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

Use view for SparsePage exclusively. (#6590)

Browse Source
This commit is contained in:
Jiaming Yuan
2021-01-11 18:04:55 +08:00
committed by GitHub
parent 78f2cd83d7
commit f2f7dd87b8
23 changed files with 151 additions and 113 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
src/linear/coordinate_common.h
View File

@@ -82,7 +82,8 @@ inline std::pair<double, double> GetGradient(int group_idx, int num_group, int f
DMatrix *p_fmat) {
double sum_grad = 0.0, sum_hess = 0.0;
for (const auto &batch : p_fmat->GetBatches<CSCPage>()) {
auto col = batch[fidx];
auto page = batch.GetView();
auto col = page[fidx];
const auto ndata = static_cast<bst_omp_uint>(col.size());
for (bst_omp_uint j = 0; j < ndata; ++j) {
const bst_float v = col[j].fvalue;
@@ -111,7 +112,8 @@ inline std::pair<double, double> GetGradientParallel(int group_idx, int num_grou
DMatrix *p_fmat) {
double sum_grad = 0.0, sum_hess = 0.0;
for (const auto &batch : p_fmat->GetBatches<CSCPage>()) {
auto col = batch[fidx];
auto page = batch.GetView();
auto col = page[fidx];
const auto ndata = static_cast<bst_omp_uint>(col.size());
#pragma omp parallel for schedule(static) reduction(+ : sum_grad, sum_hess)
for (bst_omp_uint j = 0; j < ndata; ++j) {
@@ -166,7 +168,8 @@ inline void UpdateResidualParallel(int fidx, int group_idx, int num_group,
DMatrix *p_fmat) {
if (dw == 0.0f) return;
for (const auto &batch : p_fmat->GetBatches<CSCPage>()) {
auto col = batch[fidx];
auto page = batch.GetView();
auto col = page[fidx];
// update grad value
const auto num_row = static_cast<bst_omp_uint>(col.size());
#pragma omp parallel for schedule(static)
@@ -334,9 +337,10 @@ class GreedyFeatureSelector : public FeatureSelector {
// 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>()) {
#pragma omp parallel for schedule(static)
auto page = batch.GetView();
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < nfeat; ++i) {
const auto col = batch[i];
const auto col = page[i];
const bst_uint ndata = col.size();
auto &sums = gpair_sums_[group_idx * nfeat + i];
for (bst_uint j = 0u; j < ndata; ++j) {
@@ -399,10 +403,11 @@ class ThriftyFeatureSelector : public FeatureSelector {
// 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>()) {
// column-parallel is usually faster than row-parallel
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) {
const auto col = batch[i];
const auto col = page[i];
const bst_uint ndata = col.size();
for (bst_uint gid = 0u; gid < ngroup; ++gid) {
auto &sums = gpair_sums_[gid * nfeat + i];

5
src/linear/updater_gpu_coordinate.cu
View File

@@ -60,6 +60,7 @@ class GPUCoordinateUpdater : public LinearUpdater { // NOLINT
CHECK(p_fmat->SingleColBlock());
SparsePage const& batch = *(p_fmat->GetBatches<CSCPage>().begin());
auto page = batch.GetView();
if (IsEmpty()) {
return;
@@ -72,7 +73,7 @@ class GPUCoordinateUpdater : public LinearUpdater { // NOLINT
row_ptr_ = {0};
// iterate through columns
for (size_t fidx = 0; fidx < batch.Size(); fidx++) {
common::Span<Entry const> col = batch[fidx];
common::Span<Entry const> col = page[fidx];
auto cmp = [](Entry e1, Entry e2) {
return e1.index < e2.index;
};
@@ -89,7 +90,7 @@ class GPUCoordinateUpdater : public LinearUpdater { // NOLINT
data_.resize(row_ptr_.back());
gpair_.resize(num_row_ * model_param.num_output_group);
for (size_t fidx = 0; fidx < batch.Size(); fidx++) {
auto col = batch[fidx];
auto col = page[fidx];
auto seg = column_segments[fidx];
dh::safe_cuda(cudaMemcpy(
data_.data().get() + row_ptr_[fidx],

3
src/linear/updater_shotgun.cc
View File

@@ -52,6 +52,7 @@ class ShotgunUpdater : public LinearUpdater {
selector_->Setup(*model, in_gpair->ConstHostVector(), p_fmat,
param_.reg_alpha_denorm, param_.reg_lambda_denorm, 0);
for (const auto &batch : p_fmat->GetBatches<CSCPage>()) {
auto page = batch.GetView();
const auto nfeat = static_cast<bst_omp_uint>(batch.Size());
#pragma omp parallel for schedule(static)
for (bst_omp_uint i = 0; i < nfeat; ++i) {
@@ -60,7 +61,7 @@ class ShotgunUpdater : public LinearUpdater {
param_.reg_lambda_denorm);
if (ii < 0) continue;
const bst_uint fid = ii;
auto col = batch[ii];
auto col = page[ii];
for (int gid = 0; gid < ngroup; ++gid) {
double sum_grad = 0.0, sum_hess = 0.0;
for (auto& c : col) {