ps_gpu_wrapper.cc 77.9 KB
Newer Older
T
Thunderbrook 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
// 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. */
T
Thunderbrook 已提交
28
#ifdef PADDLE_WITH_HETERPS
Y
yaoxuefeng 已提交
29

30 31
#include "paddle/fluid/framework/fleet/ps_gpu_wrapper.h"

T
Thunderbrook 已提交
32
#include <algorithm>
Y
yaoxuefeng 已提交
33
#include <deque>
L
lxsbupt 已提交
34
#include <unordered_set>
Y
yaoxuefeng 已提交
35

D
danleifeng 已提交
36
#include "paddle/fluid/framework/data_set.h"
D
danleifeng 已提交
37
#include "paddle/fluid/framework/fleet/heter_ps/gpu_graph_utils.h"
L
lxsbupt 已提交
38
#include "paddle/fluid/framework/fleet/heter_ps/graph_gpu_wrapper.h"
T
Thunderbrook 已提交
39
#include "paddle/fluid/platform/timer.h"
40 41 42
#if defined(PADDLE_WITH_PSCORE)
#include "paddle/fluid/distributed/ps/table/depends/feature_value.h"
#endif
T
Thunderbrook 已提交
43

D
danleifeng 已提交
44
DECLARE_int32(gpugraph_dedup_pull_push_mode);
L
lxsbupt 已提交
45
DECLARE_int32(gpugraph_storage_mode);
D
danleifeng 已提交
46

T
Thunderbrook 已提交
47 48 49
namespace paddle {
namespace framework {

T
Thunderbrook 已提交
50
#ifdef PADDLE_WITH_PSLIB
51 52 53 54 55 56
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());
T
Thunderbrook 已提交
57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86
  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);
}
87 88 89 90 91 92 93 94 95 96 97 98

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);
}
T
Thunderbrook 已提交
99 100
#endif

T
Thunderbrook 已提交
101 102
std::shared_ptr<PSGPUWrapper> PSGPUWrapper::s_instance_ = NULL;
bool PSGPUWrapper::is_initialized_ = false;
103
std::mutex PSGPUWrapper::ins_mutex;
T
Thunderbrook 已提交
104 105 106 107 108
#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) {
109 110
  int ret = afs_handler_.init(
      fs_name.c_str(), fs_user.c_str(), pass_wd.c_str(), conf.c_str());
T
Thunderbrook 已提交
111
  if (ret != 0) {
112
    VLOG(0) << "AFS Init Error";
T
Thunderbrook 已提交
113 114 115 116
  }
  use_afs_api_ = 1;
}
#endif
L
lxsbupt 已提交
117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220

void PSGPUWrapper::add_key_to_local(const std::vector<uint64_t>& vec_data) {
  size_t total_len = vec_data.size();
  size_t len_per_thread = total_len / thread_keys_thread_num_;
  size_t begin = 0;
  std::vector<std::thread> threads;

  int 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_) {
      threads.push_back(
          std::thread(gen_graph_data_func,
                      std::ref(vec_data),
                      begin,
                      begin + len_per_thread + (i < remain ? 1 : 0),
                      i));
    } else {
      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();
  }
}

void PSGPUWrapper::add_key_to_gputask(std::shared_ptr<HeterContext> gpu_task) {
  std::vector<std::thread> threads;
  platform::Timer timeline;
  timeline.Start();
  // 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();
  size_t slot_num = slot_vector_.size() - 1;
  // no slot_fea mode and whole_hbm mode, only keep one unique_sort action
  if (slot_num > 0 && FLAGS_gpugraph_storage_mode !=
                          paddle::framework::GpuGraphStorageMode::WHOLE_HBM) {
    gpu_task->UniqueKeys();
  }
  timeline.Pause();
  VLOG(0) << "GpuPs task unique cost " << timeline.ElapsedSec() << " seconds.";
}

void PSGPUWrapper::resize_gputask(std::shared_ptr<HeterContext> gpu_task) {
  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);
      }
      gpu_task->value_dim_ptr_[i][j].resize(
          gpu_task->feature_dim_keys_[i][j].size());
    }
  }
}

221
void PSGPUWrapper::PreBuildTask(std::shared_ptr<HeterContext> gpu_task) {
Y
yaoxuefeng 已提交
222
  VLOG(3) << "PSGPUWrapper::BuildGPUPSTask begin";
T
Thunderbrook 已提交
223 224
  platform::Timer timeline;
  timeline.Start();
225
  int device_num = heter_devices_.size();
Y
yaoxuefeng 已提交
226
  gpu_task->init(thread_keys_shard_num_, device_num, multi_mf_dim_);
227

Y
yaoxuefeng 已提交
228
  std::vector<std::thread> threads;
Y
yaoxuefeng 已提交
229 230 231 232 233 234 235
  // 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_);
236
    }
Y
yaoxuefeng 已提交
237
  }
Y
yaoxuefeng 已提交
238 239 240 241

  size_t total_len = 0;
  size_t len_per_thread = 0;
  int remain = 0;
Y
yaoxuefeng 已提交
242
  size_t begin = 0;
Y
yaoxuefeng 已提交
243 244 245

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

L
lxsbupt 已提交
246
  VLOG(1) << "gpu_graph_mode_:" << gpu_graph_mode_;
D
danleifeng 已提交
247 248 249
  if (!gpu_graph_mode_) {
    if (data_set_name.find("SlotRecordDataset") != std::string::npos) {
      VLOG(0) << "ps_gpu_wrapper use SlotRecordDataset";
250
      SlotRecordDataset* dataset = (SlotRecordDataset*)(dataset_);  // NOLINT
D
danleifeng 已提交
251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280
      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]);
              }
Y
yaoxuefeng 已提交
281
            }
282 283
          }
        }
D
danleifeng 已提交
284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303
      };
      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";
304
      MultiSlotDataset* dataset = (MultiSlotDataset*)(dataset_);  // NOLINT
D
danleifeng 已提交
305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334
      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);
335
      }
D
danleifeng 已提交
336 337 338 339 340 341
      for (std::thread& t : threads) {
        t.join();
      }
      timeline.Pause();
      VLOG(0) << "GpuPs build task cost " << timeline.ElapsedSec()
              << " seconds.";
Y
yaoxuefeng 已提交
342 343
    }
  } else {
L
lxsbupt 已提交
344
    SlotRecordDataset* dataset = reinterpret_cast<SlotRecordDataset*>(dataset_);
D
danleifeng 已提交
345
    const std::vector<uint64_t>& vec_data = dataset->GetGpuGraphTotalKeys();
L
lxsbupt 已提交
346 347 348 349 350 351 352
    timeline.Start();
    add_key_to_local(vec_data);
    timeline.Pause();
    VLOG(0) << "GpuGraphTotalKeys: " << vec_data.size()
            << ", add_key_to_local cost " << timeline.ElapsedSec()
            << " seconds.";
  }
Y
yaoxuefeng 已提交
353

L
lxsbupt 已提交
354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396
  add_key_to_gputask(gpu_task);
}

void PSGPUWrapper::add_slot_feature(std::shared_ptr<HeterContext> gpu_task) {
  platform::Timer timeline;
  platform::Timer time_stage;
  timeline.Start();
  // 8卡数据分片
  size_t device_num = heter_devices_.size();
  std::vector<std::thread> threads;
  size_t slot_num = slot_vector_.size() - 1;  // node slot 9008 in slot_vector
  auto& local_dim_keys = gpu_task->feature_dim_keys_;  // [shard_num, 0, keys]]
  double divide_nodeid_cost = 0;
  double get_feature_id_cost = 0;
  double add_feature_to_set_cost = 0;
  double add_feature_to_key_cost = 0;

  std::vector<std::vector<uint64_t>> node_ids(device_num);
  size_t node_num = 0;
  for (int i = 0; i < thread_keys_shard_num_; i++) {
    for (int j = 0; j < multi_mf_dim_; j++) {
      node_num += local_dim_keys[i][j].size();
    }
  }
  for (auto& node_id_vector : node_ids) {
    node_id_vector.reserve(node_num * 1.2 / device_num);
  }

  auto& device_dim_mutex = gpu_task->dim_mutex_;

  auto divide_nodeid_to_device =
      [this, device_num, &local_dim_keys, &node_ids, &device_dim_mutex](int i,
                                                                        int j) {
        std::vector<std::vector<uint64_t>> task_keys(device_num);
        size_t batch = 10000;
        for (size_t k = 0; k < device_num; k++) {
          task_keys[k].reserve(batch * 1.2 / device_num);
        }
        std::vector<int> shuffle_device = shuffle_int_vector(device_num);
        size_t start = 0;
        while (start < local_dim_keys[i][j].size()) {
          if (batch + start > local_dim_keys[i][j].size()) {
            batch = local_dim_keys[i][j].size() - start;
D
danleifeng 已提交
397
          }
L
lxsbupt 已提交
398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495
          for (size_t k = start; k < (start + batch); k++) {
            int shard = local_dim_keys[i][j][k] % device_num;
            task_keys[shard].push_back(local_dim_keys[i][j][k]);
          }
          // allocate local keys to devices
          for (auto dev : shuffle_device) {
            device_dim_mutex[dev][0]->lock();
            int len = task_keys[dev].size();
            for (int k = 0; k < len; ++k) {
              node_ids[dev].push_back(task_keys[dev][k]);
            }
            device_dim_mutex[dev][0]->unlock();
            task_keys[dev].clear();
          }
          start += batch;
        }
      };
  threads.resize(thread_keys_shard_num_ * multi_mf_dim_);
  time_stage.Start();

  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(divide_nodeid_to_device, i, j);
    }
  }
  for (std::thread& t : threads) {
    t.join();
  }
  threads.clear();
  time_stage.Pause();
  divide_nodeid_cost = time_stage.ElapsedSec();
  gpu_task->sub_graph_feas = new std::vector<GpuPsCommGraphFea>;
  std::vector<GpuPsCommGraphFea>& sub_graph_feas =
      *((std::vector<GpuPsCommGraphFea>*)gpu_task->sub_graph_feas);
  std::vector<std::vector<uint64_t>> feature_ids(device_num);
  std::vector<uint64_t*> feature_list(device_num);
  std::vector<size_t> feature_list_size(device_num);
  size_t batch = 40000;

  time_stage.Start();
  if (FLAGS_gpugraph_storage_mode ==
      paddle::framework::GpuGraphStorageMode::MEM_EMB_AND_GPU_GRAPH) {
    auto gpu_graph_ptr = GraphGpuWrapper::GetInstance();
    auto h_slot_feature_num_map = gpu_graph_ptr->slot_feature_num_map();
    int fea_num_per_node = 0;
    for (size_t i = 0; i < slot_num; ++i) {
      fea_num_per_node += h_slot_feature_num_map[i];
    }

    auto get_feature_id = [this,
                           slot_num,
                           batch,
                           fea_num_per_node,
                           &h_slot_feature_num_map,
                           &node_ids,
                           &feature_ids](int i) {
      platform::CUDADeviceGuard guard(resource_->dev_id(i));
      int* d_slot_feature_num_map;
      uint64_t* d_node_list_ptr;
      uint64_t* d_feature_list_ptr;
      CUDA_CHECK(cudaMalloc(&d_slot_feature_num_map, slot_num * sizeof(int)));
      CUDA_CHECK(cudaMemcpy(d_slot_feature_num_map,
                            h_slot_feature_num_map.data(),
                            sizeof(int) * slot_num,
                            cudaMemcpyHostToDevice));
      CUDA_CHECK(cudaMalloc(&d_node_list_ptr, batch * sizeof(uint64_t)));
      CUDA_CHECK(cudaMalloc(&d_feature_list_ptr,
                            batch * fea_num_per_node * sizeof(uint64_t)));
      auto gpu_graph_ptr = GraphGpuWrapper::GetInstance();
      uint64_t pos = 0;
      size_t real_batch = 0;
      feature_ids[i].resize(node_ids[i].size() * fea_num_per_node);
      while (pos < node_ids[i].size()) {
        real_batch = (pos + batch) <= node_ids[i].size()
                         ? batch
                         : node_ids[i].size() - pos;
        CUDA_CHECK(cudaMemcpy(d_node_list_ptr,
                              node_ids[i].data() + pos,
                              real_batch * sizeof(uint64_t),
                              cudaMemcpyHostToDevice));
        int ret = gpu_graph_ptr->get_feature_of_nodes(i,
                                                      d_node_list_ptr,
                                                      d_feature_list_ptr,
                                                      real_batch,
                                                      slot_num,
                                                      d_slot_feature_num_map,
                                                      fea_num_per_node);
        PADDLE_ENFORCE_EQ(
            ret,
            0,
            platform::errors::PreconditionNotMet("get_feature_of_nodes error"));

        CUDA_CHECK(cudaMemcpy(feature_ids[i].data() + pos * fea_num_per_node,
                              d_feature_list_ptr,
                              real_batch * fea_num_per_node * sizeof(uint64_t),
                              cudaMemcpyDeviceToHost));
        pos += real_batch;
D
danleifeng 已提交
496
      }
L
lxsbupt 已提交
497 498 499 500 501 502 503 504
      cudaFree(d_slot_feature_num_map);
      cudaFree(d_node_list_ptr);
      cudaFree(d_feature_list_ptr);
    };

    threads.resize(device_num);
    for (size_t i = 0; i < device_num; i++) {
      threads[i] = std::thread(get_feature_id, i);
Y
yaoxuefeng 已提交
505 506 507 508
    }
    for (std::thread& t : threads) {
      t.join();
    }
L
lxsbupt 已提交
509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605
    threads.clear();
    for (size_t i = 0; i < device_num; i++) {
      feature_list[i] = feature_ids[i].data();
      feature_list_size[i] = feature_ids[i].size();
    }
  } else if (FLAGS_gpugraph_storage_mode ==
                 paddle::framework::GpuGraphStorageMode::
                     MEM_EMB_FEATURE_AND_GPU_GRAPH ||
             FLAGS_gpugraph_storage_mode ==
                 paddle::framework::GpuGraphStorageMode::
                     SSD_EMB_AND_MEM_FEATURE_GPU_GRAPH) {
    auto gpu_graph_ptr = GraphGpuWrapper::GetInstance();
    sub_graph_feas = gpu_graph_ptr->get_sub_graph_fea(node_ids, slot_num);
    for (size_t i = 0; i < device_num; i++) {
      feature_list[i] = sub_graph_feas[i].feature_list;
      feature_list_size[i] = sub_graph_feas[i].feature_size;
    }
  } else {
    VLOG(0) << "FLAGS_gpugraph_storage_mode is not adaptived";
  }
  time_stage.Pause();
  get_feature_id_cost = time_stage.ElapsedSec();
  size_t feature_num = 0;
  for (size_t i = 0; i < device_num; i++) {
    feature_num += feature_list_size[i];
  }
  VLOG(0) << "feature_num is " << feature_num << " node_num num is "
          << node_num;

  size_t set_num = thread_keys_shard_num_;
  std::vector<std::unordered_set<uint64_t>> feature_id_set(set_num);
  std::vector<std::mutex> set_mutex(set_num);

  auto add_feature_to_set =
      [this, set_num, &feature_list, &feature_id_set, &set_mutex](
          int dev, size_t start, size_t end) {
        size_t batch = 10000 * set_num;
        std::vector<std::vector<uint64_t>> feature_list_tmp(set_num);
        for (size_t i = 0; i < set_num; i++) {
          feature_list_tmp[i].reserve((batch * 1.2) / set_num);
        }
        std::vector<int> shuffle_set_index = shuffle_int_vector(set_num);
        size_t pos = start;
        size_t real_batch = 0;
        while (pos < end) {
          real_batch = (pos + batch <= end) ? batch : end - pos;
          for (size_t i = pos; i < pos + real_batch; i++) {
            if (feature_list[dev][i] == 0) {
              continue;
            }
            int shard_num = feature_list[dev][i] % set_num;
            feature_list_tmp[shard_num].push_back(feature_list[dev][i]);
          }
          // uniq in local
          for (size_t i = 0; i < set_num; i++) {
            std::sort(feature_list_tmp[i].begin(), feature_list_tmp[i].end());
            size_t idx = 0;
            size_t total = feature_list_tmp[i].size();
            for (size_t j = 0; j < total; j++) {
              auto& k = feature_list_tmp[i][j];
              if (idx > 0 && feature_list_tmp[i][idx - 1] == k) {
                continue;
              }
              feature_list_tmp[i][idx] = k;
              ++idx;
            }
            feature_list_tmp[i].resize(idx);
          }
          // uniq in global
          for (auto set_index : shuffle_set_index) {
            set_mutex[set_index].lock();
            for (auto feature_id : feature_list_tmp[set_index]) {
              feature_id_set[set_index].insert(feature_id);
            }
            set_mutex[set_index].unlock();
            feature_list_tmp[set_index].clear();
          }
          pos += real_batch;
        }
      };
  size_t device_thread_num = 8;
  threads.resize(device_num * device_thread_num);
  time_stage.Start();
  for (size_t i = 0; i < device_num; i++) {
    size_t start = 0;
    for (size_t j = 0; j < device_thread_num; j++) {
      size_t batch = feature_list_size[i] / device_thread_num;
      if (j < feature_list_size[i] % device_thread_num) {
        batch += 1;
      }
      threads[i * device_thread_num + j] =
          std::thread(add_feature_to_set, i, start, start + batch);
      start += batch;
    }
  }
  for (std::thread& t : threads) {
    t.join();
Y
yaoxuefeng 已提交
606
  }
607
  threads.clear();
L
lxsbupt 已提交
608 609 610 611 612 613 614 615 616 617 618 619 620
  time_stage.Pause();
  add_feature_to_set_cost = time_stage.ElapsedSec();
  auto add_feature_to_key = [this,
                             device_num,
                             &feature_id_set,
                             &local_dim_keys,
                             set_num](int shard_num, int j) {
    local_dim_keys[shard_num][j].reserve(local_dim_keys[shard_num][j].size() +
                                         feature_id_set[shard_num].size());
    for (auto it = feature_id_set[shard_num].begin();
         it != feature_id_set[shard_num].end();
         it++) {
      local_dim_keys[shard_num][j].push_back(*it);
621
    }
L
lxsbupt 已提交
622
    feature_id_set[shard_num].clear();
623
  };
L
lxsbupt 已提交
624 625 626
  time_stage.Start();
  threads.resize(thread_keys_shard_num_ * multi_mf_dim_);
  for (int i = 0; i < thread_keys_shard_num_; i++) {
Y
yaoxuefeng 已提交
627
    for (int j = 0; j < multi_mf_dim_; j++) {
L
lxsbupt 已提交
628
      threads[i * multi_mf_dim_ + j] = std::thread(add_feature_to_key, i, j);
629
    }
630
  }
L
lxsbupt 已提交
631
  for (std::thread& t : threads) {
632
    t.join();
Y
yaoxuefeng 已提交
633
  }
L
lxsbupt 已提交
634 635 636
  time_stage.Pause();
  add_feature_to_key_cost = time_stage.ElapsedSec();
  threads.clear();
Y
yaoxuefeng 已提交
637
  timeline.Pause();
L
lxsbupt 已提交
638 639 640 641 642
  VLOG(0) << " add_slot_feature costs: " << timeline.ElapsedSec() << " s."
          << " divide_nodeid_cost " << divide_nodeid_cost
          << " get_feature_id_cost " << get_feature_id_cost
          << " add_feature_to_set_cost " << add_feature_to_set_cost
          << " add_feature_to_key_cost " << add_feature_to_key_cost;
643 644 645 646
}

void PSGPUWrapper::BuildPull(std::shared_ptr<HeterContext> gpu_task) {
  platform::Timer timeline;
L
lxsbupt 已提交
647 648 649 650 651 652 653 654 655 656 657 658 659 660
  size_t slot_num = slot_vector_.size() - 1;  // node slot 9008 in slot_vector
  if (slot_num > 0 && FLAGS_gpugraph_storage_mode !=
                          paddle::framework::GpuGraphStorageMode::WHOLE_HBM) {
    add_slot_feature(gpu_task);
  }

  resize_gputask(gpu_task);

  platform::Timer time_stage;
  time_stage.Start();
  gpu_task->UniqueKeys();
  time_stage.Pause();
  VLOG(0) << "BuildPull slot feature uniq and sort cost time: "
          << time_stage.ElapsedSec();
661

662 663 664 665 666
  auto& local_dim_keys = gpu_task->feature_dim_keys_;
  auto& local_dim_ptr = gpu_task->value_dim_ptr_;

  auto& device_dim_keys = gpu_task->device_dim_keys_;
  auto& device_dim_ptr = gpu_task->device_dim_ptr_;
Y
yaoxuefeng 已提交
667 668 669 670

  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_);
671
  }
Y
yaoxuefeng 已提交
672

T
Thunderbrook 已提交
673
  // auto& device_mutex = gpu_task->mutex_;
674 675 676 677 678 679

  std::vector<std::thread> threads(thread_keys_shard_num_);
#ifdef PADDLE_WITH_PSLIB
  auto fleet_ptr = FleetWrapper::GetInstance();
#endif
#ifdef PADDLE_WITH_PSCORE
680
  auto fleet_ptr = paddle::distributed::FleetWrapper::GetInstance();
681
#endif
682

683
#if (defined PADDLE_WITH_PSLIB) && (defined PADDLE_WITH_HETERPS)
684 685 686 687 688 689 690 691 692 693 694
  // 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

695
  timeline.Start();
696

697
  auto ptl_dynamic_mf_func =
L
lxsbupt 已提交
698 699
      [this, &local_dim_keys, &local_dim_ptr, &fleet_ptr, &gpu_task](int i,
                                                                     int j) {
700 701 702
        size_t key_size = local_dim_keys[i][j].size();
        int32_t status = -1;
        int32_t cnt = 0;
703
#ifdef PADDLE_WITH_PSLIB
704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730
        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);
          }
731

732 733 734 735
          if (flag) {
            break;
          }
        }
736 737
#endif
#ifdef PADDLE_WITH_PSCORE
738 739
        while (true) {
          auto tt = fleet_ptr->worker_ptr_->PullSparsePtr(
L
lxsbupt 已提交
740
              i,
741 742 743
              reinterpret_cast<char**>(local_dim_ptr[i][j].data()),
              this->table_id_,
              local_dim_keys[i][j].data(),
L
lxsbupt 已提交
744 745
              key_size,
              gpu_task->pass_id_);
746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765
          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);
          }
766

767 768 769 770
          if (flag) {
            break;
          }
        }
771
#endif
772 773 774 775 776
        if (status != 0) {
          LOG(ERROR) << "fleet pull sparse failed, status[" << status << "]";
          sleep(300);
          exit(-1);
        } else {
L
lxsbupt 已提交
777
          VLOG(1) << "FleetWrapper Pull sparse to local done with table size: "
778 779 780
                  << local_dim_keys[i][j].size();
        }
      };
Y
yaoxuefeng 已提交
781 782

  threads.resize(thread_keys_shard_num_ * multi_mf_dim_);
L
lxsbupt 已提交
783 784 785

  uint64_t total_key = 0;
  std::vector<std::future<void>> task_futures;
Y
yaoxuefeng 已提交
786 787 788 789
  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));
L
lxsbupt 已提交
790
      total_key += local_dim_keys[i][j].size();
791
    }
792
  }
Y
yaoxuefeng 已提交
793 794
  for (auto& f : task_futures) {
    f.wait();
795
  }
Y
yaoxuefeng 已提交
796
  task_futures.clear();
797
  timeline.Pause();
L
lxsbupt 已提交
798 799
  VLOG(0) << "pull sparse from CpuPS into GpuPS total keys " << total_key
          << ", cost " << timeline.ElapsedSec() << " seconds.";
Y
yaoxuefeng 已提交
800 801 802 803 804 805 806 807
  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();
  }
L
lxsbupt 已提交
808
}
809

L
lxsbupt 已提交
810 811 812 813 814 815 816
void PSGPUWrapper::divide_to_device(std::shared_ptr<HeterContext> gpu_task) {
  platform::Timer timeline;
  int device_num = heter_devices_.size();
  std::vector<std::thread> threads;
  std::vector<std::future<void>> task_futures;
  auto& local_dim_keys = gpu_task->feature_dim_keys_;
  auto& local_dim_ptr = gpu_task->value_dim_ptr_;
Y
yaoxuefeng 已提交
817

L
lxsbupt 已提交
818 819 820 821 822 823 824 825 826 827 828 829 830
  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_;
  // auto& device_mutex = gpu_task->mutex_;

  if (multi_mf_dim_) {
    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_);
    }
  }

  timeline.Start();
831 832 833 834 835
  auto build_pull_dynamic_mf_func = [this,
                                     device_num,
                                     &local_dim_keys,
                                     &local_dim_ptr,
                                     &device_dim_keys,
Y
yaoxuefeng 已提交
836 837
                                     &device_dim_ptr,
                                     &device_dim_mutex](int i, int j) {
838
    std::vector<std::vector<FeatureKey>> task_keys(device_num);
839
#ifdef PADDLE_WITH_PSLIB
840 841
    std::vector<std::vector<paddle::ps::DownpourFixedFeatureValue*>> task_ptrs(
        device_num);
842 843 844 845 846 847
#endif

#ifdef PADDLE_WITH_PSCORE
    std::vector<std::vector<paddle::distributed::FixedFeatureValue*>> task_ptrs(
        device_num);
#endif
848 849 850 851 852
    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]);
    }
Y
yaoxuefeng 已提交
853
    // allocate local keys to devices
L
lxsbupt 已提交
854 855
    std::vector<int> shuffle_device = shuffle_int_vector(device_num);
    for (auto dev : shuffle_device) {
Y
yaoxuefeng 已提交
856 857 858 859 860 861 862 863
      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];
864
      }
Y
yaoxuefeng 已提交
865
      device_dim_mutex[dev][j]->unlock();
866 867
    }
  };
L
lxsbupt 已提交
868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904

  if (multi_mf_dim_) {
    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++) {
        threads[i * multi_mf_dim_ + j] =
            std::thread(build_pull_dynamic_mf_func, i, j);
      }
    }
    for (std::thread& t : threads) {
      t.join();
    }
  }
  timeline.Pause();
  VLOG(0) << "GpuPs prepare for build hbm cost " << timeline.ElapsedSec()
          << " seconds.";
}

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

  auto& device_keys = gpu_task->device_keys_;
  auto& device_vals = gpu_task->device_values_;
  // auto& device_mutex = gpu_task->mutex_;

  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_;

905 906 907 908 909 910 911
  auto build_func = [device_num,
                     record_status,
                     &pass_values,
                     &local_keys,
                     &local_ptr,
                     &device_task_keys,
                     &device_task_ptrs](int i) {
T
Thunderbrook 已提交
912
    auto& task_keys = device_task_keys[i];
T
Thunderbrook 已提交
913
#ifdef PADDLE_WITH_PSLIB
T
Thunderbrook 已提交
914
    auto& task_ptrs = device_task_ptrs[i];
T
Thunderbrook 已提交
915 916 917
#endif

#ifdef PADDLE_WITH_PSCORE
T
Thunderbrook 已提交
918
    auto& task_ptrs = device_task_ptrs[i];
T
Thunderbrook 已提交
919
#endif
920 921 922 923 924 925

    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]);
    }
926
#ifdef PADDLE_WITH_PSLIB
Y
yaoxuefeng 已提交
927 928 929 930 931 932 933 934 935 936 937 938 939 940 941
    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);
        }
      }
    }
942
#endif
T
Thunderbrook 已提交
943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959
  };
  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;
  }
960 961 962 963
  auto calc_prefix_func = [this,
                           &prefix_sum,
                           &device_keys,
                           &device_vals,
T
Thunderbrook 已提交
964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984
                           &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";
985 986 987 988 989
  auto prepare_dev_value_func = [device_num,
                                 &prefix_sum,
                                 &device_keys,
                                 &device_vals,
                                 &device_task_keys,
T
Thunderbrook 已提交
990 991 992
                                 &device_task_ptrs](int dev, int shard_id) {
#ifdef PADDLE_WITH_PSLIB
    auto& task_ptrs = device_task_ptrs[shard_id];
993

T
Thunderbrook 已提交
994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016
    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;
Y
yaoxuefeng 已提交
1017 1018
        }
      }
T
Thunderbrook 已提交
1019
    }
T
Thunderbrook 已提交
1020
#endif
T
Thunderbrook 已提交
1021
    VLOG(3) << "GpuPs build hbmps done";
Y
yaoxuefeng 已提交
1022
  };
L
lxsbupt 已提交
1023
  if (!multi_mf_dim_) {
T
Thunderbrook 已提交
1024 1025 1026 1027 1028 1029 1030 1031 1032 1033
    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();
Y
yaoxuefeng 已提交
1034 1035
  }
  timeline.Pause();
T
Thunderbrook 已提交
1036
  VLOG(0) << "GpuPs prepare for build hbm cost " << timeline.ElapsedSec()
1037
          << " seconds.";
Y
yaoxuefeng 已提交
1038 1039
}

1040
void PSGPUWrapper::BuildGPUTask(std::shared_ptr<HeterContext> gpu_task) {
1041
  int device_num = heter_devices_.size();
L
lxsbupt 已提交
1042 1043
  platform::Timer stagetime;
  stagetime.Start();
T
Thunderbrook 已提交
1044

1045
  std::vector<size_t> feature_keys_count(device_num);
T
Thunderbrook 已提交
1046
  size_t size_max = 0;
Y
yaoxuefeng 已提交
1047 1048 1049 1050 1051 1052 1053

  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();
1054
    }
L
lxsbupt 已提交
1055
    VLOG(0) << i << " card with dynamic mf contains feasign nums total: "
Y
yaoxuefeng 已提交
1056 1057
            << feature_keys_count[i];
    size_max = std::max(size_max, feature_keys_count[i]);
T
Thunderbrook 已提交
1058
  }
1059
  if (size_max <= 0) {
1060
    VLOG(0) << "Skip build gpu ps cause feasign nums = " << size_max;
1061 1062
    return;
  }
1063
  std::vector<std::thread> threads(device_num);
D
danleifeng 已提交
1064
  auto accessor_wrapper_ptr =
D
danleifeng 已提交
1065
      GlobalAccessorFactory::GetInstance().GetAccessorWrapper();
L
lxsbupt 已提交
1066 1067 1068
  if (HeterPs_ == NULL) {
    HeterPs_ = HeterPsBase::get_instance(
        size_max, resource_, fleet_config_, accessor_class_, optimizer_type_);
F
Fan Zhang 已提交
1069
#ifdef PADDLE_WITH_CUDA
L
lxsbupt 已提交
1070 1071 1072 1073
    HeterPs_->set_nccl_comm_and_size(
        inner_comms_, inter_comms_, node_size_, rank_id_);
    HeterPs_->set_sparse_sgd(optimizer_config_);
    HeterPs_->set_embedx_sgd(optimizer_config_);
F
Fan Zhang 已提交
1074
#endif
L
lxsbupt 已提交
1075 1076 1077 1078 1079 1080
  }
  stagetime.Pause();
  VLOG(0) << "card: "
          << " BuildGPUTask create HeterPs_ costs: " << stagetime.ElapsedSec()
          << " s.";
  stagetime.Start();
Z
zmxdream 已提交
1081

L
lxsbupt 已提交
1082 1083 1084
  auto build_dynamic_mf_func = [this, &gpu_task, &accessor_wrapper_ptr](
                                   int i, int j, size_t start, size_t end) {
    // this->HeterPs_->set_multi_mf_dim(multi_mf_dim_, max_mf_dim_);
Y
yaoxuefeng 已提交
1085
    auto& device_dim_ptrs = gpu_task->device_dim_ptr_[i][j];
Z
zmxdream 已提交
1086 1087
    int mf_dim = this->index_dim_vec_[j];
    size_t feature_value_size =
D
danleifeng 已提交
1088
        accessor_wrapper_ptr->GetFeatureValueSize(mf_dim);
L
lxsbupt 已提交
1089 1090 1091 1092 1093
    size_t real_len = end - start;
    std::shared_ptr<char> build_values(new char[feature_value_size * real_len],
                                       [](char* p) { delete[] p; });
    char* test_build_values = build_values.get();
    for (size_t k = start; k < end; k++) {
D
danleifeng 已提交
1094
#ifdef PADDLE_WITH_PSLIB
L
lxsbupt 已提交
1095 1096
      float* val = reinterpret_cast<float*>(test_build_values +
                                            (k - start) * feature_value_size);
Y
yaoxuefeng 已提交
1097 1098 1099 1100 1101 1102 1103 1104 1105
      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()];
1106 1107
      val->slot = int(ptr_val[paddle::ps::DownpourCtrDymfAccessor::  // NOLINT
                              DownpourCtrDymfFeatureValue::slot_index()]);
Y
yaoxuefeng 已提交
1108 1109 1110 1111 1112
      val->lr = ptr_val[paddle::ps::DownpourCtrDymfAccessor::
                            DownpourCtrDymfFeatureValue::embed_w_index()];
      val->lr_g2sum =
          ptr_val[paddle::ps::DownpourCtrDymfAccessor::
                      DownpourCtrDymfFeatureValue::embed_g2sum_index()];
Y
yaoxuefeng 已提交
1113
      // TODO(xuefeng) set mf_dim while using DownpourCtrDymfAccessor
Y
yaoxuefeng 已提交
1114
      ptr_val[paddle::ps::DownpourCtrDymfAccessor::DownpourCtrDymfFeatureValue::
1115
                  mf_dim_index()] = float(mf_dim);  // NOLINT
Y
yaoxuefeng 已提交
1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127
      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;
        }
      }
L
lxsbupt 已提交
1128 1129 1130 1131
      VLOG(5) << "build " << k << " : "
              << feature_value_accessor_.ParseToString(
                     val,
                     feature_value_accessor_.common_feature_value.Dim(mf_dim));
D
danleifeng 已提交
1132 1133
#endif
#ifdef PADDLE_WITH_PSCORE
L
lxsbupt 已提交
1134 1135
      void* val = reinterpret_cast<float*>(test_build_values +
                                           (k - start) * feature_value_size);
D
danleifeng 已提交
1136 1137 1138
      accessor_wrapper_ptr->BuildFill(
          val, device_dim_ptrs[k], cpu_table_accessor_, mf_dim);
#endif
Y
yaoxuefeng 已提交
1139
    }
L
lxsbupt 已提交
1140 1141 1142 1143 1144 1145 1146 1147
    task_info task;
    task.build_values = build_values;
    task.offset = start;
    task.device_id = i;
    task.multi_mf_dim = j;
    task.start = 0;
    task.end = real_len;
    cpu_reday_channels_[i]->Put(task);
Z
zmxdream 已提交
1148
  };
Y
yaoxuefeng 已提交
1149

L
lxsbupt 已提交
1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162
  auto build_dymf_hbm_pool = [this,
                              &gpu_task,
                              &accessor_wrapper_ptr,
                              &feature_keys_count](int i) {
    platform::CUDADeviceGuard guard(resource_->dev_id(i));
    // reset table
    this->HeterPs_->reset_table(i,
                                feature_keys_count[i],
                                optimizer_config_,
                                optimizer_config_,
                                infer_mode_);
    // insert hbm table
    std::vector<std::thread> threads(multi_mf_dim_);
Z
zmxdream 已提交
1163
    for (int j = 0; j < multi_mf_dim_; j++) {
L
lxsbupt 已提交
1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207
      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);
      this->hbm_pools_[i * this->multi_mf_dim_ + j]->reset(len,
                                                           feature_value_size);

      auto build_ps_thread =
          [this, &gpu_task](
              int i, int j, size_t len, size_t feature_value_size) {
            auto& device_dim_keys = gpu_task->device_dim_keys_[i][j];
            this->HeterPs_->build_ps(
                i,
                device_dim_keys.data(),
                this->hbm_pools_[i * this->multi_mf_dim_ + j]->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: " << this->index_dim_vec_[j] << " len: " << len;
              HeterPs_->show_one_table(i);
            }
          };
      threads[j] = std::thread(build_ps_thread, i, j, len, feature_value_size);
    }
    // build feature table
    size_t slot_num = slot_vector_.size() - 1;  // node slot 9008 in slot_vector
    if (slot_num > 0 &&
        (FLAGS_gpugraph_storage_mode == paddle::framework::GpuGraphStorageMode::
                                            MEM_EMB_FEATURE_AND_GPU_GRAPH ||
         FLAGS_gpugraph_storage_mode ==
             paddle::framework::GpuGraphStorageMode::
                 SSD_EMB_AND_MEM_FEATURE_GPU_GRAPH)) {
      auto build_feature_table = [this, &gpu_task](int i) {
        auto gpu_graph_ptr = GraphGpuWrapper::GetInstance();
        std::vector<GpuPsCommGraphFea>* tmp =
            (std::vector<GpuPsCommGraphFea>*)gpu_task->sub_graph_feas;
        gpu_graph_ptr->build_gpu_graph_fea((*tmp)[i], i);
      };
      threads.push_back(std::thread(build_feature_table, i));
Z
zmxdream 已提交
1208
    }
Y
yaoxuefeng 已提交
1209

L
lxsbupt 已提交
1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234
    struct task_info task;
    while (cpu_reday_channels_[i]->Get(task)) {
      auto hbm = this->hbm_pools_[task.device_id * this->multi_mf_dim_ +
                                  task.multi_mf_dim]
                     ->mem();
      int mf_dim = this->index_dim_vec_[task.multi_mf_dim];
      size_t feature_value_size =
          accessor_wrapper_ptr->GetFeatureValueSize(mf_dim);
      auto hbm_start = hbm + task.offset * feature_value_size;
      CUDA_CHECK(
          cudaMemcpy(hbm_start,
                     task.build_values.get() + task.start * feature_value_size,
                     (task.end - task.start) * feature_value_size,
                     cudaMemcpyHostToDevice));
    }
    platform::Timer stagetime;
    stagetime.Start();
    for (std::thread& t : threads) {
      t.join();
    }
    stagetime.Pause();
    VLOG(0) << "card: " << i
            << " BuildGPUTask build_ps async costs: " << stagetime.ElapsedSec()
            << " s.";
  };
Y
yaoxuefeng 已提交
1235

L
lxsbupt 已提交
1236 1237 1238 1239 1240
  std::vector<std::future<void>> cpu_task_futures;
  std::vector<std::future<void>> gpu_task_futures;

  int once_gpu_copy = 64 * 1024;
  threads.resize(device_num * multi_mf_dim_);
Z
zmxdream 已提交
1241
  for (int i = 0; i < device_num; i++) {
L
lxsbupt 已提交
1242 1243 1244
    cpu_reday_channels_[i]->Open();
    gpu_task_futures.emplace_back(
        hbm_thread_pool_[i]->enqueue(build_dymf_hbm_pool, i));
Z
zmxdream 已提交
1245
    for (int j = 0; j < multi_mf_dim_; j++) {
L
lxsbupt 已提交
1246 1247 1248 1249 1250 1251 1252 1253 1254
      auto& device_dim_keys = gpu_task->device_dim_keys_[i][j];
      size_t len = device_dim_keys.size();
      size_t start = 0;
      size_t end = 0;
      while (end < len) {
        start = end;
        end = end + once_gpu_copy < len ? (end + once_gpu_copy) : len;
        cpu_task_futures.emplace_back(cpu_work_pool_[i]->enqueue(
            build_dynamic_mf_func, i, j, start, end));
Z
zmxdream 已提交
1255
      }
Y
yaoxuefeng 已提交
1256
    }
Z
zmxdream 已提交
1257
  }
L
lxsbupt 已提交
1258 1259 1260 1261

  stagetime.Start();
  for (auto& f : cpu_task_futures) {
    f.wait();
Z
zmxdream 已提交
1262
  }
L
lxsbupt 已提交
1263 1264 1265 1266
  cpu_task_futures.clear();
  stagetime.Pause();
  VLOG(0) << " BuildGPUTask build_dynamic_mf_func "
          << " cost " << stagetime.ElapsedSec() << " s.";
Y
yaoxuefeng 已提交
1267
  for (int i = 0; i < device_num; i++) {
L
lxsbupt 已提交
1268
    cpu_reday_channels_[i]->Close();
Y
yaoxuefeng 已提交
1269
  }
L
lxsbupt 已提交
1270 1271 1272
  stagetime.Start();
  for (auto& f : gpu_task_futures) {
    f.wait();
T
Thunderbrook 已提交
1273
  }
L
lxsbupt 已提交
1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286
  gpu_task_futures.clear();
  if (FLAGS_gpugraph_storage_mode == paddle::framework::GpuGraphStorageMode::
                                         MEM_EMB_FEATURE_AND_GPU_GRAPH ||
      FLAGS_gpugraph_storage_mode == paddle::framework::GpuGraphStorageMode::
                                         SSD_EMB_AND_MEM_FEATURE_GPU_GRAPH) {
    std::vector<GpuPsCommGraphFea>* tmp =
        (std::vector<GpuPsCommGraphFea>*)gpu_task->sub_graph_feas;
    delete tmp;
    gpu_task->sub_graph_feas = NULL;
  }
  stagetime.Pause();
  VLOG(0) << "  build_dymf_hbm_pool "
          << " cost " << stagetime.ElapsedSec() << " s.";
1287 1288 1289 1290 1291 1292 1293 1294 1295
}

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";
L
lxsbupt 已提交
1296 1297 1298 1299 1300
  gpu_graph_mode_ = dataset_->GetGpuGraphMode();
  if (dataset_->GetMemoryDataSize() == 0) {
    VLOG(0) << "GetMemoryDataSize == 0";
    return;
  }
1301 1302 1303 1304
  // local shuffle
  if (is_shuffle) {
    dataset_->LocalShuffle();
  }
Y
yaoxuefeng 已提交
1305

L
lxsbupt 已提交
1306 1307 1308 1309 1310 1311 1312 1313 1314
  InitSlotInfo();
  if (FLAGS_gpugraph_storage_mode != GpuGraphStorageMode::WHOLE_HBM) {
    std::shared_ptr<HeterContext> gpu_task = gpu_task_pool_.Get();
    gpu_task->Reset();
    gpu_task->pass_id_ = (uint16_t)(dataset_->GetPassID());
    data_ready_channel_->Put(gpu_task);
  } else if (hbm_sparse_table_initialized_ == false) {
    SparseTableToHbm();
  }
Y
yaoxuefeng 已提交
1315

1316 1317 1318 1319 1320
  VLOG(3) << "End LoadIntoMemory(), dataset[" << dataset_ << "]";
}

void PSGPUWrapper::start_build_thread() {
  running_ = true;
1321
  VLOG(3) << "start build CPU ps thread.";
1322
  pre_build_threads_ = std::thread([this] { pre_build_thread(); });
L
lxsbupt 已提交
1323
  buildpull_threads_ = std::thread([this] { build_pull_thread(); });
1324 1325
}

1326 1327
void PSGPUWrapper::pre_build_thread() {
  // prebuild: process load_data
1328 1329 1330 1331 1332
  while (running_) {
    std::shared_ptr<HeterContext> gpu_task = nullptr;
    if (!data_ready_channel_->Get(gpu_task)) {
      continue;
    }
1333
    VLOG(3) << "thread PreBuildTask start.";
1334 1335 1336
    platform::Timer timer;
    timer.Start();
    // build cpu ps data process
1337
    PreBuildTask(gpu_task);
1338
    timer.Pause();
1339
    VLOG(0) << "thread PreBuildTask end, cost time: " << timer.ElapsedSec()
T
Thunderbrook 已提交
1340
            << " s";
1341 1342 1343 1344 1345
    buildcpu_ready_channel_->Put(gpu_task);
  }
  VLOG(3) << "build cpu thread end";
}

L
lxsbupt 已提交
1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366
void PSGPUWrapper::build_pull_thread() {
  while (running_) {
    std::shared_ptr<HeterContext> gpu_task = nullptr;
    if (!buildcpu_ready_channel_->Get(gpu_task)) {
      continue;
    }
    VLOG(3) << "thread build pull start.";
    platform::Timer timer;
    timer.Start();
    // build cpu ps data process
    BuildPull(gpu_task);
    if (multi_mf_dim_) {
      divide_to_device(gpu_task);
    }
    timer.Pause();
    VLOG(1) << "thread BuildPull end, cost time: " << timer.ElapsedSec() << "s";
    buildpull_ready_channel_->Put(gpu_task);
  }
  VLOG(3) << "build cpu thread end";
}

1367 1368 1369 1370 1371 1372 1373 1374
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
L
lxsbupt 已提交
1375
  if (!buildpull_ready_channel_->Get(gpu_task)) {
1376
    return;
1377
  }
1378

L
lxsbupt 已提交
1379
  VLOG(0) << "PrepareGPUTask start.";
1380 1381
  platform::Timer timer;
  timer.Start();
L
lxsbupt 已提交
1382 1383 1384
  if (!multi_mf_dim_) {
    PrepareGPUTask(gpu_task);
  }
1385 1386
  BuildGPUTask(gpu_task);
  timer.Pause();
L
lxsbupt 已提交
1387 1388
  VLOG(0) << "PrepareGPUTask + BuildGPUTask end, cost time: "
          << timer.ElapsedSec() << "s";
1389 1390

  current_task_ = gpu_task;
1391 1392 1393 1394
}

void PSGPUWrapper::BeginPass() {
  platform::Timer timer;
L
lxsbupt 已提交
1395 1396 1397
  if (FLAGS_gpugraph_storage_mode == GpuGraphStorageMode::WHOLE_HBM) {
    return;
  }
1398 1399 1400 1401 1402
  timer.Start();
  if (current_task_) {
    PADDLE_THROW(
        platform::errors::Fatal("[BeginPass] current task is not ended."));
  }
1403

D
danleifeng 已提交
1404
  debug_gpu_memory_info("befor build task");
1405
  build_task();
D
danleifeng 已提交
1406
  debug_gpu_memory_info("after build task");
1407
  timer.Pause();
1408 1409 1410 1411 1412

  if (current_task_ == nullptr) {
    PADDLE_THROW(platform::errors::Fatal(
        "[BeginPass] after build_task, current task is not null."));
  }
D
danleifeng 已提交
1413 1414 1415 1416 1417 1418 1419
  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";
  }
1420 1421 1422
}

void PSGPUWrapper::EndPass() {
L
lxsbupt 已提交
1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477
  if (FLAGS_gpugraph_storage_mode == GpuGraphStorageMode::WHOLE_HBM) {
    return;
  }
  platform::Timer stagetime;
  stagetime.Start();
  HbmToSparseTable();
  stagetime.Pause();
  VLOG(0) << "EndPass HbmToSparseTable cost time: " << stagetime.ElapsedSec()
          << "s";

  gpu_task_pool_.Push(current_task_);
  current_task_ = nullptr;
  gpu_free_channel_->Put(current_task_);
  // fleet_ptr->pslib_ptr_->_worker_ptr->release_table_mutex(this->table_id_);
}

void PSGPUWrapper::SparseTableToHbm() {
  std::shared_ptr<HeterContext> gpu_task = gpu_task_pool_.Get();
  gpu_task->Reset();
  size_t device_num = heter_devices_.size();
  gpu_task->init(thread_keys_shard_num_, device_num, multi_mf_dim_);
  gpu_task->pass_id_ = (uint16_t)(dataset_->GetPassID());
  auto gpu_graph_ptr = GraphGpuWrapper::GetInstance();
  auto node_to_id = gpu_graph_ptr->feature_to_id;
  auto edge_to_id = gpu_graph_ptr->edge_to_id;
  std::vector<uint64_t> vec_data = gpu_graph_ptr->get_graph_total_keys();

  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_);
    }
  }

  add_key_to_local(vec_data);
  add_key_to_gputask(gpu_task);
  BuildPull(gpu_task);
  if (!multi_mf_dim_) {
    PrepareGPUTask(gpu_task);
  } else {
    divide_to_device(gpu_task);
  }
  BuildGPUTask(gpu_task);
  current_task_ = gpu_task;
  hbm_sparse_table_initialized_ = true;
}

void PSGPUWrapper::HbmToSparseTable() {
  // hbm no update not need dump
  if (grad_push_count_ == 0) {
    return;
  }
  grad_push_count_ = 0;

1478 1479 1480 1481 1482 1483
  if (!current_task_) {
    PADDLE_THROW(
        platform::errors::Fatal("[EndPass] current task has been ended."));
  }
  size_t keysize_max = 0;
  // in case of feasign_num = 0, skip dump_to_cpu
Y
yaoxuefeng 已提交
1484

1485
  for (size_t i = 0; i < heter_devices_.size(); i++) {
Y
yaoxuefeng 已提交
1486 1487 1488 1489 1490
    for (int j = 0; j < multi_mf_dim_; j++) {
      keysize_max =
          std::max(keysize_max, current_task_->device_dim_keys_[i][j].size());
    }
  }
D
danleifeng 已提交
1491
  auto accessor_wrapper_ptr =
D
danleifeng 已提交
1492
      GlobalAccessorFactory::GetInstance().GetAccessorWrapper();
L
lxsbupt 已提交
1493 1494 1495 1496 1497 1498

  int once_cpu_num = 16 * 1024;
  int once_gpu_copy = 8 * once_cpu_num;

  auto dump_pool_to_cpu_func = [this, &accessor_wrapper_ptr, once_cpu_num](
                                   int i, int j, size_t start, size_t end) {
Y
yaoxuefeng 已提交
1499 1500
    PADDLE_ENFORCE_GPU_SUCCESS(cudaSetDevice(this->resource_->dev_id(i)));
    auto& hbm_pool = this->hbm_pools_[i * this->multi_mf_dim_ + j];
L
lxsbupt 已提交
1501
    size_t real_len = end - start;
1502
    // ============ multi-thread process feasign============
Y
yaoxuefeng 已提交
1503 1504
    int mf_dim = this->index_dim_vec_[j];
    size_t feature_value_size =
D
danleifeng 已提交
1505
        accessor_wrapper_ptr->GetFeatureValueSize(mf_dim);
L
lxsbupt 已提交
1506 1507 1508 1509 1510 1511

    std::shared_ptr<char> build_values(new char[feature_value_size * real_len],
                                       [](char* p) { delete[] p; });
    uint64_t offset = start * feature_value_size;
    char* test_build_values = build_values.get();

1512 1513 1514 1515
    cudaMemcpy(test_build_values,
               hbm_pool->mem() + offset,
               feature_value_size * real_len,
               cudaMemcpyDeviceToHost);
L
lxsbupt 已提交
1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545
    for (size_t k = 0; k * once_cpu_num < real_len; k++) {
      struct task_info task;
      task.build_values = build_values;
      task.offset = start;
      task.device_id = i;
      task.multi_mf_dim = j;
      task.start = k * once_cpu_num;
      task.end = (k + 1) * once_cpu_num < real_len ? ((k + 1) * once_cpu_num)
                                                   : (real_len);
      cpu_reday_channels_[i]->Put(task);
    }
  };
  auto cpu_func = [this, &accessor_wrapper_ptr](int j) {
    struct task_info task;
    while (cpu_reday_channels_[j]->Get(task)) {
      auto& device_keys =
          this->current_task_
              ->device_dim_keys_[task.device_id][task.multi_mf_dim];
      char* test_build_values = task.build_values.get();
      int mf_dim = this->index_dim_vec_[task.multi_mf_dim];
      size_t feature_value_size =
          accessor_wrapper_ptr->GetFeatureValueSize(mf_dim);
      uint64_t unuse_key = std::numeric_limits<uint64_t>::max();
      for (int i = task.start; i < task.end; ++i) {
        if (device_keys[i + task.offset] == unuse_key) {
          continue;
        }
        size_t local_offset = i * feature_value_size;
        float* gpu_val =
            reinterpret_cast<float*>(test_build_values + local_offset);
1546
#ifdef PADDLE_WITH_PSLIB
L
lxsbupt 已提交
1547
        // TODO: PSLIB DumpFill
D
danleifeng 已提交
1548 1549
#endif
#ifdef PADDLE_WITH_PSCORE
L
lxsbupt 已提交
1550
        accessor_wrapper_ptr->DumpFill(gpu_val, cpu_table_accessor_, mf_dim);
D
danleifeng 已提交
1551
#endif
L
lxsbupt 已提交
1552
      }
Y
yaoxuefeng 已提交
1553 1554
    }
  };
L
lxsbupt 已提交
1555 1556 1557 1558 1559 1560
  platform::Timer timer;
  timer.Start();
  std::vector<std::future<void>> cpu_task_futures;
  std::vector<std::future<void>> gpu_task_futures;
  size_t thread_num = 16;
  size_t device_num = heter_devices_.size();
Y
yaoxuefeng 已提交
1561 1562 1563
  if (multi_mf_dim_) {
    VLOG(0) << "psgpu wrapper dump pool: multi_mf_dim_: " << multi_mf_dim_;
    for (size_t i = 0; i < device_num; i++) {
L
lxsbupt 已提交
1564
      cpu_reday_channels_[i]->Open();
Y
yaoxuefeng 已提交
1565
      for (int j = 0; j < multi_mf_dim_; j++) {
L
lxsbupt 已提交
1566 1567 1568 1569 1570 1571 1572 1573 1574
        auto& device_keys = this->current_task_->device_dim_keys_[i][j];
        size_t len = device_keys.size();
        size_t start = 0;
        size_t end = 0;
        while (end < len) {
          start = end;
          end = end + once_gpu_copy < len ? (end + once_gpu_copy) : len;
          gpu_task_futures.emplace_back(hbm_thread_pool_[i]->enqueue(
              dump_pool_to_cpu_func, i, j, start, end));
1575
        }
Y
yaoxuefeng 已提交
1576
      }
L
lxsbupt 已提交
1577 1578 1579
      for (size_t j = 0; j < thread_num; j++) {
        cpu_task_futures.emplace_back(cpu_work_pool_[i]->enqueue(cpu_func, i));
      }
Y
yaoxuefeng 已提交
1580
    }
1581
  }
L
lxsbupt 已提交
1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599
  for (auto& f : gpu_task_futures) {
    f.wait();
  }
  timer.Pause();
  VLOG(0) << " EndPass  dump_pool_to_cpu_func "
          << " cost " << timer.ElapsedSec() << " s.";
  for (size_t i = 0; i < device_num; i++) {
    cpu_reday_channels_[i]->Close();
  }
  gpu_task_futures.clear();
  timer.Start();
  for (auto& f : cpu_task_futures) {
    f.wait();
  }
  cpu_task_futures.clear();
  timer.Pause();
  VLOG(0) << " EndPass  cpu_func "
          << " cost " << timer.ElapsedSec() << " s.";
1600 1601 1602
  if (keysize_max != 0) {
    HeterPs_->end_pass();
  }
L
lxsbupt 已提交
1603
}
1604

L
lxsbupt 已提交
1605 1606 1607
void PSGPUWrapper::DumpToMem() {
  if (FLAGS_gpugraph_storage_mode == GpuGraphStorageMode::WHOLE_HBM) {
    this->HbmToSparseTable();
Y
yaoxuefeng 已提交
1608
  }
T
Thunderbrook 已提交
1609 1610 1611 1612 1613 1614 1615 1616
}

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) {
D
danleifeng 已提交
1617 1618
  VLOG(0) << "Warning:: recommand use pull_gpups_sparse op instead. This "
             "PullSparse is not used.";
Y
yaoxuefeng 已提交
1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632
}

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();

D
danleifeng 已提交
1633
  auto accessor_wrapper_ptr =
D
danleifeng 已提交
1634 1635 1636
      GlobalAccessorFactory::GetInstance().GetAccessorWrapper();
  size_t feature_value_size =
      accessor_wrapper_ptr->GetPullValueSize(max_mf_dim_);
D
danleifeng 已提交
1637
  VLOG(3) << "PullSparse max_dim:" << max_mf_dim_
D
danleifeng 已提交
1638
          << " pull_feature_value_size:" << pull_type_size_;
Y
yaoxuefeng 已提交
1639 1640 1641 1642 1643

  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)) {
D
danleifeng 已提交
1644
#ifdef PADDLE_WITH_CUDA
Y
yaoxuefeng 已提交
1645 1646
    int device_id = place.GetDeviceId();
    int devid_2_index = HeterPs_->get_index_by_devid(device_id);
D
danleifeng 已提交
1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662
    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";

1663
      auto stream = dynamic_cast<phi::GPUContext*>(
D
danleifeng 已提交
1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771
                        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 << "]";
1772
      phi::DenseTensor& total_keys_tensor = keys_tensor[devid_2_index];
D
danleifeng 已提交
1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827
      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);
Y
yaoxuefeng 已提交
1828 1829
    }
    pull_gpups_timer.Pause();
D
danleifeng 已提交
1830
#endif
F
Fan Zhang 已提交
1831 1832
  } else if (platform::is_xpu_place(place)) {
#ifdef PADDLE_WITH_XPU_KP
D
danleifeng 已提交
1833 1834 1835 1836 1837 1838
    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);
F
Fan Zhang 已提交
1839 1840 1841
    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);
1842
    phi::DenseTensor& total_keys_tensor = keys_tensor[devid_2_index];
D
danleifeng 已提交
1843 1844 1845
    uint64_t* total_keys =
        reinterpret_cast<uint64_t*>(total_keys_tensor.mutable_data<int64_t>(
            {int64_t(total_length), 1}, place));
F
Fan Zhang 已提交
1846 1847 1848 1849 1850 1851 1852

    // 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];
    }

F
Fan Zhang 已提交
1853 1854 1855 1856 1857
    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());
1858 1859
    PADDLE_ENFORCE_XPU_SUCCESS(xpu_memcpy(xpu_keys,
                                          keys.data(),
F
Fan Zhang 已提交
1860 1861
                                          keys.size() * sizeof(uint64_t*),
                                          XPU_HOST_TO_DEVICE));
1862 1863
    PADDLE_ENFORCE_XPU_SUCCESS(xpu_memcpy(xpu_len,
                                          slot_lengths_lod.data(),
F
Fan Zhang 已提交
1864 1865 1866
                                          slot_lengths.size() * sizeof(int64_t),
                                          XPU_HOST_TO_DEVICE));

1867 1868 1869 1870
    this->CopyKeys(place,
                   xpu_keys,
                   total_keys,
                   xpu_len,
F
Fan Zhang 已提交
1871 1872 1873 1874 1875
                   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();
1876 1877 1878
    HeterPs_->pull_sparse(devid_2_index,
                          total_keys,
                          total_values_gpu,
F
Fan Zhang 已提交
1879 1880 1881 1882 1883
                          static_cast<int>(total_length));
    pull_gpups_timer.Pause();

    VLOG(3) << "Begin Copy result to tensor, total_length[" << total_length
            << "]";
D
danleifeng 已提交
1884 1885 1886 1887 1888 1889 1890 1891
    accessor_wrapper_ptr->CopyForPull(place,
                                      xpu_keys,
                                      values,
                                      total_values_gpu,
                                      xpu_len,
                                      static_cast<int>(slot_lengths.size()),
                                      hidden_size,
                                      total_length,
D
danleifeng 已提交
1892
                                      feature_value_size);
F
Fan Zhang 已提交
1893
#endif
T
Thunderbrook 已提交
1894 1895
  } else {
    PADDLE_THROW(platform::errors::PreconditionNotMet(
F
Fan Zhang 已提交
1896
        "GpuPs/XpuPs: PullSparse Only Support CUDAPlace or XPUPlace Now."));
T
Thunderbrook 已提交
1897 1898
  }
  all_timer.Pause();
1899
  VLOG(3) << "GpuPs PullSparse total costs: " << all_timer.ElapsedSec()
T
Thunderbrook 已提交
1900 1901 1902 1903 1904 1905 1906 1907 1908 1909
          << " 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,
1910 1911
                                  const int hidden_size,
                                  const int batch_size) {
L
lxsbupt 已提交
1912
  ++grad_push_count_;
T
Thunderbrook 已提交
1913 1914 1915
  platform::Timer all_timer;
  platform::Timer push_gpups_timer;
  all_timer.Start();
D
danleifeng 已提交
1916
  auto accessor_wrapper_ptr =
D
danleifeng 已提交
1917
      GlobalAccessorFactory::GetInstance().GetAccessorWrapper();
D
danleifeng 已提交
1918
  size_t grad_value_size = accessor_wrapper_ptr->GetPushValueSize(max_mf_dim_);
D
danleifeng 已提交
1919

T
Thunderbrook 已提交
1920 1921 1922 1923
  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)) {
F
Fan Zhang 已提交
1924
#ifdef PADDLE_WITH_CUDA
1925
    int device_id = place.GetDeviceId();
T
Thunderbrook 已提交
1926
    int devid_2_index = HeterPs_->get_index_by_devid(device_id);
D
danleifeng 已提交
1927 1928 1929 1930 1931 1932
    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;
1933
      auto stream = dynamic_cast<phi::GPUContext*>(
D
danleifeng 已提交
1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947
                        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);
      }
T
Thunderbrook 已提交
1948

D
danleifeng 已提交
1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022
      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());

2023
      phi::DenseTensor& total_keys_tensor = keys_tensor[devid_2_index];
D
danleifeng 已提交
2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045
      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));
    }
T
Thunderbrook 已提交
2046
    push_gpups_timer.Pause();
F
Fan Zhang 已提交
2047
#endif
F
Fan Zhang 已提交
2048
  } else if (platform::is_xpu_place(place)) {
F
Fan Zhang 已提交
2049
#ifdef PADDLE_WITH_XPU_KP
F
Fan Zhang 已提交
2050 2051
    int device_id = place.GetDeviceId();
    int devid_2_index = HeterPs_->get_index_by_devid(device_id);
D
danleifeng 已提交
2052 2053 2054 2055 2056 2057 2058 2059
    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());
2060
    phi::DenseTensor& total_keys_tensor = keys_tensor[devid_2_index];
F
Fan Zhang 已提交
2061
    uint64_t* total_keys =
D
danleifeng 已提交
2062
        reinterpret_cast<uint64_t*>(total_keys_tensor.data<int64_t>());
F
Fan Zhang 已提交
2063
    VLOG(3) << "Begin copy grad tensor to xpups struct";
D
danleifeng 已提交
2064 2065 2066 2067 2068 2069 2070 2071
    accessor_wrapper_ptr->CopyForPush(place,
                                      grad_values,
                                      total_grad_values_gpu,
                                      slot_lengths,
                                      hidden_size,
                                      total_length,
                                      batch_size,
                                      slot_vector_);
F
Fan Zhang 已提交
2072 2073 2074 2075

    VLOG(3) << "Begin call PushSparseXPU in XPUPS, dev: " << devid_2_index
            << " len: " << total_length;
    push_gpups_timer.Start();
2076 2077 2078
    HeterPs_->push_sparse(devid_2_index,
                          total_keys,
                          total_grad_values_gpu,
F
Fan Zhang 已提交
2079 2080
                          static_cast<int>(total_length));
    push_gpups_timer.Pause();
F
Fan Zhang 已提交
2081
#endif
T
Thunderbrook 已提交
2082 2083 2084 2085 2086
  } else {
    PADDLE_THROW(platform::errors::PreconditionNotMet(
        "GPUPS: PushSparseGrad Only Support CUDAPlace Now."));
  }
  all_timer.Pause();
Y
yaoxuefeng 已提交
2087 2088
  time_3 += all_timer.ElapsedSec();
  time_4 += push_gpups_timer.ElapsedSec();
2089
  VLOG(3) << "PushSparseGrad total cost: " << all_timer.ElapsedSec()
T
Thunderbrook 已提交
2090 2091 2092 2093 2094
          << " s, of which GPUPS cost: " << push_gpups_timer.ElapsedSec()
          << " s";
  VLOG(3) << "End PushSparseGrad";
}

D
danleifeng 已提交
2095
}  // namespace framework
T
Thunderbrook 已提交
2096 2097
}  // end namespace paddle
#endif