diff --git a/paddle/fluid/inference/analysis/subgraph_splitter.cc b/paddle/fluid/inference/analysis/subgraph_splitter.cc index 773fceeeb2e0f353a901eba84e30dc0e6f12fe99..c3a2dbf9d18f80f93381d485a3870e43411bd992 100644 --- a/paddle/fluid/inference/analysis/subgraph_splitter.cc +++ b/paddle/fluid/inference/analysis/subgraph_splitter.cc @@ -85,6 +85,14 @@ struct BriefNode { std::vector 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 &node_map, int src_id, int dst_id) { // merge the two adjacent nodes into one node. @@ -224,8 +232,8 @@ std::vector> 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. diff --git a/paddle/fluid/inference/tensorrt/convert/activation_op.cc b/paddle/fluid/inference/tensorrt/convert/activation_op.cc index e1cace9cc1b06f036f52e82b7b86c99a02d50f50..8168cdff1b85fc05d22fbec7fac6ab8892f3a907 100644 --- a/paddle/fluid/inference/tensorrt/convert/activation_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/activation_op.cc @@ -35,6 +35,8 @@ class ReluOpConverter : public OpConverter { engine_, Activation, *const_cast(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. diff --git a/paddle/fluid/inference/tensorrt/convert/batch_norm_op.cc b/paddle/fluid/inference/tensorrt/convert/batch_norm_op.cc index 94f8b0ae5606d39a722ffe28501645c9b6fc5d2e..3330af2da6c97ad153dcecd86be4b441eac62b5e 100644 --- a/paddle/fluid/inference/tensorrt/convert/batch_norm_op.cc +++ b/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()] = diff --git a/paddle/fluid/inference/tensorrt/convert/concat_op.cc b/paddle/fluid/inference/tensorrt/convert/concat_op.cc index bb9627bf957b63993b2c8d23e7ec8122eb004eaf..2983e91cb2f209c2b086b5422cb1c90f3dfe9940 100644 --- a/paddle/fluid/inference/tensorrt/convert/concat_op.cc +++ b/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 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(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. diff --git a/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc b/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc index 841a95db38ce7cf0cb5961ff04cb569ee2633e6f..022e43a571edea1ccaaefbdbaffcfc3f4b23c3a0 100644 --- a/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc +++ b/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); diff --git a/paddle/fluid/inference/tensorrt/convert/elementwise_op.cc b/paddle/fluid/inference/tensorrt/convert/elementwise_op.cc index 60a72b4eb5c75b5cd12305f13763a9a1a567213f..0a6ce568f194f03c7259e1ebf28dd6ce4df2d594 100644 --- a/paddle/fluid/inference/tensorrt/convert/elementwise_op.cc +++ b/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(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. diff --git a/paddle/fluid/inference/tensorrt/convert/fc_op.cc b/paddle/fluid/inference/tensorrt/convert/fc_op.cc index ad98d85aae9cf594922aca00c43718ccfbce2278..7c21ecd95da07b498eed2ab1bbdcc0e8cd184787 100644 --- a/paddle/fluid/inference/tensorrt/convert/fc_op.cc +++ b/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) { diff --git a/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc b/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc index 73f1b28ddf73403862e55d102a259d7b6cf67b1f..f9bb66a6e9f81a10368db7710108c319860e940a 100644 --- a/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc +++ b/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); diff --git a/paddle/fluid/operators/tensorrt_engine_op.h b/paddle/fluid/operators/tensorrt_engine_op.h index 395d8bcc07955295b06174fa195cb95130aaadb5..79e75ea9a035b654f0bb7026d3a491bebe0b23c4 100644 --- a/paddle/fluid/operators/tensorrt_engine_op.h +++ b/paddle/fluid/operators/tensorrt_engine_op.h @@ -161,20 +161,6 @@ class TensorRTEngineKernel : public framework::OpKernel { boost::get(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(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; }