Hendrik/xgboost
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

More refactoring to take advantage of collective aggregators (#9081)

Browse Source
This commit is contained in:
Rong Ou
2023-04-25 12:36:09 -07:00
committed by GitHub
parent 49ccae7fb9
commit a320b402a5
10 changed files with 81 additions and 81 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
src/common/hist_util.cc
View File

@@ -45,20 +45,18 @@ HistogramCuts SketchOnDMatrix(DMatrix *m, int32_t max_bins, int32_t n_threads, b
if (!use_sorted) {
HostSketchContainer container(max_bins, m->Info().feature_types.ConstHostSpan(), reduced,
HostSketchContainer::UseGroup(info),
m->Info().IsColumnSplit(), n_threads);
HostSketchContainer::UseGroup(info), n_threads);
for (auto const& page : m->GetBatches<SparsePage>()) {
container.PushRowPage(page, info, hessian);
}
container.MakeCuts(&out);
container.MakeCuts(m->Info(), &out);
} else {
SortedSketchContainer container{max_bins, m->Info().feature_types.ConstHostSpan(), reduced,
HostSketchContainer::UseGroup(info),
m->Info().IsColumnSplit(), n_threads};
HostSketchContainer::UseGroup(info), n_threads};
for (auto const& page : m->GetBatches<SortedCSCPage>()) {
container.PushColPage(page, info, hessian);
}
container.MakeCuts(&out);
container.MakeCuts(m->Info(), &out);
}
return out;

26
src/common/quantile.cc
View File

@@ -6,6 +6,7 @@
#include <limits>
#include <utility>
#include "../collective/aggregator.h"
#include "../collective/communicator-inl.h"
#include "../data/adapter.h"
#include "categorical.h"
@@ -18,13 +19,12 @@ template <typename WQSketch>
SketchContainerImpl<WQSketch>::SketchContainerImpl(std::vector<bst_row_t> columns_size,
int32_t max_bins,
Span<FeatureType const> feature_types,
bool use_group, bool col_split,
bool use_group,
int32_t n_threads)
: feature_types_(feature_types.cbegin(), feature_types.cend()),
columns_size_{std::move(columns_size)},
max_bins_{max_bins},
use_group_ind_{use_group},
col_split_{col_split},
n_threads_{n_threads} {
monitor_.Init(__func__);
CHECK_NE(columns_size_.size(), 0);
@@ -202,10 +202,10 @@ void SketchContainerImpl<WQSketch>::GatherSketchInfo(
}
template <typename WQSketch>
void SketchContainerImpl<WQSketch>::AllreduceCategories() {
void SketchContainerImpl<WQSketch>::AllreduceCategories(MetaInfo const& info) {
auto world_size = collective::GetWorldSize();
auto rank = collective::GetRank();
if (world_size == 1 || col_split_) {
if (world_size == 1 || info.IsColumnSplit()) {
return;
}
@@ -273,6 +273,7 @@ void SketchContainerImpl<WQSketch>::AllreduceCategories() {
template <typename WQSketch>
void SketchContainerImpl<WQSketch>::AllReduce(
MetaInfo const& info,
std::vector<typename WQSketch::SummaryContainer> *p_reduced,
std::vector<int32_t>* p_num_cuts) {
monitor_.Start(__func__);
@@ -281,7 +282,7 @@ void SketchContainerImpl<WQSketch>::AllReduce(
collective::Allreduce<collective::Operation::kMax>(&n_columns, 1);
CHECK_EQ(n_columns, sketches_.size()) << "Number of columns differs across workers";
AllreduceCategories();
AllreduceCategories(info);
auto& num_cuts = *p_num_cuts;
CHECK_EQ(num_cuts.size(), 0);
@@ -292,10 +293,7 @@ void SketchContainerImpl<WQSketch>::AllReduce(
// Prune the intermediate num cuts for synchronization.
std::vector<bst_row_t> global_column_size(columns_size_);
if (!col_split_) {
collective::Allreduce<collective::Operation::kSum>(global_column_size.data(),
global_column_size.size());
}
collective::GlobalSum(info, &global_column_size);
ParallelFor(sketches_.size(), n_threads_, [&](size_t i) {
int32_t intermediate_num_cuts = static_cast<int32_t>(
@@ -316,7 +314,7 @@ void SketchContainerImpl<WQSketch>::AllReduce(
});
auto world = collective::GetWorldSize();
if (world == 1 || col_split_) {
if (world == 1 || info.IsColumnSplit()) {
monitor_.Stop(__func__);
return;
}
@@ -382,11 +380,11 @@ auto AddCategories(std::set<float> const &categories, HistogramCuts *cuts) {
}
template <typename WQSketch>
void SketchContainerImpl<WQSketch>::MakeCuts(HistogramCuts* cuts) {
void SketchContainerImpl<WQSketch>::MakeCuts(MetaInfo const& info, HistogramCuts* cuts) {
monitor_.Start(__func__);
std::vector<typename WQSketch::SummaryContainer> reduced;
std::vector<int32_t> num_cuts;
this->AllReduce(&reduced, &num_cuts);
this->AllReduce(info, &reduced, &num_cuts);
cuts->min_vals_.HostVector().resize(sketches_.size(), 0.0f);
std::vector<typename WQSketch::SummaryContainer> final_summaries(reduced.size());
@@ -443,8 +441,8 @@ template class SketchContainerImpl<WXQuantileSketch<float, float>>;
HostSketchContainer::HostSketchContainer(int32_t max_bins, common::Span<FeatureType const> ft,
std::vector<size_t> columns_size, bool use_group,
bool col_split, int32_t n_threads)
: SketchContainerImpl{columns_size, max_bins, ft, use_group, col_split, n_threads} {
int32_t n_threads)
: SketchContainerImpl{columns_size, max_bins, ft, use_group, n_threads} {
monitor_.Init(__func__);
ParallelFor(sketches_.size(), n_threads_, Sched::Auto(), [&](auto i) {
auto n_bins = std::min(static_cast<size_t>(max_bins_), columns_size_[i]);

16
src/common/quantile.h
View File

@@ -789,7 +789,6 @@ class SketchContainerImpl {
std::vector<bst_row_t> columns_size_;
int32_t max_bins_;
bool use_group_ind_{false};
bool col_split_;
int32_t n_threads_;
bool has_categorical_{false};
Monitor monitor_;
@@ -802,7 +801,7 @@ class SketchContainerImpl {
* \param use_group whether is assigned to group to data instance.
*/
SketchContainerImpl(std::vector<bst_row_t> columns_size, int32_t max_bins,
common::Span<FeatureType const> feature_types, bool use_group, bool col_split,
common::Span<FeatureType const> feature_types, bool use_group,
int32_t n_threads);
static bool UseGroup(MetaInfo const &info) {
@@ -829,7 +828,7 @@ class SketchContainerImpl {
std::vector<bst_row_t> *p_sketches_scan,
std::vector<typename WQSketch::Entry> *p_global_sketches);
// Merge sketches from all workers.
void AllReduce(std::vector<typename WQSketch::SummaryContainer> *p_reduced,
void AllReduce(MetaInfo const& info, std::vector<typename WQSketch::SummaryContainer> *p_reduced,
std::vector<int32_t> *p_num_cuts);
template <typename Batch, typename IsValid>
@@ -883,11 +882,11 @@ class SketchContainerImpl {
/* \brief Push a CSR matrix. */
void PushRowPage(SparsePage const &page, MetaInfo const &info, Span<float const> hessian = {});
void MakeCuts(HistogramCuts* cuts);
void MakeCuts(MetaInfo const& info, HistogramCuts* cuts);
private:
// Merge all categories from other workers.
void AllreduceCategories();
void AllreduceCategories(MetaInfo const& info);
};
class HostSketchContainer : public SketchContainerImpl<WQuantileSketch<float, float>> {
@@ -896,8 +895,7 @@ class HostSketchContainer : public SketchContainerImpl<WQuantileSketch<float, fl
public:
HostSketchContainer(int32_t max_bins, common::Span<FeatureType const> ft,
std::vector<size_t> columns_size, bool use_group, bool col_split,
int32_t n_threads);
std::vector<size_t> columns_size, bool use_group, int32_t n_threads);
template <typename Batch>
void PushAdapterBatch(Batch const &batch, size_t base_rowid, MetaInfo const &info, float missing);
@@ -993,9 +991,9 @@ class SortedSketchContainer : public SketchContainerImpl<WXQuantileSketch<float,
public:
explicit SortedSketchContainer(int32_t max_bins, common::Span<FeatureType const> ft,
std::vector<size_t> columns_size, bool use_group, bool col_split,
std::vector<size_t> columns_size, bool use_group,
int32_t n_threads)
: SketchContainerImpl{columns_size, max_bins, ft, use_group, col_split, n_threads} {
: SketchContainerImpl{columns_size, max_bins, ft, use_group, n_threads} {
monitor_.Init(__func__);
sketches_.resize(columns_size.size());
size_t i = 0;