Optimize ‘hist’ for multi-core CPU (#4529)
* Initial performance optimizations for xgboost * remove includes * revert float->double * fix for CI * fix for CI * fix for CI * fix for CI * fix for CI * fix for CI * fix for CI * fix for CI * fix for CI * fix for CI * Check existence of _mm_prefetch and __builtin_prefetch * Fix lint * optimizations for CPU * appling comments in review * add some comments, code refactoring * fixing issues in CI * adding runtime checks * remove 1 extra check * remove extra checks in BuildHist * remove checks * add debug info * added debug info * revert changes * added comments * Apply suggestions from code review Co-Authored-By: Philip Hyunsu Cho <chohyu01@cs.washington.edu> * apply review comments * Remove unused function CreateNewNodes() * Add descriptive comment on node_idx variable in QuantileHistMaker::Builder::BuildHistsBatch()
This commit is contained in:
Egor Smirnov
committed by
Philip Hyunsu Cho
Philip Hyunsu Cho
parent
abffbe014e
commit
4d6590be3c
@@ -11,13 +11,50 @@
|
||||
#include <xgboost/generic_parameters.h>
|
||||
#include <limits>
|
||||
#include <vector>
|
||||
#include <algorithm>
|
||||
#include <utility>
|
||||
#include "row_set.h"
|
||||
#include "../tree/param.h"
|
||||
#include "./quantile.h"
|
||||
#include "./timer.h"
|
||||
#include "../include/rabit/rabit.h"
|
||||
#include "random.h"
|
||||
|
||||
namespace xgboost {
|
||||
|
||||
/*!
|
||||
* \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 {
|
||||
|
||||
/*
|
||||
@@ -114,7 +151,7 @@ struct HistCutMatrix {
|
||||
};
|
||||
|
||||
/*! \brief Builds the cut matrix on the GPU.
|
||||
*
|
||||
*
|
||||
* \return The row stride across the entire dataset.
|
||||
*/
|
||||
size_t DeviceSketch
|
||||
@@ -134,9 +171,10 @@ using GHistIndexRow = Span<uint32_t const>;
|
||||
*/
|
||||
struct GHistIndexMatrix {
|
||||
/*! \brief row pointer to rows by element position */
|
||||
std::vector<size_t> row_ptr;
|
||||
// std::vector<size_t> row_ptr;
|
||||
SimpleArray<size_t> row_ptr;
|
||||
/*! \brief The index data */
|
||||
std::vector<uint32_t> index;
|
||||
SimpleArray<uint32_t> index;
|
||||
/*! \brief hit count of each index */
|
||||
std::vector<size_t> hit_count;
|
||||
/*! \brief The corresponding cuts */
|
||||
@@ -170,6 +208,11 @@ struct GHistIndexBlock {
|
||||
|
||||
inline GHistIndexBlock(const size_t* row_ptr, const uint32_t* index)
|
||||
: row_ptr(row_ptr), index(index) {}
|
||||
|
||||
// get i-th row
|
||||
inline GHistIndexRow operator[](size_t i) const {
|
||||
return {&index[0] + row_ptr[i], detail::ptrdiff_t(row_ptr[i + 1] - row_ptr[i])};
|
||||
}
|
||||
};
|
||||
|
||||
class ColumnMatrix;
|
||||
@@ -202,12 +245,63 @@ class GHistIndexBlockMatrix {
|
||||
};
|
||||
|
||||
/*!
|
||||
* \brief histogram of graident statistics for a single node.
|
||||
* Consists of multiple GradStats, each entry showing total graident statistics
|
||||
* for that particular bin
|
||||
* Uses global bin id so as to represent all features simultaneously
|
||||
* \brief used instead of GradStats to have float instead of double to reduce histograms
|
||||
* this improves performance by 10-30% and memory consumption for histograms by 2x
|
||||
* accuracy in both cases is the same
|
||||
*/
|
||||
using GHistRow = Span<tree::GradStats>;
|
||||
struct GradStatHist {
|
||||
typedef float GradType;
|
||||
/*! \brief sum gradient statistics */
|
||||
GradType sum_grad;
|
||||
/*! \brief sum hessian statistics */
|
||||
GradType sum_hess;
|
||||
|
||||
GradStatHist() : sum_grad{0}, sum_hess{0} {
|
||||
static_assert(sizeof(GradStatHist) == 8,
|
||||
"Size of GradStatHist is not 8 bytes.");
|
||||
}
|
||||
|
||||
inline void Add(const GradStatHist& b) {
|
||||
sum_grad += b.sum_grad;
|
||||
sum_hess += b.sum_hess;
|
||||
}
|
||||
|
||||
inline void Add(const tree::GradStats& b) {
|
||||
sum_grad += b.sum_grad;
|
||||
sum_hess += b.sum_hess;
|
||||
}
|
||||
|
||||
inline void Add(const GradientPair& p) {
|
||||
this->Add(p.GetGrad(), p.GetHess());
|
||||
}
|
||||
|
||||
inline void Add(const GradType& grad, const GradType& hess) {
|
||||
sum_grad += grad;
|
||||
sum_hess += hess;
|
||||
}
|
||||
|
||||
inline tree::GradStats ToGradStat() const {
|
||||
return tree::GradStats(sum_grad, sum_hess);
|
||||
}
|
||||
|
||||
inline void SetSubstract(const GradStatHist& a, const GradStatHist& b) {
|
||||
sum_grad = a.sum_grad - b.sum_grad;
|
||||
sum_hess = a.sum_hess - b.sum_hess;
|
||||
}
|
||||
|
||||
inline void SetSubstract(const tree::GradStats& a, const GradStatHist& b) {
|
||||
sum_grad = a.sum_grad - b.sum_grad;
|
||||
sum_hess = a.sum_hess - b.sum_hess;
|
||||
}
|
||||
|
||||
inline GradType GetGrad() const { return sum_grad; }
|
||||
inline GradType GetHess() const { return sum_hess; }
|
||||
inline static void Reduce(GradStatHist& a, const GradStatHist& b) { // NOLINT(*)
|
||||
a.Add(b);
|
||||
}
|
||||
};
|
||||
|
||||
using GHistRow = Span<GradStatHist>;
|
||||
|
||||
/*!
|
||||
* \brief histogram of gradient statistics for multiple nodes
|
||||
@@ -215,49 +309,43 @@ using GHistRow = Span<tree::GradStats>;
|
||||
class HistCollection {
|
||||
public:
|
||||
// access histogram for i-th node
|
||||
GHistRow operator[](bst_uint nid) const {
|
||||
constexpr uint32_t kMax = std::numeric_limits<uint32_t>::max();
|
||||
CHECK_NE(row_ptr_[nid], kMax);
|
||||
tree::GradStats* ptr =
|
||||
const_cast<tree::GradStats*>(dmlc::BeginPtr(data_) + row_ptr_[nid]);
|
||||
return {ptr, nbins_};
|
||||
inline GHistRow operator[](bst_uint nid) {
|
||||
AddHistRow(nid);
|
||||
return { const_cast<GradStatHist*>(dmlc::BeginPtr(data_arr_[nid])), nbins_};
|
||||
}
|
||||
|
||||
// have we computed a histogram for i-th node?
|
||||
bool RowExists(bst_uint nid) const {
|
||||
const uint32_t k_max = std::numeric_limits<uint32_t>::max();
|
||||
return (nid < row_ptr_.size() && row_ptr_[nid] != k_max);
|
||||
inline bool RowExists(bst_uint nid) const {
|
||||
return nid < data_arr_.size();
|
||||
}
|
||||
|
||||
// initialize histogram collection
|
||||
void Init(uint32_t nbins) {
|
||||
nbins_ = nbins;
|
||||
row_ptr_.clear();
|
||||
data_.clear();
|
||||
inline void Init(uint32_t nbins) {
|
||||
if (nbins_ != nbins) {
|
||||
data_arr_.clear();
|
||||
nbins_ = nbins;
|
||||
}
|
||||
}
|
||||
|
||||
// create an empty histogram for i-th node
|
||||
void AddHistRow(bst_uint nid) {
|
||||
constexpr uint32_t kMax = std::numeric_limits<uint32_t>::max();
|
||||
if (nid >= row_ptr_.size()) {
|
||||
row_ptr_.resize(nid + 1, kMax);
|
||||
}
|
||||
CHECK_EQ(row_ptr_[nid], kMax);
|
||||
inline void AddHistRow(bst_uint nid) {
|
||||
if (data_arr_.size() <= nid) {
|
||||
size_t prev = data_arr_.size();
|
||||
data_arr_.resize(nid + 1);
|
||||
|
||||
row_ptr_[nid] = data_.size();
|
||||
data_.resize(data_.size() + nbins_);
|
||||
for (size_t i = prev; i < data_arr_.size(); ++i) {
|
||||
data_arr_[i].resize(nbins_);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
private:
|
||||
/*! \brief number of all bins over all features */
|
||||
uint32_t nbins_;
|
||||
|
||||
std::vector<tree::GradStats> data_;
|
||||
|
||||
/*! \brief row_ptr_[nid] locates bin for historgram of node nid */
|
||||
std::vector<size_t> row_ptr_;
|
||||
uint32_t nbins_ = 0;
|
||||
std::vector<std::vector<GradStatHist>> data_arr_;
|
||||
};
|
||||
|
||||
|
||||
/*!
|
||||
* \brief builder for histograms of gradient statistics
|
||||
*/
|
||||
@@ -267,21 +355,55 @@ class GHistBuilder {
|
||||
inline void Init(size_t nthread, uint32_t nbins) {
|
||||
nthread_ = nthread;
|
||||
nbins_ = nbins;
|
||||
thread_init_.resize(nthread_);
|
||||
}
|
||||
|
||||
// construct a histogram via histogram aggregation
|
||||
void BuildHist(const std::vector<GradientPair>& gpair,
|
||||
const RowSetCollection::Elem row_indices,
|
||||
const GHistIndexMatrix& gmat,
|
||||
GHistRow hist);
|
||||
// same, with feature grouping
|
||||
void BuildBlockHist(const std::vector<GradientPair>& gpair,
|
||||
const RowSetCollection::Elem row_indices,
|
||||
const GHistIndexBlockMatrix& gmatb,
|
||||
GHistRow hist);
|
||||
// construct a histogram via subtraction trick
|
||||
void SubtractionTrick(GHistRow self, GHistRow sibling, GHistRow parent);
|
||||
const RowSetCollection::Elem row_indices,
|
||||
const GHistIndexBlockMatrix& gmatb,
|
||||
GHistRow hist) {
|
||||
constexpr int kUnroll = 8; // loop unrolling factor
|
||||
const int32_t nblock = gmatb.GetNumBlock();
|
||||
const size_t nrows = row_indices.end - row_indices.begin;
|
||||
const size_t rest = nrows % kUnroll;
|
||||
|
||||
#pragma omp parallel for
|
||||
for (int32_t bid = 0; bid < nblock; ++bid) {
|
||||
auto gmat = gmatb[bid];
|
||||
|
||||
for (size_t i = 0; i < nrows - rest; i += kUnroll) {
|
||||
size_t rid[kUnroll];
|
||||
size_t ibegin[kUnroll];
|
||||
size_t iend[kUnroll];
|
||||
GradientPair stat[kUnroll];
|
||||
for (int k = 0; k < kUnroll; ++k) {
|
||||
rid[k] = row_indices.begin[i + k];
|
||||
}
|
||||
for (int k = 0; k < kUnroll; ++k) {
|
||||
ibegin[k] = gmat.row_ptr[rid[k]];
|
||||
iend[k] = gmat.row_ptr[rid[k] + 1];
|
||||
}
|
||||
for (int k = 0; k < kUnroll; ++k) {
|
||||
stat[k] = gpair[rid[k]];
|
||||
}
|
||||
for (int k = 0; k < kUnroll; ++k) {
|
||||
for (size_t j = ibegin[k]; j < iend[k]; ++j) {
|
||||
const uint32_t bin = gmat.index[j];
|
||||
hist[bin].Add(stat[k]);
|
||||
}
|
||||
}
|
||||
}
|
||||
for (size_t i = nrows - rest; i < nrows; ++i) {
|
||||
const size_t rid = row_indices.begin[i];
|
||||
const size_t ibegin = gmat.row_ptr[rid];
|
||||
const size_t iend = gmat.row_ptr[rid + 1];
|
||||
const GradientPair stat = gpair[rid];
|
||||
for (size_t j = ibegin; j < iend; ++j) {
|
||||
const uint32_t bin = gmat.index[j];
|
||||
hist[bin].Add(stat);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
uint32_t GetNumBins() {
|
||||
return nbins_;
|
||||
@@ -292,11 +414,19 @@ class GHistBuilder {
|
||||
size_t nthread_;
|
||||
/*! \brief number of all bins over all features */
|
||||
uint32_t nbins_;
|
||||
std::vector<size_t> thread_init_;
|
||||
std::vector<tree::GradStats> data_;
|
||||
};
|
||||
|
||||
|
||||
void BuildHistLocalDense(size_t istart, size_t iend, size_t nrows, const size_t* rid,
|
||||
const uint32_t* index, const GradientPair::ValueT* pgh, const size_t* row_ptr,
|
||||
GradStatHist::GradType* data_local_hist, GradStatHist* grad_stat);
|
||||
|
||||
void BuildHistLocalSparse(size_t istart, size_t iend, size_t nrows, const size_t* rid,
|
||||
const uint32_t* index, const GradientPair::ValueT* pgh, const size_t* row_ptr,
|
||||
GradStatHist::GradType* data_local_hist, GradStatHist* grad_stat);
|
||||
|
||||
void SubtractionTrick(GHistRow self, GHistRow sibling, GHistRow parent);
|
||||
|
||||
} // namespace common
|
||||
} // namespace xgboost
|
||||
#endif // XGBOOST_COMMON_HIST_UTIL_H_
|
||||
|
||||
Reference in New Issue
Block a user