blob: a28f6690708f62cec00b35f8d0fcf405b1094c02
1 | /* |
2 | * Cinepak encoder (c) 2011 Tomas Härdin |
3 | * http://titan.codemill.se/~tomhar/cinepakenc.patch |
4 | * |
5 | * Fixes and improvements, vintage decoders compatibility |
6 | * (c) 2013, 2014 Rl, Aetey Global Technologies AB |
7 | |
8 | Permission is hereby granted, free of charge, to any person obtaining a |
9 | copy of this software and associated documentation files (the "Software"), |
10 | to deal in the Software without restriction, including without limitation |
11 | the rights to use, copy, modify, merge, publish, distribute, sublicense, |
12 | and/or sell copies of the Software, and to permit persons to whom the |
13 | Software is furnished to do so, subject to the following conditions: |
14 | |
15 | The above copyright notice and this permission notice shall be included |
16 | in all copies or substantial portions of the Software. |
17 | |
18 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
19 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
20 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL |
21 | THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR |
22 | OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, |
23 | ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR |
24 | OTHER DEALINGS IN THE SOFTWARE. |
25 | |
26 | * MAYBE: |
27 | * - "optimally" split the frame into several non-regular areas |
28 | * using a separate codebook pair for each area and approximating |
29 | * the area by several rectangular strips (generally not full width ones) |
30 | * (use quadtree splitting? a simple fixed-granularity grid?) |
31 | * |
32 | * |
33 | * version 2014-01-23 Rl |
34 | * - added option handling for flexibility |
35 | * |
36 | * version 2014-01-21 Rl |
37 | * - believe it or not, now we get even smaller files, with better quality |
38 | * (which means I missed an optimization earlier :) |
39 | * |
40 | * version 2014-01-20 Rl |
41 | * - made the encoder compatible with vintage decoders |
42 | * and added some yet unused code for possible future |
43 | * incremental codebook updates |
44 | * - fixed a small memory leak |
45 | * |
46 | * version 2013-04-28 Rl |
47 | * - bugfixed codebook optimization logic |
48 | * |
49 | * version 2013-02-14 Rl |
50 | * "Valentine's Day" version: |
51 | * - made strip division more robust |
52 | * - minimized bruteforcing the number of strips, |
53 | * (costs some R/D but speeds up compession a lot), the heuristic |
54 | * assumption is that score as a function of the number of strips has |
55 | * one wide minimum which moves slowly, of course not fully true |
56 | * - simplified codebook generation, |
57 | * the old code was meant for other optimizations than we actually do |
58 | * - optimized the codebook generation / error estimation for MODE_MC |
59 | * |
60 | * version 2013-02-12 Rl |
61 | * - separated codebook training sets, avoided the transfer of wasted bytes, |
62 | * which yields both better quality and smaller files |
63 | * - now using the correct colorspace (TODO: move conversion to libswscale) |
64 | * |
65 | * version 2013-02-08 Rl |
66 | * - fixes/optimization in multistrip encoding and codebook size choice, |
67 | * quality/bitrate is now better than that of the binary proprietary encoder |
68 | */ |
69 | |
70 | #include "libavutil/intreadwrite.h" |
71 | #include "avcodec.h" |
72 | #include "libavutil/lfg.h" |
73 | #include "elbg.h" |
74 | #include "internal.h" |
75 | |
76 | #include "libavutil/avassert.h" |
77 | #include "libavutil/opt.h" |
78 | |
79 | #define CVID_HEADER_SIZE 10 |
80 | #define STRIP_HEADER_SIZE 12 |
81 | #define CHUNK_HEADER_SIZE 4 |
82 | |
83 | #define MB_SIZE 4 //4x4 MBs |
84 | #define MB_AREA (MB_SIZE*MB_SIZE) |
85 | |
86 | #define VECTOR_MAX 6 //six or four entries per vector depending on format |
87 | #define CODEBOOK_MAX 256 //size of a codebook |
88 | |
89 | #define MAX_STRIPS 32 //Note: having fewer choices regarding the number of strips speeds up encoding (obviously) |
90 | #define MIN_STRIPS 1 //Note: having more strips speeds up encoding the frame (this is less obvious) |
91 | // MAX_STRIPS limits the maximum quality you can reach |
92 | // when you want high quality on high resolutions, |
93 | // MIN_STRIPS limits the minimum efficiently encodable bit rate |
94 | // on low resolutions |
95 | // the numbers are only used for brute force optimization for the first frame, |
96 | // for the following frames they are adaptively readjusted |
97 | // NOTE the decoder in ffmpeg has its own arbitrary limitation on the number |
98 | // of strips, currently 32 |
99 | |
100 | typedef enum { |
101 | MODE_V1_ONLY = 0, |
102 | MODE_V1_V4, |
103 | MODE_MC, |
104 | |
105 | MODE_COUNT, |
106 | } CinepakMode; |
107 | |
108 | typedef enum { |
109 | ENC_V1, |
110 | ENC_V4, |
111 | ENC_SKIP, |
112 | |
113 | ENC_UNCERTAIN |
114 | } mb_encoding; |
115 | |
116 | typedef struct { |
117 | int v1_vector; //index into v1 codebook |
118 | int v1_error; //error when using V1 encoding |
119 | int v4_vector[4]; //indices into v4 codebook |
120 | int v4_error; //error when using V4 encoding |
121 | int skip_error; //error when block is skipped (aka copied from last frame) |
122 | mb_encoding best_encoding; //last result from calculate_mode_score() |
123 | } mb_info; |
124 | |
125 | typedef struct { |
126 | int v1_codebook[CODEBOOK_MAX*VECTOR_MAX]; |
127 | int v4_codebook[CODEBOOK_MAX*VECTOR_MAX]; |
128 | int v1_size; |
129 | int v4_size; |
130 | CinepakMode mode; |
131 | } strip_info; |
132 | |
133 | typedef struct { |
134 | const AVClass *class; |
135 | AVCodecContext *avctx; |
136 | unsigned char *pict_bufs[4], *strip_buf, *frame_buf; |
137 | AVFrame *last_frame; |
138 | AVFrame *best_frame; |
139 | AVFrame *scratch_frame; |
140 | AVFrame *input_frame; |
141 | enum AVPixelFormat pix_fmt; |
142 | int w, h; |
143 | int frame_buf_size; |
144 | int curframe, keyint; |
145 | AVLFG randctx; |
146 | uint64_t lambda; |
147 | int *codebook_input; |
148 | int *codebook_closest; |
149 | mb_info *mb; //MB RD state |
150 | int min_strips; //the current limit |
151 | int max_strips; //the current limit |
152 | #ifdef CINEPAKENC_DEBUG |
153 | mb_info *best_mb; //TODO: remove. only used for printing stats |
154 | int num_v1_mode, num_v4_mode, num_mc_mode; |
155 | int num_v1_encs, num_v4_encs, num_skips; |
156 | #endif |
157 | // options |
158 | int max_extra_cb_iterations; |
159 | int skip_empty_cb; |
160 | int min_min_strips; |
161 | int max_max_strips; |
162 | int strip_number_delta_range; |
163 | } CinepakEncContext; |
164 | |
165 | #define OFFSET(x) offsetof(CinepakEncContext, x) |
166 | #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM |
167 | static const AVOption options[] = { |
168 | { "max_extra_cb_iterations", "Max extra codebook recalculation passes, more is better and slower", OFFSET(max_extra_cb_iterations), AV_OPT_TYPE_INT, { .i64 = 2 }, 0, INT_MAX, VE }, |
169 | { "skip_empty_cb", "Avoid wasting bytes, ignore vintage MacOS decoder", OFFSET(skip_empty_cb), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE }, |
170 | { "max_strips", "Limit strips/frame, vintage compatible is 1..3, otherwise the more the better", OFFSET(max_max_strips), AV_OPT_TYPE_INT, { .i64 = 3 }, MIN_STRIPS, MAX_STRIPS, VE }, |
171 | { "min_strips", "Enforce min strips/frame, more is worse and faster, must be <= max_strips", OFFSET(min_min_strips), AV_OPT_TYPE_INT, { .i64 = MIN_STRIPS }, MIN_STRIPS, MAX_STRIPS, VE }, |
172 | { "strip_number_adaptivity", "How fast the strip number adapts, more is slightly better, much slower", OFFSET(strip_number_delta_range), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, MAX_STRIPS-MIN_STRIPS, VE }, |
173 | { NULL }, |
174 | }; |
175 | |
176 | static const AVClass cinepak_class = { |
177 | .class_name = "cinepak", |
178 | .item_name = av_default_item_name, |
179 | .option = options, |
180 | .version = LIBAVUTIL_VERSION_INT, |
181 | }; |
182 | |
183 | static av_cold int cinepak_encode_init(AVCodecContext *avctx) |
184 | { |
185 | CinepakEncContext *s = avctx->priv_data; |
186 | int x, mb_count, strip_buf_size, frame_buf_size; |
187 | |
188 | if (avctx->width & 3 || avctx->height & 3) { |
189 | av_log(avctx, AV_LOG_ERROR, "width and height must be multiples of four (got %ix%i)\n", |
190 | avctx->width, avctx->height); |
191 | return AVERROR(EINVAL); |
192 | } |
193 | |
194 | if (s->min_min_strips > s->max_max_strips) { |
195 | av_log(avctx, AV_LOG_ERROR, "minimal number of strips can not exceed maximal (got %i and %i)\n", |
196 | s->min_min_strips, s->max_max_strips); |
197 | return AVERROR(EINVAL); |
198 | } |
199 | |
200 | if (!(s->last_frame = av_frame_alloc())) |
201 | return AVERROR(ENOMEM); |
202 | if (!(s->best_frame = av_frame_alloc())) |
203 | goto enomem; |
204 | if (!(s->scratch_frame = av_frame_alloc())) |
205 | goto enomem; |
206 | if (avctx->pix_fmt == AV_PIX_FMT_RGB24) |
207 | if (!(s->input_frame = av_frame_alloc())) |
208 | goto enomem; |
209 | |
210 | if (!(s->codebook_input = av_malloc(sizeof(int) * (avctx->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4) * (avctx->width * avctx->height) >> 2))) |
211 | goto enomem; |
212 | |
213 | if (!(s->codebook_closest = av_malloc(sizeof(int) * (avctx->width * avctx->height) >> 2))) |
214 | goto enomem; |
215 | |
216 | for(x = 0; x < (avctx->pix_fmt == AV_PIX_FMT_RGB24 ? 4 : 3); x++) |
217 | if(!(s->pict_bufs[x] = av_malloc((avctx->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4) * (avctx->width * avctx->height) >> 2))) |
218 | goto enomem; |
219 | |
220 | mb_count = avctx->width * avctx->height / MB_AREA; |
221 | |
222 | //the largest possible chunk is 0x31 with all MBs encoded in V4 mode |
223 | //and full codebooks being replaced in INTER mode, |
224 | // which is 34 bits per MB |
225 | //and 2*256 extra flag bits per strip |
226 | strip_buf_size = STRIP_HEADER_SIZE + 3 * CHUNK_HEADER_SIZE + 2 * VECTOR_MAX * CODEBOOK_MAX + 4 * (mb_count + (mb_count + 15) / 16) + (2 * CODEBOOK_MAX)/8; |
227 | |
228 | frame_buf_size = CVID_HEADER_SIZE + s->max_max_strips * strip_buf_size; |
229 | |
230 | if (!(s->strip_buf = av_malloc(strip_buf_size))) |
231 | goto enomem; |
232 | |
233 | if (!(s->frame_buf = av_malloc(frame_buf_size))) |
234 | goto enomem; |
235 | |
236 | if (!(s->mb = av_malloc_array(mb_count, sizeof(mb_info)))) |
237 | goto enomem; |
238 | |
239 | #ifdef CINEPAKENC_DEBUG |
240 | if (!(s->best_mb = av_malloc_array(mb_count, sizeof(mb_info)))) |
241 | goto enomem; |
242 | #endif |
243 | |
244 | av_lfg_init(&s->randctx, 1); |
245 | s->avctx = avctx; |
246 | s->w = avctx->width; |
247 | s->h = avctx->height; |
248 | s->frame_buf_size = frame_buf_size; |
249 | s->curframe = 0; |
250 | s->keyint = avctx->keyint_min; |
251 | s->pix_fmt = avctx->pix_fmt; |
252 | |
253 | //set up AVFrames |
254 | s->last_frame->data[0] = s->pict_bufs[0]; |
255 | s->last_frame->linesize[0] = s->w; |
256 | s->best_frame->data[0] = s->pict_bufs[1]; |
257 | s->best_frame->linesize[0] = s->w; |
258 | s->scratch_frame->data[0] = s->pict_bufs[2]; |
259 | s->scratch_frame->linesize[0] = s->w; |
260 | |
261 | if (s->pix_fmt == AV_PIX_FMT_RGB24) { |
262 | s->last_frame->data[1] = s->last_frame->data[0] + s->w * s->h; |
263 | s->last_frame->data[2] = s->last_frame->data[1] + ((s->w * s->h) >> 2); |
264 | s->last_frame->linesize[1] = s->last_frame->linesize[2] = s->w >> 1; |
265 | |
266 | s->best_frame->data[1] = s->best_frame->data[0] + s->w * s->h; |
267 | s->best_frame->data[2] = s->best_frame->data[1] + ((s->w * s->h) >> 2); |
268 | s->best_frame->linesize[1] = s->best_frame->linesize[2] = s->w >> 1; |
269 | |
270 | s->scratch_frame->data[1] = s->scratch_frame->data[0] + s->w * s->h; |
271 | s->scratch_frame->data[2] = s->scratch_frame->data[1] + ((s->w * s->h) >> 2); |
272 | s->scratch_frame->linesize[1] = s->scratch_frame->linesize[2] = s->w >> 1; |
273 | |
274 | s->input_frame->data[0] = s->pict_bufs[3]; |
275 | s->input_frame->linesize[0] = s->w; |
276 | s->input_frame->data[1] = s->input_frame->data[0] + s->w * s->h; |
277 | s->input_frame->data[2] = s->input_frame->data[1] + ((s->w * s->h) >> 2); |
278 | s->input_frame->linesize[1] = s->input_frame->linesize[2] = s->w >> 1; |
279 | } |
280 | |
281 | s->min_strips = s->min_min_strips; |
282 | s->max_strips = s->max_max_strips; |
283 | |
284 | #ifdef CINEPAKENC_DEBUG |
285 | s->num_v1_mode = s->num_v4_mode = s->num_mc_mode = s->num_v1_encs = s->num_v4_encs = s->num_skips = 0; |
286 | #endif |
287 | |
288 | return 0; |
289 | |
290 | enomem: |
291 | av_frame_free(&s->last_frame); |
292 | av_frame_free(&s->best_frame); |
293 | av_frame_free(&s->scratch_frame); |
294 | if (avctx->pix_fmt == AV_PIX_FMT_RGB24) |
295 | av_frame_free(&s->input_frame); |
296 | av_freep(&s->codebook_input); |
297 | av_freep(&s->codebook_closest); |
298 | av_freep(&s->strip_buf); |
299 | av_freep(&s->frame_buf); |
300 | av_freep(&s->mb); |
301 | #ifdef CINEPAKENC_DEBUG |
302 | av_freep(&s->best_mb); |
303 | #endif |
304 | |
305 | for(x = 0; x < (avctx->pix_fmt == AV_PIX_FMT_RGB24 ? 4 : 3); x++) |
306 | av_freep(&s->pict_bufs[x]); |
307 | |
308 | return AVERROR(ENOMEM); |
309 | } |
310 | |
311 | static int64_t calculate_mode_score(CinepakEncContext *s, int h, strip_info *info, int report, int *training_set_v1_shrunk, int *training_set_v4_shrunk |
312 | #ifdef CINEPAK_REPORT_SERR |
313 | , int64_t *serr |
314 | #endif |
315 | ) |
316 | { |
317 | //score = FF_LAMBDA_SCALE * error + lambda * bits |
318 | int x; |
319 | int entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4; |
320 | int mb_count = s->w * h / MB_AREA; |
321 | mb_info *mb; |
322 | int64_t score1, score2, score3; |
323 | int64_t ret = s->lambda * ((info->v1_size ? CHUNK_HEADER_SIZE + info->v1_size * entry_size : 0) + |
324 | (info->v4_size ? CHUNK_HEADER_SIZE + info->v4_size * entry_size : 0) + |
325 | CHUNK_HEADER_SIZE) << 3; |
326 | |
327 | //av_log(s->avctx, AV_LOG_INFO, "sizes %3i %3i -> %9"PRId64" score mb_count %i", info->v1_size, info->v4_size, ret, mb_count); |
328 | |
329 | #ifdef CINEPAK_REPORT_SERR |
330 | *serr = 0; |
331 | #endif |
332 | |
333 | switch(info->mode) { |
334 | case MODE_V1_ONLY: |
335 | //one byte per MB |
336 | ret += s->lambda * 8 * mb_count; |
337 | |
338 | // while calculating we assume all blocks are ENC_V1 |
339 | for(x = 0; x < mb_count; x++) { |
340 | mb = &s->mb[x]; |
341 | ret += FF_LAMBDA_SCALE * mb->v1_error; |
342 | #ifdef CINEPAK_REPORT_SERR |
343 | *serr += mb->v1_error; |
344 | #endif |
345 | // this function is never called for report in MODE_V1_ONLY |
346 | // if(!report) |
347 | mb->best_encoding = ENC_V1; |
348 | } |
349 | |
350 | break; |
351 | case MODE_V1_V4: |
352 | //9 or 33 bits per MB |
353 | if(report) { |
354 | // no moves between the corresponding training sets are allowed |
355 | *training_set_v1_shrunk = *training_set_v4_shrunk = 0; |
356 | for(x = 0; x < mb_count; x++) { |
357 | int mberr; |
358 | mb = &s->mb[x]; |
359 | if(mb->best_encoding == ENC_V1) |
360 | score1 = s->lambda * 9 + FF_LAMBDA_SCALE * (mberr=mb->v1_error); |
361 | else |
362 | score1 = s->lambda * 33 + FF_LAMBDA_SCALE * (mberr=mb->v4_error); |
363 | ret += score1; |
364 | #ifdef CINEPAK_REPORT_SERR |
365 | *serr += mberr; |
366 | #endif |
367 | } |
368 | } else { // find best mode per block |
369 | for(x = 0; x < mb_count; x++) { |
370 | mb = &s->mb[x]; |
371 | score1 = s->lambda * 9 + FF_LAMBDA_SCALE * mb->v1_error; |
372 | score2 = s->lambda * 33 + FF_LAMBDA_SCALE * mb->v4_error; |
373 | |
374 | if(score1 <= score2) { |
375 | ret += score1; |
376 | #ifdef CINEPAK_REPORT_SERR |
377 | *serr += mb->v1_error; |
378 | #endif |
379 | mb->best_encoding = ENC_V1; |
380 | } else { |
381 | ret += score2; |
382 | #ifdef CINEPAK_REPORT_SERR |
383 | *serr += mb->v4_error; |
384 | #endif |
385 | mb->best_encoding = ENC_V4; |
386 | } |
387 | } |
388 | } |
389 | |
390 | break; |
391 | case MODE_MC: |
392 | //1, 10 or 34 bits per MB |
393 | if(report) { |
394 | int v1_shrunk = 0, v4_shrunk = 0; |
395 | for(x = 0; x < mb_count; x++) { |
396 | mb = &s->mb[x]; |
397 | // it is OK to move blocks to ENC_SKIP here |
398 | // but not to any codebook encoding! |
399 | score1 = s->lambda * 1 + FF_LAMBDA_SCALE * mb->skip_error; |
400 | if(mb->best_encoding == ENC_SKIP) { |
401 | ret += score1; |
402 | #ifdef CINEPAK_REPORT_SERR |
403 | *serr += mb->skip_error; |
404 | #endif |
405 | } else if(mb->best_encoding == ENC_V1) { |
406 | if((score2=s->lambda * 10 + FF_LAMBDA_SCALE * mb->v1_error) >= score1) { |
407 | mb->best_encoding = ENC_SKIP; |
408 | ++v1_shrunk; |
409 | ret += score1; |
410 | #ifdef CINEPAK_REPORT_SERR |
411 | *serr += mb->skip_error; |
412 | #endif |
413 | } else { |
414 | ret += score2; |
415 | #ifdef CINEPAK_REPORT_SERR |
416 | *serr += mb->v1_error; |
417 | #endif |
418 | } |
419 | } else { |
420 | if((score3=s->lambda * 34 + FF_LAMBDA_SCALE * mb->v4_error) >= score1) { |
421 | mb->best_encoding = ENC_SKIP; |
422 | ++v4_shrunk; |
423 | ret += score1; |
424 | #ifdef CINEPAK_REPORT_SERR |
425 | *serr += mb->skip_error; |
426 | #endif |
427 | } else { |
428 | ret += score3; |
429 | #ifdef CINEPAK_REPORT_SERR |
430 | *serr += mb->v4_error; |
431 | #endif |
432 | } |
433 | } |
434 | } |
435 | *training_set_v1_shrunk = v1_shrunk; |
436 | *training_set_v4_shrunk = v4_shrunk; |
437 | } else { // find best mode per block |
438 | for(x = 0; x < mb_count; x++) { |
439 | mb = &s->mb[x]; |
440 | score1 = s->lambda * 1 + FF_LAMBDA_SCALE * mb->skip_error; |
441 | score2 = s->lambda * 10 + FF_LAMBDA_SCALE * mb->v1_error; |
442 | score3 = s->lambda * 34 + FF_LAMBDA_SCALE * mb->v4_error; |
443 | |
444 | if(score1 <= score2 && score1 <= score3) { |
445 | ret += score1; |
446 | #ifdef CINEPAK_REPORT_SERR |
447 | *serr += mb->skip_error; |
448 | #endif |
449 | mb->best_encoding = ENC_SKIP; |
450 | } else if(score2 <= score3) { |
451 | ret += score2; |
452 | #ifdef CINEPAK_REPORT_SERR |
453 | *serr += mb->v1_error; |
454 | #endif |
455 | mb->best_encoding = ENC_V1; |
456 | } else { |
457 | ret += score3; |
458 | #ifdef CINEPAK_REPORT_SERR |
459 | *serr += mb->v4_error; |
460 | #endif |
461 | mb->best_encoding = ENC_V4; |
462 | } |
463 | } |
464 | } |
465 | |
466 | break; |
467 | } |
468 | |
469 | return ret; |
470 | } |
471 | |
472 | static int write_chunk_header(unsigned char *buf, int chunk_type, int chunk_size) |
473 | { |
474 | buf[0] = chunk_type; |
475 | AV_WB24(&buf[1], chunk_size + CHUNK_HEADER_SIZE); |
476 | return CHUNK_HEADER_SIZE; |
477 | } |
478 | |
479 | static int encode_codebook(CinepakEncContext *s, int *codebook, int size, int chunk_type_yuv, int chunk_type_gray, unsigned char *buf) |
480 | { |
481 | int x, y, ret, entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4; |
482 | int incremental_codebook_replacement_mode = 0; // hardcoded here, |
483 | // the compiler should notice that this is a constant -- rl |
484 | |
485 | ret = write_chunk_header(buf, |
486 | s->pix_fmt == AV_PIX_FMT_RGB24 ? |
487 | chunk_type_yuv+(incremental_codebook_replacement_mode?1:0) : |
488 | chunk_type_gray+(incremental_codebook_replacement_mode?1:0), |
489 | entry_size * size |
490 | + (incremental_codebook_replacement_mode?(size+31)/32*4:0) ); |
491 | |
492 | // we do codebook encoding according to the "intra" mode |
493 | // but we keep the "dead" code for reference in case we will want |
494 | // to use incremental codebook updates (which actually would give us |
495 | // "kind of" motion compensation, especially in 1 strip/frame case) -- rl |
496 | // (of course, the code will be not useful as-is) |
497 | if(incremental_codebook_replacement_mode) { |
498 | int flags = 0; |
499 | int flagsind; |
500 | for(x = 0; x < size; x++) { |
501 | if(flags == 0) { |
502 | flagsind = ret; |
503 | ret += 4; |
504 | flags = 0x80000000; |
505 | } else |
506 | flags = ((flags>>1) | 0x80000000); |
507 | for(y = 0; y < entry_size; y++) |
508 | buf[ret++] = codebook[y + x*entry_size] ^ (y >= 4 ? 0x80 : 0); |
509 | if((flags&0xffffffff) == 0xffffffff) { |
510 | AV_WB32(&buf[flagsind], flags); |
511 | flags = 0; |
512 | } |
513 | } |
514 | if(flags) |
515 | AV_WB32(&buf[flagsind], flags); |
516 | } else |
517 | for(x = 0; x < size; x++) |
518 | for(y = 0; y < entry_size; y++) |
519 | buf[ret++] = codebook[y + x*entry_size] ^ (y >= 4 ? 0x80 : 0); |
520 | |
521 | return ret; |
522 | } |
523 | |
524 | //sets out to the sub picture starting at (x,y) in in |
525 | static void get_sub_picture(CinepakEncContext *s, int x, int y, |
526 | uint8_t * in_data[4], int in_linesize[4], |
527 | uint8_t *out_data[4], int out_linesize[4]) |
528 | { |
529 | out_data[0] = in_data[0] + x + y * in_linesize[0]; |
530 | out_linesize[0] = in_linesize[0]; |
531 | |
532 | if(s->pix_fmt == AV_PIX_FMT_RGB24) { |
533 | out_data[1] = in_data[1] + (x >> 1) + (y >> 1) * in_linesize[1]; |
534 | out_linesize[1] = in_linesize[1]; |
535 | |
536 | out_data[2] = in_data[2] + (x >> 1) + (y >> 1) * in_linesize[2]; |
537 | out_linesize[2] = in_linesize[2]; |
538 | } |
539 | } |
540 | |
541 | //decodes the V1 vector in mb into the 4x4 MB pointed to by data |
542 | static void decode_v1_vector(CinepakEncContext *s, uint8_t *data[4], |
543 | int linesize[4], int v1_vector, strip_info *info) |
544 | { |
545 | int entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4; |
546 | |
547 | data[0][0] = |
548 | data[0][1] = |
549 | data[0][ linesize[0]] = |
550 | data[0][1+ linesize[0]] = info->v1_codebook[v1_vector*entry_size]; |
551 | |
552 | data[0][2] = |
553 | data[0][3] = |
554 | data[0][2+ linesize[0]] = |
555 | data[0][3+ linesize[0]] = info->v1_codebook[v1_vector*entry_size+1]; |
556 | |
557 | data[0][2*linesize[0]] = |
558 | data[0][1+2*linesize[0]] = |
559 | data[0][ 3*linesize[0]] = |
560 | data[0][1+3*linesize[0]] = info->v1_codebook[v1_vector*entry_size+2]; |
561 | |
562 | data[0][2+2*linesize[0]] = |
563 | data[0][3+2*linesize[0]] = |
564 | data[0][2+3*linesize[0]] = |
565 | data[0][3+3*linesize[0]] = info->v1_codebook[v1_vector*entry_size+3]; |
566 | |
567 | if(s->pix_fmt == AV_PIX_FMT_RGB24) { |
568 | data[1][0] = |
569 | data[1][1] = |
570 | data[1][ linesize[1]] = |
571 | data[1][1+ linesize[1]] = info->v1_codebook[v1_vector*entry_size+4]; |
572 | |
573 | data[2][0] = |
574 | data[2][1] = |
575 | data[2][ linesize[2]] = |
576 | data[2][1+ linesize[2]] = info->v1_codebook[v1_vector*entry_size+5]; |
577 | } |
578 | } |
579 | |
580 | //decodes the V4 vectors in mb into the 4x4 MB pointed to by data |
581 | static void decode_v4_vector(CinepakEncContext *s, uint8_t *data[4], |
582 | int linesize[4], int *v4_vector, strip_info *info) |
583 | { |
584 | int i, x, y, entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4; |
585 | |
586 | for(i = y = 0; y < 4; y += 2) { |
587 | for(x = 0; x < 4; x += 2, i++) { |
588 | data[0][x + y*linesize[0]] = info->v4_codebook[v4_vector[i]*entry_size]; |
589 | data[0][x+1 + y*linesize[0]] = info->v4_codebook[v4_vector[i]*entry_size+1]; |
590 | data[0][x + (y+1)*linesize[0]] = info->v4_codebook[v4_vector[i]*entry_size+2]; |
591 | data[0][x+1 + (y+1)*linesize[0]] = info->v4_codebook[v4_vector[i]*entry_size+3]; |
592 | |
593 | if(s->pix_fmt == AV_PIX_FMT_RGB24) { |
594 | data[1][(x>>1) + (y>>1)*linesize[1]] = info->v4_codebook[v4_vector[i]*entry_size+4]; |
595 | data[2][(x>>1) + (y>>1)*linesize[2]] = info->v4_codebook[v4_vector[i]*entry_size+5]; |
596 | } |
597 | } |
598 | } |
599 | } |
600 | |
601 | static void copy_mb(CinepakEncContext *s, |
602 | uint8_t *a_data[4], int a_linesize[4], |
603 | uint8_t *b_data[4], int b_linesize[4]) |
604 | { |
605 | int y, p; |
606 | |
607 | for(y = 0; y < MB_SIZE; y++) { |
608 | memcpy(a_data[0]+y*a_linesize[0], b_data[0]+y*b_linesize[0], |
609 | MB_SIZE); |
610 | } |
611 | |
612 | if(s->pix_fmt == AV_PIX_FMT_RGB24) { |
613 | for(p = 1; p <= 2; p++) { |
614 | for(y = 0; y < MB_SIZE/2; y++) { |
615 | memcpy(a_data[p] + y*a_linesize[p], |
616 | b_data[p] + y*b_linesize[p], |
617 | MB_SIZE/2); |
618 | } |
619 | } |
620 | } |
621 | } |
622 | |
623 | static int encode_mode(CinepakEncContext *s, int h, |
624 | uint8_t *scratch_data[4], int scratch_linesize[4], |
625 | uint8_t *last_data[4], int last_linesize[4], |
626 | strip_info *info, unsigned char *buf) |
627 | { |
628 | int x, y, z, flags, bits, temp_size, header_ofs, ret = 0, mb_count = s->w * h / MB_AREA; |
629 | int needs_extra_bit, should_write_temp; |
630 | unsigned char temp[64]; //32/2 = 16 V4 blocks at 4 B each -> 64 B |
631 | mb_info *mb; |
632 | uint8_t *sub_scratch_data[4] = {0}, *sub_last_data[4] = {0}; |
633 | int sub_scratch_linesize[4] = {0}, sub_last_linesize[4] = {0}; |
634 | |
635 | //encode codebooks |
636 | ////// MacOS vintage decoder compatibility dictates the presence of |
637 | ////// the codebook chunk even when the codebook is empty - pretty dumb... |
638 | ////// and also the certain order of the codebook chunks -- rl |
639 | if(info->v4_size || !s->skip_empty_cb) |
640 | ret += encode_codebook(s, info->v4_codebook, info->v4_size, 0x20, 0x24, buf + ret); |
641 | |
642 | if(info->v1_size || !s->skip_empty_cb) |
643 | ret += encode_codebook(s, info->v1_codebook, info->v1_size, 0x22, 0x26, buf + ret); |
644 | |
645 | //update scratch picture |
646 | for(z = y = 0; y < h; y += MB_SIZE) { |
647 | for(x = 0; x < s->w; x += MB_SIZE, z++) { |
648 | mb = &s->mb[z]; |
649 | |
650 | get_sub_picture(s, x, y, scratch_data, scratch_linesize, |
651 | sub_scratch_data, sub_scratch_linesize); |
652 | |
653 | if(info->mode == MODE_MC && mb->best_encoding == ENC_SKIP) { |
654 | get_sub_picture(s, x, y, |
655 | last_data, last_linesize, |
656 | sub_last_data, sub_last_linesize); |
657 | copy_mb(s, sub_scratch_data, sub_scratch_linesize, |
658 | sub_last_data, sub_last_linesize); |
659 | } else if(info->mode == MODE_V1_ONLY || mb->best_encoding == ENC_V1) |
660 | decode_v1_vector(s, sub_scratch_data, sub_scratch_linesize, |
661 | mb->v1_vector, info); |
662 | else |
663 | decode_v4_vector(s, sub_scratch_data, sub_scratch_linesize, |
664 | mb->v4_vector, info); |
665 | } |
666 | } |
667 | |
668 | switch(info->mode) { |
669 | case MODE_V1_ONLY: |
670 | //av_log(s->avctx, AV_LOG_INFO, "mb_count = %i\n", mb_count); |
671 | ret += write_chunk_header(buf + ret, 0x32, mb_count); |
672 | |
673 | for(x = 0; x < mb_count; x++) |
674 | buf[ret++] = s->mb[x].v1_vector; |
675 | |
676 | break; |
677 | case MODE_V1_V4: |
678 | //remember header position |
679 | header_ofs = ret; |
680 | ret += CHUNK_HEADER_SIZE; |
681 | |
682 | for(x = 0; x < mb_count; x += 32) { |
683 | flags = 0; |
684 | for(y = x; y < FFMIN(x+32, mb_count); y++) |
685 | if(s->mb[y].best_encoding == ENC_V4) |
686 | flags |= 1 << (31 - y + x); |
687 | |
688 | AV_WB32(&buf[ret], flags); |
689 | ret += 4; |
690 | |
691 | for(y = x; y < FFMIN(x+32, mb_count); y++) { |
692 | mb = &s->mb[y]; |
693 | |
694 | if(mb->best_encoding == ENC_V1) |
695 | buf[ret++] = mb->v1_vector; |
696 | else |
697 | for(z = 0; z < 4; z++) |
698 | buf[ret++] = mb->v4_vector[z]; |
699 | } |
700 | } |
701 | |
702 | write_chunk_header(buf + header_ofs, 0x30, ret - header_ofs - CHUNK_HEADER_SIZE); |
703 | |
704 | break; |
705 | case MODE_MC: |
706 | //remember header position |
707 | header_ofs = ret; |
708 | ret += CHUNK_HEADER_SIZE; |
709 | flags = bits = temp_size = 0; |
710 | |
711 | for(x = 0; x < mb_count; x++) { |
712 | mb = &s->mb[x]; |
713 | flags |= (mb->best_encoding != ENC_SKIP) << (31 - bits++); |
714 | needs_extra_bit = 0; |
715 | should_write_temp = 0; |
716 | |
717 | if(mb->best_encoding != ENC_SKIP) { |
718 | if(bits < 32) |
719 | flags |= (mb->best_encoding == ENC_V4) << (31 - bits++); |
720 | else |
721 | needs_extra_bit = 1; |
722 | } |
723 | |
724 | if(bits == 32) { |
725 | AV_WB32(&buf[ret], flags); |
726 | ret += 4; |
727 | flags = bits = 0; |
728 | |
729 | if(mb->best_encoding == ENC_SKIP || needs_extra_bit) { |
730 | memcpy(&buf[ret], temp, temp_size); |
731 | ret += temp_size; |
732 | temp_size = 0; |
733 | } else |
734 | should_write_temp = 1; |
735 | } |
736 | |
737 | if(needs_extra_bit) { |
738 | flags = (mb->best_encoding == ENC_V4) << 31; |
739 | bits = 1; |
740 | } |
741 | |
742 | if(mb->best_encoding == ENC_V1) |
743 | temp[temp_size++] = mb->v1_vector; |
744 | else if(mb->best_encoding == ENC_V4) |
745 | for(z = 0; z < 4; z++) |
746 | temp[temp_size++] = mb->v4_vector[z]; |
747 | |
748 | if(should_write_temp) { |
749 | memcpy(&buf[ret], temp, temp_size); |
750 | ret += temp_size; |
751 | temp_size = 0; |
752 | } |
753 | } |
754 | |
755 | if(bits > 0) { |
756 | AV_WB32(&buf[ret], flags); |
757 | ret += 4; |
758 | memcpy(&buf[ret], temp, temp_size); |
759 | ret += temp_size; |
760 | } |
761 | |
762 | write_chunk_header(buf + header_ofs, 0x31, ret - header_ofs - CHUNK_HEADER_SIZE); |
763 | |
764 | break; |
765 | } |
766 | |
767 | return ret; |
768 | } |
769 | |
770 | //computes distortion of 4x4 MB in b compared to a |
771 | static int compute_mb_distortion(CinepakEncContext *s, |
772 | uint8_t *a_data[4], int a_linesize[4], |
773 | uint8_t *b_data[4], int b_linesize[4]) |
774 | { |
775 | int x, y, p, d, ret = 0; |
776 | |
777 | for(y = 0; y < MB_SIZE; y++) { |
778 | for(x = 0; x < MB_SIZE; x++) { |
779 | d = a_data[0][x + y*a_linesize[0]] - b_data[0][x + y*b_linesize[0]]; |
780 | ret += d*d; |
781 | } |
782 | } |
783 | |
784 | if(s->pix_fmt == AV_PIX_FMT_RGB24) { |
785 | for(p = 1; p <= 2; p++) { |
786 | for(y = 0; y < MB_SIZE/2; y++) { |
787 | for(x = 0; x < MB_SIZE/2; x++) { |
788 | d = a_data[p][x + y*a_linesize[p]] - b_data[p][x + y*b_linesize[p]]; |
789 | ret += d*d; |
790 | } |
791 | } |
792 | } |
793 | } |
794 | |
795 | return ret; |
796 | } |
797 | |
798 | // return the possibly adjusted size of the codebook |
799 | #define CERTAIN(x) ((x)!=ENC_UNCERTAIN) |
800 | static int quantize(CinepakEncContext *s, int h, |
801 | uint8_t *data[4], int linesize[4], |
802 | int v1mode, strip_info *info, |
803 | mb_encoding encoding) |
804 | { |
805 | int x, y, i, j, k, x2, y2, x3, y3, plane, shift, mbn; |
806 | int entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4; |
807 | int *codebook = v1mode ? info->v1_codebook : info->v4_codebook; |
808 | int size = v1mode ? info->v1_size : info->v4_size; |
809 | int64_t total_error = 0; |
810 | uint8_t vq_pict_buf[(MB_AREA*3)/2]; |
811 | uint8_t *sub_data [4], *vq_data [4]; |
812 | int sub_linesize[4], vq_linesize[4]; |
813 | |
814 | for(mbn = i = y = 0; y < h; y += MB_SIZE) { |
815 | for(x = 0; x < s->w; x += MB_SIZE, ++mbn) { |
816 | int *base; |
817 | |
818 | if(CERTAIN(encoding)) { |
819 | // use for the training only the blocks known to be to be encoded [sic:-] |
820 | if(s->mb[mbn].best_encoding != encoding) continue; |
821 | } |
822 | |
823 | base = s->codebook_input + i*entry_size; |
824 | if(v1mode) { |
825 | //subsample |
826 | for(j = y2 = 0; y2 < entry_size; y2 += 2) { |
827 | for(x2 = 0; x2 < 4; x2 += 2, j++) { |
828 | plane = y2 < 4 ? 0 : 1 + (x2 >> 1); |
829 | shift = y2 < 4 ? 0 : 1; |
830 | x3 = shift ? 0 : x2; |
831 | y3 = shift ? 0 : y2; |
832 | base[j] = (data[plane][((x+x3) >> shift) + ((y+y3) >> shift) * linesize[plane]] + |
833 | data[plane][((x+x3) >> shift) + 1 + ((y+y3) >> shift) * linesize[plane]] + |
834 | data[plane][((x+x3) >> shift) + (((y+y3) >> shift) + 1) * linesize[plane]] + |
835 | data[plane][((x+x3) >> shift) + 1 + (((y+y3) >> shift) + 1) * linesize[plane]]) >> 2; |
836 | } |
837 | } |
838 | } else { |
839 | //copy |
840 | for(j = y2 = 0; y2 < MB_SIZE; y2 += 2) { |
841 | for(x2 = 0; x2 < MB_SIZE; x2 += 2) { |
842 | for(k = 0; k < entry_size; k++, j++) { |
843 | plane = k >= 4 ? k - 3 : 0; |
844 | |
845 | if(k >= 4) { |
846 | x3 = (x+x2) >> 1; |
847 | y3 = (y+y2) >> 1; |
848 | } else { |
849 | x3 = x + x2 + (k & 1); |
850 | y3 = y + y2 + (k >> 1); |
851 | } |
852 | |
853 | base[j] = data[plane][x3 + y3*linesize[plane]]; |
854 | } |
855 | } |
856 | } |
857 | } |
858 | i += v1mode ? 1 : 4; |
859 | } |
860 | } |
861 | // if(i < mbn*(v1mode ? 1 : 4)) { |
862 | // av_log(s->avctx, AV_LOG_INFO, "reducing training set for %s from %i to %i (encoding %i)\n", v1mode?"v1":"v4", mbn*(v1mode ? 1 : 4), i, encoding); |
863 | // } |
864 | |
865 | if(i == 0) // empty training set, nothing to do |
866 | return 0; |
867 | if(i < size) { |
868 | //av_log(s->avctx, (CERTAIN(encoding) ? AV_LOG_ERROR : AV_LOG_INFO), "WOULD WASTE: %s cbsize %i bigger than training set size %i (encoding %i)\n", v1mode?"v1":"v4", size, i, encoding); |
869 | size = i; |
870 | } |
871 | |
872 | avpriv_init_elbg(s->codebook_input, entry_size, i, codebook, size, 1, s->codebook_closest, &s->randctx); |
873 | avpriv_do_elbg(s->codebook_input, entry_size, i, codebook, size, 1, s->codebook_closest, &s->randctx); |
874 | |
875 | //setup vq_data, which contains a single MB |
876 | vq_data[0] = vq_pict_buf; |
877 | vq_linesize[0] = MB_SIZE; |
878 | vq_data[1] = &vq_pict_buf[MB_AREA]; |
879 | vq_data[2] = vq_data[1] + (MB_AREA >> 2); |
880 | vq_linesize[1] = vq_linesize[2] = MB_SIZE >> 1; |
881 | |
882 | //copy indices |
883 | for(i = j = y = 0; y < h; y += MB_SIZE) { |
884 | for(x = 0; x < s->w; x += MB_SIZE, j++) { |
885 | mb_info *mb = &s->mb[j]; |
886 | // skip uninteresting blocks if we know their preferred encoding |
887 | if(CERTAIN(encoding) && mb->best_encoding != encoding) |
888 | continue; |
889 | |
890 | //point sub_data to current MB |
891 | get_sub_picture(s, x, y, data, linesize, sub_data, sub_linesize); |
892 | |
893 | if(v1mode) { |
894 | mb->v1_vector = s->codebook_closest[i]; |
895 | |
896 | //fill in vq_data with V1 data |
897 | decode_v1_vector(s, vq_data, vq_linesize, mb->v1_vector, info); |
898 | |
899 | mb->v1_error = compute_mb_distortion(s, sub_data, sub_linesize, |
900 | vq_data, vq_linesize); |
901 | total_error += mb->v1_error; |
902 | } else { |
903 | for(k = 0; k < 4; k++) |
904 | mb->v4_vector[k] = s->codebook_closest[i+k]; |
905 | |
906 | //fill in vq_data with V4 data |
907 | decode_v4_vector(s, vq_data, vq_linesize, mb->v4_vector, info); |
908 | |
909 | mb->v4_error = compute_mb_distortion(s, sub_data, sub_linesize, |
910 | vq_data, vq_linesize); |
911 | total_error += mb->v4_error; |
912 | } |
913 | i += v1mode ? 1 : 4; |
914 | } |
915 | } |
916 | // check that we did it right in the beginning of the function |
917 | av_assert0(i >= size); // training set is no smaller than the codebook |
918 | |
919 | //av_log(s->avctx, AV_LOG_INFO, "isv1 %i size= %i i= %i error %"PRId64"\n", v1mode, size, i, total_error); |
920 | |
921 | return size; |
922 | } |
923 | |
924 | static void calculate_skip_errors(CinepakEncContext *s, int h, |
925 | uint8_t *last_data[4], int last_linesize[4], |
926 | uint8_t *data[4], int linesize[4], |
927 | strip_info *info) |
928 | { |
929 | int x, y, i; |
930 | uint8_t *sub_last_data [4], *sub_pict_data [4]; |
931 | int sub_last_linesize[4], sub_pict_linesize[4]; |
932 | |
933 | for(i = y = 0; y < h; y += MB_SIZE) { |
934 | for(x = 0; x < s->w; x += MB_SIZE, i++) { |
935 | get_sub_picture(s, x, y, last_data, last_linesize, |
936 | sub_last_data, sub_last_linesize); |
937 | get_sub_picture(s, x, y, data, linesize, |
938 | sub_pict_data, sub_pict_linesize); |
939 | |
940 | s->mb[i].skip_error = compute_mb_distortion(s, |
941 | sub_last_data, sub_last_linesize, |
942 | sub_pict_data, sub_pict_linesize); |
943 | } |
944 | } |
945 | } |
946 | |
947 | static void write_strip_header(CinepakEncContext *s, int y, int h, int keyframe, unsigned char *buf, int strip_size) |
948 | { |
949 | // actually we are exclusively using intra strip coding (how much can we win |
950 | // otherwise? how to choose which part of a codebook to update?), |
951 | // keyframes are different only because we disallow ENC_SKIP on them -- rl |
952 | // (besides, the logic here used to be inverted: ) |
953 | // buf[0] = keyframe ? 0x11: 0x10; |
954 | buf[0] = keyframe ? 0x10: 0x11; |
955 | AV_WB24(&buf[1], strip_size + STRIP_HEADER_SIZE); |
956 | // AV_WB16(&buf[4], y); /* using absolute y values works -- rl */ |
957 | AV_WB16(&buf[4], 0); /* using relative values works as well -- rl */ |
958 | AV_WB16(&buf[6], 0); |
959 | // AV_WB16(&buf[8], y+h); /* using absolute y values works -- rl */ |
960 | AV_WB16(&buf[8], h); /* using relative values works as well -- rl */ |
961 | AV_WB16(&buf[10], s->w); |
962 | //av_log(s->avctx, AV_LOG_INFO, "write_strip_header() %x keyframe=%d\n", buf[0], keyframe); |
963 | } |
964 | |
965 | static int rd_strip(CinepakEncContext *s, int y, int h, int keyframe, |
966 | uint8_t *last_data[4], int last_linesize[4], |
967 | uint8_t *data[4], int linesize[4], |
968 | uint8_t *scratch_data[4], int scratch_linesize[4], |
969 | unsigned char *buf, int64_t *best_score |
970 | #ifdef CINEPAK_REPORT_SERR |
971 | , int64_t *best_serr |
972 | #endif |
973 | ) |
974 | { |
975 | int64_t score = 0; |
976 | #ifdef CINEPAK_REPORT_SERR |
977 | int64_t serr; |
978 | #endif |
979 | int best_size = 0; |
980 | strip_info info; |
981 | // for codebook optimization: |
982 | int v1enough, v1_size, v4enough, v4_size; |
983 | int new_v1_size, new_v4_size; |
984 | int v1shrunk, v4shrunk; |
985 | |
986 | if(!keyframe) |
987 | calculate_skip_errors(s, h, last_data, last_linesize, data, linesize, |
988 | &info); |
989 | |
990 | //try some powers of 4 for the size of the codebooks |
991 | //constraint the v4 codebook to be no bigger than v1 one, |
992 | //(and no less than v1_size/4) |
993 | //thus making v1 preferable and possibly losing small details? should be ok |
994 | #define SMALLEST_CODEBOOK 1 |
995 | for(v1enough = 0, v1_size = SMALLEST_CODEBOOK; v1_size <= CODEBOOK_MAX && !v1enough; v1_size <<= 2) { |
996 | for(v4enough = 0, v4_size = 0; v4_size <= v1_size && !v4enough; v4_size = v4_size ? v4_size << 2 : v1_size >= SMALLEST_CODEBOOK << 2 ? v1_size >> 2 : SMALLEST_CODEBOOK) { |
997 | //try all modes |
998 | for(CinepakMode mode = 0; mode < MODE_COUNT; mode++) { |
999 | //don't allow MODE_MC in intra frames |
1000 | if(keyframe && mode == MODE_MC) |
1001 | continue; |
1002 | |
1003 | if(mode == MODE_V1_ONLY) { |
1004 | info.v1_size = v1_size; |
1005 | // the size may shrink even before optimizations if the input is short: |
1006 | info.v1_size = quantize(s, h, data, linesize, 1, |
1007 | &info, ENC_UNCERTAIN); |
1008 | if(info.v1_size < v1_size) |
1009 | // too few eligible blocks, no sense in trying bigger sizes |
1010 | v1enough = 1; |
1011 | |
1012 | info.v4_size = 0; |
1013 | } else { // mode != MODE_V1_ONLY |
1014 | // if v4 codebook is empty then only allow V1-only mode |
1015 | if(!v4_size) |
1016 | continue; |
1017 | |
1018 | if(mode == MODE_V1_V4) { |
1019 | info.v4_size = v4_size; |
1020 | info.v4_size = quantize(s, h, data, linesize, 0, |
1021 | &info, ENC_UNCERTAIN); |
1022 | if(info.v4_size < v4_size) |
1023 | // too few eligible blocks, no sense in trying bigger sizes |
1024 | v4enough = 1; |
1025 | } |
1026 | } |
1027 | |
1028 | info.mode = mode; |
1029 | // choose the best encoding per block, based on current experience |
1030 | score = calculate_mode_score(s, h, &info, 0, |
1031 | &v1shrunk, &v4shrunk |
1032 | #ifdef CINEPAK_REPORT_SERR |
1033 | , &serr |
1034 | #endif |
1035 | ); |
1036 | |
1037 | if(mode != MODE_V1_ONLY){ |
1038 | int extra_iterations_limit = s->max_extra_cb_iterations; |
1039 | // recompute the codebooks, omitting the extra blocks |
1040 | // we assume we _may_ come here with more blocks to encode than before |
1041 | info.v1_size = v1_size; |
1042 | new_v1_size = quantize(s, h, data, linesize, 1, &info, ENC_V1); |
1043 | if(new_v1_size < info.v1_size){ |
1044 | //av_log(s->avctx, AV_LOG_INFO, "mode %i, %3i, %3i: cut v1 codebook to %i entries\n", mode, v1_size, v4_size, new_v1_size); |
1045 | info.v1_size = new_v1_size; |
1046 | } |
1047 | // we assume we _may_ come here with more blocks to encode than before |
1048 | info.v4_size = v4_size; |
1049 | new_v4_size = quantize(s, h, data, linesize, 0, &info, ENC_V4); |
1050 | if(new_v4_size < info.v4_size) { |
1051 | //av_log(s->avctx, AV_LOG_INFO, "mode %i, %3i, %3i: cut v4 codebook to %i entries at first iteration\n", mode, v1_size, v4_size, new_v4_size); |
1052 | info.v4_size = new_v4_size; |
1053 | } |
1054 | // calculate the resulting score |
1055 | // (do not move blocks to codebook encodings now, as some blocks may have |
1056 | // got bigger errors despite a smaller training set - but we do not |
1057 | // ever grow the training sets back) |
1058 | for(;;) { |
1059 | score = calculate_mode_score(s, h, &info, 1, |
1060 | &v1shrunk, &v4shrunk |
1061 | #ifdef CINEPAK_REPORT_SERR |
1062 | , &serr |
1063 | #endif |
1064 | ); |
1065 | // do we have a reason to reiterate? if so, have we reached the limit? |
1066 | if((!v1shrunk && !v4shrunk) || !extra_iterations_limit--) break; |
1067 | // recompute the codebooks, omitting the extra blocks |
1068 | if(v1shrunk) { |
1069 | info.v1_size = v1_size; |
1070 | new_v1_size = quantize(s, h, data, linesize, 1, &info, ENC_V1); |
1071 | if(new_v1_size < info.v1_size){ |
1072 | //av_log(s->avctx, AV_LOG_INFO, "mode %i, %3i, %3i: cut v1 codebook to %i entries\n", mode, v1_size, v4_size, new_v1_size); |
1073 | info.v1_size = new_v1_size; |
1074 | } |
1075 | } |
1076 | if(v4shrunk) { |
1077 | info.v4_size = v4_size; |
1078 | new_v4_size = quantize(s, h, data, linesize, 0, &info, ENC_V4); |
1079 | if(new_v4_size < info.v4_size) { |
1080 | //av_log(s->avctx, AV_LOG_INFO, "mode %i, %3i, %3i: cut v4 codebook to %i entries\n", mode, v1_size, v4_size, new_v4_size); |
1081 | info.v4_size = new_v4_size; |
1082 | } |
1083 | } |
1084 | } |
1085 | } |
1086 | |
1087 | //av_log(s->avctx, AV_LOG_INFO, "%3i %3i score = %"PRId64"\n", v1_size, v4_size, score); |
1088 | |
1089 | if(best_size == 0 || score < *best_score) { |
1090 | |
1091 | *best_score = score; |
1092 | #ifdef CINEPAK_REPORT_SERR |
1093 | *best_serr = serr; |
1094 | #endif |
1095 | best_size = encode_mode(s, h, |
1096 | scratch_data, scratch_linesize, |
1097 | last_data, last_linesize, &info, |
1098 | s->strip_buf + STRIP_HEADER_SIZE); |
1099 | |
1100 | //av_log(s->avctx, AV_LOG_INFO, "mode %i, %3i, %3i: %18"PRId64" %i B", mode, info.v1_size, info.v4_size, score, best_size); |
1101 | //av_log(s->avctx, AV_LOG_INFO, "\n"); |
1102 | #ifdef CINEPAK_REPORT_SERR |
1103 | av_log(s->avctx, AV_LOG_INFO, "mode %i, %3i, %3i: %18"PRId64" %i B\n", mode, v1_size, v4_size, serr, best_size); |
1104 | #endif |
1105 | |
1106 | #ifdef CINEPAKENC_DEBUG |
1107 | //save MB encoding choices |
1108 | memcpy(s->best_mb, s->mb, mb_count*sizeof(mb_info)); |
1109 | #endif |
1110 | |
1111 | //memcpy(strip_temp + STRIP_HEADER_SIZE, strip_temp, best_size); |
1112 | write_strip_header(s, y, h, keyframe, s->strip_buf, best_size); |
1113 | |
1114 | } |
1115 | } |
1116 | } |
1117 | } |
1118 | |
1119 | #ifdef CINEPAKENC_DEBUG |
1120 | //gather stats. this will only work properly of MAX_STRIPS == 1 |
1121 | if(best_info.mode == MODE_V1_ONLY) { |
1122 | s->num_v1_mode++; |
1123 | s->num_v1_encs += s->w*h/MB_AREA; |
1124 | } else { |
1125 | if(best_info.mode == MODE_V1_V4) |
1126 | s->num_v4_mode++; |
1127 | else |
1128 | s->num_mc_mode++; |
1129 | |
1130 | int x; |
1131 | for(x = 0; x < s->w*h/MB_AREA; x++) |
1132 | if(s->best_mb[x].best_encoding == ENC_V1) |
1133 | s->num_v1_encs++; |
1134 | else if(s->best_mb[x].best_encoding == ENC_V4) |
1135 | s->num_v4_encs++; |
1136 | else |
1137 | s->num_skips++; |
1138 | } |
1139 | #endif |
1140 | |
1141 | best_size += STRIP_HEADER_SIZE; |
1142 | memcpy(buf, s->strip_buf, best_size); |
1143 | |
1144 | return best_size; |
1145 | } |
1146 | |
1147 | static int write_cvid_header(CinepakEncContext *s, unsigned char *buf, int num_strips, int data_size, int isakeyframe) |
1148 | { |
1149 | buf[0] = isakeyframe ? 0 : 1; |
1150 | AV_WB24(&buf[1], data_size + CVID_HEADER_SIZE); |
1151 | AV_WB16(&buf[4], s->w); |
1152 | AV_WB16(&buf[6], s->h); |
1153 | AV_WB16(&buf[8], num_strips); |
1154 | |
1155 | return CVID_HEADER_SIZE; |
1156 | } |
1157 | |
1158 | static int rd_frame(CinepakEncContext *s, const AVFrame *frame, |
1159 | int isakeyframe, unsigned char *buf, int buf_size) |
1160 | { |
1161 | int num_strips, strip, i, y, nexty, size, temp_size; |
1162 | uint8_t *last_data [4], *data [4], *scratch_data [4]; |
1163 | int last_linesize[4], linesize[4], scratch_linesize[4]; |
1164 | int64_t best_score = 0, score, score_temp; |
1165 | #ifdef CINEPAK_REPORT_SERR |
1166 | int64_t best_serr = 0, serr, serr_temp; |
1167 | #endif |
1168 | |
1169 | int best_nstrips = -1, best_size = -1; // mark as uninitialzed |
1170 | |
1171 | if(s->pix_fmt == AV_PIX_FMT_RGB24) { |
1172 | int x; |
1173 | // build a copy of the given frame in the correct colorspace |
1174 | for(y = 0; y < s->h; y += 2) { |
1175 | for(x = 0; x < s->w; x += 2) { |
1176 | uint8_t *ir[2]; int32_t r, g, b, rr, gg, bb; |
1177 | ir[0] = frame->data[0] + x*3 + y*frame->linesize[0]; |
1178 | ir[1] = ir[0] + frame->linesize[0]; |
1179 | get_sub_picture(s, x, y, |
1180 | s->input_frame->data, s->input_frame->linesize, |
1181 | scratch_data, scratch_linesize); |
1182 | r = g = b = 0; |
1183 | for(i=0; i<4; ++i) { |
1184 | int i1, i2; |
1185 | i1 = (i&1); i2 = (i>=2); |
1186 | rr = ir[i2][i1*3+0]; |
1187 | gg = ir[i2][i1*3+1]; |
1188 | bb = ir[i2][i1*3+2]; |
1189 | r += rr; g += gg; b += bb; |
1190 | // using fixed point arithmetic for portable repeatability, scaling by 2^23 |
1191 | // "Y" |
1192 | // rr = 0.2857*rr + 0.5714*gg + 0.1429*bb; |
1193 | rr = (2396625*rr + 4793251*gg + 1198732*bb) >> 23; |
1194 | if( rr < 0) rr = 0; |
1195 | else if (rr > 255) rr = 255; |
1196 | scratch_data[0][i1 + i2*scratch_linesize[0]] = rr; |
1197 | } |
1198 | // let us scale down as late as possible |
1199 | // r /= 4; g /= 4; b /= 4; |
1200 | // "U" |
1201 | // rr = -0.1429*r - 0.2857*g + 0.4286*b; |
1202 | rr = (-299683*r - 599156*g + 898839*b) >> 23; |
1203 | if( rr < -128) rr = -128; |
1204 | else if (rr > 127) rr = 127; |
1205 | scratch_data[1][0] = rr + 128; // quantize needs unsigned |
1206 | // "V" |
1207 | // rr = 0.3571*r - 0.2857*g - 0.0714*b; |
1208 | rr = (748893*r - 599156*g - 149737*b) >> 23; |
1209 | if( rr < -128) rr = -128; |
1210 | else if (rr > 127) rr = 127; |
1211 | scratch_data[2][0] = rr + 128; // quantize needs unsigned |
1212 | } |
1213 | } |
1214 | } |
1215 | |
1216 | //would be nice but quite certainly incompatible with vintage players: |
1217 | // support encoding zero strips (meaning skip the whole frame) |
1218 | for(num_strips = s->min_strips; num_strips <= s->max_strips && num_strips <= s->h / MB_SIZE; num_strips++) { |
1219 | score = 0; |
1220 | size = 0; |
1221 | #ifdef CINEPAK_REPORT_SERR |
1222 | serr = 0; |
1223 | #endif |
1224 | |
1225 | for(y = 0, strip = 1; y < s->h; strip++, y = nexty) { |
1226 | int strip_height; |
1227 | |
1228 | nexty = strip * s->h / num_strips; // <= s->h |
1229 | //make nexty the next multiple of 4 if not already there |
1230 | if(nexty & 3) |
1231 | nexty += 4 - (nexty & 3); |
1232 | |
1233 | strip_height = nexty - y; |
1234 | if(strip_height <= 0) { // can this ever happen? |
1235 | av_log(s->avctx, AV_LOG_INFO, "skipping zero height strip %i of %i\n", strip, num_strips); |
1236 | continue; |
1237 | } |
1238 | |
1239 | if(s->pix_fmt == AV_PIX_FMT_RGB24) |
1240 | get_sub_picture(s, 0, y, |
1241 | s->input_frame->data, s->input_frame->linesize, |
1242 | data, linesize); |
1243 | else |
1244 | get_sub_picture(s, 0, y, |
1245 | (uint8_t **)frame->data, (int*)frame->linesize, |
1246 | data, linesize); |
1247 | get_sub_picture(s, 0, y, |
1248 | s->last_frame->data, s->last_frame->linesize, |
1249 | last_data, last_linesize); |
1250 | get_sub_picture(s, 0, y, |
1251 | s->scratch_frame->data, s->scratch_frame->linesize, |
1252 | scratch_data, scratch_linesize); |
1253 | |
1254 | if((temp_size = rd_strip(s, y, strip_height, isakeyframe, |
1255 | last_data, last_linesize, data, linesize, |
1256 | scratch_data, scratch_linesize, |
1257 | s->frame_buf + size + CVID_HEADER_SIZE, &score_temp |
1258 | #ifdef CINEPAK_REPORT_SERR |
1259 | , &serr_temp |
1260 | #endif |
1261 | )) < 0) |
1262 | return temp_size; |
1263 | |
1264 | score += score_temp; |
1265 | #ifdef CINEPAK_REPORT_SERR |
1266 | serr += serr_temp; |
1267 | #endif |
1268 | size += temp_size; |
1269 | //av_log(s->avctx, AV_LOG_INFO, "strip %d, isakeyframe=%d", strip, isakeyframe); |
1270 | //av_log(s->avctx, AV_LOG_INFO, "\n"); |
1271 | } |
1272 | |
1273 | if(best_score == 0 || score < best_score) { |
1274 | best_score = score; |
1275 | #ifdef CINEPAK_REPORT_SERR |
1276 | best_serr = serr; |
1277 | #endif |
1278 | best_size = size + write_cvid_header(s, s->frame_buf, num_strips, size, isakeyframe); |
1279 | //av_log(s->avctx, AV_LOG_INFO, "best number of strips so far: %2i, %12"PRId64", %i B\n", num_strips, score, best_size); |
1280 | #ifdef CINEPAK_REPORT_SERR |
1281 | av_log(s->avctx, AV_LOG_INFO, "best number of strips so far: %2i, %12"PRId64", %i B\n", num_strips, serr, best_size); |
1282 | #endif |
1283 | |
1284 | FFSWAP(AVFrame *, s->best_frame, s->scratch_frame); |
1285 | memcpy(buf, s->frame_buf, best_size); |
1286 | best_nstrips = num_strips; |
1287 | } |
1288 | // avoid trying too many strip numbers without a real reason |
1289 | // (this makes the processing of the very first frame faster) |
1290 | if(num_strips - best_nstrips > 4) |
1291 | break; |
1292 | } |
1293 | |
1294 | av_assert0(best_nstrips >= 0 && best_size >= 0); |
1295 | |
1296 | // let the number of strips slowly adapt to the changes in the contents, |
1297 | // compared to full bruteforcing every time this will occasionally lead |
1298 | // to some r/d performance loss but makes encoding up to several times faster |
1299 | if(!s->strip_number_delta_range) { |
1300 | if(best_nstrips == s->max_strips) { // let us try to step up |
1301 | s->max_strips = best_nstrips + 1; |
1302 | if(s->max_strips >= s->max_max_strips) |
1303 | s->max_strips = s->max_max_strips; |
1304 | } else { // try to step down |
1305 | s->max_strips = best_nstrips; |
1306 | } |
1307 | s->min_strips = s->max_strips - 1; |
1308 | if(s->min_strips < s->min_min_strips) |
1309 | s->min_strips = s->min_min_strips; |
1310 | } else { |
1311 | s->max_strips = best_nstrips + s->strip_number_delta_range; |
1312 | if(s->max_strips >= s->max_max_strips) |
1313 | s->max_strips = s->max_max_strips; |
1314 | s->min_strips = best_nstrips - s->strip_number_delta_range; |
1315 | if(s->min_strips < s->min_min_strips) |
1316 | s->min_strips = s->min_min_strips; |
1317 | } |
1318 | |
1319 | return best_size; |
1320 | } |
1321 | |
1322 | static int cinepak_encode_frame(AVCodecContext *avctx, AVPacket *pkt, |
1323 | const AVFrame *frame, int *got_packet) |
1324 | { |
1325 | CinepakEncContext *s = avctx->priv_data; |
1326 | int ret; |
1327 | |
1328 | s->lambda = frame->quality ? frame->quality - 1 : 2 * FF_LAMBDA_SCALE; |
1329 | |
1330 | if ((ret = ff_alloc_packet2(avctx, pkt, s->frame_buf_size, 0)) < 0) |
1331 | return ret; |
1332 | ret = rd_frame(s, frame, (s->curframe == 0), pkt->data, s->frame_buf_size); |
1333 | pkt->size = ret; |
1334 | if (s->curframe == 0) |
1335 | pkt->flags |= AV_PKT_FLAG_KEY; |
1336 | *got_packet = 1; |
1337 | |
1338 | FFSWAP(AVFrame *, s->last_frame, s->best_frame); |
1339 | |
1340 | if (++s->curframe >= s->keyint) |
1341 | s->curframe = 0; |
1342 | |
1343 | return 0; |
1344 | } |
1345 | |
1346 | static av_cold int cinepak_encode_end(AVCodecContext *avctx) |
1347 | { |
1348 | CinepakEncContext *s = avctx->priv_data; |
1349 | int x; |
1350 | |
1351 | av_frame_free(&s->last_frame); |
1352 | av_frame_free(&s->best_frame); |
1353 | av_frame_free(&s->scratch_frame); |
1354 | if (avctx->pix_fmt == AV_PIX_FMT_RGB24) |
1355 | av_frame_free(&s->input_frame); |
1356 | av_freep(&s->codebook_input); |
1357 | av_freep(&s->codebook_closest); |
1358 | av_freep(&s->strip_buf); |
1359 | av_freep(&s->frame_buf); |
1360 | av_freep(&s->mb); |
1361 | #ifdef CINEPAKENC_DEBUG |
1362 | av_freep(&s->best_mb); |
1363 | #endif |
1364 | |
1365 | for(x = 0; x < (avctx->pix_fmt == AV_PIX_FMT_RGB24 ? 4 : 3); x++) |
1366 | av_freep(&s->pict_bufs[x]); |
1367 | |
1368 | #ifdef CINEPAKENC_DEBUG |
1369 | av_log(avctx, AV_LOG_INFO, "strip coding stats: %i V1 mode, %i V4 mode, %i MC mode (%i V1 encs, %i V4 encs, %i skips)\n", |
1370 | s->num_v1_mode, s->num_v4_mode, s->num_mc_mode, s->num_v1_encs, s->num_v4_encs, s->num_skips); |
1371 | #endif |
1372 | |
1373 | return 0; |
1374 | } |
1375 | |
1376 | AVCodec ff_cinepak_encoder = { |
1377 | .name = "cinepak", |
1378 | .type = AVMEDIA_TYPE_VIDEO, |
1379 | .id = AV_CODEC_ID_CINEPAK, |
1380 | .priv_data_size = sizeof(CinepakEncContext), |
1381 | .init = cinepak_encode_init, |
1382 | .encode2 = cinepak_encode_frame, |
1383 | .close = cinepak_encode_end, |
1384 | .pix_fmts = (const enum AVPixelFormat[]){AV_PIX_FMT_RGB24, AV_PIX_FMT_GRAY8, AV_PIX_FMT_NONE}, |
1385 | .long_name = NULL_IF_CONFIG_SMALL("Cinepak"), |
1386 | .priv_class = &cinepak_class, |
1387 | }; |
1388 |