h264.c

/*
 * H.26L/H.264/AVC/JVT/14496-10/... encoder/decoder
 * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
 *
 * This file is part of FFmpeg.
 *
 * FFmpeg is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
 * version 2.1 of the License, or (at your option) any later version.
 *
 * FFmpeg is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
 * License along with FFmpeg; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 */

/**
 * @file libavcodec/h264.c
 * H.264 / AVC / MPEG4 part10 codec.
 * @author Michael Niedermayer <michaelni@gmx.at>
 */

#include "internal.h"
#include "dsputil.h"
#include "avcodec.h"
#include "mpegvideo.h"
#include "h264.h"
#include "h264data.h"
#include "h264_mvpred.h"
#include "h264_parser.h"
#include "golomb.h"
#include "mathops.h"
#include "rectangle.h"
#include "vdpau_internal.h"

#include "cabac.h"

//#undef NDEBUG
#include <assert.h>

static const uint8_t rem6[52]={
0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
};

static const uint8_t div6[52]={
0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8,
};

void ff_h264_write_back_intra_pred_mode(H264Context *h){
    const int mb_xy= h->mb_xy;

    h->intra4x4_pred_mode[mb_xy][0]= h->intra4x4_pred_mode_cache[7+8*1];
    h->intra4x4_pred_mode[mb_xy][1]= h->intra4x4_pred_mode_cache[7+8*2];
    h->intra4x4_pred_mode[mb_xy][2]= h->intra4x4_pred_mode_cache[7+8*3];
    h->intra4x4_pred_mode[mb_xy][3]= h->intra4x4_pred_mode_cache[7+8*4];
    h->intra4x4_pred_mode[mb_xy][4]= h->intra4x4_pred_mode_cache[4+8*4];
    h->intra4x4_pred_mode[mb_xy][5]= h->intra4x4_pred_mode_cache[5+8*4];
    h->intra4x4_pred_mode[mb_xy][6]= h->intra4x4_pred_mode_cache[6+8*4];
}

/**
 * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
 */
int ff_h264_check_intra4x4_pred_mode(H264Context *h){
    MpegEncContext * const s = &h->s;
    static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
    static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
    int i;

    if(!(h->top_samples_available&0x8000)){
        for(i=0; i<4; i++){
            int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
            if(status<0){
                av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
                return -1;
            } else if(status){
                h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
            }
        }
    }

    if((h->left_samples_available&0x8888)!=0x8888){
        static const int mask[4]={0x8000,0x2000,0x80,0x20};
        for(i=0; i<4; i++){
            if(!(h->left_samples_available&mask[i])){
                int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
                if(status<0){
                    av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
                    return -1;
                } else if(status){
                    h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
                }
            }
        }
    }

    return 0;
} //FIXME cleanup like ff_h264_check_intra_pred_mode

/**
 * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
 */
int ff_h264_check_intra_pred_mode(H264Context *h, int mode){
    MpegEncContext * const s = &h->s;
    static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
    static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};

    if(mode > 6U) {
        av_log(h->s.avctx, AV_LOG_ERROR, "out of range intra chroma pred mode at %d %d\n", s->mb_x, s->mb_y);
        return -1;
    }

    if(!(h->top_samples_available&0x8000)){
        mode= top[ mode ];
        if(mode<0){
            av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
            return -1;
        }
    }

    if((h->left_samples_available&0x8080) != 0x8080){
        mode= left[ mode ];
        if(h->left_samples_available&0x8080){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
            mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
        }
        if(mode<0){
            av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
            return -1;
        }
    }

    return mode;
}

const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
    int i, si, di;
    uint8_t *dst;
    int bufidx;

//    src[0]&0x80;                //forbidden bit
    h->nal_ref_idc= src[0]>>5;
    h->nal_unit_type= src[0]&0x1F;

    src++; length--;
#if 0
    for(i=0; i<length; i++)
        printf("%2X ", src[i]);
#endif

#if HAVE_FAST_UNALIGNED
# if HAVE_FAST_64BIT
#   define RS 7
    for(i=0; i+1<length; i+=9){
        if(!((~*(const uint64_t*)(src+i) & (*(const uint64_t*)(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
# else
#   define RS 3
    for(i=0; i+1<length; i+=5){
        if(!((~*(const uint32_t*)(src+i) & (*(const uint32_t*)(src+i) - 0x01000101U)) & 0x80008080U))
# endif
            continue;
        if(i>0 && !src[i]) i--;
        while(src[i]) i++;
#else
#   define RS 0
    for(i=0; i+1<length; i+=2){
        if(src[i]) continue;
        if(i>0 && src[i-1]==0) i--;
#endif
        if(i+2<length && src[i+1]==0 && src[i+2]<=3){
            if(src[i+2]!=3){
                /* startcode, so we must be past the end */
                length=i;
            }
            break;
        }
        i-= RS;
    }

    if(i>=length-1){ //no escaped 0
        *dst_length= length;
        *consumed= length+1; //+1 for the header
        return src;
    }

    bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
    av_fast_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE);
    dst= h->rbsp_buffer[bufidx];

    if (dst == NULL){
        return NULL;
    }

//printf("decoding esc\n");
    memcpy(dst, src, i);
    si=di=i;
    while(si+2<length){
        //remove escapes (very rare 1:2^22)
        if(src[si+2]>3){
            dst[di++]= src[si++];
            dst[di++]= src[si++];
        }else if(src[si]==0 && src[si+1]==0){
            if(src[si+2]==3){ //escape
                dst[di++]= 0;
                dst[di++]= 0;
                si+=3;
                continue;
            }else //next start code
                goto nsc;
        }

        dst[di++]= src[si++];
    }
    while(si<length)
        dst[di++]= src[si++];
nsc:

    memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);

    *dst_length= di;
    *consumed= si + 1;//+1 for the header
//FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
    return dst;
}

int ff_h264_decode_rbsp_trailing(H264Context *h, const uint8_t *src){
    int v= *src;
    int r;

    tprintf(h->s.avctx, "rbsp trailing %X\n", v);

    for(r=1; r<9; r++){
        if(v&1) return r;
        v>>=1;
    }
    return 0;
}

/**
 * IDCT transforms the 16 dc values and dequantizes them.
 * @param qp quantization parameter
 */
static void h264_luma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
#define stride 16
    int i;
    int temp[16]; //FIXME check if this is a good idea
    static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
    static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};

//memset(block, 64, 2*256);
//return;
    for(i=0; i<4; i++){
        const int offset= y_offset[i];
        const int z0= block[offset+stride*0] + block[offset+stride*4];
        const int z1= block[offset+stride*0] - block[offset+stride*4];
        const int z2= block[offset+stride*1] - block[offset+stride*5];
        const int z3= block[offset+stride*1] + block[offset+stride*5];

        temp[4*i+0]= z0+z3;
        temp[4*i+1]= z1+z2;
        temp[4*i+2]= z1-z2;
        temp[4*i+3]= z0-z3;
    }

    for(i=0; i<4; i++){
        const int offset= x_offset[i];
        const int z0= temp[4*0+i] + temp[4*2+i];
        const int z1= temp[4*0+i] - temp[4*2+i];
        const int z2= temp[4*1+i] - temp[4*3+i];
        const int z3= temp[4*1+i] + temp[4*3+i];

        block[stride*0 +offset]= ((((z0 + z3)*qmul + 128 ) >> 8)); //FIXME think about merging this into decode_residual
        block[stride*2 +offset]= ((((z1 + z2)*qmul + 128 ) >> 8));
        block[stride*8 +offset]= ((((z1 - z2)*qmul + 128 ) >> 8));
        block[stride*10+offset]= ((((z0 - z3)*qmul + 128 ) >> 8));
    }
}

#if 0
/**
 * DCT transforms the 16 dc values.
 * @param qp quantization parameter ??? FIXME
 */
static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
//    const int qmul= dequant_coeff[qp][0];
    int i;
    int temp[16]; //FIXME check if this is a good idea
    static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
    static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};

    for(i=0; i<4; i++){
        const int offset= y_offset[i];
        const int z0= block[offset+stride*0] + block[offset+stride*4];
        const int z1= block[offset+stride*0] - block[offset+stride*4];
        const int z2= block[offset+stride*1] - block[offset+stride*5];
        const int z3= block[offset+stride*1] + block[offset+stride*5];

        temp[4*i+0]= z0+z3;
        temp[4*i+1]= z1+z2;
        temp[4*i+2]= z1-z2;
        temp[4*i+3]= z0-z3;
    }

    for(i=0; i<4; i++){
        const int offset= x_offset[i];
        const int z0= temp[4*0+i] + temp[4*2+i];
        const int z1= temp[4*0+i] - temp[4*2+i];
        const int z2= temp[4*1+i] - temp[4*3+i];
        const int z3= temp[4*1+i] + temp[4*3+i];

        block[stride*0 +offset]= (z0 + z3)>>1;
        block[stride*2 +offset]= (z1 + z2)>>1;
        block[stride*8 +offset]= (z1 - z2)>>1;
        block[stride*10+offset]= (z0 - z3)>>1;
    }
}
#endif

#undef xStride
#undef stride

static void chroma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
    const int stride= 16*2;
    const int xStride= 16;
    int a,b,c,d,e;

    a= block[stride*0 + xStride*0];
    b= block[stride*0 + xStride*1];
    c= block[stride*1 + xStride*0];
    d= block[stride*1 + xStride*1];

    e= a-b;
    a= a+b;
    b= c-d;
    c= c+d;

    block[stride*0 + xStride*0]= ((a+c)*qmul) >> 7;
    block[stride*0 + xStride*1]= ((e+b)*qmul) >> 7;
    block[stride*1 + xStride*0]= ((a-c)*qmul) >> 7;
    block[stride*1 + xStride*1]= ((e-b)*qmul) >> 7;
}

#if 0
static void chroma_dc_dct_c(DCTELEM *block){
    const int stride= 16*2;
    const int xStride= 16;
    int a,b,c,d,e;

    a= block[stride*0 + xStride*0];
    b= block[stride*0 + xStride*1];
    c= block[stride*1 + xStride*0];
    d= block[stride*1 + xStride*1];

    e= a-b;
    a= a+b;
    b= c-d;
    c= c+d;

    block[stride*0 + xStride*0]= (a+c);
    block[stride*0 + xStride*1]= (e+b);
    block[stride*1 + xStride*0]= (a-c);
    block[stride*1 + xStride*1]= (e-b);
}
#endif

static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
                           uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
                           int src_x_offset, int src_y_offset,
                           qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op){
    MpegEncContext * const s = &h->s;
    const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
    int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
    const int luma_xy= (mx&3) + ((my&3)<<2);
    uint8_t * src_y = pic->data[0] + (mx>>2) + (my>>2)*h->mb_linesize;
    uint8_t * src_cb, * src_cr;
    int extra_width= h->emu_edge_width;
    int extra_height= h->emu_edge_height;
    int emu=0;
    const int full_mx= mx>>2;
    const int full_my= my>>2;
    const int pic_width  = 16*s->mb_width;
    const int pic_height = 16*s->mb_height >> MB_FIELD;

    if(mx&7) extra_width -= 3;
    if(my&7) extra_height -= 3;

    if(   full_mx < 0-extra_width
       || full_my < 0-extra_height
       || full_mx + 16/*FIXME*/ > pic_width + extra_width
       || full_my + 16/*FIXME*/ > pic_height + extra_height){
        ff_emulated_edge_mc(s->edge_emu_buffer, src_y - 2 - 2*h->mb_linesize, h->mb_linesize, 16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
            src_y= s->edge_emu_buffer + 2 + 2*h->mb_linesize;
        emu=1;
    }

    qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
    if(!square){
        qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
    }

    if(CONFIG_GRAY && s->flags&CODEC_FLAG_GRAY) return;

    if(MB_FIELD){
        // chroma offset when predicting from a field of opposite parity
        my += 2 * ((s->mb_y & 1) - (pic->reference - 1));
        emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
    }
    src_cb= pic->data[1] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
    src_cr= pic->data[2] + (mx>>3) + (my>>3)*h->mb_uvlinesize;

    if(emu){
        ff_emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize, 9, 9/*FIXME*/, (mx>>3), (my>>3), pic_width>>1, pic_height>>1);
            src_cb= s->edge_emu_buffer;
    }
    chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);

    if(emu){
        ff_emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize, 9, 9/*FIXME*/, (mx>>3), (my>>3), pic_width>>1, pic_height>>1);
            src_cr= s->edge_emu_buffer;
    }
    chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
}

static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
                           uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
                           int x_offset, int y_offset,
                           qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
                           qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
                           int list0, int list1){
    MpegEncContext * const s = &h->s;
    qpel_mc_func *qpix_op=  qpix_put;
    h264_chroma_mc_func chroma_op= chroma_put;

    dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
    dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
    dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
    x_offset += 8*s->mb_x;
    y_offset += 8*(s->mb_y >> MB_FIELD);

    if(list0){
        Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
        mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
                           dest_y, dest_cb, dest_cr, x_offset, y_offset,
                           qpix_op, chroma_op);

        qpix_op=  qpix_avg;
        chroma_op= chroma_avg;
    }

    if(list1){
        Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
        mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
                           dest_y, dest_cb, dest_cr, x_offset, y_offset,
                           qpix_op, chroma_op);
    }
}

static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
                           uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
                           int x_offset, int y_offset,
                           qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
                           h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
                           h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
                           int list0, int list1){
    MpegEncContext * const s = &h->s;

    dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
    dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
    dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
    x_offset += 8*s->mb_x;
    y_offset += 8*(s->mb_y >> MB_FIELD);

    if(list0 && list1){
        /* don't optimize for luma-only case, since B-frames usually
         * use implicit weights => chroma too. */
        uint8_t *tmp_cb = s->obmc_scratchpad;
        uint8_t *tmp_cr = s->obmc_scratchpad + 8;
        uint8_t *tmp_y  = s->obmc_scratchpad + 8*h->mb_uvlinesize;
        int refn0 = h->ref_cache[0][ scan8[n] ];
        int refn1 = h->ref_cache[1][ scan8[n] ];

        mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
                    dest_y, dest_cb, dest_cr,
                    x_offset, y_offset, qpix_put, chroma_put);
        mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
                    tmp_y, tmp_cb, tmp_cr,
                    x_offset, y_offset, qpix_put, chroma_put);

        if(h->use_weight == 2){
            int weight0 = h->implicit_weight[refn0][refn1];
            int weight1 = 64 - weight0;
            luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
            chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
            chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
        }else{
            luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
                            h->luma_weight[0][refn0], h->luma_weight[1][refn1],
                            h->luma_offset[0][refn0] + h->luma_offset[1][refn1]);
            chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
                            h->chroma_weight[0][refn0][0], h->chroma_weight[1][refn1][0],
                            h->chroma_offset[0][refn0][0] + h->chroma_offset[1][refn1][0]);
            chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
                            h->chroma_weight[0][refn0][1], h->chroma_weight[1][refn1][1],
                            h->chroma_offset[0][refn0][1] + h->chroma_offset[1][refn1][1]);
        }
    }else{
        int list = list1 ? 1 : 0;
        int refn = h->ref_cache[list][ scan8[n] ];
        Picture *ref= &h->ref_list[list][refn];
        mc_dir_part(h, ref, n, square, chroma_height, delta, list,
                    dest_y, dest_cb, dest_cr, x_offset, y_offset,
                    qpix_put, chroma_put);

        luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
                       h->luma_weight[list][refn], h->luma_offset[list][refn]);
        if(h->use_weight_chroma){
            chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
                             h->chroma_weight[list][refn][0], h->chroma_offset[list][refn][0]);
            chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
                             h->chroma_weight[list][refn][1], h->chroma_offset[list][refn][1]);
        }
    }
}

static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
                           uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
                           int x_offset, int y_offset,
                           qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
                           qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
                           h264_weight_func *weight_op, h264_biweight_func *weight_avg,
                           int list0, int list1){
    if((h->use_weight==2 && list0 && list1
        && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ] != 32))
       || h->use_weight==1)
        mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
                         x_offset, y_offset, qpix_put, chroma_put,
                         weight_op[0], weight_op[3], weight_avg[0], weight_avg[3], list0, list1);
    else
        mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
                    x_offset, y_offset, qpix_put, chroma_put, qpix_avg, chroma_avg, list0, list1);
}

static inline void prefetch_motion(H264Context *h, int list){
    /* fetch pixels for estimated mv 4 macroblocks ahead
     * optimized for 64byte cache lines */
    MpegEncContext * const s = &h->s;
    const int refn = h->ref_cache[list][scan8[0]];
    if(refn >= 0){
        const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
        const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
        uint8_t **src= h->ref_list[list][refn].data;
        int off= mx + (my + (s->mb_x&3)*4)*h->mb_linesize + 64;
        s->dsp.prefetch(src[0]+off, s->linesize, 4);
        off= (mx>>1) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + 64;
        s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
    }
}

static void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
                      qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
                      qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
                      h264_weight_func *weight_op, h264_biweight_func *weight_avg){
    MpegEncContext * const s = &h->s;
    const int mb_xy= h->mb_xy;
    const int mb_type= s->current_picture.mb_type[mb_xy];

    assert(IS_INTER(mb_type));

    prefetch_motion(h, 0);

    if(IS_16X16(mb_type)){
        mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
                qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
                weight_op, weight_avg,
                IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
    }else if(IS_16X8(mb_type)){
        mc_part(h, 0, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 0,
                qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
                &weight_op[1], &weight_avg[1],
                IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
        mc_part(h, 8, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 4,
                qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
                &weight_op[1], &weight_avg[1],
                IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
    }else if(IS_8X16(mb_type)){
        mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
                qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
                &weight_op[2], &weight_avg[2],
                IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
        mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
                qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
                &weight_op[2], &weight_avg[2],
                IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
    }else{
        int i;

        assert(IS_8X8(mb_type));

        for(i=0; i<4; i++){
            const int sub_mb_type= h->sub_mb_type[i];
            const int n= 4*i;
            int x_offset= (i&1)<<2;
            int y_offset= (i&2)<<1;

            if(IS_SUB_8X8(sub_mb_type)){
                mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
                    qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
                    &weight_op[3], &weight_avg[3],
                    IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
            }else if(IS_SUB_8X4(sub_mb_type)){
                mc_part(h, n  , 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset,
                    qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
                    &weight_op[4], &weight_avg[4],
                    IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
                mc_part(h, n+2, 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
                    qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
                    &weight_op[4], &weight_avg[4],
                    IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
            }else if(IS_SUB_4X8(sub_mb_type)){
                mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
                    qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
                    &weight_op[5], &weight_avg[5],
                    IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
                mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
                    qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
                    &weight_op[5], &weight_avg[5],
                    IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
            }else{
                int j;
                assert(IS_SUB_4X4(sub_mb_type));
                for(j=0; j<4; j++){
                    int sub_x_offset= x_offset + 2*(j&1);
                    int sub_y_offset= y_offset +   (j&2);
                    mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
                        qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
                        &weight_op[6], &weight_avg[6],
                        IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
                }
            }
        }
    }

    prefetch_motion(h, 1);
}


static void free_tables(H264Context *h){
    int i;
    H264Context *hx;
    av_freep(&h->intra4x4_pred_mode);
    av_freep(&h->chroma_pred_mode_table);
    av_freep(&h->cbp_table);
    av_freep(&h->mvd_table[0]);
    av_freep(&h->mvd_table[1]);
    av_freep(&h->direct_table);
    av_freep(&h->non_zero_count);
    av_freep(&h->slice_table_base);
    h->slice_table= NULL;
    av_freep(&h->list_counts);

    av_freep(&h->mb2b_xy);
    av_freep(&h->mb2b8_xy);

    for(i = 0; i < MAX_THREADS; i++) {
        hx = h->thread_context[i];
        if(!hx) continue;
        av_freep(&hx->top_borders[1]);
        av_freep(&hx->top_borders[0]);
        av_freep(&hx->s.obmc_scratchpad);
        av_freep(&hx->rbsp_buffer[1]);
        av_freep(&hx->rbsp_buffer[0]);
        hx->rbsp_buffer_size[0] = 0;
        hx->rbsp_buffer_size[1] = 0;
        if (i) av_freep(&h->thread_context[i]);
    }
}

static void init_dequant8_coeff_table(H264Context *h){
    int i,q,x;
    const int transpose = (h->s.dsp.h264_idct8_add != ff_h264_idct8_add_c); //FIXME ugly
    h->dequant8_coeff[0] = h->dequant8_buffer[0];
    h->dequant8_coeff[1] = h->dequant8_buffer[1];

    for(i=0; i<2; i++ ){
        if(i && !memcmp(h->pps.scaling_matrix8[0], h->pps.scaling_matrix8[1], 64*sizeof(uint8_t))){
            h->dequant8_coeff[1] = h->dequant8_buffer[0];
            break;
        }

        for(q=0; q<52; q++){
            int shift = div6[q];
            int idx = rem6[q];
            for(x=0; x<64; x++)
                h->dequant8_coeff[i][q][transpose ? (x>>3)|((x&7)<<3) : x] =
                    ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
                    h->pps.scaling_matrix8[i][x]) << shift;
        }
    }
}

static void init_dequant4_coeff_table(H264Context *h){
    int i,j,q,x;
    const int transpose = (h->s.dsp.h264_idct_add != ff_h264_idct_add_c); //FIXME ugly
    for(i=0; i<6; i++ ){
        h->dequant4_coeff[i] = h->dequant4_buffer[i];
        for(j=0; j<i; j++){
            if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
                h->dequant4_coeff[i] = h->dequant4_buffer[j];
                break;
            }
        }
        if(j<i)
            continue;

        for(q=0; q<52; q++){
            int shift = div6[q] + 2;
            int idx = rem6[q];
            for(x=0; x<16; x++)
                h->dequant4_coeff[i][q][transpose ? (x>>2)|((x<<2)&0xF) : x] =
                    ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
                    h->pps.scaling_matrix4[i][x]) << shift;
        }
    }
}

static void init_dequant_tables(H264Context *h){
    int i,x;
    init_dequant4_coeff_table(h);
    if(h->pps.transform_8x8_mode)
        init_dequant8_coeff_table(h);
    if(h->sps.transform_bypass){
        for(i=0; i<6; i++)
            for(x=0; x<16; x++)
                h->dequant4_coeff[i][0][x] = 1<<6;
        if(h->pps.transform_8x8_mode)
            for(i=0; i<2; i++)
                for(x=0; x<64; x++)
                    h->dequant8_coeff[i][0][x] = 1<<6;
    }
}


int ff_h264_alloc_tables(H264Context *h){
    MpegEncContext * const s = &h->s;
    const int big_mb_num= s->mb_stride * (s->mb_height+1);
    int x,y;

    FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode, big_mb_num * 8  * sizeof(uint8_t), fail)

    FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count    , big_mb_num * 32 * sizeof(uint8_t), fail)
    FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base), fail)
    FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table, big_mb_num * sizeof(uint16_t), fail)

    FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t), fail)
    FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0], 32*big_mb_num * sizeof(uint16_t), fail);
    FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1], 32*big_mb_num * sizeof(uint16_t), fail);
    FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table, 32*big_mb_num * sizeof(uint8_t) , fail);
    FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts, big_mb_num * sizeof(uint8_t), fail)

    memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
    h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;

    FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy  , big_mb_num * sizeof(uint32_t), fail);
    FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b8_xy , big_mb_num * sizeof(uint32_t), fail);
    for(y=0; y<s->mb_height; y++){
        for(x=0; x<s->mb_width; x++){
            const int mb_xy= x + y*s->mb_stride;
            const int b_xy = 4*x + 4*y*h->b_stride;
            const int b8_xy= 2*x + 2*y*h->b8_stride;

            h->mb2b_xy [mb_xy]= b_xy;
            h->mb2b8_xy[mb_xy]= b8_xy;
        }
    }

    s->obmc_scratchpad = NULL;

    if(!h->dequant4_coeff[0])
        init_dequant_tables(h);

    return 0;
fail:
    free_tables(h);
    return -1;
}

/**
 * Mimic alloc_tables(), but for every context thread.
 */
static void clone_tables(H264Context *dst, H264Context *src){
    dst->intra4x4_pred_mode       = src->intra4x4_pred_mode;
    dst->non_zero_count           = src->non_zero_count;
    dst->slice_table              = src->slice_table;
    dst->cbp_table                = src->cbp_table;
    dst->mb2b_xy                  = src->mb2b_xy;
    dst->mb2b8_xy                 = src->mb2b8_xy;
    dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
    dst->mvd_table[0]             = src->mvd_table[0];
    dst->mvd_table[1]             = src->mvd_table[1];
    dst->direct_table             = src->direct_table;
    dst->list_counts              = src->list_counts;

    dst->s.obmc_scratchpad = NULL;
    ff_h264_pred_init(&dst->hpc, src->s.codec_id);
}

/**
 * Init context
 * Allocate buffers which are not shared amongst multiple threads.
 */
static int context_init(H264Context *h){
    FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0], h->s.mb_width * (16+8+8) * sizeof(uint8_t), fail)
    FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1], h->s.mb_width * (16+8+8) * sizeof(uint8_t), fail)

    return 0;
fail:
    return -1; // free_tables will clean up for us
}

static av_cold void common_init(H264Context *h){
    MpegEncContext * const s = &h->s;

    s->width = s->avctx->width;
    s->height = s->avctx->height;
    s->codec_id= s->avctx->codec->id;

    ff_h264_pred_init(&h->hpc, s->codec_id);

    h->dequant_coeff_pps= -1;
    s->unrestricted_mv=1;
    s->decode=1; //FIXME

    dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early

    memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
    memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
}

av_cold int ff_h264_decode_init(AVCodecContext *avctx){
    H264Context *h= avctx->priv_data;
    MpegEncContext * const s = &h->s;

    MPV_decode_defaults(s);

    s->avctx = avctx;
    common_init(h);

    s->out_format = FMT_H264;
    s->workaround_bugs= avctx->workaround_bugs;

    // set defaults
//    s->decode_mb= ff_h263_decode_mb;