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

Optimize adapter element counting on GPU. (#9209)

Browse Source 
- Implement a simple `IterSpan` for passing iterators with size.
- Use shared memory for column size counts.
- Use one thread for each sample in row count to reduce atomic operations.
This commit is contained in:
Jiaming Yuan
2023-05-30 23:28:43 +08:00
committed by GitHub
parent 097f11b6e0
commit 17fd3f55e9
9 changed files with 323 additions and 61 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/data/iterative_dmatrix.cu
View File

@@ -80,7 +80,7 @@ void IterativeDMatrix::InitFromCUDA(Context const* ctx, BatchParam const& p,
}
auto batch_rows = num_rows();
accumulated_rows += batch_rows;
dh::caching_device_vector<size_t> row_counts(batch_rows + 1, 0);
dh::device_vector<size_t> row_counts(batch_rows + 1, 0);
common::Span<size_t> row_counts_span(row_counts.data().get(), row_counts.size());
row_stride = std::max(row_stride, Dispatch(proxy, [=](auto const& value) {
return GetRowCounts(value, row_counts_span, get_device(), missing);
@@ -134,7 +134,7 @@ void IterativeDMatrix::InitFromCUDA(Context const* ctx, BatchParam const& p,
init_page();
dh::safe_cuda(cudaSetDevice(get_device()));
auto rows = num_rows();
dh::caching_device_vector<size_t> row_counts(rows + 1, 0);
dh::device_vector<size_t> row_counts(rows + 1, 0);
common::Span<size_t> row_counts_span(row_counts.data().get(), row_counts.size());
Dispatch(proxy, [=](auto const& value) {
return GetRowCounts(value, row_counts_span, get_device(), missing);