Imported Debian version 2.5.3~trusty1
[deb_ffmpeg.git] / ffmpeg / libavcodec / roqvideoenc.c
CommitLineData
2ba45a60
DM
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 */
81static 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
88static 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
103static 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
113static inline int square(int x)
114{
115 return x*x;
116}
117
118static 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).
130static 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
146static 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,
166 enc->frame_to_enc->linesize, enc->last_frame->linesize,
167 size);
168}
169
170/**
171 * @return distortion between two macroblocks
172 */
173static 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
187typedef struct
188{
189 int eval_dist[4];
190 int best_bit_use;
191 int best_coding;
192
193 int subCels[4];
194 motion_vect motion;
195 int cbEntry;
196} SubcelEvaluation;
197
198typedef struct
199{
200 int eval_dist[4];
201 int best_coding;
202
203 SubcelEvaluation subCels[4];
204
205 motion_vect motion;
206 int cbEntry;
207
208 int sourceX, sourceY;
209} CelEvaluation;
210
211typedef 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 */
225typedef struct RoqTempData
226{
227 CelEvaluation *cel_evals;
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
234 int mainChunkSize;
235
236 int numCB4;
237 int numCB2;
238
239 RoqCodebooks codebooks;
240
241 int *closest_cb2;
242 int used_option[4];
243} RoqTempdata;
244
245/**
246 * Initialize cel evaluators and set their source coordinates
247 */
248static void 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
254 /* Map to the ROQ quadtree order */
255 for (y=0; y<enc->height; y+=16)
256 for (x=0; x<enc->width; x+=16)
257 for(i=0; i<4; i++) {
258 tempData->cel_evals[n ].sourceX = x + (i&1)*8;
259 tempData->cel_evals[n++].sourceY = y + (i&2)*4;
260 }
261}
262
263/**
264 * Get macroblocks from parts of the image
265 */
266static void get_frame_mb(const AVFrame *frame, int x, int y, uint8_t mb[], int dim)
267{
268 int i, j, cp;
269
270 for (cp=0; cp<3; cp++) {
271 int stride = frame->linesize[cp];
272 for (i=0; i<dim; i++)
273 for (j=0; j<dim; j++)
274 *mb++ = frame->data[cp][(y+i)*stride + x + j];
275 }
276}
277
278/**
279 * Find the codebook with the lowest distortion from an image
280 */
281static int index_mb(uint8_t cluster[], uint8_t cb[], int numCB,
282 int *outIndex, int dim)
283{
284 int i, lDiff = INT_MAX, pick=0;
285
286 /* Diff against the others */
287 for (i=0; i<numCB; i++) {
288 int diff = squared_diff_macroblock(cluster, cb + i*dim*dim*3, dim);
289 if (diff < lDiff) {
290 lDiff = diff;
291 pick = i;
292 }
293 }
294
295 *outIndex = pick;
296 return lDiff;
297}
298
299#define EVAL_MOTION(MOTION) \
300 do { \
301 diff = eval_motion_dist(enc, j, i, MOTION, blocksize); \
302 \
303 if (diff < lowestdiff) { \
304 lowestdiff = diff; \
305 bestpick = MOTION; \
306 } \
307 } while(0)
308
309static void motion_search(RoqContext *enc, int blocksize)
310{
311 static const motion_vect offsets[8] = {
312 {{ 0,-1}},
313 {{ 0, 1}},
314 {{-1, 0}},
315 {{ 1, 0}},
316 {{-1, 1}},
317 {{ 1,-1}},
318 {{-1,-1}},
319 {{ 1, 1}},
320 };
321
322 int diff, lowestdiff, oldbest;
323 int off[3];
324 motion_vect bestpick = {{0,0}};
325 int i, j, k, offset;
326
327 motion_vect *last_motion;
328 motion_vect *this_motion;
329 motion_vect vect, vect2;
330
331 int max=(enc->width/blocksize)*enc->height/blocksize;
332
333 if (blocksize == 4) {
334 last_motion = enc->last_motion4;
335 this_motion = enc->this_motion4;
336 } else {
337 last_motion = enc->last_motion8;
338 this_motion = enc->this_motion8;
339 }
340
341 for (i=0; i<enc->height; i+=blocksize)
342 for (j=0; j<enc->width; j+=blocksize) {
343 lowestdiff = eval_motion_dist(enc, j, i, (motion_vect) {{0,0}},
344 blocksize);
345 bestpick.d[0] = 0;
346 bestpick.d[1] = 0;
347
348 if (blocksize == 4)
349 EVAL_MOTION(enc->this_motion8[(i/8)*(enc->width/8) + j/8]);
350
351 offset = (i/blocksize)*enc->width/blocksize + j/blocksize;
352 if (offset < max && offset >= 0)
353 EVAL_MOTION(last_motion[offset]);
354
355 offset++;
356 if (offset < max && offset >= 0)
357 EVAL_MOTION(last_motion[offset]);
358
359 offset = (i/blocksize + 1)*enc->width/blocksize + j/blocksize;
360 if (offset < max && offset >= 0)
361 EVAL_MOTION(last_motion[offset]);
362
363 off[0]= (i/blocksize)*enc->width/blocksize + j/blocksize - 1;
364 off[1]= off[0] - enc->width/blocksize + 1;
365 off[2]= off[1] + 1;
366
367 if (i) {
368
369 for(k=0; k<2; k++)
370 vect.d[k]= mid_pred(this_motion[off[0]].d[k],
371 this_motion[off[1]].d[k],
372 this_motion[off[2]].d[k]);
373
374 EVAL_MOTION(vect);
375 for(k=0; k<3; k++)
376 EVAL_MOTION(this_motion[off[k]]);
377 } else if(j)
378 EVAL_MOTION(this_motion[off[0]]);
379
380 vect = bestpick;
381
382 oldbest = -1;
383 while (oldbest != lowestdiff) {
384 oldbest = lowestdiff;
385 for (k=0; k<8; k++) {
386 vect2 = vect;
387 vect2.d[0] += offsets[k].d[0];
388 vect2.d[1] += offsets[k].d[1];
389 EVAL_MOTION(vect2);
390 }
391 vect = bestpick;
392 }
393 offset = (i/blocksize)*enc->width/blocksize + j/blocksize;
394 this_motion[offset] = bestpick;
395 }
396}
397
398/**
399 * Get distortion for all options available to a subcel
400 */
401static void gather_data_for_subcel(SubcelEvaluation *subcel, int x,
402 int y, RoqContext *enc, RoqTempdata *tempData)
403{
404 uint8_t mb4[4*4*3];
405 uint8_t mb2[2*2*3];
406 int cluster_index;
407 int i, best_dist;
408
409 static const int bitsUsed[4] = {2, 10, 10, 34};
410
411 if (enc->framesSinceKeyframe >= 1) {
412 subcel->motion = enc->this_motion4[y*enc->width/16 + x/4];
413
414 subcel->eval_dist[RoQ_ID_FCC] =
415 eval_motion_dist(enc, x, y,
416 enc->this_motion4[y*enc->width/16 + x/4], 4);
417 } else
418 subcel->eval_dist[RoQ_ID_FCC] = INT_MAX;
419
420 if (enc->framesSinceKeyframe >= 2)
421 subcel->eval_dist[RoQ_ID_MOT] = block_sse(enc->frame_to_enc->data,
422 enc->current_frame->data, x,
423 y, x, y,
424 enc->frame_to_enc->linesize,
425 enc->current_frame->linesize,
426 4);
427 else
428 subcel->eval_dist[RoQ_ID_MOT] = INT_MAX;
429
430 cluster_index = y*enc->width/16 + x/4;
431
432 get_frame_mb(enc->frame_to_enc, x, y, mb4, 4);
433
434 subcel->eval_dist[RoQ_ID_SLD] = index_mb(mb4,
435 tempData->codebooks.unpacked_cb4,
436 tempData->codebooks.numCB4,
437 &subcel->cbEntry, 4);
438
439 subcel->eval_dist[RoQ_ID_CCC] = 0;
440
441 for(i=0;i<4;i++) {
442 subcel->subCels[i] = tempData->closest_cb2[cluster_index*4+i];
443
444 get_frame_mb(enc->frame_to_enc, x+2*(i&1),
445 y+(i&2), mb2, 2);
446
447 subcel->eval_dist[RoQ_ID_CCC] +=
448 squared_diff_macroblock(tempData->codebooks.unpacked_cb2 + subcel->subCels[i]*2*2*3, mb2, 2);
449 }
450
451 best_dist = INT_MAX;
452 for (i=0; i<4; i++)
453 if (ROQ_LAMBDA_SCALE*subcel->eval_dist[i] + enc->lambda*bitsUsed[i] <
454 best_dist) {
455 subcel->best_coding = i;
456 subcel->best_bit_use = bitsUsed[i];
457 best_dist = ROQ_LAMBDA_SCALE*subcel->eval_dist[i] +
458 enc->lambda*bitsUsed[i];
459 }
460}
461
462/**
463 * Get distortion for all options available to a cel
464 */
465static void gather_data_for_cel(CelEvaluation *cel, RoqContext *enc,
466 RoqTempdata *tempData)
467{
468 uint8_t mb8[8*8*3];
469 int index = cel->sourceY*enc->width/64 + cel->sourceX/8;
470 int i, j, best_dist, divide_bit_use;
471
472 int bitsUsed[4] = {2, 10, 10, 0};
473
474 if (enc->framesSinceKeyframe >= 1) {
475 cel->motion = enc->this_motion8[index];
476
477 cel->eval_dist[RoQ_ID_FCC] =
478 eval_motion_dist(enc, cel->sourceX, cel->sourceY,
479 enc->this_motion8[index], 8);
480 } else
481 cel->eval_dist[RoQ_ID_FCC] = INT_MAX;
482
483 if (enc->framesSinceKeyframe >= 2)
484 cel->eval_dist[RoQ_ID_MOT] = block_sse(enc->frame_to_enc->data,
485 enc->current_frame->data,
486 cel->sourceX, cel->sourceY,
487 cel->sourceX, cel->sourceY,
488 enc->frame_to_enc->linesize,
489 enc->current_frame->linesize,8);
490 else
491 cel->eval_dist[RoQ_ID_MOT] = INT_MAX;
492
493 get_frame_mb(enc->frame_to_enc, cel->sourceX, cel->sourceY, mb8, 8);
494
495 cel->eval_dist[RoQ_ID_SLD] =
496 index_mb(mb8, tempData->codebooks.unpacked_cb4_enlarged,
497 tempData->codebooks.numCB4, &cel->cbEntry, 8);
498
499 gather_data_for_subcel(cel->subCels + 0, cel->sourceX+0, cel->sourceY+0, enc, tempData);
500 gather_data_for_subcel(cel->subCels + 1, cel->sourceX+4, cel->sourceY+0, enc, tempData);
501 gather_data_for_subcel(cel->subCels + 2, cel->sourceX+0, cel->sourceY+4, enc, tempData);
502 gather_data_for_subcel(cel->subCels + 3, cel->sourceX+4, cel->sourceY+4, enc, tempData);
503
504 cel->eval_dist[RoQ_ID_CCC] = 0;
505 divide_bit_use = 0;
506 for (i=0; i<4; i++) {
507 cel->eval_dist[RoQ_ID_CCC] +=
508 cel->subCels[i].eval_dist[cel->subCels[i].best_coding];
509 divide_bit_use += cel->subCels[i].best_bit_use;
510 }
511
512 best_dist = INT_MAX;
513 bitsUsed[3] = 2 + divide_bit_use;
514
515 for (i=0; i<4; i++)
516 if (ROQ_LAMBDA_SCALE*cel->eval_dist[i] + enc->lambda*bitsUsed[i] <
517 best_dist) {
518 cel->best_coding = i;
519 best_dist = ROQ_LAMBDA_SCALE*cel->eval_dist[i] +
520 enc->lambda*bitsUsed[i];
521 }
522
523 tempData->used_option[cel->best_coding]++;
524 tempData->mainChunkSize += bitsUsed[cel->best_coding];
525
526 if (cel->best_coding == RoQ_ID_SLD)
527 tempData->codebooks.usedCB4[cel->cbEntry]++;
528
529 if (cel->best_coding == RoQ_ID_CCC)
530 for (i=0; i<4; i++) {
531 if (cel->subCels[i].best_coding == RoQ_ID_SLD)
532 tempData->codebooks.usedCB4[cel->subCels[i].cbEntry]++;
533 else if (cel->subCels[i].best_coding == RoQ_ID_CCC)
534 for (j=0; j<4; j++)
535 tempData->codebooks.usedCB2[cel->subCels[i].subCels[j]]++;
536 }
537}
538
539static void remap_codebooks(RoqContext *enc, RoqTempdata *tempData)
540{
541 int i, j, idx=0;
542
543 /* Make remaps for the final codebook usage */
544 for (i=0; i<(enc->quake3_compat ? MAX_CBS_4x4-1 : MAX_CBS_4x4); i++) {
545 if (tempData->codebooks.usedCB4[i]) {
546 tempData->i2f4[i] = idx;
547 tempData->f2i4[idx] = i;
548 for (j=0; j<4; j++)
549 tempData->codebooks.usedCB2[enc->cb4x4[i].idx[j]]++;
550 idx++;
551 }
552 }
553
554 tempData->numCB4 = idx;
555
556 idx = 0;
557 for (i=0; i<MAX_CBS_2x2; i++) {
558 if (tempData->codebooks.usedCB2[i]) {
559 tempData->i2f2[i] = idx;
560 tempData->f2i2[idx] = i;
561 idx++;
562 }
563 }
564 tempData->numCB2 = idx;
565
566}
567
568/**
569 * Write codebook chunk
570 */
571static void write_codebooks(RoqContext *enc, RoqTempdata *tempData)
572{
573 int i, j;
574 uint8_t **outp= &enc->out_buf;
575
576 if (tempData->numCB2) {
577 bytestream_put_le16(outp, RoQ_QUAD_CODEBOOK);
578 bytestream_put_le32(outp, tempData->numCB2*6 + tempData->numCB4*4);
579 bytestream_put_byte(outp, tempData->numCB4);
580 bytestream_put_byte(outp, tempData->numCB2);
581
582 for (i=0; i<tempData->numCB2; i++) {
583 bytestream_put_buffer(outp, enc->cb2x2[tempData->f2i2[i]].y, 4);
584 bytestream_put_byte(outp, enc->cb2x2[tempData->f2i2[i]].u);
585 bytestream_put_byte(outp, enc->cb2x2[tempData->f2i2[i]].v);
586 }
587
588 for (i=0; i<tempData->numCB4; i++)
589 for (j=0; j<4; j++)
590 bytestream_put_byte(outp, tempData->i2f2[enc->cb4x4[tempData->f2i4[i]].idx[j]]);
591
592 }
593}
594
595static inline uint8_t motion_arg(motion_vect mot)
596{
597 uint8_t ax = 8 - ((uint8_t) mot.d[0]);
598 uint8_t ay = 8 - ((uint8_t) mot.d[1]);
599 return ((ax&15)<<4) | (ay&15);
600}
601
602typedef struct
603{
604 int typeSpool;
605 int typeSpoolLength;
606 uint8_t argumentSpool[64];
607 uint8_t *args;
608 uint8_t **pout;
609} CodingSpool;
610
611/* NOTE: Typecodes must be spooled AFTER arguments!! */
612static void write_typecode(CodingSpool *s, uint8_t type)
613{
614 s->typeSpool |= (type & 3) << (14 - s->typeSpoolLength);
615 s->typeSpoolLength += 2;
616 if (s->typeSpoolLength == 16) {
617 bytestream_put_le16(s->pout, s->typeSpool);
618 bytestream_put_buffer(s->pout, s->argumentSpool,
619 s->args - s->argumentSpool);
620 s->typeSpoolLength = 0;
621 s->typeSpool = 0;
622 s->args = s->argumentSpool;
623 }
624}
625
626static void reconstruct_and_encode_image(RoqContext *enc, RoqTempdata *tempData, int w, int h, int numBlocks)
627{
628 int i, j, k;
629 int x, y;
630 int subX, subY;
631 int dist=0;
632
633 roq_qcell *qcell;
634 CelEvaluation *eval;
635
636 CodingSpool spool;
637
638 spool.typeSpool=0;
639 spool.typeSpoolLength=0;
640 spool.args = spool.argumentSpool;
641 spool.pout = &enc->out_buf;
642
643 if (tempData->used_option[RoQ_ID_CCC]%2)
644 tempData->mainChunkSize+=8; //FIXME
645
646 /* Write the video chunk header */
647 bytestream_put_le16(&enc->out_buf, RoQ_QUAD_VQ);
648 bytestream_put_le32(&enc->out_buf, tempData->mainChunkSize/8);
649 bytestream_put_byte(&enc->out_buf, 0x0);
650 bytestream_put_byte(&enc->out_buf, 0x0);
651
652 for (i=0; i<numBlocks; i++) {
653 eval = tempData->cel_evals + i;
654
655 x = eval->sourceX;
656 y = eval->sourceY;
657 dist += eval->eval_dist[eval->best_coding];
658
659 switch (eval->best_coding) {
660 case RoQ_ID_MOT:
661 write_typecode(&spool, RoQ_ID_MOT);
662 break;
663
664 case RoQ_ID_FCC:
665 bytestream_put_byte(&spool.args, motion_arg(eval->motion));
666
667 write_typecode(&spool, RoQ_ID_FCC);
668 ff_apply_motion_8x8(enc, x, y,
669 eval->motion.d[0], eval->motion.d[1]);
670 break;
671
672 case RoQ_ID_SLD:
673 bytestream_put_byte(&spool.args, tempData->i2f4[eval->cbEntry]);
674 write_typecode(&spool, RoQ_ID_SLD);
675
676 qcell = enc->cb4x4 + eval->cbEntry;
677 ff_apply_vector_4x4(enc, x , y , enc->cb2x2 + qcell->idx[0]);
678 ff_apply_vector_4x4(enc, x+4, y , enc->cb2x2 + qcell->idx[1]);
679 ff_apply_vector_4x4(enc, x , y+4, enc->cb2x2 + qcell->idx[2]);
680 ff_apply_vector_4x4(enc, x+4, y+4, enc->cb2x2 + qcell->idx[3]);
681 break;
682
683 case RoQ_ID_CCC:
684 write_typecode(&spool, RoQ_ID_CCC);
685
686 for (j=0; j<4; j++) {
687 subX = x + 4*(j&1);
688 subY = y + 2*(j&2);
689
690 switch(eval->subCels[j].best_coding) {
691 case RoQ_ID_MOT:
692 break;
693
694 case RoQ_ID_FCC:
695 bytestream_put_byte(&spool.args,
696 motion_arg(eval->subCels[j].motion));
697
698 ff_apply_motion_4x4(enc, subX, subY,
699 eval->subCels[j].motion.d[0],
700 eval->subCels[j].motion.d[1]);
701 break;
702
703 case RoQ_ID_SLD:
704 bytestream_put_byte(&spool.args,
705 tempData->i2f4[eval->subCels[j].cbEntry]);
706
707 qcell = enc->cb4x4 + eval->subCels[j].cbEntry;
708
709 ff_apply_vector_2x2(enc, subX , subY ,
710 enc->cb2x2 + qcell->idx[0]);
711 ff_apply_vector_2x2(enc, subX+2, subY ,
712 enc->cb2x2 + qcell->idx[1]);
713 ff_apply_vector_2x2(enc, subX , subY+2,
714 enc->cb2x2 + qcell->idx[2]);
715 ff_apply_vector_2x2(enc, subX+2, subY+2,
716 enc->cb2x2 + qcell->idx[3]);
717 break;
718
719 case RoQ_ID_CCC:
720 for (k=0; k<4; k++) {
721 int cb_idx = eval->subCels[j].subCels[k];
722 bytestream_put_byte(&spool.args,
723 tempData->i2f2[cb_idx]);
724
725 ff_apply_vector_2x2(enc, subX + 2*(k&1), subY + (k&2),
726 enc->cb2x2 + cb_idx);
727 }
728 break;
729 }
730 write_typecode(&spool, eval->subCels[j].best_coding);
731 }
732 break;
733 }
734 }
735
736 /* Flush the remainder of the argument/type spool */
737 while (spool.typeSpoolLength)
738 write_typecode(&spool, 0x0);
739
740#if 0
741 uint8_t *fdata[3] = {enc->frame_to_enc->data[0],
742 enc->frame_to_enc->data[1],
743 enc->frame_to_enc->data[2]};
744 uint8_t *cdata[3] = {enc->current_frame->data[0],
745 enc->current_frame->data[1],
746 enc->current_frame->data[2]};
747 av_log(enc->avctx, AV_LOG_ERROR, "Expected distortion: %i Actual: %i\n",
748 dist,
749 block_sse(fdata, cdata, 0, 0, 0, 0,
750 enc->frame_to_enc->linesize,
751 enc->current_frame->linesize,
752 enc->width)); //WARNING: Square dimensions implied...
753#endif
754}
755
756
757/**
758 * Create a single YUV cell from a 2x2 section of the image
759 */
760static inline void frame_block_to_cell(uint8_t *block, uint8_t * const *data,
761 int top, int left, const int *stride)
762{
763 int i, j, u=0, v=0;
764
765 for (i=0; i<2; i++)
766 for (j=0; j<2; j++) {
767 int x = (top+i)*stride[0] + left + j;
768 *block++ = data[0][x];
769 x = (top+i)*stride[1] + left + j;
770 u += data[1][x];
771 v += data[2][x];
772 }
773
774 *block++ = (u+2)/4;
775 *block++ = (v+2)/4;
776}
777
778/**
779 * Create YUV clusters for the entire image
780 */
781static void create_clusters(const AVFrame *frame, int w, int h, uint8_t *yuvClusters)
782{
783 int i, j, k, l;
784
785 for (i=0; i<h; i+=4)
786 for (j=0; j<w; j+=4) {
787 for (k=0; k < 2; k++)
788 for (l=0; l < 2; l++)
789 frame_block_to_cell(yuvClusters + (l + 2*k)*6, frame->data,
790 i+2*k, j+2*l, frame->linesize);
791 yuvClusters += 24;
792 }
793}
794
795static void generate_codebook(RoqContext *enc, RoqTempdata *tempdata,
796 int *points, int inputCount, roq_cell *results,
797 int size, int cbsize)
798{
799 int i, j, k;
800 int c_size = size*size/4;
801 int *buf;
802 int *codebook = av_malloc_array(6*c_size, cbsize*sizeof(int));
803 int *closest_cb;
804
805 if (size == 4)
806 closest_cb = av_malloc_array(6*c_size, inputCount*sizeof(int));
807 else
808 closest_cb = tempdata->closest_cb2;
809
810 avpriv_init_elbg(points, 6*c_size, inputCount, codebook, cbsize, 1, closest_cb, &enc->randctx);
811 avpriv_do_elbg(points, 6*c_size, inputCount, codebook, cbsize, 1, closest_cb, &enc->randctx);
812
813 if (size == 4)
814 av_free(closest_cb);
815
816 buf = codebook;
817 for (i=0; i<cbsize; i++)
818 for (k=0; k<c_size; k++) {
819 for(j=0; j<4; j++)
820 results->y[j] = *buf++;
821
822 results->u = (*buf++ + CHROMA_BIAS/2)/CHROMA_BIAS;
823 results->v = (*buf++ + CHROMA_BIAS/2)/CHROMA_BIAS;
824 results++;
825 }
826
827 av_free(codebook);
828}
829
830static void generate_new_codebooks(RoqContext *enc, RoqTempdata *tempData)
831{
832 int i,j;
833 RoqCodebooks *codebooks = &tempData->codebooks;
834 int max = enc->width*enc->height/16;
835 uint8_t mb2[3*4];
836 roq_cell *results4 = av_malloc(sizeof(roq_cell)*MAX_CBS_4x4*4);
837 uint8_t *yuvClusters=av_malloc_array(max, sizeof(int)*6*4);
838 int *points = av_malloc_array(max, 6*4*sizeof(int));
839 int bias;
840
841 /* Subsample YUV data */
842 create_clusters(enc->frame_to_enc, enc->width, enc->height, yuvClusters);
843
844 /* Cast to integer and apply chroma bias */
845 for (i=0; i<max*24; i++) {
846 bias = ((i%6)<4) ? 1 : CHROMA_BIAS;
847 points[i] = bias*yuvClusters[i];
848 }
849
850 /* Create 4x4 codebooks */
851 generate_codebook(enc, tempData, points, max, results4, 4, (enc->quake3_compat ? MAX_CBS_4x4-1 : MAX_CBS_4x4));
852
853 codebooks->numCB4 = (enc->quake3_compat ? MAX_CBS_4x4-1 : MAX_CBS_4x4);
854
855 tempData->closest_cb2 = av_malloc_array(max, 4*sizeof(int));
856
857 /* Create 2x2 codebooks */
858 generate_codebook(enc, tempData, points, max*4, enc->cb2x2, 2, MAX_CBS_2x2);
859
860 codebooks->numCB2 = MAX_CBS_2x2;
861
862 /* Unpack 2x2 codebook clusters */
863 for (i=0; i<codebooks->numCB2; i++)
864 unpack_roq_cell(enc->cb2x2 + i, codebooks->unpacked_cb2 + i*2*2*3);
865
866 /* Index all 4x4 entries to the 2x2 entries, unpack, and enlarge */
867 for (i=0; i<codebooks->numCB4; i++) {
868 for (j=0; j<4; j++) {
869 unpack_roq_cell(&results4[4*i + j], mb2);
870 index_mb(mb2, codebooks->unpacked_cb2, codebooks->numCB2,
871 &enc->cb4x4[i].idx[j], 2);
872 }
873 unpack_roq_qcell(codebooks->unpacked_cb2, enc->cb4x4 + i,
874 codebooks->unpacked_cb4 + i*4*4*3);
875 enlarge_roq_mb4(codebooks->unpacked_cb4 + i*4*4*3,
876 codebooks->unpacked_cb4_enlarged + i*8*8*3);
877 }
878
879 av_free(yuvClusters);
880 av_free(points);
881 av_free(results4);
882}
883
884static int roq_encode_video(RoqContext *enc)
885{
886 RoqTempdata *tempData = enc->tmpData;
887 int i;
888
889 memset(tempData, 0, sizeof(*tempData));
890
891 create_cel_evals(enc, tempData);
892
893 generate_new_codebooks(enc, tempData);
894
895 if (enc->framesSinceKeyframe >= 1) {
896 motion_search(enc, 8);
897 motion_search(enc, 4);
898 }
899
900 retry_encode:
901 for (i=0; i<enc->width*enc->height/64; i++)
902 gather_data_for_cel(tempData->cel_evals + i, enc, tempData);
903
904 /* Quake 3 can't handle chunks bigger than 65535 bytes */
905 if (tempData->mainChunkSize/8 > 65535 && enc->quake3_compat) {
906 if (enc->lambda > 100000) {
907 av_log(enc->avctx, AV_LOG_ERROR, "Cannot encode video in Quake compatible form\n");
908 return AVERROR(EINVAL);
909 }
910 av_log(enc->avctx, AV_LOG_ERROR,
911 "Warning, generated a frame too big for Quake (%d > 65535), "
912 "now switching to a bigger qscale value.\n",
913 tempData->mainChunkSize/8);
914 enc->lambda *= 1.5;
915 tempData->mainChunkSize = 0;
916 memset(tempData->used_option, 0, sizeof(tempData->used_option));
917 memset(tempData->codebooks.usedCB4, 0,
918 sizeof(tempData->codebooks.usedCB4));
919 memset(tempData->codebooks.usedCB2, 0,
920 sizeof(tempData->codebooks.usedCB2));
921
922 goto retry_encode;
923 }
924
925 remap_codebooks(enc, tempData);
926
927 write_codebooks(enc, tempData);
928
929 reconstruct_and_encode_image(enc, tempData, enc->width, enc->height,
930 enc->width*enc->height/64);
931
932 enc->avctx->coded_frame = enc->current_frame;
933
934 /* Rotate frame history */
935 FFSWAP(AVFrame *, enc->current_frame, enc->last_frame);
936 FFSWAP(motion_vect *, enc->last_motion4, enc->this_motion4);
937 FFSWAP(motion_vect *, enc->last_motion8, enc->this_motion8);
938
f6fa7814
DM
939 av_freep(&tempData->cel_evals);
940 av_freep(&tempData->closest_cb2);
2ba45a60
DM
941
942 enc->framesSinceKeyframe++;
943
944 return 0;
945}
946
947static av_cold int roq_encode_end(AVCodecContext *avctx)
948{
949 RoqContext *enc = avctx->priv_data;
950
951 av_frame_free(&enc->current_frame);
952 av_frame_free(&enc->last_frame);
953
f6fa7814
DM
954 av_freep(&enc->tmpData);
955 av_freep(&enc->this_motion4);
956 av_freep(&enc->last_motion4);
957 av_freep(&enc->this_motion8);
958 av_freep(&enc->last_motion8);
2ba45a60
DM
959
960 return 0;
961}
962
963static av_cold int roq_encode_init(AVCodecContext *avctx)
964{
965 RoqContext *enc = avctx->priv_data;
966
967 av_lfg_init(&enc->randctx, 1);
968
969 enc->framesSinceKeyframe = 0;
970 if ((avctx->width & 0xf) || (avctx->height & 0xf)) {
971 av_log(avctx, AV_LOG_ERROR, "Dimensions must be divisible by 16\n");
972 return AVERROR(EINVAL);
973 }
974
975 if (avctx->width > 65535 || avctx->height > 65535) {
976 av_log(avctx, AV_LOG_ERROR, "Dimensions are max %d\n", enc->quake3_compat ? 32768 : 65535);
977 return AVERROR(EINVAL);
978 }
979
980 if (((avctx->width)&(avctx->width-1))||((avctx->height)&(avctx->height-1)))
981 av_log(avctx, AV_LOG_ERROR, "Warning: dimensions not power of two, this is not supported by quake\n");
982
983 enc->width = avctx->width;
984 enc->height = avctx->height;
985
986 enc->framesSinceKeyframe = 0;
987 enc->first_frame = 1;
988
989 enc->last_frame = av_frame_alloc();
990 enc->current_frame = av_frame_alloc();
991 if (!enc->last_frame || !enc->current_frame) {
992 roq_encode_end(avctx);
993 return AVERROR(ENOMEM);
994 }
995
996 enc->tmpData = av_malloc(sizeof(RoqTempdata));
997
998 enc->this_motion4 =
999 av_mallocz_array((enc->width*enc->height/16), sizeof(motion_vect));
1000
1001 enc->last_motion4 =
1002 av_malloc_array ((enc->width*enc->height/16), sizeof(motion_vect));
1003
1004 enc->this_motion8 =
1005 av_mallocz_array((enc->width*enc->height/64), sizeof(motion_vect));
1006
1007 enc->last_motion8 =
1008 av_malloc_array ((enc->width*enc->height/64), sizeof(motion_vect));
1009
1010 return 0;
1011}
1012
1013static void roq_write_video_info_chunk(RoqContext *enc)
1014{
1015 /* ROQ info chunk */
1016 bytestream_put_le16(&enc->out_buf, RoQ_INFO);
1017
1018 /* Size: 8 bytes */
1019 bytestream_put_le32(&enc->out_buf, 8);
1020
1021 /* Unused argument */
1022 bytestream_put_byte(&enc->out_buf, 0x00);
1023 bytestream_put_byte(&enc->out_buf, 0x00);
1024
1025 /* Width */
1026 bytestream_put_le16(&enc->out_buf, enc->width);
1027
1028 /* Height */
1029 bytestream_put_le16(&enc->out_buf, enc->height);
1030
1031 /* Unused in Quake 3, mimics the output of the real encoder */
1032 bytestream_put_byte(&enc->out_buf, 0x08);
1033 bytestream_put_byte(&enc->out_buf, 0x00);
1034 bytestream_put_byte(&enc->out_buf, 0x04);
1035 bytestream_put_byte(&enc->out_buf, 0x00);
1036}
1037
1038static int roq_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1039 const AVFrame *frame, int *got_packet)
1040{
1041 RoqContext *enc = avctx->priv_data;
1042 int size, ret;
1043
1044 enc->avctx = avctx;
1045
1046 enc->frame_to_enc = frame;
1047
1048 if (frame->quality)
1049 enc->lambda = frame->quality - 1;
1050 else
1051 enc->lambda = 2*ROQ_LAMBDA_SCALE;
1052
1053 /* 138 bits max per 8x8 block +
1054 * 256 codebooks*(6 bytes 2x2 + 4 bytes 4x4) + 8 bytes frame header */
1055 size = ((enc->width * enc->height / 64) * 138 + 7) / 8 + 256 * (6 + 4) + 8;
1056 if ((ret = ff_alloc_packet2(avctx, pkt, size)) < 0)
1057 return ret;
1058 enc->out_buf = pkt->data;
1059
1060 /* Check for I frame */
1061 if (enc->framesSinceKeyframe == avctx->gop_size)
1062 enc->framesSinceKeyframe = 0;
1063
1064 if (enc->first_frame) {
1065 /* Alloc memory for the reconstruction data (we must know the stride
1066 for that) */
1067 if ((ret = ff_get_buffer(avctx, enc->current_frame, 0)) < 0 ||
1068 (ret = ff_get_buffer(avctx, enc->last_frame, 0)) < 0)
1069 return ret;
1070
1071 /* Before the first video frame, write a "video info" chunk */
1072 roq_write_video_info_chunk(enc);
1073
1074 enc->first_frame = 0;
1075 }
1076
1077 /* Encode the actual frame */
1078 if ((ret = roq_encode_video(enc)) < 0)
1079 return ret;
1080
1081 pkt->size = enc->out_buf - pkt->data;
1082 if (enc->framesSinceKeyframe == 1)
1083 pkt->flags |= AV_PKT_FLAG_KEY;
1084 *got_packet = 1;
1085
1086 return 0;
1087}
1088
1089#define OFFSET(x) offsetof(RoqContext, x)
1090#define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1091static const AVOption options[] = {
1092 { "quake3_compat", "Whether to respect known limitations in Quake 3 decoder", OFFSET(quake3_compat), AV_OPT_TYPE_INT, { .i64 = 1 }, 0, 1, VE },
1093 { NULL },
1094};
1095
1096static const AVClass roq_class = {
1097 .class_name = "RoQ",
1098 .item_name = av_default_item_name,
1099 .option = options,
1100 .version = LIBAVUTIL_VERSION_INT,
1101};
1102
1103AVCodec ff_roq_encoder = {
1104 .name = "roqvideo",
1105 .long_name = NULL_IF_CONFIG_SMALL("id RoQ video"),
1106 .type = AVMEDIA_TYPE_VIDEO,
1107 .id = AV_CODEC_ID_ROQ,
1108 .priv_data_size = sizeof(RoqContext),
1109 .init = roq_encode_init,
1110 .encode2 = roq_encode_frame,
1111 .close = roq_encode_end,
1112 .pix_fmts = (const enum AVPixelFormat[]){ AV_PIX_FMT_YUVJ444P,
1113 AV_PIX_FMT_NONE },
1114 .priv_class = &roq_class,
1115};