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 "elbg.h"
65 #include "internal.h"
66 #include "mathops.h"
67 
68 #define CHROMA_BIAS 1
69 
70 /**
71  * Maximum number of generated 4x4 codebooks. Can't be 256 to workaround a
72  * Quake 3 bug.
73  */
74 #define MAX_CBS_4x4 256
75 
76 #define MAX_CBS_2x2 256 ///< Maximum number of 2x2 codebooks.
77 
78 /* The cast is useful when multiplying it by INT_MAX */
79 #define ROQ_LAMBDA_SCALE ((uint64_t) FF_LAMBDA_SCALE)
80 
81 typedef struct RoqCodebooks {
82     int numCB4;
83     int numCB2;
84     int usedCB2[MAX_CBS_2x2];
85     int usedCB4[MAX_CBS_4x4];
86     uint8_t unpacked_cb2[MAX_CBS_2x2*2*2*3];
87     uint8_t unpacked_cb4[MAX_CBS_4x4*4*4*3];
88     uint8_t unpacked_cb4_enlarged[MAX_CBS_4x4*8*8*3];
89 } RoqCodebooks;
90 
91 /**
92  * Temporary vars
93  */
94 typedef struct RoqTempData
95 {
96     int f2i4[MAX_CBS_4x4];
97     int i2f4[MAX_CBS_4x4];
98     int f2i2[MAX_CBS_2x2];
99     int i2f2[MAX_CBS_2x2];
100 
101     int mainChunkSize;
102 
103     int numCB4;
104     int numCB2;
105 
106     RoqCodebooks codebooks;
107 
108     int used_option[4];
109 } RoqTempData;
110 
111 typedef struct SubcelEvaluation {
112     int eval_dist[4];
113     int best_bit_use;
114     int best_coding;
115 
116     int subCels[4];
117     motion_vect motion;
118     int cbEntry;
119 } SubcelEvaluation;
120 
121 typedef struct CelEvaluation {
122     int eval_dist[4];
123     int best_coding;
124 
125     SubcelEvaluation subCels[4];
126 
127     motion_vect motion;
128     int cbEntry;
129 
130     int sourceX, sourceY;
131 } CelEvaluation;
132 
133 typedef struct RoqEncContext {
134     RoqContext common;
135     AVLFG randctx;
136     uint64_t lambda;
137 
138     motion_vect *this_motion4;
139     motion_vect *last_motion4;
140 
141     motion_vect *this_motion8;
142     motion_vect *last_motion8;
143 
144     unsigned int framesSinceKeyframe;
145 
146     const AVFrame *frame_to_enc;
147     uint8_t *out_buf;
148     RoqTempData tmp_data;
149     roq_cell results4[4 * MAX_CBS_4x4];
150     int tmp_codebook_buf[FFMAX(24 * MAX_CBS_4x4, 6 * MAX_CBS_2x2)];
151 
152     CelEvaluation *cel_evals;
153     int *closest_cb;
154     int *points;  // Allocated together with closest_cb
155 
156     int first_frame;
157     int quake3_compat; // Quake 3 compatibility option
158 } RoqEncContext;
159 
160 /* Macroblock support functions */
unpack_roq_cell(roq_cell * cell,uint8_t u[4* 3])161 static void unpack_roq_cell(roq_cell *cell, uint8_t u[4*3])
162 {
163     memcpy(u  , cell->y, 4);
164     memset(u+4, cell->u, 4);
165     memset(u+8, cell->v, 4);
166 }
167 
unpack_roq_qcell(uint8_t cb2[],roq_qcell * qcell,uint8_t u[4* 4* 3])168 static void unpack_roq_qcell(uint8_t cb2[], roq_qcell *qcell, uint8_t u[4*4*3])
169 {
170     int i,cp;
171     static const int offsets[4] = {0, 2, 8, 10};
172 
173     for (cp=0; cp<3; cp++)
174         for (i=0; i<4; i++) {
175             u[4*4*cp + offsets[i]  ] = cb2[qcell->idx[i]*2*2*3 + 4*cp  ];
176             u[4*4*cp + offsets[i]+1] = cb2[qcell->idx[i]*2*2*3 + 4*cp+1];
177             u[4*4*cp + offsets[i]+4] = cb2[qcell->idx[i]*2*2*3 + 4*cp+2];
178             u[4*4*cp + offsets[i]+5] = cb2[qcell->idx[i]*2*2*3 + 4*cp+3];
179         }
180 }
181 
182 
enlarge_roq_mb4(uint8_t base[3* 16],uint8_t u[3* 64])183 static void enlarge_roq_mb4(uint8_t base[3*16], uint8_t u[3*64])
184 {
185     int x,y,cp;
186 
187     for(cp=0; cp<3; cp++)
188         for(y=0; y<8; y++)
189             for(x=0; x<8; x++)
190                 *u++ = base[(y/2)*4 + (x/2) + 16*cp];
191 }
192 
square(int x)193 static inline int square(int x)
194 {
195     return x*x;
196 }
197 
eval_sse(const uint8_t * a,const uint8_t * b,int count)198 static inline int eval_sse(const uint8_t *a, const uint8_t *b, int count)
199 {
200     int diff=0;
201 
202     while(count--)
203         diff += square(*b++ - *a++);
204 
205     return diff;
206 }
207 
208 // FIXME Could use DSPContext.sse, but it is not so speed critical (used
209 // just for motion estimation).
block_sse(uint8_t * const * buf1,uint8_t * const * buf2,int x1,int y1,int x2,int y2,const int * stride1,const int * stride2,int size)210 static int block_sse(uint8_t * const *buf1, uint8_t * const *buf2, int x1, int y1,
211                      int x2, int y2, const int *stride1, const int *stride2, int size)
212 {
213     int i, k;
214     int sse=0;
215 
216     for (k=0; k<3; k++) {
217         int bias = (k ? CHROMA_BIAS : 4);
218         for (i=0; i<size; i++)
219             sse += bias*eval_sse(buf1[k] + (y1+i)*stride1[k] + x1,
220                                  buf2[k] + (y2+i)*stride2[k] + x2, size);
221     }
222 
223     return sse;
224 }
225 
eval_motion_dist(RoqEncContext * enc,int x,int y,motion_vect vect,int size)226 static int eval_motion_dist(RoqEncContext *enc, int x, int y, motion_vect vect,
227                              int size)
228 {
229     RoqContext *const roq = &enc->common;
230     int mx=vect.d[0];
231     int my=vect.d[1];
232 
233     if (mx < -7 || mx > 7)
234         return INT_MAX;
235 
236     if (my < -7 || my > 7)
237         return INT_MAX;
238 
239     mx += x;
240     my += y;
241 
242     if ((unsigned) mx > roq->width-size || (unsigned) my > roq->height-size)
243         return INT_MAX;
244 
245     return block_sse(enc->frame_to_enc->data, roq->last_frame->data, x, y,
246                      mx, my,
247                      enc->frame_to_enc->linesize, roq->last_frame->linesize,
248                      size);
249 }
250 
251 /**
252  * @return distortion between two macroblocks
253  */
squared_diff_macroblock(uint8_t a[],uint8_t b[],int size)254 static inline int squared_diff_macroblock(uint8_t a[], uint8_t b[], int size)
255 {
256     int cp, sdiff=0;
257 
258     for(cp=0;cp<3;cp++) {
259         int bias = (cp ? CHROMA_BIAS : 4);
260         sdiff += bias*eval_sse(a, b, size*size);
261         a += size*size;
262         b += size*size;
263     }
264 
265     return sdiff;
266 }
267 
268 /**
269  * Initialize cel evaluators and set their source coordinates
270  */
create_cel_evals(RoqEncContext * enc)271 static int create_cel_evals(RoqEncContext *enc)
272 {
273     RoqContext *const roq = &enc->common;
274 
275     enc->cel_evals = av_malloc_array(roq->width * roq->height / 64, sizeof(CelEvaluation));
276     if (!enc->cel_evals)
277         return AVERROR(ENOMEM);
278 
279     /* Map to the ROQ quadtree order */
280     for (int y = 0, n = 0; y < roq->height; y += 16)
281         for (int x = 0; x < roq->width; x += 16)
282             for(int i = 0; i < 4; i++) {
283                 enc->cel_evals[n  ].sourceX = x + (i&1)*8;
284                 enc->cel_evals[n++].sourceY = y + (i&2)*4;
285             }
286 
287     return 0;
288 }
289 
290 /**
291  * Get macroblocks from parts of the image
292  */
get_frame_mb(const AVFrame * frame,int x,int y,uint8_t mb[],int dim)293 static void get_frame_mb(const AVFrame *frame, int x, int y, uint8_t mb[], int dim)
294 {
295     int i, j, cp;
296 
297     for (cp=0; cp<3; cp++) {
298         int stride = frame->linesize[cp];
299         for (i=0; i<dim; i++)
300             for (j=0; j<dim; j++)
301                 *mb++ = frame->data[cp][(y+i)*stride + x + j];
302     }
303 }
304 
305 /**
306  * Find the codebook with the lowest distortion from an image
307  */
index_mb(uint8_t cluster[],uint8_t cb[],int numCB,int * outIndex,int dim)308 static int index_mb(uint8_t cluster[], uint8_t cb[], int numCB,
309                     int *outIndex, int dim)
310 {
311     int i, lDiff = INT_MAX, pick=0;
312 
313     /* Diff against the others */
314     for (i=0; i<numCB; i++) {
315         int diff = squared_diff_macroblock(cluster, cb + i*dim*dim*3, dim);
316         if (diff < lDiff) {
317             lDiff = diff;
318             pick = i;
319         }
320     }
321 
322     *outIndex = pick;
323     return lDiff;
324 }
325 
326 #define EVAL_MOTION(MOTION) \
327     do { \
328         diff = eval_motion_dist(enc, j, i, MOTION, blocksize); \
329             \
330         if (diff < lowestdiff) { \
331             lowestdiff = diff; \
332             bestpick = MOTION; \
333         } \
334     } while(0)
335 
motion_search(RoqEncContext * enc,int blocksize)336 static void motion_search(RoqEncContext *enc, int blocksize)
337 {
338     static const motion_vect offsets[8] = {
339         {{ 0,-1}},
340         {{ 0, 1}},
341         {{-1, 0}},
342         {{ 1, 0}},
343         {{-1, 1}},
344         {{ 1,-1}},
345         {{-1,-1}},
346         {{ 1, 1}},
347     };
348 
349     RoqContext *const roq = &enc->common;
350     int diff, lowestdiff, oldbest;
351     int off[3];
352     motion_vect bestpick = {{0,0}};
353     int i, j, k, offset;
354 
355     motion_vect *last_motion;
356     motion_vect *this_motion;
357     motion_vect vect, vect2;
358     const int max = (roq->width / blocksize) * roq->height / blocksize;
359 
360     if (blocksize == 4) {
361         last_motion = enc->last_motion4;
362         this_motion = enc->this_motion4;
363     } else {
364         last_motion = enc->last_motion8;
365         this_motion = enc->this_motion8;
366     }
367 
368     for (i = 0; i< roq->height; i += blocksize)
369         for (j = 0; j < roq->width; j += blocksize) {
370             lowestdiff = eval_motion_dist(enc, j, i, (motion_vect) {{0,0}},
371                                           blocksize);
372             bestpick.d[0] = 0;
373             bestpick.d[1] = 0;
374 
375             if (blocksize == 4)
376                 EVAL_MOTION(enc->this_motion8[(i/8) * (roq->width/8) + j/8]);
377 
378             offset = (i/blocksize) * roq->width / blocksize + j / blocksize;
379             if (offset < max && offset >= 0)
380                 EVAL_MOTION(last_motion[offset]);
381 
382             offset++;
383             if (offset < max && offset >= 0)
384                 EVAL_MOTION(last_motion[offset]);
385 
386             offset = (i/blocksize + 1) * roq->width / blocksize + j / blocksize;
387             if (offset < max && offset >= 0)
388                 EVAL_MOTION(last_motion[offset]);
389 
390             off[0]= (i/blocksize) * roq->width / blocksize + j/blocksize - 1;
391             off[1]= off[0] - roq->width / blocksize + 1;
392             off[2]= off[1] + 1;
393 
394             if (i) {
395 
396                 for(k=0; k<2; k++)
397                     vect.d[k]= mid_pred(this_motion[off[0]].d[k],
398                                         this_motion[off[1]].d[k],
399                                         this_motion[off[2]].d[k]);
400 
401                 EVAL_MOTION(vect);
402                 for(k=0; k<3; k++)
403                     EVAL_MOTION(this_motion[off[k]]);
404             } else if(j)
405                 EVAL_MOTION(this_motion[off[0]]);
406 
407             vect = bestpick;
408 
409             oldbest = -1;
410             while (oldbest != lowestdiff) {
411                 oldbest = lowestdiff;
412                 for (k=0; k<8; k++) {
413                     vect2 = vect;
414                     vect2.d[0] += offsets[k].d[0];
415                     vect2.d[1] += offsets[k].d[1];
416                     EVAL_MOTION(vect2);
417                 }
418                 vect = bestpick;
419             }
420             offset = (i/blocksize) * roq->width / blocksize + j/blocksize;
421             this_motion[offset] = bestpick;
422         }
423 }
424 
425 /**
426  * Get distortion for all options available to a subcel
427  */
gather_data_for_subcel(SubcelEvaluation * subcel,int x,int y,RoqEncContext * enc)428 static void gather_data_for_subcel(SubcelEvaluation *subcel, int x,
429                                    int y, RoqEncContext *enc)
430 {
431     RoqContext *const roq = &enc->common;
432     RoqTempData *const tempData = &enc->tmp_data;
433     uint8_t mb4[4*4*3];
434     uint8_t mb2[2*2*3];
435     int cluster_index;
436     int i, best_dist;
437 
438     static const int bitsUsed[4] = {2, 10, 10, 34};
439 
440     if (enc->framesSinceKeyframe >= 1) {
441         subcel->motion = enc->this_motion4[y * roq->width / 16 + x / 4];
442 
443         subcel->eval_dist[RoQ_ID_FCC] =
444             eval_motion_dist(enc, x, y,
445                              enc->this_motion4[y * roq->width / 16 + x / 4], 4);
446     } else
447         subcel->eval_dist[RoQ_ID_FCC] = INT_MAX;
448 
449     if (enc->framesSinceKeyframe >= 2)
450         subcel->eval_dist[RoQ_ID_MOT] = block_sse(enc->frame_to_enc->data,
451                                                   roq->current_frame->data, x,
452                                                   y, x, y,
453                                                   enc->frame_to_enc->linesize,
454                                                   roq->current_frame->linesize,
455                                                   4);
456     else
457         subcel->eval_dist[RoQ_ID_MOT] = INT_MAX;
458 
459     cluster_index = y * roq->width / 16 + x / 4;
460 
461     get_frame_mb(enc->frame_to_enc, x, y, mb4, 4);
462 
463     subcel->eval_dist[RoQ_ID_SLD] = index_mb(mb4,
464                                              tempData->codebooks.unpacked_cb4,
465                                              tempData->codebooks.numCB4,
466                                              &subcel->cbEntry, 4);
467 
468     subcel->eval_dist[RoQ_ID_CCC] = 0;
469 
470     for(i=0;i<4;i++) {
471         subcel->subCels[i] = enc->closest_cb[cluster_index*4+i];
472 
473         get_frame_mb(enc->frame_to_enc, x+2*(i&1),
474                      y+(i&2), mb2, 2);
475 
476         subcel->eval_dist[RoQ_ID_CCC] +=
477             squared_diff_macroblock(tempData->codebooks.unpacked_cb2 + subcel->subCels[i]*2*2*3, mb2, 2);
478     }
479 
480     best_dist = INT_MAX;
481     for (i=0; i<4; i++)
482         if (ROQ_LAMBDA_SCALE*subcel->eval_dist[i] + enc->lambda*bitsUsed[i] <
483             best_dist) {
484             subcel->best_coding = i;
485             subcel->best_bit_use = bitsUsed[i];
486             best_dist = ROQ_LAMBDA_SCALE*subcel->eval_dist[i] +
487                 enc->lambda*bitsUsed[i];
488         }
489 }
490 
491 /**
492  * Get distortion for all options available to a cel
493  */
gather_data_for_cel(CelEvaluation * cel,RoqEncContext * enc)494 static void gather_data_for_cel(CelEvaluation *cel, RoqEncContext *enc)
495 {
496     RoqContext *const roq = &enc->common;
497     RoqTempData *const tempData = &enc->tmp_data;
498     uint8_t mb8[8*8*3];
499     int index = cel->sourceY * roq->width / 64 + cel->sourceX/8;
500     int i, j, best_dist, divide_bit_use;
501 
502     int bitsUsed[4] = {2, 10, 10, 0};
503 
504     if (enc->framesSinceKeyframe >= 1) {
505         cel->motion = enc->this_motion8[index];
506 
507         cel->eval_dist[RoQ_ID_FCC] =
508             eval_motion_dist(enc, cel->sourceX, cel->sourceY,
509                              enc->this_motion8[index], 8);
510     } else
511         cel->eval_dist[RoQ_ID_FCC] = INT_MAX;
512 
513     if (enc->framesSinceKeyframe >= 2)
514         cel->eval_dist[RoQ_ID_MOT] = block_sse(enc->frame_to_enc->data,
515                                                roq->current_frame->data,
516                                                cel->sourceX, cel->sourceY,
517                                                cel->sourceX, cel->sourceY,
518                                                enc->frame_to_enc->linesize,
519                                                roq->current_frame->linesize,8);
520     else
521         cel->eval_dist[RoQ_ID_MOT] = INT_MAX;
522 
523     get_frame_mb(enc->frame_to_enc, cel->sourceX, cel->sourceY, mb8, 8);
524 
525     cel->eval_dist[RoQ_ID_SLD] =
526         index_mb(mb8, tempData->codebooks.unpacked_cb4_enlarged,
527                  tempData->codebooks.numCB4, &cel->cbEntry, 8);
528 
529     gather_data_for_subcel(cel->subCels + 0, cel->sourceX+0, cel->sourceY+0, enc);
530     gather_data_for_subcel(cel->subCels + 1, cel->sourceX+4, cel->sourceY+0, enc);
531     gather_data_for_subcel(cel->subCels + 2, cel->sourceX+0, cel->sourceY+4, enc);
532     gather_data_for_subcel(cel->subCels + 3, cel->sourceX+4, cel->sourceY+4, enc);
533 
534     cel->eval_dist[RoQ_ID_CCC] = 0;
535     divide_bit_use = 0;
536     for (i=0; i<4; i++) {
537         cel->eval_dist[RoQ_ID_CCC] +=
538             cel->subCels[i].eval_dist[cel->subCels[i].best_coding];
539         divide_bit_use += cel->subCels[i].best_bit_use;
540     }
541 
542     best_dist = INT_MAX;
543     bitsUsed[3] = 2 + divide_bit_use;
544 
545     for (i=0; i<4; i++)
546         if (ROQ_LAMBDA_SCALE*cel->eval_dist[i] + enc->lambda*bitsUsed[i] <
547             best_dist) {
548             cel->best_coding = i;
549             best_dist = ROQ_LAMBDA_SCALE*cel->eval_dist[i] +
550                 enc->lambda*bitsUsed[i];
551         }
552 
553     tempData->used_option[cel->best_coding]++;
554     tempData->mainChunkSize += bitsUsed[cel->best_coding];
555 
556     if (cel->best_coding == RoQ_ID_SLD)
557         tempData->codebooks.usedCB4[cel->cbEntry]++;
558 
559     if (cel->best_coding == RoQ_ID_CCC)
560         for (i=0; i<4; i++) {
561             if (cel->subCels[i].best_coding == RoQ_ID_SLD)
562                 tempData->codebooks.usedCB4[cel->subCels[i].cbEntry]++;
563             else if (cel->subCels[i].best_coding == RoQ_ID_CCC)
564                 for (j=0; j<4; j++)
565                     tempData->codebooks.usedCB2[cel->subCels[i].subCels[j]]++;
566         }
567 }
568 
remap_codebooks(RoqEncContext * enc)569 static void remap_codebooks(RoqEncContext *enc)
570 {
571     RoqContext *const roq = &enc->common;
572     RoqTempData *const tempData = &enc->tmp_data;
573     int i, j, idx=0;
574 
575     /* Make remaps for the final codebook usage */
576     for (i=0; i<(enc->quake3_compat ? MAX_CBS_4x4-1 : MAX_CBS_4x4); i++) {
577         if (tempData->codebooks.usedCB4[i]) {
578             tempData->i2f4[i] = idx;
579             tempData->f2i4[idx] = i;
580             for (j=0; j<4; j++)
581                 tempData->codebooks.usedCB2[roq->cb4x4[i].idx[j]]++;
582             idx++;
583         }
584     }
585 
586     tempData->numCB4 = idx;
587 
588     idx = 0;
589     for (i=0; i<MAX_CBS_2x2; i++) {
590         if (tempData->codebooks.usedCB2[i]) {
591             tempData->i2f2[i] = idx;
592             tempData->f2i2[idx] = i;
593             idx++;
594         }
595     }
596     tempData->numCB2 = idx;
597 
598 }
599 
600 /**
601  * Write codebook chunk
602  */
write_codebooks(RoqEncContext * enc)603 static void write_codebooks(RoqEncContext *enc)
604 {
605     RoqContext *const roq = &enc->common;
606     RoqTempData *const tempData = &enc->tmp_data;
607     int i, j;
608     uint8_t **outp= &enc->out_buf;
609 
610     if (tempData->numCB2) {
611         bytestream_put_le16(outp, RoQ_QUAD_CODEBOOK);
612         bytestream_put_le32(outp, tempData->numCB2*6 + tempData->numCB4*4);
613         bytestream_put_byte(outp, tempData->numCB4);
614         bytestream_put_byte(outp, tempData->numCB2);
615 
616         for (i=0; i<tempData->numCB2; i++) {
617             bytestream_put_buffer(outp, roq->cb2x2[tempData->f2i2[i]].y, 4);
618             bytestream_put_byte(outp, roq->cb2x2[tempData->f2i2[i]].u);
619             bytestream_put_byte(outp, roq->cb2x2[tempData->f2i2[i]].v);
620         }
621 
622         for (i=0; i<tempData->numCB4; i++)
623             for (j=0; j<4; j++)
624                 bytestream_put_byte(outp, tempData->i2f2[roq->cb4x4[tempData->f2i4[i]].idx[j]]);
625 
626     }
627 }
628 
motion_arg(motion_vect mot)629 static inline uint8_t motion_arg(motion_vect mot)
630 {
631     uint8_t ax = 8 - ((uint8_t) mot.d[0]);
632     uint8_t ay = 8 - ((uint8_t) mot.d[1]);
633     return ((ax&15)<<4) | (ay&15);
634 }
635 
636 typedef struct CodingSpool {
637     int typeSpool;
638     int typeSpoolLength;
639     uint8_t argumentSpool[64];
640     uint8_t *args;
641     uint8_t **pout;
642 } CodingSpool;
643 
644 /* NOTE: Typecodes must be spooled AFTER arguments!! */
write_typecode(CodingSpool * s,uint8_t type)645 static void write_typecode(CodingSpool *s, uint8_t type)
646 {
647     s->typeSpool |= (type & 3) << (14 - s->typeSpoolLength);
648     s->typeSpoolLength += 2;
649     if (s->typeSpoolLength == 16) {
650         bytestream_put_le16(s->pout, s->typeSpool);
651         bytestream_put_buffer(s->pout, s->argumentSpool,
652                               s->args - s->argumentSpool);
653         s->typeSpoolLength = 0;
654         s->typeSpool = 0;
655         s->args = s->argumentSpool;
656     }
657 }
658 
reconstruct_and_encode_image(RoqEncContext * enc,int w,int h,int numBlocks)659 static void reconstruct_and_encode_image(RoqEncContext *enc,
660                                          int w, int h, int numBlocks)
661 {
662     RoqContext *const roq = &enc->common;
663     RoqTempData *const tempData = &enc->tmp_data;
664     int i, j, k;
665     int x, y;
666     int subX, subY;
667     int dist=0;
668 
669     roq_qcell *qcell;
670     CelEvaluation *eval;
671 
672     CodingSpool spool;
673 
674     spool.typeSpool=0;
675     spool.typeSpoolLength=0;
676     spool.args = spool.argumentSpool;
677     spool.pout = &enc->out_buf;
678 
679     if (tempData->used_option[RoQ_ID_CCC]%2)
680         tempData->mainChunkSize+=8; //FIXME
681 
682     /* Write the video chunk header */
683     bytestream_put_le16(&enc->out_buf, RoQ_QUAD_VQ);
684     bytestream_put_le32(&enc->out_buf, tempData->mainChunkSize/8);
685     bytestream_put_byte(&enc->out_buf, 0x0);
686     bytestream_put_byte(&enc->out_buf, 0x0);
687 
688     for (i=0; i<numBlocks; i++) {
689         eval = enc->cel_evals + i;
690 
691         x = eval->sourceX;
692         y = eval->sourceY;
693         dist += eval->eval_dist[eval->best_coding];
694 
695         switch (eval->best_coding) {
696         case RoQ_ID_MOT:
697             write_typecode(&spool, RoQ_ID_MOT);
698             break;
699 
700         case RoQ_ID_FCC:
701             bytestream_put_byte(&spool.args, motion_arg(eval->motion));
702 
703             write_typecode(&spool, RoQ_ID_FCC);
704             ff_apply_motion_8x8(roq, x, y,
705                                 eval->motion.d[0], eval->motion.d[1]);
706             break;
707 
708         case RoQ_ID_SLD:
709             bytestream_put_byte(&spool.args, tempData->i2f4[eval->cbEntry]);
710             write_typecode(&spool, RoQ_ID_SLD);
711 
712             qcell = roq->cb4x4 + eval->cbEntry;
713             ff_apply_vector_4x4(roq, x  , y  , roq->cb2x2 + qcell->idx[0]);
714             ff_apply_vector_4x4(roq, x+4, y  , roq->cb2x2 + qcell->idx[1]);
715             ff_apply_vector_4x4(roq, x  , y+4, roq->cb2x2 + qcell->idx[2]);
716             ff_apply_vector_4x4(roq, x+4, y+4, roq->cb2x2 + qcell->idx[3]);
717             break;
718 
719         case RoQ_ID_CCC:
720             write_typecode(&spool, RoQ_ID_CCC);
721 
722             for (j=0; j<4; j++) {
723                 subX = x + 4*(j&1);
724                 subY = y + 2*(j&2);
725 
726                 switch(eval->subCels[j].best_coding) {
727                 case RoQ_ID_MOT:
728                     break;
729 
730                 case RoQ_ID_FCC:
731                     bytestream_put_byte(&spool.args,
732                                         motion_arg(eval->subCels[j].motion));
733 
734                     ff_apply_motion_4x4(roq, subX, subY,
735                                         eval->subCels[j].motion.d[0],
736                                         eval->subCels[j].motion.d[1]);
737                     break;
738 
739                 case RoQ_ID_SLD:
740                     bytestream_put_byte(&spool.args,
741                                         tempData->i2f4[eval->subCels[j].cbEntry]);
742 
743                     qcell = roq->cb4x4 + eval->subCels[j].cbEntry;
744 
745                     ff_apply_vector_2x2(roq, subX  , subY  ,
746                                         roq->cb2x2 + qcell->idx[0]);
747                     ff_apply_vector_2x2(roq, subX+2, subY  ,
748                                         roq->cb2x2 + qcell->idx[1]);
749                     ff_apply_vector_2x2(roq, subX  , subY+2,
750                                         roq->cb2x2 + qcell->idx[2]);
751                     ff_apply_vector_2x2(roq, subX+2, subY+2,
752                                         roq->cb2x2 + qcell->idx[3]);
753                     break;
754 
755                 case RoQ_ID_CCC:
756                     for (k=0; k<4; k++) {
757                         int cb_idx = eval->subCels[j].subCels[k];
758                         bytestream_put_byte(&spool.args,
759                                             tempData->i2f2[cb_idx]);
760 
761                         ff_apply_vector_2x2(roq, subX + 2*(k&1), subY + (k&2),
762                                             roq->cb2x2 + cb_idx);
763                     }
764                     break;
765                 }
766                 write_typecode(&spool, eval->subCels[j].best_coding);
767             }
768             break;
769         }
770     }
771 
772     /* Flush the remainder of the argument/type spool */
773     while (spool.typeSpoolLength)
774         write_typecode(&spool, 0x0);
775 }
776 
777 
778 /**
779  * Create a single YUV cell from a 2x2 section of the image
780  */
frame_block_to_cell(int * block,uint8_t * const * data,int top,int left,const int * stride)781 static inline void frame_block_to_cell(int *block, uint8_t * const *data,
782                                        int top, int left, const int *stride)
783 {
784     int i, j, u=0, v=0;
785 
786     for (i=0; i<2; i++)
787         for (j=0; j<2; j++) {
788             int x = (top+i)*stride[0] + left + j;
789             *block++ = data[0][x];
790             x = (top+i)*stride[1] + left + j;
791             u       += data[1][x];
792             v       += data[2][x];
793         }
794 
795     *block++ = (u + 2) / 4 * CHROMA_BIAS;
796     *block++ = (v + 2) / 4 * CHROMA_BIAS;
797 }
798 
799 /**
800  * Create YUV clusters for the entire image
801  */
create_clusters(const AVFrame * frame,int w,int h,int * points)802 static void create_clusters(const AVFrame *frame, int w, int h, int *points)
803 {
804     int i, j, k, l;
805 
806     for (i=0; i<h; i+=4)
807         for (j=0; j<w; j+=4) {
808             for (k=0; k < 2; k++)
809                 for (l=0; l < 2; l++)
810                     frame_block_to_cell(points + (l + 2*k)*6, frame->data,
811                                         i+2*k, j+2*l, frame->linesize);
812             points += 24;
813         }
814 }
815 
generate_codebook(RoqEncContext * enc,int * points,int inputCount,roq_cell * results,int size,int cbsize)816 static int generate_codebook(RoqEncContext *enc,
817                              int *points, int inputCount, roq_cell *results,
818                              int size, int cbsize)
819 {
820     int i, j, k, ret = 0;
821     int c_size = size*size/4;
822     int *buf;
823     int *codebook = enc->tmp_codebook_buf;
824     int *closest_cb = enc->closest_cb;
825 
826     ret = avpriv_init_elbg(points, 6 * c_size, inputCount, codebook,
827                        cbsize, 1, closest_cb, &enc->randctx);
828     if (ret < 0)
829         return ret;
830     ret = avpriv_do_elbg(points, 6 * c_size, inputCount, codebook,
831                      cbsize, 1, closest_cb, &enc->randctx);
832     if (ret < 0)
833         return ret;
834 
835     buf = codebook;
836     for (i=0; i<cbsize; i++)
837         for (k=0; k<c_size; k++) {
838             for(j=0; j<4; j++)
839                 results->y[j] = *buf++;
840 
841             results->u =    (*buf++ + CHROMA_BIAS/2)/CHROMA_BIAS;
842             results->v =    (*buf++ + CHROMA_BIAS/2)/CHROMA_BIAS;
843             results++;
844         }
845     return 0;
846 }
847 
generate_new_codebooks(RoqEncContext * enc)848 static int generate_new_codebooks(RoqEncContext *enc)
849 {
850     int i, j, ret = 0;
851     RoqCodebooks *codebooks = &enc->tmp_data.codebooks;
852     RoqContext *const roq = &enc->common;
853     int max = roq->width * roq->height / 16;
854     uint8_t mb2[3*4];
855     int *points = enc->points;
856 
857     /* Subsample YUV data */
858     create_clusters(enc->frame_to_enc, roq->width, roq->height, points);
859 
860     codebooks->numCB4 = (enc->quake3_compat ? MAX_CBS_4x4-1 : MAX_CBS_4x4);
861 
862     /* Create 4x4 codebooks */
863     if ((ret = generate_codebook(enc, points, max, enc->results4,
864                                  4, codebooks->numCB4)) < 0)
865         return ret;
866 
867     /* Create 2x2 codebooks */
868     if ((ret = generate_codebook(enc, points, max * 4,
869                                  roq->cb2x2, 2, MAX_CBS_2x2)) < 0)
870         return ret;
871 
872     codebooks->numCB2 = MAX_CBS_2x2;
873 
874     /* Unpack 2x2 codebook clusters */
875     for (i=0; i<codebooks->numCB2; i++)
876         unpack_roq_cell(roq->cb2x2 + i, codebooks->unpacked_cb2 + i*2*2*3);
877 
878     /* Index all 4x4 entries to the 2x2 entries, unpack, and enlarge */
879     for (i=0; i<codebooks->numCB4; i++) {
880         for (j=0; j<4; j++) {
881             unpack_roq_cell(&enc->results4[4*i + j], mb2);
882             index_mb(mb2, codebooks->unpacked_cb2, codebooks->numCB2,
883                      &roq->cb4x4[i].idx[j], 2);
884         }
885         unpack_roq_qcell(codebooks->unpacked_cb2, roq->cb4x4 + i,
886                          codebooks->unpacked_cb4 + i*4*4*3);
887         enlarge_roq_mb4(codebooks->unpacked_cb4 + i*4*4*3,
888                         codebooks->unpacked_cb4_enlarged + i*8*8*3);
889     }
890 
891     return 0;
892 }
893 
roq_encode_video(RoqEncContext * enc)894 static int roq_encode_video(RoqEncContext *enc)
895 {
896     RoqTempData *const tempData = &enc->tmp_data;
897     RoqContext *const roq = &enc->common;
898     int ret;
899 
900     memset(tempData, 0, sizeof(*tempData));
901 
902     ret = generate_new_codebooks(enc);
903     if (ret < 0)
904         return ret;
905 
906     if (enc->framesSinceKeyframe >= 1) {
907         motion_search(enc, 8);
908         motion_search(enc, 4);
909     }
910 
911  retry_encode:
912     for (int i = 0; i < roq->width * roq->height / 64; i++)
913         gather_data_for_cel(enc->cel_evals + i, enc);
914 
915     /* Quake 3 can't handle chunks bigger than 65535 bytes */
916     if (tempData->mainChunkSize/8 > 65535 && enc->quake3_compat) {
917         if (enc->lambda > 100000) {
918             av_log(roq->avctx, AV_LOG_ERROR, "Cannot encode video in Quake compatible form\n");
919             return AVERROR(EINVAL);
920         }
921         av_log(roq->avctx, AV_LOG_ERROR,
922                "Warning, generated a frame too big for Quake (%d > 65535), "
923                "now switching to a bigger qscale value.\n",
924                tempData->mainChunkSize/8);
925         enc->lambda *= 1.5;
926         tempData->mainChunkSize = 0;
927         memset(tempData->used_option, 0, sizeof(tempData->used_option));
928         memset(tempData->codebooks.usedCB4, 0,
929                sizeof(tempData->codebooks.usedCB4));
930         memset(tempData->codebooks.usedCB2, 0,
931                sizeof(tempData->codebooks.usedCB2));
932 
933         goto retry_encode;
934     }
935 
936     remap_codebooks(enc);
937 
938     write_codebooks(enc);
939 
940     reconstruct_and_encode_image(enc, roq->width, roq->height,
941                                  roq->width * roq->height / 64);
942 
943     /* Rotate frame history */
944     FFSWAP(AVFrame *,    roq->current_frame,   roq->last_frame);
945     FFSWAP(motion_vect *, enc->last_motion4, enc->this_motion4);
946     FFSWAP(motion_vect *, enc->last_motion8, enc->this_motion8);
947 
948     enc->framesSinceKeyframe++;
949 
950     return 0;
951 }
952 
roq_encode_end(AVCodecContext * avctx)953 static av_cold int roq_encode_end(AVCodecContext *avctx)
954 {
955     RoqEncContext *const enc = avctx->priv_data;
956 
957     av_frame_free(&enc->common.current_frame);
958     av_frame_free(&enc->common.last_frame);
959 
960     av_freep(&enc->cel_evals);
961     av_freep(&enc->closest_cb);
962     av_freep(&enc->this_motion4);
963     av_freep(&enc->last_motion4);
964     av_freep(&enc->this_motion8);
965     av_freep(&enc->last_motion8);
966 
967     return 0;
968 }
969 
roq_encode_init(AVCodecContext * avctx)970 static av_cold int roq_encode_init(AVCodecContext *avctx)
971 {
972     RoqEncContext *const enc = avctx->priv_data;
973     RoqContext    *const roq = &enc->common;
974 
975     av_lfg_init(&enc->randctx, 1);
976 
977     roq->avctx = avctx;
978 
979     enc->framesSinceKeyframe = 0;
980     if ((avctx->width & 0xf) || (avctx->height & 0xf)) {
981         av_log(avctx, AV_LOG_ERROR, "Dimensions must be divisible by 16\n");
982         return AVERROR(EINVAL);
983     }
984 
985     if (avctx->width > 65535 || avctx->height > 65535) {
986         av_log(avctx, AV_LOG_ERROR, "Dimensions are max %d\n", enc->quake3_compat ? 32768 : 65535);
987         return AVERROR(EINVAL);
988     }
989 
990     if (((avctx->width)&(avctx->width-1))||((avctx->height)&(avctx->height-1)))
991         av_log(avctx, AV_LOG_ERROR, "Warning: dimensions not power of two, this is not supported by quake\n");
992 
993     roq->width  = avctx->width;
994     roq->height = avctx->height;
995 
996     enc->framesSinceKeyframe = 0;
997     enc->first_frame = 1;
998 
999     roq->last_frame    = av_frame_alloc();
1000     roq->current_frame = av_frame_alloc();
1001     if (!roq->last_frame || !roq->current_frame)
1002         return AVERROR(ENOMEM);
1003 
1004     enc->this_motion4 =
1005         av_mallocz_array(roq->width * roq->height / 16, sizeof(motion_vect));
1006 
1007     enc->last_motion4 =
1008         av_malloc_array (roq->width * roq->height / 16, sizeof(motion_vect));
1009 
1010     enc->this_motion8 =
1011         av_mallocz_array(roq->width * roq->height / 64, sizeof(motion_vect));
1012 
1013     enc->last_motion8 =
1014         av_malloc_array (roq->width * roq->height / 64, sizeof(motion_vect));
1015 
1016     /* 4x4 codebook needs 6 * 4 * 4 / 4 * width * height / 16 * sizeof(int);
1017      * and so does the points buffer. */
1018     enc->closest_cb   =
1019         av_malloc_array(roq->width * roq->height, 3 * sizeof(int));
1020 
1021     if (!enc->this_motion4 || !enc->last_motion4 ||
1022         !enc->this_motion8 || !enc->last_motion8 || !enc->closest_cb)
1023         return AVERROR(ENOMEM);
1024 
1025     enc->points = enc->closest_cb + roq->width * roq->height * 3 / 2;
1026 
1027     return create_cel_evals(enc);
1028 }
1029 
roq_write_video_info_chunk(RoqEncContext * enc)1030 static void roq_write_video_info_chunk(RoqEncContext *enc)
1031 {
1032     /* ROQ info chunk */
1033     bytestream_put_le16(&enc->out_buf, RoQ_INFO);
1034 
1035     /* Size: 8 bytes */
1036     bytestream_put_le32(&enc->out_buf, 8);
1037 
1038     /* Unused argument */
1039     bytestream_put_byte(&enc->out_buf, 0x00);
1040     bytestream_put_byte(&enc->out_buf, 0x00);
1041 
1042     /* Width */
1043     bytestream_put_le16(&enc->out_buf, enc->common.width);
1044 
1045     /* Height */
1046     bytestream_put_le16(&enc->out_buf, enc->common.height);
1047 
1048     /* Unused in Quake 3, mimics the output of the real encoder */
1049     bytestream_put_byte(&enc->out_buf, 0x08);
1050     bytestream_put_byte(&enc->out_buf, 0x00);
1051     bytestream_put_byte(&enc->out_buf, 0x04);
1052     bytestream_put_byte(&enc->out_buf, 0x00);
1053 }
1054 
roq_encode_frame(AVCodecContext * avctx,AVPacket * pkt,const AVFrame * frame,int * got_packet)1055 static int roq_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1056                             const AVFrame *frame, int *got_packet)
1057 {
1058     RoqEncContext *const enc = avctx->priv_data;
1059     RoqContext    *const roq = &enc->common;
1060     int size, ret;
1061 
1062     roq->avctx = avctx;
1063 
1064     enc->frame_to_enc = frame;
1065 
1066     if (frame->quality)
1067         enc->lambda = frame->quality - 1;
1068     else
1069         enc->lambda = 2*ROQ_LAMBDA_SCALE;
1070 
1071     /* 138 bits max per 8x8 block +
1072      *     256 codebooks*(6 bytes 2x2 + 4 bytes 4x4) + 8 bytes frame header */
1073     size = ((roq->width * roq->height / 64) * 138 + 7) / 8 + 256 * (6 + 4) + 8;
1074     if ((ret = ff_alloc_packet2(avctx, pkt, size, 0)) < 0)
1075         return ret;
1076     enc->out_buf = pkt->data;
1077 
1078     /* Check for I-frame */
1079     if (enc->framesSinceKeyframe == avctx->gop_size)
1080         enc->framesSinceKeyframe = 0;
1081 
1082     if (enc->first_frame) {
1083         /* Alloc memory for the reconstruction data (we must know the stride
1084          for that) */
1085         if ((ret = ff_get_buffer(avctx, roq->current_frame, 0)) < 0 ||
1086             (ret = ff_get_buffer(avctx, roq->last_frame,    0)) < 0)
1087             return ret;
1088 
1089         /* Before the first video frame, write a "video info" chunk */
1090         roq_write_video_info_chunk(enc);
1091 
1092         enc->first_frame = 0;
1093     }
1094 
1095     /* Encode the actual frame */
1096     ret = roq_encode_video(enc);
1097     if (ret < 0)
1098         return ret;
1099 
1100     pkt->size   = enc->out_buf - pkt->data;
1101     if (enc->framesSinceKeyframe == 1)
1102         pkt->flags |= AV_PKT_FLAG_KEY;
1103     *got_packet = 1;
1104 
1105     return 0;
1106 }
1107 
1108 #define OFFSET(x) offsetof(RoqEncContext, x)
1109 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1110 static const AVOption options[] = {
1111     { "quake3_compat", "Whether to respect known limitations in Quake 3 decoder", OFFSET(quake3_compat), AV_OPT_TYPE_BOOL, { .i64 = 1 }, 0, 1, VE },
1112     { NULL },
1113 };
1114 
1115 static const AVClass roq_class = {
1116     .class_name = "RoQ",
1117     .item_name  = av_default_item_name,
1118     .option     = options,
1119     .version    = LIBAVUTIL_VERSION_INT,
1120 };
1121 
1122 AVCodec ff_roq_encoder = {
1123     .name                 = "roqvideo",
1124     .long_name            = NULL_IF_CONFIG_SMALL("id RoQ video"),
1125     .type                 = AVMEDIA_TYPE_VIDEO,
1126     .id                   = AV_CODEC_ID_ROQ,
1127     .priv_data_size       = sizeof(RoqEncContext),
1128     .init                 = roq_encode_init,
1129     .encode2              = roq_encode_frame,
1130     .close                = roq_encode_end,
1131     .pix_fmts             = (const enum AVPixelFormat[]){ AV_PIX_FMT_YUVJ444P,
1132                                                         AV_PIX_FMT_NONE },
1133     .priv_class     = &roq_class,
1134     .caps_internal        = FF_CODEC_CAP_INIT_CLEANUP,
1135 };
1136