comparison h264.c @ 3215:06f98047ff26 libavcodec

prefetch pixels for future motion compensation. 2-5% faster h264.

3214:91f89a395b28

    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 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= s->mb_x + s->mb_y*s->mb_stride;
    const int mb_type= s->current_picture.mb_type[mb_xy];

    assert(IS_INTER(mb_type));

    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[0], &weight_avg[0],

                        IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
                }
            }
        }
    }
}

static void decode_init_vlc(H264Context *h){
    static int done = 0;

3215:06f98047ff26

    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)*s->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= s->mb_x + s->mb_y*s->mb_stride;
    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[0], &weight_avg[0],

                        IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
                }
            }
        }
    }

    prefetch_motion(h, 1);
}

static void decode_init_vlc(H264Context *h){
    static int done = 0;