PaddlePaddle inference 源码分析（四）

本节介绍预测处理的流程。预测处理流程主要分为3部分，包括准备输入数据、执行、获取输出数据。

一、放入输入数据

简单的使用方法如下所示：

vector<string> input_names = predictor->GetInputNames();
unique_ptr<Tensor> input_t = predictor->GetInputHandle(input_names[0]);
input_t->Reshape(input_shape);
input_t->CopyFromCpu(input.data());

我们按照这个流程一步一步来深入

1、GetInputNames

这个调用有点绕，因为对外提供的头文件是paddle_infer作用域，因此这里的实际实现是先在paddle_infer下函数调用，然后调用了实际创建出来的AnalysisPredictor::GetInputNamse。

这一步是获取输入的节点名称。这里idx2feeds_是std::map<size_t, std::string>，保存的是模型文件中op->Type==feed的名称

// 接口类实现
namespace paddle_infer {
std::vector<std::string> Predictor::GetInputNames() {
  return predictor_->GetInputNames();
}
}

// 实际实现
std::vector<std::string> AnalysisPredictor::GetInputNames() {
  std::vector<std::string> input_names;
  for (auto &item : idx2feeds_) {
    input_names.push_back(item.second);
  }
  return input_names;
}

2、GetInputHandle

作用是根据节点名称获取到对应的内存区域。前文介绍过Scope中保存了所有节点的信息，这里就是拿到输入节点Scope的内存区域.这里executor保存的scope是predictor的sub_scope。

namespace paddle_infer {
std::unique_ptr<Tensor> Predictor::GetInputHandle(const std::string &name) {
  return predictor_->GetInputTensor(name);
}
}

std::unique_ptr<ZeroCopyTensor> AnalysisPredictor::GetInputTensor(
    const std::string &name) {
  PADDLE_ENFORCE_NOT_NULL(
      executor_->scope()->FindVar(name),
      platform::errors::PreconditionNotMet(
          "The variable named %s is not found in the scope of the exector.",
          name));
  // 拿到scope
  std::unique_ptr<ZeroCopyTensor> res(
      new ZeroCopyTensor(static_cast<void *>(executor_->scope())));
  res->input_or_output_ = true;
  res->SetName(name);
  // 根据设备获取对应place
  if (platform::is_cpu_place(place_)) {
    res->SetPlace(PaddlePlace::kCPU);
  } else if (platform::is_xpu_place(place_)) {
    if (config_.lite_engine_enabled()) {
      // Currently, Paddle-Lite's XPU user interface only supports the transfer
      // of host data pointers. If it is currently used as a subgraph, execution
      // efficiency will be sacrificed, so it is temporarily set to cpu place.
      // And, the current lite engine of xpu must execute all parts of the
      // model.
      res->SetPlace(PaddlePlace::kCPU);
    } else {
      auto xpu_place = BOOST_GET_CONST(platform::XPUPlace, place_);
      res->SetPlace(PaddlePlace::kXPU, xpu_place.GetDeviceId());
    }
  } else if (platform::is_npu_place(place_)) {
    auto npu_place = BOOST_GET_CONST(platform::NPUPlace, place_);
    res->SetPlace(PaddlePlace::kNPU, npu_place.GetDeviceId());
  } else {
    auto gpu_place = BOOST_GET_CONST(platform::CUDAPlace, place_);
    res->SetPlace(PaddlePlace::kGPU, gpu_place.GetDeviceId());
  }
  return res;
}

3、ZeroCopyTensor::Reshape

这一步骤的作用就是操作输入tensort，重新确定输入数据的维度信息。这里我们会详细介绍一下tensor的操作。

3.1 基类及接口是paddle_infer::Tensor(paddle_tensor.h/inference/api/details/zero_copy_tensor.cc).ZeroCopyTensor（paddle_api.h）是paddle_infer::Tensor的子类，主要重写了copy相关的函数。会在下一小结具体讲述。

3.2 实际的reshape操作作用在Tensor::Reshape中，实际逻辑为从sub_scope中取出对应名称的Variable(framework/variable.h)并对其进行操作。

void Tensor::Reshape(const std::vector<int> &shape) {
  // 判断是否设置了name
  PADDLE_ENFORCE_EQ(
      name_.empty(), false,
      paddle::platform::errors::PreconditionNotMet(
          "Need to SetName first, so that the corresponding tensor can "
          "be retrieved."));
  // 判断是否为input，只有input才能重新设置
  PADDLE_ENFORCE_EQ(input_or_output_, true,
                    paddle::platform::errors::PermissionDenied(
                        "Can't reshape the output tensor, it is readonly"));
  // 获取scope，然后取出对应名称节点的变量并进行设置。这里使用的是sub_scope，其中保存的都是非永久性的节点
  auto *scope = static_cast<paddle::framework::Scope *>(scope_);
  auto *var = scope->FindVar(name_);
  PADDLE_ENFORCE_NOT_NULL(
      var, paddle::platform::errors::PreconditionNotMet(
               "No tensor called [%s] in the runtime scope", name_));
  auto *tensor = var->GetMutable<paddle::framework::LoDTensor>();
  tensor->Resize(paddle::framework::make_ddim(shape));
}

3.3 var->GetMutable,这里实际在Variable中创建对应类型的存储数据。存储数据用LoDTensor（lod_tensor.h）,创建一个LoDTensor的对象赋值给

  template <typename T>
  T* GetMutable() {
    if (!holder_) {
      holder_.reset(new PlaceholderImpl<T>());
    } else {
      PADDLE_ENFORCE_EQ(
          holder_->Type(), VarTypeTrait<T>::kId,
          platform::errors::InvalidArgument(
              "The Variable type must be %s, but the type it holds is %s.",
              ToTypeName(VarTypeTrait<T>::kId), ToTypeName(holder_->Type())));
    }
    return static_cast<T*>(holder_->Ptr());
  }

PlaceholderImpl是一个模板类，用于包装T，这样Variable类在构造时不需要包含模板，只需要把Placeholder指针作为成员变量即可std::shared_ptr<Placeholder> holder_;PlaceholderImpl构造时会保存obj指针，同时保存obj的类型序号，序号实际在proto::VarType中定义。对应关系实现已注册好。

  // Placeholder hides type T, so it doesn't appear as a template
  // parameter of Variable.
  template <typename T>
  struct PlaceholderImpl : public Placeholder {
    static_assert(
        IsRegisteredVarType<T>(),
        "Not registered type. Please register T inside var_type_traits.h");
    PlaceholderImpl() { this->Init(&obj_, VarTypeTrait<T>::kId); }

   private:
    T obj_;
  };

这里会检查T类型是否已注册，注册列表详见framework/var_type_traits.h

REG_PROTO_VAR_TYPE_TRAIT(LoDTensor, proto::VarType::LOD_TENSOR);
REG_PROTO_VAR_TYPE_TRAIT(SelectedRows, proto::VarType::SELECTED_ROWS);
REG_PROTO_VAR_TYPE_TRAIT(std::vector<Scope *>, proto::VarType::STEP_SCOPES);
REG_PROTO_VAR_TYPE_TRAIT(LoDRankTable, proto::VarType::LOD_RANK_TABLE);
REG_PROTO_VAR_TYPE_TRAIT(LoDTensorArray, proto::VarType::LOD_TENSOR_ARRAY);
REG_PROTO_VAR_TYPE_TRAIT(platform::PlaceList, proto::VarType::PLACE_LIST);
REG_PROTO_VAR_TYPE_TRAIT(ReaderHolder, proto::VarType::READER);
REG_PROTO_VAR_TYPE_TRAIT(FeedList, proto::VarType::FEED_LIST);
REG_PROTO_VAR_TYPE_TRAIT(FetchList, proto::VarType::FETCH_LIST);
REG_PROTO_VAR_TYPE_TRAIT(int, proto::VarType::INT32);
REG_PROTO_VAR_TYPE_TRAIT(float, proto::VarType::FP32);
REG_PROTO_VAR_TYPE_TRAIT(Vocab, proto::VarType::VOCAB);
REG_PROTO_VAR_TYPE_TRAIT(String, proto::VarType::STRING);
REG_PROTO_VAR_TYPE_TRAIT(Strings, proto::VarType::STRINGS);

3.4 LoDTensor这里命名空间为paddle::framework，注意与之前paddle_infer::Tensor区分开。LoDTensor的父类为paddle::framework::Tensor(framework/tensor.h)，Resize操作也是直接使用父类函数

Tensor& Tensor::Resize(const DDim& dims) {
  dims_ = dims;
  return *this;
}

　　paddle::framework::make_ddim(shape)这里创建DDim（ddim.h），其中rank_保存了维度长度，dim_保存了具体的维度数据，最大为9维

4. ZeroCopyTensor::CopyFromCpu 

这一步真正进行内存拷贝。我们分4步详细介绍

template <typename T>
void Tensor::CopyFromCpu(const T *data) {
  // 1
  EAGER_GET_TENSOR(paddle::framework::LoDTensor);
  PADDLE_ENFORCE_GE(tensor->numel(), 0,
                    paddle::platform::errors::PreconditionNotMet(
                        "You should call Tensor::Reshape(const "
                        "std::vector<int> &shape)"
                        "function before copying data from cpu."));
  // 2
  size_t ele_size = tensor->numel() * sizeof(T);

  // 3
  if (place_ == PlaceType::kCPU) {
    auto *t_data = tensor->mutable_data<T>(paddle::platform::CPUPlace());
    std::memcpy(static_cast<void *>(t_data), data, ele_size);
  } else if (place_ == PlaceType::kGPU) {
  // 4
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
    paddle::platform::DeviceContextPool &pool =
        paddle::platform::DeviceContextPool::Instance();
    paddle::platform::CUDAPlace gpu_place(device_);
    auto *t_data = tensor->mutable_data<T>(gpu_place);
    auto *dev_ctx = static_cast<const paddle::platform::CUDADeviceContext *>(
        pool.Get(gpu_place));

    paddle::memory::Copy(gpu_place, static_cast<void *>(t_data),
                         paddle::platform::CPUPlace(), data, ele_size,
                         dev_ctx->stream());
#else
    PADDLE_THROW(paddle::platform::errors::Unavailable(
        "Can not create tensor with CUDA place because paddle is not compiled "
        "with CUDA."));
#endif
  } else if (place_ == PlaceType::kXPU) {
   ...// 昆仑xpu相关
  } else if (place_ == PlaceType::kNPU) {
   ...// 华为昇腾相关
  } else {
    PADDLE_THROW(paddle::platform::errors::InvalidArgument(
        "The analysis predictor supports CPU, GPU, NPU and XPU now."));
  }
}

 

4.1 取出scope对应var中创建的LoDTensor指针，赋值给tensor_

// 1调用入口
EAGER_GET_TENSOR(paddle::framework::LoDTensor);
// 2 调用FindTensor获取指针
#define EAGER_GET_TENSOR(tensor_type)    \
  if (!tensor_) {                        \
    tensor_ = FindTensor<tensor_type>(); \
  }                                      \
  auto *tensor = static_cast<tensor_type *>(tensor_);
// 3 实际逻辑，在scope对应var中使用GetMutable,由于Reshape时已经调用该接口进行了创建，而且本地调用类型与创建类型一致，会直接获取之前创建的LoDTensor对象指针。
template <typename T>
void *Tensor::FindTensor() const {
  PADDLE_ENFORCE_EQ(
      name_.empty(), false,
      paddle::platform::errors::PreconditionNotMet(
          "Need to SetName first, so that the corresponding tensor can "
          "be retrieved."));
  auto *scope = static_cast<paddle::framework::Scope *>(scope_);
  auto *var = scope->FindVar(name_);
  PADDLE_ENFORCE_NOT_NULL(
      var, paddle::platform::errors::PreconditionNotMet(
               "No tensor called [%s] in the runtime scope", name_));
  auto *tensor = var->GetMutable<T>();
  return tensor;
}

 

4.2 计算需要的内存大小,实际与之前Reshap的大小以及T的类型有关。最终就是dim0*dim1*...*sizeof(T)

int64_t Tensor::numel() const { return product(dims_); }

最终通过这种递归调用模板的方式计算所有维度数据的乘积

template <size_t kStart, size_t kEnd, bool kStop>
struct UnrollProduct {
  template <typename T>
  HOSTDEVICE inline static T Run(const T *d) {
    return d[kStart] *
           UnrollProduct<kStart + 1, kEnd, kStart + 1 == kEnd>::Run(d);
  }
};

template <size_t kStart, size_t kEnd>
struct UnrollProduct<kStart, kEnd, true> {
  template <typename T>
  HOSTDEVICE inline constexpr static T Run(const T *d) {
    return 1;
  }
};

4.3 对于CPU place的数据拷贝。对于CPU比较简单，就是从tensor中拿到内存，然后将数据进行拷贝。

4.3.1 获取内存指针

// 注意，这里tensor仍将是paddle::framework::Tensor
// 1拿到内存
auto *t_data = tensor->mutable_data<T>(paddle::platform::CPUPlace());

4.3.2 获取T数据类型

// 2.1 tgensor_impl.h首先判断T是否为pod，然后获取T类型
template <typename T>
inline T* Tensor::mutable_data(const platform::Place& place,
                               size_t requested_size) {
  static_assert(std::is_pod<T>::value, "T must be POD");
  return reinterpret_cast<T*>(
      mutable_data(place, DataTypeTrait<T>::DataType(), requested_size));
}

// 2.2 具体获取类型的方法 data_type.h
  DataTypeTrait<T>::DataType()

// 工具宏，用了遍历类型
#define _ForEachDataType_(callback)                                      \
  _ForEachDataTypeHelper_(callback, float, FP32);                        \
  _ForEachDataTypeHelper_(callback, ::paddle::platform::float16, FP16);  \
  _ForEachDataTypeHelper_(callback, ::paddle::platform::bfloat16, BF16); \
  _ForEachDataTypeHelper_(callback, double, FP64);                       \
  _ForEachDataTypeHelper_(callback, int, INT32);                         \
  _ForEachDataTypeHelper_(callback, int64_t, INT64);                     \
  _ForEachDataTypeHelper_(callback, bool, BOOL);                         \
  _ForEachDataTypeHelper_(callback, uint8_t, UINT8);                     \
  _ForEachDataTypeHelper_(callback, int16_t, INT16);                     \
  _ForEachDataTypeHelper_(callback, int8_t, INT8);                       \
  _ForEachDataTypeHelper_(callback, ::paddle::platform::complex<float>,  \
                          COMPLEX64);                                    \
  _ForEachDataTypeHelper_(callback, ::paddle::platform::complex<double>, \
                          COMPLEX128);

// 首先在初始化时创建map，其中数据类型与framework.proto中VarType::Type的对应
//初始化函数，使用遍历宏调用RegisterType函数,这个函数把数据类型与proto_type对应关系写入map
static DataTypeMap* InitDataTypeMap() {
  auto retv = new DataTypeMap();

#define RegType(cc_type, proto_type) \
  RegisterType<cc_type>(retv, proto_type, #cc_type)

  _ForEachDataType_(RegType);

#undef RegType
  return retv;
}

// 然后使用宏定义创建对应的特化DataTypeTrait
// 创建特化类的宏
#define DefineDataTypeTrait(cpp_type, proto_type)                           \
  template <>                                                               \
  struct DataTypeTrait<cpp_type> {                                          \
    constexpr static proto::VarType::Type DataType() { return proto_type; } \
  }

// 使用遍历宏去调用创建宏
_ForEachDataType_(DefineDataTypeTrait);

4.3.3 申请内存

确定了具体数据类型后，会实际调用framework::Tensor->mutable_data(place, type, requested_size=0)。然后仍旧使用numel()*SizeOfType(type)获取数据大小，然后使用memory::AllocShared获取对应place的内存，保存到Tensor的holder（memory::Allocation）,同时返回内存指针

 

4.4 GPU显存的拷贝

相比与CPU内存的拷贝。GPU这里有两点不同，第一个是使用全局的DeviceContextPool，DeviceContextPool保存了place与deviceContext的对应关系。这里获取到对应的CUDADeviceContext。使用同样的方法获取内存指针后，将数据拷贝到显存中。

 

二、执行预测

执行Predictor::Run。这里实际执行的是AnalysisPredictor::ZeroCopyRun函数

bool AnalysisPredictor::ZeroCopyRun() {
  paddle::platform::SetNumThreads(config_.cpu_math_library_num_threads());
#ifdef PADDLE_WITH_MKLDNN
  ...#endif

  executor_->Run();

  if (config_.shape_range_info_collected()) {
    CollectShapeRangeInfo();
  }

  // Fix TensorArray reuse not cleaned bug.
  tensor_array_batch_cleaner_.CollectTensorArrays(sub_scope_);
  tensor_array_batch_cleaner_.ResetTensorArray();

  // recover the cpu_math_library_num_threads to 1, in order to avoid thread
  // conflict when integrating it into deployment service.
  paddle::platform::SetNumThreads(1);
#ifdef PADDLE_WITH_MKLDNN
  ...#endif
#if defined(PADDLE_WITH_MKLML)
  ...#endif
  return true;
}

1、NaiveExecutor::Run

基础逻辑是逐步调用op->Run

void NaiveExecutor::Run() {
#ifdef PADDLE_WITH_MKLDNN
  platform::AttachPointerHashToMKLDNNKey(this, place_);
#endif
  platform::ScopedFlushDenormal flush;
  for (auto &op : ops_) {
    VLOG(4) << std::this_thread::get_id() << " run "
            << op->DebugStringEx(scope_) << " on scope " << scope_;
    op->SetIsCalledByExecutor(false);
    op->Run(*scope_, place_);
  }
}

2、OP RUN

OP RUN这里会直接调用OperatorBase::Run,先从place中获取设备id，然后调用子类OP的RunImpl.所需要的资源为scope和place

void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
  try {
    VLOG(4) << place << " " << DebugStringEx(&scope);
    if (platform::is_gpu_place(place)) {
#if !defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_HIP)
      PADDLE_THROW(platform::errors::Unavailable(
          "Cannot run operator on place %s, please recompile paddle or "
          "reinstall Paddle with CUDA support.",
          place));
#else
      auto dev_id = BOOST_GET_CONST(platform::CUDAPlace, place).device;
      platform::SetDeviceId(dev_id);
#endif
    } else if (platform::is_xpu_place(place)) {
    ... //XPU相关，获取设备id
    } else if (platform::is_npu_place(place)) {
    ... //NPU相关即昇腾，获取id
    }

    {
      // TODO(wangchaochaohu) : refine code to use only one RecordEvent)
      // in order to record different op type cost time
      // and different op name cost time,we set two event.
      platform::RecordEvent op_type_record_event(Type());
      auto op_name = platform::OpName(outputs_, Type());
      platform::RecordEvent op_name_record_event(
          op_name, platform::EventRole::kUniqueOp);
      // 实际逻辑，调用子类OP的RunImpl
      RunImpl(scope, place);
    }

    VLOG(3) << GetExecutionPlace(place) << " " << DebugStringEx(&scope);
  } catch (platform::EnforceNotMet& exception) {
    ... // 错误信息
  }
}

 

OP这里分为两种，可以参考《PaddlePaddle inference 源码分析（三）》

一种OP直接继承自OperatorBase，另一种OP继承自OperatorWithKernel.

两者区别为，继承自OperatorBase的OP直接实现了RunImpl，这类OP直接运行于CPU上。继承自OperatorWithKernel的OP运行OperatorWithKernel::RunImpl,然后运行对应Kernel，这里选择Kernel的逻辑为根据place以及op_device参数。

2.1 OperatorBase类型的Runimpl

此类OP一般为功能性OP

我们以assert_op为例。它的基本逻辑为从scope中根据名称拿到输入的LoDTensor,然后从LoDTensor中获取数据,如果有数据，则取出具体数据，也就是错误信息进行打印。

void RunImpl(const framework::Scope &scope,
               const platform::Place &dev_place) const override {
    // 从scope中获取tensor
    const framework::Variable *cond_var_ptr = scope.FindVar(Input(kCond));
    PADDLE_ENFORCE_NOT_NULL(cond_var_ptr,
                            platform::errors::NotFound(
                                "Input(Condition) of AssertOp is not found."));
    const LoDTensor &cond = cond_var_ptr->Get<LoDTensor>();
    PADDLE_ENFORCE_EQ(
        cond.dims(), paddle::framework::make_ddim({1}),
        platform::errors::InvalidArgument(
            "The numel of Input(Condition) of AssertOp must be 1. But now "
            "the Condition's shape is %s.",
            cond.dims().to_str()));

    // 判断tensor中是否有数据
    bool cond_data = GetCondData(cond);
    if (cond_data) {
      return;
    }

    TensorFormatter formatter;
    formatter.SetSummarize(Attr<int64_t>(kSummarize));
    // 对数据进行处理，也就是打印错误信息
    const std::vector<std::string> &x_names = Inputs(kData);
    for (const std::string &name : x_names) {
      const framework::Variable *x_var_ptr = scope.FindVar(name);
      const framework::LoDTensor &x_tensor = x_var_ptr->Get<LoDTensor>();
      formatter.Print(x_tensor, name);
    }

    PADDLE_THROW(platform::errors::InvalidArgument(
        "The condition variable '%s' of AssertOp must be "
        "true, but received false",
        Input(kCond)));
  }

2.2 OperatorWithKernel RunImpl

绝大部分的计算型OP均为OperatorWithKernel类型。

它的调用步骤如下：OperatorWithKernel::RunImpl(scope,place)->OperatorWithKernel::RunImpl(scope,place,runtimecontext)->OperatorWithKernel::ChooseKernel->注册的kernel_func_

2.2.1 RunImpl(scope,place)

获取RuntimeContext。分为两种情况，一种允许cache rumtimecontext，那就会沿用OP保存的runtime_ctx_，否则会创建新的。

2.2.2 RunImpl(scope, place,runtime_ctx)

从单例的DeviceContextPool中根据place获取对应的DeviceContext

如果OP对象缓存的kernel_type_或kernel_func_为空，则调用ChooseKernel获取kernel_func

然后调用PrepareData获得transfer_scope

调用子类实现的InferShape设置输入输出的维度信息

调用kernel_func去实际进行计算

2.2.3 ChooseKernel(runtime_ctx,scope,place)

使用AllOpKernels获取到一个全局的map，这个map组成为<key=op名称,value=OpKernelMap>，每个OP的所有kernel实现都放在对应的OpKernelMap中。OpKernelMap的组成为<OpKernelType, OpKernelFunc>，其中OpKernelType包含了place、数据类型信息等。

首先根据Op_type拿到对应的OpKernelMap，然后根据DeviceContext获取place,并且从OP以及RuntimeContext中 获取数据类型，生成对应的OpKernelType

之后还会查看一下是否含有op_device这个参数，如果有则将place值改为op_device的属性值。然后取出对应Func

将OpKernelType以及取出的OpKernelFunc保存到op的kernen_type_以及kernel_func_

2.2.4 PrepareData

准备输入的数据，这里会创建thread local的scope,并传递出来.

然后会比较所有输入tensor.place与kernel_type_的place是否一致。如果不一致，会发送数据搬运，例如tensor为cpu数据，kernel为gpu，就会把数据从cpu搬运到gpu中。

PS：OperatorWithKernel::GetKernelTypeForVar函数用于根据tensor生成对应的OpKernelType变量kernel_type_for_var，用于与kernel_type_比较。有些OpWithKernel会自己重写GetKernelTypeForVar函数，这样返回的kernel_type_for_var与kernel_type_一致用来避免数据搬运。例如range_op（https://github.com/PaddlePaddle/Paddle/pull/25810），由于kernel运行在cpu上，在gpu模式下，保证输入数据放在cpu中，重写GetKernelTypeForVar避免搬运，来降低无谓的数据拷贝。

2.2.5 InferShape

这里会调用OP自行实现的InferShape函数，用于设置输出的维度信息

2.2.6 kernel_func

会调用op实际注册的对应类型与place的kernel函数.这里注册形式也是使用模板类，在类中实现了Compute函数，实际执行时会运行Compute函数。

 

三、获取输出数据

简单的使用方法如下所示：

std::vector<float> *out_data
auto output_names = predictor->GetOutputNames();
auto output_t = predictor->GetOutputHandle(output_names[0]);
std::vector<int> output_shape = output_t->shape();
int out_num = std::accumulate(output_shape.begin(), output_shape.end(), 1,std::multiplies<int>());

out_data->resize(out_num);
output_t->CopyToCpu(out_data->data());

1、GetOutputNames

与输入类似，都是从map中获取对应名称

//接口类实现
std::vector<std::string> Predictor::GetOutputNames() {
  return predictor_->GetOutputNames();
}

//实际实现
std::vector<std::string> AnalysisPredictor::GetOutputNames() {
  std::vector<std::string> output_names;
  for (auto &item : idx2fetches_) {
    output_names.push_back(item.second);
  }
  return output_names;
}

2、GetOutputHandle

从sub_scope中获取输出对应的tensor

//接口
std::unique_ptr<Tensor> Predictor::GetOutputHandle(const std::string &name) {
  return predictor_->GetOutputTensor(name);
}

//实际实现
std::unique_ptr<ZeroCopyTensor> AnalysisPredictor::GetOutputTensor(
    const std::string &name) {
  PADDLE_ENFORCE_NOT_NULL(
      executor_->scope()->FindVar(name),
      platform::errors::PreconditionNotMet(
          "he variable named %s is not found in the scope of the exector.",
          name));
  std::unique_ptr<ZeroCopyTensor> res(
      new ZeroCopyTensor(static_cast<void *>(executor_->scope())));
  res->input_or_output_ = false;
  res->SetName(name);
  if (platform::is_cpu_place(place_)) {
    res->SetPlace(PaddlePlace::kCPU);
  } else if (platform::is_xpu_place(place_)) {
    if (config_.lite_engine_enabled()) {
      // Currently, Paddle-Lite's XPU user interface only supports the transfer
      // of host data pointers. If it is currently used as a subgraph, execution
      // efficiency will be sacrificed, so it is temporarily set to cpu place.
      // And, the current lite engine of xpu must execute all parts of the
      // model.
      res->SetPlace(PaddlePlace::kCPU);
    } else {
      auto xpu_place = BOOST_GET_CONST(platform::XPUPlace, place_);
      res->SetPlace(PaddlePlace::kXPU, xpu_place.GetDeviceId());
    }
  } else if (platform::is_npu_place(place_)) {
    auto npu_place = BOOST_GET_CONST(platform::NPUPlace, place_);
    res->SetPlace(PaddlePlace::kNPU, npu_place.GetDeviceId());
  } else {
    auto gpu_place = BOOST_GET_CONST(platform::CUDAPlace, place_);
    res->SetPlace(PaddlePlace::kGPU, gpu_place.GetDeviceId());
  }
  return res;
}

3、CopyToCpu

实际逻辑如下：

首先拿到保存的LoDTensor，然后使用内部的memory::Allocation给data分配内存。根据place区分原数据是在cpu还是gpu上，进行拷贝

template <typename T>
void Tensor::CopyToCpuImpl(T *data, void *exec_stream, CallbackFunc cb,
                           void *cb_params) const {
  EAGER_GET_TENSOR(paddle::framework::LoDTensor);
  auto ele_num = tensor->numel();
  auto *t_data = tensor->data<T>();
  auto t_place = tensor->place();

  paddle::framework::Tensor out;
  auto mem_allocation = std::make_shared<paddle::memory::Allocation>(
      static_cast<void *>(data), ele_num * sizeof(T),
      paddle::platform::CPUPlace());
  out.ResetHolder(mem_allocation);

  if (paddle::platform::is_cpu_place(t_place)) {
#ifdef PADDLE_WITH_MKLDNN
    ...#else
    std::memcpy(static_cast<void *>(data), t_data, ele_num * sizeof(T));
#endif
  } else if (place_ == PlaceType::kGPU) {
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
    paddle::platform::DeviceContextPool &pool =
        paddle::platform::DeviceContextPool::Instance();
    auto gpu_place = BOOST_GET_CONST(paddle::platform::CUDAPlace, t_place);
    auto *dev_ctx = static_cast<const paddle::platform::CUDADeviceContext *>(
        pool.Get(gpu_place));
    paddle::memory::Copy(paddle::platform::CPUPlace(),
                         static_cast<void *>(data), gpu_place, t_data,
                         ele_num * sizeof(T), dev_ctx->stream());
#ifdef PADDLE_WITH_HIP
    hipStreamSynchronize(dev_ctx->stream());
#else
    // async, return stream
    if (nullptr != exec_stream) {
      *(static_cast<cudaStream_t *>(exec_stream)) = dev_ctx->stream();
      // async with callback
    } else if (cb) {
      cudaLaunchHostFunc(dev_ctx->stream(), cb, cb_params);
      // sync
    } else {
      cudaStreamSynchronize(dev_ctx->stream());
    }
#endif
#else
    PADDLE_THROW(paddle::platform::errors::Unavailable(
        "Can not create tensor with CUDA place because paddle is not compiled "
        "with CUDA."));
#endif
  } else if (place_ == PlaceType::kXPU) {
#ifdef PADDLE_WITH_XPU
    ...
  } else if (place_ == PlaceType::kNPU) {
#ifdef PADDLE_WITH_ASCEND_CL
    ...
  } else {
    PADDLE_THROW(paddle::platform::errors::InvalidArgument(
        "The analysis predictor supports CPU, GPU, NPU and XPU now."));
  }
}