xgboost
hist_util.h
Go to the documentation of this file.
1 
7 #ifndef XGBOOST_COMMON_HIST_UTIL_H_
8 #define XGBOOST_COMMON_HIST_UTIL_H_
9 
10 #include <xgboost/data.h>
12 #include <limits>
13 #include <vector>
14 #include <algorithm>
15 #include <memory>
16 #include <utility>
17 #include <map>
18 
19 #include "row_set.h"
20 #include "common.h"
21 #include "threading_utils.h"
22 #include "../tree/param.h"
23 #include "./quantile.h"
24 #include "./timer.h"
25 #include "../include/rabit/rabit.h"
26 
27 namespace xgboost {
28 namespace common {
34 
35 // A CSC matrix representing histogram cuts, used in CPU quantile hist.
36 // The cut values represent upper bounds of bins containing approximately equal numbers of elements
38  protected:
39  using BinIdx = uint32_t;
40 
41  public:
44  // storing minimum value in a sketch set.
46 
47  HistogramCuts();
53  cut_ptrs_.Copy(that.cut_ptrs_);
54  min_vals_.Copy(that.min_vals_);
55  }
56 
57  HistogramCuts(HistogramCuts&& that) noexcept(true) {
58  *this = std::forward<HistogramCuts&&>(that);
59  }
60 
66  cut_ptrs_.Copy(that.cut_ptrs_);
67  min_vals_.Copy(that.min_vals_);
68  return *this;
69  }
70 
71  HistogramCuts& operator=(HistogramCuts&& that) noexcept(true) {
72  cut_ptrs_ = std::move(that.cut_ptrs_);
73  cut_values_ = std::move(that.cut_values_);
74  min_vals_ = std::move(that.min_vals_);
75  return *this;
76  }
77 
78  uint32_t FeatureBins(uint32_t feature) const {
79  return cut_ptrs_.ConstHostVector().at(feature + 1) -
80  cut_ptrs_.ConstHostVector()[feature];
81  }
82 
83  // Getters. Cuts should be of no use after building histogram indices, but currently
84  // it's deeply linked with quantile_hist, gpu sketcher and gpu_hist. So we preserve
85  // these for now.
86  std::vector<uint32_t> const& Ptrs() const { return cut_ptrs_.ConstHostVector(); }
87  std::vector<float> const& Values() const { return cut_values_.ConstHostVector(); }
88  std::vector<float> const& MinValues() const { return min_vals_.ConstHostVector(); }
89 
90  size_t TotalBins() const { return cut_ptrs_.ConstHostVector().back(); }
91 
92  // Return the index of a cut point that is strictly greater than the input
93  // value, or the last available index if none exists
94  BinIdx SearchBin(float value, uint32_t column_id) const {
95  auto beg = cut_ptrs_.ConstHostVector().at(column_id);
96  auto end = cut_ptrs_.ConstHostVector().at(column_id + 1);
97  const auto &values = cut_values_.ConstHostVector();
98  auto it = std::upper_bound(values.cbegin() + beg, values.cbegin() + end, value);
99  BinIdx idx = it - values.cbegin();
100  if (idx == end) {
101  idx -= 1;
102  }
103  return idx;
104  }
105 
106  BinIdx SearchBin(Entry const& e) const {
107  return SearchBin(e.fvalue, e.index);
108  }
109 };
110 
111 inline HistogramCuts SketchOnDMatrix(DMatrix *m, int32_t max_bins) {
112  HistogramCuts out;
113  auto const& info = m->Info();
114  const auto threads = omp_get_max_threads();
115  std::vector<std::vector<bst_row_t>> column_sizes(threads);
116  for (auto& column : column_sizes) {
117  column.resize(info.num_col_, 0);
118  }
119  std::vector<bst_row_t> reduced(info.num_col_, 0);
120  for (auto const& page : m->GetBatches<SparsePage>()) {
121  auto const &entries_per_column =
122  HostSketchContainer::CalcColumnSize(page, info.num_col_, threads);
123  for (size_t i = 0; i < entries_per_column.size(); ++i) {
124  reduced[i] += entries_per_column[i];
125  }
126  }
127  HostSketchContainer container(reduced, max_bins,
129  for (auto const &page : m->GetBatches<SparsePage>()) {
130  container.PushRowPage(page, info);
131  }
132  container.MakeCuts(&out);
133  return out;
134 }
135 
140 };
141 
142 struct Index {
143  Index() {
144  SetBinTypeSize(binTypeSize_);
145  }
146  Index(const Index& i) = delete;
147  Index& operator=(Index i) = delete;
148  Index(Index&& i) = delete;
149  Index& operator=(Index&& i) = delete;
150  uint32_t operator[](size_t i) const {
151  if (offset_ptr_ != nullptr) {
152  return func_(data_ptr_, i) + offset_ptr_[i%p_];
153  } else {
154  return func_(data_ptr_, i);
155  }
156  }
157  void SetBinTypeSize(BinTypeSize binTypeSize) {
158  binTypeSize_ = binTypeSize;
159  switch (binTypeSize) {
160  case kUint8BinsTypeSize:
161  func_ = &GetValueFromUint8;
162  break;
163  case kUint16BinsTypeSize:
164  func_ = &GetValueFromUint16;
165  break;
166  case kUint32BinsTypeSize:
167  func_ = &GetValueFromUint32;
168  break;
169  default:
170  CHECK(binTypeSize == kUint8BinsTypeSize ||
171  binTypeSize == kUint16BinsTypeSize ||
172  binTypeSize == kUint32BinsTypeSize);
173  }
174  }
176  return binTypeSize_;
177  }
178  template<typename T>
179  T* data() const { // NOLINT
180  return static_cast<T*>(data_ptr_);
181  }
182  uint32_t* Offset() const {
183  return offset_ptr_;
184  }
185  size_t OffsetSize() const {
186  return offset_.size();
187  }
188  size_t Size() const {
189  return data_.size() / (binTypeSize_);
190  }
191  void Resize(const size_t nBytesData) {
192  data_.resize(nBytesData);
193  data_ptr_ = reinterpret_cast<void*>(data_.data());
194  }
195  void ResizeOffset(const size_t nDisps) {
196  offset_.resize(nDisps);
197  offset_ptr_ = offset_.data();
198  p_ = nDisps;
199  }
200  std::vector<uint8_t>::const_iterator begin() const { // NOLINT
201  return data_.begin();
202  }
203  std::vector<uint8_t>::const_iterator end() const { // NOLINT
204  return data_.end();
205  }
206 
207  private:
208  static uint32_t GetValueFromUint8(void *t, size_t i) {
209  return reinterpret_cast<uint8_t*>(t)[i];
210  }
211  static uint32_t GetValueFromUint16(void* t, size_t i) {
212  return reinterpret_cast<uint16_t*>(t)[i];
213  }
214  static uint32_t GetValueFromUint32(void* t, size_t i) {
215  return reinterpret_cast<uint32_t*>(t)[i];
216  }
217 
218  using Func = uint32_t (*)(void*, size_t);
219 
220  std::vector<uint8_t> data_;
221  std::vector<uint32_t> offset_; // size of this field is equal to number of features
222  void* data_ptr_;
223  BinTypeSize binTypeSize_ {kUint8BinsTypeSize};
224  size_t p_ {1};
225  uint32_t* offset_ptr_ {nullptr};
226  Func func_;
227 };
228 
229 
238  std::vector<size_t> row_ptr;
242  std::vector<size_t> hit_count;
246  size_t max_num_bins;
247  // Create a global histogram matrix, given cut
248  void Init(DMatrix* p_fmat, int max_num_bins);
249 
250  // specific method for sparse data as no posibility to reduce allocated memory
251  template <typename BinIdxType, typename GetOffset>
253  size_t batch_threads, const SparsePage &batch,
254  size_t rbegin, size_t nbins, GetOffset get_offset) {
255  const xgboost::Entry *data_ptr = batch.data.HostVector().data();
256  const std::vector<bst_row_t> &offset_vec = batch.offset.HostVector();
257  const size_t batch_size = batch.Size();
258  CHECK_LT(batch_size, offset_vec.size());
259  BinIdxType* index_data = index_data_span.data();
260  ParallelFor(omp_ulong(batch_size), batch_threads, [&](omp_ulong i) {
261  const int tid = omp_get_thread_num();
262  size_t ibegin = row_ptr[rbegin + i];
263  size_t iend = row_ptr[rbegin + i + 1];
264  const size_t size = offset_vec[i + 1] - offset_vec[i];
265  SparsePage::Inst inst = {data_ptr + offset_vec[i], size};
266  CHECK_EQ(ibegin + inst.size(), iend);
267  for (bst_uint j = 0; j < inst.size(); ++j) {
268  uint32_t idx = cut.SearchBin(inst[j]);
269  index_data[ibegin + j] = get_offset(idx, j);
270  ++hit_count_tloc_[tid * nbins + idx];
271  }
272  });
273  }
274 
275  void ResizeIndex(const size_t n_index,
276  const bool isDense);
277 
278  inline void GetFeatureCounts(size_t* counts) const {
279  auto nfeature = cut.Ptrs().size() - 1;
280  for (unsigned fid = 0; fid < nfeature; ++fid) {
281  auto ibegin = cut.Ptrs()[fid];
282  auto iend = cut.Ptrs()[fid + 1];
283  for (auto i = ibegin; i < iend; ++i) {
284  counts[fid] += hit_count[i];
285  }
286  }
287  }
288  inline bool IsDense() const {
289  return isDense_;
290  }
291 
292  private:
293  std::vector<size_t> hit_count_tloc_;
294  bool isDense_;
295 };
296 
297 template <typename GradientIndex>
299  GradientIndex const &data,
300  uint32_t const fidx_begin,
301  uint32_t const fidx_end) {
302  uint32_t previous_middle = std::numeric_limits<uint32_t>::max();
303  while (end != begin) {
304  auto middle = begin + (end - begin) / 2;
305  if (middle == previous_middle) {
306  break;
307  }
308  previous_middle = middle;
309 
310  auto gidx = data[middle];
311 
312  if (gidx >= fidx_begin && gidx < fidx_end) {
313  return static_cast<int32_t>(gidx);
314  } else if (gidx < fidx_begin) {
315  begin = middle;
316  } else {
317  end = middle;
318  }
319  }
320  // Value is missing
321  return -1;
322 }
323 
325  const size_t* row_ptr;
326  const uint32_t* index;
327 
328  inline GHistIndexBlock(const size_t* row_ptr, const uint32_t* index)
329  : row_ptr(row_ptr), index(index) {}
330 
331  // get i-th row
332  inline GHistIndexRow operator[](size_t i) const {
333  return {&index[0] + row_ptr[i], row_ptr[i + 1] - row_ptr[i]};
334  }
335 };
336 
337 class ColumnMatrix;
338 
340  public:
341  void Init(const GHistIndexMatrix& gmat,
342  const ColumnMatrix& colmat,
343  const tree::TrainParam& param);
344 
345  inline GHistIndexBlock operator[](size_t i) const {
346  return {blocks_[i].row_ptr_begin, blocks_[i].index_begin};
347  }
348 
349  inline size_t GetNumBlock() const {
350  return blocks_.size();
351  }
352 
353  private:
354  std::vector<size_t> row_ptr_;
355  std::vector<uint32_t> index_;
356  const HistogramCuts* cut_;
357  struct Block {
358  const size_t* row_ptr_begin;
359  const size_t* row_ptr_end;
360  const uint32_t* index_begin;
361  const uint32_t* index_end;
362  };
363  std::vector<Block> blocks_;
364 };
365 
366 template<typename GradientSumT>
368 
372 template<typename GradientSumT>
373 void InitilizeHistByZeroes(GHistRow<GradientSumT> hist, size_t begin, size_t end);
374 
378 template<typename GradientSumT>
380  size_t begin, size_t end);
381 
385 template<typename GradientSumT>
387  size_t begin, size_t end);
388 
392 template<typename GradientSumT>
394  const GHistRow<GradientSumT> src2,
395  size_t begin, size_t end);
396 
400 template<typename GradientSumT>
402  public:
405 
406  // access histogram for i-th node
408  constexpr uint32_t kMax = std::numeric_limits<uint32_t>::max();
409  const size_t id = row_ptr_[nid];
410  CHECK_NE(id, kMax);
411  GradientPairT* ptr = nullptr;
412  if (contiguous_allocation_) {
413  ptr = const_cast<GradientPairT*>(data_[0].data() + nbins_*id);
414  } else {
415  ptr = const_cast<GradientPairT*>(data_[id].data());
416  }
417  return {ptr, nbins_};
418  }
419 
420  // have we computed a histogram for i-th node?
421  bool RowExists(bst_uint nid) const {
422  const uint32_t k_max = std::numeric_limits<uint32_t>::max();
423  return (nid < row_ptr_.size() && row_ptr_[nid] != k_max);
424  }
425 
426  // initialize histogram collection
427  void Init(uint32_t nbins) {
428  if (nbins_ != nbins) {
429  nbins_ = nbins;
430  // quite expensive operation, so let's do this only once
431  data_.clear();
432  }
433  row_ptr_.clear();
434  n_nodes_added_ = 0;
435  }
436 
437  // create an empty histogram for i-th node
438  void AddHistRow(bst_uint nid) {
439  constexpr uint32_t kMax = std::numeric_limits<uint32_t>::max();
440  if (nid >= row_ptr_.size()) {
441  row_ptr_.resize(nid + 1, kMax);
442  }
443  CHECK_EQ(row_ptr_[nid], kMax);
444 
445  if (data_.size() < (nid + 1)) {
446  data_.resize((nid + 1));
447  }
448 
449  row_ptr_[nid] = n_nodes_added_;
450  n_nodes_added_++;
451  }
452  // allocate thread local memory i-th node
453  void AllocateData(bst_uint nid) {
454  if (data_[row_ptr_[nid]].size() == 0) {
455  data_[row_ptr_[nid]].resize(nbins_, {0, 0});
456  }
457  }
458  // allocate common buffer contiguously for all nodes, need for single Allreduce call
460  const size_t new_size = nbins_*data_.size();
461  contiguous_allocation_ = true;
462  if (data_[0].size() != new_size) {
463  data_[0].resize(new_size);
464  }
465  }
466 
467  private:
469  uint32_t nbins_ = 0;
471  uint32_t n_nodes_added_ = 0;
473  bool contiguous_allocation_ = false;
474 
475  std::vector<std::vector<GradientPairT>> data_;
476 
478  std::vector<size_t> row_ptr_;
479 };
480 
486 template<typename GradientSumT>
488  public:
490 
491  void Init(size_t nbins) {
492  if (nbins != nbins_) {
493  hist_buffer_.Init(nbins);
494  nbins_ = nbins;
495  }
496  }
497 
498  // Add new elements if needed, mark all hists as unused
499  // targeted_hists - already allocated hists which should contain final results after Reduce() call
500  void Reset(size_t nthreads, size_t nodes, const BlockedSpace2d& space,
501  const std::vector<GHistRowT>& targeted_hists) {
502  hist_buffer_.Init(nbins_);
503  tid_nid_to_hist_.clear();
504  threads_to_nids_map_.clear();
505 
506  targeted_hists_ = targeted_hists;
507 
508  CHECK_EQ(nodes, targeted_hists.size());
509 
510  nodes_ = nodes;
511  nthreads_ = nthreads;
512 
513  MatchThreadsToNodes(space);
516 
517  hist_was_used_.resize(nthreads * nodes_);
518  std::fill(hist_was_used_.begin(), hist_was_used_.end(), static_cast<int>(false));
519  }
520 
521  // Get specified hist, initialize hist by zeros if it wasn't used before
522  GHistRowT GetInitializedHist(size_t tid, size_t nid) {
523  CHECK_LT(nid, nodes_);
524  CHECK_LT(tid, nthreads_);
525 
526  int idx = tid_nid_to_hist_.at({tid, nid});
527  if (idx >= 0) {
528  hist_buffer_.AllocateData(idx);
529  }
530  GHistRowT hist = idx == -1 ? targeted_hists_[nid] : hist_buffer_[idx];
531 
532  if (!hist_was_used_[tid * nodes_ + nid]) {
533  InitilizeHistByZeroes(hist, 0, hist.size());
534  hist_was_used_[tid * nodes_ + nid] = static_cast<int>(true);
535  }
536 
537  return hist;
538  }
539 
540  // Reduce following bins (begin, end] for nid-node in dst across threads
541  void ReduceHist(size_t nid, size_t begin, size_t end) {
542  CHECK_GT(end, begin);
543  CHECK_LT(nid, nodes_);
544 
545  GHistRowT dst = targeted_hists_[nid];
546 
547  bool is_updated = false;
548  for (size_t tid = 0; tid < nthreads_; ++tid) {
549  if (hist_was_used_[tid * nodes_ + nid]) {
550  is_updated = true;
551 
552  int idx = tid_nid_to_hist_.at({tid, nid});
553  GHistRowT src = idx == -1 ? targeted_hists_[nid] : hist_buffer_[idx];
554 
555  if (dst.data() != src.data()) {
556  IncrementHist(dst, src, begin, end);
557  }
558  }
559  }
560  if (!is_updated) {
561  // In distributed mode - some tree nodes can be empty on local machines,
562  // So we need just set local hist by zeros in this case
563  InitilizeHistByZeroes(dst, begin, end);
564  }
565  }
566 
567  protected:
568  void MatchThreadsToNodes(const BlockedSpace2d& space) {
569  const size_t space_size = space.Size();
570  const size_t chunck_size = space_size / nthreads_ + !!(space_size % nthreads_);
571 
572  threads_to_nids_map_.resize(nthreads_ * nodes_, false);
573 
574  for (size_t tid = 0; tid < nthreads_; ++tid) {
575  size_t begin = chunck_size * tid;
576  size_t end = std::min(begin + chunck_size, space_size);
577 
578  if (begin < space_size) {
579  size_t nid_begin = space.GetFirstDimension(begin);
580  size_t nid_end = space.GetFirstDimension(end-1);
581 
582  for (size_t nid = nid_begin; nid <= nid_end; ++nid) {
583  // true - means thread 'tid' will work to compute partial hist for node 'nid'
584  threads_to_nids_map_[tid * nodes_ + nid] = true;
585  }
586  }
587  }
588  }
589 
591  size_t hist_allocated_additionally = 0;
592 
593  for (size_t nid = 0; nid < nodes_; ++nid) {
594  int nthreads_for_nid = 0;
595 
596  for (size_t tid = 0; tid < nthreads_; ++tid) {
597  if (threads_to_nids_map_[tid * nodes_ + nid]) {
598  nthreads_for_nid++;
599  }
600  }
601 
602  // In distributed mode - some tree nodes can be empty on local machines,
603  // set nthreads_for_nid to 0 in this case.
604  // In another case - allocate additional (nthreads_for_nid - 1) histograms,
605  // because one is already allocated externally (will store final result for the node).
606  hist_allocated_additionally += std::max<int>(0, nthreads_for_nid - 1);
607  }
608 
609  for (size_t i = 0; i < hist_allocated_additionally; ++i) {
610  hist_buffer_.AddHistRow(i);
611  }
612  }
613 
615  size_t hist_allocated_additionally = 0;
616 
617  for (size_t nid = 0; nid < nodes_; ++nid) {
618  bool first_hist = true;
619  for (size_t tid = 0; tid < nthreads_; ++tid) {
620  if (threads_to_nids_map_[tid * nodes_ + nid]) {
621  if (first_hist) {
622  tid_nid_to_hist_[{tid, nid}] = -1;
623  first_hist = false;
624  } else {
625  tid_nid_to_hist_[{tid, nid}] = hist_allocated_additionally++;
626  }
627  }
628  }
629  }
630  }
631 
633  size_t nbins_ = 0;
635  size_t nthreads_ = 0;
637  size_t nodes_ = 0;
645  std::vector<int> hist_was_used_;
646 
648  std::vector<bool> threads_to_nids_map_;
650  std::vector<GHistRowT> targeted_hists_;
655  std::map<std::pair<size_t, size_t>, int> tid_nid_to_hist_;
656 };
657 
661 template<typename GradientSumT>
663  public:
665 
666  GHistBuilder() = default;
667  GHistBuilder(size_t nthread, uint32_t nbins) : nthread_{nthread}, nbins_{nbins} {}
668 
669  // construct a histogram via histogram aggregation
670  void BuildHist(const std::vector<GradientPair>& gpair,
671  const RowSetCollection::Elem row_indices,
672  const GHistIndexMatrix& gmat,
673  GHistRowT hist,
674  bool isDense);
675  // same, with feature grouping
676  void BuildBlockHist(const std::vector<GradientPair>& gpair,
677  const RowSetCollection::Elem row_indices,
678  const GHistIndexBlockMatrix& gmatb,
679  GHistRowT hist);
680  // construct a histogram via subtraction trick
681  void SubtractionTrick(GHistRowT self,
682  GHistRowT sibling,
683  GHistRowT parent);
684 
685  uint32_t GetNumBins() const {
686  return nbins_;
687  }
688 
689  private:
691  size_t nthread_ { 0 };
693  uint32_t nbins_ { 0 };
694 };
695 } // namespace common
696 } // namespace xgboost
697 #endif // XGBOOST_COMMON_HIST_UTIL_H_
xgboost::common::GHistIndexBlockMatrix
Definition: hist_util.h:339
xgboost::common::HostSketchContainer::MakeCuts
void MakeCuts(HistogramCuts *cuts)
xgboost::common::GHistIndexMatrix::row_ptr
std::vector< size_t > row_ptr
row pointer to rows by element position
Definition: hist_util.h:238
xgboost::Entry::index
bst_feature_t index
feature index
Definition: data.h:186
xgboost::common::BinTypeSize
BinTypeSize
Definition: hist_util.h:136
quantile.h
util to compute quantiles
xgboost::DMatrix::Info
virtual MetaInfo & Info()=0
meta information of the dataset
XGBOOST_HOST_DEV_INLINE
#define XGBOOST_HOST_DEV_INLINE
Definition: base.h:91
xgboost::common::GHistIndexBlockMatrix::GetNumBlock
size_t GetNumBlock() const
Definition: hist_util.h:349
xgboost::common::IncrementHist
void IncrementHist(GHistRow< GradientSumT > dst, const GHistRow< GradientSumT > add, size_t begin, size_t end)
Increment hist as dst += add in range [begin, end)
xgboost::detail::GradientPairInternal
Implementation of gradient statistics pair. Template specialisation may be used to overload different...
Definition: base.h:141
xgboost::common::HistCollection::AllocateData
void AllocateData(bst_uint nid)
Definition: hist_util.h:453
xgboost::common::ParallelGHistBuilder::GetInitializedHist
GHistRowT GetInitializedHist(size_t tid, size_t nid)
Definition: hist_util.h:522
xgboost::SparsePage::Size
size_t Size() const
Definition: data.h:266
xgboost::common::HostSketchContainer
Definition: quantile.h:704
xgboost::common::ParallelGHistBuilder::nodes_
size_t nodes_
number of nodes which will be processed in parallel
Definition: hist_util.h:637
xgboost::common::HistogramCuts::TotalBins
size_t TotalBins() const
Definition: hist_util.h:90
xgboost::SparsePage
In-memory storage unit of sparse batch, stored in CSR format.
Definition: data.h:243
xgboost::common::BlockedSpace2d::Size
size_t Size() const
Definition: threading_utils.h:84
xgboost::common::HistogramCuts::HistogramCuts
HistogramCuts(HistogramCuts const &that)
Definition: hist_util.h:48
xgboost::common::kUint16BinsTypeSize
@ kUint16BinsTypeSize
Definition: hist_util.h:138
xgboost::common::HistCollection::AllocateAllData
void AllocateAllData()
Definition: hist_util.h:459
xgboost::common::HistogramCuts::BinIdx
uint32_t BinIdx
Definition: hist_util.h:39
xgboost::Entry
Element from a sparse vector.
Definition: data.h:184
row_set.h
Quick Utility to compute subset of rows.
xgboost::common::HistogramCuts
Definition: hist_util.h:37
xgboost::common::GHistIndexMatrix::p_fmat
DMatrix * p_fmat
Definition: hist_util.h:245
xgboost::common::ParallelGHistBuilder::threads_to_nids_map_
std::vector< bool > threads_to_nids_map_
Buffer for additional histograms for Parallel processing
Definition: hist_util.h:648
xgboost::HostDeviceVector< bst_float >
xgboost::common::GHistIndexBlock::row_ptr
const size_t * row_ptr
Definition: hist_util.h:325
xgboost::common::ParallelGHistBuilder::ReduceHist
void ReduceHist(size_t nid, size_t begin, size_t end)
Definition: hist_util.h:541
xgboost::common::HistogramCuts::HistogramCuts
HistogramCuts()
xgboost::common::GHistIndexMatrix::index
Index index
The index data.
Definition: hist_util.h:240
xgboost::common::ParallelGHistBuilder::Reset
void Reset(size_t nthreads, size_t nodes, const BlockedSpace2d &space, const std::vector< GHistRowT > &targeted_hists)
Definition: hist_util.h:500
xgboost::common::Index::ResizeOffset
void ResizeOffset(const size_t nDisps)
Definition: hist_util.h:195
xgboost::common::Index
Definition: hist_util.h:142
xgboost::common::GHistIndexBlockMatrix::Init
void Init(const GHistIndexMatrix &gmat, const ColumnMatrix &colmat, const tree::TrainParam &param)
xgboost::common::HistogramCuts::HistogramCuts
HistogramCuts(HistogramCuts &&that) noexcept(true)
Definition: hist_util.h:57
xgboost::common::Index::Resize
void Resize(const size_t nBytesData)
Definition: hist_util.h:191
xgboost::common::GHistIndexMatrix::cut
HistogramCuts cut
The corresponding cuts.
Definition: hist_util.h:244
xgboost::common::HistogramCuts::cut_ptrs_
HostDeviceVector< uint32_t > cut_ptrs_
Definition: hist_util.h:43
xgboost::common::BlockedSpace2d::GetFirstDimension
size_t GetFirstDimension(size_t i) const
Definition: threading_utils.h:89
xgboost::common::HistogramCuts::SearchBin
BinIdx SearchBin(float value, uint32_t column_id) const
Definition: hist_util.h:94
xgboost::common::ParallelGHistBuilder::tid_nid_to_hist_
std::map< std::pair< size_t, size_t >, int > tid_nid_to_hist_
map pair {tid, nid} to index of allocated histogram from hist_buffer_ and targeted_hists_,...
Definition: hist_util.h:655
xgboost::common::GHistIndexMatrix::GetFeatureCounts
void GetFeatureCounts(size_t *counts) const
Definition: hist_util.h:278
xgboost::common::GHistIndexMatrix::Init
void Init(DMatrix *p_fmat, int max_num_bins)
xgboost::common::CopyHist
void CopyHist(GHistRow< GradientSumT > dst, const GHistRow< GradientSumT > src, size_t begin, size_t end)
Copy hist from src to dst in range [begin, end)
xgboost::DMatrix
Internal data structured used by XGBoost during training.
Definition: data.h:449
xgboost::common::HostSketchContainer::CalcColumnSize
static std::vector< bst_row_t > CalcColumnSize(SparsePage const &page, bst_feature_t const n_columns, size_t const nthreads)
xgboost::common::ColumnMatrix
a collection of columns, with support for construction from GHistIndexMatrix.
Definition: column_matrix.h:101
xgboost::omp_ulong
dmlc::omp_ulong omp_ulong
define unsigned long for openmp loop
Definition: base.h:268
xgboost::common::GHistIndexBlock::index
const uint32_t * index
Definition: hist_util.h:326
xgboost::common::BlockedSpace2d
Definition: threading_utils.h:54
xgboost::common::Index::Offset
uint32_t * Offset() const
Definition: hist_util.h:182
xgboost::common::GHistBuilder::GetNumBins
uint32_t GetNumBins() const
Definition: hist_util.h:685
xgboost::common::ParallelGHistBuilder::AllocateAdditionalHistograms
void AllocateAdditionalHistograms()
Definition: hist_util.h:590
xgboost::common::Index::SetBinTypeSize
void SetBinTypeSize(BinTypeSize binTypeSize)
Definition: hist_util.h:157
xgboost::common::Index::Size
size_t Size() const
Definition: hist_util.h:188
xgboost::common::Index::data
T * data() const
Definition: hist_util.h:179
xgboost::HostDeviceVector::HostVector
std::vector< T > & HostVector()
xgboost::common::GHistIndexMatrix::SetIndexData
void SetIndexData(common::Span< BinIdxType > index_data_span, size_t batch_threads, const SparsePage &batch, size_t rbegin, size_t nbins, GetOffset get_offset)
Definition: hist_util.h:252
xgboost::common::GHistIndexMatrix
preprocessed global index matrix, in CSR format
Definition: hist_util.h:236
xgboost::common::HistCollection::AddHistRow
void AddHistRow(bst_uint nid)
Definition: hist_util.h:438
xgboost::common::HistogramCuts::Ptrs
std::vector< uint32_t > const & Ptrs() const
Definition: hist_util.h:86
xgboost::common::HistCollection::Init
void Init(uint32_t nbins)
Definition: hist_util.h:427
xgboost::common::HistCollection
histogram of gradient statistics for multiple nodes
Definition: hist_util.h:401
xgboost::common::HistogramCuts::min_vals_
HostDeviceVector< float > min_vals_
Definition: hist_util.h:45
xgboost::common::HistCollection::operator[]
GHistRowT operator[](bst_uint nid) const
Definition: hist_util.h:407
xgboost::common::HistogramCuts::operator=
HistogramCuts & operator=(HistogramCuts &&that) noexcept(true)
Definition: hist_util.h:71
xgboost::common::BinarySearchBin
int32_t XGBOOST_HOST_DEV_INLINE BinarySearchBin(bst_uint begin, bst_uint end, GradientIndex const &data, uint32_t const fidx_begin, uint32_t const fidx_end)
Definition: hist_util.h:298
timer.h
xgboost::common::ParallelGHistBuilder
Stores temporary histograms to compute them in parallel Supports processing multiple tree-nodes for n...
Definition: hist_util.h:487
xgboost::common::GHistBuilder::GHistBuilder
GHistBuilder(size_t nthread, uint32_t nbins)
Definition: hist_util.h:667
xgboost::common::Index::operator[]
uint32_t operator[](size_t i) const
Definition: hist_util.h:150
xgboost::common::HistCollection::RowExists
bool RowExists(bst_uint nid) const
Definition: hist_util.h:421
xgboost::common::GHistBuilder::GHistBuilder
GHistBuilder()=default
xgboost::common::GHistIndexBlock::GHistIndexBlock
GHistIndexBlock(const size_t *row_ptr, const uint32_t *index)
Definition: hist_util.h:328
xgboost::common::InitilizeHistByZeroes
void InitilizeHistByZeroes(GHistRow< GradientSumT > hist, size_t begin, size_t end)
fill a histogram by zeros
xgboost::common::HistogramCuts::FeatureBins
uint32_t FeatureBins(uint32_t feature) const
Definition: hist_util.h:78
xgboost::common::HistogramCuts::MinValues
std::vector< float > const & MinValues() const
Definition: hist_util.h:88
xgboost::common::GHistIndexBlock::operator[]
GHistIndexRow operator[](size_t i) const
Definition: hist_util.h:332
xgboost::common::ParallelGHistBuilder::nbins_
size_t nbins_
number of bins in each histogram
Definition: hist_util.h:633
xgboost::common::Index::operator=
Index & operator=(Index i)=delete
xgboost::common::GHistIndexMatrix::hit_count
std::vector< size_t > hit_count
hit count of each index
Definition: hist_util.h:242
xgboost::common::ParallelFor
void ParallelFor(Index size, size_t nthreads, Func fn)
Definition: threading_utils.h:137
xgboost::common::ParallelGHistBuilder::MatchThreadsToNodes
void MatchThreadsToNodes(const BlockedSpace2d &space)
Definition: hist_util.h:568
xgboost::HostDeviceVector::Copy
void Copy(const HostDeviceVector< T > &other)
xgboost::common::GHistIndexMatrix::ResizeIndex
void ResizeIndex(const size_t n_index, const bool isDense)
xgboost::common::SubtractionHist
void SubtractionHist(GHistRow< GradientSumT > dst, const GHistRow< GradientSumT > src1, const GHistRow< GradientSumT > src2, size_t begin, size_t end)
Compute Subtraction: dst = src1 - src2 in range [begin, end)
xgboost::common::ParallelGHistBuilder::nthreads_
size_t nthreads_
number of threads for parallel computation
Definition: hist_util.h:635
xgboost::common::ParallelGHistBuilder::MatchNodeNidPairToHist
void MatchNodeNidPairToHist()
Definition: hist_util.h:614
xgboost::HostDeviceVector::Size
size_t Size() const
xgboost::HostDeviceVector::Resize
void Resize(size_t new_size, T v=T())
xgboost::common::SketchOnDMatrix
HistogramCuts SketchOnDMatrix(DMatrix *m, int32_t max_bins)
Definition: hist_util.h:111
xgboost::common::GHistIndexMatrix::IsDense
bool IsDense() const
Definition: hist_util.h:288
xgboost::common::Index::GetBinTypeSize
BinTypeSize GetBinTypeSize() const
Definition: hist_util.h:175
xgboost::common::HostSketchContainer::UseGroup
static bool UseGroup(MetaInfo const &info)
Definition: quantile.h:725
xgboost::common::GHistIndexMatrix::max_num_bins
size_t max_num_bins
Definition: hist_util.h:246
common.h
Common utilities.
xgboost::common::GHistBuilder::BuildBlockHist
void BuildBlockHist(const std::vector< GradientPair > &gpair, const RowSetCollection::Elem row_indices, const GHistIndexBlockMatrix &gmatb, GHistRowT hist)
xgboost::common::GHistBuilder::GHistRowT
GHistRow< GradientSumT > GHistRowT
Definition: hist_util.h:664
xgboost::common::GHistBuilder::BuildHist
void BuildHist(const std::vector< GradientPair > &gpair, const RowSetCollection::Elem row_indices, const GHistIndexMatrix &gmat, GHistRowT hist, bool isDense)
generic_parameters.h
xgboost::common::Span
span class implementation, based on ISO++20 span<T>. The interface should be the same.
Definition: span.h:142
xgboost::common::kUint32BinsTypeSize
@ kUint32BinsTypeSize
Definition: hist_util.h:139
data.h
The input data structure of xgboost.
xgboost::common::GHistIndexBlockMatrix::operator[]
GHistIndexBlock operator[](size_t i) const
Definition: hist_util.h:345
xgboost::common::kUint8BinsTypeSize
@ kUint8BinsTypeSize
Definition: hist_util.h:137
xgboost::SparsePage::offset
HostDeviceVector< bst_row_t > offset
Definition: data.h:246
threading_utils.h
xgboost::common::Index::Index
Index()
Definition: hist_util.h:143
xgboost::common::ParallelGHistBuilder::Init
void Init(size_t nbins)
Definition: hist_util.h:491
xgboost::common::HistogramCuts::SearchBin
BinIdx SearchBin(Entry const &e) const
Definition: hist_util.h:106
xgboost::common::HostSketchContainer::PushRowPage
void PushRowPage(SparsePage const &page, MetaInfo const &info)
xgboost::HostDeviceVector::ConstHostVector
const std::vector< T > & ConstHostVector() const
xgboost::common::Span::size
constexpr XGBOOST_DEVICE index_type size() const __span_noexcept
Definition: span.h:547
xgboost::common::Span::data
constexpr XGBOOST_DEVICE pointer data() const __span_noexcept
Definition: span.h:542
xgboost::common::HistogramCuts::cut_values_
HostDeviceVector< bst_float > cut_values_
Definition: hist_util.h:42
xgboost::bst_uint
uint32_t bst_uint
unsigned integer type used for feature index.
Definition: base.h:113
xgboost::common::GHistIndexBlock
Definition: hist_util.h:324
xgboost::common::ParallelGHistBuilder::targeted_hists_
std::vector< GHistRowT > targeted_hists_
Contains histograms for final results
Definition: hist_util.h:650
xgboost::Entry::fvalue
bst_float fvalue
feature value
Definition: data.h:188
xgboost::common::Index::begin
std::vector< uint8_t >::const_iterator begin() const
Definition: hist_util.h:200
xgboost::common::ParallelGHistBuilder::hist_buffer_
HistCollection< GradientSumT > hist_buffer_
Buffer for additional histograms for Parallel processing
Definition: hist_util.h:639
xgboost::DMatrix::GetBatches
BatchSet< T > GetBatches(const BatchParam &param={})
Gets batches. Use range based for loop over BatchSet to access individual batches.
xgboost::common::HistogramCuts::operator=
HistogramCuts & operator=(HistogramCuts const &that)
Definition: hist_util.h:61
xgboost::common::Index::end
std::vector< uint8_t >::const_iterator end() const
Definition: hist_util.h:203
xgboost::common::HistogramCuts::Values
std::vector< float > const & Values() const
Definition: hist_util.h:87
xgboost::SparsePage::data
HostDeviceVector< Entry > data
the data of the segments
Definition: data.h:248
xgboost::common::GHistBuilder::SubtractionTrick
void SubtractionTrick(GHistRowT self, GHistRowT sibling, GHistRowT parent)
xgboost::common::ParallelGHistBuilder::hist_was_used_
std::vector< int > hist_was_used_
Marks which hists were used, it means that they should be merged. Contains only {true or false} value...
Definition: hist_util.h:645
xgboost::common::Index::OffsetSize
size_t OffsetSize() const
Definition: hist_util.h:185
xgboost::common::GHistBuilder
builder for histograms of gradient statistics
Definition: hist_util.h:662
xgboost
namespace of xgboost
Definition: base.h:110