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

Initial GPU support for the approx tree method. (#9414)

Browse Source
This commit is contained in:
Jiaming Yuan
2023-07-31 15:50:28 +08:00
committed by GitHub
parent 8f0efb4ab3
commit 912e341d57
23 changed files with 639 additions and 360 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
src/tree/constraints.h
View File

@@ -1,5 +1,5 @@
/*!
* Copyright 2018-2019 by Contributors
/**
* Copyright 2018-2023 by Contributors
*/
#ifndef XGBOOST_TREE_CONSTRAINTS_H_
#define XGBOOST_TREE_CONSTRAINTS_H_
@@ -8,10 +8,8 @@
#include <unordered_set>
#include <vector>
#include "xgboost/span.h"
#include "xgboost/base.h"
#include "param.h"
#include "xgboost/base.h"
namespace xgboost {
/*!

28
src/tree/gpu_hist/gradient_based_sampler.cu
View File

@@ -8,10 +8,10 @@
#include <xgboost/logging.h>
#include <algorithm>
#include <cstddef> // for size_t
#include <limits>
#include <utility>
#include "../../common/compressed_iterator.h"
#include "../../common/cuda_context.cuh" // for CUDAContext
#include "../../common/random.h"
#include "../param.h"
@@ -202,27 +202,27 @@ ExternalMemoryUniformSampling::ExternalMemoryUniformSampling(size_t n_rows,
GradientBasedSample ExternalMemoryUniformSampling::Sample(Context const* ctx,
common::Span<GradientPair> gpair,
DMatrix* dmat) {
auto cuctx = ctx->CUDACtx();
// Set gradient pair to 0 with p = 1 - subsample
thrust::replace_if(dh::tbegin(gpair), dh::tend(gpair),
thrust::counting_iterator<size_t>(0),
BernoulliTrial(common::GlobalRandom()(), subsample_),
GradientPair());
thrust::replace_if(cuctx->CTP(), dh::tbegin(gpair), dh::tend(gpair),
thrust::counting_iterator<std::size_t>(0),
BernoulliTrial(common::GlobalRandom()(), subsample_), GradientPair{});
// Count the sampled rows.
size_t sample_rows = thrust::count_if(dh::tbegin(gpair), dh::tend(gpair), IsNonZero());
size_t sample_rows =
thrust::count_if(cuctx->CTP(), dh::tbegin(gpair), dh::tend(gpair), IsNonZero{});
// Compact gradient pairs.
gpair_.resize(sample_rows);
thrust::copy_if(dh::tbegin(gpair), dh::tend(gpair), gpair_.begin(), IsNonZero());
thrust::copy_if(cuctx->CTP(), dh::tbegin(gpair), dh::tend(gpair), gpair_.begin(), IsNonZero{});
// Index the sample rows.
thrust::transform(dh::tbegin(gpair), dh::tend(gpair), sample_row_index_.begin(), IsNonZero());
thrust::exclusive_scan(sample_row_index_.begin(), sample_row_index_.end(),
thrust::transform(cuctx->CTP(), dh::tbegin(gpair), dh::tend(gpair), sample_row_index_.begin(),
IsNonZero());
thrust::exclusive_scan(cuctx->CTP(), sample_row_index_.begin(), sample_row_index_.end(),
sample_row_index_.begin());
thrust::transform(dh::tbegin(gpair), dh::tend(gpair),
sample_row_index_.begin(),
sample_row_index_.begin(),
ClearEmptyRows());
thrust::transform(cuctx->CTP(), dh::tbegin(gpair), dh::tend(gpair), sample_row_index_.begin(),
sample_row_index_.begin(), ClearEmptyRows());
auto batch_iterator = dmat->GetBatches<EllpackPage>(ctx, batch_param_);
auto first_page = (*batch_iterator.begin()).Impl();
@@ -232,7 +232,7 @@ GradientBasedSample ExternalMemoryUniformSampling::Sample(Context const* ctx,
first_page->row_stride, sample_rows));
// Compact the ELLPACK pages into the single sample page.
thrust::fill(dh::tbegin(page_->gidx_buffer), dh::tend(page_->gidx_buffer), 0);
thrust::fill(cuctx->CTP(), dh::tbegin(page_->gidx_buffer), dh::tend(page_->gidx_buffer), 0);
for (auto& batch : batch_iterator) {
page_->Compact(ctx->gpu_id, batch.Impl(), dh::ToSpan(sample_row_index_));
}

1
src/tree/updater_approx.cc
View File

@@ -11,7 +11,6 @@
#include "../common/random.h"
#include "../data/gradient_index.h"
#include "common_row_partitioner.h"
#include "constraints.h"
#include "driver.h"
#include "hist/evaluate_splits.h"
#include "hist/histogram.h"

231
src/tree/updater_gpu_hist.cu
View File

@@ -31,7 +31,6 @@
#include "gpu_hist/histogram.cuh"
#include "gpu_hist/row_partitioner.cuh"
#include "param.h"
#include "split_evaluator.h"
#include "updater_gpu_common.cuh"
#include "xgboost/base.h"
#include "xgboost/context.h"
@@ -49,13 +48,30 @@ DMLC_REGISTRY_FILE_TAG(updater_gpu_hist);
#endif // !defined(GTEST_TEST)
// training parameters specific to this algorithm
struct GPUHistMakerTrainParam
: public XGBoostParameter<GPUHistMakerTrainParam> {
struct GPUHistMakerTrainParam : public XGBoostParameter<GPUHistMakerTrainParam> {
bool debug_synchronize;
// declare parameters
DMLC_DECLARE_PARAMETER(GPUHistMakerTrainParam) {
DMLC_DECLARE_FIELD(debug_synchronize).set_default(false).describe(
"Check if all distributed tree are identical after tree construction.");
DMLC_DECLARE_FIELD(debug_synchronize)
.set_default(false)
.describe("Check if all distributed tree are identical after tree construction.");
}
// Only call this method for testing
void CheckTreesSynchronized(RegTree const* local_tree) const {
if (this->debug_synchronize) {
std::string s_model;
common::MemoryBufferStream fs(&s_model);
int rank = collective::GetRank();
if (rank == 0) {
local_tree->Save(&fs);
}
fs.Seek(0);
collective::Broadcast(&s_model, 0);
RegTree reference_tree{}; // rank 0 tree
reference_tree.Load(&fs);
CHECK(*local_tree == reference_tree);
}
}
};
#if !defined(GTEST_TEST)
@@ -170,16 +186,15 @@ class DeviceHistogramStorage {
};
// Manage memory for a single GPU
template <typename GradientSumT>
struct GPUHistMakerDevice {
private:
GPUHistEvaluator evaluator_;
Context const* ctx_;
std::shared_ptr<common::ColumnSampler> column_sampler_;
public:
EllpackPageImpl const* page{nullptr};
common::Span<FeatureType const> feature_types;
BatchParam batch_param;
std::unique_ptr<RowPartitioner> row_partitioner;
DeviceHistogramStorage<> hist{};
@@ -199,7 +214,6 @@ struct GPUHistMakerDevice {
dh::PinnedMemory pinned2;
common::Monitor monitor;
common::ColumnSampler column_sampler;
FeatureInteractionConstraintDevice interaction_constraints;
std::unique_ptr<GradientBasedSampler> sampler;
@@ -208,22 +222,22 @@ struct GPUHistMakerDevice {
GPUHistMakerDevice(Context const* ctx, bool is_external_memory,
common::Span<FeatureType const> _feature_types, bst_row_t _n_rows,
TrainParam _param, uint32_t column_sampler_seed, uint32_t n_features,
BatchParam _batch_param)
TrainParam _param, std::shared_ptr<common::ColumnSampler> column_sampler,
uint32_t n_features, BatchParam batch_param)
: evaluator_{_param, n_features, ctx->gpu_id},
ctx_(ctx),
feature_types{_feature_types},
param(std::move(_param)),
column_sampler(column_sampler_seed),
interaction_constraints(param, n_features),
batch_param(std::move(_batch_param)) {
sampler.reset(new GradientBasedSampler(ctx, _n_rows, batch_param, param.subsample,
param.sampling_method, is_external_memory));
column_sampler_(std::move(column_sampler)),
interaction_constraints(param, n_features) {
sampler = std::make_unique<GradientBasedSampler>(ctx, _n_rows, batch_param, param.subsample,
param.sampling_method, is_external_memory);
if (!param.monotone_constraints.empty()) {
// Copy assigning an empty vector causes an exception in MSVC debug builds
monotone_constraints = param.monotone_constraints;
}
CHECK(column_sampler_);
monitor.Init(std::string("GPUHistMakerDevice") + std::to_string(ctx_->gpu_id));
}
@@ -234,16 +248,16 @@ struct GPUHistMakerDevice {
CHECK(page);
feature_groups.reset(new FeatureGroups(page->Cuts(), page->is_dense,
dh::MaxSharedMemoryOptin(ctx_->gpu_id),
sizeof(GradientSumT)));
sizeof(GradientPairPrecise)));
}
}
// Reset values for each update iteration
void Reset(HostDeviceVector<GradientPair>* dh_gpair, DMatrix* dmat, int64_t num_columns) {
auto const& info = dmat->Info();
this->column_sampler.Init(ctx_, num_columns, info.feature_weights.HostVector(),
param.colsample_bynode, param.colsample_bylevel,
param.colsample_bytree);
this->column_sampler_->Init(ctx_, num_columns, info.feature_weights.HostVector(),
param.colsample_bynode, param.colsample_bylevel,
param.colsample_bytree);
dh::safe_cuda(cudaSetDevice(ctx_->gpu_id));
this->interaction_constraints.Reset();
@@ -275,8 +289,8 @@ struct GPUHistMakerDevice {
GPUExpandEntry EvaluateRootSplit(GradientPairInt64 root_sum) {
int nidx = RegTree::kRoot;
GPUTrainingParam gpu_param(param);
auto sampled_features = column_sampler.GetFeatureSet(0);
sampled_features->SetDevice(ctx_->gpu_id);
auto sampled_features = column_sampler_->GetFeatureSet(0);
sampled_features->SetDevice(ctx_->Device());
common::Span<bst_feature_t> feature_set =
interaction_constraints.Query(sampled_features->DeviceSpan(), nidx);
auto matrix = page->GetDeviceAccessor(ctx_->gpu_id);
@@ -316,13 +330,13 @@ struct GPUHistMakerDevice {
int right_nidx = tree[candidate.nid].RightChild();
nidx[i * 2] = left_nidx;
nidx[i * 2 + 1] = right_nidx;
auto left_sampled_features = column_sampler.GetFeatureSet(tree.GetDepth(left_nidx));
left_sampled_features->SetDevice(ctx_->gpu_id);
auto left_sampled_features = column_sampler_->GetFeatureSet(tree.GetDepth(left_nidx));
left_sampled_features->SetDevice(ctx_->Device());
feature_sets.emplace_back(left_sampled_features);
common::Span<bst_feature_t> left_feature_set =
interaction_constraints.Query(left_sampled_features->DeviceSpan(), left_nidx);
auto right_sampled_features = column_sampler.GetFeatureSet(tree.GetDepth(right_nidx));
right_sampled_features->SetDevice(ctx_->gpu_id);
auto right_sampled_features = column_sampler_->GetFeatureSet(tree.GetDepth(right_nidx));
right_sampled_features->SetDevice(ctx_->Device());
feature_sets.emplace_back(right_sampled_features);
common::Span<bst_feature_t> right_feature_set =
interaction_constraints.Query(right_sampled_features->DeviceSpan(),
@@ -657,7 +671,6 @@ struct GPUHistMakerDevice {
evaluator_.ApplyTreeSplit(candidate, p_tree);
const auto& parent = tree[candidate.nid];
std::size_t max_nidx = std::max(parent.LeftChild(), parent.RightChild());
interaction_constraints.Split(candidate.nid, parent.SplitIndex(), parent.LeftChild(),
parent.RightChild());
}
@@ -693,9 +706,8 @@ struct GPUHistMakerDevice {
return root_entry;
}
void UpdateTree(HostDeviceVector<GradientPair>* gpair_all, DMatrix* p_fmat,
ObjInfo const* task, RegTree* p_tree,
HostDeviceVector<bst_node_t>* p_out_position) {
void UpdateTree(HostDeviceVector<GradientPair>* gpair_all, DMatrix* p_fmat, ObjInfo const* task,
RegTree* p_tree, HostDeviceVector<bst_node_t>* p_out_position) {
auto& tree = *p_tree;
// Process maximum 32 nodes at a time
Driver<GPUExpandEntry> driver(param, 32);
@@ -720,7 +732,6 @@ struct GPUHistMakerDevice {
std::copy_if(expand_set.begin(), expand_set.end(), std::back_inserter(filtered_expand_set),
[&](const auto& e) { return driver.IsChildValid(e); });
auto new_candidates =
pinned.GetSpan<GPUExpandEntry>(filtered_expand_set.size() * 2, GPUExpandEntry());
@@ -753,8 +764,7 @@ class GPUHistMaker : public TreeUpdater {
using GradientSumT = GradientPairPrecise;
public:
explicit GPUHistMaker(Context const* ctx, ObjInfo const* task)
: TreeUpdater(ctx), task_{task} {};
explicit GPUHistMaker(Context const* ctx, ObjInfo const* task) : TreeUpdater(ctx), task_{task} {};
void Configure(const Args& args) override {
// Used in test to count how many configurations are performed
LOG(DEBUG) << "[GPU Hist]: Configure";
@@ -786,13 +796,10 @@ class GPUHistMaker : public TreeUpdater {
// build tree
try {
size_t t_idx{0};
std::size_t t_idx{0};
for (xgboost::RegTree* tree : trees) {
this->UpdateTree(param, gpair, dmat, tree, &out_position[t_idx]);
if (hist_maker_param_.debug_synchronize) {
this->CheckTreesSynchronized(tree);
}
this->hist_maker_param_.CheckTreesSynchronized(tree);
++t_idx;
}
dh::safe_cuda(cudaGetLastError());
@@ -809,13 +816,14 @@ class GPUHistMaker : public TreeUpdater {
// Synchronise the column sampling seed
uint32_t column_sampling_seed = common::GlobalRandom()();
collective::Broadcast(&column_sampling_seed, sizeof(column_sampling_seed), 0);
this->column_sampler_ = std::make_shared<common::ColumnSampler>(column_sampling_seed);
auto batch_param = BatchParam{param->max_bin, TrainParam::DftSparseThreshold()};
dh::safe_cuda(cudaSetDevice(ctx_->gpu_id));
info_->feature_types.SetDevice(ctx_->gpu_id);
maker.reset(new GPUHistMakerDevice<GradientSumT>(
maker = std::make_unique<GPUHistMakerDevice>(
ctx_, !dmat->SingleColBlock(), info_->feature_types.ConstDeviceSpan(), info_->num_row_,
*param, column_sampling_seed, info_->num_col_, batch_param));
*param, column_sampler_, info_->num_col_, batch_param);
p_last_fmat_ = dmat;
initialised_ = true;
@@ -830,21 +838,6 @@ class GPUHistMaker : public TreeUpdater {
p_last_tree_ = p_tree;
}
// Only call this method for testing
void CheckTreesSynchronized(RegTree* local_tree) const {
std::string s_model;
common::MemoryBufferStream fs(&s_model);
int rank = collective::GetRank();
if (rank == 0) {
local_tree->Save(&fs);
}
fs.Seek(0);
collective::Broadcast(&s_model, 0);
RegTree reference_tree{}; // rank 0 tree
reference_tree.Load(&fs);
CHECK(*local_tree == reference_tree);
}
void UpdateTree(TrainParam const* param, HostDeviceVector<GradientPair>* gpair, DMatrix* p_fmat,
RegTree* p_tree, HostDeviceVector<bst_node_t>* p_out_position) {
monitor_.Start("InitData");
@@ -868,7 +861,7 @@ class GPUHistMaker : public TreeUpdater {
MetaInfo* info_{}; // NOLINT
std::unique_ptr<GPUHistMakerDevice<GradientSumT>> maker; // NOLINT
std::unique_ptr<GPUHistMakerDevice> maker; // NOLINT
[[nodiscard]] char const* Name() const override { return "grow_gpu_hist"; }
[[nodiscard]] bool HasNodePosition() const override { return true; }
@@ -883,6 +876,7 @@ class GPUHistMaker : public TreeUpdater {
ObjInfo const* task_{nullptr};
common::Monitor monitor_;
std::shared_ptr<common::ColumnSampler> column_sampler_;
};
#if !defined(GTEST_TEST)
@@ -892,4 +886,131 @@ XGBOOST_REGISTER_TREE_UPDATER(GPUHistMaker, "grow_gpu_hist")
return new GPUHistMaker(ctx, task);
});
#endif // !defined(GTEST_TEST)
class GPUGlobalApproxMaker : public TreeUpdater {
public:
explicit GPUGlobalApproxMaker(Context const* ctx, ObjInfo const* task)
: TreeUpdater(ctx), task_{task} {};
void Configure(Args const& args) override {
// Used in test to count how many configurations are performed
LOG(DEBUG) << "[GPU Approx]: Configure";
hist_maker_param_.UpdateAllowUnknown(args);
dh::CheckComputeCapability();
initialised_ = false;
monitor_.Init(this->Name());
}
void LoadConfig(Json const& in) override {
auto const& config = get<Object const>(in);
FromJson(config.at("approx_train_param"), &this->hist_maker_param_);
initialised_ = false;
}
void SaveConfig(Json* p_out) const override {
auto& out = *p_out;
out["approx_train_param"] = ToJson(hist_maker_param_);
}
~GPUGlobalApproxMaker() override { dh::GlobalMemoryLogger().Log(); }
void Update(TrainParam const* param, HostDeviceVector<GradientPair>* gpair, DMatrix* p_fmat,
common::Span<HostDeviceVector<bst_node_t>> out_position,
const std::vector<RegTree*>& trees) override {
monitor_.Start("Update");
this->InitDataOnce(p_fmat);
// build tree
hess_.resize(gpair->Size());
auto hess = dh::ToSpan(hess_);
gpair->SetDevice(ctx_->Device());
auto d_gpair = gpair->ConstDeviceSpan();
auto cuctx = ctx_->CUDACtx();
thrust::transform(cuctx->CTP(), dh::tcbegin(d_gpair), dh::tcend(d_gpair), dh::tbegin(hess),
[=] XGBOOST_DEVICE(GradientPair const& g) { return g.GetHess(); });
auto const& info = p_fmat->Info();
info.feature_types.SetDevice(ctx_->Device());
auto batch = BatchParam{param->max_bin, hess, !task_->const_hess};
maker_ = std::make_unique<GPUHistMakerDevice>(
ctx_, !p_fmat->SingleColBlock(), info.feature_types.ConstDeviceSpan(), info.num_row_,
*param, column_sampler_, info.num_col_, batch);
std::size_t t_idx{0};
for (xgboost::RegTree* tree : trees) {
this->UpdateTree(gpair, p_fmat, tree, &out_position[t_idx]);
this->hist_maker_param_.CheckTreesSynchronized(tree);
++t_idx;
}
monitor_.Stop("Update");
}
void InitDataOnce(DMatrix* p_fmat) {
if (this->initialised_) {
return;
}
monitor_.Start(__func__);
CHECK(ctx_->IsCUDA()) << error::InvalidCUDAOrdinal();
// Synchronise the column sampling seed
uint32_t column_sampling_seed = common::GlobalRandom()();
collective::Broadcast(&column_sampling_seed, sizeof(column_sampling_seed), 0);
this->column_sampler_ = std::make_shared<common::ColumnSampler>(column_sampling_seed);
p_last_fmat_ = p_fmat;
initialised_ = true;
monitor_.Stop(__func__);
}
void InitData(DMatrix* p_fmat, RegTree const* p_tree) {
this->InitDataOnce(p_fmat);
p_last_tree_ = p_tree;
}
void UpdateTree(HostDeviceVector<GradientPair>* gpair, DMatrix* p_fmat, RegTree* p_tree,
HostDeviceVector<bst_node_t>* p_out_position) {
monitor_.Start("InitData");
this->InitData(p_fmat, p_tree);
monitor_.Stop("InitData");
gpair->SetDevice(ctx_->gpu_id);
maker_->UpdateTree(gpair, p_fmat, task_, p_tree, p_out_position);
}
bool UpdatePredictionCache(const DMatrix* data,
linalg::MatrixView<bst_float> p_out_preds) override {
if (maker_ == nullptr || p_last_fmat_ == nullptr || p_last_fmat_ != data) {
return false;
}
monitor_.Start("UpdatePredictionCache");
bool result = maker_->UpdatePredictionCache(p_out_preds, p_last_tree_);
monitor_.Stop("UpdatePredictionCache");
return result;
}
[[nodiscard]] char const* Name() const override { return "grow_gpu_approx"; }
[[nodiscard]] bool HasNodePosition() const override { return true; }
private:
bool initialised_{false};
GPUHistMakerTrainParam hist_maker_param_;
dh::device_vector<float> hess_;
std::shared_ptr<common::ColumnSampler> column_sampler_;
std::unique_ptr<GPUHistMakerDevice> maker_;
DMatrix* p_last_fmat_{nullptr};
RegTree const* p_last_tree_{nullptr};
ObjInfo const* task_{nullptr};
common::Monitor monitor_;
};
#if !defined(GTEST_TEST)
XGBOOST_REGISTER_TREE_UPDATER(GPUApproxMaker, "grow_gpu_approx")
.describe("Grow tree with GPU.")
.set_body([](Context const* ctx, ObjInfo const* task) {
return new GPUGlobalApproxMaker(ctx, task);
});
#endif // !defined(GTEST_TEST)
} // namespace xgboost::tree