Add Cinepak encoder
authorTomas Härdin <tomas.hardin at codemill.se>
Tue, 27 Jun 2017 22:27:11 +0000 (00:27 +0200)
committerDiego Biurrun <diego@biurrun.de>
Wed, 5 Jul 2017 11:05:54 +0000 (13:05 +0200)
With permission of Tomas Härdin applied by Rl aetey.se

Signed-off-by: Diego Biurrun <diego@biurrun.de>
Changelog
doc/general.texi
libavcodec/Makefile
libavcodec/allcodecs.c
libavcodec/cinepakenc.c [new file with mode: 0644]

index e44df54..adcca3f 100644 (file)
--- a/Changelog
+++ b/Changelog
@@ -16,6 +16,7 @@ version <next>:
 - FM Screen Capture Codec decoder
 - ClearVideo decoder (I-frames only)
 - support for decoding through D3D11VA in avconv
+- Cinepak encoder
 
 
 version 12:
index 905e2d1..adf5b4e 100644 (file)
@@ -631,7 +631,7 @@ following image formats are supported:
     @tab Codec used in Delphine Software International games.
 @item Discworld II BMV Video @tab     @tab  X
 @item Canopus Lossless Codec @tab     @tab  X
-@item Cinepak                @tab     @tab  X
+@item Cinepak                @tab  X  @tab  X
 @item Cirrus Logic AccuPak   @tab  X  @tab  X
     @tab fourcc: CLJR
 @item Creative YUV (CYUV)    @tab     @tab  X
index 240f7d8..2b91588 100644 (file)
@@ -186,6 +186,7 @@ OBJS-$(CONFIG_CDGRAPHICS_DECODER)      += cdgraphics.o
 OBJS-$(CONFIG_CDXL_DECODER)            += cdxl.o
 OBJS-$(CONFIG_CFHD_DECODER)            += cfhd.o cfhddata.o
 OBJS-$(CONFIG_CINEPAK_DECODER)         += cinepak.o
+OBJS-$(CONFIG_CINEPAK_ENCODER)         += cinepakenc.o
 OBJS-$(CONFIG_CLEARVIDEO_DECODER)      += clearvideo.o
 OBJS-$(CONFIG_CLJR_DECODER)            += cljrdec.o
 OBJS-$(CONFIG_CLJR_ENCODER)            += cljrenc.o
index 70c35e9..5635ae1 100644 (file)
@@ -133,7 +133,7 @@ void avcodec_register_all(void)
     REGISTER_DECODER(CDGRAPHICS,        cdgraphics);
     REGISTER_DECODER(CDXL,              cdxl);
     REGISTER_DECODER(CFHD,              cfhd);
-    REGISTER_DECODER(CINEPAK,           cinepak);
+    REGISTER_ENCDEC (CINEPAK,           cinepak);
     REGISTER_DECODER(CLEARVIDEO,        clearvideo);
     REGISTER_ENCDEC (CLJR,              cljr);
     REGISTER_DECODER(CLLC,              cllc);
diff --git a/libavcodec/cinepakenc.c b/libavcodec/cinepakenc.c
new file mode 100644 (file)
index 0000000..2809295
--- /dev/null
@@ -0,0 +1,822 @@
+/*
+ * Cinepak encoder (c) 2011 Tomas Härdin
+ * http://titan.codemill.se/~tomhar/cinepakenc.patch
+
+Permission is hereby granted, free of charge, to any person obtaining a
+copy of this software and associated documentation files (the "Software"),
+to deal in the Software without restriction, including without limitation
+the rights to use, copy, modify, merge, publish, distribute, sublicense,
+and/or sell copies of the Software, and to permit persons to whom the
+Software is furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included
+in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
+THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
+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.
+
+ */
+
+#include "libavutil/common.h"
+#include "libavutil/intreadwrite.h"
+#include "avcodec.h"
+#include "libavutil/lfg.h"
+#include "elbg.h"
+
+#define CVID_HEADER_SIZE 10
+#define STRIP_HEADER_SIZE 12
+#define CHUNK_HEADER_SIZE 4
+
+#define MB_SIZE 4           //4x4 MBs
+#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 MAX_STRIPS  1       //Note: having fewer choices regarding the number of strip speeds up encoding (obviously)
+#define MIN_STRIPS  1       //Note: having more strips speeds up encoding the frame (this is less obvious)
+
+typedef enum {
+    MODE_V1_ONLY = 0,
+    MODE_V1_V4,
+    MODE_MC,
+
+    MODE_COUNT,
+} CinepakMode;
+
+typedef enum {
+    ENC_V1,
+    ENC_V4,
+    ENC_SKIP
+} 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 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;
+} strip_info;
+
+typedef struct {
+    AVCodecContext *avctx;
+    unsigned char *pict_bufs[3], *strip_buf, *frame_buf;
+    AVFrame last_frame;
+    AVFrame best_frame;
+    AVFrame scratch_frame;
+    enum AVPixelFormat pix_fmt;
+    int w, h;
+    int curframe, keyint;
+    AVLFG randctx;
+    uint64_t lambda;
+    int *codebook_input;
+    int *codebook_closest;
+    mb_info *mb;                                //MB RD state
+#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;
+} CinepakEncContext;
+
+static av_cold int cinepak_encode_init(AVCodecContext *avctx)
+{
+    CinepakEncContext *s = avctx->priv_data;
+    int x, mb_count, strip_buf_size, frame_buf_size;
+
+    if (avctx->width & 3 || avctx->height & 3) {
+        av_log(avctx, AV_LOG_ERROR, "width and height must be multiples of four (got %ix%i)\n",
+                avctx->width, avctx->height);
+        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)))
+        return AVERROR(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)))
+            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);
+
+    frame_buf_size = CVID_HEADER_SIZE + MAX_STRIPS * strip_buf_size;
+
+    if (!(s->strip_buf = av_malloc(strip_buf_size)))
+        goto enomem;
+
+    if (!(s->frame_buf = av_malloc(frame_buf_size)))
+        goto enomem;
+
+    if (!(s->mb = av_malloc(mb_count*sizeof(mb_info))))
+        goto enomem;
+
+#ifdef CINEPAKENC_DEBUG
+    if (!(s->best_mb = av_malloc(mb_count*sizeof(mb_info))))
+        goto enomem;
+#endif
+
+    av_lfg_init(&s->randctx, 1);
+    s->avctx = avctx;
+    s->w = avctx->width;
+    s->h = avctx->height;
+    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->num_v1_mode = s->num_v4_mode = s->num_mc_mode = s->num_v1_encs = s->num_v4_encs = s->num_skips = 0;
+
+    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);
+#ifdef CINEPAKENC_DEBUG
+    av_free(s->best_mb);
+#endif
+
+    for(x = 0; x < 3; x++)
+        av_free(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)
+{
+    //score = FF_LAMBDA_SCALE * error + lambda * bits
+    int x;
+    int entry_size = s->pix_fmt == AV_PIX_FMT_YUV420P ? 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) +
+                   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);
+
+    switch(mode) {
+    case MODE_V1_ONLY:
+        //one byte per MB
+        ret += s->lambda * 8 * mb_count;
+
+        for(x = 0; x < mb_count; x++) {
+            mb = &s->mb[x];
+            ret += FF_LAMBDA_SCALE * mb->v1_error;
+            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) {
+                ret += score1;
+                mb->best_encoding = ENC_V1;
+            } else {
+                ret += score2;
+                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;
+            }
+        }
+
+        break;
+    }
+
+    return ret;
+}
+
+static int write_chunk_header(unsigned char *buf, int chunk_type, int chunk_size)
+{
+    buf[0] = chunk_type;
+    AV_WB24(&buf[1], chunk_size + CHUNK_HEADER_SIZE);
+    return CHUNK_HEADER_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);
+
+    return ret;
+}
+
+//sets out to the sub picture starting at (x,y) in in
+static void get_sub_picture(CinepakEncContext *s, int x, int y, AVPicture *in, AVPicture *out)
+{
+    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) {
+        out->data[1] = in->data[1] + (x >> 1) + (y >> 1) * in->linesize[1];
+        out->linesize[1] = in->linesize[1];
+
+        out->data[2] = in->data[2] + (x >> 1) + (y >> 1) * in->linesize[2];
+        out->linesize[2] = in->linesize[2];
+    }
+}
+
+//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)
+{
+    int entry_size = s->pix_fmt == AV_PIX_FMT_YUV420P ? 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][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][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][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];
+
+    if(s->pix_fmt == AV_PIX_FMT_YUV420P) {
+        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[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];
+    }
+}
+
+//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;
+
+    for(i = y = 0; y < 4; y += 2) {
+        for(x = 0; x < 4; x += 2, i++) {
+            sub_pict->data[0][x   +     y*sub_pict->linesize[0]] = info->v4_codebook[v4_vector[i]*entry_size];
+            sub_pict->data[0][x+1 +     y*sub_pict->linesize[0]] = info->v4_codebook[v4_vector[i]*entry_size+1];
+            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) {
+                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];
+            }
+        }
+    }
+}
+
+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)
+{
+    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;
+
+    //encode codebooks
+    if(v1_size)
+        ret += encode_codebook(s, info->v1_codebook, v1_size, 0x22, 0x26, buf + ret);
+
+    if(v4_size)
+        ret += encode_codebook(s, info->v4_codebook, v4_size, 0x20, 0x24, 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);
+        }
+    }
+
+    switch(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);
+
+        for(x = 0; x < mb_count; x++)
+            buf[ret++] = s->mb[x].v1_vector;
+
+        break;
+    case MODE_V1_V4:
+        //remember header position
+        header_ofs = ret;
+        ret += CHUNK_HEADER_SIZE;
+
+        for(x = 0; x < mb_count; x += 32) {
+            flags = 0;
+            for(y = x; y < FFMIN(x+32, mb_count); y++)
+                if(s->mb[y].best_encoding == ENC_V4)
+                    flags |= 1 << (31 - y + x);
+
+            AV_WB32(&buf[ret], flags);
+            ret += 4;
+
+            for(y = x; y < FFMIN(x+32, mb_count); y++) {
+                mb = &s->mb[y];
+
+                if(mb->best_encoding == ENC_V1)
+                    buf[ret++] = mb->v1_vector;
+                else
+                    for(z = 0; z < 4; z++)
+                        buf[ret++] = mb->v4_vector[v4][z];
+            }
+        }
+
+        write_chunk_header(buf + header_ofs, 0x30, ret - header_ofs - CHUNK_HEADER_SIZE);
+
+        break;
+    case MODE_MC:
+        //remember header position
+        header_ofs = ret;
+        ret += CHUNK_HEADER_SIZE;
+        flags = bits = temp_size = 0;
+
+        for(x = 0; x < mb_count; x++) {
+            mb = &s->mb[x];
+            flags |= (mb->best_encoding != ENC_SKIP) << (31 - bits++);
+            needs_extra_bit = 0;
+            should_write_temp = 0;
+
+            if(mb->best_encoding != ENC_SKIP) {
+                if(bits < 32)
+                    flags |= (mb->best_encoding == ENC_V4) << (31 - bits++);
+                else
+                    needs_extra_bit = 1;
+            }
+
+            if(bits == 32) {
+                AV_WB32(&buf[ret], flags);
+                ret += 4;
+                flags = bits = 0;
+
+                if(mb->best_encoding == ENC_SKIP || needs_extra_bit) {
+                    memcpy(&buf[ret], temp, temp_size);
+                    ret += temp_size;
+                    temp_size = 0;
+                } else
+                    should_write_temp = 1;
+            }
+
+            if(needs_extra_bit) {
+                flags = (mb->best_encoding == ENC_V4) << 31;
+                bits = 1;
+            }
+
+            if(mb->best_encoding == ENC_V1)
+                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];
+
+            if(should_write_temp) {
+                memcpy(&buf[ret], temp, temp_size);
+                ret += temp_size;
+                temp_size = 0;
+            }
+        }
+
+        if(bits > 0) {
+            AV_WB32(&buf[ret], flags);
+            ret += 4;
+            memcpy(&buf[ret], temp, temp_size);
+            ret += temp_size;
+        }
+
+        write_chunk_header(buf + header_ofs, 0x31, ret - header_ofs - CHUNK_HEADER_SIZE);
+
+        break;
+    }
+
+    return ret;
+}
+
+//computes distortion of 4x4 MB in b compared to a
+static int compute_mb_distortion(CinepakEncContext *s, AVPicture *a, AVPicture *b)
+{
+    int x, y, p, d, ret = 0;
+
+    for(y = 0; y < MB_SIZE; y++) {
+        for(x = 0; x < MB_SIZE; x++) {
+            d = a->data[0][x + y*a->linesize[0]] - b->data[0][x + y*b->linesize[0]];
+            ret += d*d;
+        }
+    }
+
+    if(s->pix_fmt == AV_PIX_FMT_YUV420P) {
+        for(p = 1; p <= 2; p++) {
+            for(y = 0; y < MB_SIZE/2; y++) {
+                for(x = 0; x < MB_SIZE/2; x++) {
+                    d = a->data[p][x + y*a->linesize[p]] - b->data[p][x + y*b->linesize[p]];
+                    ret += d*d;
+                }
+            }
+        }
+    }
+
+    return ret;
+}
+
+static int quantize(CinepakEncContext *s, int h, AVPicture *pict, int v1mode, int size, int v4, strip_info *info)
+{
+    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 *codebook = v1mode ? info->v1_codebook : info->v4_codebook;
+    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;
+
+            if(v1mode) {
+                //subsample
+                for(j = y2 = 0; y2 < entry_size; y2 += 2) {
+                    for(x2 = 0; x2 < 4; x2 += 2, j++) {
+                        plane = y2 < 4 ? 0 : 1 + (x2 >> 1);
+                        shift = y2 < 4 ? 0 : 1;
+                        x3 = shift ? 0 : x2;
+                        y3 = shift ? 0 : y2;
+                        base[j] = (pict->data[plane][((x+x3) >> shift) +      ((y+y3) >> shift)      * pict->linesize[plane]] +
+                                   pict->data[plane][((x+x3) >> shift) + 1 +  ((y+y3) >> shift)      * pict->linesize[plane]] +
+                                   pict->data[plane][((x+x3) >> shift) +     (((y+y3) >> shift) + 1) * pict->linesize[plane]] +
+                                   pict->data[plane][((x+x3) >> shift) + 1 + (((y+y3) >> shift) + 1) * pict->linesize[plane]]) >> 2;
+                    }
+                }
+            } else {
+                //copy
+                for(j = y2 = 0; y2 < MB_SIZE; y2 += 2) {
+                    for(x2 = 0; x2 < MB_SIZE; x2 += 2) {
+                        for(k = 0; k < entry_size; k++, j++) {
+                            plane = k >= 4 ? k - 3 : 0;
+
+                            if(k >= 4) {
+                                x3 = (x+x2) >> 1;
+                                y3 = (y+y2) >> 1;
+                            } else {
+                                x3 = x + x2 + (k & 1);
+                                y3 = y + y2 + (k >> 1);
+                            }
+
+                            base[j] = pict->data[plane][x3 + y3*pict->linesize[plane]];
+                        }
+                    }
+                }
+            }
+        }
+    }
+
+    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);
+
+    //setup vq_pict, which contains a single MB
+    vq_pict.data[0] = vq_pict_buf;
+    vq_pict.linesize[0] = MB_SIZE;
+    vq_pict.data[1] = &vq_pict_buf[MB_AREA];
+    vq_pict.data[2] = vq_pict.data[1] + (MB_AREA >> 2);
+    vq_pict.linesize[1] = vq_pict.linesize[2] = MB_SIZE >> 1;
+
+    //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) {
+            mb_info *mb = &s->mb[j];
+
+            //point sub_pict to current MB
+            get_sub_picture(s, x, y, pict, &sub_pict);
+
+            if(v1mode) {
+                mb->v1_vector = s->codebook_closest[i];
+
+                //fill in vq_pict with V1 data
+                decode_v1_vector(s, &vq_pict, mb, 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];
+
+                //fill in vq_pict with V4 data
+                decode_v4_vector(s, &vq_pict, mb->v4_vector[v4], info);
+
+                mb->v4_error[v4] = compute_mb_distortion(s, &sub_pict, &vq_pict);
+                total_error += mb->v4_error[v4];
+            }
+        }
+    }
+
+    //av_log(s->avctx, AV_LOG_INFO, "mode %i size %i i %i error %li\n", v1mode, size, i, total_error);
+
+    return 0;
+}
+
+static void calculate_skip_errors(CinepakEncContext *s, int h, AVPicture *last_pict, AVPicture *pict, strip_info *info)
+{
+    int x, y, i;
+    AVPicture sub_last, sub_pict;
+
+    for(i = y = 0; y < h; y += MB_SIZE) {
+        for(x = 0; x < s->w; x += MB_SIZE, i++) {
+            get_sub_picture(s, x, y, last_pict, &sub_last);
+            get_sub_picture(s, x, y, pict,      &sub_pict);
+
+            s->mb[i].skip_error = compute_mb_distortion(s, &sub_last, &sub_pict);
+        }
+    }
+}
+
+static void write_strip_header(CinepakEncContext *s, int y, int h, int keyframe, unsigned char *buf, int strip_size)
+{
+    buf[0] = keyframe ? 0x11: 0x10;
+    AV_WB24(&buf[1], strip_size + STRIP_HEADER_SIZE);
+    AV_WB16(&buf[4], y);
+    AV_WB16(&buf[6], 0);
+    AV_WB16(&buf[8], h);
+    AV_WB16(&buf[10], s->w);
+}
+
+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)
+{
+    int64_t score = 0;
+    int best_size = 0, v1_size, v4_size, v4, mb_count = s->w * h / MB_AREA;
+    strip_info info;
+    CinepakMode best_mode;
+    int v4_codebooks[CODEBOOK_NUM][CODEBOOK_MAX*VECTOR_MAX];
+
+    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 all modes
+            for(CinepakMode mode = 0; mode < MODE_COUNT; mode++) {
+                //don't allow MODE_MC in inter 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;
+
+                info.v4_codebook = v4 >= 0 ? v4_codebooks[v4] : NULL;
+                score = calculate_mode_score(s, mode, h, v1_size, v4_size, v4, &info);
+
+                //av_log(s->avctx, AV_LOG_INFO, "%3i %3i score = %li\n", v1_size, v4_size, score);
+
+                if(best_size == 0 || score < *best_score) {
+                    *best_score = score;
+                    best_size = encode_mode(s, mode, h, v1_size, v4_size, v4, scratch_pict, &info, s->strip_buf + STRIP_HEADER_SIZE);
+                    best_mode = mode;
+
+                    av_log(s->avctx, AV_LOG_INFO, "mode %i, %3i, %3i: %18li %i B\n", mode, v1_size, v4_size, score, best_size);
+
+#ifdef CINEPAKENC_DEBUG
+                    //save MB encoding choices
+                    memcpy(s->best_mb, s->mb, mb_count*sizeof(mb_info));
+#endif
+
+                    //memcpy(strip_temp + STRIP_HEADER_SIZE, strip_temp, best_size);
+                    write_strip_header(s, y, h, keyframe, s->strip_buf, best_size);
+                }
+            }
+        }
+    }
+
+#ifdef CINEPAKENC_DEBUG
+    //gather stats. this will only work properly of MAX_STRIPS == 1
+    if(best_mode == MODE_V1_ONLY) {
+        s->num_v1_mode++;
+        s->num_v1_encs += s->w*h/MB_AREA;
+    } else {
+        if(best_mode == MODE_V1_V4)
+            s->num_v4_mode++;
+        else
+            s->num_mc_mode++;
+
+        int x;
+        for(x = 0; x < s->w*h/MB_AREA; x++)
+            if(s->best_mb[x].best_encoding == ENC_V1)
+                s->num_v1_encs++;
+            else if(s->best_mb[x].best_encoding == ENC_V4)
+                s->num_v4_encs++;
+            else
+                s->num_skips++;
+    }
+#endif
+
+    best_size += STRIP_HEADER_SIZE;
+    memcpy(buf, s->strip_buf, best_size);
+
+
+    return best_size;
+}
+
+static int write_cvid_header(CinepakEncContext *s, unsigned char *buf, int num_strips, int data_size)
+{
+    buf[0] = 0;
+    AV_WB24(&buf[1], data_size + CVID_HEADER_SIZE);
+    AV_WB16(&buf[4], s->w);
+    AV_WB16(&buf[6], s->h);
+    AV_WB16(&buf[8], num_strips);
+
+    return CVID_HEADER_SIZE;
+}
+
+static int rd_frame(CinepakEncContext *s, AVFrame *frame, unsigned char *buf, int buf_size)
+{
+    int num_strips, strip, h, i, y, size, temp_size, best_size;
+    AVPicture last_pict, pict, scratch_pict;
+    int64_t best_score = 0, score, score_temp;
+
+    //TODO: support encoding zero strips (meaning skip the whole frame)
+    for(num_strips = MIN_STRIPS; num_strips <= MAX_STRIPS && num_strips <= s->h / MB_SIZE; num_strips++) {
+        score = 0;
+        size = 0;
+        h = s->h / num_strips;
+        //make h into next multiple of 4
+        h += 4 - (h & 3);
+
+        for(strip = 0; strip < num_strips; strip++) {
+            y = strip*h;
+
+            get_sub_picture(s, 0, y, (AVPicture*)frame,            &pict);
+            get_sub_picture(s, 0, y, (AVPicture*)&s->last_frame,    &last_pict);
+            get_sub_picture(s, 0, y, (AVPicture*)&s->scratch_frame, &scratch_pict);
+
+            if((temp_size = rd_strip(s, y, FFMIN(h, s->h - y), frame->key_frame, &last_pict, &pict, &scratch_pict, s->frame_buf + CVID_HEADER_SIZE, &score_temp)) < 0)
+                return temp_size;
+
+            score += score_temp;
+            size += temp_size;
+        }
+
+        if(best_score == 0 || score < best_score) {
+            best_score = score;
+            best_size = size + write_cvid_header(s, s->frame_buf, num_strips, size);
+            av_log(s->avctx, AV_LOG_INFO, "best number of strips so far: %2i, %12li, %i B\n", num_strips, score, best_size);
+
+            FFSWAP(AVFrame, s->best_frame, s->scratch_frame);
+        }
+    }
+
+    memcpy(buf, s->frame_buf, best_size);
+
+    return best_size;
+}
+
+static int cinepak_encode_frame(AVCodecContext *avctx, unsigned char *buf, int buf_size, void *data)
+{
+    CinepakEncContext *s = avctx->priv_data;
+    AVFrame *frame = data;
+    int ret;
+
+    s->lambda = frame->quality ? frame->quality - 1 : 2 * FF_LAMBDA_SCALE;
+
+    frame->key_frame = s->curframe == 0;
+    frame->pict_type = frame->key_frame ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
+
+    ret = rd_frame(s, frame, buf, buf_size);
+
+    FFSWAP(AVFrame, s->last_frame, s->best_frame);
+
+    if (++s->curframe >= s->keyint)
+        s->curframe = 0;
+
+    return ret;
+}
+
+static av_cold int cinepak_encode_end(AVCodecContext *avctx)
+{
+    CinepakEncContext *s = avctx->priv_data;
+    int x;
+
+    av_free(s->codebook_input);
+    av_free(s->codebook_closest);
+    av_free(s->strip_buf);
+    av_free(s->frame_buf);
+    av_free(s->mb);
+#ifdef CINEPAKENC_DEBUG
+    av_free(s->best_mb);
+#endif
+
+    for(x = 0; x < 3; x++)
+        av_free(s->pict_bufs[x]);
+
+    av_log(avctx, AV_LOG_INFO, "strip coding stats: %i V1 mode, %i V4 mode, %i MC mode (%i V1 encs, %i V4 encs, %i skips)\n",
+        s->num_v1_mode, s->num_v4_mode, s->num_mc_mode, s->num_v1_encs, s->num_v4_encs, s->num_skips);
+
+    return 0;
+}
+
+AVCodec ff_cinepak_encoder = {
+    .name           = "cinepak",
+    .type           = AVMEDIA_TYPE_VIDEO,
+    .id             = AV_CODEC_ID_CINEPAK,
+    .priv_data_size = sizeof(CinepakEncContext),
+    .init           = cinepak_encode_init,
+    .encode2        = cinepak_encode_frame,
+    .close          = cinepak_encode_end,
+    .pix_fmts       = (const enum AVPixelFormat[]){AV_PIX_FMT_RGB24, AV_PIX_FMT_GRAY8, AV_PIX_FMT_NONE},
+    .long_name      = NULL_IF_CONFIG_SMALL("Cinepak"),
+};