diff --git a/CMakeLists.txt b/CMakeLists.txt
index 9c3ff01e5d3e20923021904cdbe9008a11cc30ce..f2e3a2a4e974dcdb0431bced75a35305d4fbddfa 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -25,9 +25,10 @@ endif()
 set(CMAKE_CXX_FLAGS "-std=c++14 -O3 -s ${CMAKE_CXX_FLAGS}")
 if (DEBUGING)
     set(CMAKE_BUILD_TYPE Debug)
-    set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS}")
+    set(CMAKE_CXX_FLAGS_DEBUG "-O3 -DNDEBUG")
 else()
     set(CMAKE_BUILD_TYPE Release)
+    set(CMAKE_CXX_FLAGS_RELEASE "-O3 -DNDEBUG")
 endif ()
 
 if(DEBUGING)
diff --git a/src/operators/math/gemm.cpp b/src/operators/math/gemm.cpp
index da3dacb58a72d779d2ccd1224bbf4eab12dfbb91..7ade30600aa47feaf054d0bef043cae3c1fdd1e4 100644
--- a/src/operators/math/gemm.cpp
+++ b/src/operators/math/gemm.cpp
@@ -375,12 +375,15 @@ void AddDot4x4(int k, float alpha, const float *a, int lda, const float *b,
 
       "subs       %[kc2], %[kc2], #1  \n\t"
       "blt        end_kc2_%=          \n\t"
+      "loop_kc2_%=:                   \n\t"
       "vld1.32    {q0}, [%[a]]!       \n\t"
       "vld1.32    {q1}, [%[b]]!       \n\t"
       "vmla.f32   q10, q1, d0[0]      \n\t"
       "vmla.f32   q11, q1, d0[1]      \n\t"
       "vmla.f32   q12, q1, d1[0]      \n\t"
       "vmla.f32   q13, q1, d1[1]      \n\t"
+      "subs       %[kc2], %[kc2], #1  \n\t"
+      "bge        loop_kc2_%=         \n\t"
       "end_kc2_%=:                    \n\t"
 
       "cmp        %[mc],      #4      \n\t"