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

More support for column split in gpu predictor (#9562)

Browse Source
This commit is contained in:
Rong Ou
2023-09-13 17:13:13 -07:00
committed by GitHub
parent a343ae3b34
commit d8c3cc92ae
5 changed files with 212 additions and 124 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
tests/cpp/predictor/test_cpu_predictor.cc
View File

@@ -127,8 +127,8 @@ TEST(CpuPredictor, IterationRange) {
}
TEST(CpuPredictor, IterationRangeColmnSplit) {
Context ctx;
TestIterationRangeColumnSplit(&ctx);
auto constexpr kWorldSize = 2;
TestIterationRangeColumnSplit(kWorldSize, false);
}
TEST(CpuPredictor, ExternalMemory) {
@@ -226,23 +226,21 @@ TEST(CPUPredictor, GHistIndexTraining) {
}
TEST(CPUPredictor, CategoricalPrediction) {
Context ctx;
TestCategoricalPrediction(&ctx, false);
TestCategoricalPrediction(false, false);
}
TEST(CPUPredictor, CategoricalPredictionColumnSplit) {
Context ctx;
TestCategoricalPredictionColumnSplit(&ctx);
auto constexpr kWorldSize = 2;
RunWithInMemoryCommunicator(kWorldSize, TestCategoricalPrediction, false, true);
}
TEST(CPUPredictor, CategoricalPredictLeaf) {
Context ctx;
TestCategoricalPredictLeaf(&ctx, false);
TestCategoricalPredictLeaf(false, false);
}
TEST(CPUPredictor, CategoricalPredictLeafColumnSplit) {
Context ctx;
TestCategoricalPredictLeafColumnSplit(&ctx);
auto constexpr kWorldSize = 2;
RunWithInMemoryCommunicator(kWorldSize, TestCategoricalPredictLeaf, false, true);
}
TEST(CpuPredictor, UpdatePredictionCache) {
@@ -256,8 +254,8 @@ TEST(CpuPredictor, LesserFeatures) {
}
TEST(CpuPredictor, LesserFeaturesColumnSplit) {
Context ctx;
TestPredictionWithLesserFeaturesColumnSplit(&ctx);
auto constexpr kWorldSize = 2;
RunWithInMemoryCommunicator(kWorldSize, TestPredictionWithLesserFeaturesColumnSplit, false);
}
TEST(CpuPredictor, Sparse) {
@@ -267,9 +265,9 @@ TEST(CpuPredictor, Sparse) {
}
TEST(CpuPredictor, SparseColumnSplit) {
Context ctx;
TestSparsePredictionColumnSplit(&ctx, 0.2);
TestSparsePredictionColumnSplit(&ctx, 0.8);
auto constexpr kWorldSize = 2;
TestSparsePredictionColumnSplit(kWorldSize, false, 0.2);
TestSparsePredictionColumnSplit(kWorldSize, false, 0.8);
}
TEST(CpuPredictor, Multi) {

27
tests/cpp/predictor/test_gpu_predictor.cu
View File

@@ -206,6 +206,10 @@ TEST(GpuPredictor, LesserFeatures) {
TestPredictionWithLesserFeatures(&ctx);
}
TEST_F(MGPUPredictorTest, LesserFeaturesColumnSplit) {
RunWithInMemoryCommunicator(world_size_, TestPredictionWithLesserFeaturesColumnSplit, true);
}
// Very basic test of empty model
TEST(GPUPredictor, ShapStump) {
cudaSetDevice(0);
@@ -270,14 +274,24 @@ TEST(GPUPredictor, IterationRange) {
TestIterationRange(&ctx);
}
TEST_F(MGPUPredictorTest, IterationRangeColumnSplit) {
TestIterationRangeColumnSplit(world_size_, true);
}
TEST(GPUPredictor, CategoricalPrediction) {
auto ctx = MakeCUDACtx(0);
TestCategoricalPrediction(&ctx, false);
TestCategoricalPrediction(true, false);
}
TEST_F(MGPUPredictorTest, CategoricalPredictionColumnSplit) {
RunWithInMemoryCommunicator(world_size_, TestCategoricalPrediction, true, true);
}
TEST(GPUPredictor, CategoricalPredictLeaf) {
auto ctx = MakeCUDACtx(0);
TestCategoricalPredictLeaf(&ctx, false);
TestCategoricalPredictLeaf(true, false);
}
TEST_F(MGPUPredictorTest, CategoricalPredictionLeafColumnSplit) {
RunWithInMemoryCommunicator(world_size_, TestCategoricalPredictLeaf, true, true);
}
TEST(GPUPredictor, PredictLeafBasic) {
@@ -305,4 +319,9 @@ TEST(GPUPredictor, Sparse) {
TestSparsePrediction(&ctx, 0.2);
TestSparsePrediction(&ctx, 0.8);
}
TEST_F(MGPUPredictorTest, SparseColumnSplit) {
TestSparsePredictionColumnSplit(world_size_, true, 0.2);
TestSparsePredictionColumnSplit(world_size_, true, 0.8);
}
} // namespace xgboost::predictor

159
tests/cpp/predictor/test_predictor.cc
View File

@@ -172,16 +172,6 @@ void VerifyPredictionWithLesserFeatures(Learner *learner, bst_row_t kRows,
ASSERT_THROW({ learner->Predict(m_invalid, false, &prediction, 0, 0); }, dmlc::Error);
}
void VerifyPredictionWithLesserFeaturesColumnSplit(Learner *learner, size_t rows,
std::shared_ptr<DMatrix> m_test,
std::shared_ptr<DMatrix> m_invalid) {
auto const world_size = collective::GetWorldSize();
auto const rank = collective::GetRank();
std::shared_ptr<DMatrix> sliced_test{m_test->SliceCol(world_size, rank)};
std::shared_ptr<DMatrix> sliced_invalid{m_invalid->SliceCol(world_size, rank)};
VerifyPredictionWithLesserFeatures(learner, rows, sliced_test, sliced_invalid);
}
} // anonymous namespace
void TestPredictionWithLesserFeatures(Context const *ctx) {
@@ -229,16 +219,24 @@ void TestPredictionDeviceAccess() {
#endif // defined(XGBOOST_USE_CUDA)
}
void TestPredictionWithLesserFeaturesColumnSplit(Context const *ctx) {
size_t constexpr kRows = 256, kTrainCols = 256, kTestCols = 4, kIters = 4;
auto m_train = RandomDataGenerator(kRows, kTrainCols, 0.5).GenerateDMatrix(true);
auto learner = LearnerForTest(ctx, m_train, kIters);
void TestPredictionWithLesserFeaturesColumnSplit(bool use_gpu) {
auto const world_size = collective::GetWorldSize();
auto const rank = collective::GetRank();
std::size_t constexpr kRows = 256, kTrainCols = 256, kTestCols = 4, kIters = 4;
auto m_train = RandomDataGenerator(kRows, kTrainCols, 0.5).Seed(rank).GenerateDMatrix(true);
Context ctx;
if (use_gpu) {
ctx = MakeCUDACtx(common::AllVisibleGPUs() == 1 ? 0 : rank);
}
auto learner = LearnerForTest(&ctx, m_train, kIters);
auto m_test = RandomDataGenerator(kRows, kTestCols, 0.5).GenerateDMatrix(false);
auto m_invalid = RandomDataGenerator(kRows, kTrainCols + 1, 0.5).GenerateDMatrix(false);
auto constexpr kWorldSize = 2;
RunWithInMemoryCommunicator(kWorldSize, VerifyPredictionWithLesserFeaturesColumnSplit,
learner.get(), kRows, m_test, m_invalid);
std::shared_ptr<DMatrix> sliced_test{m_test->SliceCol(world_size, rank)};
std::shared_ptr<DMatrix> sliced_invalid{m_invalid->SliceCol(world_size, rank)};
VerifyPredictionWithLesserFeatures(learner.get(), kRows, sliced_test, sliced_invalid);
}
void GBTreeModelForTest(gbm::GBTreeModel *model, uint32_t split_ind,
@@ -260,7 +258,11 @@ void GBTreeModelForTest(gbm::GBTreeModel *model, uint32_t split_ind,
model->CommitModelGroup(std::move(trees), 0);
}
void TestCategoricalPrediction(Context const* ctx, bool is_column_split) {
void TestCategoricalPrediction(bool use_gpu, bool is_column_split) {
Context ctx;
if (use_gpu) {
ctx = MakeCUDACtx(common::AllVisibleGPUs() == 1 ? 0 : collective::GetRank());
}
size_t constexpr kCols = 10;
PredictionCacheEntry out_predictions;
@@ -270,10 +272,10 @@ void TestCategoricalPrediction(Context const* ctx, bool is_column_split) {
float left_weight = 1.3f;
float right_weight = 1.7f;
gbm::GBTreeModel model(&mparam, ctx);
gbm::GBTreeModel model(&mparam, &ctx);
GBTreeModelForTest(&model, split_ind, split_cat, left_weight, right_weight);
std::unique_ptr<Predictor> predictor{CreatePredictorForTest(ctx)};
std::unique_ptr<Predictor> predictor{CreatePredictorForTest(&ctx)};
std::vector<float> row(kCols);
row[split_ind] = split_cat;
@@ -303,12 +305,11 @@ void TestCategoricalPrediction(Context const* ctx, bool is_column_split) {
ASSERT_EQ(out_predictions.predictions.HostVector()[0], left_weight + score);
}
void TestCategoricalPredictionColumnSplit(Context const *ctx) {
auto constexpr kWorldSize = 2;
RunWithInMemoryCommunicator(kWorldSize, TestCategoricalPrediction, ctx, true);
}
void TestCategoricalPredictLeaf(Context const *ctx, bool is_column_split) {
void TestCategoricalPredictLeaf(bool use_gpu, bool is_column_split) {
Context ctx;
if (use_gpu) {
ctx = MakeCUDACtx(common::AllVisibleGPUs() == 1 ? 0 : collective::GetRank());
}
size_t constexpr kCols = 10;
PredictionCacheEntry out_predictions;
@@ -319,10 +320,10 @@ void TestCategoricalPredictLeaf(Context const *ctx, bool is_column_split) {
float left_weight = 1.3f;
float right_weight = 1.7f;
gbm::GBTreeModel model(&mparam, ctx);
gbm::GBTreeModel model(&mparam, &ctx);
GBTreeModelForTest(&model, split_ind, split_cat, left_weight, right_weight);
std::unique_ptr<Predictor> predictor{CreatePredictorForTest(ctx)};
std::unique_ptr<Predictor> predictor{CreatePredictorForTest(&ctx)};
std::vector<float> row(kCols);
row[split_ind] = split_cat;
@@ -347,11 +348,6 @@ void TestCategoricalPredictLeaf(Context const *ctx, bool is_column_split) {
ASSERT_EQ(out_predictions.predictions.HostVector()[0], 1);
}
void TestCategoricalPredictLeafColumnSplit(Context const *ctx) {
auto constexpr kWorldSize = 2;
RunWithInMemoryCommunicator(kWorldSize, TestCategoricalPredictLeaf, ctx, true);
}
void TestIterationRange(Context const* ctx) {
size_t constexpr kRows = 1000, kCols = 20, kClasses = 4, kForest = 3, kIters = 10;
auto dmat = RandomDataGenerator(kRows, kCols, 0)
@@ -411,15 +407,30 @@ void TestIterationRange(Context const* ctx) {
}
namespace {
void VerifyIterationRangeColumnSplit(DMatrix *dmat, Learner *learner, Learner *sliced,
void VerifyIterationRangeColumnSplit(bool use_gpu, Json const &ranged_model,
Json const &sliced_model, std::size_t rows, std::size_t cols,
std::size_t classes,
std::vector<float> const &expected_margin_ranged,
std::vector<float> const &expected_margin_sliced,
std::vector<float> const &expected_leaf_ranged,
std::vector<float> const &expected_leaf_sliced) {
auto const world_size = collective::GetWorldSize();
auto const rank = collective::GetRank();
Context ctx;
if (use_gpu) {
ctx = MakeCUDACtx(common::AllVisibleGPUs() == 1 ? 0 : rank);
}
auto dmat = RandomDataGenerator(rows, cols, 0).GenerateDMatrix(true, true, classes);
std::shared_ptr<DMatrix> Xy{dmat->SliceCol(world_size, rank)};
std::unique_ptr<Learner> learner{Learner::Create({Xy})};
learner->SetParam("device", ctx.DeviceName());
learner->LoadModel(ranged_model);
std::unique_ptr<Learner> sliced{Learner::Create({Xy})};
sliced->SetParam("device", ctx.DeviceName());
sliced->LoadModel(sliced_model);
HostDeviceVector<float> out_predt_sliced;
HostDeviceVector<float> out_predt_ranged;
@@ -428,11 +439,15 @@ void VerifyIterationRangeColumnSplit(DMatrix *dmat, Learner *learner, Learner *s
sliced->Predict(Xy, true, &out_predt_sliced, 0, 0, false, false, false, false, false);
learner->Predict(Xy, true, &out_predt_ranged, 0, 3, false, false, false, false, false);
auto const &h_sliced = out_predt_sliced.HostVector();
auto const &h_range = out_predt_ranged.HostVector();
ASSERT_EQ(h_sliced.size(), expected_margin_sliced.size());
ASSERT_EQ(h_sliced, expected_margin_sliced);
ASSERT_EQ(h_range.size(), expected_margin_ranged.size());
ASSERT_EQ(h_range, expected_margin_ranged);
auto const &h_ranged = out_predt_ranged.HostVector();
EXPECT_EQ(h_sliced.size(), expected_margin_sliced.size());
for (std::size_t i = 0; i < expected_margin_sliced.size(); ++i) {
ASSERT_FLOAT_EQ(h_sliced[i], expected_margin_sliced[i]) << "rank " << rank << ", i " << i;
}
EXPECT_EQ(h_ranged.size(), expected_margin_ranged.size());
for (std::size_t i = 0; i < expected_margin_ranged.size(); ++i) {
ASSERT_FLOAT_EQ(h_ranged[i], expected_margin_ranged[i]) << "rank " << rank << ", i " << i;
}
}
// Leaf
@@ -440,21 +455,27 @@ void VerifyIterationRangeColumnSplit(DMatrix *dmat, Learner *learner, Learner *s
sliced->Predict(Xy, false, &out_predt_sliced, 0, 0, false, true, false, false, false);
learner->Predict(Xy, false, &out_predt_ranged, 0, 3, false, true, false, false, false);
auto const &h_sliced = out_predt_sliced.HostVector();
auto const &h_range = out_predt_ranged.HostVector();
ASSERT_EQ(h_sliced.size(), expected_leaf_sliced.size());
ASSERT_EQ(h_sliced, expected_leaf_sliced);
ASSERT_EQ(h_range.size(), expected_leaf_ranged.size());
ASSERT_EQ(h_range, expected_leaf_ranged);
auto const &h_ranged = out_predt_ranged.HostVector();
EXPECT_EQ(h_sliced.size(), expected_leaf_sliced.size());
for (std::size_t i = 0; i < expected_leaf_sliced.size(); ++i) {
ASSERT_FLOAT_EQ(h_sliced[i], expected_leaf_sliced[i]) << "rank " << rank << ", i " << i;
}
EXPECT_EQ(h_ranged.size(), expected_leaf_ranged.size());
for (std::size_t i = 0; i < expected_leaf_ranged.size(); ++i) {
ASSERT_FLOAT_EQ(h_ranged[i], expected_leaf_ranged[i]) << "rank " << rank << ", i " << i;
}
}
}
} // anonymous namespace
void TestIterationRangeColumnSplit(Context const* ctx) {
size_t constexpr kRows = 1000, kCols = 20, kClasses = 4, kForest = 3, kIters = 10;
void TestIterationRangeColumnSplit(int world_size, bool use_gpu) {
std::size_t constexpr kRows = 1000, kCols = 20, kClasses = 4, kForest = 3, kIters = 10;
auto dmat = RandomDataGenerator(kRows, kCols, 0).GenerateDMatrix(true, true, kClasses);
auto learner = LearnerForTest(ctx, dmat, kIters, kForest);
learner->SetParam("device", ctx->DeviceName());
Context ctx;
if (use_gpu) {
ctx = MakeCUDACtx(0);
}
auto learner = LearnerForTest(&ctx, dmat, kIters, kForest);
bool bound = false;
std::unique_ptr<Learner> sliced{learner->Slice(0, 3, 1, &bound)};
@@ -476,9 +497,13 @@ void TestIterationRangeColumnSplit(Context const* ctx) {
auto const &leaf_sliced = leaf_predt_sliced.HostVector();
auto const &leaf_ranged = leaf_predt_ranged.HostVector();
auto constexpr kWorldSize = 2;
RunWithInMemoryCommunicator(kWorldSize, VerifyIterationRangeColumnSplit, dmat.get(),
learner.get(), sliced.get(), margin_ranged, margin_sliced,
Json ranged_model{Object{}};
learner->SaveModel(&ranged_model);
Json sliced_model{Object{}};
sliced->SaveModel(&sliced_model);
RunWithInMemoryCommunicator(world_size, VerifyIterationRangeColumnSplit, use_gpu, ranged_model,
sliced_model, kRows, kCols, kClasses, margin_ranged, margin_sliced,
leaf_ranged, leaf_sliced);
}
@@ -539,11 +564,20 @@ void TestSparsePrediction(Context const *ctx, float sparsity) {
}
namespace {
void VerifySparsePredictionColumnSplit(DMatrix *dmat, Learner *learner,
void VerifySparsePredictionColumnSplit(bool use_gpu, Json const &model, std::size_t rows,
std::size_t cols, float sparsity,
std::vector<float> const &expected_predt) {
std::shared_ptr<DMatrix> sliced{
dmat->SliceCol(collective::GetWorldSize(), collective::GetRank())};
Context ctx;
if (use_gpu) {
ctx = MakeCUDACtx(common::AllVisibleGPUs() == 1 ? 0 : collective::GetRank());
}
auto Xy = RandomDataGenerator(rows, cols, sparsity).GenerateDMatrix(true);
std::shared_ptr<DMatrix> sliced{Xy->SliceCol(collective::GetWorldSize(), collective::GetRank())};
HostDeviceVector<float> sparse_predt;
std::unique_ptr<Learner> learner{Learner::Create({sliced})};
learner->SetParam("device", ctx.DeviceName());
learner->LoadModel(model);
learner->Predict(sliced, false, &sparse_predt, 0, 0);
auto const &predt = sparse_predt.HostVector();
@@ -554,10 +588,14 @@ void VerifySparsePredictionColumnSplit(DMatrix *dmat, Learner *learner,
}
} // anonymous namespace
void TestSparsePredictionColumnSplit(Context const* ctx, float sparsity) {
void TestSparsePredictionColumnSplit(int world_size, bool use_gpu, float sparsity) {
Context ctx;
if (use_gpu) {
ctx = MakeCUDACtx(0);
}
size_t constexpr kRows = 512, kCols = 128, kIters = 4;
auto Xy = RandomDataGenerator(kRows, kCols, sparsity).GenerateDMatrix(true);
auto learner = LearnerForTest(ctx, Xy, kIters);
auto learner = LearnerForTest(&ctx, Xy, kIters);
HostDeviceVector<float> sparse_predt;
@@ -567,12 +605,11 @@ void TestSparsePredictionColumnSplit(Context const* ctx, float sparsity) {
learner.reset(Learner::Create({Xy}));
learner->LoadModel(model);
learner->SetParam("device", ctx->DeviceName());
learner->SetParam("device", ctx.DeviceName());
learner->Predict(Xy, false, &sparse_predt, 0, 0);
auto constexpr kWorldSize = 2;
RunWithInMemoryCommunicator(kWorldSize, VerifySparsePredictionColumnSplit, Xy.get(),
learner.get(), sparse_predt.HostVector());
RunWithInMemoryCommunicator(world_size, VerifySparsePredictionColumnSplit, use_gpu, model,
kRows, kCols, sparsity, sparse_predt.HostVector());
}
void TestVectorLeafPrediction(Context const *ctx) {

14
tests/cpp/predictor/test_predictor.h
View File

@@ -94,23 +94,19 @@ void TestPredictionWithLesserFeatures(Context const* ctx);
void TestPredictionDeviceAccess();
void TestCategoricalPrediction(Context const* ctx, bool is_column_split);
void TestCategoricalPrediction(bool use_gpu, bool is_column_split);
void TestCategoricalPredictionColumnSplit(Context const* ctx);
void TestPredictionWithLesserFeaturesColumnSplit(bool use_gpu);
void TestPredictionWithLesserFeaturesColumnSplit(Context const* ctx);
void TestCategoricalPredictLeaf(Context const* ctx, bool is_column_split);
void TestCategoricalPredictLeafColumnSplit(Context const* ctx);
void TestCategoricalPredictLeaf(bool use_gpu, bool is_column_split);
void TestIterationRange(Context const* ctx);
void TestIterationRangeColumnSplit(Context const* ctx);
void TestIterationRangeColumnSplit(int world_size, bool use_gpu);
void TestSparsePrediction(Context const* ctx, float sparsity);
void TestSparsePredictionColumnSplit(Context const* ctx, float sparsity);
void TestSparsePredictionColumnSplit(int world_size, bool use_gpu, float sparsity);
void TestVectorLeafPrediction(Context const* ctx);
} // namespace xgboost