diff --git a/src/operators/batchnorm_op.cpp b/src/operators/batchnorm_op.cpp
index d2fbd9fb6b0192b3728678ae92de7bf8e44e3620..644a27c586375bc66d327e18ac5182e8fce2893b 100644
--- a/src/operators/batchnorm_op.cpp
+++ b/src/operators/batchnorm_op.cpp
@@ -35,7 +35,6 @@ namespace ops = paddle_mobile::operators;
 REGISTER_OPERATOR_CPU(batch_norm, ops::BatchNormOp);
 #endif
 #ifdef PADDLE_MOBILE_MALI_GPU
-USE_OP_MALI_GPU(batch_norm);
 REGISTER_OPERATOR_MALI_GPU(batch_norm, ops::BatchNormOp);
 #endif
 #ifdef PADDLE_MOBILE_FPGA
diff --git a/src/operators/fusion_conv_add.cpp b/src/operators/fusion_conv_add.cpp
index be70370f9de0963bbe6625513257be890e36dacb..656d30c4e1921914d3fe80d930f4219d73f025ea 100644
--- a/src/operators/fusion_conv_add.cpp
+++ b/src/operators/fusion_conv_add.cpp
@@ -76,7 +76,7 @@ namespace ops = paddle_mobile::operators;
 REGISTER_OPERATOR_CPU(fusion_conv_add, ops::FusionConvAddOp);
 #endif
 #ifdef PADDLE_MOBILE_MALI_GPU
-REGISTER_OPERATOR_MALI_GPU(conv_add, ops::FusionConvAddOp);
+REGISTER_OPERATOR_MALI_GPU(fusion_conv_add, ops::FusionConvAddOp);
 #endif
 #ifdef PADDLE_MOBILE_FPGA
 #endif
diff --git a/src/operators/fusion_conv_add.h b/src/operators/fusion_conv_add.h
index 02c9d910b955fd9398df9406a5f730c4a7abbfee..bc623efc8c9288a6007a8141220a2d40b29d7bdb 100644
--- a/src/operators/fusion_conv_add.h
+++ b/src/operators/fusion_conv_add.h
@@ -96,7 +96,7 @@ static framework::FusionOpRegistrar convadd_registrar(
 USE_OP_CPU(fusion_conv_add);
 #endif
 #ifdef PADDLE_MOBILE_MALI_GPU
-USE_OP_MALI_GPU(conv_add);
+USE_OP_MALI_GPU(fusion_conv_add);
 #endif
 #ifdef PADDLE_MOBILE_FPGA
 #endif
diff --git a/src/operators/fusion_fc_op.cpp b/src/operators/fusion_fc_op.cpp
index 1b2a46defc520519e0fb61779cf45059f0a54913..0ca3c26c47e91771ab7019261f78815ad8463408 100644
--- a/src/operators/fusion_fc_op.cpp
+++ b/src/operators/fusion_fc_op.cpp
@@ -80,7 +80,7 @@ namespace ops = paddle_mobile::operators;
 REGISTER_OPERATOR_CPU(fusion_fc, ops::FusionFcOp);
 #endif
 #ifdef PADDLE_MOBILE_MALI_GPU
-REGISTER_OPERATOR_MALI_GPU(fc, ops::FusionFcOp);
+REGISTER_OPERATOR_MALI_GPU(fusion_fc, ops::FusionFcOp);
 #endif
 #ifdef PADDLE_MOBILE_FPGA
 #endif
diff --git a/src/operators/fusion_fc_op.h b/src/operators/fusion_fc_op.h
index c07d59e31e8124325d48a5b9ff208e1e425146a8..b545bb8380aedcc31200080beaec4c56f137ad13 100644
--- a/src/operators/fusion_fc_op.h
+++ b/src/operators/fusion_fc_op.h
@@ -78,7 +78,7 @@ extern framework::FusionOpRegistrar fc_registrar(new FusionFcMatcher());
 
 #ifndef CONV_CPU_REGISTER
 #define CONV_CPU_REGISTER
-static framework::FusionOpRegistrar fc_registrar(new FusionFcMatcher());
+extern framework::FusionOpRegistrar fc_registrar(new FusionFcMatcher());
 #endif
 
 #endif
@@ -93,7 +93,7 @@ static framework::FusionOpRegistrar fc_registrar(new FusionFcMatcher());
 USE_OP_CPU(fusion_fc);
 #endif
 #ifdef PADDLE_MOBILE_MALI_GPU
-USE_OP_MALI_GPU(fc);
+USE_OP_MALI_GPU(fusion_fc);
 #endif
 #ifdef PADDLE_MOBILE_FPGA
 #endif
diff --git a/src/operators/kernel/mali/acl_operator.cc b/src/operators/kernel/mali/acl_operator.cc
old mode 100644
new mode 100755
diff --git a/src/operators/kernel/mali/acl_operator.h b/src/operators/kernel/mali/acl_operator.h
old mode 100644
new mode 100755
index c2e13283b1c679d6dfc8972af5ace5e579d568e6..bf8200d486f91998c79540177ab1b26596a3e9dc
--- a/src/operators/kernel/mali/acl_operator.h
+++ b/src/operators/kernel/mali/acl_operator.h
@@ -225,6 +225,7 @@ class AclParameters {
 
   bool is_global_pool;
   bool is_channel_concat;
+  bool is_bypass;
 
   std::vector<framework::LoDTensor *> in_tensor;
 };
diff --git a/src/operators/kernel/mali/acl_tensor.cc b/src/operators/kernel/mali/acl_tensor.cc
old mode 100644
new mode 100755
diff --git a/src/operators/kernel/mali/acl_tensor.h b/src/operators/kernel/mali/acl_tensor.h
old mode 100644
new mode 100755
diff --git a/src/operators/kernel/mali/batchnorm_kernel.cpp b/src/operators/kernel/mali/batchnorm_kernel.cpp
old mode 100644
new mode 100755
index e749f4223e54988c0db54ac739a9f73bcd8c6240..ad648d615cd8f9134b212d484d7174c95e027551
--- a/src/operators/kernel/mali/batchnorm_kernel.cpp
+++ b/src/operators/kernel/mali/batchnorm_kernel.cpp
@@ -136,7 +136,7 @@ bool BatchNormKernel<GPU_MALI, float>::Init(BatchNormParam* param) {
     acl_op = new AclBatchNormOp<GPU_MALI, float>();
     this->SetAclOp((void*)acl_op, (void*)this);
   }
-  if (acl_op->Bypass_acl(param)) {
+  if (acl_op->Bypass_acl(*param)) {
     std::cout << "init acl failed" << std::endl;
     return false;
   }
diff --git a/src/operators/kernel/mali/concat_kernel.cpp b/src/operators/kernel/mali/concat_kernel.cpp
old mode 100644
new mode 100755
index e0c78841c30c04ff43dd56c24ecad444712fd6d7..aaa586b6d977bfca96e596261ec090637cf87207
--- a/src/operators/kernel/mali/concat_kernel.cpp
+++ b/src/operators/kernel/mali/concat_kernel.cpp
@@ -102,14 +102,14 @@ class AclConcatOp : public acl::ACLOperator {
 };
 
 template <>
-bool ConcatKernel<GPU_MALI, float>::Init(const ConcatParam& param) const {
+bool ConcatKernel<GPU_MALI, float>::Init(ConcatParam* param) {
   AclConcatOp<GPU_MALI, float>* acl_op =
       reinterpret_cast<AclConcatOp<GPU_MALI, float>*>(this->GetAclOp());
   if (acl_op == nullptr) {
     acl_op = new AclConcatOp<GPU_MALI, float>();
     this->SetAclOp((void*)acl_op, (void*)this);
   }
-  if (acl_op->Bypass_acl(param)) {
+  if (acl_op->Bypass_acl(*param)) {
     std::cout << "init acl failed" << std::endl;
     return false;
   }
diff --git a/src/operators/kernel/mali/conv_add_kernel.cpp b/src/operators/kernel/mali/conv_add_kernel.cpp
index 57b6d82ac68748c909be6da2fc2a490e4aa262dc..318db016d12f36981c07627139bcc49d07162d52 100644
--- a/src/operators/kernel/mali/conv_add_kernel.cpp
+++ b/src/operators/kernel/mali/conv_add_kernel.cpp
@@ -196,15 +196,14 @@ class AclConvAddOp : public acl::ACLOperator {
 };
 
 template <>
-bool ConvAddKernel<GPU_MALI, float>::Init(
-    const FusionConvAddParam& param) const {
+bool ConvAddKernel<GPU_MALI, float>::Init(FusionConvAddParam* param) {
   AclConvAddOp<GPU_MALI, float>* acl_op =
       reinterpret_cast<AclConvAddOp<GPU_MALI, float>*>(this->GetAclOp());
   if (acl_op == nullptr) {
     acl_op = new AclConvAddOp<GPU_MALI, float>();
     this->SetAclOp((void*)acl_op, (void*)this);
   }
-  if (acl_op->Bypass_acl(param)) {
+  if (acl_op->Bypass_acl(*param)) {
     std::cout << "init acl failed" << std::endl;
     return false;
   }
diff --git a/src/operators/kernel/mali/conv_kernel.cpp b/src/operators/kernel/mali/conv_kernel.cpp
old mode 100644
new mode 100755
index 30bb763728763785b2f66d0283743a91463d7f9a..c548977ebaa34fabc1b1fe54d6db9690bcb424f1
--- a/src/operators/kernel/mali/conv_kernel.cpp
+++ b/src/operators/kernel/mali/conv_kernel.cpp
@@ -203,7 +203,7 @@ bool ConvKernel<GPU_MALI, float>::Init(ConvParam* param) {
     acl_op = new AclConvOp<GPU_MALI, float>();
     this->SetAclOp((void*)acl_op, (void*)this);
   }
-  if (acl_op->Bypass_acl(param)) {
+  if (acl_op->Bypass_acl(*param)) {
     std::cout << "init acl failed" << std::endl;
     return false;
   }
diff --git a/src/operators/kernel/mali/elementwise_add_kernel.cpp b/src/operators/kernel/mali/elementwise_add_kernel.cpp
old mode 100644
new mode 100755
diff --git a/src/operators/kernel/mali/fushion_fc_kernel.cpp b/src/operators/kernel/mali/fushion_fc_kernel.cpp
old mode 100644
new mode 100755
diff --git a/src/operators/kernel/mali/lrn_kernel.cpp b/src/operators/kernel/mali/lrn_kernel.cpp
index c063ec8783382ccef79086368df8a97320010c23..4fb5fca8695dccc45c7169d8572618965b3d84a3 100644
--- a/src/operators/kernel/mali/lrn_kernel.cpp
+++ b/src/operators/kernel/mali/lrn_kernel.cpp
@@ -20,6 +20,7 @@ limitations under the License. */
 #ifdef PADDLE_MOBILE_MALI_GPU
 #include "acl_operator.h"
 #include "framework/operator.h"
+#include "operators/kernel/central-arm-func/lrn_arm_func.h"
 #include "operators/op_param.h"
 
 namespace paddle_mobile {
@@ -59,12 +60,15 @@ class AclLrnOp : public acl::ACLOperator {
     acl_configure(lrn, this, norm_info);
   }
 
+  void Set_bypass(bool bypass) { args.is_bypass = bypass; }
+
   void RunAcl(void* input, void* output) {
     acl::ACLOperator::acl_run(input, output);
   }
   bool Bypass_acl(const LrnParam& param) {
     bool bypass_acl = false;
     AclParametersByContext(param);
+    InitAclLayer(param);
     // for performance, more groups impact GPU performance
     if (this->force_bypass_acl_path_) {
       bypass_acl = true;
@@ -107,13 +111,18 @@ class AclLrnOp : public acl::ACLOperator {
 };
 
 template <>
-bool LrnKernel<GPU_MALI, float>::Init(const LrnParam& param) const {
+bool LrnKernel<GPU_MALI, float>::Init(LrnParam* param) {
   AclLrnOp<GPU_MALI, float>* acl_op =
       reinterpret_cast<AclLrnOp<GPU_MALI, float>*>(this->GetAclOp());
   if (acl_op == nullptr) {
     acl_op = new AclLrnOp<GPU_MALI, float>();
     this->SetAclOp((void*)acl_op, (void*)this);
   }
+  if (acl_op->Bypass_acl(*param)) {
+    acl_op->Set_bypass(true);
+    std::cout << "init acl failed" << std::endl;
+    return true;
+  }
   return true;
 }
 
@@ -125,14 +134,14 @@ void LrnKernel<GPU_MALI, float>::Compute(const LrnParam& param) const {
   if (acl_op == nullptr) {
     return;
   }
-  if (acl_op->Bypass_acl(param)) {
-    std::cout << "init acl failed" << std::endl;
+  acl::AclParameters& args = acl_op->getargs();
+  if (args.is_bypass) {
+    std::cout << "bypass op" << std::endl;
+    LrnCompute<float>(param);
     return;
   }
-  acl::AclParameters& args = acl_op->getargs();
   const float* input_data = (const float*)args.input_data;
   const float* output_data = (const float*)args.output_data;
-  acl_op->InitAclLayer(param);
   for (int n = 0; n < args.batch; ++n) {
     acl_op->RunAcl((void*)input_data, (void*)output_data);
     input_data += args.in_depth * args.in_cols * args.in_rows;
diff --git a/src/operators/kernel/mali/mul_kernel.cpp b/src/operators/kernel/mali/mul_kernel.cpp
old mode 100644
new mode 100755
diff --git a/src/operators/kernel/mali/pool_kernel.cpp b/src/operators/kernel/mali/pool_kernel.cpp
old mode 100644
new mode 100755
index 66f06e0e8f60b6bb77b800fbe9ea934af0f77ef6..1f49391341d69a0690352c69c9c208550f8e1c24
--- a/src/operators/kernel/mali/pool_kernel.cpp
+++ b/src/operators/kernel/mali/pool_kernel.cpp
@@ -180,14 +180,14 @@ class AclPoolOp : public acl::ACLOperator {
 };
 
 template <>
-bool PoolKernel<GPU_MALI, float>::Init(const PoolParam& param) const {
+bool PoolKernel<GPU_MALI, float>::Init(PoolParam* param) {
   AclPoolOp<GPU_MALI, float>* acl_op =
       reinterpret_cast<AclPoolOp<GPU_MALI, float>*>(this->GetAclOp());
   if (acl_op == nullptr) {
     acl_op = new AclPoolOp<GPU_MALI, float>();
     this->SetAclOp((void*)acl_op, (void*)this);
   }
-  if (acl_op->Bypass_acl(param)) {
+  if (acl_op->Bypass_acl(*param)) {
     std::cout << "init acl failed" << std::endl;
     return false;
   }
diff --git a/src/operators/kernel/mali/relu_kernel.cpp b/src/operators/kernel/mali/relu_kernel.cpp
old mode 100644
new mode 100755
index 760076c9dc60f5ded46a4c88729ae3f423dc1880..1a8c0f88543199e7a863cc44b5b0a6be3bc6212d
--- a/src/operators/kernel/mali/relu_kernel.cpp
+++ b/src/operators/kernel/mali/relu_kernel.cpp
@@ -100,14 +100,14 @@ class AclReluOp : public acl::ACLOperator {
 };
 
 template <>
-bool ReluKernel<GPU_MALI, float>::Init(const ReluParam& param) const {
+bool ReluKernel<GPU_MALI, float>::Init(ReluParam* param) {
   AclReluOp<GPU_MALI, float>* acl_op =
       reinterpret_cast<AclReluOp<GPU_MALI, float>*>(this->GetAclOp());
   if (acl_op == nullptr) {
     acl_op = new AclReluOp<GPU_MALI, float>();
     this->SetAclOp((void*)acl_op, (void*)this);
   }
-  if (acl_op->Bypass_acl(param)) {
+  if (acl_op->Bypass_acl(*param)) {
     std::cout << "init acl failed" << std::endl;
     return false;
   }
diff --git a/src/operators/kernel/mali/reshape_kernel.cpp b/src/operators/kernel/mali/reshape_kernel.cpp
old mode 100644
new mode 100755
diff --git a/src/operators/kernel/mali/softmax_kernel.cpp b/src/operators/kernel/mali/softmax_kernel.cpp
old mode 100644
new mode 100755
index 777e85a1dd3684e7e929662949b3264a79f9b142..37d2f2b6b1656602e5acfecd3ac79733f570844d
--- a/src/operators/kernel/mali/softmax_kernel.cpp
+++ b/src/operators/kernel/mali/softmax_kernel.cpp
@@ -97,14 +97,14 @@ class AclSoftmaxOp : public acl::ACLOperator {
 };
 
 template <>
-bool SoftmaxKernel<GPU_MALI, float>::Init(const SoftmaxParam& param) const {
+bool SoftmaxKernel<GPU_MALI, float>::Init(SoftmaxParam* param) {
   AclSoftmaxOp<GPU_MALI, float>* acl_op =
       reinterpret_cast<AclSoftmaxOp<GPU_MALI, float>*>(this->GetAclOp());
   if (acl_op == nullptr) {
     acl_op = new AclSoftmaxOp<GPU_MALI, float>();
     this->SetAclOp((void*)acl_op, (void*)this);
   }
-  if (acl_op->Bypass_acl(param)) {
+  if (acl_op->Bypass_acl(*param)) {
     std::cout << "init acl failed" << std::endl;
     return false;
   }