ps_gpu_wrapper.cc

// Copyright (c) 2019 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.

/* Copyright (c) 2018 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. */
#ifdef PADDLE_WITH_HETERPS

#include "paddle/fluid/framework/fleet/ps_gpu_wrapper.h"

#include <algorithm>
#include <deque>

#include "paddle/fluid/framework/data_set.h"
#include "paddle/fluid/framework/fleet/heter_ps/gpu_graph_utils.h"
#include "paddle/fluid/platform/timer.h"
#if defined(PADDLE_WITH_PSCORE)
#include "paddle/fluid/distributed/ps/table/depends/feature_value.h"
#endif

DECLARE_int32(gpugraph_dedup_pull_push_mode);

namespace paddle {
namespace framework {

#ifdef PADDLE_WITH_PSLIB
void AfsWrapper::init(const std::string& fs_name,
                      const std::string& fs_user,
                      const std::string& pass_wd,
                      const std::string& conf) {
  int ret = afs_handler_.init(
      fs_name.c_str(), fs_user.c_str(), pass_wd.c_str(), conf.c_str());
  if (ret != 0) {
    LOG(ERROR) << "AFS Init Error";
  }
}

int AfsWrapper::remove(const std::string& path) {
  return afs_handler_.remove(path);
}

int AfsWrapper::mkdir(const std::string& path) {
  return afs_handler_.mkdir(path);
}

std::vector<std::string> AfsWrapper::list(const std::string& path) {
  return afs_handler_.list(path);
}

int AfsWrapper::exist(const std::string& path) {
  return afs_handler_.exist(path);
}

int AfsWrapper::upload(const std::string& local_file,
                       const std::string& afs_file) {
  return afs_handler_.upload_file(local_file, afs_file);
}

int AfsWrapper::download(const std::string& local_file,
                         const std::string& afs_file) {
  return afs_handler_.download_file(local_file, afs_file);
}

int AfsWrapper::touchz(const std::string& path) {
  return afs_handler_.touchz(path);
}

std::string AfsWrapper::cat(const std::string& path) {
  return afs_handler_.cat(path);
}

int AfsWrapper::mv(const std::string& old_path, const std::string& dest_path) {
  return afs_handler_.mv(old_path, dest_path);
}
#endif

std::shared_ptr<PSGPUWrapper> PSGPUWrapper::s_instance_ = NULL;
bool PSGPUWrapper::is_initialized_ = false;
std::mutex PSGPUWrapper::ins_mutex;
#ifdef PADDLE_WITH_PSLIB
void PSGPUWrapper::InitAfsApi(const std::string& fs_name,
                              const std::string& fs_user,
                              const std::string& pass_wd,
                              const std::string& conf) {
  int ret = afs_handler_.init(
      fs_name.c_str(), fs_user.c_str(), pass_wd.c_str(), conf.c_str());
  if (ret != 0) {
    VLOG(0) << "AFS Init Error";
  }
  use_afs_api_ = 1;
}
#endif
void PSGPUWrapper::PreBuildTask(std::shared_ptr<HeterContext> gpu_task) {
  VLOG(3) << "PSGPUWrapper::BuildGPUPSTask begin";
  platform::Timer timeline;
  timeline.Start();
  int device_num = heter_devices_.size();
  gpu_task->init(thread_keys_shard_num_, device_num, multi_mf_dim_);

  std::vector<std::thread> threads;
  // data should be in input channel

  thread_dim_keys_.resize(thread_keys_thread_num_);
  for (int i = 0; i < thread_keys_thread_num_; i++) {
    thread_dim_keys_[i].resize(thread_keys_shard_num_);
    for (int j = 0; j < thread_keys_shard_num_; j++) {
      thread_dim_keys_[i][j].resize(multi_mf_dim_);
    }
  }

  size_t total_len = 0;
  size_t len_per_thread = 0;
  int remain = 0;
  size_t begin = 0;

  std::string data_set_name = std::string(typeid(*dataset_).name());

  VLOG(0) << "gpu_graph_mode_:" << gpu_graph_mode_;
  if (!gpu_graph_mode_) {
    if (data_set_name.find("SlotRecordDataset") != std::string::npos) {
      VLOG(0) << "ps_gpu_wrapper use SlotRecordDataset";
      SlotRecordDataset* dataset = (SlotRecordDataset*)(dataset_);
      auto input_channel = dataset->GetInputChannel();
      VLOG(0) << "psgpu wrapperinputslotchannle size: "
              << input_channel->Size();
      const std::deque<SlotRecord>& vec_data = input_channel->GetData();
      total_len = vec_data.size();
      len_per_thread = total_len / thread_keys_thread_num_;
      remain = total_len % thread_keys_thread_num_;
      VLOG(0) << "total len: " << total_len;
      auto gen_dynamic_mf_func = [this](
                                     const std::deque<SlotRecord>& total_data,
                                     int begin_index,
                                     int end_index,
                                     int i) {
        for (auto iter = total_data.begin() + begin_index;
             iter != total_data.begin() + end_index;
             iter++) {
          const auto& ins = *iter;
          const auto& feasign_v = ins->slot_uint64_feasigns_.slot_values;
          const auto& slot_offset = ins->slot_uint64_feasigns_.slot_offsets;
          for (size_t slot_idx = 0; slot_idx < slot_offset_vector_.size();
               slot_idx++) {
            for (size_t j = slot_offset[slot_offset_vector_[slot_idx]];
                 j < slot_offset[slot_offset_vector_[slot_idx] + 1];
                 j++) {
              int shard_id = feasign_v[j] % thread_keys_shard_num_;
              int dim_id = slot_index_vec_[slot_idx];
              if (feasign_v[j] != 0) {
                this->thread_dim_keys_[i][shard_id][dim_id].insert(
                    feasign_v[j]);
              }
            }
          }
        }
      };
      for (int i = 0; i < thread_keys_thread_num_; i++) {
        threads.push_back(
            std::thread(gen_dynamic_mf_func,
                        std::ref(vec_data),
                        begin,
                        begin + len_per_thread + (i < remain ? 1 : 0),
                        i));

        begin += len_per_thread + (i < remain ? 1 : 0);
      }
      for (std::thread& t : threads) {
        t.join();
      }
      timeline.Pause();
      VLOG(0) << "GpuPs build task cost " << timeline.ElapsedSec()
              << " seconds.";
    } else {
      CHECK(data_set_name.find("MultiSlotDataset") != std::string::npos);
      VLOG(0) << "ps_gpu_wrapper use MultiSlotDataset";
      MultiSlotDataset* dataset = (MultiSlotDataset*)(dataset_);
      auto input_channel = dataset->GetInputChannel();

      const std::deque<Record>& vec_data = input_channel->GetData();
      total_len = vec_data.size();
      len_per_thread = total_len / thread_keys_thread_num_;
      remain = total_len % thread_keys_thread_num_;
      auto gen_func = [this](const std::deque<Record>& total_data,
                             int begin_index,
                             int end_index,
                             int i) {
        for (auto iter = total_data.begin() + begin_index;
             iter != total_data.begin() + end_index;
             iter++) {
          const auto& ins = *iter;
          const auto& feasign_v = ins.uint64_feasigns_;
          for (const auto feasign : feasign_v) {
            uint64_t cur_key = feasign.sign().uint64_feasign_;
            int shard_id = cur_key % thread_keys_shard_num_;
            this->thread_keys_[i][shard_id].insert(cur_key);
          }
        }
      };
      for (int i = 0; i < thread_keys_thread_num_; i++) {
        threads.push_back(
            std::thread(gen_func,
                        std::ref(vec_data),
                        begin,
                        begin + len_per_thread + (i < remain ? 1 : 0),
                        i));
        begin += len_per_thread + (i < remain ? 1 : 0);
      }
      for (std::thread& t : threads) {
        t.join();
      }
      timeline.Pause();
      VLOG(0) << "GpuPs build task cost " << timeline.ElapsedSec()
              << " seconds.";
    }
  } else {
    VLOG(0) << "PreBuild in GpuGraph mode";
    SlotRecordDataset* dataset = (SlotRecordDataset*)(dataset_);
    const std::vector<uint64_t>& vec_data = dataset->GetGpuGraphTotalKeys();

    total_len = vec_data.size();
    len_per_thread = total_len / thread_keys_thread_num_;
    VLOG(0) << "GpuGraphTotalKeys: " << total_len;
    remain = total_len % thread_keys_thread_num_;
    auto gen_graph_data_func = [this](const std::vector<uint64_t>& total_data,
                                      int begin_index,
                                      int end_index,
                                      int i) {
      for (auto iter = total_data.begin() + begin_index;
           iter != total_data.begin() + end_index;
           iter++) {
        uint64_t cur_key = *iter;
        int shard_id = cur_key % thread_keys_shard_num_;
        this->thread_keys_[i][shard_id].insert(cur_key);
      }
    };
    auto gen_graph_dynamic_mf_func =
        [this](const std::vector<uint64_t>& total_data,
               int begin_index,
               int end_index,
               int i) {
          for (auto iter = total_data.begin() + begin_index;
               iter != total_data.begin() + end_index;
               iter++) {
            uint64_t cur_key = *iter;
            int shard_id = cur_key % thread_keys_shard_num_;
            // TODO: feasign <-> slot <-> multi_dim
            this->thread_dim_keys_[i][shard_id][0].insert(cur_key);
          }
        };
    for (int i = 0; i < thread_keys_thread_num_; i++) {
      if (!multi_mf_dim_) {
        VLOG(1) << "psgpu graph wrapper genfunc";
        threads.push_back(
            std::thread(gen_graph_data_func,
                        std::ref(vec_data),
                        begin,
                        begin + len_per_thread + (i < remain ? 1 : 0),
                        i));
      } else {
        VLOG(1) << "psgpu graph wrapper genfunc with dynamic mf";
        threads.push_back(
            std::thread(gen_graph_dynamic_mf_func,
                        std::ref(vec_data),
                        begin,
                        begin + len_per_thread + (i < remain ? 1 : 0),
                        i));
      }
      begin += len_per_thread + (i < remain ? 1 : 0);
    }
    for (std::thread& t : threads) {
      t.join();
    }
  }

  timeline.Start();

  threads.clear();
  // merge thread_keys to shard_keys
  auto merge_ins_dynamic_mf_func = [this, gpu_task](int shard_num, int dim_id) {
    for (int i = 0; i < thread_keys_thread_num_; ++i) {
      gpu_task->batch_add_keys(
          shard_num, dim_id, thread_dim_keys_[i][shard_num][dim_id]);
      thread_dim_keys_[i][shard_num][dim_id].clear();
    }
  };
  for (int i = 0; i < thread_keys_shard_num_; ++i) {
    for (int j = 0; j < multi_mf_dim_; j++) {
      threads.push_back(std::thread(merge_ins_dynamic_mf_func, i, j));
    }
  }
  for (auto& t : threads) {
    t.join();
  }
  timeline.Pause();

  VLOG(0) << "GpuPs task add keys cost " << timeline.ElapsedSec()
          << " seconds.";
  timeline.Start();
  gpu_task->UniqueKeys();
  timeline.Pause();

  VLOG(0) << "GpuPs task unique cost " << timeline.ElapsedSec() << " seconds.";
  for (int i = 0; i < thread_keys_shard_num_; i++) {
    for (int j = 0; j < multi_mf_dim_; j++) {
      if (i == 0 && j == multi_mf_dim_ - 1) {
        gpu_task->feature_dim_keys_[i][j].push_back(0);
      }
      VLOG(0) << "GpuPs shard: " << i << "mf dim: " << index_dim_vec_[j]
              << " key len: " << gpu_task->feature_dim_keys_[i][j].size();
      gpu_task->value_dim_ptr_[i][j].resize(
          gpu_task->feature_dim_keys_[i][j].size());
    }
  }
}

void PSGPUWrapper::BuildPull(std::shared_ptr<HeterContext> gpu_task) {
  platform::Timer timeline;
  std::vector<std::future<void>> task_futures;
  int device_num = heter_devices_.size();
  auto& local_keys = gpu_task->feature_keys_;
  auto& local_ptr = gpu_task->value_ptr_;

  auto& local_dim_keys = gpu_task->feature_dim_keys_;
  auto& local_dim_ptr = gpu_task->value_dim_ptr_;

  auto& device_keys = gpu_task->device_keys_;
  auto& device_vals = gpu_task->device_values_;
  auto& device_dim_keys = gpu_task->device_dim_keys_;
  auto& device_dim_ptr = gpu_task->device_dim_ptr_;
  auto& device_dim_mutex = gpu_task->dim_mutex_;

  for (size_t dev = 0; dev < device_dim_keys.size(); dev++) {
    device_dim_keys[dev].resize(multi_mf_dim_);
    device_dim_ptr[dev].resize(multi_mf_dim_);
  }

  // auto& device_mutex = gpu_task->mutex_;

  std::vector<std::thread> threads(thread_keys_shard_num_);
#ifdef PADDLE_WITH_PSLIB
  auto fleet_ptr = FleetWrapper::GetInstance();
#endif
#ifdef PADDLE_WITH_PSCORE
  auto fleet_ptr = paddle::distributed::FleetWrapper::GetInstance();
#endif

#if (defined PADDLE_WITH_PSLIB) && (defined PADDLE_WITH_HETERPS)
  // get day_id: day nums from 1970
  struct std::tm b;
  b.tm_year = year_ - 1900;
  b.tm_mon = month_ - 1;
  b.tm_mday = day_;
  b.tm_min = b.tm_hour = b.tm_sec = 0;
  std::time_t seconds_from_1970 = std::mktime(&b);
  int day_id = seconds_from_1970 / 86400;
  fleet_ptr->pslib_ptr_->_worker_ptr->set_day_id(table_id_, day_id);
#endif

  timeline.Start();

  auto ptl_dynamic_mf_func =
      [this, &local_dim_keys, &local_dim_ptr, &fleet_ptr](int i, int j) {
        size_t key_size = local_dim_keys[i][j].size();
        int32_t status = -1;
        int32_t cnt = 0;
#ifdef PADDLE_WITH_PSLIB
        while (true) {
          auto tt = fleet_ptr->pslib_ptr_->_worker_ptr->pull_sparse_ptr(
              i,
              reinterpret_cast<char**>(local_dim_ptr[i][j].data()),
              this->table_id_,
              local_dim_keys[i][j].data(),
              key_size);
          bool flag = true;

          tt.wait();

          try {
            status = tt.get();
          } catch (const std::future_error& e) {
            VLOG(0) << "Caught a future_error with code" << e.code()
                    << ", Message:" << e.what();
          }
          if (status != 0) {
            VLOG(0) << "fleet pull sparse failed, status[" << status << "]";
            sleep(sleep_seconds_before_fail_exit_);
            flag = false;
            cnt++;
          }
          if (cnt > 3) {
            VLOG(0) << "fleet pull sparse failed, retry 3 times";
            exit(-1);
          }

          if (flag) {
            break;
          }
        }
#endif
#ifdef PADDLE_WITH_PSCORE
        while (true) {
          auto tt = fleet_ptr->worker_ptr_->PullSparsePtr(
              reinterpret_cast<char**>(local_dim_ptr[i][j].data()),
              this->table_id_,
              local_dim_keys[i][j].data(),
              key_size);
          bool flag = true;

          tt.wait();

          try {
            status = tt.get();
          } catch (const std::future_error& e) {
            VLOG(0) << "Caught a future_error with code" << e.code()
                    << ", Message:" << e.what();
          }
          if (status != 0) {
            VLOG(0) << "fleet pull sparse failed, status[" << status << "]";
            sleep(sleep_seconds_before_fail_exit_);
            flag = false;
            cnt++;
          }
          if (cnt > 3) {
            VLOG(0) << "fleet pull sparse failed, retry 3 times";
            exit(-1);
          }

          if (flag) {
            break;
          }
        }
#endif
        if (status != 0) {
          LOG(ERROR) << "fleet pull sparse failed, status[" << status << "]";
          sleep(300);
          exit(-1);
        } else {
          VLOG(0) << "FleetWrapper Pull sparse to local done with table size: "
                  << local_dim_keys[i][j].size();
        }
      };

  threads.resize(thread_keys_shard_num_ * multi_mf_dim_);
  for (int i = 0; i < thread_keys_shard_num_; i++) {
    for (int j = 0; j < multi_mf_dim_; j++) {
      task_futures.emplace_back(
          pull_thread_pool_[i]->enqueue(ptl_dynamic_mf_func, i, j));
    }
  }
  for (auto& f : task_futures) {
    f.wait();
  }
  task_futures.clear();
  timeline.Pause();
  VLOG(0) << "pull sparse from CpuPS into GpuPS cost " << timeline.ElapsedSec()
          << " seconds.";
  if (multi_node_) {
    auto gloo_wrapper = paddle::framework::GlooWrapper::GetInstance();
    if (!gloo_wrapper->IsInitialized()) {
      VLOG(0) << "GLOO is not inited";
      gloo_wrapper->Init();
    }
    gloo_wrapper->Barrier();
  }

  timeline.Start();
  std::vector<std::vector<std::pair<uint64_t, char*>>> pass_values;

  bool record_status = false;
  auto& device_task_keys = gpu_task->device_task_keys_;
  auto& device_task_ptrs = gpu_task->device_task_ptr_;
  auto build_pull_dynamic_mf_func = [this,
                                     device_num,
                                     &local_dim_keys,
                                     &local_dim_ptr,
                                     &device_dim_keys,
                                     &device_dim_ptr,
                                     &device_dim_mutex](int i, int j) {
    std::vector<std::vector<FeatureKey>> task_keys(device_num);
#ifdef PADDLE_WITH_PSLIB
    std::vector<std::vector<paddle::ps::DownpourFixedFeatureValue*>> task_ptrs(
        device_num);
#endif

#ifdef PADDLE_WITH_PSCORE
    std::vector<std::vector<paddle::distributed::FixedFeatureValue*>> task_ptrs(
        device_num);
#endif
    for (size_t k = 0; k < local_dim_keys[i][j].size(); k++) {
      int shard = local_dim_keys[i][j][k] % device_num;
      task_keys[shard].push_back(local_dim_keys[i][j][k]);
      task_ptrs[shard].push_back(local_dim_ptr[i][j][k]);
    }
    // allocate local keys to devices
    for (int dev = 0; dev < device_num; dev++) {
      device_dim_mutex[dev][j]->lock();
      int len = task_keys[dev].size();
      int cur = device_dim_keys[dev][j].size();
      device_dim_keys[dev][j].resize(device_dim_keys[dev][j].size() + len);
      device_dim_ptr[dev][j].resize(device_dim_ptr[dev][j].size() + len);
      for (int k = 0; k < len; ++k) {
        device_dim_keys[dev][j][cur + k] = task_keys[dev][k];
        device_dim_ptr[dev][j][cur + k] = task_ptrs[dev][k];
      }
      device_dim_mutex[dev][j]->unlock();
    }
  };
  auto build_func = [device_num,
                     record_status,
                     &pass_values,
                     &local_keys,
                     &local_ptr,
                     &device_task_keys,
                     &device_task_ptrs](int i) {
    auto& task_keys = device_task_keys[i];
#ifdef PADDLE_WITH_PSLIB
    auto& task_ptrs = device_task_ptrs[i];
#endif

#ifdef PADDLE_WITH_PSCORE
    auto& task_ptrs = device_task_ptrs[i];
#endif

    for (size_t j = 0; j < local_keys[i].size(); j++) {
      int shard = local_keys[i][j] % device_num;
      task_keys[shard].push_back(local_keys[i][j]);
      task_ptrs[shard].push_back(local_ptr[i][j]);
    }
#ifdef PADDLE_WITH_PSLIB
    if (record_status) {
      size_t local_keys_size = local_keys.size();
      size_t pass_values_size = pass_values.size();
      for (size_t j = 0; j < pass_values_size; j += local_keys_size) {
        auto& shard_values = pass_values[j];
        for (size_t pair_idx = 0; pair_idx < pass_values[j].size();
             pair_idx++) {
          auto& cur_pair = shard_values[pair_idx];
          int shard = cur_pair.first % device_num;
          task_keys[shard].push_back(cur_pair.first);
          task_ptrs[shard].push_back(
              (paddle::ps::DownpourFixedFeatureValue*)cur_pair.second);
        }
      }
    }
#endif
  };
  if (!multi_mf_dim_) {
    for (int i = 0; i < thread_keys_shard_num_; i++) {
      task_futures.emplace_back(hbm_thread_pool_[i]->enqueue(build_func, i));
    }
    for (auto& f : task_futures) {
      f.wait();
    }
    task_futures.clear();
    VLOG(0) << "GpuPs build hbmps done";
  }
  std::vector<std::vector<int>> prefix_sum;
  prefix_sum.resize(device_num);
  for (int i = 0; i < device_num; i++) {
    prefix_sum[i].resize(thread_keys_shard_num_ + 1);
    prefix_sum[i][0] = 0;
  }
  auto calc_prefix_func = [this,
                           &prefix_sum,
                           &device_keys,
                           &device_vals,
                           &device_task_keys](int device_num) {
    for (int j = 0; j < thread_keys_shard_num_; j++) {
      prefix_sum[device_num][j + 1] =
          prefix_sum[device_num][j] + device_task_keys[j][device_num].size();
    }
    device_keys[device_num].resize(
        prefix_sum[device_num][thread_keys_shard_num_]);
    device_vals[device_num].resize(
        prefix_sum[device_num][thread_keys_shard_num_]);
  };
  if (!multi_mf_dim_) {
    for (int i = 0; i < device_num; i++) {
      task_futures.emplace_back(
          hbm_thread_pool_[i]->enqueue(calc_prefix_func, i));
    }
    for (auto& f : task_futures) {
      f.wait();
    }
    task_futures.clear();
  }
  VLOG(0) << "prefix done";
  auto prepare_dev_value_func = [device_num,
                                 &prefix_sum,
                                 &device_keys,
                                 &device_vals,
                                 &device_task_keys,
                                 &device_task_ptrs](int dev, int shard_id) {
  // auto& task_keys = device_task_keys[shard_id];
#ifdef PADDLE_WITH_PSLIB
    auto& task_ptrs = device_task_ptrs[shard_id];
#endif

    // #ifdef PADDLE_WITH_PSCORE
    //     auto& task_ptrs = device_task_ptrs[shard_id];
    // #endif

    // int len = prefix_sum[dev][shard_id + 1] - prefix_sum[dev][shard_id];
    // int cur = prefix_sum[dev][shard_id];
#ifdef PADDLE_WITH_PSLIB
    for (int j = 0; j < len; ++j) {
      device_keys[dev][cur + j] = task_keys[dev][j];
      float* ptr_val = task_ptrs[dev][j]->data();
      FeatureValue& val = device_vals[dev][cur + j];
      size_t dim = task_ptrs[dev][j]->size();

      val.delta_score = ptr_val[1];
      val.show = ptr_val[2];
      val.clk = ptr_val[3];
      val.slot = ptr_val[6];
      val.lr = ptr_val[4];
      val.lr_g2sum = ptr_val[5];
      val.cpu_ptr = (uint64_t)(task_ptrs[dev][j]);

      if (dim > 7) {
        val.mf_size = MF_DIM + 1;
        for (int x = 0; x < val.mf_size; x++) {
          val.mf[x] = ptr_val[x + 7];
        }
      } else {
        val.mf_size = 0;
        for (int x = 0; x < MF_DIM + 1; x++) {
          val.mf[x] = 0;
        }
      }
    }
#endif
    VLOG(3) << "GpuPs build hbmps done";
  };

  if (multi_mf_dim_) {
    for (int i = 0; i < thread_keys_shard_num_; i++) {
      for (int j = 0; j < multi_mf_dim_; j++) {
        threads[i * multi_mf_dim_ + j] =
            std::thread(build_pull_dynamic_mf_func, i, j);
      }
    }
    for (std::thread& t : threads) {
      t.join();
    }
  } else {
    for (int i = 0; i < thread_keys_shard_num_; i++) {
      for (int j = 0; j < device_num; j++) {
        task_futures.emplace_back(
            hbm_thread_pool_[i]->enqueue(prepare_dev_value_func, j, i));
      }
    }
    for (auto& f : task_futures) {
      f.wait();
    }
    task_futures.clear();
  }
  timeline.Pause();
  VLOG(0) << "GpuPs prepare for build hbm cost " << timeline.ElapsedSec()
          << " seconds.";
}

void PSGPUWrapper::BuildGPUTask(std::shared_ptr<HeterContext> gpu_task) {
  int device_num = heter_devices_.size();
  platform::Timer timeline;
  timeline.Start();

  std::vector<size_t> feature_keys_count(device_num);
  size_t size_max = 0;

  for (int i = 0; i < device_num; i++) {
    for (int j = 0; j < multi_mf_dim_; j++) {
      feature_keys_count[i] += gpu_task->device_dim_ptr_[i][j].size();
      VLOG(1) << i << " card with dynamic mf dim: " << index_dim_vec_[j]
              << " dim index: " << j << " contains feasign nums: "
              << gpu_task->device_dim_ptr_[i][j].size();
    }
    VLOG(1) << i << " card with dynamic mf contains feasign nums total: "
            << feature_keys_count[i];
    size_max = std::max(size_max, feature_keys_count[i]);
  }

  if (HeterPs_) {
    delete HeterPs_;
    HeterPs_ = nullptr;
  }
  if (size_max <= 0) {
    VLOG(0) << "Skip build gpu ps cause feasign nums = " << size_max;
    return;
  }
  std::vector<std::thread> threads(device_num);
  auto accessor_wrapper_ptr =
      GlobalAccessorFactory::GetInstance().GetAccessorWrapper();
  HeterPs_ = HeterPsBase::get_instance(
      size_max, resource_, fleet_config_, accessor_class_, optimizer_type_);
#ifdef PADDLE_WITH_CUDA
  HeterPs_->set_nccl_comm_and_size(inner_comms_, inter_comms_, node_size_);
  HeterPs_->set_sparse_sgd(optimizer_config_);
  HeterPs_->set_embedx_sgd(optimizer_config_);
#endif

  auto build_dymf_mem_pool = [this, &gpu_task, &accessor_wrapper_ptr](int i,
                                                                      int j) {
    this->HeterPs_->set_multi_mf_dim(multi_mf_dim_, max_mf_dim_);
    int mf_dim = this->index_dim_vec_[j];
    VLOG(0) << "building table: " << i << "with mf dim: " << mf_dim
            << " feature_value_size:"
            << accessor_wrapper_ptr->GetFeatureValueSize(mf_dim);
    size_t feature_value_size =
        accessor_wrapper_ptr->GetFeatureValueSize(mf_dim);
    auto& device_dim_keys = gpu_task->device_dim_keys_[i][j];
    auto& device_dim_ptrs = gpu_task->device_dim_ptr_[i][j];
    size_t len = device_dim_keys.size();
    CHECK(len == device_dim_ptrs.size());
    this->mem_pools_[i * this->multi_mf_dim_ + j] =
        new MemoryPool(len, feature_value_size);
  };
  auto build_dymf_hbm_pool = [this, &gpu_task, &accessor_wrapper_ptr](int i,
                                                                      int j) {
    auto& device_dim_keys = gpu_task->device_dim_keys_[i][j];
    size_t len = device_dim_keys.size();
    int mf_dim = this->index_dim_vec_[j];
    size_t feature_value_size =
        accessor_wrapper_ptr->GetFeatureValueSize(mf_dim);

    auto& mem_pool = this->mem_pools_[i * this->multi_mf_dim_ + j];
    platform::CUDADeviceGuard guard(resource_->dev_id(i));
    this->hbm_pools_[i * this->multi_mf_dim_ + j] = new HBMMemoryPool(mem_pool);
    auto& cur_pool = this->hbm_pools_[i * this->multi_mf_dim_ + j];

    this->HeterPs_->build_ps(i,
                             device_dim_keys.data(),
                             cur_pool->mem(),
                             len,
                             feature_value_size,
                             500000,
                             2);
    if (device_dim_keys.size() > 0) {
      VLOG(3) << "show table: " << i
              << " table kv size: " << device_dim_keys.size()
              << "dim: " << mf_dim << " len: " << len;
      HeterPs_->show_one_table(i);
    }
    delete mem_pool;
  };
  int thread_num = 16;
  auto build_dynamic_mf_func = [this,
                                &gpu_task,
                                thread_num,
                                &accessor_wrapper_ptr](int i, int j, int z) {
    // this->HeterPs_->set_multi_mf_dim(multi_mf_dim_, max_mf_dim_);
    int mf_dim = this->index_dim_vec_[j];
    VLOG(0) << "building table: " << i << "with mf dim: " << mf_dim;
    auto& device_dim_keys = gpu_task->device_dim_keys_[i][j];
    auto& device_dim_ptrs = gpu_task->device_dim_ptr_[i][j];
    size_t len = device_dim_keys.size();
    CHECK(len == device_dim_ptrs.size());
    // this->mem_pools_[i * this->multi_mf_dim_ + j] =
    //    new MemoryPool(len, feature_value_size);
    auto& mem_pool = this->mem_pools_[i * this->multi_mf_dim_ + j];

    // ============ add for multi-thread ================
    size_t len_per_thread = len / thread_num;
    size_t remain = len % thread_num;
    size_t left = 0, right = 0;

    size_t real_len = len_per_thread;
    if ((size_t)z < remain) real_len++;

    if ((size_t)z < remain) {
      left = z * (len_per_thread + 1);
      right = left + real_len;
    } else {
      left = remain * (len_per_thread + 1) + (z - remain) * len_per_thread;
      right = left + real_len;
    }
    // ============ add for multi-thread ================

    for (size_t k = left; k < right; k++) {
#ifdef PADDLE_WITH_PSLIB
      float* val = (float*)(mem_pool->mem_address(k));
      float* ptr_val = device_dim_ptrs[k]->data();
      size_t dim = device_dim_ptrs[k]->size();
      val->delta_score =
          ptr_val[paddle::ps::DownpourCtrDymfAccessor::
                      DownpourCtrDymfFeatureValue::delta_score_index()];
      val->show = ptr_val[paddle::ps::DownpourCtrDymfAccessor::
                              DownpourCtrDymfFeatureValue::show_index()];
      val->clk = ptr_val[paddle::ps::DownpourCtrDymfAccessor::
                             DownpourCtrDymfFeatureValue::click_index()];
      val->slot = int(ptr_val[paddle::ps::DownpourCtrDymfAccessor::
                                  DownpourCtrDymfFeatureValue::slot_index()]);
      val->lr = ptr_val[paddle::ps::DownpourCtrDymfAccessor::
                            DownpourCtrDymfFeatureValue::embed_w_index()];
      val->lr_g2sum =
          ptr_val[paddle::ps::DownpourCtrDymfAccessor::
                      DownpourCtrDymfFeatureValue::embed_g2sum_index()];
      // TODO(xuefeng) set mf_dim while using DownpourCtrDymfAccessor
      ptr_val[paddle::ps::DownpourCtrDymfAccessor::DownpourCtrDymfFeatureValue::
                  mf_dim_index()] = float(mf_dim);
      val->mf_dim = mf_dim;
      if (dim > 8) {  // CpuPS alreay expand as mf_dim
        val->mf_size = mf_dim + 1;
        for (int x = 0; x < val->mf_dim + 1; x++) {
          val->mf[x] = ptr_val[x + 8];
        }
      } else {
        val->mf_size = 0;
        for (int x = 0; x < val->mf_dim + 1; x++) {
          val->mf[x] = 0;
        }
      }
#endif
#ifdef PADDLE_WITH_PSCORE
      void* val = mem_pool->mem_address(k);
      accessor_wrapper_ptr->BuildFill(
          val, device_dim_ptrs[k], cpu_table_accessor_, mf_dim);
#endif
    }
  };

  threads.resize(device_num * multi_mf_dim_);
  for (int i = 0; i < device_num; i++) {
    for (int j = 0; j < multi_mf_dim_; j++) {
      threads[i + j * device_num] = std::thread(build_dymf_mem_pool, i, j);
    }
  }

  for (std::thread& t : threads) {
    t.join();
  }
  threads.clear();

  // multi-thread process
  threads.resize(device_num * multi_mf_dim_ * thread_num);
  for (int i = 0; i < device_num; i++) {
    for (int j = 0; j < multi_mf_dim_; j++) {
      for (int k = 0; k < thread_num; k++) {
        threads[(i + j * device_num) * thread_num + k] =
            std::thread(build_dynamic_mf_func, i, j, k);
      }
    }
  }
  for (std::thread& t : threads) {
    t.join();
  }
  threads.clear();
  threads.resize(device_num * multi_mf_dim_);
  for (int i = 0; i < device_num; i++) {
    for (int j = 0; j < multi_mf_dim_; j++) {
      threads[i + j * device_num] = std::thread(build_dymf_hbm_pool, i, j);
    }
  }
  for (std::thread& t : threads) {
    t.join();
  }
  threads.clear();

  timeline.Pause();
  VLOG(0) << "GpuPs build table total costs: " << timeline.ElapsedSec()
          << " s.";
}

void PSGPUWrapper::LoadIntoMemory(bool is_shuffle) {
  platform::Timer timer;
  VLOG(3) << "Begin LoadIntoMemory(), dataset[" << dataset_ << "]";
  timer.Start();
  dataset_->LoadIntoMemory();
  timer.Pause();
  VLOG(0) << "LoadIntoMemory cost: " << timer.ElapsedSec() << "s";

  // local shuffle
  if (is_shuffle) {
    dataset_->LocalShuffle();
  }
  InitSlotInfo();
  gpu_graph_mode_ = dataset_->GetGpuGraphMode();
  std::shared_ptr<HeterContext> gpu_task = gpu_task_pool_.Get();
  gpu_task->Reset();

  data_ready_channel_->Put(gpu_task);

  VLOG(3) << "End LoadIntoMemory(), dataset[" << dataset_ << "]";
}

void PSGPUWrapper::start_build_thread() {
  running_ = true;
  VLOG(3) << "start build CPU ps thread.";
  pre_build_threads_ = std::thread([this] { pre_build_thread(); });
}

void PSGPUWrapper::pre_build_thread() {
  // prebuild: process load_data
  while (running_) {
    std::shared_ptr<HeterContext> gpu_task = nullptr;
    if (!data_ready_channel_->Get(gpu_task)) {
      continue;
    }
    VLOG(3) << "thread PreBuildTask start.";
    platform::Timer timer;
    timer.Start();
    // build cpu ps data process
    PreBuildTask(gpu_task);
    timer.Pause();
    VLOG(0) << "thread PreBuildTask end, cost time: " << timer.ElapsedSec()
            << " s";
    buildcpu_ready_channel_->Put(gpu_task);
  }
  VLOG(3) << "build cpu thread end";
}

void PSGPUWrapper::build_task() {
  // build_task: build_pull + build_gputask
  std::shared_ptr<HeterContext> gpu_task = nullptr;
  // train end, gpu free
  if (!gpu_free_channel_->Get(gpu_task)) {
    return;
  }
  // ins and pre_build end
  if (!buildcpu_ready_channel_->Get(gpu_task)) {
    return;
  }

  VLOG(0) << "BuildPull start.";
  platform::Timer timer;
  timer.Start();
  BuildPull(gpu_task);
  BuildGPUTask(gpu_task);
  timer.Pause();
  VLOG(0) << "BuildPull + BuildGPUTask end, cost time: " << timer.ElapsedSec()
          << "s";

  current_task_ = gpu_task;
}

void PSGPUWrapper::BeginPass() {
  platform::Timer timer;
  timer.Start();
  if (current_task_) {
    PADDLE_THROW(
        platform::errors::Fatal("[BeginPass] current task is not ended."));
  }

  debug_gpu_memory_info("befor build task");
  build_task();
  debug_gpu_memory_info("after build task");
  timer.Pause();

  if (current_task_ == nullptr) {
    PADDLE_THROW(platform::errors::Fatal(
        "[BeginPass] after build_task, current task is not null."));
  }
  if (FLAGS_gpugraph_dedup_pull_push_mode) {
    VLOG(0) << "BeginPass end, cost time: " << timer.ElapsedSec()
            << "s, enable pull push dedup mode="
            << FLAGS_gpugraph_dedup_pull_push_mode;
  } else {
    VLOG(0) << "BeginPass end, cost time: " << timer.ElapsedSec() << "s";
  }
}

void PSGPUWrapper::EndPass() {
  if (!current_task_) {
    PADDLE_THROW(
        platform::errors::Fatal("[EndPass] current task has been ended."));
  }
  platform::Timer timer;
  timer.Start();
  size_t keysize_max = 0;
  // in case of feasign_num = 0, skip dump_to_cpu

  for (size_t i = 0; i < heter_devices_.size(); i++) {
    for (int j = 0; j < multi_mf_dim_; j++) {
      keysize_max =
          std::max(keysize_max, current_task_->device_dim_keys_[i][j].size());
    }
  }
  int thread_num = 8;
  auto accessor_wrapper_ptr =
      GlobalAccessorFactory::GetInstance().GetAccessorWrapper();
  auto dump_pool_to_cpu_func = [this, thread_num, &accessor_wrapper_ptr](
                                   int i, int j, int z) {
    PADDLE_ENFORCE_GPU_SUCCESS(cudaSetDevice(this->resource_->dev_id(i)));
    auto& hbm_pool = this->hbm_pools_[i * this->multi_mf_dim_ + j];
    auto& device_keys = this->current_task_->device_dim_keys_[i][j];
    size_t len = device_keys.size();
    // ====== multi-thread process feasign================
    int len_per_thread = len / thread_num;
    int remain = len % thread_num;
    int left = -1, right = -1;
    int real_len = len_per_thread;
    if (z < remain) real_len++;
    if (z < remain) {
      left = z * (len_per_thread + 1);
      right = left + real_len;
    } else {
      left = remain * (len_per_thread + 1) + (z - remain) * len_per_thread;
      right = left + real_len;
    }
    // ============ multi-thread process feasign============
    int mf_dim = this->index_dim_vec_[j];
    size_t feature_value_size =
        accessor_wrapper_ptr->GetFeatureValueSize(mf_dim);
    VLOG(0) << "dump pool to cpu table: " << i << "with mf dim: " << mf_dim
            << " key_len :" << len
            << " feature_value_size:" << feature_value_size;
    char* test_build_values = (char*)malloc(feature_value_size * real_len);
    uint64_t offset = left * feature_value_size;
    cudaMemcpy(test_build_values,
               hbm_pool->mem() + offset,
               feature_value_size * real_len,
               cudaMemcpyDeviceToHost);
    CHECK(len == hbm_pool->capacity());
    uint64_t unuse_key = std::numeric_limits<uint64_t>::max();
    for (int i = left; i < right; ++i) {
      if (device_keys[i] == unuse_key) {
        continue;
      }
      size_t local_offset = (i - left) * feature_value_size;
      float* gpu_val = (float*)(test_build_values + local_offset);
#ifdef PADDLE_WITH_PSLIB
      // TODO: PSLIB DumpFill
#endif
#ifdef PADDLE_WITH_PSCORE
      accessor_wrapper_ptr->DumpFill(gpu_val, cpu_table_accessor_, mf_dim);
#endif
    }
    free(test_build_values);
  };
  if (multi_mf_dim_) {
    VLOG(0) << "psgpu wrapper dump pool: multi_mf_dim_: " << multi_mf_dim_;
    size_t device_num = heter_devices_.size();
    std::vector<std::thread> threads(device_num * multi_mf_dim_ * thread_num);
    for (size_t i = 0; i < device_num; i++) {
      for (int j = 0; j < multi_mf_dim_; j++) {
        for (int k = 0; k < thread_num; k++) {
          threads[(i + j * device_num) * thread_num + k] =
              std::thread(dump_pool_to_cpu_func, i, j, k);
        }
      }
    }
    for (std::thread& t : threads) {
      t.join();
    }
  }
  if (keysize_max != 0) {
    HeterPs_->end_pass();
  }

  for (size_t i = 0; i < hbm_pools_.size(); i++) {
    delete hbm_pools_[i];
  }
  gpu_task_pool_.Push(current_task_);
  current_task_ = nullptr;
  gpu_free_channel_->Put(current_task_);
  timer.Pause();
  VLOG(1) << "EndPass end, cost time: " << timer.ElapsedSec() << "s";
}

void PSGPUWrapper::PullSparse(const paddle::platform::Place& place,
                              const int table_id,
                              const std::vector<const uint64_t*>& keys,
                              const std::vector<float*>& values,
                              const std::vector<int64_t>& slot_lengths,
                              const int hidden_size) {
  VLOG(0) << "Warning:: recommand use pull_gpups_sparse op instead. This "
             "PullSparse is not used.";
}

void PSGPUWrapper::PullSparse(const paddle::platform::Place& place,
                              const int table_id,
                              const std::vector<const uint64_t*>& keys,
                              const std::vector<float*>& values,
                              const std::vector<int64_t>& slot_lengths,
                              const std::vector<int>& slot_dim,
                              const int hidden_size) {
  VLOG(3) << "Begine Gpu Ps PullSparse";
  platform::Timer all_timer;
  platform::Timer pull_gpups_timer;
  all_timer.Start();

  auto accessor_wrapper_ptr =
      GlobalAccessorFactory::GetInstance().GetAccessorWrapper();
  size_t feature_value_size =
      accessor_wrapper_ptr->GetPullValueSize(max_mf_dim_);
  VLOG(3) << "PullSparse max_dim:" << max_mf_dim_
          << " pull_feature_value_size:" << pull_type_size_;

  if (platform::is_cpu_place(place)) {
    PADDLE_THROW(platform::errors::Unimplemented(
        "Warning:: CPUPlace is not supported in GpuPs now."));
  } else if (platform::is_gpu_place(place)) {
#ifdef PADDLE_WITH_CUDA
    int device_id = place.GetDeviceId();
    int devid_2_index = HeterPs_->get_index_by_devid(device_id);
    if (FLAGS_gpugraph_dedup_pull_push_mode > 0) {
      auto& dev = device_caches_[devid_2_index];
      int slot_num = static_cast<int>(slot_lengths.size());
      std::vector<int64_t> slot_lengths_lod;
      slot_lengths_lod.reserve(slot_num + 1);
      slot_lengths_lod.push_back(0);

      int64_t total_length = 0;
      for (int i = 0; i < slot_num; ++i) {
        total_length += slot_lengths[i];
        slot_lengths_lod.push_back(total_length);
      }
      dev.total_key_length = total_length;
      VLOG(3) << "[" << device_id << "]Begin copy keys, key_num["
              << total_length << "] dedup mode";

      auto stream = dynamic_cast<platform::CUDADeviceContext*>(
                        platform::DeviceContextPool::Instance().Get(place))
                        ->stream();

      uint64_t* total_keys = dev.keys_tensor.mutable_data<uint64_t>(
          (total_length * 3) * sizeof(uint64_t), place);

      int* gpu_slot_dims = dev.dims_tensor.mutable_data<int>(
          slot_dim.size() * sizeof(int), place);
      uint64_t** gpu_keys = dev.keys_ptr_tensor.mutable_data<uint64_t*>(
          keys.size() * sizeof(uint64_t*), place);

      int64_t* slot_lens = dev.slot_lens.mutable_data<int64_t>(
          (slot_num + 1) * sizeof(int64_t), place);
      cudaMemcpyAsync(gpu_keys,
                      keys.data(),
                      keys.size() * sizeof(uint64_t*),
                      cudaMemcpyHostToDevice,
                      stream);
      cudaMemcpyAsync(slot_lens,
                      slot_lengths_lod.data(),
                      slot_lengths_lod.size() * sizeof(int64_t),
                      cudaMemcpyHostToDevice,
                      stream);

      cudaMemcpyAsync(gpu_slot_dims,
                      slot_dim.data(),
                      slot_dim.size() * sizeof(int),
                      cudaMemcpyHostToDevice,
                      stream);
      float** gpu_values = dev.values_ptr_tensor.mutable_data<float*>(
          values.size() * sizeof(float*), place);
      cudaMemcpyAsync(gpu_values,
                      values.data(),
                      values.size() * sizeof(float*),
                      cudaMemcpyHostToDevice,
                      stream);

      int* key2slot = dev.keys2slot.mutable_data<int>(
          (total_length * 5) * sizeof(int), place);

      this->CopyKeys(place,
                     gpu_keys,
                     total_keys,
                     slot_lens,
                     slot_num,
                     static_cast<int>(total_length),
                     key2slot);

      uint32_t* d_restore_idx =
          reinterpret_cast<uint32_t*>(&key2slot[total_length]);
      uint32_t* d_sorted_idx =
          reinterpret_cast<uint32_t*>(&d_restore_idx[total_length]);
      uint32_t* d_offset =
          reinterpret_cast<uint32_t*>(&d_sorted_idx[total_length]);
      uint32_t* d_merged_cnts =
          reinterpret_cast<uint32_t*>(&d_offset[total_length]);
      uint64_t* d_merged_keys =
          reinterpret_cast<uint64_t*>(&total_keys[total_length]);
      uint64_t* d_sorted_keys =
          reinterpret_cast<uint64_t*>(&d_merged_keys[total_length]);

      int dedup_size = HeterPs_->dedup_keys_and_fillidx(
          devid_2_index,
          static_cast<int>(total_length),
          total_keys,     // input
          d_merged_keys,  // output
          d_sorted_keys,  // sort keys
          d_restore_idx,  // pull fill idx
          d_sorted_idx,   // sort old idx
          d_offset,       // offset
          d_merged_cnts,
          FLAGS_gpugraph_dedup_pull_push_mode & 0x02);
      //      printf("device %d, end dedup_keys_and_fillidx total %d, "
      //              "dedup_size %d, slot num: %d, value size: %d\n",
      //             device_id, int(total_length), dedup_size, slot_num,
      //             int(feature_value_size));

      PADDLE_ENFORCE_GT(dedup_size,
                        0,
                        platform::errors::PreconditionNotMet(
                            "dedup keys need more than zero failed in BoxPS."));
      dev.dedup_key_length = dedup_size;

      int64_t total_bytes = dedup_size * feature_value_size;
      float* total_values_gpu =
          dev.pull_push_tensor.mutable_data<float>(total_bytes, place);
      pull_gpups_timer.Start();
      HeterPs_->pull_sparse(
          devid_2_index, d_merged_keys, total_values_gpu, dedup_size);

      // values.size() not sure equal slot_num
      accessor_wrapper_ptr->CopyForPull(place,
                                        total_keys,
                                        gpu_values,
                                        total_values_gpu,
                                        slot_lens,
                                        key2slot,
                                        max_mf_dim_ + 3,
                                        total_length,
                                        gpu_slot_dims,
                                        d_restore_idx,
                                        feature_value_size);
    } else {
      size_t total_length =
          std::accumulate(slot_lengths.begin(), slot_lengths.end(), 0UL);
      auto buf = memory::Alloc(place, total_length * feature_value_size);
      float* total_values_gpu = reinterpret_cast<float*>(buf->ptr());
      VLOG(3) << "Begin copy keys, key_num[" << total_length << "]";
      LoDTensor& total_keys_tensor = keys_tensor[devid_2_index];
      uint64_t* total_keys =
          reinterpret_cast<uint64_t*>(total_keys_tensor.mutable_data<int64_t>(
              {int64_t(total_length), 1}, place));
      // construct slot_level lod info
      auto slot_lengths_lod = slot_lengths;
      for (size_t i = 1; i < slot_lengths_lod.size(); i++) {
        slot_lengths_lod[i] += slot_lengths_lod[i - 1];
      }
      auto buf_key = memory::Alloc(place, keys.size() * sizeof(uint64_t*));
      auto buf_length =
          memory::Alloc(place, slot_lengths.size() * sizeof(int64_t));
      uint64_t** gpu_keys = reinterpret_cast<uint64_t**>(buf_key->ptr());
      int64_t* gpu_len = reinterpret_cast<int64_t*>(buf_length->ptr());
      cudaMemcpy(gpu_keys,
                 keys.data(),
                 keys.size() * sizeof(uint64_t*),
                 cudaMemcpyHostToDevice);
      cudaMemcpy(gpu_len,
                 slot_lengths_lod.data(),
                 slot_lengths.size() * sizeof(int64_t),
                 cudaMemcpyHostToDevice);

      auto buf_dim = memory::Alloc(place, slot_dim.size() * sizeof(int));
      int* gpu_dim = reinterpret_cast<int*>(buf_dim->ptr());
      cudaMemcpy(gpu_dim,
                 slot_dim.data(),
                 slot_dim.size() * sizeof(int),
                 cudaMemcpyHostToDevice);

      this->CopyKeys(place,
                     gpu_keys,
                     total_keys,
                     gpu_len,
                     static_cast<int>(slot_lengths.size()),
                     static_cast<int>(total_length));
      VLOG(3) << "Begin call PullSparseGPU in GPUPS, dev: " << devid_2_index
              << " len: " << total_length;

      pull_gpups_timer.Start();
      HeterPs_->pull_sparse(
          devid_2_index, total_keys, total_values_gpu, total_length);

      VLOG(3) << "Begin Copy result to tensor, total_length[" << total_length
              << "]";

      accessor_wrapper_ptr->CopyForPull(place,
                                        gpu_keys,
                                        values,
                                        total_values_gpu,
                                        gpu_len,
                                        static_cast<int>(slot_lengths.size()),
                                        hidden_size,
                                        total_length,
                                        gpu_dim,
                                        feature_value_size);
    }
    pull_gpups_timer.Pause();
#endif
  } else if (platform::is_xpu_place(place)) {
#ifdef PADDLE_WITH_XPU_KP
    VLOG(3) << "Begine Xpu Ps PullSparse";
    size_t total_length =
        std::accumulate(slot_lengths.begin(), slot_lengths.end(), 0UL);
    FeatureValue* total_values_gpu = nullptr;
    xpu_malloc(reinterpret_cast<void**>(&total_values_gpu),
               total_length * feature_value_size);
    VLOG(3) << "Begin copy keys, key_num[" << total_length << "]";
    int device_id = place.GetDeviceId();
    int devid_2_index = HeterPs_->get_index_by_devid(device_id);
    LoDTensor& total_keys_tensor = keys_tensor[devid_2_index];
    uint64_t* total_keys =
        reinterpret_cast<uint64_t*>(total_keys_tensor.mutable_data<int64_t>(
            {int64_t(total_length), 1}, place));

    // construct slot_level lod info
    auto slot_lengths_lod = slot_lengths;
    for (size_t i = 1; i < slot_lengths_lod.size(); i++) {
      slot_lengths_lod[i] += slot_lengths_lod[i - 1];
    }

    auto buf_key = memory::Alloc(place, keys.size() * sizeof(uint64_t*));
    auto buf_length =
        memory::Alloc(place, slot_lengths.size() * sizeof(int64_t));
    uint64_t** xpu_keys = reinterpret_cast<uint64_t**>(buf_key->ptr());
    int64_t* xpu_len = reinterpret_cast<int64_t*>(buf_length->ptr());
    PADDLE_ENFORCE_XPU_SUCCESS(xpu_memcpy(xpu_keys,
                                          keys.data(),
                                          keys.size() * sizeof(uint64_t*),
                                          XPU_HOST_TO_DEVICE));
    PADDLE_ENFORCE_XPU_SUCCESS(xpu_memcpy(xpu_len,
                                          slot_lengths_lod.data(),
                                          slot_lengths.size() * sizeof(int64_t),
                                          XPU_HOST_TO_DEVICE));

    this->CopyKeys(place,
                   xpu_keys,
                   total_keys,
                   xpu_len,
                   static_cast<int>(slot_lengths.size()),
                   static_cast<int>(total_length));
    VLOG(3) << "Begin call PullSparseGPU in GPUPS, dev: " << devid_2_index
            << " len: " << total_length;
    pull_gpups_timer.Start();
    HeterPs_->pull_sparse(devid_2_index,
                          total_keys,
                          total_values_gpu,
                          static_cast<int>(total_length));
    pull_gpups_timer.Pause();

    VLOG(3) << "Begin Copy result to tensor, total_length[" << total_length
            << "]";
    accessor_wrapper_ptr->CopyForPull(place,
                                      xpu_keys,
                                      values,
                                      total_values_gpu,
                                      xpu_len,
                                      static_cast<int>(slot_lengths.size()),
                                      hidden_size,
                                      total_length,
                                      feature_value_size);
#endif
  } else {
    PADDLE_THROW(platform::errors::PreconditionNotMet(
        "GpuPs/XpuPs: PullSparse Only Support CUDAPlace or XPUPlace Now."));
  }
  all_timer.Pause();
  VLOG(3) << "GpuPs PullSparse total costs: " << all_timer.ElapsedSec()
          << " s, of which GPUPS costs: " << pull_gpups_timer.ElapsedSec()
          << " s";
  VLOG(3) << "End PullSparse";
}

void PSGPUWrapper::PushSparseGrad(const paddle::platform::Place& place,
                                  const int table_id,
                                  const std::vector<const uint64_t*>& keys,
                                  const std::vector<const float*>& grad_values,
                                  const std::vector<int64_t>& slot_lengths,
                                  const int hidden_size,
                                  const int batch_size) {
  platform::Timer all_timer;
  platform::Timer push_gpups_timer;
  all_timer.Start();
  auto accessor_wrapper_ptr =
      GlobalAccessorFactory::GetInstance().GetAccessorWrapper();
  size_t grad_value_size = accessor_wrapper_ptr->GetPushValueSize(max_mf_dim_);

  if (platform::is_cpu_place(place)) {
    PADDLE_THROW(platform::errors::Unimplemented(
        "Warning:: CPUPlace is not supported in GPUPS now."));
  } else if (platform::is_gpu_place(place)) {
#ifdef PADDLE_WITH_CUDA
    int device_id = place.GetDeviceId();
    int devid_2_index = HeterPs_->get_index_by_devid(device_id);
    if (FLAGS_gpugraph_dedup_pull_push_mode > 0) {
      auto& dev = device_caches_[devid_2_index];
      int64_t total_length = dev.total_key_length;
      VLOG(3) << "Begin push sparse, key_num[" << total_length
              << "] dedup mode, device:" << device_id << ", index"
              << devid_2_index;
      auto stream = dynamic_cast<platform::CUDADeviceContext*>(
                        platform::DeviceContextPool::Instance().Get(place))
                        ->stream();
      uint64_t* total_keys = dev.keys_tensor.data<uint64_t>();
      int* slot_dims = dev.dims_tensor.data<int>();
      int slot_num = static_cast<int>(slot_lengths.size());
      if (!dev.d_slot_vector.IsInitialized()) {
        int* buf_slot_vector =
            dev.d_slot_vector.mutable_data<int>(slot_num * sizeof(int), place);
        cudaMemcpyAsync(buf_slot_vector,
                        slot_vector_.data(),
                        slot_num * sizeof(int),
                        cudaMemcpyHostToDevice,
                        stream);
      }

      const int64_t* slot_lens = dev.slot_lens.data<int64_t>();
      const int* d_slot_vector = dev.d_slot_vector.data<int>();
      const int* key2slot = dev.keys2slot.data<int>();
      float** gpu_values = dev.values_ptr_tensor.data<float*>();
      cudaMemcpyAsync(gpu_values,
                      grad_values.data(),
                      grad_values.size() * sizeof(float*),
                      cudaMemcpyHostToDevice,
                      stream);

      uint64_t* d_merged_keys = &total_keys[total_length];

      int64_t dedup_size = dev.dedup_key_length;
      int64_t total_bytes = dedup_size * grad_value_size;
      float* total_grad_values_gpu =
          dev.pull_push_tensor.mutable_data<float>(total_bytes, place);
      // dedup rate more than 3
      if (total_length > dedup_size * 3) {
        const uint32_t* d_restore_idx =
            reinterpret_cast<const uint32_t*>(&key2slot[total_length]);
        accessor_wrapper_ptr->CopyForPush(place,
                                          total_keys,
                                          gpu_values,
                                          total_grad_values_gpu,
                                          d_slot_vector,
                                          slot_lens,
                                          max_mf_dim_ + 3,
                                          total_length,
                                          dedup_size,
                                          batch_size,
                                          slot_dims,
                                          key2slot,
                                          d_restore_idx,
                                          grad_value_size);
      } else {
        const uint32_t* d_sorted_idx =
            reinterpret_cast<const uint32_t*>(&key2slot[total_length * 2]);
        const uint32_t* d_offset =
            reinterpret_cast<const uint32_t*>(&d_sorted_idx[total_length]);
        const uint32_t* d_merged_cnts =
            reinterpret_cast<const uint32_t*>(&d_offset[total_length]);
        accessor_wrapper_ptr->CopyForPush(place,
                                          d_merged_keys,
                                          gpu_values,
                                          total_grad_values_gpu,
                                          d_slot_vector,
                                          slot_lens,
                                          max_mf_dim_ + 3,
                                          total_length,
                                          dedup_size,
                                          batch_size,
                                          slot_dims,
                                          key2slot,
                                          d_sorted_idx,
                                          d_offset,
                                          d_merged_cnts,
                                          grad_value_size);
      }

      push_gpups_timer.Start();
      HeterPs_->push_sparse(devid_2_index,
                            d_merged_keys,
                            total_grad_values_gpu,
                            static_cast<int>(dedup_size));
    } else {
      int64_t total_length =
          std::accumulate(slot_lengths.begin(), slot_lengths.end(), 0UL);
      VLOG(3) << "Begin GPUPS PushSparseGrad";

      auto buf = memory::Alloc(place, total_length * grad_value_size);
      VLOG(3) << "Push Sparse Max mf dimention: " << max_mf_dim_
              << "grad_value_size:" << grad_value_size;
      float* total_grad_values_gpu = reinterpret_cast<float*>(buf->ptr());

      LoDTensor& total_keys_tensor = keys_tensor[devid_2_index];
      uint64_t* total_keys =
          reinterpret_cast<uint64_t*>(total_keys_tensor.data<int64_t>());
      VLOG(3) << "Begin copy grad tensor to gpups struct";

      accessor_wrapper_ptr->CopyForPush(place,
                                        grad_values,
                                        total_grad_values_gpu,
                                        slot_lengths,
                                        total_length,
                                        batch_size,
                                        grad_value_size,
                                        slot_vector_,
                                        slot_mf_dim_vector_);

      VLOG(3) << "Begin call PushSparseGPU in GPUPS, dev: " << devid_2_index
              << " len: " << total_length;
      push_gpups_timer.Start();
      HeterPs_->push_sparse(devid_2_index,
                            total_keys,
                            total_grad_values_gpu,
                            static_cast<int>(total_length));
    }
    push_gpups_timer.Pause();
#endif
  } else if (platform::is_xpu_place(place)) {
#ifdef PADDLE_WITH_XPU_KP
    int device_id = place.GetDeviceId();
    int devid_2_index = HeterPs_->get_index_by_devid(device_id);
    int64_t total_length =
        std::accumulate(slot_lengths.begin(), slot_lengths.end(), 0UL);
    VLOG(3) << "Begin GPUPS PushSparseGrad";

    auto buf = memory::Alloc(place, total_length * grad_value_size);
    VLOG(3) << "Push Sparse Max mf dimention: " << max_mf_dim_
            << "grad_value_size:" << grad_value_size;
    float* total_grad_values_gpu = reinterpret_cast<float*>(buf->ptr());
    LoDTensor& total_keys_tensor = keys_tensor[devid_2_index];
    uint64_t* total_keys =
        reinterpret_cast<uint64_t*>(total_keys_tensor.data<int64_t>());
    VLOG(3) << "Begin copy grad tensor to xpups struct";
    accessor_wrapper_ptr->CopyForPush(place,
                                      grad_values,
                                      total_grad_values_gpu,
                                      slot_lengths,
                                      hidden_size,
                                      total_length,
                                      batch_size,
                                      slot_vector_);

    VLOG(3) << "Begin call PushSparseXPU in XPUPS, dev: " << devid_2_index
            << " len: " << total_length;
    push_gpups_timer.Start();
    HeterPs_->push_sparse(devid_2_index,
                          total_keys,
                          total_grad_values_gpu,
                          static_cast<int>(total_length));
    push_gpups_timer.Pause();
#endif
  } else {
    PADDLE_THROW(platform::errors::PreconditionNotMet(
        "GPUPS: PushSparseGrad Only Support CUDAPlace Now."));
  }
  all_timer.Pause();
  time_3 += all_timer.ElapsedSec();
  time_4 += push_gpups_timer.ElapsedSec();
  VLOG(3) << "PushSparseGrad total cost: " << all_timer.ElapsedSec()
          << " s, of which GPUPS cost: " << push_gpups_timer.ElapsedSec()
          << " s";
  VLOG(3) << "End PushSparseGrad";
}

}  // namespace framework
}  // end namespace paddle
#endif