comparison dv.c @ 1887:85fe2f4633ec libavcodec

* DV decoding/encoding now supports MultiThreading for up to 324 CPUs ;-)

1886:fbcf02596520

#include "dsputil.h"
#include "mpegvideo.h"
#include "simple_idct.h"
#include "dvdata.h"

typedef struct DVVideoDecodeContext {
    const DVprofile* sys;
    AVFrame picture;
    
    uint8_t dv_zigzag[2][64];
    uint8_t dv_idct_shift[2][22][64];
  
    void (*get_pixels)(DCTELEM *block, const uint8_t *pixels, int line_size);
    void (*fdct[2])(DCTELEM *block);
    void (*idct_put[2])(uint8_t *dest, int line_size, DCTELEM *block);
    
    GetBitContext gb;
    DCTELEM block[5*6][64] __align8;
} DVVideoDecodeContext;

#define TEX_VLC_BITS 9

#ifdef DV_CODEC_TINY_TARGET
#define DV_VLC_MAP_RUN_SIZE 15
#define DV_VLC_MAP_LEV_SIZE 23
#else
#define DV_VLC_MAP_RUN_SIZE  64 
#define DV_VLC_MAP_LEV_SIZE 512
#endif

/* XXX: also include quantization */
static RL_VLC_ELEM *dv_rl_vlc;
/* VLC encoding lookup table */
static struct dv_vlc_pair {
   uint32_t vlc;
   uint8_t  size;
} (*dv_vlc_map)[DV_VLC_MAP_LEV_SIZE] = NULL;

static void dv_build_unquantize_tables(DVVideoDecodeContext *s, uint8_t* perm)
{
    int i, q, j;

    /* NOTE: max left shift is 6 */
    for(q = 0; q < 22; q++) {

    }
}

static int dvvideo_init(AVCodecContext *avctx)
{
    DVVideoDecodeContext *s = avctx->priv_data;
    DSPContext dsp;
    static int done=0;
    int i, j;

    if (!done) {

        dv_vlc_map = av_mallocz(DV_VLC_MAP_LEV_SIZE*DV_VLC_MAP_RUN_SIZE*sizeof(struct dv_vlc_pair));
	if (!dv_vlc_map)
	    return -ENOMEM;

        /* NOTE: as a trick, we use the fact the no codes are unused
           to accelerate the parsing of partial codes */
        init_vlc(&dv_vlc, TEX_VLC_BITS, NB_DV_VLC, 
                 dv_vlc_len, 1, 1, dv_vlc_bits, 2, 2);

        dv_rl_vlc = av_malloc(dv_vlc.table_size * sizeof(RL_VLC_ELEM));
	if (!dv_rl_vlc) {
	    av_free(dv_vlc_map);
	    return -ENOMEM;
	}
        for(i = 0; i < dv_vlc.table_size; i++){
            int code= dv_vlc.table[i][0];

    avctx->coded_frame = &s->picture;
    
    return 0;
}

// #define VLC_DEBUG

typedef struct BlockInfo {
    const uint8_t *shift_table;
    const uint8_t *scan_table;

#ifndef ALT_BITSTREAM_READER
#warning only works with ALT_BITSTREAM_READER
#endif

/* decode ac coefs */
static void dv_decode_ac(DVVideoDecodeContext *s, BlockInfo *mb, DCTELEM *block)
{
    int last_index = get_bits_size(&s->gb);
    int last_re_index;
    int shift_offset = mb->shift_offset;
    const uint8_t *scan_table = mb->scan_table;
    const uint8_t *shift_table = mb->shift_table;
    int pos = mb->pos;

    int sign = 0;
#ifndef ALT_BITSTREAM_READER //FIXME
    int re_index=0; 
    int re1_index=0;
#endif
    OPEN_READER(re, &s->gb);
    
#ifdef VLC_DEBUG
    printf("start\n");
#endif

        int l;
        GetBitContext gb1;

        /* build the dummy bit buffer */
        l = 16 - partial_bit_count;
        UPDATE_CACHE(re, &s->gb);
#ifdef VLC_DEBUG
        printf("show=%04x\n", SHOW_UBITS(re, &s->gb, 16));
#endif
        v = (mb->partial_bit_buffer << l) | SHOW_UBITS(re, &s->gb, l);
        buf[0] = v >> 8;
        buf[1] = v;
#ifdef VLC_DEBUG
        printf("v=%04x cnt=%d %04x\n", 
               v, partial_bit_count, (mb->partial_bit_buffer << l));

        last_re_index = 0; /* avoid warning */
        re_index += l;
        /* by definition, if we can read the vlc, all partial bits
           will be read (otherwise we could have read the vlc before) */
        mb->partial_bit_count = 0;
        UPDATE_CACHE(re, &s->gb);
        goto handle_vlc;
    }

    /* get the AC coefficients until last_index is reached */
    for(;;) {
        UPDATE_CACHE(re, &s->gb);
#ifdef VLC_DEBUG
        printf("%2d: bits=%04x index=%d\n", 
               pos, SHOW_UBITS(re, &s->gb, 16), re_index);
#endif
        last_re_index = re_index;
        GET_RL_VLC(level, run, re, &s->gb, dv_rl_vlc, 
                   TEX_VLC_BITS, 2);
    handle_vlc:
#ifdef VLC_DEBUG
        printf("run=%d level=%d\n", run, level);
#endif
	if (level) {
	    sign = SHOW_SBITS(re, &s->gb, 1);
	    LAST_SKIP_BITS(re, &s->gb, 1);
	}
	if (re_index > last_index) {
	    /* should be < 16 bits otherwise a codeword could have been parsed */
	    re_index = last_re_index;
	    UPDATE_CACHE(re, &s->gb);
	    mb->partial_bit_count = last_index - re_index;
	    mb->partial_bit_buffer = SHOW_UBITS(re, &s->gb, mb->partial_bit_count);
	    re_index = last_index;
	    break;
	}
        
	pos += run; 	

            level = level << (shift_table[pos1] + shift_offset);
            block[pos1] = level;
            //            printf("run=%d level=%d shift=%d\n", run, level, shift_table[pos1]);
        } 
    }
    CLOSE_READER(re, &s->gb);
    mb->pos = pos;
}

static inline void bit_copy(PutBitContext *pb, GetBitContext *gb)
{

        put_bits(pb, bits_left, get_bits(gb, bits_left));
    }
}

/* mb_x and mb_y are in units of 8 pixels */
static inline void dv_decode_video_segment(DVVideoDecodeContext *s, 
                                           uint8_t *buf_ptr1, 
                                           const uint16_t *mb_pos_ptr)
{
    int quant, dc, dct_mode, class1, j;
    int mb_index, mb_x, mb_y, v, last_index;
    DCTELEM *block, *block1;
    int c_offset;
    uint8_t *y_ptr;
    BlockInfo mb_data[5 * 6], *mb, *mb1;
    void (*idct_put)(uint8_t *dest, int line_size, DCTELEM *block);
    uint8_t *buf_ptr;
    PutBitContext pb, vs_pb;
    uint8_t mb_bit_buffer[80 + 4]; /* allow some slack */
    uint8_t vs_bit_buffer[5 * 80 + 4]; /* allow some slack */
    
    memset(s->block, 0, sizeof(s->block));

    /* pass 1 : read DC and AC coefficients in blocks */
    buf_ptr = buf_ptr1;
    block1 = &s->block[0][0];
    mb1 = mb_data;
    init_put_bits(&vs_pb, vs_bit_buffer, 5 * 80);
    for(mb_index = 0; mb_index < 5; mb_index++, mb1 += 6, block1 += 6 * 64) {
        /* skip header */
        quant = buf_ptr[3] & 0x0f;

        init_put_bits(&pb, mb_bit_buffer, 80);
        mb = mb1;
        block = block1;
        for(j = 0;j < 6; j++) {
            last_index = block_sizes[j];
	    init_get_bits(&s->gb, buf_ptr, last_index);
            
            /* get the dc */
            dc = get_bits(&s->gb, 9);
            dc = (dc << (32 - 9)) >> (32 - 9);
            dct_mode = get_bits1(&s->gb);
            mb->dct_mode = dct_mode;
            mb->scan_table = s->dv_zigzag[dct_mode];
            class1 = get_bits(&s->gb, 2);
            mb->shift_offset = (class1 == 3);
            mb->shift_table = s->dv_idct_shift[dct_mode]
                [quant + dv_quant_offset[class1]];
            dc = dc << 2;
            /* convert to unsigned because 128 is not added in the

            mb->partial_bit_count = 0;

#ifdef VLC_DEBUG
            printf("MB block: %d, %d ", mb_index, j);
#endif
            dv_decode_ac(s, mb, block);

            /* write the remaining bits  in a new buffer only if the
               block is finished */
            if (mb->pos >= 64)
                bit_copy(&pb, &s->gb);
            
            block += 64;
            mb++;
        }
        

#ifdef VLC_DEBUG
        printf("***pass 2 size=%d MB#=%d\n", put_bits_count(&pb), mb_index);
#endif
        block = block1;
        mb = mb1;
        init_get_bits(&s->gb, mb_bit_buffer, put_bits_count(&pb));
	flush_put_bits(&pb);
        for(j = 0;j < 6; j++, block += 64, mb++) {
            if (mb->pos < 64 && get_bits_left(&s->gb) > 0) {
                dv_decode_ac(s, mb, block);
                /* if still not finished, no need to parse other blocks */
                if (mb->pos < 64)
                    break;
            }
        }
        /* all blocks are finished, so the extra bytes can be used at
           the video segment level */
        if (j >= 6)
	    bit_copy(&vs_pb, &s->gb);
    }

    /* we need a pass other the whole video segment */
#ifdef VLC_DEBUG
    printf("***pass 3 size=%d\n", put_bits_count(&vs_pb));
#endif
    block = &s->block[0][0];
    mb = mb_data;
    init_get_bits(&s->gb, vs_bit_buffer, put_bits_count(&vs_pb));
    flush_put_bits(&vs_pb);
    for(mb_index = 0; mb_index < 5; mb_index++) {
        for(j = 0;j < 6; j++) {
            if (mb->pos < 64) {
#ifdef VLC_DEBUG
                printf("start %d:%d\n", mb_index, j);
#endif
                dv_decode_ac(s, mb, block);
            }
	    if (mb->pos >= 64 && mb->pos < 127)
		av_log(NULL, AV_LOG_ERROR, "AC EOB marker is absent pos=%d\n", mb->pos);
            block += 64;
            mb++;
        }
    }
    
    /* compute idct and place blocks */
    block = &s->block[0][0];
    mb = mb_data;
    for(mb_index = 0; mb_index < 5; mb_index++) {
        v = *mb_pos_ptr++;
        mb_x = v & 0xff;
        mb_y = v >> 8;

/*
 * This is a very rough initial implementaion. The performance is
 * horrible and the weighting is missing. But it's missing from the 
 * decoding step also -- so at least we're on the same page with decoder ;-)
 */
static inline void dv_encode_video_segment(DVVideoDecodeContext *s, 
                                           uint8_t *dif, 
                                           const uint16_t *mb_pos_ptr)
{
    int mb_index, i, j, v;
    int mb_x, mb_y, c_offset, linesize; 
    uint8_t*  y_ptr;
    uint8_t*  data;
    uint8_t*  ptr;
    int       do_edge_wrap;
    DCTELEM   block[64] __align8;
    EncBlockInfo  enc_blks[5*6];
    PutBitContext pbs[5*6];
    PutBitContext* pb; 
    EncBlockInfo* enc_blk;
    int       vs_bit_size = 0;

	    } else {             /* Simple copy: 8x8 -> 8x8 */
	        s->get_pixels(block, data, linesize);
	    }
	  
            enc_blk->dct_mode = dv_guess_dct_mode(block);
	    enc_blk->mb = &s->block[mb_index*6+j][0];
	    enc_blk->area_q[0] = enc_blk->area_q[1] = enc_blk->area_q[2] = enc_blk->area_q[3] = 0;
	    enc_blk->partial_bit_count = 0;
	    enc_blk->partial_bit_buffer = 0;
	    enc_blk->cur_ac = 1;
	    

    for (j=0; j<5*6; j++)
       flush_put_bits(&pbs[j]);
}

/* NOTE: exactly one frame must be given (120000 bytes for NTSC,
   144000 bytes for PAL) */
static int dvvideo_decode_frame(AVCodecContext *avctx, 
                                 void *data, int *data_size,
                                 uint8_t *buf, int buf_size)
{
    DVVideoDecodeContext *s = avctx->priv_data;
    int ds, vs;
    const uint16_t *mb_pos_ptr;
  
    *data_size=0;
    /* special case for last picture */
    if(buf_size==0)
        return 0;
    
    s->sys = dv_frame_profile(buf);
    if (!s->sys || buf_size < s->sys->frame_size)
        return -1; /* NOTE: we only accept several full frames */

	
    if(s->picture.data[0])
        avctx->release_buffer(avctx, &s->picture);
    
    s->picture.reference = 0;
    avctx->pix_fmt = s->sys->pix_fmt;

        return -1;
    }
    s->picture.interlaced_frame = 1;
    s->picture.top_field_first = 0;

    /* for each DIF segment */
    mb_pos_ptr = s->sys->video_place;
    for (ds = 0; ds < s->sys->difseg_size; ds++) {
        buf += 6 * 80; /* skip DIF segment header */
        
        for(vs = 0; vs < 27; vs++) {
            if ((vs % 3) == 0)
	        buf += 80; /* skip audio block */
            
#ifdef VLC_DEBUG
            printf("********************* %d, %d **********************\n", ds, vs);
#endif
	    dv_decode_video_segment(s, buf, mb_pos_ptr);
            buf += 5 * 80;
            mb_pos_ptr += 5;
        }
    }

    emms_c();

    /* return image */
    *data_size = sizeof(AVFrame);
    *(AVFrame*)data= s->picture;

}

static int dvvideo_encode_frame(AVCodecContext *c, uint8_t *buf, int buf_size, 
                                void *data)
{
    DVVideoDecodeContext *s = c->priv_data;
    const uint16_t *mb_pos_ptr;
    int ds, vs;

    s->sys = dv_codec_profile(c);
    if (!s->sys)
	return -1;
    
    c->pix_fmt = s->sys->pix_fmt;
    s->picture = *((AVFrame *)data);

    /* for each DIF segment */
    mb_pos_ptr = s->sys->video_place;
    for (ds = 0; ds < s->sys->difseg_size; ds++) {
        buf += 6 * 80; /* skip DIF segment header */
        
        for(vs = 0; vs < 27; vs++) {
            if ((vs % 3) == 0)
	        buf += 80; /* skip audio block */

#ifdef VLC_DEBUG
            printf("********************* %d, %d **********************\n", ds, vs);
#endif
	    dv_encode_video_segment(s, buf, mb_pos_ptr);
            buf += 5 * 80;
            mb_pos_ptr += 5;
        }
    }

    emms_c();
    return s->sys->frame_size;
}

static int dvvideo_end(AVCodecContext *avctx)
{
    avcodec_default_free_buffers(avctx);    
    return 0;
}

AVCodec dvvideo_decoder = {
    "dvvideo",
    CODEC_TYPE_VIDEO,
    CODEC_ID_DVVIDEO,
    sizeof(DVVideoDecodeContext),
    dvvideo_init,
    dvvideo_encode_frame,
    dvvideo_end,
    dvvideo_decode_frame,
    CODEC_CAP_DR1,
    NULL
};

1887:85fe2f4633ec

#include "dsputil.h"
#include "mpegvideo.h"
#include "simple_idct.h"
#include "dvdata.h"

typedef struct DVVideoContext {
    const DVprofile* sys;
    AVFrame picture;
    uint8_t *buf;
    
    uint8_t dv_zigzag[2][64];
    uint8_t dv_idct_shift[2][22][64];
  
    void (*get_pixels)(DCTELEM *block, const uint8_t *pixels, int line_size);
    void (*fdct[2])(DCTELEM *block);
    void (*idct_put[2])(uint8_t *dest, int line_size, DCTELEM *block);
} DVVideoContext;

#define TEX_VLC_BITS 9

#ifdef DV_CODEC_TINY_TARGET
#define DV_VLC_MAP_RUN_SIZE 15
#define DV_VLC_MAP_LEV_SIZE 23
#else
#define DV_VLC_MAP_RUN_SIZE  64 
#define DV_VLC_MAP_LEV_SIZE 512
#endif

/* MultiThreading */
static uint8_t** dv_anchor;

/* XXX: also include quantization */
static RL_VLC_ELEM *dv_rl_vlc;
/* VLC encoding lookup table */
static struct dv_vlc_pair {
   uint32_t vlc;
   uint8_t  size;
} (*dv_vlc_map)[DV_VLC_MAP_LEV_SIZE] = NULL;

static void dv_build_unquantize_tables(DVVideoContext *s, uint8_t* perm)
{
    int i, q, j;

    /* NOTE: max left shift is 6 */
    for(q = 0; q < 22; q++) {

    }
}

static int dvvideo_init(AVCodecContext *avctx)
{
    DVVideoContext *s = avctx->priv_data;
    DSPContext dsp;
    static int done=0;
    int i, j;

    if (!done) {

        dv_vlc_map = av_mallocz(DV_VLC_MAP_LEV_SIZE*DV_VLC_MAP_RUN_SIZE*sizeof(struct dv_vlc_pair));
	if (!dv_vlc_map)
	    return -ENOMEM;

	dv_anchor = av_malloc(12*27*sizeof(void*));
	if (!dv_anchor) {
	    av_free(dv_vlc_map);
	    return -ENOMEM;
	}
	for (i=0; i<12*27; i++)
	    (int)dv_anchor[i] = i;
	
        /* NOTE: as a trick, we use the fact the no codes are unused
           to accelerate the parsing of partial codes */
        init_vlc(&dv_vlc, TEX_VLC_BITS, NB_DV_VLC, 
                 dv_vlc_len, 1, 1, dv_vlc_bits, 2, 2);

        dv_rl_vlc = av_malloc(dv_vlc.table_size * sizeof(RL_VLC_ELEM));
	if (!dv_rl_vlc) {
	    av_free(dv_anchor);
	    av_free(dv_vlc_map);
	    return -ENOMEM;
	}
        for(i = 0; i < dv_vlc.table_size; i++){
            int code= dv_vlc.table[i][0];

    avctx->coded_frame = &s->picture;
    
    return 0;
}

static int dvvideo_end(AVCodecContext *avctx)
{
    avcodec_default_free_buffers(avctx);    
    return 0;
}

// #define VLC_DEBUG

typedef struct BlockInfo {
    const uint8_t *shift_table;
    const uint8_t *scan_table;

#ifndef ALT_BITSTREAM_READER
#warning only works with ALT_BITSTREAM_READER
#endif

/* decode ac coefs */
static void dv_decode_ac(GetBitContext *gb, BlockInfo *mb, DCTELEM *block)
{
    int last_index = get_bits_size(gb);
    int last_re_index;
    int shift_offset = mb->shift_offset;
    const uint8_t *scan_table = mb->scan_table;
    const uint8_t *shift_table = mb->shift_table;
    int pos = mb->pos;

    int sign = 0;
#ifndef ALT_BITSTREAM_READER //FIXME
    int re_index=0; 
    int re1_index=0;
#endif
    OPEN_READER(re, gb);
    
#ifdef VLC_DEBUG
    printf("start\n");
#endif

        int l;
        GetBitContext gb1;

        /* build the dummy bit buffer */
        l = 16 - partial_bit_count;
        UPDATE_CACHE(re, gb);
#ifdef VLC_DEBUG
        printf("show=%04x\n", SHOW_UBITS(re, gb, 16));
#endif
        v = (mb->partial_bit_buffer << l) | SHOW_UBITS(re, gb, l);
        buf[0] = v >> 8;
        buf[1] = v;
#ifdef VLC_DEBUG
        printf("v=%04x cnt=%d %04x\n", 
               v, partial_bit_count, (mb->partial_bit_buffer << l));

        last_re_index = 0; /* avoid warning */
        re_index += l;
        /* by definition, if we can read the vlc, all partial bits
           will be read (otherwise we could have read the vlc before) */
        mb->partial_bit_count = 0;
        UPDATE_CACHE(re, gb);
        goto handle_vlc;
    }

    /* get the AC coefficients until last_index is reached */
    for(;;) {
        UPDATE_CACHE(re, gb);
#ifdef VLC_DEBUG
        printf("%2d: bits=%04x index=%d\n", 
               pos, SHOW_UBITS(re, gb, 16), re_index);
#endif
        last_re_index = re_index;
        GET_RL_VLC(level, run, re, gb, dv_rl_vlc, 
                   TEX_VLC_BITS, 2);
    handle_vlc:
#ifdef VLC_DEBUG
        printf("run=%d level=%d\n", run, level);
#endif
	if (level) {
	    sign = SHOW_SBITS(re, gb, 1);
	    LAST_SKIP_BITS(re, gb, 1);
	}
	if (re_index > last_index) {
	    /* should be < 16 bits otherwise a codeword could have been parsed */
	    re_index = last_re_index;
	    UPDATE_CACHE(re, gb);
	    mb->partial_bit_count = last_index - re_index;
	    mb->partial_bit_buffer = SHOW_UBITS(re, gb, mb->partial_bit_count);
	    re_index = last_index;
	    break;
	}
        
	pos += run; 	

            level = level << (shift_table[pos1] + shift_offset);
            block[pos1] = level;
            //            printf("run=%d level=%d shift=%d\n", run, level, shift_table[pos1]);
        } 
    }
    CLOSE_READER(re, gb);
    mb->pos = pos;
}

static inline void bit_copy(PutBitContext *pb, GetBitContext *gb)
{

        put_bits(pb, bits_left, get_bits(gb, bits_left));
    }
}

/* mb_x and mb_y are in units of 8 pixels */
static inline void dv_decode_video_segment(DVVideoContext *s, 
                                           uint8_t *buf_ptr1, 
                                           const uint16_t *mb_pos_ptr)
{
    int quant, dc, dct_mode, class1, j;
    int mb_index, mb_x, mb_y, v, last_index;
    DCTELEM *block, *block1;
    int c_offset;
    uint8_t *y_ptr;
    void (*idct_put)(uint8_t *dest, int line_size, DCTELEM *block);
    uint8_t *buf_ptr;
    PutBitContext pb, vs_pb;
    GetBitContext gb;
    BlockInfo mb_data[5 * 6], *mb, *mb1;
    DCTELEM sblock[5*6][64] __align8;
    uint8_t mb_bit_buffer[80 + 4]; /* allow some slack */
    uint8_t vs_bit_buffer[5 * 80 + 4]; /* allow some slack */
	    
    memset(sblock, 0, sizeof(sblock));

    /* pass 1 : read DC and AC coefficients in blocks */
    buf_ptr = buf_ptr1;
    block1 = &sblock[0][0];
    mb1 = mb_data;
    init_put_bits(&vs_pb, vs_bit_buffer, 5 * 80);
    for(mb_index = 0; mb_index < 5; mb_index++, mb1 += 6, block1 += 6 * 64) {
        /* skip header */
        quant = buf_ptr[3] & 0x0f;

        init_put_bits(&pb, mb_bit_buffer, 80);
        mb = mb1;
        block = block1;
        for(j = 0;j < 6; j++) {
            last_index = block_sizes[j];
	    init_get_bits(&gb, buf_ptr, last_index);
            
            /* get the dc */
            dc = get_bits(&gb, 9);
            dc = (dc << (32 - 9)) >> (32 - 9);
            dct_mode = get_bits1(&gb);
            mb->dct_mode = dct_mode;
            mb->scan_table = s->dv_zigzag[dct_mode];
            class1 = get_bits(&gb, 2);
            mb->shift_offset = (class1 == 3);
            mb->shift_table = s->dv_idct_shift[dct_mode]
                [quant + dv_quant_offset[class1]];
            dc = dc << 2;
            /* convert to unsigned because 128 is not added in the

            mb->partial_bit_count = 0;

#ifdef VLC_DEBUG
            printf("MB block: %d, %d ", mb_index, j);
#endif
            dv_decode_ac(&gb, mb, block);

            /* write the remaining bits  in a new buffer only if the
               block is finished */
            if (mb->pos >= 64)
                bit_copy(&pb, &gb);
            
            block += 64;
            mb++;
        }
        

#ifdef VLC_DEBUG
        printf("***pass 2 size=%d MB#=%d\n", put_bits_count(&pb), mb_index);
#endif
        block = block1;
        mb = mb1;
        init_get_bits(&gb, mb_bit_buffer, put_bits_count(&pb));
	flush_put_bits(&pb);
        for(j = 0;j < 6; j++, block += 64, mb++) {
            if (mb->pos < 64 && get_bits_left(&gb) > 0) {
                dv_decode_ac(&gb, mb, block);
                /* if still not finished, no need to parse other blocks */
                if (mb->pos < 64)
                    break;
            }
        }
        /* all blocks are finished, so the extra bytes can be used at
           the video segment level */
        if (j >= 6)
	    bit_copy(&vs_pb, &gb);
    }

    /* we need a pass other the whole video segment */
#ifdef VLC_DEBUG
    printf("***pass 3 size=%d\n", put_bits_count(&vs_pb));
#endif
    block = &sblock[0][0];
    mb = mb_data;
    init_get_bits(&gb, vs_bit_buffer, put_bits_count(&vs_pb));
    flush_put_bits(&vs_pb);
    for(mb_index = 0; mb_index < 5; mb_index++) {
        for(j = 0;j < 6; j++) {
            if (mb->pos < 64) {
#ifdef VLC_DEBUG
                printf("start %d:%d\n", mb_index, j);
#endif
                dv_decode_ac(&gb, mb, block);
            }
	    if (mb->pos >= 64 && mb->pos < 127)
		av_log(NULL, AV_LOG_ERROR, "AC EOB marker is absent pos=%d\n", mb->pos);
            block += 64;
            mb++;
        }
    }
    
    /* compute idct and place blocks */
    block = &sblock[0][0];
    mb = mb_data;
    for(mb_index = 0; mb_index < 5; mb_index++) {
        v = *mb_pos_ptr++;
        mb_x = v & 0xff;
        mb_y = v >> 8;

/*
 * This is a very rough initial implementaion. The performance is
 * horrible and the weighting is missing. But it's missing from the 
 * decoding step also -- so at least we're on the same page with decoder ;-)
 */
static inline void dv_encode_video_segment(DVVideoContext *s, 
                                           uint8_t *dif, 
                                           const uint16_t *mb_pos_ptr)
{
    int mb_index, i, j, v;
    int mb_x, mb_y, c_offset, linesize; 
    uint8_t*  y_ptr;
    uint8_t*  data;
    uint8_t*  ptr;
    int       do_edge_wrap;
    DCTELEM   block[64] __align8;
    DCTELEM   sblock[5*6][64] __align8;
    EncBlockInfo  enc_blks[5*6];
    PutBitContext pbs[5*6];
    PutBitContext* pb; 
    EncBlockInfo* enc_blk;
    int       vs_bit_size = 0;

	    } else {             /* Simple copy: 8x8 -> 8x8 */
	        s->get_pixels(block, data, linesize);
	    }
	  
            enc_blk->dct_mode = dv_guess_dct_mode(block);
	    enc_blk->mb = &sblock[mb_index*6+j][0];
	    enc_blk->area_q[0] = enc_blk->area_q[1] = enc_blk->area_q[2] = enc_blk->area_q[3] = 0;
	    enc_blk->partial_bit_count = 0;
	    enc_blk->partial_bit_buffer = 0;
	    enc_blk->cur_ac = 1;
	    

    for (j=0; j<5*6; j++)
       flush_put_bits(&pbs[j]);
}

static int dv_decode_mt(AVCodecContext *avctx, void* sl)
{
    DVVideoContext *s = avctx->priv_data;
    int slice = (int)sl;
    dv_decode_video_segment(s, &s->buf[((slice/27)*6+(slice/3)+slice*5+7)*80],
	                    &s->sys->video_place[slice*5]);
    return 0;
}

static int dv_encode_mt(AVCodecContext *avctx, void* sl)
{
    DVVideoContext *s = avctx->priv_data;
    int slice = (int)sl;
    dv_encode_video_segment(s, &s->buf[((slice/27)*6+(slice/3)+slice*5+7)*80],
	                    &s->sys->video_place[slice*5]);
    return 0;
}

/* NOTE: exactly one frame must be given (120000 bytes for NTSC,
   144000 bytes for PAL) */
static int dvvideo_decode_frame(AVCodecContext *avctx, 
                                 void *data, int *data_size,
                                 uint8_t *buf, int buf_size)
{
    DVVideoContext *s = avctx->priv_data;
  
    *data_size=0;
    /* special case for last picture */
    if(buf_size==0)
        return 0;
    
    s->sys = dv_frame_profile(buf);
    if (!s->sys || buf_size < s->sys->frame_size)
        return -1; /* NOTE: we only accept several full frames */

    if(s->picture.data[0])
        avctx->release_buffer(avctx, &s->picture);
    
    s->picture.reference = 0;
    avctx->pix_fmt = s->sys->pix_fmt;

        return -1;
    }
    s->picture.interlaced_frame = 1;
    s->picture.top_field_first = 0;

    s->buf = buf;
    avctx->execute(avctx, dv_decode_mt, (void**)&dv_anchor[0], NULL, 
	           s->sys->difseg_size * 27);
    
    emms_c();

    /* return image */
    *data_size = sizeof(AVFrame);
    *(AVFrame*)data= s->picture;

}

static int dvvideo_encode_frame(AVCodecContext *c, uint8_t *buf, int buf_size, 
                                void *data)
{
    DVVideoContext *s = c->priv_data;

    s->sys = dv_codec_profile(c);
    if (!s->sys)
	return -1;
    
    c->pix_fmt = s->sys->pix_fmt;
    s->picture = *((AVFrame *)data);

    s->buf = buf;
    c->execute(c, dv_encode_mt, (void**)&dv_anchor[0], NULL, 
	       s->sys->difseg_size * 27);

    emms_c();
    return s->sys->frame_size;
}

AVCodec dvvideo_encoder = {
    "dvvideo",
    CODEC_TYPE_VIDEO,
    CODEC_ID_DVVIDEO,
    sizeof(DVVideoContext),
    dvvideo_init,
    dvvideo_encode_frame,
    dvvideo_end,
    NULL,
    CODEC_CAP_DR1,
    NULL
};

AVCodec dvvideo_decoder = {
    "dvvideo",
    CODEC_TYPE_VIDEO,
    CODEC_ID_DVVIDEO,
    sizeof(DVVideoContext),
    dvvideo_init,
    NULL,
    dvvideo_end,
    dvvideo_decode_frame,
    CODEC_CAP_DR1,
    NULL
};