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

Support column split in gpu hist updater (#9384)

Browse Source
This commit is contained in:
Rong Ou
2023-08-31 03:09:35 -07:00
committed by GitHub
parent ccfc90e4c6
commit 9bab06cbca
10 changed files with 187 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
tests/cpp/tree/gpu_hist/test_evaluate_splits.cu
View File

@@ -24,7 +24,7 @@ auto ZeroParam() {
inline GradientQuantiser DummyRoundingFactor() {
thrust::device_vector<GradientPair> gpair(1);
gpair[0] = {1000.f, 1000.f}; // Tests should not exceed sum of 1000
return GradientQuantiser(dh::ToSpan(gpair));
return {dh::ToSpan(gpair), MetaInfo()};
}
thrust::device_vector<GradientPairInt64> ConvertToInteger(std::vector<GradientPairPrecise> x) {

6
tests/cpp/tree/gpu_hist/test_histogram.cu
View File

@@ -39,7 +39,7 @@ void TestDeterministicHistogram(bool is_dense, int shm_size) {
FeatureGroups feature_groups(page->Cuts(), page->is_dense, shm_size,
sizeof(GradientPairInt64));
auto quantiser = GradientQuantiser(gpair.DeviceSpan());
auto quantiser = GradientQuantiser(gpair.DeviceSpan(), MetaInfo());
BuildGradientHistogram(ctx.CUDACtx(), page->GetDeviceAccessor(0),
feature_groups.DeviceAccessor(0), gpair.DeviceSpan(), ridx, d_histogram,
quantiser);
@@ -53,7 +53,7 @@ void TestDeterministicHistogram(bool is_dense, int shm_size) {
dh::device_vector<GradientPairInt64> new_histogram(num_bins);
auto d_new_histogram = dh::ToSpan(new_histogram);
auto quantiser = GradientQuantiser(gpair.DeviceSpan());
auto quantiser = GradientQuantiser(gpair.DeviceSpan(), MetaInfo());
BuildGradientHistogram(ctx.CUDACtx(), page->GetDeviceAccessor(0),
feature_groups.DeviceAccessor(0), gpair.DeviceSpan(), ridx,
d_new_histogram, quantiser);
@@ -131,7 +131,7 @@ void TestGPUHistogramCategorical(size_t num_categories) {
dh::device_vector<GradientPairInt64> cat_hist(num_categories);
auto gpair = GenerateRandomGradients(kRows, 0, 2);
gpair.SetDevice(0);
auto quantiser = GradientQuantiser(gpair.DeviceSpan());
auto quantiser = GradientQuantiser(gpair.DeviceSpan(), MetaInfo());
/**
* Generate hist with cat data.
*/

6
tests/cpp/tree/gpu_hist/test_row_partitioner.cu
View File

@@ -30,7 +30,7 @@ void TestUpdatePositionBatch() {
std::vector<int> extra_data = {0};
// Send the first five training instances to the right node
// and the second 5 to the left node
rp.UpdatePositionBatch({0}, {1}, {2}, extra_data, [=] __device__(RowPartitioner::RowIndexT ridx, int) {
rp.UpdatePositionBatch({0}, {1}, {2}, extra_data, [=] __device__(RowPartitioner::RowIndexT ridx, int, int) {
return ridx > 4;
});
rows = rp.GetRowsHost(1);
@@ -43,7 +43,7 @@ void TestUpdatePositionBatch() {
}
// Split the left node again
rp.UpdatePositionBatch({1}, {3}, {4}, extra_data,[=] __device__(RowPartitioner::RowIndexT ridx, int) {
rp.UpdatePositionBatch({1}, {3}, {4}, extra_data,[=] __device__(RowPartitioner::RowIndexT ridx, int, int) {
return ridx < 7;
});
EXPECT_EQ(rp.GetRows(3).size(), 2);
@@ -57,7 +57,7 @@ void TestSortPositionBatch(const std::vector<int>& ridx_in, const std::vector<Se
thrust::device_vector<uint32_t> ridx_tmp(ridx_in.size());
thrust::device_vector<bst_uint> counts(segments.size());
auto op = [=] __device__(auto ridx, int data) { return ridx % 2 == 0; };
auto op = [=] __device__(auto ridx, int split_index, int data) { return ridx % 2 == 0; };
std::vector<int> op_data(segments.size());
std::vector<PerNodeData<int>> h_batch_info(segments.size());
dh::TemporaryArray<PerNodeData<int>> d_batch_info(segments.size());