diff --git a/src/operators/kernel/arm/relu_kernel.cpp b/src/operators/kernel/arm/relu_kernel.cpp
index 2237f6187b94620923e460fbb0785f71e8e9d28f..e7e0941a4d0bf48d86525cc52ee33301cdcbf67e 100644
--- a/src/operators/kernel/arm/relu_kernel.cpp
+++ b/src/operators/kernel/arm/relu_kernel.cpp
@@ -40,62 +40,63 @@ void ReluKernel<CPU, float>::Compute(const ReluParam &param) const {
   int numel = input_x->numel();
   if (numel > 32) {
     asm volatile(
-    "pld        [%[input_x_ptr], #0]        \n\t"
-    "vmov.f32   q8,    #0.0                 \n\t"
-    "subs %[num], %[num], #32                \n\t"
-    "blt        end_num_%=                  \n\t"
-    "loop_num_%=:                           \n\t"
-    "pld        [%[input_x_ptr], #1024]      \n\t"
+        "pld        [%[input_x_ptr], #0]        \n\t"
+        "vmov.f32   q8,    #0.0                 \n\t"
+        "subs %[num], %[num], #32                \n\t"
+        "blt        end_num_%=                  \n\t"
+        "loop_num_%=:                           \n\t"
+        "pld        [%[input_x_ptr], #1024]      \n\t"
 
-    "vld1.32 {q0, q1}, [%[input_x_ptr]]!    \n\t"
-    "vld1.32 {q2, q3}, [%[input_x_ptr]]!    \n\t"
-    "vld1.32 {q4, q5}, [%[input_x_ptr]]!    \n\t"
-    "vld1.32 {q6, q7}, [%[input_x_ptr]]!    \n\t"
+        "vld1.32 {q0, q1}, [%[input_x_ptr]]!    \n\t"
+        "vld1.32 {q2, q3}, [%[input_x_ptr]]!    \n\t"
+        "vld1.32 {q4, q5}, [%[input_x_ptr]]!    \n\t"
+        "vld1.32 {q6, q7}, [%[input_x_ptr]]!    \n\t"
 
-    "vmax.f32 q0, q0, q8                   \n\t"
-    "vmax.f32 q1, q1, q8                    \n\t"
-    "vmax.f32 q2, q2, q8                   \n\t"
-    "vmax.f32 q3, q3, q8                   \n\t"
-    "vmax.f32 q4, q4, q8                   \n\t"
-    "vmax.f32 q5, q5, q8                   \n\t"
-    "vmax.f32 q6, q6, q8                   \n\t"
-    "vmax.f32 q7, q7, q8                   \n\t"
+        "vmax.f32 q0, q0, q8                   \n\t"
+        "vmax.f32 q1, q1, q8                    \n\t"
+        "vmax.f32 q2, q2, q8                   \n\t"
+        "vmax.f32 q3, q3, q8                   \n\t"
+        "vmax.f32 q4, q4, q8                   \n\t"
+        "vmax.f32 q5, q5, q8                   \n\t"
+        "vmax.f32 q6, q6, q8                   \n\t"
+        "vmax.f32 q7, q7, q8                   \n\t"
 
-    "vst1.32 {q0, q1}, [%[out_ptr]]!        \n\t"
-    "vst1.32 {q2, q3}, [%[out_ptr]]!       \n\t"
-    "vst1.32 {q4, q5}, [%[out_ptr]]!       \n\t"
-    "vst1.32 {q6, q7}, [%[out_ptr]]!       \n\t"
+        "vst1.32 {q0, q1}, [%[out_ptr]]!        \n\t"
+        "vst1.32 {q2, q3}, [%[out_ptr]]!       \n\t"
+        "vst1.32 {q4, q5}, [%[out_ptr]]!       \n\t"
+        "vst1.32 {q6, q7}, [%[out_ptr]]!       \n\t"
 
-    "subs %[num], %[num], #32              \n\t"
-    "bge        loop_num_%=                \n\t"
-    "end_num_%=:                           \n\t"
-    "cmp %[num], #0                         \n\t"
-    "bge   end_%=                          \n\t"
-    "mov r6, #4                             \n\t"
-    "mul r5, %[num], r6                     \n\t"
-    "add %[input_x_ptr], %[input_x_ptr], r5     \n\t"
-    "vld1.32 {q0, q1}, [%[input_x_ptr]]!    \n\t"
-    "vld1.32 {q2, q3}, [%[input_x_ptr]]!    \n\t"
-    "vld1.32 {q4, q5}, [%[input_x_ptr]]!    \n\t"
-    "vld1.32 {q6, q7}, [%[input_x_ptr]]!    \n\t"
-    "vmax.f32 q0, q0, q8                   \n\t"
-    "vmax.f32 q1, q1, q8                    \n\t"
-    "vmax.f32 q2, q2, q8                   \n\t"
-    "vmax.f32 q3, q3, q8                   \n\t"
-    "vmax.f32 q4, q4, q8                   \n\t"
-    "vmax.f32 q5, q5, q8                   \n\t"
-    "vmax.f32 q6, q6, q8                   \n\t"
-    "vmax.f32 q7, q7, q8                   \n\t"
-    "add %[out_ptr], %[out_ptr], r5       \n\t"
-    "vst1.32 {q0, q1}, [%[out_ptr]]!        \n\t"
-    "vst1.32 {q2, q3}, [%[out_ptr]]!       \n\t"
-    "vst1.32 {q4, q5}, [%[out_ptr]]!       \n\t"
-    "vst1.32 {q6, q7}, [%[out_ptr]]!       \n\t"
-    "end_%=:                                \n\t"
-    :
-    :[out_ptr]"r"(out_ptr), [input_x_ptr]"r"(input_x_ptr), [num]"r"(numel)
-    :"memory", "q0", "q1", "q2", "q3", "q4","q5","q6", "q7", "q8", "r5", "r6"
-    );
+        "subs %[num], %[num], #32              \n\t"
+        "bge        loop_num_%=                \n\t"
+        "end_num_%=:                           \n\t"
+        "cmp %[num], #0                         \n\t"
+        "bge   end_%=                          \n\t"
+        "mov r6, #4                             \n\t"
+        "mul r5, %[num], r6                     \n\t"
+        "add %[input_x_ptr], %[input_x_ptr], r5     \n\t"
+        "vld1.32 {q0, q1}, [%[input_x_ptr]]!    \n\t"
+        "vld1.32 {q2, q3}, [%[input_x_ptr]]!    \n\t"
+        "vld1.32 {q4, q5}, [%[input_x_ptr]]!    \n\t"
+        "vld1.32 {q6, q7}, [%[input_x_ptr]]!    \n\t"
+        "vmax.f32 q0, q0, q8                   \n\t"
+        "vmax.f32 q1, q1, q8                    \n\t"
+        "vmax.f32 q2, q2, q8                   \n\t"
+        "vmax.f32 q3, q3, q8                   \n\t"
+        "vmax.f32 q4, q4, q8                   \n\t"
+        "vmax.f32 q5, q5, q8                   \n\t"
+        "vmax.f32 q6, q6, q8                   \n\t"
+        "vmax.f32 q7, q7, q8                   \n\t"
+        "add %[out_ptr], %[out_ptr], r5       \n\t"
+        "vst1.32 {q0, q1}, [%[out_ptr]]!        \n\t"
+        "vst1.32 {q2, q3}, [%[out_ptr]]!       \n\t"
+        "vst1.32 {q4, q5}, [%[out_ptr]]!       \n\t"
+        "vst1.32 {q6, q7}, [%[out_ptr]]!       \n\t"
+        "end_%=:                                \n\t"
+        :
+        :
+        [out_ptr] "r"(out_ptr), [input_x_ptr] "r"(input_x_ptr), [num] "r"(numel)
+        : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "r5",
+          "r6");
   } else {
     ReluFunctor<float> func_;
     math::Transform trans;