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

Fix CPU hist init for sparse dataset. (#4625)

Browse Source 
* Fix CPU hist init for sparse dataset.

* Implement sparse histogram cut.
* Allow empty features.

* Fix windows build, don't use sparse in distributed environment.

* Comments.

* Smaller threshold.

* Fix windows omp.

* Fix msvc lambda capture.

* Fix MSVC macro.

* Fix MSVC initialization list.

* Fix MSVC initialization list x2.

* Preserve categorical feature behavior.

* Rename matrix to sparse cuts.
* Reuse UseGroup.
* Check for categorical data when adding cut.

Co-Authored-By: Philip Hyunsu Cho <chohyu01@cs.washington.edu>

* Sanity check.

* Fix comments.

* Fix comment.
This commit is contained in:
Jiaming Yuan
2019-07-04 19:27:03 -04:00
committed by Philip Hyunsu Cho
parent b7a1f22d24
commit d9a47794a5
33 changed files with 681 additions and 299 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/predictor/gpu_predictor.cu
View File

@@ -238,7 +238,7 @@ class GPUPredictor : public xgboost::Predictor {
auto& offsets = *out_offsets;
size_t n_shards = devices_.Size();
offsets.resize(n_shards + 2);
size_t rows_per_shard = dh::DivRoundUp(batch_size, n_shards);
size_t rows_per_shard = common::DivRoundUp(batch_size, n_shards);
for (size_t shard = 0; shard < devices_.Size(); ++shard) {
size_t n_rows = std::min(batch_size, shard * rows_per_shard);
offsets[shard] = batch_offset + n_rows * n_classes;
@@ -284,7 +284,7 @@ class GPUPredictor : public xgboost::Predictor {
dh::safe_cuda(cudaSetDevice(device_));
const int BLOCK_THREADS = 128;
size_t num_rows = batch.offset.DeviceSize(device_) - 1;
const int GRID_SIZE = static_cast<int>(dh::DivRoundUp(num_rows, BLOCK_THREADS));
const int GRID_SIZE = static_cast<int>(common::DivRoundUp(num_rows, BLOCK_THREADS));
int shared_memory_bytes = static_cast<int>
(sizeof(float) * num_features * BLOCK_THREADS);