Merge remote branch 'origin/master' into experimental

Change-Id: I993021d0b2d7fbe44d6371464f2686eed3ccfaae

vp8/common/x86/recon_sse2.asm

@@ -251,15 +251,8 @@ sym(vp8_intra_pred_uv_dc_mmx2):
     movsxd      rax,        dword ptr arg(3) ;src_stride;
     sub         rsi,        rax
     pxor        mm0,        mm0
-    movd        mm1,        [rsi]
-    movd        mm2,        [rsi+4]
-    punpcklbw   mm1,        mm0
-    punpcklbw   mm2,        mm0
-    paddw       mm1,        mm2
-    pshufw      mm2,        mm1, 0x0e
-    paddw       mm1,        mm2
-    pshufw      mm2,        mm1, 0x01
-    paddw       mm1,        mm2
+    movq        mm1,        [rsi]
+    psadbw      mm1,        mm0
 
     ; from left
     dec         rsi
@@ -331,15 +324,8 @@ sym(vp8_intra_pred_uv_dctop_mmx2):
     movsxd      rax,        dword ptr arg(3) ;src_stride;
     sub         rsi,        rax
     pxor        mm0,        mm0
-    movd        mm1,        [rsi]
-    movd        mm2,        [rsi+4]
-    punpcklbw   mm1,        mm0
-    punpcklbw   mm2,        mm0
-    paddw       mm1,        mm2
-    pshufw      mm2,        mm1, 0x0e
-    paddw       mm1,        mm2
-    pshufw      mm2,        mm1, 0x01
-    paddw       mm1,        mm2
+    movq        mm1,        [rsi]
+    psadbw      mm1,        mm0
 
     ; add up
     paddw       mm1,        [GLOBAL(dc_4)]

vp8/common/x86/recon_wrapper_sse2.c

@@ -27,7 +27,7 @@ extern build_intra_predictors_mbuv_prototype(vp8_intra_pred_uv_ve_mmx);
 extern build_intra_predictors_mbuv_prototype(vp8_intra_pred_uv_tm_sse2);
 extern build_intra_predictors_mbuv_prototype(vp8_intra_pred_uv_tm_ssse3);
 
-static inline void vp8_build_intra_predictors_mbuv_x86(MACROBLOCKD *x,
+static void vp8_build_intra_predictors_mbuv_x86(MACROBLOCKD *x,
                                                 unsigned char *dst_u,
                                                 unsigned char *dst_v,
                                                 int dst_stride,

vpx/src/vpx_encoder.c

@@ -16,6 +16,7 @@
 #include <limits.h>
 #include <string.h>
 #include "vpx/internal/vpx_codec_internal.h"
+#include "vpx_config.h"
 
 #define SAVE_STATUS(ctx,var) (ctx?(ctx->err = var):var)