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

Optimize GPU evaluation function for categorical data. (#7705)

Browse Source 
* Use transform and cache.
This commit is contained in:
Jiaming Yuan
2022-02-28 17:46:29 +08:00
committed by GitHub
parent 18a4af63aa
commit 1d468e20a4
3 changed files with 77 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
src/tree/gpu_hist/evaluate_splits.cuh
View File

@@ -51,6 +51,12 @@ class GPUHistEvaluator {
dh::CUDAStream copy_stream_;
// storage for sorted index of feature histogram, used for sort based splits.
dh::device_vector<bst_feature_t> cat_sorted_idx_;
// cached input for sorting the histogram, used for sort based splits.
using SortPair = thrust::tuple<uint32_t, double>;
dh::device_vector<SortPair> sort_input_;
// cache for feature index
dh::device_vector<bst_feature_t> feature_idx_;
// Training param used for evaluation
TrainParam param_;
// whether the input data requires sort based split, which is more complicated so we try
// to avoid it if possible.
@@ -95,6 +101,13 @@ class GPUHistEvaluator {
return dh::ToSpan(cat_sorted_idx_);
}
auto SortInput(EvaluateSplitInputs<GradientSumT> left) {
if (left.nidx == RegTree::kRoot && !cat_sorted_idx_.empty()) {
return dh::ToSpan(sort_input_).first(left.feature_values.size());
}
return dh::ToSpan(sort_input_);
}
public:
GPUHistEvaluator(TrainParam const &param, bst_feature_t n_features, int32_t device)
: tree_evaluator_{param, n_features, device}, param_{param} {}