comparison snow.c @ 5633:873ea64637d9 libavcodec

code to do halfpel interpolation per frame (unfinished and under ifdef but it should be faster when its finished)

5632:cdc40afde7b8

    DSPContext dsp;
    AVFrame new_picture;
    AVFrame input_picture;              ///< new_picture with the internal linesizes
    AVFrame current_picture;
    AVFrame last_picture[MAX_REF_FRAMES];
    AVFrame mconly_picture;
//     uint8_t q_context[16];
    uint8_t header_state[32];
    uint8_t block_state[128 + 32*128];
    int keyframe;

    }

    return 0;
}

static int frame_start(SnowContext *s){
   AVFrame tmp;
   int w= s->avctx->width; //FIXME round up to x16 ?
   int h= s->avctx->height;

        draw_edges(s->current_picture.data[2], s->current_picture.linesize[2], w>>1, h>>1, EDGE_WIDTH/2);
    }

    tmp= s->last_picture[s->max_ref_frames-1];
    memmove(s->last_picture+1, s->last_picture, (s->max_ref_frames-1)*sizeof(AVFrame));
    s->last_picture[0]= s->current_picture;
    s->current_picture= tmp;

    if(s->keyframe){
        s->ref_frames= 0;

            s->avctx->error[plane_index] += error;
            s->current_picture.error[plane_index] = error;
        }
    }

    if(s->last_picture[s->max_ref_frames-1].data[0])
        avctx->release_buffer(avctx, &s->last_picture[s->max_ref_frames-1]);

    s->current_picture.coded_picture_number = avctx->frame_number;
    s->current_picture.pict_type = pict->pict_type;
    s->current_picture.quality = pict->quality;
    s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);

static int decode_frame(AVCodecContext *avctx, void *data, int *data_size, uint8_t *buf, int buf_size){
    SnowContext *s = avctx->priv_data;
    RangeCoder * const c= &s->c;
    int bytes_read;
    AVFrame *picture = data;
    int level, orientation, plane_index;

    ff_init_range_decoder(c, buf, buf_size);
    ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);

    s->current_picture.pict_type= FF_I_TYPE; //FIXME I vs. P

STOP_TIMER("idwt + predict_slices")}
    }

    emms_c();

    if(s->last_picture[s->max_ref_frames-1].data[0])
        avctx->release_buffer(avctx, &s->last_picture[s->max_ref_frames-1]);

if(!(s->avctx->debug&2048))
    *picture= s->current_picture;
else
    *picture= s->mconly_picture;

5633:873ea64637d9

    DSPContext dsp;
    AVFrame new_picture;
    AVFrame input_picture;              ///< new_picture with the internal linesizes
    AVFrame current_picture;
    AVFrame last_picture[MAX_REF_FRAMES];
    uint8_t *halfpel_plane[MAX_REF_FRAMES][4][4];
    AVFrame mconly_picture;
//     uint8_t q_context[16];
    uint8_t header_state[32];
    uint8_t block_state[128 + 32*128];
    int keyframe;

    }

    return 0;
}

static void halfpel_interpol(SnowContext *s, uint8_t *halfpel[4][4], AVFrame *frame){
    int p,x,y;

    assert(!(s->avctx->flags & CODEC_FLAG_EMU_EDGE));

    for(p=0; p<3; p++){
        int is_chroma= !!p;
        int w= s->avctx->width  >>is_chroma;
        int h= s->avctx->height >>is_chroma;
        int ls= frame->linesize[p];
        uint8_t *src= frame->data[p];

        halfpel[1][p]= (uint8_t*)av_malloc(ls * (h+2*EDGE_WIDTH)) + EDGE_WIDTH*(1+ls);
        halfpel[2][p]= (uint8_t*)av_malloc(ls * (h+2*EDGE_WIDTH)) + EDGE_WIDTH*(1+ls);
        halfpel[3][p]= (uint8_t*)av_malloc(ls * (h+2*EDGE_WIDTH)) + EDGE_WIDTH*(1+ls);

        halfpel[0][p]= src;
        for(y=0; y<h; y++){
            for(x=0; x<w; x++){
                int i= y*ls + x;

                halfpel[1][p][i]= (20*(src[i] + src[i+1]) - 5*(src[i-1] + src[i+2]) + (src[i-2] + src[i+3]) + 16 )>>5;
            }
        }
        for(y=0; y<h; y++){
            for(x=0; x<w; x++){
                int i= y*ls + x;

                halfpel[2][p][i]= (20*(src[i] + src[i+ls]) - 5*(src[i-ls] + src[i+2*ls]) + (src[i-2*ls] + src[i+3*ls]) + 16 )>>5;
            }
        }
        src= halfpel[1][p];
        for(y=0; y<h; y++){
            for(x=0; x<w; x++){
                int i= y*ls + x;

                halfpel[3][p][i]= (20*(src[i] + src[i+ls]) - 5*(src[i-ls] + src[i+2*ls]) + (src[i-2*ls] + src[i+3*ls]) + 16 )>>5;
            }
        }

//FIXME border!
    }
}

static int frame_start(SnowContext *s){
   AVFrame tmp;
   int w= s->avctx->width; //FIXME round up to x16 ?
   int h= s->avctx->height;

        draw_edges(s->current_picture.data[2], s->current_picture.linesize[2], w>>1, h>>1, EDGE_WIDTH/2);
    }

    tmp= s->last_picture[s->max_ref_frames-1];
    memmove(s->last_picture+1, s->last_picture, (s->max_ref_frames-1)*sizeof(AVFrame));
    memmove(s->halfpel_plane+1, s->halfpel_plane, (s->max_ref_frames-1)*sizeof(void*)*4*4);
#ifdef USE_HALFPEL_PLANE
    if(s->current_picture.data[0])
        halfpel_interpol(s, s->halfpel_plane[0], &s->current_picture);
#endif
    s->last_picture[0]= s->current_picture;
    s->current_picture= tmp;

    if(s->keyframe){
        s->ref_frames= 0;

            s->avctx->error[plane_index] += error;
            s->current_picture.error[plane_index] = error;
        }
    }

    if(s->last_picture[s->max_ref_frames-1].data[0]){
        avctx->release_buffer(avctx, &s->last_picture[s->max_ref_frames-1]);
        for(i=0; i<9; i++)
            if(s->halfpel_plane[s->max_ref_frames-1][1+i/3][i%3])
                av_free(s->halfpel_plane[s->max_ref_frames-1][1+i/3][i%3] - EDGE_WIDTH*(1+s->current_picture.linesize[i%3]));
    }

    s->current_picture.coded_picture_number = avctx->frame_number;
    s->current_picture.pict_type = pict->pict_type;
    s->current_picture.quality = pict->quality;
    s->m.frame_bits = 8*(s->c.bytestream - s->c.bytestream_start);

static int decode_frame(AVCodecContext *avctx, void *data, int *data_size, uint8_t *buf, int buf_size){
    SnowContext *s = avctx->priv_data;
    RangeCoder * const c= &s->c;
    int bytes_read;
    AVFrame *picture = data;
    int level, orientation, plane_index, i;

    ff_init_range_decoder(c, buf, buf_size);
    ff_build_rac_states(c, 0.05*(1LL<<32), 256-8);

    s->current_picture.pict_type= FF_I_TYPE; //FIXME I vs. P

STOP_TIMER("idwt + predict_slices")}
    }

    emms_c();

    if(s->last_picture[s->max_ref_frames-1].data[0]){
        avctx->release_buffer(avctx, &s->last_picture[s->max_ref_frames-1]);
        for(i=0; i<9; i++)
            if(s->halfpel_plane[s->max_ref_frames-1][1+i/3][i%3])
                av_free(s->halfpel_plane[s->max_ref_frames-1][1+i/3][i%3] - EDGE_WIDTH*(1+s->current_picture.linesize[i%3]));
    }

if(!(s->avctx->debug&2048))
    *picture= s->current_picture;
else
    *picture= s->mconly_picture;