Merge pull request #1438 from zhangyang0701/develop

Avoid memory leakage for FPGA track close #1372

src/fpga/V1/api.cpp

@@ -623,7 +623,7 @@ void fill_deconv_arg(struct DeconvArgs *arg, framework::Tensor *input,
   fpga::format_fp16_ofm(out, dims_out_new);
   auto out_ptr = out->data<half>();
   arg->output.address =
-      (half *)out_ptr +  // NOLINT
+      out_ptr +
       omit_size * sizeof(half) *
           (align_to_x(real_out_width * arg->filter_num, IMAGE_ALIGNMENT));
   arg->output.scale_address = out->scale;
@@ -713,7 +713,6 @@ 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].output.activation.activation_type =
           activation_enable;
       arg->split_conv_args[i]
@@ -759,8 +758,8 @@ void fill_deconv_arg(struct DeconvArgs *arg, framework::Tensor *input,
           align_to_x(arg->split_conv_args[i]->conv_arg[j].filter_num,
                      FILTER_NUM_ALIGNMENT) *
           sizeof(int8_t);
-      auto filter_head = &((
-          int8_t *)filter_ptr)[j * element_num * filter_num_per_div +  // NOLINT
+      auto filter_head =
+          &filter_ptr[j * element_num * filter_num_per_div +  // NOLINT
                       i * filter_sub_conv_offset];
       arg->split_conv_args[i]->conv_arg[j].filter_address =
           fpga_malloc(filter_size);
@@ -836,6 +835,10 @@ void fill_dwconv_arg(struct DWconvArgs *arg, framework::Tensor *input,
                      int16_t leaky_relu_negative_slope, int stride_h,
                      int stride_w, int padding_h, int padding_w,
                      float *bias_ptr) {
+  auto deleter = [](void *p) { fpga_free(p); };
+  arg->vector_dwconv_space.push_back(
+      std::shared_ptr<char>(reinterpret_cast<char *>(bias_ptr), deleter));
+
   auto filter_ptr = filter->data<uint8_t>();
   auto input_ptr = input->data<half>();
   auto output_ptr = out->mutable_data<half>();

src/fpga/V1/deconv_filter.cpp

@@ -247,6 +247,7 @@ void deconv_format_filter(float** data_in, int num, int channel, int height,
     fpga_copy(ptr_space + i * align_offset, ptr_tmp, align_offset);
     fpga_free(ptr_tmp);
   }
+  fpga_free(ptr_ptr_data);
   *data_in = reinterpret_cast<float*>(ptr_space);
 
   /*    {

src/fpga/common/fpga_common.h

@@ -187,6 +187,7 @@ struct SplitArgs {
   uint32_t* out_channel_nums;
   uint32_t height;
   uint32_t width;
+  std::vector<std::shared_ptr<char>> vector_split_space;
 };
 
 struct PoolingArgs {
@@ -236,6 +237,7 @@ struct DWconvArgs {
   struct KernelArgs kernel;
   struct ImageInputArgs image;
   struct ImageOutputArgs output;
+  std::vector<std::shared_ptr<char>> vector_dwconv_space;
 };
 
 struct DWDeconvArgs {

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

@@ -59,8 +59,6 @@ bool ConvAddBNKernel<FPGA, float>::Init(FusionConvAddBNParam<FPGA> *param) {
     bs_ptr[i + channel] = new_scale_ptr[i];
     bs_ptr[i] = new_bias_ptr[i];
   }
-  param->SetNewScale(new_scale);
-  param->SetNewBias(new_bias);
 
   fpga::format_conv_data(filter, out, &bs_ptr, param->Groups());
   fpga::SplitConvArgs conv_arg = {0};
@@ -70,6 +68,9 @@ bool ConvAddBNKernel<FPGA, float>::Init(FusionConvAddBNParam<FPGA> *param) {
                        param->Paddings()[0], param->Paddings()[1], bs_ptr);
   param->SetFpgaArgs(conv_arg);
 
+  delete new_scale;
+  delete new_bias;
+
   return true;
 }
 

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

@@ -60,8 +60,6 @@ bool ConvAddBNReluKernel<FPGA, float>::Init(
     bs_ptr[i + channel] = new_scale_ptr[i];
     bs_ptr[i] = new_bias_ptr[i];
   }
-  param->SetNewScale(new_scale);
-  param->SetNewBias(new_bias);
 
   const int groups = param->Groups();
   if (groups == channel) {
@@ -71,6 +69,8 @@ bool ConvAddBNReluKernel<FPGA, float>::Init(
                           leaky_relu_negative_slope, strides[0], strides[1],
                           paddings[0], paddings[1], new_bias_ptr);
     param->SetFpgaArgs(dwconv_arg);
+    fpga::fpga_free(new_scale_ptr);
+    fpga::fpga_free(bs_ptr);
   } else {
     fpga::format_conv_data(filter, out, &bs_ptr, param->Groups());
     fpga::SplitConvArgs conv_arg = {0};
@@ -78,6 +78,8 @@ bool ConvAddBNReluKernel<FPGA, float>::Init(
                          leaky_relu_negative_slope, param->Groups(), strides[0],
                          strides[1], paddings[0], paddings[1], bs_ptr);
     param->SetFpgaArgs(conv_arg);
+    delete new_scale;
+    delete new_bias;
   }
   return true;
 }

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

@@ -51,8 +51,6 @@ bool ConvBNKernel<FPGA, float>::Init(FusionConvBNParam<FPGA> *param) {
     bs_ptr[i + channel] = new_scale_ptr[i];
     bs_ptr[i] = new_bias_ptr[i];
   }
-  param->SetNewScale(new_scale);
-  param->SetNewBias(new_bias);
 
   fpga::format_conv_data(filter, out, &bs_ptr, param->Groups());
   fpga::SplitConvArgs conv_arg = {0};
@@ -61,6 +59,8 @@ bool ConvBNKernel<FPGA, float>::Init(FusionConvBNParam<FPGA> *param) {
                        param->Strides()[0], param->Strides()[1],
                        param->Paddings()[0], param->Paddings()[1], bs_ptr);
   param->SetFpgaArgs(conv_arg);
+  delete new_scale;
+  delete new_bias;
   return true;
 }
 

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

@@ -51,8 +51,6 @@ bool ConvBNReluKernel<FPGA, float>::Init(FusionConvBNReluParam<FPGA> *param) {
     bs_ptr[i + channel] = new_scale_ptr[i];
     bs_ptr[i] = new_bias_ptr[i];
   }
-  param->SetNewScale(new_scale);
-  param->SetNewBias(new_bias);
 
   fpga::format_conv_data(filter, out, &bs_ptr, param->Groups());
   fpga::SplitConvArgs conv_arg = {0};
@@ -61,6 +59,9 @@ bool ConvBNReluKernel<FPGA, float>::Init(FusionConvBNReluParam<FPGA> *param) {
                        param->Strides()[0], param->Strides()[1],
                        param->Paddings()[0], param->Paddings()[1], bs_ptr);
   param->SetFpgaArgs(conv_arg);
+
+  delete new_scale;
+  delete new_bias;
   return true;
 }
 

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

@@ -43,6 +43,8 @@ bool SplitKernel<FPGA, float>::Init(SplitParam<FPGA> *param) {
     out_channels[i] = (uint32_t)sections[i];
   }
 
+  auto deleter = [](void *p) { fpga::fpga_free(p); };
+
   fpga::SplitArgs arg = {0};
   arg.image_num = image_num;
   arg.image_in = in->data<half>();
@@ -52,6 +54,12 @@ bool SplitKernel<FPGA, float>::Init(SplitParam<FPGA> *param) {
   arg.out_channel_nums = out_channels;
   arg.height = (uint32_t)in->dims()[2];
   arg.width = (uint32_t)in->dims()[3];
+  arg.vector_split_space.push_back(
+      std::shared_ptr<char>(reinterpret_cast<char *>(images_out), deleter));
+  arg.vector_split_space.push_back(
+      std::shared_ptr<char>(reinterpret_cast<char *>(scales_out), deleter));
+  arg.vector_split_space.push_back(
+      std::shared_ptr<char>(reinterpret_cast<char *>(out_channels), deleter));
 
   param->SetFpgaArgs(arg);
   return true;

test/fpga/test_rfcn.cpp

@@ -39,7 +39,7 @@ void readStream(std::string filename, uint8_t *buf) {
 }
 
 static const char *g_rfcn_combine = "../models/rfcn";
-const std::string g_image_src_float = "../models/rfcn/data.bin";
+static const char *g_image_src_float = "../models/rfcn/data.bin";
 int main() {
   paddle_mobile::fpga::open_device();
   paddle_mobile::PaddleMobile<paddle_mobile::FPGA> paddle_mobile;
@@ -55,6 +55,7 @@ int main() {
     paddle_mobile.Predict_To(-1);
     paddle_mobile.GetResults(&v);
     DLOG << "Computation done";
+    fpga::fpga_free(img);
   }
 
   return 0;