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

Improve OpenMP exception handling (#6680)

Browse Source
This commit is contained in:
Louis Desreumaux
2021-02-25 06:56:16 +01:00
committed by GitHub
parent c375173dca
commit 9b530e5697
26 changed files with 610 additions and 475 deletions
Download Patch File Download Diff File Expand all files Collapse all files

63
src/linear/coordinate_common.h
View File

@@ -14,6 +14,7 @@
#include "./param.h"
#include "../gbm/gblinear_model.h"
#include "../common/random.h"
#include "../common/threading_utils.h"
namespace xgboost {
namespace linear {
@@ -115,14 +116,18 @@ inline std::pair<double, double> GetGradientParallel(int group_idx, int num_grou
auto page = batch.GetView();
auto col = page[fidx];
const auto ndata = static_cast<bst_omp_uint>(col.size());
dmlc::OMPException exc;
#pragma omp parallel for schedule(static) reduction(+ : sum_grad, sum_hess)
for (bst_omp_uint j = 0; j < ndata; ++j) {
const bst_float v = col[j].fvalue;
auto &p = gpair[col[j].index * num_group + group_idx];
if (p.GetHess() < 0.0f) continue;
sum_grad += p.GetGrad() * v;
sum_hess += p.GetHess() * v * v;
exc.Run([&]() {
const bst_float v = col[j].fvalue;
auto &p = gpair[col[j].index * num_group + group_idx];
if (p.GetHess() < 0.0f) return;
sum_grad += p.GetGrad() * v;
sum_hess += p.GetHess() * v * v;
});
}
exc.Rethrow();
}
return std::make_pair(sum_grad, sum_hess);
}
@@ -142,14 +147,18 @@ inline std::pair<double, double> GetBiasGradientParallel(int group_idx, int num_
DMatrix *p_fmat) {
double sum_grad = 0.0, sum_hess = 0.0;
const auto ndata = static_cast<bst_omp_uint>(p_fmat->Info().num_row_);
dmlc::OMPException exc;
#pragma omp parallel for schedule(static) reduction(+ : sum_grad, sum_hess)
for (bst_omp_uint i = 0; i < ndata; ++i) {
auto &p = gpair[i * num_group + group_idx];
if (p.GetHess() >= 0.0f) {
sum_grad += p.GetGrad();
sum_hess += p.GetHess();
}
exc.Run([&]() {
auto &p = gpair[i * num_group + group_idx];
if (p.GetHess() >= 0.0f) {
sum_grad += p.GetGrad();
sum_hess += p.GetHess();
}
});
}
exc.Rethrow();
return std::make_pair(sum_grad, sum_hess);
}
@@ -172,12 +181,16 @@ inline void UpdateResidualParallel(int fidx, int group_idx, int num_group,
auto col = page[fidx];
// update grad value
const auto num_row = static_cast<bst_omp_uint>(col.size());
dmlc::OMPException exc;
#pragma omp parallel for schedule(static)
for (bst_omp_uint j = 0; j < num_row; ++j) {
GradientPair &p = (*in_gpair)[col[j].index * num_group + group_idx];
if (p.GetHess() < 0.0f) continue;
p += GradientPair(p.GetHess() * col[j].fvalue * dw, 0);
exc.Run([&]() {
GradientPair &p = (*in_gpair)[col[j].index * num_group + group_idx];
if (p.GetHess() < 0.0f) return;
p += GradientPair(p.GetHess() * col[j].fvalue * dw, 0);
});
}
exc.Rethrow();
}
}
@@ -195,12 +208,16 @@ inline void UpdateBiasResidualParallel(int group_idx, int num_group, float dbias
DMatrix *p_fmat) {
if (dbias == 0.0f) return;
const auto ndata = static_cast<bst_omp_uint>(p_fmat->Info().num_row_);
dmlc::OMPException exc;
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < ndata; ++i) {
GradientPair &g = (*in_gpair)[i * num_group + group_idx];
if (g.GetHess() < 0.0f) continue;
g += GradientPair(g.GetHess() * dbias, 0);
exc.Run([&]() {
GradientPair &g = (*in_gpair)[i * num_group + group_idx];
if (g.GetHess() < 0.0f) return;
g += GradientPair(g.GetHess() * dbias, 0);
});
}
exc.Rethrow();
}
/**
@@ -336,10 +353,9 @@ class GreedyFeatureSelector : public FeatureSelector {
const bst_omp_uint nfeat = model.learner_model_param->num_feature;
// Calculate univariate gradient sums
std::fill(gpair_sums_.begin(), gpair_sums_.end(), std::make_pair(0., 0.));
for (const auto &batch : p_fmat->GetBatches<CSCPage>()) {
auto page = batch.GetView();
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < nfeat; ++i) {
for (const auto &batch : p_fmat->GetBatches<CSCPage>()) {
auto page = batch.GetView();
common::ParallelFor(nfeat, [&](bst_omp_uint i) {
const auto col = page[i];
const bst_uint ndata = col.size();
auto &sums = gpair_sums_[group_idx * nfeat + i];
@@ -350,7 +366,7 @@ class GreedyFeatureSelector : public FeatureSelector {
sums.first += p.GetGrad() * v;
sums.second += p.GetHess() * v * v;
}
}
});
}
// Find a feature with the largest magnitude of weight change
int best_fidx = 0;
@@ -405,8 +421,7 @@ class ThriftyFeatureSelector : public FeatureSelector {
for (const auto &batch : p_fmat->GetBatches<CSCPage>()) {
auto page = batch.GetView();
// column-parallel is usually fastaer than row-parallel
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < nfeat; ++i) {
common::ParallelFor(nfeat, [&](bst_omp_uint i) {
const auto col = page[i];
const bst_uint ndata = col.size();
for (bst_uint gid = 0u; gid < ngroup; ++gid) {
@@ -419,7 +434,7 @@ class ThriftyFeatureSelector : public FeatureSelector {
sums.second += p.GetHess() * v * v;
}
}
}
});
}
// rank by descending weight magnitude within the groups
std::fill(deltaw_.begin(), deltaw_.end(), 0.f);