diff --git a/README.chromium b/README.chromium
index 0412b0daf..143508ad0 100644
--- a/README.chromium
+++ b/README.chromium
@@ -1,6 +1,6 @@
 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 373
+Version: 374
 License: BSD
 License File: LICENSE
 
diff --git a/include/libyuv/version.h b/include/libyuv/version.h
index b8fe8d8dd..68317b35a 100644
--- a/include/libyuv/version.h
+++ b/include/libyuv/version.h
@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 373
+#define LIBYUV_VERSION 374
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT
diff --git a/source/convert.cc b/source/convert.cc
index 66127e9dd..ca2bf7fa9 100644
--- a/source/convert.cc
+++ b/source/convert.cc
@@ -962,8 +962,7 @@ int V210ToI420(const uint8* src_v210, int src_stride_v210,
       }
     }
   }
-#endif
-#if defined(HAS_UYVYTOYROW_NEON)
+#elif defined(HAS_UYVYTOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     if (width > 8) {
       UYVYToYRow = UYVYToYRow_Any_NEON;
diff --git a/source/convert_from.cc b/source/convert_from.cc
index f5b9dda3e..839587798 100644
--- a/source/convert_from.cc
+++ b/source/convert_from.cc
@@ -998,8 +998,7 @@ int I420ToRAW(const uint8* src_y, int src_stride_y,
       ARGBToRAWRow = ARGBToRAWRow_SSSE3;
     }
   }
-#endif
-#if defined(HAS_ARGBTORAWROW_NEON)
+#elif defined(HAS_ARGBTORAWROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     if (width * 3 <= kMaxStride) {
       ARGBToRAWRow = ARGBToRAWRow_Any_NEON;
diff --git a/source/cpu_id.cc b/source/cpu_id.cc
index 61c2f93de..9e7ac55c4 100644
--- a/source/cpu_id.cc
+++ b/source/cpu_id.cc
@@ -58,18 +58,36 @@ void CpuId(int cpu_info[4], int) {
 }
 #endif
 
-// Low level cpuid for X86.  Returns zeros on other CPUs.
-#if !defined(__CLR_VER) && defined(_M_IX86)
-// TODO(fbarchard): Port to GCC and 64 bit Visual C.
+// X86 CPUs have xgetbv to detect OS saves high parts of ymm registers.
+#if !defined(__CLR_VER) && defined(_MSC_VER) && (_MSC_FULL_VER >= 160040219)
 #define HAS_XGETBV
-// Return low 32 bits of BV - OS support for register saving.
-__declspec(naked)
-static uint32 XGetBV32(void) {
-  _asm _emit 0x0f
-  _asm _emit 0x01
-  _asm _emit 0xd0  // xgetbv
-  _asm ret
+static uint32 XGetBV(unsigned int xcr) {
+  return static_cast<uint32>(_xgetbv(xcr));
 }
+#elif !defined(__CLR_VER) && defined(_M_IX86)
+#define HAS_XGETBV
+__declspec(naked) __declspec(align(16))
+static uint32 XGetBV(unsigned int xcr) {
+  __asm {
+    mov        ecx, [esp + 4]    // xcr
+    xgetbv  // updates eax and edx.  edx unused/
+    ret
+  }
+}
+#elif defined(__i386__) || defined(__x86_64__)
+#define HAS_XGETBV
+static uint32 XGetBV(unsigned int xcr) {
+  uint32 xcr_feature_mask;
+  asm volatile (
+    ".byte 0x0f, 0x01, 0xd0\n"
+    : "=a"(xcr_feature_mask)
+    : "c"(xcr)
+    : "memory", "cc", "edx");  // edx unused.
+  return xcr_feature_mask;
+}
+#endif
+#ifdef HAS_XGETBV
+static const int kXCR_XFEATURE_ENABLED_MASK = 0;
 #endif
 
 // based on libvpx arm_cpudetect.c
@@ -105,13 +123,13 @@ int InitCpuFlags() {
               ((cpu_info[2] & 0x00000200) ? kCpuHasSSSE3 : 0) |
               ((cpu_info[2] & 0x00080000) ? kCpuHasSSE41 : 0) |
               ((cpu_info[2] & 0x00100000) ? kCpuHasSSE42 : 0) |
-              // TODO(fbarchard): AVX test BV same as AVX2.
               (((cpu_info[2] & 0x18000000) == 0x18000000) ? kCpuHasAVX : 0) |
               kCpuInitialized | kCpuHasX86;
 #ifdef HAS_XGETBV
   if (cpu_info_ & kCpuHasAVX) {
     __cpuid(cpu_info, 7);
-    if ((cpu_info[1] & 0x00000020) && ((XGetBV32() & 0x06) == 0x06)) {
+    if ((cpu_info[1] & 0x00000020) &&
+        ((XGetBV(kXCR_XFEATURE_ENABLED_MASK) & 0x06) == 0x06)) {
       cpu_info_ |= kCpuHasAVX2;
     }
   }
diff --git a/source/row_neon.cc b/source/row_neon.cc
index 465c5e737..6988766be 100644
--- a/source/row_neon.cc
+++ b/source/row_neon.cc
@@ -459,7 +459,7 @@ void ARGBToRGBARow_NEON(const uint8* src_argb, uint8* dst_rgba, int pix) {
   "1:                                          \n"
     "vld4.8     {d5, d6, d7, d8}, [%0]!        \n"  // load 8 pixels of ARGB.
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vmov.u8    d4, d8                         \n"
+    "vmov.u8    d4, d8                         \n"  // move A before RGB.
     "vst4.8     {d4, d5, d6, d7}, [%1]!        \n"  // store 8 pixels of RGBA.
     "bgt        1b                             \n"
   : "+r"(src_argb),  // %0