Mercurial > libavcodec.hg
annotate roqvideoenc.c @ 11989:176c5deb6756 libavcodec
Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but
we apply them as 16x8/8x16/8x8 subblocks where possible. Since this allows
us to use width=8/16 instead of width=4 MC functions, we can now take more
advantage of SSE2/SSSE3 optimizations, leading to a total speedup for splitMV
filter of about 10%.
author | rbultje |
---|---|
date | Mon, 28 Jun 2010 13:50:55 +0000 |
parents | 7dd2a45249a9 |
children | fdafbcef52f5 |
rev | line source |
---|---|
5184 | 1 /* |
2 * RoQ Video Encoder. | |
3 * | |
5219 | 4 * Copyright (C) 2007 Vitor Sessak <vitor1001@gmail.com> |
5184 | 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 /** | |
11644
7dd2a45249a9
Remove explicit filename from Doxygen @file commands.
diego
parents:
11560
diff
changeset
|
26 * @file |
6812
0d01bae8d207
cosmetics: s/Id/id/ in libavcodec where Id refers to id Software.
diego
parents:
6788
diff
changeset
|
27 * id RoQ encoder by Vitor. Based on the Switchblade3 library and the |
5184 | 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 "roqvideo.h" | |
60 #include "bytestream.h" | |
61 #include "elbg.h" | |
8627
d6bab465b82c
moves mid_pred() into mathops.h (with arch specific code split by directory)
aurel
parents:
8296
diff
changeset
|
62 #include "mathops.h" |
5184 | 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, | |
5198 | 128 int y2, int *stride1, int *stride2, int size) |
5184 | 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++) | |
5198 | 136 sse += bias*eval_sse(buf1[k] + (y1+i)*stride1[k] + x1, |
137 buf2[k] + (y2+i)*stride2[k] + x2, size); | |
5184 | 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, | |
5198 | 162 mx, my, |
163 enc->frame_to_enc->linesize, enc->last_frame->linesize, | |
164 size); | |
5184 | 165 } |
166 | |
167 /** | |
168 * Returns distortion between two macroblocks | |
169 */ | |
170 static inline int squared_diff_macroblock(uint8_t a[], uint8_t b[], int size) | |
171 { | |
172 int cp, sdiff=0; | |
173 | |
174 for(cp=0;cp<3;cp++) { | |
175 int bias = (cp ? CHROMA_BIAS : 4); | |
176 sdiff += bias*eval_sse(a, b, size*size); | |
177 a += size*size; | |
178 b += size*size; | |
179 } | |
180 | |
181 return sdiff; | |
182 } | |
183 | |
184 typedef struct | |
185 { | |
186 int eval_dist[4]; | |
187 int best_bit_use; | |
188 int best_coding; | |
189 | |
190 int subCels[4]; | |
191 motion_vect motion; | |
192 int cbEntry; | |
8296 | 193 } SubcelEvaluation; |
5184 | 194 |
195 typedef struct | |
196 { | |
197 int eval_dist[4]; | |
198 int best_coding; | |
199 | |
8296 | 200 SubcelEvaluation subCels[4]; |
5184 | 201 |
202 motion_vect motion; | |
203 int cbEntry; | |
204 | |
205 int sourceX, sourceY; | |
8296 | 206 } CelEvaluation; |
5184 | 207 |
208 typedef struct | |
209 { | |
210 int numCB4; | |
211 int numCB2; | |
212 int usedCB2[MAX_CBS_2x2]; | |
213 int usedCB4[MAX_CBS_4x4]; | |
214 uint8_t unpacked_cb2[MAX_CBS_2x2*2*2*3]; | |
215 uint8_t unpacked_cb4[MAX_CBS_4x4*4*4*3]; | |
216 uint8_t unpacked_cb4_enlarged[MAX_CBS_4x4*8*8*3]; | |
8296 | 217 } RoqCodebooks; |
5184 | 218 |
219 /** | |
220 * Temporary vars | |
221 */ | |
8643 | 222 typedef struct RoqTempData |
5184 | 223 { |
8296 | 224 CelEvaluation *cel_evals; |
5184 | 225 |
226 int f2i4[MAX_CBS_4x4]; | |
227 int i2f4[MAX_CBS_4x4]; | |
228 int f2i2[MAX_CBS_2x2]; | |
229 int i2f2[MAX_CBS_2x2]; | |
230 | |
231 int mainChunkSize; | |
232 | |
233 int numCB4; | |
234 int numCB2; | |
235 | |
8296 | 236 RoqCodebooks codebooks; |
5184 | 237 |
238 int *closest_cb2; | |
239 int used_option[4]; | |
8296 | 240 } RoqTempdata; |
5184 | 241 |
242 /** | |
243 * Initializes cel evaluators and sets their source coordinates | |
244 */ | |
8296 | 245 static void create_cel_evals(RoqContext *enc, RoqTempdata *tempData) |
5184 | 246 { |
247 int n=0, x, y, i; | |
248 | |
8296 | 249 tempData->cel_evals = av_malloc(enc->width*enc->height/64 * sizeof(CelEvaluation)); |
5184 | 250 |
251 /* Map to the ROQ quadtree order */ | |
252 for (y=0; y<enc->height; y+=16) | |
253 for (x=0; x<enc->width; x+=16) | |
254 for(i=0; i<4; i++) { | |
255 tempData->cel_evals[n ].sourceX = x + (i&1)*8; | |
256 tempData->cel_evals[n++].sourceY = y + (i&2)*4; | |
257 } | |
258 } | |
259 | |
260 /** | |
261 * Get macroblocks from parts of the image | |
262 */ | |
263 static void get_frame_mb(AVFrame *frame, int x, int y, uint8_t mb[], int dim) | |
264 { | |
265 int i, j, cp; | |
266 | |
5191 | 267 for (cp=0; cp<3; cp++) { |
268 int stride = frame->linesize[cp]; | |
5184 | 269 for (i=0; i<dim; i++) |
270 for (j=0; j<dim; j++) | |
271 *mb++ = frame->data[cp][(y+i)*stride + x + j]; | |
5191 | 272 } |
5184 | 273 } |
274 | |
275 /** | |
276 * Find the codebook with the lowest distortion from an image | |
277 */ | |
278 static int index_mb(uint8_t cluster[], uint8_t cb[], int numCB, | |
279 int *outIndex, int dim) | |
280 { | |
281 int i, lDiff = INT_MAX, pick=0; | |
282 | |
283 /* Diff against the others */ | |
284 for (i=0; i<numCB; i++) { | |
285 int diff = squared_diff_macroblock(cluster, cb + i*dim*dim*3, dim); | |
286 if (diff < lDiff) { | |
287 lDiff = diff; | |
288 pick = i; | |
289 } | |
290 } | |
291 | |
292 *outIndex = pick; | |
293 return lDiff; | |
294 } | |
295 | |
296 #define EVAL_MOTION(MOTION) \ | |
297 do { \ | |
298 diff = eval_motion_dist(enc, j, i, MOTION, blocksize); \ | |
299 \ | |
300 if (diff < lowestdiff) { \ | |
301 lowestdiff = diff; \ | |
302 bestpick = MOTION; \ | |
303 } \ | |
304 } while(0) | |
305 | |
306 static void motion_search(RoqContext *enc, int blocksize) | |
307 { | |
308 static const motion_vect offsets[8] = { | |
309 {{ 0,-1}}, | |
310 {{ 0, 1}}, | |
311 {{-1, 0}}, | |
312 {{ 1, 0}}, | |
313 {{-1, 1}}, | |
314 {{ 1,-1}}, | |
315 {{-1,-1}}, | |
316 {{ 1, 1}}, | |
317 }; | |
318 | |
319 int diff, lowestdiff, oldbest; | |
320 int off[3]; | |
321 motion_vect bestpick = {{0,0}}; | |
322 int i, j, k, offset; | |
323 | |
324 motion_vect *last_motion; | |
325 motion_vect *this_motion; | |
326 motion_vect vect, vect2; | |
327 | |
328 int max=(enc->width/blocksize)*enc->height/blocksize; | |
329 | |
330 if (blocksize == 4) { | |
331 last_motion = enc->last_motion4; | |
332 this_motion = enc->this_motion4; | |
333 } else { | |
334 last_motion = enc->last_motion8; | |
335 this_motion = enc->this_motion8; | |
336 } | |
337 | |
338 for (i=0; i<enc->height; i+=blocksize) | |
339 for (j=0; j<enc->width; j+=blocksize) { | |
340 lowestdiff = eval_motion_dist(enc, j, i, (motion_vect) {{0,0}}, | |
341 blocksize); | |
342 bestpick.d[0] = 0; | |
343 bestpick.d[1] = 0; | |
344 | |
345 if (blocksize == 4) | |
346 EVAL_MOTION(enc->this_motion8[(i/8)*(enc->width/8) + j/8]); | |
347 | |
348 offset = (i/blocksize)*enc->width/blocksize + j/blocksize; | |
349 if (offset < max && offset >= 0) | |
350 EVAL_MOTION(last_motion[offset]); | |
351 | |
352 offset++; | |
353 if (offset < max && offset >= 0) | |
354 EVAL_MOTION(last_motion[offset]); | |
355 | |
356 offset = (i/blocksize + 1)*enc->width/blocksize + j/blocksize; | |
357 if (offset < max && offset >= 0) | |
358 EVAL_MOTION(last_motion[offset]); | |
359 | |
360 off[0]= (i/blocksize)*enc->width/blocksize + j/blocksize - 1; | |
361 off[1]= off[0] - enc->width/blocksize + 1; | |
362 off[2]= off[1] + 1; | |
363 | |
364 if (i) { | |
365 | |
366 for(k=0; k<2; k++) | |
367 vect.d[k]= mid_pred(this_motion[off[0]].d[k], | |
368 this_motion[off[1]].d[k], | |
369 this_motion[off[2]].d[k]); | |
370 | |
371 EVAL_MOTION(vect); | |
372 for(k=0; k<3; k++) | |
373 EVAL_MOTION(this_motion[off[k]]); | |
374 } else if(j) | |
375 EVAL_MOTION(this_motion[off[0]]); | |
376 | |
377 vect = bestpick; | |
378 | |
379 oldbest = -1; | |
380 while (oldbest != lowestdiff) { | |
381 oldbest = lowestdiff; | |
382 for (k=0; k<8; k++) { | |
383 vect2 = vect; | |
384 vect2.d[0] += offsets[k].d[0]; | |
385 vect2.d[1] += offsets[k].d[1]; | |
386 EVAL_MOTION(vect2); | |
387 } | |
388 vect = bestpick; | |
389 } | |
390 offset = (i/blocksize)*enc->width/blocksize + j/blocksize; | |
391 this_motion[offset] = bestpick; | |
392 } | |
393 } | |
394 | |
395 /** | |
396 * Gets distortion for all options available to a subcel | |
397 */ | |
8296 | 398 static void gather_data_for_subcel(SubcelEvaluation *subcel, int x, |
399 int y, RoqContext *enc, RoqTempdata *tempData) | |
5184 | 400 { |
401 uint8_t mb4[4*4*3]; | |
402 uint8_t mb2[2*2*3]; | |
403 int cluster_index; | |
404 int i, best_dist; | |
405 | |
406 static const int bitsUsed[4] = {2, 10, 10, 34}; | |
407 | |
408 if (enc->framesSinceKeyframe >= 1) { | |
409 subcel->motion = enc->this_motion4[y*enc->width/16 + x/4]; | |
410 | |
411 subcel->eval_dist[RoQ_ID_FCC] = | |
412 eval_motion_dist(enc, x, y, | |
413 enc->this_motion4[y*enc->width/16 + x/4], 4); | |
414 } else | |
415 subcel->eval_dist[RoQ_ID_FCC] = INT_MAX; | |
416 | |
417 if (enc->framesSinceKeyframe >= 2) | |
418 subcel->eval_dist[RoQ_ID_MOT] = block_sse(enc->frame_to_enc->data, | |
419 enc->current_frame->data, x, | |
5191 | 420 y, x, y, |
421 enc->frame_to_enc->linesize, | |
5198 | 422 enc->current_frame->linesize, |
5191 | 423 4); |
5184 | 424 else |
425 subcel->eval_dist[RoQ_ID_MOT] = INT_MAX; | |
426 | |
427 cluster_index = y*enc->width/16 + x/4; | |
428 | |
429 get_frame_mb(enc->frame_to_enc, x, y, mb4, 4); | |
430 | |
431 subcel->eval_dist[RoQ_ID_SLD] = index_mb(mb4, | |
432 tempData->codebooks.unpacked_cb4, | |
433 tempData->codebooks.numCB4, | |
434 &subcel->cbEntry, 4); | |
435 | |
436 subcel->eval_dist[RoQ_ID_CCC] = 0; | |
437 | |
438 for(i=0;i<4;i++) { | |
439 subcel->subCels[i] = tempData->closest_cb2[cluster_index*4+i]; | |
440 | |
441 get_frame_mb(enc->frame_to_enc, x+2*(i&1), | |
442 y+(i&2), mb2, 2); | |
443 | |
444 subcel->eval_dist[RoQ_ID_CCC] += | |
445 squared_diff_macroblock(tempData->codebooks.unpacked_cb2 + subcel->subCels[i]*2*2*3, mb2, 2); | |
446 } | |
447 | |
448 best_dist = INT_MAX; | |
449 for (i=0; i<4; i++) | |
450 if (ROQ_LAMBDA_SCALE*subcel->eval_dist[i] + enc->lambda*bitsUsed[i] < | |
451 best_dist) { | |
452 subcel->best_coding = i; | |
453 subcel->best_bit_use = bitsUsed[i]; | |
454 best_dist = ROQ_LAMBDA_SCALE*subcel->eval_dist[i] + | |
455 enc->lambda*bitsUsed[i]; | |
456 } | |
457 } | |
458 | |
459 /** | |
460 * Gets distortion for all options available to a cel | |
461 */ | |
8296 | 462 static void gather_data_for_cel(CelEvaluation *cel, RoqContext *enc, |
463 RoqTempdata *tempData) | |
5184 | 464 { |
465 uint8_t mb8[8*8*3]; | |
466 int index = cel->sourceY*enc->width/64 + cel->sourceX/8; | |
467 int i, j, best_dist, divide_bit_use; | |
468 | |
469 int bitsUsed[4] = {2, 10, 10, 0}; | |
470 | |
471 if (enc->framesSinceKeyframe >= 1) { | |
472 cel->motion = enc->this_motion8[index]; | |
473 | |
474 cel->eval_dist[RoQ_ID_FCC] = | |
475 eval_motion_dist(enc, cel->sourceX, cel->sourceY, | |
476 enc->this_motion8[index], 8); | |
477 } else | |
478 cel->eval_dist[RoQ_ID_FCC] = INT_MAX; | |
479 | |
480 if (enc->framesSinceKeyframe >= 2) | |
481 cel->eval_dist[RoQ_ID_MOT] = block_sse(enc->frame_to_enc->data, | |
482 enc->current_frame->data, | |
483 cel->sourceX, cel->sourceY, | |
484 cel->sourceX, cel->sourceY, | |
5198 | 485 enc->frame_to_enc->linesize, |
486 enc->current_frame->linesize,8); | |
5184 | 487 else |
488 cel->eval_dist[RoQ_ID_MOT] = INT_MAX; | |
489 | |
490 get_frame_mb(enc->frame_to_enc, cel->sourceX, cel->sourceY, mb8, 8); | |
491 | |
492 cel->eval_dist[RoQ_ID_SLD] = | |
493 index_mb(mb8, tempData->codebooks.unpacked_cb4_enlarged, | |
494 tempData->codebooks.numCB4, &cel->cbEntry, 8); | |
495 | |
496 gather_data_for_subcel(cel->subCels + 0, cel->sourceX+0, cel->sourceY+0, enc, tempData); | |
497 gather_data_for_subcel(cel->subCels + 1, cel->sourceX+4, cel->sourceY+0, enc, tempData); | |
498 gather_data_for_subcel(cel->subCels + 2, cel->sourceX+0, cel->sourceY+4, enc, tempData); | |
499 gather_data_for_subcel(cel->subCels + 3, cel->sourceX+4, cel->sourceY+4, enc, tempData); | |
500 | |
501 cel->eval_dist[RoQ_ID_CCC] = 0; | |
502 divide_bit_use = 0; | |
503 for (i=0; i<4; i++) { | |
504 cel->eval_dist[RoQ_ID_CCC] += | |
505 cel->subCels[i].eval_dist[cel->subCels[i].best_coding]; | |
506 divide_bit_use += cel->subCels[i].best_bit_use; | |
507 } | |
508 | |
509 best_dist = INT_MAX; | |
510 bitsUsed[3] = 2 + divide_bit_use; | |
511 | |
512 for (i=0; i<4; i++) | |
513 if (ROQ_LAMBDA_SCALE*cel->eval_dist[i] + enc->lambda*bitsUsed[i] < | |
514 best_dist) { | |
515 cel->best_coding = i; | |
516 best_dist = ROQ_LAMBDA_SCALE*cel->eval_dist[i] + | |
517 enc->lambda*bitsUsed[i]; | |
518 } | |
519 | |
520 tempData->used_option[cel->best_coding]++; | |
521 tempData->mainChunkSize += bitsUsed[cel->best_coding]; | |
522 | |
523 if (cel->best_coding == RoQ_ID_SLD) | |
524 tempData->codebooks.usedCB4[cel->cbEntry]++; | |
525 | |
526 if (cel->best_coding == RoQ_ID_CCC) | |
527 for (i=0; i<4; i++) { | |
528 if (cel->subCels[i].best_coding == RoQ_ID_SLD) | |
529 tempData->codebooks.usedCB4[cel->subCels[i].cbEntry]++; | |
530 else if (cel->subCels[i].best_coding == RoQ_ID_CCC) | |
531 for (j=0; j<4; j++) | |
532 tempData->codebooks.usedCB2[cel->subCels[i].subCels[j]]++; | |
533 } | |
534 } | |
535 | |
8296 | 536 static void remap_codebooks(RoqContext *enc, RoqTempdata *tempData) |
5184 | 537 { |
538 int i, j, idx=0; | |
539 | |
540 /* Make remaps for the final codebook usage */ | |
541 for (i=0; i<MAX_CBS_4x4; i++) { | |
542 if (tempData->codebooks.usedCB4[i]) { | |
543 tempData->i2f4[i] = idx; | |
544 tempData->f2i4[idx] = i; | |
545 for (j=0; j<4; j++) | |
546 tempData->codebooks.usedCB2[enc->cb4x4[i].idx[j]]++; | |
547 idx++; | |
548 } | |
549 } | |
550 | |
551 tempData->numCB4 = idx; | |
552 | |
553 idx = 0; | |
554 for (i=0; i<MAX_CBS_2x2; i++) { | |
555 if (tempData->codebooks.usedCB2[i]) { | |
556 tempData->i2f2[i] = idx; | |
557 tempData->f2i2[idx] = i; | |
558 idx++; | |
559 } | |
560 } | |
561 tempData->numCB2 = idx; | |
562 | |
563 } | |
564 | |
565 /** | |
566 * Write codebook chunk | |
567 */ | |
8296 | 568 static void write_codebooks(RoqContext *enc, RoqTempdata *tempData) |
5184 | 569 { |
570 int i, j; | |
571 uint8_t **outp= &enc->out_buf; | |
572 | |
573 if (tempData->numCB2) { | |
574 bytestream_put_le16(outp, RoQ_QUAD_CODEBOOK); | |
575 bytestream_put_le32(outp, tempData->numCB2*6 + tempData->numCB4*4); | |
576 bytestream_put_byte(outp, tempData->numCB4); | |
577 bytestream_put_byte(outp, tempData->numCB2); | |
578 | |
579 for (i=0; i<tempData->numCB2; i++) { | |
580 bytestream_put_buffer(outp, enc->cb2x2[tempData->f2i2[i]].y, 4); | |
581 bytestream_put_byte(outp, enc->cb2x2[tempData->f2i2[i]].u); | |
582 bytestream_put_byte(outp, enc->cb2x2[tempData->f2i2[i]].v); | |
583 } | |
584 | |
585 for (i=0; i<tempData->numCB4; i++) | |
586 for (j=0; j<4; j++) | |
587 bytestream_put_byte(outp, tempData->i2f2[enc->cb4x4[tempData->f2i4[i]].idx[j]]); | |
588 | |
589 } | |
590 } | |
591 | |
592 static inline uint8_t motion_arg(motion_vect mot) | |
593 { | |
594 uint8_t ax = 8 - ((uint8_t) mot.d[0]); | |
595 uint8_t ay = 8 - ((uint8_t) mot.d[1]); | |
596 return ((ax&15)<<4) | (ay&15); | |
597 } | |
598 | |
599 typedef struct | |
600 { | |
601 int typeSpool; | |
602 int typeSpoolLength; | |
603 uint8_t argumentSpool[64]; | |
604 uint8_t *args; | |
605 uint8_t **pout; | |
606 } CodingSpool; | |
607 | |
608 /* NOTE: Typecodes must be spooled AFTER arguments!! */ | |
609 static void write_typecode(CodingSpool *s, uint8_t type) | |
610 { | |
5205 | 611 s->typeSpool |= (type & 3) << (14 - s->typeSpoolLength); |
5184 | 612 s->typeSpoolLength += 2; |
613 if (s->typeSpoolLength == 16) { | |
614 bytestream_put_le16(s->pout, s->typeSpool); | |
615 bytestream_put_buffer(s->pout, s->argumentSpool, | |
616 s->args - s->argumentSpool); | |
617 s->typeSpoolLength = 0; | |
618 s->typeSpool = 0; | |
619 s->args = s->argumentSpool; | |
620 } | |
621 } | |
622 | |
8296 | 623 static void reconstruct_and_encode_image(RoqContext *enc, RoqTempdata *tempData, int w, int h, int numBlocks) |
5184 | 624 { |
625 int i, j, k; | |
626 int x, y; | |
627 int subX, subY; | |
628 int dist=0; | |
629 | |
630 roq_qcell *qcell; | |
8296 | 631 CelEvaluation *eval; |
5184 | 632 |
633 CodingSpool spool; | |
634 | |
635 spool.typeSpool=0; | |
636 spool.typeSpoolLength=0; | |
637 spool.args = spool.argumentSpool; | |
638 spool.pout = &enc->out_buf; | |
639 | |
640 if (tempData->used_option[RoQ_ID_CCC]%2) | |
641 tempData->mainChunkSize+=8; //FIXME | |
642 | |
643 /* Write the video chunk header */ | |
644 bytestream_put_le16(&enc->out_buf, RoQ_QUAD_VQ); | |
645 bytestream_put_le32(&enc->out_buf, tempData->mainChunkSize/8); | |
646 bytestream_put_byte(&enc->out_buf, 0x0); | |
647 bytestream_put_byte(&enc->out_buf, 0x0); | |
648 | |
649 for (i=0; i<numBlocks; i++) { | |
650 eval = tempData->cel_evals + i; | |
651 | |
652 x = eval->sourceX; | |
653 y = eval->sourceY; | |
654 dist += eval->eval_dist[eval->best_coding]; | |
655 | |
656 switch (eval->best_coding) { | |
657 case RoQ_ID_MOT: | |
658 write_typecode(&spool, RoQ_ID_MOT); | |
659 break; | |
660 | |
661 case RoQ_ID_FCC: | |
662 bytestream_put_byte(&spool.args, motion_arg(eval->motion)); | |
663 | |
664 write_typecode(&spool, RoQ_ID_FCC); | |
665 ff_apply_motion_8x8(enc, x, y, | |
666 eval->motion.d[0], eval->motion.d[1]); | |
667 break; | |
668 | |
669 case RoQ_ID_SLD: | |
670 bytestream_put_byte(&spool.args, tempData->i2f4[eval->cbEntry]); | |
671 write_typecode(&spool, RoQ_ID_SLD); | |
672 | |
673 qcell = enc->cb4x4 + eval->cbEntry; | |
674 ff_apply_vector_4x4(enc, x , y , enc->cb2x2 + qcell->idx[0]); | |
675 ff_apply_vector_4x4(enc, x+4, y , enc->cb2x2 + qcell->idx[1]); | |
676 ff_apply_vector_4x4(enc, x , y+4, enc->cb2x2 + qcell->idx[2]); | |
677 ff_apply_vector_4x4(enc, x+4, y+4, enc->cb2x2 + qcell->idx[3]); | |
678 break; | |
679 | |
680 case RoQ_ID_CCC: | |
681 write_typecode(&spool, RoQ_ID_CCC); | |
682 | |
683 for (j=0; j<4; j++) { | |
684 subX = x + 4*(j&1); | |
685 subY = y + 2*(j&2); | |
686 | |
687 switch(eval->subCels[j].best_coding) { | |
688 case RoQ_ID_MOT: | |
689 break; | |
690 | |
691 case RoQ_ID_FCC: | |
692 bytestream_put_byte(&spool.args, | |
693 motion_arg(eval->subCels[j].motion)); | |
694 | |
695 ff_apply_motion_4x4(enc, subX, subY, | |
696 eval->subCels[j].motion.d[0], | |
697 eval->subCels[j].motion.d[1]); | |
698 break; | |
699 | |
700 case RoQ_ID_SLD: | |
701 bytestream_put_byte(&spool.args, | |
702 tempData->i2f4[eval->subCels[j].cbEntry]); | |
703 | |
704 qcell = enc->cb4x4 + eval->subCels[j].cbEntry; | |
705 | |
706 ff_apply_vector_2x2(enc, subX , subY , | |
707 enc->cb2x2 + qcell->idx[0]); | |
708 ff_apply_vector_2x2(enc, subX+2, subY , | |
709 enc->cb2x2 + qcell->idx[1]); | |
710 ff_apply_vector_2x2(enc, subX , subY+2, | |
711 enc->cb2x2 + qcell->idx[2]); | |
712 ff_apply_vector_2x2(enc, subX+2, subY+2, | |
713 enc->cb2x2 + qcell->idx[3]); | |
714 break; | |
715 | |
716 case RoQ_ID_CCC: | |
717 for (k=0; k<4; k++) { | |
718 int cb_idx = eval->subCels[j].subCels[k]; | |
719 bytestream_put_byte(&spool.args, | |
720 tempData->i2f2[cb_idx]); | |
721 | |
722 ff_apply_vector_2x2(enc, subX + 2*(k&1), subY + (k&2), | |
723 enc->cb2x2 + cb_idx); | |
724 } | |
725 break; | |
726 } | |
727 write_typecode(&spool, eval->subCels[j].best_coding); | |
728 } | |
729 break; | |
730 } | |
731 } | |
732 | |
733 /* Flush the remainder of the argument/type spool */ | |
734 while (spool.typeSpoolLength) | |
735 write_typecode(&spool, 0x0); | |
736 | |
737 #if 0 | |
738 uint8_t *fdata[3] = {enc->frame_to_enc->data[0], | |
739 enc->frame_to_enc->data[1], | |
740 enc->frame_to_enc->data[2]}; | |
741 uint8_t *cdata[3] = {enc->current_frame->data[0], | |
742 enc->current_frame->data[1], | |
743 enc->current_frame->data[2]}; | |
744 av_log(enc->avctx, AV_LOG_ERROR, "Expected distortion: %i Actual: %i\n", | |
745 dist, | |
746 block_sse(fdata, cdata, 0, 0, 0, 0, | |
5198 | 747 enc->frame_to_enc->linesize, |
748 enc->current_frame->linesize, | |
749 enc->width)); //WARNING: Square dimensions implied... | |
5184 | 750 #endif |
751 } | |
752 | |
753 | |
754 /** | |
755 * Create a single YUV cell from a 2x2 section of the image | |
756 */ | |
757 static inline void frame_block_to_cell(uint8_t *block, uint8_t **data, | |
758 int top, int left, int *stride) | |
759 { | |
760 int i, j, u=0, v=0; | |
761 | |
762 for (i=0; i<2; i++) | |
763 for (j=0; j<2; j++) { | |
764 int x = (top+i)*stride[0] + left + j; | |
765 *block++ = data[0][x]; | |
5191 | 766 x = (top+i)*stride[1] + left + j; |
5184 | 767 u += data[1][x]; |
768 v += data[2][x]; | |
769 } | |
770 | |
771 *block++ = (u+2)/4; | |
772 *block++ = (v+2)/4; | |
773 } | |
774 | |
775 /** | |
776 * Creates YUV clusters for the entire image | |
777 */ | |
778 static void create_clusters(AVFrame *frame, int w, int h, uint8_t *yuvClusters) | |
779 { | |
780 int i, j, k, l; | |
781 | |
782 for (i=0; i<h; i+=4) | |
783 for (j=0; j<w; j+=4) { | |
784 for (k=0; k < 2; k++) | |
785 for (l=0; l < 2; l++) | |
786 frame_block_to_cell(yuvClusters + (l + 2*k)*6, frame->data, | |
787 i+2*k, j+2*l, frame->linesize); | |
788 yuvClusters += 24; | |
789 } | |
790 } | |
791 | |
8296 | 792 static void generate_codebook(RoqContext *enc, RoqTempdata *tempdata, |
5184 | 793 int *points, int inputCount, roq_cell *results, |
794 int size, int cbsize) | |
795 { | |
796 int i, j, k; | |
797 int c_size = size*size/4; | |
9460
12c8d0e11f39
Remove useless assignment in generate_codebook(). Found by Clang static analyser.
vitor
parents:
9154
diff
changeset
|
798 int *buf; |
5184 | 799 int *codebook = av_malloc(6*c_size*cbsize*sizeof(int)); |
800 int *closest_cb; | |
801 | |
802 if (size == 4) | |
803 closest_cb = av_malloc(6*c_size*inputCount*sizeof(int)); | |
804 else | |
805 closest_cb = tempdata->closest_cb2; | |
806 | |
807 ff_init_elbg(points, 6*c_size, inputCount, codebook, cbsize, 1, closest_cb, &enc->randctx); | |
808 ff_do_elbg(points, 6*c_size, inputCount, codebook, cbsize, 1, closest_cb, &enc->randctx); | |
809 | |
810 if (size == 4) | |
811 av_free(closest_cb); | |
812 | |
813 buf = codebook; | |
814 for (i=0; i<cbsize; i++) | |
815 for (k=0; k<c_size; k++) { | |
816 for(j=0; j<4; j++) | |
817 results->y[j] = *buf++; | |
818 | |
819 results->u = (*buf++ + CHROMA_BIAS/2)/CHROMA_BIAS; | |
820 results->v = (*buf++ + CHROMA_BIAS/2)/CHROMA_BIAS; | |
821 results++; | |
822 } | |
823 | |
824 av_free(codebook); | |
825 } | |
826 | |
8296 | 827 static void generate_new_codebooks(RoqContext *enc, RoqTempdata *tempData) |
5184 | 828 { |
829 int i,j; | |
8296 | 830 RoqCodebooks *codebooks = &tempData->codebooks; |
5184 | 831 int max = enc->width*enc->height/16; |
832 uint8_t mb2[3*4]; | |
833 roq_cell *results4 = av_malloc(sizeof(roq_cell)*MAX_CBS_4x4*4); | |
834 uint8_t *yuvClusters=av_malloc(sizeof(int)*max*6*4); | |
835 int *points = av_malloc(max*6*4*sizeof(int)); | |
836 int bias; | |
837 | |
838 /* Subsample YUV data */ | |
839 create_clusters(enc->frame_to_enc, enc->width, enc->height, yuvClusters); | |
840 | |
841 /* Cast to integer and apply chroma bias */ | |
842 for (i=0; i<max*24; i++) { | |
843 bias = ((i%6)<4) ? 1 : CHROMA_BIAS; | |
844 points[i] = bias*yuvClusters[i]; | |
845 } | |
846 | |
847 /* Create 4x4 codebooks */ | |
848 generate_codebook(enc, tempData, points, max, results4, 4, MAX_CBS_4x4); | |
849 | |
850 codebooks->numCB4 = MAX_CBS_4x4; | |
851 | |
852 tempData->closest_cb2 = av_malloc(max*4*sizeof(int)); | |
853 | |
854 /* Create 2x2 codebooks */ | |
855 generate_codebook(enc, tempData, points, max*4, enc->cb2x2, 2, MAX_CBS_2x2); | |
856 | |
857 codebooks->numCB2 = MAX_CBS_2x2; | |
858 | |
859 /* Unpack 2x2 codebook clusters */ | |
860 for (i=0; i<codebooks->numCB2; i++) | |
861 unpack_roq_cell(enc->cb2x2 + i, codebooks->unpacked_cb2 + i*2*2*3); | |
862 | |
863 /* Index all 4x4 entries to the 2x2 entries, unpack, and enlarge */ | |
864 for (i=0; i<codebooks->numCB4; i++) { | |
865 for (j=0; j<4; j++) { | |
866 unpack_roq_cell(&results4[4*i + j], mb2); | |
867 index_mb(mb2, codebooks->unpacked_cb2, codebooks->numCB2, | |
868 &enc->cb4x4[i].idx[j], 2); | |
869 } | |
870 unpack_roq_qcell(codebooks->unpacked_cb2, enc->cb4x4 + i, | |
871 codebooks->unpacked_cb4 + i*4*4*3); | |
872 enlarge_roq_mb4(codebooks->unpacked_cb4 + i*4*4*3, | |
873 codebooks->unpacked_cb4_enlarged + i*8*8*3); | |
874 } | |
875 | |
876 av_free(yuvClusters); | |
877 av_free(points); | |
878 av_free(results4); | |
879 } | |
880 | |
881 static void roq_encode_video(RoqContext *enc) | |
882 { | |
8643 | 883 RoqTempdata *tempData = enc->tmpData; |
5184 | 884 int i; |
885 | |
8643 | 886 memset(tempData, 0, sizeof(*tempData)); |
5184 | 887 |
8643 | 888 create_cel_evals(enc, tempData); |
5184 | 889 |
8643 | 890 generate_new_codebooks(enc, tempData); |
5184 | 891 |
892 if (enc->framesSinceKeyframe >= 1) { | |
893 motion_search(enc, 8); | |
894 motion_search(enc, 4); | |
895 } | |
896 | |
897 retry_encode: | |
898 for (i=0; i<enc->width*enc->height/64; i++) | |
8643 | 899 gather_data_for_cel(tempData->cel_evals + i, enc, tempData); |
5184 | 900 |
901 /* Quake 3 can't handle chunks bigger than 65536 bytes */ | |
8643 | 902 if (tempData->mainChunkSize/8 > 65536) { |
5184 | 903 enc->lambda *= .8; |
904 goto retry_encode; | |
905 } | |
906 | |
8643 | 907 remap_codebooks(enc, tempData); |
5184 | 908 |
8643 | 909 write_codebooks(enc, tempData); |
5184 | 910 |
8643 | 911 reconstruct_and_encode_image(enc, tempData, enc->width, enc->height, |
5184 | 912 enc->width*enc->height/64); |
913 | |
7308
29319d07aeb4
Set avctx->coded_frame in RoQ encoder. At some point in
vitor
parents:
7040
diff
changeset
|
914 enc->avctx->coded_frame = enc->current_frame; |
29319d07aeb4
Set avctx->coded_frame in RoQ encoder. At some point in
vitor
parents:
7040
diff
changeset
|
915 |
5184 | 916 /* Rotate frame history */ |
917 FFSWAP(AVFrame *, enc->current_frame, enc->last_frame); | |
918 FFSWAP(motion_vect *, enc->last_motion4, enc->this_motion4); | |
919 FFSWAP(motion_vect *, enc->last_motion8, enc->this_motion8); | |
920 | |
8643 | 921 av_free(tempData->cel_evals); |
922 av_free(tempData->closest_cb2); | |
5184 | 923 |
924 enc->framesSinceKeyframe++; | |
925 } | |
926 | |
927 static int roq_encode_init(AVCodecContext *avctx) | |
928 { | |
929 RoqContext *enc = avctx->priv_data; | |
930 | |
9154
aa459306ee59
Use FLG pseudo-random number generator in RoQ and ELBG
vitor
parents:
8718
diff
changeset
|
931 av_lfg_init(&enc->randctx, 1); |
5184 | 932 |
933 enc->framesSinceKeyframe = 0; | |
934 if ((avctx->width & 0xf) || (avctx->height & 0xf)) { | |
935 av_log(avctx, AV_LOG_ERROR, "Dimensions must be divisible by 16\n"); | |
936 return -1; | |
937 } | |
938 | |
939 if (((avctx->width)&(avctx->width-1))||((avctx->height)&(avctx->height-1))) | |
940 av_log(avctx, AV_LOG_ERROR, "Warning: dimensions not power of two\n"); | |
941 | |
942 enc->width = avctx->width; | |
943 enc->height = avctx->height; | |
944 | |
945 enc->framesSinceKeyframe = 0; | |
946 enc->first_frame = 1; | |
947 | |
948 enc->last_frame = &enc->frames[0]; | |
949 enc->current_frame = &enc->frames[1]; | |
950 | |
8643 | 951 enc->tmpData = av_malloc(sizeof(RoqTempdata)); |
952 | |
5184 | 953 enc->this_motion4 = |
954 av_mallocz((enc->width*enc->height/16)*sizeof(motion_vect)); | |
955 | |
956 enc->last_motion4 = | |
957 av_malloc ((enc->width*enc->height/16)*sizeof(motion_vect)); | |
958 | |
959 enc->this_motion8 = | |
960 av_mallocz((enc->width*enc->height/64)*sizeof(motion_vect)); | |
961 | |
962 enc->last_motion8 = | |
963 av_malloc ((enc->width*enc->height/64)*sizeof(motion_vect)); | |
964 | |
965 return 0; | |
966 } | |
967 | |
968 static void roq_write_video_info_chunk(RoqContext *enc) | |
969 { | |
970 /* ROQ info chunk */ | |
971 bytestream_put_le16(&enc->out_buf, RoQ_INFO); | |
972 | |
973 /* Size: 8 bytes */ | |
974 bytestream_put_le32(&enc->out_buf, 8); | |
975 | |
976 /* Unused argument */ | |
977 bytestream_put_byte(&enc->out_buf, 0x00); | |
978 bytestream_put_byte(&enc->out_buf, 0x00); | |
979 | |
980 /* Width */ | |
981 bytestream_put_le16(&enc->out_buf, enc->width); | |
982 | |
983 /* Height */ | |
984 bytestream_put_le16(&enc->out_buf, enc->height); | |
985 | |
986 /* Unused in Quake 3, mimics the output of the real encoder */ | |
987 bytestream_put_byte(&enc->out_buf, 0x08); | |
988 bytestream_put_byte(&enc->out_buf, 0x00); | |
989 bytestream_put_byte(&enc->out_buf, 0x04); | |
990 bytestream_put_byte(&enc->out_buf, 0x00); | |
991 } | |
992 | |
993 static int roq_encode_frame(AVCodecContext *avctx, unsigned char *buf, int buf_size, void *data) | |
994 { | |
995 RoqContext *enc = avctx->priv_data; | |
996 AVFrame *frame= data; | |
997 uint8_t *buf_start = buf; | |
998 | |
999 enc->out_buf = buf; | |
1000 enc->avctx = avctx; | |
1001 | |
1002 enc->frame_to_enc = frame; | |
1003 | |
1004 if (frame->quality) | |
1005 enc->lambda = frame->quality - 1; | |
1006 else | |
1007 enc->lambda = 2*ROQ_LAMBDA_SCALE; | |
1008 | |
1009 /* 138 bits max per 8x8 block + | |
1010 * 256 codebooks*(6 bytes 2x2 + 4 bytes 4x4) + 8 bytes frame header */ | |
1011 if (((enc->width*enc->height/64)*138+7)/8 + 256*(6+4) + 8 > buf_size) { | |
1012 av_log(avctx, AV_LOG_ERROR, " RoQ: Output buffer too small!\n"); | |
1013 return -1; | |
1014 } | |
1015 | |
1016 /* Check for I frame */ | |
1017 if (enc->framesSinceKeyframe == avctx->gop_size) | |
1018 enc->framesSinceKeyframe = 0; | |
1019 | |
1020 if (enc->first_frame) { | |
1021 /* Alloc memory for the reconstruction data (we must know the stride | |
1022 for that) */ | |
1023 if (avctx->get_buffer(avctx, enc->current_frame) || | |
1024 avctx->get_buffer(avctx, enc->last_frame)) { | |
1025 av_log(avctx, AV_LOG_ERROR, " RoQ: get_buffer() failed\n"); | |
1026 return -1; | |
1027 } | |
1028 | |
1029 /* Before the first video frame, write a "video info" chunk */ | |
1030 roq_write_video_info_chunk(enc); | |
1031 | |
1032 enc->first_frame = 0; | |
1033 } | |
1034 | |
1035 /* Encode the actual frame */ | |
1036 roq_encode_video(enc); | |
1037 | |
1038 return enc->out_buf - buf_start; | |
1039 } | |
1040 | |
1041 static int roq_encode_end(AVCodecContext *avctx) | |
1042 { | |
1043 RoqContext *enc = avctx->priv_data; | |
1044 | |
1045 avctx->release_buffer(avctx, enc->last_frame); | |
1046 avctx->release_buffer(avctx, enc->current_frame); | |
1047 | |
8643 | 1048 av_free(enc->tmpData); |
5184 | 1049 av_free(enc->this_motion4); |
1050 av_free(enc->last_motion4); | |
1051 av_free(enc->this_motion8); | |
1052 av_free(enc->last_motion8); | |
1053 | |
1054 return 0; | |
1055 } | |
1056 | |
1057 AVCodec roq_encoder = | |
1058 { | |
1059 "roqvideo", | |
11560
8a4984c5cacc
Define AVMediaType enum, and use it instead of enum CodecType, which
stefano
parents:
10397
diff
changeset
|
1060 AVMEDIA_TYPE_VIDEO, |
5184 | 1061 CODEC_ID_ROQ, |
1062 sizeof(RoqContext), | |
1063 roq_encode_init, | |
1064 roq_encode_frame, | |
1065 roq_encode_end, | |
10146
38cfe222e1a4
Mark all pix_fmts and supported_framerates compound literals as const.
reimar
parents:
9981
diff
changeset
|
1066 .supported_framerates = (const AVRational[]){{30,1}, {0,0}}, |
38cfe222e1a4
Mark all pix_fmts and supported_framerates compound literals as const.
reimar
parents:
9981
diff
changeset
|
1067 .pix_fmts = (const enum PixelFormat[]){PIX_FMT_YUV444P, PIX_FMT_NONE}, |
7040
e943e1409077
Make AVCodec long_names definition conditional depending on CONFIG_SMALL.
stefano
parents:
6812
diff
changeset
|
1068 .long_name = NULL_IF_CONFIG_SMALL("id RoQ video"), |
5184 | 1069 }; |