X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libavcodec%2Fmotion_est.c;h=16db0f87c9e75511225636b3db4a229dbc8b850e;hb=f0a838e52505fb48d15fb23554107cbd282f8132;hp=a43cd13619525a08af7c98e560be54ffa139eb81;hpb=155ec6edf82692bcf3a5f87d2bc697404f4e5aaf;p=ffmpeg

diff --git a/libavcodec/motion_est.c b/libavcodec/motion_est.c
index a43cd136195..16db0f87c9e 100644
--- a/libavcodec/motion_est.c
+++ b/libavcodec/motion_est.c
@@ -1,31 +1,32 @@
 /*
- * Motion estimation 
+ * Motion estimation
  * Copyright (c) 2000,2001 Fabrice Bellard.
  * Copyright (c) 2002-2004 Michael Niedermayer
- * 
  *
- * This library is free software; you can redistribute it and/or
+ * new motion estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
  * modify it under the terms of the GNU Lesser General Public
  * License as published by the Free Software Foundation; either
- * version 2 of the License, or (at your option) any later version.
+ * version 2.1 of the License, or (at your option) any later version.
  *
- * This library is distributed in the hope that it will be useful,
+ * FFmpeg 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
  * Lesser General Public License for more details.
  *
  * You should have received a copy of the GNU Lesser General Public
- * License along with this library; if not, write to the Free Software
- * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
- *
- * new Motion Estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  */
- 
+
 /**
  * @file motion_est.c
  * Motion estimation.
  */
- 
+
 #include <stdlib.h>
 #include <stdio.h>
 #include <limits.h>
@@ -45,7 +46,7 @@
 #define P_MV1 P[9]
 
 static inline int sad_hpel_motion_search(MpegEncContext * s,
-				  int *mx_ptr, int *my_ptr, int dmin,
+                                  int *mx_ptr, int *my_ptr, int dmin,
                                   int src_index, int ref_index,
                                   int size, int h);
 
@@ -69,7 +70,7 @@ typedef struct Minima{
 static int minima_cmp(const void *a, const void *b){
     const Minima *da = (const Minima *) a;
     const Minima *db = (const Minima *) b;
-    
+
     return da->height - db->height;
 }
 
@@ -97,11 +98,14 @@ static inline void init_ref(MotionEstContext *c, uint8_t *src[3], uint8_t *ref[3
 
 static int get_flags(MotionEstContext *c, int direct, int chroma){
     return   ((c->avctx->flags&CODEC_FLAG_QPEL) ? FLAG_QPEL : 0)
-           + (direct ? FLAG_DIRECT : 0) 
+           + (direct ? FLAG_DIRECT : 0)
            + (chroma ? FLAG_CHROMA : 0);
 }
 
-static always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
+/*! \brief compares a block (either a full macroblock or a partition thereof)
+    against a proposed motion-compensated prediction of that block
+ */
+static av_always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
                       const int size, const int h, int ref_index, int src_index,
                       me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
     MotionEstContext * const c= &s->me;
@@ -117,6 +121,7 @@ static always_inline int cmp(MpegEncContext *s, const int x, const int y, const
     int d;
     //FIXME check chroma 4mv, (no crashes ...)
     if(flags&FLAG_DIRECT){
+        assert(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1));
         if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
             const int time_pp= s->pp_time;
             const int time_pb= s->pb_time;
@@ -130,7 +135,7 @@ static always_inline int cmp(MpegEncContext *s, const int x, const int y, const
                     int by = hy ? fy - c->co_located_mv[i][1] : c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(qpel+4));
                     int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
                     int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
-        
+
                     uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
                     if(qpel){
                         c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
@@ -147,7 +152,7 @@ static always_inline int cmp(MpegEncContext *s, const int x, const int y, const
                 int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
                 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
                 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
-                
+
                 if(qpel){
                     c->qpel_put[1][fxy](c->temp               , ref[0] + (fx>>2) + (fy>>2)*stride               , stride);
                     c->qpel_put[1][fxy](c->temp + 8           , ref[0] + (fx>>2) + (fy>>2)*stride + 8           , stride);
@@ -157,7 +162,7 @@ static always_inline int cmp(MpegEncContext *s, const int x, const int y, const
                     c->qpel_avg[1][bxy](c->temp + 8           , ref[8] + (bx>>2) + (by>>2)*stride + 8           , stride);
                     c->qpel_avg[1][bxy](c->temp     + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride     + 8*stride, stride);
                     c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
-                }else{            
+                }else{
                     assert((fx>>1) + 16*s->mb_x >= -16);
                     assert((fy>>1) + 16*s->mb_y >= -16);
                     assert((fx>>1) + 16*s->mb_x <= s->width);
@@ -175,7 +180,7 @@ static always_inline int cmp(MpegEncContext *s, const int x, const int y, const
         }else
             d= 256*256*256*32;
     }else{
-        int uvdxy;
+        int uvdxy;              /* no, it might not be used uninitialized */
         if(dxy){
             if(qpel){
                 c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
@@ -192,9 +197,9 @@ static always_inline int cmp(MpegEncContext *s, const int x, const int y, const
                 if(chroma)
                     uvdxy= dxy | (x&1) | (2*(y&1));
             }
-            d = cmp_func(s, c->temp, src[0], stride, h); 
+            d = cmp_func(s, c->temp, src[0], stride, h);
         }else{
-            d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h); 
+            d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
             if(chroma)
                 uvdxy= (x&1) + 2*(y&1);
         }
@@ -202,8 +207,8 @@ static always_inline int cmp(MpegEncContext *s, const int x, const int y, const
             uint8_t * const uvtemp= c->temp + 16*stride;
             c->hpel_put[size+1][uvdxy](uvtemp  , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
             c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
-            d += chroma_cmp_func(s, uvtemp  , src[1], uvstride, h>>1); 
-            d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1); 
+            d += chroma_cmp_func(s, uvtemp  , src[1], uvstride, h>>1);
+            d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
         }
     }
 #if 0
@@ -228,13 +233,19 @@ static void zero_hpel(uint8_t *a, const uint8_t *b, int stride, int h){
 
 void ff_init_me(MpegEncContext *s){
     MotionEstContext * const c= &s->me;
+    int cache_size= FFMIN(ME_MAP_SIZE>>ME_MAP_SHIFT, 1<<ME_MAP_SHIFT);
+    int dia_size= FFMAX(FFABS(s->avctx->dia_size)&255, FFABS(s->avctx->pre_dia_size)&255);
     c->avctx= s->avctx;
 
+    if(cache_size < 2*dia_size && !c->stride){
+        av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n");
+    }
+
     ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, c->avctx->me_pre_cmp);
     ff_set_cmp(&s->dsp, s->dsp.me_cmp, c->avctx->me_cmp);
     ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, c->avctx->me_sub_cmp);
     ff_set_cmp(&s->dsp, s->dsp.mb_cmp, c->avctx->mb_cmp);
-    
+
     c->flags    = get_flags(c, 0, c->avctx->me_cmp    &FF_CMP_CHROMA);
     c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
     c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp    &FF_CMP_CHROMA);
@@ -248,8 +259,8 @@ void ff_init_me(MpegEncContext *s){
     }else{
         if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
             c->sub_motion_search= hpel_motion_search;
-        else if(   c->avctx->me_sub_cmp == FF_CMP_SAD 
-                && c->avctx->    me_cmp == FF_CMP_SAD 
+        else if(   c->avctx->me_sub_cmp == FF_CMP_SAD
+                && c->avctx->    me_cmp == FF_CMP_SAD
                 && c->avctx->    mb_cmp == FF_CMP_SAD)
             c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
         else
@@ -260,16 +271,18 @@ void ff_init_me(MpegEncContext *s){
     else               c->hpel_put= s->dsp.put_pixels_tab;
 
     if(s->linesize){
-        c->stride  = s->linesize; 
+        c->stride  = s->linesize;
         c->uvstride= s->uvlinesize;
     }else{
         c->stride  = 16*s->mb_width + 32;
         c->uvstride=  8*s->mb_width + 16;
     }
 
-    // 8x8 fullpel search would need a 4x4 chroma compare, which we dont have yet, and even if we had the motion estimation code doesnt expect it
+    /* 8x8 fullpel search would need a 4x4 chroma compare, which we do
+     * not have yet, and even if we had, the motion estimation code
+     * does not expect it. */
     if(s->codec_id != CODEC_ID_SNOW){
-        if((c->avctx->me_cmp&FF_CMP_CHROMA) && !s->dsp.me_cmp[2]){
+        if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){
             s->dsp.me_cmp[2]= zero_cmp;
         }
         if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){
@@ -279,9 +292,13 @@ void ff_init_me(MpegEncContext *s){
         c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
     }
 
+    if(s->codec_id == CODEC_ID_H261){
+        c->sub_motion_search= no_sub_motion_search;
+    }
+
     c->temp= c->scratchpad;
 }
-      
+
 #if 0
 static int pix_dev(uint8_t * pix, int line_size, int mean)
 {
@@ -289,30 +306,31 @@ static int pix_dev(uint8_t * pix, int line_size, int mean)
 
     s = 0;
     for (i = 0; i < 16; i++) {
-	for (j = 0; j < 16; j += 8) {
-	    s += ABS(pix[0]-mean);
-	    s += ABS(pix[1]-mean);
-	    s += ABS(pix[2]-mean);
-	    s += ABS(pix[3]-mean);
-	    s += ABS(pix[4]-mean);
-	    s += ABS(pix[5]-mean);
-	    s += ABS(pix[6]-mean);
-	    s += ABS(pix[7]-mean);
-	    pix += 8;
-	}
-	pix += line_size - 16;
+        for (j = 0; j < 16; j += 8) {
+            s += FFABS(pix[0]-mean);
+            s += FFABS(pix[1]-mean);
+            s += FFABS(pix[2]-mean);
+            s += FFABS(pix[3]-mean);
+            s += FFABS(pix[4]-mean);
+            s += FFABS(pix[5]-mean);
+            s += FFABS(pix[6]-mean);
+            s += FFABS(pix[7]-mean);
+            pix += 8;
+        }
+        pix += line_size - 16;
     }
     return s;
 }
 #endif
 
 static inline void no_motion_search(MpegEncContext * s,
-				    int *mx_ptr, int *my_ptr)
+                                    int *mx_ptr, int *my_ptr)
 {
     *mx_ptr = 16 * s->mb_x;
     *my_ptr = 16 * s->mb_y;
 }
 
+#if 0  /* the use of these functions is inside #if 0 */
 static int full_motion_search(MpegEncContext * s,
                               int *mx_ptr, int *my_ptr, int range,
                               int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
@@ -323,35 +341,35 @@ static int full_motion_search(MpegEncContext * s,
 
     xx = 16 * s->mb_x;
     yy = 16 * s->mb_y;
-    x1 = xx - range + 1;	/* we loose one pixel to avoid boundary pb with half pixel pred */
+    x1 = xx - range + 1;        /* we loose one pixel to avoid boundary pb with half pixel pred */
     if (x1 < xmin)
-	x1 = xmin;
+        x1 = xmin;
     x2 = xx + range - 1;
     if (x2 > xmax)
-	x2 = xmax;
+        x2 = xmax;
     y1 = yy - range + 1;
     if (y1 < ymin)
-	y1 = ymin;
+        y1 = ymin;
     y2 = yy + range - 1;
     if (y2 > ymax)
-	y2 = ymax;
+        y2 = ymax;
     pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
     dmin = 0x7fffffff;
     mx = 0;
     my = 0;
     for (y = y1; y <= y2; y++) {
-	for (x = x1; x <= x2; x++) {
-	    d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x,
-			     s->linesize, 16);
-	    if (d < dmin ||
-		(d == dmin &&
-		 (abs(x - xx) + abs(y - yy)) <
-		 (abs(mx - xx) + abs(my - yy)))) {
-		dmin = d;
-		mx = x;
-		my = y;
-	    }
-	}
+        for (x = x1; x <= x2; x++) {
+            d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x,
+                             s->linesize, 16);
+            if (d < dmin ||
+                (d == dmin &&
+                 (abs(x - xx) + abs(y - yy)) <
+                 (abs(mx - xx) + abs(my - yy)))) {
+                dmin = d;
+                mx = x;
+                my = y;
+            }
+        }
     }
 
     *mx_ptr = mx;
@@ -359,8 +377,8 @@ static int full_motion_search(MpegEncContext * s,
 
 #if 0
     if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
-	*my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
-	fprintf(stderr, "error %d %d\n", *mx_ptr, *my_ptr);
+        *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
+        av_log(NULL, AV_LOG_ERROR, "error %d %d\n", *mx_ptr, *my_ptr);
     }
 #endif
     return dmin;
@@ -381,22 +399,22 @@ static int log_motion_search(MpegEncContext * s,
     /* Left limit */
     x1 = xx - range;
     if (x1 < xmin)
-	x1 = xmin;
+        x1 = xmin;
 
     /* Right limit */
     x2 = xx + range;
     if (x2 > xmax)
-	x2 = xmax;
+        x2 = xmax;
 
     /* Upper limit */
     y1 = yy - range;
     if (y1 < ymin)
-	y1 = ymin;
+        y1 = ymin;
 
     /* Lower limit */
     y2 = yy + range;
     if (y2 > ymax)
-	y2 = ymax;
+        y2 = ymax;
 
     pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
     dmin = 0x7fffffff;
@@ -404,39 +422,39 @@ static int log_motion_search(MpegEncContext * s,
     my = 0;
 
     do {
-	for (y = y1; y <= y2; y += range) {
-	    for (x = x1; x <= x2; x += range) {
-		d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
-		if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
-		    dmin = d;
-		    mx = x;
-		    my = y;
-		}
-	    }
-	}
-
-	range = range >> 1;
-
-	x1 = mx - range;
-	if (x1 < xmin)
-	    x1 = xmin;
-
-	x2 = mx + range;
-	if (x2 > xmax)
-	    x2 = xmax;
-
-	y1 = my - range;
-	if (y1 < ymin)
-	    y1 = ymin;
-
-	y2 = my + range;
-	if (y2 > ymax)
-	    y2 = ymax;
+        for (y = y1; y <= y2; y += range) {
+            for (x = x1; x <= x2; x += range) {
+                d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
+                if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
+                    dmin = d;
+                    mx = x;
+                    my = y;
+                }
+            }
+        }
+
+        range = range >> 1;
+
+        x1 = mx - range;
+        if (x1 < xmin)
+            x1 = xmin;
+
+        x2 = mx + range;
+        if (x2 > xmax)
+            x2 = xmax;
+
+        y1 = my - range;
+        if (y1 < ymin)
+            y1 = ymin;
+
+        y2 = my + range;
+        if (y2 > ymax)
+            y2 = ymax;
 
     } while (range >= 1);
 
 #ifdef DEBUG
-    fprintf(stderr, "log       - MX: %d\tMY: %d\n", mx, my);
+    av_log(s->avctx, AV_LOG_DEBUG, "log       - MX: %d\tMY: %d\n", mx, my);
 #endif
     *mx_ptr = mx;
     *my_ptr = my;
@@ -457,22 +475,22 @@ static int phods_motion_search(MpegEncContext * s,
     /* Left limit */
     x1 = xx - range;
     if (x1 < xmin)
-	x1 = xmin;
+        x1 = xmin;
 
     /* Right limit */
     x2 = xx + range;
     if (x2 > xmax)
-	x2 = xmax;
+        x2 = xmax;
 
     /* Upper limit */
     y1 = yy - range;
     if (y1 < ymin)
-	y1 = ymin;
+        y1 = ymin;
 
     /* Lower limit */
     y2 = yy + range;
     if (y2 > ymax)
-	y2 = ymax;
+        y2 = ymax;
 
     pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
     mx = 0;
@@ -484,48 +502,48 @@ static int phods_motion_search(MpegEncContext * s,
         dminx = 0x7fffffff;
         dminy = 0x7fffffff;
 
-	lastx = x;
-	for (x = x1; x <= x2; x += range) {
-	    d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
-	    if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
-		dminx = d;
-		mx = x;
-	    }
-	}
-
-	x = lastx;
-	for (y = y1; y <= y2; y += range) {
-	    d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
-	    if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
-		dminy = d;
-		my = y;
-	    }
-	}
-
-	range = range >> 1;
-
-	x = mx;
-	y = my;
-	x1 = mx - range;
-	if (x1 < xmin)
-	    x1 = xmin;
-
-	x2 = mx + range;
-	if (x2 > xmax)
-	    x2 = xmax;
-
-	y1 = my - range;
-	if (y1 < ymin)
-	    y1 = ymin;
-
-	y2 = my + range;
-	if (y2 > ymax)
-	    y2 = ymax;
+        lastx = x;
+        for (x = x1; x <= x2; x += range) {
+            d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
+            if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
+                dminx = d;
+                mx = x;
+            }
+        }
+
+        x = lastx;
+        for (y = y1; y <= y2; y += range) {
+            d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
+            if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
+                dminy = d;
+                my = y;
+            }
+        }
+
+        range = range >> 1;
+
+        x = mx;
+        y = my;
+        x1 = mx - range;
+        if (x1 < xmin)
+            x1 = xmin;
+
+        x2 = mx + range;
+        if (x2 > xmax)
+            x2 = xmax;
+
+        y1 = my - range;
+        if (y1 < ymin)
+            y1 = ymin;
+
+        y2 = my + range;
+        if (y2 > ymax)
+            y2 = ymax;
 
     } while (range >= 1);
 
 #ifdef DEBUG
-    fprintf(stderr, "phods     - MX: %d\tMY: %d\n", mx, my);
+    av_log(s->avctx, AV_LOG_DEBUG, "phods     - MX: %d\tMY: %d\n", mx, my);
 #endif
 
     /* half pixel search */
@@ -533,7 +551,7 @@ static int phods_motion_search(MpegEncContext * s,
     *my_ptr = my;
     return dminy;
 }
-
+#endif /* 0 */
 
 #define Z_THRESHOLD 256
 
@@ -545,7 +563,7 @@ static int phods_motion_search(MpegEncContext * s,
 }
 
 static inline int sad_hpel_motion_search(MpegEncContext * s,
-				  int *mx_ptr, int *my_ptr, int dmin,
+                                  int *mx_ptr, int *my_ptr, int dmin,
                                   int src_index, int ref_index,
                                   int size, int h)
 {
@@ -556,7 +574,7 @@ static inline int sad_hpel_motion_search(MpegEncContext * s,
     int stride= c->stride;
     const int flags= c->sub_flags;
     LOAD_COMMON
-    
+
     assert(flags == 0);
 
     if(c->skip){
@@ -566,19 +584,19 @@ static inline int sad_hpel_motion_search(MpegEncContext * s,
         return dmin;
     }
 //    printf("N");
-        
+
     pix = c->src[src_index][0];
 
     mx = *mx_ptr;
     my = *my_ptr;
     ptr = c->ref[ref_index][0] + (my * stride) + mx;
-    
+
     dminh = dmin;
 
-    if (mx > xmin && mx < xmax && 
+    if (mx > xmin && mx < xmax &&
         my > ymin && my < ymax) {
         int dx=0, dy=0;
-        int d, pen_x, pen_y; 
+        int d, pen_x, pen_y;
         const int index= (my<<ME_MAP_SHIFT) + mx;
         const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
         const int l= score_map[(index- 1               )&(ME_MAP_SIZE-1)];
@@ -587,7 +605,7 @@ static inline int sad_hpel_motion_search(MpegEncContext * s,
         mx<<=1;
         my<<=1;
 
-        
+
         pen_x= pred_x + mx;
         pen_y= pred_y + my;
 
@@ -655,11 +673,11 @@ static inline int sad_hpel_motion_search(MpegEncContext * s,
 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
 {
     const int xy= s->mb_x + s->mb_y*s->mb_stride;
-    
+
     s->p_mv_table[xy][0] = mx;
     s->p_mv_table[xy][1] = my;
 
-    /* has allready been set to the 4 MV if 4MV is done */
+    /* has already been set to the 4 MV if 4MV is done */
     if(mv4){
         int mot_xy= s->block_index[0];
 
@@ -682,6 +700,7 @@ static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
 static inline void get_limits(MpegEncContext *s, int x, int y)
 {
     MotionEstContext * const c= &s->me;
+    int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
 /*
     if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
     else                   c->range= 16;
@@ -691,12 +710,24 @@ static inline void get_limits(MpegEncContext *s, int x, int y)
         c->ymin = - y - 16;
         c->xmax = - x + s->mb_width *16;
         c->ymax = - y + s->mb_height*16;
+    } else if (s->out_format == FMT_H261){
+        // Search range of H261 is different from other codec standards
+        c->xmin = (x > 15) ? - 15 : 0;
+        c->ymin = (y > 15) ? - 15 : 0;
+        c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
+        c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
     } else {
         c->xmin = - x;
         c->ymin = - y;
         c->xmax = - x + s->mb_width *16 - 16;
         c->ymax = - y + s->mb_height*16 - 16;
     }
+    if(range){
+        c->xmin = FFMAX(c->xmin,-range);
+        c->xmax = FFMIN(c->xmax, range);
+        c->ymin = FFMAX(c->ymin,-range);
+        c->ymax = FFMIN(c->ymax, range);
+    }
 }
 
 static inline void init_mv4_ref(MotionEstContext *c){
@@ -720,11 +751,10 @@ static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
     int dmin_sum=0, mx4_sum=0, my4_sum=0;
     int same=1;
     const int stride= c->stride;
-    const int uvstride= c->uvstride;
     uint8_t *mv_penalty= c->current_mv_penalty;
 
     init_mv4_ref(c);
-    
+
     for(block=0; block<4; block++){
         int mx4, my4;
         int pred_x4, pred_y4;
@@ -751,7 +781,7 @@ static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
             if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
             if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
             if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
-    
+
             P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
             P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
 
@@ -764,7 +794,7 @@ static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
         dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
 
         dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
-        
+
         if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
             int dxy;
             const int offset= ((block&1) + (block>>1)*stride)*8;
@@ -797,20 +827,20 @@ static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
             mx4_sum+= mx4;
             my4_sum+= my4;
         }
-            
+
         s->current_picture.motion_val[0][ s->block_index[block] ][0]= mx4;
         s->current_picture.motion_val[0][ s->block_index[block] ][1]= my4;
 
         if(mx4 != mx || my4 != my) same=0;
     }
-    
+
     if(same)
         return INT_MAX;
-    
+
     if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
         dmin_sum += s->dsp.mb_cmp[0](s, s->new_picture.data[0] + s->mb_x*16 + s->mb_y*16*stride, c->scratchpad, stride, 16);
     }
-    
+
     if(c->avctx->mb_cmp&FF_CMP_CHROMA){
         int dxy;
         int mx, my;
@@ -819,9 +849,9 @@ static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
         mx= ff_h263_round_chroma(mx4_sum);
         my= ff_h263_round_chroma(my4_sum);
         dxy = ((my & 1) << 1) | (mx & 1);
-        
+
         offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
-       
+
         if(s->no_rounding){
             s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad    , s->last_picture.data[1] + offset, s->uvlinesize, 8);
             s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad+8  , s->last_picture.data[2] + offset, s->uvlinesize, 8);
@@ -833,7 +863,7 @@ static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
         dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.data[1] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, c->scratchpad  , s->uvlinesize, 8);
         dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.data[2] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, c->scratchpad+8, s->uvlinesize, 8);
     }
-    
+
     c->pred_x= mx;
     c->pred_y= my;
 
@@ -860,7 +890,7 @@ static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
     }
 }
 
-static int interlaced_search(MpegEncContext *s, int ref_index, 
+static int interlaced_search(MpegEncContext *s, int ref_index,
                              int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
 {
     MotionEstContext * const c= &s->me;
@@ -871,17 +901,16 @@ static int interlaced_search(MpegEncContext *s, int ref_index,
     uint8_t * const mv_penalty= c->current_mv_penalty;
     int same=1;
     const int stride= 2*s->linesize;
-    const int uvstride= 2*s->uvlinesize;
     int dmin_sum= 0;
     const int mot_stride= s->mb_stride;
     const int xy= s->mb_x + s->mb_y*mot_stride;
-    
+
     c->ymin>>=1;
     c->ymax>>=1;
     c->stride<<=1;
     c->uvstride<<=1;
     init_interlaced_ref(s, ref_index);
-    
+
     for(block=0; block<2; block++){
         int field_select;
         int best_dmin= INT_MAX;
@@ -890,19 +919,21 @@ static int interlaced_search(MpegEncContext *s, int ref_index,
         for(field_select=0; field_select<2; field_select++){
             int dmin, mx_i, my_i;
             int16_t (*mv_table)[2]= mv_tables[block][field_select];
-            
+
             if(user_field_select){
+                assert(field_select==0 || field_select==1);
+                assert(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
                 if(field_select_tables[block][xy] != field_select)
                     continue;
             }
-            
+
             P_LEFT[0] = mv_table[xy - 1][0];
             P_LEFT[1] = mv_table[xy - 1][1];
             if(P_LEFT[0]       > (c->xmax<<1)) P_LEFT[0]       = (c->xmax<<1);
-            
+
             c->pred_x= P_LEFT[0];
             c->pred_y= P_LEFT[1];
-            
+
             if(!s->first_slice_line){
                 P_TOP[0]      = mv_table[xy - mot_stride][0];
                 P_TOP[1]      = mv_table[xy - mot_stride][1];
@@ -912,20 +943,20 @@ static int interlaced_search(MpegEncContext *s, int ref_index,
                 if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
                 if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
                 if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
-    
+
                 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
                 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
             }
             P_MV1[0]= mx; //FIXME not correct if block != field_select
             P_MV1[1]= my / 2;
-            
+
             dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
 
             dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
-            
+
             mv_table[xy][0]= mx_i;
             mv_table[xy][1]= my_i;
-            
+
             if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
                 int dxy;
 
@@ -942,9 +973,9 @@ static int interlaced_search(MpegEncContext *s, int ref_index,
                 dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
             }else
                 dmin+= c->mb_penalty_factor; //field_select bits
-                
+
             dmin += field_select != block; //slightly prefer same field
-            
+
             if(dmin < best_dmin){
                 best_dmin= dmin;
                 best_field= field_select;
@@ -955,14 +986,14 @@ static int interlaced_search(MpegEncContext *s, int ref_index,
 
             if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
             if(mv_table[xy][1]&1) same=0;
-            if(mv_table[xy][1]*2 != my) same=0; 
+            if(mv_table[xy][1]*2 != my) same=0;
             if(best_field != block) same=0;
         }
 
         field_select_tables[block][xy]= best_field;
         dmin_sum += best_dmin;
     }
-    
+
     c->ymin<<=1;
     c->ymax<<=1;
     c->stride>>=1;
@@ -970,7 +1001,7 @@ static int interlaced_search(MpegEncContext *s, int ref_index,
 
     if(same)
         return INT_MAX;
-    
+
     switch(c->avctx->mb_cmp&0xFF){
     /*case FF_CMP_SSE:
         return dmin_sum+ 32*s->qscale*s->qscale;*/
@@ -984,7 +1015,7 @@ static int interlaced_search(MpegEncContext *s, int ref_index,
 static void clip_input_mv(MpegEncContext * s, int16_t *mv, int interlaced){
     int ymax= s->me.ymax>>interlaced;
     int ymin= s->me.ymin>>interlaced;
-    
+
     if(mv[0] < s->me.xmin) mv[0] = s->me.xmin;
     if(mv[0] > s->me.xmax) mv[0] = s->me.xmax;
     if(mv[1] <       ymin) mv[1] =       ymin;
@@ -1004,13 +1035,13 @@ static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int
     int d=0;
     me_cmp_func cmpf= s->dsp.sse[0];
     me_cmp_func chroma_cmpf= s->dsp.sse[1];
-    
+
     if(p_type && USES_LIST(mb_type, 1)){
         av_log(c->avctx, AV_LOG_ERROR, "backward motion vector in P frame\n");
-        return INT_MAX;
+        return INT_MAX/2;
     }
     assert(IS_INTRA(mb_type) || USES_LIST(mb_type,0) || USES_LIST(mb_type,1));
-    
+
     for(i=0; i<4; i++){
         int xy= s->block_index[i];
         clip_input_mv(s, p->motion_val[0][xy], !!IS_INTERLACED(mb_type));
@@ -1022,10 +1053,10 @@ static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int
         s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
         c->stride<<=1;
         c->uvstride<<=1;
-        
+
         if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){
             av_log(c->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");
-            return INT_MAX;
+            return INT_MAX/2;
         }
 
         if(USES_LIST(mb_type, 0)){
@@ -1049,10 +1080,10 @@ static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int
                 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_FORWARD_I;
             }
 
-            x= p->motion_val[0][xy ][0]; 
+            x= p->motion_val[0][xy ][0];
             y= p->motion_val[0][xy ][1];
             d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0, 0, cmpf, chroma_cmpf, flags);
-            x= p->motion_val[0][xy2][0]; 
+            x= p->motion_val[0][xy2][0];
             y= p->motion_val[0][xy2][1];
             d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1, 1, cmpf, chroma_cmpf, flags);
         }
@@ -1073,10 +1104,10 @@ static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int
                 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BACKWARD_I;
             }
 
-            x= p->motion_val[1][xy ][0]; 
+            x= p->motion_val[1][xy ][0];
             y= p->motion_val[1][xy ][1];
             d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0+2, 0, cmpf, chroma_cmpf, flags);
-            x= p->motion_val[1][xy2][0]; 
+            x= p->motion_val[1][xy2][0];
             y= p->motion_val[1][xy2][1];
             d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1+2, 1, cmpf, chroma_cmpf, flags);
             //FIXME bidir scores
@@ -1086,14 +1117,14 @@ static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int
     }else if(IS_8X8(mb_type)){
         if(!(s->flags & CODEC_FLAG_4MV)){
             av_log(c->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");
-            return INT_MAX;
+            return INT_MAX/2;
         }
         cmpf= s->dsp.sse[1];
         chroma_cmpf= s->dsp.sse[1];
         init_mv4_ref(c);
         for(i=0; i<4; i++){
             xy= s->block_index[i];
-            x= p->motion_val[0][xy][0]; 
+            x= p->motion_val[0][xy][0];
             y= p->motion_val[0][xy][1];
             d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);
         }
@@ -1111,14 +1142,14 @@ static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int
                 *(uint32_t*)s->b_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
                 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_FORWARD;
             }
-            x= p->motion_val[0][xy][0]; 
+            x= p->motion_val[0][xy][0];
             y= p->motion_val[0][xy][1];
             d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 0, 0, cmpf, chroma_cmpf, flags);
         }else if(USES_LIST(mb_type, 1)){
             *(uint32_t*)s->b_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
             s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BACKWARD;
-           
-            x= p->motion_val[1][xy][0]; 
+
+            x= p->motion_val[1][xy][0];
             y= p->motion_val[1][xy][1];
             d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 2, 0, cmpf, chroma_cmpf, flags);
         }else
@@ -1132,12 +1163,14 @@ void ff_estimate_p_frame_motion(MpegEncContext * s,
 {
     MotionEstContext * const c= &s->me;
     uint8_t *pix, *ppix;
-    int sum, varc, vard, mx, my, dmin;
+    int sum, mx, my, dmin;
+    int varc;            ///< the variance of the block (sum of squared (p[y][x]-average))
+    int vard;            ///< sum of squared differences with the estimated motion vector
     int P[10][2];
     const int shift= 1+s->quarter_sample;
     int mb_type=0;
     Picture * const pic= &s->current_picture;
-    
+
     init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
 
     assert(s->quarter_sample==0 || s->quarter_sample==1);
@@ -1155,50 +1188,48 @@ void ff_estimate_p_frame_motion(MpegEncContext * s,
     /* intra / predictive decision */
     pix = c->src[0][0];
     sum = s->dsp.pix_sum(pix, s->linesize);
-    varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
+    varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
 
     pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
-    pic->mb_var [s->mb_stride * mb_y + mb_x] = varc;
-    c->mb_var_sum_temp += varc;
+    pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
+    c->mb_var_sum_temp += (varc+128)>>8;
 
     if(c->avctx->me_threshold){
-        vard= (check_input_motion(s, mb_x, mb_y, 1)+128)>>8;
-        
-        if(vard<c->avctx->me_threshold){
-            pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
-            c->mc_mb_var_sum_temp += vard;
-            if (vard <= 64 || vard < varc) { //FIXME
-                c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
-            }else{
-                c->scene_change_score+= s->qscale;
-            }
+        vard= check_input_motion(s, mb_x, mb_y, 1);
+
+        if((vard+128)>>8 < c->avctx->me_threshold){
+            int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
+            int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
+            pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
+            c->mc_mb_var_sum_temp += (vard+128)>>8;
+            c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
             return;
         }
-        if(vard<c->avctx->mb_threshold)
+        if((vard+128)>>8 < c->avctx->mb_threshold)
             mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];
     }
 
     switch(s->me_method) {
     case ME_ZERO:
     default:
-	no_motion_search(s, &mx, &my);
+        no_motion_search(s, &mx, &my);
         mx-= mb_x*16;
         my-= mb_y*16;
         dmin = 0;
         break;
 #if 0
     case ME_FULL:
-	dmin = full_motion_search(s, &mx, &my, range, ref_picture);
+        dmin = full_motion_search(s, &mx, &my, range, ref_picture);
         mx-= mb_x*16;
         my-= mb_y*16;
         break;
     case ME_LOG:
-	dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
+        dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
         mx-= mb_x*16;
         my-= mb_y*16;
         break;
     case ME_PHODS:
-	dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
+        dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
         mx-= mb_x*16;
         my-= mb_y*16;
         break;
@@ -1222,7 +1253,7 @@ void ff_estimate_p_frame_motion(MpegEncContext * s,
                 if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
                 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
                 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
-        
+
                 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
                 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
 
@@ -1239,29 +1270,28 @@ void ff_estimate_p_frame_motion(MpegEncContext * s,
             }
 
         }
-        dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);       
+        dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
 
         break;
     }
 
     /* At this point (mx,my) are full-pell and the relative displacement */
     ppix = c->ref[0][0] + (my * s->linesize) + mx;
-        
-    vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16)+128)>>8;
 
-    pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
-//    pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin; 
-    c->mc_mb_var_sum_temp += vard;
-    
+    vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);
+
+    pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
+//    pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
+    c->mc_mb_var_sum_temp += (vard+128)>>8;
+
 #if 0
     printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
-	   varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
+           varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
 #endif
     if(mb_type){
-        if (vard <= 64 || vard < varc)
-            c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
-        else
-            c->scene_change_score+= s->qscale;
+        int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
+        int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
+        c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
 
         if(mb_type == CANDIDATE_MB_TYPE_INTER){
             c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
@@ -1279,25 +1309,25 @@ void ff_estimate_p_frame_motion(MpegEncContext * s,
             interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
         }
     }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
-        if (vard <= 64 || vard < varc)
-            c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
-        else
-            c->scene_change_score+= s->qscale;
+        int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
+        int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
+        c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
 
-        if (vard*2 + 200 > varc)
+        if (vard*2 + 200*256 > varc)
             mb_type|= CANDIDATE_MB_TYPE_INTRA;
-        if (varc*2 + 200 > vard){
+        if (varc*2 + 200*256 > vard || s->qscale > 24){
+//        if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
             mb_type|= CANDIDATE_MB_TYPE_INTER;
             c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
             if(s->flags&CODEC_FLAG_MV0)
                 if(mx || my)
-                    mb_type |= CANDIDATE_MB_TYPE_SKIPED; //FIXME check difference
+                    mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
         }else{
             mx <<=shift;
             my <<=shift;
         }
         if((s->flags&CODEC_FLAG_4MV)
-           && !c->skip && varc>50 && vard>10){
+           && !c->skip && varc>50<<8 && vard>10<<8){
             if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
                 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
 
@@ -1318,7 +1348,7 @@ void ff_estimate_p_frame_motion(MpegEncContext * s,
             dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
 
         if((s->flags&CODEC_FLAG_4MV)
-           && !c->skip && varc>50 && vard>10){
+           && !c->skip && varc>50<<8 && vard>10<<8){
             int dmin4= h263_mv4_search(s, mx, my, shift);
             if(dmin4 < dmin){
                 mb_type= CANDIDATE_MB_TYPE_INTER4V;
@@ -1333,17 +1363,17 @@ void ff_estimate_p_frame_motion(MpegEncContext * s,
                 dmin= dmin_i;
             }
         }
-                
-//        pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin; 
+
+//        pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
         set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
 
         /* get intra luma score */
         if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
-            intra_score= (varc<<8) - 500; //FIXME dont scale it down so we dont have to fix it
+            intra_score= varc - 500;
         }else{
             int mean= (sum+128)>>8;
             mean*= 0x01010101;
-            
+
             for(i=0; i<16; i++){
                 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
                 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
@@ -1359,36 +1389,36 @@ void ff_estimate_p_frame_motion(MpegEncContext * s,
             for(i=1; i<3; i++){
                 uint8_t *dest_c;
                 int mean;
-                
+
                 if(s->out_format == FMT_H263){
                     mean= (s->dc_val[i][mb_x + mb_y*s->b8_stride] + 4)>>3; //FIXME not exact but simple ;)
                 }else{
                     mean= (s->last_dc[i] + 4)>>3;
                 }
                 dest_c = s->new_picture.data[i] + (mb_y * 8  * (s->uvlinesize)) + mb_x * 8;
-                
+
                 mean*= 0x01010101;
                 for(i=0; i<8; i++){
                     *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 0]) = mean;
                     *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 4]) = mean;
                 }
-                
+
                 intra_score+= s->dsp.mb_cmp[1](s, c->scratchpad, dest_c, s->uvlinesize);
-            }                
+            }
         }
 #endif
         intra_score += c->mb_penalty_factor*16;
-        
+
         if(intra_score < dmin){
             mb_type= CANDIDATE_MB_TYPE_INTRA;
             s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
         }else
             s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;
-        
-        if (vard <= 64 || vard < varc) { //FIXME
-            c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
-        }else{
-            c->scene_change_score+= s->qscale;
+
+        {
+            int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
+            int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
+            c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
         }
     }
 
@@ -1404,7 +1434,7 @@ int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
     const int shift= 1+s->quarter_sample;
     const int xy= mb_x + mb_y*s->mb_stride;
     init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
-    
+
     assert(s->quarter_sample==0 || s->quarter_sample==1);
 
     c->pre_penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
@@ -1423,7 +1453,7 @@ int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
         c->pred_x= P_LEFT[0];
         c->pred_y= P_LEFT[1];
         P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
-        P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME 
+        P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
     } else {
         P_TOP[0]      = s->p_mv_table[xy + s->mb_stride    ][0];
         P_TOP[1]      = s->p_mv_table[xy + s->mb_stride    ][1];
@@ -1432,7 +1462,7 @@ int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
         if(P_TOP[1]      < (c->ymin<<shift)) P_TOP[1]     = (c->ymin<<shift);
         if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
         if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
-    
+
         P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
         P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
 
@@ -1440,11 +1470,11 @@ int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
         c->pred_y = P_MEDIAN[1];
     }
 
-    dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);       
+    dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
 
     s->p_mv_table[xy][0] = mx<<shift;
     s->p_mv_table[xy][1] = my<<shift;
-    
+
     return dmin;
 }
 
@@ -1459,7 +1489,7 @@ static int ff_estimate_motion_b(MpegEncContext * s,
     const int mot_xy = mb_y*mot_stride + mb_x;
     uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
     int mv_scale;
-        
+
     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
@@ -1470,24 +1500,24 @@ static int ff_estimate_motion_b(MpegEncContext * s,
     switch(s->me_method) {
     case ME_ZERO:
     default:
-	no_motion_search(s, &mx, &my);
+        no_motion_search(s, &mx, &my);
         dmin = 0;
         mx-= mb_x*16;
         my-= mb_y*16;
         break;
 #if 0
     case ME_FULL:
-	dmin = full_motion_search(s, &mx, &my, range, ref_picture);
+        dmin = full_motion_search(s, &mx, &my, range, ref_picture);
         mx-= mb_x*16;
         my-= mb_y*16;
         break;
     case ME_LOG:
-	dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
+        dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
         mx-= mb_x*16;
         my-= mb_y*16;
         break;
     case ME_PHODS:
-	dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
+        dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
         mx-= mb_x*16;
         my-= mb_y*16;
         break;
@@ -1509,27 +1539,27 @@ static int ff_estimate_motion_b(MpegEncContext * s,
                 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1]= (c->ymax<<shift);
                 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
                 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
-        
+
                 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
                 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
             }
             c->pred_x= P_LEFT[0];
             c->pred_y= P_LEFT[1];
         }
-        
+
         if(mv_table == s->b_forw_mv_table){
             mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
         }else{
             mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
         }
-        
+
         dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
- 
+
         break;
     }
-    
+
     dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
-                                   
+
     if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
         dmin= ff_get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
 
@@ -1552,9 +1582,9 @@ static inline int check_bidir_mv(MpegEncContext * s,
     //FIXME better f_code prediction (max mv & distance)
     //FIXME pointers
     MotionEstContext * const c= &s->me;
-    uint8_t * const mv_penalty= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
+    uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
+    uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
     int stride= c->stride;
-    int uvstride= c->uvstride;
     uint8_t *dest_y = c->scratchpad;
     uint8_t *ptr;
     int dxy;
@@ -1575,7 +1605,7 @@ static inline int check_bidir_mv(MpegEncContext * s,
         dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
         src_x = motion_bx >> 2;
         src_y = motion_by >> 2;
-    
+
         ptr = ref2_data[0] + (src_y * stride) + src_x;
         s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y    , ptr    , stride);
     }else{
@@ -1589,25 +1619,26 @@ static inline int check_bidir_mv(MpegEncContext * s,
         dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
         src_x = motion_bx >> 1;
         src_y = motion_by >> 1;
-    
+
         ptr = ref2_data[0] + (src_y * stride) + src_x;
         s->dsp.avg_pixels_tab[size][dxy](dest_y    , ptr    , stride, h);
     }
 
-    fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*c->mb_penalty_factor
-           +(mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*c->mb_penalty_factor
+    fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
+           +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
            + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
-           
+
     if(c->avctx->mb_cmp&FF_CMP_CHROMA){
     }
     //FIXME CHROMA !!!
-           
+
     return fbmin;
 }
 
 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
 {
+    MotionEstContext * const c= &s->me;
     const int mot_stride = s->mb_stride;
     const int xy = mb_y *mot_stride + mb_x;
     int fbmin;
@@ -1619,16 +1650,87 @@ static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
     int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
     int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
     int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
+    const int flags= c->sub_flags;
+    const int qpel= flags&FLAG_QPEL;
+    const int shift= 1+qpel;
+    const int xmin= c->xmin<<shift;
+    const int ymin= c->ymin<<shift;
+    const int xmax= c->xmax<<shift;
+    const int ymax= c->ymax<<shift;
+    uint8_t map[8][8][8][8];
+
+    memset(map,0,sizeof(map));
+#define BIDIR_MAP(fx,fy,bx,by) \
+    map[(motion_fx+fx)&7][(motion_fy+fy)&7][(motion_bx+bx)&7][(motion_by+by)&7]
+    BIDIR_MAP(0,0,0,0) = 1;
 
-    //FIXME do refinement and add flag
-    
     fbmin= check_bidir_mv(s, motion_fx, motion_fy,
                           motion_bx, motion_by,
                           pred_fx, pred_fy,
                           pred_bx, pred_by,
                           0, 16);
 
-   return fbmin;
+    if(s->avctx->bidir_refine){
+        int score, end;
+#define CHECK_BIDIR(fx,fy,bx,by)\
+    if( !BIDIR_MAP(fx,fy,bx,by)\
+       &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
+       &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
+        BIDIR_MAP(fx,fy,bx,by) = 1;\
+        score= check_bidir_mv(s, motion_fx+fx, motion_fy+fy, motion_bx+bx, motion_by+by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);\
+        if(score < fbmin){\
+            fbmin= score;\
+            motion_fx+=fx;\
+            motion_fy+=fy;\
+            motion_bx+=bx;\
+            motion_by+=by;\
+            end=0;\
+        }\
+    }
+#define CHECK_BIDIR2(a,b,c,d)\
+CHECK_BIDIR(a,b,c,d)\
+CHECK_BIDIR(-(a),-(b),-(c),-(d))
+
+#define CHECK_BIDIRR(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)
+
+        do{
+            end=1;
+
+            CHECK_BIDIRR( 0, 0, 0, 1)
+            if(s->avctx->bidir_refine > 1){
+                CHECK_BIDIRR( 0, 0, 1, 1)
+                CHECK_BIDIR2( 0, 1, 0, 1)
+                CHECK_BIDIR2( 1, 0, 1, 0)
+                CHECK_BIDIRR( 0, 0,-1, 1)
+                CHECK_BIDIR2( 0,-1, 0, 1)
+                CHECK_BIDIR2(-1, 0, 1, 0)
+                if(s->avctx->bidir_refine > 2){
+                    CHECK_BIDIRR( 0, 1, 1, 1)
+                    CHECK_BIDIRR( 0,-1, 1, 1)
+                    CHECK_BIDIRR( 0, 1,-1, 1)
+                    CHECK_BIDIRR( 0, 1, 1,-1)
+                    if(s->avctx->bidir_refine > 3){
+                        CHECK_BIDIR2( 1, 1, 1, 1)
+                        CHECK_BIDIRR( 1, 1, 1,-1)
+                        CHECK_BIDIR2( 1, 1,-1,-1)
+                        CHECK_BIDIR2( 1,-1,-1, 1)
+                        CHECK_BIDIR2( 1,-1, 1,-1)
+                    }
+                }
+            }
+        }while(!end);
+    }
+
+    s->b_bidir_forw_mv_table[xy][0]= motion_fx;
+    s->b_bidir_forw_mv_table[xy][1]= motion_fy;
+    s->b_bidir_back_mv_table[xy][0]= motion_bx;
+    s->b_bidir_back_mv_table[xy][1]= motion_by;
+
+    return fbmin;
 }
 
 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
@@ -1643,7 +1745,7 @@ static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
     const int time_pb= s->pb_time;
     int mx, my, xmin, xmax, ymin, ymax;
     int16_t (*mv_table)[2]= s->b_direct_mv_table;
-    
+
     c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
     ymin= xmin=(-32)>>shift;
     ymax= xmax=   31>>shift;
@@ -1657,7 +1759,7 @@ static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
     for(i=0; i<4; i++){
         int index= s->block_index[i];
         int min, max;
-    
+
         c->co_located_mv[i][0]= s->next_picture.motion_val[0][index][0];
         c->co_located_mv[i][1]= s->next_picture.motion_val[0][index][1];
         c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
@@ -1678,19 +1780,19 @@ static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
         min+= 16*mb_y - 1;
         ymax= FFMIN(ymax, s->height - max);
         ymin= FFMAX(ymin, - 16      - min);
-        
+
         if(s->mv_type == MV_TYPE_16X16) break;
     }
-    
+
     assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
-    
+
     if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
         s->b_direct_mv_table[mot_xy][0]= 0;
         s->b_direct_mv_table[mot_xy][1]= 0;
 
         return 256*256*256*64;
     }
-    
+
     c->xmin= xmin;
     c->ymin= ymin;
     c->xmax= xmax;
@@ -1700,33 +1802,33 @@ static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
     c->pred_x=0;
     c->pred_y=0;
 
-    P_LEFT[0]        = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
-    P_LEFT[1]        = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
+    P_LEFT[0]        = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
+    P_LEFT[1]        = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
 
     /* special case for first line */
-    if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as its cliped
-        P_TOP[0]      = clip(mv_table[mot_xy - mot_stride             ][0], xmin<<shift, xmax<<shift);
-        P_TOP[1]      = clip(mv_table[mot_xy - mot_stride             ][1], ymin<<shift, ymax<<shift);
-        P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1         ][0], xmin<<shift, xmax<<shift);
-        P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1         ][1], ymin<<shift, ymax<<shift);
-    
+    if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
+        P_TOP[0]      = av_clip(mv_table[mot_xy - mot_stride             ][0], xmin<<shift, xmax<<shift);
+        P_TOP[1]      = av_clip(mv_table[mot_xy - mot_stride             ][1], ymin<<shift, ymax<<shift);
+        P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1         ][0], xmin<<shift, xmax<<shift);
+        P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1         ][1], ymin<<shift, ymax<<shift);
+
         P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
         P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
     }
- 
+
     dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
-    if(c->sub_flags&FLAG_QPEL) 
+    if(c->sub_flags&FLAG_QPEL)
         dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
     else
         dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
-    
+
     if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
         dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
-    
+
     get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
 
-    s->b_direct_mv_table[mot_xy][0]= mx;
-    s->b_direct_mv_table[mot_xy][1]= my;
+    mv_table[mot_xy][0]= mx;
+    mv_table[mot_xy][1]= my;
     c->flags     &= ~FLAG_DIRECT;
     c->sub_flags &= ~FLAG_DIRECT;
 
@@ -1744,29 +1846,41 @@ void ff_estimate_b_frame_motion(MpegEncContext * s,
     init_ref(c, s->new_picture.data, s->last_picture.data, s->next_picture.data, 16*mb_x, 16*mb_y, 2);
 
     get_limits(s, 16*mb_x, 16*mb_y);
-    
+
     c->skip=0;
+
+    if(s->codec_id == CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]){
+        int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
+
+        score= ((unsigned)(score*score + 128*256))>>16;
+        c->mc_mb_var_sum_temp += score;
+        s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
+        s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
+
+        return;
+    }
+
     if(c->avctx->me_threshold){
-        int vard= (check_input_motion(s, mb_x, mb_y, 0)+128)>>8;
-        
-        if(vard<c->avctx->me_threshold){
+        int vard= check_input_motion(s, mb_x, mb_y, 0);
+
+        if((vard+128)>>8 < c->avctx->me_threshold){
 //            pix = c->src[0][0];
 //            sum = s->dsp.pix_sum(pix, s->linesize);
-//            varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
-        
-//            pic->mb_var   [s->mb_stride * mb_y + mb_x] = varc;
-             s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
+//            varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
+
+//            pic->mb_var   [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
+             s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
 /*            pic->mb_mean  [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
-            c->mb_var_sum_temp    += varc;*/
-            c->mc_mb_var_sum_temp += vard;
-/*            if (vard <= 64 || vard < varc) {
+            c->mb_var_sum_temp    += (varc+128)>>8;*/
+            c->mc_mb_var_sum_temp += (vard+128)>>8;
+/*            if (vard <= 64<<8 || vard < varc) {
                 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
             }else{
-                c->scene_change_score+= s->qscale;
+                c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
             }*/
             return;
         }
-        if(vard<c->avctx->mb_threshold){
+        if((vard+128)>>8 < c->avctx->mb_threshold){
             type= s->mb_type[mb_y*s->mb_stride + mb_x];
             if(type == CANDIDATE_MB_TYPE_DIRECT){
                 direct_search(s, mb_x, mb_y);
@@ -1804,7 +1918,7 @@ void ff_estimate_b_frame_motion(MpegEncContext * s,
 //FIXME penalty stuff for non mpeg4
     c->skip=0;
     fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
-    
+
     c->skip=0;
     bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
 //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
@@ -1812,7 +1926,7 @@ void ff_estimate_b_frame_motion(MpegEncContext * s,
     c->skip=0;
     fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
 //printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
-    
+
     if(s->flags & CODEC_FLAG_INTERLACED_ME){
 //FIXME mb type penalty
         c->skip=0;
@@ -1830,14 +1944,14 @@ void ff_estimate_b_frame_motion(MpegEncContext * s,
     {
         int score= fmin;
         type = CANDIDATE_MB_TYPE_FORWARD;
-        
+
         if (dmin <= score){
             score = dmin;
             type = CANDIDATE_MB_TYPE_DIRECT;
         }
         if(bmin<score){
             score=bmin;
-            type= CANDIDATE_MB_TYPE_BACKWARD; 
+            type= CANDIDATE_MB_TYPE_BACKWARD;
         }
         if(fbmin<score){
             score=fbmin;
@@ -1851,7 +1965,7 @@ void ff_estimate_b_frame_motion(MpegEncContext * s,
             score=bimin;
             type= CANDIDATE_MB_TYPE_BACKWARD_I;
         }
-        
+
         score= ((unsigned)(score*score + 128*256))>>16;
         c->mc_mb_var_sum_temp += score;
         s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
@@ -1867,8 +1981,10 @@ void ff_estimate_b_frame_motion(MpegEncContext * s,
             type |= CANDIDATE_MB_TYPE_BIDIR_I;
         }
          //FIXME something smarter
-        if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB
-#if 0        
+        if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
+        if(s->codec_id == CODEC_ID_MPEG4 && type&CANDIDATE_MB_TYPE_DIRECT && s->flags&CODEC_FLAG_MV0 && *(uint32_t*)s->b_direct_mv_table[xy])
+            type |= CANDIDATE_MB_TYPE_DIRECT0;
+#if 0
         if(s->out_format == FMT_MPEG1)
             type |= CANDIDATE_MB_TYPE_INTRA;
 #endif
@@ -1882,11 +1998,16 @@ int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
 {
     if(s->me_method>=ME_EPZS){
         int score[8];
-        int i, y;
+        int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
         uint8_t * fcode_tab= s->fcode_tab;
         int best_fcode=-1;
         int best_score=-10000000;
 
+        if(s->msmpeg4_version)
+            range= FFMIN(range, 16);
+        else if(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
+            range= FFMIN(range, 256);
+
         for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
 
         for(y=0; y<s->mb_height; y++){
@@ -1894,10 +2015,16 @@ int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
             int xy= y*s->mb_stride;
             for(x=0; x<s->mb_width; x++){
                 if(s->mb_type[xy] & type){
-                    int fcode= FFMAX(fcode_tab[mv_table[xy][0] + MAX_MV],
-                                     fcode_tab[mv_table[xy][1] + MAX_MV]);
+                    int mx= mv_table[xy][0];
+                    int my= mv_table[xy][1];
+                    int fcode= FFMAX(fcode_tab[mx + MAX_MV],
+                                     fcode_tab[my + MAX_MV]);
                     int j;
-                    
+
+                        if(mx >= range || mx < -range ||
+                           my >= range || my < -range)
+                            continue;
+
                     for(j=0; j<fcode && j<8; j++){
                         if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
                             score[j]-= 170;
@@ -1906,7 +2033,7 @@ int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
                 xy++;
             }
         }
-        
+
         for(i=1; i<8; i++){
             if(score[i] > best_score){
                 best_score= score[i];
@@ -1933,12 +2060,13 @@ void ff_fix_long_p_mvs(MpegEncContext * s)
     int y, range;
     assert(s->pict_type==P_TYPE);
 
-    range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);
-    
-    if(s->msmpeg4_version) range= 16;
-    
+    range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
+
+    assert(range <= 16 || !s->msmpeg4_version);
+    assert(range <=256 || !(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
+
     if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
-    
+
 //printf("%d no:%d %d//\n", clip, noclip, f_code);
     if(s->flags&CODEC_FLAG_4MV){
         const int wrap= s->b8_stride;
@@ -1976,16 +2104,15 @@ void ff_fix_long_p_mvs(MpegEncContext * s)
  *
  * @param truncate 1 for truncation, 0 for using intra
  */
-void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select, 
+void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
                      int16_t (*mv_table)[2], int f_code, int type, int truncate)
 {
     MotionEstContext * const c= &s->me;
     int y, h_range, v_range;
 
     // RAL: 8 in MPEG-1, 16 in MPEG-4
-    int range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);
+    int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
 
-    if(s->msmpeg4_version) range= 16;
     if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
 
     h_range= range;