bugfix/trt engine op (#11487)

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

@@ -64,7 +64,8 @@ class OpConverter {
     (*it)(op, scope, test_mode);
   }
 
-  // convert fluid block to tensorrt network
+  // Convert a fluid block to tensorrt network, NOTE it just convert operators,
+  // the INetwork's inputs and outputs should specified in some other modules.
   void ConvertBlock(const framework::proto::BlockDesc& block,
                     const std::unordered_set<std::string>& parameters,
                     const framework::Scope& scope, TensorRTEngine* engine) {

paddle/fluid/inference/tensorrt/engine.h

@@ -51,11 +51,12 @@ class TensorRTEngine : public EngineBase {
     nvinfer1::Weights w_;
   };
 
-  TensorRTEngine(int max_batch, int max_workspace, cudaStream_t* stream,
+  TensorRTEngine(int max_batch, int max_workspace,
+                 cudaStream_t* stream = nullptr,
                  nvinfer1::ILogger& logger = NaiveLogger::Global())
       : max_batch_(max_batch),
         max_workspace_(max_workspace),
-        stream_(stream),
+        stream_(stream ? stream : &default_stream_),
         logger_(logger) {}
 
   virtual ~TensorRTEngine();
@@ -121,6 +122,8 @@ class TensorRTEngine : public EngineBase {
   // the max memory size the engine uses
   int max_workspace_;
   cudaStream_t* stream_;
+  // If stream_ is not set from outside, hold its own stream.
+  cudaStream_t default_stream_;
   nvinfer1::ILogger& logger_;
 
   std::vector<Buffer> buffers_;
@@ -165,20 +168,31 @@ class TensorRTEngine : public EngineBase {
  */
 class TRT_EngineManager {
  public:
-  TensorRTEngine* Create(int max_batch, int max_workspace,
-                         cudaStream_t* stream) {
-    engines_.emplace_back(new TensorRTEngine(max_batch, max_workspace, stream));
-    return engines_.back().get();
+  bool HasEngine(const std::string& name) const {
+    return engines_.count(name) != 0;
+  }
+
+  // Get an engine called `name`.
+  TensorRTEngine* Get(const std::string& name) const {
+    return engines_.at(name).get();
+  }
+
+  // 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);
+    engines_[name].reset(p);
+    return p;
   }
 
   void DeleteALl() {
-    for (auto& ptr : engines_) {
-      ptr.reset(nullptr);
+    for (auto& item : engines_) {
+      item.second.reset(nullptr);
     }
   }
 
  private:
-  std::vector<std::unique_ptr<TensorRTEngine>> engines_;
+  std::unordered_map<std::string, std::unique_ptr<TensorRTEngine>> engines_;
 };
 
 }  // namespace tensorrt

paddle/fluid/operators/tensorrt_engine_op.cc

@@ -66,17 +66,25 @@ nvinfer1::Dims Vec2TRT_Dims(const std::vector<int64_t> &shape) {
 }  // namespace
 
 template <typename DeviceContext, typename T>
-void paddle::operators::TensorRTEngineKernel<DeviceContext, T>::Prepare(
+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"));
-  max_batch_ = context.Attr<int>("max_batch");
+  int max_batch = context.Attr<int>("max_batch");
   auto max_workspace = context.Attr<int>("max_workspace");
-  engine_ = Singleton<TRT_EngineManager>::Global().Create(
-      max_batch_, max_workspace, &stream_);
-  engine_->InitNetwork();
+  auto params = context.Attr<std::vector<std::string>>("parameters");
+  std::unordered_set<std::string> parameters;
+  for (const auto &param : params) {
+    parameters.insert(param);
+  }
+
+  // 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);
   // Add inputs
@@ -87,24 +95,23 @@ void paddle::operators::TensorRTEngineKernel<DeviceContext, T>::Prepare(
     PADDLE_ENFORCE_EQ(var->GetType(), FluidDT::VarType_Type_LOD_TENSOR,
                       "TensorRT engine only takes LoDTensor as input");
     auto shape = var->GetShape();
-    engine_->DeclareInput(
+    engine->DeclareInput(
         input, FluidDataType2TRT(
                    var->Proto()->type().lod_tensor().tensor().data_type()),
         Vec2TRT_Dims(var->GetShape()));
   }
 
-  // TODO(Superjomn) parameters should be passed after analysised from outside.
   inference::Singleton<inference::tensorrt::OpConverter>::Global().ConvertBlock(
-      block_desc, {}, context.scope(), engine_);
+      block_desc, parameters, context.scope(), engine);
 
   // Add outputs
   VLOG(4) << "declare outputs";
   for (auto &output : context.Outputs("Ys")) {
     VLOG(4) << "declare output " << output;
-    engine_->DeclareOutput(output);
+    engine->DeclareOutput(output);
   }
 
-  engine_->FreezeNetwork();
+  engine->FreezeNetwork();
 }
 
 class TensorRTEngineOpMaker : public framework::OpProtoAndCheckerMaker {
@@ -113,6 +120,7 @@ class TensorRTEngineOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("Xs", "A list of inputs.").AsDuplicable();
     AddOutput("Ys", "A list of outputs").AsDuplicable();
     AddAttr<std::string>("subgraph", "the subgraph.");
+    AddAttr<std::string>("engine_uniq_key", "unique key for the TRT engine.");
     AddAttr<int>("max_batch", "the maximum batch size.");
     AddAttr<int>("max_workspace", "the maximum batch size.");
     AddComment("TensorRT engine operator.");

paddle/fluid/operators/tensorrt_engine_op.h

@@ -19,10 +19,14 @@
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/inference/analysis/helper.h"
 #include "paddle/fluid/inference/tensorrt/engine.h"
+#include "paddle/fluid/inference/tensorrt/engine.h"
 
 namespace paddle {
 namespace operators {
 
+using inference::Singleton;
+using inference::tensorrt::TRT_EngineManager;
+
 class TensorRTEngineOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -47,16 +51,18 @@ template <typename DeviceContext, typename T>
 class TensorRTEngineKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    if (!engine_) {
+    auto engine_name = context.Attr<std::string>("engine_uniq_key");
+    if (!Singleton<TRT_EngineManager>::Global().HasEngine(engine_name)) {
       Prepare(context);
     }
+    auto* engine = Singleton<TRT_EngineManager>::Global().Get(engine_name);
     auto input_names = context.op().Inputs("Xs");
     PADDLE_ENFORCE(!input_names.empty(), "should pass more than one inputs");
     // Try to determine a batch_size
     auto& tensor0 = inference::analysis::GetFromScope<framework::LoDTensor>(
         context.scope(), input_names.front());
     int batch_size = tensor0.dims()[0];
-    PADDLE_ENFORCE_LE(batch_size, max_batch_);
+    PADDLE_ENFORCE_LE(batch_size, context.Attr<int>("max_batch"));
 
     // Convert input tensor from fluid to engine.
     for (const auto& x : context.Inputs("Xs")) {
@@ -64,20 +70,20 @@ class TensorRTEngineKernel : public framework::OpKernel<T> {
       auto& t = inference::analysis::GetFromScope<framework::LoDTensor>(
           context.scope(), x);
       if (platform::is_cpu_place(t.place())) {
-        engine_->SetInputFromCPU(x, static_cast<const void*>(t.data<void>()),
-                                 t.memory_size());
+        engine->SetInputFromCPU(x, static_cast<const void*>(t.data<void>()),
+                                t.memory_size());
       } else {
-        engine_->SetInputFromGPU(x, static_cast<const void*>(t.data<void>()),
-                                 t.memory_size());
+        engine->SetInputFromGPU(x, static_cast<const void*>(t.data<void>()),
+                                t.memory_size());
       }
     }
     // Execute the engine.
     PADDLE_ENFORCE_GT(batch_size, 0);
-    engine_->Execute(batch_size);
+    engine->Execute(batch_size);
     // Convert output tensor from engine to fluid
     for (const auto& y : context.Outputs("Ys")) {
       // convert output and copy to fluid.
-      nvinfer1::ITensor* trt_t = engine_->GetITensor(y);
+      nvinfer1::ITensor* trt_t = engine->GetITensor(y);
       auto dims = trt_t->getDimensions();
       // Use the output ITensor's dims to reshape the Fluid Tensor.
       std::vector<int> ddim(dims.d, dims.d + dims.nbDims);
@@ -89,27 +95,22 @@ class TensorRTEngineKernel : public framework::OpKernel<T> {
       auto size = inference::analysis::AccuDims(dims.d, dims.nbDims);
       if (platform::is_cpu_place(fluid_t->place())) {
         // TODO(Superjomn) change this float to dtype size.
-        engine_->GetOutputInCPU(
+        engine->GetOutputInCPU(
             y, fluid_t->mutable_data<float>(platform::CPUPlace()),
             size * sizeof(float));
       } else {
-        engine_->GetOutputInGPU(
+        engine->GetOutputInGPU(
             y, fluid_t->mutable_data<float>(platform::CUDAPlace()),
             size * sizeof(float));
       }
     }
 
-    cudaStreamSynchronize(stream_);
+    cudaStreamSynchronize(*engine->stream());
   }
 
  protected:
   // Build the engine.
   void Prepare(const framework::ExecutionContext& context) const;
-
- private:
-  mutable cudaStream_t stream_;
-  mutable inference::tensorrt::TensorRTEngine* engine_{nullptr};
-  mutable int max_batch_{0};
 };
 
 }  // namespace operators

paddle/fluid/operators/tensorrt_engine_op_test.cc

@@ -79,6 +79,17 @@ void SetAttr<int64_t>(framework::proto::OpDesc* op, const std::string& name,
   attr->set_type(paddle::framework::proto::AttrType::LONG);
   attr->set_l(data);
 }
+template <>
+void SetAttr<std::vector<std::string>>(framework::proto::OpDesc* op,
+                                       const std::string& name,
+                                       const std::vector<std::string>& data) {
+  auto* attr = op->add_attrs();
+  attr->set_name(name);
+  attr->set_type(paddle::framework::proto::AttrType::STRINGS);
+  for (const auto& s : data) {
+    attr->add_strings(s.c_str());
+  }
+}
 
 }  // namespace
 
@@ -123,11 +134,15 @@ TEST(TensorRTEngineOp, manual) {
   engine_op_desc.SetOutput("Ys", std::vector<std::string>({"z0"}));
   SetAttr<std::string>(engine_op_desc.Proto(), "subgraph",
                        block_->SerializeAsString());
-  SetAttr<int>(engine_op_desc.Proto(), "max_batch", 30);
+  SetAttr<int>(engine_op_desc.Proto(), "max_batch", 100);
   SetAttr<int>(engine_op_desc.Proto(), "max_workspace", 1 << 10);
+  SetAttr<std::string>(engine_op_desc.Proto(), "engine_uniq_key", "a_engine");
+  SetAttr<std::vector<std::string>>(engine_op_desc.Proto(), "parameters",
+                                    std::vector<std::string>({}));
 
   LOG(INFO) << "create engine op";
   auto engine_op = framework::OpRegistry::CreateOp(*engine_op_desc.Proto());
+  LOG(INFO) << "engine_op " << engine_op.get();
 
   framework::Scope scope;
   platform::CPUPlace place;
@@ -145,6 +160,88 @@ TEST(TensorRTEngineOp, manual) {
   engine_op->Run(scope, place);
 }
 
+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);
+
+  auto* block_ = program.Proto()->add_blocks();
+  block_->set_idx(0);
+  block_->set_parent_idx(-1);
+
+  using shape_t = std::vector<int64_t>;
+
+  LOG(INFO) << "create block desc";
+  framework::BlockDesc block_desc(&program, block_);
+
+  auto AddFCLayer = [&](const std::string& x_name, const std::string& y_name,
+                        const std::string& z_name, bool x_created,
+                        const shape_t& x_shape, const shape_t& y_shape,
+                        const shape_t& z_shape) {
+
+    LOG(INFO) << "create fc op";
+    auto* fc = block_desc.AppendOp();
+    fc->SetType("mul");
+    fc->SetInput("X", std::vector<std::string>({x_name}));
+    fc->SetInput("Y", std::vector<std::string>({y_name}));
+    fc->SetOutput("Out", std::vector<std::string>({z_name}));
+
+    // Set inputs' variable shape in BlockDesc
+    if (!x_created) {
+      AddTensorToBlockDesc(block_, x_name,
+                           std::vector<int64_t>({batch_size, input_dim, 1, 1}));
+    }
+    AddTensorToBlockDesc(block_, y_name,
+                         std::vector<int64_t>({input_dim, output_dim}));
+    AddTensorToBlockDesc(block_, z_name,
+                         std::vector<int64_t>({batch_size, output_dim}));
+
+    // Prepare variables.
+    if (!x_created) {
+      CreateCPUTensor(&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));
+
+    // It is wired, need to copy manually.
+    *block_->add_ops() = *fc->Proto();
+  };
+
+  // Test with 4 layer FC
+  AddFCLayer("x0", "y0", "z0", false, {batch_size, input_dim},
+             {input_dim, output_dim}, {batch_size, output_dim});
+  AddFCLayer("z0", "y1", "z1", true, {}, {output_dim, output_dim},
+             {batch_size, output_dim});
+  AddFCLayer("z1", "y2", "z2", true, {}, {output_dim, output_dim},
+             {batch_size, output_dim});
+  AddFCLayer("z2", "y3", "z3", true, {}, {output_dim, output_dim},
+             {batch_size, output_dim});
+
+  LOG(INFO) << "create tensorrt desc";
+  framework::OpDesc engine_op_desc(nullptr);
+  engine_op_desc.SetType("tensorrt_engine");
+  engine_op_desc.SetInput("Xs", std::vector<std::string>({"x0"}));
+  engine_op_desc.SetOutput("Ys", std::vector<std::string>({"z3"}));
+
+  SetAttr<std::string>(engine_op_desc.Proto(), "subgraph",
+                       block_->SerializeAsString());
+  SetAttr<int>(engine_op_desc.Proto(), "max_batch", batch_size);
+  SetAttr<int>(engine_op_desc.Proto(), "max_workspace", 2 << 10);
+  SetAttr<std::vector<std::string>>(
+      engine_op_desc.Proto(), "parameters",
+      std::vector<std::string>({"y0", "y1", "y2", "y3"}));
+  SetAttr<std::string>(engine_op_desc.Proto(), "engine_uniq_key", "b_engine");
+
+  auto engine_op = framework::OpRegistry::CreateOp(*engine_op_desc.Proto());
+
+  // Execute them.
+  engine_op->Run(scope, place);
+}
+
+// Test with a larger FC layer.
+TEST(TensorRTEngineOp, fc) { Execute(40, 256, 256); }
+
 }  // namespace operators
 }  // namespace paddle