Merge branch 'develop' of upstream into fix_docs

Author: Yibing Liu

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/activation_op.cc

@@ -112,7 +112,7 @@ Sigmoid Activation Operator
 __attribute__((unused)) constexpr char LogSigmoidDoc[] = R"DOC(
 Logsigmoid Activation Operator
 
-$$out = \log \frac{1}{1 + e^{-x}}$$
+$$out = \\log \\frac{1}{1 + e^{-x}}$$
 
 )DOC";
 
@@ -252,15 +252,14 @@ class SoftShrinkOpMaker : public framework::OpProtoAndCheckerMaker {
     AddOutput("Out", "Output of Softshrink operator");
     AddAttr<float>("lambda", "non-negative offset").SetDefault(0.5f);
     AddComment(R"DOC(
-Softshrink Activation Operator.
-
-$$
-out = \begin{cases} 
-    x - \lambda, \text{if } x > \lambda \\
-    x + \lambda, \text{if } x < -\lambda \\
-    0,  \text{otherwise}
-    \end{cases}
-$$
+:strong:`Softshrink Activation Operator`
+
+..  math::
+    out = \begin{cases} 
+         x - \lambda, \text{if } x > \lambda \\
+         x + \lambda, \text{if } x < -\lambda \\
+         0,  \text{otherwise}
+         \end{cases}
 
 )DOC");
   }

paddle/fluid/operators/detection/box_coder_op.cc

@@ -106,23 +106,36 @@ class BoxCoderOpMaker : public framework::OpProtoAndCheckerMaker {
               "and M represents the number of deocded boxes.");
 
     AddComment(R"DOC(
-Bounding Box Coder Operator.
+
+Bounding Box Coder.
+
 Encode/Decode the target bounding box with the priorbox information.
+
 The Encoding schema described below:
-ox = (tx - px) / pw / pxv
-oy = (ty - py) / ph / pyv
-ow = log(abs(tw / pw)) / pwv 
-oh = log(abs(th / ph)) / phv 
+
+    ox = (tx - px) / pw / pxv
+
+    oy = (ty - py) / ph / pyv
+
+    ow = log(abs(tw / pw)) / pwv 
+
+    oh = log(abs(th / ph)) / phv 
+
 The Decoding schema described below:
-ox = (pw * pxv * tx * + px) - tw / 2
-oy = (ph * pyv * ty * + py) - th / 2
-ow = exp(pwv * tw) * pw + tw / 2
-oh = exp(phv * th) * ph + th / 2
-where tx, ty, tw, th denote the target box's center coordinates, width and
-height respectively. Similarly, px, py, pw, ph denote the priorbox's(anchor)
-center coordinates, width and height. pxv, pyv, pwv, phv denote the variance
-of the priorbox and ox, oy, ow, oh denote the encoded/decoded coordinates,
-width and height.
+
+    ox = (pw * pxv * tx * + px) - tw / 2
+
+    oy = (ph * pyv * ty * + py) - th / 2
+
+    ow = exp(pwv * tw) * pw + tw / 2
+
+    oh = exp(phv * th) * ph + th / 2
+
+where `tx`, `ty`, `tw`, `th` denote the target box's center coordinates, width
+and height respectively. Similarly, `px`, `py`, `pw`, `ph` denote the
+priorbox's (anchor) center coordinates, width and height. `pxv`, `pyv`, `pwv`,
+`phv` denote the variance of the priorbox and `ox`, `oy`, `ow`, `oh` denote the
+encoded/decoded coordinates, width and height.
 )DOC");
   }
 };

paddle/fluid/operators/gaussian_random_batch_size_like_op.cc

@@ -36,11 +36,12 @@ class GaussianRandomBatchSizeLikeOpMaker : public BatchSizeLikeOpMaker {
   void Apply() override {
     AddAttr<float>("mean",
                    "(float, default 0.0) "
-                   "mean of random tensor.")
+                   "The mean (or center) of the gaussian distribution.")
         .SetDefault(.0f);
     AddAttr<float>("std",
                    "(float, default 1.0) "
-                   "std of random tensor.")
+                   "The standard deviation (std, or spread) of the "
+                   "gaussian distribution.")
         .SetDefault(1.0f);
     AddAttr<int>("seed",
                  "(int, default 0) "
@@ -55,9 +56,11 @@ class GaussianRandomBatchSizeLikeOpMaker : public BatchSizeLikeOpMaker {
         .SetDefault(framework::proto::VarType::FP32);
 
     AddComment(R"DOC(
-GaussianRandom Operator.
 
 Used to initialize tensors with gaussian random generator.
+The defalut mean of the distribution is 0. and defalut standard
+deviation (std) of the distribution is 1.. Uers can set mean and std
+by input arguments.
 )DOC");
   }
 };

paddle/fluid/operators/listen_and_serv_op.cc

@@ -348,7 +348,8 @@ class ListenAndServOpMaker : public framework::OpProtoAndCheckerMaker {
 };
 
 void SignalHandler::StopAndExit(int signal_num) {
-  VLOG(3) << "Catch interrupt signal: " << signal_num << ", program will exit";
+  // Do not use VLOG here for the device for printing maybe already released.
+  // exit will release interal allocated resoureces.
   exit(0);
 }
 

paddle/fluid/operators/mean_op.cc

@@ -33,12 +33,10 @@ class MeanOp : public framework::OperatorWithKernel {
 class MeanOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
-    AddInput("X", "The input of mean op");
-    AddOutput("Out", "The output of mean op").Reuse("X");
+    AddInput("X", "(Tensor) The input of mean op");
+    AddOutput("Out", "(Tensor) The output of mean op").Reuse("X");
     AddComment(R"DOC(
-Mean Operator.
-
-Out is a scalar which is the mean of all elements in X. 
+Mean Operator calculates the mean of all elements in X.
 
 )DOC");
   }

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;
@@ -47,37 +51,39 @@ 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")) {
       // convert input and copy to TRT engine's buffer
       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