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

Unify CPU hist sketching (#5880)

Browse Source
This commit is contained in:
Jiaming Yuan
2020-08-12 01:33:06 +08:00
committed by GitHub
parent bd6b7f4aa7
commit ee70a2380b
18 changed files with 648 additions and 677 deletions
Download Patch File Download Diff File Expand all files Collapse all files

349
src/common/hist_util.cc
View File

@@ -113,346 +113,12 @@ void GHistIndexMatrix::ResizeIndex(const size_t rbegin, const SparsePage& batch,
}
HistogramCuts::HistogramCuts() {
monitor_.Init(__FUNCTION__);
cut_ptrs_.HostVector().emplace_back(0);
}
// Dispatch to specific builder.
void HistogramCuts::Build(DMatrix* dmat, uint32_t const max_num_bins) {
auto const& info = dmat->Info();
size_t const total = info.num_row_ * info.num_col_;
size_t const nnz = info.num_nonzero_;
float const sparsity = static_cast<float>(nnz) / static_cast<float>(total);
// Use a small number to avoid calling `dmat->GetColumnBatches'.
float constexpr kSparsityThreshold = 0.0005;
// FIXME(trivialfis): Distributed environment is not supported.
if (sparsity < kSparsityThreshold && (!rabit::IsDistributed())) {
LOG(INFO) << "Building quantile cut on a sparse dataset.";
SparseCuts cuts(this);
cuts.Build(dmat, max_num_bins);
} else {
LOG(INFO) << "Building quantile cut on a dense dataset or distributed environment.";
DenseCuts cuts(this);
cuts.Build(dmat, max_num_bins);
}
LOG(INFO) << "Total number of hist bins: " << cut_ptrs_.HostVector().back();
}
bool CutsBuilder::UseGroup(DMatrix* dmat) {
auto& info = dmat->Info();
return CutsBuilder::UseGroup(info);
}
bool CutsBuilder::UseGroup(MetaInfo const& info) {
size_t const num_groups = info.group_ptr_.size() == 0 ?
0 : info.group_ptr_.size() - 1;
// Use group index for weights?
bool const use_group_ind = num_groups != 0 &&
(info.weights_.Size() != info.num_row_);
return use_group_ind;
}
void SparseCuts::SingleThreadBuild(SparsePage const& page, MetaInfo const& info,
uint32_t max_num_bins,
bool const use_group_ind,
uint32_t beg_col, uint32_t end_col,
uint32_t thread_id) {
CHECK_GE(end_col, beg_col);
// Data groups, used in ranking.
std::vector<bst_uint> const& group_ptr = info.group_ptr_;
auto &local_min_vals = p_cuts_->min_vals_.HostVector();
auto &local_cuts = p_cuts_->cut_values_.HostVector();
auto &local_ptrs = p_cuts_->cut_ptrs_.HostVector();
local_min_vals.resize(end_col - beg_col, 0);
for (uint32_t col_id = beg_col; col_id < page.Size() && col_id < end_col; ++col_id) {
// Using a local variable makes things easier, but at the cost of memory trashing.
WQSketch sketch;
common::Span<xgboost::Entry const> const column = page[col_id];
uint32_t const n_bins = std::min(static_cast<uint32_t>(column.size()),
max_num_bins);
if (n_bins == 0) {
// cut_ptrs_ is initialized with a zero, so there's always an element at the back
CHECK_GE(local_ptrs.size(), 1);
local_ptrs.emplace_back(local_ptrs.back());
continue;
}
sketch.Init(info.num_row_, 1.0 / (n_bins * WQSketch::kFactor));
for (auto const& entry : column) {
uint32_t weight_ind = 0;
if (use_group_ind) {
auto row_idx = entry.index;
uint32_t group_ind =
this->SearchGroupIndFromRow(group_ptr, page.base_rowid + row_idx);
weight_ind = group_ind;
} else {
weight_ind = entry.index;
}
sketch.Push(entry.fvalue, info.GetWeight(weight_ind));
}
WQSketch::SummaryContainer out_summary;
sketch.GetSummary(&out_summary);
WQSketch::SummaryContainer summary;
summary.Reserve(n_bins + 1);
summary.SetPrune(out_summary, n_bins + 1);
// Can be use data[1] as the min values so that we don't need to
// store another array?
float mval = summary.data[0].value;
local_min_vals[col_id - beg_col] = mval - (fabs(mval) + 1e-5);
this->AddCutPoint(summary, max_num_bins);
bst_float cpt = (summary.size > 0) ?
summary.data[summary.size - 1].value :
local_min_vals[col_id - beg_col];
cpt += fabs(cpt) + 1e-5;
local_cuts.emplace_back(cpt);
local_ptrs.emplace_back(local_cuts.size());
}
}
std::vector<size_t> SparseCuts::LoadBalance(SparsePage const& page,
size_t const nthreads) {
/* Some sparse datasets have their mass concentrating on small
* number of features. To avoid wating for a few threads running
* forever, we here distirbute different number of columns to
* different threads according to number of entries. */
size_t const total_entries = page.data.Size();
size_t const entries_per_thread = common::DivRoundUp(total_entries, nthreads);
std::vector<size_t> cols_ptr(nthreads+1, 0);
size_t count {0};
size_t current_thread {1};
for (size_t col_id = 0; col_id < page.Size(); ++col_id) {
auto const column = page[col_id];
cols_ptr[current_thread]++; // add one column to thread
count += column.size();
if (count > entries_per_thread + 1) {
current_thread++;
count = 0;
cols_ptr[current_thread] = cols_ptr[current_thread-1];
}
}
// Idle threads.
for (; current_thread < cols_ptr.size() - 1; ++current_thread) {
cols_ptr[current_thread+1] = cols_ptr[current_thread];
}
return cols_ptr;
}
void SparseCuts::Build(DMatrix* dmat, uint32_t const max_num_bins) {
monitor_.Start(__FUNCTION__);
// Use group index for weights?
auto use_group = UseGroup(dmat);
uint32_t nthreads = omp_get_max_threads();
CHECK_GT(nthreads, 0);
std::vector<HistogramCuts> cuts_containers(nthreads);
std::vector<std::unique_ptr<SparseCuts>> sparse_cuts(nthreads);
for (size_t i = 0; i < nthreads; ++i) {
sparse_cuts[i].reset(new SparseCuts(&cuts_containers[i]));
}
for (auto const& page : dmat->GetBatches<CSCPage>()) {
CHECK_LE(page.Size(), dmat->Info().num_col_);
monitor_.Start("Load balance");
std::vector<size_t> col_ptr = LoadBalance(page, nthreads);
monitor_.Stop("Load balance");
// We here decouples the logic between build and parallelization
// to simplify things a bit.
#pragma omp parallel for num_threads(nthreads) schedule(static)
for (omp_ulong i = 0; i < nthreads; ++i) {
common::Monitor t_monitor;
t_monitor.Init("SingleThreadBuild: " + std::to_string(i));
t_monitor.Start(std::to_string(i));
sparse_cuts[i]->SingleThreadBuild(page, dmat->Info(), max_num_bins, use_group,
col_ptr[i], col_ptr[i+1], i);
t_monitor.Stop(std::to_string(i));
}
this->Concat(sparse_cuts, dmat->Info().num_col_);
}
monitor_.Stop(__FUNCTION__);
}
void SparseCuts::Concat(
std::vector<std::unique_ptr<SparseCuts>> const& cuts, uint32_t n_cols) {
monitor_.Start(__FUNCTION__);
uint32_t nthreads = omp_get_max_threads();
auto &local_min_vals = p_cuts_->min_vals_.HostVector();
auto &local_cuts = p_cuts_->cut_values_.HostVector();
auto &local_ptrs = p_cuts_->cut_ptrs_.HostVector();
local_min_vals.resize(n_cols, std::numeric_limits<float>::max());
size_t min_vals_tail = 0;
for (uint32_t t = 0; t < nthreads; ++t) {
auto& thread_min_vals = cuts[t]->p_cuts_->min_vals_.HostVector();
auto& thread_cuts = cuts[t]->p_cuts_->cut_values_.HostVector();
auto& thread_ptrs = cuts[t]->p_cuts_->cut_ptrs_.HostVector();
// concat csc pointers.
size_t const old_ptr_size = local_ptrs.size();
local_ptrs.resize(
thread_ptrs.size() + local_ptrs.size() - 1);
size_t const new_icp_size = local_ptrs.size();
auto tail = local_ptrs[old_ptr_size-1];
for (size_t j = old_ptr_size; j < new_icp_size; ++j) {
local_ptrs[j] = tail + thread_ptrs[j-old_ptr_size+1];
}
// concat csc values
size_t const old_iv_size = local_cuts.size();
local_cuts.resize(
thread_cuts.size() + local_cuts.size());
size_t const new_iv_size = local_cuts.size();
for (size_t j = old_iv_size; j < new_iv_size; ++j) {
local_cuts[j] = thread_cuts[j-old_iv_size];
}
// merge min values
for (size_t j = 0; j < thread_min_vals.size(); ++j) {
local_min_vals.at(min_vals_tail + j) =
std::min(local_min_vals.at(min_vals_tail + j), thread_min_vals.at(j));
}
min_vals_tail += thread_min_vals.size();
}
monitor_.Stop(__FUNCTION__);
}
void DenseCuts::Build(DMatrix* p_fmat, uint32_t max_num_bins) {
monitor_.Start(__FUNCTION__);
const MetaInfo& info = p_fmat->Info();
// safe factor for better accuracy
std::vector<WQSketch> sketchs;
const int nthread = omp_get_max_threads();
unsigned const nstep =
static_cast<unsigned>((info.num_col_ + nthread - 1) / nthread);
unsigned const ncol = static_cast<unsigned>(info.num_col_);
sketchs.resize(info.num_col_);
for (auto& s : sketchs) {
s.Init(info.num_row_, 1.0 / (max_num_bins * WQSketch::kFactor));
}
// Data groups, used in ranking.
std::vector<bst_uint> const& group_ptr = info.group_ptr_;
size_t const num_groups = group_ptr.size() == 0 ? 0 : group_ptr.size() - 1;
// Use group index for weights?
bool const use_group = UseGroup(p_fmat);
const bool isDense = p_fmat->IsDense();
for (const auto &batch : p_fmat->GetBatches<SparsePage>()) {
size_t group_ind = 0;
if (use_group) {
group_ind = this->SearchGroupIndFromRow(group_ptr, batch.base_rowid);
}
#pragma omp parallel num_threads(nthread) firstprivate(group_ind, use_group)
{
CHECK_EQ(nthread, omp_get_num_threads());
auto tid = static_cast<unsigned>(omp_get_thread_num());
unsigned begin = std::min(nstep * tid, ncol);
unsigned end = std::min(nstep * (tid + 1), ncol);
// do not iterate if no columns are assigned to the thread
if (begin < end && end <= ncol) {
for (size_t i = 0; i < batch.Size(); ++i) { // NOLINT(*)
size_t const ridx = batch.base_rowid + i;
SparsePage::Inst const inst = batch[i];
if (use_group &&
group_ptr[group_ind] == ridx &&
// maximum equals to weights.size() - 1
group_ind < num_groups - 1) {
// move to next group
group_ind++;
}
size_t w_idx = use_group ? group_ind : ridx;
auto w = info.GetWeight(w_idx);
if (isDense) {
auto data = inst.data();
for (size_t ii = begin; ii < end; ii++) {
sketchs[ii].Push(data[ii].fvalue, w);
}
} else {
for (auto const& entry : inst) {
if (entry.index >= begin && entry.index < end) {
sketchs[entry.index].Push(entry.fvalue, w);
}
}
}
}
}
}
}
Init(&sketchs, max_num_bins, info.num_row_);
monitor_.Stop(__FUNCTION__);
}
/**
* \param [in,out] in_sketchs
* \param max_num_bins The maximum number bins.
* \param max_rows Number of rows in this DMatrix.
*/
void DenseCuts::Init
(std::vector<WQSketch>* in_sketchs, uint32_t max_num_bins, size_t max_rows) {
monitor_.Start(__func__);
std::vector<WQSketch>& sketchs = *in_sketchs;
// Compute how many cuts samples we need at each node
// Do not require more than the number of total rows in training data
// This allows efficient training on wide data
size_t global_max_rows = max_rows;
rabit::Allreduce<rabit::op::Sum>(&global_max_rows, 1);
size_t intermediate_num_cuts =
std::min(global_max_rows, static_cast<size_t>(max_num_bins * WQSketch::kFactor));
// gather the histogram data
rabit::SerializeReducer<WQSketch::SummaryContainer> sreducer;
std::vector<WQSketch::SummaryContainer> summary_array;
summary_array.resize(sketchs.size());
for (size_t i = 0; i < sketchs.size(); ++i) {
WQSketch::SummaryContainer out;
sketchs[i].GetSummary(&out);
summary_array[i].Reserve(intermediate_num_cuts);
summary_array[i].SetPrune(out, intermediate_num_cuts);
}
CHECK_EQ(summary_array.size(), in_sketchs->size());
size_t nbytes = WQSketch::SummaryContainer::CalcMemCost(intermediate_num_cuts);
// TODO(chenqin): rabit failure recovery assumes no boostrap onetime call after loadcheckpoint
// we need to move this allreduce before loadcheckpoint call in future
sreducer.Allreduce(dmlc::BeginPtr(summary_array), nbytes, summary_array.size());
p_cuts_->min_vals_.HostVector().resize(sketchs.size());
for (size_t fid = 0; fid < summary_array.size(); ++fid) {
WQSketch::SummaryContainer a;
a.Reserve(max_num_bins + 1);
a.SetPrune(summary_array[fid], max_num_bins + 1);
const bst_float mval = a.data[0].value;
p_cuts_->min_vals_.HostVector()[fid] = mval - (fabs(mval) + 1e-5);
AddCutPoint(a, max_num_bins);
// push a value that is greater than anything
const bst_float cpt
= (a.size > 0) ? a.data[a.size - 1].value : p_cuts_->min_vals_.HostVector()[fid];
// this must be bigger than last value in a scale
const bst_float last = cpt + (fabs(cpt) + 1e-5);
p_cuts_->cut_values_.HostVector().push_back(last);
// Ensure that every feature gets at least one quantile point
CHECK_LE(p_cuts_->cut_values_.HostVector().size(), std::numeric_limits<uint32_t>::max());
auto cut_size = static_cast<uint32_t>(p_cuts_->cut_values_.HostVector().size());
CHECK_GT(cut_size, p_cuts_->cut_ptrs_.HostVector().back());
p_cuts_->cut_ptrs_.HostVector().push_back(cut_size);
}
monitor_.Stop(__func__);
}
void GHistIndexMatrix::Init(DMatrix* p_fmat, int max_bins) {
cut.Build(p_fmat, max_bins);
cut = SketchOnDMatrix(p_fmat, max_bins);
max_num_bins = max_bins;
const int32_t nthread = omp_get_max_threads();
const uint32_t nbins = cut.Ptrs().back();
@@ -1048,12 +714,11 @@ void BuildHistKernel(const std::vector<GradientPair>& gpair,
}
}
template<typename GradientSumT>
void GHistBuilder<GradientSumT>::BuildHist(const std::vector<GradientPair>& gpair,
const RowSetCollection::Elem row_indices,
const GHistIndexMatrix& gmat,
GHistRowT hist,
bool isDense) {
template <typename GradientSumT>
void GHistBuilder<GradientSumT>::BuildHist(
const std::vector<GradientPair> &gpair,
const RowSetCollection::Elem row_indices, const GHistIndexMatrix &gmat,
GHistRowT hist, bool isDense) {
const size_t nrows = row_indices.Size();
const size_t no_prefetch_size = Prefetch::NoPrefetchSize(nrows);