Preliminary adaptive B-frame decision (not yet tuned).

Fix flushing of delayed frames when the encode finishes.


git-svn-id: svn://svn.videolan.org/x264/trunk@137 df754926-b1dd-0310-bc7b-ec298dee348c

common/common.c

@@ -64,6 +64,8 @@ void    x264_param_default( x264_param_t *param )
     param->i_keyint_min = 100;
     param->i_bframe = 0;
     param->i_scenecut_threshold = 40;
+    param->b_bframe_adaptive = 1;
+    param->i_bframe_bias = 0;
 
     param->b_deblocking_filter = 1;
     param->i_deblocking_filter_alphac0 = 0;

common/common.h

@@ -52,6 +52,8 @@
 #define X264_MIN(a,b) ( (a)<(b) ? (a) : (b) )
 #define X264_MAX(a,b) ( (a)>(b) ? (a) : (b) )
 #define X264_ABS(a)   ( (a)< 0 ? -(a) : (a) )
+#define X264_MIN3(a,b,c) X264_MIN((a),X264_MIN((b),(c)))
+#define X264_MIN4(a,b,c,d) X264_MIN((a),X264_MIN3((b),(c),(d)))
 
 /****************************************************************************
  * Generals functions
@@ -80,6 +82,23 @@ static inline float x264_clip3f( float v, float f_min, float f_max )
     return ( (v < f_min) ? f_min : (v > f_max) ? f_max : v );
 }
 
+static inline int x264_median( int a, int b, int c )
+{
+    int min = a, max =a;
+    if( b < min )
+        min = b;
+    else
+        max = b;    /* no need to do 'b > max' (more consuming than always doing affectation) */
+
+    if( c < min )
+        min = c;
+    else if( c > max )
+        max = c;
+
+    return a + b + c - min - max;
+}
+
+
 /****************************************************************************
  *
  ****************************************************************************/
@@ -215,6 +234,8 @@ struct x264_t
         x264_frame_t *next[X264_BFRAME_MAX+1];
         /* Unused frames */
         x264_frame_t *unused[X264_BFRAME_MAX+1];
+        /* For adaptive B decision */
+        x264_frame_t *last_nonb;
 
         /* frames used for reference +1 for decoding */
         x264_frame_t *reference[16+1];
@@ -222,6 +243,8 @@ struct x264_t
         int i_last_idr; /* Frame number of the last IDR */
 
         int i_input;    /* Number of input frames already accepted */
+        
+        int i_delay;    /* Number of frames buffered for B reordering */
     } frames;
 
     /* current frame being encoded */
@@ -266,6 +289,9 @@ struct x264_t
         int     i_mb_xy;
         int     i_b8_xy;
         int     i_b4_xy;
+        
+        /* Search parameters */
+        int     i_subpel_refine;
 
         /* Allowed qpel MV range to stay within the picture + emulated edge pixels */
         int     mv_min[2];

common/frame.c

@@ -73,9 +73,18 @@ x264_frame_t *x264_frame_new( x264_t *h )
                                         ( frame->i_lines[0] + 64 ) );
 
         frame->filtered[i+1] = ((uint8_t*)frame->buffer[4+i]) +
-                          frame->i_stride[0] * 32 + 32;
+                                frame->i_stride[0] * 32 + 32;
     }
 
+    frame->i_stride_lowres = frame->i_stride[0]/2;
+    frame->i_lines_lowres = frame->i_lines[0]/2;
+    for( i = 0; i < 4; i++ )
+    {
+        frame->buffer[7+i] = x264_malloc( frame->i_stride_lowres *
+                                        ( frame->i_lines[0]/2 + 64 ) );
+        frame->lowres[i] = ((uint8_t*)frame->buffer[7+i]) +
+                            frame->i_stride_lowres * 32 + 32;
+    }
 
     frame->i_poc = -1;
     frame->i_type = X264_TYPE_AUTO;
@@ -106,7 +115,7 @@ void x264_frame_delete( x264_frame_t *frame )
     {
         x264_free( frame->buffer[i] );
     }
-    for( i = 4; i < 7; i++ ) /* filtered planes */
+    for( i = 4; i < 11; i++ ) /* filtered planes */
     {
         x264_free( frame->buffer[i] );
     }
@@ -235,6 +244,43 @@ void x264_frame_expand_border_filtered( x264_frame_t *frame )
     }
 }
 
+void x264_frame_expand_lowres( x264_frame_t *frame )
+{
+    int w = 32;
+    int i, y;
+    for( i = 0; i < 4; i++ )
+    {
+        int i_stride = frame->i_stride_lowres;
+        int i_lines = frame->i_lines_lowres;
+#define PPIXEL(x, y) ( frame->lowres[i] + (x) +(y)*i_stride )
+        for( y = 0; y < w; y++ )
+        {
+            /* upper band */
+            memcpy( PPIXEL(0,-y-1), PPIXEL(0,0), i_stride - 2 * w);
+            /* up left corner */
+            memset( PPIXEL(-w,-y-1 ), PPIXEL(0,0)[0], w );
+            /* up right corner */
+            memset( PPIXEL(i_stride - 2*w,-y-1), PPIXEL( i_stride-1-2*w,0)[0], w );
+
+            /* lower band */
+            memcpy( PPIXEL(0, i_lines+y), PPIXEL(0,i_lines-1), i_stride - 2 * w );
+            /* low left corner */
+            memset( PPIXEL(-w, i_lines+y), PPIXEL(0,i_lines-1)[0], w);
+            /* low right corner */
+            memset( PPIXEL(i_stride-2*w, i_lines+y), PPIXEL(i_stride-1-2*w,i_lines-1)[0], w);
+
+        }
+        for( y = 0; y < i_lines; y++ )
+        {
+            /* left band */
+            memset( PPIXEL( -w, y ), PPIXEL( 0, y )[0], w );
+            /* right band */
+            memset( PPIXEL( i_stride-2*w, y ), PPIXEL( i_stride - 1-2*w, y )[0], w );
+        }
+#undef PPIXEL
+    }
+}
+
 
 /* FIXME theses tables are duplicated with the ones in macroblock.c */
 static const uint8_t block_idx_xy[4][4] =

common/frame.h

@@ -37,12 +37,15 @@ typedef struct
     int     i_plane;
     int     i_stride[4];
     int     i_lines[4];
+    int     i_stride_lowres;
+    int     i_lines_lowres;
     uint8_t *plane[4];
     uint8_t *filtered[4]; /* plane[0], H, V, HV */
+    uint8_t *lowres[4]; /* half-size copy of input frame: Orig, H, V, HV */
 
     /* for unrestricted mv we allocate more data than needed
      * allocated data are stored in buffer */
-    void    *buffer[7];
+    void    *buffer[11];
 
     /* motion data */
     int16_t (*mv[2])[2];
@@ -50,6 +53,12 @@ typedef struct
     int     i_ref[2];
     int     ref_poc[2][16];
 
+    /* for adaptive B-frame decision.
+     * contains the SATD cost of the lowres frame encoded in various modes
+     * FIXME: how big an array do we need? */
+    int     i_cost_est[16][16];
+    int     i_intra_mbs[16];
+
 } x264_frame_t;
 
 x264_frame_t *x264_frame_new( x264_t *h );
@@ -58,11 +67,12 @@ void          x264_frame_delete( x264_frame_t *frame );
 void          x264_frame_copy_picture( x264_t *h, x264_frame_t *dst, x264_picture_t *src );
 
 void          x264_frame_expand_border( x264_frame_t *frame );
-
 void          x264_frame_expand_border_filtered( x264_frame_t *frame );
+void          x264_frame_expand_border_lowres( x264_frame_t *frame );
 
 void          x264_frame_deblocking_filter( x264_t *h, int i_slice_type );
 
 void          x264_frame_filter( int cpu, x264_frame_t *frame );
+void          x264_frame_init_lowres( int cpu, x264_frame_t *frame );
 
 #endif

common/macroblock.c

@@ -166,22 +166,6 @@ void x264_mb_dequant_4x4( int16_t dct[4][4], int i_qscale )
     }
 }
 
-static inline int x264_median( int a, int b, int c )
-{
-    int min = a, max =a;
-    if( b < min )
-        min = b;
-    else
-        max = b;    /* no need to do 'b > max' (more consuming than always doing affectation) */
-
-    if( c < min )
-        min = c;
-    else if( c > max )
-        max = c;
-
-    return a + b + c - min - max;
-}
-
 void x264_mb_predict_mv( x264_t *h, int i_list, int idx, int i_width, int mvp[2] )
 {
     const int i8 = x264_scan8[idx];

common/mc.c

@@ -444,3 +444,40 @@ void x264_frame_filter( int cpu, x264_frame_t *frame )
         }
     }
 }
+
+void x264_frame_init_lowres( int cpu, x264_frame_t *frame )
+{
+    // FIXME: tapfilter?
+    const int i_stride = frame->i_stride[0];
+    const int i_stride2 = frame->i_stride_lowres;
+    const int i_width2 = i_stride2 - 64;
+    int x, y, i;
+    for( y = 0; y < frame->i_lines_lowres - 1; y++ )
+    {
+        uint8_t *src0 = &frame->plane[0][2*y*i_stride];
+        uint8_t *src1 = src0+i_stride;
+        uint8_t *src2 = src1+i_stride;
+        uint8_t *dst0 = &frame->lowres[0][y*i_stride2];
+        uint8_t *dsth = &frame->lowres[1][y*i_stride2];
+        uint8_t *dstv = &frame->lowres[2][y*i_stride2];
+        uint8_t *dstc = &frame->lowres[3][y*i_stride2];
+        for( x = 0; x < i_width2 - 1; x++ )
+        {
+            dst0[x] = (src0[2*x  ] + src0[2*x+1] + src1[2*x  ] + src1[2*x+1] + 2) >> 2;
+            dsth[x] = (src0[2*x+1] + src0[2*x+2] + src1[2*x+1] + src1[2*x+2] + 2) >> 2;
+            dstv[x] = (src1[2*x  ] + src1[2*x+1] + src2[2*x  ] + src2[2*x+1] + 2) >> 2;
+            dstc[x] = (src1[2*x+1] + src1[2*x+2] + src2[2*x+1] + src2[2*x+2] + 2) >> 2;
+        }
+        dst0[x] = (src0[2*x  ] + src0[2*x+1] + src1[2*x  ] + src1[2*x+1] + 2) >> 2;
+        dstv[x] = (src1[2*x  ] + src1[2*x+1] + src2[2*x  ] + src2[2*x+1] + 2) >> 2;
+        dsth[x] = (src0[2*x+1] + src1[2*x+1] + 1) >> 1;
+        dstc[x] = (src1[2*x+1] + src2[2*x+1] + 1) >> 1;
+    }
+    for( i = 0; i < 4; i++ )
+        memcpy( &frame->lowres[i][y*i_stride2], &frame->lowres[i][(y-1)*i_stride2], i_width2 );
+
+    for( y = 0; y < 16; y++ )
+        for( x = 0; x < 16; x++ )
+            frame->i_cost_est[x][y] = -1;
+}
+

encoder/analyse.c

@@ -176,6 +176,8 @@ static void x264_mb_analyse_init( x264_t *h, x264_mb_analysis_t *a, int i_qp )
         int i;
         int i_fmv_range = h->param.analyse.i_mv_range - 16;
 
+        h->mb.i_subpel_refine = h->param.analyse.i_subpel_refine;
+
         /* Calculate max allowed MV range */
 #define CLIP_FMV(mv) x264_clip3( mv, -i_fmv_range, i_fmv_range )
         h->mb.mv_min_fpel[0] = CLIP_FMV( -16*h->mb.i_mb_x - 8 );
@@ -1795,3 +1797,5 @@ void x264_macroblock_analyse( x264_t *h )
     }
 }
 
+#include "voptype_decision.c"
+

encoder/analyse.h

@@ -25,5 +25,6 @@
 #define _ANALYSE_H 1
 
 void x264_macroblock_analyse( x264_t *h );
+void x264_voptype_analyse( x264_t *h );
 
 #endif

encoder/encoder.c

@@ -341,7 +341,10 @@ x264_t *x264_encoder_open   ( x264_param_t *param )
     if( h->param.i_keyint_max <= 0 )
         h->param.i_keyint_max = 1;
     h->param.i_keyint_min = x264_clip3( h->param.i_keyint_min, 1, h->param.i_keyint_max/2+1 );
+
     h->param.i_bframe = x264_clip3( h->param.i_bframe, 0, X264_BFRAME_MAX );
+    h->param.i_bframe_bias = x264_clip3( h->param.i_bframe_bias, -90, 100 );
+    h->frames.i_delay = h->param.i_bframe;
 
     h->param.i_deblocking_filter_alphac0 = x264_clip3( h->param.i_deblocking_filter_alphac0, -6, 6 );
     h->param.i_deblocking_filter_beta    = x264_clip3( h->param.i_deblocking_filter_beta, -6, 6 );
@@ -424,7 +427,7 @@ x264_t *x264_encoder_open   ( x264_param_t *param )
         h->frames.next[i]    = NULL;
         h->frames.unused[i]  = NULL;
     }
-    for( i = 0; i < 1 + h->param.i_bframe; i++ )
+    for( i = 0; i < 1 + h->frames.i_delay; i++ )
     {
         h->frames.unused[i] =  x264_frame_new( h );
     }
@@ -678,6 +681,18 @@ static inline void x264_reference_update( x264_t *h )
     /* expand border of filtered images */
     x264_frame_expand_border_filtered( h->fdec );
 
+    /* move lowres copy of the image to the ref frame */
+    for( i = 0; i < 4; i++)
+    {
+        uint8_t *tmp = h->fdec->lowres[i];
+        h->fdec->lowres[i] = h->fenc->lowres[i];
+        h->fenc->lowres[i] = tmp;
+    }
+
+    /* adaptive B decision needs a pointer, since it can't use the ref lists */
+    if( h->sh.i_type != SLICE_TYPE_B )
+        h->frames.last_nonb = h->fdec;
+
     /* move frame in the buffer */
     h->fdec = h->frames.reference[h->param.i_frame_reference+1];
     for( i = h->param.i_frame_reference+1; i > 0; i-- )
@@ -911,7 +926,8 @@ static inline void x264_slice_write( x264_t *h, int i_nal_type, int i_nal_ref_id
  ****************************************************************************/
 int     x264_encoder_encode( x264_t *h,
                              x264_nal_t **pp_nal, int *pi_nal,
-                             x264_picture_t *pic )
+                             x264_picture_t *pic_in,
+                             x264_picture_t *pic_out )
 {
     x264_frame_t   *frame_psnr = h->fdec; /* just to keep the current decoded frame for psnr calculation */
     int     i_nal_type;
@@ -931,22 +947,24 @@ int     x264_encoder_encode( x264_t *h,
 
     /* ------------------- Setup new frame from picture -------------------- */
     TIMER_START( i_mtime_encode_frame );
-    if( pic != NULL )
+    if( pic_in != NULL )
     {
         /* 1: Copy the picture to a frame and move it to a buffer */
         x264_frame_t *fenc = x264_frame_get( h->frames.unused );
 
-        x264_frame_copy_picture( h, fenc, pic );
+        x264_frame_copy_picture( h, fenc, pic_in );
 
         fenc->i_frame = h->frames.i_input++;
 
         x264_frame_put( h->frames.next, fenc );
 
-        if( h->frames.i_input <= h->param.i_bframe )
+        x264_frame_init_lowres( h->param.cpu, fenc );
+
+        if( h->frames.i_input <= h->frames.i_delay )
         {
             /* Nothing yet to encode */
             /* waiting for filling bframe buffer */
-            pic->i_type = X264_TYPE_AUTO;
+            pic_out->i_type = X264_TYPE_AUTO;
             return 0;
         }
     }
@@ -967,6 +985,8 @@ int     x264_encoder_encode( x264_t *h,
                 h->frames.next[i]->i_type =
                     x264_ratecontrol_slice_type( h, h->frames.next[i]->i_frame );
         }
+        else if( h->param.i_bframe && h->param.b_bframe_adaptive )
+            x264_voptype_analyse( h );
 
         for( bframes = 0;; bframes++ )
         {
@@ -1027,7 +1047,7 @@ int     x264_encoder_encode( x264_t *h,
     {
         /* Nothing yet to encode (ex: waiting for I/P with B frames) */
         /* waiting for filling bframe buffer */
-        pic->i_type = X264_TYPE_AUTO;
+        pic_out->i_type = X264_TYPE_AUTO;
         return 0;
     }
 
@@ -1069,7 +1089,6 @@ do_encode:
         i_slice_type = SLICE_TYPE_B;
     }
 
-    pic->i_type     =
     h->fdec->i_type = h->fenc->i_type;
     h->fdec->i_poc  = h->fenc->i_poc;
     h->fdec->i_frame = h->fenc->i_frame;
@@ -1080,10 +1099,8 @@ do_encode:
     /* Init the rate control */
     x264_ratecontrol_start( h, i_slice_type );
     i_global_qp = x264_ratecontrol_qp( h );
-    if( h->fenc->i_qpplus1 > 0 )
-    {
-        i_global_qp = x264_clip3( h->fenc->i_qpplus1 - 1, 0, 51 );
-    }
+    pic_out->i_qpplus1 =
+    h->fdec->i_qpplus1 = i_global_qp + 1;
 
     /* build ref list 0/1 */
     x264_reference_build_list( h, h->fdec->i_poc );
@@ -1251,12 +1268,12 @@ do_encode:
 
     /* Set output picture properties */
     if( i_slice_type == SLICE_TYPE_I )
-        pic->i_type = i_nal_type == NAL_SLICE_IDR ? X264_TYPE_IDR : X264_TYPE_I;
+        pic_out->i_type = i_nal_type == NAL_SLICE_IDR ? X264_TYPE_IDR : X264_TYPE_I;
     else if( i_slice_type == SLICE_TYPE_P )
-        pic->i_type = X264_TYPE_P;
+        pic_out->i_type = X264_TYPE_P;
     else
-        pic->i_type = X264_TYPE_B;
-    pic->i_pts = h->fenc->i_pts;
+        pic_out->i_type = X264_TYPE_B;
+    pic_out->i_pts = h->fenc->i_pts;
 
     /* ---------------------- Update encoder state ------------------------- */
     /* update cabac */

encoder/me.c

@@ -99,7 +99,7 @@ void x264_me_search_ref( x264_t *h, x264_me_t *m, int (*mvc)[2], int i_mvc, int
     
     COST_MV( 0, 0 );
 
-    if( h->param.analyse.i_subpel_refine >= 2 )
+    if( h->mb.i_subpel_refine >= 2 )
     {
         /* hexagon search */
         /* Don't need to test mv_range each time, we won't go outside picture+padding */
@@ -157,7 +157,7 @@ void x264_me_search_ref( x264_t *h, x264_me_t *m, int (*mvc)[2], int i_mvc, int
             + m->cost_mv;
 
     /* subpel refine */
-    if( h->param.analyse.i_subpel_refine >= 3 )
+    if( h->mb.i_subpel_refine >= 3 )
     {
         int hpel, qpel;
 
@@ -172,8 +172,8 @@ void x264_me_search_ref( x264_t *h, x264_me_t *m, int (*mvc)[2], int i_mvc, int
                 *p_fullpel_thresh = m->cost;
         }
 
-        hpel = subpel_iterations[h->param.analyse.i_subpel_refine][2];
-        qpel = subpel_iterations[h->param.analyse.i_subpel_refine][3];
+        hpel = subpel_iterations[h->mb.i_subpel_refine][2];
+        qpel = subpel_iterations[h->mb.i_subpel_refine][3];
         refine_subpel( h, m, hpel, qpel );
     }
 }
@@ -181,8 +181,8 @@ void x264_me_search_ref( x264_t *h, x264_me_t *m, int (*mvc)[2], int i_mvc, int
 
 void x264_me_refine_qpel( x264_t *h, x264_me_t *m )
 {
-    int hpel = subpel_iterations[h->param.analyse.i_subpel_refine][0];
-    int qpel = subpel_iterations[h->param.analyse.i_subpel_refine][1];
+    int hpel = subpel_iterations[h->mb.i_subpel_refine][0];
+    int qpel = subpel_iterations[h->mb.i_subpel_refine][1];
 //  if( hpel || qpel )
 	refine_subpel( h, m, hpel, qpel );
 }

encoder/voptype_decision.c

+/*****************************************************************************
+ * voptype_decision.c: h264 encoder library
+ *****************************************************************************
+ * Copyright (C) 2005 Loren Merritt
+ *
+ * Authors: Loren Merritt <lorenm@u.washington.edu>
+ *
+ * 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., 59 Temple Place - Suite 330, Boston, MA  02111, USA.
+ *****************************************************************************/
+
+#include <stdlib.h>
+#include <string.h>
+#include <math.h>
+#include <limits.h>
+
+#include "common/common.h"
+#include "common/macroblock.h"
+#include "common/cpu.h"
+#include "macroblock.h"
+#include "me.h"
+
+
+static void x264_mb_analyse_load_costs_lowres( x264_t *h, x264_mb_analysis_t *a )
+{
+    static int16_t *p_cost_mv;
+
+    if( !p_cost_mv )
+    {
+        int i;
+        x264_cpu_restore( h->param.cpu );
+        p_cost_mv = x264_malloc( (2*2*h->param.analyse.i_mv_range + 1) * sizeof(int16_t) );
+        p_cost_mv += 2*h->param.analyse.i_mv_range;
+        for( i = 0; i <= 2*h->param.analyse.i_mv_range; i++ )
+            p_cost_mv[-i] =
+            p_cost_mv[i]  = (int)( a->i_lambda * (1 + 2*log(2*i+1)/log(2)) );
+    }
+
+    a->p_cost_mv = p_cost_mv;
+}
+
+
+int x264_voptype_mb_cost( x264_t *h, x264_mb_analysis_t *a,
+                          x264_frame_t **frames, int p0, int p1, int b )
+{
+    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->sps->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 dist_scale_factor = ( ((b-p0) << 8) + ((p1-p0) >> 1) ) / (p1-p0);
+
+    uint8_t pix1[9*9], pix2[8*8];
+    x264_me_t m[2];
+    int mvc[3][2], i_mvc;
+    int i_bcost = INT_MAX;
+    int i_cost_bak;
+    int l, i;
+
+#define SAVE_MVS( mv0, mv1 ) \
+    { \
+        fenc->mv[0][i_mb_xy][0] = mv0[0]; \
+        fenc->mv[0][i_mb_xy][1] = mv0[1]; \
+        fenc->mv[1][i_mb_xy][0] = mv1[0]; \
+        fenc->mv[1][i_mb_xy][1] = mv1[1]; \
+    }
+#define TRY_BIDIR( mv0, mv1, penalty ) \
+    { \
+        int stride2 = 8; \
+        uint8_t *src2; \
+        int i_cost; \
+        h->mc.mc_luma( m[0].p_fref, m[0].i_stride, pix1, 8, \
+                       (mv0)[0], (mv0)[1], 8, 8 ); \
+        src2 = h->mc.get_ref( m[1].p_fref, m[1].i_stride, pix2, &stride2, \
+                       (mv1)[0], (mv1)[1], 8, 8 ); \
+        h->pixf.avg[PIXEL_8x8]( pix1, 8, src2, stride2 ); \
+        i_cost = penalty + h->pixf.satd[PIXEL_8x8]( \
+                           m[0].p_fenc, m[0].i_stride, pix1, 8 ); \
+        if( i_bcost > i_cost ) \
+        { \
+            i_bcost = i_cost; \
+            SAVE_MVS( mv0, mv1 ); \
+        } \
+    }
+
+    m[0].i_pixel = PIXEL_8x8;
+    m[0].p_cost_mv = a->p_cost_mv;
+    m[0].i_stride = i_stride;
+    m[0].p_fenc = &fenc->lowres[0][ i_pel_offset ];
+    LOAD_HPELS( m[0].p_fref, fref0->lowres, i_pel_offset );
+
+    if( b_bidir )
+    {
+        int16_t *mvr = fref1->mv[0][i_mb_xy];
+        int dmv[2][2];
+        int mv0[2] = {0,0};
+
+        m[1] = m[0];
+        LOAD_HPELS( m[1].p_fref, fref1->lowres, i_pel_offset );
+
+        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];
+
+        TRY_BIDIR( dmv[0], dmv[1], 0 );
+        TRY_BIDIR( mv0, mv0, 0 );
+//      if( i_bcost < 60 ) // arbitrary threshold
+//          return i_bcost;
+    }
+
+    i_cost_bak = i_bcost;
+    for( l = 0; l < 1 + b_bidir; l++ )
+    {
+        int16_t (*fenc_mv)[2] = &fenc->mv[0][i_mb_xy];
+        mvc[0][0] = fenc_mv[-1][0];
+        mvc[0][1] = fenc_mv[-1][1];
+        mvc[1][0] = fenc_mv[-i_mb_stride][0];
+        mvc[1][1] = fenc_mv[-i_mb_stride][1];
+        mvc[2][0] = fenc_mv[-i_mb_stride+1][0];
+        mvc[2][1] = fenc_mv[-i_mb_stride+1][1];
+        mvc[3][0] = fenc_mv[-i_mb_stride-1][0];
+        mvc[3][1] = fenc_mv[-i_mb_stride-1][1];
+        m[l].mvp[0] = x264_median( mvc[0][0], mvc[1][0], mvc[2][0] );
+        m[l].mvp[1] = x264_median( mvc[0][1], mvc[1][1], mvc[2][1] );
+        i_mvc = 4;
+
+        x264_me_search( h, &m[l], mvc, i_mvc );
+
+        i_bcost = X264_MIN( i_bcost, m[l].cost + 3 );
+    }
+
+    if( b_bidir )
+        TRY_BIDIR( m[0].mv, m[1].mv, 5 );
+
+    if( i_bcost < i_cost_bak )
+        SAVE_MVS( m[0].mv, m[1].mv );
+
+    /* intra */
+//  if( i_bcost > 100 ) // arbitrary threshold
+    {
+        uint8_t *src = &fenc->lowres[0][ i_pel_offset - i_stride - 1 ];
+        int intra_penalty = 5 + 10 * b_bidir;
+        i_cost_bak = i_bcost;
+
+        memcpy( pix1, src, 9 );
+        for( i=1; i<9; i++, src += i_stride )
+            pix1[i*9] = src[0];
+        src =