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

merge changes Jun092023

Browse Source
This commit is contained in:
amdsc21
2023-06-09 22:41:33 +02:00
parent 35cde3b1b2 152e2fb072
commit 5f78360949
63 changed files with 1299 additions and 243 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
src/tree/updater_gpu_hist.cu
View File

@@ -302,7 +302,7 @@ struct GPUHistMakerDevice {
matrix.feature_segments,
matrix.gidx_fvalue_map,
matrix.min_fvalue,
matrix.is_dense
matrix.is_dense && !collective::IsDistributed()
};
auto split = this->evaluator_.EvaluateSingleSplit(inputs, shared_inputs);
return split;
@@ -316,11 +316,11 @@ struct GPUHistMakerDevice {
std::vector<bst_node_t> nidx(2 * candidates.size());
auto h_node_inputs = pinned2.GetSpan<EvaluateSplitInputs>(2 * candidates.size());
auto matrix = page->GetDeviceAccessor(ctx_->gpu_id);
EvaluateSplitSharedInputs shared_inputs{
GPUTrainingParam{param}, *quantiser, feature_types, matrix.feature_segments,
matrix.gidx_fvalue_map, matrix.min_fvalue,
matrix.is_dense
};
EvaluateSplitSharedInputs shared_inputs{GPUTrainingParam{param}, *quantiser, feature_types,
matrix.feature_segments, matrix.gidx_fvalue_map,
matrix.min_fvalue,
// is_dense represents the local data
matrix.is_dense && !collective::IsDistributed()};
dh::TemporaryArray<GPUExpandEntry> entries(2 * candidates.size());
// Store the feature set ptrs so they dont go out of scope before the kernel is called
std::vector<std::shared_ptr<HostDeviceVector<bst_feature_t>>> feature_sets;

7
src/tree/updater_quantile_hist.cc
View File

@@ -419,6 +419,7 @@ class HistBuilder {
CPUExpandEntry InitRoot(DMatrix *p_fmat, linalg::MatrixView<GradientPair const> gpair,
RegTree *p_tree) {
monitor_->Start(__func__);
CPUExpandEntry node(RegTree::kRoot, p_tree->GetDepth(0));
std::size_t page_id = 0;
@@ -434,7 +435,7 @@ class HistBuilder {
{
GradientPairPrecise grad_stat;
if (p_fmat->IsDense()) {
if (p_fmat->IsDense() && !collective::IsDistributed()) {
/**
* Specialized code for dense data: For dense data (with no missing value), the sum
* of gradient histogram is equal to snode[nid]
@@ -475,12 +476,14 @@ class HistBuilder {
node = entries.front();
}
monitor_->Stop(__func__);
return node;
}
void BuildHistogram(DMatrix *p_fmat, RegTree *p_tree,
std::vector<CPUExpandEntry> const &valid_candidates,
linalg::MatrixView<GradientPair const> gpair) {
monitor_->Start(__func__);
std::vector<CPUExpandEntry> nodes_to_build(valid_candidates.size());
std::vector<CPUExpandEntry> nodes_to_sub(valid_candidates.size());
@@ -508,6 +511,7 @@ class HistBuilder {
nodes_to_sub, gpair.Values());
++page_id;
}
monitor_->Stop(__func__);
}
void UpdatePosition(DMatrix *p_fmat, RegTree const *p_tree,
@@ -525,6 +529,7 @@ class HistBuilder {
std::vector<bst_node_t> *p_out_position) {
monitor_->Start(__func__);
if (!task_->UpdateTreeLeaf()) {
monitor_->Stop(__func__);
return;
}
for (auto const &part : partitioner_) {