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

Optimizations for RNG in InitData kernel (#5522)

Browse Source 
* optimizations for subsampling in InitData

* optimizations for subsampling in InitData

Co-authored-by: SHVETS, KIRILL <kirill.shvets@intel.com>
This commit is contained in:
ShvetsKS
2020-04-16 18:24:32 +03:00
committed by GitHub
parent e268fb0093
commit a2d86b8e4b
3 changed files with 110 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

69
src/tree/updater_quantile_hist.cc
View File

@@ -536,6 +536,63 @@ bool QuantileHistMaker::Builder::UpdatePredictionCache(
return true;
}
void QuantileHistMaker::Builder::InitSampling(const std::vector<GradientPair>& gpair,
const DMatrix& fmat,
std::vector<size_t>* row_indices) {
const auto& info = fmat.Info();
auto& rnd = common::GlobalRandom();
std::vector<size_t>& row_indices_local = *row_indices;
size_t* p_row_indices = row_indices_local.data();
#if XGBOOST_CUSTOMIZE_GLOBAL_PRNG
std::bernoulli_distribution coin_flip(param_.subsample);
size_t j = 0;
for (size_t i = 0; i < info.num_row_; ++i) {
if (gpair[i].GetHess() >= 0.0f && coin_flip(rnd)) {
p_row_indices[j++] = i;
}
}
/* resize row_indices to reduce memory */
row_indices_local.resize(j);
#else
const size_t nthread = this->nthread_;
std::vector<size_t> row_offsets(nthread, 0);
/* usage of mt19937_64 give 2x speed up for subsampling */
std::vector<std::mt19937> rnds(nthread);
/* create engine for each thread */
for (std::mt19937& r : rnds) {
r = rnd;
}
const size_t discard_size = info.num_row_ / nthread;
#pragma omp parallel num_threads(nthread)
{
const size_t tid = omp_get_thread_num();
const size_t ibegin = tid * discard_size;
const size_t iend = (tid == (nthread - 1)) ?
info.num_row_ : ibegin + discard_size;
std::bernoulli_distribution coin_flip(param_.subsample);
rnds[tid].discard(2*discard_size * tid);
for (size_t i = ibegin; i < iend; ++i) {
if (gpair[i].GetHess() >= 0.0f && coin_flip(rnds[tid])) {
p_row_indices[ibegin + row_offsets[tid]++] = i;
}
}
}
/* discard global engine */
rnd = rnds[nthread - 1];
size_t prefix_sum = row_offsets[0];
for (size_t i = 1; i < nthread; ++i) {
const size_t ibegin = i * discard_size;
for (size_t k = 0; k < row_offsets[i]; ++k) {
row_indices_local[prefix_sum + k] = row_indices_local[ibegin + k];
}
prefix_sum += row_offsets[i];
}
/* resize row_indices to reduce memory */
row_indices_local.resize(prefix_sum);
#endif // XGBOOST_CUSTOMIZE_GLOBAL_PRNG
}
void QuantileHistMaker::Builder::InitData(const GHistIndexMatrix& gmat,
const std::vector<GradientPair>& gpair,
const DMatrix& fmat,
@@ -569,22 +626,14 @@ void QuantileHistMaker::Builder::InitData(const GHistIndexMatrix& gmat,
std::vector<size_t>& row_indices = *row_set_collection_.Data();
row_indices.resize(info.num_row_);
auto* p_row_indices = row_indices.data();
size_t* p_row_indices = row_indices.data();
// mark subsample and build list of member rows
if (param_.subsample < 1.0f) {
CHECK_EQ(param_.sampling_method, TrainParam::kUniform)
<< "Only uniform sampling is supported, "
<< "gradient-based sampling is only support by GPU Hist.";
std::bernoulli_distribution coin_flip(param_.subsample);
auto& rnd = common::GlobalRandom();
size_t j = 0;
for (size_t i = 0; i < info.num_row_; ++i) {
if (gpair[i].GetHess() >= 0.0f && coin_flip(rnd)) {
p_row_indices[j++] = i;
}
}
row_indices.resize(j);
InitSampling(gpair, fmat, &row_indices);
} else {
MemStackAllocator<bool, 128> buff(this->nthread_);
bool* p_buff = buff.Get();

3
src/tree/updater_quantile_hist.h
View File

@@ -202,6 +202,9 @@ class QuantileHistMaker: public TreeUpdater {
const DMatrix& fmat,
const RegTree& tree);
void InitSampling(const std::vector<GradientPair>& gpair,
const DMatrix& fmat, std::vector<size_t>* row_indices);
void EvaluateSplits(const std::vector<ExpandEntry>& nodes_set,
const GHistIndexMatrix& gmat,
const HistCollection& hist,