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

Cleanup on device sketch. (#5874)

Browse Source 
* Remove old functions.

* Merge weighted and un-weighted into a common interface.
This commit is contained in:
Jiaming Yuan
2020-07-14 10:15:54 +08:00
committed by GitHub
parent 9f85e92602
commit dd445af56e
10 changed files with 97 additions and 209 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
tests/cpp/common/test_hist_util.cc
View File

@@ -164,7 +164,12 @@ TEST(CutsBuilder, SearchGroupInd) {
group_ind = CutsBuilder::SearchGroupIndFromRow(p_mat->Info().group_ptr_, 5);
ASSERT_EQ(group_ind, 2);
EXPECT_ANY_THROW(CutsBuilder::SearchGroupIndFromRow(p_mat->Info().group_ptr_, 17));
p_mat->Info().Validate(-1);
EXPECT_THROW(CutsBuilder::SearchGroupIndFromRow(p_mat->Info().group_ptr_, 17),
dmlc::Error);
std::vector<bst_uint> group_ptr {0, 1, 2};
CHECK_EQ(CutsBuilder::SearchGroupIndFromRow(group_ptr, 1), 1);
}
TEST(SparseCuts, SingleThreadedBuild) {

53
tests/cpp/common/test_hist_util.cu
View File

@@ -227,16 +227,23 @@ TEST(HistUtil, DeviceSketchExternalMemoryWithWeights) {
}
template <typename Adapter>
void ValidateBatchedCuts(Adapter adapter, int num_bins, int num_columns, int num_rows,
DMatrix* dmat) {
auto MakeUnweightedCutsForTest(Adapter adapter, int32_t num_bins, float missing, size_t batch_size = 0) {
common::HistogramCuts batched_cuts;
SketchContainer sketch_container(num_bins, num_columns, num_rows, 0);
AdapterDeviceSketch(adapter.Value(), num_bins, std::numeric_limits<float>::quiet_NaN(),
SketchContainer sketch_container(num_bins, adapter.NumColumns(), adapter.NumRows(), 0);
MetaInfo info;
AdapterDeviceSketch(adapter.Value(), num_bins, info, std::numeric_limits<float>::quiet_NaN(),
&sketch_container);
sketch_container.MakeCuts(&batched_cuts);
ValidateCuts(batched_cuts, dmat, num_bins);
return batched_cuts;
}
template <typename Adapter>
void ValidateBatchedCuts(Adapter adapter, int num_bins, int num_columns, int num_rows,
DMatrix* dmat, size_t batch_size = 0) {
common::HistogramCuts batched_cuts = MakeUnweightedCutsForTest(
adapter, num_bins, std::numeric_limits<float>::quiet_NaN());
ValidateCuts(batched_cuts, dmat, num_bins);
}
TEST(HistUtil, AdapterDeviceSketch) {
int rows = 5;
@@ -251,7 +258,7 @@ TEST(HistUtil, AdapterDeviceSketch) {
data::CupyAdapter adapter(str);
auto device_cuts = AdapterDeviceSketch(&adapter, num_bins, missing);
auto device_cuts = MakeUnweightedCutsForTest(adapter, num_bins, missing);
auto host_cuts = GetHostCuts(&adapter, num_bins, missing);
EXPECT_EQ(device_cuts.Values(), host_cuts.Values());
@@ -269,8 +276,7 @@ TEST(HistUtil, AdapterDeviceSketchMemory) {
dh::GlobalMemoryLogger().Clear();
ConsoleLogger::Configure({{"verbosity", "3"}});
auto cuts = AdapterDeviceSketch(&adapter, num_bins,
std::numeric_limits<float>::quiet_NaN());
auto cuts = MakeUnweightedCutsForTest(adapter, num_bins, std::numeric_limits<float>::quiet_NaN());
ConsoleLogger::Configure({{"verbosity", "0"}});
size_t bytes_constant = 1000;
size_t bytes_required = detail::RequiredMemory(
@@ -286,12 +292,13 @@ TEST(HistUtil, AdapterSketchSlidingWindowMemory) {
auto x = GenerateRandom(num_rows, num_columns);
auto x_device = thrust::device_vector<float>(x);
auto adapter = AdapterFromData(x_device, num_rows, num_columns);
MetaInfo info;
dh::GlobalMemoryLogger().Clear();
ConsoleLogger::Configure({{"verbosity", "3"}});
common::HistogramCuts batched_cuts;
SketchContainer sketch_container(num_bins, num_columns, num_rows, 0);
AdapterDeviceSketch(adapter.Value(), num_bins, std::numeric_limits<float>::quiet_NaN(),
AdapterDeviceSketch(adapter.Value(), num_bins, info, std::numeric_limits<float>::quiet_NaN(),
&sketch_container);
HistogramCuts cuts;
sketch_container.MakeCuts(&cuts);
@@ -318,9 +325,9 @@ TEST(HistUtil, AdapterSketchSlidingWindowWeightedMemory) {
ConsoleLogger::Configure({{"verbosity", "3"}});
common::HistogramCuts batched_cuts;
SketchContainer sketch_container(num_bins, num_columns, num_rows, 0);
AdapterDeviceSketchWeighted(adapter.Value(), num_bins, info,
std::numeric_limits<float>::quiet_NaN(),
&sketch_container);
AdapterDeviceSketch(adapter.Value(), num_bins, info,
std::numeric_limits<float>::quiet_NaN(),
&sketch_container);
HistogramCuts cuts;
sketch_container.MakeCuts(&cuts);
ConsoleLogger::Configure({{"verbosity", "0"}});
@@ -340,9 +347,8 @@ TEST(HistUtil, AdapterDeviceSketchCategorical) {
auto dmat = GetDMatrixFromData(x, n, 1);
auto x_device = thrust::device_vector<float>(x);
auto adapter = AdapterFromData(x_device, n, 1);
auto cuts = AdapterDeviceSketch(&adapter, num_bins,
std::numeric_limits<float>::quiet_NaN());
ValidateCuts(cuts, dmat.get(), num_bins);
ValidateBatchedCuts(adapter, num_bins, adapter.NumColumns(),
adapter.NumRows(), dmat.get());
}
}
}
@@ -357,9 +363,6 @@ TEST(HistUtil, AdapterDeviceSketchMultipleColumns) {
auto x_device = thrust::device_vector<float>(x);
for (auto num_bins : bin_sizes) {
auto adapter = AdapterFromData(x_device, num_rows, num_columns);
auto cuts = AdapterDeviceSketch(&adapter, num_bins,
std::numeric_limits<float>::quiet_NaN());
ValidateCuts(cuts, dmat.get(), num_bins);
ValidateBatchedCuts(adapter, num_bins, num_columns, num_rows, dmat.get());
}
}
@@ -375,11 +378,7 @@ TEST(HistUtil, AdapterDeviceSketchBatches) {
auto dmat = GetDMatrixFromData(x, num_rows, num_columns);
auto x_device = thrust::device_vector<float>(x);
auto adapter = AdapterFromData(x_device, num_rows, num_columns);
auto cuts = AdapterDeviceSketch(&adapter, num_bins,
std::numeric_limits<float>::quiet_NaN(),
batch_size);
ValidateCuts(cuts, dmat.get(), num_bins);
ValidateBatchedCuts(adapter, num_bins, num_columns, num_rows, dmat.get());
ValidateBatchedCuts(adapter, num_bins, num_columns, num_rows, dmat.get(), batch_size);
}
}
@@ -396,8 +395,8 @@ TEST(HistUtil, SketchingEquivalent) {
auto dmat_cuts = DeviceSketch(0, dmat.get(), num_bins);
auto x_device = thrust::device_vector<float>(x);
auto adapter = AdapterFromData(x_device, num_rows, num_columns);
auto adapter_cuts = AdapterDeviceSketch(
&adapter, num_bins, std::numeric_limits<float>::quiet_NaN());
common::HistogramCuts adapter_cuts = MakeUnweightedCutsForTest(
adapter, num_bins, std::numeric_limits<float>::quiet_NaN());
EXPECT_EQ(dmat_cuts.Values(), adapter_cuts.Values());
EXPECT_EQ(dmat_cuts.Ptrs(), adapter_cuts.Ptrs());
EXPECT_EQ(dmat_cuts.MinValues(), adapter_cuts.MinValues());
@@ -467,8 +466,8 @@ void TestAdapterSketchFromWeights(bool with_group) {
data::CupyAdapter adapter(m);
auto const& batch = adapter.Value();
SketchContainer sketch_container(kBins, kCols, kRows, 0);
AdapterDeviceSketchWeighted(adapter.Value(), kBins, info, std::numeric_limits<float>::quiet_NaN(),
&sketch_container);
AdapterDeviceSketch(adapter.Value(), kBins, info, std::numeric_limits<float>::quiet_NaN(),
&sketch_container);
common::HistogramCuts cuts;
sketch_container.MakeCuts(&cuts);

49
tests/cpp/common/test_quantile.cu
View File

@@ -53,8 +53,8 @@ void TestSketchUnique(float sparsity) {
.Device(0)
.GenerateArrayInterface(&storage);
data::CupyAdapter adapter(interface_str);
AdapterDeviceSketchWeighted(adapter.Value(), n_bins, info,
std::numeric_limits<float>::quiet_NaN(), &sketch);
AdapterDeviceSketch(adapter.Value(), n_bins, info,
std::numeric_limits<float>::quiet_NaN(), &sketch);
auto n_cuts = detail::RequiredSampleCutsPerColumn(n_bins, kRows);
dh::caching_device_vector<size_t> column_sizes_scan;
@@ -127,8 +127,8 @@ TEST(GPUQuantile, Prune) {
.Seed(seed)
.GenerateArrayInterface(&storage);
data::CupyAdapter adapter(interface_str);
AdapterDeviceSketchWeighted(adapter.Value(), n_bins, info,
std::numeric_limits<float>::quiet_NaN(), &sketch);
AdapterDeviceSketch(adapter.Value(), n_bins, info,
std::numeric_limits<float>::quiet_NaN(), &sketch);
auto n_cuts = detail::RequiredSampleCutsPerColumn(n_bins, kRows);
ASSERT_EQ(sketch.Data().size(), n_cuts * kCols);
@@ -158,7 +158,8 @@ TEST(GPUQuantile, MergeEmpty) {
RandomDataGenerator{kRows, kCols, 0}.Device(0).GenerateArrayInterface(
&storage_0);
data::CupyAdapter adapter_0(interface_str_0);
AdapterDeviceSketch(adapter_0.Value(), n_bins,
MetaInfo info;
AdapterDeviceSketch(adapter_0.Value(), n_bins, info,
std::numeric_limits<float>::quiet_NaN(), &sketch_0);
std::vector<SketchEntry> entries_before(sketch_0.Data().size());
@@ -197,8 +198,8 @@ TEST(GPUQuantile, MergeBasic) {
.Seed(seed)
.GenerateArrayInterface(&storage_0);
data::CupyAdapter adapter_0(interface_str_0);
AdapterDeviceSketchWeighted(adapter_0.Value(), n_bins, info,
std::numeric_limits<float>::quiet_NaN(), &sketch_0);
AdapterDeviceSketch(adapter_0.Value(), n_bins, info,
std::numeric_limits<float>::quiet_NaN(), &sketch_0);
SketchContainer sketch_1(n_bins, kCols, kRows * kRows, 0);
HostDeviceVector<float> storage_1;
@@ -207,8 +208,8 @@ TEST(GPUQuantile, MergeBasic) {
.Seed(seed)
.GenerateArrayInterface(&storage_1);
data::CupyAdapter adapter_1(interface_str_1);
AdapterDeviceSketchWeighted(adapter_1.Value(), n_bins, info,
std::numeric_limits<float>::quiet_NaN(), &sketch_1);
AdapterDeviceSketch(adapter_1.Value(), n_bins, info,
std::numeric_limits<float>::quiet_NaN(), &sketch_1);
size_t size_before_merge = sketch_0.Data().size();
sketch_0.Merge(sketch_1.ColumnsPtr(), sketch_1.Data());
@@ -243,9 +244,9 @@ void TestMergeDuplicated(int32_t n_bins, size_t cols, size_t rows, float frac) {
.Seed(seed)
.GenerateArrayInterface(&storage_0);
data::CupyAdapter adapter_0(interface_str_0);
AdapterDeviceSketchWeighted(adapter_0.Value(), n_bins, info,
std::numeric_limits<float>::quiet_NaN(),
&sketch_0);
AdapterDeviceSketch(adapter_0.Value(), n_bins, info,
std::numeric_limits<float>::quiet_NaN(),
&sketch_0);
size_t f_rows = rows * frac;
SketchContainer sketch_1(n_bins, cols, f_rows, 0);
@@ -269,9 +270,9 @@ void TestMergeDuplicated(int32_t n_bins, size_t cols, size_t rows, float frac) {
}
});
data::CupyAdapter adapter_1(interface_str_1);
AdapterDeviceSketchWeighted(adapter_1.Value(), n_bins, info,
std::numeric_limits<float>::quiet_NaN(),
&sketch_1);
AdapterDeviceSketch(adapter_1.Value(), n_bins, info,
std::numeric_limits<float>::quiet_NaN(),
&sketch_1);
size_t size_before_merge = sketch_0.Data().size();
sketch_0.Merge(sketch_1.ColumnsPtr(), sketch_1.Data());
@@ -344,9 +345,9 @@ TEST(GPUQuantile, AllReduceBasic) {
.GenerateArrayInterface(&storage);
data::CupyAdapter adapter(interface_str);
containers.emplace_back(n_bins, kCols, kRows, 0);
AdapterDeviceSketchWeighted(adapter.Value(), n_bins, info,
std::numeric_limits<float>::quiet_NaN(),
&containers.back());
AdapterDeviceSketch(adapter.Value(), n_bins, info,
std::numeric_limits<float>::quiet_NaN(),
&containers.back());
}
for (auto& sketch : containers) {
sketch.Prune(intermediate_num_cuts);
@@ -367,9 +368,9 @@ TEST(GPUQuantile, AllReduceBasic) {
.Seed(rank + seed)
.GenerateArrayInterface(&storage);
data::CupyAdapter adapter(interface_str);
AdapterDeviceSketchWeighted(adapter.Value(), n_bins, info,
std::numeric_limits<float>::quiet_NaN(),
&sketch_distributed);
AdapterDeviceSketch(adapter.Value(), n_bins, info,
std::numeric_limits<float>::quiet_NaN(),
&sketch_distributed);
sketch_distributed.AllReduce();
sketch_distributed.Unique();
@@ -426,9 +427,9 @@ TEST(GPUQuantile, SameOnAllWorkers) {
.Seed(rank + seed)
.GenerateArrayInterface(&storage);
data::CupyAdapter adapter(interface_str);
AdapterDeviceSketchWeighted(adapter.Value(), n_bins, info,
std::numeric_limits<float>::quiet_NaN(),
&sketch_distributed);
AdapterDeviceSketch(adapter.Value(), n_bins, info,
std::numeric_limits<float>::quiet_NaN(),
&sketch_distributed);
sketch_distributed.AllReduce();
sketch_distributed.Unique();
TestQuantileElemRank(0, sketch_distributed.Data(), sketch_distributed.ColumnsPtr());