diff --git a/encoder/me.c b/encoder/me.c
index 25e1195..0f1a9e1 100644
--- a/encoder/me.c
+++ b/encoder/me.c
@@ -24,6 +24,7 @@
 
 #include "common/common.h"
 #include "me.h"
+#include <bench.h>
 
 /* presets selected from good points on the speed-vs-quality curve of several test videos
  * subpel_iters[i_subpel_refine] = { refine_hpel, refine_qpel, me_hpel, me_qpel }
@@ -86,6 +87,7 @@ static void refine_subpel( x264_t *h, x264_me_t *m, int hpel_iters, int qpel_ite
 
 #define COST_MV_X4( m0x, m0y, m1x, m1y, m2x, m2y, m3x, m3y )\
 {\
+    START_TIMER;\
     uint8_t *pix_base = p_fref + omx + omy*m->i_stride[0];\
     h->pixf.fpelcmp_x4[i_pixel]( m->p_fenc[0],\
         pix_base + (m0x) + (m0y)*m->i_stride[0],\
@@ -93,14 +95,37 @@ static void refine_subpel( x264_t *h, x264_me_t *m, int hpel_iters, int qpel_ite
         pix_base + (m2x) + (m2y)*m->i_stride[0],\
         pix_base + (m3x) + (m3y)*m->i_stride[0],\
         m->i_stride[0], costs );\
-    costs[0] += BITS_MVD( omx+(m0x), omy+(m0y) );\
-    costs[1] += BITS_MVD( omx+(m1x), omy+(m1y) );\
-    costs[2] += BITS_MVD( omx+(m2x), omy+(m2y) );\
-    costs[3] += BITS_MVD( omx+(m3x), omy+(m3y) );\
-    COPY3_IF_LT( bcost, costs[0], bmx, omx+(m0x), bmy, omy+(m0y) );\
-    COPY3_IF_LT( bcost, costs[1], bmx, omx+(m1x), bmy, omy+(m1y) );\
-    COPY3_IF_LT( bcost, costs[2], bmx, omx+(m2x), bmy, omy+(m2y) );\
-    COPY3_IF_LT( bcost, costs[3], bmx, omx+(m3x), bmy, omy+(m3y) );\
+    if( __builtin_constant_p(m0x) )\
+    {\
+        const int8_t deltas[8][2] = {{0,0}, {m0x,m0y}, {m1x,m1y}, {0,0}, {m2x,m2y}, {0,0}, {m3x,m3y}, {0,0}};\
+        costs[0] = bcost + ((costs[0] + BITS_MVD( omx+(m0x), omy+(m0y) )) << 16);\
+        costs[1] += BITS_MVD( omx+(m1x), omy+(m1y) ) + (-1<<16);\
+        costs[2] += BITS_MVD( omx+(m2x), omy+(m2y) ) + (-1<<16);\
+        costs[3] += BITS_MVD( omx+(m3x), omy+(m3y) ) + (-1<<16);\
+        asm volatile(\
+            "phminposuw %1, %%xmm0 \n"\
+            "movd %%xmm0, %0 \n"\
+            :"=r"(bcost)\
+            :"m"(costs[0]), "m"(costs[1]), "m"(costs[2]), "m"(costs[3])\
+        );\
+        int d = bcost >> 16;\
+        bcost &= 0xffff;\
+        bmx += deltas[d][0];\
+        bmy += deltas[d][1];\
+        STOP_TIMER("x4_sse4");\
+    }\
+    else\
+    {\
+        costs[0] += BITS_MVD( omx+(m0x), omy+(m0y) );\
+        costs[1] += BITS_MVD( omx+(m1x), omy+(m1y) );\
+        costs[2] += BITS_MVD( omx+(m2x), omy+(m2y) );\
+        costs[3] += BITS_MVD( omx+(m3x), omy+(m3y) );\
+        COPY3_IF_LT( bcost, costs[0], bmx, omx+(m0x), bmy, omy+(m0y) );\
+        COPY3_IF_LT( bcost, costs[1], bmx, omx+(m1x), bmy, omy+(m1y) );\
+        COPY3_IF_LT( bcost, costs[2], bmx, omx+(m2x), bmy, omy+(m2y) );\
+        COPY3_IF_LT( bcost, costs[3], bmx, omx+(m3x), bmy, omy+(m3y) );\
+        STOP_TIMER("x4");\
+    }\
 }
 
 #define COST_MV_X3_ABS( m0x, m0y, m1x, m1y, m2x, m2y )\
@@ -167,7 +192,7 @@ void x264_me_search_ref( x264_t *h, x264_me_t *m, int16_t (*mvc)[2], int i_mvc,
 
     int i = 0, j;
     int dir;
-    int costs[6];
+    DECLARE_ALIGNED_16( int costs[6] );
 
     int mv_x_min = h->mb.mv_min_fpel[0];
     int mv_y_min = h->mb.mv_min_fpel[1];