diff --git a/lite/backends/fpga/KD/debugger.hpp b/lite/backends/fpga/KD/debugger.hpp
index 81ae57fc88b25dd907c21efab7f79dfe7e524d98..6445b6ce41e75cecbacafb90aaa805e847f9478a 100755
--- a/lite/backends/fpga/KD/debugger.hpp
+++ b/lite/backends/fpga/KD/debugger.hpp
@@ -14,6 +14,8 @@
 
 #pragma once
 
+#include <fstream>
+#include <iostream>
 #include <string>
 #include <unordered_map>
 
@@ -37,8 +39,18 @@ class Debugger {
     }
   }
 
+  void tick(std::string key) {
+    float value = 0;
+    if (tick_tock_map.count(key) > 0) {
+      value += tick_tock_map[key] = value;
+    }
+  }
+
+  void tock(std::string key) {}
+
  private:
   std::unordered_map<std::string, bool> op_config;
+  std::unordered_map<std::string, float> tick_tock_map;
   Debugger() {
     op_config["concat"] = true;
     op_config["pooling"] = true;
diff --git a/lite/backends/fpga/KD/dl_engine.hpp b/lite/backends/fpga/KD/dl_engine.hpp
index eddf5ca454cdc9e91f87d6e4f2c8dfc13f35fdc6..fe66c84143fbc05f0b52a11e2e315b7f3db9054c 100755
--- a/lite/backends/fpga/KD/dl_engine.hpp
+++ b/lite/backends/fpga/KD/dl_engine.hpp
@@ -15,6 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include <stdio.h>
+
 #include "lite/backends/fpga/KD/llapi/filter.h"
 #include "lite/backends/fpga/KD/llapi/zynqmp_api.h"
 
diff --git a/lite/backends/fpga/KD/io.cpp b/lite/backends/fpga/KD/io.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..705faefe7443270c40b7a7c88f517f5381d83919
--- /dev/null
+++ b/lite/backends/fpga/KD/io.cpp
@@ -0,0 +1,43 @@
+/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "io.hpp"
+
+namespace paddle {
+namespace zynqmp {
+
+// FpgaIO::FpgaIO() {}
+// void FpgaIO::setMutex(std::mutex* mtx) { mtx_ = mtx; }
+
+// void FpgaIO::setConditionVariable(std::condition_variable* condition) {
+//   condition_ = condition;
+// }
+
+// void FpgaIO::lock() {
+//   if (mtx_ != nullptr && !locked_) {
+//     mtx_->lock();
+//     locked_ = true;
+//   }
+// }
+
+// void FpgaIO::unlock() {
+//   if (mtx_ != nullptr) {
+//     mtx_->unlock();
+//     condition_->notify_one();
+//   }
+//   locked_ = false;
+// }
+
+}  // namespace zynqmp
+}  // namespace paddle
diff --git a/lite/backends/fpga/KD/io.hpp b/lite/backends/fpga/KD/io.hpp
new file mode 100644
index 0000000000000000000000000000000000000000..408c6c1765a83b3f54329ac0baea841ae17996e6
--- /dev/null
+++ b/lite/backends/fpga/KD/io.hpp
@@ -0,0 +1,50 @@
+/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include <stdio.h>
+// #include <condition_variable>
+// #include <mutex>
+
+namespace paddle {
+namespace zynqmp {
+
+class FpgaIO {
+ public:
+  static FpgaIO& get_instance() {
+    static FpgaIO s_instance;
+    return s_instance;
+  }
+
+  void allocData(size_t s) { data_ = new float[s]; }
+
+  float* getData() { return data_; }
+
+  // void setMutex(std::mutex* mtx);
+  // void setConditionVariable(std::condition_variable* condition);
+  // void lock();
+  // void unlock();
+
+ private:
+  std::mutex* mtx_ = nullptr;
+  std::condition_variable* condition_ = nullptr;
+  bool locked_ = false;
+
+  float* data_ = nullptr;
+
+  FpgaIO();
+};
+}  // namespace zynqmp
+}  // namespace paddle
diff --git a/lite/backends/fpga/KD/pes/fully_connected_pe.hpp b/lite/backends/fpga/KD/pes/fully_connected_pe.hpp
index db3e05276171607da4cea421dd554846a00314a6..99ae89e29cf41daf3a7e28e3d391abefd0945c51 100755
--- a/lite/backends/fpga/KD/pes/fully_connected_pe.hpp
+++ b/lite/backends/fpga/KD/pes/fully_connected_pe.hpp
@@ -14,6 +14,8 @@ limitations under the License. */
 
 #pragma once
 
+#include <math.h>
+#include <cmath>
 #include <vector>
 
 #include "lite/backends/fpga/KD/pe.hpp"
@@ -37,10 +39,9 @@ class FullyConnectedPE : public PE {
     ConvParam& convParam_ = convPE_.param();
     Tensor* input = param_.input;
     convParam_.input = param_.input;
-    num_ = param_.input->shape().num();
-
     convParam_.output = param_.output;
-
+    convParam_.relu = param_.relu;
+    // convParam_.activeParam.type = param_.activeParam.type;
     convParam_.groups = 1;
     convParam_.strides = {1, 1};
     convParam_.paddings = {0, 0};
@@ -49,34 +50,54 @@ class FullyConnectedPE : public PE {
 
     int num = param_.filter->shape().channel();
     int chw = param_.filter->shape().num();
+    int align = 32;
+    int chw_aligned = ((chw + align - 1) / align) * align;
+
+    int infer_num = 1;
+    Shape in_shape(NCHW, {infer_num, chw_aligned, 1, 1});
+    aligned_input_.mutableData<float16>(FP16, in_shape);
+    convParam_.input = &aligned_input_;
+
+    Shape out_shape(NCHW, {infer_num, num, 1, 1});
+    aligned_output_.mutableData<float16>(FP16, out_shape);
+    convParam_.output = &aligned_output_;
 
     int height = param_.input->shape().height();
     int width = param_.input->shape().width();
-    int filter_channel = chw / height / width;
+    // int filter_channel = chw / height / width;
 
     int channel = param_.output->shape().channel();
-    Shape shape(NCHW, {num, filter_channel, height, width});
-    Tensor* conv_filter = new Tensor();
-    float* new_filter_data = conv_filter->mutableData<float>(FP32, shape);
+    Shape shape(NCHW, {num, chw_aligned, 1, 1});
+    float* new_filter_data = conv_filter_.mutableData<float>(FP32, shape);
     float* filter_data = param_.filter->data<float>();
 
+    memset(new_filter_data, 0, num * chw_aligned * sizeof(float));
+
     for (int i = 0; i < num; i++) {
       for (int j = 0; j < chw; j++) {
         float scale = filter_data[j * num + i];
-        new_filter_data[i * chw + j] = scale;
+        new_filter_data[i * chw_aligned + j] = scale;
       }
     }
-    conv_filter->flush();
-    convParam_.filter = conv_filter;
 
-    Shape sb_shape(N, {channel});
+    conv_filter_.flush();
+    convParam_.filter = &conv_filter_;
+    // param_.filter->saveToFile("param_filter", true);
+    // conv_filter->saveToFile("conv_filter", true);
+    // exit(-1);
+
+    Shape sb_shape(N, {num});
     float* scale_data = convParam_.scale()->mutableData<float>(FP32, sb_shape);
     float* bias_data = convParam_.bias()->mutableData<float>(FP32, sb_shape);
 
-    for (int i = 0; i < channel; i++) {
+    for (int i = 0; i < num; i++) {
       scale_data[i] = 1.0f;
       bias_data[i] = param_.bias->data<float>()[i];
     }
+    // for (int i = 0; i < num; i++) {
+    //   scale_data[i] = 1.0f;
+    //   bias_data[i] = param_.bias->data<float>()[i];
+    // }
     convParam_.scale()->flush();
     convParam_.bias()->flush();
 
@@ -84,15 +105,197 @@ class FullyConnectedPE : public PE {
     convPE_.apply();
   }
 
-  bool dispatch() { return convPE_.dispatch(); }
+  void cpu_compute() {
+    int num = param_.filter->shape().channel();
+    int chw = param_.filter->shape().num();
+
+    float* filter_data = param_.filter->data<float>();
+    float max = 0.0f;
+    Tensor* input = param_.input;
+    Tensor* output = param_.output;
+    float16* input_data = input->data<float16>();
+    float16* output_data = output->data<float16>();
+
+    for (int n = 0; n < input->shape().num(); n++) {
+      float16* input_data = input->data<float16>() + n * chw;
+      float16* output_data =
+          output->data<float16>() + n * output->shape().channel();
+      for (int i = 0; i < num; i++) {
+        float sum = 0;
+        float bias = param_.bias->data<float>()[i];
+        for (int j = 0; j < chw; j++) {
+          float scale = filter_data[j * num + i];
+          float data = half_to_float(input_data[j]);
+          sum += scale * data;
+        }
+        output_data[i] = float_to_half(sum + bias);
+        if (max < output_data[i]) {
+          max = output_data[i];
+        }
+      }
+    }
+
+    output->flush();
+    output->scale()[0] = max / 127.0f;
+    output->scale()[1] = 127.0f / max;
+  }
+
+  void cpu_compute1() {
+    int num = conv_filter_.shape().num();
+    int chw = conv_filter_.shape().channel();
+    // chw = 336;
+
+    float* filter_data = conv_filter_.data<float>();
+    float max = 0.0f;
+    Tensor* input = param_.input;
+    Tensor* output = param_.output;
+    float16* input_data = input->data<float16>();
+    float16* output_data = output->data<float16>();
+
+    for (int n = 0; n < input->shape().num(); n++) {
+      float16* input_data = input->data<float16>() + n * chw;
+      float16* output_data =
+          output->data<float16>() + n * output->shape().channel();
+      for (int i = 0; i < num; i++) {
+        float sum = 0;
+        float bias = param_.bias->data<float>()[i];
+        for (int j = 0; j < chw; j++) {
+          float scale = filter_data[i * chw + j];
+          float data = half_to_float(input_data[j]);
+          sum += scale * data;
+        }
+        float value = sum + bias;
+        if (std::isinf(value) || i > 321) {
+          std::cout << "i:" << i << " sum:" << sum << " bias:" << bias
+                    << std::endl;
+          // exit(-1);
+        }
+        if (i > 321) {
+          std::cout << "i:" << i << " sum:" << sum << " bias:" << bias
+                    << std::endl;
+          // exit(-1);
+        }
+
+        output_data[i] = float_to_half(value);
+        if (max < value) {
+          max = value;
+        }
+      }
+    }
+
+    output->flush();
+    output->scale()[0] = max / 127.0f;
+    output->scale()[1] = 127.0f / max;
+
+    output->saveToFile("cpu_compute", true);
+    // exit(-1);
+  }
+
+  void batch_to_w() {
+    ConvParam& convParam_ = convPE_.param();
+
+    int channel = param_.input->shape().channel();
+    param_.input->invalidate();
+
+    int remainder =
+        aligned_input_.shape().channel() - param_.input->shape().channel();
+
+    float max = 0;
+    for (int n = 0; n < param_.input->shape().num(); n++) {
+      memset(aligned_input_.data<float16>(),
+             0,
+             aligned_input_.shape().channel() * sizeof(float16));
+      memcpy(
+          aligned_input_.data<float16>() + n * aligned_input_.shape().channel(),
+          param_.input->data<float16>() + n * channel,
+          channel * sizeof(float16));
+      aligned_input_.copyScaleFrom(param_.input);
+      aligned_input_.flush();
+    }
+
+    convPE_.dispatch();
+  }
+
+  bool dispatch() {
+    // batch_to_w();
+    // return 1;
+    // cpu_compute1();
+    // return 1;
+
+    // int num = param_.filter->shape().channel();
+    // if (num == 2) {
+    //   cpu_compute();
+    //   return 1;
+    // } else {
+    // return convPE_.dispatch();
+    // }
+    ConvParam& convParam_ = convPE_.param();
+
+    if (param_.input->shape().channel() == 321 &&
+        param_.output->shape().channel() == 384) {
+      // conv_filter_.saveToFile("conv_filter", true);
+      // cpu_compute1();
+      // return 1;
+    }
+
+    int channel = param_.input->shape().channel();
+    param_.input->invalidate();
+
+    int remainder =
+        aligned_input_.shape().channel() - param_.input->shape().channel();
+
+    float max = 0;
+    for (int n = 0; n < param_.input->shape().num(); n++) {
+      memset(aligned_input_.data<float16>(),
+             0,
+             aligned_input_.shape().channel() * sizeof(float16));
+      memcpy(aligned_input_.data<float16>(),
+             param_.input->data<float16>() + n * channel,
+             channel * sizeof(float16));
+      aligned_input_.copyScaleFrom(param_.input);
+      aligned_input_.flush();
+
+      if (param_.input->shape().channel() == 321 &&
+          param_.output->shape().channel() == 384) {
+        // aligned_input_.saveToFile("aligned_input_", true);
+        // convParam_.filter->saveToFile("conv_filter", true);
+      }
+
+      convPE_.dispatch();
+      aligned_output_.invalidate();
+      if (param_.input->shape().num() == 230) {
+        // aligned_output_.saveToFile("ao", true);
+      }
+      //
+      float16* src = aligned_output_.data<float16>();
+      float16* dst =
+          param_.output->data<float16>() + n * param_.output->shape().channel();
+      memcpy(dst, src, param_.output->shape().channel() * sizeof(float16));
+      if (aligned_output_.scale()[0] > max) {
+        max = aligned_output_.scale()[0];
+      }
+    }
+    param_.output->flush();
+    param_.output->scale()[0] = max / 127.0f;
+    param_.output->scale()[1] = 127.0f / max;
+    // param_.output->saveToFile("out", true);
+
+    // exit(-1);
+    // cpu_compute();
+
+    // ConvParam& convParam_ = convPE_.param();
+    // convParam_.scale()->saveToFile("scale", true);
+    return true;
+  }
 
   FullyConnectedParam& param() { return param_; }
 
  private:
   FullyConnectedParam param_;
+  Tensor aligned_input_;
+  Tensor aligned_output_;
   ConvPE convPE_;
-  Tensor tempOut_;
-  int num_ = 1;
+  Tensor conv_filter_;
 };
 }  // namespace zynqmp
 }  // namespace paddle
diff --git a/lite/backends/fpga/KD/pes/input_pe.hpp b/lite/backends/fpga/KD/pes/input_pe.hpp
index 380c85e17e7de63486bbc93bb62ae8728286dd7a..abae094d97e66f1ecff78f1e41199893456e6722 100755
--- a/lite/backends/fpga/KD/pes/input_pe.hpp
+++ b/lite/backends/fpga/KD/pes/input_pe.hpp
@@ -29,19 +29,28 @@ class InputPE : public PE {
   }
 
   bool dispatch() {
+    std::cout << "input_dispatch()\n";
     Tensor* input = param_.input;
     Tensor* output = param_.output;
 
     Tensor* src = input;
+    // std::cout << "input:" << input << std::endl;
     input->flush();
+    // std::cout << "input_flush()\n";
     Tensor half_tensor;
     if (input->dataType() == DataType::FP32) {
+      // std::cout << "2()\n";
       half_tensor.mutableData<void*>(DataType::FP16, input->shape());
+      // std::cout << "3()\n";
       half_tensor.copyFrom(input);
+      // std::cout << "4()\n";
       src = &half_tensor;
     }
+    // std::cout << "5()\n";
     output->mutableData<void>();
+    // std::cout << "6()\n";
     src->alignImage(output, true);
+    // std::cout << "7()\n";
     return true;
   }
 
diff --git a/lite/backends/fpga/KD/pes/output_pe.hpp b/lite/backends/fpga/KD/pes/output_pe.hpp
index 2944691693b135a2d2df7b91ecbe0ef249b015d8..53da0c5be768d05d1ca58abf26a4a4ab0425cc9f 100755
--- a/lite/backends/fpga/KD/pes/output_pe.hpp
+++ b/lite/backends/fpga/KD/pes/output_pe.hpp
@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include "lite/backends/fpga/KD/llapi/zynqmp_api.h"
 #include "lite/backends/fpga/KD/pe.hpp"
 #include "lite/backends/fpga/KD/pe_params.hpp"
 
@@ -52,6 +53,10 @@ class OutputPE : public PE {
     memcpy(DLEngine::get_instance().out_data,
            output->data<void>(),
            output->shape().numel() * sizeof(float));
+
+    // auto max = fpga_get_memory_size_max();
+    // std::cout << "===== Max: ===== :: " << max << std::endl;
+
     return true;
   }
 
diff --git a/lite/backends/fpga/KD/tensor.hpp b/lite/backends/fpga/KD/tensor.hpp
index f247741a02758a3eb0cfa7f6c653d21e4263601d..e906ced5ff31538a1078f5013364502cb2b66c57 100755
--- a/lite/backends/fpga/KD/tensor.hpp
+++ b/lite/backends/fpga/KD/tensor.hpp
@@ -103,12 +103,18 @@ class Tensor {
     return reinterpret_cast<Dtype*>(ptr);
   }
 
+  void releaseData() {
+    released = true;
+    placeHolder_.reset();
+  }
+
   template <typename Dtype>
   Dtype* mutableData(DataType dataType, const Shape& shape) {
-    if (this->shape_ != nullptr) {
-      delete shape_;
-    }
-    this->shape_ = new Shape(shape);
+    // std::cout << "enter \n";
+    // std::cout << "before new shape\n";
+    // this->shape_ = new Shape(shape);
+    this->shape_.reset(new Shape(shape));
+    // std::cout << "new shape \n";
     this->dataType_ = dataType;
     return mutableData<Dtype>();
   }
@@ -117,11 +123,14 @@ class Tensor {
   Dtype* mutableData() {
     size_t memorySize =
         shape_->memorySize(CellSize(dataType_)) * mem_scale_factor_;
+    // std::cout << "mem_size:" << memorySize << std::endl;
     if (placeHolder_ != nullptr) {
+      // std::cout << "placeHolder_ not null"<< std::endl;
       if (memorySize > placeHolder_->memorySize()) {
         placeHolder_.reset(new PlaceHolder(memorySize));
       }
     } else {
+      // std::cout << "placeHolder_ null"<< std::endl;
       placeHolder_.reset(new PlaceHolder(memorySize));
     }
     return data<Dtype>();
@@ -138,7 +147,7 @@ class Tensor {
 
   DataType dataType() { return this->dataType_; }
 
-  Shape& shape() { return *shape_; }
+  Shape& shape() { return *(shape_.get()); }
 
   bool aligned() { return this->aligned_; }
 
@@ -247,15 +256,17 @@ class Tensor {
   void shareDataWith(Tensor* src) { shareDataWith(src, src->shape()); }
 
   void shareDataWith(Tensor* src, const Shape& shape, int offset = 0) {
-    if (shape_ != nullptr) {
-      delete shape_;
-    }
+    // if (shape_ != nullptr) {
+    //   delete shape_;
+    // }
+
     this->placeHolder_ = src->placeHolder_;
     this->dataType_ = src->dataType_;
     this->aligned_ = src->aligned_;
     this->dateLocation_ = src->dateLocation_;
     this->offset = offset;
-    shape_ = new Shape(const_cast<Shape&>(shape));
+    // shape_ = new Shape(const_cast<Shape&>(shape));
+    shape_.reset(new Shape(shape));
   }
 
   void copyFrom(Tensor* src) {
@@ -300,6 +311,14 @@ class Tensor {
   }
 
   void flush() {
+    // std::cout << "released:" << released << std::endl;
+    // std::cout << "placeHolder_" << placeHolder_.get() << std::endl;
+
+    if (released) {
+      // std::cout << "flush::" << this << std::endl;
+      return;
+    }
+
     size_t memorySize =
         shape_->memorySize(CellSize(dataType_)) * mem_scale_factor_;
     fpga_flush(placeHolder_->data(), memorySize);
@@ -463,18 +482,13 @@ class Tensor {
     return os;
   }
 
-  ~Tensor() {
-    if (shape_ != nullptr) {
-      delete shape_;
-      shape_ = nullptr;
-    }
-  }
-
  private:
+  bool released = false;
   int offset = 0;
   float mem_scale_factor_ = 1.0f;
   std::shared_ptr<PlaceHolder> placeHolder_;
-  Shape* shape_ = nullptr;
+  std::shared_ptr<Shape> shape_;
+  // Shape* shape_ = nullptr;
   DataType dataType_ = FP32;
   bool aligned_ = false;
   DataSyncStatus synchedStatus_ = Synched;
diff --git a/lite/backends/fpga/lite_tensor.cc b/lite/backends/fpga/lite_tensor.cc
index 7f1e8d3e17f97315e77532b77bbcfcc8331edd4f..6ec9f6866ab859460ebfa56a0e13800d45ba5d52 100755
--- a/lite/backends/fpga/lite_tensor.cc
+++ b/lite/backends/fpga/lite_tensor.cc
@@ -69,7 +69,7 @@ std::string DDimLite::repr() const {
 }
 
 void TensorLite::ShareDataWith(const TensorLite &other) {
-  buffer_ = other.buffer_;
+  buffer_ = other.buffer_;  // TODO(chonwhite) delete buffer;
   dims_ = other.dims_;
   zynq_tensor_ = other.zynq_tensor_;
   target_ = other.target_;
@@ -79,10 +79,10 @@ void TensorLite::ShareDataWith(const TensorLite &other) {
 }
 
 void *TensorLite::mutable_data(size_t memory_size) {
-  memory_size_ = memory_size;
+  memory_size_ = memory_size;  // TODO(chonwhite) delete buffer;
   buffer_->ResetLazy(target_, memory_size_);
   // throw -1;
-  std::cout << memory_size << std::endl;
+  // std::cout << memory_size << std::endl;
   return buffer_->data();
 }
 
@@ -92,16 +92,34 @@ void *TensorLite::mutable_data(TargetType target, size_t memory_size) {
 }
 
 void TensorLite::CopyDataFrom(const TensorLite &other) {
+  // std::cout << "other11:: "<< &other << std::endl;
   dims_ = other.dims_;
   target_ = other.target_;
   lod_ = other.lod_;
-  auto dt = zynq_tensor_->dataType();
+  // std::cout << "before dataType\n";
 
-  auto shape = other.zynq_tensor_->shape();
+  if (zynq_tensor_.get() == nullptr) {
+    zynq_tensor_.reset(new zynqmp::Tensor());
+  }
 
+  auto dt = zynq_tensor_->dataType();
+  // std::cout << "after dataType\n";
+
+  // std::cout << "before resize\n";
   Resize(other.dims());
+  auto shape = other.zynq_tensor_->shape();
+  // std::cout << "after resize\n";
   zynq_tensor_->mutableData<void>(zynq_tensor_->dataType(), shape);
-  this->ZynqTensor()->copyFrom(other.ZynqTensor());
+  // std::cout << "after mutableData\n";
+  // std::cout << "ZynqTensor():" << this->ZynqTensor() << std::endl;
+  // std::cout << "other Tensor():" << other.ZynqTensor() << std::endl;
+
+  // this->ZynqTensor()->copyFrom(other.ZynqTensor());
+  memcpy(this->ZynqTensor()->data<void>(),
+         other.ZynqTensor()->data<void>(),
+         other.ZynqTensor()->shape().numel() * sizeof(float));
+  // memcpy()
+  // std::cout << "after copyFrom\n";
 }
 
 }  // namespace lite
diff --git a/lite/backends/fpga/lite_tensor.h b/lite/backends/fpga/lite_tensor.h
index 49aded3d7d7db6d293e13298d98c2f3b165f411f..698438e90b27848f19c70abc814d5ad2308764ab 100755
--- a/lite/backends/fpga/lite_tensor.h
+++ b/lite/backends/fpga/lite_tensor.h
@@ -81,6 +81,10 @@ class DDimLite {
     return !(a == b);
   }
 
+  ~DDimLite() {
+    // std::cout << "free DDimLite\n";
+  }
+
  private:
   std::vector<value_type> data_;
 };
@@ -109,7 +113,12 @@ class TensorLite {
     return zynq_tensor_->data<R>() + offset_;
   }
 
-  void Resize(const DDimLite &ddim) { dims_ = ddim; }
+  void Resize(const DDimLite &ddim) {
+    // std::cout << "Resize \n";
+    // std::cout << "ddim:" << & ddim << std::endl;
+    dims_ = ddim;
+    // std::cout << "after Reize \n";
+  }
   void Resize(const std::vector<int64_t> &x) { dims_ = DDimLite(x); }
 
   const DDimLite &dims() const { return dims_; }
@@ -142,7 +151,9 @@ class TensorLite {
   void *mutable_data(size_t memory_size);
   void *mutable_data(TargetType target, size_t memory_size);
 
-  const void *raw_data() const { return buffer_->data(); }
+  const void *raw_data() const {
+    return buffer_->data();
+  }  // TODO(chonwhite) delete buffer;
 
   size_t data_size() const { return this->dims().production(); }
 
@@ -150,7 +161,9 @@ class TensorLite {
 
   size_t offset() const { return offset_; }
 
-  bool IsInitialized() const { return buffer_->data(); }
+  bool IsInitialized() const {
+    return buffer_->data();
+  }  // TODO(chonwhite) delete buffer;
 
   // Other share data to this.
   void ShareDataWith(const TensorLite &other);
@@ -168,7 +181,7 @@ class TensorLite {
   // template <typename T>
   // TensorLite Slice(int64_t begin, int64_t end) const;
 
-  zynqmp::Tensor *ZynqTensor() const { return zynq_tensor_; }
+  zynqmp::Tensor *ZynqTensor() const { return zynq_tensor_.get(); }
 
   friend std::ostream &operator<<(std::ostream &os, const TensorLite &tensor) {
     os << "Tensor:" << '\n';
@@ -197,7 +210,8 @@ class TensorLite {
   size_t memory_size_{};
   size_t offset_{0};
 
-  zynqmp::Tensor *zynq_tensor_ = new zynqmp::Tensor();
+  // zynqmp::Tensor *zynq_tensor_ = new zynqmp::Tensor();
+  std::shared_ptr<zynqmp::Tensor> zynq_tensor_;
 
   template <typename T>
   void mutable_data_internal();
@@ -206,6 +220,7 @@ class TensorLite {
 template <typename T, typename R>
 R *TensorLite::mutable_data() {
   std::vector<int> v;
+  // std::cout << "mutable_data \n";
   for (int i = 0; i < dims_.size(); i++) {
     v.push_back(dims_[i]);
   }
@@ -228,7 +243,7 @@ R *TensorLite::mutable_data() {
       break;
   }
   zynqmp::Shape input_shape(layout_type, v);
-
+  // std::cout << "input_shape \n";
   zynqmp::DataType data_type = zynqmp::FP32;
   if (typeid(T) == typeid(float)) {
     data_type = zynqmp::FP32;
@@ -236,6 +251,13 @@ R *TensorLite::mutable_data() {
   if (typeid(T) == typeid(zynqmp::float16)) {
     data_type = zynqmp::FP16;
   }
+  // std::cout << "mutableData \n";
+  // std::cout << "zynq_tensor_:" << zynq_tensor_.get() << std::endl;
+
+  if (zynq_tensor_.get() == nullptr) {
+    zynq_tensor_.reset(new zynqmp::Tensor());
+  }
+
   return zynq_tensor_->mutableData<R>(data_type, input_shape);
 }
 
@@ -276,6 +298,7 @@ TensorLite TensorLite::Slice(int64_t begin, int64_t end) const {
 
 template <typename T>
 void TensorLite::Slice(TensorLite &dst, int64_t begin, int64_t end) const {
+  // TODO(chonwhite) delete this function;
   CHECK_GE(begin, 0);
   CHECK_LE(end, dims_[0]);
   CHECK_LT(begin, end);
diff --git a/lite/kernels/fpga/feed_compute.cc b/lite/kernels/fpga/feed_compute.cc
index 79329e99a3e5e812dca487c17452f3f5d1e96449..9ca2424bc2f8a748c348cac4aafd219e538c7a17 100755
--- a/lite/kernels/fpga/feed_compute.cc
+++ b/lite/kernels/fpga/feed_compute.cc
@@ -40,8 +40,8 @@ void FeedCompute::PrepareForRun() {
 void FeedCompute::Run() {
   auto& param = this->Param<param_t>();
   Tensor& x = param.feed_list->at(param.col);
+  pe_.param().input = x.ZynqTensor();
   pe_.dispatch();
-
   auto out_lod = param.out->mutable_lod();
   *out_lod = x.lod();
 
diff --git a/lite/kernels/fpga/mul_compute.cc b/lite/kernels/fpga/mul_compute.cc
index c27600d9f773ff0aae04a2ee519905bc0e58785c..659c8dfb653c1707105a7337493ee4f3b3357b76 100755
--- a/lite/kernels/fpga/mul_compute.cc
+++ b/lite/kernels/fpga/mul_compute.cc
@@ -80,7 +80,8 @@ void mul(MulCompute* k) {
 }
 
 void MulCompute::Run() {
-  pe_.dispatch();
+  // pe_.dispatch();
+  mul(this);
 #ifdef FPGA_PRINT_TENSOR
   zynqmp::FullyConnectedParam& fc_param = pe_.param();
   Debugger::get_instance().registerOutput("mul", fc_param.output);