fix comments and set name for trt layer and ITensor

Author: NHZlX

paddle/fluid/inference/analysis/subgraph_splitter.cc

@@ -85,6 +85,14 @@ struct BriefNode {
   std::vector<BriefNode *> outlinks;
 };
 
+// Union two adjacent BriefNode.
+// Suppose we have two adjacent nodes src and dst.
+// We will perform the following operations:
+// 1. add all inputs(except src) of dst to src inlinks.
+// 2. add all outputs of dst to src outlinks.
+// 3. change all the dst's inputs and outputs
+// corresponding inlinks and outlinks to src node.
+// 4. delete all dst's inlinks and outlinks.
 void UnionContractedNodes(const std::unordered_map<int, BriefNode *> &node_map,
                           int src_id, int dst_id) {
   // merge the two adjacent nodes into one node.
@@ -224,8 +232,8 @@ std::vector<std::vector<Node *>> SubGraphSplitter::ExtractSubGraphs() {
     //  Our algorithm must guarantee that:
     //  1. The graph is always directed acyclic graph（DAG）.
     //  2. If there is a path in the subgraph from X to Y (X and Y are both
-    //  nodes
-    //     in the subgraph), then all paths from X to Y are in the subgraph.
+    //  nodes in the subgraph), then all paths from X to Y are in the
+    //  subgraph.
     //
     //  In order to achieve the above guarantee.
     //  For adjacent nodes src -> dst.

paddle/fluid/inference/tensorrt/convert/activation_op.cc

@@ -35,6 +35,8 @@ class ReluOpConverter : public OpConverter {
         engine_, Activation, *const_cast<nvinfer1::ITensor*>(input_tensor),
         nvinfer1::ActivationType::kRELU);
     auto output_name = op_desc.Output("Out")[0];
+    layer->setName(("relu (Output: " + output_name + ")").c_str());
+    layer->getOutput(0)->setName(output_name.c_str());
     engine_->SetITensor(output_name, layer->getOutput(0));
     if (test_mode) {  // the test framework can not determine which is the
                       // output, so place the declaration inside.

paddle/fluid/inference/tensorrt/convert/batch_norm_op.cc

@@ -116,6 +116,8 @@ class BatchNormOpConverter : public OpConverter {
                              scale_weights.get(), power_weights.get());
 
     auto output_name = op_desc.Output("Y").front();
+    layer->setName(("batch_norm (Output: " + output_name + ")").c_str());
+    layer->getOutput(0)->setName(output_name.c_str());
     engine_->weight_map[op_desc.Input("Bias").front()] =
         std::move(combile_bias_tensor);
     engine_->weight_map[op_desc.Input("Scale").front()] =

paddle/fluid/inference/tensorrt/convert/concat_op.cc

@@ -30,7 +30,9 @@ class ConcatOpConverter : public OpConverter {
     framework::OpDesc op_desc(op, nullptr);
     // Declare inputs
     std::vector<nvinfer1::ITensor*> itensors;
+    std::cout << "Concat op: " << std::endl;
     for (auto& input_name : op_desc.Input("X")) {
+      std::cout << input_name << std::endl;
       itensors.push_back(engine_->GetITensor(input_name));
     }
     int axis = boost::get<int>(op_desc.GetAttr("axis"));
@@ -42,6 +44,8 @@ class ConcatOpConverter : public OpConverter {
     axis = axis - 1;  // Remove batch dim
     layer->setAxis(axis);
     auto output_name = op_desc.Output("Out")[0];
+    layer->setName(("concat (Output: " + output_name + ")").c_str());
+    layer->getOutput(0)->setName(output_name.c_str());
     engine_->SetITensor(output_name, layer->getOutput(0));
     if (test_mode) {  // the test framework can not determine which is the
                       // output, so place the declaration inside.

paddle/fluid/inference/tensorrt/convert/conv2d_op.cc

@@ -26,6 +26,9 @@ class Conv2dOpConverter : public OpConverter {
         << "convert a fluid conv2d op to tensorrt conv layer without bias";
 
     framework::OpDesc op_desc(op, nullptr);
+    std::cout << "Conv op: " << std::endl;
+    std::cout << op_desc.Input("Input").front() << std::endl;
+    std::cout << op_desc.Output("Output").front() << std::endl;
     PADDLE_ENFORCE_EQ(op_desc.Input("Input").size(), 1);
     PADDLE_ENFORCE_EQ(op_desc.Input("Filter").size(), 1);  // Y is a weight
     PADDLE_ENFORCE_EQ(op_desc.Output("Output").size(), 1);
@@ -78,8 +81,10 @@ class Conv2dOpConverter : public OpConverter {
     layer->setNbGroups(groups);
 
     auto output_name = op_desc.Output("Output").front();
+    layer->setName(("conv2d (Output: " + output_name + ")").c_str());
     engine_->weight_map[op_desc.Input("Filter").front()] =
         std::move(weight_tensor);
+    layer->getOutput(0)->setName(output_name.c_str());
     engine_->SetITensor(output_name, layer->getOutput(0));
     if (test_mode) {
       engine_->DeclareOutput(output_name);

paddle/fluid/inference/tensorrt/convert/elementwise_op.cc

@@ -89,6 +89,8 @@ class ElementwiseWeightOpConverter : public OpConverter {
         shift_weights.get(), scale_weights.get(), power_weights.get());
     auto output_name = op_desc.Output("Out")[0];
 
+    layer->setName(("elementwise_add (Output: " + output_name + ")").c_str());
+    layer->getOutput(0)->setName(output_name.c_str());
     engine_->weight_map[op_desc.Input("Y").front()] = std::move(weight_tensor);
     engine_->SetITensor(output_name, layer->getOutput(0));
     if (test_mode) {  // the test framework can not determine which is the
@@ -137,6 +139,8 @@ class ElementwiseTensorOpConverter : public OpConverter {
         *const_cast<nvinfer1::ITensor*>(Y), op_pair->second);
 
     auto output_name = op_desc.Output("Out")[0];
+    layer->setName(("elementwise (Output: " + output_name + ")").c_str());
+    layer->getOutput(0)->setName(output_name.c_str());
     engine_->SetITensor(output_name, layer->getOutput(0));
     if (test_mode) {  // the test framework can not determine which is the
                       // output, so place the declaration inside.

paddle/fluid/inference/tensorrt/convert/fc_op.cc

@@ -107,6 +107,8 @@ class FcOpConverter : public OpConverter {
                                        n_output, tmp_weight.get(), bias.get());
 
     auto output_name = op_desc.Output("Out").front();
+    layer->setName(("fc (Output: " + output_name + ")").c_str());
+    layer->getOutput(0)->setName(output_name.c_str());
     engine_->SetITensor(output_name, layer->getOutput(0));
     engine_->weight_map[op_desc.Input("Y").front()] = std::move(tmp);
     if (test_mode) {

paddle/fluid/inference/tensorrt/convert/pool2d_op.cc

@@ -72,6 +72,8 @@ class Pool2dOpConverter : public OpConverter {
     layer->setPadding(nv_paddings);
 
     auto output_name = op_desc.Output("Out")[0];
+    layer->setName(("pool2d (Output: " + output_name + ")").c_str());
+    layer->getOutput(0)->setName(output_name.c_str());
     engine_->SetITensor(output_name, layer->getOutput(0));
     if (test_mode) {
       engine_->DeclareOutput(output_name);

paddle/fluid/operators/tensorrt_engine_op.h

@@ -161,20 +161,6 @@ class TensorRTEngineKernel : public framework::OpKernel<T> {
               boost::get<platform::CUDAPlace>(context.GetPlace()).device)),
           size * sizeof(float));
 
-      // TODO(zhaolong) : delete it sometimes
-      /* THIS CODE JUST FOR TEST
-      std::cout << output_maps[output_index] << std::endl;
-      platform::CPUPlace cpu_place;
-      framework::LoDTensor temp_tensor;
-      temp_tensor.Resize(framework::make_ddim(ddim));
-      auto* temp_data = temp_tensor.mutable_data<float>(cpu_place);
-
-      TensorCopySync(*fluid_t, cpu_place ,&temp_tensor);
-      for(int i = 0; i < size; i++) {
-        std::cout << temp_data[i] <<  " " ;
-      }
-      std::cout << std::endl;
-      */
       output_index += 1;
     }