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

[SYCL] Add nodes initialisation (#10269)

Browse Source 
---------

Co-authored-by: Dmitry Razdoburdin <>
Co-authored-by: Jiaming Yuan <jm.yuan@outlook.com>
This commit is contained in:
Dmitry Razdoburdin
2024-05-21 17:38:52 +02:00
committed by GitHub
parent 7a54ca41c9
commit c7e7ce7569
7 changed files with 342 additions and 33 deletions
Download Patch File Download Diff File Expand all files Collapse all files

106
tests/cpp/plugin/test_sycl_hist_updater.cc
View File

@@ -16,11 +16,13 @@ namespace xgboost::sycl::tree {
template <typename GradientSumT>
class TestHistUpdater : public HistUpdater<GradientSumT> {
public:
TestHistUpdater(::sycl::queue qu,
TestHistUpdater(const Context* ctx,
::sycl::queue qu,
const xgboost::tree::TrainParam& param,
std::unique_ptr<TreeUpdater> pruner,
FeatureInteractionConstraintHost int_constraints_,
DMatrix const* fmat) : HistUpdater<GradientSumT>(qu, param, std::move(pruner),
DMatrix const* fmat) : HistUpdater<GradientSumT>(ctx, qu, param,
std::move(pruner),
int_constraints_, fmat) {}
void TestInitSampling(const USMVector<GradientPair, MemoryType::on_device> &gpair,
@@ -28,12 +30,11 @@ class TestHistUpdater : public HistUpdater<GradientSumT> {
HistUpdater<GradientSumT>::InitSampling(gpair, row_indices);
}
auto* TestInitData(Context const * ctx,
const common::GHistIndexMatrix& gmat,
const USMVector<GradientPair, MemoryType::on_device> &gpair,
const DMatrix& fmat,
const RegTree& tree) {
HistUpdater<GradientSumT>::InitData(ctx, gmat, gpair, fmat, tree);
auto* TestInitData(const common::GHistIndexMatrix& gmat,
const USMVector<GradientPair, MemoryType::on_device> &gpair,
const DMatrix& fmat,
const RegTree& tree) {
HistUpdater<GradientSumT>::InitData(gmat, gpair, fmat, tree);
return &(HistUpdater<GradientSumT>::row_set_collection_);
}
@@ -44,6 +45,15 @@ class TestHistUpdater : public HistUpdater<GradientSumT> {
HistUpdater<GradientSumT>::BuildHistogramsLossGuide(entry, gmat, p_tree, gpair);
return &(HistUpdater<GradientSumT>::hist_);
}
auto TestInitNewNode(int nid,
const common::GHistIndexMatrix& gmat,
const USMVector<GradientPair, MemoryType::on_device> &gpair,
const DMatrix& fmat,
const RegTree& tree) {
HistUpdater<GradientSumT>::InitNewNode(nid, gmat, gpair, fmat, tree);
return HistUpdater<GradientSumT>::snode_host_[nid];
}
};
void GenerateRandomGPairs(::sycl::queue* qu, GradientPair* gpair_ptr, size_t num_rows, bool has_neg_hess) {
@@ -79,7 +89,7 @@ void TestHistUpdaterSampling(const xgboost::tree::TrainParam& param) {
FeatureInteractionConstraintHost int_constraints;
std::unique_ptr<TreeUpdater> pruner{TreeUpdater::Create("prune", &ctx, &task)};
TestHistUpdater<GradientSumT> updater(qu, param, std::move(pruner), int_constraints, p_fmat.get());
TestHistUpdater<GradientSumT> updater(&ctx, qu, param, std::move(pruner), int_constraints, p_fmat.get());
USMVector<size_t, MemoryType::on_device> row_indices_0(&qu, num_rows);
USMVector<size_t, MemoryType::on_device> row_indices_1(&qu, num_rows);
@@ -135,7 +145,7 @@ void TestHistUpdaterInitData(const xgboost::tree::TrainParam& param, bool has_ne
FeatureInteractionConstraintHost int_constraints;
std::unique_ptr<TreeUpdater> pruner{TreeUpdater::Create("prune", &ctx, &task)};
TestHistUpdater<GradientSumT> updater(qu, param, std::move(pruner), int_constraints, p_fmat.get());
TestHistUpdater<GradientSumT> updater(&ctx, qu, param, std::move(pruner), int_constraints, p_fmat.get());
USMVector<GradientPair, MemoryType::on_device> gpair(&qu, num_rows);
GenerateRandomGPairs(&qu, gpair.Data(), num_rows, has_neg_hess);
@@ -146,7 +156,7 @@ void TestHistUpdaterInitData(const xgboost::tree::TrainParam& param, bool has_ne
gmat.Init(qu, &ctx, dmat, n_bins);
RegTree tree;
auto* row_set_collection = updater.TestInitData(&ctx, gmat, gpair, *p_fmat, tree);
auto* row_set_collection = updater.TestInitData(gmat, gpair, *p_fmat, tree);
auto& row_indices = row_set_collection->Data();
std::vector<size_t> row_indices_host(row_indices.Size());
@@ -191,7 +201,7 @@ void TestHistUpdaterBuildHistogramsLossGuide(const xgboost::tree::TrainParam& pa
FeatureInteractionConstraintHost int_constraints;
std::unique_ptr<TreeUpdater> pruner{TreeUpdater::Create("prune", &ctx, &task)};
TestHistUpdater<GradientSumT> updater(qu, param, std::move(pruner), int_constraints, p_fmat.get());
TestHistUpdater<GradientSumT> updater(&ctx, qu, param, std::move(pruner), int_constraints, p_fmat.get());
updater.SetHistSynchronizer(new BatchHistSynchronizer<GradientSumT>());
updater.SetHistRowsAdder(new BatchHistRowsAdder<GradientSumT>());
@@ -213,7 +223,7 @@ void TestHistUpdaterBuildHistogramsLossGuide(const xgboost::tree::TrainParam& pa
ExpandEntry node1(1, tree.GetDepth(1));
ExpandEntry node2(2, tree.GetDepth(2));
auto* row_set_collection = updater.TestInitData(&ctx, gmat, gpair, *p_fmat, tree);
auto* row_set_collection = updater.TestInitData(gmat, gpair, *p_fmat, tree);
row_set_collection->AddSplit(0, 1, 2, 42, num_rows - 42);
updater.TestBuildHistogramsLossGuide(node0, gmat, &tree, gpair);
@@ -237,6 +247,66 @@ void TestHistUpdaterBuildHistogramsLossGuide(const xgboost::tree::TrainParam& pa
}
}
template <typename GradientSumT>
void TestHistUpdaterInitNewNode(const xgboost::tree::TrainParam& param, float sparsity) {
const size_t num_rows = 1u << 8;
const size_t num_columns = 1;
const size_t n_bins = 32;
Context ctx;
ctx.UpdateAllowUnknown(Args{{"device", "sycl"}});
DeviceManager device_manager;
auto qu = device_manager.GetQueue(ctx.Device());
ObjInfo task{ObjInfo::kRegression};
auto p_fmat = RandomDataGenerator{num_rows, num_columns, sparsity}.GenerateDMatrix();
FeatureInteractionConstraintHost int_constraints;
std::unique_ptr<TreeUpdater> pruner{TreeUpdater::Create("prune", &ctx, &task)};
TestHistUpdater<GradientSumT> updater(&ctx, qu, param, std::move(pruner), int_constraints, p_fmat.get());
updater.SetHistSynchronizer(new BatchHistSynchronizer<GradientSumT>());
updater.SetHistRowsAdder(new BatchHistRowsAdder<GradientSumT>());
USMVector<GradientPair, MemoryType::on_device> gpair(&qu, num_rows);
auto* gpair_ptr = gpair.Data();
GenerateRandomGPairs(&qu, gpair_ptr, num_rows, false);
DeviceMatrix dmat;
dmat.Init(qu, p_fmat.get());
common::GHistIndexMatrix gmat;
gmat.Init(qu, &ctx, dmat, n_bins);
RegTree tree;
tree.ExpandNode(0, 0, 0, false, 0, 0, 0, 0, 0, 0, 0);
ExpandEntry node(ExpandEntry::kRootNid, tree.GetDepth(ExpandEntry::kRootNid));
auto* row_set_collection = updater.TestInitData(gmat, gpair, *p_fmat, tree);
auto& row_idxs = row_set_collection->Data();
const size_t* row_idxs_ptr = row_idxs.DataConst();
updater.TestBuildHistogramsLossGuide(node, gmat, &tree, gpair);
const auto snode = updater.TestInitNewNode(ExpandEntry::kRootNid, gmat, gpair, *p_fmat, tree);
GradStats<GradientSumT> grad_stat;
{
::sycl::buffer<GradStats<GradientSumT>> buff(&grad_stat, 1);
qu.submit([&](::sycl::handler& cgh) {
auto buff_acc = buff.template get_access<::sycl::access::mode::read_write>(cgh);
cgh.single_task<>([=]() {
for (size_t i = 0; i < num_rows; ++i) {
size_t row_idx = row_idxs_ptr[i];
buff_acc[0] += GradStats<GradientSumT>(gpair_ptr[row_idx].GetGrad(),
gpair_ptr[row_idx].GetHess());
}
});
}).wait_and_throw();
}
EXPECT_NEAR(snode.stats.GetGrad(), grad_stat.GetGrad(), 1e-6 * grad_stat.GetGrad());
EXPECT_NEAR(snode.stats.GetHess(), grad_stat.GetHess(), 1e-6 * grad_stat.GetHess());
}
TEST(SyclHistUpdater, Sampling) {
xgboost::tree::TrainParam param;
param.UpdateAllowUnknown(Args{{"subsample", "0.7"}});
@@ -266,4 +336,14 @@ TEST(SyclHistUpdater, BuildHistogramsLossGuide) {
TestHistUpdaterBuildHistogramsLossGuide<double>(param, 0.5);
}
TEST(SyclHistUpdater, InitNewNode) {
xgboost::tree::TrainParam param;
param.UpdateAllowUnknown(Args{{"max_depth", "3"}});
TestHistUpdaterInitNewNode<float>(param, 0.0);
TestHistUpdaterInitNewNode<float>(param, 0.5);
TestHistUpdaterInitNewNode<double>(param, 0.0);
TestHistUpdaterInitNewNode<double>(param, 0.5);
}
} // namespace xgboost::sycl::tree