diff --git a/test/test_intra_pred_speed.cc b/test/test_intra_pred_speed.cc
index c270072..5f4c48f 100644
--- a/test/test_intra_pred_speed.cc
+++ b/test/test_intra_pred_speed.cc
@@ -187,10 +187,15 @@
                 vpx_d153_predictor_4x4_c, vpx_d207_predictor_4x4_c,
                 vpx_d63_predictor_4x4_c, vpx_tm_predictor_4x4_c)
 
+#if HAVE_SSE && CONFIG_USE_X86INC
+INTRA_PRED_TEST(SSE, TestIntraPred4, vpx_dc_predictor_4x4_sse,
+                vpx_dc_left_predictor_4x4_sse, vpx_dc_top_predictor_4x4_sse,
+                vpx_dc_128_predictor_4x4_sse, vpx_v_predictor_4x4_sse, NULL,
+                NULL, NULL, NULL, NULL, NULL, NULL, NULL)
+#endif  // HAVE_SSE && CONFIG_USE_X86INC
+
 #if HAVE_SSE2 && CONFIG_USE_X86INC
-INTRA_PRED_TEST(SSE2, TestIntraPred4, vpx_dc_predictor_4x4_sse2,
-                vpx_dc_left_predictor_4x4_sse2, vpx_dc_top_predictor_4x4_sse2,
-                vpx_dc_128_predictor_4x4_sse2, vpx_v_predictor_4x4_sse2,
+INTRA_PRED_TEST(SSE2, TestIntraPred4, NULL, NULL, NULL, NULL, NULL,
                 vpx_h_predictor_4x4_sse2, NULL, NULL, NULL, NULL, NULL, NULL,
                 vpx_tm_predictor_4x4_sse2)
 #endif  // HAVE_SSE2 && CONFIG_USE_X86INC
diff --git a/vp10/encoder/ratectrl.c b/vp10/encoder/ratectrl.c
index d4c25c0..3ff2476 100644
--- a/vp10/encoder/ratectrl.c
+++ b/vp10/encoder/ratectrl.c
@@ -1210,8 +1210,12 @@
     rc->frames_since_golden = 0;
 
     // If we are not using alt ref in the up and coming group clear the arf
-    // active flag.
-    if (!rc->source_alt_ref_pending) {
+    // active flag. In multi arf group case, if the index is not 0 then
+    // we are overlaying a mid group arf so should not reset the flag.
+    if (cpi->oxcf.pass == 2) {
+      if (!rc->source_alt_ref_pending && (cpi->twopass.gf_group.index == 0))
+        rc->source_alt_ref_active = 0;
+    } else if (!rc->source_alt_ref_pending) {
       rc->source_alt_ref_active = 0;
     }
 
diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index f6152f5..296212e 100644
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -91,7 +91,7 @@
 specialize qw/vpx_d153_predictor_4x4/, "$ssse3_x86inc";
 
 add_proto qw/void vpx_v_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_v_predictor_4x4 neon msa/, "$sse2_x86inc";
+specialize qw/vpx_v_predictor_4x4 neon msa/, "$sse_x86inc";
 
 add_proto qw/void vpx_ve_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vpx_ve_predictor_4x4/;
@@ -100,16 +100,16 @@
 specialize qw/vpx_tm_predictor_4x4 neon dspr2 msa/, "$sse2_x86inc";
 
 add_proto qw/void vpx_dc_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_dc_predictor_4x4 dspr2 msa neon/, "$sse2_x86inc";
+specialize qw/vpx_dc_predictor_4x4 dspr2 msa neon/, "$sse_x86inc";
 
 add_proto qw/void vpx_dc_top_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_dc_top_predictor_4x4 msa neon/, "$sse2_x86inc";
+specialize qw/vpx_dc_top_predictor_4x4 msa neon/, "$sse_x86inc";
 
 add_proto qw/void vpx_dc_left_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_dc_left_predictor_4x4 msa neon/, "$sse2_x86inc";
+specialize qw/vpx_dc_left_predictor_4x4 msa neon/, "$sse_x86inc";
 
 add_proto qw/void vpx_dc_128_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_dc_128_predictor_4x4 msa neon/, "$sse2_x86inc";
+specialize qw/vpx_dc_128_predictor_4x4 msa neon/, "$sse_x86inc";
 
 add_proto qw/void vpx_d207_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vpx_d207_predictor_8x8/, "$ssse3_x86inc";
diff --git a/vpx_dsp/x86/intrapred_sse2.asm b/vpx_dsp/x86/intrapred_sse2.asm
index 7971e27..13fe140 100644
--- a/vpx_dsp/x86/intrapred_sse2.asm
+++ b/vpx_dsp/x86/intrapred_sse2.asm
@@ -23,7 +23,7 @@
 
 SECTION .text
 
-INIT_XMM sse2
+INIT_MMX sse
 cglobal dc_predictor_4x4, 4, 5, 2, dst, stride, above, left, goffset
   GET_GOT     goffsetq
 
@@ -33,7 +33,7 @@
   psadbw                m0, m1
   paddw                 m0, [GLOBAL(pw_4)]
   psraw                 m0, 3
-  pshuflw               m0, m0, 0x0
+  pshufw                m0, m0, 0x0
   packuswb              m0, m0
   movd      [dstq        ], m0
   movd      [dstq+strideq], m0
@@ -44,17 +44,16 @@
   RESTORE_GOT
   RET
 
-INIT_XMM sse2
-cglobal dc_left_predictor_4x4, 2, 5, 2, dst, stride, above, left, goffset
+INIT_MMX sse
+cglobal dc_left_predictor_4x4, 4, 5, 2, dst, stride, above, left, goffset
   GET_GOT     goffsetq
 
-  movifnidn          leftq, leftmp
   pxor                  m1, m1
   movd                  m0, [leftq]
   psadbw                m0, m1
   paddw                 m0, [GLOBAL(pw2_4)]
   psraw                 m0, 2
-  pshuflw               m0, m0, 0x0
+  pshufw                m0, m0, 0x0
   packuswb              m0, m0
   movd      [dstq        ], m0
   movd      [dstq+strideq], m0
@@ -65,8 +64,8 @@
   RESTORE_GOT
   RET
 
-INIT_XMM sse2
-cglobal dc_top_predictor_4x4, 3, 5, 2, dst, stride, above, left, goffset
+INIT_MMX sse
+cglobal dc_top_predictor_4x4, 4, 5, 2, dst, stride, above, left, goffset
   GET_GOT     goffsetq
 
   pxor                  m1, m1
@@ -74,7 +73,7 @@
   psadbw                m0, m1
   paddw                 m0, [GLOBAL(pw2_4)]
   psraw                 m0, 2
-  pshuflw               m0, m0, 0x0
+  pshufw                m0, m0, 0x0
   packuswb              m0, m0
   movd      [dstq        ], m0
   movd      [dstq+strideq], m0
@@ -167,8 +166,8 @@
   RESTORE_GOT
   RET
 
-INIT_XMM sse2
-cglobal dc_128_predictor_4x4, 2, 5, 1, dst, stride, above, left, goffset
+INIT_MMX sse
+cglobal dc_128_predictor_4x4, 4, 5, 3, dst, stride, above, left, goffset
   GET_GOT     goffsetq
 
   DEFINE_ARGS dst, stride, stride3
@@ -454,7 +453,7 @@
   RESTORE_GOT
   RET
 
-INIT_XMM sse2
+INIT_MMX sse
 cglobal v_predictor_4x4, 3, 3, 1, dst, stride, above
   movd                  m0, [aboveq]
   movd      [dstq        ], m0