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 
27 /*
28  * TODO:
29  * - optimize: color space conversion (move conversion to libswscale), ...
30  * MAYBE:
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?)
35  */
36 
37 #include <string.h>
38 
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"
45 
46 #include "avcodec.h"
47 #include "elbg.h"
48 #include "internal.h"
49 
50 #define CVID_HEADER_SIZE 10
51 #define STRIP_HEADER_SIZE 12
52 #define CHUNK_HEADER_SIZE 4
53 
54 #define MB_SIZE 4           //4x4 MBs
55 #define MB_AREA (MB_SIZE * MB_SIZE)
56 
57 #define VECTOR_MAX     6    // six or four entries per vector depending on format
58 #define CODEBOOK_MAX 256    // size of a codebook
59 
60 #define MAX_STRIPS  32      // Note: having fewer choices regarding the number of strips speeds up encoding (obviously)
61 #define MIN_STRIPS   1      // Note: having more strips speeds up encoding the frame (this is less obvious)
62 // MAX_STRIPS limits the maximum quality you can reach
63 //            when you want high quality on high resolutions,
64 // MIN_STRIPS limits the minimum efficiently encodable bit rate
65 //            on low resolutions
66 // the numbers are only used for brute force optimization for the first frame,
67 // for the following frames they are adaptively readjusted
68 // NOTE the decoder in ffmpeg has its own arbitrary limitation on the number
69 // of strips, currently 32
70 
71 typedef enum CinepakMode {
72     MODE_V1_ONLY = 0,
73     MODE_V1_V4,
74     MODE_MC,
75 
76     MODE_COUNT,
77 } CinepakMode;
78 
79 typedef enum mb_encoding {
80     ENC_V1,
81     ENC_V4,
82     ENC_SKIP,
83 
84     ENC_UNCERTAIN
85 } mb_encoding;
86 
87 typedef struct mb_info {
88     int v1_vector;              // index into v1 codebook
89     int v1_error;               // error when using V1 encoding
90     int v4_vector[4];           // indices into v4 codebook
91     int v4_error;               // error when using V4 encoding
92     int skip_error;             // error when block is skipped (aka copied from last frame)
93     mb_encoding best_encoding;  // last result from calculate_mode_score()
94 } mb_info;
95 
96 typedef struct strip_info {
97     int v1_codebook[CODEBOOK_MAX * VECTOR_MAX];
98     int v4_codebook[CODEBOOK_MAX * VECTOR_MAX];
99     int v1_size;
100     int v4_size;
101     CinepakMode mode;
102 } strip_info;
103 
104 typedef struct CinepakEncContext {
105     const AVClass *class;
106     AVCodecContext *avctx;
107     unsigned char *pict_bufs[4], *strip_buf, *frame_buf;
108     AVFrame *last_frame;
109     AVFrame *best_frame;
110     AVFrame *scratch_frame;
111     AVFrame *input_frame;
112     enum AVPixelFormat pix_fmt;
113     int w, h;
114     int frame_buf_size;
115     int curframe, keyint;
116     AVLFG randctx;
117     uint64_t lambda;
118     int *codebook_input;
119     int *codebook_closest;
120     mb_info *mb;                // MB RD state
121     int min_strips;             // the current limit
122     int max_strips;             // the current limit
123     // options
124     int max_extra_cb_iterations;
125     int skip_empty_cb;
126     int min_min_strips;
127     int max_max_strips;
128     int strip_number_delta_range;
129 } CinepakEncContext;
130 
131 #define OFFSET(x) offsetof(CinepakEncContext, x)
132 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
133 static const AVOption options[] = {
134     { "max_extra_cb_iterations", "Max extra codebook recalculation passes, more is better and slower",
135       OFFSET(max_extra_cb_iterations),  AV_OPT_TYPE_INT, { .i64 =          2 },          0, INT_MAX,                 VE },
136     { "skip_empty_cb",           "Avoid wasting bytes, ignore vintage MacOS decoder",
137       OFFSET(skip_empty_cb),            AV_OPT_TYPE_BOOL, { .i64 =         0 },          0, 1,                       VE },
138     { "max_strips",              "Limit strips/frame, vintage compatible is 1..3, otherwise the more the better",
139       OFFSET(max_max_strips),           AV_OPT_TYPE_INT, { .i64 =          3 }, MIN_STRIPS, MAX_STRIPS,              VE },
140     { "min_strips",              "Enforce min strips/frame, more is worse and faster, must be <= max_strips",
141       OFFSET(min_min_strips),           AV_OPT_TYPE_INT, { .i64 = MIN_STRIPS }, MIN_STRIPS, MAX_STRIPS,              VE },
142     { "strip_number_adaptivity", "How fast the strip number adapts, more is slightly better, much slower",
143       OFFSET(strip_number_delta_range), AV_OPT_TYPE_INT, { .i64 =          0 },          0, MAX_STRIPS - MIN_STRIPS, VE },
144     { NULL },
145 };
146 
147 static const AVClass cinepak_class = {
148     .class_name = "cinepak",
149     .item_name  = av_default_item_name,
150     .option     = options,
151     .version    = LIBAVUTIL_VERSION_INT,
152 };
153 
cinepak_encode_init(AVCodecContext * avctx)154 static av_cold int cinepak_encode_init(AVCodecContext *avctx)
155 {
156     CinepakEncContext *s = avctx->priv_data;
157     int x, mb_count, strip_buf_size, frame_buf_size;
158 
159     if (avctx->width & 3 || avctx->height & 3) {
160         av_log(avctx, AV_LOG_ERROR, "width and height must be multiples of four (got %ix%i)\n",
161                avctx->width, avctx->height);
162         return AVERROR(EINVAL);
163     }
164 
165     if (s->min_min_strips > s->max_max_strips) {
166         av_log(avctx, AV_LOG_ERROR, "minimum number of strips must not exceed maximum (got %i and %i)\n",
167                s->min_min_strips, s->max_max_strips);
168         return AVERROR(EINVAL);
169     }
170 
171     if (!(s->last_frame = av_frame_alloc()))
172         return AVERROR(ENOMEM);
173     if (!(s->best_frame = av_frame_alloc()))
174         return AVERROR(ENOMEM);
175     if (!(s->scratch_frame = av_frame_alloc()))
176         return AVERROR(ENOMEM);
177     if (avctx->pix_fmt == AV_PIX_FMT_RGB24)
178         if (!(s->input_frame = av_frame_alloc()))
179             return AVERROR(ENOMEM);
180 
181     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))))
182         return AVERROR(ENOMEM);
183 
184     if (!(s->codebook_closest = av_malloc_array((avctx->width * avctx->height) >> 2, sizeof(*s->codebook_closest))))
185         return AVERROR(ENOMEM);;
186 
187     for (x = 0; x < (avctx->pix_fmt == AV_PIX_FMT_RGB24 ? 4 : 3); x++)
188         if (!(s->pict_bufs[x] = av_malloc((avctx->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4) * (avctx->width * avctx->height) >> 2)))
189             return AVERROR(ENOMEM);
190 
191     mb_count = avctx->width * avctx->height / MB_AREA;
192 
193     // the largest possible chunk is 0x31 with all MBs encoded in V4 mode
194     // and full codebooks being replaced in INTER mode,
195     // which is 34 bits per MB
196     // and 2*256 extra flag bits per strip
197     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;
198 
199     frame_buf_size = CVID_HEADER_SIZE + s->max_max_strips * strip_buf_size;
200 
201     if (!(s->strip_buf = av_malloc(strip_buf_size)))
202         return AVERROR(ENOMEM);
203 
204     if (!(s->frame_buf = av_malloc(frame_buf_size)))
205         return AVERROR(ENOMEM);
206 
207     if (!(s->mb = av_malloc_array(mb_count, sizeof(mb_info))))
208         return AVERROR(ENOMEM);
209 
210     av_lfg_init(&s->randctx, 1);
211     s->avctx          = avctx;
212     s->w              = avctx->width;
213     s->h              = avctx->height;
214     s->frame_buf_size = frame_buf_size;
215     s->curframe       = 0;
216     s->keyint         = avctx->keyint_min;
217     s->pix_fmt        = avctx->pix_fmt;
218 
219     // set up AVFrames
220     s->last_frame->data[0]        = s->pict_bufs[0];
221     s->last_frame->linesize[0]    = s->w;
222     s->best_frame->data[0]        = s->pict_bufs[1];
223     s->best_frame->linesize[0]    = s->w;
224     s->scratch_frame->data[0]     = s->pict_bufs[2];
225     s->scratch_frame->linesize[0] = s->w;
226 
227     if (s->pix_fmt == AV_PIX_FMT_RGB24) {
228         s->last_frame->data[1]     = s->last_frame->data[0] +   s->w * s->h;
229         s->last_frame->data[2]     = s->last_frame->data[1] + ((s->w * s->h) >> 2);
230         s->last_frame->linesize[1] =
231         s->last_frame->linesize[2] = s->w >> 1;
232 
233         s->best_frame->data[1]     = s->best_frame->data[0] +   s->w * s->h;
234         s->best_frame->data[2]     = s->best_frame->data[1] + ((s->w * s->h) >> 2);
235         s->best_frame->linesize[1] =
236         s->best_frame->linesize[2] = s->w >> 1;
237 
238         s->scratch_frame->data[1]     = s->scratch_frame->data[0] +   s->w * s->h;
239         s->scratch_frame->data[2]     = s->scratch_frame->data[1] + ((s->w * s->h) >> 2);
240         s->scratch_frame->linesize[1] =
241         s->scratch_frame->linesize[2] = s->w >> 1;
242 
243         s->input_frame->data[0]     = s->pict_bufs[3];
244         s->input_frame->linesize[0] = s->w;
245         s->input_frame->data[1]     = s->input_frame->data[0] +   s->w * s->h;
246         s->input_frame->data[2]     = s->input_frame->data[1] + ((s->w * s->h) >> 2);
247         s->input_frame->linesize[1] =
248         s->input_frame->linesize[2] = s->w >> 1;
249     }
250 
251     s->min_strips = s->min_min_strips;
252     s->max_strips = s->max_max_strips;
253 
254     return 0;
255 }
256 
calculate_mode_score(CinepakEncContext * s,int h,strip_info * info,int report,int * training_set_v1_shrunk,int * training_set_v4_shrunk)257 static int64_t calculate_mode_score(CinepakEncContext *s, int h,
258                                     strip_info *info, int report,
259                                     int *training_set_v1_shrunk,
260                                     int *training_set_v4_shrunk)
261 {
262     // score = FF_LAMBDA_SCALE * error + lambda * bits
263     int x;
264     int entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4;
265     int mb_count   = s->w * h / MB_AREA;
266     mb_info *mb;
267     int64_t score1, score2, score3;
268     int64_t ret = s->lambda * ((info->v1_size ? CHUNK_HEADER_SIZE + info->v1_size * entry_size : 0) +
269                                (info->v4_size ? CHUNK_HEADER_SIZE + info->v4_size * entry_size : 0) +
270                                CHUNK_HEADER_SIZE) << 3;
271 
272     switch (info->mode) {
273     case MODE_V1_ONLY:
274         // one byte per MB
275         ret += s->lambda * 8 * mb_count;
276 
277         // while calculating we assume all blocks are ENC_V1
278         for (x = 0; x < mb_count; x++) {
279             mb   = &s->mb[x];
280             ret += FF_LAMBDA_SCALE * mb->v1_error;
281             // this function is never called for report in MODE_V1_ONLY
282             // if (!report)
283             mb->best_encoding = ENC_V1;
284         }
285 
286         break;
287     case MODE_V1_V4:
288         // 9 or 33 bits per MB
289         if (report) {
290             // no moves between the corresponding training sets are allowed
291             *training_set_v1_shrunk = *training_set_v4_shrunk = 0;
292             for (x = 0; x < mb_count; x++) {
293                 int mberr;
294                 mb = &s->mb[x];
295                 if (mb->best_encoding == ENC_V1)
296                     score1 = s->lambda * 9 + FF_LAMBDA_SCALE * (mberr = mb->v1_error);
297                 else
298                     score1 = s->lambda * 33 + FF_LAMBDA_SCALE * (mberr = mb->v4_error);
299                 ret += score1;
300             }
301         } else { // find best mode per block
302             for (x = 0; x < mb_count; x++) {
303                 mb     = &s->mb[x];
304                 score1 = s->lambda * 9 + FF_LAMBDA_SCALE * mb->v1_error;
305                 score2 = s->lambda * 33 + FF_LAMBDA_SCALE * mb->v4_error;
306 
307                 if (score1 <= score2) {
308                     ret += score1;
309                     mb->best_encoding = ENC_V1;
310                 } else {
311                     ret += score2;
312                     mb->best_encoding = ENC_V4;
313                 }
314             }
315         }
316 
317         break;
318     case MODE_MC:
319         // 1, 10 or 34 bits per MB
320         if (report) {
321             int v1_shrunk = 0, v4_shrunk = 0;
322             for (x = 0; x < mb_count; x++) {
323                 mb = &s->mb[x];
324                 // it is OK to move blocks to ENC_SKIP here
325                 // but not to any codebook encoding!
326                 score1 = s->lambda * 1 + FF_LAMBDA_SCALE * mb->skip_error;
327                 if (mb->best_encoding == ENC_SKIP) {
328                     ret += score1;
329                 } else if (mb->best_encoding == ENC_V1) {
330                     if ((score2 = s->lambda * 10 + FF_LAMBDA_SCALE * mb->v1_error) >= score1) {
331                         mb->best_encoding = ENC_SKIP;
332                         ++v1_shrunk;
333                         ret += score1;
334                     } else {
335                         ret += score2;
336                     }
337                 } else {
338                     if ((score3 = s->lambda * 34 + FF_LAMBDA_SCALE * mb->v4_error) >= score1) {
339                         mb->best_encoding = ENC_SKIP;
340                         ++v4_shrunk;
341                         ret += score1;
342                     } else {
343                         ret += score3;
344                     }
345                 }
346             }
347             *training_set_v1_shrunk = v1_shrunk;
348             *training_set_v4_shrunk = v4_shrunk;
349         } else { // find best mode per block
350             for (x = 0; x < mb_count; x++) {
351                 mb     = &s->mb[x];
352                 score1 = s->lambda * 1 + FF_LAMBDA_SCALE * mb->skip_error;
353                 score2 = s->lambda * 10 + FF_LAMBDA_SCALE * mb->v1_error;
354                 score3 = s->lambda * 34 + FF_LAMBDA_SCALE * mb->v4_error;
355 
356                 if (score1 <= score2 && score1 <= score3) {
357                     ret += score1;
358                     mb->best_encoding = ENC_SKIP;
359                 } else if (score2 <= score3) {
360                     ret += score2;
361                     mb->best_encoding = ENC_V1;
362                 } else {
363                     ret += score3;
364                     mb->best_encoding = ENC_V4;
365                 }
366             }
367         }
368 
369         break;
370     }
371 
372     return ret;
373 }
374 
write_chunk_header(unsigned char * buf,int chunk_type,int chunk_size)375 static int write_chunk_header(unsigned char *buf, int chunk_type, int chunk_size)
376 {
377     buf[0] = chunk_type;
378     AV_WB24(&buf[1], chunk_size + CHUNK_HEADER_SIZE);
379     return CHUNK_HEADER_SIZE;
380 }
381 
encode_codebook(CinepakEncContext * s,int * codebook,int size,int chunk_type_yuv,int chunk_type_gray,unsigned char * buf)382 static int encode_codebook(CinepakEncContext *s, int *codebook, int size,
383                            int chunk_type_yuv, int chunk_type_gray,
384                            unsigned char *buf)
385 {
386     int x, y, ret, entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4;
387     int incremental_codebook_replacement_mode = 0; // hardcoded here,
388     // the compiler should notice that this is a constant -- rl
389 
390     ret = write_chunk_header(buf,
391                              s->pix_fmt == AV_PIX_FMT_RGB24 ?
392                              chunk_type_yuv  + (incremental_codebook_replacement_mode ? 1 : 0) :
393                              chunk_type_gray + (incremental_codebook_replacement_mode ? 1 : 0),
394                              entry_size * size +
395                              (incremental_codebook_replacement_mode ? (size + 31) / 32 * 4 : 0));
396 
397     // we do codebook encoding according to the "intra" mode
398     // but we keep the "dead" code for reference in case we will want
399     // to use incremental codebook updates (which actually would give us
400     // "kind of" motion compensation, especially in 1 strip/frame case) -- rl
401     // (of course, the code will be not useful as-is)
402     if (incremental_codebook_replacement_mode) {
403         int flags = 0;
404         int flagsind;
405         for (x = 0; x < size; x++) {
406             if (flags == 0) {
407                 flagsind = ret;
408                 ret     += 4;
409                 flags    = 0x80000000;
410             } else
411                 flags = ((flags >> 1) | 0x80000000);
412             for (y = 0; y < entry_size; y++)
413                 buf[ret++] = codebook[y + x * entry_size] ^ (y >= 4 ? 0x80 : 0);
414             if ((flags & 0xffffffff) == 0xffffffff) {
415                 AV_WB32(&buf[flagsind], flags);
416                 flags = 0;
417             }
418         }
419         if (flags)
420             AV_WB32(&buf[flagsind], flags);
421     } else
422         for (x = 0; x < size; x++)
423             for (y = 0; y < entry_size; y++)
424                 buf[ret++] = codebook[y + x * entry_size] ^ (y >= 4 ? 0x80 : 0);
425 
426     return ret;
427 }
428 
429 // sets out to the sub picture starting at (x,y) in in
get_sub_picture(CinepakEncContext * s,int x,int y,uint8_t * in_data[4],int in_linesize[4],uint8_t * out_data[4],int out_linesize[4])430 static void get_sub_picture(CinepakEncContext *s, int x, int y,
431                             uint8_t * in_data[4], int  in_linesize[4],
432                             uint8_t *out_data[4], int out_linesize[4])
433 {
434     out_data[0]     = in_data[0] + x + y * in_linesize[0];
435     out_linesize[0] = in_linesize[0];
436 
437     if (s->pix_fmt == AV_PIX_FMT_RGB24) {
438         out_data[1]     = in_data[1] + (x >> 1) + (y >> 1) * in_linesize[1];
439         out_linesize[1] = in_linesize[1];
440 
441         out_data[2]     = in_data[2] + (x >> 1) + (y >> 1) * in_linesize[2];
442         out_linesize[2] = in_linesize[2];
443     }
444 }
445 
446 // decodes the V1 vector in mb into the 4x4 MB pointed to by data
decode_v1_vector(CinepakEncContext * s,uint8_t * data[4],int linesize[4],int v1_vector,strip_info * info)447 static void decode_v1_vector(CinepakEncContext *s, uint8_t *data[4],
448                              int linesize[4], int v1_vector, strip_info *info)
449 {
450     int entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4;
451 
452     data[0][0] =
453     data[0][1] =
454     data[0][    linesize[0]] =
455     data[0][1 + linesize[0]] = info->v1_codebook[v1_vector * entry_size];
456 
457     data[0][2] =
458     data[0][3] =
459     data[0][2 + linesize[0]] =
460     data[0][3 + linesize[0]] = info->v1_codebook[v1_vector * entry_size + 1];
461 
462     data[0][    2 * linesize[0]] =
463     data[0][1 + 2 * linesize[0]] =
464     data[0][    3 * linesize[0]] =
465     data[0][1 + 3 * linesize[0]] = info->v1_codebook[v1_vector * entry_size + 2];
466 
467     data[0][2 + 2 * linesize[0]] =
468     data[0][3 + 2 * linesize[0]] =
469     data[0][2 + 3 * linesize[0]] =
470     data[0][3 + 3 * linesize[0]] = info->v1_codebook[v1_vector * entry_size + 3];
471 
472     if (s->pix_fmt == AV_PIX_FMT_RGB24) {
473         data[1][0] =
474         data[1][1] =
475         data[1][    linesize[1]] =
476         data[1][1 + linesize[1]] = info->v1_codebook[v1_vector * entry_size + 4];
477 
478         data[2][0] =
479         data[2][1] =
480         data[2][    linesize[2]] =
481         data[2][1 + linesize[2]] = info->v1_codebook[v1_vector * entry_size + 5];
482     }
483 }
484 
485 // decodes the V4 vectors in mb into the 4x4 MB pointed to by data
decode_v4_vector(CinepakEncContext * s,uint8_t * data[4],int linesize[4],int * v4_vector,strip_info * info)486 static void decode_v4_vector(CinepakEncContext *s, uint8_t *data[4],
487                              int linesize[4], int *v4_vector, strip_info *info)
488 {
489     int i, x, y, entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4;
490 
491     for (i = y = 0; y < 4; y += 2) {
492         for (x = 0; x < 4; x += 2, i++) {
493             data[0][x     +  y      * linesize[0]] = info->v4_codebook[v4_vector[i] * entry_size];
494             data[0][x + 1 +  y      * linesize[0]] = info->v4_codebook[v4_vector[i] * entry_size + 1];
495             data[0][x     + (y + 1) * linesize[0]] = info->v4_codebook[v4_vector[i] * entry_size + 2];
496             data[0][x + 1 + (y + 1) * linesize[0]] = info->v4_codebook[v4_vector[i] * entry_size + 3];
497 
498             if (s->pix_fmt == AV_PIX_FMT_RGB24) {
499                 data[1][(x >> 1) + (y >> 1) * linesize[1]] = info->v4_codebook[v4_vector[i] * entry_size + 4];
500                 data[2][(x >> 1) + (y >> 1) * linesize[2]] = info->v4_codebook[v4_vector[i] * entry_size + 5];
501             }
502         }
503     }
504 }
505 
copy_mb(CinepakEncContext * s,uint8_t * a_data[4],int a_linesize[4],uint8_t * b_data[4],int b_linesize[4])506 static void copy_mb(CinepakEncContext *s,
507                     uint8_t *a_data[4], int a_linesize[4],
508                     uint8_t *b_data[4], int b_linesize[4])
509 {
510     int y, p;
511 
512     for (y = 0; y < MB_SIZE; y++)
513         memcpy(a_data[0] + y * a_linesize[0], b_data[0] + y * b_linesize[0],
514                MB_SIZE);
515 
516     if (s->pix_fmt == AV_PIX_FMT_RGB24) {
517         for (p = 1; p <= 2; p++)
518             for (y = 0; y < MB_SIZE / 2; y++)
519                 memcpy(a_data[p] + y * a_linesize[p],
520                        b_data[p] + y * b_linesize[p],
521                        MB_SIZE / 2);
522     }
523 }
524 
encode_mode(CinepakEncContext * s,int h,uint8_t * scratch_data[4],int scratch_linesize[4],uint8_t * last_data[4],int last_linesize[4],strip_info * info,unsigned char * buf)525 static int encode_mode(CinepakEncContext *s, int h,
526                        uint8_t *scratch_data[4], int scratch_linesize[4],
527                        uint8_t *last_data[4], int last_linesize[4],
528                        strip_info *info, unsigned char *buf)
529 {
530     int x, y, z, bits, temp_size, header_ofs, ret = 0, mb_count = s->w * h / MB_AREA;
531     int needs_extra_bit, should_write_temp;
532     uint32_t flags;
533     unsigned char temp[64]; // 32/2 = 16 V4 blocks at 4 B each -> 64 B
534     mb_info *mb;
535     uint8_t *sub_scratch_data[4] = { 0 }, *sub_last_data[4] = { 0 };
536     int sub_scratch_linesize[4] = { 0 }, sub_last_linesize[4] = { 0 };
537 
538     // encode codebooks
539     ////// MacOS vintage decoder compatibility dictates the presence of
540     ////// the codebook chunk even when the codebook is empty - pretty dumb...
541     ////// and also the certain order of the codebook chunks -- rl
542     if (info->v4_size || !s->skip_empty_cb)
543         ret += encode_codebook(s, info->v4_codebook, info->v4_size, 0x20, 0x24, buf + ret);
544 
545     if (info->v1_size || !s->skip_empty_cb)
546         ret += encode_codebook(s, info->v1_codebook, info->v1_size, 0x22, 0x26, buf + ret);
547 
548     // update scratch picture
549     for (z = y = 0; y < h; y += MB_SIZE)
550         for (x = 0; x < s->w; x += MB_SIZE, z++) {
551             mb = &s->mb[z];
552 
553             get_sub_picture(s, x, y, scratch_data, scratch_linesize,
554                             sub_scratch_data, sub_scratch_linesize);
555 
556             if (info->mode == MODE_MC && mb->best_encoding == ENC_SKIP) {
557                 get_sub_picture(s, x, y, last_data, last_linesize,
558                                 sub_last_data, sub_last_linesize);
559                 copy_mb(s, sub_scratch_data, sub_scratch_linesize,
560                         sub_last_data, sub_last_linesize);
561             } else if (info->mode == MODE_V1_ONLY || mb->best_encoding == ENC_V1)
562                 decode_v1_vector(s, sub_scratch_data, sub_scratch_linesize,
563                                  mb->v1_vector, info);
564             else
565                 decode_v4_vector(s, sub_scratch_data, sub_scratch_linesize,
566                                  mb->v4_vector, info);
567         }
568 
569     switch (info->mode) {
570     case MODE_V1_ONLY:
571         ret += write_chunk_header(buf + ret, 0x32, mb_count);
572 
573         for (x = 0; x < mb_count; x++)
574             buf[ret++] = s->mb[x].v1_vector;
575 
576         break;
577     case MODE_V1_V4:
578         // remember header position
579         header_ofs = ret;
580         ret       += CHUNK_HEADER_SIZE;
581 
582         for (x = 0; x < mb_count; x += 32) {
583             flags = 0;
584             for (y = x; y < FFMIN(x + 32, mb_count); y++)
585                 if (s->mb[y].best_encoding == ENC_V4)
586                     flags |= 1U << (31 - y + x);
587 
588             AV_WB32(&buf[ret], flags);
589             ret += 4;
590 
591             for (y = x; y < FFMIN(x + 32, mb_count); y++) {
592                 mb = &s->mb[y];
593 
594                 if (mb->best_encoding == ENC_V1)
595                     buf[ret++] = mb->v1_vector;
596                 else
597                     for (z = 0; z < 4; z++)
598                         buf[ret++] = mb->v4_vector[z];
599             }
600         }
601 
602         write_chunk_header(buf + header_ofs, 0x30, ret - header_ofs - CHUNK_HEADER_SIZE);
603 
604         break;
605     case MODE_MC:
606         // remember header position
607         header_ofs = ret;
608         ret       += CHUNK_HEADER_SIZE;
609         flags      = bits = temp_size = 0;
610 
611         for (x = 0; x < mb_count; x++) {
612             mb                = &s->mb[x];
613             flags            |= (uint32_t)(mb->best_encoding != ENC_SKIP) << (31 - bits++);
614             needs_extra_bit   = 0;
615             should_write_temp = 0;
616 
617             if (mb->best_encoding != ENC_SKIP) {
618                 if (bits < 32)
619                     flags |= (uint32_t)(mb->best_encoding == ENC_V4) << (31 - bits++);
620                 else
621                     needs_extra_bit = 1;
622             }
623 
624             if (bits == 32) {
625                 AV_WB32(&buf[ret], flags);
626                 ret  += 4;
627                 flags = bits = 0;
628 
629                 if (mb->best_encoding == ENC_SKIP || needs_extra_bit) {
630                     memcpy(&buf[ret], temp, temp_size);
631                     ret      += temp_size;
632                     temp_size = 0;
633                 } else
634                     should_write_temp = 1;
635             }
636 
637             if (needs_extra_bit) {
638                 flags = (uint32_t)(mb->best_encoding == ENC_V4) << 31;
639                 bits  = 1;
640             }
641 
642             if (mb->best_encoding == ENC_V1)
643                 temp[temp_size++] = mb->v1_vector;
644             else if (mb->best_encoding == ENC_V4)
645                 for (z = 0; z < 4; z++)
646                     temp[temp_size++] = mb->v4_vector[z];
647 
648             if (should_write_temp) {
649                 memcpy(&buf[ret], temp, temp_size);
650                 ret      += temp_size;
651                 temp_size = 0;
652             }
653         }
654 
655         if (bits > 0) {
656             AV_WB32(&buf[ret], flags);
657             ret += 4;
658             memcpy(&buf[ret], temp, temp_size);
659             ret += temp_size;
660         }
661 
662         write_chunk_header(buf + header_ofs, 0x31, ret - header_ofs - CHUNK_HEADER_SIZE);
663 
664         break;
665     }
666 
667     return ret;
668 }
669 
670 // computes distortion of 4x4 MB in b compared to a
compute_mb_distortion(CinepakEncContext * s,uint8_t * a_data[4],int a_linesize[4],uint8_t * b_data[4],int b_linesize[4])671 static int compute_mb_distortion(CinepakEncContext *s,
672                                  uint8_t *a_data[4], int a_linesize[4],
673                                  uint8_t *b_data[4], int b_linesize[4])
674 {
675     int x, y, p, d, ret = 0;
676 
677     for (y = 0; y < MB_SIZE; y++)
678         for (x = 0; x < MB_SIZE; x++) {
679             d = a_data[0][x + y * a_linesize[0]] - b_data[0][x + y * b_linesize[0]];
680             ret += d * d;
681         }
682 
683     if (s->pix_fmt == AV_PIX_FMT_RGB24) {
684         for (p = 1; p <= 2; p++) {
685             for (y = 0; y < MB_SIZE / 2; y++)
686                 for (x = 0; x < MB_SIZE / 2; x++) {
687                     d = a_data[p][x + y * a_linesize[p]] - b_data[p][x + y * b_linesize[p]];
688                     ret += d * d;
689                 }
690         }
691     }
692 
693     return ret;
694 }
695 
696 // return the possibly adjusted size of the codebook
697 #define CERTAIN(x) ((x) != ENC_UNCERTAIN)
quantize(CinepakEncContext * s,int h,uint8_t * data[4],int linesize[4],int v1mode,strip_info * info,mb_encoding encoding)698 static int quantize(CinepakEncContext *s, int h, uint8_t *data[4],
699                     int linesize[4], int v1mode, strip_info *info,
700                     mb_encoding encoding)
701 {
702     int x, y, i, j, k, x2, y2, x3, y3, plane, shift, mbn;
703     int entry_size      = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4;
704     int *codebook       = v1mode ? info->v1_codebook : info->v4_codebook;
705     int size            = v1mode ? info->v1_size : info->v4_size;
706     int64_t total_error = 0;
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];
710 
711     for (mbn = i = y = 0; y < h; y += MB_SIZE) {
712         for (x = 0; x < s->w; x += MB_SIZE, ++mbn) {
713             int *base;
714 
715             if (CERTAIN(encoding)) {
716                 // use for the training only the blocks known to be to be encoded [sic:-]
717                 if (s->mb[mbn].best_encoding != encoding)
718                     continue;
719             }
720 
721             base = s->codebook_input + i * entry_size;
722             if (v1mode) {
723                 // subsample
724                 for (j = y2 = 0; y2 < entry_size; y2 += 2)
725                     for (x2 = 0; x2 < 4; x2 += 2, j++) {
726                         plane   = y2 < 4 ? 0 : 1 + (x2 >> 1);
727                         shift   = y2 < 4 ? 0 : 1;
728                         x3      = shift ? 0 : x2;
729                         y3      = shift ? 0 : y2;
730                         base[j] = (data[plane][((x + x3) >> shift) +      ((y + y3) >> shift)      * linesize[plane]] +
731                                    data[plane][((x + x3) >> shift) + 1 +  ((y + y3) >> shift)      * linesize[plane]] +
732                                    data[plane][((x + x3) >> shift) +     (((y + y3) >> shift) + 1) * linesize[plane]] +
733                                    data[plane][((x + x3) >> shift) + 1 + (((y + y3) >> shift) + 1) * linesize[plane]]) >> 2;
734                     }
735             } else {
736                 // copy
737                 for (j = y2 = 0; y2 < MB_SIZE; y2 += 2) {
738                     for (x2 = 0; x2 < MB_SIZE; x2 += 2)
739                         for (k = 0; k < entry_size; k++, j++) {
740                             plane = k >= 4 ? k - 3 : 0;
741 
742                             if (k >= 4) {
743                                 x3 = (x + x2) >> 1;
744                                 y3 = (y + y2) >> 1;
745                             } else {
746                                 x3 = x + x2 + (k & 1);
747                                 y3 = y + y2 + (k >> 1);
748                             }
749 
750                             base[j] = data[plane][x3 + y3 * linesize[plane]];
751                         }
752                 }
753             }
754             i += v1mode ? 1 : 4;
755         }
756     }
757 
758     if (i == 0) // empty training set, nothing to do
759         return 0;
760     if (i < size)
761         size = i;
762 
763     avpriv_init_elbg(s->codebook_input, entry_size, i, codebook, size, 1, s->codebook_closest, &s->randctx);
764     avpriv_do_elbg(s->codebook_input, entry_size, i, codebook, size, 1, s->codebook_closest, &s->randctx);
765 
766     // set up vq_data, which contains a single MB
767     vq_data[0]     = vq_pict_buf;
768     vq_linesize[0] = MB_SIZE;
769     vq_data[1]     = &vq_pict_buf[MB_AREA];
770     vq_data[2]     = vq_data[1] + (MB_AREA >> 2);
771     vq_linesize[1] =
772     vq_linesize[2] = MB_SIZE >> 1;
773 
774     // copy indices
775     for (i = j = y = 0; y < h; y += MB_SIZE)
776         for (x = 0; x < s->w; x += MB_SIZE, j++) {
777             mb_info *mb = &s->mb[j];
778             // skip uninteresting blocks if we know their preferred encoding
779             if (CERTAIN(encoding) && mb->best_encoding != encoding)
780                 continue;
781 
782             // point sub_data to current MB
783             get_sub_picture(s, x, y, data, linesize, sub_data, sub_linesize);
784 
785             if (v1mode) {
786                 mb->v1_vector = s->codebook_closest[i];
787 
788                 // fill in vq_data with V1 data
789                 decode_v1_vector(s, vq_data, vq_linesize, mb->v1_vector, info);
790 
791                 mb->v1_error = compute_mb_distortion(s, sub_data, sub_linesize,
792                                                      vq_data, vq_linesize);
793                 total_error += mb->v1_error;
794             } else {
795                 for (k = 0; k < 4; k++)
796                     mb->v4_vector[k] = s->codebook_closest[i + k];
797 
798                 // fill in vq_data with V4 data
799                 decode_v4_vector(s, vq_data, vq_linesize, mb->v4_vector, info);
800 
801                 mb->v4_error = compute_mb_distortion(s, sub_data, sub_linesize,
802                                                      vq_data, vq_linesize);
803                 total_error += mb->v4_error;
804             }
805             i += v1mode ? 1 : 4;
806         }
807     // check that we did it right in the beginning of the function
808     av_assert0(i >= size); // training set is no smaller than the codebook
809 
810     return size;
811 }
812 
calculate_skip_errors(CinepakEncContext * s,int h,uint8_t * last_data[4],int last_linesize[4],uint8_t * data[4],int linesize[4],strip_info * info)813 static void calculate_skip_errors(CinepakEncContext *s, int h,
814                                   uint8_t *last_data[4], int last_linesize[4],
815                                   uint8_t *data[4], int linesize[4],
816                                   strip_info *info)
817 {
818     int x, y, i;
819     uint8_t *sub_last_data    [4], *sub_pict_data    [4];
820     int      sub_last_linesize[4],  sub_pict_linesize[4];
821 
822     for (i = y = 0; y < h; y += MB_SIZE)
823         for (x = 0; x < s->w; x += MB_SIZE, i++) {
824             get_sub_picture(s, x, y, last_data, last_linesize,
825                             sub_last_data, sub_last_linesize);
826             get_sub_picture(s, x, y, data, linesize,
827                             sub_pict_data, sub_pict_linesize);
828 
829             s->mb[i].skip_error =
830                 compute_mb_distortion(s,
831                                       sub_last_data, sub_last_linesize,
832                                       sub_pict_data, sub_pict_linesize);
833         }
834 }
835 
write_strip_header(CinepakEncContext * s,int y,int h,int keyframe,unsigned char * buf,int strip_size)836 static void write_strip_header(CinepakEncContext *s, int y, int h, int keyframe,
837                                unsigned char *buf, int strip_size)
838 {
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;
845     AV_WB24(&buf[1], strip_size + STRIP_HEADER_SIZE);
846     // AV_WB16(&buf[4], y); /* using absolute y values works -- rl */
847     AV_WB16(&buf[4], 0); /* using relative values works as well -- rl */
848     AV_WB16(&buf[6], 0);
849     // AV_WB16(&buf[8], y + h); /* using absolute y values works -- rl */
850     AV_WB16(&buf[8], h); /* using relative values works as well -- rl */
851     AV_WB16(&buf[10], s->w);
852 }
853 
rd_strip(CinepakEncContext * s,int y,int h,int keyframe,uint8_t * last_data[4],int last_linesize[4],uint8_t * data[4],int linesize[4],uint8_t * scratch_data[4],int scratch_linesize[4],unsigned char * buf,int64_t * best_score)854 static int rd_strip(CinepakEncContext *s, int y, int h, int keyframe,
855                     uint8_t *last_data[4], int last_linesize[4],
856                     uint8_t *data[4], int linesize[4],
857                     uint8_t *scratch_data[4], int scratch_linesize[4],
858                     unsigned char *buf, int64_t *best_score)
859 {
860     int64_t score = 0;
861     int best_size = 0;
862     strip_info info;
863     // for codebook optimization:
864     int v1enough, v1_size, v4enough, v4_size;
865     int new_v1_size, new_v4_size;
866     int v1shrunk, v4shrunk;
867 
868     if (!keyframe)
869         calculate_skip_errors(s, h, last_data, last_linesize, data, linesize,
870                               &info);
871 
872     // try some powers of 4 for the size of the codebooks
873     // constraint the v4 codebook to be no bigger than v1 one,
874     // (and no less than v1_size/4)
875     // thus making v1 preferable and possibly losing small details? should be ok
876 #define SMALLEST_CODEBOOK 1
877     for (v1enough = 0, v1_size = SMALLEST_CODEBOOK; v1_size <= CODEBOOK_MAX && !v1enough; v1_size <<= 2) {
878         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) {
879             CinepakMode mode;
880             // try all modes
881             for (mode = 0; mode < MODE_COUNT; mode++) {
882                 // don't allow MODE_MC in intra frames
883                 if (keyframe && mode == MODE_MC)
884                     continue;
885 
886                 if (mode == MODE_V1_ONLY) {
887                     info.v1_size = v1_size;
888                     // the size may shrink even before optimizations if the input is short:
889                     info.v1_size = quantize(s, h, data, linesize, 1,
890                                             &info, ENC_UNCERTAIN);
891                     if (info.v1_size < v1_size)
892                         // too few eligible blocks, no sense in trying bigger sizes
893                         v1enough = 1;
894 
895                     info.v4_size = 0;
896                 } else { // mode != MODE_V1_ONLY
897                     // if v4 codebook is empty then only allow V1-only mode
898                     if (!v4_size)
899                         continue;
900 
901                     if (mode == MODE_V1_V4) {
902                         info.v4_size = v4_size;
903                         info.v4_size = quantize(s, h, data, linesize, 0,
904                                                 &info, ENC_UNCERTAIN);
905                         if (info.v4_size < v4_size)
906                             // too few eligible blocks, no sense in trying bigger sizes
907                             v4enough = 1;
908                     }
909                 }
910 
911                 info.mode = mode;
912                 // choose the best encoding per block, based on current experience
913                 score = calculate_mode_score(s, h, &info, 0,
914                                              &v1shrunk, &v4shrunk);
915 
916                 if (mode != MODE_V1_ONLY) {
917                     int extra_iterations_limit = s->max_extra_cb_iterations;
918                     // recompute the codebooks, omitting the extra blocks
919                     // we assume we _may_ come here with more blocks to encode than before
920                     info.v1_size = v1_size;
921                     new_v1_size = quantize(s, h, data, linesize, 1, &info, ENC_V1);
922                     if (new_v1_size < info.v1_size)
923                         info.v1_size = new_v1_size;
924                     // we assume we _may_ come here with more blocks to encode than before
925                     info.v4_size = v4_size;
926                     new_v4_size = quantize(s, h, data, linesize, 0, &info, ENC_V4);
927                     if (new_v4_size < info.v4_size)
928                         info.v4_size = new_v4_size;
929                     // calculate the resulting score
930                     // (do not move blocks to codebook encodings now, as some blocks may have
931                     // got bigger errors despite a smaller training set - but we do not
932                     // ever grow the training sets back)
933                     for (;;) {
934                         score = calculate_mode_score(s, h, &info, 1,
935                                                      &v1shrunk, &v4shrunk);
936                         // do we have a reason to reiterate? if so, have we reached the limit?
937                         if ((!v1shrunk && !v4shrunk) || !extra_iterations_limit--)
938                             break;
939                         // recompute the codebooks, omitting the extra blocks
940                         if (v1shrunk) {
941                             info.v1_size = v1_size;
942                             new_v1_size = quantize(s, h, data, linesize, 1, &info, ENC_V1);
943                             if (new_v1_size < info.v1_size)
944                                 info.v1_size = new_v1_size;
945                         }
946                         if (v4shrunk) {
947                             info.v4_size = v4_size;
948                             new_v4_size = quantize(s, h, data, linesize, 0, &info, ENC_V4);
949                             if (new_v4_size < info.v4_size)
950                                 info.v4_size = new_v4_size;
951                         }
952                     }
953                 }
954 
955                 if (best_size == 0 || score < *best_score) {
956                     *best_score = score;
957                     best_size = encode_mode(s, h,
958                                             scratch_data, scratch_linesize,
959                                             last_data, last_linesize, &info,
960                                             s->strip_buf + STRIP_HEADER_SIZE);
961 
962                     write_strip_header(s, y, h, keyframe, s->strip_buf, best_size);
963                 }
964             }
965         }
966     }
967 
968     best_size += STRIP_HEADER_SIZE;
969     memcpy(buf, s->strip_buf, best_size);
970 
971     return best_size;
972 }
973 
write_cvid_header(CinepakEncContext * s,unsigned char * buf,int num_strips,int data_size,int isakeyframe)974 static int write_cvid_header(CinepakEncContext *s, unsigned char *buf,
975                              int num_strips, int data_size, int isakeyframe)
976 {
977     buf[0] = isakeyframe ? 0 : 1;
978     AV_WB24(&buf[1], data_size + CVID_HEADER_SIZE);
979     AV_WB16(&buf[4], s->w);
980     AV_WB16(&buf[6], s->h);
981     AV_WB16(&buf[8], num_strips);
982 
983     return CVID_HEADER_SIZE;
984 }
985 
rd_frame(CinepakEncContext * s,const AVFrame * frame,int isakeyframe,unsigned char * buf,int buf_size)986 static int rd_frame(CinepakEncContext *s, const AVFrame *frame,
987                     int isakeyframe, unsigned char *buf, int buf_size)
988 {
989     int num_strips, strip, i, y, nexty, size, temp_size, best_size;
990     uint8_t *last_data    [4], *data    [4], *scratch_data    [4];
991     int      last_linesize[4],  linesize[4],  scratch_linesize[4];
992     int64_t best_score = 0, score, score_temp;
993     int best_nstrips;
994 
995     if (s->pix_fmt == AV_PIX_FMT_RGB24) {
996         int x;
997         // build a copy of the given frame in the correct colorspace
998         for (y = 0; y < s->h; y += 2)
999             for (x = 0; x < s->w; x += 2) {
1000                 uint8_t *ir[2];
1001                 int32_t r, g, b, rr, gg, bb;
1002                 ir[0] = frame->data[0] + x * 3 + y * frame->linesize[0];
1003                 ir[1] = ir[0] + frame->linesize[0];
1004                 get_sub_picture(s, x, y,
1005                                 s->input_frame->data, s->input_frame->linesize,
1006                                 scratch_data, scratch_linesize);
1007                 r = g = b = 0;
1008                 for (i = 0; i < 4; ++i) {
1009                     int i1, i2;
1010                     i1 = (i & 1);
1011                     i2 = (i >= 2);
1012                     rr = ir[i2][i1 * 3 + 0];
1013                     gg = ir[i2][i1 * 3 + 1];
1014                     bb = ir[i2][i1 * 3 + 2];
1015                     r += rr;
1016                     g += gg;
1017                     b += bb;
1018                     // using fixed point arithmetic for portable repeatability, scaling by 2^23
1019                     // "Y"
1020                     // rr = 0.2857 * rr + 0.5714 * gg + 0.1429 * bb;
1021                     rr = (2396625 * rr + 4793251 * gg + 1198732 * bb) >> 23;
1022                     if (rr < 0)
1023                         rr = 0;
1024                     else if (rr > 255)
1025                         rr = 255;
1026                     scratch_data[0][i1 + i2 * scratch_linesize[0]] = rr;
1027                 }
1028                 // let us scale down as late as possible
1029                 //                r /= 4; g /= 4; b /= 4;
1030                 // "U"
1031                 // rr = -0.1429 * r - 0.2857 * g + 0.4286 * b;
1032                 rr = (-299683 * r - 599156 * g + 898839 * b) >> 23;
1033                 if (rr < -128)
1034                     rr = -128;
1035                 else if (rr > 127)
1036                     rr = 127;
1037                 scratch_data[1][0] = rr + 128; // quantize needs unsigned
1038                 // "V"
1039                 // rr = 0.3571 * r - 0.2857 * g - 0.0714 * b;
1040                 rr = (748893 * r - 599156 * g - 149737 * b) >> 23;
1041                 if (rr < -128)
1042                     rr = -128;
1043                 else if (rr > 127)
1044                     rr = 127;
1045                 scratch_data[2][0] = rr + 128; // quantize needs unsigned
1046             }
1047     }
1048 
1049     // would be nice but quite certainly incompatible with vintage players:
1050     // support encoding zero strips (meaning skip the whole frame)
1051     for (num_strips = s->min_strips; num_strips <= s->max_strips && num_strips <= s->h / MB_SIZE; num_strips++) {
1052         score = 0;
1053         size  = 0;
1054 
1055         for (y = 0, strip = 1; y < s->h; strip++, y = nexty) {
1056             int strip_height;
1057 
1058             nexty = strip * s->h / num_strips; // <= s->h
1059             // make nexty the next multiple of 4 if not already there
1060             if (nexty & 3)
1061                 nexty += 4 - (nexty & 3);
1062 
1063             strip_height = nexty - y;
1064             if (strip_height <= 0) { // can this ever happen?
1065                 av_log(s->avctx, AV_LOG_INFO, "skipping zero height strip %i of %i\n", strip, num_strips);
1066                 continue;
1067             }
1068 
1069             if (s->pix_fmt == AV_PIX_FMT_RGB24)
1070                 get_sub_picture(s, 0, y,
1071                                 s->input_frame->data, s->input_frame->linesize,
1072                                 data, linesize);
1073             else
1074                 get_sub_picture(s, 0, y,
1075                                 (uint8_t **)frame->data, (int *)frame->linesize,
1076                                 data, linesize);
1077             get_sub_picture(s, 0, y,
1078                             s->last_frame->data, s->last_frame->linesize,
1079                             last_data, last_linesize);
1080             get_sub_picture(s, 0, y,
1081                             s->scratch_frame->data, s->scratch_frame->linesize,
1082                             scratch_data, scratch_linesize);
1083 
1084             if ((temp_size = rd_strip(s, y, strip_height, isakeyframe,
1085                                       last_data, last_linesize, data, linesize,
1086                                       scratch_data, scratch_linesize,
1087                                       s->frame_buf + size + CVID_HEADER_SIZE,
1088                                       &score_temp)) < 0)
1089                 return temp_size;
1090 
1091             score += score_temp;
1092             size += temp_size;
1093         }
1094 
1095         if (best_score == 0 || score < best_score) {
1096             best_score = score;
1097             best_size = size + write_cvid_header(s, s->frame_buf, num_strips, size, isakeyframe);
1098 
1099             FFSWAP(AVFrame *, s->best_frame, s->scratch_frame);
1100             memcpy(buf, s->frame_buf, best_size);
1101             best_nstrips = num_strips;
1102         }
1103         // avoid trying too many strip numbers without a real reason
1104         // (this makes the processing of the very first frame faster)
1105         if (num_strips - best_nstrips > 4)
1106             break;
1107     }
1108 
1109     // let the number of strips slowly adapt to the changes in the contents,
1110     // compared to full bruteforcing every time this will occasionally lead
1111     // to some r/d performance loss but makes encoding up to several times faster
1112     if (!s->strip_number_delta_range) {
1113         if (best_nstrips == s->max_strips) { // let us try to step up
1114             s->max_strips = best_nstrips + 1;
1115             if (s->max_strips >= s->max_max_strips)
1116                 s->max_strips = s->max_max_strips;
1117         } else { // try to step down
1118             s->max_strips = best_nstrips;
1119         }
1120         s->min_strips = s->max_strips - 1;
1121         if (s->min_strips < s->min_min_strips)
1122             s->min_strips = s->min_min_strips;
1123     } else {
1124         s->max_strips = best_nstrips + s->strip_number_delta_range;
1125         if (s->max_strips >= s->max_max_strips)
1126             s->max_strips = s->max_max_strips;
1127         s->min_strips = best_nstrips - s->strip_number_delta_range;
1128         if (s->min_strips < s->min_min_strips)
1129             s->min_strips = s->min_min_strips;
1130     }
1131 
1132     return best_size;
1133 }
1134 
cinepak_encode_frame(AVCodecContext * avctx,AVPacket * pkt,const AVFrame * frame,int * got_packet)1135 static int cinepak_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1136                                 const AVFrame *frame, int *got_packet)
1137 {
1138     CinepakEncContext *s = avctx->priv_data;
1139     int ret;
1140 
1141     s->lambda = frame->quality ? frame->quality - 1 : 2 * FF_LAMBDA_SCALE;
1142 
1143     if ((ret = ff_alloc_packet2(avctx, pkt, s->frame_buf_size, 0)) < 0)
1144         return ret;
1145     ret       = rd_frame(s, frame, (s->curframe == 0), pkt->data, s->frame_buf_size);
1146     pkt->size = ret;
1147     if (s->curframe == 0)
1148         pkt->flags |= AV_PKT_FLAG_KEY;
1149     *got_packet = 1;
1150 
1151     FFSWAP(AVFrame *, s->last_frame, s->best_frame);
1152 
1153     if (++s->curframe >= s->keyint)
1154         s->curframe = 0;
1155 
1156     return 0;
1157 }
1158 
cinepak_encode_end(AVCodecContext * avctx)1159 static av_cold int cinepak_encode_end(AVCodecContext *avctx)
1160 {
1161     CinepakEncContext *s = avctx->priv_data;
1162     int x;
1163 
1164     av_frame_free(&s->last_frame);
1165     av_frame_free(&s->best_frame);
1166     av_frame_free(&s->scratch_frame);
1167     if (avctx->pix_fmt == AV_PIX_FMT_RGB24)
1168         av_frame_free(&s->input_frame);
1169     av_freep(&s->codebook_input);
1170     av_freep(&s->codebook_closest);
1171     av_freep(&s->strip_buf);
1172     av_freep(&s->frame_buf);
1173     av_freep(&s->mb);
1174 
1175     for (x = 0; x < (avctx->pix_fmt == AV_PIX_FMT_RGB24 ? 4 : 3); x++)
1176         av_freep(&s->pict_bufs[x]);
1177 
1178     return 0;
1179 }
1180 
1181 AVCodec ff_cinepak_encoder = {
1182     .name           = "cinepak",
1183     .long_name      = NULL_IF_CONFIG_SMALL("Cinepak"),
1184     .type           = AVMEDIA_TYPE_VIDEO,
1185     .id             = AV_CODEC_ID_CINEPAK,
1186     .priv_data_size = sizeof(CinepakEncContext),
1187     .init           = cinepak_encode_init,
1188     .encode2        = cinepak_encode_frame,
1189     .close          = cinepak_encode_end,
1190     .pix_fmts       = (const enum AVPixelFormat[]) { AV_PIX_FMT_RGB24, AV_PIX_FMT_GRAY8, AV_PIX_FMT_NONE },
1191     .priv_class     = &cinepak_class,
1192     .caps_internal  = FF_CODEC_CAP_INIT_CLEANUP,
1193 };
1194