comparison libmpeg2/slice.c @ 36:846535ace7a2

libmpeg2-0.2.0 merge

35:25f148e9890a

/*
 * slice.c
 * Copyright (C) 1999-2000 Aaron Holtzman <[e-mail]>
 *
 * This file is part of mpeg2dec, a free MPEG-2 video stream decoder.
 *
 * mpeg2dec is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by

#include "config.h"

#include <string.h>
#include <inttypes.h>

#include "mpeg2_internal.h"
#include "attributes.h"

extern mc_functions_t mc_functions;
extern void (* idct_block_copy) (int16_t * block, uint8_t * dest, int stride);
extern void (* idct_block_add) (int16_t * block, uint8_t * dest, int stride);

static int16_t DCTblock[64] ATTR_ALIGN(16);

#include "vlc.h"

static int non_linear_quantizer_scale [] = {
     0,  1,  2,  3,  4,  5,   6,   7,
     8, 10, 12, 14, 16, 18,  20,  22,
    24, 28, 32, 36, 40, 44,  48,  52,
    56, 64, 72, 80, 88, 96, 104, 112
};

static inline int get_macroblock_modes (slice_t * slice, int picture_structure,
					int picture_coding_type,
					int frame_pred_frame_dct)
{
#define bit_buf (slice->bitstream_buf)
#define bits (slice->bitstream_bits)
#define bit_ptr (slice->bitstream_ptr)
    int macroblock_modes;
    MBtab * tab;

    switch (picture_coding_type) {
    case I_TYPE:

	tab = MB_I + UBITS (bit_buf, 1);
	DUMPBITS (bit_buf, bits, tab->len);
	macroblock_modes = tab->modes;

	if ((! frame_pred_frame_dct) && (picture_structure == FRAME_PICTURE)) {
	    macroblock_modes |= UBITS (bit_buf, 1) * DCT_TYPE_INTERLACED;
	    DUMPBITS (bit_buf, bits, 1);
	}

	return macroblock_modes;

	tab = MB_P + UBITS (bit_buf, 5);
	DUMPBITS (bit_buf, bits, tab->len);
	macroblock_modes = tab->modes;

	if (picture_structure != FRAME_PICTURE) {
	    if (macroblock_modes & MACROBLOCK_MOTION_FORWARD) {
		macroblock_modes |= UBITS (bit_buf, 2) * MOTION_TYPE_BASE;
		DUMPBITS (bit_buf, bits, 2);
	    }
	    return macroblock_modes;
	} else if (frame_pred_frame_dct) {
	    if (macroblock_modes & MACROBLOCK_MOTION_FORWARD)
		macroblock_modes |= MC_FRAME;
	    return macroblock_modes;
	} else {
	    if (macroblock_modes & MACROBLOCK_MOTION_FORWARD) {

	tab = MB_B + UBITS (bit_buf, 6);
	DUMPBITS (bit_buf, bits, tab->len);
	macroblock_modes = tab->modes;

	if (picture_structure != FRAME_PICTURE) {
	    if (! (macroblock_modes & MACROBLOCK_INTRA)) {
		macroblock_modes |= UBITS (bit_buf, 2) * MOTION_TYPE_BASE;
		DUMPBITS (bit_buf, bits, 2);
	    }
	    return macroblock_modes;
	} else if (frame_pred_frame_dct) {
	    //if (! (macroblock_modes & MACROBLOCK_INTRA))
	    macroblock_modes |= MC_FRAME;
	    return macroblock_modes;
	} else {
	    if (macroblock_modes & MACROBLOCK_INTRA)
		goto intra;

#undef bit_buf
#undef bits
#undef bit_ptr
}

static inline int get_quantizer_scale (slice_t * slice, int q_scale_type)
{
#define bit_buf (slice->bitstream_buf)
#define bits (slice->bitstream_bits)
#define bit_ptr (slice->bitstream_ptr)

    int quantizer_scale_code;

    quantizer_scale_code = UBITS (bit_buf, 5);
    DUMPBITS (bit_buf, bits, 5);

    if (q_scale_type)
	return non_linear_quantizer_scale [quantizer_scale_code];
    else
	return quantizer_scale_code << 1;
#undef bit_buf
#undef bits
#undef bit_ptr
}

static inline int get_motion_delta (slice_t * slice, int f_code)
{
#define bit_buf (slice->bitstream_buf)
#define bits (slice->bitstream_bits)
#define bit_ptr (slice->bitstream_ptr)

    int delta;
    int sign;
    MVtab * tab;

#else
    return (vector << (27 - f_code)) >> (27 - f_code);
#endif
}

static inline int get_dmv (slice_t * slice)
{
#define bit_buf (slice->bitstream_buf)
#define bits (slice->bitstream_bits)
#define bit_ptr (slice->bitstream_ptr)

    DMVtab * tab;

    tab = DMV_2 + UBITS (bit_buf, 2);
    DUMPBITS (bit_buf, bits, tab->len);

#undef bit_buf
#undef bits
#undef bit_ptr
}

static inline int get_coded_block_pattern (slice_t * slice)
{
#define bit_buf (slice->bitstream_buf)
#define bits (slice->bitstream_bits)
#define bit_ptr (slice->bitstream_ptr)

    CBPtab * tab;

    NEEDBITS (bit_buf, bits, bit_ptr);

#undef bit_buf
#undef bits
#undef bit_ptr
}

static inline int get_luma_dc_dct_diff (slice_t * slice)
{
#define bit_buf (slice->bitstream_buf)
#define bits (slice->bitstream_bits)
#define bit_ptr (slice->bitstream_ptr)
    DCtab * tab;
    int size;
    int dc_diff;

    if (bit_buf < 0xf8000000) {

#undef bit_buf
#undef bits
#undef bit_ptr
}

static inline int get_chroma_dc_dct_diff (slice_t * slice)
{
#define bit_buf (slice->bitstream_buf)
#define bits (slice->bitstream_bits)
#define bit_ptr (slice->bitstream_ptr)
    DCtab * tab;
    int size;
    int dc_diff;

    if (bit_buf < 0xf8000000) {

#undef bit_buf
#undef bits
#undef bit_ptr
}

#define SATURATE(val)		\
do {				\
    if (val > 2047)		\
	val = 2047;		\
    else if (val < -2048)	\
	val = -2048;		\
} while (0)

static void get_intra_block_B14 (picture_t * picture, slice_t * slice,
				 int16_t * dest)
{
    int i;
    int j;
    int val;
    uint8_t * scan = picture->scan;
    uint8_t * quant_matrix = picture->intra_quantizer_matrix;
    int quantizer_scale = slice->quantizer_scale;
    int mismatch;
    DCTtab * tab;
    uint32_t bit_buf;
    int bits;
    uint8_t * bit_ptr;

    i = 0;
    mismatch = ~dest[0];

    bit_buf = slice->bitstream_buf;
    bits = slice->bitstream_bits;
    bit_ptr = slice->bitstream_ptr;

    NEEDBITS (bit_buf, bits, bit_ptr);

    while (1) {
	if (bit_buf >= 0x28000000) {

	    tab = DCT_B14AC_5 - 5 + UBITS (bit_buf, 5);

	    i += tab->run;
	    if (i >= 64)
		break;	// end of block

	normal_code:
	    j = scan[i];
	    bit_buf <<= tab->len;
	    bits += tab->len + 1;
	    val = (tab->level * quantizer_scale * quant_matrix[j]) >> 4;

	    // if (bitstream_get (1)) val = -val;
	    val = (val ^ SBITS (bit_buf, 1)) - SBITS (bit_buf, 1);

	    SATURATE (val);
	    dest[j] = val;
	    mismatch ^= val;

	    i += tab->run;
	    if (i < 64)
		goto normal_code;

	    // escape code

	    i += UBITS (bit_buf << 6, 6) - 64;
	    if (i >= 64)
		break;	// illegal, but check needed to avoid buffer overflow

	    j = scan[i];

	    DUMPBITS (bit_buf, bits, 12);
	    NEEDBITS (bit_buf, bits, bit_ptr);

	    GETWORD (bit_buf, bits + 16, bit_ptr);
	    i += tab->run;
	    if (i < 64)
		goto normal_code;
	}
	break;	// illegal, but check needed to avoid buffer overflow
    }
    dest[63] ^= mismatch & 1;
    DUMPBITS (bit_buf, bits, 2);	// dump end of block code
    slice->bitstream_buf = bit_buf;
    slice->bitstream_bits = bits;
    slice->bitstream_ptr = bit_ptr;
}

static void get_intra_block_B15 (picture_t * picture, slice_t * slice,
				 int16_t * dest)
{
    int i;
    int j;
    int val;
    uint8_t * scan = picture->scan;
    uint8_t * quant_matrix = picture->intra_quantizer_matrix;
    int quantizer_scale = slice->quantizer_scale;
    int mismatch;
    DCTtab * tab;
    uint32_t bit_buf;
    int bits;
    uint8_t * bit_ptr;

    i = 0;
    mismatch = ~dest[0];

    bit_buf = slice->bitstream_buf;
    bits = slice->bitstream_bits;
    bit_ptr = slice->bitstream_ptr;

    NEEDBITS (bit_buf, bits, bit_ptr);

    while (1) {
	if (bit_buf >= 0x04000000) {

		j = scan[i];
		bit_buf <<= tab->len;
		bits += tab->len + 1;
		val = (tab->level * quantizer_scale * quant_matrix[j]) >> 4;

		// if (bitstream_get (1)) val = -val;
		val = (val ^ SBITS (bit_buf, 1)) - SBITS (bit_buf, 1);

		SATURATE (val);
		dest[j] = val;
		mismatch ^= val;

		continue;

	    } else {

		// end of block. I commented out this code because if we
		// dont exit here we will still exit at the later test :)

		//if (i >= 128) break;	// end of block

		// escape code

		i += UBITS (bit_buf << 6, 6) - 64;
		if (i >= 64)
		    break;	// illegal, but check against buffer overflow

		j = scan[i];

		DUMPBITS (bit_buf, bits, 12);
		NEEDBITS (bit_buf, bits, bit_ptr);

	    GETWORD (bit_buf, bits + 16, bit_ptr);
	    i += tab->run;
	    if (i < 64)
		goto normal_code;
	}
	break;	// illegal, but check needed to avoid buffer overflow
    }
    dest[63] ^= mismatch & 1;
    DUMPBITS (bit_buf, bits, 4);	// dump end of block code
    slice->bitstream_buf = bit_buf;
    slice->bitstream_bits = bits;
    slice->bitstream_ptr = bit_ptr;
}

static void get_non_intra_block (picture_t * picture, slice_t * slice,
				 int16_t * dest)
{
    int i;
    int j;
    int val;
    uint8_t * scan = picture->scan;
    uint8_t * quant_matrix = picture->non_intra_quantizer_matrix;
    int quantizer_scale = slice->quantizer_scale;
    int mismatch;
    DCTtab * tab;
    uint32_t bit_buf;
    int bits;
    uint8_t * bit_ptr;

    i = -1;
    mismatch = 1;

    bit_buf = slice->bitstream_buf;
    bits = slice->bitstream_bits;
    bit_ptr = slice->bitstream_ptr;

    NEEDBITS (bit_buf, bits, bit_ptr);
    if (bit_buf >= 0x28000000) {
	tab = DCT_B14DC_5 - 5 + UBITS (bit_buf, 5);
	goto entry_1;

	    tab = DCT_B14AC_5 - 5 + UBITS (bit_buf, 5);

	entry_1:
	    i += tab->run;
	    if (i >= 64)
		break;	// end of block

	normal_code:
	    j = scan[i];
	    bit_buf <<= tab->len;
	    bits += tab->len + 1;
	    val = ((2*tab->level+1) * quantizer_scale * quant_matrix[j]) >> 5;

	    // if (bitstream_get (1)) val = -val;
	    val = (val ^ SBITS (bit_buf, 1)) - SBITS (bit_buf, 1);

	    SATURATE (val);
	    dest[j] = val;
	    mismatch ^= val;

	    i += tab->run;
	    if (i < 64)
		goto normal_code;

	    // escape code

	    i += UBITS (bit_buf << 6, 6) - 64;
	    if (i >= 64)
		break;	// illegal, but check needed to avoid buffer overflow

	    j = scan[i];

	    DUMPBITS (bit_buf, bits, 12);
	    NEEDBITS (bit_buf, bits, bit_ptr);

	    GETWORD (bit_buf, bits + 16, bit_ptr);
	    i += tab->run;
	    if (i < 64)
		goto normal_code;
	}
	break;	// illegal, but check needed to avoid buffer overflow
    }
    dest[63] ^= mismatch & 1;
    DUMPBITS (bit_buf, bits, 2);	// dump end of block code
    slice->bitstream_buf = bit_buf;
    slice->bitstream_bits = bits;
    slice->bitstream_ptr = bit_ptr;
}

static void get_mpeg1_intra_block (picture_t * picture, slice_t * slice,
				   int16_t * dest)
{
    int i;
    int j;
    int val;
    uint8_t * scan = picture->scan;
    uint8_t * quant_matrix = picture->intra_quantizer_matrix;
    int quantizer_scale = slice->quantizer_scale;
    DCTtab * tab;
    uint32_t bit_buf;
    int bits;
    uint8_t * bit_ptr;

    i = 0;

    bit_buf = slice->bitstream_buf;
    bits = slice->bitstream_bits;
    bit_ptr = slice->bitstream_ptr;

    NEEDBITS (bit_buf, bits, bit_ptr);

    while (1) {
	if (bit_buf >= 0x28000000) {

	    tab = DCT_B14AC_5 - 5 + UBITS (bit_buf, 5);

	    i += tab->run;
	    if (i >= 64)
		break;	// end of block

	normal_code:
	    j = scan[i];
	    bit_buf <<= tab->len;
	    bits += tab->len + 1;
	    val = (tab->level * quantizer_scale * quant_matrix[j]) >> 4;

	    // oddification
	    val = (val - 1) | 1;

	    // if (bitstream_get (1)) val = -val;
	    val = (val ^ SBITS (bit_buf, 1)) - SBITS (bit_buf, 1);

	    SATURATE (val);
	    dest[j] = val;

	    i += tab->run;
	    if (i < 64)
		goto normal_code;

	    // escape code

	    i += UBITS (bit_buf << 6, 6) - 64;
	    if (i >= 64)
		break;	// illegal, but check needed to avoid buffer overflow

	    j = scan[i];

	    DUMPBITS (bit_buf, bits, 12);
	    NEEDBITS (bit_buf, bits, bit_ptr);

		DUMPBITS (bit_buf, bits, 8);
		val = UBITS (bit_buf, 8) + 2 * val;
	    }
	    val = (val * quantizer_scale * quant_matrix[j]) / 16;

	    // oddification
	    val = (val + ~SBITS (val, 1)) | 1;

	    SATURATE (val);
	    dest[j] = val;

	    GETWORD (bit_buf, bits + 16, bit_ptr);
	    i += tab->run;
	    if (i < 64)
		goto normal_code;
	}
	break;	// illegal, but check needed to avoid buffer overflow
    }
    DUMPBITS (bit_buf, bits, 2);	// dump end of block code
    slice->bitstream_buf = bit_buf;
    slice->bitstream_bits = bits;
    slice->bitstream_ptr = bit_ptr;
}

static void get_mpeg1_non_intra_block (picture_t * picture, slice_t * slice,
				       int16_t * dest)
{
    int i;
    int j;
    int val;
    uint8_t * scan = picture->scan;
    uint8_t * quant_matrix = picture->non_intra_quantizer_matrix;
    int quantizer_scale = slice->quantizer_scale;
    DCTtab * tab;
    uint32_t bit_buf;
    int bits;
    uint8_t * bit_ptr;

    i = -1;

    bit_buf = slice->bitstream_buf;
    bits = slice->bitstream_bits;
    bit_ptr = slice->bitstream_ptr;

    NEEDBITS (bit_buf, bits, bit_ptr);
    if (bit_buf >= 0x28000000) {
	tab = DCT_B14DC_5 - 5 + UBITS (bit_buf, 5);
	goto entry_1;

	    tab = DCT_B14AC_5 - 5 + UBITS (bit_buf, 5);

	entry_1:
	    i += tab->run;
	    if (i >= 64)
		break;	// end of block

	normal_code:
	    j = scan[i];
	    bit_buf <<= tab->len;
	    bits += tab->len + 1;
	    val = ((2*tab->level+1) * quantizer_scale * quant_matrix[j]) >> 5;

	    // oddification
	    val = (val - 1) | 1;

	    // if (bitstream_get (1)) val = -val;
	    val = (val ^ SBITS (bit_buf, 1)) - SBITS (bit_buf, 1);

	    SATURATE (val);
	    dest[j] = val;

	    i += tab->run;
	    if (i < 64)
		goto normal_code;

	    // escape code

	    i += UBITS (bit_buf << 6, 6) - 64;
	    if (i >= 64)
		break;	// illegal, but check needed to avoid buffer overflow

	    j = scan[i];

	    DUMPBITS (bit_buf, bits, 12);
	    NEEDBITS (bit_buf, bits, bit_ptr);

		val = UBITS (bit_buf, 8) + 2 * val;
	    }
	    val = 2 * (val + SBITS (val, 1)) + 1;
	    val = (val * quantizer_scale * quant_matrix[j]) / 32;

	    // oddification
	    val = (val + ~SBITS (val, 1)) | 1;

	    SATURATE (val);
	    dest[j] = val;

	    GETWORD (bit_buf, bits + 16, bit_ptr);
	    i += tab->run;
	    if (i < 64)
		goto normal_code;
	}
	break;	// illegal, but check needed to avoid buffer overflow
    }
    DUMPBITS (bit_buf, bits, 2);	// dump end of block code
    slice->bitstream_buf = bit_buf;
    slice->bitstream_bits = bits;
    slice->bitstream_ptr = bit_ptr;
}

static inline int get_macroblock_address_increment (slice_t * slice)
{
#define bit_buf (slice->bitstream_buf)
#define bits (slice->bitstream_bits)
#define bit_ptr (slice->bitstream_ptr)

    MBAtab * tab;
    int mba;

    mba = 0;

	} else if (bit_buf >= 0x03000000) {
	    tab = MBA_11 - 24 + UBITS (bit_buf, 11);
	    DUMPBITS (bit_buf, bits, tab->len);
	    return mba + tab->mba;
	} else switch (UBITS (bit_buf, 11)) {
	case 8:		// macroblock_escape
	    mba += 33;
	    // no break here on purpose
	case 15:	// macroblock_stuffing (MPEG1 only)
	    DUMPBITS (bit_buf, bits, 11);
	    NEEDBITS (bit_buf, bits, bit_ptr);
	    break;
	default:	// end of slice, or error
	    return 0;
	}
    }

#undef bit_buf
#undef bits
#undef bit_ptr
}

static inline void slice_intra_DCT (picture_t * picture, slice_t * slice,
				    int cc, uint8_t * dest, int stride)
{
#define bit_buf (slice->bitstream_buf)
#define bits (slice->bitstream_bits)  
#define bit_ptr (slice->bitstream_ptr)
    NEEDBITS (bit_buf, bits, bit_ptr);
    //Get the intra DC coefficient and inverse quantize it
    if (cc == 0)
	slice->dc_dct_pred[0] += get_luma_dc_dct_diff (slice);
    else
	slice->dc_dct_pred[cc] += get_chroma_dc_dct_diff (slice);
    DCTblock[0] = slice->dc_dct_pred[cc] << (3 - picture->intra_dc_precision);

    if (picture->mpeg1) {
	if (picture->picture_coding_type != D_TYPE)
	    get_mpeg1_intra_block (picture, slice, DCTblock);
    } else if (picture->intra_vlc_format)
	get_intra_block_B15 (picture, slice, DCTblock);
    else
	get_intra_block_B14 (picture, slice, DCTblock);
    idct_block_copy (DCTblock, dest, stride);
    memset (DCTblock, 0, sizeof (DCTblock));
#undef bit_buf
#undef bits
#undef bit_ptr
}

static inline void slice_non_intra_DCT (picture_t * picture, slice_t * slice,
					uint8_t * dest, int stride)
{
    if (picture->mpeg1)
	get_mpeg1_non_intra_block (picture, slice, DCTblock);
    else
	get_non_intra_block (picture, slice, DCTblock);
    idct_block_add (DCTblock, dest, stride);
    memset (DCTblock, 0, sizeof (DCTblock));
}

static inline void motion_block (void (** table) (uint8_t *, uint8_t *,
						  int32_t, int32_t), 
				 int x_pred, int y_pred,

    table[4+xy_half] (dest[1] + dest_offset, src1, stride, height);
    table[4+xy_half] (dest[2] + dest_offset, src2, stride, height);
}


static void motion_mp1 (slice_t * slice, motion_t * motion,
			uint8_t * dest[3], int offset, int width,
			void (** table) (uint8_t *, uint8_t *, int, int))
{
#define bit_buf (slice->bitstream_buf)
#define bits (slice->bitstream_bits)
#define bit_ptr (slice->bitstream_ptr)
    int motion_x, motion_y;

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_x = motion->pmv[0][0] + get_motion_delta (slice, motion->f_code[0]);
    motion_x = bound_motion_vector (motion_x, motion->f_code[0]);
    motion->pmv[0][0] = motion_x;

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_y = motion->pmv[0][1] + get_motion_delta (slice, motion->f_code[0]);
    motion_y = bound_motion_vector (motion_y, motion->f_code[0]);
    motion->pmv[0][1] = motion_y;

    if (motion->f_code[1]) {
	motion_x <<= 1;
	motion_y <<= 1;
    }

    motion_block (table, motion_x, motion_y, dest, offset,
		  motion->ref[0], offset, width, 16, 0);
#undef bit_buf
#undef bits
#undef bit_ptr
}

static void motion_mp1_reuse (slice_t * slice, motion_t * motion,
			      uint8_t * dest[3], int offset, int width,
			      void (** table) (uint8_t *, uint8_t *, int, int))
{
    int motion_x, motion_y;

    motion_x = motion->pmv[0][0];

	motion_x <<= 1;
	motion_y <<= 1;
    }

    motion_block (table, motion_x, motion_y, dest, offset,
		  motion->ref[0], offset, width, 16, 0);
}

static void motion_fr_frame (slice_t * slice, motion_t * motion,
			     uint8_t * dest[3], int offset, int width,
			     void (** table) (uint8_t *, uint8_t *, int, int))
{
#define bit_buf (slice->bitstream_buf)
#define bits (slice->bitstream_bits)
#define bit_ptr (slice->bitstream_ptr)
    int motion_x, motion_y;

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_x = motion->pmv[0][0] + get_motion_delta (slice, motion->f_code[0]);
    motion_x = bound_motion_vector (motion_x, motion->f_code[0]);
    motion->pmv[1][0] = motion->pmv[0][0] = motion_x;

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_y = motion->pmv[0][1] + get_motion_delta (slice, motion->f_code[1]);
    motion_y = bound_motion_vector (motion_y, motion->f_code[1]);
    motion->pmv[1][1] = motion->pmv[0][1] = motion_y;

    motion_block (table, motion_x, motion_y, dest, offset,
		  motion->ref[0], offset, width, 16, 0);
#undef bit_buf
#undef bits
#undef bit_ptr
}

static void motion_fr_field (slice_t * slice, motion_t * motion,
			     uint8_t * dest[3], int offset, int width,
			     void (** table) (uint8_t *, uint8_t *, int, int))
{
#define bit_buf (slice->bitstream_buf)
#define bits (slice->bitstream_bits)
#define bit_ptr (slice->bitstream_ptr)
    int motion_x, motion_y;
    int field_select;

    NEEDBITS (bit_buf, bits, bit_ptr);
    field_select = SBITS (bit_buf, 1);
    DUMPBITS (bit_buf, bits, 1);

    motion_x = motion->pmv[0][0] + get_motion_delta (slice, motion->f_code[0]);
    motion_x = bound_motion_vector (motion_x, motion->f_code[0]);
    motion->pmv[0][0] = motion_x;

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_y = (motion->pmv[0][1] >> 1) + get_motion_delta (slice,
							    motion->f_code[1]);
    //motion_y = bound_motion_vector (motion_y, motion->f_code[1]);
    motion->pmv[0][1] = motion_y << 1;

    motion_block (table, motion_x, motion_y, dest, offset,
		  motion->ref[0], offset + (field_select & width),
		  width * 2, 8, 0);

    NEEDBITS (bit_buf, bits, bit_ptr);
    field_select = SBITS (bit_buf, 1);
    DUMPBITS (bit_buf, bits, 1);

    motion_x = motion->pmv[1][0] + get_motion_delta (slice, motion->f_code[0]);
    motion_x = bound_motion_vector (motion_x, motion->f_code[0]);
    motion->pmv[1][0] = motion_x;

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_y = (motion->pmv[1][1] >> 1) + get_motion_delta (slice,
							    motion->f_code[1]);
    //motion_y = bound_motion_vector (motion_y, motion->f_code[1]);
    motion->pmv[1][1] = motion_y << 1;

    motion_block (table, motion_x, motion_y, dest, offset + width,
		  motion->ref[0], offset + (field_select & width),
		  width * 2, 8, 0);
#undef bit_buf
#undef bits
#undef bit_ptr
}

static int motion_dmv_top_field_first;
static void motion_fr_dmv (slice_t * slice, motion_t * motion,
			   uint8_t * dest[3], int offset, int width,
			   void (** table) (uint8_t *, uint8_t *, int, int))
{
#define bit_buf (slice->bitstream_buf)
#define bits (slice->bitstream_bits)
#define bit_ptr (slice->bitstream_ptr)
    int motion_x, motion_y;
    int dmv_x, dmv_y;
    int m;
    int other_x, other_y;

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_x = motion->pmv[0][0] + get_motion_delta (slice, motion->f_code[0]);
    motion_x = bound_motion_vector (motion_x, motion->f_code[0]);
    motion->pmv[1][0] = motion->pmv[0][0] = motion_x;

    NEEDBITS (bit_buf, bits, bit_ptr);
    dmv_x = get_dmv (slice);

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_y = (motion->pmv[0][1] >> 1) + get_motion_delta (slice,
							    motion->f_code[1]);
    //motion_y = bound_motion_vector (motion_y, motion->f_code[1]);
    motion->pmv[1][1] = motion->pmv[0][1] = motion_y << 1;

    NEEDBITS (bit_buf, bits, bit_ptr);
    dmv_y = get_dmv (slice);

    motion_block (mc_functions.put, motion_x, motion_y, dest, offset,
		  motion->ref[0], offset, width * 2, 8, 0);

    m = motion_dmv_top_field_first ? 1 : 3;
    other_x = ((motion_x * m + (motion_x > 0)) >> 1) + dmv_x;
    other_y = ((motion_y * m + (motion_y > 0)) >> 1) + dmv_y - 1;
    motion_block (mc_functions.avg, other_x, other_y, dest, offset,
		  motion->ref[0], offset + width, width * 2, 8, 0);

    motion_block (mc_functions.put, motion_x, motion_y, dest, offset + width,
		  motion->ref[0], offset + width, width * 2, 8, 0);

    m = motion_dmv_top_field_first ? 3 : 1;
    other_x = ((motion_x * m + (motion_x > 0)) >> 1) + dmv_x;
    other_y = ((motion_y * m + (motion_y > 0)) >> 1) + dmv_y + 1;
    motion_block (mc_functions.avg, other_x, other_y, dest, offset + width,
		  motion->ref[0], offset, width * 2, 8, 0);
#undef bit_buf
#undef bits
#undef bit_ptr
}

// like motion_frame, but reuse previous motion vectors
static void motion_fr_reuse (slice_t * slice, motion_t * motion,
			     uint8_t * dest[3], int offset, int width,
			     void (** table) (uint8_t *, uint8_t *, int, int))
{
    motion_block (table, motion->pmv[0][0], motion->pmv[0][1], dest, offset,
		  motion->ref[0], offset, width, 16, 0);
}

// like motion_frame, but use null motion vectors
static void motion_fr_zero (slice_t * slice, motion_t * motion,
			    uint8_t * dest[3], int offset, int width,
			    void (** table) (uint8_t *, uint8_t *, int, int))
{
    motion_block (table, 0, 0, dest, offset,
		  motion->ref[0], offset, width, 16, 0);
}

// like motion_frame, but parsing without actual motion compensation
static void motion_fr_conceal (slice_t * slice, motion_t * motion)
{
#define bit_buf (slice->bitstream_buf)
#define bits (slice->bitstream_bits)
#define bit_ptr (slice->bitstream_ptr)
    int tmp;

    NEEDBITS (bit_buf, bits, bit_ptr);
    tmp = motion->pmv[0][0] + get_motion_delta (slice, motion->f_code[0]);
    tmp = bound_motion_vector (tmp, motion->f_code[0]);
    motion->pmv[1][0] = motion->pmv[0][0] = tmp;

    NEEDBITS (bit_buf, bits, bit_ptr);
    tmp = motion->pmv[0][1] + get_motion_delta (slice, motion->f_code[1]);
    tmp = bound_motion_vector (tmp, motion->f_code[1]);
    motion->pmv[1][1] = motion->pmv[0][1] = tmp;

    DUMPBITS (bit_buf, bits, 1); // remove marker_bit
#undef bit_buf
#undef bits
#undef bit_ptr
}

static void motion_fi_field (slice_t * slice, motion_t * motion,
			     uint8_t * dest[3], int offset, int width,
			     void (** table) (uint8_t *, uint8_t *, int, int))
{
#define bit_buf (slice->bitstream_buf)
#define bits (slice->bitstream_bits)
#define bit_ptr (slice->bitstream_ptr)
    int motion_x, motion_y;
    int field_select;

    NEEDBITS (bit_buf, bits, bit_ptr);
    field_select = UBITS (bit_buf, 1);
    DUMPBITS (bit_buf, bits, 1);

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_x = motion->pmv[0][0] + get_motion_delta (slice, motion->f_code[0]);
    motion_x = bound_motion_vector (motion_x, motion->f_code[0]);
    motion->pmv[1][0] = motion->pmv[0][0] = motion_x;

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_y = motion->pmv[0][1] + get_motion_delta (slice, motion->f_code[1]);
    motion_y = bound_motion_vector (motion_y, motion->f_code[1]);
    motion->pmv[1][1] = motion->pmv[0][1] = motion_y;

    motion_block (table, motion_x, motion_y, dest, offset,
		  motion->ref[field_select], offset, width, 16, 0);
#undef bit_buf
#undef bits
#undef bit_ptr
}

static void motion_fi_16x8 (slice_t * slice, motion_t * motion,
			    uint8_t * dest[3], int offset, int width,
			    void (** table) (uint8_t *, uint8_t *, int, int))
{
#define bit_buf (slice->bitstream_buf)
#define bits (slice->bitstream_bits)
#define bit_ptr (slice->bitstream_ptr)
    int motion_x, motion_y;
    int field_select;

    NEEDBITS (bit_buf, bits, bit_ptr);
    field_select = UBITS (bit_buf, 1);
    DUMPBITS (bit_buf, bits, 1);

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_x = motion->pmv[0][0] + get_motion_delta (slice, motion->f_code[0]);
    motion_x = bound_motion_vector (motion_x, motion->f_code[0]);
    motion->pmv[0][0] = motion_x;

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_y = motion->pmv[0][1] + get_motion_delta (slice, motion->f_code[1]);
    motion_y = bound_motion_vector (motion_y, motion->f_code[1]);
    motion->pmv[0][1] = motion_y;

    motion_block (table, motion_x, motion_y, dest, offset,
		  motion->ref[field_select], offset, width, 8, 0);

    NEEDBITS (bit_buf, bits, bit_ptr);
    field_select = UBITS (bit_buf, 1);
    DUMPBITS (bit_buf, bits, 1);

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_x = motion->pmv[1][0] + get_motion_delta (slice, motion->f_code[0]);
    motion_x = bound_motion_vector (motion_x, motion->f_code[0]);
    motion->pmv[1][0] = motion_x;

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_y = motion->pmv[1][1] + get_motion_delta (slice, motion->f_code[1]);
    motion_y = bound_motion_vector (motion_y, motion->f_code[1]);
    motion->pmv[1][1] = motion_y;

    motion_block (table, motion_x, motion_y, dest, offset,
		  motion->ref[field_select], offset, width, 8, 1);
#undef bit_buf
#undef bits
#undef bit_ptr
}

static int current_field = 0;
static void motion_fi_dmv (slice_t * slice, motion_t * motion,
			   uint8_t * dest[3], int offset, int width,
			   void (** table) (uint8_t *, uint8_t *, int, int))
{
#define bit_buf (slice->bitstream_buf)
#define bits (slice->bitstream_bits)
#define bit_ptr (slice->bitstream_ptr)
    int motion_x, motion_y;
    int dmv_x, dmv_y;

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_x = motion->pmv[0][0] + get_motion_delta (slice, motion->f_code[0]);
    motion_x = bound_motion_vector (motion_x, motion->f_code[0]);
    motion->pmv[1][0] = motion->pmv[0][0] = motion_x;

    NEEDBITS (bit_buf, bits, bit_ptr);
    dmv_x = get_dmv (slice);

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_y = motion->pmv[0][1] + get_motion_delta (slice, motion->f_code[1]);
    motion_y = bound_motion_vector (motion_y, motion->f_code[1]);
    motion->pmv[1][1] = motion->pmv[0][1] = motion_y;

    NEEDBITS (bit_buf, bits, bit_ptr);
    dmv_y = get_dmv (slice);

    motion_block (mc_functions.put, motion_x, motion_y, dest, offset,
		  motion->ref[current_field], offset, width, 16, 0);

    motion_x = ((motion_x + (motion_x > 0)) >> 1) + dmv_x;
    motion_y = ((motion_y + (motion_y > 0)) >> 1) + dmv_y +
	2 * current_field - 1;
    motion_block (mc_functions.avg, motion_x, motion_y, dest, offset,
		  motion->ref[!current_field], offset, width, 16, 0);
#undef bit_buf
#undef bits
#undef bit_ptr
}

static void motion_fi_reuse (slice_t * slice, motion_t * motion,
			     uint8_t * dest[3], int offset, int width,
			     void (** table) (uint8_t *, uint8_t *, int, int))
{
    motion_block (table, motion->pmv[0][0], motion->pmv[0][1], dest, offset,
		  motion->ref[current_field], offset, width, 16, 0);
}

static void motion_fi_zero (slice_t * slice, motion_t * motion,
			    uint8_t * dest[3], int offset, int width,
			    void (** table) (uint8_t *, uint8_t *, int, int))
{
    motion_block (table, 0, 0, dest, offset,
		  motion->ref[current_field], offset, width, 16, 0);
}

static void motion_fi_conceal (slice_t * slice, motion_t * motion)
{
#define bit_buf (slice->bitstream_buf)
#define bits (slice->bitstream_bits)
#define bit_ptr (slice->bitstream_ptr)
    int tmp;

    NEEDBITS (bit_buf, bits, bit_ptr);
    DUMPBITS (bit_buf, bits, 1); // remove field_select

    NEEDBITS (bit_buf, bits, bit_ptr);
    tmp = motion->pmv[0][0] + get_motion_delta (slice, motion->f_code[0]);
    tmp = bound_motion_vector (tmp, motion->f_code[0]);
    motion->pmv[1][0] = motion->pmv[0][0] = tmp;

    NEEDBITS (bit_buf, bits, bit_ptr);
    tmp = motion->pmv[0][1] + get_motion_delta (slice, motion->f_code[1]);
    tmp = bound_motion_vector (tmp, motion->f_code[1]);
    motion->pmv[1][1] = motion->pmv[0][1] = tmp;

    DUMPBITS (bit_buf, bits, 1); // remove marker_bit
#undef bit_buf
#undef bits
#undef bit_ptr
}

#define MOTION(routine,direction,slice,dest,offset,stride)		\
do {									\
    if ((direction) & MACROBLOCK_MOTION_FORWARD)			\
	routine (&slice, &((slice).f_motion), dest, offset, stride,	\
		 mc_functions.put);					\
    if ((direction) & MACROBLOCK_MOTION_BACKWARD)			\
	routine (&slice, &((slice).b_motion), dest, offset, stride,	\
		 ((direction) & MACROBLOCK_MOTION_FORWARD ?		\
		  mc_functions.avg : mc_functions.put));		\
} while (0)

#define CHECK_DISPLAY					\
do {							\
    if (offset == width) {				\
	slice.f_motion.ref[0][0] += 16 * offset;	\
	slice.f_motion.ref[0][1] += 4 * offset;		\
	slice.f_motion.ref[0][2] += 4 * offset;		\
	slice.b_motion.ref[0][0] += 16 * offset;	\
	slice.b_motion.ref[0][1] += 4 * offset;		\
	slice.b_motion.ref[0][2] += 4 * offset;		\
	dest[0] += 16 * offset;				\
	dest[1] += 4 * offset;				\
	dest[2] += 4 * offset;				\
	offset = 0; ++ypos;				\
    }							\
} while (0)

int slice_process (picture_t * picture, uint8_t code, uint8_t * buffer)
{
#define bit_buf (slice.bitstream_buf)
#define bits (slice.bitstream_bits)
#define bit_ptr (slice.bitstream_ptr)
    slice_t slice;
    int macroblock_modes;
    int width;
    int ypos=code-1;
    uint8_t * dest[3];
    int offset;
    uint8_t ** forward_ref[2];

    width = picture->coded_picture_width;
    offset = ypos * width * 4;

    forward_ref[0] = picture->forward_reference_frame;
    if (picture->picture_structure != FRAME_PICTURE) {
	offset <<= 1;
	forward_ref[1] = picture->forward_reference_frame;
	current_field = (picture->picture_structure == BOTTOM_FIELD);
	if ((picture->second_field) &&
	    (picture->picture_coding_type != B_TYPE))
	    forward_ref[picture->picture_structure == TOP_FIELD] =
		picture->current_frame;
	slice.f_motion.ref[1][0] = forward_ref[1][0] + offset * 4 + width;
	slice.f_motion.ref[1][1] = forward_ref[1][1] + offset + (width >> 1);
	slice.f_motion.ref[1][2] = forward_ref[1][2] + offset + (width >> 1);
	slice.b_motion.ref[1][0] =
	    picture->backward_reference_frame[0] + offset * 4 + width;
	slice.b_motion.ref[1][1] =
	    picture->backward_reference_frame[1] + offset + (width >> 1);
	slice.b_motion.ref[1][2] =
	    picture->backward_reference_frame[2] + offset + (width >> 1);
    }

    slice.f_motion.ref[0][0] = forward_ref[0][0] + offset * 4;
    slice.f_motion.ref[0][1] = forward_ref[0][1] + offset;
    slice.f_motion.ref[0][2] = forward_ref[0][2] + offset;
    slice.f_motion.f_code[0] = picture->f_code[0][0];
    slice.f_motion.f_code[1] = picture->f_code[0][1];
    slice.f_motion.pmv[0][0] = slice.f_motion.pmv[0][1] = 0;
    slice.f_motion.pmv[1][0] = slice.f_motion.pmv[1][1] = 0;
    slice.b_motion.ref[0][0] =
	picture->backward_reference_frame[0] + offset * 4;
    slice.b_motion.ref[0][1] =
	picture->backward_reference_frame[1] + offset;
    slice.b_motion.ref[0][2] =
	picture->backward_reference_frame[2] + offset;
    slice.b_motion.f_code[0] = picture->f_code[1][0];
    slice.b_motion.f_code[1] = picture->f_code[1][1];
    slice.b_motion.pmv[0][0] = slice.b_motion.pmv[0][1] = 0;
    slice.b_motion.pmv[1][0] = slice.b_motion.pmv[1][1] = 0;

    if ((! HACK_MODE) && (!picture->mpeg1) &&
	(picture->picture_coding_type == B_TYPE))
	offset = 0;

    dest[0] = picture->current_frame[0] + offset * 4;
    dest[1] = picture->current_frame[1] + offset;
    dest[2] = picture->current_frame[2] + offset;

    switch (picture->picture_structure) {
    case BOTTOM_FIELD:
	dest[0] += width;
	dest[1] += width >> 1;
	dest[2] += width >> 1;
	// follow thru
    case TOP_FIELD:
	width <<= 1;
    }

    //reset intra dc predictor
    slice.dc_dct_pred[0]=slice.dc_dct_pred[1]=slice.dc_dct_pred[2]= 
	1<< (picture->intra_dc_precision + 7) ;

    bitstream_init (&slice, buffer);

    slice.quantizer_scale = get_quantizer_scale (&slice,
						 picture->q_scale_type);

    //Ignore intra_slice and all the extra data
    while (bit_buf & 0x80000000) {
	DUMPBITS (bit_buf, bits, 9);
	NEEDBITS (bit_buf, bits, bit_ptr);
    }
    DUMPBITS (bit_buf, bits, 1);

    NEEDBITS (bit_buf, bits, bit_ptr);
    offset = get_macroblock_address_increment (&slice) << 4;

    while (1) {
	NEEDBITS (bit_buf, bits, bit_ptr);

	macroblock_modes =
	    get_macroblock_modes (&slice, picture->picture_structure,
				  picture->picture_coding_type,
				  picture->frame_pred_frame_dct);

	// maybe integrate MACROBLOCK_QUANT test into get_macroblock_modes ?
	if (macroblock_modes & MACROBLOCK_QUANT)
	    slice.quantizer_scale =
		get_quantizer_scale (&slice, picture->q_scale_type);

	if (macroblock_modes & MACROBLOCK_INTRA) {

	    int DCT_offset, DCT_stride;

	    if (picture->concealment_motion_vectors) {
		if (picture->picture_structure == FRAME_PICTURE)
		    motion_fr_conceal (&slice, &slice.f_motion);
		else
		    motion_fi_conceal (&slice, &slice.f_motion);
	    } else {
		slice.f_motion.pmv[0][0] = slice.f_motion.pmv[0][1] = 0;
		slice.f_motion.pmv[1][0] = slice.f_motion.pmv[1][1] = 0;
		slice.b_motion.pmv[0][0] = slice.b_motion.pmv[0][1] = 0;
		slice.b_motion.pmv[1][0] = slice.b_motion.pmv[1][1] = 0;
	    }

	    if (macroblock_modes & DCT_TYPE_INTERLACED) {
		DCT_offset = width;
		DCT_stride = width * 2;
	    } else {
		DCT_offset = width * 8;
		DCT_stride = width;
	    }

	    // Decode lum blocks
	    slice_intra_DCT (picture, &slice, 0,
			     dest[0] + offset, DCT_stride);
	    slice_intra_DCT (picture, &slice, 0,
			     dest[0] + offset + 8, DCT_stride);
	    slice_intra_DCT (picture, &slice, 0,
			     dest[0] + offset + DCT_offset, DCT_stride);
	    slice_intra_DCT (picture, &slice, 0,
			     dest[0] + offset + DCT_offset + 8, DCT_stride);

	    // Decode chroma blocks
	    slice_intra_DCT (picture, &slice, 1,
			     dest[1] + (offset>>1), width>>1);
	    slice_intra_DCT (picture, &slice, 2,
			     dest[2] + (offset>>1), width>>1);

	    if (picture->picture_coding_type == D_TYPE) {
		NEEDBITS (bit_buf, bits, bit_ptr);
		DUMPBITS (bit_buf, bits, 1);
	    }
	} else {

	    if (picture->mpeg1) {
		if ((macroblock_modes & MOTION_TYPE_MASK) == MC_FRAME)
		    MOTION (motion_mp1, macroblock_modes, slice,
			    dest, offset,width);
		else {
		    // non-intra mb without forward mv in a P picture
		    slice.f_motion.pmv[0][0] = slice.f_motion.pmv[0][1] = 0;
		    slice.f_motion.pmv[1][0] = slice.f_motion.pmv[1][1] = 0;

		    MOTION (motion_fr_zero, MACROBLOCK_MOTION_FORWARD, slice,
			    dest, offset, width);
		}
	    } else if (picture->picture_structure == FRAME_PICTURE)
		switch (macroblock_modes & MOTION_TYPE_MASK) {
		case MC_FRAME:
		    MOTION (motion_fr_frame, macroblock_modes, slice,
			    dest, offset, width);
		    break;

		case MC_FIELD:
		    MOTION (motion_fr_field, macroblock_modes, slice,
			    dest, offset, width);
		    break;

		case MC_DMV:
		    motion_dmv_top_field_first = picture->top_field_first;
		    MOTION (motion_fr_dmv, MACROBLOCK_MOTION_FORWARD, slice,
			    dest, offset, width);
		    break;

		case 0:
		    // non-intra mb without forward mv in a P picture
		    slice.f_motion.pmv[0][0] = slice.f_motion.pmv[0][1] = 0;
		    slice.f_motion.pmv[1][0] = slice.f_motion.pmv[1][1] = 0;

		    MOTION (motion_fr_zero, MACROBLOCK_MOTION_FORWARD, slice,
			    dest, offset, width);
		    break;
		}
	    else
		switch (macroblock_modes & MOTION_TYPE_MASK) {
		case MC_FIELD:
		    MOTION (motion_fi_field, macroblock_modes, slice,
			    dest, offset, width);
		    break;

		case MC_16X8:
		    MOTION (motion_fi_16x8, macroblock_modes, slice,
			    dest, offset, width);
		    break;

		case MC_DMV:
		    motion_dmv_top_field_first = picture->top_field_first;
		    MOTION (motion_fi_dmv, MACROBLOCK_MOTION_FORWARD, slice,
			    dest, offset, width);
		    break;

		case 0:
		    // non-intra mb without forward mv in a P picture
		    slice.f_motion.pmv[0][0] = slice.f_motion.pmv[0][1] = 0;
		    slice.f_motion.pmv[1][0] = slice.f_motion.pmv[1][1] = 0;

		    MOTION (motion_fi_zero, MACROBLOCK_MOTION_FORWARD, slice,
			    dest, offset, width);
		    break;
		}

	    //6.3.17.4 Coded block pattern
	    if (macroblock_modes & MACROBLOCK_PATTERN) {
		int coded_block_pattern;
		int DCT_offset, DCT_stride;

		if (macroblock_modes & DCT_TYPE_INTERLACED) {
		    DCT_offset = width;
		    DCT_stride = width * 2;
		} else {
		    DCT_offset = width * 8;
		    DCT_stride = width;
		}

		coded_block_pattern = get_coded_block_pattern (&slice);

		// Decode lum blocks

		if (coded_block_pattern & 0x20)
		    slice_non_intra_DCT (picture, &slice,
					 dest[0] + offset, DCT_stride);
		if (coded_block_pattern & 0x10)
		    slice_non_intra_DCT (picture, &slice,
					 dest[0] + offset + 8, DCT_stride);
		if (coded_block_pattern & 0x08)
		    slice_non_intra_DCT (picture, &slice,
					 dest[0] + offset + DCT_offset,
					 DCT_stride);
		if (coded_block_pattern & 0x04)
		    slice_non_intra_DCT (picture, &slice,
					 dest[0] + offset + DCT_offset + 8,
					 DCT_stride);

		// Decode chroma blocks

		if (coded_block_pattern & 0x2)
		    slice_non_intra_DCT (picture, &slice,
					 dest[1] + (offset>>1), width >> 1);
		if (coded_block_pattern & 0x1)
		    slice_non_intra_DCT (picture, &slice,
					 dest[2] + (offset>>1), width >> 1);
	    }

	    slice.dc_dct_pred[0]=slice.dc_dct_pred[1]=slice.dc_dct_pred[2]=
		1 << (picture->intra_dc_precision + 7);
	}

//        printf("[%d]",slice.quantizer_scale);
//        printf("[%d,%d]",offset>>4,ypos);
	quant_store[ypos+1][(offset>>4)+1] = slice.quantizer_scale;

	offset += 16;
	CHECK_DISPLAY;

	NEEDBITS (bit_buf, bits, bit_ptr);

	if (bit_buf & 0x80000000) {
	    DUMPBITS (bit_buf, bits, 1);
	} else {
	    int mba_inc;

	    mba_inc = get_macroblock_address_increment (&slice);
	    if (!mba_inc)
		break;

	    //reset intra dc predictor on skipped block
	    slice.dc_dct_pred[0]=slice.dc_dct_pred[1]=slice.dc_dct_pred[2]=
		1<< (picture->intra_dc_precision + 7);

	    //handling of skipped mb's differs between P_TYPE and B_TYPE
	    //pictures
	    if (picture->picture_coding_type == P_TYPE) {
		slice.f_motion.pmv[0][0] = slice.f_motion.pmv[0][1] = 0;
		slice.f_motion.pmv[1][0] = slice.f_motion.pmv[1][1] = 0;

		do {
		    if (picture->picture_structure == FRAME_PICTURE)
			MOTION (motion_fr_zero, MACROBLOCK_MOTION_FORWARD,
				slice, dest, offset, width);
		    else
			MOTION (motion_fi_zero, MACROBLOCK_MOTION_FORWARD,
				slice, dest, offset, width);

	quant_store[ypos+1][(offset>>4)+1] = slice.quantizer_scale;
//        printf("[%d,%d]",offset>>4,ypos);
		    offset += 16;
		    CHECK_DISPLAY;
		} while (--mba_inc);
	    } else {
		do {
		    if (picture->mpeg1)
			MOTION (motion_mp1_reuse, macroblock_modes,
				slice, dest, offset, width);
		    else if (picture->picture_structure == FRAME_PICTURE)
			MOTION (motion_fr_reuse, macroblock_modes,
				slice, dest, offset, width);
		    else
			MOTION (motion_fi_reuse, macroblock_modes,
				slice, dest, offset, width);

	quant_store[ypos+1][(offset>>4)+1] = slice.quantizer_scale;
//        printf("[%d,%d]",offset>>4,ypos);
		    offset += 16;
		    CHECK_DISPLAY;
		} while (--mba_inc);
	    }
	}

36:846535ace7a2

/*
 * slice.c
 * Copyright (C) 1999-2001 Aaron Holtzman <[e-mail]>
 *
 * This file is part of mpeg2dec, a free MPEG-2 video stream decoder.
 *
 * mpeg2dec is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by

#include "config.h"

#include <string.h>
#include <inttypes.h>

#include "video_out.h"
#include "mpeg2_internal.h"
#include "attributes.h"

extern mc_functions_t mc_functions;
extern void (* idct_block_copy) (int16_t * block, uint8_t * dest, int stride);
extern void (* idct_block_add) (int16_t * block, uint8_t * dest, int stride);

#include "vlc.h"

static int non_linear_quantizer_scale [] = {
     0,  1,  2,  3,  4,  5,   6,   7,
     8, 10, 12, 14, 16, 18,  20,  22,
    24, 28, 32, 36, 40, 44,  48,  52,
    56, 64, 72, 80, 88, 96, 104, 112
};

static inline int get_macroblock_modes (picture_t * picture)
{
#define bit_buf (picture->bitstream_buf)
#define bits (picture->bitstream_bits)
#define bit_ptr (picture->bitstream_ptr)
    int macroblock_modes;
    MBtab * tab;

    switch (picture->picture_coding_type) {
    case I_TYPE:

	tab = MB_I + UBITS (bit_buf, 1);
	DUMPBITS (bit_buf, bits, tab->len);
	macroblock_modes = tab->modes;

	if ((! (picture->frame_pred_frame_dct)) &&
	    (picture->picture_structure == FRAME_PICTURE)) {
	    macroblock_modes |= UBITS (bit_buf, 1) * DCT_TYPE_INTERLACED;
	    DUMPBITS (bit_buf, bits, 1);
	}

	return macroblock_modes;

	tab = MB_P + UBITS (bit_buf, 5);
	DUMPBITS (bit_buf, bits, tab->len);
	macroblock_modes = tab->modes;

	if (picture->picture_structure != FRAME_PICTURE) {
	    if (macroblock_modes & MACROBLOCK_MOTION_FORWARD) {
		macroblock_modes |= UBITS (bit_buf, 2) * MOTION_TYPE_BASE;
		DUMPBITS (bit_buf, bits, 2);
	    }
	    return macroblock_modes;
	} else if (picture->frame_pred_frame_dct) {
	    if (macroblock_modes & MACROBLOCK_MOTION_FORWARD)
		macroblock_modes |= MC_FRAME;
	    return macroblock_modes;
	} else {
	    if (macroblock_modes & MACROBLOCK_MOTION_FORWARD) {

	tab = MB_B + UBITS (bit_buf, 6);
	DUMPBITS (bit_buf, bits, tab->len);
	macroblock_modes = tab->modes;

	if (picture->picture_structure != FRAME_PICTURE) {
	    if (! (macroblock_modes & MACROBLOCK_INTRA)) {
		macroblock_modes |= UBITS (bit_buf, 2) * MOTION_TYPE_BASE;
		DUMPBITS (bit_buf, bits, 2);
	    }
	    return macroblock_modes;
	} else if (picture->frame_pred_frame_dct) {
	    /* if (! (macroblock_modes & MACROBLOCK_INTRA)) */
	    macroblock_modes |= MC_FRAME;
	    return macroblock_modes;
	} else {
	    if (macroblock_modes & MACROBLOCK_INTRA)
		goto intra;

#undef bit_buf
#undef bits
#undef bit_ptr
}

static inline int get_quantizer_scale (picture_t * picture)
{
#define bit_buf (picture->bitstream_buf)
#define bits (picture->bitstream_bits)
#define bit_ptr (picture->bitstream_ptr)

    int quantizer_scale_code;

    quantizer_scale_code = UBITS (bit_buf, 5);
    DUMPBITS (bit_buf, bits, 5);

    if (picture->q_scale_type)
	return non_linear_quantizer_scale [quantizer_scale_code];
    else
	return quantizer_scale_code << 1;
#undef bit_buf
#undef bits
#undef bit_ptr
}

static inline int get_motion_delta (picture_t * picture, int f_code)
{
#define bit_buf (picture->bitstream_buf)
#define bits (picture->bitstream_bits)
#define bit_ptr (picture->bitstream_ptr)

    int delta;
    int sign;
    MVtab * tab;

#else
    return (vector << (27 - f_code)) >> (27 - f_code);
#endif
}

static inline int get_dmv (picture_t * picture)
{
#define bit_buf (picture->bitstream_buf)
#define bits (picture->bitstream_bits)
#define bit_ptr (picture->bitstream_ptr)

    DMVtab * tab;

    tab = DMV_2 + UBITS (bit_buf, 2);
    DUMPBITS (bit_buf, bits, tab->len);

#undef bit_buf
#undef bits
#undef bit_ptr
}

static inline int get_coded_block_pattern (picture_t * picture)
{
#define bit_buf (picture->bitstream_buf)
#define bits (picture->bitstream_bits)
#define bit_ptr (picture->bitstream_ptr)

    CBPtab * tab;

    NEEDBITS (bit_buf, bits, bit_ptr);

#undef bit_buf
#undef bits
#undef bit_ptr
}

static inline int get_luma_dc_dct_diff (picture_t * picture)
{
#define bit_buf (picture->bitstream_buf)
#define bits (picture->bitstream_bits)
#define bit_ptr (picture->bitstream_ptr)
    DCtab * tab;
    int size;
    int dc_diff;

    if (bit_buf < 0xf8000000) {

#undef bit_buf
#undef bits
#undef bit_ptr
}

static inline int get_chroma_dc_dct_diff (picture_t * picture)
{
#define bit_buf (picture->bitstream_buf)
#define bits (picture->bitstream_bits)
#define bit_ptr (picture->bitstream_ptr)
    DCtab * tab;
    int size;
    int dc_diff;

    if (bit_buf < 0xf8000000) {

#undef bit_buf
#undef bits
#undef bit_ptr
}

#define SATURATE(val)			\
do {					\
    if ((uint32_t)(val + 2048) > 4095)	\
	val = (val > 0) ? 2047 : -2048;	\
} while (0)

static void get_intra_block_B14 (picture_t * picture)
{
    int i;
    int j;
    int val;
    uint8_t * scan = picture->scan;
    uint8_t * quant_matrix = picture->intra_quantizer_matrix;
    int quantizer_scale = picture->quantizer_scale;
    int mismatch;
    DCTtab * tab;
    uint32_t bit_buf;
    int bits;
    uint8_t * bit_ptr;
    int16_t * dest;

    dest = picture->DCTblock;
    i = 0;
    mismatch = ~dest[0];

    bit_buf = picture->bitstream_buf;
    bits = picture->bitstream_bits;
    bit_ptr = picture->bitstream_ptr;

    NEEDBITS (bit_buf, bits, bit_ptr);

    while (1) {
	if (bit_buf >= 0x28000000) {

	    tab = DCT_B14AC_5 - 5 + UBITS (bit_buf, 5);

	    i += tab->run;
	    if (i >= 64)
		break;	/* end of block */

	normal_code:
	    j = scan[i];
	    bit_buf <<= tab->len;
	    bits += tab->len + 1;
	    val = (tab->level * quantizer_scale * quant_matrix[j]) >> 4;

	    /* if (bitstream_get (1)) val = -val; */
	    val = (val ^ SBITS (bit_buf, 1)) - SBITS (bit_buf, 1);

	    SATURATE (val);
	    dest[j] = val;
	    mismatch ^= val;

	    i += tab->run;
	    if (i < 64)
		goto normal_code;

	    /* escape code */

	    i += UBITS (bit_buf << 6, 6) - 64;
	    if (i >= 64)
		break;	/* illegal, check needed to avoid buffer overflow */

	    j = scan[i];

	    DUMPBITS (bit_buf, bits, 12);
	    NEEDBITS (bit_buf, bits, bit_ptr);

	    GETWORD (bit_buf, bits + 16, bit_ptr);
	    i += tab->run;
	    if (i < 64)
		goto normal_code;
	}
	break;	/* illegal, check needed to avoid buffer overflow */
    }
    dest[63] ^= mismatch & 1;
    DUMPBITS (bit_buf, bits, 2);	/* dump end of block code */
    picture->bitstream_buf = bit_buf;
    picture->bitstream_bits = bits;
    picture->bitstream_ptr = bit_ptr;
}

static void get_intra_block_B15 (picture_t * picture)
{
    int i;
    int j;
    int val;
    uint8_t * scan = picture->scan;
    uint8_t * quant_matrix = picture->intra_quantizer_matrix;
    int quantizer_scale = picture->quantizer_scale;
    int mismatch;
    DCTtab * tab;
    uint32_t bit_buf;
    int bits;
    uint8_t * bit_ptr;
    int16_t * dest;

    dest = picture->DCTblock;
    i = 0;
    mismatch = ~dest[0];

    bit_buf = picture->bitstream_buf;
    bits = picture->bitstream_bits;
    bit_ptr = picture->bitstream_ptr;

    NEEDBITS (bit_buf, bits, bit_ptr);

    while (1) {
	if (bit_buf >= 0x04000000) {

		j = scan[i];
		bit_buf <<= tab->len;
		bits += tab->len + 1;
		val = (tab->level * quantizer_scale * quant_matrix[j]) >> 4;

		/* if (bitstream_get (1)) val = -val; */
		val = (val ^ SBITS (bit_buf, 1)) - SBITS (bit_buf, 1);

		SATURATE (val);
		dest[j] = val;
		mismatch ^= val;

		continue;

	    } else {

		/* end of block. I commented out this code because if we */
		/* dont exit here we will still exit at the later test :) */

		/* if (i >= 128) break;	*/	/* end of block */

		/* escape code */

		i += UBITS (bit_buf << 6, 6) - 64;
		if (i >= 64)
		    break;	/* illegal, check against buffer overflow */

		j = scan[i];

		DUMPBITS (bit_buf, bits, 12);
		NEEDBITS (bit_buf, bits, bit_ptr);

	    GETWORD (bit_buf, bits + 16, bit_ptr);
	    i += tab->run;
	    if (i < 64)
		goto normal_code;
	}
	break;	/* illegal, check needed to avoid buffer overflow */
    }
    dest[63] ^= mismatch & 1;
    DUMPBITS (bit_buf, bits, 4);	/* dump end of block code */
    picture->bitstream_buf = bit_buf;
    picture->bitstream_bits = bits;
    picture->bitstream_ptr = bit_ptr;
}

static void get_non_intra_block (picture_t * picture)
{
    int i;
    int j;
    int val;
    uint8_t * scan = picture->scan;
    uint8_t * quant_matrix = picture->non_intra_quantizer_matrix;
    int quantizer_scale = picture->quantizer_scale;
    int mismatch;
    DCTtab * tab;
    uint32_t bit_buf;
    int bits;
    uint8_t * bit_ptr;
    int16_t * dest;

    i = -1;
    mismatch = 1;
    dest = picture->DCTblock;

    bit_buf = picture->bitstream_buf;
    bits = picture->bitstream_bits;
    bit_ptr = picture->bitstream_ptr;

    NEEDBITS (bit_buf, bits, bit_ptr);
    if (bit_buf >= 0x28000000) {
	tab = DCT_B14DC_5 - 5 + UBITS (bit_buf, 5);
	goto entry_1;

	    tab = DCT_B14AC_5 - 5 + UBITS (bit_buf, 5);

	entry_1:
	    i += tab->run;
	    if (i >= 64)
		break;	/* end of block */

	normal_code:
	    j = scan[i];
	    bit_buf <<= tab->len;
	    bits += tab->len + 1;
	    val = ((2*tab->level+1) * quantizer_scale * quant_matrix[j]) >> 5;

	    /* if (bitstream_get (1)) val = -val; */
	    val = (val ^ SBITS (bit_buf, 1)) - SBITS (bit_buf, 1);

	    SATURATE (val);
	    dest[j] = val;
	    mismatch ^= val;

	    i += tab->run;
	    if (i < 64)
		goto normal_code;

	    /* escape code */

	    i += UBITS (bit_buf << 6, 6) - 64;
	    if (i >= 64)
		break;	/* illegal, check needed to avoid buffer overflow */

	    j = scan[i];

	    DUMPBITS (bit_buf, bits, 12);
	    NEEDBITS (bit_buf, bits, bit_ptr);

	    GETWORD (bit_buf, bits + 16, bit_ptr);
	    i += tab->run;
	    if (i < 64)
		goto normal_code;
	}
	break;	/* illegal, check needed to avoid buffer overflow */
    }
    dest[63] ^= mismatch & 1;
    DUMPBITS (bit_buf, bits, 2);	/* dump end of block code */
    picture->bitstream_buf = bit_buf;
    picture->bitstream_bits = bits;
    picture->bitstream_ptr = bit_ptr;
}

static void get_mpeg1_intra_block (picture_t * picture)
{
    int i;
    int j;
    int val;
    uint8_t * scan = picture->scan;
    uint8_t * quant_matrix = picture->intra_quantizer_matrix;
    int quantizer_scale = picture->quantizer_scale;
    DCTtab * tab;
    uint32_t bit_buf;
    int bits;
    uint8_t * bit_ptr;
    int16_t * dest;

    i = 0;
    dest = picture->DCTblock;

    bit_buf = picture->bitstream_buf;
    bits = picture->bitstream_bits;
    bit_ptr = picture->bitstream_ptr;

    NEEDBITS (bit_buf, bits, bit_ptr);

    while (1) {
	if (bit_buf >= 0x28000000) {

	    tab = DCT_B14AC_5 - 5 + UBITS (bit_buf, 5);

	    i += tab->run;
	    if (i >= 64)
		break;	/* end of block */

	normal_code:
	    j = scan[i];
	    bit_buf <<= tab->len;
	    bits += tab->len + 1;
	    val = (tab->level * quantizer_scale * quant_matrix[j]) >> 4;

	    /* oddification */
	    val = (val - 1) | 1;

	    /* if (bitstream_get (1)) val = -val; */
	    val = (val ^ SBITS (bit_buf, 1)) - SBITS (bit_buf, 1);

	    SATURATE (val);
	    dest[j] = val;

	    i += tab->run;
	    if (i < 64)
		goto normal_code;

	    /* escape code */

	    i += UBITS (bit_buf << 6, 6) - 64;
	    if (i >= 64)
		break;	/* illegal, check needed to avoid buffer overflow */

	    j = scan[i];

	    DUMPBITS (bit_buf, bits, 12);
	    NEEDBITS (bit_buf, bits, bit_ptr);

		DUMPBITS (bit_buf, bits, 8);
		val = UBITS (bit_buf, 8) + 2 * val;
	    }
	    val = (val * quantizer_scale * quant_matrix[j]) / 16;

	    /* oddification */
	    val = (val + ~SBITS (val, 1)) | 1;

	    SATURATE (val);
	    dest[j] = val;

	    GETWORD (bit_buf, bits + 16, bit_ptr);
	    i += tab->run;
	    if (i < 64)
		goto normal_code;
	}
	break;	/* illegal, check needed to avoid buffer overflow */
    }
    DUMPBITS (bit_buf, bits, 2);	/* dump end of block code */
    picture->bitstream_buf = bit_buf;
    picture->bitstream_bits = bits;
    picture->bitstream_ptr = bit_ptr;
}

static void get_mpeg1_non_intra_block (picture_t * picture)
{
    int i;
    int j;
    int val;
    uint8_t * scan = picture->scan;
    uint8_t * quant_matrix = picture->non_intra_quantizer_matrix;
    int quantizer_scale = picture->quantizer_scale;
    DCTtab * tab;
    uint32_t bit_buf;
    int bits;
    uint8_t * bit_ptr;
    int16_t * dest;

    i = -1;
    dest = picture->DCTblock;

    bit_buf = picture->bitstream_buf;
    bits = picture->bitstream_bits;
    bit_ptr = picture->bitstream_ptr;

    NEEDBITS (bit_buf, bits, bit_ptr);
    if (bit_buf >= 0x28000000) {
	tab = DCT_B14DC_5 - 5 + UBITS (bit_buf, 5);
	goto entry_1;

	    tab = DCT_B14AC_5 - 5 + UBITS (bit_buf, 5);

	entry_1:
	    i += tab->run;
	    if (i >= 64)
		break;	/* end of block */

	normal_code:
	    j = scan[i];
	    bit_buf <<= tab->len;
	    bits += tab->len + 1;
	    val = ((2*tab->level+1) * quantizer_scale * quant_matrix[j]) >> 5;

	    /* oddification */
	    val = (val - 1) | 1;

	    /* if (bitstream_get (1)) val = -val; */
	    val = (val ^ SBITS (bit_buf, 1)) - SBITS (bit_buf, 1);

	    SATURATE (val);
	    dest[j] = val;

	    i += tab->run;
	    if (i < 64)
		goto normal_code;

	    /* escape code */

	    i += UBITS (bit_buf << 6, 6) - 64;
	    if (i >= 64)
		break;	/* illegal, check needed to avoid buffer overflow */

	    j = scan[i];

	    DUMPBITS (bit_buf, bits, 12);
	    NEEDBITS (bit_buf, bits, bit_ptr);

		val = UBITS (bit_buf, 8) + 2 * val;
	    }
	    val = 2 * (val + SBITS (val, 1)) + 1;
	    val = (val * quantizer_scale * quant_matrix[j]) / 32;

	    /* oddification */
	    val = (val + ~SBITS (val, 1)) | 1;

	    SATURATE (val);
	    dest[j] = val;

	    GETWORD (bit_buf, bits + 16, bit_ptr);
	    i += tab->run;
	    if (i < 64)
		goto normal_code;
	}
	break;	/* illegal, check needed to avoid buffer overflow */
    }
    DUMPBITS (bit_buf, bits, 2);	/* dump end of block code */
    picture->bitstream_buf = bit_buf;
    picture->bitstream_bits = bits;
    picture->bitstream_ptr = bit_ptr;
}

static inline int get_macroblock_address_increment (picture_t * picture)
{
#define bit_buf (picture->bitstream_buf)
#define bits (picture->bitstream_bits)
#define bit_ptr (picture->bitstream_ptr)

    MBAtab * tab;
    int mba;

    mba = 0;

	} else if (bit_buf >= 0x03000000) {
	    tab = MBA_11 - 24 + UBITS (bit_buf, 11);
	    DUMPBITS (bit_buf, bits, tab->len);
	    return mba + tab->mba;
	} else switch (UBITS (bit_buf, 11)) {
	case 8:		/* macroblock_escape */
	    mba += 33;
	    /* no break here on purpose */
	case 15:	/* macroblock_stuffing (MPEG1 only) */
	    DUMPBITS (bit_buf, bits, 11);
	    NEEDBITS (bit_buf, bits, bit_ptr);
	    break;
	default:	/* end of slice, or error */
	    return 0;
	}
    }

#undef bit_buf
#undef bits
#undef bit_ptr
}

static inline void slice_intra_DCT (picture_t * picture, int cc,
				    uint8_t * dest, int stride)
{
#define bit_buf (picture->bitstream_buf)
#define bits (picture->bitstream_bits)  
#define bit_ptr (picture->bitstream_ptr)
    NEEDBITS (bit_buf, bits, bit_ptr);
    /* Get the intra DC coefficient and inverse quantize it */
    if (cc == 0)
	picture->dc_dct_pred[0] += get_luma_dc_dct_diff (picture);
    else
	picture->dc_dct_pred[cc] += get_chroma_dc_dct_diff (picture);
    picture->DCTblock[0] =
	picture->dc_dct_pred[cc] << (3 - picture->intra_dc_precision);
    memset (picture->DCTblock + 1, 0, 63 * sizeof (int16_t));

    if (picture->mpeg1) {
	if (picture->picture_coding_type != D_TYPE)
	    get_mpeg1_intra_block (picture);
    } else if (picture->intra_vlc_format)
	get_intra_block_B15 (picture);
    else
	get_intra_block_B14 (picture);
    idct_block_copy (picture->DCTblock, dest, stride);
#undef bit_buf
#undef bits
#undef bit_ptr
}

static inline void slice_non_intra_DCT (picture_t * picture, uint8_t * dest,
					int stride)
{
    memset (picture->DCTblock, 0, 64 * sizeof (int16_t));
    if (picture->mpeg1)
	get_mpeg1_non_intra_block (picture);
    else
	get_non_intra_block (picture);
    idct_block_add (picture->DCTblock, dest, stride);
}

static inline void motion_block (void (** table) (uint8_t *, uint8_t *,
						  int32_t, int32_t), 
				 int x_pred, int y_pred,

    table[4+xy_half] (dest[1] + dest_offset, src1, stride, height);
    table[4+xy_half] (dest[2] + dest_offset, src2, stride, height);
}


static void motion_mp1 (picture_t * picture, motion_t * motion,
			uint8_t * dest[3], int offset, int stride,
			void (** table) (uint8_t *, uint8_t *, int, int))
{
#define bit_buf (picture->bitstream_buf)
#define bits (picture->bitstream_bits)
#define bit_ptr (picture->bitstream_ptr)
    int motion_x, motion_y;

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_x = motion->pmv[0][0] + get_motion_delta (picture,
						     motion->f_code[0]);
    motion_x = bound_motion_vector (motion_x, motion->f_code[0]);
    motion->pmv[0][0] = motion_x;

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_y = motion->pmv[0][1] + get_motion_delta (picture,
						     motion->f_code[0]);
    motion_y = bound_motion_vector (motion_y, motion->f_code[0]);
    motion->pmv[0][1] = motion_y;

    if (motion->f_code[1]) {
	motion_x <<= 1;
	motion_y <<= 1;
    }

    motion_block (table, motion_x, motion_y, dest, offset,
		  motion->ref[0], offset, stride, 16, 0);
#undef bit_buf
#undef bits
#undef bit_ptr
}

static void motion_mp1_reuse (picture_t * picture, motion_t * motion,
			      uint8_t * dest[3], int offset, int stride,
			      void (** table) (uint8_t *, uint8_t *, int, int))
{
    int motion_x, motion_y;

    motion_x = motion->pmv[0][0];

	motion_x <<= 1;
	motion_y <<= 1;
    }

    motion_block (table, motion_x, motion_y, dest, offset,
		  motion->ref[0], offset, stride, 16, 0);
}

static void motion_fr_frame (picture_t * picture, motion_t * motion,
			     uint8_t * dest[3], int offset, int stride,
			     void (** table) (uint8_t *, uint8_t *, int, int))
{
#define bit_buf (picture->bitstream_buf)
#define bits (picture->bitstream_bits)
#define bit_ptr (picture->bitstream_ptr)
    int motion_x, motion_y;

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_x = motion->pmv[0][0] + get_motion_delta (picture,
						     motion->f_code[0]);
    motion_x = bound_motion_vector (motion_x, motion->f_code[0]);
    motion->pmv[1][0] = motion->pmv[0][0] = motion_x;

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_y = motion->pmv[0][1] + get_motion_delta (picture,
						     motion->f_code[1]);
    motion_y = bound_motion_vector (motion_y, motion->f_code[1]);
    motion->pmv[1][1] = motion->pmv[0][1] = motion_y;

    motion_block (table, motion_x, motion_y, dest, offset,
		  motion->ref[0], offset, stride, 16, 0);
#undef bit_buf
#undef bits
#undef bit_ptr
}

static void motion_fr_field (picture_t * picture, motion_t * motion,
			     uint8_t * dest[3], int offset, int stride,
			     void (** table) (uint8_t *, uint8_t *, int, int))
{
#define bit_buf (picture->bitstream_buf)
#define bits (picture->bitstream_bits)
#define bit_ptr (picture->bitstream_ptr)
    int motion_x, motion_y;
    int field_select;

    NEEDBITS (bit_buf, bits, bit_ptr);
    field_select = SBITS (bit_buf, 1);
    DUMPBITS (bit_buf, bits, 1);

    motion_x = motion->pmv[0][0] + get_motion_delta (picture,
						     motion->f_code[0]);
    motion_x = bound_motion_vector (motion_x, motion->f_code[0]);
    motion->pmv[0][0] = motion_x;

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_y = (motion->pmv[0][1] >> 1) + get_motion_delta (picture,
							    motion->f_code[1]);
    /* motion_y = bound_motion_vector (motion_y, motion->f_code[1]); */
    motion->pmv[0][1] = motion_y << 1;

    motion_block (table, motion_x, motion_y, dest, offset,
		  motion->ref[0], offset + (field_select & stride),
		  stride * 2, 8, 0);

    NEEDBITS (bit_buf, bits, bit_ptr);
    field_select = SBITS (bit_buf, 1);
    DUMPBITS (bit_buf, bits, 1);

    motion_x = motion->pmv[1][0] + get_motion_delta (picture,
						     motion->f_code[0]);
    motion_x = bound_motion_vector (motion_x, motion->f_code[0]);
    motion->pmv[1][0] = motion_x;

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_y = (motion->pmv[1][1] >> 1) + get_motion_delta (picture,
							    motion->f_code[1]);
    /* motion_y = bound_motion_vector (motion_y, motion->f_code[1]); */
    motion->pmv[1][1] = motion_y << 1;

    motion_block (table, motion_x, motion_y, dest, offset + stride,
		  motion->ref[0], offset + (field_select & stride),
		  stride * 2, 8, 0);
#undef bit_buf
#undef bits
#undef bit_ptr
}

static void motion_fr_dmv (picture_t * picture, motion_t * motion,
			   uint8_t * dest[3], int offset, int stride,
			   void (** table) (uint8_t *, uint8_t *, int, int))
{
#define bit_buf (picture->bitstream_buf)
#define bits (picture->bitstream_bits)
#define bit_ptr (picture->bitstream_ptr)
    int motion_x, motion_y;
    int dmv_x, dmv_y;
    int m;
    int other_x, other_y;

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_x = motion->pmv[0][0] + get_motion_delta (picture,
						     motion->f_code[0]);
    motion_x = bound_motion_vector (motion_x, motion->f_code[0]);
    motion->pmv[1][0] = motion->pmv[0][0] = motion_x;

    NEEDBITS (bit_buf, bits, bit_ptr);
    dmv_x = get_dmv (picture);

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_y = (motion->pmv[0][1] >> 1) + get_motion_delta (picture,
							    motion->f_code[1]);
    /* motion_y = bound_motion_vector (motion_y, motion->f_code[1]); */
    motion->pmv[1][1] = motion->pmv[0][1] = motion_y << 1;

    NEEDBITS (bit_buf, bits, bit_ptr);
    dmv_y = get_dmv (picture);

    motion_block (mc_functions.put, motion_x, motion_y, dest, offset,
		  motion->ref[0], offset, stride * 2, 8, 0);

    m = picture->top_field_first ? 1 : 3;
    other_x = ((motion_x * m + (motion_x > 0)) >> 1) + dmv_x;
    other_y = ((motion_y * m + (motion_y > 0)) >> 1) + dmv_y - 1;
    motion_block (mc_functions.avg, other_x, other_y, dest, offset,
		  motion->ref[0], offset + stride, stride * 2, 8, 0);

    motion_block (mc_functions.put, motion_x, motion_y, dest, offset + stride,
		  motion->ref[0], offset + stride, stride * 2, 8, 0);

    m = picture->top_field_first ? 3 : 1;
    other_x = ((motion_x * m + (motion_x > 0)) >> 1) + dmv_x;
    other_y = ((motion_y * m + (motion_y > 0)) >> 1) + dmv_y + 1;
    motion_block (mc_functions.avg, other_x, other_y, dest, offset + stride,
		  motion->ref[0], offset, stride * 2, 8, 0);
#undef bit_buf
#undef bits
#undef bit_ptr
}

/* like motion_frame, but reuse previous motion vectors */
static void motion_fr_reuse (picture_t * picture, motion_t * motion,
			     uint8_t * dest[3], int offset, int stride,
			     void (** table) (uint8_t *, uint8_t *, int, int))
{
    motion_block (table, motion->pmv[0][0], motion->pmv[0][1], dest, offset,
		  motion->ref[0], offset, stride, 16, 0);
}

/* like motion_frame, but use null motion vectors */
static void motion_fr_zero (picture_t * picture, motion_t * motion,
			    uint8_t * dest[3], int offset, int stride,
			    void (** table) (uint8_t *, uint8_t *, int, int))
{
    motion_block (table, 0, 0, dest, offset,
		  motion->ref[0], offset, stride, 16, 0);
}

/* like motion_frame, but parsing without actual motion compensation */
static void motion_fr_conceal (picture_t * picture)
{
#define bit_buf (picture->bitstream_buf)
#define bits (picture->bitstream_bits)
#define bit_ptr (picture->bitstream_ptr)
    int tmp;

    NEEDBITS (bit_buf, bits, bit_ptr);
    tmp = (picture->f_motion.pmv[0][0] +
	   get_motion_delta (picture, picture->f_motion.f_code[0]));
    tmp = bound_motion_vector (tmp, picture->f_motion.f_code[0]);
    picture->f_motion.pmv[1][0] = picture->f_motion.pmv[0][0] = tmp;

    NEEDBITS (bit_buf, bits, bit_ptr);
    tmp = (picture->f_motion.pmv[0][1] +
	   get_motion_delta (picture, picture->f_motion.f_code[1]));
    tmp = bound_motion_vector (tmp, picture->f_motion.f_code[1]);
    picture->f_motion.pmv[1][1] = picture->f_motion.pmv[0][1] = tmp;

    DUMPBITS (bit_buf, bits, 1); /* remove marker_bit */
#undef bit_buf
#undef bits
#undef bit_ptr
}

static void motion_fi_field (picture_t * picture, motion_t * motion,
			     uint8_t * dest[3], int offset, int stride,
			     void (** table) (uint8_t *, uint8_t *, int, int))
{
#define bit_buf (picture->bitstream_buf)
#define bits (picture->bitstream_bits)
#define bit_ptr (picture->bitstream_ptr)
    int motion_x, motion_y;
    int field_select;

    NEEDBITS (bit_buf, bits, bit_ptr);
    field_select = UBITS (bit_buf, 1);
    DUMPBITS (bit_buf, bits, 1);

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_x = motion->pmv[0][0] + get_motion_delta (picture,
						     motion->f_code[0]);
    motion_x = bound_motion_vector (motion_x, motion->f_code[0]);
    motion->pmv[1][0] = motion->pmv[0][0] = motion_x;

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_y = motion->pmv[0][1] + get_motion_delta (picture,
						     motion->f_code[1]);
    motion_y = bound_motion_vector (motion_y, motion->f_code[1]);
    motion->pmv[1][1] = motion->pmv[0][1] = motion_y;

    motion_block (table, motion_x, motion_y, dest, offset,
		  motion->ref[field_select], offset, stride, 16, 0);
#undef bit_buf
#undef bits
#undef bit_ptr
}

static void motion_fi_16x8 (picture_t * picture, motion_t * motion,
			    uint8_t * dest[3], int offset, int stride,
			    void (** table) (uint8_t *, uint8_t *, int, int))
{
#define bit_buf (picture->bitstream_buf)
#define bits (picture->bitstream_bits)
#define bit_ptr (picture->bitstream_ptr)
    int motion_x, motion_y;
    int field_select;

    NEEDBITS (bit_buf, bits, bit_ptr);
    field_select = UBITS (bit_buf, 1);
    DUMPBITS (bit_buf, bits, 1);

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_x = motion->pmv[0][0] + get_motion_delta (picture,
						     motion->f_code[0]);
    motion_x = bound_motion_vector (motion_x, motion->f_code[0]);
    motion->pmv[0][0] = motion_x;

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_y = motion->pmv[0][1] + get_motion_delta (picture,
						     motion->f_code[1]);
    motion_y = bound_motion_vector (motion_y, motion->f_code[1]);
    motion->pmv[0][1] = motion_y;

    motion_block (table, motion_x, motion_y, dest, offset,
		  motion->ref[field_select], offset, stride, 8, 0);

    NEEDBITS (bit_buf, bits, bit_ptr);
    field_select = UBITS (bit_buf, 1);
    DUMPBITS (bit_buf, bits, 1);

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_x = motion->pmv[1][0] + get_motion_delta (picture,
						     motion->f_code[0]);
    motion_x = bound_motion_vector (motion_x, motion->f_code[0]);
    motion->pmv[1][0] = motion_x;

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_y = motion->pmv[1][1] + get_motion_delta (picture,
						     motion->f_code[1]);
    motion_y = bound_motion_vector (motion_y, motion->f_code[1]);
    motion->pmv[1][1] = motion_y;

    motion_block (table, motion_x, motion_y, dest, offset,
		  motion->ref[field_select], offset, stride, 8, 1);
#undef bit_buf
#undef bits
#undef bit_ptr
}

static void motion_fi_dmv (picture_t * picture, motion_t * motion,
			   uint8_t * dest[3], int offset, int stride,
			   void (** table) (uint8_t *, uint8_t *, int, int))
{
#define bit_buf (picture->bitstream_buf)
#define bits (picture->bitstream_bits)
#define bit_ptr (picture->bitstream_ptr)
    int motion_x, motion_y;
    int dmv_x, dmv_y;

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_x = motion->pmv[0][0] + get_motion_delta (picture,
						     motion->f_code[0]);
    motion_x = bound_motion_vector (motion_x, motion->f_code[0]);
    motion->pmv[1][0] = motion->pmv[0][0] = motion_x;

    NEEDBITS (bit_buf, bits, bit_ptr);
    dmv_x = get_dmv (picture);

    NEEDBITS (bit_buf, bits, bit_ptr);
    motion_y = motion->pmv[0][1] + get_motion_delta (picture,
						     motion->f_code[1]);
    motion_y = bound_motion_vector (motion_y, motion->f_code[1]);
    motion->pmv[1][1] = motion->pmv[0][1] = motion_y;

    NEEDBITS (bit_buf, bits, bit_ptr);
    dmv_y = get_dmv (picture);

    motion_block (mc_functions.put, motion_x, motion_y, dest, offset,
		  motion->ref[picture->current_field], offset, stride, 16, 0);

    motion_x = ((motion_x + (motion_x > 0)) >> 1) + dmv_x;
    motion_y = ((motion_y + (motion_y > 0)) >> 1) + dmv_y +
	2 * picture->current_field - 1;
    motion_block (mc_functions.avg, motion_x, motion_y, dest, offset,
		  motion->ref[!picture->current_field], offset, stride, 16, 0);
#undef bit_buf
#undef bits
#undef bit_ptr
}

static void motion_fi_reuse (picture_t * picture, motion_t * motion,
			     uint8_t * dest[3], int offset, int stride,
			     void (** table) (uint8_t *, uint8_t *, int, int))
{
    motion_block (table, motion->pmv[0][0], motion->pmv[0][1], dest, offset,
		  motion->ref[picture->current_field], offset, stride, 16, 0);
}

static void motion_fi_zero (picture_t * picture, motion_t * motion,
			    uint8_t * dest[3], int offset, int stride,
			    void (** table) (uint8_t *, uint8_t *, int, int))
{
    motion_block (table, 0, 0, dest, offset,
		  motion->ref[picture->current_field], offset, stride, 16, 0);
}

static void motion_fi_conceal (picture_t * picture)
{
#define bit_buf (picture->bitstream_buf)
#define bits (picture->bitstream_bits)
#define bit_ptr (picture->bitstream_ptr)
    int tmp;

    NEEDBITS (bit_buf, bits, bit_ptr);
    DUMPBITS (bit_buf, bits, 1); /* remove field_select */

    NEEDBITS (bit_buf, bits, bit_ptr);
    tmp = (picture->f_motion.pmv[0][0] +
	   get_motion_delta (picture, picture->f_motion.f_code[0]));
    tmp = bound_motion_vector (tmp, picture->f_motion.f_code[0]);
    picture->f_motion.pmv[1][0] = picture->f_motion.pmv[0][0] = tmp;

    NEEDBITS (bit_buf, bits, bit_ptr);
    tmp = (picture->f_motion.pmv[0][1] +
	   get_motion_delta (picture, picture->f_motion.f_code[1]));
    tmp = bound_motion_vector (tmp, picture->f_motion.f_code[1]);
    picture->f_motion.pmv[1][1] = picture->f_motion.pmv[0][1] = tmp;

    DUMPBITS (bit_buf, bits, 1); /* remove marker_bit */
#undef bit_buf
#undef bits
#undef bit_ptr
}

#define MOTION(routine,direction)					\
do {									\
    if ((direction) & MACROBLOCK_MOTION_FORWARD)			\
	routine (picture, &(picture->f_motion), dest, offset, stride,	\
		 mc_functions.put);					\
    if ((direction) & MACROBLOCK_MOTION_BACKWARD)			\
	routine (picture, &(picture->b_motion), dest, offset, stride,	\
		 ((direction) & MACROBLOCK_MOTION_FORWARD ?		\
		  mc_functions.avg : mc_functions.put));		\
} while (0)

#define CHECK_DISPLAY							\
do {									\
    if (offset == picture->coded_picture_width) {			\
	picture->f_motion.ref[0][0] += 16 * stride;			\
	picture->f_motion.ref[0][1] += 4 * stride;			\
	picture->f_motion.ref[0][2] += 4 * stride;			\
	picture->b_motion.ref[0][0] += 16 * stride;			\
	picture->b_motion.ref[0][1] += 4 * stride;			\
	picture->b_motion.ref[0][2] += 4 * stride;			\
	do { /* just so we can use the break statement */		\
	    if (picture->current_frame->copy) {				\
		picture->current_frame->copy (picture->current_frame,	\
					      dest);			\
		if (picture->picture_coding_type == B_TYPE)		\
		    break;						\
	    }								\
	    dest[0] += 16 * stride;					\
	    dest[1] += 4 * stride;					\
	    dest[2] += 4 * stride;					\
	} while (0);							\
	offset = 0;							\
    }									\
} while (0)

int slice_process (picture_t * picture, uint8_t code, uint8_t * buffer)
{
#define bit_buf (picture->bitstream_buf)
#define bits (picture->bitstream_bits)
#define bit_ptr (picture->bitstream_ptr)
    int macroblock_modes;
    int stride;
    uint8_t * dest[3];
    int offset;
    uint8_t ** forward_ref[2];

    stride = picture->coded_picture_width;
    offset = (code - 1) * stride * 4;

    forward_ref[0] = picture->forward_reference_frame->base;
    if (picture->picture_structure != FRAME_PICTURE) {
	offset <<= 1;
	forward_ref[1] = picture->forward_reference_frame->base;
	picture->current_field = (picture->picture_structure == BOTTOM_FIELD);
	if ((picture->second_field) &&
	    (picture->picture_coding_type != B_TYPE))
	    forward_ref[picture->picture_structure == TOP_FIELD] =
		picture->current_frame->base;
	picture->f_motion.ref[1][0] = forward_ref[1][0] + offset * 4 + stride;
	picture->f_motion.ref[1][1] =
	    forward_ref[1][1] + offset + (stride >> 1);
	picture->f_motion.ref[1][2] =
	    forward_ref[1][2] + offset + (stride >> 1);
	picture->b_motion.ref[1][0] = 
	    picture->backward_reference_frame->base[0] + offset * 4 + stride;
	picture->b_motion.ref[1][1] =
	    (picture->backward_reference_frame->base[1] +
	     offset + (stride >> 1));
	picture->b_motion.ref[1][2] =
	    (picture->backward_reference_frame->base[2] +
	     offset + (stride >> 1));
    }

    picture->f_motion.ref[0][0] = forward_ref[0][0] + offset * 4;
    picture->f_motion.ref[0][1] = forward_ref[0][1] + offset;
    picture->f_motion.ref[0][2] = forward_ref[0][2] + offset;
    picture->f_motion.f_code[0] = picture->f_code[0][0];
    picture->f_motion.f_code[1] = picture->f_code[0][1];
    picture->f_motion.pmv[0][0] = picture->f_motion.pmv[0][1] = 0;
    picture->f_motion.pmv[1][0] = picture->f_motion.pmv[1][1] = 0;
    picture->b_motion.ref[0][0] =
	picture->backward_reference_frame->base[0] + offset * 4;
    picture->b_motion.ref[0][1] =
	picture->backward_reference_frame->base[1] + offset;
    picture->b_motion.ref[0][2] =
	picture->backward_reference_frame->base[2] + offset;
    picture->b_motion.f_code[0] = picture->f_code[1][0];
    picture->b_motion.f_code[1] = picture->f_code[1][1];
    picture->b_motion.pmv[0][0] = picture->b_motion.pmv[0][1] = 0;
    picture->b_motion.pmv[1][0] = picture->b_motion.pmv[1][1] = 0;

    if ((picture->current_frame->copy) &&
	(picture->picture_coding_type == B_TYPE))
	offset = 0;

    dest[0] = picture->current_frame->base[0] + offset * 4;
    dest[1] = picture->current_frame->base[1] + offset;
    dest[2] = picture->current_frame->base[2] + offset;

    switch (picture->picture_structure) {
    case BOTTOM_FIELD:
	dest[0] += stride;
	dest[1] += stride >> 1;
	dest[2] += stride >> 1;
	/* follow thru */
    case TOP_FIELD:
	stride <<= 1;
    }

    /* reset intra dc predictor */
    picture->dc_dct_pred[0] = picture->dc_dct_pred[1] =
	picture->dc_dct_pred[2] = 1 << (picture->intra_dc_precision + 7);

    bitstream_init (picture, buffer);

    picture->quantizer_scale = get_quantizer_scale (picture);

    /* ignore intra_slice and all the extra data */
    while (bit_buf & 0x80000000) {
	DUMPBITS (bit_buf, bits, 9);
	NEEDBITS (bit_buf, bits, bit_ptr);
    }
    DUMPBITS (bit_buf, bits, 1);

    NEEDBITS (bit_buf, bits, bit_ptr);
    offset = get_macroblock_address_increment (picture) << 4;

    while (1) {
	NEEDBITS (bit_buf, bits, bit_ptr);

	macroblock_modes = get_macroblock_modes (picture);

	/* maybe integrate MACROBLOCK_QUANT test into get_macroblock_modes ? */
	if (macroblock_modes & MACROBLOCK_QUANT)
	    picture->quantizer_scale = get_quantizer_scale (picture);

	if (macroblock_modes & MACROBLOCK_INTRA) {

	    int DCT_offset, DCT_stride;

	    if (picture->concealment_motion_vectors) {
		if (picture->picture_structure == FRAME_PICTURE)
		    motion_fr_conceal (picture);
		else
		    motion_fi_conceal (picture);
	    } else {
		picture->f_motion.pmv[0][0] = picture->f_motion.pmv[0][1] = 0;
		picture->f_motion.pmv[1][0] = picture->f_motion.pmv[1][1] = 0;
		picture->b_motion.pmv[0][0] = picture->b_motion.pmv[0][1] = 0;
		picture->b_motion.pmv[1][0] = picture->b_motion.pmv[1][1] = 0;
	    }

	    if (macroblock_modes & DCT_TYPE_INTERLACED) {
		DCT_offset = stride;
		DCT_stride = stride * 2;
	    } else {
		DCT_offset = stride * 8;
		DCT_stride = stride;
	    }

	    /* Decode lum blocks */
	    slice_intra_DCT (picture, 0, dest[0] + offset, DCT_stride);
	    slice_intra_DCT (picture, 0, dest[0] + offset + 8, DCT_stride);
	    slice_intra_DCT (picture, 0, dest[0] + offset + DCT_offset,
			     DCT_stride);
	    slice_intra_DCT (picture, 0, dest[0] + offset + DCT_offset + 8,
			     DCT_stride);

	    /* Decode chroma blocks */
	    slice_intra_DCT (picture, 1, dest[1] + (offset >> 1), stride >> 1);
	    slice_intra_DCT (picture, 2, dest[2] + (offset >> 1), stride >> 1);

	    if (picture->picture_coding_type == D_TYPE) {
		NEEDBITS (bit_buf, bits, bit_ptr);
		DUMPBITS (bit_buf, bits, 1);
	    }
	} else {

	    if (picture->mpeg1) {
		if ((macroblock_modes & MOTION_TYPE_MASK) == MC_FRAME)
		    MOTION (motion_mp1, macroblock_modes);
		else {
		    /* non-intra mb without forward mv in a P picture */
		    picture->f_motion.pmv[0][0] = 0;
		    picture->f_motion.pmv[0][1] = 0;
		    picture->f_motion.pmv[1][0] = 0;
		    picture->f_motion.pmv[1][1] = 0;
		    MOTION (motion_fr_zero, MACROBLOCK_MOTION_FORWARD);
		}
	    } else if (picture->picture_structure == FRAME_PICTURE)
		switch (macroblock_modes & MOTION_TYPE_MASK) {
		case MC_FRAME:
		    MOTION (motion_fr_frame, macroblock_modes);
		    break;

		case MC_FIELD:
		    MOTION (motion_fr_field, macroblock_modes);
		    break;

		case MC_DMV:
		    MOTION (motion_fr_dmv, MACROBLOCK_MOTION_FORWARD);
		    break;

		case 0:
		    /* non-intra mb without forward mv in a P picture */
		    picture->f_motion.pmv[0][0] = 0;
		    picture->f_motion.pmv[0][1] = 0;
		    picture->f_motion.pmv[1][0] = 0;
		    picture->f_motion.pmv[1][1] = 0;
		    MOTION (motion_fr_zero, MACROBLOCK_MOTION_FORWARD);
		    break;
		}
	    else
		switch (macroblock_modes & MOTION_TYPE_MASK) {
		case MC_FIELD:
		    MOTION (motion_fi_field, macroblock_modes);
		    break;

		case MC_16X8:
		    MOTION (motion_fi_16x8, macroblock_modes);
		    break;

		case MC_DMV:
		    MOTION (motion_fi_dmv, MACROBLOCK_MOTION_FORWARD);
		    break;

		case 0:
		    /* non-intra mb without forward mv in a P picture */
		    picture->f_motion.pmv[0][0] = 0;
		    picture->f_motion.pmv[0][1] = 0;
		    picture->f_motion.pmv[1][0] = 0;
		    picture->f_motion.pmv[1][1] = 0;
		    MOTION (motion_fi_zero, MACROBLOCK_MOTION_FORWARD);
		    break;
		}

	    /* 6.3.17.4 Coded block pattern */
	    if (macroblock_modes & MACROBLOCK_PATTERN) {
		int coded_block_pattern;
		int DCT_offset, DCT_stride;

		if (macroblock_modes & DCT_TYPE_INTERLACED) {
		    DCT_offset = stride;
		    DCT_stride = stride * 2;
		} else {
		    DCT_offset = stride * 8;
		    DCT_stride = stride;
		}

		coded_block_pattern = get_coded_block_pattern (picture);

		/* Decode lum blocks */

		if (coded_block_pattern & 0x20)
		    slice_non_intra_DCT (picture, dest[0] + offset,
					 DCT_stride);
		if (coded_block_pattern & 0x10)
		    slice_non_intra_DCT (picture, dest[0] + offset + 8,
					 DCT_stride);
		if (coded_block_pattern & 0x08)
		    slice_non_intra_DCT (picture,
					 dest[0] + offset + DCT_offset,
					 DCT_stride);
		if (coded_block_pattern & 0x04)
		    slice_non_intra_DCT (picture,
					 dest[0] + offset + DCT_offset + 8,
					 DCT_stride);

		/* Decode chroma blocks */

		if (coded_block_pattern & 0x2)
		    slice_non_intra_DCT (picture, dest[1] + (offset >> 1),
					 stride >> 1);
		if (coded_block_pattern & 0x1)
		    slice_non_intra_DCT (picture, dest[2] + (offset >> 1),
					 stride >> 1);
	    }

	    picture->dc_dct_pred[0] = picture->dc_dct_pred[1] =
		picture->dc_dct_pred[2] = 1 << (picture->intra_dc_precision+7);
	}

	offset += 16;
	CHECK_DISPLAY;

	NEEDBITS (bit_buf, bits, bit_ptr);

	if (bit_buf & 0x80000000) {
	    DUMPBITS (bit_buf, bits, 1);
	} else {
	    int mba_inc;

	    mba_inc = get_macroblock_address_increment (picture);
	    if (!mba_inc)
		break;

	    /* reset intra dc predictor on skipped block */
	    picture->dc_dct_pred[0] = picture->dc_dct_pred[1] =
		picture->dc_dct_pred[2] = 1 << (picture->intra_dc_precision+7);

	    /* handling of skipped mb's differs between P_TYPE and B_TYPE */
	    /* pictures */
	    if (picture->picture_coding_type == P_TYPE) {
		picture->f_motion.pmv[0][0] = picture->f_motion.pmv[0][1] = 0;
		picture->f_motion.pmv[1][0] = picture->f_motion.pmv[1][1] = 0;

		do {
		    if (picture->picture_structure == FRAME_PICTURE)
			MOTION (motion_fr_zero, MACROBLOCK_MOTION_FORWARD);
		    else
			MOTION (motion_fi_zero, MACROBLOCK_MOTION_FORWARD);

		    offset += 16;
		    CHECK_DISPLAY;
		} while (--mba_inc);
	    } else {
		do {
		    if (picture->mpeg1)
			MOTION (motion_mp1_reuse, macroblock_modes);
		    else if (picture->picture_structure == FRAME_PICTURE)
			MOTION (motion_fr_reuse, macroblock_modes);
		    else
			MOTION (motion_fi_reuse, macroblock_modes);

		    offset += 16;
		    CHECK_DISPLAY;
		} while (--mba_inc);
	    }
	}