Merge pull request #12753 from NHZlX/add_benchmark

modify tensorrt engine op from cpu mode to gpu 

paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.cc

@@ -23,7 +23,7 @@
 namespace paddle {
 namespace inference {
 
-DEFINE_int32(tensorrt_max_batchsize, 3, "TensorRT maximum batch size");
+DEFINE_int32(tensorrt_max_batchsize, 1, "TensorRT maximum batch size");
 DEFINE_int32(tensorrt_workspace_size, 2048, "TensorRT workspace size");
 
 namespace analysis {
@@ -52,7 +52,6 @@ bool DataFlowGraphToFluidPass::Initialize(Argument *argument) {
 bool DataFlowGraphToFluidPass::Finalize() { return true; }
 
 void DataFlowGraphToFluidPass::Run(DataFlowGraph *graph) {
-  FilterRedundantOutputOfSubGraph(graph);
   LOG(INFO) << "graph.inputs " << graph->inputs.size();
   for (auto &node : GraphTraits<DataFlowGraph>(graph).nodes_in_TS()) {
     if (node.deleted()) continue;

paddle/fluid/inference/analysis/subgraph_splitter.cc

@@ -153,6 +153,7 @@ void SubGraphFuse::ReplaceNodesWithSubGraphs() {
       inlink_or_outlink_cleaner(o->inlinks);
     }
   }
+  FilterRedundantOutputOfSubGraph(graph_);
 }
 
 }  // namespace analysis

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

@@ -35,12 +35,20 @@ class Conv2dOpConverter : public OpConverter {
     auto* Y_v = scope.FindVar(op_desc.Input("Filter").front());
     PADDLE_ENFORCE_NOT_NULL(Y_v);
     auto* Y_t = Y_v->GetMutable<framework::LoDTensor>();
-    auto* weight_data = Y_t->mutable_data<float>(platform::CPUPlace());
 
-    PADDLE_ENFORCE_EQ(Y_t->dims().size(), 4UL);
-    const int n_output = Y_t->dims()[0];
-    const int filter_h = Y_t->dims()[2];
-    const int filter_w = Y_t->dims()[3];
+    platform::CPUPlace cpu_place;
+    std::unique_ptr<framework::LoDTensor> weight_tensor(
+        new framework::LoDTensor());
+    weight_tensor->Resize(Y_t->dims());
+    TensorCopySync((*Y_t), cpu_place, weight_tensor.get());
+
+    auto* weight_data =
+        weight_tensor->mutable_data<float>(platform::CPUPlace());
+
+    PADDLE_ENFORCE_EQ(weight_tensor->dims().size(), 4UL);
+    const int n_output = weight_tensor->dims()[0];
+    const int filter_h = weight_tensor->dims()[2];
+    const int filter_w = weight_tensor->dims()[3];
 
     const int groups = boost::get<int>(op_desc.GetAttr("groups"));
     const std::vector<int> dilations =
@@ -57,7 +65,7 @@ class Conv2dOpConverter : public OpConverter {
 
     TensorRTEngine::Weight weight{nvinfer1::DataType::kFLOAT,
                                   static_cast<void*>(weight_data),
-                                  Y_t->memory_size() / sizeof(float)};
+                                  weight_tensor->memory_size() / sizeof(float)};
 
     TensorRTEngine::Weight bias{nvinfer1::DataType::kFLOAT, nullptr, 0};
     auto* layer = TRT_ENGINE_ADD_LAYER(
@@ -70,6 +78,8 @@ class Conv2dOpConverter : public OpConverter {
     layer->setNbGroups(groups);
 
     auto output_name = op_desc.Output("Output").front();
+    engine_->weight_map[op_desc.Input("Filter").front()] =
+        std::move(weight_tensor);
     engine_->SetITensor(output_name, layer->getOutput(0));
     if (test_mode) {
       engine_->DeclareOutput(output_name);

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

@@ -12,7 +12,6 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
 
 namespace paddle {
@@ -40,10 +39,17 @@ class ElementwiseWeightOpConverter : public OpConverter {
     auto* Y_v = scope.FindVar(op_desc.Input("Y").front());
     PADDLE_ENFORCE_NOT_NULL(Y_v);
     auto* Y_t = Y_v->GetMutable<framework::LoDTensor>();
-    auto* weight_data = Y_t->mutable_data<float>(platform::CPUPlace());
+
+    platform::CPUPlace cpu_place;
+    std::unique_ptr<framework::LoDTensor> weight_tensor(
+        new framework::LoDTensor());
+    weight_tensor->Resize(Y_t->dims());
+    TensorCopySync((*Y_t), cpu_place, weight_tensor.get());
+    auto* weight_data =
+        weight_tensor->mutable_data<float>(platform::CPUPlace());
     auto scale_mode = nvinfer1::ScaleMode::kELEMENTWISE;
 
-    std::vector<int> dims_y = framework::vectorize2int(Y_t->dims());
+    std::vector<int> dims_y = framework::vectorize2int(weight_tensor->dims());
     if (static_cast<int>(dims_y.size()) == dims_x.nbDims + 1) {
       if (dims_y[0] == 1) dims_y.erase(dims_y.begin());
     }
@@ -70,9 +76,9 @@ class ElementwiseWeightOpConverter : public OpConverter {
       PADDLE_THROW("TensorRT unsupported weight Shape for Elementwise op!");
     }
 
-    TensorRTEngine::Weight shift_weights{nvinfer1::DataType::kFLOAT,
-                                         static_cast<void*>(weight_data),
-                                         Y_t->memory_size() / sizeof(float)};
+    TensorRTEngine::Weight shift_weights{
+        nvinfer1::DataType::kFLOAT, static_cast<void*>(weight_data),
+        weight_tensor->memory_size() / sizeof(float)};
     TensorRTEngine::Weight scale_weights{nvinfer1::DataType::kFLOAT, nullptr,
                                          0};
     TensorRTEngine::Weight power_weights{nvinfer1::DataType::kFLOAT, nullptr,
@@ -82,6 +88,8 @@ class ElementwiseWeightOpConverter : public OpConverter {
         engine_, Scale, *const_cast<nvinfer1::ITensor*>(X), scale_mode,
         shift_weights.get(), scale_weights.get(), power_weights.get());
     auto output_name = op_desc.Output("Out")[0];
+
+    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
                       // output, so place the declaration inside.

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

@@ -12,12 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/fluid/framework/eigen.h"
-#include "paddle/fluid/framework/lod_tensor.h"
-#include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
-#include "paddle/fluid/inference/tensorrt/engine.h"
-#include "paddle/fluid/platform/place.h"
 
 namespace paddle {
 namespace inference {
@@ -73,19 +68,26 @@ class FcOpConverter : public OpConverter {
     auto* Y_t = Y_v->GetMutable<framework::LoDTensor>();
     // This may trigger a GPU->CPU copy, because TRT's weight can only be
     // assigned from CPU memory, that can't be avoided.
-    auto* weight_data = Y_t->mutable_data<float>(platform::CPUPlace());
-    PADDLE_ENFORCE_EQ(Y_t->dims().size(), 2UL);  // a matrix
-    size_t n_output = Y_t->dims()[1];
+    platform::CPUPlace cpu_place;
+    framework::LoDTensor weight_tensor;
+    weight_tensor.Resize(Y_t->dims());
+    TensorCopySync((*Y_t), cpu_place, &weight_tensor);
 
-    framework::LoDTensor tmp;
-    tmp.Resize(Y_t->dims());
-    memcpy(tmp.mutable_data<float>(platform::CPUPlace()), weight_data,
+    auto* weight_data = weight_tensor.mutable_data<float>(platform::CPUPlace());
+
+    PADDLE_ENFORCE_EQ(weight_tensor.dims().size(), 2UL);  // a matrix
+    size_t n_output = weight_tensor.dims()[1];
+
+    std::unique_ptr<framework::Tensor> tmp(new framework::LoDTensor());
+    tmp->Resize(weight_tensor.dims());
+
+    memcpy(tmp->mutable_data<float>(platform::CPUPlace()), weight_data,
            Y_t->dims()[0] * Y_t->dims()[1] * sizeof(float));
     TensorRTEngine::Weight weight{nvinfer1::DataType::kFLOAT,
                                   static_cast<void*>(weight_data),
                                   Y_t->memory_size() / sizeof(float)};
     TensorRTEngine::Weight tmp_weight(nvinfer1::DataType::kFLOAT,
-                                      static_cast<void*>(tmp.data<float>()),
+                                      static_cast<void*>(tmp->data<float>()),
                                       Y_t->memory_size() / sizeof(float));
     weight.dims.assign({Y_t->dims()[0], Y_t->dims()[1]});
     tmp_weight.dims = weight.dims;
@@ -106,6 +108,7 @@ class FcOpConverter : public OpConverter {
 
     auto output_name = op_desc.Output("Out").front();
     engine_->SetITensor(output_name, layer->getOutput(0));
+    engine_->weight_map[op_desc.Input("Y").front()] = std::move(tmp);
     if (test_mode) {
       engine_->DeclareOutput(output_name);
     }

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

@@ -57,6 +57,7 @@ TEST(OpConverter, ConvertBlock) {
   auto* x = scope.Var("conv2d-Y");
   auto* x_tensor = x->GetMutable<framework::LoDTensor>();
   x_tensor->Resize(framework::make_ddim(dim_vec));
+  x_tensor->mutable_data<float>(platform::CUDAPlace(0));
 
   OpConverter converter;
   converter.ConvertBlock(*block->Proto(), {"conv2d-Y"}, scope,

paddle/fluid/inference/tensorrt/convert/ut_helper.h

@@ -24,6 +24,7 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/tensor_util.h"
 #include "paddle/fluid/inference/analysis/helper.h"
 #include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
 #include "paddle/fluid/inference/tensorrt/engine.h"
@@ -48,11 +49,17 @@ void RandomizeTensor(framework::LoDTensor* tensor, const platform::Place& place,
   auto dims = tensor->dims();
   size_t num_elements = analysis::AccuDims(dims, dims.size());
   PADDLE_ENFORCE_GT(num_elements, 0);
-  auto* data = tensor->mutable_data<float>(place);
+
+  platform::CPUPlace cpu_place;
+  framework::LoDTensor temp_tensor;
+  temp_tensor.Resize(dims);
+  auto* temp_data = temp_tensor.mutable_data<float>(cpu_place);
 
   for (size_t i = 0; i < num_elements; i++) {
-    *(data + i) = random(0., 1.);
+    *(temp_data + i) = random(0., 1.);
   }
+
+  TensorCopySync(temp_tensor, place, tensor);
 }
 
 /*
@@ -101,8 +108,8 @@ class TRTConvertValidation {
   }
 
   void DeclVar(const std::string& name, const std::vector<int> dim_vec) {
-    platform::CPUPlace place;
-    platform::CPUDeviceContext ctx(place);
+    platform::CUDAPlace place;
+    platform::CUDADeviceContext ctx(place);
 
     auto* x = scope_.Var(name);
     auto* x_tensor = x->GetMutable<framework::LoDTensor>();
@@ -141,7 +148,7 @@ class TRTConvertValidation {
       PADDLE_ENFORCE(var);
       auto tensor = var->GetMutable<framework::LoDTensor>();
 
-      engine_->SetInputFromCPU(
+      engine_->SetInputFromGPU(
           input, static_cast<void*>(tensor->data<void>()),
           sizeof(float) *
               analysis::AccuDims(tensor->dims(), tensor->dims().size()));
@@ -151,8 +158,8 @@ class TRTConvertValidation {
   void Execute(int batch_size) {
     // Execute Fluid Op
     PADDLE_ENFORCE_LE(batch_size, max_batch_size_);
-    platform::CPUPlace place;
-    platform::CPUDeviceContext ctx(place);
+    platform::CUDAPlace place;
+    platform::CUDADeviceContext ctx(place);
     op_->Run(scope_, place);
     // Execute TRT.
     engine_->Execute(batch_size);

paddle/fluid/inference/tensorrt/engine.cc

@@ -33,6 +33,7 @@ void TensorRTEngine::Build(const DescType &paddle_model) {
 }
 
 void TensorRTEngine::Execute(int batch_size) {
+  freshDeviceId();
   batch_size_ = batch_size;
   std::vector<void *> buffers;
   for (auto &buf : buffers_) {
@@ -60,6 +61,7 @@ TensorRTEngine::~TensorRTEngine() {
 }
 
 void TensorRTEngine::FreezeNetwork() {
+  freshDeviceId();
   PADDLE_ENFORCE(infer_builder_ != nullptr,
                  "Call InitNetwork first to initialize network.");
   PADDLE_ENFORCE(infer_network_ != nullptr,
@@ -241,6 +243,13 @@ void TensorRTEngine::SetRuntimeBatch(size_t batch_size) {
 
 int TensorRTEngine::GetRuntimeBatch() { return runtime_batch_; }
 
+void TensorRTEngine::freshDeviceId() {
+  int count;
+  cudaGetDeviceCount(&count);
+  PADDLE_ENFORCE_LT(device_, count);
+  cudaSetDevice(device_);
+}
+
 }  // namespace tensorrt
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/tensorrt/engine.h

@@ -19,6 +19,7 @@ limitations under the License. */
 #include <string>
 #include <unordered_map>
 #include <vector>
+#include "paddle/fluid/framework/tensor.h"
 #include "paddle/fluid/inference/engine.h"
 #include "paddle/fluid/inference/tensorrt/helper.h"
 #include "paddle/fluid/inference/utils/singleton.h"
@@ -52,13 +53,15 @@ class TensorRTEngine : public EngineBase {
   };
 
   TensorRTEngine(int max_batch, int max_workspace,
-                 cudaStream_t* stream = nullptr,
+                 cudaStream_t* stream = nullptr, int device = 0,
                  nvinfer1::ILogger& logger = NaiveLogger::Global())
       : max_batch_(max_batch),
         max_workspace_(max_workspace),
         stream_(stream ? stream : &default_stream_),
-        logger_(logger) {
-    cudaStreamCreate(&default_stream_);
+        logger_(logger),
+        device_(device) {
+    freshDeviceId();
+    cudaStreamCreate(stream_);
   }
 
   virtual ~TensorRTEngine();
@@ -119,6 +122,15 @@ class TensorRTEngine : public EngineBase {
   nvinfer1::INetworkDefinition* network() { return infer_network_.get(); }
   void SetRuntimeBatch(size_t batch_size);
   int GetRuntimeBatch();
+  int GetDevice() { return device_; }
+
+  // A pointer to CPU memory is needed of the TRT weight.
+  // Before TRT runs, fluid loads weight into GPU storage.
+  // so we need to copy the weights from GPU to CPU in our op converter.
+  // We use a map to store these weights for the weight memory is not released
+  // in advance, which affecting the construction of TRT Op.
+  std::unordered_map<std::string /*name*/, std::unique_ptr<framework::Tensor>>
+      weight_map;
 
  private:
   // the max batch size
@@ -140,6 +152,8 @@ class TensorRTEngine : public EngineBase {
   std::unordered_map<std::string /*name*/, size_t /*max size*/> buffer_sizes_;
   std::unordered_map<std::string /*name*/, nvinfer1::ITensor* /*ITensor*/>
       itensor_map_;
+  // The specific GPU id that the TensorRTEngine bounded to.
+  int device_;
 
   // TensorRT related internal members
   template <typename T>
@@ -156,6 +170,10 @@ class TensorRTEngine : public EngineBase {
   infer_ptr<nvinfer1::INetworkDefinition> infer_network_;
   infer_ptr<nvinfer1::ICudaEngine> infer_engine_;
   infer_ptr<nvinfer1::IExecutionContext> infer_context_;
+  // Each ICudaEngine object is bound to a specific GPU when it is instantiated,
+  // ensure that the thread is associated with the correct device by calling
+  // freshDeviceId().
+  void freshDeviceId();
 };  // class TensorRTEngine
 
 // Add an layer__ into engine__ with args ARGS.
@@ -188,8 +206,8 @@ class TRT_EngineManager {
 
   // Create or get an engine called `name`
   TensorRTEngine* Create(int max_batch, int max_workspace, cudaStream_t* stream,
-                         const std::string& name) {
-    auto* p = new TensorRTEngine(max_batch, max_workspace, stream);
+                         const std::string& name, int gpu_device = 0) {
+    auto* p = new TensorRTEngine(max_batch, max_workspace, stream, gpu_device);
     engines_[name].reset(p);
     return p;
   }

paddle/fluid/inference/tensorrt/test_engine.cc

@@ -27,7 +27,7 @@ namespace tensorrt {
 class TensorRTEngineTest : public ::testing::Test {
  protected:
   void SetUp() override {
-    ASSERT_EQ(0, cudaStreamCreate(&stream_));
+    // ASSERT_EQ(0, cudaStreamCreate(&stream_));
     engine_ = new TensorRTEngine(10, 1 << 10, &stream_);
     engine_->InitNetwork();
   }

paddle/fluid/operators/CMakeLists.txt

@@ -100,7 +100,8 @@ function(op_library TARGET)
     endif()
 
     # Define operators that don't need pybind here.
-    foreach(manual_pybind_op "compare_op" "logical_op" "nccl_op" "tensor_array_read_write_op")
+    foreach(manual_pybind_op "compare_op" "logical_op" "nccl_op"
+"tensor_array_read_write_op" "tensorrt_engine_op")
         if ("${TARGET}" STREQUAL "${manual_pybind_op}")
             set(pybind_flag 1)
         endif()
@@ -248,6 +249,7 @@ op_library(softmax_op DEPS softmax)
 op_library(sequence_softmax_op DEPS softmax)
 if (WITH_GPU AND TENSORRT_FOUND)
     op_library(tensorrt_engine_op DEPS tensorrt_engine tensorrt_converter)
+    file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(tensorrt_engine);\n")
     nv_test(test_tensorrt_engine_op SRCS tensorrt_engine_op_test.cc
       DEPS tensorrt_engine_op
       analysis)

paddle/fluid/operators/tensorrt_engine_op.cc

@@ -17,10 +17,6 @@
 #include <string>
 #include <vector>
 
-#include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
-#include "paddle/fluid/inference/tensorrt/engine.h"
-#include "paddle/fluid/inference/utils/singleton.h"
 #include "paddle/fluid/operators/tensorrt_engine_op.h"
 
 namespace paddle {
@@ -29,100 +25,6 @@ DEFINE_int32(tensorrt_engine_batch_size, 1, "the batch_size of TensorRT");
 
 namespace operators {
 
-using inference::Singleton;
-using inference::tensorrt::TRT_EngineManager;
-
-using FluidDT = framework::proto::VarType_Type;
-using TRT_DT = nvinfer1::DataType;
-
-namespace {
-
-TRT_DT FluidDataType2TRT(FluidDT type) {
-  switch (type) {
-    case FluidDT::VarType_Type_FP32:
-      return TRT_DT::kFLOAT;
-    case FluidDT::VarType_Type_INT32:
-      return TRT_DT::kINT32;
-    default:
-      return TRT_DT::kINT32;
-  }
-  PADDLE_THROW("unkown type");
-  return TRT_DT::kINT32;
-}
-
-nvinfer1::Dims Vec2TRT_Dims(const std::vector<int64_t> &shape) {
-  PADDLE_ENFORCE_GT(shape.size(), 1UL,
-                    "TensorRT' tensor input requires at least 2 dimensions");
-  PADDLE_ENFORCE_LE(shape.size(), 4UL,
-                    "TensorRT' tensor input requires at most 4 dimensions");
-  PADDLE_ENFORCE_EQ(shape.size(), 4UL);
-  return nvinfer1::DimsCHW(shape[1], shape[2], shape[3]);
-}
-
-}  // namespace
-
-template <typename DeviceContext, typename T>
-void TensorRTEngineKernel<DeviceContext, T>::Prepare(
-    const framework::ExecutionContext &context) const {
-  VLOG(4) << "Prepare engine";
-  // Get the ProgramDesc and pass to convert.
-  framework::proto::BlockDesc block_desc;
-  block_desc.ParseFromString(context.Attr<std::string>("subgraph"));
-  int max_batch = context.Attr<int>("max_batch");
-  auto max_workspace = context.Attr<int>("max_workspace");
-  auto params = context.Attr<std::vector<std::string>>("parameters");
-  std::unordered_set<std::string> parameters;
-  for (const auto &param : params) {
-    parameters.insert(param);
-  }
-
-  std::vector<std::string> output_maps =
-      context.Attr<std::vector<std::string>>("output_name_mapping");
-
-  // TODO(Superjomn) replace this with a different stream
-  auto *engine = Singleton<TRT_EngineManager>::Global().Create(
-      max_batch, max_workspace, nullptr /*engine hold its own stream*/,
-      context.Attr<std::string>("engine_uniq_key"));
-  engine->InitNetwork();
-
-  framework::BlockDesc block(nullptr /*programdesc*/, &block_desc);
-  VLOG(4) << "parsed var size " << block.AllVars().size();
-  // Add inputs
-  VLOG(4) << "declare inputs";
-  for (auto &input : context.Inputs("Xs")) {
-    if (parameters.count(input)) continue;
-    VLOG(4) << "declare input " << input;
-    auto *var = block.FindVar(input);
-    // TensorRT engine need to create parameters. The parameter's description
-    // should be set in
-    PADDLE_ENFORCE(var, "no variable called %s", input);
-    PADDLE_ENFORCE_EQ(var->GetType(), FluidDT::VarType_Type_LOD_TENSOR,
-                      "TensorRT engine only takes LoDTensor as input");
-    auto shape = var->GetShape();
-    // For the special batch_size placeholder -1, drop it and pass the real
-    // shape of data.
-    // TODO(Superjomn) fix this with batch broadcast, or it can't handle
-    // variational batch size.
-    if (shape[0] == -1) {
-      shape[0] = FLAGS_tensorrt_engine_batch_size;
-    }
-    engine->DeclareInput(
-        input, FluidDataType2TRT(
-                   var->Proto()->type().lod_tensor().tensor().data_type()),
-        Vec2TRT_Dims(shape));
-  }
-
-  inference::Singleton<inference::tensorrt::OpConverter>::Global().ConvertBlock(
-      block_desc, parameters, context.scope(), engine);
-
-  // Add outputs
-  for (auto &output : output_maps) {
-    engine->DeclareOutput(output);
-  }
-
-  engine->FreezeNetwork();
-}
-
 class TensorRTEngineOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
@@ -150,11 +52,4 @@ namespace ops = paddle::operators;
 REGISTER_OPERATOR(tensorrt_engine, ops::TensorRTEngineOp,
                   ops::TensorRTEngineOpMaker, ops::TensorRTEngineOpMaker);
 
-REGISTER_OP_CPU_KERNEL(
-    tensorrt_engine,
-    ops::TensorRTEngineKernel<paddle::platform::CPUDeviceContext, float>,
-    ops::TensorRTEngineKernel<paddle::platform::CPUDeviceContext, double>,
-    ops::TensorRTEngineKernel<paddle::platform::CPUDeviceContext, int>,
-    ops::TensorRTEngineKernel<paddle::platform::CPUDeviceContext, int64_t>);
-
 #endif  // PADDLE_WITH_CUDA

paddle/fluid/operators/tensorrt_engine_op.cu.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/operators/tensorrt_engine_op.h"
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_CUDA_KERNEL(
+    tensorrt_engine,
+    ops::TensorRTEngineKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::TensorRTEngineKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::TensorRTEngineKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::TensorRTEngineKernel<paddle::platform::CUDADeviceContext, int64_t>);

paddle/fluid/operators/tensorrt_engine_op.h

@@ -19,8 +19,10 @@
 #include <string>
 #include <vector>
 
+#include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/inference/analysis/helper.h"
+#include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
 #include "paddle/fluid/inference/tensorrt/engine.h"
 
 namespace paddle {
@@ -29,6 +31,35 @@ DECLARE_int32(tensorrt_engine_batch_size);
 
 namespace operators {
 
+using FluidDT = framework::proto::VarType_Type;
+using TRT_DT = nvinfer1::DataType;
+
+namespace {
+
+TRT_DT FluidDataType2TRT(FluidDT type) {
+  switch (type) {
+    case FluidDT::VarType_Type_FP32:
+      return TRT_DT::kFLOAT;
+    case FluidDT::VarType_Type_INT32:
+      return TRT_DT::kINT32;
+    default:
+      return TRT_DT::kINT32;
+  }
+  PADDLE_THROW("unkown type");
+  return TRT_DT::kINT32;
+}
+
+nvinfer1::Dims Vec2TRT_Dims(const std::vector<int64_t>& shape) {
+  PADDLE_ENFORCE_GT(shape.size(), 1UL,
+                    "TensorRT' tensor input requires at least 2 dimensions");
+  PADDLE_ENFORCE_LE(shape.size(), 4UL,
+                    "TensorRT' tensor input requires at most 4 dimensions");
+  PADDLE_ENFORCE_EQ(shape.size(), 4UL);
+  return nvinfer1::DimsCHW(shape[1], shape[2], shape[3]);
+}
+
+}  // namespace
+
 using inference::Singleton;
 using inference::tensorrt::TRT_EngineManager;
 
@@ -47,7 +78,7 @@ class TensorRTEngineOp : public framework::OperatorWithKernel {
                                   .FindVar(input0)
                                   ->GetMutable<framework::LoDTensor>()
                                   ->type()),
-        platform::CPUPlace());
+        ctx.GetPlace());
     return kt;
   }
 };
@@ -94,7 +125,9 @@ class TensorRTEngineKernel : public framework::OpKernel<T> {
 
     // Convert output tensor from engine to fluid
     int output_index = 0;
+    VLOG(4) << "TensorRT Engine Op Outputs:";
     for (const auto& y : context.Outputs("Ys")) {
+      VLOG(4) << y;
       // convert output and copy to fluid.
       nvinfer1::ITensor* trt_t = engine->GetITensor(output_maps[output_index]);
       auto dims = trt_t->getDimensions();
@@ -113,8 +146,10 @@ class TensorRTEngineKernel : public framework::OpKernel<T> {
       // TODO(Superjomn) change this float to dtype size.
       auto size = inference::analysis::AccuDims(dims.d, dims.nbDims) *
                   FLAGS_tensorrt_engine_batch_size;
-      engine->GetOutputInCPU(output_maps[output_index],
-                             fluid_t->mutable_data<float>(platform::CPUPlace()),
+      engine->GetOutputInGPU(
+          output_maps[output_index],
+          fluid_t->mutable_data<float>(platform::CUDAPlace(
+              boost::get<platform::CUDAPlace>(context.GetPlace()).device)),
           size * sizeof(float));
       //} else {
       // engine->GetOutputInGPU(
@@ -128,8 +163,67 @@ class TensorRTEngineKernel : public framework::OpKernel<T> {
   }
 
  protected:
-  // Build the engine.
-  void Prepare(const framework::ExecutionContext& context) const;
+  void Prepare(const framework::ExecutionContext& context) const {
+    VLOG(4) << "Prepare engine";
+    // Get the ProgramDesc and pass to convert.
+    framework::proto::BlockDesc block_desc;
+    block_desc.ParseFromString(context.Attr<std::string>("subgraph"));
+    int max_batch = context.Attr<int>("max_batch");
+    auto max_workspace = context.Attr<int>("max_workspace");
+    auto params = context.Attr<std::vector<std::string>>("parameters");
+    std::unordered_set<std::string> parameters;
+    for (const auto& param : params) {
+      parameters.insert(param);
+    }
+
+    std::vector<std::string> output_maps =
+        context.Attr<std::vector<std::string>>("output_name_mapping");
+
+    // TODO(Superjomn) replace this with a different stream
+    auto* engine = Singleton<TRT_EngineManager>::Global().Create(
+        max_batch, max_workspace, nullptr /*engine hold its own stream*/,
+        context.Attr<std::string>("engine_uniq_key"),
+        boost::get<platform::CUDAPlace>(context.GetPlace()).device);
+
+    engine->InitNetwork();
+
+    framework::BlockDesc block(nullptr /*programdesc*/, &block_desc);
+    VLOG(4) << "parsed var size " << block.AllVars().size();
+    // Add inputs
+    VLOG(4) << "declare inputs";
+    for (auto& input : context.Inputs("Xs")) {
+      if (parameters.count(input)) continue;
+      VLOG(4) << "declare input " << input;
+      auto* var = block.FindVar(input);
+      // TensorRT engine need to create parameters. The parameter's description
+      // should be set in
+      PADDLE_ENFORCE(var, "no variable called %s", input);
+      PADDLE_ENFORCE_EQ(var->GetType(), FluidDT::VarType_Type_LOD_TENSOR,
+                        "TensorRT engine only takes LoDTensor as input");
+      auto shape = var->GetShape();
+      // For the special batch_size placeholder -1, drop it and pass the real
+      // shape of data.
+      // TODO(Superjomn) fix this with batch broadcast, or it can't handle
+      // variational batch size.
+      if (shape[0] == -1) {
+        shape[0] = FLAGS_tensorrt_engine_batch_size;
+      }
+      engine->DeclareInput(
+          input, FluidDataType2TRT(
+                     var->Proto()->type().lod_tensor().tensor().data_type()),
+          Vec2TRT_Dims(shape));
+    }
+
+    inference::Singleton<inference::tensorrt::OpConverter>::Global()
+        .ConvertBlock(block_desc, parameters, context.scope(), engine);
+
+    // Add outputs
+    for (auto& output : output_maps) {
+      engine->DeclareOutput(output);
+    }
+
+    engine->FreezeNetwork();
+  }
 };
 
 }  // namespace operators

paddle/fluid/operators/tensorrt_engine_op_test.cc

@@ -23,20 +23,20 @@ limitations under the License. */
 #include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
 #include "paddle/fluid/inference/tensorrt/convert/ut_helper.h"
 
-USE_CPU_ONLY_OP(tensorrt_engine);
+USE_CUDA_ONLY_OP(tensorrt_engine);
 
 namespace paddle {
 namespace operators {
 
 namespace {
-void CreateCPUTensor(framework::Scope* scope, const std::string& name,
+void CreateCUDATensor(framework::Scope* scope, const std::string& name,
                       const std::vector<int64_t>& shape) {
   auto* var = scope->Var(name);
   auto* tensor = var->GetMutable<framework::LoDTensor>();
   auto dims = framework::make_ddim(shape);
   tensor->Resize(dims);
-  platform::CPUPlace place;
-  platform::CPUDeviceContext ctx(place);
+  platform::CUDAPlace place;
+  platform::CUDADeviceContext ctx(place);
   inference::tensorrt::RandomizeTensor(tensor, place, ctx);
 }
 
@@ -112,15 +112,15 @@ TEST(TensorRTEngineOp, manual) {
   LOG(INFO) << "engine_op " << engine_op.get();
 
   framework::Scope scope;
-  platform::CPUPlace place;
-  platform::CPUDeviceContext ctx(place);
+  platform::CUDAPlace place;
+  platform::CUDADeviceContext ctx(place);
   // Prepare variables.
-  CreateCPUTensor(&scope, "x", std::vector<int64_t>({2, 4}));
-  CreateCPUTensor(&scope, "y", std::vector<int64_t>({4, 6}));
-  CreateCPUTensor(&scope, "z", std::vector<int64_t>({2, 6}));
+  CreateCUDATensor(&scope, "x", std::vector<int64_t>({2, 4}));
+  CreateCUDATensor(&scope, "y", std::vector<int64_t>({4, 6}));
+  CreateCUDATensor(&scope, "z", std::vector<int64_t>({2, 6}));
 
-  CreateCPUTensor(&scope, "y0", std::vector<int64_t>({6, 8}));
-  CreateCPUTensor(&scope, "z0", std::vector<int64_t>({2, 8}));
+  CreateCUDATensor(&scope, "y0", std::vector<int64_t>({6, 8}));
+  CreateCUDATensor(&scope, "z0", std::vector<int64_t>({2, 8}));
 
   // Execute them.
   LOG(INFO) << "engine_op run";
@@ -130,8 +130,8 @@ TEST(TensorRTEngineOp, manual) {
 void Execute(int batch_size, int input_dim, int output_dim, int nlayers = 1) {
   framework::ProgramDesc program;
   framework::Scope scope;
-  platform::CPUPlace place;
-  platform::CPUDeviceContext ctx(place);
+  platform::CUDAPlace place;
+  platform::CUDADeviceContext ctx(place);
 
   auto* block_ = program.Proto()->add_blocks();
   block_->set_idx(0);
@@ -165,10 +165,10 @@ void Execute(int batch_size, int input_dim, int output_dim, int nlayers = 1) {
 
     // Prepare variables.
     if (!x_created) {
-      CreateCPUTensor(&scope, x_name, std::vector<int64_t>(x_shape));
+      CreateCUDATensor(&scope, x_name, std::vector<int64_t>(x_shape));
     }
-    CreateCPUTensor(&scope, y_name, std::vector<int64_t>(y_shape));
-    CreateCPUTensor(&scope, z_name, std::vector<int64_t>(z_shape));
+    CreateCUDATensor(&scope, y_name, std::vector<int64_t>(y_shape));
+    CreateCUDATensor(&scope, z_name, std::vector<int64_t>(z_shape));
 
     // It is wired, need to copy manually.
     *block_->add_ops() = *fc->Proto();