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

Remove internal use of gpu_id. (#9568)

Browse Source
This commit is contained in:
Jiaming Yuan
2023-09-20 23:29:51 +08:00
committed by GitHub
parent 38ac52dd87
commit 8c676c889d
121 changed files with 1012 additions and 1044 deletions
Download Patch File Download Diff File Expand all files Collapse all files

101
tests/cpp/tree/gpu_hist/test_evaluate_splits.cu
View File

@@ -1,5 +1,5 @@
/*!
* Copyright 2020-2022 by XGBoost contributors
/**
* Copyright 2020-2023, XGBoost contributors
*/
#include <gtest/gtest.h>
#include <thrust/host_vector.h>
@@ -9,9 +9,7 @@
#include "../../histogram_helpers.h"
#include "../test_evaluate_splits.h" // TestPartitionBasedSplit
namespace xgboost {
namespace tree {
namespace xgboost::tree {
namespace {
auto ZeroParam() {
auto args = Args{{"min_child_weight", "0"}, {"lambda", "0"}};
@@ -37,11 +35,12 @@ thrust::device_vector<GradientPairInt64> ConvertToInteger(std::vector<GradientPa
}
TEST_F(TestCategoricalSplitWithMissing, GPUHistEvaluator) {
auto ctx = MakeCUDACtx(0);
thrust::device_vector<bst_feature_t> feature_set = std::vector<bst_feature_t>{0};
GPUTrainingParam param{param_};
cuts_.cut_ptrs_.SetDevice(0);
cuts_.cut_values_.SetDevice(0);
cuts_.min_vals_.SetDevice(0);
cuts_.cut_ptrs_.SetDevice(ctx.Device());
cuts_.cut_values_.SetDevice(ctx.Device());
cuts_.min_vals_.SetDevice(ctx.Device());
thrust::device_vector<GradientPairInt64> feature_histogram{ConvertToInteger(feature_histogram_)};
dh::device_vector<FeatureType> feature_types(feature_set.size(), FeatureType::kCategorical);
@@ -57,9 +56,10 @@ TEST_F(TestCategoricalSplitWithMissing, GPUHistEvaluator) {
cuts_.min_vals_.ConstDeviceSpan(),
false};
GPUHistEvaluator evaluator{param_, static_cast<bst_feature_t>(feature_set.size()), 0};
GPUHistEvaluator evaluator{param_, static_cast<bst_feature_t>(feature_set.size()), ctx.Device()};
evaluator.Reset(cuts_, dh::ToSpan(feature_types), feature_set.size(), param_, false, 0);
evaluator.Reset(cuts_, dh::ToSpan(feature_types), feature_set.size(), param_, false,
ctx.Device());
DeviceSplitCandidate result = evaluator.EvaluateSingleSplit(input, shared_inputs).split;
ASSERT_EQ(result.thresh, 1);
@@ -69,6 +69,7 @@ TEST_F(TestCategoricalSplitWithMissing, GPUHistEvaluator) {
}
TEST(GpuHist, PartitionBasic) {
auto ctx = MakeCUDACtx(0);
TrainParam tparam = ZeroParam();
tparam.max_cat_to_onehot = 0;
GPUTrainingParam param{tparam};
@@ -77,9 +78,9 @@ TEST(GpuHist, PartitionBasic) {
cuts.cut_values_.HostVector() = std::vector<float>{0.0, 1.0, 2.0};
cuts.cut_ptrs_.HostVector() = std::vector<uint32_t>{0, 3};
cuts.min_vals_.HostVector() = std::vector<float>{0.0};
cuts.cut_ptrs_.SetDevice(0);
cuts.cut_values_.SetDevice(0);
cuts.min_vals_.SetDevice(0);
cuts.cut_ptrs_.SetDevice(ctx.Device());
cuts.cut_values_.SetDevice(ctx.Device());
cuts.min_vals_.SetDevice(ctx.Device());
thrust::device_vector<bst_feature_t> feature_set = std::vector<bst_feature_t>{0};
thrust::device_vector<int> monotonic_constraints(feature_set.size(), 0);
@@ -100,8 +101,8 @@ TEST(GpuHist, PartitionBasic) {
false,
};
GPUHistEvaluator evaluator{tparam, static_cast<bst_feature_t>(feature_set.size()), 0};
evaluator.Reset(cuts, dh::ToSpan(feature_types), feature_set.size(), tparam, false, 0);
GPUHistEvaluator evaluator{tparam, static_cast<bst_feature_t>(feature_set.size()), ctx.Device()};
evaluator.Reset(cuts, dh::ToSpan(feature_types), feature_set.size(), tparam, false, ctx.Device());
{
// -1.0s go right
@@ -183,6 +184,7 @@ TEST(GpuHist, PartitionBasic) {
}
TEST(GpuHist, PartitionTwoFeatures) {
auto ctx = MakeCUDACtx(0);
TrainParam tparam = ZeroParam();
tparam.max_cat_to_onehot = 0;
GPUTrainingParam param{tparam};
@@ -191,9 +193,9 @@ TEST(GpuHist, PartitionTwoFeatures) {
cuts.cut_values_.HostVector() = std::vector<float>{0.0, 1.0, 2.0, 0.0, 1.0, 2.0};
cuts.cut_ptrs_.HostVector() = std::vector<uint32_t>{0, 3, 6};
cuts.min_vals_.HostVector() = std::vector<float>{0.0, 0.0};
cuts.cut_ptrs_.SetDevice(0);
cuts.cut_values_.SetDevice(0);
cuts.min_vals_.SetDevice(0);
cuts.cut_ptrs_.SetDevice(ctx.Device());
cuts.cut_values_.SetDevice(ctx.Device());
cuts.min_vals_.SetDevice(ctx.Device());
thrust::device_vector<bst_feature_t> feature_set = std::vector<bst_feature_t>{0, 1};
thrust::device_vector<int> monotonic_constraints(feature_set.size(), 0);
@@ -212,8 +214,8 @@ TEST(GpuHist, PartitionTwoFeatures) {
cuts.min_vals_.ConstDeviceSpan(),
false};
GPUHistEvaluator evaluator{tparam, static_cast<bst_feature_t>(feature_set.size()), 0};
evaluator.Reset(cuts, dh::ToSpan(feature_types), feature_set.size(), tparam, false, 0);
GPUHistEvaluator evaluator{tparam, static_cast<bst_feature_t>(feature_set.size()), ctx.Device()};
evaluator.Reset(cuts, dh::ToSpan(feature_types), feature_set.size(), tparam, false, ctx.Device());
{
auto parent_sum = quantiser.ToFixedPoint(GradientPairPrecise{-6.0, 3.0});
@@ -243,6 +245,7 @@ TEST(GpuHist, PartitionTwoFeatures) {
}
TEST(GpuHist, PartitionTwoNodes) {
auto ctx = MakeCUDACtx(0);
TrainParam tparam = ZeroParam();
tparam.max_cat_to_onehot = 0;
GPUTrainingParam param{tparam};
@@ -251,9 +254,9 @@ TEST(GpuHist, PartitionTwoNodes) {
cuts.cut_values_.HostVector() = std::vector<float>{0.0, 1.0, 2.0};
cuts.cut_ptrs_.HostVector() = std::vector<uint32_t>{0, 3};
cuts.min_vals_.HostVector() = std::vector<float>{0.0};
cuts.cut_ptrs_.SetDevice(0);
cuts.cut_values_.SetDevice(0);
cuts.min_vals_.SetDevice(0);
cuts.cut_ptrs_.SetDevice(ctx.Device());
cuts.cut_values_.SetDevice(ctx.Device());
cuts.min_vals_.SetDevice(ctx.Device());
thrust::device_vector<bst_feature_t> feature_set = std::vector<bst_feature_t>{0};
thrust::device_vector<int> monotonic_constraints(feature_set.size(), 0);
@@ -272,8 +275,10 @@ TEST(GpuHist, PartitionTwoNodes) {
cuts.min_vals_.ConstDeviceSpan(),
false};
GPUHistEvaluator evaluator{tparam, static_cast<bst_feature_t>(feature_set.size()), 0};
evaluator.Reset(cuts, dh::ToSpan(feature_types), feature_set.size(), tparam, false, 0);
GPUHistEvaluator evaluator{tparam, static_cast<bst_feature_t>(feature_set.size()),
ctx.Device()};
evaluator.Reset(cuts, dh::ToSpan(feature_types), feature_set.size(), tparam, false,
ctx.Device());
{
auto parent_sum = quantiser.ToFixedPoint(GradientPairPrecise{-6.0, 3.0});
@@ -295,12 +300,14 @@ TEST(GpuHist, PartitionTwoNodes) {
}
void TestEvaluateSingleSplit(bool is_categorical) {
auto ctx = MakeCUDACtx(0);
auto quantiser = DummyRoundingFactor();
auto parent_sum = quantiser.ToFixedPoint(GradientPairPrecise{0.0, 1.0});
TrainParam tparam = ZeroParam();
GPUTrainingParam param{tparam};
common::HistogramCuts cuts{MakeCutsForTest({1.0, 2.0, 11.0, 12.0}, {0, 2, 4}, {0.0, 0.0}, 0)};
common::HistogramCuts cuts{
MakeCutsForTest({1.0, 2.0, 11.0, 12.0}, {0, 2, 4}, {0.0, 0.0}, ctx.Device())};
thrust::device_vector<bst_feature_t> feature_set = std::vector<bst_feature_t>{0, 1};
// Setup gradients so that second feature gets higher gain
@@ -325,8 +332,10 @@ void TestEvaluateSingleSplit(bool is_categorical) {
cuts.min_vals_.ConstDeviceSpan(),
false};
GPUHistEvaluator evaluator{tparam, static_cast<bst_feature_t>(feature_set.size()), 0};
evaluator.Reset(cuts, dh::ToSpan(feature_types), feature_set.size(), tparam, false, 0);
GPUHistEvaluator evaluator{tparam, static_cast<bst_feature_t>(feature_set.size()),
ctx.Device()};
evaluator.Reset(cuts, dh::ToSpan(feature_types), feature_set.size(), tparam, false,
ctx.Device());
DeviceSplitCandidate result = evaluator.EvaluateSingleSplit(input, shared_inputs).split;
EXPECT_EQ(result.findex, 1);
@@ -363,7 +372,7 @@ TEST(GpuHist, EvaluateSingleSplitMissing) {
dh::ToSpan(feature_min_values),
false};
GPUHistEvaluator evaluator(tparam, feature_set.size(), 0);
GPUHistEvaluator evaluator(tparam, feature_set.size(), FstCU());
DeviceSplitCandidate result = evaluator.EvaluateSingleSplit(input, shared_inputs).split;
EXPECT_EQ(result.findex, 0);
@@ -375,7 +384,7 @@ TEST(GpuHist, EvaluateSingleSplitMissing) {
TEST(GpuHist, EvaluateSingleSplitEmpty) {
TrainParam tparam = ZeroParam();
GPUHistEvaluator evaluator(tparam, 1, 0);
GPUHistEvaluator evaluator(tparam, 1, FstCU());
DeviceSplitCandidate result =
evaluator
.EvaluateSingleSplit(
@@ -410,7 +419,7 @@ TEST(GpuHist, EvaluateSingleSplitFeatureSampling) {
dh::ToSpan(feature_min_values),
false};
GPUHistEvaluator evaluator(tparam, feature_min_values.size(), 0);
GPUHistEvaluator evaluator(tparam, feature_min_values.size(), FstCU());
DeviceSplitCandidate result = evaluator.EvaluateSingleSplit(input, shared_inputs).split;
EXPECT_EQ(result.findex, 1);
@@ -442,7 +451,7 @@ TEST(GpuHist, EvaluateSingleSplitBreakTies) {
dh::ToSpan(feature_min_values),
false};
GPUHistEvaluator evaluator(tparam, feature_min_values.size(), 0);
GPUHistEvaluator evaluator(tparam, feature_min_values.size(), FstCU());
DeviceSplitCandidate result = evaluator.EvaluateSingleSplit(input, shared_inputs).split;
EXPECT_EQ(result.findex, 0);
@@ -477,7 +486,8 @@ TEST(GpuHist, EvaluateSplits) {
dh::ToSpan(feature_min_values),
false};
GPUHistEvaluator evaluator{tparam, static_cast<bst_feature_t>(feature_min_values.size()), 0};
GPUHistEvaluator evaluator{tparam, static_cast<bst_feature_t>(feature_min_values.size()),
FstCU()};
dh::device_vector<EvaluateSplitInputs> inputs =
std::vector<EvaluateSplitInputs>{input_left, input_right};
evaluator.LaunchEvaluateSplits(input_left.feature_set.size(), dh::ToSpan(inputs), shared_inputs,
@@ -493,14 +503,15 @@ TEST(GpuHist, EvaluateSplits) {
}
TEST_F(TestPartitionBasedSplit, GpuHist) {
auto ctx = MakeCUDACtx(0);
dh::device_vector<FeatureType> ft{std::vector<FeatureType>{FeatureType::kCategorical}};
GPUHistEvaluator evaluator{param_, static_cast<bst_feature_t>(info_.num_col_), 0};
GPUHistEvaluator evaluator{param_, static_cast<bst_feature_t>(info_.num_col_), ctx.Device()};
cuts_.cut_ptrs_.SetDevice(0);
cuts_.cut_values_.SetDevice(0);
cuts_.min_vals_.SetDevice(0);
cuts_.cut_ptrs_.SetDevice(ctx.Device());
cuts_.cut_values_.SetDevice(ctx.Device());
cuts_.min_vals_.SetDevice(ctx.Device());
evaluator.Reset(cuts_, dh::ToSpan(ft), info_.num_col_, param_, false, 0);
evaluator.Reset(cuts_, dh::ToSpan(ft), info_.num_col_, param_, false, ctx.Device());
// Convert the sample histogram to fixed point
auto quantiser = DummyRoundingFactor();
@@ -528,15 +539,16 @@ class MGPUHistTest : public BaseMGPUTest {};
namespace {
void VerifyColumnSplitEvaluateSingleSplit(bool is_categorical) {
auto ctx = MakeCUDACtx(GPUIDX);
auto rank = collective::GetRank();
auto quantiser = DummyRoundingFactor();
auto parent_sum = quantiser.ToFixedPoint(GradientPairPrecise{0.0, 1.0});
TrainParam tparam = ZeroParam();
GPUTrainingParam param{tparam};
common::HistogramCuts cuts{rank == 0
? MakeCutsForTest({1.0, 2.0}, {0, 2, 2}, {0.0, 0.0}, GPUIDX)
: MakeCutsForTest({11.0, 12.0}, {0, 0, 2}, {0.0, 0.0}, GPUIDX)};
common::HistogramCuts cuts{
rank == 0 ? MakeCutsForTest({1.0, 2.0}, {0, 2, 2}, {0.0, 0.0}, ctx.Device())
: MakeCutsForTest({11.0, 12.0}, {0, 0, 2}, {0.0, 0.0}, ctx.Device())};
thrust::device_vector<bst_feature_t> feature_set = std::vector<bst_feature_t>{0, 1};
// Setup gradients so that second feature gets higher gain
@@ -562,8 +574,8 @@ void VerifyColumnSplitEvaluateSingleSplit(bool is_categorical) {
cuts.min_vals_.ConstDeviceSpan(),
false};
GPUHistEvaluator evaluator{tparam, static_cast<bst_feature_t>(feature_set.size()), GPUIDX};
evaluator.Reset(cuts, dh::ToSpan(feature_types), feature_set.size(), tparam, true, GPUIDX);
GPUHistEvaluator evaluator{tparam, static_cast<bst_feature_t>(feature_set.size()), ctx.Device()};
evaluator.Reset(cuts, dh::ToSpan(feature_types), feature_set.size(), tparam, true, ctx.Device());
DeviceSplitCandidate result = evaluator.EvaluateSingleSplit(input, shared_inputs).split;
EXPECT_EQ(result.findex, 1) << "rank: " << rank;
@@ -583,5 +595,4 @@ TEST_F(MGPUHistTest, ColumnSplitEvaluateSingleSplit) {
TEST_F(MGPUHistTest, ColumnSplitEvaluateSingleCategoricalSplit) {
DoTest(VerifyColumnSplitEvaluateSingleSplit, true);
}
} // namespace tree
} // namespace xgboost
} // namespace xgboost::tree