Commit | Line | Data |
---|---|---|
03e3e765 VS |
1 | /* |
2 | * RoQ Video Encoder. | |
3 | * | |
4 | * Copyright (C) 2007 Vitor <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 roqvideoenc.c | |
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 | #include <unistd.h> | |
59 | ||
60 | #include "roqvideo.h" | |
61 | #include "bytestream.h" | |
62 | #include "elbg.h" | |
63 | ||
64 | #define CHROMA_BIAS 1 | |
65 | ||
66 | /** | |
67 | * Maximum number of generated 4x4 codebooks. Can't be 256 to workaround a | |
68 | * Quake 3 bug. | |
69 | */ | |
70 | #define MAX_CBS_4x4 255 | |
71 | ||
72 | #define MAX_CBS_2x2 256 ///< Maximum number of 2x2 codebooks. | |
73 | ||
74 | /* The cast is useful when multiplying it by INT_MAX */ | |
75 | #define ROQ_LAMBDA_SCALE ((uint64_t) FF_LAMBDA_SCALE) | |
76 | ||
77 | /* Macroblock support functions */ | |
78 | static void unpack_roq_cell(roq_cell *cell, uint8_t u[4*3]) | |
79 | { | |
80 | memcpy(u , cell->y, 4); | |
81 | memset(u+4, cell->u, 4); | |
82 | memset(u+8, cell->v, 4); | |
83 | } | |
84 | ||
85 | static void unpack_roq_qcell(uint8_t cb2[], roq_qcell *qcell, uint8_t u[4*4*3]) | |
86 | { | |
87 | int i,cp; | |
88 | static const int offsets[4] = {0, 2, 8, 10}; | |
89 | ||
90 | for (cp=0; cp<3; cp++) | |
91 | for (i=0; i<4; i++) { | |
92 | u[4*4*cp + offsets[i] ] = cb2[qcell->idx[i]*2*2*3 + 4*cp ]; | |
93 | u[4*4*cp + offsets[i]+1] = cb2[qcell->idx[i]*2*2*3 + 4*cp+1]; | |
94 | u[4*4*cp + offsets[i]+4] = cb2[qcell->idx[i]*2*2*3 + 4*cp+2]; | |
95 | u[4*4*cp + offsets[i]+5] = cb2[qcell->idx[i]*2*2*3 + 4*cp+3]; | |
96 | } | |
97 | } | |
98 | ||
99 | ||
100 | static void enlarge_roq_mb4(uint8_t base[3*16], uint8_t u[3*64]) | |
101 | { | |
102 | int x,y,cp; | |
103 | ||
104 | for(cp=0; cp<3; cp++) | |
105 | for(y=0; y<8; y++) | |
106 | for(x=0; x<8; x++) | |
107 | *u++ = base[(y/2)*4 + (x/2) + 16*cp]; | |
108 | } | |
109 | ||
110 | static inline int square(int x) | |
111 | { | |
112 | return x*x; | |
113 | } | |
114 | ||
115 | static inline int eval_sse(uint8_t *a, uint8_t *b, int count) | |
116 | { | |
117 | int diff=0; | |
118 | ||
119 | while(count--) | |
120 | diff += square(*b++ - *a++); | |
121 | ||
122 | return diff; | |
123 | } | |
124 | ||
125 | // FIXME Could use DSPContext.sse, but it is not so speed critical (used | |
126 | // just for motion estimation). | |
127 | static int block_sse(uint8_t **buf1, uint8_t **buf2, int x1, int y1, int x2, | |
128 | int y2, int stride, int size) | |
129 | { | |
130 | int i, k; | |
131 | int sse=0; | |
132 | ||
133 | for (k=0; k<3; k++) { | |
134 | int bias = (k ? CHROMA_BIAS : 4); | |
135 | for (i=0; i<size; i++) | |
136 | sse += bias*eval_sse(buf1[k] + (y1+i)*stride + x1, | |
137 | buf2[k] + (y2+i)*stride + x2, size); | |
138 | } | |
139 | ||
140 | return sse; | |
141 | } | |
142 | ||
143 | static int eval_motion_dist(RoqContext *enc, int x, int y, motion_vect vect, | |
144 | int size) | |
145 | { | |
146 | int mx=vect.d[0]; | |
147 | int my=vect.d[1]; | |
148 | ||
149 | if (mx < -7 || mx > 7) | |
150 | return INT_MAX; | |
151 | ||
152 | if (my < -7 || my > 7) | |
153 | return INT_MAX; | |
154 | ||
155 | mx += x; | |
156 | my += y; | |
157 | ||
158 | if ((unsigned) mx > enc->width-size || (unsigned) my > enc->height-size) | |
159 | return INT_MAX; | |
160 | ||
161 | return block_sse(enc->frame_to_enc->data, enc->last_frame->data, x, y, | |
162 | mx, my, enc->y_stride, size); | |
163 | } | |
164 | ||
165 | /** | |
166 | * Returns distortion between two macroblocks | |
167 | */ | |
168 | static inline int squared_diff_macroblock(uint8_t a[], uint8_t b[], int size) | |
169 | { | |
170 | int cp, sdiff=0; | |
171 | ||
172 | for(cp=0;cp<3;cp++) { | |
173 | int bias = (cp ? CHROMA_BIAS : 4); | |
174 | sdiff += bias*eval_sse(a, b, size*size); | |
175 | a += size*size; | |
176 | b += size*size; | |
177 | } | |
178 | ||
179 | return sdiff; | |
180 | } | |
181 | ||
182 | typedef struct | |
183 | { | |
184 | int eval_dist[4]; | |
185 | int best_bit_use; | |
186 | int best_coding; | |
187 | ||
188 | int subCels[4]; | |
189 | motion_vect motion; | |
190 | int cbEntry; | |
191 | } subcel_evaluation_t; | |
192 | ||
193 | typedef struct | |
194 | { | |
195 | int eval_dist[4]; | |
196 | int best_coding; | |
197 | ||
198 | subcel_evaluation_t subCels[4]; | |
199 | ||
200 | motion_vect motion; | |
201 | int cbEntry; | |
202 | ||
203 | int sourceX, sourceY; | |
204 | } cel_evaluation_t; | |
205 | ||
206 | typedef struct | |
207 | { | |
208 | int numCB4; | |
209 | int numCB2; | |
210 | int usedCB2[MAX_CBS_2x2]; | |
211 | int usedCB4[MAX_CBS_4x4]; | |
212 | uint8_t unpacked_cb2[MAX_CBS_2x2*2*2*3]; | |
213 | uint8_t unpacked_cb4[MAX_CBS_4x4*4*4*3]; | |
214 | uint8_t unpacked_cb4_enlarged[MAX_CBS_4x4*8*8*3]; | |
215 | } roq_codebooks_t; | |
216 | ||
217 | /** | |
218 | * Temporary vars | |
219 | */ | |
220 | typedef struct | |
221 | { | |
222 | cel_evaluation_t *cel_evals; | |
223 | ||
224 | int f2i4[MAX_CBS_4x4]; | |
225 | int i2f4[MAX_CBS_4x4]; | |
226 | int f2i2[MAX_CBS_2x2]; | |
227 | int i2f2[MAX_CBS_2x2]; | |
228 | ||
229 | int mainChunkSize; | |
230 | ||
231 | int numCB4; | |
232 | int numCB2; | |
233 | ||
234 | roq_codebooks_t codebooks; | |
235 | ||
236 | int *closest_cb2; | |
237 | int used_option[4]; | |
238 | } roq_tempdata_t; | |
239 | ||
240 | /** | |
241 | * Initializes cel evaluators and sets their source coordinates | |
242 | */ | |
243 | static void create_cel_evals(RoqContext *enc, roq_tempdata_t *tempData) | |
244 | { | |
245 | int n=0, x, y, i; | |
246 | ||
247 | tempData->cel_evals = av_malloc(enc->width*enc->height/64 * sizeof(cel_evaluation_t)); | |
248 | ||
249 | /* Map to the ROQ quadtree order */ | |
250 | for (y=0; y<enc->height; y+=16) | |
251 | for (x=0; x<enc->width; x+=16) | |
252 | for(i=0; i<4; i++) { | |
253 | tempData->cel_evals[n ].sourceX = x + (i&1)*8; | |
254 | tempData->cel_evals[n++].sourceY = y + (i&2)*4; | |
255 | } | |
256 | } | |
257 | ||
258 | /** | |
259 | * Get macroblocks from parts of the image | |
260 | */ | |
261 | static void get_frame_mb(AVFrame *frame, int x, int y, uint8_t mb[], int dim) | |
262 | { | |
263 | int i, j, cp; | |
264 | int stride = frame->linesize[0]; | |
265 | ||
266 | for (cp=0; cp<3; cp++) | |
267 | for (i=0; i<dim; i++) | |
268 | for (j=0; j<dim; j++) | |
269 | *mb++ = frame->data[cp][(y+i)*stride + x + j]; | |
270 | } | |
271 | ||
272 | /** | |
273 | * Find the codebook with the lowest distortion from an image | |
274 | */ | |
275 | static int index_mb(uint8_t cluster[], uint8_t cb[], int numCB, | |
276 | int *outIndex, int dim) | |
277 | { | |
278 | int i, lDiff = INT_MAX, pick=0; | |
279 | ||
280 | /* Diff against the others */ | |
281 | for (i=0; i<numCB; i++) { | |
282 | int diff = squared_diff_macroblock(cluster, cb + i*dim*dim*3, dim); | |
283 | if (diff < lDiff) { | |
284 | lDiff = diff; | |
285 | pick = i; | |
286 | } | |
287 | } | |
288 | ||
289 | *outIndex = pick; | |
290 | return lDiff; | |
291 | } | |
292 | ||
293 | #define EVAL_MOTION(MOTION) \ | |
294 | do { \ | |
295 | diff = eval_motion_dist(enc, j, i, MOTION, blocksize); \ | |
296 | \ | |
297 | if (diff < lowestdiff) { \ | |
298 | lowestdiff = diff; \ | |
299 | bestpick = MOTION; \ | |
300 | } \ | |
301 | } while(0) | |
302 | ||
303 | static void motion_search(RoqContext *enc, int blocksize) | |
304 | { | |
305 | static const motion_vect offsets[8] = { | |
306 | {{ 0,-1}}, | |
307 | {{ 0, 1}}, | |
308 | {{-1, 0}}, | |
309 | {{ 1, 0}}, | |
310 | {{-1, 1}}, | |
311 | {{ 1,-1}}, | |
312 | {{-1,-1}}, | |
313 | {{ 1, 1}}, | |
314 | }; | |
315 | ||
316 | int diff, lowestdiff, oldbest; | |
317 | int off[3]; | |
318 | motion_vect bestpick = {{0,0}}; | |
319 | int i, j, k, offset; | |
320 | ||
321 | motion_vect *last_motion; | |
322 | motion_vect *this_motion; | |
323 | motion_vect vect, vect2; | |
324 | ||
325 | int max=(enc->width/blocksize)*enc->height/blocksize; | |
326 | ||
327 | if (blocksize == 4) { | |
328 | last_motion = enc->last_motion4; | |
329 | this_motion = enc->this_motion4; | |
330 | } else { | |
331 | last_motion = enc->last_motion8; | |
332 | this_motion = enc->this_motion8; | |
333 | } | |
334 | ||
335 | for (i=0; i<enc->height; i+=blocksize) | |
336 | for (j=0; j<enc->width; j+=blocksize) { | |
337 | lowestdiff = eval_motion_dist(enc, j, i, (motion_vect) {{0,0}}, | |
338 | blocksize); | |
339 | bestpick.d[0] = 0; | |
340 | bestpick.d[1] = 0; | |
341 | ||
342 | if (blocksize == 4) | |
343 | EVAL_MOTION(enc->this_motion8[(i/8)*(enc->width/8) + j/8]); | |
344 | ||
345 | offset = (i/blocksize)*enc->width/blocksize + j/blocksize; | |
346 | if (offset < max && offset >= 0) | |
347 | EVAL_MOTION(last_motion[offset]); | |
348 | ||
349 | offset++; | |
350 | if (offset < max && offset >= 0) | |
351 | EVAL_MOTION(last_motion[offset]); | |
352 | ||
353 | offset = (i/blocksize + 1)*enc->width/blocksize + j/blocksize; | |
354 | if (offset < max && offset >= 0) | |
355 | EVAL_MOTION(last_motion[offset]); | |
356 | ||
357 | off[0]= (i/blocksize)*enc->width/blocksize + j/blocksize - 1; | |
358 | off[1]= off[0] - enc->width/blocksize + 1; | |
359 | off[2]= off[1] + 1; | |
360 | ||
361 | if (i) { | |
362 | ||
363 | for(k=0; k<2; k++) | |
364 | vect.d[k]= mid_pred(this_motion[off[0]].d[k], | |
365 | this_motion[off[1]].d[k], | |
366 | this_motion[off[2]].d[k]); | |
367 | ||
368 | EVAL_MOTION(vect); | |
369 | for(k=0; k<3; k++) | |
370 | EVAL_MOTION(this_motion[off[k]]); | |
371 | } else if(j) | |
372 | EVAL_MOTION(this_motion[off[0]]); | |
373 | ||
374 | vect = bestpick; | |
375 | ||
376 | oldbest = -1; | |
377 | while (oldbest != lowestdiff) { | |
378 | oldbest = lowestdiff; | |
379 | for (k=0; k<8; k++) { | |
380 | vect2 = vect; | |
381 | vect2.d[0] += offsets[k].d[0]; | |
382 | vect2.d[1] += offsets[k].d[1]; | |
383 | EVAL_MOTION(vect2); | |
384 | } | |
385 | vect = bestpick; | |
386 | } | |
387 | offset = (i/blocksize)*enc->width/blocksize + j/blocksize; | |
388 | this_motion[offset] = bestpick; | |
389 | } | |
390 | } | |
391 | ||
392 | /** | |
393 | * Gets distortion for all options available to a subcel | |
394 | */ | |
395 | static void gather_data_for_subcel(subcel_evaluation_t *subcel, int x, | |
396 | int y, RoqContext *enc, roq_tempdata_t *tempData) | |
397 | { | |
398 | uint8_t mb4[4*4*3]; | |
399 | uint8_t mb2[2*2*3]; | |
400 | int cluster_index; | |
401 | int i, best_dist; | |
402 | ||
403 | static const int bitsUsed[4] = {2, 10, 10, 34}; | |
404 | ||
405 | if (enc->framesSinceKeyframe >= 1) { | |
406 | subcel->motion = enc->this_motion4[y*enc->width/16 + x/4]; | |
407 | ||
408 | subcel->eval_dist[RoQ_ID_FCC] = | |
409 | eval_motion_dist(enc, x, y, | |
410 | enc->this_motion4[y*enc->width/16 + x/4], 4); | |
411 | } else | |
412 | subcel->eval_dist[RoQ_ID_FCC] = INT_MAX; | |
413 | ||
414 | if (enc->framesSinceKeyframe >= 2) | |
415 | subcel->eval_dist[RoQ_ID_MOT] = block_sse(enc->frame_to_enc->data, | |
416 | enc->current_frame->data, x, | |
417 | y, x, y, enc->y_stride, 4); | |
418 | else | |
419 | subcel->eval_dist[RoQ_ID_MOT] = INT_MAX; | |
420 | ||
421 | cluster_index = y*enc->width/16 + x/4; | |
422 | ||
423 | get_frame_mb(enc->frame_to_enc, x, y, mb4, 4); | |
424 | ||
425 | subcel->eval_dist[RoQ_ID_SLD] = index_mb(mb4, | |
426 | tempData->codebooks.unpacked_cb4, | |
427 | tempData->codebooks.numCB4, | |
428 | &subcel->cbEntry, 4); | |
429 | ||
430 | subcel->eval_dist[RoQ_ID_CCC] = 0; | |
431 | ||
432 | for(i=0;i<4;i++) { | |
433 | subcel->subCels[i] = tempData->closest_cb2[cluster_index*4+i]; | |
434 | ||
435 | get_frame_mb(enc->frame_to_enc, x+2*(i&1), | |
436 | y+(i&2), mb2, 2); | |
437 | ||
438 | subcel->eval_dist[RoQ_ID_CCC] += | |
439 | squared_diff_macroblock(tempData->codebooks.unpacked_cb2 + subcel->subCels[i]*2*2*3, mb2, 2); | |
440 | } | |
441 | ||
442 | best_dist = INT_MAX; | |
443 | for (i=0; i<4; i++) | |
444 | if (ROQ_LAMBDA_SCALE*subcel->eval_dist[i] + enc->lambda*bitsUsed[i] < | |
445 | best_dist) { | |
446 | subcel->best_coding = i; | |
447 | subcel->best_bit_use = bitsUsed[i]; | |
448 | best_dist = ROQ_LAMBDA_SCALE*subcel->eval_dist[i] + | |
449 | enc->lambda*bitsUsed[i]; | |
450 | } | |
451 | } | |
452 | ||
453 | /** | |
454 | * Gets distortion for all options available to a cel | |
455 | */ | |
456 | static void gather_data_for_cel(cel_evaluation_t *cel, RoqContext *enc, | |
457 | roq_tempdata_t *tempData) | |
458 | { | |
459 | uint8_t mb8[8*8*3]; | |
460 | int index = cel->sourceY*enc->width/64 + cel->sourceX/8; | |
461 | int i, j, best_dist, divide_bit_use; | |
462 | ||
463 | int bitsUsed[4] = {2, 10, 10, 0}; | |
464 | ||
465 | if (enc->framesSinceKeyframe >= 1) { | |
466 | cel->motion = enc->this_motion8[index]; | |
467 | ||
468 | cel->eval_dist[RoQ_ID_FCC] = | |
469 | eval_motion_dist(enc, cel->sourceX, cel->sourceY, | |
470 | enc->this_motion8[index], 8); | |
471 | } else | |
472 | cel->eval_dist[RoQ_ID_FCC] = INT_MAX; | |
473 | ||
474 | if (enc->framesSinceKeyframe >= 2) | |
475 | cel->eval_dist[RoQ_ID_MOT] = block_sse(enc->frame_to_enc->data, | |
476 | enc->current_frame->data, | |
477 | cel->sourceX, cel->sourceY, | |
478 | cel->sourceX, cel->sourceY, | |
479 | enc->y_stride, 8); | |
480 | else | |
481 | cel->eval_dist[RoQ_ID_MOT] = INT_MAX; | |
482 | ||
483 | get_frame_mb(enc->frame_to_enc, cel->sourceX, cel->sourceY, mb8, 8); | |
484 | ||
485 | cel->eval_dist[RoQ_ID_SLD] = | |
486 | index_mb(mb8, tempData->codebooks.unpacked_cb4_enlarged, | |
487 | tempData->codebooks.numCB4, &cel->cbEntry, 8); | |
488 | ||
489 | gather_data_for_subcel(cel->subCels + 0, cel->sourceX+0, cel->sourceY+0, enc, tempData); | |
490 | gather_data_for_subcel(cel->subCels + 1, cel->sourceX+4, cel->sourceY+0, enc, tempData); | |
491 | gather_data_for_subcel(cel->subCels + 2, cel->sourceX+0, cel->sourceY+4, enc, tempData); | |
492 | gather_data_for_subcel(cel->subCels + 3, cel->sourceX+4, cel->sourceY+4, enc, tempData); | |
493 | ||
494 | cel->eval_dist[RoQ_ID_CCC] = 0; | |
495 | divide_bit_use = 0; | |
496 | for (i=0; i<4; i++) { | |
497 | cel->eval_dist[RoQ_ID_CCC] += | |
498 | cel->subCels[i].eval_dist[cel->subCels[i].best_coding]; | |
499 | divide_bit_use += cel->subCels[i].best_bit_use; | |
500 | } | |
501 | ||
502 | best_dist = INT_MAX; | |
503 | bitsUsed[3] = 2 + divide_bit_use; | |
504 | ||
505 | for (i=0; i<4; i++) | |
506 | if (ROQ_LAMBDA_SCALE*cel->eval_dist[i] + enc->lambda*bitsUsed[i] < | |
507 | best_dist) { | |
508 | cel->best_coding = i; | |
509 | best_dist = ROQ_LAMBDA_SCALE*cel->eval_dist[i] + | |
510 | enc->lambda*bitsUsed[i]; | |
511 | } | |
512 | ||
513 | tempData->used_option[cel->best_coding]++; | |
514 | tempData->mainChunkSize += bitsUsed[cel->best_coding]; | |
515 | ||
516 | if (cel->best_coding == RoQ_ID_SLD) | |
517 | tempData->codebooks.usedCB4[cel->cbEntry]++; | |
518 | ||
519 | if (cel->best_coding == RoQ_ID_CCC) | |
520 | for (i=0; i<4; i++) { | |
521 | if (cel->subCels[i].best_coding == RoQ_ID_SLD) | |
522 | tempData->codebooks.usedCB4[cel->subCels[i].cbEntry]++; | |
523 | else if (cel->subCels[i].best_coding == RoQ_ID_CCC) | |
524 | for (j=0; j<4; j++) | |
525 | tempData->codebooks.usedCB2[cel->subCels[i].subCels[j]]++; | |
526 | } | |
527 | } | |
528 | ||
529 | static void remap_codebooks(RoqContext *enc, roq_tempdata_t *tempData) | |
530 | { | |
531 | int i, j, idx=0; | |
532 | ||
533 | /* Make remaps for the final codebook usage */ | |
534 | for (i=0; i<MAX_CBS_4x4; i++) { | |
535 | if (tempData->codebooks.usedCB4[i]) { | |
536 | tempData->i2f4[i] = idx; | |
537 | tempData->f2i4[idx] = i; | |
538 | for (j=0; j<4; j++) | |
539 | tempData->codebooks.usedCB2[enc->cb4x4[i].idx[j]]++; | |
540 | idx++; | |
541 | } | |
542 | } | |
543 | ||
544 | tempData->numCB4 = idx; | |
545 | ||
546 | idx = 0; | |
547 | for (i=0; i<MAX_CBS_2x2; i++) { | |
548 | if (tempData->codebooks.usedCB2[i]) { | |
549 | tempData->i2f2[i] = idx; | |
550 | tempData->f2i2[idx] = i; | |
551 | idx++; | |
552 | } | |
553 | } | |
554 | tempData->numCB2 = idx; | |
555 | ||
556 | } | |
557 | ||
558 | /** | |
559 | * Write codebook chunk | |
560 | */ | |
561 | static void write_codebooks(RoqContext *enc, roq_tempdata_t *tempData) | |
562 | { | |
563 | int i, j; | |
564 | uint8_t **outp= &enc->out_buf; | |
565 | ||
566 | if (tempData->numCB2) { | |
567 | bytestream_put_le16(outp, RoQ_QUAD_CODEBOOK); | |
568 | bytestream_put_le32(outp, tempData->numCB2*6 + tempData->numCB4*4); | |
569 | bytestream_put_byte(outp, tempData->numCB4); | |
570 | bytestream_put_byte(outp, tempData->numCB2); | |
571 | ||
572 | for (i=0; i<tempData->numCB2; i++) { | |
573 | bytestream_put_buffer(outp, enc->cb2x2[tempData->f2i2[i]].y, 4); | |
574 | bytestream_put_byte(outp, enc->cb2x2[tempData->f2i2[i]].u); | |
575 | bytestream_put_byte(outp, enc->cb2x2[tempData->f2i2[i]].v); | |
576 | } | |
577 | ||
578 | for (i=0; i<tempData->numCB4; i++) | |
579 | for (j=0; j<4; j++) | |
580 | bytestream_put_byte(outp, tempData->i2f2[enc->cb4x4[tempData->f2i4[i]].idx[j]]); | |
581 | ||
582 | } | |
583 | } | |
584 | ||
585 | static inline uint8_t motion_arg(motion_vect mot) | |
586 | { | |
587 | uint8_t ax = 8 - ((uint8_t) mot.d[0]); | |
588 | uint8_t ay = 8 - ((uint8_t) mot.d[1]); | |
589 | return ((ax&15)<<4) | (ay&15); | |
590 | } | |
591 | ||
592 | typedef struct | |
593 | { | |
594 | int typeSpool; | |
595 | int typeSpoolLength; | |
596 | uint8_t argumentSpool[64]; | |
597 | uint8_t *args; | |
598 | uint8_t **pout; | |
599 | } CodingSpool; | |
600 | ||
601 | /* NOTE: Typecodes must be spooled AFTER arguments!! */ | |
602 | static void write_typecode(CodingSpool *s, uint8_t type) | |
603 | { | |
604 | s->typeSpool |= ((type) & 3) << (14 - s->typeSpoolLength); | |
605 | s->typeSpoolLength += 2; | |
606 | if (s->typeSpoolLength == 16) { | |
607 | bytestream_put_le16(s->pout, s->typeSpool); | |
608 | bytestream_put_buffer(s->pout, s->argumentSpool, | |
609 | s->args - s->argumentSpool); | |
610 | s->typeSpoolLength = 0; | |
611 | s->typeSpool = 0; | |
612 | s->args = s->argumentSpool; | |
613 | } | |
614 | } | |
615 | ||
616 | static void reconstruct_and_encode_image(RoqContext *enc, roq_tempdata_t *tempData, int w, int h, int numBlocks) | |
617 | { | |
618 | int i, j, k; | |
619 | int x, y; | |
620 | int subX, subY; | |
621 | int dist=0; | |
622 | ||
623 | roq_qcell *qcell; | |
624 | cel_evaluation_t *eval; | |
625 | ||
626 | CodingSpool spool; | |
627 | ||
628 | spool.typeSpool=0; | |
629 | spool.typeSpoolLength=0; | |
630 | spool.args = spool.argumentSpool; | |
631 | spool.pout = &enc->out_buf; | |
632 | ||
633 | if (tempData->used_option[RoQ_ID_CCC]%2) | |
634 | tempData->mainChunkSize+=8; //FIXME | |
635 | ||
636 | /* Write the video chunk header */ | |
637 | bytestream_put_le16(&enc->out_buf, RoQ_QUAD_VQ); | |
638 | bytestream_put_le32(&enc->out_buf, tempData->mainChunkSize/8); | |
639 | bytestream_put_byte(&enc->out_buf, 0x0); | |
640 | bytestream_put_byte(&enc->out_buf, 0x0); | |
641 | ||
642 | for (i=0; i<numBlocks; i++) { | |
643 | eval = tempData->cel_evals + i; | |
644 | ||
645 | x = eval->sourceX; | |
646 | y = eval->sourceY; | |
647 | dist += eval->eval_dist[eval->best_coding]; | |
648 | ||
649 | switch (eval->best_coding) { | |
650 | case RoQ_ID_MOT: | |
651 | write_typecode(&spool, RoQ_ID_MOT); | |
652 | break; | |
653 | ||
654 | case RoQ_ID_FCC: | |
655 | bytestream_put_byte(&spool.args, motion_arg(eval->motion)); | |
656 | ||
657 | write_typecode(&spool, RoQ_ID_FCC); | |
658 | ff_apply_motion_8x8(enc, x, y, | |
659 | eval->motion.d[0], eval->motion.d[1]); | |
660 | break; | |
661 | ||
662 | case RoQ_ID_SLD: | |
663 | bytestream_put_byte(&spool.args, tempData->i2f4[eval->cbEntry]); | |
664 | write_typecode(&spool, RoQ_ID_SLD); | |
665 | ||
666 | qcell = enc->cb4x4 + eval->cbEntry; | |
667 | ff_apply_vector_4x4(enc, x , y , enc->cb2x2 + qcell->idx[0]); | |
668 | ff_apply_vector_4x4(enc, x+4, y , enc->cb2x2 + qcell->idx[1]); | |
669 | ff_apply_vector_4x4(enc, x , y+4, enc->cb2x2 + qcell->idx[2]); | |
670 | ff_apply_vector_4x4(enc, x+4, y+4, enc->cb2x2 + qcell->idx[3]); | |
671 | break; | |
672 | ||
673 | case RoQ_ID_CCC: | |
674 | write_typecode(&spool, RoQ_ID_CCC); | |
675 | ||
676 | for (j=0; j<4; j++) { | |
677 | subX = x + 4*(j&1); | |
678 | subY = y + 2*(j&2); | |
679 | ||
680 | switch(eval->subCels[j].best_coding) { | |
681 | case RoQ_ID_MOT: | |
682 | break; | |
683 | ||
684 | case RoQ_ID_FCC: | |
685 | bytestream_put_byte(&spool.args, | |
686 | motion_arg(eval->subCels[j].motion)); | |
687 | ||
688 | ff_apply_motion_4x4(enc, subX, subY, | |
689 | eval->subCels[j].motion.d[0], | |
690 | eval->subCels[j].motion.d[1]); | |
691 | break; | |
692 | ||
693 | case RoQ_ID_SLD: | |
694 | bytestream_put_byte(&spool.args, | |
695 | tempData->i2f4[eval->subCels[j].cbEntry]); | |
696 | ||
697 | qcell = enc->cb4x4 + eval->subCels[j].cbEntry; | |
698 | ||
699 | ff_apply_vector_2x2(enc, subX , subY , | |
700 | enc->cb2x2 + qcell->idx[0]); | |
701 | ff_apply_vector_2x2(enc, subX+2, subY , | |
702 | enc->cb2x2 + qcell->idx[1]); | |
703 | ff_apply_vector_2x2(enc, subX , subY+2, | |
704 | enc->cb2x2 + qcell->idx[2]); | |
705 | ff_apply_vector_2x2(enc, subX+2, subY+2, | |
706 | enc->cb2x2 + qcell->idx[3]); | |
707 | break; | |
708 | ||
709 | case RoQ_ID_CCC: | |
710 | for (k=0; k<4; k++) { | |
711 | int cb_idx = eval->subCels[j].subCels[k]; | |
712 | bytestream_put_byte(&spool.args, | |
713 | tempData->i2f2[cb_idx]); | |
714 | ||
715 | ff_apply_vector_2x2(enc, subX + 2*(k&1), subY + (k&2), | |
716 | enc->cb2x2 + cb_idx); | |
717 | } | |
718 | break; | |
719 | } | |
720 | write_typecode(&spool, eval->subCels[j].best_coding); | |
721 | } | |
722 | break; | |
723 | } | |
724 | } | |
725 | ||
726 | /* Flush the remainder of the argument/type spool */ | |
727 | while (spool.typeSpoolLength) | |
728 | write_typecode(&spool, 0x0); | |
729 | ||
730 | #if 0 | |
731 | uint8_t *fdata[3] = {enc->frame_to_enc->data[0], | |
732 | enc->frame_to_enc->data[1], | |
733 | enc->frame_to_enc->data[2]}; | |
734 | uint8_t *cdata[3] = {enc->current_frame->data[0], | |
735 | enc->current_frame->data[1], | |
736 | enc->current_frame->data[2]}; | |
737 | av_log(enc->avctx, AV_LOG_ERROR, "Expected distortion: %i Actual: %i\n", | |
738 | dist, | |
739 | block_sse(fdata, cdata, 0, 0, 0, 0, | |
740 | enc->y_stride, enc->width)); //WARNING: Square dimensions implied... | |
741 | #endif | |
742 | } | |
743 | ||
744 | ||
745 | /** | |
746 | * Create a single YUV cell from a 2x2 section of the image | |
747 | */ | |
748 | static inline void frame_block_to_cell(uint8_t *block, uint8_t **data, | |
749 | int top, int left, int *stride) | |
750 | { | |
751 | int i, j, u=0, v=0; | |
752 | ||
753 | for (i=0; i<2; i++) | |
754 | for (j=0; j<2; j++) { | |
755 | int x = (top+i)*stride[0] + left + j; | |
756 | *block++ = data[0][x]; | |
757 | u += data[1][x]; | |
758 | v += data[2][x]; | |
759 | } | |
760 | ||
761 | *block++ = (u+2)/4; | |
762 | *block++ = (v+2)/4; | |
763 | } | |
764 | ||
765 | /** | |
766 | * Creates YUV clusters for the entire image | |
767 | */ | |
768 | static void create_clusters(AVFrame *frame, int w, int h, uint8_t *yuvClusters) | |
769 | { | |
770 | int i, j, k, l; | |
771 | ||
772 | for (i=0; i<h; i+=4) | |
773 | for (j=0; j<w; j+=4) { | |
774 | for (k=0; k < 2; k++) | |
775 | for (l=0; l < 2; l++) | |
776 | frame_block_to_cell(yuvClusters + (l + 2*k)*6, frame->data, | |
777 | i+2*k, j+2*l, frame->linesize); | |
778 | yuvClusters += 24; | |
779 | } | |
780 | } | |
781 | ||
782 | static void generate_codebook(RoqContext *enc, roq_tempdata_t *tempdata, | |
783 | int *points, int inputCount, roq_cell *results, | |
784 | int size, int cbsize) | |
785 | { | |
786 | int i, j, k; | |
787 | int c_size = size*size/4; | |
788 | int *buf = points; | |
789 | int *codebook = av_malloc(6*c_size*cbsize*sizeof(int)); | |
790 | int *closest_cb; | |
791 | ||
792 | if (size == 4) | |
793 | closest_cb = av_malloc(6*c_size*inputCount*sizeof(int)); | |
794 | else | |
795 | closest_cb = tempdata->closest_cb2; | |
796 | ||
797 | ff_init_elbg(points, 6*c_size, inputCount, codebook, cbsize, 1, closest_cb, &enc->randctx); | |
798 | ff_do_elbg(points, 6*c_size, inputCount, codebook, cbsize, 1, closest_cb, &enc->randctx); | |
799 | ||
800 | if (size == 4) | |
801 | av_free(closest_cb); | |
802 | ||
803 | buf = codebook; | |
804 | for (i=0; i<cbsize; i++) | |
805 | for (k=0; k<c_size; k++) { | |
806 | for(j=0; j<4; j++) | |
807 | results->y[j] = *buf++; | |
808 | ||
809 | results->u = (*buf++ + CHROMA_BIAS/2)/CHROMA_BIAS; | |
810 | results->v = (*buf++ + CHROMA_BIAS/2)/CHROMA_BIAS; | |
811 | results++; | |
812 | } | |
813 | ||
814 | av_free(codebook); | |
815 | } | |
816 | ||
817 | static void generate_new_codebooks(RoqContext *enc, roq_tempdata_t *tempData) | |
818 | { | |
819 | int i,j; | |
820 | roq_codebooks_t *codebooks = &tempData->codebooks; | |
821 | int max = enc->width*enc->height/16; | |
822 | uint8_t mb2[3*4]; | |
823 | roq_cell *results4 = av_malloc(sizeof(roq_cell)*MAX_CBS_4x4*4); | |
824 | uint8_t *yuvClusters=av_malloc(sizeof(int)*max*6*4); | |
825 | int *points = av_malloc(max*6*4*sizeof(int)); | |
826 | int bias; | |
827 | ||
828 | /* Subsample YUV data */ | |
829 | create_clusters(enc->frame_to_enc, enc->width, enc->height, yuvClusters); | |
830 | ||
831 | /* Cast to integer and apply chroma bias */ | |
832 | for (i=0; i<max*24; i++) { | |
833 | bias = ((i%6)<4) ? 1 : CHROMA_BIAS; | |
834 | points[i] = bias*yuvClusters[i]; | |
835 | } | |
836 | ||
837 | /* Create 4x4 codebooks */ | |
838 | generate_codebook(enc, tempData, points, max, results4, 4, MAX_CBS_4x4); | |
839 | ||
840 | codebooks->numCB4 = MAX_CBS_4x4; | |
841 | ||
842 | tempData->closest_cb2 = av_malloc(max*4*sizeof(int)); | |
843 | ||
844 | /* Create 2x2 codebooks */ | |
845 | generate_codebook(enc, tempData, points, max*4, enc->cb2x2, 2, MAX_CBS_2x2); | |
846 | ||
847 | codebooks->numCB2 = MAX_CBS_2x2; | |
848 | ||
849 | /* Unpack 2x2 codebook clusters */ | |
850 | for (i=0; i<codebooks->numCB2; i++) | |
851 | unpack_roq_cell(enc->cb2x2 + i, codebooks->unpacked_cb2 + i*2*2*3); | |
852 | ||
853 | /* Index all 4x4 entries to the 2x2 entries, unpack, and enlarge */ | |
854 | for (i=0; i<codebooks->numCB4; i++) { | |
855 | for (j=0; j<4; j++) { | |
856 | unpack_roq_cell(&results4[4*i + j], mb2); | |
857 | index_mb(mb2, codebooks->unpacked_cb2, codebooks->numCB2, | |
858 | &enc->cb4x4[i].idx[j], 2); | |
859 | } | |
860 | unpack_roq_qcell(codebooks->unpacked_cb2, enc->cb4x4 + i, | |
861 | codebooks->unpacked_cb4 + i*4*4*3); | |
862 | enlarge_roq_mb4(codebooks->unpacked_cb4 + i*4*4*3, | |
863 | codebooks->unpacked_cb4_enlarged + i*8*8*3); | |
864 | } | |
865 | ||
866 | av_free(yuvClusters); | |
867 | av_free(points); | |
868 | av_free(results4); | |
869 | } | |
870 | ||
871 | static void roq_encode_video(RoqContext *enc) | |
872 | { | |
873 | roq_tempdata_t tempData; | |
874 | int i; | |
875 | ||
876 | memset(&tempData, 0, sizeof(tempData)); | |
877 | ||
878 | create_cel_evals(enc, &tempData); | |
879 | ||
880 | generate_new_codebooks(enc, &tempData); | |
881 | ||
882 | if (enc->framesSinceKeyframe >= 1) { | |
883 | motion_search(enc, 8); | |
884 | motion_search(enc, 4); | |
885 | } | |
886 | ||
887 | retry_encode: | |
888 | for (i=0; i<enc->width*enc->height/64; i++) | |
889 | gather_data_for_cel(tempData.cel_evals + i, enc, &tempData); | |
890 | ||
891 | /* Quake 3 can't handle chunks bigger than 65536 bytes */ | |
892 | if (tempData.mainChunkSize/8 > 65536) { | |
893 | enc->lambda *= .8; | |
894 | goto retry_encode; | |
895 | } | |
896 | ||
897 | remap_codebooks(enc, &tempData); | |
898 | ||
899 | write_codebooks(enc, &tempData); | |
900 | ||
901 | reconstruct_and_encode_image(enc, &tempData, enc->width, enc->height, | |
902 | enc->width*enc->height/64); | |
903 | ||
904 | /* Rotate frame history */ | |
905 | FFSWAP(AVFrame *, enc->current_frame, enc->last_frame); | |
906 | FFSWAP(motion_vect *, enc->last_motion4, enc->this_motion4); | |
907 | FFSWAP(motion_vect *, enc->last_motion8, enc->this_motion8); | |
908 | ||
909 | av_free(tempData.cel_evals); | |
910 | av_free(tempData.closest_cb2); | |
911 | ||
912 | enc->framesSinceKeyframe++; | |
913 | } | |
914 | ||
915 | static int roq_encode_init(AVCodecContext *avctx) | |
916 | { | |
917 | RoqContext *enc = avctx->priv_data; | |
918 | ||
919 | av_init_random(1, &enc->randctx); | |
920 | ||
921 | enc->framesSinceKeyframe = 0; | |
922 | if ((avctx->width & 0xf) || (avctx->height & 0xf)) { | |
923 | av_log(avctx, AV_LOG_ERROR, "Dimensions must be divisible by 16\n"); | |
924 | return -1; | |
925 | } | |
926 | ||
927 | if (((avctx->width)&(avctx->width-1))||((avctx->height)&(avctx->height-1))) | |
928 | av_log(avctx, AV_LOG_ERROR, "Warning: dimensions not power of two\n"); | |
929 | ||
930 | if (avcodec_check_dimensions(avctx, avctx->width, avctx->height)) { | |
931 | av_log(avctx, AV_LOG_ERROR, "Invalid dimensions (%dx%d)\n", | |
932 | avctx->width, avctx->height); | |
933 | return -1; | |
934 | } | |
935 | ||
936 | enc->width = avctx->width; | |
937 | enc->height = avctx->height; | |
938 | ||
939 | enc->framesSinceKeyframe = 0; | |
940 | enc->first_frame = 1; | |
941 | ||
942 | enc->last_frame = &enc->frames[0]; | |
943 | enc->current_frame = &enc->frames[1]; | |
944 | ||
945 | enc->this_motion4 = | |
946 | av_mallocz((enc->width*enc->height/16)*sizeof(motion_vect)); | |
947 | ||
948 | enc->last_motion4 = | |
949 | av_malloc ((enc->width*enc->height/16)*sizeof(motion_vect)); | |
950 | ||
951 | enc->this_motion8 = | |
952 | av_mallocz((enc->width*enc->height/64)*sizeof(motion_vect)); | |
953 | ||
954 | enc->last_motion8 = | |
955 | av_malloc ((enc->width*enc->height/64)*sizeof(motion_vect)); | |
956 | ||
957 | return 0; | |
958 | } | |
959 | ||
960 | static void roq_write_video_info_chunk(RoqContext *enc) | |
961 | { | |
962 | /* ROQ info chunk */ | |
963 | bytestream_put_le16(&enc->out_buf, RoQ_INFO); | |
964 | ||
965 | /* Size: 8 bytes */ | |
966 | bytestream_put_le32(&enc->out_buf, 8); | |
967 | ||
968 | /* Unused argument */ | |
969 | bytestream_put_byte(&enc->out_buf, 0x00); | |
970 | bytestream_put_byte(&enc->out_buf, 0x00); | |
971 | ||
972 | /* Width */ | |
973 | bytestream_put_le16(&enc->out_buf, enc->width); | |
974 | ||
975 | /* Height */ | |
976 | bytestream_put_le16(&enc->out_buf, enc->height); | |
977 | ||
978 | /* Unused in Quake 3, mimics the output of the real encoder */ | |
979 | bytestream_put_byte(&enc->out_buf, 0x08); | |
980 | bytestream_put_byte(&enc->out_buf, 0x00); | |
981 | bytestream_put_byte(&enc->out_buf, 0x04); | |
982 | bytestream_put_byte(&enc->out_buf, 0x00); | |
983 | } | |
984 | ||
985 | static int roq_encode_frame(AVCodecContext *avctx, unsigned char *buf, int buf_size, void *data) | |
986 | { | |
987 | RoqContext *enc = avctx->priv_data; | |
988 | AVFrame *frame= data; | |
989 | uint8_t *buf_start = buf; | |
990 | ||
991 | enc->out_buf = buf; | |
992 | enc->y_stride = frame->linesize[0]; | |
993 | enc->c_stride = frame->linesize[1]; | |
994 | enc->avctx = avctx; | |
995 | ||
996 | enc->frame_to_enc = frame; | |
997 | ||
998 | if (frame->quality) | |
999 | enc->lambda = frame->quality - 1; | |
1000 | else | |
1001 | enc->lambda = 2*ROQ_LAMBDA_SCALE; | |
1002 | ||
1003 | /* 138 bits max per 8x8 block + | |
1004 | * 256 codebooks*(6 bytes 2x2 + 4 bytes 4x4) + 8 bytes frame header */ | |
1005 | if (((enc->width*enc->height/64)*138+7)/8 + 256*(6+4) + 8 > buf_size) { | |
1006 | av_log(avctx, AV_LOG_ERROR, " RoQ: Output buffer too small!\n"); | |
1007 | return -1; | |
1008 | } | |
1009 | ||
1010 | /* Check for I frame */ | |
1011 | if (enc->framesSinceKeyframe == avctx->gop_size) | |
1012 | enc->framesSinceKeyframe = 0; | |
1013 | ||
1014 | if (enc->first_frame) { | |
1015 | /* Alloc memory for the reconstruction data (we must know the stride | |
1016 | for that) */ | |
1017 | if (avctx->get_buffer(avctx, enc->current_frame) || | |
1018 | avctx->get_buffer(avctx, enc->last_frame)) { | |
1019 | av_log(avctx, AV_LOG_ERROR, " RoQ: get_buffer() failed\n"); | |
1020 | return -1; | |
1021 | } | |
1022 | ||
1023 | /* Before the first video frame, write a "video info" chunk */ | |
1024 | roq_write_video_info_chunk(enc); | |
1025 | ||
1026 | enc->first_frame = 0; | |
1027 | } | |
1028 | ||
1029 | /* Encode the actual frame */ | |
1030 | roq_encode_video(enc); | |
1031 | ||
1032 | return enc->out_buf - buf_start; | |
1033 | } | |
1034 | ||
1035 | static int roq_encode_end(AVCodecContext *avctx) | |
1036 | { | |
1037 | RoqContext *enc = avctx->priv_data; | |
1038 | ||
1039 | avctx->release_buffer(avctx, enc->last_frame); | |
1040 | avctx->release_buffer(avctx, enc->current_frame); | |
1041 | ||
1042 | av_free(enc->this_motion4); | |
1043 | av_free(enc->last_motion4); | |
1044 | av_free(enc->this_motion8); | |
1045 | av_free(enc->last_motion8); | |
1046 | ||
1047 | return 0; | |
1048 | } | |
1049 | ||
1050 | AVCodec roq_encoder = | |
1051 | { | |
1052 | "roqvideo", | |
1053 | CODEC_TYPE_VIDEO, | |
1054 | CODEC_ID_ROQ, | |
1055 | sizeof(RoqContext), | |
1056 | roq_encode_init, | |
1057 | roq_encode_frame, | |
1058 | roq_encode_end, | |
1059 | .supported_framerates = (AVRational[]){{30,1}, {0,0}}, | |
1060 | .pix_fmts = (enum PixelFormat[]){PIX_FMT_YUV444P, -1}, | |
1061 | }; |