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/lfg.h"
61#include "libavutil/opt.h"
62#include "roqvideo.h"
63#include "bytestream.h"
64#include "codec_internal.h"
65#include "elbg.h"
66#include "encode.h"
67#include "internal.h"
68#include "mathops.h"
69
70#define CHROMA_BIAS 1
71
72/**
73 * Maximum number of generated 4x4 codebooks. Can't be 256 to workaround a
74 * Quake 3 bug.
75 */
76#define MAX_CBS_4x4 256
77
78#define MAX_CBS_2x2 256 ///< Maximum number of 2x2 codebooks.
79
80/* The cast is useful when multiplying it by INT_MAX */
81#define ROQ_LAMBDA_SCALE ((uint64_t) FF_LAMBDA_SCALE)
82
83typedef struct RoqCodebooks {
84    int numCB4;
85    int numCB2;
86    int usedCB2[MAX_CBS_2x2];
87    int usedCB4[MAX_CBS_4x4];
88    uint8_t unpacked_cb2[MAX_CBS_2x2*2*2*3];
89    uint8_t unpacked_cb4[MAX_CBS_4x4*4*4*3];
90    uint8_t unpacked_cb4_enlarged[MAX_CBS_4x4*8*8*3];
91} RoqCodebooks;
92
93/**
94 * Temporary vars
95 */
96typedef struct RoqTempData
97{
98    int f2i4[MAX_CBS_4x4];
99    int i2f4[MAX_CBS_4x4];
100    int f2i2[MAX_CBS_2x2];
101    int i2f2[MAX_CBS_2x2];
102
103    int mainChunkSize;
104
105    int numCB4;
106    int numCB2;
107
108    RoqCodebooks codebooks;
109
110    int used_option[4];
111} RoqTempData;
112
113typedef struct SubcelEvaluation {
114    int eval_dist[4];
115    int best_bit_use;
116    int best_coding;
117
118    int subCels[4];
119    motion_vect motion;
120    int cbEntry;
121} SubcelEvaluation;
122
123typedef struct CelEvaluation {
124    int eval_dist[4];
125    int best_coding;
126
127    SubcelEvaluation subCels[4];
128
129    motion_vect motion;
130    int cbEntry;
131
132    int sourceX, sourceY;
133} CelEvaluation;
134
135typedef struct RoqEncContext {
136    RoqContext common;
137    struct ELBGContext *elbg;
138    AVLFG randctx;
139    uint64_t lambda;
140
141    motion_vect *this_motion4;
142    motion_vect *last_motion4;
143
144    motion_vect *this_motion8;
145    motion_vect *last_motion8;
146
147    unsigned int framesSinceKeyframe;
148
149    const AVFrame *frame_to_enc;
150    uint8_t *out_buf;
151    RoqTempData tmp_data;
152    roq_cell results4[4 * MAX_CBS_4x4];
153    int tmp_codebook_buf[FFMAX(24 * MAX_CBS_4x4, 6 * MAX_CBS_2x2)];
154
155    CelEvaluation *cel_evals;
156    int *closest_cb;
157    int *points;  // Allocated together with closest_cb
158
159    int first_frame;
160    int quake3_compat; // Quake 3 compatibility option
161} RoqEncContext;
162
163/* Macroblock support functions */
164static void unpack_roq_cell(roq_cell *cell, uint8_t u[4*3])
165{
166    memcpy(u  , cell->y, 4);
167    memset(u+4, cell->u, 4);
168    memset(u+8, cell->v, 4);
169}
170
171static void unpack_roq_qcell(uint8_t cb2[], roq_qcell *qcell, uint8_t u[4*4*3])
172{
173    int i,cp;
174    static const int offsets[4] = {0, 2, 8, 10};
175
176    for (cp=0; cp<3; cp++)
177        for (i=0; i<4; i++) {
178            u[4*4*cp + offsets[i]  ] = cb2[qcell->idx[i]*2*2*3 + 4*cp  ];
179            u[4*4*cp + offsets[i]+1] = cb2[qcell->idx[i]*2*2*3 + 4*cp+1];
180            u[4*4*cp + offsets[i]+4] = cb2[qcell->idx[i]*2*2*3 + 4*cp+2];
181            u[4*4*cp + offsets[i]+5] = cb2[qcell->idx[i]*2*2*3 + 4*cp+3];
182        }
183}
184
185
186static void enlarge_roq_mb4(uint8_t base[3*16], uint8_t u[3*64])
187{
188    int x,y,cp;
189
190    for(cp=0; cp<3; cp++)
191        for(y=0; y<8; y++)
192            for(x=0; x<8; x++)
193                *u++ = base[(y/2)*4 + (x/2) + 16*cp];
194}
195
196static inline int square(int x)
197{
198    return x*x;
199}
200
201static inline int eval_sse(const uint8_t *a, const uint8_t *b, int count)
202{
203    int diff=0;
204
205    while(count--)
206        diff += square(*b++ - *a++);
207
208    return diff;
209}
210
211// FIXME Could use DSPContext.sse, but it is not so speed critical (used
212// just for motion estimation).
213static int block_sse(uint8_t * const *buf1, uint8_t * const *buf2, int x1, int y1,
214                     int x2, int y2, const int *stride1, const int *stride2, int size)
215{
216    int i, k;
217    int sse=0;
218
219    for (k=0; k<3; k++) {
220        int bias = (k ? CHROMA_BIAS : 4);
221        for (i=0; i<size; i++)
222            sse += bias*eval_sse(buf1[k] + (y1+i)*stride1[k] + x1,
223                                 buf2[k] + (y2+i)*stride2[k] + x2, size);
224    }
225
226    return sse;
227}
228
229static int eval_motion_dist(RoqEncContext *enc, int x, int y, motion_vect vect,
230                             int size)
231{
232    RoqContext *const roq = &enc->common;
233    int mx=vect.d[0];
234    int my=vect.d[1];
235
236    if (mx < -7 || mx > 7)
237        return INT_MAX;
238
239    if (my < -7 || my > 7)
240        return INT_MAX;
241
242    mx += x;
243    my += y;
244
245    if ((unsigned) mx > roq->width-size || (unsigned) my > roq->height-size)
246        return INT_MAX;
247
248    return block_sse(enc->frame_to_enc->data, roq->last_frame->data, x, y,
249                     mx, my,
250                     enc->frame_to_enc->linesize, roq->last_frame->linesize,
251                     size);
252}
253
254/**
255 * @return distortion between two macroblocks
256 */
257static inline int squared_diff_macroblock(uint8_t a[], uint8_t b[], int size)
258{
259    int cp, sdiff=0;
260
261    for(cp=0;cp<3;cp++) {
262        int bias = (cp ? CHROMA_BIAS : 4);
263        sdiff += bias*eval_sse(a, b, size*size);
264        a += size*size;
265        b += size*size;
266    }
267
268    return sdiff;
269}
270
271/**
272 * Initialize cel evaluators and set their source coordinates
273 */
274static int create_cel_evals(RoqEncContext *enc)
275{
276    RoqContext *const roq = &enc->common;
277
278    enc->cel_evals = av_malloc_array(roq->width * roq->height / 64, sizeof(CelEvaluation));
279    if (!enc->cel_evals)
280        return AVERROR(ENOMEM);
281
282    /* Map to the ROQ quadtree order */
283    for (int y = 0, n = 0; y < roq->height; y += 16)
284        for (int x = 0; x < roq->width; x += 16)
285            for(int i = 0; i < 4; i++) {
286                enc->cel_evals[n  ].sourceX = x + (i&1)*8;
287                enc->cel_evals[n++].sourceY = y + (i&2)*4;
288            }
289
290    return 0;
291}
292
293/**
294 * Get macroblocks from parts of the image
295 */
296static void get_frame_mb(const AVFrame *frame, int x, int y, uint8_t mb[], int dim)
297{
298    int i, j, cp;
299
300    for (cp=0; cp<3; cp++) {
301        int stride = frame->linesize[cp];
302        for (i=0; i<dim; i++)
303            for (j=0; j<dim; j++)
304                *mb++ = frame->data[cp][(y+i)*stride + x + j];
305    }
306}
307
308/**
309 * Find the codebook with the lowest distortion from an image
310 */
311static int index_mb(uint8_t cluster[], uint8_t cb[], int numCB,
312                    int *outIndex, int dim)
313{
314    int i, lDiff = INT_MAX, pick=0;
315
316    /* Diff against the others */
317    for (i=0; i<numCB; i++) {
318        int diff = squared_diff_macroblock(cluster, cb + i*dim*dim*3, dim);
319        if (diff < lDiff) {
320            lDiff = diff;
321            pick = i;
322        }
323    }
324
325    *outIndex = pick;
326    return lDiff;
327}
328
329#define EVAL_MOTION(MOTION) \
330    do { \
331        diff = eval_motion_dist(enc, j, i, MOTION, blocksize); \
332            \
333        if (diff < lowestdiff) { \
334            lowestdiff = diff; \
335            bestpick = MOTION; \
336        } \
337    } while(0)
338
339static void motion_search(RoqEncContext *enc, int blocksize)
340{
341    static const motion_vect offsets[8] = {
342        {{ 0,-1}},
343        {{ 0, 1}},
344        {{-1, 0}},
345        {{ 1, 0}},
346        {{-1, 1}},
347        {{ 1,-1}},
348        {{-1,-1}},
349        {{ 1, 1}},
350    };
351
352    RoqContext *const roq = &enc->common;
353    int diff, lowestdiff, oldbest;
354    int off[3];
355    motion_vect bestpick = {{0,0}};
356    int i, j, k, offset;
357
358    motion_vect *last_motion;
359    motion_vect *this_motion;
360    motion_vect vect, vect2;
361    const int max = (roq->width / blocksize) * roq->height / blocksize;
362
363    if (blocksize == 4) {
364        last_motion = enc->last_motion4;
365        this_motion = enc->this_motion4;
366    } else {
367        last_motion = enc->last_motion8;
368        this_motion = enc->this_motion8;
369    }
370
371    for (i = 0; i< roq->height; i += blocksize)
372        for (j = 0; j < roq->width; j += blocksize) {
373            lowestdiff = eval_motion_dist(enc, j, i, (motion_vect) {{0,0}},
374                                          blocksize);
375            bestpick.d[0] = 0;
376            bestpick.d[1] = 0;
377
378            if (blocksize == 4)
379                EVAL_MOTION(enc->this_motion8[(i/8) * (roq->width/8) + j/8]);
380
381            offset = (i/blocksize) * roq->width / blocksize + j / blocksize;
382            if (offset < max && offset >= 0)
383                EVAL_MOTION(last_motion[offset]);
384
385            offset++;
386            if (offset < max && offset >= 0)
387                EVAL_MOTION(last_motion[offset]);
388
389            offset = (i/blocksize + 1) * roq->width / blocksize + j / blocksize;
390            if (offset < max && offset >= 0)
391                EVAL_MOTION(last_motion[offset]);
392
393            off[0]= (i/blocksize) * roq->width / blocksize + j/blocksize - 1;
394            off[1]= off[0] - roq->width / blocksize + 1;
395            off[2]= off[1] + 1;
396
397            if (i) {
398
399                for(k=0; k<2; k++)
400                    vect.d[k]= mid_pred(this_motion[off[0]].d[k],
401                                        this_motion[off[1]].d[k],
402                                        this_motion[off[2]].d[k]);
403
404                EVAL_MOTION(vect);
405                for(k=0; k<3; k++)
406                    EVAL_MOTION(this_motion[off[k]]);
407            } else if(j)
408                EVAL_MOTION(this_motion[off[0]]);
409
410            vect = bestpick;
411
412            oldbest = -1;
413            while (oldbest != lowestdiff) {
414                oldbest = lowestdiff;
415                for (k=0; k<8; k++) {
416                    vect2 = vect;
417                    vect2.d[0] += offsets[k].d[0];
418                    vect2.d[1] += offsets[k].d[1];
419                    EVAL_MOTION(vect2);
420                }
421                vect = bestpick;
422            }
423            offset = (i/blocksize) * roq->width / blocksize + j/blocksize;
424            this_motion[offset] = bestpick;
425        }
426}
427
428/**
429 * Get distortion for all options available to a subcel
430 */
431static void gather_data_for_subcel(SubcelEvaluation *subcel, int x,
432                                   int y, RoqEncContext *enc)
433{
434    RoqContext *const roq = &enc->common;
435    RoqTempData *const tempData = &enc->tmp_data;
436    uint8_t mb4[4*4*3];
437    uint8_t mb2[2*2*3];
438    int cluster_index;
439    int i, best_dist;
440
441    static const int bitsUsed[4] = {2, 10, 10, 34};
442
443    if (enc->framesSinceKeyframe >= 1) {
444        subcel->motion = enc->this_motion4[y * roq->width / 16 + x / 4];
445
446        subcel->eval_dist[RoQ_ID_FCC] =
447            eval_motion_dist(enc, x, y,
448                             enc->this_motion4[y * roq->width / 16 + x / 4], 4);
449    } else
450        subcel->eval_dist[RoQ_ID_FCC] = INT_MAX;
451
452    if (enc->framesSinceKeyframe >= 2)
453        subcel->eval_dist[RoQ_ID_MOT] = block_sse(enc->frame_to_enc->data,
454                                                  roq->current_frame->data, x,
455                                                  y, x, y,
456                                                  enc->frame_to_enc->linesize,
457                                                  roq->current_frame->linesize,
458                                                  4);
459    else
460        subcel->eval_dist[RoQ_ID_MOT] = INT_MAX;
461
462    cluster_index = y * roq->width / 16 + x / 4;
463
464    get_frame_mb(enc->frame_to_enc, x, y, mb4, 4);
465
466    subcel->eval_dist[RoQ_ID_SLD] = index_mb(mb4,
467                                             tempData->codebooks.unpacked_cb4,
468                                             tempData->codebooks.numCB4,
469                                             &subcel->cbEntry, 4);
470
471    subcel->eval_dist[RoQ_ID_CCC] = 0;
472
473    for(i=0;i<4;i++) {
474        subcel->subCels[i] = enc->closest_cb[cluster_index*4+i];
475
476        get_frame_mb(enc->frame_to_enc, x+2*(i&1),
477                     y+(i&2), mb2, 2);
478
479        subcel->eval_dist[RoQ_ID_CCC] +=
480            squared_diff_macroblock(tempData->codebooks.unpacked_cb2 + subcel->subCels[i]*2*2*3, mb2, 2);
481    }
482
483    best_dist = INT_MAX;
484    for (i=0; i<4; i++)
485        if (ROQ_LAMBDA_SCALE*subcel->eval_dist[i] + enc->lambda*bitsUsed[i] <
486            best_dist) {
487            subcel->best_coding = i;
488            subcel->best_bit_use = bitsUsed[i];
489            best_dist = ROQ_LAMBDA_SCALE*subcel->eval_dist[i] +
490                enc->lambda*bitsUsed[i];
491        }
492}
493
494/**
495 * Get distortion for all options available to a cel
496 */
497static void gather_data_for_cel(CelEvaluation *cel, RoqEncContext *enc)
498{
499    RoqContext *const roq = &enc->common;
500    RoqTempData *const tempData = &enc->tmp_data;
501    uint8_t mb8[8*8*3];
502    int index = cel->sourceY * roq->width / 64 + cel->sourceX/8;
503    int i, j, best_dist, divide_bit_use;
504
505    int bitsUsed[4] = {2, 10, 10, 0};
506
507    if (enc->framesSinceKeyframe >= 1) {
508        cel->motion = enc->this_motion8[index];
509
510        cel->eval_dist[RoQ_ID_FCC] =
511            eval_motion_dist(enc, cel->sourceX, cel->sourceY,
512                             enc->this_motion8[index], 8);
513    } else
514        cel->eval_dist[RoQ_ID_FCC] = INT_MAX;
515
516    if (enc->framesSinceKeyframe >= 2)
517        cel->eval_dist[RoQ_ID_MOT] = block_sse(enc->frame_to_enc->data,
518                                               roq->current_frame->data,
519                                               cel->sourceX, cel->sourceY,
520                                               cel->sourceX, cel->sourceY,
521                                               enc->frame_to_enc->linesize,
522                                               roq->current_frame->linesize,8);
523    else
524        cel->eval_dist[RoQ_ID_MOT] = INT_MAX;
525
526    get_frame_mb(enc->frame_to_enc, cel->sourceX, cel->sourceY, mb8, 8);
527
528    cel->eval_dist[RoQ_ID_SLD] =
529        index_mb(mb8, tempData->codebooks.unpacked_cb4_enlarged,
530                 tempData->codebooks.numCB4, &cel->cbEntry, 8);
531
532    gather_data_for_subcel(cel->subCels + 0, cel->sourceX+0, cel->sourceY+0, enc);
533    gather_data_for_subcel(cel->subCels + 1, cel->sourceX+4, cel->sourceY+0, enc);
534    gather_data_for_subcel(cel->subCels + 2, cel->sourceX+0, cel->sourceY+4, enc);
535    gather_data_for_subcel(cel->subCels + 3, cel->sourceX+4, cel->sourceY+4, enc);
536
537    cel->eval_dist[RoQ_ID_CCC] = 0;
538    divide_bit_use = 0;
539    for (i=0; i<4; i++) {
540        cel->eval_dist[RoQ_ID_CCC] +=
541            cel->subCels[i].eval_dist[cel->subCels[i].best_coding];
542        divide_bit_use += cel->subCels[i].best_bit_use;
543    }
544
545    best_dist = INT_MAX;
546    bitsUsed[3] = 2 + divide_bit_use;
547
548    for (i=0; i<4; i++)
549        if (ROQ_LAMBDA_SCALE*cel->eval_dist[i] + enc->lambda*bitsUsed[i] <
550            best_dist) {
551            cel->best_coding = i;
552            best_dist = ROQ_LAMBDA_SCALE*cel->eval_dist[i] +
553                enc->lambda*bitsUsed[i];
554        }
555
556    tempData->used_option[cel->best_coding]++;
557    tempData->mainChunkSize += bitsUsed[cel->best_coding];
558
559    if (cel->best_coding == RoQ_ID_SLD)
560        tempData->codebooks.usedCB4[cel->cbEntry]++;
561
562    if (cel->best_coding == RoQ_ID_CCC)
563        for (i=0; i<4; i++) {
564            if (cel->subCels[i].best_coding == RoQ_ID_SLD)
565                tempData->codebooks.usedCB4[cel->subCels[i].cbEntry]++;
566            else if (cel->subCels[i].best_coding == RoQ_ID_CCC)
567                for (j=0; j<4; j++)
568                    tempData->codebooks.usedCB2[cel->subCels[i].subCels[j]]++;
569        }
570}
571
572static void remap_codebooks(RoqEncContext *enc)
573{
574    RoqContext *const roq = &enc->common;
575    RoqTempData *const tempData = &enc->tmp_data;
576    int i, j, idx=0;
577
578    /* Make remaps for the final codebook usage */
579    for (i=0; i<(enc->quake3_compat ? MAX_CBS_4x4-1 : MAX_CBS_4x4); i++) {
580        if (tempData->codebooks.usedCB4[i]) {
581            tempData->i2f4[i] = idx;
582            tempData->f2i4[idx] = i;
583            for (j=0; j<4; j++)
584                tempData->codebooks.usedCB2[roq->cb4x4[i].idx[j]]++;
585            idx++;
586        }
587    }
588
589    tempData->numCB4 = idx;
590
591    idx = 0;
592    for (i=0; i<MAX_CBS_2x2; i++) {
593        if (tempData->codebooks.usedCB2[i]) {
594            tempData->i2f2[i] = idx;
595            tempData->f2i2[idx] = i;
596            idx++;
597        }
598    }
599    tempData->numCB2 = idx;
600
601}
602
603/**
604 * Write codebook chunk
605 */
606static void write_codebooks(RoqEncContext *enc)
607{
608    RoqContext *const roq = &enc->common;
609    RoqTempData *const tempData = &enc->tmp_data;
610    int i, j;
611    uint8_t **outp= &enc->out_buf;
612
613    if (tempData->numCB2) {
614        bytestream_put_le16(outp, RoQ_QUAD_CODEBOOK);
615        bytestream_put_le32(outp, tempData->numCB2*6 + tempData->numCB4*4);
616        bytestream_put_byte(outp, tempData->numCB4);
617        bytestream_put_byte(outp, tempData->numCB2);
618
619        for (i=0; i<tempData->numCB2; i++) {
620            bytestream_put_buffer(outp, roq->cb2x2[tempData->f2i2[i]].y, 4);
621            bytestream_put_byte(outp, roq->cb2x2[tempData->f2i2[i]].u);
622            bytestream_put_byte(outp, roq->cb2x2[tempData->f2i2[i]].v);
623        }
624
625        for (i=0; i<tempData->numCB4; i++)
626            for (j=0; j<4; j++)
627                bytestream_put_byte(outp, tempData->i2f2[roq->cb4x4[tempData->f2i4[i]].idx[j]]);
628
629    }
630}
631
632static inline uint8_t motion_arg(motion_vect mot)
633{
634    uint8_t ax = 8 - ((uint8_t) mot.d[0]);
635    uint8_t ay = 8 - ((uint8_t) mot.d[1]);
636    return ((ax&15)<<4) | (ay&15);
637}
638
639typedef struct CodingSpool {
640    int typeSpool;
641    int typeSpoolLength;
642    uint8_t argumentSpool[64];
643    uint8_t *args;
644    uint8_t **pout;
645} CodingSpool;
646
647/* NOTE: Typecodes must be spooled AFTER arguments!! */
648static void write_typecode(CodingSpool *s, uint8_t type)
649{
650    s->typeSpool |= (type & 3) << (14 - s->typeSpoolLength);
651    s->typeSpoolLength += 2;
652    if (s->typeSpoolLength == 16) {
653        bytestream_put_le16(s->pout, s->typeSpool);
654        bytestream_put_buffer(s->pout, s->argumentSpool,
655                              s->args - s->argumentSpool);
656        s->typeSpoolLength = 0;
657        s->typeSpool = 0;
658        s->args = s->argumentSpool;
659    }
660}
661
662static void reconstruct_and_encode_image(RoqEncContext *enc,
663                                         int w, int h, int numBlocks)
664{
665    RoqContext *const roq = &enc->common;
666    RoqTempData *const tempData = &enc->tmp_data;
667    int i, j, k;
668    int x, y;
669    int subX, subY;
670
671    roq_qcell *qcell;
672    CelEvaluation *eval;
673
674    CodingSpool spool;
675
676    spool.typeSpool=0;
677    spool.typeSpoolLength=0;
678    spool.args = spool.argumentSpool;
679    spool.pout = &enc->out_buf;
680
681    if (tempData->used_option[RoQ_ID_CCC]%2)
682        tempData->mainChunkSize+=8; //FIXME
683
684    /* Write the video chunk header */
685    bytestream_put_le16(&enc->out_buf, RoQ_QUAD_VQ);
686    bytestream_put_le32(&enc->out_buf, tempData->mainChunkSize/8);
687    bytestream_put_byte(&enc->out_buf, 0x0);
688    bytestream_put_byte(&enc->out_buf, 0x0);
689
690    for (i=0; i<numBlocks; i++) {
691        eval = enc->cel_evals + i;
692
693        x = eval->sourceX;
694        y = eval->sourceY;
695
696        switch (eval->best_coding) {
697        case RoQ_ID_MOT:
698            write_typecode(&spool, RoQ_ID_MOT);
699            break;
700
701        case RoQ_ID_FCC:
702            bytestream_put_byte(&spool.args, motion_arg(eval->motion));
703
704            write_typecode(&spool, RoQ_ID_FCC);
705            ff_apply_motion_8x8(roq, x, y,
706                                eval->motion.d[0], eval->motion.d[1]);
707            break;
708
709        case RoQ_ID_SLD:
710            bytestream_put_byte(&spool.args, tempData->i2f4[eval->cbEntry]);
711            write_typecode(&spool, RoQ_ID_SLD);
712
713            qcell = roq->cb4x4 + eval->cbEntry;
714            ff_apply_vector_4x4(roq, x  , y  , roq->cb2x2 + qcell->idx[0]);
715            ff_apply_vector_4x4(roq, x+4, y  , roq->cb2x2 + qcell->idx[1]);
716            ff_apply_vector_4x4(roq, x  , y+4, roq->cb2x2 + qcell->idx[2]);
717            ff_apply_vector_4x4(roq, x+4, y+4, roq->cb2x2 + qcell->idx[3]);
718            break;
719
720        case RoQ_ID_CCC:
721            write_typecode(&spool, RoQ_ID_CCC);
722
723            for (j=0; j<4; j++) {
724                subX = x + 4*(j&1);
725                subY = y + 2*(j&2);
726
727                switch(eval->subCels[j].best_coding) {
728                case RoQ_ID_MOT:
729                    break;
730
731                case RoQ_ID_FCC:
732                    bytestream_put_byte(&spool.args,
733                                        motion_arg(eval->subCels[j].motion));
734
735                    ff_apply_motion_4x4(roq, subX, subY,
736                                        eval->subCels[j].motion.d[0],
737                                        eval->subCels[j].motion.d[1]);
738                    break;
739
740                case RoQ_ID_SLD:
741                    bytestream_put_byte(&spool.args,
742                                        tempData->i2f4[eval->subCels[j].cbEntry]);
743
744                    qcell = roq->cb4x4 + eval->subCels[j].cbEntry;
745
746                    ff_apply_vector_2x2(roq, subX  , subY  ,
747                                        roq->cb2x2 + qcell->idx[0]);
748                    ff_apply_vector_2x2(roq, subX+2, subY  ,
749                                        roq->cb2x2 + qcell->idx[1]);
750                    ff_apply_vector_2x2(roq, subX  , subY+2,
751                                        roq->cb2x2 + qcell->idx[2]);
752                    ff_apply_vector_2x2(roq, subX+2, subY+2,
753                                        roq->cb2x2 + qcell->idx[3]);
754                    break;
755
756                case RoQ_ID_CCC:
757                    for (k=0; k<4; k++) {
758                        int cb_idx = eval->subCels[j].subCels[k];
759                        bytestream_put_byte(&spool.args,
760                                            tempData->i2f2[cb_idx]);
761
762                        ff_apply_vector_2x2(roq, subX + 2*(k&1), subY + (k&2),
763                                            roq->cb2x2 + cb_idx);
764                    }
765                    break;
766                }
767                write_typecode(&spool, eval->subCels[j].best_coding);
768            }
769            break;
770        }
771    }
772
773    /* Flush the remainder of the argument/type spool */
774    while (spool.typeSpoolLength)
775        write_typecode(&spool, 0x0);
776}
777
778
779/**
780 * Create a single YUV cell from a 2x2 section of the image
781 */
782static inline void frame_block_to_cell(int *block, uint8_t * const *data,
783                                       int top, int left, const int *stride)
784{
785    int i, j, u=0, v=0;
786
787    for (i=0; i<2; i++)
788        for (j=0; j<2; j++) {
789            int x = (top+i)*stride[0] + left + j;
790            *block++ = data[0][x];
791            x = (top+i)*stride[1] + left + j;
792            u       += data[1][x];
793            v       += data[2][x];
794        }
795
796    *block++ = (u + 2) / 4 * CHROMA_BIAS;
797    *block++ = (v + 2) / 4 * CHROMA_BIAS;
798}
799
800/**
801 * Create YUV clusters for the entire image
802 */
803static void create_clusters(const AVFrame *frame, int w, int h, int *points)
804{
805    int i, j, k, l;
806
807    for (i=0; i<h; i+=4)
808        for (j=0; j<w; j+=4) {
809            for (k=0; k < 2; k++)
810                for (l=0; l < 2; l++)
811                    frame_block_to_cell(points + (l + 2*k)*6, frame->data,
812                                        i+2*k, j+2*l, frame->linesize);
813            points += 24;
814        }
815}
816
817static int generate_codebook(RoqEncContext *enc,
818                             int *points, int inputCount, roq_cell *results,
819                             int size, int cbsize)
820{
821    int i, j, k, ret = 0;
822    int c_size = size*size/4;
823    int *buf;
824    int *codebook = enc->tmp_codebook_buf;
825    int *closest_cb = enc->closest_cb;
826
827    ret = avpriv_elbg_do(&enc->elbg, points, 6 * c_size, inputCount, codebook,
828                         cbsize, 1, closest_cb, &enc->randctx, 0);
829    if (ret < 0)
830        return ret;
831
832    buf = codebook;
833    for (i=0; i<cbsize; i++)
834        for (k=0; k<c_size; k++) {
835            for(j=0; j<4; j++)
836                results->y[j] = *buf++;
837
838            results->u =    (*buf++ + CHROMA_BIAS/2)/CHROMA_BIAS;
839            results->v =    (*buf++ + CHROMA_BIAS/2)/CHROMA_BIAS;
840            results++;
841        }
842    return 0;
843}
844
845static int generate_new_codebooks(RoqEncContext *enc)
846{
847    int i, j, ret = 0;
848    RoqCodebooks *codebooks = &enc->tmp_data.codebooks;
849    RoqContext *const roq = &enc->common;
850    int max = roq->width * roq->height / 16;
851    uint8_t mb2[3*4];
852    int *points = enc->points;
853
854    /* Subsample YUV data */
855    create_clusters(enc->frame_to_enc, roq->width, roq->height, points);
856
857    codebooks->numCB4 = (enc->quake3_compat ? MAX_CBS_4x4-1 : MAX_CBS_4x4);
858
859    /* Create 4x4 codebooks */
860    if ((ret = generate_codebook(enc, points, max, enc->results4,
861                                 4, codebooks->numCB4)) < 0)
862        return ret;
863
864    /* Create 2x2 codebooks */
865    if ((ret = generate_codebook(enc, points, max * 4,
866                                 roq->cb2x2, 2, MAX_CBS_2x2)) < 0)
867        return ret;
868
869    codebooks->numCB2 = MAX_CBS_2x2;
870
871    /* Unpack 2x2 codebook clusters */
872    for (i=0; i<codebooks->numCB2; i++)
873        unpack_roq_cell(roq->cb2x2 + i, codebooks->unpacked_cb2 + i*2*2*3);
874
875    /* Index all 4x4 entries to the 2x2 entries, unpack, and enlarge */
876    for (i=0; i<codebooks->numCB4; i++) {
877        for (j=0; j<4; j++) {
878            unpack_roq_cell(&enc->results4[4*i + j], mb2);
879            index_mb(mb2, codebooks->unpacked_cb2, codebooks->numCB2,
880                     &roq->cb4x4[i].idx[j], 2);
881        }
882        unpack_roq_qcell(codebooks->unpacked_cb2, roq->cb4x4 + i,
883                         codebooks->unpacked_cb4 + i*4*4*3);
884        enlarge_roq_mb4(codebooks->unpacked_cb4 + i*4*4*3,
885                        codebooks->unpacked_cb4_enlarged + i*8*8*3);
886    }
887
888    return 0;
889}
890
891static int roq_encode_video(RoqEncContext *enc)
892{
893    RoqTempData *const tempData = &enc->tmp_data;
894    RoqContext *const roq = &enc->common;
895    int ret;
896
897    memset(tempData, 0, sizeof(*tempData));
898
899    ret = generate_new_codebooks(enc);
900    if (ret < 0)
901        return ret;
902
903    if (enc->framesSinceKeyframe >= 1) {
904        motion_search(enc, 8);
905        motion_search(enc, 4);
906    }
907
908 retry_encode:
909    for (int i = 0; i < roq->width * roq->height / 64; i++)
910        gather_data_for_cel(enc->cel_evals + i, enc);
911
912    /* Quake 3 can't handle chunks bigger than 65535 bytes */
913    if (tempData->mainChunkSize/8 > 65535 && enc->quake3_compat) {
914        if (enc->lambda > 100000) {
915            av_log(roq->avctx, AV_LOG_ERROR, "Cannot encode video in Quake compatible form\n");
916            return AVERROR(EINVAL);
917        }
918        av_log(roq->avctx, AV_LOG_ERROR,
919               "Warning, generated a frame too big for Quake (%d > 65535), "
920               "now switching to a bigger qscale value.\n",
921               tempData->mainChunkSize/8);
922        enc->lambda *= 1.5;
923        tempData->mainChunkSize = 0;
924        memset(tempData->used_option, 0, sizeof(tempData->used_option));
925        memset(tempData->codebooks.usedCB4, 0,
926               sizeof(tempData->codebooks.usedCB4));
927        memset(tempData->codebooks.usedCB2, 0,
928               sizeof(tempData->codebooks.usedCB2));
929
930        goto retry_encode;
931    }
932
933    remap_codebooks(enc);
934
935    write_codebooks(enc);
936
937    reconstruct_and_encode_image(enc, roq->width, roq->height,
938                                 roq->width * roq->height / 64);
939
940    /* Rotate frame history */
941    FFSWAP(AVFrame *,    roq->current_frame,   roq->last_frame);
942    FFSWAP(motion_vect *, enc->last_motion4, enc->this_motion4);
943    FFSWAP(motion_vect *, enc->last_motion8, enc->this_motion8);
944
945    enc->framesSinceKeyframe++;
946
947    return 0;
948}
949
950static av_cold int roq_encode_end(AVCodecContext *avctx)
951{
952    RoqEncContext *const enc = avctx->priv_data;
953
954    av_frame_free(&enc->common.current_frame);
955    av_frame_free(&enc->common.last_frame);
956
957    av_freep(&enc->cel_evals);
958    av_freep(&enc->closest_cb);
959    av_freep(&enc->this_motion4);
960    av_freep(&enc->last_motion4);
961    av_freep(&enc->this_motion8);
962    av_freep(&enc->last_motion8);
963
964    avpriv_elbg_free(&enc->elbg);
965
966    return 0;
967}
968
969static av_cold int roq_encode_init(AVCodecContext *avctx)
970{
971    RoqEncContext *const enc = avctx->priv_data;
972    RoqContext    *const roq = &enc->common;
973
974    av_lfg_init(&enc->randctx, 1);
975
976    roq->avctx = avctx;
977
978    enc->framesSinceKeyframe = 0;
979    if ((avctx->width & 0xf) || (avctx->height & 0xf)) {
980        av_log(avctx, AV_LOG_ERROR, "Dimensions must be divisible by 16\n");
981        return AVERROR(EINVAL);
982    }
983
984    if (avctx->width > 65535 || avctx->height > 65535) {
985        av_log(avctx, AV_LOG_ERROR, "Dimensions are max %d\n", enc->quake3_compat ? 32768 : 65535);
986        return AVERROR(EINVAL);
987    }
988
989    if (((avctx->width)&(avctx->width-1))||((avctx->height)&(avctx->height-1)))
990        av_log(avctx, AV_LOG_ERROR, "Warning: dimensions not power of two, this is not supported by quake\n");
991
992    roq->width  = avctx->width;
993    roq->height = avctx->height;
994
995    enc->framesSinceKeyframe = 0;
996    enc->first_frame = 1;
997
998    roq->last_frame    = av_frame_alloc();
999    roq->current_frame = av_frame_alloc();
1000    if (!roq->last_frame || !roq->current_frame)
1001        return AVERROR(ENOMEM);
1002
1003    enc->this_motion4 =
1004        av_calloc(roq->width * roq->height / 16, sizeof(*enc->this_motion4));
1005
1006    enc->last_motion4 =
1007        av_malloc_array (roq->width * roq->height / 16, sizeof(motion_vect));
1008
1009    enc->this_motion8 =
1010        av_calloc(roq->width * roq->height / 64, sizeof(*enc->this_motion8));
1011
1012    enc->last_motion8 =
1013        av_malloc_array (roq->width * roq->height / 64, sizeof(motion_vect));
1014
1015    /* 4x4 codebook needs 6 * 4 * 4 / 4 * width * height / 16 * sizeof(int);
1016     * and so does the points buffer. */
1017    enc->closest_cb   =
1018        av_malloc_array(roq->width * roq->height, 3 * sizeof(int));
1019
1020    if (!enc->this_motion4 || !enc->last_motion4 ||
1021        !enc->this_motion8 || !enc->last_motion8 || !enc->closest_cb)
1022        return AVERROR(ENOMEM);
1023
1024    enc->points = enc->closest_cb + roq->width * roq->height * 3 / 2;
1025
1026    return create_cel_evals(enc);
1027}
1028
1029static void roq_write_video_info_chunk(RoqEncContext *enc)
1030{
1031    /* ROQ info chunk */
1032    bytestream_put_le16(&enc->out_buf, RoQ_INFO);
1033
1034    /* Size: 8 bytes */
1035    bytestream_put_le32(&enc->out_buf, 8);
1036
1037    /* Unused argument */
1038    bytestream_put_byte(&enc->out_buf, 0x00);
1039    bytestream_put_byte(&enc->out_buf, 0x00);
1040
1041    /* Width */
1042    bytestream_put_le16(&enc->out_buf, enc->common.width);
1043
1044    /* Height */
1045    bytestream_put_le16(&enc->out_buf, enc->common.height);
1046
1047    /* Unused in Quake 3, mimics the output of the real encoder */
1048    bytestream_put_byte(&enc->out_buf, 0x08);
1049    bytestream_put_byte(&enc->out_buf, 0x00);
1050    bytestream_put_byte(&enc->out_buf, 0x04);
1051    bytestream_put_byte(&enc->out_buf, 0x00);
1052}
1053
1054static int roq_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1055                            const AVFrame *frame, int *got_packet)
1056{
1057    RoqEncContext *const enc = avctx->priv_data;
1058    RoqContext    *const roq = &enc->common;
1059    int size, ret;
1060
1061    roq->avctx = avctx;
1062
1063    enc->frame_to_enc = frame;
1064
1065    if (frame->quality)
1066        enc->lambda = frame->quality - 1;
1067    else
1068        enc->lambda = 2*ROQ_LAMBDA_SCALE;
1069
1070    /* 138 bits max per 8x8 block +
1071     *     256 codebooks*(6 bytes 2x2 + 4 bytes 4x4) + 8 bytes frame header */
1072    size = ((roq->width * roq->height / 64) * 138 + 7) / 8 + 256 * (6 + 4) + 8;
1073    if ((ret = ff_alloc_packet(avctx, pkt, size)) < 0)
1074        return ret;
1075    enc->out_buf = pkt->data;
1076
1077    /* Check for I-frame */
1078    if (enc->framesSinceKeyframe == avctx->gop_size)
1079        enc->framesSinceKeyframe = 0;
1080
1081    if (enc->first_frame) {
1082        /* Alloc memory for the reconstruction data (we must know the stride
1083         for that) */
1084        if ((ret = ff_encode_alloc_frame(avctx, roq->current_frame)) < 0 ||
1085            (ret = ff_encode_alloc_frame(avctx, roq->last_frame   )) < 0)
1086            return ret;
1087
1088        /* Before the first video frame, write a "video info" chunk */
1089        roq_write_video_info_chunk(enc);
1090
1091        enc->first_frame = 0;
1092    }
1093
1094    /* Encode the actual frame */
1095    ret = roq_encode_video(enc);
1096    if (ret < 0)
1097        return ret;
1098
1099    pkt->size   = enc->out_buf - pkt->data;
1100    if (enc->framesSinceKeyframe == 1)
1101        pkt->flags |= AV_PKT_FLAG_KEY;
1102    *got_packet = 1;
1103
1104    return 0;
1105}
1106
1107#define OFFSET(x) offsetof(RoqEncContext, x)
1108#define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1109static const AVOption options[] = {
1110    { "quake3_compat", "Whether to respect known limitations in Quake 3 decoder", OFFSET(quake3_compat), AV_OPT_TYPE_BOOL, { .i64 = 1 }, 0, 1, VE },
1111    { NULL },
1112};
1113
1114static const AVClass roq_class = {
1115    .class_name = "RoQ",
1116    .item_name  = av_default_item_name,
1117    .option     = options,
1118    .version    = LIBAVUTIL_VERSION_INT,
1119};
1120
1121const FFCodec ff_roq_encoder = {
1122    .p.name               = "roqvideo",
1123    .p.long_name          = NULL_IF_CONFIG_SMALL("id RoQ video"),
1124    .p.type               = AVMEDIA_TYPE_VIDEO,
1125    .p.id                 = AV_CODEC_ID_ROQ,
1126    .priv_data_size       = sizeof(RoqEncContext),
1127    .init                 = roq_encode_init,
1128    FF_CODEC_ENCODE_CB(roq_encode_frame),
1129    .close                = roq_encode_end,
1130    .p.pix_fmts           = (const enum AVPixelFormat[]){ AV_PIX_FMT_YUVJ444P,
1131                                                        AV_PIX_FMT_NONE },
1132    .p.priv_class   = &roq_class,
1133    .caps_internal        = FF_CODEC_CAP_INIT_THREADSAFE | FF_CODEC_CAP_INIT_CLEANUP,
1134};
1135