/* Copyright (c) 2020 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. */ #include "paddle/fluid/framework/device_worker.h" #include "paddle/fluid/framework/device_worker_factory.h" #include "paddle/fluid/framework/fleet/fleet_wrapper.h" #include "paddle/fluid/framework/fleet/heter_wrapper.h" #include "paddle/fluid/platform/cpu_helper.h" #include "paddle/fluid/string/string_helper.h" #if (defined PADDLE_WITH_NCCL) && (defined PADDLE_WITH_PSLIB) #include "paddle/fluid/platform/cuda_device_guard.h" #if defined _WIN32 || defined __APPLE__ #else #define _LINUX #endif namespace paddle { namespace framework { void PSGPUWorker::Initialize(const TrainerDesc& desc) { param_ = desc.downpour_param(); mpi_rank_ = desc.mpi_rank(); trainer_desc_ = desc; /* for (int i = 0; i < trainer_desc_.xpu_recv_list_size(); ++i) { send_var_list_.push_back(trainer_desc_.xpu_recv_list(i)); } */ for (int i = 0; i < param_.sparse_table_size(); ++i) { uint64_t table_id = static_cast(param_.sparse_table(i).table_id()); TableParameter table = param_.sparse_table(i); sparse_key_names_[table_id].resize(table.sparse_key_name_size()); for (int j = 0; j < table.sparse_key_name_size(); ++j) { sparse_key_names_[table_id][j] = table.sparse_key_name(j); } sparse_value_names_[table_id].resize(table.sparse_value_name_size()); for (int j = 0; j < table.sparse_value_name_size(); ++j) { sparse_value_names_[table_id][j] = table.sparse_value_name(j); } sparse_grad_names_[table_id].resize(table.sparse_grad_name_size()); for (int j = 0; j < table.sparse_grad_name_size(); ++j) { sparse_grad_names_[table_id][j] = table.sparse_grad_name(j); } label_var_name_[table_id] = table.label_var_name(); sparse_push_keys_[table_id] = std::vector(); } for (int i = 0; i < param_.dense_table_size(); ++i) { uint64_t table_id = static_cast(param_.dense_table(i).table_id()); auto table = param_.dense_table(i); dense_value_names_[table_id].resize(table.dense_value_name_size()); for (int j = 0; j < table.dense_value_name_size(); ++j) { dense_value_names_[table_id][j] = table.dense_value_name(j); } dense_grad_names_[table_id].resize(table.dense_grad_name_size()); for (int j = 0; j < table.dense_grad_name_size(); ++j) { dense_grad_names_[table_id][j] = table.dense_grad_name(j); } } skip_ops_.resize(param_.skip_ops_size()); for (int i = 0; i < param_.skip_ops_size(); ++i) { skip_ops_[i] = param_.skip_ops(i); } for (int i = 0; i < param_.stat_var_names_size(); ++i) { stat_var_name_map_[param_.stat_var_names(i)] = 1; } need_to_push_sparse_ = param_.push_sparse(); need_to_push_dense_ = param_.push_dense(); fetch_config_ = desc.fetch_config(); use_cvm_ = desc.use_cvm(); // for sparse value accessor, embedding only no_cvm_ = desc.no_cvm(); scale_datanorm_ = desc.scale_datanorm(); dump_slot_ = desc.dump_slot(); dump_fields_.resize(desc.dump_fields_size()); for (int i = 0; i < desc.dump_fields_size(); ++i) { dump_fields_[i] = desc.dump_fields(i); } adjust_ins_weight_config_ = desc.adjust_ins_weight_config(); need_dump_param_ = false; dump_param_.resize(desc.dump_param_size()); for (int i = 0; i < desc.dump_param_size(); ++i) { dump_param_[i] = desc.dump_param(i); } if (desc.dump_param_size() != 0) { need_dump_param_ = true; } for (int i = 0; i < desc.check_nan_var_names_size(); ++i) { check_nan_var_names_.push_back(desc.check_nan_var_names(i)); } copy_table_config_ = desc.copy_table_config(); for (int i = 0; i < copy_table_config_.src_sparse_tables_size(); ++i) { uint64_t src_table = copy_table_config_.src_sparse_tables(i); uint64_t dest_table = copy_table_config_.dest_sparse_tables(i); VLOG(3) << "copy_sparse_tables_ push back " << src_table << "->" << dest_table; copy_sparse_tables_.push_back(std::make_pair(src_table, dest_table)); } for (int i = 0; i < copy_table_config_.src_dense_tables_size(); ++i) { uint64_t src_table = copy_table_config_.src_dense_tables(i); uint64_t dest_table = copy_table_config_.dest_dense_tables(i); VLOG(3) << "copy_dense_tables_ push back " << src_table << "->" << dest_table; copy_dense_tables_.push_back(std::make_pair(src_table, dest_table)); } for (auto& m : copy_table_config_.table_denpendency_map()) { if (sparse_key_names_.find(m.key()) != sparse_key_names_.end()) { // currently only support one dependency for (auto& value : m.values()) { table_dependency_[m.key()] = value; } } } // pull_queue_ = paddle::framework::MakeChannel>(); // push_queue_ = paddle::framework::MakeChannel>(); } void PSGPUWorker::SetChannelWriter(ChannelObject* queue) { writer_.Reset(queue); } void PSGPUWorker::SetNeedDump(bool need_dump_field) { need_dump_field_ = need_dump_field; } void PSGPUWorker::DumpParam() {} void PSGPUWorker::TrainFiles() { VLOG(3) << "train file A"; platform::SetNumThreads(1); VLOG(3) << "train file B"; // how to accumulate fetched values here device_reader_->Start(); VLOG(3) << "train file C"; int cur_batch; while ((cur_batch = device_reader_->Next()) > 0) { VLOG(3) << "train file D"; for (auto& op : ops_) { bool need_skip = false; for (auto t = 0u; t < skip_ops_.size(); ++t) { if (op->Type().find(skip_ops_[t]) != std::string::npos) { need_skip = true; break; } } if (!need_skip) { op->Run(*thread_scope_, place_); } } PrintFetchVars(); thread_scope_->DropKids(); } return; } void PSGPUWorker::ResetStat() { total_time_ = 0; read_time_ = 0; pack_time_ = 0; pull_sparse_local_time_ = 0; op_all_time_ = 0; xpu_op_time_ = 0; xpu_wait_time_ = 0; cpu_op_time_ = 0; collect_label_time_ = 0; fill_sparse_time_ = 0; push_sparse_time_ = 0; gpu_2_cpu_time_ = 0; cpu_2_gpu_time_ = 0; total_inst_ = 0; } void PSGPUWorker::ProduceTasks() { return; } } // end namespace framework } // end namespace paddle #endif