@@ -1,6 +1,9 @@

 /*

  * Cinepak encoder (c) 2011 Tomas Härdin

  * http://titan.codemill.se/~tomhar/cinepakenc.patch

+ *

+ * Fixes and improvements, vintage decoders compatibility

+ *  (c) 2013, 2014 Rl, Aetey Global Technologies AB

 Permission is hereby granted, free of charge, to any person obtaining a

 copy of this software and associated documentation files (the "Software"),

@@ -20,13 +23,27 @@ OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,

 ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR

 OTHER DEALINGS IN THE SOFTWARE.

+ * TODO:

+ * - optimize: color space conversion (move conversion to libswscale), ...

+ * - implement options to set the min/max number of strips?

+ * MAYBE:

+ * - "optimally" split the frame into several non-regular areas

+ *   using a separate codebook pair for each area and approximating

+ *   the area by several rectangular strips (generally not full width ones)

+ *   (use quadtree splitting? a simple fixed-granularity grid?)

  */

+#include <string.h>

+

 #include "libavutil/common.h"

+#include "libavutil/internal.h"

 #include "libavutil/intreadwrite.h"

 #include "avcodec.h"

 #include "libavutil/lfg.h"

 #include "elbg.h"

+#include "internal.h"

+

+#include "libavutil/avassert.h"

 #define CVID_HEADER_SIZE 10

 #define STRIP_HEADER_SIZE 12

@@ -36,11 +53,20 @@ OTHER DEALINGS IN THE SOFTWARE.

 #define MB_AREA (MB_SIZE*MB_SIZE)

 #define VECTOR_MAX 6        //six or four entries per vector depending on format

-#define CODEBOOK_MAX 256

-#define CODEBOOK_NUM 5      //five potential codebooks (1, 4, 16, 64, 256) for V1 and V4

+#define CODEBOOK_MAX 256    //size of a codebook

-#define MAX_STRIPS  1       //Note: having fewer choices regarding the number of strip speeds up encoding (obviously)

+//#define MAX_STRIPS  32      //Note: having fewer choices regarding the number of strips speeds up encoding (obviously)

+#define MAX_STRIPS  3       // This seems to be max for vintage players! -- rl

+// TODO: we might want to have a "vintage compatibilty" switch

 #define MIN_STRIPS  1       //Note: having more strips speeds up encoding the frame (this is less obvious)

+// MAX_STRIPS limits the maximum quality you can reach

+//            when you want high quality on high resolutions,

+// MIN_STRIPS limits the minimum efficiently encodable bit rate

+//            on low resolutions

+// the numbers are only used for brute force optimization for the first frame,

+// for the following frames they are adaptively readjusted

+// NOTE the decoder in ffmpeg has its own arbitrary limitation on the number

+// of strips, currently 32

 typedef enum {

     MODE_V1_ONLY = 0,

@@ -53,42 +79,51 @@ typedef enum {

 typedef enum {

     ENC_V1,

     ENC_V4,

-    ENC_SKIP

+    ENC_SKIP,

+

+    ENC_UNCERTAIN

 } mb_encoding;

 typedef struct {

     int v1_vector;                  //index into v1 codebook

     int v1_error;                   //error when using V1 encoding

-    int v4_vector[CODEBOOK_NUM][4]; //indices into v4 codebooks

-    int v4_error[CODEBOOK_NUM];     //error when using V4 encodings

+    int v4_vector[4];               //indices into v4 codebook

+    int v4_error;                   //error when using V4 encoding

     int skip_error;                 //error when block is skipped (aka copied from last frame)

     mb_encoding best_encoding;      //last result from calculate_mode_score()

 } mb_info;

 typedef struct {

     int v1_codebook[CODEBOOK_MAX*VECTOR_MAX];

-    int *v4_codebook;

+    int v4_codebook[CODEBOOK_MAX*VECTOR_MAX];

+    int v1_size;

+    int v4_size;

+    CinepakMode mode;

 } strip_info;

 typedef struct {

     AVCodecContext *avctx;

-    unsigned char *pict_bufs[3], *strip_buf, *frame_buf;

-    AVFrame last_frame;

-    AVFrame best_frame;

-    AVFrame scratch_frame;

+    unsigned char *pict_bufs[4], *strip_buf, *frame_buf;

+    AVFrame *last_frame;

+    AVFrame *best_frame;

+    AVFrame *scratch_frame;

+    AVFrame *input_frame;

     enum AVPixelFormat pix_fmt;

     int w, h;

+    int frame_buf_size;

     int curframe, keyint;

     AVLFG randctx;

     uint64_t lambda;

     int *codebook_input;

     int *codebook_closest;

     mb_info *mb;                                //MB RD state

+    int min_strips;          //the current limit

+    int max_strips;          //the current limit

 #ifdef CINEPAKENC_DEBUG

     mb_info *best_mb;                           //TODO: remove. only used for printing stats

-#endif

     int num_v1_mode, num_v4_mode, num_mc_mode;

     int num_v1_encs, num_v4_encs, num_skips;

+#endif

 } CinepakEncContext;

 static av_cold int cinepak_encode_init(AVCodecContext *avctx)

@@ -102,20 +137,33 @@ static av_cold int cinepak_encode_init(AVCodecContext *avctx)

         return AVERROR(EINVAL);

     }

-    if (!(s->codebook_input = av_malloc(sizeof(int) * (avctx->pix_fmt == AV_PIX_FMT_YUV420P ? 6 : 4) * (avctx->width * avctx->height) >> 2)))

+    if (!(s->last_frame = av_frame_alloc()))

         return AVERROR(ENOMEM);

+    if (!(s->best_frame = av_frame_alloc()))

+        goto enomem;

+    if (!(s->scratch_frame = av_frame_alloc()))

+        goto enomem;

+    if (avctx->pix_fmt == AV_PIX_FMT_RGB24)

+        if (!(s->input_frame = av_frame_alloc()))

+            goto enomem;

+

+    if (!(s->codebook_input = av_malloc(sizeof(int) * (avctx->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4) * (avctx->width * avctx->height) >> 2)))

+        goto enomem;

     if (!(s->codebook_closest = av_malloc(sizeof(int) * (avctx->width * avctx->height) >> 2)))

         goto enomem;

-    for(x = 0; x < 3; x++)

-        if(!(s->pict_bufs[x] = av_malloc((avctx->pix_fmt == AV_PIX_FMT_YUV420P ? 6 : 4) * (avctx->width * avctx->height) >> 2)))

+    for(x = 0; x < (avctx->pix_fmt == AV_PIX_FMT_RGB24 ? 4 : 3); x++)

+        if(!(s->pict_bufs[x] = av_malloc((avctx->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4) * (avctx->width * avctx->height) >> 2)))

             goto enomem;

     mb_count = avctx->width * avctx->height / MB_AREA;

-    //the largest possible chunk is 0x31 with all MBs encoded in V4 mode, which is 34 bits per MB

-    strip_buf_size = STRIP_HEADER_SIZE + 3 * CHUNK_HEADER_SIZE + 2 * VECTOR_MAX * CODEBOOK_MAX + 4 * (mb_count + (mb_count + 15) / 16);

+    //the largest possible chunk is 0x31 with all MBs encoded in V4 mode

+    //and full codebooks being replaced in INTER mode,

+    // which is 34 bits per MB

+    //and 2*256 extra flag bits per strip

+    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;

     frame_buf_size = CVID_HEADER_SIZE + MAX_STRIPS * strip_buf_size;

@@ -137,113 +185,221 @@ static av_cold int cinepak_encode_init(AVCodecContext *avctx)

     s->avctx = avctx;

     s->w = avctx->width;

     s->h = avctx->height;

+    s->frame_buf_size = frame_buf_size;

     s->curframe = 0;

     s->keyint = avctx->keyint_min;

     s->pix_fmt = avctx->pix_fmt;

     //set up AVFrames

-    s->last_frame.data[0]        = s->pict_bufs[0];

-    s->last_frame.linesize[0]    = s->w;

-    s->best_frame.data[0]        = s->pict_bufs[1];

-    s->best_frame.linesize[0]    = s->w;

-    s->scratch_frame.data[0]     = s->pict_bufs[2];

-    s->scratch_frame.linesize[0] = s->w;

-

-    if(s->pix_fmt == AV_PIX_FMT_YUV420P) {

-        s->last_frame.data[1]        = s->last_frame.data[0] + s->w * s->h;

-        s->last_frame.data[2]        = s->last_frame.data[1] + ((s->w * s->h) >> 2);

-        s->last_frame.linesize[1]    = s->last_frame.linesize[2] = s->w >> 1;

-

-        s->best_frame.data[1]        = s->best_frame.data[0] + s->w * s->h;

-        s->best_frame.data[2]        = s->best_frame.data[1] + ((s->w * s->h) >> 2);

-        s->best_frame.linesize[1]    = s->best_frame.linesize[2] = s->w >> 1;

-

-        s->scratch_frame.data[1]     = s->scratch_frame.data[0] + s->w * s->h;

-        s->scratch_frame.data[2]     = s->scratch_frame.data[1] + ((s->w * s->h) >> 2);

-        s->scratch_frame.linesize[1] = s->scratch_frame.linesize[2] = s->w >> 1;

+    s->last_frame->data[0]        = s->pict_bufs[0];

+    s->last_frame->linesize[0]    = s->w;

+    s->best_frame->data[0]        = s->pict_bufs[1];

+    s->best_frame->linesize[0]    = s->w;

+    s->scratch_frame->data[0]     = s->pict_bufs[2];

+    s->scratch_frame->linesize[0] = s->w;

+

+    if (s->pix_fmt == AV_PIX_FMT_RGB24) {

+        s->last_frame->data[1]        = s->last_frame->data[0] + s->w * s->h;

+        s->last_frame->data[2]        = s->last_frame->data[1] + ((s->w * s->h) >> 2);

+        s->last_frame->linesize[1]    = s->last_frame->linesize[2] = s->w >> 1;

+

+        s->best_frame->data[1]        = s->best_frame->data[0] + s->w * s->h;

+        s->best_frame->data[2]        = s->best_frame->data[1] + ((s->w * s->h) >> 2);

+        s->best_frame->linesize[1]    = s->best_frame->linesize[2] = s->w >> 1;

+

+        s->scratch_frame->data[1]     = s->scratch_frame->data[0] + s->w * s->h;

+        s->scratch_frame->data[2]     = s->scratch_frame->data[1] + ((s->w * s->h) >> 2);

+        s->scratch_frame->linesize[1] = s->scratch_frame->linesize[2] = s->w >> 1;

+

+        s->input_frame->data[0]       = s->pict_bufs[3];

+        s->input_frame->linesize[0]   = s->w;

+        s->input_frame->data[1]       = s->input_frame->data[0] + s->w * s->h;

+        s->input_frame->data[2]       = s->input_frame->data[1] + ((s->w * s->h) >> 2);

+        s->input_frame->linesize[1]   = s->input_frame->linesize[2] = s->w >> 1;

     }

+    s->min_strips = MIN_STRIPS;

+    s->max_strips = MAX_STRIPS;

+

+#ifdef CINEPAKENC_DEBUG

     s->num_v1_mode = s->num_v4_mode = s->num_mc_mode = s->num_v1_encs = s->num_v4_encs = s->num_skips = 0;

+#endif

     return 0;

 enomem:

-    av_free(s->codebook_input);

-    av_free(s->codebook_closest);

-    av_free(s->strip_buf);

-    av_free(s->frame_buf);

-    av_free(s->mb);

+    av_frame_free(&s->last_frame);

+    av_frame_free(&s->best_frame);

+    av_frame_free(&s->scratch_frame);

+    if (avctx->pix_fmt == AV_PIX_FMT_RGB24)

+        av_frame_free(&s->input_frame);

+    av_freep(&s->codebook_input);

+    av_freep(&s->codebook_closest);

+    av_freep(&s->strip_buf);

+    av_freep(&s->frame_buf);

+    av_freep(&s->mb);

 #ifdef CINEPAKENC_DEBUG

-    av_free(s->best_mb);

+    av_freep(&s->best_mb);

 #endif

-    for(x = 0; x < 3; x++)

-        av_free(s->pict_bufs[x]);

+    for(x = 0; x < (avctx->pix_fmt == AV_PIX_FMT_RGB24 ? 4 : 3); x++)

+        av_freep(&s->pict_bufs[x]);

     return AVERROR(ENOMEM);

 }

-static int64_t calculate_mode_score(CinepakEncContext *s, CinepakMode mode, int h, int v1_size, int v4_size, int v4, strip_info *info)

+static int64_t calculate_mode_score(CinepakEncContext *s, int h, strip_info *info, int report, int *training_set_v1_shrunk, int *training_set_v4_shrunk

+#ifdef CINEPAK_REPORT_SERR

+, int64_t *serr

+#endif

+)

 {

     //score = FF_LAMBDA_SCALE * error + lambda * bits

     int x;

-    int entry_size = s->pix_fmt == AV_PIX_FMT_YUV420P ? 6 : 4;

+    int entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4;

     int mb_count = s->w * h / MB_AREA;

     mb_info *mb;

     int64_t score1, score2, score3;

-    int64_t ret = s->lambda * ((v1_size ? CHUNK_HEADER_SIZE + v1_size * entry_size : 0) +

-                   (v4_size ? CHUNK_HEADER_SIZE + v4_size * entry_size : 0) +

+    int64_t ret = s->lambda * ((info->v1_size ? CHUNK_HEADER_SIZE + info->v1_size * entry_size : 0) +

+                   (info->v4_size ? CHUNK_HEADER_SIZE + info->v4_size * entry_size : 0) +

                    CHUNK_HEADER_SIZE) << 3;

-    //av_log(s->avctx, AV_LOG_INFO, "sizes %3i %3i -> %9li score mb_count %i", v1_size, v4_size, ret, mb_count);

+    //av_log(s->avctx, AV_LOG_INFO, "sizes %3i %3i -> %9lli score mb_count %i", info->v1_size, info->v4_size, (long long int)ret, mb_count);

+

+#ifdef CINEPAK_REPORT_SERR

+    *serr = 0;

+#endif

-    switch(mode) {

+    switch(info->mode) {

     case MODE_V1_ONLY:

         //one byte per MB

         ret += s->lambda * 8 * mb_count;

+// while calculating we assume all blocks are ENC_V1

         for(x = 0; x < mb_count; x++) {

             mb = &s->mb[x];

             ret += FF_LAMBDA_SCALE * mb->v1_error;

+#ifdef CINEPAK_REPORT_SERR

+            *serr += mb->v1_error;

+#endif

+// this function is never called for report in MODE_V1_ONLY

+//            if(!report)

             mb->best_encoding = ENC_V1;

         }

         break;

     case MODE_V1_V4:

         //9 or 33 bits per MB

-        for(x = 0; x < mb_count; x++) {

-            mb = &s->mb[x];

-            score1 = s->lambda * 9  + FF_LAMBDA_SCALE * mb->v1_error;

-            score2 = s->lambda * 33 + FF_LAMBDA_SCALE * mb->v4_error[v4];

-

-            if(score1 <= score2) {

+        if(report) {

+// no moves between the corresponding training sets are allowed

+            *training_set_v1_shrunk = *training_set_v4_shrunk = 0;

+            for(x = 0; x < mb_count; x++) {

+                int mberr;

+                mb = &s->mb[x];

+                if(mb->best_encoding == ENC_V1)

+                    score1 = s->lambda * 9  + FF_LAMBDA_SCALE * (mberr=mb->v1_error);

+                else

+                    score1 = s->lambda * 33 + FF_LAMBDA_SCALE * (mberr=mb->v4_error);

                 ret += score1;

-                mb->best_encoding = ENC_V1;

-            } else {

-                ret += score2;

-                mb->best_encoding = ENC_V4;

+#ifdef CINEPAK_REPORT_SERR

+                *serr += mberr;

+#endif

+            }

+        } else { // find best mode per block

+            for(x = 0; x < mb_count; x++) {

+                mb = &s->mb[x];

+                score1 = s->lambda * 9  + FF_LAMBDA_SCALE * mb->v1_error;

+                score2 = s->lambda * 33 + FF_LAMBDA_SCALE * mb->v4_error;

+

+                if(score1 <= score2) {

+                    ret += score1;

+#ifdef CINEPAK_REPORT_SERR

+                    *serr += mb->v1_error;

+#endif

+                    mb->best_encoding = ENC_V1;

+                } else {

+                    ret += score2;

+#ifdef CINEPAK_REPORT_SERR

+                    *serr += mb->v4_error;

+#endif

+                    mb->best_encoding = ENC_V4;

+                }

             }

         }

         break;

     case MODE_MC:

         //1, 10 or 34 bits per MB

-        for(x = 0; x < mb_count; x++) {

-            mb = &s->mb[x];

-            score1 = s->lambda * 1  + FF_LAMBDA_SCALE * mb->skip_error;

-            score2 = s->lambda * 10 + FF_LAMBDA_SCALE * mb->v1_error;

-            score3 = s->lambda * 34 + FF_LAMBDA_SCALE * mb->v4_error[v4];

-

-

-            if(score1 <= score2 && score1 <= score3) {

-                ret += score1;

-                mb->best_encoding = ENC_SKIP;

-            } else if(score2 <= score1 && score2 <= score3) {

-                ret += score2;

-                mb->best_encoding = ENC_V1;

-            } else {

-                ret += score3;

-                mb->best_encoding = ENC_V4;

+        if(report) {

+            int v1_shrunk = 0, v4_shrunk = 0;

+            for(x = 0; x < mb_count; x++) {

+                mb = &s->mb[x];

+// it is OK to move blocks to ENC_SKIP here

+// but not to any codebook encoding!

+                score1 = s->lambda * 1  + FF_LAMBDA_SCALE * mb->skip_error;

+                if(mb->best_encoding == ENC_SKIP) {

+                    ret += score1;

+#ifdef CINEPAK_REPORT_SERR

+                    *serr += mb->skip_error;

+#endif

+                } else if(mb->best_encoding == ENC_V1) {

+                    if((score2=s->lambda * 10 + FF_LAMBDA_SCALE * mb->v1_error) >= score1) {

+                        mb->best_encoding = ENC_SKIP;

+                        ++v1_shrunk;

+                        ret += score1;

+#ifdef CINEPAK_REPORT_SERR

+                        *serr += mb->skip_error;

+#endif

+                    } else {

+                        ret += score2;

+#ifdef CINEPAK_REPORT_SERR

+                        *serr += mb->v1_error;

+#endif

+                    }

+                } else {

+                    if((score3=s->lambda * 34 + FF_LAMBDA_SCALE * mb->v4_error) >= score1) {

+                        mb->best_encoding = ENC_SKIP;

+                        ++v4_shrunk;

+                        ret += score1;

+#ifdef CINEPAK_REPORT_SERR

+                        *serr += mb->skip_error;

+#endif

+                    } else {

+                        ret += score3;

+#ifdef CINEPAK_REPORT_SERR

+                        *serr += mb->v4_error;

+#endif

+                    }

+                }

+            }

+            *training_set_v1_shrunk = v1_shrunk;

+            *training_set_v4_shrunk = v4_shrunk;

+        } else { // find best mode per block

+            for(x = 0; x < mb_count; x++) {

+                mb = &s->mb[x];

+                score1 = s->lambda * 1  + FF_LAMBDA_SCALE * mb->skip_error;

+                score2 = s->lambda * 10 + FF_LAMBDA_SCALE * mb->v1_error;

+                score3 = s->lambda * 34 + FF_LAMBDA_SCALE * mb->v4_error;

+

+                if(score1 <= score2 && score1 <= score3) {

+                    ret += score1;

+#ifdef CINEPAK_REPORT_SERR

+                    *serr += mb->skip_error;

+#endif

+                    mb->best_encoding = ENC_SKIP;

+                } else if(score2 <= score3) {

+                    ret += score2;

+#ifdef CINEPAK_REPORT_SERR

+                    *serr += mb->v1_error;

+#endif

+                    mb->best_encoding = ENC_V1;

+                } else {

+                    ret += score3;

+#ifdef CINEPAK_REPORT_SERR

+                    *serr += mb->v4_error;

+#endif

+                    mb->best_encoding = ENC_V4;

+                }

             }

         }

@@ -262,13 +418,45 @@ static int write_chunk_header(unsigned char *buf, int chunk_type, int chunk_size

 static int encode_codebook(CinepakEncContext *s, int *codebook, int size, int chunk_type_yuv, int chunk_type_gray, unsigned char *buf)

 {

-    int x, y, ret, entry_size = s->pix_fmt == AV_PIX_FMT_YUV420P ? 6 : 4;

-

-    ret = write_chunk_header(buf, s->pix_fmt == AV_PIX_FMT_YUV420P ? chunk_type_yuv : chunk_type_gray, entry_size * size);

-

-    for(x = 0; x < size; x++)

-        for(y = 0; y < entry_size; y++)

-            buf[ret++] = codebook[y + x*entry_size] ^ (y >= 4 ? 0x80 : 0);

+    int x, y, ret, entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4;

+    int incremental_codebook_replacement_mode = 0; // hardcoded here,

+                // the compiler should notice that this is a constant -- rl

+

+    ret = write_chunk_header(buf,

+          s->pix_fmt == AV_PIX_FMT_RGB24 ?

+           chunk_type_yuv+(incremental_codebook_replacement_mode?1:0) :

+           chunk_type_gray+(incremental_codebook_replacement_mode?1:0),

+          entry_size * size

+           + (incremental_codebook_replacement_mode?(size+31)/32*4:0) );

+

+// we do codebook encoding according to the "intra" mode

+// but we keep the "dead" code for reference in case we will want

+// to use incremental codebook updates (which actually would give us

+// "kind of" motion compensation, especially in 1 strip/frame case) -- rl

+// (of course, the code will be not useful as-is)

+    if(incremental_codebook_replacement_mode) {

+        int flags = 0;

+        int flagsind;

+        for(x = 0; x < size; x++) {

+            if(flags == 0) {

+                flagsind = ret;

+                ret += 4;

+                flags = 0x80000000;

+            } else

+                flags = ((flags>>1) | 0x80000000);

+            for(y = 0; y < entry_size; y++)

+                buf[ret++] = codebook[y + x*entry_size] ^ (y >= 4 ? 0x80 : 0);

+            if((flags&0xffffffff) == 0xffffffff) {

+                AV_WB32(&buf[flagsind], flags);

+                flags = 0;

+            }

+        }

+        if(flags)

+            AV_WB32(&buf[flagsind], flags);

+    } else

+        for(x = 0; x < size; x++)

+            for(y = 0; y < entry_size; y++)

+                buf[ret++] = codebook[y + x*entry_size] ^ (y >= 4 ? 0x80 : 0);

     return ret;

 }

@@ -279,7 +467,7 @@ static void get_sub_picture(CinepakEncContext *s, int x, int y, AVPicture *in, A

     out->data[0] = in->data[0] + x + y * in->linesize[0];

     out->linesize[0] = in->linesize[0];

-    if(s->pix_fmt == AV_PIX_FMT_YUV420P) {

+    if(s->pix_fmt == AV_PIX_FMT_RGB24) {

         out->data[1] = in->data[1] + (x >> 1) + (y >> 1) * in->linesize[1];

         out->linesize[1] = in->linesize[1];

@@ -289,47 +477,47 @@ static void get_sub_picture(CinepakEncContext *s, int x, int y, AVPicture *in, A

 }

 //decodes the V1 vector in mb into the 4x4 MB pointed to by sub_pict

-static void decode_v1_vector(CinepakEncContext *s, AVPicture *sub_pict, mb_info *mb, strip_info *info)

+static void decode_v1_vector(CinepakEncContext *s, AVPicture *sub_pict, int v1_vector, strip_info *info)

 {

-    int entry_size = s->pix_fmt == AV_PIX_FMT_YUV420P ? 6 : 4;

+    int entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4;

     sub_pict->data[0][0] =

             sub_pict->data[0][1] =

             sub_pict->data[0][    sub_pict->linesize[0]] =

-            sub_pict->data[0][1+  sub_pict->linesize[0]] = info->v1_codebook[mb->v1_vector*entry_size];

+            sub_pict->data[0][1+  sub_pict->linesize[0]] = info->v1_codebook[v1_vector*entry_size];

     sub_pict->data[0][2] =

             sub_pict->data[0][3] =

             sub_pict->data[0][2+  sub_pict->linesize[0]] =

-            sub_pict->data[0][3+  sub_pict->linesize[0]] = info->v1_codebook[mb->v1_vector*entry_size+1];

+            sub_pict->data[0][3+  sub_pict->linesize[0]] = info->v1_codebook[v1_vector*entry_size+1];

     sub_pict->data[0][2*sub_pict->linesize[0]] =

             sub_pict->data[0][1+2*sub_pict->linesize[0]] =

             sub_pict->data[0][  3*sub_pict->linesize[0]] =

-            sub_pict->data[0][1+3*sub_pict->linesize[0]] = info->v1_codebook[mb->v1_vector*entry_size+2];

+            sub_pict->data[0][1+3*sub_pict->linesize[0]] = info->v1_codebook[v1_vector*entry_size+2];

     sub_pict->data[0][2+2*sub_pict->linesize[0]] =

             sub_pict->data[0][3+2*sub_pict->linesize[0]] =

             sub_pict->data[0][2+3*sub_pict->linesize[0]] =

-            sub_pict->data[0][3+3*sub_pict->linesize[0]] = info->v1_codebook[mb->v1_vector*entry_size+3];

+            sub_pict->data[0][3+3*sub_pict->linesize[0]] = info->v1_codebook[v1_vector*entry_size+3];

-    if(s->pix_fmt == AV_PIX_FMT_YUV420P) {

+    if(s->pix_fmt == AV_PIX_FMT_RGB24) {

         sub_pict->data[1][0] =

             sub_pict->data[1][1] =

             sub_pict->data[1][    sub_pict->linesize[1]] =

-            sub_pict->data[1][1+  sub_pict->linesize[1]] = info->v1_codebook[mb->v1_vector*entry_size+4];

+            sub_pict->data[1][1+  sub_pict->linesize[1]] = info->v1_codebook[v1_vector*entry_size+4];

         sub_pict->data[2][0] =

             sub_pict->data[2][1] =

             sub_pict->data[2][    sub_pict->linesize[2]] =

-            sub_pict->data[2][1+  sub_pict->linesize[2]] = info->v1_codebook[mb->v1_vector*entry_size+5];

+            sub_pict->data[2][1+  sub_pict->linesize[2]] = info->v1_codebook[v1_vector*entry_size+5];

     }

 }

 //decodes the V4 vectors in mb into the 4x4 MB pointed to by sub_pict

 static void decode_v4_vector(CinepakEncContext *s, AVPicture *sub_pict, int *v4_vector, strip_info *info)

 {

-    int i, x, y, entry_size = s->pix_fmt == AV_PIX_FMT_YUV420P ? 6 : 4;

+    int i, x, y, entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4;

     for(i = y = 0; y < 4; y += 2) {

         for(x = 0; x < 4; x += 2, i++) {

@@ -338,7 +526,7 @@ static void decode_v4_vector(CinepakEncContext *s, AVPicture *sub_pict, int *v4_

             sub_pict->data[0][x   + (y+1)*sub_pict->linesize[0]] = info->v4_codebook[v4_vector[i]*entry_size+2];

             sub_pict->data[0][x+1 + (y+1)*sub_pict->linesize[0]] = info->v4_codebook[v4_vector[i]*entry_size+3];

-            if(s->pix_fmt == AV_PIX_FMT_YUV420P) {

+            if(s->pix_fmt == AV_PIX_FMT_RGB24) {

                 sub_pict->data[1][(x>>1) + (y>>1)*sub_pict->linesize[1]] = info->v4_codebook[v4_vector[i]*entry_size+4];

                 sub_pict->data[2][(x>>1) + (y>>1)*sub_pict->linesize[2]] = info->v4_codebook[v4_vector[i]*entry_size+5];

             }

@@ -346,39 +534,62 @@ static void decode_v4_vector(CinepakEncContext *s, AVPicture *sub_pict, int *v4_

     }

 }

-static int encode_mode(CinepakEncContext *s, CinepakMode mode, int h, int v1_size, int v4_size, int v4, AVPicture *scratch_pict, strip_info *info, unsigned char *buf)

+static void copy_mb(CinepakEncContext *s, AVPicture *a, AVPicture *b)

+{

+    int y, p;

+

+    for(y = 0; y < MB_SIZE; y++) {

+        memcpy(a->data[0]+y*a->linesize[0], b->data[0]+y*b->linesize[0],

+               MB_SIZE);

+    }

+

+    if(s->pix_fmt == AV_PIX_FMT_RGB24) {

+        for(p = 1; p <= 2; p++) {

+            for(y = 0; y < MB_SIZE/2; y++) {

+                memcpy(a->data[p] + y*a->linesize[p],

+                       b->data[p] + y*b->linesize[p],

+                       MB_SIZE/2);

+            }

+        }

+    }

+}

+

+static int encode_mode(CinepakEncContext *s, int h, AVPicture *scratch_pict, AVPicture *last_pict, strip_info *info, unsigned char *buf)

 {

     int x, y, z, flags, bits, temp_size, header_ofs, ret = 0, mb_count = s->w * h / MB_AREA;

     int needs_extra_bit, should_write_temp;

     unsigned char temp[64]; //32/2 = 16 V4 blocks at 4 B each -> 64 B

     mb_info *mb;

-    AVPicture sub_scratch;

+    AVPicture sub_scratch, sub_last;

     //encode codebooks

-    if(v1_size)

-        ret += encode_codebook(s, info->v1_codebook, v1_size, 0x22, 0x26, buf + ret);

+////// MacOS vintage decoder compatibility dictates the presence of

+////// the codebook chunk even when the codebook is empty - pretty dumb...

+////// and also the certain order of the codebook chunks -- rl

+//    if(info->v4_size)

+        ret += encode_codebook(s, info->v4_codebook, info->v4_size, 0x20, 0x24, buf + ret);

-    if(v4_size)

-        ret += encode_codebook(s, info->v4_codebook, v4_size, 0x20, 0x24, buf + ret);

+//    if(info->v1_size)

+        ret += encode_codebook(s, info->v1_codebook, info->v1_size, 0x22, 0x26, buf + ret);

     //update scratch picture

     for(z = y = 0; y < h; y += MB_SIZE) {

         for(x = 0; x < s->w; x += MB_SIZE, z++) {

             mb = &s->mb[z];

-            if(mode == MODE_MC && mb->best_encoding == ENC_SKIP)

-                continue;

-

             get_sub_picture(s, x, y, scratch_pict, &sub_scratch);

-            if(mode == MODE_V1_ONLY || mb->best_encoding == ENC_V1)

-                decode_v1_vector(s, &sub_scratch, mb, info);

-            else if(mode != MODE_V1_ONLY && mb->best_encoding == ENC_V4)

-                decode_v4_vector(s, &sub_scratch, mb->v4_vector[v4], info);

+            if(info->mode == MODE_MC && mb->best_encoding == ENC_SKIP) {

+                get_sub_picture(s, x, y, last_pict, &sub_last);

+                copy_mb(s, &sub_scratch, &sub_last);

+            } else if(info->mode == MODE_V1_ONLY || mb->best_encoding == ENC_V1)

+                decode_v1_vector(s, &sub_scratch, mb->v1_vector, info);

+            else

+                decode_v4_vector(s, &sub_scratch, mb->v4_vector, info);

         }

     }

-    switch(mode) {

+    switch(info->mode) {

     case MODE_V1_ONLY:

         //av_log(s->avctx, AV_LOG_INFO, "mb_count = %i\n", mb_count);

         ret += write_chunk_header(buf + ret, 0x32, mb_count);

@@ -408,7 +619,7 @@ static int encode_mode(CinepakEncContext *s, CinepakMode mode, int h, int v1_siz

                     buf[ret++] = mb->v1_vector;

                 else

                     for(z = 0; z < 4; z++)

-                        buf[ret++] = mb->v4_vector[v4][z];

+                        buf[ret++] = mb->v4_vector[z];

             }

         }

@@ -456,7 +667,7 @@ static int encode_mode(CinepakEncContext *s, CinepakMode mode, int h, int v1_siz

                 temp[temp_size++] = mb->v1_vector;

             else if(mb->best_encoding == ENC_V4)

                 for(z = 0; z < 4; z++)

-                    temp[temp_size++] = mb->v4_vector[v4][z];

+                    temp[temp_size++] = mb->v4_vector[z];

             if(should_write_temp) {

                 memcpy(&buf[ret], temp, temp_size);

@@ -492,7 +703,7 @@ static int compute_mb_distortion(CinepakEncContext *s, AVPicture *a, AVPicture *

         }

     }

-    if(s->pix_fmt == AV_PIX_FMT_YUV420P) {

+    if(s->pix_fmt == AV_PIX_FMT_RGB24) {

         for(p = 1; p <= 2; p++) {

             for(y = 0; y < MB_SIZE/2; y++) {

                 for(x = 0; x < MB_SIZE/2; x++) {

@@ -506,19 +717,30 @@ static int compute_mb_distortion(CinepakEncContext *s, AVPicture *a, AVPicture *

     return ret;

 }

-static int quantize(CinepakEncContext *s, int h, AVPicture *pict, int v1mode, int size, int v4, strip_info *info)

+// return the possibly adjusted size of the codebook

+#define CERTAIN(x) ((x)!=ENC_UNCERTAIN)

+static int quantize(CinepakEncContext *s, int h, AVPicture *pict,

+                    int v1mode, strip_info *info,

+                    mb_encoding encoding)

 {

-    int x, y, i, j, k, x2, y2, x3, y3, plane, shift;

-    int entry_size = s->pix_fmt == AV_PIX_FMT_YUV420P ? 6 : 4;

+    int x, y, i, j, k, x2, y2, x3, y3, plane, shift, mbn;

+    int entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4;

     int *codebook = v1mode ? info->v1_codebook : info->v4_codebook;

+    int size = v1mode ? info->v1_size : info->v4_size;

     int64_t total_error = 0;

     uint8_t vq_pict_buf[(MB_AREA*3)/2];

     AVPicture sub_pict, vq_pict;

-    for(i = y = 0; y < h; y += MB_SIZE) {

-        for(x = 0; x < s->w; x += MB_SIZE, i += v1mode ? 1 : 4) {

-            int *base = s->codebook_input + i*entry_size;

+    for(mbn = i = y = 0; y < h; y += MB_SIZE) {

+        for(x = 0; x < s->w; x += MB_SIZE, ++mbn) {

+            int *base;

+            if(CERTAIN(encoding)) {

+// use for the training only the blocks known to be to be encoded [sic:-]

+               if(s->mb[mbn].best_encoding != encoding) continue;

+            }

+

+            base = s->codebook_input + i*entry_size;

             if(v1mode) {

                 //subsample

                 for(j = y2 = 0; y2 < entry_size; y2 += 2) {

@@ -553,8 +775,19 @@ static int quantize(CinepakEncContext *s, int h, AVPicture *pict, int v1mode, in

                     }

                 }

             }

+            i += v1mode ? 1 : 4;

         }

     }

+//    if(i < mbn*(v1mode ? 1 : 4)) {

+//        av_log(s->avctx, AV_LOG_INFO, "reducing training set for %s from %i to %i (encoding %i)\n", v1mode?"v1":"v4", mbn*(v1mode ? 1 : 4), i, encoding);

+//    }

+

+    if(i == 0) // empty training set, nothing to do

+        return 0;

+    if(i < size) {

+        //av_log(s->avctx, (CERTAIN(encoding) ? AV_LOG_ERROR : AV_LOG_INFO), "WOULD WASTE: %s cbsize %i bigger than training set size %i (encoding %i)\n", v1mode?"v1":"v4", size, i, encoding);

+        size = i;

+    }

     ff_init_elbg(s->codebook_input, entry_size, i, codebook, size, 1, s->codebook_closest, &s->randctx);

     ff_do_elbg(s->codebook_input, entry_size, i, codebook, size, 1, s->codebook_closest, &s->randctx);

@@ -568,8 +801,11 @@ static int quantize(CinepakEncContext *s, int h, AVPicture *pict, int v1mode, in

     //copy indices

     for(i = j = y = 0; y < h; y += MB_SIZE) {

-        for(x = 0; x < s->w; x += MB_SIZE, j++, i += v1mode ? 1 : 4) {

+        for(x = 0; x < s->w; x += MB_SIZE, j++) {

             mb_info *mb = &s->mb[j];

+// skip uninteresting blocks if we know their preferred encoding

+            if(CERTAIN(encoding) && mb->best_encoding != encoding)

+                continue;

             //point sub_pict to current MB

             get_sub_picture(s, x, y, pict, &sub_pict);

@@ -578,26 +814,29 @@ static int quantize(CinepakEncContext *s, int h, AVPicture *pict, int v1mode, in

                 mb->v1_vector = s->codebook_closest[i];

                 //fill in vq_pict with V1 data

-                decode_v1_vector(s, &vq_pict, mb, info);

+                decode_v1_vector(s, &vq_pict, mb->v1_vector, info);

                 mb->v1_error = compute_mb_distortion(s, &sub_pict, &vq_pict);

                 total_error += mb->v1_error;

             } else {

                 for(k = 0; k < 4; k++)

-                    mb->v4_vector[v4][k] = s->codebook_closest[i+k];

+                    mb->v4_vector[k] = s->codebook_closest[i+k];

                 //fill in vq_pict with V4 data

-                decode_v4_vector(s, &vq_pict, mb->v4_vector[v4], info);

+                decode_v4_vector(s, &vq_pict, mb->v4_vector, info);

-                mb->v4_error[v4] = compute_mb_distortion(s, &sub_pict, &vq_pict);

-                total_error += mb->v4_error[v4];

+                mb->v4_error = compute_mb_distortion(s, &sub_pict, &vq_pict);

+                total_error += mb->v4_error;

             }

+            i += v1mode ? 1 : 4;

         }

     }

+// check that we did it right in the beginning of the function

+    av_assert0(i >= size); // training set is no smaller than the codebook

-    //av_log(s->avctx, AV_LOG_INFO, "mode %i size %i i %i error %li\n", v1mode, size, i, total_error);

+    //av_log(s->avctx, AV_LOG_INFO, "isv1 %i size= %i i= %i error %lli\n", v1mode, size, i, (long long int)total_error);

-    return 0;

+    return size;

 }

 static void calculate_skip_errors(CinepakEncContext *s, int h, AVPicture *last_pict, AVPicture *pict, strip_info *info)

@@ -617,59 +856,151 @@ static void calculate_skip_errors(CinepakEncContext *s, int h, AVPicture *last_p

 static void write_strip_header(CinepakEncContext *s, int y, int h, int keyframe, unsigned char *buf, int strip_size)

 {

-    buf[0] = keyframe ? 0x11: 0x10;

+// actually we are exclusively using intra strip coding (how much can we win

+// otherwise? how to choose which part of a codebook to update?),

+// keyframes are different only because we disallow ENC_SKIP on them -- rl

+// (besides, the logic here used to be inverted: )

+//    buf[0] = keyframe ? 0x11: 0x10;

+    buf[0] = keyframe ? 0x10: 0x11;

     AV_WB24(&buf[1], strip_size + STRIP_HEADER_SIZE);

-    AV_WB16(&buf[4], y);

+//    AV_WB16(&buf[4], y); /* using absolute y values works -- rl */

+    AV_WB16(&buf[4], 0); /* using relative values works as well -- rl */

     AV_WB16(&buf[6], 0);

-    AV_WB16(&buf[8], h);

+//    AV_WB16(&buf[8], y+h); /* using absolute y values works -- rl */

+    AV_WB16(&buf[8], h); /* using relative values works as well -- rl */

     AV_WB16(&buf[10], s->w);

+    //av_log(s->avctx, AV_LOG_INFO, "write_strip_header() %x keyframe=%d\n", buf[0], keyframe);

 }

-static int rd_strip(CinepakEncContext *s, int y, int h, int keyframe, AVPicture *last_pict, AVPicture *pict, AVPicture *scratch_pict, unsigned char *buf, int64_t *best_score)

+static int rd_strip(CinepakEncContext *s, int y, int h, int keyframe, AVPicture *last_pict, AVPicture *pict, AVPicture *scratch_pict, unsigned char *buf, int64_t *best_score

+#ifdef CINEPAK_REPORT_SERR

+, int64_t *best_serr

+#endif

+)

 {

     int64_t score = 0;

-    int best_size = 0, v1_size, v4_size, v4, mb_count = s->w * h / MB_AREA;

+#ifdef CINEPAK_REPORT_SERR

+    int64_t serr;

+#endif

+    int best_size = 0;

     strip_info info;

-    CinepakMode best_mode;

-    int v4_codebooks[CODEBOOK_NUM][CODEBOOK_MAX*VECTOR_MAX];

+// for codebook optimization:

+    int v1enough, v1_size, v4enough, v4_size;

+    int new_v1_size, new_v4_size;

+    int v1shrunk, v4shrunk;

     if(!keyframe)

         calculate_skip_errors(s, h, last_pict, pict, &info);

-    //precompute V4 codebooks

-    for(v4_size = 1, v4 = 0; v4_size <= 256; v4_size <<= 2, v4++) {

-        info.v4_codebook = v4_codebooks[v4];

-        quantize(s, h, pict, 0, v4_size, v4, &info);

-    }

-

-    //try all powers of 4 for the size of the codebooks

-    //constraint the v4 codebook to be no bigger than the v1 codebook

-    for(v1_size = 1; v1_size <= 256; v1_size <<= 2) {

-        //compute V1 codebook

-        quantize(s, h, pict, 1, v1_size, -1, &info);

-

-        for(v4_size = 0, v4 = -1; v4_size <= v1_size; v4_size = v4_size ? v4_size << 2 : v1_size >= 4 ? v1_size >> 2 : 1, v4++) {

+    //try some powers of 4 for the size of the codebooks

+    //constraint the v4 codebook to be no bigger than v1 one,

+    //(and no less than v1_size/4)

+    //thus making v1 preferable and possibly losing small details? should be ok

+#define SMALLEST_CODEBOOK 1

+    for(v1enough = 0, v1_size = SMALLEST_CODEBOOK; v1_size <= CODEBOOK_MAX && !v1enough; v1_size <<= 2) {

+        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) {

             //try all modes

             for(CinepakMode mode = 0; mode < MODE_COUNT; mode++) {

-                //don't allow MODE_MC in inter frames

+                //don't allow MODE_MC in intra frames

                 if(keyframe && mode == MODE_MC)

                     continue;

-                //only allow V1-only mode if v4 codebook is empty

-                if(!v4_size && mode != MODE_V1_ONLY)

-                    continue;

+                if(mode == MODE_V1_ONLY) {

+                    info.v1_size = v1_size;

+// the size may shrink even before optimizations if the input is short:

+                    info.v1_size = quantize(s, h, pict, 1, &info, ENC_UNCERTAIN);

+                    if(info.v1_size < v1_size)

+// too few eligible blocks, no sense in trying bigger sizes

+                        v1enough = 1;

+

+                    info.v4_size = 0;

+                } else { // mode != MODE_V1_ONLY

+                    // if v4 codebook is empty then only allow V1-only mode

+                    if(!v4_size)

+                        continue;

+

+                    if(mode == MODE_V1_V4) {

+                        info.v4_size = v4_size;

+                        info.v4_size = quantize(s, h, pict, 0, &info, ENC_UNCERTAIN);

+                        if(info.v4_size < v4_size)

+// too few eligible blocks, no sense in trying bigger sizes

+                            v4enough = 1;

+                    }

+                }

-                info.v4_codebook = v4 >= 0 ? v4_codebooks[v4] : NULL;

-                score = calculate_mode_score(s, mode, h, v1_size, v4_size, v4, &info);

+                info.mode = mode;

+// choose the best encoding per block, based on current experience

+                score = calculate_mode_score(s, h, &info, 0,