LCOV - code coverage report
Current view: top level - libavcodec - cinepakenc.c (source / functions) Hit Total Coverage
Test: coverage.info Lines: 529 597 88.6 %
Date: 2017-12-13 18:07:29 Functions: 18 18 100.0 %

          Line data    Source code
       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             : 
     154           3 : static av_cold int cinepak_encode_init(AVCodecContext *avctx)
     155             : {
     156           3 :     CinepakEncContext *s = avctx->priv_data;
     157             :     int x, mb_count, strip_buf_size, frame_buf_size;
     158             : 
     159           3 :     if (avctx->width & 3 || avctx->height & 3) {
     160           0 :         av_log(avctx, AV_LOG_ERROR, "width and height must be multiples of four (got %ix%i)\n",
     161             :                avctx->width, avctx->height);
     162           0 :         return AVERROR(EINVAL);
     163             :     }
     164             : 
     165           3 :     if (s->min_min_strips > s->max_max_strips) {
     166           0 :         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           0 :         return AVERROR(EINVAL);
     169             :     }
     170             : 
     171           3 :     if (!(s->last_frame = av_frame_alloc()))
     172           0 :         return AVERROR(ENOMEM);
     173           3 :     if (!(s->best_frame = av_frame_alloc()))
     174           0 :         goto enomem;
     175           3 :     if (!(s->scratch_frame = av_frame_alloc()))
     176           0 :         goto enomem;
     177           3 :     if (avctx->pix_fmt == AV_PIX_FMT_RGB24)
     178           3 :         if (!(s->input_frame = av_frame_alloc()))
     179           0 :             goto enomem;
     180             : 
     181           3 :     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           0 :         goto enomem;
     183             : 
     184           3 :     if (!(s->codebook_closest = av_malloc_array((avctx->width * avctx->height) >> 2, sizeof(*s->codebook_closest))))
     185           0 :         goto enomem;
     186             : 
     187          15 :     for (x = 0; x < (avctx->pix_fmt == AV_PIX_FMT_RGB24 ? 4 : 3); x++)
     188          12 :         if (!(s->pict_bufs[x] = av_malloc((avctx->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4) * (avctx->width * avctx->height) >> 2)))
     189           0 :             goto enomem;
     190             : 
     191           3 :     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           3 :     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           3 :     frame_buf_size = CVID_HEADER_SIZE + s->max_max_strips * strip_buf_size;
     200             : 
     201           3 :     if (!(s->strip_buf = av_malloc(strip_buf_size)))
     202           0 :         goto enomem;
     203             : 
     204           3 :     if (!(s->frame_buf = av_malloc(frame_buf_size)))
     205           0 :         goto enomem;
     206             : 
     207           3 :     if (!(s->mb = av_malloc_array(mb_count, sizeof(mb_info))))
     208           0 :         goto enomem;
     209             : 
     210           3 :     av_lfg_init(&s->randctx, 1);
     211           3 :     s->avctx          = avctx;
     212           3 :     s->w              = avctx->width;
     213           3 :     s->h              = avctx->height;
     214           3 :     s->frame_buf_size = frame_buf_size;
     215           3 :     s->curframe       = 0;
     216           3 :     s->keyint         = avctx->keyint_min;
     217           3 :     s->pix_fmt        = avctx->pix_fmt;
     218             : 
     219             :     // set up AVFrames
     220           3 :     s->last_frame->data[0]        = s->pict_bufs[0];
     221           3 :     s->last_frame->linesize[0]    = s->w;
     222           3 :     s->best_frame->data[0]        = s->pict_bufs[1];
     223           3 :     s->best_frame->linesize[0]    = s->w;
     224           3 :     s->scratch_frame->data[0]     = s->pict_bufs[2];
     225           3 :     s->scratch_frame->linesize[0] = s->w;
     226             : 
     227           3 :     if (s->pix_fmt == AV_PIX_FMT_RGB24) {
     228           3 :         s->last_frame->data[1]     = s->last_frame->data[0] +   s->w * s->h;
     229           3 :         s->last_frame->data[2]     = s->last_frame->data[1] + ((s->w * s->h) >> 2);
     230           6 :         s->last_frame->linesize[1] =
     231           6 :         s->last_frame->linesize[2] = s->w >> 1;
     232             : 
     233           3 :         s->best_frame->data[1]     = s->best_frame->data[0] +   s->w * s->h;
     234           3 :         s->best_frame->data[2]     = s->best_frame->data[1] + ((s->w * s->h) >> 2);
     235           6 :         s->best_frame->linesize[1] =
     236           6 :         s->best_frame->linesize[2] = s->w >> 1;
     237             : 
     238           3 :         s->scratch_frame->data[1]     = s->scratch_frame->data[0] +   s->w * s->h;
     239           3 :         s->scratch_frame->data[2]     = s->scratch_frame->data[1] + ((s->w * s->h) >> 2);
     240           6 :         s->scratch_frame->linesize[1] =
     241           6 :         s->scratch_frame->linesize[2] = s->w >> 1;
     242             : 
     243           3 :         s->input_frame->data[0]     = s->pict_bufs[3];
     244           3 :         s->input_frame->linesize[0] = s->w;
     245           3 :         s->input_frame->data[1]     = s->input_frame->data[0] +   s->w * s->h;
     246           3 :         s->input_frame->data[2]     = s->input_frame->data[1] + ((s->w * s->h) >> 2);
     247           6 :         s->input_frame->linesize[1] =
     248           6 :         s->input_frame->linesize[2] = s->w >> 1;
     249             :     }
     250             : 
     251           3 :     s->min_strips = s->min_min_strips;
     252           3 :     s->max_strips = s->max_max_strips;
     253             : 
     254           3 :     return 0;
     255             : 
     256           0 : enomem:
     257           0 :     av_frame_free(&s->last_frame);
     258           0 :     av_frame_free(&s->best_frame);
     259           0 :     av_frame_free(&s->scratch_frame);
     260           0 :     if (avctx->pix_fmt == AV_PIX_FMT_RGB24)
     261           0 :         av_frame_free(&s->input_frame);
     262           0 :     av_freep(&s->codebook_input);
     263           0 :     av_freep(&s->codebook_closest);
     264           0 :     av_freep(&s->strip_buf);
     265           0 :     av_freep(&s->frame_buf);
     266           0 :     av_freep(&s->mb);
     267             : 
     268           0 :     for (x = 0; x < (avctx->pix_fmt == AV_PIX_FMT_RGB24 ? 4 : 3); x++)
     269           0 :         av_freep(&s->pict_bufs[x]);
     270             : 
     271           0 :     return AVERROR(ENOMEM);
     272             : }
     273             : 
     274       48249 : static int64_t calculate_mode_score(CinepakEncContext *s, int h,
     275             :                                     strip_info *info, int report,
     276             :                                     int *training_set_v1_shrunk,
     277             :                                     int *training_set_v4_shrunk)
     278             : {
     279             :     // score = FF_LAMBDA_SCALE * error + lambda * bits
     280             :     int x;
     281       48249 :     int entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4;
     282       48249 :     int mb_count   = s->w * h / MB_AREA;
     283             :     mb_info *mb;
     284             :     int64_t score1, score2, score3;
     285      192996 :     int64_t ret = s->lambda * ((info->v1_size ? CHUNK_HEADER_SIZE + info->v1_size * entry_size : 0) +
     286       96498 :                                (info->v4_size ? CHUNK_HEADER_SIZE + info->v4_size * entry_size : 0) +
     287       48249 :                                CHUNK_HEADER_SIZE) << 3;
     288             : 
     289       48249 :     switch (info->mode) {
     290       10542 :     case MODE_V1_ONLY:
     291             :         // one byte per MB
     292       10542 :         ret += s->lambda * 8 * mb_count;
     293             : 
     294             :         // while calculating we assume all blocks are ENC_V1
     295     3268398 :         for (x = 0; x < mb_count; x++) {
     296     3257856 :             mb   = &s->mb[x];
     297     3257856 :             ret += FF_LAMBDA_SCALE * mb->v1_error;
     298             :             // this function is never called for report in MODE_V1_ONLY
     299             :             // if (!report)
     300     3257856 :             mb->best_encoding = ENC_V1;
     301             :         }
     302             : 
     303       10542 :         break;
     304       13554 :     case MODE_V1_V4:
     305             :         // 9 or 33 bits per MB
     306       13554 :         if (report) {
     307             :             // no moves between the corresponding training sets are allowed
     308        6777 :             *training_set_v1_shrunk = *training_set_v4_shrunk = 0;
     309     2101113 :             for (x = 0; x < mb_count; x++) {
     310             :                 int mberr;
     311     2094336 :                 mb = &s->mb[x];
     312     2094336 :                 if (mb->best_encoding == ENC_V1)
     313      664899 :                     score1 = s->lambda * 9 + FF_LAMBDA_SCALE * (mberr = mb->v1_error);
     314             :                 else
     315     1429437 :                     score1 = s->lambda * 33 + FF_LAMBDA_SCALE * (mberr = mb->v4_error);
     316     2094336 :                 ret += score1;
     317             :             }
     318             :         } else { // find best mode per block
     319     2101113 :             for (x = 0; x < mb_count; x++) {
     320     2094336 :                 mb     = &s->mb[x];
     321     2094336 :                 score1 = s->lambda * 9 + FF_LAMBDA_SCALE * mb->v1_error;
     322     2094336 :                 score2 = s->lambda * 33 + FF_LAMBDA_SCALE * mb->v4_error;
     323             : 
     324     2094336 :                 if (score1 <= score2) {
     325      664899 :                     ret += score1;
     326      664899 :                     mb->best_encoding = ENC_V1;
     327             :                 } else {
     328     1429437 :                     ret += score2;
     329     1429437 :                     mb->best_encoding = ENC_V4;
     330             :                 }
     331             :             }
     332             :         }
     333             : 
     334       13554 :         break;
     335       24153 :     case MODE_MC:
     336             :         // 1, 10 or 34 bits per MB
     337       24153 :         if (report) {
     338       17673 :             int v1_shrunk = 0, v4_shrunk = 0;
     339     5471753 :             for (x = 0; x < mb_count; x++) {
     340     5454080 :                 mb = &s->mb[x];
     341             :                 // it is OK to move blocks to ENC_SKIP here
     342             :                 // but not to any codebook encoding!
     343     5454080 :                 score1 = s->lambda * 1 + FF_LAMBDA_SCALE * mb->skip_error;
     344     5454080 :                 if (mb->best_encoding == ENC_SKIP) {
     345     2516362 :                     ret += score1;
     346     2937718 :                 } else if (mb->best_encoding == ENC_V1) {
     347      830323 :                     if ((score2 = s->lambda * 10 + FF_LAMBDA_SCALE * mb->v1_error) >= score1) {
     348       24137 :                         mb->best_encoding = ENC_SKIP;
     349       24137 :                         ++v1_shrunk;
     350       24137 :                         ret += score1;
     351             :                     } else {
     352      806186 :                         ret += score2;
     353             :                     }
     354             :                 } else {
     355     2107395 :                     if ((score3 = s->lambda * 34 + FF_LAMBDA_SCALE * mb->v4_error) >= score1) {
     356       75551 :                         mb->best_encoding = ENC_SKIP;
     357       75551 :                         ++v4_shrunk;
     358       75551 :                         ret += score1;
     359             :                     } else {
     360     2031844 :                         ret += score3;
     361             :                     }
     362             :                 }
     363             :             }
     364       17673 :             *training_set_v1_shrunk = v1_shrunk;
     365       17673 :             *training_set_v4_shrunk = v4_shrunk;
     366             :         } else { // find best mode per block
     367     1997136 :             for (x = 0; x < mb_count; x++) {
     368     1990656 :                 mb     = &s->mb[x];
     369     1990656 :                 score1 = s->lambda * 1 + FF_LAMBDA_SCALE * mb->skip_error;
     370     1990656 :                 score2 = s->lambda * 10 + FF_LAMBDA_SCALE * mb->v1_error;
     371     1990656 :                 score3 = s->lambda * 34 + FF_LAMBDA_SCALE * mb->v4_error;
     372             : 
     373     1990656 :                 if (score1 <= score2 && score1 <= score3) {
     374      842199 :                     ret += score1;
     375      842199 :                     mb->best_encoding = ENC_SKIP;
     376     1148457 :                 } else if (score2 <= score3) {
     377      314890 :                     ret += score2;
     378      314890 :                     mb->best_encoding = ENC_V1;
     379             :                 } else {
     380      833567 :                     ret += score3;
     381      833567 :                     mb->best_encoding = ENC_V4;
     382             :                 }
     383             :             }
     384             :         }
     385             : 
     386       24153 :         break;
     387             :     }
     388             : 
     389       48249 :     return ret;
     390             : }
     391             : 
     392       30759 : static int write_chunk_header(unsigned char *buf, int chunk_type, int chunk_size)
     393             : {
     394       30759 :     buf[0] = chunk_type;
     395       30759 :     AV_WB24(&buf[1], chunk_size + CHUNK_HEADER_SIZE);
     396       30759 :     return CHUNK_HEADER_SIZE;
     397             : }
     398             : 
     399       20506 : static int encode_codebook(CinepakEncContext *s, int *codebook, int size,
     400             :                            int chunk_type_yuv, int chunk_type_gray,
     401             :                            unsigned char *buf)
     402             : {
     403       20506 :     int x, y, ret, entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4;
     404       20506 :     int incremental_codebook_replacement_mode = 0; // hardcoded here,
     405             :     // the compiler should notice that this is a constant -- rl
     406             : 
     407       61518 :     ret = write_chunk_header(buf,
     408       20506 :                              s->pix_fmt == AV_PIX_FMT_RGB24 ?
     409       20506 :                              chunk_type_yuv  + (incremental_codebook_replacement_mode ? 1 : 0) :
     410           0 :                              chunk_type_gray + (incremental_codebook_replacement_mode ? 1 : 0),
     411       20506 :                              entry_size * size +
     412       20506 :                              (incremental_codebook_replacement_mode ? (size + 31) / 32 * 4 : 0));
     413             : 
     414             :     // we do codebook encoding according to the "intra" mode
     415             :     // but we keep the "dead" code for reference in case we will want
     416             :     // to use incremental codebook updates (which actually would give us
     417             :     // "kind of" motion compensation, especially in 1 strip/frame case) -- rl
     418             :     // (of course, the code will be not useful as-is)
     419       20506 :     if (incremental_codebook_replacement_mode) {
     420           0 :         int flags = 0;
     421             :         int flagsind;
     422           0 :         for (x = 0; x < size; x++) {
     423           0 :             if (flags == 0) {
     424           0 :                 flagsind = ret;
     425           0 :                 ret     += 4;
     426           0 :                 flags    = 0x80000000;
     427             :             } else
     428           0 :                 flags = ((flags >> 1) | 0x80000000);
     429           0 :             for (y = 0; y < entry_size; y++)
     430           0 :                 buf[ret++] = codebook[y + x * entry_size] ^ (y >= 4 ? 0x80 : 0);
     431           0 :             if ((flags & 0xffffffff) == 0xffffffff) {
     432           0 :                 AV_WB32(&buf[flagsind], flags);
     433           0 :                 flags = 0;
     434             :             }
     435             :         }
     436           0 :         if (flags)
     437           0 :             AV_WB32(&buf[flagsind], flags);
     438             :     } else
     439      751672 :         for (x = 0; x < size; x++)
     440     5118162 :             for (y = 0; y < entry_size; y++)
     441     4386996 :                 buf[ret++] = codebook[y + x * entry_size] ^ (y >= 4 ? 0x80 : 0);
     442             : 
     443       20506 :     return ret;
     444             : }
     445             : 
     446             : // sets out to the sub picture starting at (x,y) in in
     447    14989880 : static void get_sub_picture(CinepakEncContext *s, int x, int y,
     448             :                             uint8_t * in_data[4], int  in_linesize[4],
     449             :                             uint8_t *out_data[4], int out_linesize[4])
     450             : {
     451    14989880 :     out_data[0]     = in_data[0] + x + y * in_linesize[0];
     452    14989880 :     out_linesize[0] = in_linesize[0];
     453             : 
     454    14989880 :     if (s->pix_fmt == AV_PIX_FMT_RGB24) {
     455    14989880 :         out_data[1]     = in_data[1] + (x >> 1) + (y >> 1) * in_linesize[1];
     456    14989880 :         out_linesize[1] = in_linesize[1];
     457             : 
     458    14989880 :         out_data[2]     = in_data[2] + (x >> 1) + (y >> 1) * in_linesize[2];
     459    14989880 :         out_linesize[2] = in_linesize[2];
     460             :     }
     461    14989880 : }
     462             : 
     463             : // decodes the V1 vector in mb into the 4x4 MB pointed to by data
     464     5323426 : static void decode_v1_vector(CinepakEncContext *s, uint8_t *data[4],
     465             :                              int linesize[4], int v1_vector, strip_info *info)
     466             : {
     467     5323426 :     int entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4;
     468             : 
     469    10646852 :     data[0][0] =
     470    15970278 :     data[0][1] =
     471    15970278 :     data[0][    linesize[0]] =
     472    10646852 :     data[0][1 + linesize[0]] = info->v1_codebook[v1_vector * entry_size];
     473             : 
     474    10646852 :     data[0][2] =
     475    15970278 :     data[0][3] =
     476    15970278 :     data[0][2 + linesize[0]] =
     477    10646852 :     data[0][3 + linesize[0]] = info->v1_codebook[v1_vector * entry_size + 1];
     478             : 
     479    10646852 :     data[0][    2 * linesize[0]] =
     480    15970278 :     data[0][1 + 2 * linesize[0]] =
     481    15970278 :     data[0][    3 * linesize[0]] =
     482    10646852 :     data[0][1 + 3 * linesize[0]] = info->v1_codebook[v1_vector * entry_size + 2];
     483             : 
     484    10646852 :     data[0][2 + 2 * linesize[0]] =
     485    15970278 :     data[0][3 + 2 * linesize[0]] =
     486    15970278 :     data[0][2 + 3 * linesize[0]] =
     487    10646852 :     data[0][3 + 3 * linesize[0]] = info->v1_codebook[v1_vector * entry_size + 3];
     488             : 
     489     5323426 :     if (s->pix_fmt == AV_PIX_FMT_RGB24) {
     490    10646852 :         data[1][0] =
     491    15970278 :         data[1][1] =
     492    15970278 :         data[1][    linesize[1]] =
     493    10646852 :         data[1][1 + linesize[1]] = info->v1_codebook[v1_vector * entry_size + 4];
     494             : 
     495    10646852 :         data[2][0] =
     496    15970278 :         data[2][1] =
     497    15970278 :         data[2][    linesize[2]] =
     498    10646852 :         data[2][1 + linesize[2]] = info->v1_codebook[v1_vector * entry_size + 5];
     499             :     }
     500     5323426 : }
     501             : 
     502             : // decodes the V4 vectors in mb into the 4x4 MB pointed to by data
     503     6993789 : static void decode_v4_vector(CinepakEncContext *s, uint8_t *data[4],
     504             :                              int linesize[4], int *v4_vector, strip_info *info)
     505             : {
     506     6993789 :     int i, x, y, entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4;
     507             : 
     508    20981367 :     for (i = y = 0; y < 4; y += 2) {
     509    41962734 :         for (x = 0; x < 4; x += 2, i++) {
     510    27975156 :             data[0][x     +  y      * linesize[0]] = info->v4_codebook[v4_vector[i] * entry_size];
     511    27975156 :             data[0][x + 1 +  y      * linesize[0]] = info->v4_codebook[v4_vector[i] * entry_size + 1];
     512    27975156 :             data[0][x     + (y + 1) * linesize[0]] = info->v4_codebook[v4_vector[i] * entry_size + 2];
     513    27975156 :             data[0][x + 1 + (y + 1) * linesize[0]] = info->v4_codebook[v4_vector[i] * entry_size + 3];
     514             : 
     515    27975156 :             if (s->pix_fmt == AV_PIX_FMT_RGB24) {
     516    27975156 :                 data[1][(x >> 1) + (y >> 1) * linesize[1]] = info->v4_codebook[v4_vector[i] * entry_size + 4];
     517    27975156 :                 data[2][(x >> 1) + (y >> 1) * linesize[2]] = info->v4_codebook[v4_vector[i] * entry_size + 5];
     518             :             }
     519             :         }
     520             :     }
     521     6993789 : }
     522             : 
     523      883619 : static void copy_mb(CinepakEncContext *s,
     524             :                     uint8_t *a_data[4], int a_linesize[4],
     525             :                     uint8_t *b_data[4], int b_linesize[4])
     526             : {
     527             :     int y, p;
     528             : 
     529     4418095 :     for (y = 0; y < MB_SIZE; y++)
     530     3534476 :         memcpy(a_data[0] + y * a_linesize[0], b_data[0] + y * b_linesize[0],
     531             :                MB_SIZE);
     532             : 
     533      883619 :     if (s->pix_fmt == AV_PIX_FMT_RGB24) {
     534     2650857 :         for (p = 1; p <= 2; p++)
     535     5301714 :             for (y = 0; y < MB_SIZE / 2; y++)
     536     3534476 :                 memcpy(a_data[p] + y * a_linesize[p],
     537     3534476 :                        b_data[p] + y * b_linesize[p],
     538             :                        MB_SIZE / 2);
     539             :     }
     540      883619 : }
     541             : 
     542       10253 : static int encode_mode(CinepakEncContext *s, int h,
     543             :                        uint8_t *scratch_data[4], int scratch_linesize[4],
     544             :                        uint8_t *last_data[4], int last_linesize[4],
     545             :                        strip_info *info, unsigned char *buf)
     546             : {
     547       10253 :     int x, y, z, flags, bits, temp_size, header_ofs, ret = 0, mb_count = s->w * h / MB_AREA;
     548             :     int needs_extra_bit, should_write_temp;
     549             :     unsigned char temp[64]; // 32/2 = 16 V4 blocks at 4 B each -> 64 B
     550             :     mb_info *mb;
     551       10253 :     uint8_t *sub_scratch_data[4] = { 0 }, *sub_last_data[4] = { 0 };
     552       10253 :     int sub_scratch_linesize[4] = { 0 }, sub_last_linesize[4] = { 0 };
     553             : 
     554             :     // encode codebooks
     555             :     ////// MacOS vintage decoder compatibility dictates the presence of
     556             :     ////// the codebook chunk even when the codebook is empty - pretty dumb...
     557             :     ////// and also the certain order of the codebook chunks -- rl
     558       10253 :     if (info->v4_size || !s->skip_empty_cb)
     559       10253 :         ret += encode_codebook(s, info->v4_codebook, info->v4_size, 0x20, 0x24, buf + ret);
     560             : 
     561       10253 :     if (info->v1_size || !s->skip_empty_cb)
     562       10253 :         ret += encode_codebook(s, info->v1_codebook, info->v1_size, 0x22, 0x26, buf + ret);
     563             : 
     564             :     // update scratch picture
     565      109405 :     for (z = y = 0; y < h; y += MB_SIZE)
     566     3272016 :         for (x = 0; x < s->w; x += MB_SIZE, z++) {
     567     3172864 :             mb = &s->mb[z];
     568             : 
     569     3172864 :             get_sub_picture(s, x, y, scratch_data, scratch_linesize,
     570             :                             sub_scratch_data, sub_scratch_linesize);
     571             : 
     572     3172864 :             if (info->mode == MODE_MC && mb->best_encoding == ENC_SKIP) {
     573      883619 :                 get_sub_picture(s, x, y, last_data, last_linesize,
     574             :                                 sub_last_data, sub_last_linesize);
     575      883619 :                 copy_mb(s, sub_scratch_data, sub_scratch_linesize,
     576             :                         sub_last_data, sub_last_linesize);
     577     2289245 :             } else if (info->mode == MODE_V1_ONLY || mb->best_encoding == ENC_V1)
     578      828313 :                 decode_v1_vector(s, sub_scratch_data, sub_scratch_linesize,
     579             :                                  mb->v1_vector, info);
     580             :             else
     581     1460932 :                 decode_v4_vector(s, sub_scratch_data, sub_scratch_linesize,
     582     1460932 :                                  mb->v4_vector, info);
     583             :         }
     584             : 
     585       10253 :     switch (info->mode) {
     586        1018 :     case MODE_V1_ONLY:
     587        1018 :         ret += write_chunk_header(buf + ret, 0x32, mb_count);
     588             : 
     589      316922 :         for (x = 0; x < mb_count; x++)
     590      315904 :             buf[ret++] = s->mb[x].v1_vector;
     591             : 
     592        1018 :         break;
     593        3665 :     case MODE_V1_V4:
     594             :         // remember header position
     595        3665 :         header_ofs = ret;
     596        3665 :         ret       += CHUNK_HEADER_SIZE;
     597             : 
     598       39285 :         for (x = 0; x < mb_count; x += 32) {
     599       35620 :             flags = 0;
     600     1175460 :             for (y = x; y < FFMIN(x + 32, mb_count); y++)
     601     1139840 :                 if (s->mb[y].best_encoding == ENC_V4)
     602      861679 :                     flags |= 1 << (31 - y + x);
     603             : 
     604       35620 :             AV_WB32(&buf[ret], flags);
     605       35620 :             ret += 4;
     606             : 
     607     1175460 :             for (y = x; y < FFMIN(x + 32, mb_count); y++) {
     608     1139840 :                 mb = &s->mb[y];
     609             : 
     610     1139840 :                 if (mb->best_encoding == ENC_V1)
     611      278161 :                     buf[ret++] = mb->v1_vector;
     612             :                 else
     613     4308395 :                     for (z = 0; z < 4; z++)
     614     3446716 :                         buf[ret++] = mb->v4_vector[z];
     615             :             }
     616             :         }
     617             : 
     618        3665 :         write_chunk_header(buf + header_ofs, 0x30, ret - header_ofs - CHUNK_HEADER_SIZE);
     619             : 
     620        3665 :         break;
     621        5570 :     case MODE_MC:
     622             :         // remember header position
     623        5570 :         header_ofs = ret;
     624        5570 :         ret       += CHUNK_HEADER_SIZE;
     625        5570 :         flags      = bits = temp_size = 0;
     626             : 
     627     1722690 :         for (x = 0; x < mb_count; x++) {
     628     1717120 :             mb                = &s->mb[x];
     629     1717120 :             flags            |= (mb->best_encoding != ENC_SKIP) << (31 - bits++);
     630     1717120 :             needs_extra_bit   = 0;
     631     1717120 :             should_write_temp = 0;
     632             : 
     633     1717120 :             if (mb->best_encoding != ENC_SKIP) {
     634      833501 :                 if (bits < 32)
     635      810194 :                     flags |= (mb->best_encoding == ENC_V4) << (31 - bits++);
     636             :                 else
     637       23307 :                     needs_extra_bit = 1;
     638             :             }
     639             : 
     640     1717120 :             if (bits == 32) {
     641       76945 :                 AV_WB32(&buf[ret], flags);
     642       76945 :                 ret  += 4;
     643       76945 :                 flags = bits = 0;
     644             : 
     645       76945 :                 if (mb->best_encoding == ENC_SKIP || needs_extra_bit) {
     646       50298 :                     memcpy(&buf[ret], temp, temp_size);
     647       50298 :                     ret      += temp_size;
     648       50298 :                     temp_size = 0;
     649             :                 } else
     650       26647 :                     should_write_temp = 1;
     651             :             }
     652             : 
     653     1717120 :             if (needs_extra_bit) {
     654       23307 :                 flags = (mb->best_encoding == ENC_V4) << 31;
     655       23307 :                 bits  = 1;
     656             :             }
     657             : 
     658     1717120 :             if (mb->best_encoding == ENC_V1)
     659      234248 :                 temp[temp_size++] = mb->v1_vector;
     660     1482872 :             else if (mb->best_encoding == ENC_V4)
     661     2996265 :                 for (z = 0; z < 4; z++)
     662     2397012 :                     temp[temp_size++] = mb->v4_vector[z];
     663             : 
     664     1717120 :             if (should_write_temp) {
     665       26647 :                 memcpy(&buf[ret], temp, temp_size);
     666       26647 :                 ret      += temp_size;
     667       26647 :                 temp_size = 0;
     668             :             }
     669             :         }
     670             : 
     671        5570 :         if (bits > 0) {
     672        5386 :             AV_WB32(&buf[ret], flags);
     673        5386 :             ret += 4;
     674        5386 :             memcpy(&buf[ret], temp, temp_size);
     675        5386 :             ret += temp_size;
     676             :         }
     677             : 
     678        5570 :         write_chunk_header(buf + header_ofs, 0x31, ret - header_ofs - CHUNK_HEADER_SIZE);
     679             : 
     680        5570 :         break;
     681             :     }
     682             : 
     683       10253 :     return ret;
     684             : }
     685             : 
     686             : // computes distortion of 4x4 MB in b compared to a
     687    10249154 : static int compute_mb_distortion(CinepakEncContext *s,
     688             :                                  uint8_t *a_data[4], int a_linesize[4],
     689             :                                  uint8_t *b_data[4], int b_linesize[4])
     690             : {
     691    10249154 :     int x, y, p, d, ret = 0;
     692             : 
     693    51245770 :     for (y = 0; y < MB_SIZE; y++)
     694   204983080 :         for (x = 0; x < MB_SIZE; x++) {
     695   163986464 :             d = a_data[0][x + y * a_linesize[0]] - b_data[0][x + y * b_linesize[0]];
     696   163986464 :             ret += d * d;
     697             :         }
     698             : 
     699    10249154 :     if (s->pix_fmt == AV_PIX_FMT_RGB24) {
     700    30747462 :         for (p = 1; p <= 2; p++) {
     701    61494924 :             for (y = 0; y < MB_SIZE / 2; y++)
     702   122989848 :                 for (x = 0; x < MB_SIZE / 2; x++) {
     703    81993232 :                     d = a_data[p][x + y * a_linesize[p]] - b_data[p][x + y * b_linesize[p]];
     704    81993232 :                     ret += d * d;
     705             :                 }
     706             :         }
     707             :     }
     708             : 
     709    10249154 :     return ret;
     710             : }
     711             : 
     712             : // return the possibly adjusted size of the codebook
     713             : #define CERTAIN(x) ((x) != ENC_UNCERTAIN)
     714       58459 : static int quantize(CinepakEncContext *s, int h, uint8_t *data[4],
     715             :                     int linesize[4], int v1mode, strip_info *info,
     716             :                     mb_encoding encoding)
     717             : {
     718             :     int x, y, i, j, k, x2, y2, x3, y3, plane, shift, mbn;
     719       58459 :     int entry_size      = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4;
     720       58459 :     int *codebook       = v1mode ? info->v1_codebook : info->v4_codebook;
     721       58459 :     int size            = v1mode ? info->v1_size : info->v4_size;
     722       58459 :     int64_t total_error = 0;
     723             :     uint8_t vq_pict_buf[(MB_AREA * 3) / 2];
     724             :     uint8_t     *sub_data[4],     *vq_data[4];
     725             :     int      sub_linesize[4],  vq_linesize[4];
     726             : 
     727      623443 :     for (mbn = i = y = 0; y < h; y += MB_SIZE) {
     728    18644472 :         for (x = 0; x < s->w; x += MB_SIZE, ++mbn) {
     729             :             int *base;
     730             : 
     731    18079488 :             if (CERTAIN(encoding)) {
     732             :                 // use for the training only the blocks known to be to be encoded [sic:-]
     733    12727296 :                 if (s->mb[mbn].best_encoding != encoding)
     734     8051518 :                     continue;
     735             :             }
     736             : 
     737    10027970 :             base = s->codebook_input + i * entry_size;
     738    10027970 :             if (v1mode) {
     739             :                 // subsample
     740    17980452 :                 for (j = y2 = 0; y2 < entry_size; y2 += 2)
     741    40456017 :                     for (x2 = 0; x2 < 4; x2 += 2, j++) {
     742    26970678 :                         plane   = y2 < 4 ? 0 : 1 + (x2 >> 1);
     743    26970678 :                         shift   = y2 < 4 ? 0 : 1;
     744    26970678 :                         x3      = shift ? 0 : x2;
     745    26970678 :                         y3      = shift ? 0 : y2;
     746   107882712 :                         base[j] = (data[plane][((x + x3) >> shift) +      ((y + y3) >> shift)      * linesize[plane]] +
     747    53941356 :                                    data[plane][((x + x3) >> shift) + 1 +  ((y + y3) >> shift)      * linesize[plane]] +
     748    53941356 :                                    data[plane][((x + x3) >> shift) +     (((y + y3) >> shift) + 1) * linesize[plane]] +
     749    53941356 :                                    data[plane][((x + x3) >> shift) + 1 + (((y + y3) >> shift) + 1) * linesize[plane]]) >> 2;
     750             :                     }
     751             :             } else {
     752             :                 // copy
     753    16598571 :                 for (j = y2 = 0; y2 < MB_SIZE; y2 += 2) {
     754    33197142 :                     for (x2 = 0; x2 < MB_SIZE; x2 += 2)
     755   154919996 :                         for (k = 0; k < entry_size; k++, j++) {
     756   132788568 :                             plane = k >= 4 ? k - 3 : 0;
     757             : 
     758   132788568 :                             if (k >= 4) {
     759    44262856 :                                 x3 = (x + x2) >> 1;
     760    44262856 :                                 y3 = (y + y2) >> 1;
     761             :                             } else {
     762    88525712 :                                 x3 = x + x2 + (k & 1);
     763    88525712 :                                 y3 = y + y2 + (k >> 1);
     764             :                             }
     765             : 
     766   132788568 :                             base[j] = data[plane][x3 + y3 * linesize[plane]];
     767             :                         }
     768             :                 }
     769             :             }
     770    10027970 :             i += v1mode ? 1 : 4;
     771             :         }
     772             :     }
     773             : 
     774       58459 :     if (i == 0) // empty training set, nothing to do
     775          34 :         return 0;
     776       58425 :     if (i < size)
     777        5040 :         size = i;
     778             : 
     779       58425 :     avpriv_init_elbg(s->codebook_input, entry_size, i, codebook, size, 1, s->codebook_closest, &s->randctx);
     780       58425 :     avpriv_do_elbg(s->codebook_input, entry_size, i, codebook, size, 1, s->codebook_closest, &s->randctx);
     781             : 
     782             :     // set up vq_data, which contains a single MB
     783       58425 :     vq_data[0]     = vq_pict_buf;
     784       58425 :     vq_linesize[0] = MB_SIZE;
     785       58425 :     vq_data[1]     = &vq_pict_buf[MB_AREA];
     786       58425 :     vq_data[2]     = vq_data[1] + (MB_AREA >> 2);
     787       58425 :     vq_linesize[1] =
     788       58425 :     vq_linesize[2] = MB_SIZE >> 1;
     789             : 
     790             :     // copy indices
     791      623121 :     for (i = j = y = 0; y < h; y += MB_SIZE)
     792    18634968 :         for (x = 0; x < s->w; x += MB_SIZE, j++) {
     793    18070272 :             mb_info *mb = &s->mb[j];
     794             :             // skip uninteresting blocks if we know their preferred encoding
     795    18070272 :             if (CERTAIN(encoding) && mb->best_encoding != encoding)
     796     8042302 :                 continue;
     797             : 
     798             :             // point sub_data to current MB
     799    10027970 :             get_sub_picture(s, x, y, data, linesize, sub_data, sub_linesize);
     800             : 
     801    10027970 :             if (v1mode) {
     802     4495113 :                 mb->v1_vector = s->codebook_closest[i];
     803             : 
     804             :                 // fill in vq_data with V1 data
     805     4495113 :                 decode_v1_vector(s, vq_data, vq_linesize, mb->v1_vector, info);
     806             : 
     807     4495113 :                 mb->v1_error = compute_mb_distortion(s, sub_data, sub_linesize,
     808             :                                                      vq_data, vq_linesize);
     809     4495113 :                 total_error += mb->v1_error;
     810             :             } else {
     811    27664285 :                 for (k = 0; k < 4; k++)
     812    22131428 :                     mb->v4_vector[k] = s->codebook_closest[i + k];
     813             : 
     814             :                 // fill in vq_data with V4 data
     815     5532857 :                 decode_v4_vector(s, vq_data, vq_linesize, mb->v4_vector, info);
     816             : 
     817     5532857 :                 mb->v4_error = compute_mb_distortion(s, sub_data, sub_linesize,
     818             :                                                      vq_data, vq_linesize);
     819     5532857 :                 total_error += mb->v4_error;
     820             :             }
     821    10027970 :             i += v1mode ? 1 : 4;
     822             :         }
     823             :     // check that we did it right in the beginning of the function
     824       58425 :     av_assert0(i >= size); // training set is no smaller than the codebook
     825             : 
     826       58425 :     return size;
     827             : }
     828             : 
     829         720 : static void calculate_skip_errors(CinepakEncContext *s, int h,
     830             :                                   uint8_t *last_data[4], int last_linesize[4],
     831             :                                   uint8_t *data[4], int linesize[4],
     832             :                                   strip_info *info)
     833             : {
     834             :     int x, y, i;
     835             :     uint8_t *sub_last_data    [4], *sub_pict_data    [4];
     836             :     int      sub_last_linesize[4],  sub_pict_linesize[4];
     837             : 
     838        7632 :     for (i = y = 0; y < h; y += MB_SIZE)
     839      228096 :         for (x = 0; x < s->w; x += MB_SIZE, i++) {
     840      221184 :             get_sub_picture(s, x, y, last_data, last_linesize,
     841             :                             sub_last_data, sub_last_linesize);
     842      221184 :             get_sub_picture(s, x, y, data, linesize,
     843             :                             sub_pict_data, sub_pict_linesize);
     844             : 
     845      442368 :             s->mb[i].skip_error =
     846      221184 :                 compute_mb_distortion(s,
     847             :                                       sub_last_data, sub_last_linesize,
     848             :                                       sub_pict_data, sub_pict_linesize);
     849             :         }
     850         720 : }
     851             : 
     852       10253 : static void write_strip_header(CinepakEncContext *s, int y, int h, int keyframe,
     853             :                                unsigned char *buf, int strip_size)
     854             : {
     855             :     // actually we are exclusively using intra strip coding (how much can we win
     856             :     // otherwise? how to choose which part of a codebook to update?),
     857             :     // keyframes are different only because we disallow ENC_SKIP on them -- rl
     858             :     // (besides, the logic here used to be inverted: )
     859             :     //    buf[0] = keyframe ? 0x11: 0x10;
     860       10253 :     buf[0] = keyframe ? 0x10 : 0x11;
     861       10253 :     AV_WB24(&buf[1], strip_size + STRIP_HEADER_SIZE);
     862             :     // AV_WB16(&buf[4], y); /* using absolute y values works -- rl */
     863       10253 :     AV_WB16(&buf[4], 0); /* using relative values works as well -- rl */
     864       10253 :     AV_WB16(&buf[6], 0);
     865             :     // AV_WB16(&buf[8], y + h); /* using absolute y values works -- rl */
     866       10253 :     AV_WB16(&buf[8], h); /* using relative values works as well -- rl */
     867       10253 :     AV_WB16(&buf[10], s->w);
     868       10253 : }
     869             : 
     870         753 : static int rd_strip(CinepakEncContext *s, int y, int h, int keyframe,
     871             :                     uint8_t *last_data[4], int last_linesize[4],
     872             :                     uint8_t *data[4], int linesize[4],
     873             :                     uint8_t *scratch_data[4], int scratch_linesize[4],
     874             :                     unsigned char *buf, int64_t *best_score)
     875             : {
     876         753 :     int64_t score = 0;
     877         753 :     int best_size = 0;
     878             :     strip_info info;
     879             :     // for codebook optimization:
     880             :     int v1enough, v1_size, v4enough, v4_size;
     881             :     int new_v1_size, new_v4_size;
     882             :     int v1shrunk, v4shrunk;
     883             : 
     884         753 :     if (!keyframe)
     885         720 :         calculate_skip_errors(s, h, last_data, last_linesize, data, linesize,
     886             :                               &info);
     887             : 
     888             :     // try some powers of 4 for the size of the codebooks
     889             :     // constraint the v4 codebook to be no bigger than v1 one,
     890             :     // (and no less than v1_size/4)
     891             :     // thus making v1 preferable and possibly losing small details? should be ok
     892             : #define SMALLEST_CODEBOOK 1
     893        4518 :     for (v1enough = 0, v1_size = SMALLEST_CODEBOOK; v1_size <= CODEBOOK_MAX && !v1enough; v1_size <<= 2) {
     894       14307 :         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) {
     895             :             CinepakMode mode;
     896             :             // try all modes
     897       42168 :             for (mode = 0; mode < MODE_COUNT; mode++) {
     898             :                 // don't allow MODE_MC in intra frames
     899       31626 :                 if (keyframe && mode == MODE_MC)
     900         462 :                     continue;
     901             : 
     902       31164 :                 if (mode == MODE_V1_ONLY) {
     903       10542 :                     info.v1_size = v1_size;
     904             :                     // the size may shrink even before optimizations if the input is short:
     905       10542 :                     info.v1_size = quantize(s, h, data, linesize, 1,
     906             :                                             &info, ENC_UNCERTAIN);
     907       10542 :                     if (info.v1_size < v1_size)
     908             :                         // too few eligible blocks, no sense in trying bigger sizes
     909           0 :                         v1enough = 1;
     910             : 
     911       10542 :                     info.v4_size = 0;
     912             :                 } else { // mode != MODE_V1_ONLY
     913             :                     // if v4 codebook is empty then only allow V1-only mode
     914       20622 :                     if (!v4_size)
     915        7365 :                         continue;
     916             : 
     917       13257 :                     if (mode == MODE_V1_V4) {
     918        6777 :                         info.v4_size = v4_size;
     919        6777 :                         info.v4_size = quantize(s, h, data, linesize, 0,
     920             :                                                 &info, ENC_UNCERTAIN);
     921        6777 :                         if (info.v4_size < v4_size)
     922             :                             // too few eligible blocks, no sense in trying bigger sizes
     923           0 :                             v4enough = 1;
     924             :                     }
     925             :                 }
     926             : 
     927       23799 :                 info.mode = mode;
     928             :                 // choose the best encoding per block, based on current experience
     929       23799 :                 score = calculate_mode_score(s, h, &info, 0,
     930             :                                              &v1shrunk, &v4shrunk);
     931             : 
     932       23799 :                 if (mode != MODE_V1_ONLY) {
     933       13257 :                     int extra_iterations_limit = s->max_extra_cb_iterations;
     934             :                     // recompute the codebooks, omitting the extra blocks
     935             :                     // we assume we _may_ come here with more blocks to encode than before
     936       13257 :                     info.v1_size = v1_size;
     937       13257 :                     new_v1_size = quantize(s, h, data, linesize, 1, &info, ENC_V1);
     938       13257 :                     if (new_v1_size < info.v1_size)
     939        5041 :                         info.v1_size = new_v1_size;
     940             :                     // we assume we _may_ come here with more blocks to encode than before
     941       13257 :                     info.v4_size = v4_size;
     942       13257 :                     new_v4_size = quantize(s, h, data, linesize, 0, &info, ENC_V4);
     943       13257 :                     if (new_v4_size < info.v4_size)
     944          19 :                         info.v4_size = new_v4_size;
     945             :                     // calculate the resulting score
     946             :                     // (do not move blocks to codebook encodings now, as some blocks may have
     947             :                     // got bigger errors despite a smaller training set - but we do not
     948             :                     // ever grow the training sets back)
     949             :                     for (;;) {
     950       35643 :                         score = calculate_mode_score(s, h, &info, 1,
     951             :                                                      &v1shrunk, &v4shrunk);
     952             :                         // do we have a reason to reiterate? if so, have we reached the limit?
     953       24450 :                         if ((!v1shrunk && !v4shrunk) || !extra_iterations_limit--)
     954             :                             break;
     955             :                         // recompute the codebooks, omitting the extra blocks
     956       11193 :                         if (v1shrunk) {
     957        4984 :                             info.v1_size = v1_size;
     958        4984 :                             new_v1_size = quantize(s, h, data, linesize, 1, &info, ENC_V1);
     959        4984 :                             if (new_v1_size < info.v1_size)
     960          12 :                                 info.v1_size = new_v1_size;
     961             :                         }
     962       11193 :                         if (v4shrunk) {
     963        9642 :                             info.v4_size = v4_size;
     964        9642 :                             new_v4_size = quantize(s, h, data, linesize, 0, &info, ENC_V4);
     965        9642 :                             if (new_v4_size < info.v4_size)
     966           2 :                                 info.v4_size = new_v4_size;
     967             :                         }
     968             :                     }
     969             :                 }
     970             : 
     971       23799 :                 if (best_size == 0 || score < *best_score) {
     972       10253 :                     *best_score = score;
     973       10253 :                     best_size = encode_mode(s, h,
     974             :                                             scratch_data, scratch_linesize,
     975             :                                             last_data, last_linesize, &info,
     976       10253 :                                             s->strip_buf + STRIP_HEADER_SIZE);
     977             : 
     978       10253 :                     write_strip_header(s, y, h, keyframe, s->strip_buf, best_size);
     979             :                 }
     980             :             }
     981             :         }
     982             :     }
     983             : 
     984         753 :     best_size += STRIP_HEADER_SIZE;
     985         753 :     memcpy(buf, s->strip_buf, best_size);
     986             : 
     987         753 :     return best_size;
     988             : }
     989             : 
     990         303 : static int write_cvid_header(CinepakEncContext *s, unsigned char *buf,
     991             :                              int num_strips, int data_size, int isakeyframe)
     992             : {
     993         303 :     buf[0] = isakeyframe ? 0 : 1;
     994         303 :     AV_WB24(&buf[1], data_size + CVID_HEADER_SIZE);
     995         303 :     AV_WB16(&buf[4], s->w);
     996         303 :     AV_WB16(&buf[6], s->h);
     997         303 :     AV_WB16(&buf[8], num_strips);
     998             : 
     999         303 :     return CVID_HEADER_SIZE;
    1000             : }
    1001             : 
    1002         150 : static int rd_frame(CinepakEncContext *s, const AVFrame *frame,
    1003             :                     int isakeyframe, unsigned char *buf, int buf_size)
    1004             : {
    1005             :     int num_strips, strip, i, y, nexty, size, temp_size, best_size;
    1006             :     uint8_t *last_data    [4], *data    [4], *scratch_data    [4];
    1007             :     int      last_linesize[4],  linesize[4],  scratch_linesize[4];
    1008         150 :     int64_t best_score = 0, score, score_temp;
    1009             :     int best_nstrips;
    1010             : 
    1011         150 :     if (s->pix_fmt == AV_PIX_FMT_RGB24) {
    1012             :         int x;
    1013             :         // build a copy of the given frame in the correct colorspace
    1014        7350 :         for (y = 0; y < s->h; y += 2)
    1015      468000 :             for (x = 0; x < s->w; x += 2) {
    1016             :                 uint8_t *ir[2];
    1017             :                 int32_t r, g, b, rr, gg, bb;
    1018      460800 :                 ir[0] = frame->data[0] + x * 3 + y * frame->linesize[0];
    1019      460800 :                 ir[1] = ir[0] + frame->linesize[0];
    1020      460800 :                 get_sub_picture(s, x, y,
    1021      460800 :                                 s->input_frame->data, s->input_frame->linesize,
    1022             :                                 scratch_data, scratch_linesize);
    1023      460800 :                 r = g = b = 0;
    1024     2304000 :                 for (i = 0; i < 4; ++i) {
    1025             :                     int i1, i2;
    1026     1843200 :                     i1 = (i & 1);
    1027     1843200 :                     i2 = (i >= 2);
    1028     1843200 :                     rr = ir[i2][i1 * 3 + 0];
    1029     1843200 :                     gg = ir[i2][i1 * 3 + 1];
    1030     1843200 :                     bb = ir[i2][i1 * 3 + 2];
    1031     1843200 :                     r += rr;
    1032     1843200 :                     g += gg;
    1033     1843200 :                     b += bb;
    1034             :                     // using fixed point arithmetic for portable repeatability, scaling by 2^23
    1035             :                     // "Y"
    1036             :                     // rr = 0.2857 * rr + 0.5714 * gg + 0.1429 * bb;
    1037     1843200 :                     rr = (2396625 * rr + 4793251 * gg + 1198732 * bb) >> 23;
    1038     1843200 :                     if (rr < 0)
    1039           0 :                         rr = 0;
    1040     1843200 :                     else if (rr > 255)
    1041           0 :                         rr = 255;
    1042     1843200 :                     scratch_data[0][i1 + i2 * scratch_linesize[0]] = rr;
    1043             :                 }
    1044             :                 // let us scale down as late as possible
    1045             :                 //                r /= 4; g /= 4; b /= 4;
    1046             :                 // "U"
    1047             :                 // rr = -0.1429 * r - 0.2857 * g + 0.4286 * b;
    1048      460800 :                 rr = (-299683 * r - 599156 * g + 898839 * b) >> 23;
    1049      460800 :                 if (rr < -128)
    1050           0 :                     rr = -128;
    1051      460800 :                 else if (rr > 127)
    1052           0 :                     rr = 127;
    1053      460800 :                 scratch_data[1][0] = rr + 128; // quantize needs unsigned
    1054             :                 // "V"
    1055             :                 // rr = 0.3571 * r - 0.2857 * g - 0.0714 * b;
    1056      460800 :                 rr = (748893 * r - 599156 * g - 149737 * b) >> 23;
    1057      460800 :                 if (rr < -128)
    1058           0 :                     rr = -128;
    1059      460800 :                 else if (rr > 127)
    1060           0 :                     rr = 127;
    1061      460800 :                 scratch_data[2][0] = rr + 128; // quantize needs unsigned
    1062             :             }
    1063             :     }
    1064             : 
    1065             :     // would be nice but quite certainly incompatible with vintage players:
    1066             :     // support encoding zero strips (meaning skip the whole frame)
    1067         453 :     for (num_strips = s->min_strips; num_strips <= s->max_strips && num_strips <= s->h / MB_SIZE; num_strips++) {
    1068         303 :         score = 0;
    1069         303 :         size  = 0;
    1070             : 
    1071        1056 :         for (y = 0, strip = 1; y < s->h; strip++, y = nexty) {
    1072             :             int strip_height;
    1073             : 
    1074         753 :             nexty = strip * s->h / num_strips; // <= s->h
    1075             :             // make nexty the next multiple of 4 if not already there
    1076         753 :             if (nexty & 3)
    1077           0 :                 nexty += 4 - (nexty & 3);
    1078             : 
    1079         753 :             strip_height = nexty - y;
    1080         753 :             if (strip_height <= 0) { // can this ever happen?
    1081           0 :                 av_log(s->avctx, AV_LOG_INFO, "skipping zero height strip %i of %i\n", strip, num_strips);
    1082           0 :                 continue;
    1083             :             }
    1084             : 
    1085         753 :             if (s->pix_fmt == AV_PIX_FMT_RGB24)
    1086         753 :                 get_sub_picture(s, 0, y,
    1087         753 :                                 s->input_frame->data, s->input_frame->linesize,
    1088             :                                 data, linesize);
    1089             :             else
    1090           0 :                 get_sub_picture(s, 0, y,
    1091           0 :                                 (uint8_t **)frame->data, (int *)frame->linesize,
    1092             :                                 data, linesize);
    1093         753 :             get_sub_picture(s, 0, y,
    1094         753 :                             s->last_frame->data, s->last_frame->linesize,
    1095             :                             last_data, last_linesize);
    1096         753 :             get_sub_picture(s, 0, y,
    1097         753 :                             s->scratch_frame->data, s->scratch_frame->linesize,
    1098             :                             scratch_data, scratch_linesize);
    1099             : 
    1100         753 :             if ((temp_size = rd_strip(s, y, strip_height, isakeyframe,
    1101             :                                       last_data, last_linesize, data, linesize,
    1102             :                                       scratch_data, scratch_linesize,
    1103         753 :                                       s->frame_buf + size + CVID_HEADER_SIZE,
    1104             :                                       &score_temp)) < 0)
    1105           0 :                 return temp_size;
    1106             : 
    1107         753 :             score += score_temp;
    1108         753 :             size += temp_size;
    1109             :         }
    1110             : 
    1111         303 :         if (best_score == 0 || score < best_score) {
    1112         303 :             best_score = score;
    1113         303 :             best_size = size + write_cvid_header(s, s->frame_buf, num_strips, size, isakeyframe);
    1114             : 
    1115         303 :             FFSWAP(AVFrame *, s->best_frame, s->scratch_frame);
    1116         303 :             memcpy(buf, s->frame_buf, best_size);
    1117         303 :             best_nstrips = num_strips;
    1118             :         }
    1119             :         // avoid trying too many strip numbers without a real reason
    1120             :         // (this makes the processing of the very first frame faster)
    1121         303 :         if (num_strips - best_nstrips > 4)
    1122           0 :             break;
    1123             :     }
    1124             : 
    1125             :     // let the number of strips slowly adapt to the changes in the contents,
    1126             :     // compared to full bruteforcing every time this will occasionally lead
    1127             :     // to some r/d performance loss but makes encoding up to several times faster
    1128         150 :     if (!s->strip_number_delta_range) {
    1129           0 :         if (best_nstrips == s->max_strips) { // let us try to step up
    1130           0 :             s->max_strips = best_nstrips + 1;
    1131           0 :             if (s->max_strips >= s->max_max_strips)
    1132           0 :                 s->max_strips = s->max_max_strips;
    1133             :         } else { // try to step down
    1134           0 :             s->max_strips = best_nstrips;
    1135             :         }
    1136           0 :         s->min_strips = s->max_strips - 1;
    1137           0 :         if (s->min_strips < s->min_min_strips)
    1138           0 :             s->min_strips = s->min_min_strips;
    1139             :     } else {
    1140         150 :         s->max_strips = best_nstrips + s->strip_number_delta_range;
    1141         150 :         if (s->max_strips >= s->max_max_strips)
    1142         150 :             s->max_strips = s->max_max_strips;
    1143         150 :         s->min_strips = best_nstrips - s->strip_number_delta_range;
    1144         150 :         if (s->min_strips < s->min_min_strips)
    1145           0 :             s->min_strips = s->min_min_strips;
    1146             :     }
    1147             : 
    1148         150 :     return best_size;
    1149             : }
    1150             : 
    1151         150 : static int cinepak_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
    1152             :                                 const AVFrame *frame, int *got_packet)
    1153             : {
    1154         150 :     CinepakEncContext *s = avctx->priv_data;
    1155             :     int ret;
    1156             : 
    1157         150 :     s->lambda = frame->quality ? frame->quality - 1 : 2 * FF_LAMBDA_SCALE;
    1158             : 
    1159         150 :     if ((ret = ff_alloc_packet2(avctx, pkt, s->frame_buf_size, 0)) < 0)
    1160           0 :         return ret;
    1161         150 :     ret       = rd_frame(s, frame, (s->curframe == 0), pkt->data, s->frame_buf_size);
    1162         150 :     pkt->size = ret;
    1163         150 :     if (s->curframe == 0)
    1164           6 :         pkt->flags |= AV_PKT_FLAG_KEY;
    1165         150 :     *got_packet = 1;
    1166             : 
    1167         150 :     FFSWAP(AVFrame *, s->last_frame, s->best_frame);
    1168             : 
    1169         150 :     if (++s->curframe >= s->keyint)
    1170           6 :         s->curframe = 0;
    1171             : 
    1172         150 :     return 0;
    1173             : }
    1174             : 
    1175           3 : static av_cold int cinepak_encode_end(AVCodecContext *avctx)
    1176             : {
    1177           3 :     CinepakEncContext *s = avctx->priv_data;
    1178             :     int x;
    1179             : 
    1180           3 :     av_frame_free(&s->last_frame);
    1181           3 :     av_frame_free(&s->best_frame);
    1182           3 :     av_frame_free(&s->scratch_frame);
    1183           3 :     if (avctx->pix_fmt == AV_PIX_FMT_RGB24)
    1184           3 :         av_frame_free(&s->input_frame);
    1185           3 :     av_freep(&s->codebook_input);
    1186           3 :     av_freep(&s->codebook_closest);
    1187           3 :     av_freep(&s->strip_buf);
    1188           3 :     av_freep(&s->frame_buf);
    1189           3 :     av_freep(&s->mb);
    1190             : 
    1191          15 :     for (x = 0; x < (avctx->pix_fmt == AV_PIX_FMT_RGB24 ? 4 : 3); x++)
    1192          12 :         av_freep(&s->pict_bufs[x]);
    1193             : 
    1194           3 :     return 0;
    1195             : }
    1196             : 
    1197             : AVCodec ff_cinepak_encoder = {
    1198             :     .name           = "cinepak",
    1199             :     .long_name      = NULL_IF_CONFIG_SMALL("Cinepak"),
    1200             :     .type           = AVMEDIA_TYPE_VIDEO,
    1201             :     .id             = AV_CODEC_ID_CINEPAK,
    1202             :     .priv_data_size = sizeof(CinepakEncContext),
    1203             :     .init           = cinepak_encode_init,
    1204             :     .encode2        = cinepak_encode_frame,
    1205             :     .close          = cinepak_encode_end,
    1206             :     .pix_fmts       = (const enum AVPixelFormat[]) { AV_PIX_FMT_RGB24, AV_PIX_FMT_GRAY8, AV_PIX_FMT_NONE },
    1207             :     .priv_class     = &cinepak_class,
    1208             : };

Generated by: LCOV version 1.13