2 * Cinepak encoder (c) 2011 Tomas Härdin
3 * http://titan.codemill.se/~tomhar/cinepakenc.patch
5 * Fixes and improvements, vintage decoders compatibility
6 * (c) 2013, 2014 Rl, Aetey Global Technologies AB
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:
15 * The above copyright notice and this permission notice shall be included
16 * in all copies or substantial portions of the Software.
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.
29 * - optimize: color space conversion (move conversion to libswscale), ...
31 * - "optimally" split the frame into several non-regular areas
32 * using a separate codebook pair for each area and approximating
33 * the area by several rectangular strips (generally not full width ones)
34 * (use quadtree splitting? a simple fixed-granularity grid?)
39 #include "libavutil/avassert.h"
40 #include "libavutil/common.h"
41 #include "libavutil/internal.h"
42 #include "libavutil/intreadwrite.h"
43 #include "libavutil/lfg.h"
44 #include "libavutil/opt.h"
47 #include "codec_internal.h"
51 #define CVID_HEADER_SIZE 10
52 #define STRIP_HEADER_SIZE 12
53 #define CHUNK_HEADER_SIZE 4
55 #define MB_SIZE 4 //4x4 MBs
56 #define MB_AREA (MB_SIZE * MB_SIZE)
58 #define VECTOR_MAX 6 // six or four entries per vector depending on format
59 #define CODEBOOK_MAX 256 // size of a codebook
61 #define MAX_STRIPS 32 // Note: having fewer choices regarding the number of strips speeds up encoding (obviously)
62 #define MIN_STRIPS 1 // Note: having more strips speeds up encoding the frame (this is less obvious)
63 // MAX_STRIPS limits the maximum quality you can reach
64 // when you want high quality on high resolutions,
65 // MIN_STRIPS limits the minimum efficiently encodable bit rate
67 // the numbers are only used for brute force optimization for the first frame,
68 // for the following frames they are adaptively readjusted
69 // NOTE the decoder in ffmpeg has its own arbitrary limitation on the number
70 // of strips, currently 32
72 typedef enum CinepakMode {
80 typedef enum mb_encoding {
88 typedef struct mb_info {
89 int v1_vector; // index into v1 codebook
90 int v1_error; // error when using V1 encoding
91 int v4_vector[4]; // indices into v4 codebook
92 int v4_error; // error when using V4 encoding
93 int skip_error; // error when block is skipped (aka copied from last frame)
94 mb_encoding best_encoding; // last result from calculate_mode_score()
97 typedef struct strip_info {
98 int v1_codebook[CODEBOOK_MAX * VECTOR_MAX];
99 int v4_codebook[CODEBOOK_MAX * VECTOR_MAX];
105 typedef struct CinepakEncContext {
106 const AVClass *class;
107 AVCodecContext *avctx;
108 unsigned char *pict_bufs[4], *strip_buf, *frame_buf;
111 AVFrame *scratch_frame;
112 AVFrame *input_frame;
113 enum AVPixelFormat pix_fmt;
120 int *codebook_closest;
121 mb_info *mb; // MB RD state
122 int min_strips; // the current limit
123 int max_strips; // the current limit
125 int max_extra_cb_iterations;
129 int strip_number_delta_range;
130 struct ELBGContext *elbg;
133 #define OFFSET(x) offsetof(CinepakEncContext, x)
134 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
135 static const AVOption options[] = {
136 { "max_extra_cb_iterations", "Max extra codebook recalculation passes, more is better and slower",
137 OFFSET(max_extra_cb_iterations), AV_OPT_TYPE_INT, { .i64 = 2 }, 0, INT_MAX, VE },
138 { "skip_empty_cb", "Avoid wasting bytes, ignore vintage MacOS decoder",
139 OFFSET(skip_empty_cb), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE },
140 { "max_strips", "Limit strips/frame, vintage compatible is 1..3, otherwise the more the better",
141 OFFSET(max_max_strips), AV_OPT_TYPE_INT, { .i64 = 3 }, MIN_STRIPS, MAX_STRIPS, VE },
142 { "min_strips", "Enforce min strips/frame, more is worse and faster, must be <= max_strips",
143 OFFSET(min_min_strips), AV_OPT_TYPE_INT, { .i64 = MIN_STRIPS }, MIN_STRIPS, MAX_STRIPS, VE },
144 { "strip_number_adaptivity", "How fast the strip number adapts, more is slightly better, much slower",
145 OFFSET(strip_number_delta_range), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, MAX_STRIPS - MIN_STRIPS, VE },
149 static const AVClass cinepak_class = {
150 .class_name = "cinepak",
151 .item_name = av_default_item_name,
153 .version = LIBAVUTIL_VERSION_INT,
156 static av_cold int cinepak_encode_init(AVCodecContext *avctx)
158 CinepakEncContext *s = avctx->priv_data;
159 int x, mb_count, strip_buf_size, frame_buf_size;
161 if (avctx->width & 3 || avctx->height & 3) {
162 av_log(avctx, AV_LOG_ERROR, "width and height must be multiples of four (got %ix%i)\n",
163 avctx->width, avctx->height);
164 return AVERROR(EINVAL);
167 if (s->min_min_strips > s->max_max_strips) {
168 av_log(avctx, AV_LOG_ERROR, "minimum number of strips must not exceed maximum (got %i and %i)\n",
169 s->min_min_strips, s->max_max_strips);
170 return AVERROR(EINVAL);
173 if (!(s->last_frame = av_frame_alloc()))
174 return AVERROR(ENOMEM);
175 if (!(s->best_frame = av_frame_alloc()))
176 return AVERROR(ENOMEM);
177 if (!(s->scratch_frame = av_frame_alloc()))
178 return AVERROR(ENOMEM);
179 if (avctx->pix_fmt == AV_PIX_FMT_RGB24)
180 if (!(s->input_frame = av_frame_alloc()))
181 return AVERROR(ENOMEM);
183 if (!(s->codebook_input = av_malloc_array((avctx->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4) * (avctx->width * avctx->height) >> 2, sizeof(*s->codebook_input))))
184 return AVERROR(ENOMEM);
186 if (!(s->codebook_closest = av_malloc_array((avctx->width * avctx->height) >> 2, sizeof(*s->codebook_closest))))
187 return AVERROR(ENOMEM);
189 for (x = 0; x < (avctx->pix_fmt == AV_PIX_FMT_RGB24 ? 4 : 3); x++)
190 if (!(s->pict_bufs[x] = av_malloc((avctx->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4) * (avctx->width * avctx->height) >> 2)))
191 return AVERROR(ENOMEM);
193 mb_count = avctx->width * avctx->height / MB_AREA;
195 // the largest possible chunk is 0x31 with all MBs encoded in V4 mode
196 // and full codebooks being replaced in INTER mode,
197 // which is 34 bits per MB
198 // and 2*256 extra flag bits per strip
199 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;
201 frame_buf_size = CVID_HEADER_SIZE + s->max_max_strips * strip_buf_size;
203 if (!(s->strip_buf = av_malloc(strip_buf_size)))
204 return AVERROR(ENOMEM);
206 if (!(s->frame_buf = av_malloc(frame_buf_size)))
207 return AVERROR(ENOMEM);
209 if (!(s->mb = av_malloc_array(mb_count, sizeof(mb_info))))
210 return AVERROR(ENOMEM);
212 av_lfg_init(&s->randctx, 1);
215 s->h = avctx->height;
216 s->frame_buf_size = frame_buf_size;
218 s->pix_fmt = avctx->pix_fmt;
221 s->last_frame->data[0] = s->pict_bufs[0];
222 s->last_frame->linesize[0] = s->w;
223 s->best_frame->data[0] = s->pict_bufs[1];
224 s->best_frame->linesize[0] = s->w;
225 s->scratch_frame->data[0] = s->pict_bufs[2];
226 s->scratch_frame->linesize[0] = s->w;
228 if (s->pix_fmt == AV_PIX_FMT_RGB24) {
229 s->last_frame->data[1] = s->last_frame->data[0] + s->w * s->h;
230 s->last_frame->data[2] = s->last_frame->data[1] + ((s->w * s->h) >> 2);
231 s->last_frame->linesize[1] =
232 s->last_frame->linesize[2] = s->w >> 1;
234 s->best_frame->data[1] = s->best_frame->data[0] + s->w * s->h;
235 s->best_frame->data[2] = s->best_frame->data[1] + ((s->w * s->h) >> 2);
236 s->best_frame->linesize[1] =
237 s->best_frame->linesize[2] = s->w >> 1;
239 s->scratch_frame->data[1] = s->scratch_frame->data[0] + s->w * s->h;
240 s->scratch_frame->data[2] = s->scratch_frame->data[1] + ((s->w * s->h) >> 2);
241 s->scratch_frame->linesize[1] =
242 s->scratch_frame->linesize[2] = s->w >> 1;
244 s->input_frame->data[0] = s->pict_bufs[3];
245 s->input_frame->linesize[0] = s->w;
246 s->input_frame->data[1] = s->input_frame->data[0] + s->w * s->h;
247 s->input_frame->data[2] = s->input_frame->data[1] + ((s->w * s->h) >> 2);
248 s->input_frame->linesize[1] =
249 s->input_frame->linesize[2] = s->w >> 1;
252 s->min_strips = s->min_min_strips;
253 s->max_strips = s->max_max_strips;
258 static int64_t calculate_mode_score(CinepakEncContext *s, int h,
259 strip_info *info, int report,
260 int *training_set_v1_shrunk,
261 int *training_set_v4_shrunk)
263 // score = FF_LAMBDA_SCALE * error + lambda * bits
265 int entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4;
266 int mb_count = s->w * h / MB_AREA;
268 int64_t score1, score2, score3;
269 int64_t ret = s->lambda * ((info->v1_size ? CHUNK_HEADER_SIZE + info->v1_size * entry_size : 0) +
270 (info->v4_size ? CHUNK_HEADER_SIZE + info->v4_size * entry_size : 0) +
271 CHUNK_HEADER_SIZE) << 3;
273 switch (info->mode) {
276 ret += s->lambda * 8 * mb_count;
278 // while calculating we assume all blocks are ENC_V1
279 for (x = 0; x < mb_count; x++) {
281 ret += FF_LAMBDA_SCALE * mb->v1_error;
282 // this function is never called for report in MODE_V1_ONLY
284 mb->best_encoding = ENC_V1;
289 // 9 or 33 bits per MB
291 // no moves between the corresponding training sets are allowed
292 *training_set_v1_shrunk = *training_set_v4_shrunk = 0;
293 for (x = 0; x < mb_count; x++) {
296 if (mb->best_encoding == ENC_V1)
297 score1 = s->lambda * 9 + FF_LAMBDA_SCALE * (mberr = mb->v1_error);
299 score1 = s->lambda * 33 + FF_LAMBDA_SCALE * (mberr = mb->v4_error);
302 } else { // find best mode per block
303 for (x = 0; x < mb_count; x++) {
305 score1 = s->lambda * 9 + FF_LAMBDA_SCALE * mb->v1_error;
306 score2 = s->lambda * 33 + FF_LAMBDA_SCALE * mb->v4_error;
308 if (score1 <= score2) {
310 mb->best_encoding = ENC_V1;
313 mb->best_encoding = ENC_V4;
320 // 1, 10 or 34 bits per MB
322 int v1_shrunk = 0, v4_shrunk = 0;
323 for (x = 0; x < mb_count; x++) {
325 // it is OK to move blocks to ENC_SKIP here
326 // but not to any codebook encoding!
327 score1 = s->lambda * 1 + FF_LAMBDA_SCALE * mb->skip_error;
328 if (mb->best_encoding == ENC_SKIP) {
330 } else if (mb->best_encoding == ENC_V1) {
331 if ((score2 = s->lambda * 10 + FF_LAMBDA_SCALE * mb->v1_error) >= score1) {
332 mb->best_encoding = ENC_SKIP;
339 if ((score3 = s->lambda * 34 + FF_LAMBDA_SCALE * mb->v4_error) >= score1) {
340 mb->best_encoding = ENC_SKIP;
348 *training_set_v1_shrunk = v1_shrunk;
349 *training_set_v4_shrunk = v4_shrunk;
350 } else { // find best mode per block
351 for (x = 0; x < mb_count; x++) {
353 score1 = s->lambda * 1 + FF_LAMBDA_SCALE * mb->skip_error;
354 score2 = s->lambda * 10 + FF_LAMBDA_SCALE * mb->v1_error;
355 score3 = s->lambda * 34 + FF_LAMBDA_SCALE * mb->v4_error;
357 if (score1 <= score2 && score1 <= score3) {
359 mb->best_encoding = ENC_SKIP;
360 } else if (score2 <= score3) {
362 mb->best_encoding = ENC_V1;
365 mb->best_encoding = ENC_V4;
376 static int write_chunk_header(unsigned char *buf, int chunk_type, int chunk_size)
379 AV_WB24(&buf[1], chunk_size + CHUNK_HEADER_SIZE);
380 return CHUNK_HEADER_SIZE;
383 static int encode_codebook(CinepakEncContext *s, int *codebook, int size,
384 int chunk_type_yuv, int chunk_type_gray,
387 int x, y, ret, entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4;
388 int incremental_codebook_replacement_mode = 0; // hardcoded here,
389 // the compiler should notice that this is a constant -- rl
391 ret = write_chunk_header(buf,
392 s->pix_fmt == AV_PIX_FMT_RGB24 ?
393 chunk_type_yuv + (incremental_codebook_replacement_mode ? 1 : 0) :
394 chunk_type_gray + (incremental_codebook_replacement_mode ? 1 : 0),
396 (incremental_codebook_replacement_mode ? (size + 31) / 32 * 4 : 0));
398 // we do codebook encoding according to the "intra" mode
399 // but we keep the "dead" code for reference in case we will want
400 // to use incremental codebook updates (which actually would give us
401 // "kind of" motion compensation, especially in 1 strip/frame case) -- rl
402 // (of course, the code will be not useful as-is)
403 if (incremental_codebook_replacement_mode) {
406 for (x = 0; x < size; x++) {
412 flags = ((flags >> 1) | 0x80000000);
413 for (y = 0; y < entry_size; y++)
414 buf[ret++] = codebook[y + x * entry_size] ^ (y >= 4 ? 0x80 : 0);
415 if ((flags & 0xffffffff) == 0xffffffff) {
416 AV_WB32(&buf[flagsind], flags);
421 AV_WB32(&buf[flagsind], flags);
423 for (x = 0; x < size; x++)
424 for (y = 0; y < entry_size; y++)
425 buf[ret++] = codebook[y + x * entry_size] ^ (y >= 4 ? 0x80 : 0);
430 // sets out to the sub picture starting at (x,y) in in
431 static void get_sub_picture(CinepakEncContext *s, int x, int y,
432 uint8_t *const in_data[4], const int in_linesize[4],
433 uint8_t *out_data[4], int out_linesize[4])
435 out_data[0] = in_data[0] + x + y * in_linesize[0];
436 out_linesize[0] = in_linesize[0];
438 if (s->pix_fmt == AV_PIX_FMT_RGB24) {
439 out_data[1] = in_data[1] + (x >> 1) + (y >> 1) * in_linesize[1];
440 out_linesize[1] = in_linesize[1];
442 out_data[2] = in_data[2] + (x >> 1) + (y >> 1) * in_linesize[2];
443 out_linesize[2] = in_linesize[2];
447 // decodes the V1 vector in mb into the 4x4 MB pointed to by data
448 static void decode_v1_vector(CinepakEncContext *s, uint8_t *data[4],
449 int linesize[4], int v1_vector, strip_info *info)
451 int entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4;
455 data[0][ linesize[0]] =
456 data[0][1 + linesize[0]] = info->v1_codebook[v1_vector * entry_size];
460 data[0][2 + linesize[0]] =
461 data[0][3 + linesize[0]] = info->v1_codebook[v1_vector * entry_size + 1];
463 data[0][ 2 * linesize[0]] =
464 data[0][1 + 2 * linesize[0]] =
465 data[0][ 3 * linesize[0]] =
466 data[0][1 + 3 * linesize[0]] = info->v1_codebook[v1_vector * entry_size + 2];
468 data[0][2 + 2 * linesize[0]] =
469 data[0][3 + 2 * linesize[0]] =
470 data[0][2 + 3 * linesize[0]] =
471 data[0][3 + 3 * linesize[0]] = info->v1_codebook[v1_vector * entry_size + 3];
473 if (s->pix_fmt == AV_PIX_FMT_RGB24) {
476 data[1][ linesize[1]] =
477 data[1][1 + linesize[1]] = info->v1_codebook[v1_vector * entry_size + 4];
481 data[2][ linesize[2]] =
482 data[2][1 + linesize[2]] = info->v1_codebook[v1_vector * entry_size + 5];
486 // decodes the V4 vectors in mb into the 4x4 MB pointed to by data
487 static void decode_v4_vector(CinepakEncContext *s, uint8_t *data[4],
488 int linesize[4], int *v4_vector, strip_info *info)
490 int i, x, y, entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4;
492 for (i = y = 0; y < 4; y += 2) {
493 for (x = 0; x < 4; x += 2, i++) {
494 data[0][x + y * linesize[0]] = info->v4_codebook[v4_vector[i] * entry_size];
495 data[0][x + 1 + y * linesize[0]] = info->v4_codebook[v4_vector[i] * entry_size + 1];
496 data[0][x + (y + 1) * linesize[0]] = info->v4_codebook[v4_vector[i] * entry_size + 2];
497 data[0][x + 1 + (y + 1) * linesize[0]] = info->v4_codebook[v4_vector[i] * entry_size + 3];
499 if (s->pix_fmt == AV_PIX_FMT_RGB24) {
500 data[1][(x >> 1) + (y >> 1) * linesize[1]] = info->v4_codebook[v4_vector[i] * entry_size + 4];
501 data[2][(x >> 1) + (y >> 1) * linesize[2]] = info->v4_codebook[v4_vector[i] * entry_size + 5];
507 static void copy_mb(CinepakEncContext *s,
508 uint8_t *a_data[4], int a_linesize[4],
509 uint8_t *b_data[4], int b_linesize[4])
513 for (y = 0; y < MB_SIZE; y++)
514 memcpy(a_data[0] + y * a_linesize[0], b_data[0] + y * b_linesize[0],
517 if (s->pix_fmt == AV_PIX_FMT_RGB24) {
518 for (p = 1; p <= 2; p++)
519 for (y = 0; y < MB_SIZE / 2; y++)
520 memcpy(a_data[p] + y * a_linesize[p],
521 b_data[p] + y * b_linesize[p],
526 static int encode_mode(CinepakEncContext *s, int h,
527 uint8_t *scratch_data[4], int scratch_linesize[4],
528 uint8_t *last_data[4], int last_linesize[4],
529 strip_info *info, unsigned char *buf)
531 int x, y, z, bits, temp_size, header_ofs, ret = 0, mb_count = s->w * h / MB_AREA;
532 int needs_extra_bit, should_write_temp;
534 unsigned char temp[64]; // 32/2 = 16 V4 blocks at 4 B each -> 64 B
536 uint8_t *sub_scratch_data[4] = { 0 }, *sub_last_data[4] = { 0 };
537 int sub_scratch_linesize[4] = { 0 }, sub_last_linesize[4] = { 0 };
540 ////// MacOS vintage decoder compatibility dictates the presence of
541 ////// the codebook chunk even when the codebook is empty - pretty dumb...
542 ////// and also the certain order of the codebook chunks -- rl
543 if (info->v4_size || !s->skip_empty_cb)
544 ret += encode_codebook(s, info->v4_codebook, info->v4_size, 0x20, 0x24, buf + ret);
546 if (info->v1_size || !s->skip_empty_cb)
547 ret += encode_codebook(s, info->v1_codebook, info->v1_size, 0x22, 0x26, buf + ret);
549 // update scratch picture
550 for (z = y = 0; y < h; y += MB_SIZE)
551 for (x = 0; x < s->w; x += MB_SIZE, z++) {
554 get_sub_picture(s, x, y, scratch_data, scratch_linesize,
555 sub_scratch_data, sub_scratch_linesize);
557 if (info->mode == MODE_MC && mb->best_encoding == ENC_SKIP) {
558 get_sub_picture(s, x, y, last_data, last_linesize,
559 sub_last_data, sub_last_linesize);
560 copy_mb(s, sub_scratch_data, sub_scratch_linesize,
561 sub_last_data, sub_last_linesize);
562 } else if (info->mode == MODE_V1_ONLY || mb->best_encoding == ENC_V1)
563 decode_v1_vector(s, sub_scratch_data, sub_scratch_linesize,
564 mb->v1_vector, info);
566 decode_v4_vector(s, sub_scratch_data, sub_scratch_linesize,
567 mb->v4_vector, info);
570 switch (info->mode) {
572 ret += write_chunk_header(buf + ret, 0x32, mb_count);
574 for (x = 0; x < mb_count; x++)
575 buf[ret++] = s->mb[x].v1_vector;
579 // remember header position
581 ret += CHUNK_HEADER_SIZE;
583 for (x = 0; x < mb_count; x += 32) {
585 for (y = x; y < FFMIN(x + 32, mb_count); y++)
586 if (s->mb[y].best_encoding == ENC_V4)
587 flags |= 1U << (31 - y + x);
589 AV_WB32(&buf[ret], flags);
592 for (y = x; y < FFMIN(x + 32, mb_count); y++) {
595 if (mb->best_encoding == ENC_V1)
596 buf[ret++] = mb->v1_vector;
598 for (z = 0; z < 4; z++)
599 buf[ret++] = mb->v4_vector[z];
603 write_chunk_header(buf + header_ofs, 0x30, ret - header_ofs - CHUNK_HEADER_SIZE);
607 // remember header position
609 ret += CHUNK_HEADER_SIZE;
610 flags = bits = temp_size = 0;
612 for (x = 0; x < mb_count; x++) {
614 flags |= (uint32_t)(mb->best_encoding != ENC_SKIP) << (31 - bits++);
616 should_write_temp = 0;
618 if (mb->best_encoding != ENC_SKIP) {
620 flags |= (uint32_t)(mb->best_encoding == ENC_V4) << (31 - bits++);
626 AV_WB32(&buf[ret], flags);
630 if (mb->best_encoding == ENC_SKIP || needs_extra_bit) {
631 memcpy(&buf[ret], temp, temp_size);
635 should_write_temp = 1;
638 if (needs_extra_bit) {
639 flags = (uint32_t)(mb->best_encoding == ENC_V4) << 31;
643 if (mb->best_encoding == ENC_V1)
644 temp[temp_size++] = mb->v1_vector;
645 else if (mb->best_encoding == ENC_V4)
646 for (z = 0; z < 4; z++)
647 temp[temp_size++] = mb->v4_vector[z];
649 if (should_write_temp) {
650 memcpy(&buf[ret], temp, temp_size);
657 AV_WB32(&buf[ret], flags);
659 memcpy(&buf[ret], temp, temp_size);
663 write_chunk_header(buf + header_ofs, 0x31, ret - header_ofs - CHUNK_HEADER_SIZE);
671 // computes distortion of 4x4 MB in b compared to a
672 static int compute_mb_distortion(CinepakEncContext *s,
673 uint8_t *a_data[4], int a_linesize[4],
674 uint8_t *b_data[4], int b_linesize[4])
676 int x, y, p, d, ret = 0;
678 for (y = 0; y < MB_SIZE; y++)
679 for (x = 0; x < MB_SIZE; x++) {
680 d = a_data[0][x + y * a_linesize[0]] - b_data[0][x + y * b_linesize[0]];
684 if (s->pix_fmt == AV_PIX_FMT_RGB24) {
685 for (p = 1; p <= 2; p++) {
686 for (y = 0; y < MB_SIZE / 2; y++)
687 for (x = 0; x < MB_SIZE / 2; x++) {
688 d = a_data[p][x + y * a_linesize[p]] - b_data[p][x + y * b_linesize[p]];
697 // return the possibly adjusted size of the codebook
698 #define CERTAIN(x) ((x) != ENC_UNCERTAIN)
699 static int quantize(CinepakEncContext *s, int h, uint8_t *data[4],
700 int linesize[4], int v1mode, strip_info *info,
701 mb_encoding encoding)
703 int x, y, i, j, k, x2, y2, x3, y3, plane, shift, mbn;
704 int entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4;
705 int *codebook = v1mode ? info->v1_codebook : info->v4_codebook;
706 int size = v1mode ? info->v1_size : info->v4_size;
707 uint8_t vq_pict_buf[(MB_AREA * 3) / 2];
708 uint8_t *sub_data[4], *vq_data[4];
709 int sub_linesize[4], vq_linesize[4];
712 for (mbn = i = y = 0; y < h; y += MB_SIZE) {
713 for (x = 0; x < s->w; x += MB_SIZE, ++mbn) {
716 if (CERTAIN(encoding)) {
717 // use for the training only the blocks known to be to be encoded [sic:-]
718 if (s->mb[mbn].best_encoding != encoding)
722 base = s->codebook_input + i * entry_size;
725 for (j = y2 = 0; y2 < entry_size; y2 += 2)
726 for (x2 = 0; x2 < 4; x2 += 2, j++) {
727 plane = y2 < 4 ? 0 : 1 + (x2 >> 1);
728 shift = y2 < 4 ? 0 : 1;
731 base[j] = (data[plane][((x + x3) >> shift) + ((y + y3) >> shift) * linesize[plane]] +
732 data[plane][((x + x3) >> shift) + 1 + ((y + y3) >> shift) * linesize[plane]] +
733 data[plane][((x + x3) >> shift) + (((y + y3) >> shift) + 1) * linesize[plane]] +
734 data[plane][((x + x3) >> shift) + 1 + (((y + y3) >> shift) + 1) * linesize[plane]]) >> 2;
738 for (j = y2 = 0; y2 < MB_SIZE; y2 += 2) {
739 for (x2 = 0; x2 < MB_SIZE; x2 += 2)
740 for (k = 0; k < entry_size; k++, j++) {
741 plane = k >= 4 ? k - 3 : 0;
747 x3 = x + x2 + (k & 1);
748 y3 = y + y2 + (k >> 1);
751 base[j] = data[plane][x3 + y3 * linesize[plane]];
759 if (i == 0) // empty training set, nothing to do
764 ret = avpriv_elbg_do(&s->elbg, s->codebook_input, entry_size, i, codebook,
765 size, 1, s->codebook_closest, &s->randctx, 0);
769 // set up vq_data, which contains a single MB
770 vq_data[0] = vq_pict_buf;
771 vq_linesize[0] = MB_SIZE;
772 vq_data[1] = &vq_pict_buf[MB_AREA];
773 vq_data[2] = vq_data[1] + (MB_AREA >> 2);
775 vq_linesize[2] = MB_SIZE >> 1;
778 for (i = j = y = 0; y < h; y += MB_SIZE)
779 for (x = 0; x < s->w; x += MB_SIZE, j++) {
780 mb_info *mb = &s->mb[j];
781 // skip uninteresting blocks if we know their preferred encoding
782 if (CERTAIN(encoding) && mb->best_encoding != encoding)
785 // point sub_data to current MB
786 get_sub_picture(s, x, y, data, linesize, sub_data, sub_linesize);
789 mb->v1_vector = s->codebook_closest[i];
791 // fill in vq_data with V1 data
792 decode_v1_vector(s, vq_data, vq_linesize, mb->v1_vector, info);
794 mb->v1_error = compute_mb_distortion(s, sub_data, sub_linesize,
795 vq_data, vq_linesize);
797 for (k = 0; k < 4; k++)
798 mb->v4_vector[k] = s->codebook_closest[i + k];
800 // fill in vq_data with V4 data
801 decode_v4_vector(s, vq_data, vq_linesize, mb->v4_vector, info);
803 mb->v4_error = compute_mb_distortion(s, sub_data, sub_linesize,
804 vq_data, vq_linesize);
808 // check that we did it right in the beginning of the function
809 av_assert0(i >= size); // training set is no smaller than the codebook
814 static void calculate_skip_errors(CinepakEncContext *s, int h,
815 uint8_t *last_data[4], int last_linesize[4],
816 uint8_t *data[4], int linesize[4],
820 uint8_t *sub_last_data [4], *sub_pict_data [4];
821 int sub_last_linesize[4], sub_pict_linesize[4];
823 for (i = y = 0; y < h; y += MB_SIZE)
824 for (x = 0; x < s->w; x += MB_SIZE, i++) {
825 get_sub_picture(s, x, y, last_data, last_linesize,
826 sub_last_data, sub_last_linesize);
827 get_sub_picture(s, x, y, data, linesize,
828 sub_pict_data, sub_pict_linesize);
830 s->mb[i].skip_error =
831 compute_mb_distortion(s,
832 sub_last_data, sub_last_linesize,
833 sub_pict_data, sub_pict_linesize);
837 static void write_strip_keyframe(unsigned char *buf, int keyframe)
839 // actually we are exclusively using intra strip coding (how much can we win
840 // otherwise? how to choose which part of a codebook to update?),
841 // keyframes are different only because we disallow ENC_SKIP on them -- rl
842 // (besides, the logic here used to be inverted: )
843 // buf[0] = keyframe ? 0x11: 0x10;
844 buf[0] = keyframe ? 0x10 : 0x11;
847 static void write_strip_header(CinepakEncContext *s, int y, int h, int keyframe,
848 unsigned char *buf, int strip_size)
850 write_strip_keyframe(buf, keyframe);
851 AV_WB24(&buf[1], strip_size + STRIP_HEADER_SIZE);
852 // AV_WB16(&buf[4], y); /* using absolute y values works -- rl */
853 AV_WB16(&buf[4], 0); /* using relative values works as well -- rl */
855 // AV_WB16(&buf[8], y + h); /* using absolute y values works -- rl */
856 AV_WB16(&buf[8], h); /* using relative values works as well -- rl */
857 AV_WB16(&buf[10], s->w);
860 static int rd_strip(CinepakEncContext *s, int y, int h, int keyframe,
861 uint8_t *last_data[4], int last_linesize[4],
862 uint8_t *data[4], int linesize[4],
863 uint8_t *scratch_data[4], int scratch_linesize[4],
864 unsigned char *buf, int64_t *best_score, int *no_skip)
869 // for codebook optimization:
870 int v1enough, v1_size, v4enough, v4_size;
871 int new_v1_size, new_v4_size;
872 int v1shrunk, v4shrunk;
875 calculate_skip_errors(s, h, last_data, last_linesize, data, linesize,
878 // try some powers of 4 for the size of the codebooks
879 // constraint the v4 codebook to be no bigger than v1 one,
880 // (and no less than v1_size/4)
881 // thus making v1 preferable and possibly losing small details? should be ok
882 #define SMALLEST_CODEBOOK 1
883 for (v1enough = 0, v1_size = SMALLEST_CODEBOOK; v1_size <= CODEBOOK_MAX && !v1enough; v1_size <<= 2) {
884 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) {
887 for (mode = 0; mode < MODE_COUNT; mode++) {
888 // don't allow MODE_MC in intra frames
889 if (keyframe && mode == MODE_MC)
892 if (mode == MODE_V1_ONLY) {
893 info.v1_size = v1_size;
894 // the size may shrink even before optimizations if the input is short:
895 if ((new_v1_size = quantize(s, h, data, linesize, 1,
896 &info, ENC_UNCERTAIN)) < 0)
898 info.v1_size = new_v1_size;
899 if (info.v1_size < v1_size)
900 // too few eligible blocks, no sense in trying bigger sizes
904 } else { // mode != MODE_V1_ONLY
905 // if v4 codebook is empty then only allow V1-only mode
909 if (mode == MODE_V1_V4) {
910 info.v4_size = v4_size;
911 new_v4_size = quantize(s, h, data, linesize, 0,
912 &info, ENC_UNCERTAIN);
915 info.v4_size = new_v4_size;
916 if (info.v4_size < v4_size)
917 // too few eligible blocks, no sense in trying bigger sizes
923 // choose the best encoding per block, based on current experience
924 score = calculate_mode_score(s, h, &info, 0,
925 &v1shrunk, &v4shrunk);
927 if (mode != MODE_V1_ONLY) {
928 int extra_iterations_limit = s->max_extra_cb_iterations;
929 // recompute the codebooks, omitting the extra blocks
930 // we assume we _may_ come here with more blocks to encode than before
931 info.v1_size = v1_size;
932 new_v1_size = quantize(s, h, data, linesize, 1, &info, ENC_V1);
935 if (new_v1_size < info.v1_size)
936 info.v1_size = new_v1_size;
937 // we assume we _may_ come here with more blocks to encode than before
938 info.v4_size = v4_size;
939 new_v4_size = quantize(s, h, data, linesize, 0, &info, ENC_V4);
942 if (new_v4_size < info.v4_size)
943 info.v4_size = new_v4_size;
944 // calculate the resulting score
945 // (do not move blocks to codebook encodings now, as some blocks may have
946 // got bigger errors despite a smaller training set - but we do not
947 // ever grow the training sets back)
949 score = calculate_mode_score(s, h, &info, 1,
950 &v1shrunk, &v4shrunk);
951 // do we have a reason to reiterate? if so, have we reached the limit?
952 if ((!v1shrunk && !v4shrunk) || !extra_iterations_limit--)
954 // recompute the codebooks, omitting the extra blocks
956 info.v1_size = v1_size;
957 new_v1_size = quantize(s, h, data, linesize, 1, &info, ENC_V1);
960 if (new_v1_size < info.v1_size)
961 info.v1_size = new_v1_size;
964 info.v4_size = v4_size;
965 new_v4_size = quantize(s, h, data, linesize, 0, &info, ENC_V4);
968 if (new_v4_size < info.v4_size)
969 info.v4_size = new_v4_size;
974 if (best_size == 0 || score < *best_score) {
976 best_size = encode_mode(s, h,
977 scratch_data, scratch_linesize,
978 last_data, last_linesize, &info,
979 s->strip_buf + STRIP_HEADER_SIZE);
980 // in theory we could have MODE_MC without ENC_SKIP,
981 // but MODE_V1_V4 will always be more efficient
982 *no_skip = info.mode != MODE_MC;
984 write_strip_header(s, y, h, keyframe, s->strip_buf, best_size);
990 best_size += STRIP_HEADER_SIZE;
991 memcpy(buf, s->strip_buf, best_size);
996 static int write_cvid_header(CinepakEncContext *s, unsigned char *buf,
997 int num_strips, int data_size, int isakeyframe)
999 buf[0] = isakeyframe ? 0 : 1;
1000 AV_WB24(&buf[1], data_size + CVID_HEADER_SIZE);
1001 AV_WB16(&buf[4], s->w);
1002 AV_WB16(&buf[6], s->h);
1003 AV_WB16(&buf[8], num_strips);
1005 return CVID_HEADER_SIZE;
1008 static int rd_frame(CinepakEncContext *s, const AVFrame *frame,
1009 int isakeyframe, unsigned char *buf, int buf_size, int *got_keyframe)
1011 int num_strips, strip, i, y, nexty, size, temp_size, best_size;
1012 uint8_t *last_data [4], *data [4], *scratch_data [4];
1013 int last_linesize[4], linesize[4], scratch_linesize[4];
1014 int64_t best_score = 0, score, score_temp;
1015 int best_nstrips, best_strip_offsets[MAX_STRIPS];
1017 if (s->pix_fmt == AV_PIX_FMT_RGB24) {
1019 // build a copy of the given frame in the correct colorspace
1020 for (y = 0; y < s->h; y += 2)
1021 for (x = 0; x < s->w; x += 2) {
1022 const uint8_t *ir[2];
1023 int32_t r, g, b, rr, gg, bb;
1024 ir[0] = frame->data[0] + x * 3 + y * frame->linesize[0];
1025 ir[1] = ir[0] + frame->linesize[0];
1026 get_sub_picture(s, x, y,
1027 s->input_frame->data, s->input_frame->linesize,
1028 scratch_data, scratch_linesize);
1030 for (i = 0; i < 4; ++i) {
1034 rr = ir[i2][i1 * 3 + 0];
1035 gg = ir[i2][i1 * 3 + 1];
1036 bb = ir[i2][i1 * 3 + 2];
1040 // using fixed point arithmetic for portable repeatability, scaling by 2^23
1042 // rr = 0.2857 * rr + 0.5714 * gg + 0.1429 * bb;
1043 rr = (2396625 * rr + 4793251 * gg + 1198732 * bb) >> 23;
1048 scratch_data[0][i1 + i2 * scratch_linesize[0]] = rr;
1050 // let us scale down as late as possible
1051 // r /= 4; g /= 4; b /= 4;
1053 // rr = -0.1429 * r - 0.2857 * g + 0.4286 * b;
1054 rr = (-299683 * r - 599156 * g + 898839 * b) >> 23;
1059 scratch_data[1][0] = rr + 128; // quantize needs unsigned
1061 // rr = 0.3571 * r - 0.2857 * g - 0.0714 * b;
1062 rr = (748893 * r - 599156 * g - 149737 * b) >> 23;
1067 scratch_data[2][0] = rr + 128; // quantize needs unsigned
1071 // would be nice but quite certainly incompatible with vintage players:
1072 // support encoding zero strips (meaning skip the whole frame)
1073 for (num_strips = s->min_strips; num_strips <= s->max_strips && num_strips <= s->h / MB_SIZE; num_strips++) {
1074 int strip_offsets[MAX_STRIPS];
1075 int all_no_skip = 1;
1079 for (y = 0, strip = 1; y < s->h; strip++, y = nexty) {
1080 int strip_height, no_skip;
1082 strip_offsets[strip-1] = size + CVID_HEADER_SIZE;
1083 nexty = strip * s->h / num_strips; // <= s->h
1084 // make nexty the next multiple of 4 if not already there
1086 nexty += 4 - (nexty & 3);
1088 strip_height = nexty - y;
1089 if (strip_height <= 0) { // can this ever happen?
1090 av_log(s->avctx, AV_LOG_INFO, "skipping zero height strip %i of %i\n", strip, num_strips);
1094 if (s->pix_fmt == AV_PIX_FMT_RGB24)
1095 get_sub_picture(s, 0, y,
1096 s->input_frame->data, s->input_frame->linesize,
1099 get_sub_picture(s, 0, y,
1100 frame->data, frame->linesize,
1102 get_sub_picture(s, 0, y,
1103 s->last_frame->data, s->last_frame->linesize,
1104 last_data, last_linesize);
1105 get_sub_picture(s, 0, y,
1106 s->scratch_frame->data, s->scratch_frame->linesize,
1107 scratch_data, scratch_linesize);
1109 if ((temp_size = rd_strip(s, y, strip_height, isakeyframe,
1110 last_data, last_linesize, data, linesize,
1111 scratch_data, scratch_linesize,
1112 s->frame_buf + strip_offsets[strip-1],
1113 &score_temp, &no_skip)) < 0)
1116 score += score_temp;
1118 all_no_skip &= no_skip;
1121 if (best_score == 0 || score < best_score) {
1123 best_size = size + write_cvid_header(s, s->frame_buf, num_strips, size, all_no_skip);
1125 FFSWAP(AVFrame *, s->best_frame, s->scratch_frame);
1126 memcpy(buf, s->frame_buf, best_size);
1127 best_nstrips = num_strips;
1128 *got_keyframe = all_no_skip; // no skip MBs in any strip -> keyframe
1129 memcpy(best_strip_offsets, strip_offsets, sizeof(strip_offsets));
1131 // avoid trying too many strip numbers without a real reason
1132 // (this makes the processing of the very first frame faster)
1133 if (num_strips - best_nstrips > 4)
1137 // update strip headers
1138 for (i = 0; i < best_nstrips; i++) {
1139 write_strip_keyframe(s->frame_buf + best_strip_offsets[i], *got_keyframe);
1142 // let the number of strips slowly adapt to the changes in the contents,
1143 // compared to full bruteforcing every time this will occasionally lead
1144 // to some r/d performance loss but makes encoding up to several times faster
1145 if (!s->strip_number_delta_range) {
1146 if (best_nstrips == s->max_strips) { // let us try to step up
1147 s->max_strips = best_nstrips + 1;
1148 if (s->max_strips >= s->max_max_strips)
1149 s->max_strips = s->max_max_strips;
1150 } else { // try to step down
1151 s->max_strips = best_nstrips;
1153 s->min_strips = s->max_strips - 1;
1154 if (s->min_strips < s->min_min_strips)
1155 s->min_strips = s->min_min_strips;
1157 s->max_strips = best_nstrips + s->strip_number_delta_range;
1158 if (s->max_strips >= s->max_max_strips)
1159 s->max_strips = s->max_max_strips;
1160 s->min_strips = best_nstrips - s->strip_number_delta_range;
1161 if (s->min_strips < s->min_min_strips)
1162 s->min_strips = s->min_min_strips;
1168 static int cinepak_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1169 const AVFrame *frame, int *got_packet)
1171 CinepakEncContext *s = avctx->priv_data;
1172 int ret, got_keyframe;
1174 s->lambda = frame->quality ? frame->quality - 1 : 2 * FF_LAMBDA_SCALE;
1176 if ((ret = ff_alloc_packet(avctx, pkt, s->frame_buf_size)) < 0)
1178 ret = rd_frame(s, frame, (s->curframe == 0), pkt->data, s->frame_buf_size, &got_keyframe);
1181 pkt->flags |= AV_PKT_FLAG_KEY;
1186 FFSWAP(AVFrame *, s->last_frame, s->best_frame);
1188 if (++s->curframe >= avctx->gop_size)
1194 static av_cold int cinepak_encode_end(AVCodecContext *avctx)
1196 CinepakEncContext *s = avctx->priv_data;
1199 avpriv_elbg_free(&s->elbg);
1200 av_frame_free(&s->last_frame);
1201 av_frame_free(&s->best_frame);
1202 av_frame_free(&s->scratch_frame);
1203 if (avctx->pix_fmt == AV_PIX_FMT_RGB24)
1204 av_frame_free(&s->input_frame);
1205 av_freep(&s->codebook_input);
1206 av_freep(&s->codebook_closest);
1207 av_freep(&s->strip_buf);
1208 av_freep(&s->frame_buf);
1211 for (x = 0; x < (avctx->pix_fmt == AV_PIX_FMT_RGB24 ? 4 : 3); x++)
1212 av_freep(&s->pict_bufs[x]);
1217 const FFCodec ff_cinepak_encoder = {
1218 .p.name = "cinepak",
1219 CODEC_LONG_NAME("Cinepak"),
1220 .p.type = AVMEDIA_TYPE_VIDEO,
1221 .p.id = AV_CODEC_ID_CINEPAK,
1222 .p.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_ENCODER_REORDERED_OPAQUE,
1223 .priv_data_size = sizeof(CinepakEncContext),
1224 .init = cinepak_encode_init,
1225 FF_CODEC_ENCODE_CB(cinepak_encode_frame),
1226 .close = cinepak_encode_end,
1227 .p.pix_fmts = (const enum AVPixelFormat[]) { AV_PIX_FMT_RGB24, AV_PIX_FMT_GRAY8, AV_PIX_FMT_NONE },
1228 .p.priv_class = &cinepak_class,
1229 .caps_internal = FF_CODEC_CAP_INIT_CLEANUP,