// Copyright (c) 2022 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. #pragma once #include "paddle/fluid/distributed/collective/ProcessGroup.h" namespace paddle { namespace distributed { // NOTE(liyurui): Notice that some backends use `stream` as an abstract // conception of hardward resource. We provide this base class allowing users to // put communications on calculation stream. In some scenorios, we found this // will save the time of switching streams. class ProcessGroupStream : public ProcessGroup { public: class TaskStream : public ProcessGroup::Task { public: // TODO(liyurui): This constructor is temporary here for compatible reason, // will be deleted soon. TaskStream(int rank, const std::vector& inputs, CommType comm_type) : Task(rank, inputs, comm_type) {} TaskStream(int rank, const std::vector& inputs, CommType comm_type, bool sync_op, bool use_calc_stream) : Task(rank, inputs, comm_type, sync_op), use_calc_stream_(use_calc_stream) {} virtual ~TaskStream() = default; protected: bool UseCalcStream() const { return use_calc_stream_; } private: bool use_calc_stream_{false}; }; ProcessGroupStream(int rank, int size, const platform::Place& place, int gid); virtual ~ProcessGroupStream() = default; virtual phi::DeviceContext* GetDeviceContext(const Place& place, bool use_calc_stream) const; std::shared_ptr AllGather( std::vector& in_tensors, // NOLINT std::vector& out_tensors, // NOLINT bool sync_op) override; virtual std::shared_ptr AllGather( std::vector& in_tensors, // NOLINT std::vector& out_tensors, // NOLINT bool sync_op, bool use_calc_stream); std::shared_ptr AllReduce( std::vector& input_tensors, // NOLINT std::vector& output_tensors, // NOLINT const AllreduceOptions& options, bool sync_op) override; virtual std::shared_ptr AllReduce( std::vector& input_tensors, // NOLINT std::vector& output_tensors, // NOLINT const AllreduceOptions& options, bool sync_op, bool use_calc_stream); std::shared_ptr AllToAll( std::vector& in_tensors, // NOLINT std::vector& out_tensors, // NOLINT bool sync_op) override; virtual std::shared_ptr AllToAll( std::vector& in_tensors, // NOLINT std::vector& out_tensors, // NOLINT bool sync_op, bool use_calc_stream); std::shared_ptr AllToAllSingle( std::vector& in_tensors, // NOLINT std::vector& out_tensors, // NOLINT std::vector& in_sizes, // NOLINT std::vector& out_sizes, // NOLINT bool sync_op) override; virtual std::shared_ptr AllToAllSingle( std::vector& in_tensors, // NOLINT std::vector& out_tensors, // NOLINT std::vector& in_sizes, // NOLINT std::vector& out_sizes, // NOLINT bool sync_op, bool use_calc_stream); std::shared_ptr Broadcast( std::vector& in_tensors, // NOLINT std::vector& out_tensors, // NOLINT const BroadcastOptions& opts, bool sync_op) override; virtual std::shared_ptr Broadcast( std::vector& in_tensors, // NOLINT std::vector& out_tensors, // NOLINT const BroadcastOptions& opts, bool sync_op, bool use_calc_stream); std::shared_ptr Reduce( std::vector& in_tensors, // NOLINT std::vector& out_tensors, // NOLINT const ReduceOptions& opts, bool sync_op) override; virtual std::shared_ptr Reduce( std::vector& in_tensors, // NOLINT std::vector& out_tensors, // NOLINT const ReduceOptions& opts, bool sync_op, bool use_calc_stream); std::shared_ptr ReduceScatter( std::vector& in_tensors, // NOLINT std::vector& out_tensors, // NOLINT const ReduceScatterOptions& opts, bool sync_op) override; virtual std::shared_ptr ReduceScatter( std::vector& in_tensors, // NOLINT std::vector& out_tensors, // NOLINT const ReduceScatterOptions& opts, bool sync_op, bool use_calc_stream); std::shared_ptr Scatter( std::vector& in_tensors, // NOLINT std::vector& out_tensors, // NOLINT const ScatterOptions& opts, bool sync_op) override; virtual std::shared_ptr Scatter( std::vector& in_tensors, // NOLINT std::vector& out_tensors, // NOLINT const ScatterOptions& opts, bool sync_op, bool use_calc_stream); std::shared_ptr Send( std::vector& tensors, // NOLINT int dst_rank, bool sync_op) override; virtual std::shared_ptr Send( std::vector& tensors, // NOLINT int dst_rank, bool sync_op, bool use_calc_stream); std::shared_ptr Send_Partial( phi::DenseTensor& tensors, // NOLINT int dst_rank, int64_t offset, int64_t length, bool sync_op) override; virtual std::shared_ptr Send_Partial( phi::DenseTensor& tensors, // NOLINT int dst_rank, int64_t offset, int64_t length, bool sync_op, bool use_calc_stream); std::shared_ptr Recv( std::vector& tensors, // NOLINT int src_rank, bool sync_op) override; virtual std::shared_ptr Recv( std::vector& tensors, // NOLINT int src_rank, bool sync_op, bool use_calc_stream); std::shared_ptr Recv_Partial( phi::DenseTensor& tensors, // NOLINT int src_rank, int64_t offset, int64_t length, bool sync_op) override; virtual std::shared_ptr Recv_Partial( phi::DenseTensor& tensors, // NOLINT int src_rank, int64_t offset, int64_t length, bool sync_op, bool use_calc_stream); std::shared_ptr AllGather_Partial( std::vector& in_tensors, std::vector& out_tensors, int64_t offset, int64_t length, bool sync_op) override; virtual std::shared_ptr AllGather_Partial( std::vector& in_tensors, // NOLINT std::vector& out_tensors, // NOLINT int64_t offset, int64_t length, bool sync_op, bool use_calc_stream); }; } // namespace distributed } // namespace paddle