}


        // rc will use sad value
        t->mb.sad = t->cmp[MB_16x16](t->mb.src_y, t->stride, ref, 16);
        if (t->mb.sad < sad_t)
        {
            return FALSE;
        }

        {
            // use foreman.cif, qp = 30, waste 35497 times encode vs. total times 80266
            DECLARE_ALIGNED_MATRIX(dct, 16, 16, int16_t, 16);

            int32_t qp = t->qp_y;
            int32_t i, j;
            int16_t* curdct;
            // we will count coeff cost, from jm80
            int32_t run, k;
            int32_t coeff_cost, total_cost;

            total_cost = 0;

            t->expand8to16sub(ref, 16 / 4, 16 / 4, dct, t->mb.src_y, t->stride);
            for(i = 0 ; i < 4 ; i ++)
            {
                coeff_cost = 0;
                for(j = 0 ; j < 4 ; j ++)
                {
                    int32_t idx = 4 * i + j;
                    int32_t idx_r = luma_index[idx];
                    curdct = dct + 16 * idx_r;

                    t->fdct4x4(curdct);

                    t->quant4x4(curdct, qp, FALSE);
                    scan_zig_4x4(t->mb.dct_y_z[idx], curdct);
                    {
                        run = -1;
                        for(k = 0 ; k < 16 ; k ++)
                        {
                            run ++;
                            if (t->mb.dct_y_z[idx][k] != 0)
                            {
                                if (ABS(t->mb.dct_y_z[idx][k]) > 1)
                                {
                                    return FALSE;
                                }
                                else
                                {
                                    coeff_cost += COEFF_COST[run];
                                    run = -1;
                                }
                            }
                        }
                    }
                }
                if (coeff_cost <= t->param.luma_coeff_cost)
                {
                    int32_t idx_r = luma_index[4 * i];
                    memset(t->mb.dct_y_z[4 * i], 0, 16 * sizeof(int16_t));
                    memset(dct + 16 * idx_r, 0, 16 * sizeof(int16_t));

                    idx_r = luma_index[4 * i + 1];
                    memset(t->mb.dct_y_z[4 * i + 1], 0, 16 * sizeof(int16_t));
                    memset(dct + 16 * idx_r, 0, 16 * sizeof(int16_t));

                    idx_r = luma_index[4 * i + 2];
                    memset(t->mb.dct_y_z[4 * i + 2], 0, 16 * sizeof(int16_t));
                    memset(dct + 16 * idx_r, 0, 16 * sizeof(int16_t));

                    idx_r = luma_index[4 * i + 3];
                    memset(t->mb.dct_y_z[4 * i + 3], 0, 16 * sizeof(int16_t));
                    memset(dct + 16 * idx_r, 0, 16 * sizeof(int16_t));
                    coeff_cost = 0;
                }
                else
                {
                    total_cost += coeff_cost;
                    if (total_cost > 5)
                        return FALSE;
                }
            }

            memset(dct, 0, 16 * 16 * sizeof(int16_t));
            memset(t->mb.dct_y_z, 0, sizeof(int16_t) * 16 * 16);

            t->contract16to8add(dct, 16 / 4, 16 / 4, ref, t->mb.dst_y, t->edged_stride);
        }

        {
            DECLARE_ALIGNED_MATRIX(pred_u, 8, 8, uint8_t, CACHE_SIZE);
            DECLARE_ALIGNED_MATRIX(pred_v, 8, 8, uint8_t, CACHE_SIZE);
            DECLARE_ALIGNED_MATRIX(dct, 10, 8, int16_t, CACHE_SIZE);

            int32_t qp = t->qp_uv;
            int16_t* curdct;
            uint8_t* start;
            uint8_t* dst;

            uint8_t* src;

            // we will count coeff cost, from jm80
            int32_t run, k;
            int32_t coeff_cost;

            src = t->ref[0][vec.refno]->U + ((t->mb.mb_y << 3) + (vec.y >> 3)) * t->edged_stride_uv + (t->mb.mb_x << 3) + (vec.x >> 3);
            t->eighth_pixel_mc_u(src, t->edged_stride_uv, pred_u, vec.x, vec.y, 8, 8);
            src = t->ref[0][vec.refno]->V + ((t->mb.mb_y << 3) + (vec.y >> 3)) * t->edged_stride_uv + (t->mb.mb_x << 3) + (vec.x >> 3);
            t->eighth_pixel_mc_u(src, t->edged_stride_uv, pred_v, vec.x, vec.y, 8, 8);

            start = pred_u;
            src   = t->mb.src_u;
            dst   = t->mb.dst_u;
            for(j = 0 ; j < 2 ; j ++)
            {
                coeff_cost = 0;

                t->expand8to16sub(start, 8 / 4, 8 / 4, dct, src, t->stride_uv);
                curdct = dct;
                for(i = 0 ; i < 4 ; i ++)
                {
                    run = -1;

                    t->fdct4x4(curdct);
                    dct[64 + i] = curdct[0];

                    t->quant4x4(curdct, qp, FALSE);
                    scan_zig_4x4(t->mb.dct_uv_z[j][i], curdct);
                    {
                        for(k = 1 ; k < 16 ; k ++)
                        {
                            run ++;
                            if (t->mb.dct_uv_z[j][i][k] != 0)
                            {
                                if (ABS(t->mb.dct_uv_z[j][i][k]) > 1)
                                {
                                    coeff_cost += 16 * 16 * 256;
                                    return FALSE;
                                }
                                else
                                {
                                    coeff_cost += COEFF_COST[run];
                                    run = -1;
                                }
                            }
                        }
                    }

                    curdct += 16;
                }
                if (coeff_cost < CHROMA_COEFF_COST)
                {
                    memset(&t->mb.dct_uv_z[j][0][0], 0, 4 * 16 * sizeof(int16_t));
                    memset(dct, 0, 8 * 8 * sizeof(int16_t));
                }
                else
                {
                    return FALSE;
                }
                t->fdct2x2dc(curdct);
                t->quant2x2dc(curdct, qp, FALSE);
                scan_zig_2x2(t->mb.dc2x2_z[j], curdct);
                if (array_non_zero_count(t->mb.dc2x2_z[j], 4) != 0)
                {
                    return FALSE;
                }

                t->iquant2x2dc(curdct, qp);
                t->idct2x2dc(curdct);

                curdct = dct;
                for(i = 0 ; i < 4 ; i ++)
                {
                    curdct[0] = dct[64 + i];
                    t->idct4x4(curdct);
                    curdct += 16;
                }
                t->contract16to8add(dct, 8 / 4, 8 / 4, start, dst, t->edged_stride_uv);

                //
                // change to v
                //
                start = pred_v;
                dst   = t->mb.dst_v;
                src   = t->mb.src_v;
            }
        }

        t->mb.mb_mode = P_SKIP;
        t->mb.mb_part = MB_16x16;
        memcpy(t->mb.pred_p16x16, ref, sizeof(uint8_t) * 16 * 16);
        copy_nvec(&vec, &t->mb.vec[0][0], 4, 4, 4);

        return TRUE;
    }
    
    return FALSE;
}

uint32_t
T264_mode_decision_interp_y(_RW T264_t* t)
{
    uint32_t sad;
    uint32_t sad_min = -1;
    uint8_t best_mode;
    uint8_t sub_part[4];
    uint8_t part;
    int32_t i, n;
    int32_t preds[9];	
    int32_t modes;
    search_data_t s0;
    subpart_search_data_t s1;

    typedef uint32_t (*p16x16_function_t)(T264_t*, search_data_t* s);
    static const p16x16_function_t p16x16_function[] = 
    {
        T264_mode_decision_inter_16x16p,
        T264_mode_decision_inter_16x8p,
        T264_mode_decision_inter_8x16p
    };

    // xxx
#ifdef USE_PREV_DETECT
    uint32_t sad_median;    // for skip detect
    sad_median = T264_predict_sad(t, 0);
    // p skip detection
    if (T264_detect_pskip(t, sad_median))
        return t->mb.sad;
#endif

    T264_inter_p16x16_mode_available(t, preds, &modes);

    best_mode = P_MODE;
    s0.list_index = 0;
    s1.list_index = 0;
	
    for(n = 0 ; n < modes ; n ++)
    {
        int32_t mode = preds[n];
        sad = p16x16_function[mode](t, &s0);

        if (sad < sad_min)
        {
			part = mode;
            sad_min = sad;
        }
    }

    if (t->flags & USE_SUBBLOCK)
    {
        uint32_t sub_sad_all = 0;

        typedef uint32_t (*p8x8_function_t)(T264_t*, int32_t, subpart_search_data_t* s);
        static const p8x8_function_t p8x8_function[] = 
        {
            T264_mode_decision_inter_8x8p,
            T264_mode_decision_inter_8x8p,
            T264_mode_decision_inter_8x4p,
            T264_mode_decision_inter_4x8p,
            T264_mode_decision_inter_4x4p
        };
        s1.vec[0][0].refno = s0.vec[0].refno;

        for(i = 0 ; i < 4 ; i ++)
        {
            uint32_t sub_sad;
            uint32_t sub_sad_min = -1;

            T264_inter_p8x8_mode_available(t, preds, &modes, i);

            for(n = 0 ; n < modes ; n ++)
            {
                int32_t mode = preds[n];
                T264_vector_t vec_bak[4];

                vec_bak[0] = t->mb.vec_ref[VEC_LUMA + i / 2 * 16 + i % 2 * 2 + 0].vec[0];
                vec_bak[1] = t->mb.vec_ref[VEC_LUMA + i / 2 * 16 + i % 2 * 2 + 1].vec[0];
                vec_bak[2] = t->mb.vec_ref[VEC_LUMA + i / 2 * 16 + i % 2 * 2 + 8].vec[0];
                vec_bak[3] = t->mb.vec_ref[VEC_LUMA + i / 2 * 16 + i % 2 * 2 + 9].vec[0];

                sub_sad = p8x8_function[mode - MB_8x8](t, i, &s1);

                if (sub_sad < sub_sad_min)
                {
                    sub_part[i] = mode;
                    sub_sad_min = sub_sad;
                }
                else
                {
                    // restore current best mode
                    t->mb.vec_ref[VEC_LUMA + i / 2 * 16 + i % 2 * 2 + 0].vec[0] = vec_bak[0];
                    t->mb.vec_ref[VEC_LUMA + i / 2 * 16 + i % 2 * 2 + 1].vec[0] = vec_bak[1];
                    t->mb.vec_ref[VEC_LUMA + i / 2 * 16 + i % 2 * 2 + 8].vec[0] = vec_bak[2];
                    t->mb.vec_ref[VEC_LUMA + i / 2 * 16 + i % 2 * 2 + 9].vec[0] = vec_bak[3];
                }
            }

            sub_sad_all += sub_sad_min;
        }

        if (sub_sad_all < sad_min)
        {
            part = MB_8x8;
            sad_min = sub_sad_all;
        }
    }

    switch (part)
    {
    case MB_16x8:
        sad_min  = T264_quarter_pixel_search(t, 0, s0.src[1], s0.ref[1], s0.offset[1], &s0.vec[1], &s0.vec_median[1], s0.sad[1], 16, 8, t->mb.pred_p16x16, MB_16x8);
        sad_min += T264_quarter_pixel_search(t, 0, s0.src[2], s0.ref[2], s0.offset[2], &s0.vec[2], &s0.vec_median[2], s0.sad[2], 16, 8, t->mb.pred_p16x16 + 16 * 8, MB_16x8);
        copy_nvec(&s0.vec[1], &t->mb.vec[0][0], 4, 2, 4);
        copy_nvec(&s0.vec[2], &t->mb.vec[0][8], 4, 2, 4);
        break;
    case MB_8x16:
        sad_min  = T264_quarter_pixel_search(t, 0, s0.src[3], s0.ref[3], s0.offset[3], &s0.vec[3], &s0.vec_median[3], s0.sad[3], 8, 16, t->mb.pred_p16x16, MB_8x16);
        sad_min += T264_quarter_pixel_search(t, 0, s0.src[4], s0.ref[4], s0.offset[4], &s0.vec[4], &s0.vec_median[4], s0.sad[4], 8, 16, t->mb.pred_p16x16 + 8, MB_8x16);
        copy_nvec(&s0.vec[3], &t->mb.vec[0][0], 2, 4, 4);
        copy_nvec(&s0.vec[4], &t->mb.vec[0][2], 2, 4, 4);
        break;
    case MB_8x8:
    case MB_8x8ref0:
        sad_min = 0;
        for(i = 0 ; i < 4 ; i ++)
        {
            switch(sub_part[i])
            {
            case MB_8x8:
                sad_min += T264_quarter_pixel_search(t, 0, s1.src[i][0], s1.ref[i][0], s1.offset[i][0], &s1.vec[i][0], &s1.vec_median[i][0], s1.sad[i][0], 8, 8, t->mb.pred_p16x16 + i / 2 * 16 * 8 + i % 2 * 8, MB_8x8);
                t->mb.vec[0][i / 2 * 8 + i % 2 * 2 + 0] = s1.vec[i][0];
                t->mb.vec[0][i / 2 * 8 + i % 2 * 2 + 1] = s1.vec[i][0];
                t->mb.vec[0][i / 2 * 8 + i % 2 * 2 + 4] = s1.vec[i][0];
                t->mb.vec[0][i / 2 * 8 + i % 2 * 2 + 5] = s1.vec[i][0];
                sad_min += eg_size_ue(t->bs, MB_8x8);
        	    break;
            case MB_8x4:
                sad_min += T264_quarter_pixel_search(t, 0, s1.src[i][1], s1.ref[i][1], s1.offset[i][1], &s1.vec[i][1], &s1.vec_median[i][1], s1.sad[i][1], 8, 4, t->mb.pred_p16x16 + i / 2 * 16 * 8 + i % 2 * 8, MB_8x4);
                sad_min += T264_quarter_pixel_search(t, 0, s1.src[i][2], s1.ref[i][2], s1.offset[i][2], &s1.vec[i][2], &s1.vec_median[i][2], s1.sad[i][2], 8, 4, t->mb.pred_p16x16 + i / 2 * 16 * 8 + i % 2 * 8 + 16 * 4, MB_8x4);
                t->mb.vec[0][i / 2 * 8 + i % 2 * 2 + 0] = s1.vec[i][1];
                t->mb.vec[0][i / 2 * 8 + i % 2 * 2 + 1] = s1.vec[i][1];
                t->mb.vec[0][i / 2 * 8 + i % 2 * 2 + 4] = s1.vec[i][2];
                t->mb.vec[0][i / 2 * 8 + i % 2 * 2 + 5] = s1.vec[i][2];
                sad_min += eg_size_ue(t->bs, MB_8x4);
        	    break;
            case MB_4x8:
                sad_min += T264_quarter_pixel_search(t, 0, s1.src[i][3], s1.ref[i][3], s1.offset[i][3], &s1.vec[i][3], &s1.vec[i][3], s1.sad[i][3], 4, 8, t->mb.pred_p16x16 + i / 2 * 16 * 8 + i % 2 * 8, MB_4x8);
                sad_min += T264_quarter_pixel_search(t, 0, s1.src[i][4], s1.ref[i][4], s1.offset[i][4], &s1.vec[i][4], &s1.vec[i][4], s1.sad[i][4], 4, 8, t->mb.pred_p16x16 + i / 2 * 16 * 8 + i % 2 * 8 + 4, MB_4x8);
                t->mb.vec[0][i / 2 * 8 + i % 2 * 2 + 0] = s1.vec[i][3];
                t->mb.vec[0][i / 2 * 8 + i % 2 * 2 + 1] = s1.vec[i][4];
                t->mb.vec[0][i / 2 * 8 + i % 2 * 2 + 4] = s1.vec[i][3];
                t->mb.vec[0][i / 2 * 8 + i % 2 * 2 + 5] = s1.vec[i][4];
                sad_min += eg_size_ue(t->bs, MB_4x8);
                break;
            case MB_4x4:
                sad_min += T264_quarter_pixel_search(t, 0, s1.src[i][5], s1.ref[i][5], s1.offset[i][5], &s1.vec[i][5], &s1.vec[i][5], s1.sad[i][5], 4, 4, t->mb.pred_p16x16 + i / 2 * 16 * 8 + i % 2 * 8, MB_4x4);
                sad_min += T264_quarter_pixel_search(t, 0, s1.src[i][6], s1.ref[i][6], s1.offset[i][6], &s1.vec[i][6], &s1.vec[i][6], s1.sad[i][6], 4, 4, t->mb.pred_p16x16 + i / 2 * 16 * 8 + i % 2 * 8 + 4, MB_4x4);
                sad_min += T264_quarter_pixel_search(t, 0, s1.src[i][7], s1.ref[i][7], s1.offset[i][7], &s1.vec[i][7], &s1.vec[i][7], s1.sad[i][7], 4, 4, t->mb.pred_p16x16 + i / 2 * 16 * 8 + i % 2 * 8 + 16 * 4, MB_4x4);
                sad_min += T264_quarter_pixel_search(t, 0, s1.src[i][8], s1.ref[i][8], s1.offset[i][8], &s1.vec[i][8], &s1.vec[i][8], s1.sad[i][8], 4, 4, t->mb.pred_p16x16 + i / 2 * 16 * 8 + i % 2 * 8 + 16 * 4 + 4, MB_4x4);
                t->mb.vec[0][i / 2 * 8 + i % 2 * 2 + 0] = s1.vec[i][5];
                t->mb.vec[0][i / 2 * 8 + i % 2 * 2 + 1] = s1.vec[i][6];
                t->mb.vec[0][i / 2 * 8 + i % 2 * 2 + 4] = s1.vec[i][7];
                t->mb.vec[0][i / 2 * 8 + i % 2 * 2 + 5] = s1.vec[i][8];
                sad_min += eg_size_ue(t->bs, MB_4x4);
                break;
            default:
                break;
            }
            t->mb.submb_part[i / 2 * 8 + i % 2 * 2 + 0] = sub_part[i];
            t->mb.submb_part[i / 2 * 8 + i % 2 * 2 + 1] = sub_part[i];
            t->mb.submb_part[i / 2 * 8 + i % 2 * 2 + 4] = sub_part[i];
            t->mb.submb_part[i / 2 * 8 + i % 2 * 2 + 5] = sub_part[i];
        }
    	break;
    default:
        break;
    }

    // 3ks chenm
    if (t->flags & USE_INTRAININTER)
        sad = T264_mode_decision_intra_y(t);
    else
        sad = -1;
    if (sad <= sad_min)
    {
        best_mode = t->mb.mb_mode;
        sad_min = sad;
    }
    else