diff --git a/src/fpga/V1/api.cpp b/src/fpga/V1/api.cpp
index a8540e72f3ad077ac4aa49e34b535675f04dcd16..9a408a8f2fbe3c600679ddb2e3eadb493f323165 100644
--- a/src/fpga/V1/api.cpp
+++ b/src/fpga/V1/api.cpp
@@ -346,9 +346,9 @@ void expand_conv_arg(ConvArgs *arg) {
   auto filter_pad_width_mul_channel =
       args.image.pad_width * args.image.channels;
   auto image_amount_per_row_multi_win_first =
-      image_amount_per_row * (4 * args.kernel.stride_h - args.image.pad_height);
+      image_amount_per_row * (2 * args.kernel.stride_h - args.image.pad_height);
   auto image_amount_per_row_multi_win =
-      image_amount_per_row * (4 * args.kernel.stride_h);
+      image_amount_per_row * (2 * args.kernel.stride_h);
 
   auto image_block_num = block_num;
   auto image_block_len =
@@ -375,7 +375,8 @@ void expand_conv_arg(ConvArgs *arg) {
       (512 / (align_to_x(args.filter_num, 4) / 4 * 2) > 2)
           ? (512 / (align_to_x(args.filter_num, 4) / 4 * 2) - 2)
           : 0;
-  auto cmd = 0UL | (args.relu_enabled ? USE_RELU : 0) | USE_BIAS;
+  // auto cmd = 0UL | (args.relu_enabled ? USE_RELU : 0) | USE_BIAS;
+  auto cmd = 0UL | USE_BIAS;
 
   auto deconv_param = ((args.deconv_tx_param.deconv_en) << 24) |
                       ((args.deconv_tx_param.sub_conv_num) << 16) |
@@ -413,7 +414,8 @@ void expand_conv_arg(ConvArgs *arg) {
 
 void expand_EW_arg(EWAddArgs *arg) {
   EWAddArgs args = *arg;
-  uint64_t cmd = args.relu_enabled ? USE_RELU : 0;
+  // uint64_t cmd = args.relu_enabled ? USE_RELU : 0;
+  uint64_t cmd = 0;
   uint64_t datalen = (uint64_t)args.image0.width *
                      (uint64_t)args.image0.height *
                      (uint64_t)args.image0.channels;
@@ -441,8 +443,10 @@ void expand_EW_arg(EWAddArgs *arg) {
 
 void fill_split_arg(struct SplitConvArgs *arg, framework::Tensor *input,
                     framework::Tensor *out, framework::Tensor *filter,
-                    bool relu_enabled, int group_num, int stride_h,
-                    int stride_w, int padding_h, int padding_w, float *bs_ptr) {
+                    ActivationType activation_enable,
+                    int16_t leaky_relu_negative_slope, int group_num,
+                    int stride_h, int stride_w, int padding_h, int padding_w,
+                    float *bs_ptr) {
   auto input_ptr = input->data<float>();
   auto filter_ptr = filter->data<float>();
   auto out_ptr = out->data<float>();
@@ -488,7 +492,10 @@ void fill_split_arg(struct SplitConvArgs *arg, framework::Tensor *input,
             filter->dims()[3]));
 
   for (int i = 0; i < n; i++) {
-    arg->conv_arg[i].relu_enabled = relu_enabled;
+    // arg->conv_arg[i].relu_enabled = relu_enabled;
+    arg->conv_arg[i].output.activation.activation_type = activation_enable;
+    arg->conv_arg[i].output.activation.leaky_relu_negative_slope =
+        leaky_relu_negative_slope;
     arg->conv_arg[i].group_num = (uint32_t)group_num;
     arg->conv_arg[i].kernel.stride_h = (uint32_t)stride_h;
     arg->conv_arg[i].kernel.stride_w = (uint32_t)stride_w;
@@ -560,8 +567,9 @@ void fill_split_arg(struct SplitConvArgs *arg, framework::Tensor *input,
 
 void fill_deconv_arg(struct DeconvArgs *arg, framework::Tensor *input,
                      framework::Tensor *out, framework::Tensor *filter,
-                     bool relu_enabled, int group_num, int stride_h,
-                     int stride_w, int padding_h, int padding_w,
+                     ActivationType activation_enable,
+                     int16_t leaky_relu_negative_slope, int group_num,
+                     int stride_h, int stride_w, int padding_h, int padding_w,
                      float *bs_ptr) {
   auto input_ptr = input->data<float>();
   auto filter_ptr = filter->data<float>();
@@ -687,7 +695,13 @@ void fill_deconv_arg(struct DeconvArgs *arg, framework::Tensor *input,
     }
 
     for (int j = 0; j < split_num; ++j) {
-      arg->split_conv_args[i]->conv_arg[j].relu_enabled = relu_enabled;
+      // arg->split_conv_args[i]->conv_arg[j].relu_enabled = relu_enabled;
+      arg->split_conv_args[i]->conv_arg[j].output.activation.activation_type =
+          activation_enable;
+      arg->split_conv_args[i]
+          ->conv_arg[j]
+          .output.activation.leaky_relu_negative_slope =
+          leaky_relu_negative_slope;
       arg->split_conv_args[i]->conv_arg[j].group_num = (uint32_t)group_num;
 
       arg->split_conv_args[i]->conv_arg[j].kernel.width =
@@ -800,13 +814,17 @@ void fill_deconv_arg(struct DeconvArgs *arg, framework::Tensor *input,
 
 void fill_dwconv_arg(struct DWconvArgs *arg, framework::Tensor *input,
                      framework::Tensor *out, framework::Tensor *filter,
-                     bool relu_enabled, int stride_h, int stride_w,
-                     int padding_h, int padding_w, float *bias_ptr) {
+                     ActivationType activation_enable,
+                     int16_t leaky_relu_negative_slope, int stride_h,
+                     int stride_w, int padding_h, int padding_w,
+                     float *bias_ptr) {
   auto filter_ptr = filter->data<float>();
   auto input_ptr = input->data<float>();
   auto output_ptr = out->mutable_data<float>();
   arg->sub_conv_num = 1;
-  arg->relu_enabled = relu_enabled;
+  // arg->relu_enabled = relu_enabled;
+  arg->output.activation.activation_type = activation_enable;
+  arg->output.activation.leaky_relu_negative_slope = leaky_relu_negative_slope;
   arg->bias_address = bias_ptr;
   arg->filter_address = filter_ptr;
   arg->kernel.height = (uint32_t)filter->dims()[2];
@@ -826,8 +844,10 @@ void fill_dwconv_arg(struct DWconvArgs *arg, framework::Tensor *input,
 
 void fill_DWDeconv_arg(struct DWDeconvArgs *arg, framework::Tensor *input,
                        framework::Tensor *out, framework::Tensor *filter,
-                       bool relu_enabled, int stride_h, int stride_w,
-                       int padding_h, int padding_w, float *bias_ptr) {
+                       ActivationType activation_enable,
+                       int16_t leaky_relu_negative_slope, int stride_h,
+                       int stride_w, int padding_h, int padding_w,
+                       float *bias_ptr) {
   auto filter_ptr = filter->data<float>();
   auto input_ptr = input->data<float>();
   auto output_ptr = out->mutable_data<float>();
@@ -884,7 +904,10 @@ void fill_DWDeconv_arg(struct DWDeconvArgs *arg, framework::Tensor *input,
     arg->dw_conv_args.push_back(std::make_shared<DWconvArgs>());
 
     arg->dw_conv_args[i]->sub_conv_num = sub_conv_num;
-    arg->dw_conv_args[i]->relu_enabled = relu_enabled;
+    // arg->dw_conv_args[i]->relu_enabled = relu_enabled;
+    arg->dw_conv_args[i]->output.activation.activation_type = activation_enable;
+    arg->dw_conv_args[i]->output.activation.leaky_relu_negative_slope =
+        leaky_relu_negative_slope;
     arg->dw_conv_args[i]->bias_address = bias_ptr;
 
     arg->dw_conv_args[i]->filter_address =
diff --git a/src/fpga/V1/api.h b/src/fpga/V1/api.h
index 05d6a938c85f14770b97cd477580d0e6103fa777..05a30ddce4828bf8ac0f049ea0db4f18dc1dba79 100644
--- a/src/fpga/V1/api.h
+++ b/src/fpga/V1/api.h
@@ -47,20 +47,28 @@ void format_concat_output(framework::Tensor* out, int height, int width,
 
 void fill_split_arg(struct SplitConvArgs* arg, framework::Tensor* input,
                     framework::Tensor* out, framework::Tensor* filter,
-                    bool relu_enabled, int group_num, int stride_h,
-                    int stride_w, int padding_h, int padding_w, float* bs_ptr);
+                    ActivationType activation_enable,
+                    int16_t leaky_relu_negative_slope, int group_num,
+                    int stride_h, int stride_w, int padding_h, int padding_w,
+                    float* bs_ptr);
 void fill_deconv_arg(struct DeconvArgs* arg, framework::Tensor* input,
                      framework::Tensor* out, framework::Tensor* filter,
-                     bool relu_enabled, int group_num, int stride_h,
-                     int stride_w, int padding_h, int padding_w, float* bs_ptr);
+                     ActivationType activation_enable,
+                     int16_t leaky_relu_negative_slope, int group_num,
+                     int stride_h, int stride_w, int padding_h, int padding_w,
+                     float* bs_ptr);
 void fill_dwconv_arg(struct DWconvArgs* arg, framework::Tensor* input,
                      framework::Tensor* out, framework::Tensor* filter,
-                     bool relu_enabled, int stride_h, int stride_w,
-                     int padding_h, int padding_w, float* bias_ptr);
+                     ActivationType activation_enable,
+                     int16_t leaky_relu_negative_slope, int stride_h,
+                     int stride_w, int padding_h, int padding_w,
+                     float* bias_ptr);
 void fill_DWDeconv_arg(struct DWDeconvArgs* arg, framework::Tensor* input,
                        framework::Tensor* out, framework::Tensor* filter,
-                       bool relu_enabled, int stride_h, int stride_w,
-                       int padding_h, int padding_w, float* bs_ptr);
+                       ActivationType activation_enable,
+                       int16_t leaky_relu_negative_slope, int stride_h,
+                       int stride_w, int padding_h, int padding_w,
+                       float* bs_ptr);
 
 void format_deconv_filter(framework::Tensor* filter_tensor, float max_value,
                           int group_num, int stride);
diff --git a/src/fpga/V1/deconv_filter.cpp b/src/fpga/V1/deconv_filter.cpp
index 4c484a45d0a36db4aac677377ae11b5235603ac6..7c87452f5a7264ad069d8508cb1e9dc24f5cdc3d 100644
--- a/src/fpga/V1/deconv_filter.cpp
+++ b/src/fpga/V1/deconv_filter.cpp
@@ -19,7 +19,6 @@ limitations under the License. */
 #include "fpga/V1/filter.h"
 // #include "filter.h"
 #include "fpga/V1/api.h"
-// #include "fpga_api.h"
 
 namespace paddle_mobile {
 namespace fpga {
diff --git a/src/fpga/V1/pe.cpp b/src/fpga/V1/pe.cpp
index 16d3bc793389f49ad0b6e3bf3b064a880e4a927a..5a81e2422979f08b2113bd9b46022fe4d77154cb 100644
--- a/src/fpga/V1/pe.cpp
+++ b/src/fpga/V1/pe.cpp
@@ -63,6 +63,7 @@ using namespace std;     // NOLINT
 #define REG_TIMER_COUNTER 0x070
 
 #define REG_SCALE_PARAMETER 0x080
+#define REG_ACTIVATION_MODE_AND_LEAKY_RELU_FACTOR 0x090
 
 #define REG_FLASH_CMD 0x200
 #define REG_FLASH_DATA 0x208
@@ -189,8 +190,8 @@ int ComputeFpgaConv(const struct SplitConvArgs &args) {
 int ComputeBasicConv(const struct ConvArgs &args) {
 #ifdef FPGA_PRINT_MODE
   DLOG << "======Compute Basic Conv======";
-  DLOG << "   relu_enabled:" << args.relu_enabled
-       << "   sb_address:" << args.sb_address
+  // DLOG << "   relu_enabled:" << args.relu_enabled
+  DLOG << "   sb_address:" << args.sb_address
        << "   filter_address:" << args.filter_address
        << "   filter_num:" << args.filter_num
        << "   group_num:" << args.group_num;
@@ -212,6 +213,25 @@ int ComputeBasicConv(const struct ConvArgs &args) {
 #ifdef PADDLE_MOBILE_ZU5
   int ret = 0;
   uint64_t output_scale = 0;
+
+  uint64_t reg_ActivationArgs = 0;
+  // active function:{none,leakeyrelu,sigmoid,tanh}
+  ActivationArgs active_args;
+  // active_args.activation_type = LEAKYRELU;
+
+  active_args.activation_type = args.output.activation.activation_type;
+
+  active_args.leaky_relu_negative_slope =
+      args.output.activation.leaky_relu_negative_slope;
+
+  reg_ActivationArgs = (uint64_t(active_args.activation_type) << 32) |
+                       active_args.leaky_relu_negative_slope;
+
+  DLOG << "   activation_type:" << active_args.activation_type
+       << "   leaky_relu_negative_slope:"
+       << active_args.leaky_relu_negative_slope;
+  DLOG << "   reg_ActivationArgs:" << reg_ActivationArgs;
+
   pthread_mutex_lock(&g_fpgainfo.pe_data->mutex);
   if (ERROR == g_fpgainfo.pe_data->pes[PE_IDX_CONV]->status) {
     ret = -EIO;
@@ -219,6 +239,10 @@ int ComputeBasicConv(const struct ConvArgs &args) {
     pthread_mutex_unlock(&g_fpgainfo.pe_data->mutex);
     return ret;
   }
+
+  reg_writeq(reg_ActivationArgs,
+             REG_ACTIVATION_MODE_AND_LEAKY_RELU_FACTOR);  // active functoion
+
   reg_writeq(output_scale, REG_SCALE_PARAMETER);
   reg_writeq(
       ((uint64_t)args.image.height) | (((uint64_t)args.image.width) << 32),
@@ -278,6 +302,9 @@ int ComputeBasicConv(const struct ConvArgs &args) {
   output_scale = (output_scale << 32) | (output_scale >> 32);
   fpga_copy(args.output.scale_address, &output_scale, sizeof(float) * 2);
 
+  active_args.activation_type = NONE;
+  reg_writeq(reg_ActivationArgs, REG_ACTIVATION_MODE_AND_LEAKY_RELU_FACTOR);
+
   pthread_mutex_unlock(&g_fpgainfo.pe_data->mutex);
 
   return ret;
@@ -314,6 +341,23 @@ int ComputeFpgaPool(const struct PoolingArgs &args) {
   uint64_t image_physical_address = 0;
   uint64_t output_physical_address = 0;
 
+  uint64_t reg_ActivationArgs = 0;
+  // active function:{none,leakeyrelu,sigmoid,tanh}
+  ActivationArgs active_args;
+  // active_args.activation_type = LEAKYRELU;
+  active_args.activation_type = args.output.activation.activation_type;
+
+  active_args.leaky_relu_negative_slope =
+      args.output.activation.leaky_relu_negative_slope;
+
+  reg_ActivationArgs = (uint64_t(active_args.activation_type) << 32) |
+                       active_args.leaky_relu_negative_slope;
+
+  DLOG << "   activation_type:" << active_args.activation_type
+       << "   leaky_relu_negative_slope:"
+       << active_args.leaky_relu_negative_slope;
+  DLOG << "   reg_ActivationArgs:" << reg_ActivationArgs;
+
   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)(
@@ -364,6 +408,9 @@ int ComputeFpgaPool(const struct PoolingArgs &args) {
     return ret;
   }
 
+  reg_writeq(reg_ActivationArgs,
+             REG_ACTIVATION_MODE_AND_LEAKY_RELU_FACTOR);  // active functoion
+
   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);
@@ -408,6 +455,10 @@ int ComputeFpgaPool(const struct PoolingArgs &args) {
   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);
+
+  active_args.activation_type = NONE;
+  reg_writeq(reg_ActivationArgs, REG_ACTIVATION_MODE_AND_LEAKY_RELU_FACTOR);
+
   pthread_mutex_unlock(&g_fpgainfo.pe_data->mutex);
 
   return ret;
@@ -418,8 +469,8 @@ int ComputeFpgaPool(const struct PoolingArgs &args) {
 int ComputeFpgaEWAdd(const struct EWAddArgs &args) {
 #ifdef FPGA_PRINT_MODE
   DLOG << "=============ComputeFpgaEWAdd===========";
-  DLOG << "   relu_enabled:" << args.relu_enabled
-       << "   const0:" << fp16_2_fp32(int16_t(args.const0))
+  // DLOG << "   relu_enabled:" << args.relu_enabled
+  DLOG << "   const0:" << fp16_2_fp32(int16_t(args.const0))
        << "   const1:" << fp16_2_fp32(int16_t(args.const1));
   DLOG << "   image0_address:" << args.image0.address
        << "   image0_scale_address:" << args.image0.scale_address
@@ -441,6 +492,19 @@ int ComputeFpgaEWAdd(const struct EWAddArgs &args) {
 #ifdef PADDLE_MOBILE_ZU5
   int ret = 0;
   uint64_t output_scale = 0;
+
+  uint64_t reg_ActivationArgs = 0;
+  ActivationArgs active_args;
+  active_args.activation_type = args.output.activation.activation_type;
+  active_args.leaky_relu_negative_slope =
+      args.output.activation.leaky_relu_negative_slope;
+  reg_ActivationArgs = (uint64_t(active_args.activation_type) << 32) |
+                       active_args.leaky_relu_negative_slope;
+  DLOG << "    activation_type:" << active_args.activation_type
+       << "    leaky_relu_negative_slope:"
+       << active_args.leaky_relu_negative_slope;
+  DLOG << "    reg_ActivationArgs:" << reg_ActivationArgs;
+
   pthread_mutex_lock(&g_fpgainfo.pe_data->mutex);
   if (ERROR == g_fpgainfo.pe_data->pes[PE_IDX_EW]->status) {
     ret = -EIO;
@@ -449,6 +513,9 @@ int ComputeFpgaEWAdd(const struct EWAddArgs &args) {
     return ret;
   }
 
+  reg_writeq(reg_ActivationArgs,
+             REG_ACTIVATION_MODE_AND_LEAKY_RELU_FACTOR);  // active functoion
+
   reg_writeq(output_scale, REG_SCALE_PARAMETER);
   reg_writeq(args.driver.image0_address_phy, REG_EW_IMAGE0_BASE_ADDR);
   reg_writeq(args.driver.image1_address_phy, REG_EW_IMAGE1_BASE_ADDR);
@@ -468,6 +535,9 @@ int ComputeFpgaEWAdd(const struct EWAddArgs &args) {
   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);
+  active_args.activation_type = NONE;
+  reg_writeq(reg_ActivationArgs, REG_ACTIVATION_MODE_AND_LEAKY_RELU_FACTOR);
+
   pthread_mutex_unlock(&g_fpgainfo.pe_data->mutex);
   return ret;
 #endif
@@ -501,6 +571,17 @@ int PerformBypass(const struct BypassArgs &args) {
   uint8_t data_cell_in = 0;
   uint8_t data_cell_out = 0;
   int ret = 0;
+
+  uint64_t reg_ActivationArgs = 0;
+  ActivationArgs active_args;
+  active_args.activation_type = args.output.activation.activation_type;
+
+  active_args.leaky_relu_negative_slope =
+      args.output.activation.leaky_relu_negative_slope;
+
+  reg_ActivationArgs = (uint64_t(active_args.activation_type) << 32) |
+                       active_args.leaky_relu_negative_slope;
+
   datalen = (uint64_t)args.image.width * (uint64_t)args.image.height *
             (uint64_t)args.image.channels;
   datalen = align_to_x(datalen, 16);
@@ -559,7 +640,6 @@ int PerformBypass(const struct BypassArgs &args) {
       (data_cell_out != SIZE_FP16 && data_cell_out != SIZE_FP32)) {
     return -EFAULT;
   }
-
   pthread_mutex_lock(&g_fpgainfo.pe_data->mutex);
   if (ERROR == g_fpgainfo.pe_data->pes[PE_IDX_BYPASS]->status) {
     ret = -EIO;
@@ -567,7 +647,8 @@ int PerformBypass(const struct BypassArgs &args) {
     pthread_mutex_unlock(&g_fpgainfo.pe_data->mutex);
     return ret;
   }
-
+  reg_writeq(reg_ActivationArgs,
+             REG_ACTIVATION_MODE_AND_LEAKY_RELU_FACTOR);  // active functoion
   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);
@@ -585,6 +666,7 @@ int PerformBypass(const struct BypassArgs &args) {
   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);
+  reg_writeq(reg_ActivationArgs, REG_ACTIVATION_MODE_AND_LEAKY_RELU_FACTOR);
   pthread_mutex_unlock(&g_fpgainfo.pe_data->mutex);
   return ret;
 #endif
@@ -808,7 +890,7 @@ int ComputeFPGASplit(const struct SplitArgs &args) {
 int ComputeDWConv(const struct DWconvArgs &args) {
 #ifdef FPGA_PRINT_MODE
   DLOG << "=============ComputeDWConv===========";
-  DLOG << "   mode:" << args.relu_enabled;
+  // DLOG << "   mode:" << args.relu_enabled;
   DLOG << "   image_address:" << args.image.address
        << "   image_scale_address:" << args.image.scale_address
        << "   image_channels:" << args.image.channels
@@ -831,7 +913,8 @@ int ComputeDWConv(const struct DWconvArgs &args) {
   uint64_t output_scale = 0;
   uint64_t timer_cnt = 0;
   int ret = 0;
-  uint64_t cmd = args.relu_enabled;
+  // uint64_t cmd = args.relu_enabled;
+  uint64_t cmd = 0;
   uint64_t image_physical_address = 0;
   uint64_t output_physical_address = 0;
   uint64_t filter_physical_address = 0;
diff --git a/src/fpga/common/driver.cpp b/src/fpga/common/driver.cpp
index 18a310b09cad4a741eb83453a09f3c94d4f0db05..b1d3559dbbb238ae24cc6224e2d253dab744dce1 100644
--- a/src/fpga/common/driver.cpp
+++ b/src/fpga/common/driver.cpp
@@ -154,7 +154,6 @@ int memory_request(struct fpga_memory *memory, size_t size, uint64_t *addr) {
   unsigned int nr = (unsigned int)_nr;
   int ret = 0;
   uint64_t a_size = FPGA_PAGE_SIZE * nr;
-  DLOG << a_size;
 
   pthread_mutex_lock(&memory->mutex);
 
@@ -391,9 +390,6 @@ int fpga_invalidate_driver(void *address, size_t size) {
 
 void fpga_copy_driver(void *dest, const void *src, size_t num) {
   uint64_t i;
-
-  DLOG << "dest:" << dest << " src:" << src << " size:" << num;
-
   for (i = 0; i < num; i++) {
     *((int8_t *)dest + i) = *((int8_t *)src + i);  // NOLINT
   }
diff --git a/src/fpga/common/driver.h b/src/fpga/common/driver.h
index 4fa83b776e7b3df5df5e536de91093fd18ca67a1..d35627cd46b3f233255a98d1e1fbca27469f715c 100644
--- a/src/fpga/common/driver.h
+++ b/src/fpga/common/driver.h
@@ -29,7 +29,7 @@ namespace driver {
 
 #define DIV_ROUND_UP(n, d) (((n) + (d)-1) / (d))
 
-#define FPGA_REG_PHY_ADDR 0xa0000000
+#define FPGA_REG_PHY_ADDR 0x80000000
 #define FPGA_REG_SIZE 0x1000
 #define FPGA_MEM_PHY_ADDR 0x40000000
 #define FPGA_MEM_SIZE 0x80000000
diff --git a/src/fpga/common/fpga_common.h b/src/fpga/common/fpga_common.h
index 25ca99613e91dcbab4ffedf3802f2025afdc040f..60753e5cde1e39a1dbf4a1016667db748fc6b9f9 100644
--- a/src/fpga/common/fpga_common.h
+++ b/src/fpga/common/fpga_common.h
@@ -45,6 +45,7 @@ enum ActivationType {
   LEAKYRELU = 1,
   SIGMOID = 2,
   TANH = 3,
+  SOFTMAX = 4,
 };
 
 struct ActivationArgs {
@@ -132,7 +133,7 @@ struct DeconvTxParm {
 #endif
 
 struct ConvArgs {
-  bool relu_enabled;
+  // bool relu_enabled;
   void* sb_address;  // scale and bias
   void* filter_address;
   float* filter_scale_address;
@@ -198,7 +199,7 @@ struct PoolingArgs {
 };
 
 struct EWAddArgs {
-  bool relu_enabled;
+  // bool relu_enabled;
   uint32_t const0;  // output0 = const0 x input0 + const1 x input1;
   uint32_t const1;
   struct ImageInputArgs image0;
@@ -230,7 +231,7 @@ struct DeconvArgs {
 };
 struct DWconvArgs {
   uint32_t sub_conv_num;
-  bool relu_enabled;
+  // bool relu_enabled;
   void* bias_address;
   void* filter_address;
   struct KernelArgs kernel;
diff --git a/src/operators/activation_op.cpp b/src/operators/activation_op.cpp
index bcff87c9276721c19a970eb328fc0a183ed6c003..76c9e1a014bc0e51b032d8516ba9448fa25b2aa5 100644
--- a/src/operators/activation_op.cpp
+++ b/src/operators/activation_op.cpp
@@ -31,6 +31,10 @@ DEFINE_ACTIVATION_INFERSHAPE(Relu6);
 
 #ifdef SIGMOID_OP
 DEFINE_ACTIVATION_INFERSHAPE(Sigmoid);
+namespace ops = paddle_mobile::operators;
+#ifdef PADDLE_MOBILE_FPGA
+REGISTER_OPERATOR_FPGA(sigmoid, ops::SigmoidOp);
+#endif
 #endif  // SIGMOID_OP
 
 #ifdef TANH_OP
diff --git a/src/operators/kernel/fpga/V1/conv_add_bn_kernel.cpp b/src/operators/kernel/fpga/V1/conv_add_bn_kernel.cpp
index 30ff3155a47c813f303dc59191edd8b60e6d8ce3..3e41efdf76ed5b14d408a1278c7dba0bd1f30a1f 100644
--- a/src/operators/kernel/fpga/V1/conv_add_bn_kernel.cpp
+++ b/src/operators/kernel/fpga/V1/conv_add_bn_kernel.cpp
@@ -22,7 +22,10 @@ namespace operators {
 
 template <>
 bool ConvAddBNKernel<FPGA, float>::Init(FusionConvAddBNParam<FPGA> *param) {
-  bool relu_enabled = false;
+  // bool relu_enabled = false;
+  paddle_mobile::fpga::ActivationType activation_enable =
+      paddle_mobile::fpga::NONE;
+  int16_t leaky_relu_negative_slope = 0;
   auto input = const_cast<Tensor *>(param->Input());
 
   auto bias = param->Bias();
@@ -61,10 +64,10 @@ bool ConvAddBNKernel<FPGA, float>::Init(FusionConvAddBNParam<FPGA> *param) {
 
   fpga::format_conv_data(filter, out, &bs_ptr, param->Groups());
   fpga::SplitConvArgs conv_arg = {0};
-  fpga::fill_split_arg(&conv_arg, input, out, filter, relu_enabled,
-                       param->Groups(), param->Strides()[0],
-                       param->Strides()[1], param->Paddings()[0],
-                       param->Paddings()[1], bs_ptr);
+  fpga::fill_split_arg(&conv_arg, input, out, filter, activation_enable,
+                       leaky_relu_negative_slope, param->Groups(),
+                       param->Strides()[0], param->Strides()[1],
+                       param->Paddings()[0], param->Paddings()[1], bs_ptr);
   param->SetFpgaArgs(conv_arg);
 
   return true;
diff --git a/src/operators/kernel/fpga/V1/conv_add_bn_relu_kernel.cpp b/src/operators/kernel/fpga/V1/conv_add_bn_relu_kernel.cpp
index 7f720323253fff53f7d1bb92f8bfeec77bf0da14..b7b99be78acae80c46b9d1bd1f3cb72d5f4a7cfb 100644
--- a/src/operators/kernel/fpga/V1/conv_add_bn_relu_kernel.cpp
+++ b/src/operators/kernel/fpga/V1/conv_add_bn_relu_kernel.cpp
@@ -23,7 +23,10 @@ namespace operators {
 template <>
 bool ConvAddBNReluKernel<FPGA, float>::Init(
     FusionConvAddBNReluParam<FPGA> *param) {
-  bool relu_enabled = true;
+  // bool relu_enabled = true;
+  paddle_mobile::fpga::ActivationType activation_enable =
+      paddle_mobile::fpga::LEAKYRELU;
+  int16_t leaky_relu_negative_slope = 0;
   auto input = const_cast<Tensor *>(param->Input());
   auto bias = param->Bias();
   auto bias_ptr = bias->data<float>();
@@ -64,16 +67,16 @@ bool ConvAddBNReluKernel<FPGA, float>::Init(
   if (groups == channel) {
     fpga::format_dwconv_data(filter, out, new_scale_ptr, &new_bias_ptr);
     fpga::DWconvArgs dwconv_arg = {0};
-    fpga::fill_dwconv_arg(&dwconv_arg, input, out, filter, relu_enabled,
-                          strides[0], strides[1], paddings[0], paddings[1],
-                          new_bias_ptr);
+    fpga::fill_dwconv_arg(&dwconv_arg, input, out, filter, activation_enable,
+                          leaky_relu_negative_slope, strides[0], strides[1],
+                          paddings[0], paddings[1], new_bias_ptr);
     param->SetFpgaArgs(dwconv_arg);
   } else {
     fpga::format_conv_data(filter, out, &bs_ptr, param->Groups());
     fpga::SplitConvArgs conv_arg = {0};
-    fpga::fill_split_arg(&conv_arg, input, out, filter, relu_enabled,
-                         param->Groups(), strides[0], strides[1], paddings[0],
-                         paddings[1], bs_ptr);
+    fpga::fill_split_arg(&conv_arg, input, out, filter, activation_enable,
+                         leaky_relu_negative_slope, param->Groups(), strides[0],
+                         strides[1], paddings[0], paddings[1], bs_ptr);
     param->SetFpgaArgs(conv_arg);
   }
   return true;
diff --git a/src/operators/kernel/fpga/V1/conv_add_kernel.cpp b/src/operators/kernel/fpga/V1/conv_add_kernel.cpp
old mode 100755
new mode 100644
index e566dc9b165811a3e8a9f78d040cc8c571fd93a9..153be5a4f888c2a39a7b05b9a7fbb72e305acb8d
--- a/src/operators/kernel/fpga/V1/conv_add_kernel.cpp
+++ b/src/operators/kernel/fpga/V1/conv_add_kernel.cpp
@@ -21,7 +21,10 @@ namespace operators {
 
 template <>
 bool ConvAddKernel<FPGA, float>::Init(FusionConvAddParam<FPGA> *param) {
-  bool relu_enabled = false;
+  // bool relu_enabled = false;
+  paddle_mobile::fpga::ActivationType activation_enable =
+      paddle_mobile::fpga::NONE;
+  int16_t leaky_relu_negative_slope = 0;
   auto input = const_cast<Tensor *>(param->Input());
   const Tensor *bias = param->Bias();
   auto bias_ptr = bias->data<float>();
@@ -40,10 +43,10 @@ bool ConvAddKernel<FPGA, float>::Init(FusionConvAddParam<FPGA> *param) {
 
   fpga::format_conv_data(filter, out, &bs_ptr, param->Groups());
   fpga::SplitConvArgs conv_arg = {0};
-  fpga::fill_split_arg(&conv_arg, input, out, filter, relu_enabled,
-                       param->Groups(), param->Strides()[0],
-                       param->Strides()[1], param->Paddings()[0],
-                       param->Paddings()[1], bs_ptr);
+  fpga::fill_split_arg(&conv_arg, input, out, filter, activation_enable,
+                       leaky_relu_negative_slope, param->Groups(),
+                       param->Strides()[0], param->Strides()[1],
+                       param->Paddings()[0], param->Paddings()[1], bs_ptr);
   param->SetFpgaArgs(conv_arg);
   return true;
 }
diff --git a/src/operators/kernel/fpga/V1/conv_add_relu_kernel.cpp b/src/operators/kernel/fpga/V1/conv_add_relu_kernel.cpp
old mode 100755
new mode 100644
index 6b2a2d77c0df29b4c319061776491b0583157d6f..eef35bf74b6b28e3ec0c49d6b7ace0a350f3f194
--- a/src/operators/kernel/fpga/V1/conv_add_relu_kernel.cpp
+++ b/src/operators/kernel/fpga/V1/conv_add_relu_kernel.cpp
@@ -21,7 +21,10 @@ namespace operators {
 
 template <>
 bool ConvAddReluKernel<FPGA, float>::Init(FusionConvAddReluParam<FPGA> *param) {
-  bool relu_enabled = true;
+  // bool relu_enabled = true;
+  paddle_mobile::fpga::ActivationType activation_enable =
+      paddle_mobile::fpga::LEAKYRELU;
+  int16_t leaky_relu_negative_slope = 0;
   auto input = const_cast<Tensor *>(param->Input());
   const Tensor *bias = param->Bias();
   auto bias_ptr = bias->data<float>();
@@ -40,10 +43,10 @@ bool ConvAddReluKernel<FPGA, float>::Init(FusionConvAddReluParam<FPGA> *param) {
 
   fpga::format_conv_data(filter, out, &bs_ptr, param->Groups());
   fpga::SplitConvArgs conv_arg = {0};
-  fpga::fill_split_arg(&conv_arg, input, out, filter, relu_enabled,
-                       param->Groups(), param->Strides()[0],
-                       param->Strides()[1], param->Paddings()[0],
-                       param->Paddings()[1], bs_ptr);
+  fpga::fill_split_arg(&conv_arg, input, out, filter, activation_enable,
+                       leaky_relu_negative_slope, param->Groups(),
+                       param->Strides()[0], param->Strides()[1],
+                       param->Paddings()[0], param->Paddings()[1], bs_ptr);
   param->SetFpgaArgs(conv_arg);
   return true;
 }
diff --git a/src/operators/kernel/fpga/V1/conv_bn_kernel.cpp b/src/operators/kernel/fpga/V1/conv_bn_kernel.cpp
index 492d418b9023a3c4c802da099a5da5ebf5568649..c4c2bf184d536ace31e52defb59e97c154386464 100644
--- a/src/operators/kernel/fpga/V1/conv_bn_kernel.cpp
+++ b/src/operators/kernel/fpga/V1/conv_bn_kernel.cpp
@@ -22,7 +22,10 @@ namespace operators {
 
 template <>
 bool ConvBNKernel<FPGA, float>::Init(FusionConvBNParam<FPGA> *param) {
-  bool relu_enabled = false;
+  // bool relu_enabled = false;
+  paddle_mobile::fpga::ActivationType activation_enable =
+      paddle_mobile::fpga::NONE;
+  int16_t leaky_relu_negative_slope = 0;
   auto input = const_cast<Tensor *>(param->Input());
   auto filter = const_cast<Tensor *>(param->Filter());
   auto out = param->Output();
@@ -53,10 +56,10 @@ bool ConvBNKernel<FPGA, float>::Init(FusionConvBNParam<FPGA> *param) {
 
   fpga::format_conv_data(filter, out, &bs_ptr, param->Groups());
   fpga::SplitConvArgs conv_arg = {0};
-  fpga::fill_split_arg(&conv_arg, input, out, filter, relu_enabled,
-                       param->Groups(), param->Strides()[0],
-                       param->Strides()[1], param->Paddings()[0],
-                       param->Paddings()[1], bs_ptr);
+  fpga::fill_split_arg(&conv_arg, input, out, filter, activation_enable,
+                       leaky_relu_negative_slope, param->Groups(),
+                       param->Strides()[0], param->Strides()[1],
+                       param->Paddings()[0], param->Paddings()[1], bs_ptr);
   param->SetFpgaArgs(conv_arg);
   return true;
 }
diff --git a/src/operators/kernel/fpga/V1/conv_bn_relu_kernel.cpp b/src/operators/kernel/fpga/V1/conv_bn_relu_kernel.cpp
index 337b25ffa5d3ba00cd60935f8643213cb5ea70d3..463c90d1bb0dcd48a7b41aff73b830d14f989c73 100644
--- a/src/operators/kernel/fpga/V1/conv_bn_relu_kernel.cpp
+++ b/src/operators/kernel/fpga/V1/conv_bn_relu_kernel.cpp
@@ -22,7 +22,10 @@ namespace operators {
 
 template <>
 bool ConvBNReluKernel<FPGA, float>::Init(FusionConvBNReluParam<FPGA> *param) {
-  bool relu_enabled = true;
+  // bool relu_enabled = true;
+  paddle_mobile::fpga::ActivationType activation_enable =
+      paddle_mobile::fpga::LEAKYRELU;
+  int16_t leaky_relu_negative_slope = 0;
   auto input = const_cast<Tensor *>(param->Input());
   auto filter = const_cast<Tensor *>(param->Filter());
   auto out = param->Output();
@@ -53,10 +56,10 @@ bool ConvBNReluKernel<FPGA, float>::Init(FusionConvBNReluParam<FPGA> *param) {
 
   fpga::format_conv_data(filter, out, &bs_ptr, param->Groups());
   fpga::SplitConvArgs conv_arg = {0};
-  fpga::fill_split_arg(&conv_arg, input, out, filter, relu_enabled,
-                       param->Groups(), param->Strides()[0],
-                       param->Strides()[1], param->Paddings()[0],
-                       param->Paddings()[1], bs_ptr);
+  fpga::fill_split_arg(&conv_arg, input, out, filter, activation_enable,
+                       leaky_relu_negative_slope, param->Groups(),
+                       param->Strides()[0], param->Strides()[1],
+                       param->Paddings()[0], param->Paddings()[1], bs_ptr);
   param->SetFpgaArgs(conv_arg);
   return true;
 }
diff --git a/src/operators/kernel/fpga/V1/deconv_add_kernel.cpp b/src/operators/kernel/fpga/V1/deconv_add_kernel.cpp
index 1e21d374cb3651e582f43b2875a9c302ae86cdfb..97a4d5516b52939a3a1d90a22c8050679810d405 100644
--- a/src/operators/kernel/fpga/V1/deconv_add_kernel.cpp
+++ b/src/operators/kernel/fpga/V1/deconv_add_kernel.cpp
@@ -23,7 +23,10 @@ namespace operators {
 
 template <>
 bool DeconvAddKernel<FPGA, float>::Init(FusionDeconvAddParam<FPGA> *param) {
-  bool relu_enabled = false;
+  // bool relu_enabled = false;
+  paddle_mobile::fpga::ActivationType activation_enable =
+      paddle_mobile::fpga::NONE;
+  int16_t leaky_relu_negative_slope = 0;
   auto input = const_cast<Tensor *>(param->Input());
   const Tensor *bias = param->Bias();
   auto bias_ptr = bias->data<float>();
@@ -53,17 +56,18 @@ bool DeconvAddKernel<FPGA, float>::Init(FusionDeconvAddParam<FPGA> *param) {
     fpga::format_DWDeconv_data(filter, out, &bs_ptr, param->Groups(),
                                sub_conv_n);
     fpga::DWDeconvArgs DWDeconv_arg = {0};
-    fpga::fill_DWDeconv_arg(&DWDeconv_arg, input, out, filter, relu_enabled,
+    fpga::fill_DWDeconv_arg(&DWDeconv_arg, input, out, filter,
+                            activation_enable, leaky_relu_negative_slope,
                             param->Strides()[0], param->Strides()[1],
                             param->Paddings()[0], param->Paddings()[1], bs_ptr);
     param->SetFpgaArgs(DWDeconv_arg);
   } else {
     fpga::format_deconv_data(filter, out, &bs_ptr, param->Groups(), sub_conv_n);
     fpga::DeconvArgs deconv_arg = {0};
-    fpga::fill_deconv_arg(&deconv_arg, input, out, filter, relu_enabled,
-                          param->Groups(), param->Strides()[0],
-                          param->Strides()[1], param->Paddings()[0],
-                          param->Paddings()[1], bs_ptr);
+    fpga::fill_deconv_arg(&deconv_arg, input, out, filter, activation_enable,
+                          leaky_relu_negative_slope, param->Groups(),
+                          param->Strides()[0], param->Strides()[1],
+                          param->Paddings()[0], param->Paddings()[1], bs_ptr);
     param->SetFpgaArgs(deconv_arg);
   }
 
diff --git a/src/operators/kernel/fpga/V1/deconv_add_relu_kernel.cpp b/src/operators/kernel/fpga/V1/deconv_add_relu_kernel.cpp
index ca77b2fd440fdfcfb61498205739b3ded6346ebc..f0b29943d7731d716a19cff1e3cfc904d7610c0b 100644
--- a/src/operators/kernel/fpga/V1/deconv_add_relu_kernel.cpp
+++ b/src/operators/kernel/fpga/V1/deconv_add_relu_kernel.cpp
@@ -24,7 +24,10 @@ namespace operators {
 template <>
 bool DeconvAddReluKernel<FPGA, float>::Init(
     FusionDeconvAddReluParam<FPGA> *param) {
-  bool relu_enabled = true;
+  // bool relu_enabled = true;
+  paddle_mobile::fpga::ActivationType activation_enable =
+      paddle_mobile::fpga::LEAKYRELU;
+  int16_t leaky_relu_negative_slope = 0;
   auto input = const_cast<Tensor *>(param->Input());
   const Tensor *bias = param->Bias();
   auto bias_ptr = bias->data<float>();
@@ -54,17 +57,18 @@ bool DeconvAddReluKernel<FPGA, float>::Init(
     fpga::format_DWDeconv_data(filter, out, &bs_ptr, param->Groups(),
                                sub_conv_n);
     fpga::DWDeconvArgs DWDeconv_arg = {0};
-    fpga::fill_DWDeconv_arg(&DWDeconv_arg, input, out, filter, relu_enabled,
+    fpga::fill_DWDeconv_arg(&DWDeconv_arg, input, out, filter,
+                            activation_enable, leaky_relu_negative_slope,
                             param->Strides()[0], param->Strides()[1],
                             param->Paddings()[0], param->Paddings()[1], bs_ptr);
     param->SetFpgaArgs(DWDeconv_arg);
   } else {
     fpga::format_deconv_data(filter, out, &bs_ptr, param->Groups(), sub_conv_n);
     fpga::DeconvArgs deconv_arg = {0};
-    fpga::fill_deconv_arg(&deconv_arg, input, out, filter, relu_enabled,
-                          param->Groups(), param->Strides()[0],
-                          param->Strides()[1], param->Paddings()[0],
-                          param->Paddings()[1], bs_ptr);
+    fpga::fill_deconv_arg(&deconv_arg, input, out, filter, activation_enable,
+                          leaky_relu_negative_slope, param->Groups(),
+                          param->Strides()[0], param->Strides()[1],
+                          param->Paddings()[0], param->Paddings()[1], bs_ptr);
     param->SetFpgaArgs(deconv_arg);
   }
   return true;
diff --git a/src/operators/kernel/fpga/V1/elementwise_add_kernel.cpp b/src/operators/kernel/fpga/V1/elementwise_add_kernel.cpp
index be773412f099410b02f24b1d38d2a44d6ca77689..27eee7e5ba7045473ff035f45236d04e080a692e 100644
--- a/src/operators/kernel/fpga/V1/elementwise_add_kernel.cpp
+++ b/src/operators/kernel/fpga/V1/elementwise_add_kernel.cpp
@@ -20,7 +20,10 @@ namespace operators {
 
 template <>
 bool ElementwiseAddKernel<FPGA, float>::Init(ElementwiseAddParam<FPGA> *param) {
-  bool relu_enabled = false;
+  // bool relu_enabled = false;
+  paddle_mobile::fpga::ActivationType activation_enable =
+      paddle_mobile::fpga::NONE;
+  int16_t leaky_relu_negative_slope = 0;
   auto *input_x = const_cast<LoDTensor *>(param->InputX());
   auto *input_y = const_cast<LoDTensor *>(param->InputY());
   auto *out = param->Out();
@@ -30,7 +33,10 @@ bool ElementwiseAddKernel<FPGA, float>::Init(ElementwiseAddParam<FPGA> *param) {
   auto out_ptr = out->mutable_data<float>();
 
   fpga::EWAddArgs ewaddArgs = {0};
-  ewaddArgs.relu_enabled = relu_enabled;
+  // ewaddArgs.relu_enabled = relu_enabled;
+  ewaddArgs.output.activation.activation_type = activation_enable;
+  ewaddArgs.output.activation.leaky_relu_negative_slope =
+      leaky_relu_negative_slope;
   ewaddArgs.const0 = 0x3c00;  // =1
   ewaddArgs.const1 = 0x3c00;  // =1
   ewaddArgs.image0.address = input_x_ptr;
diff --git a/src/operators/kernel/fpga/V1/elementwise_add_relu_kernel.cpp b/src/operators/kernel/fpga/V1/elementwise_add_relu_kernel.cpp
index 541bb6126509dc7da59fa6bed5c46aff3442928b..fbbe679d4b6a6d4b0ca0a25ebb7aacf93a133943 100644
--- a/src/operators/kernel/fpga/V1/elementwise_add_relu_kernel.cpp
+++ b/src/operators/kernel/fpga/V1/elementwise_add_relu_kernel.cpp
@@ -21,7 +21,10 @@ namespace operators {
 template <>
 bool ElementwiseAddReluKernel<FPGA, float>::Init(
     ElementwiseAddReluParam<FPGA> *param) {
-  bool relu_enabled = true;
+  // bool relu_enabled = true;
+  paddle_mobile::fpga::ActivationType activation_enable =
+      paddle_mobile::fpga::LEAKYRELU;
+  int16_t leaky_relu_negative_slope = 0;
   auto *input_x = const_cast<LoDTensor *>(param->InputX());
   auto *input_y = const_cast<LoDTensor *>(param->InputY());
   auto *out = param->Out();
@@ -31,7 +34,10 @@ bool ElementwiseAddReluKernel<FPGA, float>::Init(
   auto out_ptr = out->mutable_data<float>();
 
   fpga::EWAddArgs ewaddArgs = {0};
-  ewaddArgs.relu_enabled = relu_enabled;
+  // ewaddArgs.relu_enabled = relu_enabled;
+  ewaddArgs.output.activation.activation_type = activation_enable;
+  ewaddArgs.output.activation.leaky_relu_negative_slope =
+      leaky_relu_negative_slope;
   ewaddArgs.const0 = 0x3c00;  // =1
   ewaddArgs.const1 = 0x3c00;  // =1
   ewaddArgs.image0.address = input_x_ptr;
diff --git a/src/operators/kernel/fpga/V1/fetch_kernel.cpp b/src/operators/kernel/fpga/V1/fetch_kernel.cpp
index e6e4591168b90cbe19b207cd9e77eaf5cd07de80..c00bdf57a259e24669c33f011d7b77eb20d4b308 100644
--- a/src/operators/kernel/fpga/V1/fetch_kernel.cpp
+++ b/src/operators/kernel/fpga/V1/fetch_kernel.cpp
@@ -19,12 +19,34 @@ namespace operators {
 
 template <>
 bool FetchKernel<FPGA, float>::Init(FetchParam<FPGA> *param) {
+  Tensor *output = param->Out();
+  // fpga::format_fp16_ofm(output);
   return true;
 }
 
 template <>
 void FetchKernel<FPGA, float>::Compute(const FetchParam<FPGA> &param) {
   param.Out()->ShareDataWith(*(param.InputX()));
+  /*auto input =
+          reinterpret_cast<Tensor *>(const_cast<Tensor *>(param.InputX()));
+  fpga::format_image(input);
+  auto input_ptr = input->data<float>();
+  Tensor *output = param.Out();
+  auto output_ptr = output->data<float>();
+
+  fpga::BypassArgs args = {fpga::DATA_TYPE_FP16};
+
+  args.input_data_type = fpga::DATA_TYPE_FP16;
+  args.output_data_type = fpga::DATA_TYPE_FP32;
+  args.input_layout_type = fpga::LAYOUT_CHW;
+  args.output_layout_type = fpga::LAYOUT_HWC;
+  args.image.address = reinterpret_cast<void *>(input_ptr);
+  args.image.channels = (uint32_t)input->dims()[1];
+  args.image.height = (input->dims().size() == 4) ? (uint32_t)input->dims()[2] :
+  1; args.image.width = (input->dims().size() == 4) ? (uint32_t)input->dims()[3]
+  : 1; args.image.pad_height = 0; args.image.pad_width = 0; args.output.address
+  = output_ptr; args.output.scale_address = output->scale;
+  fpga::PerformBypass(args);*/
 }
 
 template class FetchKernel<FPGA, float>;
diff --git a/src/operators/kernel/fpga/V1/fusion_fc_kernel.cpp b/src/operators/kernel/fpga/V1/fusion_fc_kernel.cpp
index 9258fb90e1e6bf9a597a387843ce781858628139..fadeae324ff8f5160bc5ff410c2e02b09539a01e 100644
--- a/src/operators/kernel/fpga/V1/fusion_fc_kernel.cpp
+++ b/src/operators/kernel/fpga/V1/fusion_fc_kernel.cpp
@@ -20,7 +20,10 @@ namespace operators {
 
 template <>
 bool FusionFcKernel<FPGA, float>::Init(FusionFcParam<FPGA> *param) {
-  bool relu_enabled = false;
+  // bool relu_enabled = false;
+  paddle_mobile::fpga::ActivationType activation_enable =
+      paddle_mobile::fpga::NONE;
+  int16_t leaky_relu_negative_slope = 0;
   auto input_x = const_cast<LoDTensor *>(param->InputX());
   auto filter = const_cast<Tensor *>(param->InputY());
   const Tensor *input_z = param->InputZ();
@@ -55,8 +58,8 @@ bool FusionFcKernel<FPGA, float>::Init(FusionFcParam<FPGA> *param) {
   fpga::format_fp16_ofm(out);
 
   fpga::SplitConvArgs conv_arg = {0};
-  fpga::fill_split_arg(&conv_arg, input_x, out, filter, relu_enabled, 1, 1, 1,
-                       0, 0, bs_ptr);
+  fpga::fill_split_arg(&conv_arg, input_x, out, filter, activation_enable,
+                       leaky_relu_negative_slope, 1, 1, 1, 0, 0, bs_ptr);
   param->SetFpgaArgs(conv_arg);
   return true;
 }
diff --git a/src/operators/kernel/fpga/V1/reshape_kernel.cpp b/src/operators/kernel/fpga/V1/reshape_kernel.cpp
index f5495e6d005f7f7c14ebd3d290ea9be02b9f0951..5e01bb74bab6996ca59632ae31f37ecfeafc918c 100644
--- a/src/operators/kernel/fpga/V1/reshape_kernel.cpp
+++ b/src/operators/kernel/fpga/V1/reshape_kernel.cpp
@@ -22,6 +22,12 @@ namespace operators {
 template <>
 bool ReshapeKernel<FPGA, float>::Init(ReshapeParam<FPGA> *param) {
   param->Out()->ShareDataWith(*param->InputX());
+  const int in_n = param->InputX()->dims()[0];
+  const int in_c = param->InputX()->dims()[1];
+  const int in_h = param->InputX()->dims()[2];
+  const int in_w = param->InputX()->dims()[3];
+  auto out = param->Out();
+  out->Resize(framework::make_ddim({in_n, in_c * in_h * in_w}));
   return true;
 }
 
diff --git a/src/operators/kernel/fpga/V1/sigmoid_kernel.cpp b/src/operators/kernel/fpga/V1/sigmoid_kernel.cpp
index 276a8fef62edfabfabb116fada145eedbf23ffa3..6c836e2776891f283677287eae54019f0dbef39b 100644
--- a/src/operators/kernel/fpga/V1/sigmoid_kernel.cpp
+++ b/src/operators/kernel/fpga/V1/sigmoid_kernel.cpp
@@ -15,73 +15,41 @@ limitations under the License. */
 #ifdef SIGMOID_OP
 
 #include "operators/kernel/activation_kernel.h"
+
 namespace paddle_mobile {
 namespace operators {
 
-using framework::DDim;
-using framework::Tensor;
-
 template <>
 bool SigmoidKernel<FPGA, float>::Init(SigmoidParam<FPGA> *param) {
+  paddle_mobile::fpga::ActivationType activation_enable =
+      paddle_mobile::fpga::SIGMOID;
+  int16_t leaky_relu_negative_slope = 0;
   auto input = const_cast<Tensor *>(param->InputX());
   auto input_ptr = input->data<float>();
   auto out = param->Out();
-  fpga::format_fp32_ofm(out);
+  fpga::format_fp16_ofm(out);
 
-  auto float_input = new Tensor;
-  if (input->dims().size() == 2) {
-    float_input->mutable_data<float>({1, input->dims()[1]});
-  } else if (input->dims().size() == 4) {
-    float_input->mutable_data<float>(
-        {1, input->dims()[2], input->dims()[3], input->dims()[1]});
-  } else {
-    DLOG << "wrong dimension of softmax input";
-  }
-
-  fpga::format_fp32_ofm(float_input);
   fpga::BypassArgs args = {fpga::DATA_TYPE_FP16};
-  args.input_layout_type = fpga::LAYOUT_HWC;
-  args.output_layout_type = fpga::LAYOUT_CHW;
   args.input_data_type = fpga::DATA_TYPE_FP16;
-  args.output_data_type = fpga::DATA_TYPE_FP32;
+  args.output_data_type = fpga::DATA_TYPE_FP16;
   args.image.address = input_ptr;
   args.image.height =
       (input->dims().size() == 4) ? (uint32_t)input->dims()[2] : 1;
   args.image.width =
       (input->dims().size() == 4) ? (uint32_t)input->dims()[3] : 1;
   args.image.channels = (uint32_t)input->dims()[1];
-  args.output.address = float_input->data<float>();
-  args.output.scale_address = float_input->scale;
-  param->SetFloatInput(float_input);
+  args.output.address = out->data<float>();
+  args.output.scale_address = out->scale;
+  args.output.activation.activation_type = activation_enable;
+  args.output.activation.leaky_relu_negative_slope = leaky_relu_negative_slope;
   param->SetFpgaArgs(args);
-
   return true;
 }
-template <typename T>
-T Sigmoid(const T a) {
-  T tmp = -1.0f * a;
-  return (1.0 / (1.0 + exp(tmp)));
-}
-template <typename T>
-void sigmoidFuntor(Tensor *input, Tensor *output) {
-  auto *input_ptr = input->data<T>();
-  auto *output_ptr = output->mutable_data<T>();
-  for (int i = 0; i < input->numel(); i++) {
-    *(output_ptr + i) = Sigmoid<T>(*(input_ptr + i));
-  }
-}
 template <>
 void SigmoidKernel<FPGA, float>::Compute(const SigmoidParam<FPGA> &param) {
-  Tensor *in_x = param.FloatInput();
-  Tensor *out = param.Out();
-
   fpga::PerformBypass(param.FpgaArgs());
-  fpga::fpga_invalidate((void *)in_x->data<float>(),  // NOLINT
-                        in_x->numel() * sizeof(float));
-  // TODO: In general case, 0 should be squeezed before softmax input  // NOLINT
-  sigmoidFuntor<float>(in_x, out);
-  fpga::fpga_flush(out->data<float>(), out->memory_size());
 }
+
 }  // namespace operators
 }  // namespace paddle_mobile
 
diff --git a/src/operators/kernel/fpga/V1/softmax_kernel.cpp b/src/operators/kernel/fpga/V1/softmax_kernel.cpp
index e5ada795b120c1438688089078be20e03f078cbb..2698fdece49409aec017112e8613a706c248cf48 100644
--- a/src/operators/kernel/fpga/V1/softmax_kernel.cpp
+++ b/src/operators/kernel/fpga/V1/softmax_kernel.cpp
@@ -26,7 +26,6 @@ bool SoftmaxKernel<FPGA, float>::Init(SoftmaxParam<FPGA> *param) {
   auto input_ptr = input->data<float>();
   auto out = param->Out();
   fpga::format_fp32_ofm(out);
-
   auto float_input = new Tensor;
   if (input->dims().size() == 2) {
     float_input->mutable_data<float>({1, input->dims()[1]});
@@ -36,7 +35,6 @@ bool SoftmaxKernel<FPGA, float>::Init(SoftmaxParam<FPGA> *param) {
   } else {
     DLOG << "wrong dimension of softmax input";
   }
-
   fpga::format_fp32_ofm(float_input);
   fpga::BypassArgs args = {fpga::DATA_TYPE_FP16};
   args.input_layout_type = fpga::LAYOUT_HWC;
@@ -53,6 +51,7 @@ bool SoftmaxKernel<FPGA, float>::Init(SoftmaxParam<FPGA> *param) {
   args.output.scale_address = float_input->scale;
   param->SetFloatInput(float_input);
   param->SetFpgaArgs(args);
+
   return true;
 }
 
diff --git a/src/operators/op_param.h b/src/operators/op_param.h
index d90dff2d7e919f736b5cfd0531074944938f2a8a..e3da6724a8c33501f50bc463ee25a88166f4351b 100644
--- a/src/operators/op_param.h
+++ b/src/operators/op_param.h
@@ -1081,14 +1081,9 @@ class SigmoidParam : public OpParam {
 #ifdef PADDLE_MOBILE_FPGA
 
  private:
-  std::shared_ptr<RType> float_input_x_;
   fpga::BypassArgs fpga_bypass_args;
 
  public:
-  RType *FloatInput() const {
-    return float_input_x_ == nullptr ? input_x_ : float_input_x_.get();
-  }
-  void SetFloatInput(Tensor *input) { float_input_x_.reset(input); }
   const fpga::BypassArgs &FpgaArgs() const { return fpga_bypass_args; }
   void SetFpgaArgs(const fpga::BypassArgs &args) { fpga_bypass_args = args; }
 #endif
@@ -1214,6 +1209,20 @@ class FetchParam : public OpParam {
  private:
   RType *input_x_;
   Tensor *out_;
+#ifdef PADDLE_MOBILE_FPGA
+
+ private:
+  std::shared_ptr<RType> float_input_x_;
+  fpga::BypassArgs fpga_bypass_args;
+
+ public:
+  RType *FloatInput() const {
+    return float_input_x_ == nullptr ? input_x_ : float_input_x_.get();
+  }
+  void SetFloatInput(Tensor *input) { float_input_x_.reset(input); }
+  const fpga::BypassArgs &FpgaArgs() const { return fpga_bypass_args; }
+  void SetFpgaArgs(const fpga::BypassArgs &args) { fpga_bypass_args = args; }
+#endif
 };
 
 #ifdef FILL_CONSTANT_OP