/*****************************************************************************
 * me.c: h264 encoder library (Motion Estimation)
 *****************************************************************************
 * Copyright (C) 2003-2008 x264 project
 *
 * Authors: Loren Merritt <lorenm@u.washington.edu>
 *          Laurent Aimar <fenrir@via.ecp.fr>
 *          Jason Garrett-Glaser <darkshikari@gmail.com>
 *
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation; either version 2 of the License, or
 * (at your option) any later version.
 *
 * This program 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 General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02111, USA.
 *****************************************************************************/

#include "common/common.h"
#include "me.h"

/* presets selected from good points on the speed-vs-quality curve of several test videos
 * subpel_iters[i_subpel_refine] = { refine_hpel, refine_qpel, me_hpel, me_qpel }
 * where me_* are the number of EPZS iterations run on all candidate block types,
 * and refine_* are run only on the winner.
 * the subme=8,9 values are much higher because any amount of satd search makes
 * up its time by reducing the number of qpel-rd iterations. */
static const int subpel_iterations[][4] =
   {{0,0,0,0},
    {1,1,0,0},
    {0,1,1,0},
    {0,2,1,0},
    {0,2,1,1},
    {0,2,1,2},
    {0,0,2,2},
    {0,0,2,2},
    {0,0,4,10},
    {0,0,4,10}};

/* (x-1)%6 */
static const int mod6m1[8] = {5,0,1,2,3,4,5,0};
/* radius 2 hexagon. repeated entries are to avoid having to compute mod6 every time. */
static const int hex2[8][2] = {{-1,-2}, {-2,0}, {-1,2}, {1,2}, {2,0}, {1,-2}, {-1,-2}, {-2,0}};
static const int square1[8][2] = {{0,-1}, {0,1}, {-1,0}, {1,0}, {-1,-1}, {1,1}, {-1,1}, {1,-1}};

static void refine_subpel( x264_t *h, x264_me_t *m, int hpel_iters, int qpel_iters, int *p_halfpel_thresh, int b_refine_qpel );

#define BITS_MVD( mx, my )\
    (p_cost_mvx[(mx)<<2] + p_cost_mvy[(my)<<2])

#define COST_MV( mx, my )\
{\
    int cost = h->pixf.fpelcmp[i_pixel]( m->p_fenc[0], FENC_STRIDE,\
                   &p_fref[(my)*m->i_stride[0]+(mx)], m->i_stride[0] )\
             + BITS_MVD(mx,my);\
    COPY3_IF_LT( bcost, cost, bmx, mx, bmy, my );\
}

#define COST_MV_HPEL( mx, my ) \
{ \
    int stride = 16; \
    uint8_t *src = h->mc.get_ref( pix, &stride, m->p_fref, m->i_stride[0], mx, my, bw, bh ); \
    int cost = h->pixf.fpelcmp[i_pixel]( m->p_fenc[0], FENC_STRIDE, src, stride ) \
             + p_cost_mvx[ mx ] + p_cost_mvy[ my ]; \
    COPY3_IF_LT( bpred_cost, cost, bpred_mx, mx, bpred_my, my ); \
}

#define COST_MV_X3_DIR( m0x, m0y, m1x, m1y, m2x, m2y, costs )\
{\
    uint8_t *pix_base = p_fref + bmx + bmy*m->i_stride[0];\
    h->pixf.fpelcmp_x3[i_pixel]( m->p_fenc[0],\
        pix_base + (m0x) + (m0y)*m->i_stride[0],\
        pix_base + (m1x) + (m1y)*m->i_stride[0],\
        pix_base + (m2x) + (m2y)*m->i_stride[0],\
        m->i_stride[0], costs );\
    (costs)[0] += BITS_MVD( bmx+(m0x), bmy+(m0y) );\
    (costs)[1] += BITS_MVD( bmx+(m1x), bmy+(m1y) );\
    (costs)[2] += BITS_MVD( bmx+(m2x), bmy+(m2y) );\
}

#define COST_MV_X4( m0x, m0y, m1x, m1y, m2x, m2y, m3x, m3y )\
{\
    uint8_t *pix_base = p_fref + omx + omy*m->i_stride[0];\
    h->pixf.fpelcmp_x4[i_pixel]( m->p_fenc[0],\
        pix_base + (m0x) + (m0y)*m->i_stride[0],\
        pix_base + (m1x) + (m1y)*m->i_stride[0],\
        pix_base + (m2x) + (m2y)*m->i_stride[0],\
        pix_base + (m3x) + (m3y)*m->i_stride[0],\
        m->i_stride[0], costs );\
    costs[0] += BITS_MVD( omx+(m0x), omy+(m0y) );\
    costs[1] += BITS_MVD( omx+(m1x), omy+(m1y) );\
    costs[2] += BITS_MVD( omx+(m2x), omy+(m2y) );\
    costs[3] += BITS_MVD( omx+(m3x), omy+(m3y) );\
    COPY3_IF_LT( bcost, costs[0], bmx, omx+(m0x), bmy, omy+(m0y) );\
    COPY3_IF_LT( bcost, costs[1], bmx, omx+(m1x), bmy, omy+(m1y) );\
    COPY3_IF_LT( bcost, costs[2], bmx, omx+(m2x), bmy, omy+(m2y) );\
    COPY3_IF_LT( bcost, costs[3], bmx, omx+(m3x), bmy, omy+(m3y) );\
}

#define COST_MV_X3_ABS( m0x, m0y, m1x, m1y, m2x, m2y )\
{\
    h->pixf.fpelcmp_x3[i_pixel]( m->p_fenc[0],\
        p_fref + (m0x) + (m0y)*m->i_stride[0],\
        p_fref + (m1x) + (m1y)*m->i_stride[0],\
        p_fref + (m2x) + (m2y)*m->i_stride[0],\
        m->i_stride[0], costs );\
    costs[0] += p_cost_mvx[(m0x)<<2]; /* no cost_mvy */\
    costs[1] += p_cost_mvx[(m1x)<<2];\
    costs[2] += p_cost_mvx[(m2x)<<2];\
    COPY3_IF_LT( bcost, costs[0], bmx, m0x, bmy, m0y );\
    COPY3_IF_LT( bcost, costs[1], bmx, m1x, bmy, m1y );\
    COPY3_IF_LT( bcost, costs[2], bmx, m2x, bmy, m2y );\
}

/*  1  */
/* 101 */
/*  1  */
#define DIA1_ITER( mx, my )\
{\
    omx = mx; omy = my;\
    COST_MV_X4( 0,-1, 0,1, -1,0, 1,0 );\
}

#define CROSS( start, x_max, y_max )\
{\
    i = start;\
    if( x_max <= X264_MIN(mv_x_max-omx, omx-mv_x_min) )\
        for( ; i < x_max-2; i+=4 )\
            COST_MV_X4( i,0, -i,0, i+2,0, -i-2,0 );\
    for( ; i < x_max; i+=2 )\
    {\
        if( omx+i <= mv_x_max )\
            COST_MV( omx+i, omy );\
        if( omx-i >= mv_x_min )\
            COST_MV( omx-i, omy );\
    }\
    i = start;\
    if( y_max <= X264_MIN(mv_y_max-omy, omy-mv_y_min) )\
        for( ; i < y_max-2; i+=4 )\
            COST_MV_X4( 0,i, 0,-i, 0,i+2, 0,-i-2 );\
    for( ; i < y_max; i+=2 )\
    {\
        if( omy+i <= mv_y_max )\
            COST_MV( omx, omy+i );\
        if( omy-i >= mv_y_min )\
            COST_MV( omx, omy-i );\
    }\
}

void x264_me_search_ref( x264_t *h, x264_me_t *m, int16_t (*mvc)[2], int i_mvc, int *p_halfpel_thresh )
{
    const int bw = x264_pixel_size[m->i_pixel].w;
    const int bh = x264_pixel_size[m->i_pixel].h;
    const int i_pixel = m->i_pixel;
    int i_me_range = h->param.analyse.i_me_range;
    int bmx, bmy, bcost;
    int bpred_mx = 0, bpred_my = 0, bpred_cost = COST_MAX;
    int omx, omy, pmx, pmy;
    uint8_t *p_fref = m->p_fref[0];
    DECLARE_ALIGNED_16( uint8_t pix[16*16] );

    int i, j;
    int dir;
    int costs[6];

    int mv_x_min = h->mb.mv_min_fpel[0];
    int mv_y_min = h->mb.mv_min_fpel[1];
    int mv_x_max = h->mb.mv_max_fpel[0];
    int mv_y_max = h->mb.mv_max_fpel[1];

#define CHECK_MVRANGE(mx,my) ( mx >= mv_x_min && mx <= mv_x_max && my >= mv_y_min && my <= mv_y_max )

    const int16_t *p_cost_mvx = m->p_cost_mv - m->mvp[0];
    const int16_t *p_cost_mvy = m->p_cost_mv - m->mvp[1];

    bmx = x264_clip3( m->mvp[0], mv_x_min*4, mv_x_max*4 );
    bmy = x264_clip3( m->mvp[1], mv_y_min*4, mv_y_max*4 );
    pmx = ( bmx + 2 ) >> 2;
    pmy = ( bmy + 2 ) >> 2;
    bcost = COST_MAX;

    /* try extra predictors if provided */
    if( h->mb.i_subpel_refine >= 3 )
    {
        uint32_t bmv = pack16to32_mask(bmx,bmy);
        COST_MV_HPEL( bmx, bmy );
        for( i = 0; i < i_mvc; i++ )
        {
            if( *(uint32_t*)mvc[i] && (bmv - *(uint32_t*)mvc[i]) )
            {
                int mx = x264_clip3( mvc[i][0], mv_x_min*4, mv_x_max*4 );
                int my = x264_clip3( mvc[i][1], mv_y_min*4, mv_y_max*4 );
                COST_MV_HPEL( mx, my );
            }
        }
        bmx = ( bpred_mx + 2 ) >> 2;
        bmy = ( bpred_my + 2 ) >> 2;
        COST_MV( bmx, bmy );
    }
    else
    {
        /* check the MVP */
        COST_MV( pmx, pmy );
        /* Because we are rounding the predicted motion vector to fullpel, there will be
         * an extra MV cost in 15 out of 16 cases.  However, when the predicted MV is
         * chosen as the best predictor, it is often the case that the subpel search will
         * result in a vector at or next to the predicted motion vector.  Therefore, it is
         * sensible to remove the cost of the MV from the rounded MVP to avoid unfairly
         * biasing against use of the predicted motion vector. */
        bcost -= BITS_MVD( pmx, pmy );
        for( i = 0; i < i_mvc; i++ )
        {
            int mx = (mvc[i][0] + 2) >> 2;
            int my = (mvc[i][1] + 2) >> 2;
            if( (mx | my) && ((mx-bmx) | (my-bmy)) )
            {
                mx = x264_clip3( mx, mv_x_min, mv_x_max );
                my = x264_clip3( my, mv_y_min, mv_y_max );
                COST_MV( mx, my );
            }
        }
    }
    COST_MV( 0, 0 );

    switch( h->mb.i_me_method )
    {
    case X264_ME_DIA:
        /* diamond search, radius 1 */
        i = 0;
        do
        {
            DIA1_ITER( bmx, bmy );
            if( (bmx == omx) & (bmy == omy) )
                break;
            if( !CHECK_MVRANGE(bmx, bmy) )
                break;
        } while( ++i < i_me_range );
        break;

    case X264_ME_HEX:
me_hex2:
        /* hexagon search, radius 2 */
#if 0
        for( i = 0; i < i_me_range/2; i++ )
        {
            omx = bmx; omy = bmy;
            COST_MV( omx-2, omy   );
            COST_MV( omx-1, omy+2 );
            COST_MV( omx+1, omy+2 );
            COST_MV( omx+2, omy   );
            COST_MV( omx+1, omy-2 );
            COST_MV( omx-1, omy-2 );
            if( bmx == omx && bmy == omy )
                break;
            if( !CHECK_MVRANGE(bmx, bmy) )
                break;
        }
#else
        /* equivalent to the above, but eliminates duplicate candidates */
        dir = -2;

        /* hexagon */
        COST_MV_X3_DIR( -2,0, -1, 2,  1, 2, costs   );
        COST_MV_X3_DIR(  2,0,  1,-2, -1,-2, costs+3 );
        COPY2_IF_LT( bcost, costs[0], dir, 0 );
        COPY2_IF_LT( bcost, costs[1], dir, 1 );
        COPY2_IF_LT( bcost, costs[2], dir, 2 );
        COPY2_IF_LT( bcost, costs[3], dir, 3 );
        COPY2_IF_LT( bcost, costs[4], dir, 4 );
        COPY2_IF_LT( bcost, costs[5], dir, 5 );

        if( dir != -2 )
        {
            bmx += hex2[dir+1][0];
            bmy += hex2[dir+1][1];
            /* half hexagon, not overlapping the previous iteration */
            for( i = 1; i < i_me_range/2 && CHECK_MVRANGE(bmx, bmy); i++ )
            {
                const int odir = mod6m1[dir+1];
                COST_MV_X3_DIR( hex2[odir+0][0], hex2[odir+0][1],
                                hex2[odir+1][0], hex2[odir+1][1],
                                hex2[odir+2][0], hex2[odir+2][1],
                                costs );
                dir = -2;
                COPY2_IF_LT( bcost, costs[0], dir, odir-1 );
                COPY2_IF_LT( bcost, costs[1], dir, odir   );
                COPY2_IF_LT( bcost, costs[2], dir, odir+1 );
                if( dir == -2 )
                    break;
                bmx += hex2[dir+1][0];
                bmy += hex2[dir+1][1];
            }
        }
#endif
        /* square refine */
        omx = bmx; omy = bmy;
        COST_MV_X4(  0,-1,  0,1, -1,0, 1,0 );
        COST_MV_X4( -1,-1, -1,1, 1,-1, 1,1 );
        break;

    case X264_ME_UMH:
        {
            /* Uneven-cross Multi-Hexagon-grid Search
             * as in JM, except with different early termination */

            static const int x264_pixel_size_shift[7] = { 0, 1, 1, 2, 3, 3, 4 };

            int ucost1, ucost2;
            int cross_start = 1;

            /* refine predictors */
            ucost1 = bcost;
            DIA1_ITER( pmx, pmy );
            if( pmx | pmy )
                DIA1_ITER( 0, 0 );

            if(i_pixel == PIXEL_4x4)
                goto me_hex2;

            ucost2 = bcost;
            if( (bmx | bmy) && ((bmx-pmx) | (bmy-pmy)) )
                DIA1_ITER( bmx, bmy );
            if( bcost == ucost2 )
                cross_start = 3;
            omx = bmx; omy = bmy;

            /* early termination */
#define SAD_THRESH(v) ( bcost < ( v >> x264_pixel_size_shift[i_pixel] ) )
            if( bcost == ucost2 && SAD_THRESH(2000) )
            {
                COST_MV_X4( 0,-2, -1,-1, 1,-1, -2,0 );
                COST_MV_X4( 2, 0, -1, 1, 1, 1,  0,2 );
                if( bcost == ucost1 && SAD_THRESH(500) )
                    break;
                if( bcost == ucost2 )
                {
                    int range = (i_me_range>>1) | 1;
                    CROSS( 3, range, range );
                    COST_MV_X4( -1,-2, 1,-2, -2,-1, 2,-1 );
                    COST_MV_X4( -2, 1, 2, 1, -1, 2, 1, 2 );
                    if( bcost == ucost2 )
                        break;
                    cross_start = range + 2;
                }
            }

            /* adaptive search range */
            if( i_mvc )
            {
                /* range multipliers based on casual inspection of some statistics of
                 * average distance between current predictor and final mv found by ESA.
                 * these have not been tuned much by actual encoding. */
                static const int range_mul[4][4] =
                {
                    { 3, 3, 4, 4 },
                    { 3, 4, 4, 4 },
                    { 4, 4, 4, 5 },
                    { 4, 4, 5, 6 },
                };
                int mvd;
                int sad_ctx, mvd_ctx;
                int denom = 1;

                if( i_mvc == 1 )
                {
                    if( i_pixel == PIXEL_16x16 )
                        /* mvc is probably the same as mvp, so the difference isn't meaningful.
                         * but prediction usually isn't too bad, so just use medium range */
                        mvd = 25;
                    else
                        mvd = abs( m->mvp[0] - mvc[0][0] )
                            + abs( m->mvp[1] - mvc[0][1] );
                }
                else
                {
                    /* calculate the degree of agreement between predictors. */
                    /* in 16x16, mvc includes all the neighbors used to make mvp,
                     * so don't count mvp separately. */
                    denom = i_mvc - 1;
                    mvd = 0;
                    if( i_pixel != PIXEL_16x16 )
                    {
                        mvd = abs( m->mvp[0] - mvc[0][0] )
                            + abs( m->mvp[1] - mvc[0][1] );
                        denom++;
                    }
                    mvd += x264_predictor_difference( mvc, i_mvc );
                }

                sad_ctx = SAD_THRESH(1000) ? 0
                        : SAD_THRESH(2000) ? 1
                        : SAD_THRESH(4000) ? 2 : 3;
                mvd_ctx = mvd < 10*denom ? 0
                        : mvd < 20*denom ? 1
                        : mvd < 40*denom ? 2 : 3;

                i_me_range = i_me_range * range_mul[mvd_ctx][sad_ctx] / 4;
            }

            /* FIXME if the above DIA2/OCT2/CROSS found a new mv, it has not updated omx/omy.
             * we are still centered on the same place as the DIA2. is this desirable? */
            CROSS( cross_start, i_me_range, i_me_range/2 );

            COST_MV_X4( -2,-2, -2,2, 2,-2, 2,2 );

            /* hexagon grid */
            omx = bmx; omy = bmy;

            i = 1;
            do
            {
                static const int hex4[16][2] = {
                    {-4, 2}, {-4, 1}, {-4, 0}, {-4,-1}, {-4,-2},
                    { 4,-2}, { 4,-1}, { 4, 0}, { 4, 1}, { 4, 2},
                    { 2, 3}, { 0, 4}, {-2, 3},
                    {-2,-3}, { 0,-4}, { 2,-3},
                };

                if( 4*i > X264_MIN4( mv_x_max-omx, omx-mv_x_min,
                                     mv_y_max-omy, omy-mv_y_min ) )
                {
                    for( j = 0; j < 16; j++ )
                    {
                        int mx = omx + hex4[j][0]*i;
                        int my = omy + hex4[j][1]*i;
                        if( CHECK_MVRANGE(mx, my) )
                            COST_MV( mx, my );
                    }
                }
                else
                {
                    COST_MV_X4( -4*i, 2*i, -4*i, 1*i, -4*i, 0*i, -4*i,-1*i );
                    COST_MV_X4( -4*i,-2*i,  4*i,-2*i,  4*i,-1*i,  4*i, 0*i );
                    COST_MV_X4(  4*i, 1*i,  4*i, 2*i,  2*i, 3*i,  0*i, 4*i );
                    COST_MV_X4( -2*i, 3*i, -2*i,-3*i,  0*i,-4*i,  2*i,-3*i );
                }
            } while( ++i <= i_me_range/4 );
            if( bmy <= mv_y_max )
                goto me_hex2;
            break;
        }

    case X264_ME_ESA:
    case X264_ME_TESA:
        {
            const int min_x = X264_MAX( bmx - i_me_range, mv_x_min );
            const int min_y = X264_MAX( bmy - i_me_range, mv_y_min );
            const int max_x = X264_MIN( bmx + i_me_range, mv_x_max );
            const int max_y = X264_MIN( bmy + i_me_range, mv_y_max );
            /* SEA is fastest in multiples of 4 */
            const int width = (max_x - min_x + 3) & ~3;
            int my;
#if 0
            /* plain old exhaustive search */
            int mx;
            for( my = min_y; my <= max_y; my++ )
                for( mx = min_x; mx <= max_x; mx++ )
                    COST_MV( mx, my );
#else
            /* successive elimination by comparing DC before a full SAD,
             * because sum(abs(diff)) >= abs(diff(sum)). */
            const int stride = m->i_stride[0];
            uint16_t *sums_base = m->integral;
            /* due to a GCC bug on some platforms (win32?), zero[] may not actually be aligned.
             * unlike the similar case in ratecontrol.c, this is not a problem because it is not used for any
             * SSE instructions and the only loss is a tiny bit of performance. */
            DECLARE_ALIGNED_16( static uint8_t zero[8*FENC_STRIDE] );
            DECLARE_ALIGNED_16( int enc_dc[4] );
            int sad_size = i_pixel <= PIXEL_8x8 ? PIXEL_8x8 : PIXEL_4x4;
            int delta = x264_pixel_size[sad_size].w;
            int16_t xs_buf[64];
            int16_t *xs = width<=64 ? xs_buf : x264_malloc( (width+15)*sizeof(int16_t) );
            int xn;
            uint16_t *cost_fpel_mvx = x264_cost_mv_fpel[h->mb.i_qp][-m->mvp[0]&3] + (-m->mvp[0]>>2);

            h->pixf.sad_x4[sad_size]( zero, m->p_fenc[0], m->p_fenc[0]+delta,
                m->p_fenc[0]+delta*FENC_STRIDE, m->p_fenc[0]+delta+delta*FENC_STRIDE,
                FENC_STRIDE, enc_dc );
            if( delta == 4 )
                sums_base += stride * (h->fenc->i_lines[0] + PADV*2);
            if( i_pixel == PIXEL_16x16 || i_pixel == PIXEL_8x16 || i_pixel == PIXEL_4x8 )
                delta *= stride;
            if( i_pixel == PIXEL_8x16 || i_pixel == PIXEL_4x8 )
                enc_dc[1] = enc_dc[2];

            if( h->mb.i_me_method == X264_ME_TESA )
            {
                // ADS threshold, then SAD threshold, then keep the best few SADs, then SATD
                typedef struct {
                    int sad;
                    int16_t mx, my;
                } mvsad_t;
                mvsad_t *mvsads = x264_malloc( width*(max_y-min_y+1)*sizeof(mvsad_t) );
                int nmvsad = 0, limit;
                int sad_thresh = i_me_range <= 16 ? 10 : i_me_range <= 24 ? 11 : 12;
                int bsad = h->pixf.sad[i_pixel]( m->p_fenc[0], FENC_STRIDE, p_fref+bmy*stride+bmx, stride )
                         + BITS_MVD( bmx, bmy );
                for( my = min_y; my <= max_y; my++ )
                {
                    int ycost = p_cost_mvy[my<<2];
                    if( bsad <= ycost )
                        continue;
                    bsad -= ycost;
                    xn = h->pixf.ads[i_pixel]( enc_dc, sums_base + min_x + my * stride, delta,
                                               cost_fpel_mvx+min_x, xs, width, bsad*17/16 );
                    for( i=0; i<xn-2; i+=3 )
                    {
                        uint8_t *ref = p_fref+min_x+my*stride;
                        int sads[3];
                        h->pixf.sad_x3[i_pixel]( m->p_fenc[0], ref+xs[i], ref+xs[i+1], ref+xs[i+2], stride, sads );
                        for( j=0; j<3; j++ )
                        {
                            int sad = sads[j] + cost_fpel_mvx[xs[i+j]];
                            if( sad < bsad*sad_thresh>>3 )
                            {
                                COPY1_IF_LT( bsad, sad );
                                mvsads[nmvsad].sad = sad + ycost;
                                mvsads[nmvsad].mx = min_x+xs[i+j];
                                mvsads[nmvsad].my = my;
                                nmvsad++;
                            }
                        }
                    }
                    for( ; i<xn; i++ )
                    {
                        int mx = min_x+xs[i];
                        int sad = h->pixf.sad[i_pixel]( m->p_fenc[0], FENC_STRIDE, p_fref+mx+my*stride, stride )
                                + cost_fpel_mvx[xs[i]];
                        if( sad < bsad*sad_thresh>>3 )
                        {
                            COPY1_IF_LT( bsad, sad );
                            mvsads[nmvsad].sad = sad + ycost;
                            mvsads[nmvsad].mx = mx;
                            mvsads[nmvsad].my = my;
                            nmvsad++;
                        }
                    }
                    bsad += ycost;
                }

                limit = i_me_range / 2;
                if( nmvsad > limit*2 )
                {
                    // halve the range if the domain is too large... eh, close enough
                    bsad = bsad*(sad_thresh+8)>>4;
                    for( i=0; i<nmvsad && mvsads[i].sad <= bsad; i++ );
                    for( j=i; j<nmvsad; j++ )
                        if( mvsads[j].sad <= bsad )
                        {
                            /* mvsad_t is not guaranteed to be 8 bytes on all archs, so check before using explicit write-combining */
                            if( sizeof( mvsad_t ) == sizeof( uint64_t ) )
                                *(uint64_t*)&mvsads[i++] = *(uint64_t*)&mvsads[j];
                            else
                                mvsads[i++] = mvsads[j];
                        }
                    nmvsad = i;
                }
                if( nmvsad > limit )
                {
                    for( i=0; i<limit; i++ )
                    {
                        int bj = i;
                        int bsad = mvsads[bj].sad;
                        for( j=i+1; j<nmvsad; j++ )
                            COPY2_IF_LT( bsad, mvsads[j].sad, bj, j );
                        if( bj > i )
                        {
                            if( sizeof( mvsad_t ) == sizeof( uint64_t ) )
                                XCHG( uint64_t, *(uint64_t*)&mvsads[i], *(uint64_t*)&mvsads[bj] );
                            else
                                XCHG( mvsad_t, mvsads[i], mvsads[bj] );
                        }
                    }
                    nmvsad = limit;
                }
                for( i=0; i<nmvsad; i++ )
                    COST_MV( mvsads[i].mx, mvsads[i].my );
                x264_free( mvsads );
            }
            else
            {
                // just ADS and SAD
                for( my = min_y; my <= max_y; my++ )
                {
                    int ycost = p_cost_mvy[my<<2];
                    if( bcost <= ycost )
                        continue;
                    bcost -= ycost;
                    xn = h->pixf.ads[i_pixel]( enc_dc, sums_base + min_x + my * stride, delta,
                                               cost_fpel_mvx+min_x, xs, width, bcost );
                    for( i=0; i<xn-2; i+=3 )
                        COST_MV_X3_ABS( min_x+xs[i],my, min_x+xs[i+1],my, min_x+xs[i+2],my );
                    bcost += ycost;
                    for( ; i<xn; i++ )
                        COST_MV( min_x+xs[i], my );
                }
            }

            if( xs != xs_buf )
                x264_free( xs );
#endif
        }
        break;
    }

    /* -> qpel mv */
    if( bpred_cost < bcost )
    {
        m->mv[0] = bpred_mx;
        m->mv[1] = bpred_my;
        m->cost = bpred_cost;
    }
    else
    {
        m->mv[0] = bmx << 2;
        m->mv[1] = bmy << 2;
        m->cost = bcost;
    }

    /* compute the real cost */
    m->cost_mv = p_cost_mvx[ m->mv[0] ] + p_cost_mvy[ m->mv[1] ];
    if( bmx == pmx && bmy == pmy && h->mb.i_subpel_refine < 3 )
        m->cost += m->cost_mv;

    /* subpel refine */
    if( h->mb.i_subpel_refine >= 2 )
    {
        int hpel = subpel_iterations[h->mb.i_subpel_refine][2];
        int qpel = subpel_iterations[h->mb.i_subpel_refine][3];
        refine_subpel( h, m, hpel, qpel, p_halfpel_thresh, 0 );
    }
    else if( m->mv[1] > h->mb.mv_max_spel[1] )
        m->mv[1] = h->mb.mv_max_spel[1];
}
#undef COST_MV

void x264_me_refine_qpel( x264_t *h, x264_me_t *m )
{
    int hpel = subpel_iterations[h->mb.i_subpel_refine][0];
    int qpel = subpel_iterations[h->mb.i_subpel_refine][1];

    if( m->i_pixel <= PIXEL_8x8 && h->sh.i_type == SLICE_TYPE_P )
        m->cost -= m->i_ref_cost;
	
    refine_subpel( h, m, hpel, qpel, NULL, 1 );
}

#define COST_MV_SAD( mx, my ) \
{ \
    int stride = 16; \
    uint8_t *src = h->mc.get_ref( pix[0], &stride, m->p_fref, m->i_stride[0], mx, my, bw, bh ); \
    int cost = h->pixf.fpelcmp[i_pixel]( m->p_fenc[0], FENC_STRIDE, src, stride ) \
             + p_cost_mvx[ mx ] + p_cost_mvy[ my ]; \
    COPY3_IF_LT( bcost, cost, bmx, mx, bmy, my ); \
}

#define COST_MV_SATD( mx, my, dir ) \
if( b_refine_qpel || (dir^1) != odir ) \
{ \
    int stride = 16; \
    uint8_t *src = h->mc.get_ref( pix[0], &stride, m->p_fref, m->i_stride[0], mx, my, bw, bh ); \
    int cost = h->pixf.mbcmp_unaligned[i_pixel]( m->p_fenc[0], FENC_STRIDE, src, stride ) \
             + p_cost_mvx[ mx ] + p_cost_mvy[ my ]; \
    if( b_chroma_me && cost < bcost ) \
    { \
        h->mc.mc_chroma( pix[0], 8, m->p_fref[4], m->i_stride[1], mx, my, bw/2, bh/2 ); \
        cost += h->pixf.mbcmp[i_pixel+3]( m->p_fenc[1], FENC_STRIDE, pix[0], 8 ); \
        if( cost < bcost ) \
        { \
            h->mc.mc_chroma( pix[0], 8, m->p_fref[5], m->i_stride[1], mx, my, bw/2, bh/2 ); \
            cost += h->pixf.mbcmp[i_pixel+3]( m->p_fenc[2], FENC_STRIDE, pix[0], 8 ); \
        } \
    } \
    if( cost < bcost ) \
    {                  \
        bcost = cost;  \
        bmx = mx;      \
        bmy = my;      \
        bdir = dir;    \
    } \
}

static void refine_subpel( x264_t *h, x264_me_t *m, int hpel_iters, int qpel_iters, int *p_halfpel_thresh, int b_refine_qpel )
{
    const int bw = x264_pixel_size[m->i_pixel].w;
    const int bh = x264_pixel_size[m->i_pixel].h;
    const int16_t *p_cost_mvx = m->p_cost_mv - m->mvp[0];
    const int16_t *p_cost_mvy = m->p_cost_mv - m->mvp[1];
    const int i_pixel = m->i_pixel;
    const int b_chroma_me = h->mb.b_chroma_me && i_pixel <= PIXEL_8x8;

    DECLARE_ALIGNED_16( uint8_t pix[2][32*18] ); // really 17x17, but round up for alignment
    int omx, omy;
    int i;

    int bmx = m->mv[0];
    int bmy = m->mv[1];
    int bcost = m->cost;
    int odir = -1, bdir;

    /* try the subpel component of the predicted mv */
    if( hpel_iters && h->mb.i_subpel_refine < 3 )
    {
        int mx = x264_clip3( m->mvp[0], h->mb.mv_min_spel[0], h->mb.mv_max_spel[0] );
        int my = x264_clip3( m->mvp[1], h->mb.mv_min_spel[1], h->mb.mv_max_spel[1] );
        if( (mx-bmx)|(my-bmy) )
            COST_MV_SAD( mx, my );
    }

    /* halfpel diamond search */
    for( i = hpel_iters; i > 0; i-- )
    {
        int omx = bmx, omy = bmy;
        int costs[4];
        int stride = 32; // candidates are either all hpel or all qpel, so one stride is enough
        uint8_t *src0, *src1, *src2, *src3;
        src0 = h->mc.get_ref( pix[0], &stride, m->p_fref, m->i_stride[0], omx, omy-2, bw, bh+1 );
        src2 = h->mc.get_ref( pix[1], &stride, m->p_fref, m->i_stride[0], omx-2, omy, bw+4, bh );
        src1 = src0 + stride;
        src3 = src2 + 1;
        h->pixf.fpelcmp_x4[i_pixel]( m->p_fenc[0], src0, src1, src2, src3, stride, costs );
        COPY2_IF_LT( bcost, costs[0] + p_cost_mvx[omx  ] + p_cost_mvy[omy-2], bmy, omy-2 );
        COPY2_IF_LT( bcost, costs[1] + p_cost_mvx[omx  ] + p_cost_mvy[omy+2], bmy, omy+2 );
        COPY3_IF_LT( bcost, costs[2] + p_cost_mvx[omx-2] + p_cost_mvy[omy  ], bmx, omx-2, bmy, omy );
        COPY3_IF_LT( bcost, costs[3] + p_cost_mvx[omx+2] + p_cost_mvy[omy  ], bmx, omx+2, bmy, omy );
        if( (bmx == omx) & (bmy == omy) )
            break;
    }

    if( !b_refine_qpel )
    {
        /* check for mvrange */
        if( bmy > h->mb.mv_max_spel[1] )
            bmy = h->mb.mv_max_spel[1];
        bcost = COST_MAX;
        COST_MV_SATD( bmx, bmy, -1 );
    }

    /* early termination when examining multiple reference frames */
    if( p_halfpel_thresh )
    {
        if( (bcost*7)>>3 > *p_halfpel_thresh )
        {
            m->cost = bcost;
            m->mv[0] = bmx;
            m->mv[1] = bmy;
            // don't need cost_mv
            return;
        }
        else if( bcost < *p_halfpel_thresh )
            *p_halfpel_thresh = bcost;
    }

    /* quarterpel diamond search */
    bdir = -1;
    for( i = qpel_iters; i > 0; i-- )
    {
        odir = bdir;
        omx = bmx;
        omy = bmy;
        COST_MV_SATD( omx, omy - 1, 0 );
        COST_MV_SATD( omx, omy + 1, 1 );
        COST_MV_SATD( omx - 1, omy, 2 );
        COST_MV_SATD( omx + 1, omy, 3 );
        if( bmx == omx && bmy == omy )
            break;
    }

    /* check for mvrange */
    if( bmy > h->mb.mv_max_spel[1] )
    {
        bmy = h->mb.mv_max_spel[1];
        bcost = COST_MAX;
        COST_MV_SATD( bmx, bmy, -1 );
    }

    m->cost = bcost;
    m->mv[0] = bmx;
    m->mv[1] = bmy;
    m->cost_mv = p_cost_mvx[ bmx ] + p_cost_mvy[ bmy ];
}

#define BIME_CACHE( dx, dy ) \
{ \
    int i = 4 + 3*dx + dy; \
    stride0[i] = bw;\
    stride1[i] = bw;\
    src0[i] = h->mc.get_ref( pix0[i], &stride0[i], m0->p_fref, m0->i_stride[0], om0x+dx, om0y+dy, bw, bh ); \
    src1[i] = h->mc.get_ref( pix1[i], &stride1[i], m1->p_fref, m1->i_stride[0], om1x+dx, om1y+dy, bw, bh ); \
}

#define BIME_CACHE2(a,b) \
    BIME_CACHE(a,b) \
    BIME_CACHE(-(a),-(b))

#define SATD_THRESH 17/16

#define COST_BIMV_SATD( m0x, m0y, m1x, m1y ) \
if( pass == 0 || !((visited[(m0x)&7][(m0y)&7][(m1x)&7] & (1<<((m1y)&7)))) ) \
{ \
    int cost; \
    int i0 = 4 + 3*(m0x-om0x) + (m0y-om0y); \
    int i1 = 4 + 3*(m1x-om1x) + (m1y-om1y); \
    visited[(m0x)&7][(m0y)&7][(m1x)&7] |= (1<<((m1y)&7));\
    h->mc.avg[i_pixel]( pix, bw, src0[i0], stride0[i0], src1[i1], stride1[i1], i_weight ); \
    cost = h->pixf.mbcmp[i_pixel]( m0->p_fenc[0], FENC_STRIDE, pix, bw ) \
         + p_cost_m0x[ m0x ] + p_cost_m0y[ m0y ] \
         + p_cost_m1x[ m1x ] + p_cost_m1y[ m1y ]; \
    if( rd ) \
    { \
        if( cost < bcost * SATD_THRESH ) \
        { \
            uint64_t costrd; \
            if( cost < bcost ) \
                bcost = cost; \
            *(uint32_t*)cache0_mv = *(uint32_t*)cache0_mv2 = pack16to32_mask(m0x,m0y); \
            *(uint32_t*)cache1_mv = *(uint32_t*)cache1_mv2 = pack16to32_mask(m1x,m1y); \
            costrd = x264_rd_cost_part( h, i_lambda2, i8, m0->i_pixel ); \
            if( costrd < bcostrd ) \
            {\
                bcostrd = costrd;\
                bm0x = m0x;      \
                bm0y = m0y;      \
                bm1x = m1x;      \
                bm1y = m1y;      \
            }\
        } \
    } \
    else if( cost < bcost ) \
    {                  \
        bcost = cost;  \
        bm0x = m0x;    \
        bm0y = m0y;    \
        bm1x = m1x;    \
        bm1y = m1y;    \
    } \
}

#define CHECK_BIDIR(a,b,c,d) \
    COST_BIMV_SATD(om0x+a, om0y+b, om1x+c, om1y+d)

#define CHECK_BIDIR2(a,b,c,d) \
    CHECK_BIDIR(a,b,c,d) \
    CHECK_BIDIR(-(a),-(b),-(c),-(d))

#define CHECK_BIDIR8(a,b,c,d) \
    CHECK_BIDIR2(a,b,c,d) \
    CHECK_BIDIR2(b,c,d,a) \
    CHECK_BIDIR2(c,d,a,b) \
    CHECK_BIDIR2(d,a,b,c)

static void ALWAYS_INLINE x264_me_refine_bidir( x264_t *h, x264_me_t *m0, x264_me_t *m1, int i_weight, int i8, int i_lambda2, int rd )
{
    static const int pixel_mv_offs[] = { 0, 4, 4*8, 0 };
    int16_t *cache0_mv = h->mb.cache.mv[0][x264_scan8[i8*4]];
    int16_t *cache0_mv2 = cache0_mv + pixel_mv_offs[m0->i_pixel];
    int16_t *cache1_mv = h->mb.cache.mv[1][x264_scan8[i8*4]];
    int16_t *cache1_mv2 = cache1_mv + pixel_mv_offs[m0->i_pixel];
    const int i_pixel = m0->i_pixel;
    const int bw = x264_pixel_size[i_pixel].w;
    const int bh = x264_pixel_size[i_pixel].h;
    const int16_t *p_cost_m0x = m0->p_cost_mv - x264_clip3( m0->mvp[0], h->mb.mv_min_spel[0], h->mb.mv_max_spel[0] );
    const int16_t *p_cost_m0y = m0->p_cost_mv - x264_clip3( m0->mvp[1], h->mb.mv_min_spel[0], h->mb.mv_max_spel[0] );
    const int16_t *p_cost_m1x = m1->p_cost_mv - x264_clip3( m1->mvp[0], h->mb.mv_min_spel[0], h->mb.mv_max_spel[0] );
    const int16_t *p_cost_m1y = m1->p_cost_mv - x264_clip3( m1->mvp[1], h->mb.mv_min_spel[0], h->mb.mv_max_spel[0] );
    DECLARE_ALIGNED_16( uint8_t pix0[9][16*16] );
    DECLARE_ALIGNED_16( uint8_t pix1[9][16*16] );
    DECLARE_ALIGNED_16( uint8_t pix[16*16] );
    uint8_t *src0[9];
    uint8_t *src1[9];
    int stride0[9];
    int stride1[9];
    int bm0x = m0->mv[0], om0x = bm0x;
    int bm0y = m0->mv[1], om0y = bm0y;
    int bm1x = m1->mv[0], om1x = bm1x;
    int bm1y = m1->mv[1], om1y = bm1y;
    int bcost = COST_MAX;
    int pass = 0;
    uint64_t bcostrd = COST_MAX64;

    /* each byte of visited represents 8 possible m1y positions, so a 4D array isn't needed */
    DECLARE_ALIGNED_16( uint8_t visited[8][8][8] );

    if( bm0y > h->mb.mv_max_spel[1] - 8 ||
        bm1y > h->mb.mv_max_spel[1] - 8 )
        return;

    h->mc.memzero_aligned( visited, sizeof(visited) );

    BIME_CACHE( 0, 0 );
    CHECK_BIDIR( 0, 0, 0, 0 );

    for( pass = 0; pass < 8; pass++ )
    {
        /* check all mv pairs that differ in at most 2 components from the current mvs. */
        /* doesn't do chroma ME. this probably doesn't matter, as the gains
         * from bidir ME are the same with and without chroma ME. */

        BIME_CACHE2( 1, 0 );
        BIME_CACHE2( 0, 1 );
        BIME_CACHE2( 1, 1 );
        BIME_CACHE2( 1,-1 );

        CHECK_BIDIR8( 0, 0, 0, 1 );
        CHECK_BIDIR8( 0, 0, 1, 1 );
        CHECK_BIDIR2( 0, 1, 0, 1 );
        CHECK_BIDIR2( 1, 0, 1, 0 );
        CHECK_BIDIR8( 0, 0,-1, 1 );
        CHECK_BIDIR2( 0,-1, 0, 1 );
        CHECK_BIDIR2(-1, 0, 1, 0 );

        if( om0x == bm0x && om0y == bm0y && om1x == bm1x && om1y == bm1y )
            break;

        om0x = bm0x;
        om0y = bm0y;
        om1x = bm1x;
        om1y = bm1y;
        BIME_CACHE( 0, 0 );
    }

    m0->mv[0] = bm0x;
    m0->mv[1] = bm0y;
    m1->mv[0] = bm1x;
    m1->mv[1] = bm1y;
}

void x264_me_refine_bidir_satd( x264_t *h, x264_me_t *m0, x264_me_t *m1, int i_weight )
{
    x264_me_refine_bidir( h, m0, m1, i_weight, 0, 0, 0 );
}

void x264_me_refine_bidir_rd( x264_t *h, x264_me_t *m0, x264_me_t *m1, int i_weight, int i8, int i_lambda2 )
{
    x264_me_refine_bidir( h, m0, m1, i_weight, i8, i_lambda2, 1 );
}

#undef COST_MV_SATD
#define COST_MV_SATD( mx, my, dst, avoid_mvp ) \
{ \
    if( !avoid_mvp || !(mx == pmx && my == pmy) ) \
    { \
        int stride = 16; \
        uint8_t *src = h->mc.get_ref( pix, &stride, m->p_fref, m->i_stride[0], mx, my, bw*4, bh*4 ); \
        dst = h->pixf.mbcmp_unaligned[i_pixel]( m->p_fenc[0], FENC_STRIDE, src, stride ) \
            + p_cost_mvx[mx] + p_cost_mvy[my]; \
        COPY1_IF_LT( bsatd, dst ); \
    } \
    else \
        dst = COST_MAX; \
}

#define COST_MV_RD( mx, my, satd, do_dir, mdir ) \
{ \
    if( satd <= bsatd * SATD_THRESH ) \
    { \
        uint64_t cost; \
        *(uint32_t*)cache_mv = *(uint32_t*)cache_mv2 = pack16to32_mask(mx,my); \
        cost = x264_rd_cost_part( h, i_lambda2, i4, m->i_pixel ); \
        COPY4_IF_LT( bcost, cost, bmx, mx, bmy, my, dir, do_dir?mdir:dir ); \
    } \
}

void x264_me_refine_qpel_rd( x264_t *h, x264_me_t *m, int i_lambda2, int i4, int i_list )
{
    // don't have to fill the whole mv cache rectangle
    static const int pixel_mv_offs[] = { 0, 4, 4*8, 0, 2, 2*8, 0 };
    int16_t *cache_mv = h->mb.cache.mv[i_list][x264_scan8[i4]];
    int16_t *cache_mv2 = cache_mv + pixel_mv_offs[m->i_pixel];
    const int16_t *p_cost_mvx, *p_cost_mvy;
    const int bw = x264_pixel_size[m->i_pixel].w>>2;
    const int bh = x264_pixel_size[m->i_pixel].h>>2;
    const int i_pixel = m->i_pixel;

    DECLARE_ALIGNED_16( uint8_t pix[16*16] );
    uint64_t bcost = m->i_pixel == PIXEL_16x16 ? m->cost : COST_MAX64;
    int bmx = m->mv[0];
    int bmy = m->mv[1];
    int omx = bmx;
    int omy = bmy;
    int pmx, pmy, i, j;
    unsigned bsatd;
    int satd = 0;
    int dir = -2;
    int satds[8];

    if( m->i_pixel != PIXEL_16x16 && i4 != 0 )
        x264_mb_predict_mv( h, i_list, i4, bw, m->mvp );
    pmx = m->mvp[0];
    pmy = m->mvp[1];
    p_cost_mvx = m->p_cost_mv - pmx;
    p_cost_mvy = m->p_cost_mv - pmy;
    COST_MV_SATD( bmx, bmy, bsatd, 0 );
    COST_MV_RD( bmx, bmy, 0, 0, 0 );

    /* check the predicted mv */
    if( (bmx != pmx || bmy != pmy)
        && pmx >= h->mb.mv_min_spel[0] && pmx <= h->mb.mv_max_spel[0]
        && pmy >= h->mb.mv_min_spel[1] && pmy <= h->mb.mv_max_spel[1] )
    {
        COST_MV_SATD( pmx, pmy, satd, 0 );
        COST_MV_RD( pmx, pmy, satd, 0,0 );
        /* The hex motion search is guaranteed to not repeat the center candidate,
         * so if pmv is chosen, set the "MV to avoid checking" to bmv instead. */
        if( bmx == pmx && bmy == pmy )
        {
            pmx = m->mv[0];
            pmy = m->mv[1];
        }
    }

    /* subpel hex search, same pattern as ME HEX. */
    dir = -2;
    omx = bmx;
    omy = bmy;
    for( j=0; j<6; j++ ) COST_MV_SATD( omx + hex2[j+1][0], omy + hex2[j+1][1], satds[j], 1 );
    for( j=0; j<6; j++ ) COST_MV_RD  ( omx + hex2[j+1][0], omy + hex2[j+1][1], satds[j], 1,j );

    if( dir != -2 )
    {
        /* half hexagon, not overlapping the previous iteration */
        for( i = 1; i < 10; i++ )
        {
            const int odir = mod6m1[dir+1];
            if( bmy > h->mb.mv_max_spel[1] - 2 ||
                bmy < h->mb.mv_min_spel[1] - 2 )
                break;
            dir = -2;
            omx = bmx;
            omy = bmy;
            for( j=0; j<3; j++ ) COST_MV_SATD( omx + hex2[odir+j][0], omy + hex2[odir+j][1], satds[j], 1 );
            for( j=0; j<3; j++ ) COST_MV_RD  ( omx + hex2[odir+j][0], omy + hex2[odir+j][1], satds[j], 1, odir-1+j );
            if( dir == -2 )
                break;
        }
    }

    /* square refine, same as pattern as ME HEX. */
    omx = bmx;
    omy = bmy;
    for( i=0; i<8; i++ ) COST_MV_SATD( omx + square1[i][0], omy  + square1[i][1], satds[i], 1 );
    for( i=0; i<8; i++ ) COST_MV_RD  ( omx + square1[i][0], omy  + square1[i][1], satds[i], 0,0 );

    bmy = x264_clip3( bmy, h->mb.mv_min_spel[1],  h->mb.mv_max_spel[1] );
    m->cost = bcost;
    m->mv[0] = bmx;
    m->mv[1] = bmy;
    x264_macroblock_cache_mv ( h, block_idx_x[i4], block_idx_y[i4], bw, bh, i_list, pack16to32_mask(bmx, bmy) );
    x264_macroblock_cache_mvd( h, block_idx_x[i4], block_idx_y[i4], bw, bh, i_list, pack16to32_mask(bmx - m->mvp[0], bmy - m->mvp[1]) );
}