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

Optimized BuildHist function (#5156)

Browse Source
This commit is contained in:
Egor Smirnov
2020-01-30 10:32:57 +03:00
committed by GitHub
parent 4240daed4e
commit c67163250e
8 changed files with 610 additions and 184 deletions
Download Patch File Download Diff File Expand all files Collapse all files

134
src/common/hist_util.cc
View File

@@ -659,13 +659,59 @@ void GHistIndexBlockMatrix::Init(const GHistIndexMatrix& gmat,
}
}
/*!
* \brief fill a histogram by zeroes
*/
void InitilizeHistByZeroes(GHistRow hist, size_t begin, size_t end) {
memset(hist.data() + begin, '\0', (end-begin)*sizeof(tree::GradStats));
}
/*!
* \brief Increment hist as dst += add in range [begin, end)
*/
void IncrementHist(GHistRow dst, const GHistRow add, size_t begin, size_t end) {
using FPType = decltype(tree::GradStats::sum_grad);
FPType* pdst = reinterpret_cast<FPType*>(dst.data());
const FPType* padd = reinterpret_cast<const FPType*>(add.data());
for (size_t i = 2 * begin; i < 2 * end; ++i) {
pdst[i] += padd[i];
}
}
/*!
* \brief Copy hist from src to dst in range [begin, end)
*/
void CopyHist(GHistRow dst, const GHistRow src, size_t begin, size_t end) {
using FPType = decltype(tree::GradStats::sum_grad);
FPType* pdst = reinterpret_cast<FPType*>(dst.data());
const FPType* psrc = reinterpret_cast<const FPType*>(src.data());
for (size_t i = 2 * begin; i < 2 * end; ++i) {
pdst[i] = psrc[i];
}
}
/*!
* \brief Compute Subtraction: dst = src1 - src2 in range [begin, end)
*/
void SubtractionHist(GHistRow dst, const GHistRow src1, const GHistRow src2,
size_t begin, size_t end) {
using FPType = decltype(tree::GradStats::sum_grad);
FPType* pdst = reinterpret_cast<FPType*>(dst.data());
const FPType* psrc1 = reinterpret_cast<const FPType*>(src1.data());
const FPType* psrc2 = reinterpret_cast<const FPType*>(src2.data());
for (size_t i = 2 * begin; i < 2 * end; ++i) {
pdst[i] = psrc1[i] - psrc2[i];
}
}
void GHistBuilder::BuildHist(const std::vector<GradientPair>& gpair,
const RowSetCollection::Elem row_indices,
const GHistIndexMatrix& gmat,
GHistRow hist) {
const size_t nthread = static_cast<size_t>(this->nthread_);
data_.resize(nbins_ * nthread_);
const size_t* rid = row_indices.begin;
const size_t nrows = row_indices.Size();
const uint32_t* index = gmat.index.data();
@@ -673,79 +719,27 @@ void GHistBuilder::BuildHist(const std::vector<GradientPair>& gpair,
const float* pgh = reinterpret_cast<const float*>(gpair.data());
double* hist_data = reinterpret_cast<double*>(hist.data());
double* data = reinterpret_cast<double*>(data_.data());
const size_t block_size = 512;
size_t n_blocks = nrows/block_size;
n_blocks += !!(nrows - n_blocks*block_size);
const size_t nthread_to_process = std::min(nthread, n_blocks);
memset(thread_init_.data(), '\0', nthread_to_process*sizeof(size_t));
const size_t cache_line_size = 64;
const size_t prefetch_offset = 10;
size_t no_prefetch_size = prefetch_offset + cache_line_size/sizeof(*rid);
no_prefetch_size = no_prefetch_size > nrows ? nrows : no_prefetch_size;
#pragma omp parallel for num_threads(nthread_to_process) schedule(guided)
for (bst_omp_uint iblock = 0; iblock < n_blocks; iblock++) {
dmlc::omp_uint tid = omp_get_thread_num();
double* data_local_hist = ((nthread_to_process == 1) ? hist_data :
reinterpret_cast<double*>(data_.data() + tid * nbins_));
for (size_t i = 0; i < nrows; ++i) {
const size_t icol_start = row_ptr[rid[i]];
const size_t icol_end = row_ptr[rid[i]+1];
if (!thread_init_[tid]) {
memset(data_local_hist, '\0', 2*nbins_*sizeof(double));
thread_init_[tid] = true;
if (i < nrows - no_prefetch_size) {
PREFETCH_READ_T0(row_ptr + rid[i + prefetch_offset]);
PREFETCH_READ_T0(pgh + 2*rid[i + prefetch_offset]);
}
const size_t istart = iblock*block_size;
const size_t iend = (((iblock+1)*block_size > nrows) ? nrows : istart + block_size);
for (size_t i = istart; i < iend; ++i) {
const size_t icol_start = row_ptr[rid[i]];
const size_t icol_end = row_ptr[rid[i]+1];
for (size_t j = icol_start; j < icol_end; ++j) {
const uint32_t idx_bin = 2*index[j];
const size_t idx_gh = 2*rid[i];
if (i < nrows - no_prefetch_size) {
PREFETCH_READ_T0(row_ptr + rid[i + prefetch_offset]);
PREFETCH_READ_T0(pgh + 2*rid[i + prefetch_offset]);
}
for (size_t j = icol_start; j < icol_end; ++j) {
const uint32_t idx_bin = 2*index[j];
const size_t idx_gh = 2*rid[i];
data_local_hist[idx_bin] += pgh[idx_gh];
data_local_hist[idx_bin+1] += pgh[idx_gh+1];
}
}
}
if (nthread_to_process > 1) {
const size_t size = (2*nbins_);
const size_t block_size = 1024;
size_t n_blocks = size/block_size;
n_blocks += !!(size - n_blocks*block_size);
size_t n_worked_bins = 0;
for (size_t i = 0; i < nthread_to_process; ++i) {
if (thread_init_[i]) {
thread_init_[n_worked_bins++] = i;
}
}
#pragma omp parallel for num_threads(std::min(nthread, n_blocks)) schedule(guided)
for (bst_omp_uint iblock = 0; iblock < n_blocks; iblock++) {
const size_t istart = iblock * block_size;
const size_t iend = (((iblock + 1) * block_size > size) ? size : istart + block_size);
const size_t bin = 2 * thread_init_[0] * nbins_;
memcpy(hist_data + istart, (data + bin + istart), sizeof(double) * (iend - istart));
for (size_t i_bin_part = 1; i_bin_part < n_worked_bins; ++i_bin_part) {
const size_t bin = 2 * thread_init_[i_bin_part] * nbins_;
for (size_t i = istart; i < iend; i++) {
hist_data[i] += data[bin + i];
}
}
hist_data[idx_bin] += pgh[idx_gh];
hist_data[idx_bin+1] += pgh[idx_gh+1];
}
}
}
@@ -801,10 +795,6 @@ void GHistBuilder::BuildBlockHist(const std::vector<GradientPair>& gpair,
}
void GHistBuilder::SubtractionTrick(GHistRow self, GHistRow sibling, GHistRow parent) {
tree::GradStats* p_self = self.data();
tree::GradStats* p_sibling = sibling.data();
tree::GradStats* p_parent = parent.data();
const size_t size = self.size();
CHECK_EQ(sibling.size(), size);
CHECK_EQ(parent.size(), size);
@@ -816,9 +806,7 @@ void GHistBuilder::SubtractionTrick(GHistRow self, GHistRow sibling, GHistRow pa
for (omp_ulong iblock = 0; iblock < n_blocks; ++iblock) {
const size_t ibegin = iblock*block_size;
const size_t iend = (((iblock+1)*block_size > size) ? size : ibegin + block_size);
for (bst_omp_uint bin_id = ibegin; bin_id < iend; bin_id++) {
p_self[bin_id].SetSubstract(p_parent[bin_id], p_sibling[bin_id]);
}
SubtractionHist(self, parent, sibling, ibegin, iend);
}
}

229
src/common/hist_util.h
View File

@@ -14,8 +14,10 @@
#include <algorithm>
#include <memory>
#include <utility>
#include <map>
#include "row_set.h"
#include "threading_utils.h"
#include "../tree/param.h"
#include "./quantile.h"
#include "./timer.h"
@@ -254,7 +256,7 @@ class DenseCuts : public CutsBuilder {
// FIXME(trivialfis): Merge this into generic cut builder.
/*! \brief Builds the cut matrix on the GPU.
*
*
* \return The row stride across the entire dataset.
*/
size_t DeviceSketch(int device,
@@ -343,13 +345,34 @@ class GHistIndexBlockMatrix {
};
/*!
* \brief histogram of graident statistics for a single node.
* Consists of multiple GradStats, each entry showing total graident statistics
* \brief histogram of gradient statistics for a single node.
* Consists of multiple GradStats, each entry showing total gradient statistics
* for that particular bin
* Uses global bin id so as to represent all features simultaneously
*/
using GHistRow = Span<tree::GradStats>;
/*!
* \brief fill a histogram by zeros
*/
void InitilizeHistByZeroes(GHistRow hist, size_t begin, size_t end);
/*!
* \brief Increment hist as dst += add in range [begin, end)
*/
void IncrementHist(GHistRow dst, const GHistRow add, size_t begin, size_t end);
/*!
* \brief Copy hist from src to dst in range [begin, end)
*/
void CopyHist(GHistRow dst, const GHistRow src, size_t begin, size_t end);
/*!
* \brief Compute Subtraction: dst = src1 - src2 in range [begin, end)
*/
void SubtractionHist(GHistRow dst, const GHistRow src1, const GHistRow src2,
size_t begin, size_t end);
/*!
* \brief histogram of gradient statistics for multiple nodes
*/
@@ -372,9 +395,13 @@ class HistCollection {
// initialize histogram collection
void Init(uint32_t nbins) {
nbins_ = nbins;
if (nbins_ != nbins) {
nbins_ = nbins;
// quite expensive operation, so let's do this only once
data_.clear();
}
row_ptr_.clear();
data_.clear();
n_nodes_added_ = 0;
}
// create an empty histogram for i-th node
@@ -385,20 +412,201 @@ class HistCollection {
}
CHECK_EQ(row_ptr_[nid], kMax);
row_ptr_[nid] = data_.size();
data_.resize(data_.size() + nbins_);
if (data_.size() < nbins_ * (nid + 1)) {
data_.resize(nbins_ * (nid + 1));
}
row_ptr_[nid] = nbins_ * n_nodes_added_;
n_nodes_added_++;
}
private:
/*! \brief number of all bins over all features */
uint32_t nbins_;
uint32_t nbins_ = 0;
/*! \brief amount of active nodes in hist collection */
uint32_t n_nodes_added_ = 0;
std::vector<tree::GradStats> data_;
/*! \brief row_ptr_[nid] locates bin for historgram of node nid */
/*! \brief row_ptr_[nid] locates bin for histogram of node nid */
std::vector<size_t> row_ptr_;
};
/*!
* \brief Stores temporary histograms to compute them in parallel
* Supports processing multiple tree-nodes for nested parallelism
* Able to reduce histograms across threads in efficient way
*/
class ParallelGHistBuilder {
public:
void Init(size_t nbins) {
if (nbins != nbins_) {
hist_buffer_.Init(nbins);
nbins_ = nbins;
}
}
// Add new elements if needed, mark all hists as unused
// targeted_hists - already allocated hists which should contain final results after Reduce() call
void Reset(size_t nthreads, size_t nodes, const BlockedSpace2d& space,
const std::vector<GHistRow>& targeted_hists) {
hist_buffer_.Init(nbins_);
tid_nid_to_hist_.clear();
hist_memory_.clear();
threads_to_nids_map_.clear();
targeted_hists_ = targeted_hists;
CHECK_EQ(nodes, targeted_hists.size());
nodes_ = nodes;
nthreads_ = nthreads;
MatchThreadsToNodes(space);
AllocateAdditionalHistograms();
MatchNodeNidPairToHist();
hist_was_used_.resize(nthreads * nodes_);
std::fill(hist_was_used_.begin(), hist_was_used_.end(), static_cast<int>(false));
}
// Get specified hist, initialize hist by zeros if it wasn't used before
GHistRow GetInitializedHist(size_t tid, size_t nid) {
CHECK_LT(nid, nodes_);
CHECK_LT(tid, nthreads_);
size_t idx = tid_nid_to_hist_.at({tid, nid});
GHistRow hist = hist_memory_[idx];
if (!hist_was_used_[tid * nodes_ + nid]) {
InitilizeHistByZeroes(hist, 0, hist.size());
hist_was_used_[tid * nodes_ + nid] = static_cast<int>(true);
}
return hist;
}
// Reduce following bins (begin, end] for nid-node in dst across threads
void ReduceHist(size_t nid, size_t begin, size_t end) {
CHECK_GT(end, begin);
CHECK_LT(nid, nodes_);
GHistRow dst = targeted_hists_[nid];
bool is_updated = false;
for (size_t tid = 0; tid < nthreads_; ++tid) {
if (hist_was_used_[tid * nodes_ + nid]) {
is_updated = true;
const size_t idx = tid_nid_to_hist_.at({tid, nid});
GHistRow src = hist_memory_[idx];
if (dst.data() != src.data()) {
IncrementHist(dst, src, begin, end);
}
}
}
if (!is_updated) {
// In distributed mode - some tree nodes can be empty on local machines,
// So we need just set local hist by zeros in this case
InitilizeHistByZeroes(dst, begin, end);
}
}
protected:
void MatchThreadsToNodes(const BlockedSpace2d& space) {
const size_t space_size = space.Size();
const size_t chunck_size = space_size / nthreads_ + !!(space_size % nthreads_);
threads_to_nids_map_.resize(nthreads_ * nodes_, false);
for (size_t tid = 0; tid < nthreads_; ++tid) {
size_t begin = chunck_size * tid;
size_t end = std::min(begin + chunck_size, space_size);
if (begin < space_size) {
size_t nid_begin = space.GetFirstDimension(begin);
size_t nid_end = space.GetFirstDimension(end-1);
for (size_t nid = nid_begin; nid <= nid_end; ++nid) {
// true - means thread 'tid' will work to compute partial hist for node 'nid'
threads_to_nids_map_[tid * nodes_ + nid] = true;
}
}
}
}
void AllocateAdditionalHistograms() {
size_t hist_allocated_additionally = 0;
for (size_t nid = 0; nid < nodes_; ++nid) {
int nthreads_for_nid = 0;
for (size_t tid = 0; tid < nthreads_; ++tid) {
if (threads_to_nids_map_[tid * nodes_ + nid]) {
nthreads_for_nid++;
}
}
// In distributed mode - some tree nodes can be empty on local machines,
// set nthreads_for_nid to 0 in this case.
// In another case - allocate additional (nthreads_for_nid - 1) histograms,
// because one is already allocated externally (will store final result for the node).
hist_allocated_additionally += std::max<int>(0, nthreads_for_nid - 1);
}
for (size_t i = 0; i < hist_allocated_additionally; ++i) {
hist_buffer_.AddHistRow(i);
}
}
void MatchNodeNidPairToHist() {
size_t hist_total = 0;
size_t hist_allocated_additionally = 0;
for (size_t nid = 0; nid < nodes_; ++nid) {
bool first_hist = true;
for (size_t tid = 0; tid < nthreads_; ++tid) {
if (threads_to_nids_map_[tid * nodes_ + nid]) {
if (first_hist) {
hist_memory_.push_back(targeted_hists_[nid]);
first_hist = false;
} else {
hist_memory_.push_back(hist_buffer_[hist_allocated_additionally]);
hist_allocated_additionally++;
}
// map pair {tid, nid} to index of allocated histogram from hist_memory_
tid_nid_to_hist_[{tid, nid}] = hist_total++;
CHECK_EQ(hist_total, hist_memory_.size());
}
}
}
}
/*! \brief number of bins in each histogram */
size_t nbins_ = 0;
/*! \brief number of threads for parallel computation */
size_t nthreads_ = 0;
/*! \brief number of nodes which will be processed in parallel */
size_t nodes_ = 0;
/*! \brief Buffer for additional histograms for Parallel processing */
HistCollection hist_buffer_;
/*!
* \brief Marks which hists were used, it means that they should be merged.
* Contains only {true or false} values
* but 'int' is used instead of 'bool', because std::vector<bool> isn't thread safe
*/
std::vector<int> hist_was_used_;
/*! \brief Buffer for additional histograms for Parallel processing */
std::vector<bool> threads_to_nids_map_;
/*! \brief Contains histograms for final results */
std::vector<GHistRow> targeted_hists_;
/*! \brief Allocated memory for histograms used for construction */
std::vector<GHistRow> hist_memory_;
/*! \brief map pair {tid, nid} to index of allocated histogram from hist_memory_ */
std::map<std::pair<size_t, size_t>, size_t> tid_nid_to_hist_;
};
/*!
* \brief builder for histograms of gradient statistics
*/
@@ -408,7 +616,6 @@ 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
@@ -433,8 +640,6 @@ 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_;
};

17
src/common/threading_utils.h
View File

@@ -108,12 +108,19 @@ class BlockedSpace2d {
// Wrapper to implement nested parallelism with simple omp parallel for
template<typename Func>
void ParallelFor2d(const BlockedSpace2d& space, Func func) {
const int num_blocks_in_space = static_cast<int>(space.Size());
void ParallelFor2d(const BlockedSpace2d& space, const int nthreads, Func func) {
const size_t num_blocks_in_space = space.Size();
#pragma omp parallel for
for (auto i = 0; i < num_blocks_in_space; i++) {
func(space.GetFirstDimension(i), space.GetRange(i));
#pragma omp parallel num_threads(nthreads)
{
size_t tid = omp_get_thread_num();
size_t chunck_size = num_blocks_in_space / nthreads + !!(num_blocks_in_space % nthreads);
size_t begin = chunck_size * tid;
size_t end = std::min(begin + chunck_size, num_blocks_in_space);
for (auto i = begin; i < end; i++) {
func(space.GetFirstDimension(i), space.GetRange(i));
}
}
}