From d2ee4e7d277923c8eabe1e6ac6c94e7dd40e965f Mon Sep 17 00:00:00 2001
From: nihuini <nihuini@tencent.com>
Date: Fri, 19 Jan 2018 16:15:13 +0800
Subject: [PATCH] ld1 and st1 handle data endian mode per element

---
 src/mat.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/src/mat.cpp b/src/mat.cpp
index 0fe7ac6dc..c0ae337ea 100644
--- a/src/mat.cpp
+++ b/src/mat.cpp
@@ -272,10 +272,10 @@ Mat Mat::from_float16(const unsigned short* data, int size)
     {
     asm volatile(
         "0:                             \n"
-        "ldr    d0, [%1], #8            \n"
+        "ld1    {v0.4h}, [%1], #8       \n"
         "fcvtl  v1.4s, v0.4h            \n"
         "subs   %w0, %w0, #1            \n"
-        "str    q1, [%2], #16           \n"
+        "st1    {v1.4s}, [%2], #16      \n"
         "bne    0b                      \n"
         : "=r"(nn),     // %0
           "=r"(data),   // %1