diff --git a/source/blender/render/SConscript b/source/blender/render/SConscript
index fb210afa545..2228ff9ebb9 100644
--- a/source/blender/render/SConscript
+++ b/source/blender/render/SConscript
@@ -1,7 +1,8 @@
 #!/usr/bin/python
 Import ('env')
 
-cflags=''
+cflags = ['-O3']
+cxxflags = ['-O3','-msse2','-mfpmath=sse']
 sources = env.Glob('intern/source/*.c')
 raysources = env.Glob('intern/raytrace/*.cpp')
 
@@ -22,7 +23,7 @@ if env['WITH_BF_OPENEXR']:
     defs.append('WITH_OPENEXR')
 
 if env['OURPLATFORM']=='linux2':
-    cflags='-pthread'
+    cflags += ['-pthread']
 
 env.BlenderLib ( libname = 'bf_render', sources = sources, includes = Split(incs), defines=defs, libtype='core', priority=145, compileflags=cflags )
-env.BlenderLib ( libname = 'bf_render_raytrace', sources = raysources, includes = Split(incs), defines=defs, libtype='core', priority=145, compileflags=cflags )
+env.BlenderLib ( libname = 'bf_render_raytrace', sources = raysources, includes = Split(incs), defines=defs, libtype='core', priority=145, compileflags=cflags, cxx_compileflags=cxxflags )
diff --git a/source/blender/render/intern/raytrace/bvh.h b/source/blender/render/intern/raytrace/bvh.h
index 2d75cd800a7..f6c12f4b3fb 100644
--- a/source/blender/render/intern/raytrace/bvh.h
+++ b/source/blender/render/intern/raytrace/bvh.h
@@ -201,8 +201,8 @@ static int bvh_node_stack_raycast_simd(Node *root, Isect *isec)
 			t_bb[0] = _mm_shuffle_ps( x0y0x1y1, x2y2x3y3, _MM_SHUFFLE(2,0,2,0) );
 			t_bb[1] = _mm_shuffle_ps( x0y0x1y1, x2y2x3y3, _MM_SHUFFLE(3,1,3,1) );
 
-			const __m128 z0X0z1X1 = _mm_shuffle_ps( _mm_load_ps(bb0), _mm_load_ps(bb1), _MM_SHUFFLE(3,1,3,1) );
-			const __m128 z2X2z3X3 = _mm_shuffle_ps( _mm_load_ps(bb2), _mm_load_ps(bb3), _MM_SHUFFLE(3,1,3,1) );
+			const __m128 z0X0z1X1 = _mm_shuffle_ps( _mm_load_ps(bb0), _mm_load_ps(bb1), _MM_SHUFFLE(3,2,3,2) );
+			const __m128 z2X2z3X3 = _mm_shuffle_ps( _mm_load_ps(bb2), _mm_load_ps(bb3), _MM_SHUFFLE(3,2,3,2) );
 			t_bb[2] = _mm_shuffle_ps( z0X0z1X1, z2X2z3X3, _MM_SHUFFLE(2,0,2,0) );
 			t_bb[3] = _mm_shuffle_ps( z0X0z1X1, z2X2z3X3, _MM_SHUFFLE(3,1,3,1) );