data_parallel_tree_learner.cpp 23.6 KB
Newer Older
1
2
3
4
/*!
 * Copyright (c) 2016 Microsoft Corporation. All rights reserved.
 * Licensed under the MIT License. See LICENSE file in the project root for license information.
 */
Guolin Ke's avatar
Guolin Ke committed
5
6
7
8
#include <cstring>
#include <tuple>
#include <vector>

9
10
#include "parallel_tree_learner.h"

Guolin Ke's avatar
Guolin Ke committed
11
12
namespace LightGBM {

13
template <typename TREELEARNER_T>
Guolin Ke's avatar
Guolin Ke committed
14
15
DataParallelTreeLearner<TREELEARNER_T>::DataParallelTreeLearner(const Config* config)
  :TREELEARNER_T(config) {
Guolin Ke's avatar
Guolin Ke committed
16
17
}

18
19
template <typename TREELEARNER_T>
DataParallelTreeLearner<TREELEARNER_T>::~DataParallelTreeLearner() {
Guolin Ke's avatar
Guolin Ke committed
20
21
}

22
23
template <typename TREELEARNER_T>
void DataParallelTreeLearner<TREELEARNER_T>::Init(const Dataset* train_data, bool is_constant_hessian) {
Guolin Ke's avatar
Guolin Ke committed
24
  // initialize SerialTreeLearner
25
  TREELEARNER_T::Init(train_data, is_constant_hessian);
Guolin Ke's avatar
Guolin Ke committed
26
27
28
  // Get local rank and global machine size
  rank_ = Network::rank();
  num_machines_ = Network::num_machines();
29
30
31
32

  auto max_cat_threshold = this->config_->max_cat_threshold;
  // need to be able to hold smaller and larger best splits in SyncUpGlobalBestSplit
  size_t split_info_size = static_cast<size_t>(SplitInfo::Size(max_cat_threshold) * 2);
33
34
35
  size_t histogram_size = this->config_->use_quantized_grad ?
    static_cast<size_t>(this->share_state_->num_hist_total_bin() * kInt32HistEntrySize) :
    static_cast<size_t>(this->share_state_->num_hist_total_bin() * kHistEntrySize);
36

Guolin Ke's avatar
Guolin Ke committed
37
  // allocate buffer for communication
38
  size_t buffer_size = std::max(histogram_size, split_info_size);
Guolin Ke's avatar
Guolin Ke committed
39

Guolin Ke's avatar
Guolin Ke committed
40
41
  input_buffer_.resize(buffer_size);
  output_buffer_.resize(buffer_size);
Guolin Ke's avatar
Guolin Ke committed
42

43
  is_feature_aggregated_.resize(this->num_features_);
Guolin Ke's avatar
Guolin Ke committed
44

Guolin Ke's avatar
Guolin Ke committed
45
46
  block_start_.resize(num_machines_);
  block_len_.resize(num_machines_);
Guolin Ke's avatar
Guolin Ke committed
47

48
49
50
51
52
  if (this->config_->use_quantized_grad) {
    block_start_int16_.resize(num_machines_);
    block_len_int16_.resize(num_machines_);
  }

53
54
  buffer_write_start_pos_.resize(this->num_features_);
  buffer_read_start_pos_.resize(this->num_features_);
55
56
57
58
59
60

  if (this->config_->use_quantized_grad) {
    buffer_write_start_pos_int16_.resize(this->num_features_);
    buffer_read_start_pos_int16_.resize(this->num_features_);
  }

Guolin Ke's avatar
Guolin Ke committed
61
  global_data_count_in_leaf_.resize(this->config_->num_leaves);
Guolin Ke's avatar
Guolin Ke committed
62
63
}

64
template <typename TREELEARNER_T>
Guolin Ke's avatar
Guolin Ke committed
65
66
67
void DataParallelTreeLearner<TREELEARNER_T>::ResetConfig(const Config* config) {
  TREELEARNER_T::ResetConfig(config);
  global_data_count_in_leaf_.resize(this->config_->num_leaves);
Guolin Ke's avatar
Guolin Ke committed
68
}
Guolin Ke's avatar
Guolin Ke committed
69

70
template <typename TREELEARNER_T>
71
72
73
74
75
76
77
78
void DataParallelTreeLearner<TREELEARNER_T>::PrepareBufferPos(
  const std::vector<std::vector<int>>& feature_distribution,
  std::vector<comm_size_t>* block_start,
  std::vector<comm_size_t>* block_len,
  std::vector<comm_size_t>* buffer_write_start_pos,
  std::vector<comm_size_t>* buffer_read_start_pos,
  comm_size_t* reduce_scatter_size,
  size_t hist_entry_size) {
Guolin Ke's avatar
Guolin Ke committed
79
  // get block start and block len for reduce scatter
80
  *reduce_scatter_size = 0;
Guolin Ke's avatar
Guolin Ke committed
81
  for (int i = 0; i < num_machines_; ++i) {
82
    (*block_len)[i] = 0;
Guolin Ke's avatar
Guolin Ke committed
83
    for (auto fid : feature_distribution[i]) {
84
      auto num_bin = this->train_data_->FeatureNumBin(fid);
Guolin Ke's avatar
Guolin Ke committed
85
      if (this->train_data_->FeatureBinMapper(fid)->GetMostFreqBin() == 0) {
Guolin Ke's avatar
Guolin Ke committed
86
87
        num_bin -= 1;
      }
88
      (*block_len)[i] += num_bin * hist_entry_size;
Guolin Ke's avatar
Guolin Ke committed
89
    }
90
    *reduce_scatter_size += (*block_len)[i];
Guolin Ke's avatar
Guolin Ke committed
91
92
  }

93
  (*block_start)[0] = 0;
Guolin Ke's avatar
Guolin Ke committed
94
  for (int i = 1; i < num_machines_; ++i) {
95
    (*block_start)[i] = (*block_start)[i - 1] + (*block_len)[i - 1];
Guolin Ke's avatar
Guolin Ke committed
96
97
  }

98
  // get buffer_write_start_pos
Guolin Ke's avatar
Guolin Ke committed
99
100
101
  int bin_size = 0;
  for (int i = 0; i < num_machines_; ++i) {
    for (auto fid : feature_distribution[i]) {
102
      (*buffer_write_start_pos)[fid] = bin_size;
103
      auto num_bin = this->train_data_->FeatureNumBin(fid);
Guolin Ke's avatar
Guolin Ke committed
104
      if (this->train_data_->FeatureBinMapper(fid)->GetMostFreqBin() == 0) {
Guolin Ke's avatar
Guolin Ke committed
105
106
        num_bin -= 1;
      }
107
      bin_size += num_bin * hist_entry_size;
Guolin Ke's avatar
Guolin Ke committed
108
109
110
    }
  }

111
  // get buffer_read_start_pos
Guolin Ke's avatar
Guolin Ke committed
112
113
  bin_size = 0;
  for (auto fid : feature_distribution[rank_]) {
114
    (*buffer_read_start_pos)[fid] = bin_size;
115
    auto num_bin = this->train_data_->FeatureNumBin(fid);
Guolin Ke's avatar
Guolin Ke committed
116
    if (this->train_data_->FeatureBinMapper(fid)->GetMostFreqBin() == 0) {
Guolin Ke's avatar
Guolin Ke committed
117
118
      num_bin -= 1;
    }
119
    bin_size += num_bin * hist_entry_size;
Guolin Ke's avatar
Guolin Ke committed
120
  }
121
}
Guolin Ke's avatar
Guolin Ke committed
122

123
124
125
126
127
128
129
130
template <typename TREELEARNER_T>
void DataParallelTreeLearner<TREELEARNER_T>::BeforeTrain() {
  TREELEARNER_T::BeforeTrain();
  // generate feature partition for current tree
  std::vector<std::vector<int>> feature_distribution(num_machines_, std::vector<int>());
  std::vector<int> num_bins_distributed(num_machines_, 0);
  for (int i = 0; i < this->train_data_->num_total_features(); ++i) {
    int inner_feature_index = this->train_data_->InnerFeatureIndex(i);
131
132
133
    if (inner_feature_index == -1) {
      continue;
    }
134
135
136
137
138
139
140
141
    if (this->col_sampler_.is_feature_used_bytree()[inner_feature_index]) {
      int cur_min_machine = static_cast<int>(ArrayArgs<int>::ArgMin(num_bins_distributed));
      feature_distribution[cur_min_machine].push_back(inner_feature_index);
      auto num_bin = this->train_data_->FeatureNumBin(inner_feature_index);
      if (this->train_data_->FeatureBinMapper(inner_feature_index)->GetMostFreqBin() == 0) {
        num_bin -= 1;
      }
      num_bins_distributed[cur_min_machine] += num_bin;
Guolin Ke's avatar
Guolin Ke committed
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
221
    is_feature_aggregated_[inner_feature_index] = false;
  }
  // get local used feature
  for (auto fid : feature_distribution[rank_]) {
    is_feature_aggregated_[fid] = true;
  }

  // get block start and block len for reduce scatter
  if (this->config_->use_quantized_grad) {
    PrepareBufferPos(feature_distribution, &block_start_, &block_len_, &buffer_write_start_pos_,
      &buffer_read_start_pos_, &reduce_scatter_size_, kInt32HistEntrySize);
    PrepareBufferPos(feature_distribution, &block_start_int16_, &block_len_int16_, &buffer_write_start_pos_int16_,
      &buffer_read_start_pos_int16_, &reduce_scatter_size_int16_, kInt16HistEntrySize);
  } else {
    PrepareBufferPos(feature_distribution, &block_start_, &block_len_, &buffer_write_start_pos_,
      &buffer_read_start_pos_, &reduce_scatter_size_, kHistEntrySize);
  }

  if (this->config_->use_quantized_grad) {
    // sync global data sumup info
    std::tuple<data_size_t, double, double, int64_t> data(this->smaller_leaf_splits_->num_data_in_leaf(),
                                                          this->smaller_leaf_splits_->sum_gradients(), this->smaller_leaf_splits_->sum_hessians(),
                                                          this->smaller_leaf_splits_->int_sum_gradients_and_hessians());
    int size = sizeof(data);
    std::memcpy(input_buffer_.data(), &data, size);
    // global sumup reduce
    Network::Allreduce(input_buffer_.data(), size, sizeof(std::tuple<data_size_t, double, double, int64_t>), output_buffer_.data(), [](const char *src, char *dst, int type_size, comm_size_t len) {
      comm_size_t used_size = 0;
      const std::tuple<data_size_t, double, double, int64_t> *p1;
      std::tuple<data_size_t, double, double, int64_t> *p2;
      while (used_size < len) {
        p1 = reinterpret_cast<const std::tuple<data_size_t, double, double, int64_t> *>(src);
        p2 = reinterpret_cast<std::tuple<data_size_t, double, double, int64_t> *>(dst);
        std::get<0>(*p2) = std::get<0>(*p2) + std::get<0>(*p1);
        std::get<1>(*p2) = std::get<1>(*p2) + std::get<1>(*p1);
        std::get<2>(*p2) = std::get<2>(*p2) + std::get<2>(*p1);
        std::get<3>(*p2) = std::get<3>(*p2) + std::get<3>(*p1);
        src += type_size;
        dst += type_size;
        used_size += type_size;
      }
    });
    // copy back
    std::memcpy(reinterpret_cast<void*>(&data), output_buffer_.data(), size);
    // set global sumup info
    this->smaller_leaf_splits_->Init(std::get<1>(data), std::get<2>(data), std::get<3>(data));
    // init global data count in leaf
    global_data_count_in_leaf_[0] = std::get<0>(data);
    // reset hist num bits according to global num data
    this->gradient_discretizer_->template SetNumBitsInHistogramBin<true>(0, -1, GetGlobalDataCountInLeaf(0), 0);
  } else {
    // sync global data sumup info
    std::tuple<data_size_t, double, double> data(this->smaller_leaf_splits_->num_data_in_leaf(),
                                                this->smaller_leaf_splits_->sum_gradients(), this->smaller_leaf_splits_->sum_hessians());
    int size = sizeof(data);
    std::memcpy(input_buffer_.data(), &data, size);
    // global sumup reduce
    Network::Allreduce(input_buffer_.data(), size, sizeof(std::tuple<data_size_t, double, double>), output_buffer_.data(), [](const char *src, char *dst, int type_size, comm_size_t len) {
      comm_size_t used_size = 0;
      const std::tuple<data_size_t, double, double> *p1;
      std::tuple<data_size_t, double, double> *p2;
      while (used_size < len) {
        p1 = reinterpret_cast<const std::tuple<data_size_t, double, double> *>(src);
        p2 = reinterpret_cast<std::tuple<data_size_t, double, double> *>(dst);
        std::get<0>(*p2) = std::get<0>(*p2) + std::get<0>(*p1);
        std::get<1>(*p2) = std::get<1>(*p2) + std::get<1>(*p1);
        std::get<2>(*p2) = std::get<2>(*p2) + std::get<2>(*p1);
        src += type_size;
        dst += type_size;
        used_size += type_size;
      }
    });
    // copy back
    std::memcpy(reinterpret_cast<void*>(&data), output_buffer_.data(), size);
    // set global sumup info
    this->smaller_leaf_splits_->Init(std::get<1>(data), std::get<2>(data));
    // init global data count in leaf
    global_data_count_in_leaf_[0] = std::get<0>(data);
  }
Guolin Ke's avatar
Guolin Ke committed
222
223
}

224
template <typename TREELEARNER_T>
225
void DataParallelTreeLearner<TREELEARNER_T>::FindBestSplits(const Tree* tree) {
226
227
  TREELEARNER_T::ConstructHistograms(
      this->col_sampler_.is_feature_used_bytree(), true);
228
229
230
231
232
  const int smaller_leaf_index = this->smaller_leaf_splits_->leaf_index();
  const data_size_t local_data_on_smaller_leaf = this->data_partition_->leaf_count(smaller_leaf_index);
  if (local_data_on_smaller_leaf <= 0) {
    // clear histogram buffer before synchronizing
    // otherwise histogram contents from the previous iteration will be sent
233
    #pragma omp parallel for num_threads(OMP_NUM_THREADS()) schedule(static)
234
235
236
237
238
239
    for (int feature_index = 0; feature_index < this->num_features_; ++feature_index) {
      if (this->col_sampler_.is_feature_used_bytree()[feature_index] == false)
        continue;
      const BinMapper* feature_bin_mapper = this->train_data_->FeatureBinMapper(feature_index);
      const int offset = static_cast<int>(feature_bin_mapper->GetMostFreqBin() == 0);
      const int num_bin = feature_bin_mapper->num_bin();
240
241
242
243
244
245
246
247
248
      if (this->config_->use_quantized_grad) {
        int32_t* hist_ptr = this->smaller_leaf_histogram_array_[feature_index].RawDataInt32();
        std::memset(reinterpret_cast<void*>(hist_ptr), 0, (num_bin - offset) * kInt32HistEntrySize);
        int16_t* hist_ptr_int16 = this->smaller_leaf_histogram_array_[feature_index].RawDataInt16();
        std::memset(reinterpret_cast<void*>(hist_ptr_int16), 0, (num_bin - offset) * kInt16HistEntrySize);
      } else {
        hist_t* hist_ptr = this->smaller_leaf_histogram_array_[feature_index].RawData();
        std::memset(reinterpret_cast<void*>(hist_ptr), 0, (num_bin - offset) * kHistEntrySize);
      }
249
250
    }
  }
Guolin Ke's avatar
Guolin Ke committed
251
  // construct local histograms
252
253
  global_timer.Start("DataParallelTreeLearner::ReduceHistogram");
  global_timer.Start("DataParallelTreeLearner::ReduceHistogram::Copy");
254
  #pragma omp parallel for num_threads(OMP_NUM_THREADS()) schedule(static)
255
  for (int feature_index = 0; feature_index < this->num_features_; ++feature_index) {
256
257
    if (this->col_sampler_.is_feature_used_bytree()[feature_index] == false)
      continue;
Guolin Ke's avatar
Guolin Ke committed
258
    // copy to buffer
259
260
261
262
263
264
    if (this->config_->use_quantized_grad) {
      const uint8_t local_smaller_leaf_num_bits = this->gradient_discretizer_->template GetHistBitsInLeaf<false>(this->smaller_leaf_splits_->leaf_index());
      const uint8_t smaller_leaf_num_bits = this->gradient_discretizer_->template GetHistBitsInLeaf<true>(this->smaller_leaf_splits_->leaf_index());
      if (smaller_leaf_num_bits <= 16) {
        std::memcpy(input_buffer_.data() + buffer_write_start_pos_int16_[feature_index],
                    this->smaller_leaf_histogram_array_[feature_index].RawDataInt16(),
265
                    this->smaller_leaf_histogram_array_[feature_index].SizeOfInt16Histogram());
266
267
268
269
      } else {
        if (local_smaller_leaf_num_bits == 32) {
          std::memcpy(input_buffer_.data() + buffer_write_start_pos_[feature_index],
                      this->smaller_leaf_histogram_array_[feature_index].RawDataInt32(),
270
                      this->smaller_leaf_histogram_array_[feature_index].SizeOfInt32Histogram());
271
272
273
274
275
276
277
        } else {
          this->smaller_leaf_histogram_array_[feature_index].CopyFromInt16ToInt32(
            input_buffer_.data() + buffer_write_start_pos_[feature_index]);
        }
      }
    } else {
      std::memcpy(input_buffer_.data() + buffer_write_start_pos_[feature_index],
278
                this->smaller_leaf_histogram_array_[feature_index].RawData(),
279
                this->smaller_leaf_histogram_array_[feature_index].SizeOfHistogram());
280
    }
Guolin Ke's avatar
Guolin Ke committed
281
  }
282
  global_timer.Stop("DataParallelTreeLearner::ReduceHistogram::Copy");
Guolin Ke's avatar
Guolin Ke committed
283
  // Reduce scatter for histogram
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
  global_timer.Start("DataParallelTreeLearner::ReduceHistogram::ReduceScatter");
  if (!this->config_->use_quantized_grad) {
    Network::ReduceScatter(input_buffer_.data(), reduce_scatter_size_, sizeof(hist_t), block_start_.data(),
                           block_len_.data(), output_buffer_.data(), static_cast<comm_size_t>(output_buffer_.size()), &HistogramSumReducer);
  } else {
    const uint8_t smaller_leaf_num_bits = this->gradient_discretizer_->template GetHistBitsInLeaf<true>(this->smaller_leaf_splits_->leaf_index());
    if (smaller_leaf_num_bits <= 16) {
      Network::ReduceScatter(input_buffer_.data(), reduce_scatter_size_int16_, sizeof(int16_t), block_start_int16_.data(),
                            block_len_int16_.data(), output_buffer_.data(), static_cast<comm_size_t>(output_buffer_.size()), &Int16HistogramSumReducer);
    } else {
      Network::ReduceScatter(input_buffer_.data(), reduce_scatter_size_, sizeof(int_hist_t), block_start_.data(),
                            block_len_.data(), output_buffer_.data(), static_cast<comm_size_t>(output_buffer_.size()), &Int32HistogramSumReducer);
    }
  }
  global_timer.Stop("DataParallelTreeLearner::ReduceHistogram::ReduceScatter");
  global_timer.Stop("DataParallelTreeLearner::ReduceHistogram");
300
  this->FindBestSplitsFromHistograms(
301
      this->col_sampler_.is_feature_used_bytree(), true, tree);
Guolin Ke's avatar
Guolin Ke committed
302
303
304
}

template <typename TREELEARNER_T>
305
void DataParallelTreeLearner<TREELEARNER_T>::FindBestSplitsFromHistograms(const std::vector<int8_t>&, bool, const Tree* tree) {
306
307
  std::vector<SplitInfo> smaller_bests_per_thread(this->share_state_->num_threads);
  std::vector<SplitInfo> larger_bests_per_thread(this->share_state_->num_threads);
308
  std::vector<int8_t> smaller_node_used_features =
309
      this->col_sampler_.GetByNode(tree, this->smaller_leaf_splits_->leaf_index());
310
  std::vector<int8_t> larger_node_used_features =
311
      this->col_sampler_.GetByNode(tree, this->larger_leaf_splits_->leaf_index());
312
313
  double smaller_leaf_parent_output = this->GetParentOutput(tree, this->smaller_leaf_splits_.get());
  double larger_leaf_parent_output = this->GetParentOutput(tree, this->larger_leaf_splits_.get());
314
315
316
317
318
319
320
321
322

  if (this->config_->use_quantized_grad && this->larger_leaf_splits_ != nullptr && this->larger_leaf_splits_->leaf_index() >= 0) {
    const int parent_index = std::min(this->smaller_leaf_splits_->leaf_index(), this->larger_leaf_splits_->leaf_index());
    const uint8_t parent_num_bits = this->gradient_discretizer_->template GetHistBitsInNode<true>(parent_index);
    const uint8_t larger_leaf_num_bits = this->gradient_discretizer_->template GetHistBitsInLeaf<true>(this->larger_leaf_splits_->leaf_index());
    const uint8_t smaller_leaf_num_bits = this->gradient_discretizer_->template GetHistBitsInLeaf<true>(this->smaller_leaf_splits_->leaf_index());
    if (parent_num_bits > 16 && larger_leaf_num_bits <= 16) {
      CHECK_LE(smaller_leaf_num_bits, 16);
      OMP_INIT_EX();
323
      #pragma omp parallel for num_threads(OMP_NUM_THREADS()) schedule(static)
324
325
326
327
328
329
330
331
332
333
      for (int feature_index = 0; feature_index < this->num_features_; ++feature_index) {
        OMP_LOOP_EX_BEGIN();
        if (!is_feature_aggregated_[feature_index]) continue;
        this->larger_leaf_histogram_array_[feature_index].CopyToBuffer(this->gradient_discretizer_->GetChangeHistBitsBuffer(feature_index));
        OMP_LOOP_EX_END();
      }
      OMP_THROW_EX();
    }
  }

334
  OMP_INIT_EX();
335
  #pragma omp parallel for num_threads(OMP_NUM_THREADS()) schedule(static)
336
  for (int feature_index = 0; feature_index < this->num_features_; ++feature_index) {
337
    OMP_LOOP_EX_BEGIN();
Guolin Ke's avatar
Guolin Ke committed
338
    if (!is_feature_aggregated_[feature_index]) continue;
Guolin Ke's avatar
Guolin Ke committed
339
    const int tid = omp_get_thread_num();
Guolin Ke's avatar
Guolin Ke committed
340
    const int real_feature_index = this->train_data_->RealFeatureIndex(feature_index);
Guolin Ke's avatar
Guolin Ke committed
341
    // restore global histograms from buffer
342
343
344
345
346
347
348
349
350
351
352
353
354
    if (this->config_->use_quantized_grad) {
      const uint8_t smaller_leaf_num_bits = this->gradient_discretizer_->template GetHistBitsInLeaf<true>(this->smaller_leaf_splits_->leaf_index());
      if (smaller_leaf_num_bits <= 16) {
        this->smaller_leaf_histogram_array_[feature_index].FromMemoryInt16(
          output_buffer_.data() + buffer_read_start_pos_int16_[feature_index]);
      } else {
        this->smaller_leaf_histogram_array_[feature_index].FromMemoryInt32(
          output_buffer_.data() + buffer_read_start_pos_[feature_index]);
      }
    } else {
      this->smaller_leaf_histogram_array_[feature_index].FromMemory(
        output_buffer_.data() + buffer_read_start_pos_[feature_index]);
    }
Guolin Ke's avatar
Guolin Ke committed
355

356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
    if (this->config_->use_quantized_grad) {
      const uint8_t smaller_leaf_num_bits = this->gradient_discretizer_->template GetHistBitsInLeaf<true>(this->smaller_leaf_splits_->leaf_index());
      const int64_t int_sum_gradient_and_hessian = this->smaller_leaf_splits_->int_sum_gradients_and_hessians();
      if (smaller_leaf_num_bits <= 16) {
        this->train_data_->template FixHistogramInt<int32_t, int32_t, 16, 16>(
          feature_index,
          int_sum_gradient_and_hessian,
          reinterpret_cast<hist_t*>(this->smaller_leaf_histogram_array_[feature_index].RawDataInt16()));
      } else {
        this->train_data_->template FixHistogramInt<int64_t, int64_t, 32, 32>(
          feature_index,
          int_sum_gradient_and_hessian,
          reinterpret_cast<hist_t*>(this->smaller_leaf_histogram_array_[feature_index].RawDataInt32()));
      }
    } else {
      this->train_data_->FixHistogram(feature_index,
                                      this->smaller_leaf_splits_->sum_gradients(), this->smaller_leaf_splits_->sum_hessians(),
                                      this->smaller_leaf_histogram_array_[feature_index].RawData());
    }
375
376
377
378
379
380

    this->ComputeBestSplitForFeature(
        this->smaller_leaf_histogram_array_, feature_index, real_feature_index,
        smaller_node_used_features[feature_index],
        GetGlobalDataCountInLeaf(this->smaller_leaf_splits_->leaf_index()),
        this->smaller_leaf_splits_.get(),
381
382
        &smaller_bests_per_thread[tid],
        smaller_leaf_parent_output);
Guolin Ke's avatar
Guolin Ke committed
383
384

    // only root leaf
385
    if (this->larger_leaf_splits_ == nullptr || this->larger_leaf_splits_->leaf_index() < 0) continue;
Guolin Ke's avatar
Guolin Ke committed
386
387

    // construct histgroms for large leaf, we init larger leaf as the parent, so we can just subtract the smaller leaf's histograms
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
    if (this->config_->use_quantized_grad) {
      const int parent_index = std::min(this->smaller_leaf_splits_->leaf_index(), this->larger_leaf_splits_->leaf_index());
      const uint8_t parent_num_bits = this->gradient_discretizer_->template GetHistBitsInNode<true>(parent_index);
      const uint8_t larger_leaf_num_bits = this->gradient_discretizer_->template GetHistBitsInLeaf<true>(this->larger_leaf_splits_->leaf_index());
      const uint8_t smaller_leaf_num_bits = this->gradient_discretizer_->template GetHistBitsInLeaf<true>(this->smaller_leaf_splits_->leaf_index());
      if (parent_num_bits <= 16) {
        CHECK_LE(smaller_leaf_num_bits, 16);
        CHECK_LE(larger_leaf_num_bits, 16);
        this->larger_leaf_histogram_array_[feature_index].template Subtract<true, int32_t, int32_t, int32_t, 16, 16, 16>(
              this->smaller_leaf_histogram_array_[feature_index]);
      } else if (larger_leaf_num_bits <= 16) {
        CHECK_LE(smaller_leaf_num_bits, 16);
        this->larger_leaf_histogram_array_[feature_index].template Subtract<true, int64_t, int32_t, int32_t, 32, 16, 16>(
            this->smaller_leaf_histogram_array_[feature_index], this->gradient_discretizer_->GetChangeHistBitsBuffer(feature_index));
      } else if (smaller_leaf_num_bits <= 16) {
        this->larger_leaf_histogram_array_[feature_index].template Subtract<true, int64_t, int32_t, int64_t, 32, 16, 32>(
              this->smaller_leaf_histogram_array_[feature_index]);
      } else {
        this->larger_leaf_histogram_array_[feature_index].template Subtract<true, int64_t, int64_t, int64_t, 32, 32, 32>(
              this->smaller_leaf_histogram_array_[feature_index]);
      }
    } else {
      this->larger_leaf_histogram_array_[feature_index].Subtract(
        this->smaller_leaf_histogram_array_[feature_index]);
    }
413
414
415
416
417
418

    this->ComputeBestSplitForFeature(
        this->larger_leaf_histogram_array_, feature_index, real_feature_index,
        larger_node_used_features[feature_index],
        GetGlobalDataCountInLeaf(this->larger_leaf_splits_->leaf_index()),
        this->larger_leaf_splits_.get(),
419
420
        &larger_bests_per_thread[tid],
        larger_leaf_parent_output);
421
    OMP_LOOP_EX_END();
Guolin Ke's avatar
Guolin Ke committed
422
  }
423
  OMP_THROW_EX();
Guolin Ke's avatar
Guolin Ke committed
424

Guolin Ke's avatar
Guolin Ke committed
425
  auto smaller_best_idx = ArrayArgs<SplitInfo>::ArgMax(smaller_bests_per_thread);
426
  int leaf = this->smaller_leaf_splits_->leaf_index();
Guolin Ke's avatar
Guolin Ke committed
427
  this->best_split_per_leaf_[leaf] = smaller_bests_per_thread[smaller_best_idx];
Guolin Ke's avatar
Guolin Ke committed
428

429
430
  if (this->larger_leaf_splits_ != nullptr &&  this->larger_leaf_splits_->leaf_index() >= 0) {
    leaf = this->larger_leaf_splits_->leaf_index();
Guolin Ke's avatar
Guolin Ke committed
431
432
433
    auto larger_best_idx = ArrayArgs<SplitInfo>::ArgMax(larger_bests_per_thread);
    this->best_split_per_leaf_[leaf] = larger_bests_per_thread[larger_best_idx];
  }
Guolin Ke's avatar
Guolin Ke committed
434

Guolin Ke's avatar
Guolin Ke committed
435
  SplitInfo smaller_best_split, larger_best_split;
436
  smaller_best_split = this->best_split_per_leaf_[this->smaller_leaf_splits_->leaf_index()];
Guolin Ke's avatar
Guolin Ke committed
437
  // find local best split for larger leaf
438
439
  if (this->larger_leaf_splits_->leaf_index() >= 0) {
    larger_best_split = this->best_split_per_leaf_[this->larger_leaf_splits_->leaf_index()];
Guolin Ke's avatar
Guolin Ke committed
440
441
442
  }

  // sync global best info
Guolin Ke's avatar
Guolin Ke committed
443
  SyncUpGlobalBestSplit(input_buffer_.data(), input_buffer_.data(), &smaller_best_split, &larger_best_split, this->config_->max_cat_threshold);
Guolin Ke's avatar
Guolin Ke committed
444
445

  // set best split
446
447
448
  this->best_split_per_leaf_[this->smaller_leaf_splits_->leaf_index()] = smaller_best_split;
  if (this->larger_leaf_splits_->leaf_index() >= 0) {
    this->best_split_per_leaf_[this->larger_leaf_splits_->leaf_index()] = larger_best_split;
Guolin Ke's avatar
Guolin Ke committed
449
450
451
  }
}

452
453
template <typename TREELEARNER_T>
void DataParallelTreeLearner<TREELEARNER_T>::Split(Tree* tree, int best_Leaf, int* left_leaf, int* right_leaf) {
454
  TREELEARNER_T::SplitInner(tree, best_Leaf, left_leaf, right_leaf, false);
455
  const SplitInfo& best_split_info = this->best_split_per_leaf_[best_Leaf];
Guolin Ke's avatar
Guolin Ke committed
456
457
458
  // need update global number of data in leaf
  global_data_count_in_leaf_[*left_leaf] = best_split_info.left_count;
  global_data_count_in_leaf_[*right_leaf] = best_split_info.right_count;
459
460
461
462
  // reset hist num bits according to global num data
  if (this->config_->use_quantized_grad) {
    this->gradient_discretizer_->template SetNumBitsInHistogramBin<true>(*left_leaf, *right_leaf, GetGlobalDataCountInLeaf(*left_leaf), GetGlobalDataCountInLeaf(*right_leaf));
  }
Guolin Ke's avatar
Guolin Ke committed
463
464
}

465
466
467
// instantiate template classes, otherwise linker cannot find the code
template class DataParallelTreeLearner<GPUTreeLearner>;
template class DataParallelTreeLearner<SerialTreeLearner>;
Guolin Ke's avatar
Guolin Ke committed
468
469

}  // namespace LightGBM