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

Move GHistIndex into DMatrix. (#7064)

Browse Source
This commit is contained in:
Jiaming Yuan
2021-07-01 00:44:49 +08:00
committed by GitHub
parent 1c8fdf2218
commit 1cd20efe68
17 changed files with 386 additions and 320 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
src/common/column_matrix.h
View File

@@ -12,6 +12,7 @@
#include <vector>
#include <memory>
#include "hist_util.h"
#include "../data/gradient_index.h"
namespace xgboost {
namespace common {
@@ -262,9 +263,10 @@ class ColumnMatrix {
return res;
}
template<typename T>
inline void SetIndexAllDense(T* index, const GHistIndexMatrix& gmat, const size_t nrow,
const size_t nfeature, const bool noMissingValues) {
template <typename T>
inline void SetIndexAllDense(T *index, const GHistIndexMatrix &gmat,
const size_t nrow, const size_t nfeature,
const bool noMissingValues) {
T* local_index = reinterpret_cast<T*>(&index_[0]);
/* missing values make sense only for column with type kDenseColumn,

205
src/common/hist_util.cc
View File

@@ -16,6 +16,7 @@
#include "column_matrix.h"
#include "quantile.h"
#include "./../tree/updater_quantile_hist.h"
#include "../data/gradient_index.h"
#if defined(XGBOOST_MM_PREFETCH_PRESENT)
#include <xmmintrin.h>
@@ -29,164 +30,10 @@
namespace xgboost {
namespace common {
void GHistIndexMatrix::ResizeIndex(const size_t n_index,
const bool isDense) {
if ((max_num_bins - 1 <= static_cast<int>(std::numeric_limits<uint8_t>::max())) && isDense) {
index.SetBinTypeSize(kUint8BinsTypeSize);
index.Resize((sizeof(uint8_t)) * n_index);
} else if ((max_num_bins - 1 > static_cast<int>(std::numeric_limits<uint8_t>::max()) &&
max_num_bins - 1 <= static_cast<int>(std::numeric_limits<uint16_t>::max())) && isDense) {
index.SetBinTypeSize(kUint16BinsTypeSize);
index.Resize((sizeof(uint16_t)) * n_index);
} else {
index.SetBinTypeSize(kUint32BinsTypeSize);
index.Resize((sizeof(uint32_t)) * n_index);
}
}
HistogramCuts::HistogramCuts() {
cut_ptrs_.HostVector().emplace_back(0);
}
void GHistIndexMatrix::Init(DMatrix* p_fmat, int 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();
hit_count.resize(nbins, 0);
hit_count_tloc_.resize(nthread * nbins, 0);
this->p_fmat = p_fmat;
size_t new_size = 1;
for (const auto &batch : p_fmat->GetBatches<SparsePage>()) {
new_size += batch.Size();
}
row_ptr.resize(new_size);
row_ptr[0] = 0;
size_t rbegin = 0;
size_t prev_sum = 0;
const bool isDense = p_fmat->IsDense();
this->isDense_ = isDense;
for (const auto &batch : p_fmat->GetBatches<SparsePage>()) {
// The number of threads is pegged to the batch size. If the OMP
// block is parallelized on anything other than the batch/block size,
// it should be reassigned
const size_t batch_threads = std::max(
size_t(1),
std::min(batch.Size(), static_cast<size_t>(omp_get_max_threads())));
auto page = batch.GetView();
MemStackAllocator<size_t, 128> partial_sums(batch_threads);
size_t* p_part = partial_sums.Get();
size_t block_size = batch.Size() / batch_threads;
dmlc::OMPException exc;
#pragma omp parallel num_threads(batch_threads)
{
#pragma omp for
for (omp_ulong tid = 0; tid < batch_threads; ++tid) {
exc.Run([&]() {
size_t ibegin = block_size * tid;
size_t iend = (tid == (batch_threads-1) ? batch.Size() : (block_size * (tid+1)));
size_t sum = 0;
for (size_t i = ibegin; i < iend; ++i) {
sum += page[i].size();
row_ptr[rbegin + 1 + i] = sum;
}
});
}
#pragma omp single
{
exc.Run([&]() {
p_part[0] = prev_sum;
for (size_t i = 1; i < batch_threads; ++i) {
p_part[i] = p_part[i - 1] + row_ptr[rbegin + i*block_size];
}
});
}
#pragma omp for
for (omp_ulong tid = 0; tid < batch_threads; ++tid) {
exc.Run([&]() {
size_t ibegin = block_size * tid;
size_t iend = (tid == (batch_threads-1) ? batch.Size() : (block_size * (tid+1)));
for (size_t i = ibegin; i < iend; ++i) {
row_ptr[rbegin + 1 + i] += p_part[tid];
}
});
}
}
exc.Rethrow();
const size_t n_offsets = cut.Ptrs().size() - 1;
const size_t n_index = row_ptr[rbegin + batch.Size()];
ResizeIndex(n_index, isDense);
CHECK_GT(cut.Values().size(), 0U);
uint32_t* offsets = nullptr;
if (isDense) {
index.ResizeOffset(n_offsets);
offsets = index.Offset();
for (size_t i = 0; i < n_offsets; ++i) {
offsets[i] = cut.Ptrs()[i];
}
}
if (isDense) {
BinTypeSize curent_bin_size = index.GetBinTypeSize();
if (curent_bin_size == kUint8BinsTypeSize) {
common::Span<uint8_t> index_data_span = {index.data<uint8_t>(),
n_index};
SetIndexData(index_data_span, batch_threads, batch, rbegin, nbins,
[offsets](auto idx, auto j) {
return static_cast<uint8_t>(idx - offsets[j]);
});
} else if (curent_bin_size == kUint16BinsTypeSize) {
common::Span<uint16_t> index_data_span = {index.data<uint16_t>(),
n_index};
SetIndexData(index_data_span, batch_threads, batch, rbegin, nbins,
[offsets](auto idx, auto j) {
return static_cast<uint16_t>(idx - offsets[j]);
});
} else {
CHECK_EQ(curent_bin_size, kUint32BinsTypeSize);
common::Span<uint32_t> index_data_span = {index.data<uint32_t>(),
n_index};
SetIndexData(index_data_span, batch_threads, batch, rbegin, nbins,
[offsets](auto idx, auto j) {
return static_cast<uint32_t>(idx - offsets[j]);
});
}
/* For sparse DMatrix we have to store index of feature for each bin
in index field to chose right offset. So offset is nullptr and index is not reduced */
} else {
common::Span<uint32_t> index_data_span = {index.data<uint32_t>(), n_index};
SetIndexData(index_data_span, batch_threads, batch, rbegin, nbins,
[](auto idx, auto) { return idx; });
}
ParallelFor(bst_omp_uint(nbins), nthread, [&](bst_omp_uint idx) {
for (int32_t tid = 0; tid < nthread; ++tid) {
hit_count[idx] += hit_count_tloc_[tid * nbins + idx];
hit_count_tloc_[tid * nbins + idx] = 0; // reset for next batch
}
});
prev_sum = row_ptr[rbegin + batch.Size()];
rbegin += batch.Size();
}
}
/*!
* \brief fill a histogram by zeros in range [begin, end)
*/
@@ -289,9 +136,9 @@ constexpr size_t Prefetch::kNoPrefetchSize;
template<typename FPType, bool do_prefetch, typename BinIdxType, bool any_missing = true>
void BuildHistKernel(const std::vector<GradientPair>& gpair,
const RowSetCollection::Elem row_indices,
const GHistIndexMatrix& gmat,
GHistRow<FPType> hist) {
const RowSetCollection::Elem row_indices,
const GHistIndexMatrix& gmat,
GHistRow<FPType> hist) {
const size_t size = row_indices.Size();
const size_t* rid = row_indices.begin;
const float* pgh = reinterpret_cast<const float*>(gpair.data());
@@ -337,8 +184,8 @@ void BuildHistKernel(const std::vector<GradientPair>& gpair,
template<typename FPType, bool do_prefetch, bool any_missing>
void BuildHistDispatch(const std::vector<GradientPair>& gpair,
const RowSetCollection::Elem row_indices,
const GHistIndexMatrix& gmat, GHistRow<FPType> hist) {
const RowSetCollection::Elem row_indices,
const GHistIndexMatrix& gmat, GHistRow<FPType> hist) {
switch (gmat.index.GetBinTypeSize()) {
case kUint8BinsTypeSize:
BuildHistKernel<FPType, do_prefetch, uint8_t, any_missing>(gpair, row_indices,
@@ -382,26 +229,26 @@ void GHistBuilder<GradientSumT>::BuildHist(
BuildHistDispatch<GradientSumT, false, any_missing>(gpair, span2, gmat, hist);
}
}
template
void GHistBuilder<float>::BuildHist<true>(const std::vector<GradientPair>& gpair,
const RowSetCollection::Elem row_indices,
const GHistIndexMatrix& gmat,
GHistRow<float> hist);
template
void GHistBuilder<float>::BuildHist<false>(const std::vector<GradientPair>& gpair,
const RowSetCollection::Elem row_indices,
const GHistIndexMatrix& gmat,
GHistRow<float> hist);
template
void GHistBuilder<double>::BuildHist<true>(const std::vector<GradientPair>& gpair,
const RowSetCollection::Elem row_indices,
const GHistIndexMatrix& gmat,
GHistRow<double> hist);
template
void GHistBuilder<double>::BuildHist<false>(const std::vector<GradientPair>& gpair,
const RowSetCollection::Elem row_indices,
const GHistIndexMatrix& gmat,
GHistRow<double> hist);
template void
GHistBuilder<float>::BuildHist<true>(const std::vector<GradientPair> &gpair,
const RowSetCollection::Elem row_indices,
const GHistIndexMatrix &gmat,
GHistRow<float> hist);
template void
GHistBuilder<float>::BuildHist<false>(const std::vector<GradientPair> &gpair,
const RowSetCollection::Elem row_indices,
const GHistIndexMatrix &gmat,
GHistRow<float> hist);
template void
GHistBuilder<double>::BuildHist<true>(const std::vector<GradientPair> &gpair,
const RowSetCollection::Elem row_indices,
const GHistIndexMatrix &gmat,
GHistRow<double> hist);
template void
GHistBuilder<double>::BuildHist<false>(const std::vector<GradientPair> &gpair,
const RowSetCollection::Elem row_indices,
const GHistIndexMatrix &gmat,
GHistRow<double> hist);
template<typename GradientSumT>
void GHistBuilder<GradientSumT>::SubtractionTrick(GHistRowT self,

106
src/common/hist_util.h
View File

@@ -25,6 +25,8 @@
#include "../include/rabit/rabit.h"
namespace xgboost {
class GHistIndexMatrix;
namespace common {
/*!
* \brief A single row in global histogram index.
@@ -226,74 +228,6 @@ struct Index {
Func func_;
};
/*!
* \brief preprocessed global index matrix, in CSR format
*
* Transform floating values to integer index in histogram This is a global histogram
* index for CPU histogram. On GPU ellpack page is used.
*/
struct GHistIndexMatrix {
/*! \brief row pointer to rows by element position */
std::vector<size_t> row_ptr;
/*! \brief The index data */
Index index;
/*! \brief hit count of each index */
std::vector<size_t> hit_count;
/*! \brief The corresponding cuts */
HistogramCuts cut;
DMatrix* p_fmat;
size_t max_num_bins;
// Create a global histogram matrix, given cut
void Init(DMatrix* p_fmat, int max_num_bins);
// specific method for sparse data as no possibility to reduce allocated memory
template <typename BinIdxType, typename GetOffset>
void SetIndexData(common::Span<BinIdxType> index_data_span,
size_t batch_threads, const SparsePage &batch,
size_t rbegin, size_t nbins, GetOffset get_offset) {
const xgboost::Entry *data_ptr = batch.data.HostVector().data();
const std::vector<bst_row_t> &offset_vec = batch.offset.HostVector();
const size_t batch_size = batch.Size();
CHECK_LT(batch_size, offset_vec.size());
BinIdxType* index_data = index_data_span.data();
ParallelFor(omp_ulong(batch_size), batch_threads, [&](omp_ulong i) {
const int tid = omp_get_thread_num();
size_t ibegin = row_ptr[rbegin + i];
size_t iend = row_ptr[rbegin + i + 1];
const size_t size = offset_vec[i + 1] - offset_vec[i];
SparsePage::Inst inst = {data_ptr + offset_vec[i], size};
CHECK_EQ(ibegin + inst.size(), iend);
for (bst_uint j = 0; j < inst.size(); ++j) {
uint32_t idx = cut.SearchBin(inst[j]);
index_data[ibegin + j] = get_offset(idx, j);
++hit_count_tloc_[tid * nbins + idx];
}
});
}
void ResizeIndex(const size_t n_index,
const bool isDense);
inline void GetFeatureCounts(size_t* counts) const {
auto nfeature = cut.Ptrs().size() - 1;
for (unsigned fid = 0; fid < nfeature; ++fid) {
auto ibegin = cut.Ptrs()[fid];
auto iend = cut.Ptrs()[fid + 1];
for (auto i = ibegin; i < iend; ++i) {
counts[fid] += hit_count[i];
}
}
}
inline bool IsDense() const {
return isDense_;
}
private:
std::vector<size_t> hit_count_tloc_;
bool isDense_;
};
template <typename GradientIndex>
int32_t XGBOOST_HOST_DEV_INLINE BinarySearchBin(size_t begin, size_t end,
GradientIndex const &data,
@@ -647,6 +581,42 @@ class GHistBuilder {
/*! \brief number of all bins over all features */
uint32_t nbins_ { 0 };
};
/*!
* \brief A C-style array with in-stack allocation. As long as the array is smaller than
* MaxStackSize, it will be allocated inside the stack. Otherwise, it will be
* heap-allocated.
*/
template<typename T, size_t MaxStackSize>
class MemStackAllocator {
public:
explicit MemStackAllocator(size_t required_size): required_size_(required_size) {
}
T* Get() {
if (!ptr_) {
if (MaxStackSize >= required_size_) {
ptr_ = stack_mem_;
} else {
ptr_ = reinterpret_cast<T*>(malloc(required_size_ * sizeof(T)));
do_free_ = true;
}
}
return ptr_;
}
~MemStackAllocator() {
if (do_free_) free(ptr_);
}
private:
T* ptr_ = nullptr;
bool do_free_ = false;
size_t required_size_;
T stack_mem_[MaxStackSize];
};
} // namespace common
} // namespace xgboost
#endif // XGBOOST_COMMON_HIST_UTIL_H_