ",
1140  .item_name = av_default_item_name,
1141  .option = options,
1142  .version = LIBAVUTIL_VERSION_INT,
1143 };
1144 
1146  .name = "roqvideo",
1147  .long_name = NULL_IF_CONFIG_SMALL("id RoQ video"),
1148  .type = AVMEDIA_TYPE_VIDEO,
1149  .id = AV_CODEC_ID_ROQ,
1150  .priv_data_size = sizeof(RoqContext),
1151  .init = roq_encode_init,
1152  .encode2 = roq_encode_frame,
1153  .close = roq_encode_end,
1154  .pix_fmts = (const enum AVPixelFormat[]){ AV_PIX_FMT_YUVJ444P,
1155  AV_PIX_FMT_NONE },
1156  .priv_class = &roq_class,
1157 };