diff --git a/encoder/analyse.c b/encoder/analyse.c
index 8457c38..b2bc9b2 100644
--- a/encoder/analyse.c
+++ b/encoder/analyse.c
@@ -167,27 +167,27 @@ static const int i_sub_mb_p_cost_table[4] = {
 
 static void x264_analyse_update_cache( x264_t *h, x264_mb_analysis_t *a );
 
+static int16_t *x264_cost_mv[52];
 uint16_t *x264_cost_mv_fpel[52][4];
 
 /* initialize an array of lambda*nbits for all possible mvs */
 static void x264_mb_analyse_load_costs( x264_t *h, x264_mb_analysis_t *a )
 {
-    static int16_t *p_cost_mv[52];
     int i, j;
 
-    if( !p_cost_mv[a->i_qp] )
+    if( !x264_cost_mv[a->i_qp] )
     {
         /* could be faster, but isn't called many times */
         /* factor of 4 from qpel, 2 from sign, and 2 because mv can be opposite from mvp */
-        p_cost_mv[a->i_qp] = x264_malloc( (4*4*2048 + 1) * sizeof(int16_t) );
-        p_cost_mv[a->i_qp] += 2*4*2048;
+        x264_cost_mv[a->i_qp] = x264_malloc( (4*4*2048 + 1) * sizeof(int16_t) );
+        x264_cost_mv[a->i_qp] += 2*4*2048;
         for( i = 0; i <= 2*4*2048; i++ )
         {
-            p_cost_mv[a->i_qp][-i] =
-            p_cost_mv[a->i_qp][i]  = a->i_lambda * bs_size_se( i );
+            x264_cost_mv[a->i_qp][-i] =
+            x264_cost_mv[a->i_qp][i]  = a->i_lambda * bs_size_se( i );
         }
     }
-    a->p_cost_mv = p_cost_mv[a->i_qp];
+    a->p_cost_mv = x264_cost_mv[a->i_qp];
 
     /* FIXME is this useful for all me methods? */
     if( h->param.analyse.i_me_method >= X264_ME_ESA && !x264_cost_mv_fpel[a->i_qp][0] )
@@ -197,11 +197,32 @@ static void x264_mb_analyse_load_costs( x264_t *h, x264_mb_analysis_t *a )
             x264_cost_mv_fpel[a->i_qp][j] = x264_malloc( (4*2048 + 1) * sizeof(int16_t) );
             x264_cost_mv_fpel[a->i_qp][j] += 2*2048;
             for( i = -2*2048; i < 2*2048; i++ )
-                x264_cost_mv_fpel[a->i_qp][j][i] = p_cost_mv[a->i_qp][i*4+j];
+                x264_cost_mv_fpel[a->i_qp][j][i] = x264_cost_mv[a->i_qp][i*4+j];
         }
     }
 }
 
+void x264_mb_analyse_free_costs( void )
+{
+    int i, j;
+
+    for( i=0; i<52; i++ )
+    {
+        if( x264_cost_mv[i] )
+        {
+            x264_free( x264_cost_mv[i] - 2*4*2048 );
+            x264_cost_mv[i] = NULL;
+        }
+
+        for( j=0; j<4; j++ )
+            if( x264_cost_mv_fpel[i][j] )
+            {
+                x264_free( x264_cost_mv_fpel[i][j] - 2*2048 );
+                x264_cost_mv_fpel[i][j] = NULL;
+            }
+    }
+}
+
 static void x264_mb_analyse_init( x264_t *h, x264_mb_analysis_t *a, int i_qp )
 {
     int i = h->param.analyse.i_subpel_refine - (h->sh.i_type == SLICE_TYPE_B);
diff --git a/encoder/analyse.h b/encoder/analyse.h
index b8c828f..4cf503c 100644
--- a/encoder/analyse.h
+++ b/encoder/analyse.h
@@ -26,5 +26,6 @@
 
 void x264_macroblock_analyse( x264_t *h );
 void x264_slicetype_decide( x264_t *h );
+void x264_mb_analyse_free_costs( void );
 
 #endif
diff --git a/encoder/encoder.c b/encoder/encoder.c
index e462a1f..aef5f97 100644
--- a/encoder/encoder.c
+++ b/encoder/encoder.c
@@ -2129,4 +2129,6 @@ void    x264_encoder_close  ( x264_t *h )
         x264_free( h->thread[i]->out.p_bitstream );
         x264_free( h->thread[i] );
     }
+
+    x264_mb_analyse_free_costs();
 }