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

Support column split in GPU evaluate splits (#9511)

Browse Source
This commit is contained in:
Rong Ou
2023-08-23 01:33:43 -07:00
committed by GitHub
parent 8c10af45a0
commit 6103dca0bb
11 changed files with 240 additions and 113 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
src/tree/gpu_hist/evaluate_splits.cu
View File

@@ -5,8 +5,8 @@
#include <vector>
#include <limits>
#include "../../collective/communicator-inl.cuh"
#include "../../common/categorical.h"
#include "../../common/device_helpers.cuh"
#include "../../data/ellpack_page.cuh"
#include "evaluate_splits.cuh"
#include "expand_entry.cuh"
@@ -409,6 +409,23 @@ void GPUHistEvaluator::EvaluateSplits(
this->LaunchEvaluateSplits(max_active_features, d_inputs, shared_inputs,
evaluator, out_splits);
if (is_column_split_) {
// With column-wise data split, we gather the split candidates from all the workers and find the
// global best candidates.
auto const world_size = collective::GetWorldSize();
dh::TemporaryArray<DeviceSplitCandidate> all_candidate_storage(out_splits.size() * world_size);
auto all_candidates = dh::ToSpan(all_candidate_storage);
collective::AllGather(device_, out_splits.data(), all_candidates.data(),
out_splits.size() * sizeof(DeviceSplitCandidate));
// Reduce to get the best candidate from all workers.
dh::LaunchN(out_splits.size(), [world_size, all_candidates, out_splits] __device__(size_t i) {
for (auto rank = 0; rank < world_size; rank++) {
out_splits[i] = out_splits[i] + all_candidates[rank * out_splits.size() + i];
}
});
}
auto d_sorted_idx = this->SortedIdx(d_inputs.size(), shared_inputs.feature_values.size());
auto d_entries = out_entries;
auto device_cats_accessor = this->DeviceCatStorage(nidx);

6
src/tree/gpu_hist/evaluate_splits.cuh
View File

@@ -83,6 +83,9 @@ class GPUHistEvaluator {
// Number of elements of categorical storage type
// needed to hold categoricals for a single mode
std::size_t node_categorical_storage_size_ = 0;
// Is the data split column-wise?
bool is_column_split_ = false;
int32_t device_;
// Copy the categories from device to host asynchronously.
void CopyToHost( const std::vector<bst_node_t>& nidx);
@@ -136,7 +139,8 @@ class GPUHistEvaluator {
* \brief Reset the evaluator, should be called before any use.
*/
void Reset(common::HistogramCuts const &cuts, common::Span<FeatureType const> ft,
bst_feature_t n_features, TrainParam const &param, int32_t device);
bst_feature_t n_features, TrainParam const &param, bool is_column_split,
int32_t device);
/**
* \brief Get host category storage for nidx. Different from the internal version, this

9
src/tree/gpu_hist/evaluator.cu
View File

@@ -14,10 +14,9 @@
namespace xgboost {
namespace tree {
void GPUHistEvaluator::Reset(common::HistogramCuts const &cuts,
common::Span<FeatureType const> ft,
bst_feature_t n_features, TrainParam const &param,
int32_t device) {
void GPUHistEvaluator::Reset(common::HistogramCuts const &cuts, common::Span<FeatureType const> ft,
bst_feature_t n_features, TrainParam const &param,
bool is_column_split, int32_t device) {
param_ = param;
tree_evaluator_ = TreeEvaluator{param, n_features, device};
has_categoricals_ = cuts.HasCategorical();
@@ -65,6 +64,8 @@ void GPUHistEvaluator::Reset(common::HistogramCuts const &cuts,
return fidx;
});
}
is_column_split_ = is_column_split;
device_ = device;
}
common::Span<bst_feature_t const> GPUHistEvaluator::SortHistogram(

3
src/tree/updater_gpu_hist.cu
View File

@@ -242,7 +242,8 @@ struct GPUHistMakerDevice {
page = sample.page;
gpair = sample.gpair;
this->evaluator_.Reset(page->Cuts(), feature_types, dmat->Info().num_col_, param, ctx_->gpu_id);
this->evaluator_.Reset(page->Cuts(), feature_types, dmat->Info().num_col_, param,
dmat->Info().IsColumnSplit(), ctx_->gpu_id);
quantiser.reset(new GradientQuantiser(this->gpair));