Merge branch 'ocr_ctc' of https://github.com/hjchen2/paddle-mobile into ocr_ctc

CMakeLists.txt

@@ -10,6 +10,7 @@ option(LOG_PROFILE "log profile" OFF)
 option(CPU "armv7 with neon" ON)
 option(GPU_MALI "mali gpu" OFF)
 option(GPU_CL "opencl gpu" OFF)
+
 option(FPGA "fpga" OFF)
 if(FPGA)
     option(FPGAV1 "fpga v1" ON)
@@ -144,7 +145,7 @@ if(FPGA)
         endforeach()
         file(GLOB_RECURSE _tmp_list src/operators/kernel/fpga/V2/*.h src/fpga/V2/*.h)
         foreach(f ${_tmp_list})
-            list(REMOVE_ITEM PADDLE_MOBILE_CC ${f})
+            list(REMOVE_ITEM PADDLE_MOBILE_H ${f})
         endforeach()
     endif()
     if(FPGAV2)
@@ -156,7 +157,7 @@ if(FPGA)
         endforeach()
         file(GLOB_RECURSE _tmp_list src/operators/kernel/fpga/V1/*.h src/fpga/V1/*.h)
         foreach(f ${_tmp_list})
-            list(REMOVE_ITEM PADDLE_MOBILE_CC ${f})
+            list(REMOVE_ITEM PADDLE_MOBILE_H ${f})
         endforeach()
     endif()
 

README.md

@@ -7,11 +7,21 @@
 <!--[![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle-Mobile.svg)](https://github.com/PaddlePaddle/Paddle-Mobile/releases)
 [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)-->
 
+Welcome to Paddle-Mobile GitHub project。Paddle-Mobile is a project of PaddlePaddle as well as a deep learning framework for embedded platforms.
 
 欢迎来到 Paddle-Mobile GitHub 项目。Paddle-Mobile是PaddlePaddle组织下的项目，是一个致力于嵌入式平台的深度学习的框架。
 
 ## Features
 
+- high performance in support of ARM CPU 
+- support Mali GPU
+- support Andreno GPU
+- support the realization of GPU Metal on Apple devices
+- support implementation on ZU5、ZU9 and other FPGA-based development boards
+- support implementation on Raspberry Pi and other arm-linux development boards
+
+## Features
+
 - 高性能支持ARM CPU 
 - 支持Mali GPU
 - 支持Andreno GPU
@@ -19,6 +29,7 @@
 - 支持ZU5、ZU9等FPGA开发板
 - 支持树莓派等arm-linux开发板
 
+
 ## Demo
 - [ANDROID](https://github.com/xiebaiyuan/paddle-mobile-demo)
 
@@ -26,6 +37,27 @@
 
 [https://github.com/PaddlePaddle/paddle-mobile/tree/develop/demo](https://github.com/PaddlePaddle/paddle-mobile/tree/develop/demo)
 
+## Documentation
+
+### Documentation of design
+
+If you want to know more details about the documentation of paddle-mobile design, please refer to the link as follows. There are many previous designs and discussion: [issue](https://github.com/PaddlePaddle/paddle-mobile/issues).
+
+[link of documentation of design](https://github.com/PaddlePaddle/paddle-mobile/blob/develop/doc/design_doc.md)
+
+### Documentation of development
+
+Documentation of development is mainly about building, running and other tasks.As a developer,you can use it with the help of contributed documents.
+* [iOS](https://github.com/PaddlePaddle/paddle-mobile/blob/develop/doc/development_ios.md)
+* [Android_CPU](https://github.com/PaddlePaddle/paddle-mobile/blob/develop/doc/development_android.md)
+* [Android_GPU](https://github.com/PaddlePaddle/paddle-mobile/blob/develop/doc/development_android_GPU.md)
+* [FPGA](https://github.com/PaddlePaddle/paddle-mobile/blob/develop/doc/development_fpga.md)
+* [ARM_LINUX](https://github.com/PaddlePaddle/paddle-mobile/blob/develop/doc/development_arm_linux.md)
+
+### How to contribute your documents
+- [tutorial link to contribute documents](https://github.com/PaddlePaddle/paddle-mobile/blob/develop/CONTRIBUTING.md)
+- Main procedure of contributing code is covered in the document above.If you have other problems during the procedure,please send them as [issue](https://github.com/PaddlePaddle/paddle-mobile/issues). We will deal with it as quickly as possible.
+
 ## 文档
 
 ### 设计文档
@@ -46,6 +78,24 @@
 - [贡献文档链接](https://github.com/PaddlePaddle/paddle-mobile/blob/develop/CONTRIBUTING.md)
 - 上面文档中涵盖了主要的贡献代码流程，如果在实践中您还遇到了其他问题，可以发[issue](https://github.com/PaddlePaddle/paddle-mobile/issues)。我们看到后会尽快处理。
 
+## Acquision of Models
+At present Paddle-Mobile only supports Paddle fluid training model. Models wiil be operated regularly after transformation if you have various models.
+### 1. Use Paddle Fluid directly to train
+It is the most reliable method to be recommanded
+### 2. Transform Caffe to Paddle Fluid model
+[https://github.com/PaddlePaddle/models/tree/develop/fluid/image_classification/caffe2fluid](https://github.com/PaddlePaddle/models/tree/develop/fluid/image_classification/caffe2fluid)
+### 3. ONNX
+ONNX is expanded as Open Neural Network Exchange. The project is aimed to make a full communication and usage among diffrent nerual network development frameworks.
+
+Except for directly using fluid models trained by PaddlePaddle,you can also get certain Paddle fluid models through onnx transformation.
+
+At present，work in support of onnx is also under operation in Baidu. Related tranformation project can be referred to here：
+[https://github.com/PaddlePaddle/paddle-onnx](https://github.com/PaddlePaddle/paddle-onnx)
+
+### 4. Download parts of testing models and testing pictures
+[http://mms-graph.bj.bcebos.com/paddle-mobile%2FmodelsAndImages.zip](http://mms-graph.bj.bcebos.com/paddle-mobile%2FmodelsAndImages.zip)
+
+
 
 ## 模型获得
 目前Paddle-Mobile仅支持Paddle fluid训练的模型。如果你手中的模型是不同种类的模型，需要进行模型转换才可以运行。
@@ -64,6 +114,22 @@ ONNX全称为“Open Neural Network Exchange”，即“开放的神经网络切
 ### 4. 部分测试模型和测试图片下载
 [http://mms-graph.bj.bcebos.com/paddle-mobile%2FmodelsAndImages.zip](http://mms-graph.bj.bcebos.com/paddle-mobile%2FmodelsAndImages.zip)
 
+<!--## Online output of simple search
+
+Gif as following is the application output of online main part detection of simple search app
+![ezgif-1-050a733dfb](http://otkwwi4x8.bkt.clouddn.com/2018-07-05-ezgif-1-050a733dfb.gif)-->
+
+## Ask Question
+
+Welcome to put forward or tackle with our problems,You can post your question in our issue modular on github. [Github Issues](https://github.com/PaddlePaddle/paddle-mobile/issues).
+
+## Copyright and License
+Paddle-Mobile provide relatively unstricted Apache-2.0 Open source agreement [Apache-2.0 license](LICENSE).
+
+
+## Old version Mobile-Deep-Learning
+Original MDL(Mobile-Deep-Learning) project has been transferred to [Mobile-Deep-Learning](https://github.com/allonli/mobile-deep-learning) 
+
 <!--## 简单搜索线上效果
 
 如下gif是简单搜索app的线上主体检测应用效果

src/fpga/V1/api.cpp

@@ -24,8 +24,6 @@ namespace fpga {
 #define USE_RELU 1
 #define USE_BIAS 2
 
-int get_align_image_cw(int cw) { return align_to_x(cw, IMAGE_ALIGNMENT); }
-
 void format_image(framework::Tensor *image_tensor) {
   auto dims = image_tensor->dims();
   auto channel = dims[1], height = dims[2], width = dims[3];
@@ -96,10 +94,6 @@ int get_aligned_filter_element_num(int chw) {
   return align_to_x(chw, FILTER_ELEMENT_ALIGNMENT);
 }
 
-int get_aligned_filter_num(int num) {
-  return align_to_x(num, FILTER_NUM_ALIGNMENT);
-}
-
 void format_filter(framework::Tensor *filter_tensor, float max_value,
                    int group_num) {
   filter_tensor->scale[0] = float(max_value / 127.0);  // NOLINT
@@ -177,46 +171,37 @@ void format_concat_output(framework::Tensor *out, int height, int width,
 
 void expand_conv_arg(ConvArgs *arg) {
   ConvArgs args = *arg;
-  uint64_t filterlen = (uint64_t)args.kernel.width *
-                       (uint64_t)args.kernel.height *
-                       (uint64_t)args.image.channels;
-  filterlen = align_to_x(filterlen, FILTER_ELEMENT_ALIGNMENT);
-  filterlen *= align_to_x((uint64_t)args.filter_num, FILTER_NUM_ALIGNMENT);
-  uint64_t fpga_bias_scale_len =
+
+  auto fpga_bias_scale_len =
       align_to_x(args.filter_num / args.group_num, 8) * args.group_num;
 
-  uint64_t output_height =
+  auto output_height =
       (args.image.height + args.image.pad_height * 2 - args.kernel.height) /
           args.kernel.stride_h +
       1;
-  uint64_t output_width =
+  auto output_width =
       (args.image.width + args.image.pad_width * 2 - args.kernel.width) /
           args.kernel.stride_w +
       1;
-  uint64_t output_size =
-      output_height * output_width * (uint64_t)args.filter_num;
-
-  auto filter_per_group = (uint64_t)(args.filter_num / args.group_num);
-  auto channel_per_group = (uint64_t)(args.image.channels / args.group_num);
-
-  uint64_t image_row_count = ((uint64_t)args.image.width) *
-                             ((uint64_t)args.image.channels);  // without align
-  uint64_t image_amount_per_row = align_to_x(image_row_count, IMAGE_ALIGNMENT);
-  uint64_t image_one_pad_per_row =
-      align_to_x(image_row_count, IMAGE_ALIGNMENT) +
-      ((uint64_t)args.image.pad_width) * ((uint64_t)args.image.channels);
-  uint64_t filter_amount_all =
-      align_to_x(((uint64_t)args.kernel.height) *
-                     ((uint64_t)args.kernel.width) * channel_per_group,
+
+  auto filter_per_group = args.filter_num / args.group_num;
+  auto channel_per_group = args.image.channels / args.group_num;
+
+  auto image_row_count = args.image.width * args.image.channels;
+  auto image_amount_per_row = align_to_x(image_row_count, IMAGE_ALIGNMENT);
+  auto image_one_pad_per_row = align_to_x(image_row_count, IMAGE_ALIGNMENT) +
+                               args.image.pad_width * args.image.channels;
+  auto filter_amount_all =
+      align_to_x(args.kernel.height * args.kernel.width * channel_per_group,
                  FILTER_ELEMENT_ALIGNMENT);
 
-  uint64_t output_amount_per_row =
-      align_to_x(output_width * ((uint64_t)args.filter_num), IMAGE_ALIGNMENT);
+  auto output_amount_per_row =
+      align_to_x(output_width * args.filter_num, IMAGE_ALIGNMENT);
 
   // find the opt partition strategy
   uint64_t res_win;
   uint64_t res_fit = 0;
-  for (res_win = 1; res_win <= output_width; res_win = res_win + 1) {
+  for (res_win = 1; res_win <= output_width; res_win++) {
     if ((align_to_x(
              (args.image.channels *
               (args.kernel.width + (res_win - 1) * args.kernel.stride_w)),
@@ -238,48 +223,48 @@ void expand_conv_arg(ConvArgs *arg) {
   }
   res_fit = res_win;
 
-  uint64_t block_num = (output_width + res_fit - 1) / res_fit;
-  uint64_t block_len = res_fit;
-  uint64_t block_last = output_width - res_fit * (block_num - 1);
+  auto block_num = (output_width + res_fit - 1) / res_fit;
+  auto block_len = res_fit;
+  auto block_last = output_width - res_fit * (block_num - 1);
 
-  uint64_t res_amount_per_row = output_width * args.filter_num;
-  uint64_t res_amount_per_row_pad = output_amount_per_row - res_amount_per_row;
+  auto res_amount_per_row = output_width * args.filter_num;
+  auto res_amount_per_row_pad = output_amount_per_row - res_amount_per_row;
 
-  uint64_t image_block_amount_per_row =
-      args.kernel.stride_w * (res_fit)*args.image.channels;
-  uint64_t filter_pad_width_mul_channel =
+  auto image_block_amount_per_row =
+      args.kernel.stride_w * res_fit * args.image.channels;
+  auto filter_pad_width_mul_channel =
       args.image.pad_width * args.image.channels;
-  uint64_t image_amount_per_row_multi_win_first =
+  auto image_amount_per_row_multi_win_first =
       image_amount_per_row * (4 * args.kernel.stride_h - args.image.pad_height);
-  uint64_t image_amount_per_row_multi_win =
+  auto image_amount_per_row_multi_win =
       image_amount_per_row * (4 * args.kernel.stride_h);
 
-  uint64_t image_block_num = block_num;
-  uint64_t image_block_len =
+  auto image_block_num = block_num;
+  auto image_block_len =
       align_to_x((args.image.channels *
                   (args.kernel.width + (block_len - 1) * args.kernel.stride_w)),
                  IMAGE_ALIGNMENT) /
           16 +
       1;
-  uint64_t image_block_len_last =
+  auto image_block_len_last =
       align_to_x(
           (args.image.channels *
            (args.kernel.width + (block_last - 1) * args.kernel.stride_w)),
           IMAGE_ALIGNMENT) /
           16 +
       1;
-  uint64_t image_win_cnt = block_len;
-  uint64_t image_win_cnt_last = block_last;
-  uint64_t res_row_data_align4_pad = res_amount_per_row_pad / 8;
-  uint64_t prog_full_cnt = 2048 / (filter_amount_all / 16 * 2) - 1;
+  auto image_win_cnt = block_len;
+  auto image_win_cnt_last = block_last;
+  auto res_row_data_align4_pad = res_amount_per_row_pad / 8;
+  auto prog_full_cnt = 2048 / (filter_amount_all / 16 * 2) - 1;
   if (prog_full_cnt == 1023) {
     prog_full_cnt--;
   }
-  uint64_t post_prog_full_cnt =
+  auto post_prog_full_cnt =
       (512 / (align_to_x(args.filter_num, 4) / 4 * 2) > 2)
           ? (512 / (align_to_x(args.filter_num, 4) / 4 * 2) - 2)
           : 0;
-  uint64_t cmd = 0UL | (args.relu_enabled ? USE_RELU : 0) | USE_BIAS;
+  auto cmd = 0UL | (args.relu_enabled ? USE_RELU : 0) | USE_BIAS;
 
   (*arg).driver.image_address_phy = vaddr_to_paddr(args.image.address);
   (*arg).driver.sb_address_phy = vaddr_to_paddr(args.sb_address);
@@ -449,7 +434,6 @@ void fill_deconv_arg(struct DeconvArgs *arg, framework::Tensor *input,
   arg->sub_conv_num = (uint32_t)stride_h;
   arg->filter_num = (uint32_t)filter->dims()[0];
   int sub_conv_num = arg->sub_conv_num;
-  int sub_stride = 1;
   int sub_pad = deconv_filter::deconv_calc_sub_pad((int)filter->dims()[3],
                                                    padding_w, stride_w);
   int sub_filter_width = deconv_filter::deconv_get_sub_filter_axis(
@@ -466,16 +450,12 @@ void fill_deconv_arg(struct DeconvArgs *arg, framework::Tensor *input,
       stride_w, (int)filter->dims()[3], padding_w);
   arg->conv_args = (ConvArgs *)fpga_malloc(sub_conv_num * sizeof(ConvArgs));
 
-  int sub_channels = (int)input->dims()[1];
-  int omit_size = arg->omit_size;
-  int real_out_width = sub_output_width * sub_conv_num - 2 * omit_size;
-  int real_out_height = sub_output_height * sub_conv_num - 2 * omit_size;
+  auto sub_channels = (int)input->dims()[1];
   int sub_filter_num = sub_conv_num * (arg->filter_num);
 
   int conv_output_size =
       (align_to_x(sub_output_width * sub_filter_num, IMAGE_ALIGNMENT)) *
       sub_output_height;
-  int ouput_size = conv_output_size * sub_conv_num;
 
   int align_sub_filter_num = align_to_x(sub_filter_num, FILTER_NUM_ALIGNMENT);
   int align_sub_filter_count =
@@ -485,7 +465,7 @@ void fill_deconv_arg(struct DeconvArgs *arg, framework::Tensor *input,
       align_sub_filter_count * align_sub_filter_num;
 
   for (int i = 0; i < sub_conv_num; ++i) {
-    arg->conv_args[i].filter_num = (arg->sub_conv_num) * (arg->filter_num);
+    arg->conv_args[i].filter_num = arg->sub_conv_num * arg->filter_num;
     arg->conv_args[i].group_num = (uint32_t)group_num;
 
     arg->conv_args[i].filter_scale_address = filter->scale;
@@ -496,7 +476,6 @@ void fill_deconv_arg(struct DeconvArgs *arg, framework::Tensor *input,
     arg->conv_args[i].kernel.stride_w = 1;
     arg->conv_args[i].kernel.stride_h = 1;
 
-    // DeconvParam.conv_args[i].image.address = (void*)ptr_image;
     arg->conv_args[i].image.scale_address = input->scale;
     arg->conv_args[i].image.channels = (uint32_t)sub_channels;
     arg->conv_args[i].image.width = (uint32_t)input->dims()[3];
@@ -504,30 +483,31 @@ void fill_deconv_arg(struct DeconvArgs *arg, framework::Tensor *input,
     arg->conv_args[i].image.pad_width = (uint32_t)sub_pad;
     arg->conv_args[i].image.pad_height = (uint32_t)sub_pad;
     arg->conv_args[i].image.address = input_ptr;
-    arg->conv_args[i].sb_address = (void *)bs_ptr;
+    arg->conv_args[i].sb_address = bs_ptr;
 
     auto filter_sub_space =
         (char *)fpga_malloc(align_conv_sub_filter_count * sizeof(char));
     fpga_copy(filter_sub_space,
               (char *)filter_ptr + i * align_conv_sub_filter_count,
               (size_t)align_conv_sub_filter_count);
-    arg->conv_args[i].filter_address = (void *)(filter_sub_space);
+    arg->conv_args[i].filter_address = filter_sub_space;
     fpga_flush(filter_sub_space, (size_t)align_conv_sub_filter_count);
 
     if (sub_conv_num == 1) {
       arg->conv_args[i].output.address = out_ptr;
       arg->conv_args[i].output.scale_address = out->scale;
     } else {
-      auto ptr_output = (half *)fpga_malloc(conv_output_size * sizeof(half));
-      arg->conv_args[i].output.address = (void *)((half *)ptr_output);
+      auto ptr_output = fpga_malloc(conv_output_size * sizeof(half));
+      arg->conv_args[i].output.address = ptr_output;
       auto ptr_output_scale = (float *)fpga_malloc(2 * sizeof(float));
       arg->conv_args[i].output.scale_address = ptr_output_scale;
     }
+    expand_conv_arg(&arg->conv_args[i]);
   }
 
   arg->output.address = out_ptr;
   arg->output.scale_address = out->scale;
-  // fpga_free(filter_ptr);
+  filter->reset_data_ptr(nullptr);
 }  // fill_deconv_arg
 
 }  // namespace fpga

src/fpga/V1/api.h

@@ -21,7 +21,6 @@ limitations under the License. */
 namespace paddle_mobile {
 namespace fpga {
 
-int get_align_image_cw(int cw);
 void format_image(framework::Tensor* image_tensor);
 void format_fp16_ofm(framework::Tensor* ofm_tensor);  // only allocate memory
 void format_fp32_ofm(framework::Tensor* ofm_tensor);
@@ -30,7 +29,6 @@ float filter_find_max(framework::Tensor* filter_tensor);
 int get_filter_num_per_div(framework::Tensor* filter_tensor, int group_num);
 int get_plit_num(framework::Tensor* filter_tensor);
 int get_aligned_filter_element_num(int chw);
-int get_aligned_filter_num(int num);
 void format_filter(framework::Tensor* filter_tensor, float max_value,
                    int group_num);
 void format_fc_filter(framework::Tensor* filter_tensor, float max_value);

src/fpga/V1/deconv_filter.cpp

@@ -40,10 +40,9 @@ inverse kernel weights of each channel for every filter
 void deconv_inverse_filter(float** data_in, int num, int channel, int width,
                            int height) {
   float* tmp = *data_in;
-  // float fix_range = 127;//  float scale = fix_range / max;
   int data_size = num * channel * width * height;
   int hw_len = height * width;
-  float* tmp_data = (float*)fpga_malloc(data_size * sizeof(float));
+  auto tmp_data = (float*)fpga_malloc(data_size * sizeof(float));
   for (int i = 0; i < num; ++i) {
     for (int j = 0; j < channel; ++j) {
       for (int k = 0; k < hw_len; ++k) {
@@ -52,7 +51,7 @@ void deconv_inverse_filter(float** data_in, int num, int channel, int width,
       }
     }
   }
-  *data_in = (float*)tmp_data;  //
+  *data_in = tmp_data;
   fpga_free(tmp);
 }
 
@@ -61,8 +60,7 @@ void deconv_inverse_filter(float** data_in, int num, int channel, int width,
 */
 int deconv_calc_sub_pad(int filter_axis, int pad, int stride) {
   if (stride == 0 || ((filter_axis - pad - 1) < 0)) {
-    // error
-    return 0;
+    PADDLE_MOBILE_ENFORCE(false, "Wrong deconv parameters");
   }
   return (filter_axis - pad - 1) / stride;
 }
@@ -79,11 +77,8 @@ int deconv_get_sub_out_axis(int image_axis, int sub_pad, int sub_filter_axis) {
    position. so the omit rows or columns is (stride - )
 */
 int deconv_get_omit(int stride, int filter_width, int pad) {
-  if (((filter_width - pad) <= 0)) {  // ((filter_width-pad) > stride) ||
-    // error
-    return 0;
-  }
-  int idx = 1;
+  PADDLE_MOBILE_ENFORCE(filter_width > pad, "Wrong deconv parameters");
+  int idx;
   bool flag = false;
   for (idx = 1; idx <= stride; ++idx) {
     int j = idx;
@@ -102,10 +97,6 @@ int deconv_get_omit(int stride, int filter_width, int pad) {
   return (stride - idx);
 }
 
-int deconv_get_sub_filter_num(int filter_num, int stride) {
-  return filter_num * stride;
-}
-
 void deconv_get_sub_filter(char** data_in, int height, int width,
                            int sub_conv_n, int kernel_num, int channel) {
   char* ptr_tmp = *data_in;
@@ -245,7 +236,6 @@ void deconv_format_filter(float** data_in, int num, int channel, int height,
   char* ptr_space = (char*)fpga_malloc(sub_conv_n * align_offset *
                                        sizeof(char));  // continuous space
   for (int i = 0; i < sub_conv_n; ++i) {
-    int offset = i * origin_offset;
     char* ptr_tmp = (ptr_ptr_data)[i];
 
     filter::align_element(&ptr_tmp, sub_num, sub_chw);

src/fpga/V1/deconv_filter.h

@@ -21,7 +21,6 @@ namespace deconv_filter {
 void deconv_inverse_filter(float** data_in, int num, int channel, int width,
                            int height);
 int deconv_calc_sub_pad(int filter_axis, int pad, int stride);
-int deconv_get_sub_filter_num(int filter_num, int stride);
 int deconv_get_sub_filter_axis(int filter_axis, int stride);
 int deconv_get_sub_out_axis(int image_axis, int sub_pad, int sub_filter_axis);
 int deconv_get_omit(int stride, int filter_width, int pad);

src/fpga/V1/pe.cpp

@@ -13,23 +13,15 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "fpga/common/pe.h"
-#include <unistd.h>
-#include <iomanip>
-#include <iostream>
-#include "common/types.h"
 #include "fpga/V1/filter.h"
 #include "fpga/V1/image.h"
 #include "fpga/common/config.h"
 #include "fpga/common/driver.h"
-using namespace std;
-using namespace paddle_mobile::fpga::driver;  // NOLINT
 
 namespace paddle_mobile {
 namespace fpga {
 
-#define IMAGE_ALIGN 16
-#define FILTER_ALIGN 16
-#define FILTER_NUM_ALIGN 32
+using namespace driver;  // NOLINT
 #define USE_RELU 1
 #define USE_BIAS 2
 
@@ -207,129 +199,6 @@ int ComputeBasicConv(const struct ConvArgs &args) {
 #ifdef PADDLE_MOBILE_ZU5
   int ret = 0;
   uint64_t output_scale = 0;
-  /*
-  uint64_t output_scale;
-  uint64_t image_scale;
-  uint64_t filter_scale;
-  uint64_t image_address_phy = 0;
-  uint64_t sb_address_phy = 0;
-  uint64_t filter_address_phy = 0;
-  uint64_t output_address_phy = 0;
-
-
-  fpga_copy(&image_scale, args.image.scale_address, 2 * sizeof(float));
-  fpga_copy(&filter_scale, args.filter_scale_address, 2 * sizeof(float));
-  uint64_t filterlen = (uint64_t)args.kernel.width *
-                       (uint64_t)args.kernel.height *
-                       (uint64_t)args.image.channels;
-  filterlen = align_to_x(filterlen, FILTER_ALIGN);
-  filterlen *= align_to_x((uint64_t)args.filter_num, FILTER_NUM_ALIGN);
-  uint64_t fpga_bias_scale_len =
-      align_to_x(args.filter_num / args.group_num, 8) * args.group_num;
-
-  uint64_t output_height =
-      (args.image.height + args.image.pad_height * 2 - args.kernel.height) /
-          args.kernel.stride_h +
-      1;
-  uint64_t output_width =
-      (args.image.width + args.image.pad_width * 2 - args.kernel.width) /
-          args.kernel.stride_w +
-      1;
-  uint64_t output_size =
-      output_height * output_width * (uint64_t)args.filter_num;
-
-  uint64_t filter_per_group = (uint64_t)(args.filter_num / args.group_num);
-  uint64_t channel_per_group = (uint64_t)(args.image.channels / args.group_num);
-
-  uint64_t image_row_count = ((uint64_t)args.image.width) *
-                             ((uint64_t)args.image.channels);  // without align
-  uint64_t image_amount_per_row = align_to_x(image_row_count, IMAGE_ALIGN);
-  uint64_t image_one_pad_per_row =
-      align_to_x(image_row_count, IMAGE_ALIGN) +
-      ((uint64_t)args.image.pad_width) * ((uint64_t)args.image.channels);
-  uint64_t filter_amount_all =
-      align_to_x(((uint64_t)args.kernel.height) *
-                     ((uint64_t)args.kernel.width) * channel_per_group,
-                 FILTER_ALIGN);
-
-  uint64_t output_amount_per_row =
-      align_to_x(output_width * ((uint64_t)args.filter_num), IMAGE_ALIGN);
-
-  // find the opt partition strategy
-  uint64_t res_win;
-  uint64_t res_fit = 0;
-  for (res_win = 1; res_win <= output_width; res_win = res_win + 1) {
-    if ((align_to_x(
-             (args.image.channels *
-              (args.kernel.width + (res_win - 1) * args.kernel.stride_w)),
-             IMAGE_ALIGN) /
-             16 +
-         1) *
-            args.kernel.height >
-        2048) {
-      break;
-    }
-  }
-
-  if (res_win != output_width) {
-    res_win -= 1;
-  }
-
-  if (((res_win % 2) != 0) && (res_win != 1)) {
-    res_win = res_win - 1;
-  }
-  res_fit = res_win;
-
-  uint64_t block_num = (output_width + res_fit - 1) / res_fit;
-  uint64_t block_len = res_fit;
-  uint64_t block_last = output_width - res_fit * (block_num - 1);
-
-  uint64_t res_amount_per_row = output_width * args.filter_num;
-  uint64_t res_amount_per_row_pad = output_amount_per_row - res_amount_per_row;
-
-  uint64_t image_block_amount_per_row =
-      args.kernel.stride_w * (res_fit)*args.image.channels;
-  uint64_t filter_pad_width_mul_channel =
-      args.image.pad_width * args.image.channels;
-  uint64_t image_amount_per_row_multi_win_first =
-      image_amount_per_row * (4 * args.kernel.stride_h - args.image.pad_height);
-  uint64_t image_amount_per_row_multi_win =
-      image_amount_per_row * (4 * args.kernel.stride_h);
-
-  uint64_t image_block_num = block_num;
-  uint64_t image_block_len =
-      align_to_x((args.image.channels *
-                  (args.kernel.width + (block_len - 1) * args.kernel.stride_w)),
-                 IMAGE_ALIGN) /
-          16 +
-      1;
-  uint64_t image_block_len_last =
-      align_to_x(
-          (args.image.channels *
-           (args.kernel.width + (block_last - 1) * args.kernel.stride_w)),
-          IMAGE_ALIGN) /
-          16 +
-      1;
-  uint64_t image_win_cnt = block_len;
-  uint64_t image_win_cnt_last = block_last;
-  uint64_t res_row_data_align4_pad = res_amount_per_row_pad / 8;
-  uint64_t prog_full_cnt = 2048 / (filter_amount_all / 16 * 2) - 1;
-  if (prog_full_cnt == 1023) {
-    prog_full_cnt--;
-  }
-  uint64_t post_prog_full_cnt =
-      (512 / (align_to_x(args.filter_num, 4) / 4 * 2) > 2)
-          ? (512 / (align_to_x(args.filter_num, 4) / 4 * 2) - 2)
-          : 0;
-
-  image_address_phy = vaddr_to_paddr(args.image.address);
-  sb_address_phy = vaddr_to_paddr(args.sb_address);
-  filter_address_phy = vaddr_to_paddr(args.filter_address);
-  output_address_phy = vaddr_to_paddr(args.output.address);
-
-  uint64_t cmd = 0UL | (args.relu_enabled ? USE_RELU : 0) | USE_BIAS;
-*/
-
   pthread_mutex_lock(&g_fpgainfo.pe_data->mutex);
   if (ERROR == g_fpgainfo.pe_data->pes[PE_IDX_CONV]->status) {
     ret = -EIO;
@@ -357,7 +226,6 @@ int ComputeBasicConv(const struct ConvArgs &args) {
   reg_writeq((uint64_t)args.image.channels, REG_CONV_CHANNEL_NUMBER);
   reg_writeq(*(uint64_t *)args.image.scale_address, REG_CONV_IMAGE_SCALE);
   reg_writeq(*(uint64_t *)args.filter_scale_address, REG_CONV_FILTER_SCALE);
-
   reg_writeq(args.driver.image_address_phy, REG_CONV_IMAGE_BASE_ADDR);
   reg_writeq(args.driver.filter_address_phy, REG_CONV_FILTER_BASE_ADDR);
   reg_writeq(args.driver.sb_address_phy, REG_CONV_SB_BASE_ADDR);
@@ -381,7 +249,6 @@ int ComputeBasicConv(const struct ConvArgs &args) {
   reg_writeq(args.driver.prog_full_cnt, 0xd08);
   reg_writeq(args.driver.post_prog_full_cnt, 0xd10);
   reg_writeq(args.driver.fpga_bias_scale_len / 4, 0xd20);
-
   reg_writeq(args.driver.cmd, REG_CONV_CMD);
 
   if (0 != fpga_regpoll(REG_INTERRUPT, INTERRUPT_CONV, PE_IRQ_TIMEOUT)) {
@@ -398,9 +265,9 @@ int ComputeBasicConv(const struct ConvArgs &args) {
 
   return ret;
 #endif
-
   return 0;
-}
+
+}  // ComputeBasicConv
 
 int ComputeFpgaPool(const struct PoolingArgs &args) {
 #ifdef FPGA_PRINT_MODE
@@ -422,19 +289,15 @@ int ComputeFpgaPool(const struct PoolingArgs &args) {
        << "   out_scale_address:" << args.output.scale_address;
 #endif
 #ifdef PADDLE_MOBILE_ZU5
-  DLOG << "Polling";
-  // return 0;
   uint64_t output_scale = 0;
   uint64_t timer_cnt = 0;
   int ret = 0;
   uint64_t cmd = 0;
-
   uint64_t image_physical_address = 0;
   uint64_t output_physical_address = 0;
 
-  image_physical_address = vaddr_to_paddr(args.image.address);
-  output_physical_address = vaddr_to_paddr(args.output.address);
-
+  image_physical_address = vaddr_to_paddr_driver(args.image.address);
+  output_physical_address = vaddr_to_paddr_driver(args.output.address);
   uint32_t output_height = (uint32_t)(
       (args.image.height + args.image.pad_height * 2 - args.kernel.height) /
           args.kernel.stride_h +
@@ -443,37 +306,35 @@ int ComputeFpgaPool(const struct PoolingArgs &args) {
       (args.image.width + args.image.pad_width * 2 - args.kernel.width) /
           args.kernel.stride_w +
       1);
-
-  uint64_t image_amount_per_row = align_to_x(
-      (uint64_t)args.image.width * (uint64_t)args.image.channels, IMAGE_ALIGN);
+  uint64_t image_amount_per_row =
+      align_to_x((uint64_t)args.image.width * (uint64_t)args.image.channels,
+                 IMAGE_ALIGNMENT);
   uint64_t image_one_pad_per_row =
       align_to_x((uint64_t)args.image.width * (uint64_t)args.image.channels,
-                 FILTER_ALIGN) +
+                 FILTER_ELEMENT_ALIGNMENT) +
       (uint64_t)args.image.pad_width * (uint64_t)args.image.channels;
   uint64_t image_two_pad_per_row = align_to_x(
       ((uint64_t)args.image.width + (uint64_t)args.image.pad_width * 2) *
           (uint64_t)args.image.channels,
-      IMAGE_ALIGN);
+      IMAGE_ALIGNMENT);
   uint64_t image_row_mul_pooling_hight =
       image_amount_per_row * (uint64_t)args.kernel.height;
   uint64_t image_row_mul_pad_hight =
       image_amount_per_row * (uint64_t)args.image.pad_height;
   uint64_t image_row_mul_step_hight =
       image_amount_per_row * (uint64_t)args.kernel.stride_h;
-  uint64_t result_amount_align_32 = align_to_x(
-      (uint64_t)output_width * (uint64_t)args.image.channels, FILTER_ALIGN);
+  uint64_t result_amount_align_32 =
+      align_to_x((uint64_t)output_width * (uint64_t)args.image.channels,
+                 FILTER_ELEMENT_ALIGNMENT);
   uint64_t result_amount_align_64 = align_to_x(
-      (uint64_t)output_width * (uint64_t)args.image.channels, IMAGE_ALIGN);
+      (uint64_t)output_width * (uint64_t)args.image.channels, IMAGE_ALIGNMENT);
   uint64_t image_calcu_height =
       (uint64_t)args.kernel.height +
       ((uint64_t)output_height - 1) * (uint64_t)args.kernel.stride_h;
-
   uint64_t image_pad_left = args.image.channels * args.image.pad_width;
   uint64_t image_skip_window = args.image.channels * args.kernel.stride_w;
-
   uint64_t image_padleft_skipwindow =
       (image_skip_window << 32) | image_pad_left;
-
   uint64_t mode_reciprocal = (uint64_t)0 | ((uint64_t)args.mode) << 16 |
                              (((uint64_t)args.kernel_reciprocal));
 
@@ -485,50 +346,36 @@ int ComputeFpgaPool(const struct PoolingArgs &args) {
     return ret;
   }
 
-  /*restart scale*/
   reg_writeq(output_scale, REG_SCALE_PARAMETER);
-
   reg_writeq(image_physical_address, REG_POOLING_IMAGE_BASE_ADDR);
   reg_writeq(output_physical_address, REG_POOLING_RESULT_BASE_ADDR);
-
   reg_writeq(
       ((uint64_t)args.image.height) | (((uint64_t)args.image.width) << 32),
       REG_POOLING_IMAGE_PIXEL);
   reg_writeq(
       ((uint64_t)args.kernel.height) | (((uint64_t)args.kernel.width) << 32),
       REG_POOLING_WINDOW_SIZE);
-
   reg_writeq(((uint64_t)output_height) | (((uint64_t)output_width) << 32),
              REG_POOLING_RESULT_PIXEL);
-
   reg_writeq(((uint64_t)args.image.pad_height) |
                  (((uint64_t)args.image.pad_width) << 32),
              REG_POOLING_PAD_PIXEL);
   reg_writeq(((uint64_t)args.kernel.stride_h) |
                  (((uint64_t)args.kernel.stride_w) << 32),
              REG_POOLING_STEP_PIXEL);
-
   reg_writeq((uint64_t)args.image.channels, REG_POOLING_CHANNEL_NUMBER);
-
   reg_writeq(image_amount_per_row, REG_POOLING_IMAGE_AMOUNT_PER_ROW);
   reg_writeq(image_one_pad_per_row, REG_POOLING_IMAGE_ONE_PAD_PER_ROW);
   reg_writeq(image_two_pad_per_row, REG_POOLING_IMAGE_TWO_PAD_PER_ROW);
-
   reg_writeq(image_row_mul_pooling_hight,
              REG_POOLING_IMAGE_ROW_MUL_WINDOW_HEIGHT);
   reg_writeq(image_row_mul_pad_hight, REG_POOLING_IMAGE_ROW_MUL_PAD_HEIGHT);
   reg_writeq(image_row_mul_step_hight, REG_POOLING_IMAGE_ROW_MUL_STEP_HEIGHT);
-
   reg_writeq(result_amount_align_32, REG_POOLING_RESULT_AMOUNT_ALIGN_32);
   reg_writeq(result_amount_align_64, REG_POOLING_RESULT_AMOUNT_ALIGN_64);
-
   reg_writeq(image_calcu_height, REG_POOLING_IMAGE_CALCU_HEIGHT);
-
   reg_writeq(image_padleft_skipwindow, REG_POOLING_IMAGE_PADLEFT_SKIPWINDOW);
   reg_writeq(mode_reciprocal, REG_POOLING_MODE_RECIPROCAL);
-
-  /*SDK刷Cache保证数据一致性*/
-
   reg_writeq(cmd, REG_POOLING_CMD);
 
   DLOG << "before reg poll";
@@ -549,7 +396,8 @@ int ComputeFpgaPool(const struct PoolingArgs &args) {
   return ret;
 #endif
   return 0;
-}
+
+}  // ComputeFpgaPool
 
 int ComputeFpgaEWAdd(const struct EWAddArgs &args) {
 #ifdef FPGA_PRINT_MODE
@@ -577,27 +425,6 @@ int ComputeFpgaEWAdd(const struct EWAddArgs &args) {
 #ifdef PADDLE_MOBILE_ZU5
   int ret = 0;
   uint64_t output_scale = 0;
-  /*uint64_t timer_cnt = 0;
-  uint64_t image0_address_phy = 0;
-  uint64_t image1_address_phy = 0;
-  uint64_t output_address_phy = 0;
-
-  uint64_t cmd = args.relu_enabled ? USE_RELU : 0;
-  uint64_t datalen = (uint64_t)args.image0.width *
-                     (uint64_t)args.image0.height *
-                     (uint64_t)args.image0.channels;
-  uint64_t coefficient = (uint64_t)args.const0 << 32 | (uint64_t)args.const1;
-  image0_address_phy = vaddr_to_paddr(args.image0.address);
-  image1_address_phy = vaddr_to_paddr(args.image1.address);
-  output_address_phy = vaddr_to_paddr(args.output.address);
-
-  uint64_t image_amount_per_row =
-  align_to_x((uint64_t)args.image0.width * (uint64_t)args.image0.channels,
-             IMAGE_ALIGN);
-  uint64_t image_image_pixel = ((uint64_t)args.image0.channels << 32) |
-                               ((uint64_t)args.image0.width << 16) |
-                               (uint64_t)args.image0.height;*/
-
   pthread_mutex_lock(&g_fpgainfo.pe_data->mutex);
   if (ERROR == g_fpgainfo.pe_data->pes[PE_IDX_EW]->status) {
     ret = -EIO;
@@ -631,7 +458,8 @@ int ComputeFpgaEWAdd(const struct EWAddArgs &args) {
   return ret;
 #endif
   return 0;
-}
+
+}  // ComputeFpgaEWAdd
 
 int PerformBypass(const struct BypassArgs &args) {
 #ifdef FPGA_PRINT_MODE
@@ -651,9 +479,6 @@ int PerformBypass(const struct BypassArgs &args) {
        << "   out_scale_address:" << args.output.scale_address;
 #endif
 #ifdef PADDLE_MOBILE_ZU5
-  DLOG << "Bypass";
-  // return 0;
-  struct fpga_pe *pe;
   uint64_t output_scale = 0;
   uint64_t timer_cnt = 0;
   uint64_t cmd = 0;
@@ -662,15 +487,12 @@ int PerformBypass(const struct BypassArgs &args) {
   uint64_t output_address_phy = 0;
   uint8_t data_cell_in = 0;
   uint8_t data_cell_out = 0;
-
   int ret = 0;
-
   datalen = (uint64_t)args.image.width * (uint64_t)args.image.height *
             (uint64_t)args.image.channels;
   datalen = align_to_x(datalen, 16);
-
-  input_address_phy = vaddr_to_paddr(args.image.address);
-  output_address_phy = vaddr_to_paddr(args.output.address);
+  input_address_phy = vaddr_to_paddr_driver(args.image.address);
+  output_address_phy = vaddr_to_paddr_driver(args.output.address);
   DLOG << "input_phy:" << input_address_phy;
   DLOG << "output_phy:" << output_address_phy;
 
@@ -733,36 +555,29 @@ int PerformBypass(const struct BypassArgs &args) {
     return ret;
   }
 
-  /*restart scale*/
   reg_writeq(output_scale, REG_SCALE_PARAMETER);
-
   reg_writeq(input_address_phy, REG_CONVERT_SRC_ADDR);
   reg_writeq(output_address_phy, REG_CONVERT_DST_ADDR);
   reg_writeq(datalen, REG_CONVERT_LENGTH);
-
-  /*SDK刷Cache保证数据一致性*/
   reg_writeq(cmd, REG_CONVERT_CMD);
 
-  DLOG << "before reg poll";
   if (0 != fpga_regpoll(REG_INTERRUPT, INTERRUPT_BYPASS, PE_IRQ_TIMEOUT)) {
     g_fpgainfo.pe_data->pes[PE_IDX_BYPASS]->status = ERROR;
     ret = -EIO;
     DLOG << "BYPASS Wait Irq Timeout!";
   }
-  DLOG << "after reg poll";
 
   output_scale = reg_readq(REG_SCALE_PARAMETER);
   output_scale = (output_scale << 32) | (output_scale >> 32);
   fpga_copy(args.output.scale_address, &output_scale, sizeof(float) * 2);
-
   //*(args.output.scale_address) = reg_readq(REG_SCALE_PARAMETER);
   //*(args.output.timer_cnt) = reg_readq(REG_TIMER_COUNTER);
   pthread_mutex_unlock(&g_fpgainfo.pe_data->mutex);
   return ret;
 #endif
-
   return 0;
-}
+
+}  // PerformBypass
 
 int ComputeFPGAConcat(const struct ConcatArgs &args) {
 #ifdef FPGA_PRINT_MODE
@@ -776,7 +591,7 @@ int ComputeFPGAConcat(const struct ConcatArgs &args) {
     DLOG << "   " << i << "th:        ";
     DLOG << "   channel_num:"
          << args.channel_num[i]
-         // << "   aligned_channel_num:" << args.aligned_channel_num[i]
+         //<< "   aligned_channel_num:" << args.aligned_channel_num[i]
          << "   image_address:" << args.images_in[i]
          << "   image_scale_address:" << args.scales_in[i];
   }
@@ -786,10 +601,15 @@ int ComputeFPGAConcat(const struct ConcatArgs &args) {
                        args.scale_out, args.image_num, args.channel_num,
                        args.height, args.width);
   return 0;
-}
-
-void deconv_post_process(half **data_in, int sub_conv_n, int num, int channel,
-                         int sub_height, int sub_width, int omit_size) {
+}  // ComputeFPGAConcat
+
+void deconv_post_process(const struct DeconvArgs &args) {
+  int sub_conv_n = args.sub_conv_num;
+  int sub_height = args.sub_output_height;
+  int sub_width = args.sub_output_width;
+  int omit_size = args.omit_size;
+  int channel = args.filter_num;
+  int num = 1;
   int origin_h = sub_height * sub_conv_n;
   int origin_w = sub_width * sub_conv_n;
   int align_origin_w = align_to_x(origin_w * channel, 16);
@@ -797,60 +617,68 @@ void deconv_post_process(half **data_in, int sub_conv_n, int num, int channel,
   int deconv_w = origin_w - 2 * omit_size;
   int deconv_row_len = deconv_w * channel;
   int align_deconv_row_len = align_to_x(deconv_row_len, 16);
-  half *ptr_tmp = *data_in;
-  half *ptr_deconv =
-      (half *)fpga_malloc(num * align_deconv_row_len * deconv_h * sizeof(half));
-  memset(ptr_deconv, 0, num * align_deconv_row_len * deconv_h * sizeof(half));
+
+  for (int idx = 0; idx < sub_conv_n; ++idx) {
+    fpga_invalidate(args.conv_args[idx].output.address,
+                    align_origin_w * origin_h * sizeof(int16_t));
+  }
+
+  auto ptr_deconv = (int16_t *)fpga_malloc(num * align_deconv_row_len *
+                                           deconv_h * sizeof(int16_t));
+  memset(ptr_deconv, 0,
+         num * align_deconv_row_len * deconv_h * sizeof(int16_t));
   int deconv_idx = 0;
   for (int nn = 0; nn < num; ++nn) {
     for (int hh = 0; hh < origin_h; ++hh) {
       int hx = (hh % sub_conv_n);
-      half *sub_t = ptr_tmp + hx * sub_height * align_origin_w;  // sub(hx,:);
-
+      auto sub_t =
+          (int16_t *)(args.conv_args[sub_conv_n - hx - 1].output.address);
       int hi = (hh / sub_conv_n);
-
       if ((hh < omit_size) || (hh >= (origin_h - omit_size))) continue;
-
-      // for (int ww = 0; ww < origin_w; ++ww){
-
-      //   if((ww < omit_size)  )// || (ww >= (origin_w-omit_size))
-      //      continue;
-
       int sidx = (nn * origin_h * align_origin_w + hi * align_origin_w +
                   omit_size * channel);
 
       fpga_copy(ptr_deconv + deconv_idx, sub_t + sidx,
-                sizeof(half) * deconv_row_len);
+                sizeof(int16_t) * deconv_row_len);
       deconv_idx += align_deconv_row_len;
-      //}
     }
   }
+  fpga_copy(args.output.address, ptr_deconv,
+            num * align_deconv_row_len * deconv_h * sizeof(int16_t));
+  fpga_flush(args.output.address,
+             num * align_deconv_row_len * deconv_h * sizeof(int16_t));
+  fpga_free(ptr_deconv);
 
-  *data_in = ptr_deconv;
-  fpga_free(ptr_tmp);
-}
+}  // deconv_post_process
 
 int ComputeFpgaDeconv(const struct DeconvArgs &args) {
 #ifdef FPGA_PRINT_MODE
   DLOG << "=============ComputeFPGADeConv===========";
   DLOG << "   filter_num:" << args.filter_num
-       << "   group_num:" << args.group_num
+       << "   group_num:" << args.group_num << "omit_size:" << args.omit_size
+       << "sub_output_width: " << args.sub_output_width
+       << "sub_output_height: " << args.sub_output_height
        << "   sub_conv_num:" << args.sub_conv_num;
+  DLOG << "args.output.address: " << args.output.address
+       << "args.output.scale_address: " << args.output.scale_address;
+  DLOG << "args.conv_args.sb_address: " << (args.conv_args)->sb_address
+
+       << "args.conv_args.filter_address: " << (args.conv_args)->filter_address;
+#endif
+#ifndef PADDLE_MOBILE_ZU5
+  return 0;
 #endif
 
   int sub_conv_num = args.sub_conv_num;
-
   for (int i = 0; i < sub_conv_num; i++) {
-    //#if CPU_SIMULATE
-
-    //#else
     ComputeBasicConv(args.conv_args[i]);
-    //#endif
   }
 
   if (sub_conv_num > 1) {
-    float max_scale = -1.0;
+    float max_scale = -1.0f;
     for (int i = 0; i < sub_conv_num; i++) {
+      paddle_mobile::fpga::fpga_invalidate(
+          args.conv_args[i].output.scale_address, 2 * sizeof(float));
       float ptr_scale = (args.conv_args[i].output.scale_address)[0];
       if (ptr_scale > max_scale) {
         args.output.scale_address[0] = ptr_scale;
@@ -858,12 +686,11 @@ int ComputeFpgaDeconv(const struct DeconvArgs &args) {
             (args.conv_args[i].output.scale_address)[1];
       }
     }
-    deconv_post_process((half **)(&(args.output.address)), args.sub_conv_num, 1,
-                        args.filter_num, (args.sub_output_height),
-                        (args.sub_output_width), args.omit_size);
+    deconv_post_process(args);
   }
+
   return 0;
-}
+}  // ComputeFpgaDeconv
 
 int ComputeFPGASplit(const struct SplitArgs &args) {
 #ifdef FPGA_PRINT_MODE
@@ -883,6 +710,7 @@ int ComputeFPGASplit(const struct SplitArgs &args) {
                      args.scales_out, args.image_num, args.out_channel_nums,
                      args.height, args.width);
   return 0;
-}
+}  // ComputeFPGASplit
+
 }  // namespace fpga
 }  // namespace paddle_mobile

src/fpga/common/driver.cpp

@@ -153,10 +153,6 @@ int memory_request(struct fpga_memory *memory, size_t size, uint64_t *addr) {
   uint64_t _nr = DIV_ROUND_UP(size, FPGA_PAGE_SIZE);
   unsigned int nr = (unsigned int)_nr;
   int ret = 0;
-  DLOG << size;
-  DLOG << _nr;
-  DLOG << nr;
-
   uint64_t a_size = FPGA_PAGE_SIZE * nr;
   DLOG << a_size;
 
@@ -283,7 +279,7 @@ int fpga_memory_add() {
   return 0;
 }
 
-uint64_t vaddr_to_paddr(void *address) {
+uint64_t vaddr_to_paddr_driver(void *address) {
   uint64_t paddr = 0;
   auto iter = g_fpgainfo.fpga_vaddr2paddr_map.find(address);
   if (iter != g_fpgainfo.fpga_vaddr2paddr_map.end()) {
@@ -315,7 +311,7 @@ void *fpga_reg_free(void *ptr) {
     g_fpgainfo.fpga_addr2size_map.erase(iter);
     munmap(ptr, size);
   } else {
-    DLOG << "Invalid pointer";
+    DLOG << "Invalid pointer" << ptr;
   }
 }
 
@@ -347,7 +343,7 @@ void fpga_free_driver(void *ptr) {
     g_fpgainfo.fpga_addr2size_map.erase(iter);
     munmap(ptr, size);
 
-    p_addr = vaddr_to_paddr(ptr);
+    p_addr = vaddr_to_paddr_driver(ptr);
     pos = (p_addr - g_fpgainfo.memory_info->mem_start) / FPGA_PAGE_SIZE;
 
     /*clear bitmap*/
@@ -361,7 +357,7 @@ void fpga_free_driver(void *ptr) {
       g_fpgainfo.fpga_vaddr2paddr_map.erase(iter);
     }
   } else {
-    DLOG << "Invalid pointer";
+    DLOG << "Invalid pointer" << ptr;
   }
 }
 
@@ -373,7 +369,7 @@ int fpga_flush_driver(void *address, size_t size) {
   struct MemoryCacheArgs args;
   uint64_t p_addr;
 
-  p_addr = vaddr_to_paddr(address);
+  p_addr = vaddr_to_paddr_driver(address);
 
   args.offset = (void *)(p_addr - FPGA_MEM_PHY_ADDR);  // NOLINT
   args.size = size;
@@ -385,7 +381,7 @@ int fpga_invalidate_driver(void *address, size_t size) {
   struct MemoryCacheArgs args;
   uint64_t p_addr;
 
-  p_addr = vaddr_to_paddr(address);
+  p_addr = vaddr_to_paddr_driver(address);
 
   args.offset = (void *)(p_addr - FPGA_MEM_PHY_ADDR);  // NOLINT
   args.size = size;

src/fpga/common/driver.h

@@ -31,8 +31,8 @@ namespace driver {
 
 #define FPGA_REG_PHY_ADDR 0xa0000000
 #define FPGA_REG_SIZE 0x1000
-#define FPGA_MEM_PHY_ADDR 0x20000000
-#define FPGA_MEM_SIZE 0x20000000
+#define FPGA_MEM_PHY_ADDR 0x40000000
+#define FPGA_MEM_SIZE 0x80000000
 
 #define FPGA_PAGE_SIZE (16UL * 1024UL)
 
@@ -122,15 +122,11 @@ void *fpga_malloc_driver(size_t size);
 
 void fpga_free_driver(void *ptr);
 
-void fpga_copy_driver(void *dest, const void *src, size_t num);
-
 int fpga_flush_driver(void *address, size_t size);
 
 int fpga_invalidate_driver(void *address, size_t size);
 
-/*pe*/
-
-uint64_t vaddr_to_paddr(void *address);
+uint64_t vaddr_to_paddr_driver(void *address);
 
 int fpga_regpoll(uint64_t reg, uint64_t val, int time);
 

src/fpga/common/fpga_common.cpp

@@ -115,7 +115,7 @@ int fpga_invalidate(void *address, size_t size) {
 }
 uint64_t vaddr_to_paddr(void *address) {
 #ifdef PADDLE_MOBILE_ZU5
-  return driver::vaddr_to_paddr(address);
+  return driver::vaddr_to_paddr_driver(address);
 #else
   return 0;
 #endif

src/fpga/common/fpga_common.h

@@ -37,6 +37,18 @@ enum LayoutType {
   LAYOUT_HWC = 0,
 };
 
+enum ActivationType {
+  NONE = 0,
+  LEAKYRELU = 1,
+  SIGMOID = 2,
+  TANH = 3,
+};
+
+struct ActivationArgs {
+  enum ActivationType activation_type;
+  int16_t leaky_relu_negative_slope;
+};
+
 struct KernelArgs {
   uint32_t width;
   uint32_t height;
@@ -58,7 +70,10 @@ struct ImageOutputArgs {
   void* address;         // output result address;
   float* scale_address;  // output scale address;
   uint64_t timer_cnt;    // time counter for FPGA computation
+  struct ActivationArgs
+      activation;  // To select activation and specify (Leaky)Relu parameter.
 };
+
 #ifdef PADDLE_MOBILE_FPGA_V1
 struct ConvDriverParam {
   uint64_t image_address_phy;
@@ -198,7 +213,11 @@ struct DeconvArgs {
   struct ConvArgs* conv_args;
 };
 
-static inline int align_to_x(int num, int x) { return (num + x - 1) / x * x; }
+// static inline int align_to_x(int num, int x) { return (num + x - 1) / x * x;
+// }
+static inline uint32_t align_to_x(int64_t num, int64_t x) {
+  return ((uint32_t)(num + x) - 1) / (uint32_t)x * (uint32_t)x;
+}
 
 int16_t fp32_2_fp16(float fp32_num);
 float fp16_2_fp32(int16_t fp16_num);

src/framework/executor.cpp

@@ -456,9 +456,8 @@ void Executor<Device, T>::LoadMemory(const VarDesc var_desc, float *tensorInput,
                                      char **data) {}
 
 template <>
-void Executor<GPU_CL, Precision::FP32>::LoadMemory(const VarDesc var_desc,
-                                                   float *tensorInput,
-                                                   char **data) {
+void Executor<GPU_CL, float>::LoadMemory(const VarDesc var_desc,
+                                         float *tensorInput, char **data) {
   // 1. version
   uint32_t version = *reinterpret_cast<uint32_t *>(*data);
 

src/io/paddle_mobile.cpp

@@ -202,50 +202,50 @@ double PaddleMobile<CPU, float>::GetPredictTime() {
 #endif
 
 #ifdef PADDLE_MOBILE_FPGA
-template <typename Device, T P>
-void PaddleMobile<Device, P>::InjectVariable(const framework::Tensor &t,
+template <typename Device, typename T>
+void PaddleMobile<Device, T>::InjectVariable(const framework::Tensor &t,
                                              std::string var_name) {
   executor_->InjectVariable(t, var_name);
 }
 
-template <typename Device, T P>
-void PaddleMobile<Device, P>::FeedData(const framework::Tensor &t) {
+template <typename Device, typename T>
+void PaddleMobile<Device, T>::FeedData(const framework::Tensor &t) {
   executor_->FeedData(t);
 }
 
-template <typename Device, T P>
-std::shared_ptr<framework::Tensor> PaddleMobile<Device, P>::FetchResult(
+template <typename Device, typename T>
+std::shared_ptr<framework::Tensor> PaddleMobile<Device, T>::FetchResult(
     int id) {
   return executor_->FetchResult(id);
 }
 
-template <typename Device, T P>
-void PaddleMobile<Device, P>::Predict_From_To(int start, int end) {
+template <typename Device, typename T>
+void PaddleMobile<Device, T>::Predict_From_To(int start, int end) {
   executor_->Predict_From_To(start, end);
 }
 
-template <typename Device, T P>
-void PaddleMobile<Device, P>::Predict_From(int start) {
+template <typename Device, typename T>
+void PaddleMobile<Device, T>::Predict_From(int start) {
   executor_->Predict_From(start);
 }
 
-template <typename Device, T P>
-void PaddleMobile<Device, P>::Predict_To(int end) {
+template <typename Device, typename T>
+void PaddleMobile<Device, T>::Predict_To(int end) {
   executor_->Predict_To(end);
 }
 #endif
 
 #ifdef PADDLE_MOBILE_CL
 static std::mutex lc;
-template <typename Device, T P>
-void PaddleMobile<Device, P>::SetCLPath(std::string path) {
+template <typename Device, typename T>
+void PaddleMobile<Device, T>::SetCLPath(std::string path) {
   std::lock_guard<std::mutex> lock(lc);
   if (framework::CLEngine::Instance()->GetCLPath() == "") {
     framework::CLEngine::Instance()->setClPath(path);
   }
 }
 template <>
-double PaddleMobile<GPU_CL, T::FP32>::GetPredictTime() {
+double PaddleMobile<GPU_CL, float>::GetPredictTime() {
   cl_int status;
   cl_uint nPlatform;
   clGetPlatformIDs(0, NULL, &nPlatform);
@@ -443,8 +443,8 @@ double PaddleMobile<GPU_CL, T::FP32>::GetPredictTime() {
     return -1;
   }
 }
-template <typename Device, T P>
-int PaddleMobile<Device, P>::readText(
+template <typename Device, typename T>
+int PaddleMobile<Device, T>::readText(
     const char *kernelPath,
     char **pcode) {  // 读取文本文件放入 pcode，返回字符串长度
   FILE *fp;

src/operators/kernel/cl/feed_kernel.cpp

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #include "operators/kernel/feed_kernel.h"
 #include "framework/cl/cl_tensor.h"
+
 namespace paddle_mobile {
 namespace operators {
 
@@ -43,7 +44,7 @@ void FeedKernel<GPU_CL, float>::Compute(const FeedParam<GPU_CL> &param) {
   const int Stride2 = out_C * out_H * out_W;
   const int Stride1 = out_H * out_W;
   const int Stride0 = out_W;
-  CLTensor input_cl_tensor(this->cl_helper_.CLContext(),
+  framework::CLTensor input_cl_tensor(this->cl_helper_.CLContext(),
                                       this->cl_helper_.CLCommandQueue());
   input_cl_tensor.Resize(input->dims());
   cl_mem inputBuffer = input_cl_tensor.mutable_with_data<float>(input_data);

src/operators/kernel/cl/fusion_fc_kernel.cpp

@@ -94,8 +94,9 @@ void FusionFcCompute(const FusionFcParam<GPU_CL> &param, cl_context context,
     memory::Copy(out_data + i * classes, input_z_data, sizeof(float) * classes);
   }
 
-  math::MatMul<float>(x_matrix, false, y_matrix, false, static_cast<float>(1),
-                      out, static_cast<float>(1), false);
+  math::MatMul<float, float>(x_matrix, false, y_matrix, false,
+                             static_cast<float>(1), out, static_cast<float>(1),
+                             false);
 
   out_image->InitEmptyImage(context, commandQueue, out->dims());
   framework::TensorToCLImage(out, out_image, context, commandQueue, kernel1);

src/operators/kernel/fpga/V1/transpose2_kernel.cpp

@@ -14,7 +14,6 @@ limitations under the License. */
 #ifdef TRANSPOSE2_OP
 
 #include "operators/kernel/transpose2_kernel.h"
-#include "operators/kernel/central-arm-func/transpose2_arm_func.h"
 
 namespace paddle_mobile {
 namespace operators {

src/operators/math/gemm.cpp

@@ -3160,7 +3160,8 @@ void Gemm::Sgemm_omp(int m, int n, int k, float alpha, const float *A, int lda,
 #endif
 
   //  int L1 = 64 / max_threads * 1024;
-  int L1 = 32 / max_threads * 1024;
+  int L = (max_threads > 2) ? 64 : 32;
+  int L1 = L / max_threads * 1024;
   KC = k;
   zero = static_cast<float *>(paddle_mobile::memory::Alloc(sizeof(float) * KC));
   memset(static_cast<void *>(zero), 0, sizeof(float) * KC);

tools/op.cmake

@@ -213,8 +213,6 @@ if(NOT FOUND_MATCH)
   set(FUSION_CONVADD_OP ON)
   set(FUSION_CONVADDPRELU_OP ON)
   set(FUSION_CONVADDRELU_OP ON)
-#  set(FUSION_CONVADDRELU_INT8_OP ON)
-#  set(FUSION_FC_INT8_OP ON)
   set(FUSION_FC_OP ON)
   set(LRN_OP ON)
   set(MUL_OP ON)
@@ -316,9 +314,6 @@ endif()
 if (FUSION_CONVADDRELU_OP)
   add_definitions(-DFUSION_CONVADDRELU_OP)
 endif()
-if (FUSION_CONVADDRELU_INT8_OP)
-  add_definitions(-DFUSION_CONVADDRELU_INT8_OP)
-endif()
 if (FUSION_CONVADDPRELU_OP)
   add_definitions(-DFUSION_CONVADDPRELU_OP)
 endif()
@@ -328,9 +323,6 @@ endif()
 if (FUSION_FC_OP)
   add_definitions(-DFUSION_FC_OP)
 endif()
-if(FUSION_FC_INT8_OP)
-  add_definitions(-DFUSION_FC_INT8_OP)
-endif()
 if (LRN_OP)
   add_definitions(-DLRN_OP)
 endif()
@@ -490,7 +482,6 @@ if (FUSION_DEQUANT_ADD_BN_RELU_QUANT_OP)
 #  add_definitions(-DFUSION_DEQUANT_ADD_BN_RELU_QUANT_OP)
 endif()
 
-
 if (TANH_OP)
   add_definitions(-DTANH_OP)
 endif()