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

Reduce device synchronisation (#5631)

Browse Source 
* Reduce device synchronisation

* Initialise pinned memory
This commit is contained in:
Rory Mitchell
2020-05-07 21:19:46 +12:00
committed by GitHub
parent 9910265064
commit fcf57823b6
7 changed files with 260 additions and 118 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
src/tree/gpu_hist/row_partitioner.cuh
View File

@@ -61,6 +61,7 @@ class RowPartitioner {
dh::caching_device_vector<int64_t>
left_counts_; // Useful to keep a bunch of zeroed memory for sort position
std::vector<cudaStream_t> streams_;
dh::PinnedMemory pinned_;
public:
RowPartitioner(int device_idx, size_t num_rows);
@@ -129,12 +130,12 @@ class RowPartitioner {
d_position[idx] = new_position;
});
// Overlap device to host memory copy (left_count) with sort
int64_t left_count;
int64_t &left_count = pinned_.GetSpan<int64_t>(1)[0];
dh::safe_cuda(cudaMemcpyAsync(&left_count, d_left_count, sizeof(int64_t),
cudaMemcpyDeviceToHost, streams_[0]));
SortPositionAndCopy(segment, left_nidx, right_nidx, d_left_count,
streams_[1]);
SortPositionAndCopy(segment, left_nidx, right_nidx, d_left_count, streams_[1]
);
dh::safe_cuda(cudaStreamSynchronize(streams_[0]));
CHECK_LE(left_count, segment.Size());