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 "roqvideo.h"
61 #include "bytestream.h"
62 #include "elbg.h"
63 #include "internal.h"
64 #include "mathops.h"
65 
66 #define CHROMA_BIAS 1
67 
68 /**
69  * Maximum number of generated 4x4 codebooks. Can't be 256 to workaround a
70  * Quake 3 bug.
71  */
72 #define MAX_CBS_4x4 255
73 
74 #define MAX_CBS_2x2 256 ///< Maximum number of 2x2 codebooks.
75 
76 /* The cast is useful when multiplying it by INT_MAX */
77 #define ROQ_LAMBDA_SCALE ((uint64_t) FF_LAMBDA_SCALE)
78 
79 /* Macroblock support functions */
80 static void unpack_roq_cell(roq_cell *cell, uint8_t u[4*3])
81 {
82  memcpy(u , cell->y, 4);
83  memset(u+4, cell->u, 4);
84  memset(u+8, cell->v, 4);
85 }
86 
87 static void unpack_roq_qcell(uint8_t cb2[], roq_qcell *qcell, uint8_t u[4*4*3])
88 {
89  int i,cp;
90  static const int offsets[4] = {0, 2, 8, 10};
91 
92  for (cp=0; cp<3; cp++)
93  for (i=0; i<4; i++) {
94  u[4*4*cp + offsets[i] ] = cb2[qcell->idx[i]*2*2*3 + 4*cp ];
95  u[4*4*cp + offsets[i]+1] = cb2[qcell->idx[i]*2*2*3 + 4*cp+1];
96  u[4*4*cp + offsets[i]+4] = cb2[qcell->idx[i]*2*2*3 + 4*cp+2];
97  u[4*4*cp + offsets[i]+5] = cb2[qcell->idx[i]*2*2*3 + 4*cp+3];
98  }
99 }
100 
101 
102 static void enlarge_roq_mb4(uint8_t base[3*16], uint8_t u[3*64])
103 {
104  int x,y,cp;
105 
106  for(cp=0; cp<3; cp++)
107  for(y=0; y<8; y++)
108  for(x=0; x<8; x++)
109  *u++ = base[(y/2)*4 + (x/2) + 16*cp];
110 }
111 
112 static inline int square(int x)
113 {
114  return x*x;
115 }
116 
117 static inline int eval_sse(const uint8_t *a, const uint8_t *b, int count)
118 {
119  int diff=0;
120 
121  while(count--)
122  diff += square(*b++ - *a++);
123 
124  return diff;
125 }
126 
127 // FIXME Could use DSPContext.sse, but it is not so speed critical (used
128 // just for motion estimation).
129 static int block_sse(uint8_t * const *buf1, uint8_t * const *buf2, int x1, int y1,
130  int x2, int y2, const int *stride1, const int *stride2, int size)
131 {
132  int i, k;
133  int sse=0;
134 
135  for (k=0; k<3; k++) {
136  int bias = (k ? CHROMA_BIAS : 4);
137  for (i=0; i<size; i++)
138  sse += bias*eval_sse(buf1[k] + (y1+i)*stride1[k] + x1,
139  buf2[k] + (y2+i)*stride2[k] + x2, size);
140  }
141 
142  return sse;
143 }
144 
145 static int eval_motion_dist(RoqContext *enc, int x, int y, motion_vect vect,
146  int size)
147 {
148  int mx=vect.d[0];
149  int my=vect.d[1];
150 
151  if (mx < -7 || mx > 7)
152  return INT_MAX;
153 
154  if (my < -7 || my > 7)
155  return INT_MAX;
156 
157  mx += x;
158  my += y;
159 
160  if ((unsigned) mx > enc->width-size || (unsigned) my > enc->height-size)
161  return INT_MAX;
162 
163  return block_sse(enc->frame_to_enc->data, enc->last_frame->data, x, y,
164  mx, my,
166  size);
167 }
168 
169 /**
170  * @return distortion between two macroblocks
171  */
172 static inline int squared_diff_macroblock(uint8_t a[], uint8_t b[], int size)
173 {
174  int cp, sdiff=0;
175 
176  for(cp=0;cp<3;cp++) {
177  int bias = (cp ? CHROMA_BIAS : 4);
178  sdiff += bias*eval_sse(a, b, size*size);
179  a += size*size;
180  b += size*size;
181  }
182 
183  return sdiff;
184 }
185 
186 typedef struct
187 {
188  int eval_dist[4];
191 
192  int subCels[4];
194  int cbEntry;
196 
197 typedef struct
198 {
199  int eval_dist[4];
201 
202  SubcelEvaluation subCels[4];
203 
205  int cbEntry;
206 
207  int sourceX, sourceY;
208 } CelEvaluation;
209 
210 typedef struct
211 {
212  int numCB4;
213  int numCB2;
214  int usedCB2[MAX_CBS_2x2];
215  int usedCB4[MAX_CBS_4x4];
216  uint8_t unpacked_cb2[MAX_CBS_2x2*2*2*3];
217  uint8_t unpacked_cb4[MAX_CBS_4x4*4*4*3];
218  uint8_t unpacked_cb4_enlarged[MAX_CBS_4x4*8*8*3];
219 } RoqCodebooks;
220 
221 /**
222  * Temporary vars
223  */
224 typedef struct RoqTempData
225 {
227 
228  int f2i4[MAX_CBS_4x4];
229  int i2f4[MAX_CBS_4x4];
230  int f2i2[MAX_CBS_2x2];
231  int i2f2[MAX_CBS_2x2];
232 
234 
235  int numCB4;
236  int numCB2;
237 
239 
241  int used_option[4];
242 } RoqTempdata;
243 
244 /**
245  * Initialize cel evaluators and set their source coordinates
246  */
247 static void create_cel_evals(RoqContext *enc, RoqTempdata *tempData)
248 {
249  int n=0, x, y, i;
250 
251  tempData->cel_evals = av_malloc(enc->width*enc->height/64 * sizeof(CelEvaluation));
252 
253  /* Map to the ROQ quadtree order */
254  for (y=0; y<enc->height; y+=16)
255  for (x=0; x<enc->width; x+=16)
256  for(i=0; i<4; i++) {
257  tempData->cel_evals[n ].sourceX = x + (i&1)*8;
258  tempData->cel_evals[n++].sourceY = y + (i&2)*4;
259  }
260 }
261 
262 /**
263  * Get macroblocks from parts of the image
264  */
265 static void get_frame_mb(const AVFrame *frame, int x, int y, uint8_t mb[], int dim)
266 {
267  int i, j, cp;
268 
269  for (cp=0; cp<3; cp++) {
270  int stride = frame->linesize[cp];
271  for (i=0; i<dim; i++)
272  for (j=0; j<dim; j++)
273  *mb++ = frame->data[cp][(y+i)*stride + x + j];
274  }
275 }
276 
277 /**
278  * Find the codebook with the lowest distortion from an image
279  */
280 static int index_mb(uint8_t cluster[], uint8_t cb[], int numCB,
281  int *outIndex, int dim)
282 {
283  int i, lDiff = INT_MAX, pick=0;
284 
285  /* Diff against the others */
286  for (i=0; i<numCB; i++) {
287  int diff = squared_diff_macroblock(cluster, cb + i*dim*dim*3, dim);
288  if (diff < lDiff) {
289  lDiff = diff;
290  pick = i;
291  }
292  }
293 
294  *outIndex = pick;
295  return lDiff;
296 }
297 
298 #define EVAL_MOTION(MOTION) \
299  do { \
300  diff = eval_motion_dist(enc, j, i, MOTION, blocksize); \
301  \
302  if (diff < lowestdiff) { \
303  lowestdiff = diff; \
304  bestpick = MOTION; \
305  } \
306  } while(0)
307 
308 static void motion_search(RoqContext *enc, int blocksize)
309 {
310  static const motion_vect offsets[8] = {
311  {{ 0,-1}},
312  {{ 0, 1}},
313  {{-1, 0}},
314  {{ 1, 0}},
315  {{-1, 1}},
316  {{ 1,-1}},
317  {{-1,-1}},
318  {{ 1, 1}},
319  };
320 
321  int diff, lowestdiff, oldbest;
322  int off[3];
323  motion_vect bestpick = {{0,0}};
324  int i, j, k, offset;
325 
326  motion_vect *last_motion;
327  motion_vect *this_motion;
328  motion_vect vect, vect2;
329 
330  int max=(enc->width/blocksize)*enc->height/blocksize;
331 
332  if (blocksize == 4) {
333  last_motion = enc->last_motion4;
334  this_motion = enc->this_motion4;
335  } else {
336  last_motion = enc->last_motion8;
337  this_motion = enc->this_motion8;
338  }
339 
340  for (i=0; i<enc->height; i+=blocksize)
341  for (j=0; j<enc->width; j+=blocksize) {
342  lowestdiff = eval_motion_dist(enc, j, i, (motion_vect) {{0,0}},
343  blocksize);
344  bestpick.d[0] = 0;
345  bestpick.d[1] = 0;
346 
347  if (blocksize == 4)
348  EVAL_MOTION(enc->this_motion8[(i/8)*(enc->width/8) + j/8]);
349 
350  offset = (i/blocksize)*enc->width/blocksize + j/blocksize;
351  if (offset < max && offset >= 0)
352  EVAL_MOTION(last_motion[offset]);
353 
354  offset++;
355  if (offset < max && offset >= 0)
356  EVAL_MOTION(last_motion[offset]);
357 
358  offset = (i/blocksize + 1)*enc->width/blocksize + j/blocksize;
359  if (offset < max && offset >= 0)
360  EVAL_MOTION(last_motion[offset]);
361 
362  off[0]= (i/blocksize)*enc->width/blocksize + j/blocksize - 1;
363  off[1]= off[0] - enc->width/blocksize + 1;
364  off[2]= off[1] + 1;
365 
366  if (i) {
367 
368  for(k=0; k<2; k++)
369  vect.d[k]= mid_pred(this_motion[off[0]].d[k],
370  this_motion[off[1]].d[k],
371  this_motion[off[2]].d[k]);
372 
373  EVAL_MOTION(vect);
374  for(k=0; k<3; k++)
375  EVAL_MOTION(this_motion[off[k]]);
376  } else if(j)
377  EVAL_MOTION(this_motion[off[0]]);
378 
379  vect = bestpick;
380 
381  oldbest = -1;
382  while (oldbest != lowestdiff) {
383  oldbest = lowestdiff;
384  for (k=0; k<8; k++) {
385  vect2 = vect;
386  vect2.d[0] += offsets[k].d[0];
387  vect2.d[1] += offsets[k].d[1];
388  EVAL_MOTION(vect2);
389  }
390  vect = bestpick;
391  }
392  offset = (i/blocksize)*enc->width/blocksize + j/blocksize;
393  this_motion[offset] = bestpick;
394  }
395 }
396 
397 /**
398  * Get distortion for all options available to a subcel
399  */
400 static void gather_data_for_subcel(SubcelEvaluation *subcel, int x,
401  int y, RoqContext *enc, RoqTempdata *tempData)
402 {
403  uint8_t mb4[4*4*3];
404  uint8_t mb2[2*2*3];
405  int cluster_index;
406  int i, best_dist;
407 
408  static const int bitsUsed[4] = {2, 10, 10, 34};
409 
410  if (enc->framesSinceKeyframe >= 1) {
411  subcel->motion = enc->this_motion4[y*enc->width/16 + x/4];
412 
413  subcel->eval_dist[RoQ_ID_FCC] =
414  eval_motion_dist(enc, x, y,
415  enc->this_motion4[y*enc->width/16 + x/4], 4);
416  } else
417  subcel->eval_dist[RoQ_ID_FCC] = INT_MAX;
418 
419  if (enc->framesSinceKeyframe >= 2)
421  enc->current_frame->data, x,
422  y, x, y,
423  enc->frame_to_enc->linesize,
424  enc->current_frame->linesize,
425  4);
426  else
427  subcel->eval_dist[RoQ_ID_MOT] = INT_MAX;
428 
429  cluster_index = y*enc->width/16 + x/4;
430 
431  get_frame_mb(enc->frame_to_enc, x, y, mb4, 4);
432 
433  subcel->eval_dist[RoQ_ID_SLD] = index_mb(mb4,
434  tempData->codebooks.unpacked_cb4,
435  tempData->codebooks.numCB4,
436  &subcel->cbEntry, 4);
437 
438  subcel->eval_dist[RoQ_ID_CCC] = 0;
439 
440  for(i=0;i<4;i++) {
441  subcel->subCels[i] = tempData->closest_cb2[cluster_index*4+i];
442 
443  get_frame_mb(enc->frame_to_enc, x+2*(i&1),
444  y+(i&2), mb2, 2);
445 
446  subcel->eval_dist[RoQ_ID_CCC] +=
447  squared_diff_macroblock(tempData->codebooks.unpacked_cb2 + subcel->subCels[i]*2*2*3, mb2, 2);
448  }
449 
450  best_dist = INT_MAX;
451  for (i=0; i<4; i++)
452  if (ROQ_LAMBDA_SCALE*subcel->eval_dist[i] + enc->lambda*bitsUsed[i] <
453  best_dist) {
454  subcel->best_coding = i;
455  subcel->best_bit_use = bitsUsed[i];
456  best_dist = ROQ_LAMBDA_SCALE*subcel->eval_dist[i] +
457  enc->lambda*bitsUsed[i];
458  }
459 }
460 
461 /**
462  * Get distortion for all options available to a cel
463  */
465  RoqTempdata *tempData)
466 {
467  uint8_t mb8[8*8*3];
468  int index = cel->sourceY*enc->width/64 + cel->sourceX/8;
469  int i, j, best_dist, divide_bit_use;
470 
471  int bitsUsed[4] = {2, 10, 10, 0};
472 
473  if (enc->framesSinceKeyframe >= 1) {
474  cel->motion = enc->this_motion8[index];
475 
476  cel->eval_dist[RoQ_ID_FCC] =
477  eval_motion_dist(enc, cel->sourceX, cel->sourceY,
478  enc->this_motion8[index], 8);
479  } else
480  cel->eval_dist[RoQ_ID_FCC] = INT_MAX;
481 
482  if (enc->framesSinceKeyframe >= 2)
484  enc->current_frame->data,
485  cel->sourceX, cel->sourceY,
486  cel->sourceX, cel->sourceY,
487  enc->frame_to_enc->linesize,
488  enc->current_frame->linesize,8);
489  else
490  cel->eval_dist[RoQ_ID_MOT] = INT_MAX;
491 
492  get_frame_mb(enc->frame_to_enc, cel->sourceX, cel->sourceY, mb8, 8);
493 
494  cel->eval_dist[RoQ_ID_SLD] =
496  tempData->codebooks.numCB4, &cel->cbEntry, 8);
497 
498  gather_data_for_subcel(cel->subCels + 0, cel->sourceX+0, cel->sourceY+0, enc, tempData);
499  gather_data_for_subcel(cel->subCels + 1, cel->sourceX+4, cel->sourceY+0, enc, tempData);
500  gather_data_for_subcel(cel->subCels + 2, cel->sourceX+0, cel->sourceY+4, enc, tempData);
501  gather_data_for_subcel(cel->subCels + 3, cel->sourceX+4, cel->sourceY+4, enc, tempData);
502 
503  cel->eval_dist[RoQ_ID_CCC] = 0;
504  divide_bit_use = 0;
505  for (i=0; i<4; i++) {
506  cel->eval_dist[RoQ_ID_CCC] +=
507  cel->subCels[i].eval_dist[cel->subCels[i].best_coding];
508  divide_bit_use += cel->subCels[i].best_bit_use;
509  }
510 
511  best_dist = INT_MAX;
512  bitsUsed[3] = 2 + divide_bit_use;
513 
514  for (i=0; i<4; i++)
515  if (ROQ_LAMBDA_SCALE*cel->eval_dist[i] + enc->lambda*bitsUsed[i] <
516  best_dist) {
517  cel->best_coding = i;
518  best_dist = ROQ_LAMBDA_SCALE*cel->eval_dist[i] +
519  enc->lambda*bitsUsed[i];
520  }
521 
522  tempData->used_option[cel->best_coding]++;
523  tempData->mainChunkSize += bitsUsed[cel->best_coding];
524 
525  if (cel->best_coding == RoQ_ID_SLD)
526  tempData->codebooks.usedCB4[cel->cbEntry]++;
527 
528  if (cel->best_coding == RoQ_ID_CCC)
529  for (i=0; i<4; i++) {
530  if (cel->subCels[i].best_coding == RoQ_ID_SLD)
531  tempData->codebooks.usedCB4[cel->subCels[i].cbEntry]++;
532  else if (cel->subCels[i].best_coding == RoQ_ID_CCC)
533  for (j=0; j<4; j++)
534  tempData->codebooks.usedCB2[cel->subCels[i].subCels[j]]++;
535  }
536 }
537 
538 static void remap_codebooks(RoqContext *enc, RoqTempdata *tempData)
539 {
540  int i, j, idx=0;
541 
542  /* Make remaps for the final codebook usage */
543  for (i=0; i<MAX_CBS_4x4; i++) {
544  if (tempData->codebooks.usedCB4[i]) {
545  tempData->i2f4[i] = idx;
546  tempData->f2i4[idx] = i;
547  for (j=0; j<4; j++)
548  tempData->codebooks.usedCB2[enc->cb4x4[i].idx[j]]++;
549  idx++;
550  }
551  }
552 
553  tempData->numCB4 = idx;
554 
555  idx = 0;
556  for (i=0; i<MAX_CBS_2x2; i++) {
557  if (tempData->codebooks.usedCB2[i]) {
558  tempData->i2f2[i] = idx;
559  tempData->f2i2[idx] = i;
560  idx++;
561  }
562  }
563  tempData->numCB2 = idx;
564 
565 }
566 
567 /**
568  * Write codebook chunk
569  */
570 static void write_codebooks(RoqContext *enc, RoqTempdata *tempData)
571 {
572  int i, j;
573  uint8_t **outp= &enc->out_buf;
574 
575  if (tempData->numCB2) {
576  bytestream_put_le16(outp, RoQ_QUAD_CODEBOOK);
577  bytestream_put_le32(outp, tempData->numCB2*6 + tempData->numCB4*4);
578  bytestream_put_byte(outp, tempData->numCB4);
579  bytestream_put_byte(outp, tempData->numCB2);
580 
581  for (i=0; i<tempData->numCB2; i++) {
582  bytestream_put_buffer(outp, enc->cb2x2[tempData->f2i2[i]].y, 4);
583  bytestream_put_byte(outp, enc->cb2x2[tempData->f2i2[i]].u);
584  bytestream_put_byte(outp, enc->cb2x2[tempData->f2i2[i]].v);
585  }
586 
587  for (i=0; i<tempData->numCB4; i++)
588  for (j=0; j<4; j++)
589  bytestream_put_byte(outp, tempData->i2f2[enc->cb4x4[tempData->f2i4[i]].idx[j]]);
590 
591  }
592 }
593 
594 static inline uint8_t motion_arg(motion_vect mot)
595 {
596  uint8_t ax = 8 - ((uint8_t) mot.d[0]);
597  uint8_t ay = 8 - ((uint8_t) mot.d[1]);
598  return ((ax&15)<<4) | (ay&15);
599 }
600 
601 typedef struct
602 {
605  uint8_t argumentSpool[64];
608 } CodingSpool;
609 
610 /* NOTE: Typecodes must be spooled AFTER arguments!! */
611 static void write_typecode(CodingSpool *s, uint8_t type)
612 {
613  s->typeSpool |= (type & 3) << (14 - s->typeSpoolLength);
614  s->typeSpoolLength += 2;
615  if (s->typeSpoolLength == 16) {
616  bytestream_put_le16(s->pout, s->typeSpool);
618  s->args - s->argumentSpool);
619  s->typeSpoolLength = 0;
620  s->typeSpool = 0;
621  s->args = s->argumentSpool;
622  }
623 }
624 
625 static void reconstruct_and_encode_image(RoqContext *enc, RoqTempdata *tempData, int w, int h, int numBlocks)
626 {
627  int i, j, k;
628  int x, y;
629  int subX, subY;
630  int dist=0;
631 
632  roq_qcell *qcell;
633  CelEvaluation *eval;
634 
635  CodingSpool spool;
636 
637  spool.typeSpool=0;
638  spool.typeSpoolLength=0;
639  spool.args = spool.argumentSpool;
640  spool.pout = &enc->out_buf;
641 
642  if (tempData->used_option[RoQ_ID_CCC]%2)
643  tempData->mainChunkSize+=8; //FIXME
644 
645  /* Write the video chunk header */
646  bytestream_put_le16(&enc->out_buf, RoQ_QUAD_VQ);
647  bytestream_put_le32(&enc->out_buf, tempData->mainChunkSize/8);
648  bytestream_put_byte(&enc->out_buf, 0x0);
649  bytestream_put_byte(&enc->out_buf, 0x0);
650 
651  for (i=0; i<numBlocks; i++) {
652  eval = tempData->cel_evals + i;
653 
654  x = eval->sourceX;
655  y = eval->sourceY;
656  dist += eval->eval_dist[eval->best_coding];
657 
658  switch (eval->best_coding) {
659  case RoQ_ID_MOT:
660  write_typecode(&spool, RoQ_ID_MOT);
661  break;
662 
663  case RoQ_ID_FCC:
664  bytestream_put_byte(&spool.args, motion_arg(eval->motion));
665 
666  write_typecode(&spool, RoQ_ID_FCC);
667  ff_apply_motion_8x8(enc, x, y,
668  eval->motion.d[0], eval->motion.d[1]);
669  break;
670 
671  case RoQ_ID_SLD:
672  bytestream_put_byte(&spool.args, tempData->i2f4[eval->cbEntry]);
673  write_typecode(&spool, RoQ_ID_SLD);
674 
675  qcell = enc->cb4x4 + eval->cbEntry;
676  ff_apply_vector_4x4(enc, x , y , enc->cb2x2 + qcell->idx[0]);
677  ff_apply_vector_4x4(enc, x+4, y , enc->cb2x2 + qcell->idx[1]);
678  ff_apply_vector_4x4(enc, x , y+4, enc->cb2x2 + qcell->idx[2]);
679  ff_apply_vector_4x4(enc, x+4, y+4, enc->cb2x2 + qcell->idx[3]);
680  break;
681 
682  case RoQ_ID_CCC:
683  write_typecode(&spool, RoQ_ID_CCC);
684 
685  for (j=0; j<4; j++) {
686  subX = x + 4*(j&1);
687  subY = y + 2*(j&2);
688 
689  switch(eval->subCels[j].best_coding) {
690  case RoQ_ID_MOT:
691  break;
692 
693  case RoQ_ID_FCC:
694  bytestream_put_byte(&spool.args,
695  motion_arg(eval->subCels[j].motion));
696 
697  ff_apply_motion_4x4(enc, subX, subY,
698  eval->subCels[j].motion.d[0],
699  eval->subCels[j].motion.d[1]);
700  break;
701 
702  case RoQ_ID_SLD:
703  bytestream_put_byte(&spool.args,
704  tempData->i2f4[eval->subCels[j].cbEntry]);
705 
706  qcell = enc->cb4x4 + eval->subCels[j].cbEntry;
707 
708  ff_apply_vector_2x2(enc, subX , subY ,
709  enc->cb2x2 + qcell->idx[0]);
710  ff_apply_vector_2x2(enc, subX+2, subY ,
711  enc->cb2x2 + qcell->idx[1]);
712  ff_apply_vector_2x2(enc, subX , subY+2,
713  enc->cb2x2 + qcell->idx[2]);
714  ff_apply_vector_2x2(enc, subX+2, subY+2,
715  enc->cb2x2 + qcell->idx[3]);
716  break;
717 
718  case RoQ_ID_CCC:
719  for (k=0; k<4; k++) {
720  int cb_idx = eval->subCels[j].subCels[k];
721  bytestream_put_byte(&spool.args,
722  tempData->i2f2[cb_idx]);
723 
724  ff_apply_vector_2x2(enc, subX + 2*(k&1), subY + (k&2),
725  enc->cb2x2 + cb_idx);
726  }
727  break;
728  }
729  write_typecode(&spool, eval->subCels[j].best_coding);
730  }
731  break;
732  }
733  }
734 
735  /* Flush the remainder of the argument/type spool */
736  while (spool.typeSpoolLength)
737  write_typecode(&spool, 0x0);
738 
739 #if 0
740  uint8_t *fdata[3] = {enc->frame_to_enc->data[0],
741  enc->frame_to_enc->data[1],
742  enc->frame_to_enc->data[2]};
743  uint8_t *cdata[3] = {enc->current_frame->data[0],
744  enc->current_frame->data[1],
745  enc->current_frame->data[2]};
746  av_log(enc->avctx, AV_LOG_ERROR, "Expected distortion: %i Actual: %i\n",
747  dist,
748  block_sse(fdata, cdata, 0, 0, 0, 0,
749  enc->frame_to_enc->linesize,
750  enc->current_frame->linesize,
751  enc->width)); //WARNING: Square dimensions implied...
752 #endif
753 }
754 
755 
756 /**
757  * Create a single YUV cell from a 2x2 section of the image
758  */
759 static inline void frame_block_to_cell(uint8_t *block, uint8_t * const *data,
760  int top, int left, const int *stride)
761 {
762  int i, j, u=0, v=0;
763 
764  for (i=0; i<2; i++)
765  for (j=0; j<2; j++) {
766  int x = (top+i)*stride[0] + left + j;
767  *block++ = data[0][x];
768  x = (top+i)*stride[1] + left + j;
769  u += data[1][x];
770  v += data[2][x];
771  }
772 
773  *block++ = (u+2)/4;
774  *block++ = (v+2)/4;
775 }
776 
777 /**
778  * Create YUV clusters for the entire image
779  */
780 static void create_clusters(const AVFrame *frame, int w, int h, uint8_t *yuvClusters)
781 {
782  int i, j, k, l;
783 
784  for (i=0; i<h; i+=4)
785  for (j=0; j<w; j+=4) {
786  for (k=0; k < 2; k++)
787  for (l=0; l < 2; l++)
788  frame_block_to_cell(yuvClusters + (l + 2*k)*6, frame->data,
789  i+2*k, j+2*l, frame->linesize);
790  yuvClusters += 24;
791  }
792 }
793 
794 static void generate_codebook(RoqContext *enc, RoqTempdata *tempdata,
795  int *points, int inputCount, roq_cell *results,
796  int size, int cbsize)
797 {
798  int i, j, k;
799  int c_size = size*size/4;
800  int *buf;
801  int *codebook = av_malloc(6*c_size*cbsize*sizeof(int));
802  int *closest_cb;
803 
804  if (size == 4)
805  closest_cb = av_malloc(6*c_size*inputCount*sizeof(int));
806  else
807  closest_cb = tempdata->closest_cb2;
808 
809  ff_init_elbg(points, 6*c_size, inputCount, codebook, cbsize, 1, closest_cb, &enc->randctx);
810  ff_do_elbg(points, 6*c_size, inputCount, codebook, cbsize, 1, closest_cb, &enc->randctx);
811 
812  if (size == 4)
813  av_free(closest_cb);
814 
815  buf = codebook;
816  for (i=0; i<cbsize; i++)
817  for (k=0; k<c_size; k++) {
818  for(j=0; j<4; j++)
819  results->y[j] = *buf++;
820 
821  results->u = (*buf++ + CHROMA_BIAS/2)/CHROMA_BIAS;
822  results->v = (*buf++ + CHROMA_BIAS/2)/CHROMA_BIAS;
823  results++;
824  }
825 
826  av_free(codebook);
827 }
828 
829 static void generate_new_codebooks(RoqContext *enc, RoqTempdata *tempData)
830 {
831  int i,j;
832  RoqCodebooks *codebooks = &tempData->codebooks;
833  int max = enc->width*enc->height/16;
834  uint8_t mb2[3*4];
835  roq_cell *results4 = av_malloc(sizeof(roq_cell)*MAX_CBS_4x4*4);
836  uint8_t *yuvClusters=av_malloc(sizeof(int)*max*6*4);
837  int *points = av_malloc(max*6*4*sizeof(int));
838  int bias;
839 
840  /* Subsample YUV data */
841  create_clusters(enc->frame_to_enc, enc->width, enc->height, yuvClusters);
842 
843  /* Cast to integer and apply chroma bias */
844  for (i=0; i<max*24; i++) {
845  bias = ((i%6)<4) ? 1 : CHROMA_BIAS;
846  points[i] = bias*yuvClusters[i];
847  }
848 
849  /* Create 4x4 codebooks */
850  generate_codebook(enc, tempData, points, max, results4, 4, MAX_CBS_4x4);
851 
852  codebooks->numCB4 = MAX_CBS_4x4;
853 
854  tempData->closest_cb2 = av_malloc(max*4*sizeof(int));
855 
856  /* Create 2x2 codebooks */
857  generate_codebook(enc, tempData, points, max*4, enc->cb2x2, 2, MAX_CBS_2x2);
858 
859  codebooks->numCB2 = MAX_CBS_2x2;
860 
861  /* Unpack 2x2 codebook clusters */
862  for (i=0; i<codebooks->numCB2; i++)
863  unpack_roq_cell(enc->cb2x2 + i, codebooks->unpacked_cb2 + i*2*2*3);
864 
865  /* Index all 4x4 entries to the 2x2 entries, unpack, and enlarge */
866  for (i=0; i<codebooks->numCB4; i++) {
867  for (j=0; j<4; j++) {
868  unpack_roq_cell(&results4[4*i + j], mb2);
869  index_mb(mb2, codebooks->unpacked_cb2, codebooks->numCB2,
870  &enc->cb4x4[i].idx[j], 2);
871  }
872  unpack_roq_qcell(codebooks->unpacked_cb2, enc->cb4x4 + i,
873  codebooks->unpacked_cb4 + i*4*4*3);
874  enlarge_roq_mb4(codebooks->unpacked_cb4 + i*4*4*3,
875  codebooks->unpacked_cb4_enlarged + i*8*8*3);
876  }
877 
878  av_free(yuvClusters);
879  av_free(points);
880  av_free(results4);
881 }
882 
883 static void roq_encode_video(RoqContext *enc)
884 {
885  RoqTempdata *tempData = enc->tmpData;
886  int i;
887 
888  memset(tempData, 0, sizeof(*tempData));
889 
890  create_cel_evals(enc, tempData);
891 
892  generate_new_codebooks(enc, tempData);
893 
894  if (enc->framesSinceKeyframe >= 1) {
895  motion_search(enc, 8);
896  motion_search(enc, 4);
897  }
898 
899  retry_encode:
900  for (i=0; i<enc->width*enc->height/64; i++)
901  gather_data_for_cel(tempData->cel_evals + i, enc, tempData);
902 
903  /* Quake 3 can't handle chunks bigger than 65535 bytes */
904  if (tempData->mainChunkSize/8 > 65535) {
905  av_log(enc->avctx, AV_LOG_ERROR,
906  "Warning, generated a frame too big (%d > 65535), "
907  "try using a smaller qscale value.\n",
908  tempData->mainChunkSize/8);
909  enc->lambda *= 1.5;
910  tempData->mainChunkSize = 0;
911  memset(tempData->used_option, 0, sizeof(tempData->used_option));
912  memset(tempData->codebooks.usedCB4, 0,
913  sizeof(tempData->codebooks.usedCB4));
914  memset(tempData->codebooks.usedCB2, 0,
915  sizeof(tempData->codebooks.usedCB2));
916 
917  goto retry_encode;
918  }
919 
920  remap_codebooks(enc, tempData);
921 
922  write_codebooks(enc, tempData);
923 
924  reconstruct_and_encode_image(enc, tempData, enc->width, enc->height,
925  enc->width*enc->height/64);
926 
927  enc->avctx->coded_frame = enc->current_frame;
928 
929  /* Rotate frame history */
930  FFSWAP(AVFrame *, enc->current_frame, enc->last_frame);
933 
934  av_free(tempData->cel_evals);
935  av_free(tempData->closest_cb2);
936 
937  enc->framesSinceKeyframe++;
938 }
939 
941 {
942  RoqContext *enc = avctx->priv_data;
943 
945  av_frame_free(&enc->last_frame);
946 
947  av_free(enc->tmpData);
948  av_free(enc->this_motion4);
949  av_free(enc->last_motion4);
950  av_free(enc->this_motion8);
951  av_free(enc->last_motion8);
952 
953  return 0;
954 }
955 
957 {
958  RoqContext *enc = avctx->priv_data;
959 
960  av_lfg_init(&enc->randctx, 1);
961 
962  enc->framesSinceKeyframe = 0;
963  if ((avctx->width & 0xf) || (avctx->height & 0xf)) {
964  av_log(avctx, AV_LOG_ERROR, "Dimensions must be divisible by 16\n");
965  return -1;
966  }
967 
968  if (((avctx->width)&(avctx->width-1))||((avctx->height)&(avctx->height-1)))
969  av_log(avctx, AV_LOG_ERROR, "Warning: dimensions not power of two\n");
970 
971  enc->width = avctx->width;
972  enc->height = avctx->height;
973 
974  enc->framesSinceKeyframe = 0;
975  enc->first_frame = 1;
976 
977  enc->last_frame = av_frame_alloc();
978  enc->current_frame = av_frame_alloc();
979  if (!enc->last_frame || !enc->current_frame) {
980  roq_encode_end(avctx);
981  return AVERROR(ENOMEM);
982  }
983 
984  enc->tmpData = av_malloc(sizeof(RoqTempdata));
985 
986  enc->this_motion4 =
987  av_mallocz((enc->width*enc->height/16)*sizeof(motion_vect));
988 
989  enc->last_motion4 =
990  av_malloc ((enc->width*enc->height/16)*sizeof(motion_vect));
991 
992  enc->this_motion8 =
993  av_mallocz((enc->width*enc->height/64)*sizeof(motion_vect));
994 
995  enc->last_motion8 =
996  av_malloc ((enc->width*enc->height/64)*sizeof(motion_vect));
997 
998  return 0;
999 }
1000 
1002 {
1003  /* ROQ info chunk */
1004  bytestream_put_le16(&enc->out_buf, RoQ_INFO);
1005 
1006  /* Size: 8 bytes */
1007  bytestream_put_le32(&enc->out_buf, 8);
1008 
1009  /* Unused argument */
1010  bytestream_put_byte(&enc->out_buf, 0x00);
1011  bytestream_put_byte(&enc->out_buf, 0x00);
1012 
1013  /* Width */
1014  bytestream_put_le16(&enc->out_buf, enc->width);
1015 
1016  /* Height */
1017  bytestream_put_le16(&enc->out_buf, enc->height);
1018 
1019  /* Unused in Quake 3, mimics the output of the real encoder */
1020  bytestream_put_byte(&enc->out_buf, 0x08);
1021  bytestream_put_byte(&enc->out_buf, 0x00);
1022  bytestream_put_byte(&enc->out_buf, 0x04);
1023  bytestream_put_byte(&enc->out_buf, 0x00);
1024 }
1025 
1027  const AVFrame *frame, int *got_packet)
1028 {
1029  RoqContext *enc = avctx->priv_data;
1030  int size, ret;
1031 
1032  enc->avctx = avctx;
1033 
1034  enc->frame_to_enc = frame;
1035 
1036  if (frame->quality)
1037  enc->lambda = frame->quality - 1;
1038  else
1039  enc->lambda = 2*ROQ_LAMBDA_SCALE;
1040 
1041  /* 138 bits max per 8x8 block +
1042  * 256 codebooks*(6 bytes 2x2 + 4 bytes 4x4) + 8 bytes frame header */
1043  size = ((enc->width * enc->height / 64) * 138 + 7) / 8 + 256 * (6 + 4) + 8;
1044  if ((ret = ff_alloc_packet2(avctx, pkt, size)) < 0)
1045  return ret;
1046  enc->out_buf = pkt->data;
1047 
1048  /* Check for I frame */
1049  if (enc->framesSinceKeyframe == avctx->gop_size)
1050  enc->framesSinceKeyframe = 0;
1051 
1052  if (enc->first_frame) {
1053  /* Alloc memory for the reconstruction data (we must know the stride
1054  for that) */
1055  if ((ret = ff_get_buffer(avctx, enc->current_frame, 0)) < 0 ||
1056  (ret = ff_get_buffer(avctx, enc->last_frame, 0)) < 0)
1057  return ret;
1058 
1059  /* Before the first video frame, write a "video info" chunk */
1061 
1062  enc->first_frame = 0;
1063  }
1064 
1065  /* Encode the actual frame */
1066  roq_encode_video(enc);
1067 
1068  pkt->size = enc->out_buf - pkt->data;
1069  if (enc->framesSinceKeyframe == 1)
1070  pkt->flags |= AV_PKT_FLAG_KEY;
1071  *got_packet = 1;
1072 
1073  return 0;
1074 }
1075 
1077  .name = "roqvideo",
1078  .type = AVMEDIA_TYPE_VIDEO,
1079  .id = AV_CODEC_ID_ROQ,
1080  .priv_data_size = sizeof(RoqContext),
1081  .init = roq_encode_init,
1082  .encode2 = roq_encode_frame,
1083  .close = roq_encode_end,
1084  .supported_framerates = (const AVRational[]){ {30,1}, {0,0} },
1085  .pix_fmts = (const enum AVPixelFormat[]){ AV_PIX_FMT_YUV444P,
1086  AV_PIX_FMT_NONE },
1087  .long_name = NULL_IF_CONFIG_SMALL("id RoQ video"),
1088 };