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

fix gpu predictor when dmatrix is mismatched with model (#4613)

Browse Source
This commit is contained in:
Rong Ou
2019-06-27 16:03:02 -07:00
committed by Rory Mitchell
parent 4d6590be3c
commit 63ec95623d
2 changed files with 24 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
src/predictor/gpu_predictor.cu
View File

@@ -278,7 +278,7 @@ class GPUPredictor : public xgboost::Predictor {
}
void PredictInternal
(const SparsePage& batch, const MetaInfo& info,
(const SparsePage& batch, size_t num_features,
HostDeviceVector<bst_float>* predictions) {
if (predictions->DeviceSize(device_) == 0) { return; }
dh::safe_cuda(cudaSetDevice(device_));
@@ -287,7 +287,7 @@ class GPUPredictor : public xgboost::Predictor {
const int GRID_SIZE = static_cast<int>(dh::DivRoundUp(num_rows, BLOCK_THREADS));
int shared_memory_bytes = static_cast<int>
(sizeof(float) * info.num_col_ * BLOCK_THREADS);
(sizeof(float) * num_features * BLOCK_THREADS);
bool use_shared = true;
if (shared_memory_bytes > max_shared_memory_bytes_) {
shared_memory_bytes = 0;
@@ -300,7 +300,7 @@ class GPUPredictor : public xgboost::Predictor {
PredictKernel<BLOCK_THREADS><<<GRID_SIZE, BLOCK_THREADS, shared_memory_bytes>>>
(dh::ToSpan(nodes_), predictions->DeviceSpan(device_), dh::ToSpan(tree_segments_),
dh::ToSpan(tree_group_), batch.offset.DeviceSpan(device_),
batch.data.DeviceSpan(device_), this->tree_begin_, this->tree_end_, info.num_col_,
batch.data.DeviceSpan(device_), this->tree_begin_, this->tree_end_, num_features,
num_rows, entry_start, use_shared, this->num_group_);
}
@@ -363,7 +363,7 @@ class GPUPredictor : public xgboost::Predictor {
batch.data.Reshard(GPUDistribution::Explicit(devices_, device_offsets));
dh::ExecuteIndexShards(&shards_, [&](int idx, DeviceShard& shard) {
shard.PredictInternal(batch, dmat->Info(), out_preds);
shard.PredictInternal(batch, model.param.num_feature, out_preds);
});
batch_offset += batch.Size() * model.param.num_output_group;
}