operator.h

/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

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. */

#pragma once

#include <boost/variant.hpp>
#include <string>
#include <unordered_map>
#include <vector>

#include "paddle/framework/attr_checker.h"
#include "paddle/framework/op_desc.pb.h"
#include "paddle/framework/op_proto.pb.h"
#include "paddle/framework/scope.h"
#include "paddle/framework/tensor.h"
#include "paddle/platform/device_context.h"
#include "paddle/platform/place.h"
#include "paddle/utils/Error.h"

namespace paddle {
namespace framework {

template <typename T>
struct EigenDeviceConverter;

template <>
struct EigenDeviceConverter<platform::CPUPlace> {
  using EigenDeviceType = Eigen::DefaultDevice;
};

#ifndef PADDLE_ONLY_CPU
template <>
struct EigenDeviceConverter<platform::GPUPlace> {
  using EigenDeviceType = Eigen::GpuDevice;
};
#endif

class OperatorBase;
using OperatorPtr = std::shared_ptr<OperatorBase>;
/**
 * OperatorBase has the basic element that Net will call to do computation.
 * Only CreateOperator from OpRegistry will new Operator directly. User
 * should always construct a proto message OpDesc and call
 * OpRegistry::CreateOp(op_desc) to get an Operator instance.
 */
class OperatorBase {
 public:
  /// If a variable is a empty variable, that name will be used.
  static std::string EMPTY_VAR_NAME() { return "@EMPTY@"; }

  /// If a variable is a temporary variable, that name will be set in Python,
  /// but it will be convert to a unique name in scope after OpCreator.
  static std::string TMP_VAR_NAME() { return "@TEMP@"; }

  virtual ~OperatorBase() {}

  template <typename T>
  inline const T& GetAttr(const std::string& name) const {
    PADDLE_ENFORCE(attrs_.count(name) != 0, "%s should be in AttributeMap",
                   name);
    return boost::get<T>(attrs_.at(name));
  }

  std::string DebugString() const;

  /// Init will be called after CreateOperator, you can put some initialization
  /// logic here.
  virtual void Init() {}

  /// InferShape infer the size of Variables used by this Operator with
  /// information inside scope
  virtual void InferShape(const ScopePtr& scope) const = 0;

  /// Net will call this function to Run an op.
  virtual void Run(const ScopePtr& scope,
                   const platform::DeviceContext& dev_ctx) const = 0;

  // Get a input with argument's name described in `op_proto`
  const std::string& Input(const std::string& name) const;
  // Get a input which has multiple variables.
  // TODO add a vector_view to prevent memory copy.
  std::vector<std::string> Inputs(const std::string& name) const;
  // Get a output with argument's name described in `op_proto`
  const std::string& Output(const std::string& name) const;
  // Get an output which has multiple variables.
  // TODO add a vector_view to prevent memory copy.
  std::vector<std::string> Outputs(const std::string& name) const;

  // init in_out_idxs_ to accelerate argument's offset lookup.
  void CreateInOutOffsetMap(const OpProto& proto);

 public:
  std::string type_;
  std::vector<std::string> inputs_;
  std::vector<std::string> outputs_;
  AttributeMap attrs_;
  // store the arguments' offset described in op_desc.
  std::unordered_map<std::string, int> in_out_idxs_;
};

class KernelContext {
 public:
  KernelContext(const OperatorBase* op, const std::shared_ptr<Scope>& scope,
                const platform::DeviceContext& device_context)
      : op_(*op), scope_(scope), device_context_(device_context) {}

  const Variable* Input(int index) const {
    return scope_->GetVariable(op_.inputs_[index]);
  }

  Variable* Output(int index) const {
    return scope_->GetVariable(op_.outputs_[index]);
  }

  const Variable* Input(const std::string& name) const {
    return scope_->GetVariable(op_.Input(name));
  }

  const Variable* Output(const std::string& name) const {
    return scope_->GetVariable(op_.Output(name));
  }

  const std::vector<const Variable*> Inputs(const std::string& name) const {
    auto names = op_.Inputs(name);
    std::vector<const Variable*> res;
    std::transform(
        names.begin(), names.end(), res.begin(),
        [this](const std::string& name) { return scope_->GetVariable(name); });
    return res;
  }

  const std::vector<const Variable*> Outputs(const std::string& name) const {
    auto names = op_.Outputs(name);
    std::vector<const Variable*> res;
    std::transform(
        names.begin(), names.end(), res.begin(),
        [this](const std::string& name) { return scope_->GetVariable(name); });
    return res;
  }

  template <typename PlaceType,
            typename DeviceType =
                typename EigenDeviceConverter<PlaceType>::EigenDeviceType>
  DeviceType* GetEigenDevice() const;

  platform::Place GetPlace() const { return device_context_.GetPlace(); }

  const OperatorBase& op_;
  const std::shared_ptr<Scope>& scope_;
  const platform::DeviceContext& device_context_;
};

class OpKernel {
 public:
  /**
   * KernelContext is the only parameter of Kernel Run function.
   * Run will get input/output variables, state such as momentum and
   * device resource such as CUDA stream, cublas handle, etc. from
   * KernelContext. User should construct it before run the Operator.
   */

  virtual void Compute(const KernelContext& context) const = 0;

  virtual ~OpKernel() {}
};

template <typename T>
struct VarToTensor {};

template <>
struct VarToTensor<Tensor*> {
  Tensor* operator()(Variable* var) { return var->GetMutable<Tensor>(); }
};

template <>
struct VarToTensor<const Tensor*> {
  const Tensor* operator()(Variable* var) { return &var->Get<Tensor>(); }
};

class OperatorWithKernel : public OperatorBase {
 public:
  struct OpKernelKey {
    platform::Place place_;

    OpKernelKey() = default;
    OpKernelKey(const platform::DeviceContext& dev_ctx) {
      place_ = dev_ctx.GetPlace();
    }

    bool operator==(const OpKernelKey& o) const { return place_ == o.place_; }
  };

  struct OpKernelHash {
    std::hash<bool> hash_;
    size_t operator()(const OpKernelKey& key) const {
      return hash_(platform::is_gpu_place(key.place_));
    }
  };

  using OpKernelMap =
      std::unordered_map<OpKernelKey, std::unique_ptr<OpKernel>, OpKernelHash>;

  void Run(const ScopePtr& scope,
           const platform::DeviceContext& dev_ctx) const final {
    auto& opKernel = AllOpKernels().at(type_).at(OpKernelKey(dev_ctx));
    opKernel->Compute(KernelContext(this, scope, dev_ctx));
  }

  static std::unordered_map<std::string /* op_type */, OpKernelMap>&
  AllOpKernels() {
    static std::unordered_map<std::string, OpKernelMap> g_all_op_kernels;
    return g_all_op_kernels;
  }

  void InferShape(const std::shared_ptr<Scope>& scope) const final {
    std::vector<const Tensor*> ins;
    VarNamesToTensors(scope, inputs_, &ins);
    std::vector<Tensor*> outs;
    VarNamesToTensors(scope, outputs_, &outs);
    InferShape(ins, outs);
  };

 private:
  template <typename T>
  void VarNamesToTensors(const std::shared_ptr<Scope>& scope,
                         const std::vector<std::string>& var_names,
                         std::vector<T>* container) const {
    container->reserve(var_names.size());
    VarToTensor<T> convert;
    for (auto& name : var_names) {
      auto var = scope->GetVariable(name);
      if (var != nullptr) {
        container->push_back(convert(var));
      } else {
        container->push_back(nullptr);
      }
    }
  }

 protected:
  virtual void InferShape(const std::vector<const Tensor*>& inputs,
                          const std::vector<Tensor*>& outputs) const = 0;
};

}  // namespace framework
}  // namespace paddle