diff --git a/lite/backends/arm/math/prior_box.cc b/lite/backends/arm/math/prior_box.cc
old mode 100644
new mode 100755
index 4ef7356e67cee4c47ddf3eb16ed5286b4271b41a..fd0729d8fce344f0360526fd38a9d7ce4064d01d
--- a/lite/backends/arm/math/prior_box.cc
+++ b/lite/backends/arm/math/prior_box.cc
@@ -21,7 +21,7 @@ namespace lite {
 namespace arm {
 namespace math {
 
-const int MALLOC_ALIGN = 16;
+const int MALLOC_ALIGN = 64;
 
 void* fast_malloc(size_t size) {
   size_t offset = sizeof(void*) + MALLOC_ALIGN - 1;
@@ -46,8 +46,8 @@ void fast_free(void* ptr) {
 
 void density_prior_box(const lite::Tensor* input,
                        const lite::Tensor* image,
-                       lite::Tensor** boxes,
-                       lite::Tensor** variances,
+                       lite::Tensor* boxes,
+                       lite::Tensor* variances,
                        const std::vector<float>& min_size_,
                        const std::vector<float>& fixed_size_,
                        const std::vector<float>& fixed_ratio_,
@@ -69,11 +69,11 @@ void density_prior_box(const lite::Tensor* input,
   int win1 = input->dims()[3];
   int hin1 = input->dims()[2];
   DDim shape_out({hin1, win1, prior_num_, 4});
-  (*boxes)->Resize(shape_out);
-  (*variances)->Resize(shape_out);
+  boxes->Resize(shape_out);
+  variances->Resize(shape_out);
 
-  float* _cpu_data = (*boxes)->mutable_data<float>();
-  float* _variance_data = (*variances)->mutable_data<float>();
+  float* _cpu_data = boxes->mutable_data<float>();
+  float* _variance_data = variances->mutable_data<float>();
 
   const int width = win1;
   const int height = hin1;
@@ -329,8 +329,8 @@ void density_prior_box(const lite::Tensor* input,
 
 void prior_box(const lite::Tensor* input,
                const lite::Tensor* image,
-               lite::Tensor** boxes,
-               lite::Tensor** variances,
+               lite::Tensor* boxes,
+               lite::Tensor* variances,
                const std::vector<float>& min_size,
                const std::vector<float>& max_size,
                const std::vector<float>& aspect_ratio,
diff --git a/lite/backends/arm/math/prior_box.h b/lite/backends/arm/math/prior_box.h
old mode 100644
new mode 100755
index 03fd62751081e491ddfb23f196d52153db5d3a5f..2a22854a6194818b23a355cc0111402312c1fcf2
--- a/lite/backends/arm/math/prior_box.h
+++ b/lite/backends/arm/math/prior_box.h
@@ -25,8 +25,8 @@ namespace math {
 
 void density_prior_box(const lite::Tensor* input,
                        const lite::Tensor* image,
-                       lite::Tensor** boxes,
-                       lite::Tensor** variances,
+                       lite::Tensor* boxes,
+                       lite::Tensor* variances,
                        const std::vector<float>& min_size_,
                        const std::vector<float>& fixed_size_,
                        const std::vector<float>& fixed_ratio_,
@@ -47,8 +47,8 @@ void density_prior_box(const lite::Tensor* input,
 
 void prior_box(const lite::Tensor* input,
                const lite::Tensor* image,
-               lite::Tensor** boxes,
-               lite::Tensor** variances,
+               lite::Tensor* boxes,
+               lite::Tensor* variances,
                const std::vector<float>& min_size,
                const std::vector<float>& max_size,
                const std::vector<float>& aspect_ratio,
diff --git a/lite/kernels/arm/prior_box_compute.cc b/lite/kernels/arm/prior_box_compute.cc
index 48ae1e94dd74453c9160604a94ecfabd0e516034..7ac1c5c2e009d33897ce6bc5c2176a648f517a3c 100644
--- a/lite/kernels/arm/prior_box_compute.cc
+++ b/lite/kernels/arm/prior_box_compute.cc
@@ -46,7 +46,7 @@ inline void ExpandAspectRatios(const std::vector<float>& input_aspect_ratior,
   }
 }
 
-void PriorBoxCompute::Run() {
+void PriorBoxCompute::PrepareForRun() {
   auto& param = Param<operators::PriorBoxParam>();
 
   bool is_flip = param.flip;
@@ -69,8 +69,8 @@ void PriorBoxCompute::Run() {
 
   lite::arm::math::prior_box(param.input,
                              param.image,
-                             &param.boxes,
-                             &param.variances,
+                             &out_boxes,
+                             &variances,
                              min_size,
                              max_size,
                              aspect_ratios_vec,
@@ -85,6 +85,22 @@ void PriorBoxCompute::Run() {
                              is_clip,
                              order,
                              min_max_aspect_ratios_order);
+  this->_flag_init = true;
+}
+
+void PriorBoxCompute::Run() {
+  if (!this->_flag_init) {
+    LOG(FATAL) << "ERROR: init priorbox first\n";
+  }
+  auto& param = Param<operators::PriorBoxParam>();
+  param.boxes->Resize(out_boxes.dims());
+  param.variances->Resize(out_boxes.dims());
+  memcpy(param.boxes->mutable_data<float>(),
+         out_boxes.data<float>(),
+         param.boxes->numel());
+  memcpy(param.variances->mutable_data<float>(),
+         variances.data<float>(),
+         param.variances->numel());
 }
 
 }  // namespace arm
diff --git a/lite/kernels/arm/prior_box_compute.h b/lite/kernels/arm/prior_box_compute.h
index d44406a98fa4d6ec142fb8cf1a83205599eeb5b2..277c3ed37c78a9e79bbb874b662f4ce100d88551 100644
--- a/lite/kernels/arm/prior_box_compute.h
+++ b/lite/kernels/arm/prior_box_compute.h
@@ -24,10 +24,15 @@ namespace arm {
 class PriorBoxCompute : public KernelLite<TARGET(kARM), PRECISION(kFloat)> {
  public:
   using param_t = operators::PriorBoxParam;
-
+  void PrepareForRun() override;
   void Run() override;
 
   virtual ~PriorBoxCompute() = default;
+
+ private:
+  bool _flag_init{false};
+  lite::Tensor out_boxes;
+  lite::Tensor variances;
 };
 
 }  // namespace arm