slicetype.c

/*****************************************************************************
 * slicetype.c: lookahead analysis
 *****************************************************************************
 * Copyright (C) 2005-2011 x264 project
 *
 * Authors: Fiona Glaser <fiona@x264.com>
 *          Loren Merritt <lorenm@u.washington.edu>
 *          Dylan Yudaken <dyudaken@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.
 *
 * This program is also available under a commercial proprietary license.
 * For more information, contact us at licensing@x264.com.
 *****************************************************************************/

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

// Indexed by pic_struct values
static const uint8_t delta_tfi_divisor[10] = { 0, 2, 1, 1, 2, 2, 3, 3, 4, 6 };

static int x264_slicetype_frame_cost( x264_t *h, x264_mb_analysis_t *a,
                                      x264_frame_t **frames, int p0, int p1, int b,
                                      int b_intra_penalty );

static void x264_lowres_context_init( x264_t *h, x264_mb_analysis_t *a )
{
    a->i_qp = X264_LOOKAHEAD_QP;
    a->i_lambda = x264_lambda_tab[ a->i_qp ];
    x264_mb_analyse_load_costs( h, a );
    if( h->param.analyse.i_subpel_refine > 1 )
    {
        h->mb.i_me_method = X264_MIN( X264_ME_HEX, h->param.analyse.i_me_method );
        h->mb.i_subpel_refine = 4;
    }
    else
    {
        h->mb.i_me_method = X264_ME_DIA;
        h->mb.i_subpel_refine = 2;
    }
    h->mb.b_chroma_me = 0;
}

/* makes a non-h264 weight (i.e. fix7), into an h264 weight */
static void x264_weight_get_h264( int weight_nonh264, int offset, x264_weight_t *w )
{
    w->i_offset = offset;
    w->i_denom = 7;
    w->i_scale = weight_nonh264;
    while( w->i_denom > 0 && (w->i_scale > 127 || !(w->i_scale & 1)) )
    {
        w->i_denom--;
        w->i_scale >>= 1;
    }
    w->i_scale = X264_MIN( w->i_scale, 127 );
}

static NOINLINE pixel *x264_weight_cost_init_luma( x264_t *h, x264_frame_t *fenc, x264_frame_t *ref, pixel *dest )
{
    int ref0_distance = fenc->i_frame - ref->i_frame - 1;
    /* Note: this will never run during lookahead as weights_analyse is only called if no
     * motion search has been done. */
    if( fenc->lowres_mvs[0][ref0_distance][0][0] != 0x7FFF )
    {
        int i_stride = fenc->i_stride_lowres;
        int i_lines = fenc->i_lines_lowres;
        int i_width = fenc->i_width_lowres;
        int i_mb_xy = 0;
        pixel *p = dest;

        for( int y = 0; y < i_lines; y += 8, p += i_stride*8 )
            for( int x = 0; x < i_width; x += 8, i_mb_xy++ )
            {
                int mvx = fenc->lowres_mvs[0][ref0_distance][i_mb_xy][0];
                int mvy = fenc->lowres_mvs[0][ref0_distance][i_mb_xy][1];
                h->mc.mc_luma( p+x, i_stride, ref->lowres, i_stride,
                               mvx+(x<<2), mvy+(y<<2), 8, 8, x264_weight_none );
            }
        x264_emms();
        return dest;
    }
    x264_emms();
    return ref->lowres[0];
}

/* How data is organized for chroma weightp:
 * [U: ref] [U: fenc]
 * [V: ref] [V: fenc]
 * fenc = ref + offset
 * v = u + stride * chroma height
 * We'll need more room if we do 4:2:2 or 4:4:4. */

static NOINLINE void x264_weight_cost_init_chroma( x264_t *h, x264_frame_t *fenc, x264_frame_t *ref, pixel *dstu, pixel *dstv )
{
    int ref0_distance = fenc->i_frame - ref->i_frame - 1;
    int i_stride = fenc->i_stride[1];
    int i_offset = i_stride / 2;
    int i_lines = fenc->i_lines[1];
    int i_width = fenc->i_width[1];
    int cw = h->mb.i_mb_width  << 3;
    int ch = h->mb.i_mb_height << 3;

    if( fenc->lowres_mvs[0][ref0_distance][0][0] != 0x7FFF )
    {
        x264_frame_expand_border_chroma( h, ref, 1 );
        for( int y = 0, mb_xy = 0, pel_offset_y = 0; y < i_lines; y += 8, pel_offset_y = y*i_stride )
            for( int x = 0, pel_offset_x = 0; x < i_width; x += 8, mb_xy++, pel_offset_x += 8 )
            {
                pixel *pixu = dstu + pel_offset_y + pel_offset_x;
                pixel *pixv = dstv + pel_offset_y + pel_offset_x;
                pixel *src1 =  ref->plane[1] + pel_offset_y + pel_offset_x*2; /* NV12 */
                int mvx = fenc->lowres_mvs[0][ref0_distance][mb_xy][0];
                int mvy = fenc->lowres_mvs[0][ref0_distance][mb_xy][1];
                h->mc.mc_chroma( pixu, pixv, i_stride, src1, i_stride, mvx, mvy, 8, 8 );
            }
    }
    else
        h->mc.plane_copy_deinterleave( dstu, i_stride, dstv, i_stride, ref->plane[1], i_stride, cw, ch );
    h->mc.plane_copy_deinterleave( dstu+i_offset, i_stride, dstv+i_offset, i_stride, fenc->plane[1], i_stride, cw, ch );
    x264_emms();
}

static NOINLINE pixel *x264_weight_cost_init_chroma444( x264_t *h, x264_frame_t *fenc, x264_frame_t *ref, pixel *dst, int p )
{
    int ref0_distance = fenc->i_frame - ref->i_frame - 1;
    int i_stride = fenc->i_stride[p];
    int i_lines = fenc->i_lines[p];
    int i_width = fenc->i_width[p];

    if( fenc->lowres_mvs[0][ref0_distance][0][0] != 0x7FFF )
    {
        x264_frame_expand_border_chroma( h, ref, p );
        for( int y = 0, mb_xy = 0, pel_offset_y = 0; y < i_lines; y += 16, pel_offset_y = y*i_stride )
            for( int x = 0, pel_offset_x = 0; x < i_width; x += 16, mb_xy++, pel_offset_x += 16 )
            {
                pixel *pix = dst + pel_offset_y + pel_offset_x;
                pixel *src = ref->plane[p] + pel_offset_y + pel_offset_x;
                int mvx = fenc->lowres_mvs[0][ref0_distance][mb_xy][0] / 2;
                int mvy = fenc->lowres_mvs[0][ref0_distance][mb_xy][1] / 2;
                /* We don't want to calculate hpels for fenc frames, so we round the motion
                 * vectors to fullpel here.  It's not too bad, I guess? */
                h->mc.copy_16x16_unaligned( pix, i_stride, src+mvx+mvy*i_stride, i_stride, 16 );
            }
        x264_emms();
        return dst;
    }
    x264_emms();
    return ref->plane[p];
}

static int x264_weight_slice_header_cost( x264_t *h, x264_weight_t *w, int b_chroma )
{
    /* Add cost of weights in the slice header. */
    int lambda = x264_lambda_tab[X264_LOOKAHEAD_QP];
    /* 4 times higher, because chroma is analyzed at full resolution. */
    if( b_chroma )
        lambda *= 4;
    int numslices;
    if( h->param.i_slice_count )
        numslices = h->param.i_slice_count;
    else if( h->param.i_slice_max_mbs )
        numslices = (h->mb.i_mb_width * h->mb.i_mb_height + h->param.i_slice_max_mbs-1) / h->param.i_slice_max_mbs;
    else
        numslices = 1;
    /* FIXME: find a way to account for --slice-max-size?
     * Multiply by 2 as there will be a duplicate. 10 bits added as if there is a weighted frame, then an additional duplicate is used.
     * Cut denom cost in half if chroma, since it's shared between the two chroma planes. */
    int denom_cost = bs_size_ue( w[0].i_denom ) * (2 - b_chroma);
    return lambda * numslices * ( 10 + denom_cost + 2 * (bs_size_se( w[0].i_scale ) + bs_size_se( w[0].i_offset )) );
}

static NOINLINE unsigned int x264_weight_cost_luma( x264_t *h, x264_frame_t *fenc, pixel *src, x264_weight_t *w )
{
    unsigned int cost = 0;
    int i_stride = fenc->i_stride_lowres;
    int i_lines = fenc->i_lines_lowres;
    int i_width = fenc->i_width_lowres;
    pixel *fenc_plane = fenc->lowres[0];
    ALIGNED_ARRAY_16( pixel, buf,[8*8] );
    int pixoff = 0;
    int i_mb = 0;

    if( w )
    {
        for( int y = 0; y < i_lines; y += 8, pixoff = y*i_stride )
            for( int x = 0; x < i_width; x += 8, i_mb++, pixoff += 8)
            {
                w->weightfn[8>>2]( buf, 8, &src[pixoff], i_stride, w, 8 );
                int cmp = h->pixf.mbcmp[PIXEL_8x8]( buf, 8, &fenc_plane[pixoff], i_stride );
                cost += X264_MIN( cmp, fenc->i_intra_cost[i_mb] );
            }
        cost += x264_weight_slice_header_cost( h, w, 0 );
    }
    else
        for( int y = 0; y < i_lines; y += 8, pixoff = y*i_stride )
            for( int x = 0; x < i_width; x += 8, i_mb++, pixoff += 8 )
            {
                int cmp = h->pixf.mbcmp[PIXEL_8x8]( &src[pixoff], i_stride, &fenc_plane[pixoff], i_stride );
                cost += X264_MIN( cmp, fenc->i_intra_cost[i_mb] );
            }
    x264_emms();
    return cost;
}

static NOINLINE unsigned int x264_weight_cost_chroma( x264_t *h, x264_frame_t *fenc, pixel *ref, x264_weight_t *w )
{
    unsigned int cost = 0;
    int i_stride = fenc->i_stride[1];
    int i_offset = i_stride / 2;
    int i_lines = fenc->i_lines[1];
    int i_width = fenc->i_width[1];
    pixel *src = ref + i_offset;
    ALIGNED_ARRAY_16( pixel, buf, [8*8] );
    int pixoff = 0;
    ALIGNED_16( static pixel flat[8] ) = {0};
    if( w )
    {
        for( int y = 0; y < i_lines; y += 8, pixoff = y*i_stride )
            for( int x = 0; x < i_width; x += 8, pixoff += 8 )
            {
                w->weightfn[8>>2]( buf, 8, &ref[pixoff], i_stride, w, 8 );
                /* The naive and seemingly sensible algorithm is to use mbcmp as in luma.
                 * But testing shows that for chroma the DC coefficient is by far the most
                 * important part of the coding cost.  Thus a more useful chroma weight is
                 * obtained by comparing each block's DC coefficient instead of the actual
                 * pixels.
                 *
                 * FIXME: add a (faster) asm sum function to replace sad. */
                cost += abs( h->pixf.sad_aligned[PIXEL_8x8](          buf,        8, flat, 0 ) -
                             h->pixf.sad_aligned[PIXEL_8x8]( &src[pixoff], i_stride, flat, 0 ) );
            }
        cost += x264_weight_slice_header_cost( h, w, 1 );
    }
    else
        for( int y = 0; y < i_lines; y += 8, pixoff = y*i_stride )
            for( int x = 0; x < i_width; x += 8, pixoff += 8 )
                cost += abs( h->pixf.sad_aligned[PIXEL_8x8]( &ref[pixoff], i_stride, flat, 0 ) -
                             h->pixf.sad_aligned[PIXEL_8x8]( &src[pixoff], i_stride, flat, 0 ) );
    x264_emms();
    return cost;
}

static NOINLINE unsigned int x264_weight_cost_chroma444( x264_t *h, x264_frame_t *fenc, pixel *ref, x264_weight_t *w, int p )
{
    unsigned int cost = 0;
    int i_stride = fenc->i_stride[p];
    int i_lines = fenc->i_lines[p];
    int i_width = fenc->i_width[p];
    pixel *src = fenc->plane[p];
    ALIGNED_ARRAY_16( pixel, buf, [16*16] );
    int pixoff = 0;
    if( w )
    {
        for( int y = 0; y < i_lines; y += 16, pixoff = y*i_stride )
            for( int x = 0; x < i_width; x += 16, pixoff += 16 )
            {
                w->weightfn[16>>2]( buf, 16, &ref[pixoff], i_stride, w, 16 );
                cost += h->pixf.mbcmp[PIXEL_16x16]( buf, 16, &src[pixoff], i_stride );
            }
        cost += x264_weight_slice_header_cost( h, w, 1 );
    }
    else
        for( int y = 0; y < i_lines; y += 16, pixoff = y*i_stride )
            for( int x = 0; x < i_width; x += 16, pixoff += 16 )
                cost += h->pixf.mbcmp[PIXEL_16x16]( &ref[pixoff], i_stride, &src[pixoff], i_stride );
    x264_emms();
    return cost;
}

void x264_weights_analyse( x264_t *h, x264_frame_t *fenc, x264_frame_t *ref, int b_lookahead )
{
    int i_delta_index = fenc->i_frame - ref->i_frame - 1;
    /* epsilon is chosen to require at least a numerator of 127 (with denominator = 128) */
    const float epsilon = 1.f/128.f;
    x264_weight_t *weights = fenc->weight[0];
    SET_WEIGHT( weights[0], 0, 1, 0, 0 );
    SET_WEIGHT( weights[1], 0, 1, 0, 0 );
    SET_WEIGHT( weights[2], 0, 1, 0, 0 );
    /* Don't check chroma in lookahead, or if there wasn't a luma weight. */
    for( int plane = 0; plane <= 2 && !( plane && ( !weights[0].weightfn || b_lookahead ) ); plane++ )
    {
        int cur_offset, start_offset, end_offset;
        int minoff, minscale, mindenom;
        unsigned int minscore, origscore;
        int found;
        float fenc_var = fenc->i_pixel_ssd[plane] + !ref->i_pixel_ssd[plane];
        float ref_var  =  ref->i_pixel_ssd[plane] + !ref->i_pixel_ssd[plane];
        float guess_scale = sqrtf( fenc_var / ref_var );
        float fenc_mean = (float)fenc->i_pixel_sum[plane] / (fenc->i_lines[!!plane] * fenc->i_width[!!plane]) / (1 << (BIT_DEPTH - 8));
        float ref_mean  = (float) ref->i_pixel_sum[plane] / (fenc->i_lines[!!plane] * fenc->i_width[!!plane]) / (1 << (BIT_DEPTH - 8));

        //early termination
        if( fabsf( ref_mean - fenc_mean ) < 0.5f && fabsf( 1.f - guess_scale ) < epsilon )
        {
            SET_WEIGHT( weights[plane], 0, 1, 0, 0 );
            continue;
        }

        if( plane )
        {
            weights[plane].i_denom = 6;
            weights[plane].i_scale = x264_clip3( round( guess_scale * 64 ), 0, 255 );
            if( weights[plane].i_scale > 127 )
            {
                weights[1].weightfn = weights[2].weightfn = NULL;
                break;
            }
        }
        else
            x264_weight_get_h264( round( guess_scale * 128 ), 0, &weights[plane] );

        found = 0;
        mindenom = weights[plane].i_denom;
        minscale = weights[plane].i_scale;
        minoff = 0;

        pixel *mcbuf;
        if( !plane )
        {
            if( !fenc->b_intra_calculated )
            {
                x264_mb_analysis_t a;
                x264_lowres_context_init( h, &a );
                x264_slicetype_frame_cost( h, &a, &fenc, 0, 0, 0, 0 );
            }
            mcbuf = x264_weight_cost_init_luma( h, fenc, ref, h->mb.p_weight_buf[0] );
            origscore = minscore = x264_weight_cost_luma( h, fenc, mcbuf, NULL );
        }
        else
        {
            if( CHROMA444 )
            {
                mcbuf = x264_weight_cost_init_chroma444( h, fenc, ref, h->mb.p_weight_buf[0], plane );
                origscore = minscore = x264_weight_cost_chroma444( h, fenc, mcbuf, NULL, plane );
            }
            else
            {
                pixel *dstu = h->mb.p_weight_buf[0];
                pixel *dstv = h->mb.p_weight_buf[0]+fenc->i_stride[1]*fenc->i_lines[1];
                /* Only initialize chroma data once. */
                if( plane == 1 )
                    x264_weight_cost_init_chroma( h, fenc, ref, dstu, dstv );
                mcbuf = plane == 1 ? dstu : dstv;
                origscore = minscore = x264_weight_cost_chroma( h, fenc, mcbuf, NULL );
            }
        }

        if( !minscore )
            continue;

        // This gives a slight improvement due to rounding errors but only tests one offset in lookahead.
        // Currently only searches within +/- 1 of the best offset found so far.
        // TODO: Try other offsets/multipliers/combinations thereof?
        cur_offset = fenc_mean - ref_mean * minscale / (1 << mindenom) + 0.5f * b_lookahead;
        start_offset = x264_clip3( cur_offset - !b_lookahead, -128, 127 );
        end_offset   = x264_clip3( cur_offset + !b_lookahead, -128, 127 );
        for( int i_off = start_offset; i_off <= end_offset; i_off++ )
        {
            SET_WEIGHT( weights[plane], 1, minscale, mindenom, i_off );
            unsigned int s;
            if( plane )
            {
                if( CHROMA444 )
                    s = x264_weight_cost_chroma444( h, fenc, mcbuf, &weights[plane], plane );
                else
                    s = x264_weight_cost_chroma( h, fenc, mcbuf, &weights[plane] );
            }
            else
                s = x264_weight_cost_luma( h, fenc, mcbuf, &weights[plane] );
            COPY3_IF_LT( minscore, s, minoff, i_off, found, 1 );

            // Don't check any more offsets if the previous one had a lower cost than the current one
            if( minoff == start_offset && i_off != start_offset )
                break;
        }
        x264_emms();

        /* FIXME: More analysis can be done here on SAD vs. SATD termination. */
        /* 0.2% termination derived experimentally to avoid weird weights in frames that are mostly intra. */
        if( !found || (minscale == 1 << mindenom && minoff == 0) || (float)minscore / origscore > 0.998f )
        {
            SET_WEIGHT( weights[plane], 0, 1, 0, 0 );
            continue;
        }
        else
            SET_WEIGHT( weights[plane], 1, minscale, mindenom, minoff );

        if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_FAKE && weights[0].weightfn && !plane )
            fenc->f_weighted_cost_delta[i_delta_index] = (float)minscore / origscore;
    }

    //FIXME, what is the correct way to deal with this?
    if( weights[1].weightfn && weights[2].weightfn && weights[1].i_denom != weights[2].i_denom )
    {
        int denom = X264_MIN( weights[1].i_denom, weights[2].i_denom );
        int i;
        for( i = 1; i <= 2; i++ )
        {
            weights[i].i_scale = x264_clip3( weights[i].i_scale >> ( weights[i].i_denom - denom ), 0, 255 );
            weights[i].i_denom = denom;
            h->mc.weight_cache( h, &weights[i] );
        }
    }

    if( weights[0].weightfn && b_lookahead )
    {
        //scale lowres in lookahead for slicetype_frame_cost
        pixel *src = ref->buffer_lowres[0];
        pixel *dst = h->mb.p_weight_buf[0];
        int width = ref->i_width_lowres + PADH*2;
        int height = ref->i_lines_lowres + PADV*2;
        x264_weight_scale_plane( h, dst, ref->i_stride_lowres, src, ref->i_stride_lowres,
                                 width, height, &weights[0] );
        fenc->weighted[0] = h->mb.p_weight_buf[0] + PADH + ref->i_stride_lowres * PADV;
    }
}

static void x264_slicetype_mb_cost( x264_t *h, x264_mb_analysis_t *a,
                                    x264_frame_t **frames, int p0, int p1, int b,
                                    int dist_scale_factor, int do_search[2], const x264_weight_t *w )
{
    x264_frame_t *fref0 = frames[p0];
    x264_frame_t *fref1 = frames[p1];
    x264_frame_t *fenc  = frames[b];
    const int b_bidir = (b < p1);
    const int i_mb_x = h->mb.i_mb_x;
    const int i_mb_y = h->mb.i_mb_y;
    const int i_mb_stride = h->mb.i_mb_width;
    const int i_mb_xy = i_mb_x + i_mb_y * i_mb_stride;
    const int i_stride = fenc->i_stride_lowres;
    const int i_pel_offset = 8 * (i_mb_x + i_mb_y * i_stride);
    const int i_bipred_weight = h->param.analyse.b_weighted_bipred ? 64 - (dist_scale_factor>>2) : 32;
    int16_t (*fenc_mvs[2])[2] = { &fenc->lowres_mvs[0][b-p0-1][i_mb_xy], &fenc->lowres_mvs[1][p1-b-1][i_mb_xy] };
    int (*fenc_costs[2]) = { &fenc->lowres_mv_costs[0][b-p0-1][i_mb_xy], &fenc->lowres_mv_costs[1][p1-b-1][i_mb_xy] };
    int b_frame_score_mb = (i_mb_x > 0 && i_mb_x < h->mb.i_mb_width - 1 &&
                            i_mb_y > 0 && i_mb_y < h->mb.i_mb_height - 1) ||
                            h->mb.i_mb_width <= 2 || h->mb.i_mb_height <= 2;

    ALIGNED_ARRAY_16( pixel, pix1,[9*FDEC_STRIDE] );
    pixel *pix2 = pix1+8;
    x264_me_t m[2];
    int i_bcost = COST_MAX;
    int list_used = 0;

    h->mb.pic.p_fenc[0] = h->mb.pic.fenc_buf;
    h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fenc[0], FENC_STRIDE, &fenc->lowres[0][i_pel_offset], i_stride, 8 );

    if( p0 == p1 )
        goto lowres_intra_mb;

    // no need for h->mb.mv_min[]
    h->mb.mv_min_fpel[0] = -8*h->mb.i_mb_x - 4;
    h->mb.mv_max_fpel[0] = 8*( h->mb.i_mb_width - h->mb.i_mb_x - 1 ) + 4;
    h->mb.mv_min_spel[0] = 4*( h->mb.mv_min_fpel[0] - 8 );
    h->mb.mv_max_spel[0] = 4*( h->mb.mv_max_fpel[0] + 8 );
    if( h->mb.i_mb_x >= h->mb.i_mb_width - 2 )
    {
        h->mb.mv_min_fpel[1] = -8*h->mb.i_mb_y - 4;
        h->mb.mv_max_fpel[1] = 8*( h->mb.i_mb_height - h->mb.i_mb_y - 1 ) + 4;
        h->mb.mv_min_spel[1] = 4*( h->mb.mv_min_fpel[1] - 8 );
        h->mb.mv_max_spel[1] = 4*( h->mb.mv_max_fpel[1] + 8 );
    }

#define LOAD_HPELS_LUMA(dst, src) \
    { \
        (dst)[0] = &(src)[0][i_pel_offset]; \
        (dst)[1] = &(src)[1][i_pel_offset]; \
        (dst)[2] = &(src)[2][i_pel_offset]; \
        (dst)[3] = &(src)[3][i_pel_offset]; \
    }
#define LOAD_WPELS_LUMA(dst,src) \
    (dst) = &(src)[i_pel_offset];

#define CLIP_MV( mv ) \
    { \
        mv[0] = x264_clip3( mv[0], h->mb.mv_min_spel[0], h->mb.mv_max_spel[0] ); \
        mv[1] = x264_clip3( mv[1], h->mb.mv_min_spel[1], h->mb.mv_max_spel[1] ); \
    }
#define TRY_BIDIR( mv0, mv1, penalty ) \
    { \
        int i_cost; \
        if( h->param.analyse.i_subpel_refine <= 1 ) \
        { \
            int hpel_idx1 = (((mv0)[0]&2)>>1) + ((mv0)[1]&2); \
            int hpel_idx2 = (((mv1)[0]&2)>>1) + ((mv1)[1]&2); \
            pixel *src1 = m[0].p_fref[hpel_idx1] + ((mv0)[0]>>2) + ((mv0)[1]>>2) * m[0].i_stride[0]; \
            pixel *src2 = m[1].p_fref[hpel_idx2] + ((mv1)[0]>>2) + ((mv1)[1]>>2) * m[1].i_stride[0]; \
            h->mc.avg[PIXEL_8x8]( pix1, 16, src1, m[0].i_stride[0], src2, m[1].i_stride[0], i_bipred_weight ); \
        } \
        else \
        { \
            int stride1 = 16, stride2 = 16; \
            pixel *src1, *src2; \
            src1 = h->mc.get_ref( pix1, &stride1, m[0].p_fref, m[0].i_stride[0], \
                                  (mv0)[0], (mv0)[1], 8, 8, w ); \
            src2 = h->mc.get_ref( pix2, &stride2, m[1].p_fref, m[1].i_stride[0], \
                                  (mv1)[0], (mv1)[1], 8, 8, w ); \
            h->mc.avg[PIXEL_8x8]( pix1, 16, src1, stride1, src2, stride2, i_bipred_weight ); \
        } \
        i_cost = penalty * a->i_lambda + h->pixf.mbcmp[PIXEL_8x8]( \
                           m[0].p_fenc[0], FENC_STRIDE, pix1, 16 ); \
        COPY2_IF_LT( i_bcost, i_cost, list_used, 3 ); \
    }

    m[0].i_pixel = PIXEL_8x8;
    m[0].p_cost_mv = a->p_cost_mv;
    m[0].i_stride[0] = i_stride;
    m[0].p_fenc[0] = h->mb.pic.p_fenc[0];
    m[0].weight = w;
    m[0].i_ref = 0;
    LOAD_HPELS_LUMA( m[0].p_fref, fref0->lowres );
    m[0].p_fref_w = m[0].p_fref[0];
    if( w[0].weightfn )
        LOAD_WPELS_LUMA( m[0].p_fref_w, fenc->weighted[0] );

    if( b_bidir )
    {
        int16_t *mvr = fref1->lowres_mvs[0][p1-p0-1][i_mb_xy];
        ALIGNED_ARRAY_8( int16_t, dmv,[2],[2] );

        m[1].i_pixel = PIXEL_8x8;
        m[1].p_cost_mv = a->p_cost_mv;
        m[1].i_stride[0] = i_stride;
        m[1].p_fenc[0] = h->mb.pic.p_fenc[0];
        m[1].i_ref = 0;
        m[1].weight = x264_weight_none;
        LOAD_HPELS_LUMA( m[1].p_fref, fref1->lowres );
        m[1].p_fref_w = m[1].p_fref[0];

        dmv[0][0] = ( mvr[0] * dist_scale_factor + 128 ) >> 8;
        dmv[0][1] = ( mvr[1] * dist_scale_factor + 128 ) >> 8;
        dmv[1][0] = dmv[0][0] - mvr[0];
        dmv[1][1] = dmv[0][1] - mvr[1];
        CLIP_MV( dmv[0] );
        CLIP_MV( dmv[1] );
        if( h->param.analyse.i_subpel_refine <= 1 )
            M64( dmv ) &= ~0x0001000100010001ULL; /* mv & ~1 */

        TRY_BIDIR( dmv[0], dmv[1], 0 );
        if( M64( dmv ) )
        {
            int i_cost;
            h->mc.avg[PIXEL_8x8]( pix1, 16, m[0].p_fref[0], m[0].i_stride[0], m[1].p_fref[0], m[1].i_stride[0], i_bipred_weight );
            i_cost = h->pixf.mbcmp[PIXEL_8x8]( m[0].p_fenc[0], FENC_STRIDE, pix1, 16 );
            COPY2_IF_LT( i_bcost, i_cost, list_used, 3 );
        }
    }

    for( int l = 0; l < 1 + b_bidir; l++ )
    {
        if( do_search[l] )
        {
            int i_mvc = 0;
            int16_t (*fenc_mv)[2] = fenc_mvs[l];
            ALIGNED_4( int16_t mvc[4][2] );

            /* Reverse-order MV prediction. */
            M32( mvc[0] ) = 0;
            M32( mvc[2] ) = 0;
#define MVC(mv) { CP32( mvc[i_mvc], mv ); i_mvc++; }
            if( i_mb_x < h->mb.i_mb_width - 1 )
                MVC( fenc_mv[1] );
            if( i_mb_y < h->mb.i_mb_height - 1 )
            {
                MVC( fenc_mv[i_mb_stride] );
                if( i_mb_x > 0 )
                    MVC( fenc_mv[i_mb_stride-1] );
                if( i_mb_x < h->mb.i_mb_width - 1 )
                    MVC( fenc_mv[i_mb_stride+1] );
            }
#undef MVC
            if( i_mvc <= 1 )
                CP32( m[l].mvp, mvc[0] );
            else
                x264_median_mv( m[l].mvp, mvc[0], mvc[1], mvc[2] );

            /* Fast skip for cases of near-zero residual.  Shortcut: don't bother except in the mv0 case,
             * since anything else is likely to have enough residual to not trigger the skip. */
            if( !M32( m[l].mvp ) )
            {
                m[l].cost = h->pixf.mbcmp[PIXEL_8x8]( m[l].p_fenc[0], FENC_STRIDE, m[l].p_fref[0], m[l].i_stride[0] );
                if( m[l].cost < 64 )
                {
                    M32( m[l].mv ) = 0;
                    goto skip_motionest;
                }
            }

            x264_me_search( h, &m[l], mvc, i_mvc );
            m[l].cost -= a->p_cost_mv[0]; // remove mvcost from skip mbs
            if( M32( m[l].mv ) )
                m[l].cost += 5 * a->i_lambda;

skip_motionest:
            CP32( fenc_mvs[l], m[l].mv );
            *fenc_costs[l] = m[l].cost;
        }
        else
        {
            CP32( m[l].mv, fenc_mvs[l] );
            m[l].cost = *fenc_costs[l];
        }
        COPY2_IF_LT( i_bcost, m[l].cost, list_used, l+1 );
    }

    if( b_bidir && ( M32( m[0].mv ) || M32( m[1].mv ) ) )
        TRY_BIDIR( m[0].mv, m[1].mv, 5 );

lowres_intra_mb:
    if( !fenc->b_intra_calculated )
    {
        ALIGNED_ARRAY_16( pixel, edge,[33] );
        pixel *pix = &pix1[8+FDEC_STRIDE - 1];
        pixel *src = &fenc->lowres[0][i_pel_offset - 1];
        const int intra_penalty = 5 * a->i_lambda;
        int satds[3];

        memcpy( pix-FDEC_STRIDE, src-i_stride, 17 * sizeof(pixel) );
        for( int i = 0; i < 8; i++ )
            pix[i*FDEC_STRIDE] = src[i*i_stride];
        pix++;

        h->pixf.intra_mbcmp_x3_8x8c( h->mb.pic.p_fenc[0], pix, satds );
        int i_icost = X264_MIN3( satds[0], satds[1], satds[2] );

        if( h->param.analyse.i_subpel_refine > 1 )
        {
            h->predict_8x8c[I_PRED_CHROMA_P]( pix );
            int satd = h->pixf.mbcmp[PIXEL_8x8]( pix, FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE );
            i_icost = X264_MIN( i_icost, satd );
            h->predict_8x8_filter( pix, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
            for( int i = 3; i < 9; i++ )
            {
                h->predict_8x8[i]( pix, edge );
                satd = h->pixf.mbcmp[PIXEL_8x8]( pix, FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE );
                i_icost = X264_MIN( i_icost, satd );
            }
        }

        i_icost += intra_penalty;
        fenc->i_intra_cost[i_mb_xy] = i_icost;
        int i_icost_aq = i_icost;
        if( h->param.rc.i_aq_mode )
            i_icost_aq = (i_icost_aq * fenc->i_inv_qscale_factor[i_mb_xy] + 128) >> 8;
        fenc->i_row_satds[0][0][h->mb.i_mb_y] += i_icost_aq;
        if( b_frame_score_mb )
        {
            fenc->i_cost_est[0][0] += i_icost;
            fenc->i_cost_est_aq[0][0] += i_icost_aq;
        }
    }

    /* forbid intra-mbs in B-frames, because it's rare and not worth checking */
    /* FIXME: Should we still forbid them now that we cache intra scores? */
    if( !b_bidir )
    {
        int i_icost = fenc->i_intra_cost[i_mb_xy];
        int b_intra = i_icost < i_bcost;
        if( b_intra )
        {
            i_bcost = i_icost;
            list_used = 0;
        }
        if( b_frame_score_mb )
            fenc->i_intra_mbs[b-p0] += b_intra;
    }

    /* In an I-frame, we've already added the results above in the intra section. */
    if( p0 != p1 )
    {
        int i_bcost_aq = i_bcost;
        if( h->param.rc.i_aq_mode )
            i_bcost_aq = (i_bcost_aq * fenc->i_inv_qscale_factor[i_mb_xy] + 128) >> 8;
        fenc->i_row_satds[b-p0][p1-b][h->mb.i_mb_y] += i_bcost_aq;
        if( b_frame_score_mb )
        {
            /* Don't use AQ-weighted costs for slicetype decision, only for ratecontrol. */
            fenc->i_cost_est[b-p0][p1-b] += i_bcost;
            fenc->i_cost_est_aq[b-p0][p1-b] += i_bcost_aq;
        }
    }

    fenc->lowres_costs[b-p0][p1-b][i_mb_xy] = X264_MIN( i_bcost, LOWRES_COST_MASK ) + (list_used << LOWRES_COST_SHIFT);
}
#undef TRY_BIDIR

#define NUM_MBS\
   (h->mb.i_mb_width > 2 && h->mb.i_mb_height > 2 ?\
   (h->mb.i_mb_width - 2) * (h->mb.i_mb_height - 2) :\
    h->mb.i_mb_width * h->mb.i_mb_height)

static int x264_slicetype_frame_cost( x264_t *h, x264_mb_analysis_t *a,
                                      x264_frame_t **frames, int p0, int p1, int b,
                                      int b_intra_penalty )
{
    int i_score = 0;
    int do_search[2];
    const x264_weight_t *w = x264_weight_none;
    /* Check whether we already evaluated this frame
     * If we have tried this frame as P, then we have also tried
     * the preceding frames as B. (is this still true?) */
    /* Also check that we already calculated the row SATDs for the current frame. */
    if( frames[b]->i_cost_est[b-p0][p1-b] >= 0 && (!h->param.rc.i_vbv_buffer_size || frames[b]->i_row_satds[b-p0][p1-b][0] != -1) )
        i_score = frames[b]->i_cost_est[b-p0][p1-b];
    else
    {
        int dist_scale_factor = 128;
        int *row_satd = frames[b]->i_row_satds[b-p0][p1-b];
        int *row_satd_intra = frames[b]->i_row_satds[0][0];

        /* For each list, check to see whether we have lowres motion-searched this reference frame before. */
        do_search[0] = b != p0 && frames[b]->lowres_mvs[0][b-p0-1][0][0] == 0x7FFF;
        do_search[1] = b != p1 && frames[b]->lowres_mvs[1][p1-b-1][0][0] == 0x7FFF;
        if( do_search[0] )
        {
            if( h->param.analyse.i_weighted_pred && b == p1 )
            {
                x264_emms();
                x264_weights_analyse( h, frames[b], frames[p0], 1 );
                w = frames[b]->weight[0];
            }
            frames[b]->lowres_mvs[0][b-p0-1][0][0] = 0;
        }
        if( do_search[1] ) frames[b]->lowres_mvs[1][p1-b-1][0][0] = 0;

        if( b == p1 )
            frames[b]->i_intra_mbs[b-p0] = 0;
        if( !frames[b]->b_intra_calculated )
        {
            frames[b]->i_cost_est[0][0] = 0;
            frames[b]->i_cost_est_aq[0][0] = 0;
        }
        if( p1 != p0 )
            dist_scale_factor = ( ((b-p0) << 8) + ((p1-p0) >> 1) ) / (p1-p0);

        frames[b]->i_cost_est[b-p0][p1-b] = 0;
        frames[b]->i_cost_est_aq[b-p0][p1-b] = 0;

        /* Lowres lookahead goes backwards because the MVs are used as predictors in the main encode.
         * This considerably improves MV prediction overall. */

        /* The edge mbs seem to reduce the predictive quality of the
         * whole frame's score, but are needed for a spatial distribution. */
        if( h->param.rc.b_mb_tree || h->param.rc.i_vbv_buffer_size ||
            h->mb.i_mb_width <= 2 || h->mb.i_mb_height <= 2 )
        {
            for( h->mb.i_mb_y = h->mb.i_mb_height - 1; h->mb.i_mb_y >= 0; h->mb.i_mb_y-- )
            {
                row_satd[h->mb.i_mb_y] = 0;
                if( !frames[b]->b_intra_calculated )
                    row_satd_intra[h->mb.i_mb_y] = 0;
                for( h->mb.i_mb_x = h->mb.i_mb_width - 1; h->mb.i_mb_x >= 0; h->mb.i_mb_x-- )
                    x264_slicetype_mb_cost( h, a, frames, p0, p1, b, dist_scale_factor, do_search, w );
            }
        }
        else
        {
            for( h->mb.i_mb_y = h->mb.i_mb_height - 2; h->mb.i_mb_y >= 1; h->mb.i_mb_y-- )
                for( h->mb.i_mb_x = h->mb.i_mb_width - 2; h->mb.i_mb_x >= 1; h->mb.i_mb_x-- )
                    x264_slicetype_mb_cost( h, a, frames, p0, p1, b, dist_scale_factor, do_search, w );
        }

        i_score = frames[b]->i_cost_est[b-p0][p1-b];
        if( b != p1 )
            i_score = (uint64_t)i_score * 100 / (120 + h->param.i_bframe_bias);
        else
            frames[b]->b_intra_calculated = 1;

        frames[b]->i_cost_est[b-p0][p1-b] = i_score;
        x264_emms();
    }

    if( b_intra_penalty )
    {
        // arbitrary penalty for I-blocks after B-frames
        int nmb = NUM_MBS;
        i_score += i_score * frames[b]->i_intra_mbs[b-p0] / (nmb * 8);
    }
    return i_score;
}

/* If MB-tree changes the quantizers, we need to recalculate the frame cost without
 * re-running lookahead. */
static int x264_slicetype_frame_cost_recalculate( x264_t *h, x264_frame_t **frames, int p0, int p1, int b )
{
    int i_score = 0;
    int *row_satd = frames[b]->i_row_satds[b-p0][p1-b];
    float *qp_offset = IS_X264_TYPE_B(frames[b]->i_type) ? frames[b]->f_qp_offset_aq : frames[b]->f_qp_offset;
    x264_emms();
    for( h->mb.i_mb_y = h->mb.i_mb_height - 1; h->mb.i_mb_y >= 0; h->mb.i_mb_y-- )
    {
        row_satd[ h->mb.i_mb_y ] = 0;
        for( h->mb.i_mb_x = h->mb.i_mb_width - 1; h->mb.i_mb_x >= 0; h->mb.i_mb_x-- )
        {
            int i_mb_xy = h->mb.i_mb_x + h->mb.i_mb_y*h->mb.i_mb_stride;
            int i_mb_cost = frames[b]->lowres_costs[b-p0][p1-b][i_mb_xy] & LOWRES_COST_MASK;
            float qp_adj = qp_offset[i_mb_xy];
            i_mb_cost = (i_mb_cost * x264_exp2fix8(qp_adj) + 128) >> 8;
            row_satd[ h->mb.i_mb_y ] += i_mb_cost;
            if( (h->mb.i_mb_y > 0 && h->mb.i_mb_y < h->mb.i_mb_height - 1 &&
                 h->mb.i_mb_x > 0 && h->mb.i_mb_x < h->mb.i_mb_width - 1) ||
                 h->mb.i_mb_width <= 2 || h->mb.i_mb_height <= 2 )
            {
                i_score += i_mb_cost;
            }
        }
    }
    return i_score;
}

static void x264_macroblock_tree_finish( x264_t *h, x264_frame_t *frame, float average_duration, int ref0_distance )
{
    int fps_factor = round( CLIP_DURATION(average_duration) / CLIP_DURATION(frame->f_duration) * 256 );
    float weightdelta = 0.0;
    if( ref0_distance && frame->f_weighted_cost_delta[ref0_distance-1] > 0 )
        weightdelta = (1.0 - frame->f_weighted_cost_delta[ref0_distance-1]);

    /* Allow the strength to be adjusted via qcompress, since the two
     * concepts are very similar. */
    float strength = 5.0f * (1.0f - h->param.rc.f_qcompress);
    for( int mb_index = 0; mb_index < h->mb.i_mb_count; mb_index++ )
    {
        int intra_cost = (frame->i_intra_cost[mb_index] * frame->i_inv_qscale_factor[mb_index] + 128) >> 8;
        if( intra_cost )
        {
            int propagate_cost = (frame->i_propagate_cost[mb_index] * fps_factor + 128) >> 8;
            float log2_ratio = x264_log2(intra_cost + propagate_cost) - x264_log2(intra_cost) + weightdelta;
            frame->f_qp_offset[mb_index] = frame->f_qp_offset_aq[mb_index] - strength * log2_ratio;
        }
    }
}

static void x264_macroblock_tree_propagate( x264_t *h, x264_frame_t **frames, float average_duration, int p0, int p1, int b, int referenced )
{
    uint16_t *ref_costs[2] = {frames[p0]->i_propagate_cost,frames[p1]->i_propagate_cost};
    int dist_scale_factor = ( ((b-p0) << 8) + ((p1-p0) >> 1) ) / (p1-p0);
    int i_bipred_weight = h->param.analyse.b_weighted_bipred ? 64 - (dist_scale_factor>>2) : 32;
    int16_t (*mvs[2])[2] = { frames[b]->lowres_mvs[0][b-p0-1], frames[b]->lowres_mvs[1][p1-b-1] };
    int bipred_weights[2] = {i_bipred_weight, 64 - i_bipred_weight};
    int *buf = h->scratch_buffer;
    uint16_t *propagate_cost = frames[b]->i_propagate_cost;

    x264_emms();
    float fps_factor = CLIP_DURATION(frames[b]->f_duration) / CLIP_DURATION(average_duration);

    /* For non-reffed frames the source costs are always zero, so just memset one row and re-use it. */
    if( !referenced )
        memset( frames[b]->i_propagate_cost, 0, h->mb.i_mb_width * sizeof(uint16_t) );

    for( h->mb.i_mb_y = 0; h->mb.i_mb_y < h->mb.i_mb_height; h->mb.i_mb_y++ )
    {
        int mb_index = h->mb.i_mb_y*h->mb.i_mb_stride;
        h->mc.mbtree_propagate_cost( buf, propagate_cost,
            frames[b]->i_intra_cost+mb_index, frames[b]->lowres_costs[b-p0][p1-b]+mb_index,
            frames[b]->i_inv_qscale_factor+mb_index, &fps_factor, h->mb.i_mb_width );
        if( referenced )
            propagate_cost += h->mb.i_mb_width;
        for( h->mb.i_mb_x = 0; h->mb.i_mb_x < h->mb.i_mb_width; h->mb.i_mb_x++, mb_index++ )
        {
            int propagate_amount = buf[h->mb.i_mb_x];
            /* Don't propagate for an intra block. */
            if( propagate_amount > 0 )
            {
                /* Access width-2 bitfield. */
                int lists_used = frames[b]->lowres_costs[b-p0][p1-b][mb_index] >> LOWRES_COST_SHIFT;
                /* Follow the MVs to the previous frame(s). */
                for( int list = 0; list < 2; list++ )
                    if( (lists_used >> list)&1 )
                    {
#define CLIP_ADD(s,x) (s) = X264_MIN((s)+(x),(1<<16)-1)
                        int listamount = propagate_amount;
                        /* Apply bipred weighting. */
                        if( lists_used == 3 )
                            listamount = (listamount * bipred_weights[list] + 32) >> 6;

                        /* Early termination for simple case of mv0. */
                        if( !M32( mvs[list][mb_index] ) )
                        {
                            CLIP_ADD( ref_costs[list][mb_index], listamount );
                            continue;
                        }

                        int x = mvs[list][mb_index][0];
                        int y = mvs[list][mb_index][1];
                        int mbx = (x>>5)+h->mb.i_mb_x;
                        int mby = (y>>5)+h->mb.i_mb_y;
                        int idx0 = mbx + mby * h->mb.i_mb_stride;
                        int idx1 = idx0 + 1;
                        int idx2 = idx0 + h->mb.i_mb_stride;
                        int idx3 = idx0 + h->mb.i_mb_stride + 1;
                        x &= 31;
                        y &= 31;
                        int idx0weight = (32-y)*(32-x);
                        int idx1weight = (32-y)*x;
                        int idx2weight = y*(32-x);
                        int idx3weight = y*x;

                        /* We could just clip the MVs, but pixels that lie outside the frame probably shouldn't
                         * be counted. */
                        if( mbx < h->mb.i_mb_width-1 && mby < h->mb.i_mb_height-1 && mbx >= 0 && mby >= 0 )
                        {