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

add namespace aliases to reduce code

Browse Source
This commit is contained in:
Hui Liu
2023-10-27 09:11:55 -07:00
parent e00131c465
commit 4a4b528d54
19 changed files with 110 additions and 407 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
src/tree/fit_stump.cu
View File

@@ -21,6 +21,12 @@
#include "xgboost/logging.h" // CHECK_EQ
#include "xgboost/span.h" // span
#if defined(XGBOOST_USE_HIP)
namespace thrust {
namespace cuda = thrust::hip;
}
#endif
namespace xgboost::tree::cuda_impl {
void FitStump(Context const* ctx, MetaInfo const& info,
linalg::TensorView<GradientPair const, 2> gpair, linalg::VectorView<float> out) {
@@ -45,11 +51,7 @@ void FitStump(Context const* ctx, MetaInfo const& info,
dh::XGBCachingDeviceAllocator<char> alloc;
#if defined(XGBOOST_USE_CUDA)
auto policy = thrust::cuda::par(alloc);
#elif defined(XGBOOST_USE_HIP)
auto policy = thrust::hip::par(alloc);
#endif
thrust::reduce_by_key(policy, key_it, key_it + gpair.Size(), grad_it,
thrust::make_discard_iterator(), dh::tbegin(d_sum.Values()));

76
src/tree/gpu_hist/evaluator.cu
View File

@@ -12,6 +12,12 @@
#include "evaluate_splits.cuh"
#include "xgboost/data.h"
#if defined(XGBOOST_USE_HIP)
namespace thrust {
namespace cuda = thrust::hip;
}
#endif
namespace xgboost::tree {
void GPUHistEvaluator::Reset(common::HistogramCuts const &cuts, common::Span<FeatureType const> ft,
bst_feature_t n_features, TrainParam const &param,
@@ -28,7 +34,6 @@ void GPUHistEvaluator::Reset(common::HistogramCuts const &cuts, common::Span<Fea
// This condition avoids sort-based split function calls if the users want
// onehot-encoding-based splits.
// For some reason, any_of adds 1.5 minutes to compilation time for CUDA 11.x.
#if defined(XGBOOST_USE_CUDA)
need_sort_histogram_ =
thrust::any_of(thrust::cuda::par(alloc), beg, end, [=] XGBOOST_DEVICE(size_t i) {
auto idx = i - 1;
@@ -39,18 +44,6 @@ void GPUHistEvaluator::Reset(common::HistogramCuts const &cuts, common::Span<Fea
}
return false;
});
#elif defined(XGBOOST_USE_HIP)
need_sort_histogram_ =
thrust::any_of(thrust::hip::par(alloc), beg, end, [=] XGBOOST_DEVICE(size_t i) {
auto idx = i - 1;
if (common::IsCat(ft, idx)) {
auto n_bins = ptrs[i] - ptrs[idx];
bool use_sort = !common::UseOneHot(n_bins, to_onehot);
return use_sort;
}
return false;
});
#endif
node_categorical_storage_size_ =
common::CatBitField::ComputeStorageSize(cuts.MaxCategory() + 1);
@@ -72,19 +65,11 @@ void GPUHistEvaluator::Reset(common::HistogramCuts const &cuts, common::Span<Fea
auto it = thrust::make_counting_iterator(0ul);
auto values = cuts.cut_values_.ConstDeviceSpan();
#if defined(XGBOOST_USE_CUDA)
thrust::transform(thrust::cuda::par(alloc), it, it + feature_idx_.size(), feature_idx_.begin(),
[=] XGBOOST_DEVICE(size_t i) {
auto fidx = dh::SegmentId(ptrs, i);
return fidx;
});
#elif defined(XGBOOST_USE_HIP)
thrust::transform(thrust::hip::par(alloc), it, it + feature_idx_.size(), feature_idx_.begin(),
[=] XGBOOST_DEVICE(size_t i) {
auto fidx = dh::SegmentId(ptrs, i);
return fidx;
});
#endif
}
is_column_split_ = is_column_split;
device_ = device;
@@ -101,7 +86,6 @@ common::Span<bst_feature_t const> GPUHistEvaluator::SortHistogram(
auto d_feature_idx = dh::ToSpan(feature_idx_);
auto total_bins = shared_inputs.feature_values.size();
#if defined(XGBOOST_USE_CUDA)
thrust::transform(thrust::cuda::par(alloc), it, it + data.size(), dh::tbegin(data),
[=] XGBOOST_DEVICE(uint32_t i) {
auto const &input = d_inputs[i / total_bins];
@@ -115,27 +99,11 @@ common::Span<bst_feature_t const> GPUHistEvaluator::SortHistogram(
}
return thrust::make_tuple(i, 0.0f);
});
#elif defined(XGBOOST_USE_HIP)
thrust::transform(thrust::hip::par(alloc), it, it + data.size(), dh::tbegin(data),
[=] XGBOOST_DEVICE(uint32_t i) {
auto const &input = d_inputs[i / total_bins];
auto j = i % total_bins;
auto fidx = d_feature_idx[j];
if (common::IsCat(shared_inputs.feature_types, fidx)) {
auto grad =
shared_inputs.rounding.ToFloatingPoint(input.gradient_histogram[j]);
auto lw = evaluator.CalcWeightCat(shared_inputs.param, grad);
return thrust::make_tuple(i, lw);
}
return thrust::make_tuple(i, 0.0f);
});
#endif
// Sort an array segmented according to
// - nodes
// - features within each node
// - gradients within each feature
#if defined(XGBOOST_USE_CUDA)
thrust::stable_sort_by_key(thrust::cuda::par(alloc), dh::tbegin(data), dh::tend(data),
dh::tbegin(sorted_idx),
[=] XGBOOST_DEVICE(SortPair const &l, SortPair const &r) {
@@ -166,38 +134,6 @@ common::Span<bst_feature_t const> GPUHistEvaluator::SortHistogram(
}
return li < ri;
});
#elif defined(XGBOOST_USE_HIP)
thrust::stable_sort_by_key(thrust::hip::par(alloc), dh::tbegin(data), dh::tend(data),
dh::tbegin(sorted_idx),
[=] XGBOOST_DEVICE(SortPair const &l, SortPair const &r) {
auto li = thrust::get<0>(l);
auto ri = thrust::get<0>(r);
auto l_node = li / total_bins;
auto r_node = ri / total_bins;
if (l_node != r_node) {
return l_node < r_node; // not the same node
}
li = li % total_bins;
ri = ri % total_bins;
auto lfidx = d_feature_idx[li];
auto rfidx = d_feature_idx[ri];
if (lfidx != rfidx) {
return lfidx < rfidx; // not the same feature
}
if (common::IsCat(shared_inputs.feature_types, lfidx)) {
auto lw = thrust::get<1>(l);
auto rw = thrust::get<1>(r);
return lw < rw;
}
return li < ri;
});
#endif
return dh::ToSpan(cat_sorted_idx_);
}
} // namespace xgboost::tree

11
src/tree/gpu_hist/histogram.cu
View File

@@ -16,6 +16,12 @@
#include "row_partitioner.cuh"
#include "xgboost/base.h"
#if defined(XGBOOST_USE_HIP)
namespace thrust {
namespace cuda = thrust::hip;
}
#endif
namespace xgboost {
namespace tree {
namespace {
@@ -60,13 +66,8 @@ GradientQuantiser::GradientQuantiser(common::Span<GradientPair const> gpair, Met
thrust::device_ptr<GradientPair const> gpair_beg{gpair.data()};
auto beg = thrust::make_transform_iterator(gpair_beg, Clip());
#if defined(XGBOOST_USE_CUDA)
Pair p =
dh::Reduce(thrust::cuda::par(alloc), beg, beg + gpair.size(), Pair{}, thrust::plus<Pair>{});
#elif defined(XGBOOST_USE_HIP)
Pair p =
dh::Reduce(thrust::hip::par(alloc), beg, beg + gpair.size(), Pair{}, thrust::plus<Pair>{});
#endif
// Treat pair as array of 4 primitive types to allreduce
using ReduceT = typename decltype(p.first)::ValueT;