path: root/libavcodec/roqvideoenc.c (plain)
blob: ac05123dc66cde0952b257d3f7cae14218be60ec

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,
166	enc->frame_to_enc->linesize, enc->last_frame->linesize,
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 SubcelEvaluation {
188	int eval_dist[4];
189	int best_bit_use;
190	int best_coding;
191
192	int subCels[4];
193	motion_vect motion;
194	int cbEntry;
195	} SubcelEvaluation;
196
197	typedef struct CelEvaluation {
198	int eval_dist[4];
199	int best_coding;
200
201	SubcelEvaluation subCels[4];
202
203	motion_vect motion;
204	int cbEntry;
205
206	int sourceX, sourceY;
207	} CelEvaluation;
208
209	typedef struct RoqCodebooks {
210	int numCB4;
211	int numCB2;
212	int usedCB2[MAX_CBS_2x2];
213	int usedCB4[MAX_CBS_4x4];
214	uint8_t unpacked_cb2[MAX_CBS_2x2*2*2*3];
215	uint8_t unpacked_cb4[MAX_CBS_4x4*4*4*3];
216	uint8_t unpacked_cb4_enlarged[MAX_CBS_4x4*8*8*3];
217	} RoqCodebooks;
218
219	/**
220	* Temporary vars
221	*/
222	typedef struct RoqTempData
223	{
224	CelEvaluation *cel_evals;
225
226	int f2i4[MAX_CBS_4x4];
227	int i2f4[MAX_CBS_4x4];
228	int f2i2[MAX_CBS_2x2];
229	int i2f2[MAX_CBS_2x2];
230
231	int mainChunkSize;
232
233	int numCB4;
234	int numCB2;
235
236	RoqCodebooks codebooks;
237
238	int *closest_cb2;
239	int used_option[4];
240	} RoqTempdata;
241
242	/**
243	* Initialize cel evaluators and set their source coordinates
244	*/
245	static int create_cel_evals(RoqContext *enc, RoqTempdata *tempData)
246	{
247	int n=0, x, y, i;
248
249	tempData->cel_evals = av_malloc_array(enc->width*enc->height/64, sizeof(CelEvaluation));
250	if (!tempData->cel_evals)
251	return AVERROR(ENOMEM);
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	return 0;
262	}
263
264	/**
265	* Get macroblocks from parts of the image
266	*/
267	static void get_frame_mb(const AVFrame *frame, int x, int y, uint8_t mb[], int dim)
268	{
269	int i, j, cp;
270
271	for (cp=0; cp<3; cp++) {
272	int stride = frame->linesize[cp];
273	for (i=0; i<dim; i++)
274	for (j=0; j<dim; j++)
275	*mb++ = frame->data[cp][(y+i)*stride + x + j];
276	}
277	}
278
279	/**
280	* Find the codebook with the lowest distortion from an image
281	*/
282	static int index_mb(uint8_t cluster[], uint8_t cb[], int numCB,
283	int *outIndex, int dim)
284	{
285	int i, lDiff = INT_MAX, pick=0;
286
287	/* Diff against the others */
288	for (i=0; i<numCB; i++) {
289	int diff = squared_diff_macroblock(cluster, cb + i*dim*dim*3, dim);
290	if (diff < lDiff) {
291	lDiff = diff;
292	pick = i;
293	}
294	}
295
296	*outIndex = pick;
297	return lDiff;
298	}
299
300	#define EVAL_MOTION(MOTION) \
301	do { \
302	diff = eval_motion_dist(enc, j, i, MOTION, blocksize); \
303	\
304	if (diff < lowestdiff) { \
305	lowestdiff = diff; \
306	bestpick = MOTION; \
307	} \
308	} while(0)
309
310	static void motion_search(RoqContext *enc, int blocksize)
311	{
312	static const motion_vect offsets[8] = {
313	{{ 0,-1}},
314	{{ 0, 1}},
315	{{-1, 0}},
316	{{ 1, 0}},
317	{{-1, 1}},
318	{{ 1,-1}},
319	{{-1,-1}},
320	{{ 1, 1}},
321	};
322
323	int diff, lowestdiff, oldbest;
324	int off[3];
325	motion_vect bestpick = {{0,0}};
326	int i, j, k, offset;
327
328	motion_vect *last_motion;
329	motion_vect *this_motion;
330	motion_vect vect, vect2;
331
332	int max=(enc->width/blocksize)*enc->height/blocksize;
333
334	if (blocksize == 4) {
335	last_motion = enc->last_motion4;
336	this_motion = enc->this_motion4;
337	} else {
338	last_motion = enc->last_motion8;
339	this_motion = enc->this_motion8;
340	}
341
342	for (i=0; i<enc->height; i+=blocksize)
343	for (j=0; j<enc->width; j+=blocksize) {
344	lowestdiff = eval_motion_dist(enc, j, i, (motion_vect) {{0,0}},
345	blocksize);
346	bestpick.d[0] = 0;
347	bestpick.d[1] = 0;
348
349	if (blocksize == 4)
350	EVAL_MOTION(enc->this_motion8[(i/8)*(enc->width/8) + j/8]);
351
352	offset = (i/blocksize)*enc->width/blocksize + j/blocksize;
353	if (offset < max && offset >= 0)
354	EVAL_MOTION(last_motion[offset]);
355
356	offset++;
357	if (offset < max && offset >= 0)
358	EVAL_MOTION(last_motion[offset]);
359
360	offset = (i/blocksize + 1)*enc->width/blocksize + j/blocksize;
361	if (offset < max && offset >= 0)
362	EVAL_MOTION(last_motion[offset]);
363
364	off[0]= (i/blocksize)*enc->width/blocksize + j/blocksize - 1;
365	off[1]= off[0] - enc->width/blocksize + 1;
366	off[2]= off[1] + 1;
367
368	if (i) {
369
370	for(k=0; k<2; k++)
371	vect.d[k]= mid_pred(this_motion[off[0]].d[k],
372	this_motion[off[1]].d[k],
373	this_motion[off[2]].d[k]);
374
375	EVAL_MOTION(vect);
376	for(k=0; k<3; k++)
377	EVAL_MOTION(this_motion[off[k]]);
378	} else if(j)
379	EVAL_MOTION(this_motion[off[0]]);
380
381	vect = bestpick;
382
383	oldbest = -1;
384	while (oldbest != lowestdiff) {
385	oldbest = lowestdiff;
386	for (k=0; k<8; k++) {
387	vect2 = vect;
388	vect2.d[0] += offsets[k].d[0];
389	vect2.d[1] += offsets[k].d[1];
390	EVAL_MOTION(vect2);
391	}
392	vect = bestpick;
393	}
394	offset = (i/blocksize)*enc->width/blocksize + j/blocksize;
395	this_motion[offset] = bestpick;
396	}
397	}
398
399	/**
400	* Get distortion for all options available to a subcel
401	*/
402	static void gather_data_for_subcel(SubcelEvaluation *subcel, int x,
403	int y, RoqContext *enc, RoqTempdata *tempData)
404	{
405	uint8_t mb4[4*4*3];
406	uint8_t mb2[2*2*3];
407	int cluster_index;
408	int i, best_dist;
409
410	static const int bitsUsed[4] = {2, 10, 10, 34};
411
412	if (enc->framesSinceKeyframe >= 1) {
413	subcel->motion = enc->this_motion4[y*enc->width/16 + x/4];
414
415	subcel->eval_dist[RoQ_ID_FCC] =
416	eval_motion_dist(enc, x, y,
417	enc->this_motion4[y*enc->width/16 + x/4], 4);
418	} else
419	subcel->eval_dist[RoQ_ID_FCC] = INT_MAX;
420
421	if (enc->framesSinceKeyframe >= 2)
422	subcel->eval_dist[RoQ_ID_MOT] = block_sse(enc->frame_to_enc->data,
423	enc->current_frame->data, x,
424	y, x, y,
425	enc->frame_to_enc->linesize,
426	enc->current_frame->linesize,
427	4);
428	else
429	subcel->eval_dist[RoQ_ID_MOT] = INT_MAX;
430
431	cluster_index = y*enc->width/16 + x/4;
432
433	get_frame_mb(enc->frame_to_enc, x, y, mb4, 4);
434
435	subcel->eval_dist[RoQ_ID_SLD] = index_mb(mb4,
436	tempData->codebooks.unpacked_cb4,
437	tempData->codebooks.numCB4,
438	&subcel->cbEntry, 4);
439
440	subcel->eval_dist[RoQ_ID_CCC] = 0;
441
442	for(i=0;i<4;i++) {
443	subcel->subCels[i] = tempData->closest_cb2[cluster_index*4+i];
444
445	get_frame_mb(enc->frame_to_enc, x+2*(i&1),
446	y+(i&2), mb2, 2);
447
448	subcel->eval_dist[RoQ_ID_CCC] +=
449	squared_diff_macroblock(tempData->codebooks.unpacked_cb2 + subcel->subCels[i]*2*2*3, mb2, 2);
450	}
451
452	best_dist = INT_MAX;
453	for (i=0; i<4; i++)
454	if (ROQ_LAMBDA_SCALE*subcel->eval_dist[i] + enc->lambda*bitsUsed[i] <
455	best_dist) {
456	subcel->best_coding = i;
457	subcel->best_bit_use = bitsUsed[i];
458	best_dist = ROQ_LAMBDA_SCALE*subcel->eval_dist[i] +
459	enc->lambda*bitsUsed[i];
460	}
461	}
462
463	/**
464	* Get distortion for all options available to a cel
465	*/
466	static void gather_data_for_cel(CelEvaluation *cel, RoqContext *enc,
467	RoqTempdata *tempData)
468	{
469	uint8_t mb8[8*8*3];
470	int index = cel->sourceY*enc->width/64 + cel->sourceX/8;
471	int i, j, best_dist, divide_bit_use;
472
473	int bitsUsed[4] = {2, 10, 10, 0};
474
475	if (enc->framesSinceKeyframe >= 1) {
476	cel->motion = enc->this_motion8[index];
477
478	cel->eval_dist[RoQ_ID_FCC] =
479	eval_motion_dist(enc, cel->sourceX, cel->sourceY,
480	enc->this_motion8[index], 8);
481	} else
482	cel->eval_dist[RoQ_ID_FCC] = INT_MAX;
483
484	if (enc->framesSinceKeyframe >= 2)
485	cel->eval_dist[RoQ_ID_MOT] = block_sse(enc->frame_to_enc->data,
486	enc->current_frame->data,
487	cel->sourceX, cel->sourceY,
488	cel->sourceX, cel->sourceY,
489	enc->frame_to_enc->linesize,
490	enc->current_frame->linesize,8);
491	else
492	cel->eval_dist[RoQ_ID_MOT] = INT_MAX;
493
494	get_frame_mb(enc->frame_to_enc, cel->sourceX, cel->sourceY, mb8, 8);
495
496	cel->eval_dist[RoQ_ID_SLD] =
497	index_mb(mb8, tempData->codebooks.unpacked_cb4_enlarged,
498	tempData->codebooks.numCB4, &cel->cbEntry, 8);
499
500	gather_data_for_subcel(cel->subCels + 0, cel->sourceX+0, cel->sourceY+0, enc, tempData);
501	gather_data_for_subcel(cel->subCels + 1, cel->sourceX+4, cel->sourceY+0, enc, tempData);
502	gather_data_for_subcel(cel->subCels + 2, cel->sourceX+0, cel->sourceY+4, enc, tempData);
503	gather_data_for_subcel(cel->subCels + 3, cel->sourceX+4, cel->sourceY+4, enc, tempData);
504
505	cel->eval_dist[RoQ_ID_CCC] = 0;
506	divide_bit_use = 0;
507	for (i=0; i<4; i++) {
508	cel->eval_dist[RoQ_ID_CCC] +=
509	cel->subCels[i].eval_dist[cel->subCels[i].best_coding];
510	divide_bit_use += cel->subCels[i].best_bit_use;
511	}
512
513	best_dist = INT_MAX;
514	bitsUsed[3] = 2 + divide_bit_use;
515
516	for (i=0; i<4; i++)
517	if (ROQ_LAMBDA_SCALE*cel->eval_dist[i] + enc->lambda*bitsUsed[i] <
518	best_dist) {
519	cel->best_coding = i;
520	best_dist = ROQ_LAMBDA_SCALE*cel->eval_dist[i] +
521	enc->lambda*bitsUsed[i];
522	}
523
524	tempData->used_option[cel->best_coding]++;
525	tempData->mainChunkSize += bitsUsed[cel->best_coding];
526
527	if (cel->best_coding == RoQ_ID_SLD)
528	tempData->codebooks.usedCB4[cel->cbEntry]++;
529
530	if (cel->best_coding == RoQ_ID_CCC)
531	for (i=0; i<4; i++) {
532	if (cel->subCels[i].best_coding == RoQ_ID_SLD)
533	tempData->codebooks.usedCB4[cel->subCels[i].cbEntry]++;
534	else if (cel->subCels[i].best_coding == RoQ_ID_CCC)
535	for (j=0; j<4; j++)
536	tempData->codebooks.usedCB2[cel->subCels[i].subCels[j]]++;
537	}
538	}
539
540	static void remap_codebooks(RoqContext *enc, RoqTempdata *tempData)
541	{
542	int i, j, idx=0;
543
544	/* Make remaps for the final codebook usage */
545	for (i=0; i<(enc->quake3_compat ? MAX_CBS_4x4-1 : MAX_CBS_4x4); i++) {
546	if (tempData->codebooks.usedCB4[i]) {
547	tempData->i2f4[i] = idx;
548	tempData->f2i4[idx] = i;
549	for (j=0; j<4; j++)
550	tempData->codebooks.usedCB2[enc->cb4x4[i].idx[j]]++;
551	idx++;
552	}
553	}
554
555	tempData->numCB4 = idx;
556
557	idx = 0;
558	for (i=0; i<MAX_CBS_2x2; i++) {
559	if (tempData->codebooks.usedCB2[i]) {
560	tempData->i2f2[i] = idx;
561	tempData->f2i2[idx] = i;
562	idx++;
563	}
564	}
565	tempData->numCB2 = idx;
566
567	}
568
569	/**
570	* Write codebook chunk
571	*/
572	static void write_codebooks(RoqContext *enc, RoqTempdata *tempData)
573	{
574	int i, j;
575	uint8_t **outp= &enc->out_buf;
576
577	if (tempData->numCB2) {
578	bytestream_put_le16(outp, RoQ_QUAD_CODEBOOK);
579	bytestream_put_le32(outp, tempData->numCB2*6 + tempData->numCB4*4);
580	bytestream_put_byte(outp, tempData->numCB4);
581	bytestream_put_byte(outp, tempData->numCB2);
582
583	for (i=0; i<tempData->numCB2; i++) {
584	bytestream_put_buffer(outp, enc->cb2x2[tempData->f2i2[i]].y, 4);
585	bytestream_put_byte(outp, enc->cb2x2[tempData->f2i2[i]].u);
586	bytestream_put_byte(outp, enc->cb2x2[tempData->f2i2[i]].v);
587	}
588
589	for (i=0; i<tempData->numCB4; i++)
590	for (j=0; j<4; j++)
591	bytestream_put_byte(outp, tempData->i2f2[enc->cb4x4[tempData->f2i4[i]].idx[j]]);
592
593	}
594	}
595
596	static inline uint8_t motion_arg(motion_vect mot)
597	{
598	uint8_t ax = 8 - ((uint8_t) mot.d[0]);
599	uint8_t ay = 8 - ((uint8_t) mot.d[1]);
600	return ((ax&15)<<4) | (ay&15);
601	}
602
603	typedef struct CodingSpool {
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!! */
612	static 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
626	static 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
741
742	/**
743	* Create a single YUV cell from a 2x2 section of the image
744	*/
745	static inline void frame_block_to_cell(uint8_t *block, uint8_t * const *data,
746	int top, int left, const int *stride)
747	{
748	int i, j, u=0, v=0;
749
750	for (i=0; i<2; i++)
751	for (j=0; j<2; j++) {
752	int x = (top+i)*stride[0] + left + j;
753	*block++ = data[0][x];
754	x = (top+i)*stride[1] + left + j;
755	u += data[1][x];
756	v += data[2][x];
757	}
758
759	*block++ = (u+2)/4;
760	*block++ = (v+2)/4;
761	}
762
763	/**
764	* Create YUV clusters for the entire image
765	*/
766	static void create_clusters(const AVFrame *frame, int w, int h, uint8_t *yuvClusters)
767	{
768	int i, j, k, l;
769
770	for (i=0; i<h; i+=4)
771	for (j=0; j<w; j+=4) {
772	for (k=0; k < 2; k++)
773	for (l=0; l < 2; l++)
774	frame_block_to_cell(yuvClusters + (l + 2*k)*6, frame->data,
775	i+2*k, j+2*l, frame->linesize);
776	yuvClusters += 24;
777	}
778	}
779
780	static int generate_codebook(RoqContext *enc, RoqTempdata *tempdata,
781	int *points, int inputCount, roq_cell *results,
782	int size, int cbsize)
783	{
784	int i, j, k, ret = 0;
785	int c_size = size*size/4;
786	int *buf;
787	int *codebook = av_malloc_array(6*c_size, cbsize*sizeof(int));
788	int *closest_cb;
789
790	if (!codebook)
791	return AVERROR(ENOMEM);
792
793	if (size == 4) {
794	closest_cb = av_malloc_array(6*c_size, inputCount*sizeof(int));
795	if (!closest_cb) {
796	ret = AVERROR(ENOMEM);
797	goto out;
798	}
799	} else
800	closest_cb = tempdata->closest_cb2;
801
802	ret = avpriv_init_elbg(points, 6 * c_size, inputCount, codebook,
803	cbsize, 1, closest_cb, &enc->randctx);
804	if (ret < 0)
805	goto out;
806	ret = avpriv_do_elbg(points, 6 * c_size, inputCount, codebook,
807	cbsize, 1, closest_cb, &enc->randctx);
808	if (ret < 0)
809	goto out;
810
811	buf = codebook;
812	for (i=0; i<cbsize; i++)
813	for (k=0; k<c_size; k++) {
814	for(j=0; j<4; j++)
815	results->y[j] = *buf++;
816
817	results->u = (*buf++ + CHROMA_BIAS/2)/CHROMA_BIAS;
818	results->v = (*buf++ + CHROMA_BIAS/2)/CHROMA_BIAS;
819	results++;
820	}
821	out:
822	if (size == 4)
823	av_free(closest_cb);
824	av_free(codebook);
825	return ret;
826	}
827
828	static int generate_new_codebooks(RoqContext *enc, RoqTempdata *tempData)
829	{
830	int i, j, ret = 0;
831	RoqCodebooks *codebooks = &tempData->codebooks;
832	int max = enc->width*enc->height/16;
833	uint8_t mb2[3*4];
834	roq_cell *results4 = av_malloc(sizeof(roq_cell)*MAX_CBS_4x4*4);
835	uint8_t *yuvClusters=av_malloc_array(max, sizeof(int)*6*4);
836	int *points = av_malloc_array(max, 6*4*sizeof(int));
837	int bias;
838
839	if (!results4 || !yuvClusters || !points) {
840	ret = AVERROR(ENOMEM);
841	goto out;
842	}
843
844	/* Subsample YUV data */
845	create_clusters(enc->frame_to_enc, enc->width, enc->height, yuvClusters);
846
847	/* Cast to integer and apply chroma bias */
848	for (i=0; i<max*24; i++) {
849	bias = ((i%6)<4) ? 1 : CHROMA_BIAS;
850	points[i] = bias*yuvClusters[i];
851	}
852
853	/* Create 4x4 codebooks */
854	if ((ret = generate_codebook(enc, tempData, points, max,
855	results4, 4, (enc->quake3_compat ? MAX_CBS_4x4-1 : MAX_CBS_4x4))) < 0)
856	goto out;
857
858	codebooks->numCB4 = (enc->quake3_compat ? MAX_CBS_4x4-1 : MAX_CBS_4x4);
859
860	tempData->closest_cb2 = av_malloc_array(max, 4*sizeof(int));
861	if (!tempData->closest_cb2) {
862	ret = AVERROR(ENOMEM);
863	goto out;
864	}
865
866	/* Create 2x2 codebooks */
867	if ((ret = generate_codebook(enc, tempData, points, max * 4,
868	enc->cb2x2, 2, MAX_CBS_2x2)) < 0)
869	goto out;
870
871	codebooks->numCB2 = MAX_CBS_2x2;
872
873	/* Unpack 2x2 codebook clusters */
874	for (i=0; i<codebooks->numCB2; i++)
875	unpack_roq_cell(enc->cb2x2 + i, codebooks->unpacked_cb2 + i*2*2*3);
876
877	/* Index all 4x4 entries to the 2x2 entries, unpack, and enlarge */
878	for (i=0; i<codebooks->numCB4; i++) {
879	for (j=0; j<4; j++) {
880	unpack_roq_cell(&results4[4*i + j], mb2);
881	index_mb(mb2, codebooks->unpacked_cb2, codebooks->numCB2,
882	&enc->cb4x4[i].idx[j], 2);
883	}
884	unpack_roq_qcell(codebooks->unpacked_cb2, enc->cb4x4 + i,
885	codebooks->unpacked_cb4 + i*4*4*3);
886	enlarge_roq_mb4(codebooks->unpacked_cb4 + i*4*4*3,
887	codebooks->unpacked_cb4_enlarged + i*8*8*3);
888	}
889	out:
890	av_free(yuvClusters);
891	av_free(points);
892	av_free(results4);
893	return ret;
894	}
895
896	static int roq_encode_video(RoqContext *enc)
897	{
898	RoqTempdata *tempData = enc->tmpData;
899	int i, ret;
900
901	memset(tempData, 0, sizeof(*tempData));
902
903	ret = create_cel_evals(enc, tempData);
904	if (ret < 0)
905	return ret;
906
907	ret = generate_new_codebooks(enc, tempData);
908	if (ret < 0)
909	return ret;
910
911	if (enc->framesSinceKeyframe >= 1) {
912	motion_search(enc, 8);
913	motion_search(enc, 4);
914	}
915
916	retry_encode:
917	for (i=0; i<enc->width*enc->height/64; i++)
918	gather_data_for_cel(tempData->cel_evals + i, enc, tempData);
919
920	/* Quake 3 can't handle chunks bigger than 65535 bytes */
921	if (tempData->mainChunkSize/8 > 65535 && enc->quake3_compat) {
922	if (enc->lambda > 100000) {
923	av_log(enc->avctx, AV_LOG_ERROR, "Cannot encode video in Quake compatible form\n");
924	return AVERROR(EINVAL);
925	}
926	av_log(enc->avctx, AV_LOG_ERROR,
927	"Warning, generated a frame too big for Quake (%d > 65535), "
928	"now switching to a bigger qscale value.\n",
929	tempData->mainChunkSize/8);
930	enc->lambda *= 1.5;
931	tempData->mainChunkSize = 0;
932	memset(tempData->used_option, 0, sizeof(tempData->used_option));
933	memset(tempData->codebooks.usedCB4, 0,
934	sizeof(tempData->codebooks.usedCB4));
935	memset(tempData->codebooks.usedCB2, 0,
936	sizeof(tempData->codebooks.usedCB2));
937
938	goto retry_encode;
939	}
940
941	remap_codebooks(enc, tempData);
942
943	write_codebooks(enc, tempData);
944
945	reconstruct_and_encode_image(enc, tempData, enc->width, enc->height,
946	enc->width*enc->height/64);
947
948	/* Rotate frame history */
949	FFSWAP(AVFrame *, enc->current_frame, enc->last_frame);
950	FFSWAP(motion_vect *, enc->last_motion4, enc->this_motion4);
951	FFSWAP(motion_vect *, enc->last_motion8, enc->this_motion8);
952
953	av_freep(&tempData->cel_evals);
954	av_freep(&tempData->closest_cb2);
955
956	enc->framesSinceKeyframe++;
957
958	return 0;
959	}
960
961	static av_cold int roq_encode_end(AVCodecContext *avctx)
962	{
963	RoqContext *enc = avctx->priv_data;
964
965	av_frame_free(&enc->current_frame);
966	av_frame_free(&enc->last_frame);
967
968	av_freep(&enc->tmpData);
969	av_freep(&enc->this_motion4);
970	av_freep(&enc->last_motion4);
971	av_freep(&enc->this_motion8);
972	av_freep(&enc->last_motion8);
973
974	return 0;
975	}
976
977	static av_cold int roq_encode_init(AVCodecContext *avctx)
978	{
979	RoqContext *enc = avctx->priv_data;
980
981	av_lfg_init(&enc->randctx, 1);
982
983	enc->avctx = avctx;
984
985	enc->framesSinceKeyframe = 0;
986	if ((avctx->width & 0xf) || (avctx->height & 0xf)) {
987	av_log(avctx, AV_LOG_ERROR, "Dimensions must be divisible by 16\n");
988	return AVERROR(EINVAL);
989	}
990
991	if (avctx->width > 65535 || avctx->height > 65535) {
992	av_log(avctx, AV_LOG_ERROR, "Dimensions are max %d\n", enc->quake3_compat ? 32768 : 65535);
993	return AVERROR(EINVAL);
994	}
995
996	if (((avctx->width)&(avctx->width-1))||((avctx->height)&(avctx->height-1)))
997	av_log(avctx, AV_LOG_ERROR, "Warning: dimensions not power of two, this is not supported by quake\n");
998
999	enc->width = avctx->width;
1000	enc->height = avctx->height;
1001
1002	enc->framesSinceKeyframe = 0;
1003	enc->first_frame = 1;
1004
1005	enc->last_frame = av_frame_alloc();
1006	enc->current_frame = av_frame_alloc();
1007	if (!enc->last_frame || !enc->current_frame) {
1008	roq_encode_end(avctx);
1009	return AVERROR(ENOMEM);
1010	}
1011
1012	enc->tmpData = av_malloc(sizeof(RoqTempdata));
1013
1014	enc->this_motion4 =
1015	av_mallocz_array((enc->width*enc->height/16), sizeof(motion_vect));
1016
1017	enc->last_motion4 =
1018	av_malloc_array ((enc->width*enc->height/16), sizeof(motion_vect));
1019
1020	enc->this_motion8 =
1021	av_mallocz_array((enc->width*enc->height/64), sizeof(motion_vect));
1022
1023	enc->last_motion8 =
1024	av_malloc_array ((enc->width*enc->height/64), sizeof(motion_vect));
1025
1026	if (!enc->tmpData || !enc->this_motion4 || !enc->last_motion4 ||
1027	!enc->this_motion8 || !enc->last_motion8) {
1028	roq_encode_end(avctx);
1029	return AVERROR(ENOMEM);
1030	}
1031
1032	return 0;
1033	}
1034
1035	static void roq_write_video_info_chunk(RoqContext *enc)
1036	{
1037	/* ROQ info chunk */
1038	bytestream_put_le16(&enc->out_buf, RoQ_INFO);
1039
1040	/* Size: 8 bytes */
1041	bytestream_put_le32(&enc->out_buf, 8);
1042
1043	/* Unused argument */
1044	bytestream_put_byte(&enc->out_buf, 0x00);
1045	bytestream_put_byte(&enc->out_buf, 0x00);
1046
1047	/* Width */
1048	bytestream_put_le16(&enc->out_buf, enc->width);
1049
1050	/* Height */
1051	bytestream_put_le16(&enc->out_buf, enc->height);
1052
1053	/* Unused in Quake 3, mimics the output of the real encoder */
1054	bytestream_put_byte(&enc->out_buf, 0x08);
1055	bytestream_put_byte(&enc->out_buf, 0x00);
1056	bytestream_put_byte(&enc->out_buf, 0x04);
1057	bytestream_put_byte(&enc->out_buf, 0x00);
1058	}
1059
1060	static int roq_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1061	const AVFrame *frame, int *got_packet)
1062	{
1063	RoqContext *enc = avctx->priv_data;
1064	int size, ret;
1065
1066	enc->avctx = avctx;
1067
1068	enc->frame_to_enc = frame;
1069
1070	if (frame->quality)
1071	enc->lambda = frame->quality - 1;
1072	else
1073	enc->lambda = 2*ROQ_LAMBDA_SCALE;
1074
1075	/* 138 bits max per 8x8 block +
1076	* 256 codebooks*(6 bytes 2x2 + 4 bytes 4x4) + 8 bytes frame header */
1077	size = ((enc->width * enc->height / 64) * 138 + 7) / 8 + 256 * (6 + 4) + 8;
1078	if ((ret = ff_alloc_packet2(avctx, pkt, size, 0)) < 0)
1079	return ret;
1080	enc->out_buf = pkt->data;
1081
1082	/* Check for I-frame */
1083	if (enc->framesSinceKeyframe == avctx->gop_size)
1084	enc->framesSinceKeyframe = 0;
1085
1086	if (enc->first_frame) {
1087	/* Alloc memory for the reconstruction data (we must know the stride
1088	for that) */
1089	if ((ret = ff_get_buffer(avctx, enc->current_frame, 0)) < 0 ||
1090	(ret = ff_get_buffer(avctx, enc->last_frame, 0)) < 0)
1091	return ret;
1092
1093	/* Before the first video frame, write a "video info" chunk */
1094	roq_write_video_info_chunk(enc);
1095
1096	enc->first_frame = 0;
1097	}
1098
1099	/* Encode the actual frame */
1100	ret = roq_encode_video(enc);
1101	if (ret < 0)
1102	return ret;
1103
1104	pkt->size = enc->out_buf - pkt->data;
1105	if (enc->framesSinceKeyframe == 1)
1106	pkt->flags |= AV_PKT_FLAG_KEY;
1107	*got_packet = 1;
1108
1109	return 0;
1110	}
1111
1112	#define OFFSET(x) offsetof(RoqContext, x)
1113	#define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1114	static const AVOption options[] = {
1115	{ "quake3_compat", "Whether to respect known limitations in Quake 3 decoder", OFFSET(quake3_compat), AV_OPT_TYPE_BOOL, { .i64 = 1 }, 0, 1, VE },
1116	{ NULL },
1117	};
1118
1119	static const AVClass roq_class = {
1120	.class_name = "RoQ",
1121	.item_name = av_default_item_name,
1122	.option = options,
1123	.version = LIBAVUTIL_VERSION_INT,
1124	};
1125
1126	AVCodec ff_roq_encoder = {
1127	.name = "roqvideo",
1128	.long_name = NULL_IF_CONFIG_SMALL("id RoQ video"),
1129	.type = AVMEDIA_TYPE_VIDEO,
1130	.id = AV_CODEC_ID_ROQ,
1131	.priv_data_size = sizeof(RoqContext),
1132	.init = roq_encode_init,
1133	.encode2 = roq_encode_frame,
1134	.close = roq_encode_end,
1135	.pix_fmts = (const enum AVPixelFormat[]){ AV_PIX_FMT_YUVJ444P,
1136	AV_PIX_FMT_NONE },
1137	.priv_class = &roq_class,
1138	};
1139