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

Resolve GPU bug on large files (#3472)

Browse Source 
Remove calls to thrust copy, fix indexing bug
This commit is contained in:
Rory Mitchell 2018-07-16 20:43:45 +12:00 committed by GitHub
parent 1b59316444
commit 0f145a0365
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
3 changed files with 28 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
src/linear/updater_gpu_coordinate.cu
View File

@ -132,9 +132,10 @@ class DeviceShard {
for (int fidx = 0; fidx < batch.Size(); fidx++) { for (int fidx = 0; fidx < batch.Size(); fidx++) {
auto col = batch[fidx]; auto col = batch[fidx];
thrust::copy(col.data + column_segments[fidx].first, auto seg = column_segments[fidx];
col.data + column_segments[fidx].second, dh::safe_cuda(cudaMemcpy(
data_.tbegin() + row_ptr_[fidx]); data_.Data() + row_ptr_[fidx], col.data + seg.first,
sizeof(Entry) * (seg.second - seg.first), cudaMemcpyHostToDevice));
} }
// Rescale indices with respect to current shard // Rescale indices with respect to current shard
RescaleIndices(ridx_begin_, &data_); RescaleIndices(ridx_begin_, &data_);

24
src/predictor/gpu_predictor.cu
View File

@ -66,16 +66,18 @@ struct DeviceMatrix {
while (iter->Next()) { while (iter->Next()) {
auto batch = iter->Value(); auto batch = iter->Value();
// Copy row ptr // Copy row ptr
thrust::copy(batch.offset.data(), batch.offset.data() + batch.Size() + 1, dh::safe_cuda(cudaMemcpy(
row_ptr.tbegin() + batch.base_rowid); row_ptr.Data() + batch.base_rowid, batch.offset.data(),
sizeof(size_t) * batch.offset.size(), cudaMemcpyHostToDevice));
if (batch.base_rowid > 0) { if (batch.base_rowid > 0) {
auto begin_itr = row_ptr.tbegin() + batch.base_rowid; auto begin_itr = row_ptr.tbegin() + batch.base_rowid;
auto end_itr = begin_itr + batch.Size() + 1; auto end_itr = begin_itr + batch.Size() + 1;
IncrementOffset(begin_itr, end_itr, batch.base_rowid); IncrementOffset(begin_itr, end_itr, batch.base_rowid);
} }
dh::safe_cuda(cudaMemcpy(data.Data() + data_offset, batch.data.data(),
sizeof(Entry) * batch.data.size(),
cudaMemcpyHostToDevice));
// Copy data // Copy data
thrust::copy(batch.data.begin(), batch.data.end(),
data.tbegin() + data_offset);
data_offset += batch.data.size(); data_offset += batch.data.size();
} }
} }
@ -301,13 +303,17 @@ class GPUPredictor : public xgboost::Predictor {
} }
nodes.resize(h_nodes.size()); nodes.resize(h_nodes.size());
thrust::copy(h_nodes.begin(), h_nodes.end(), nodes.begin()); dh::safe_cuda(cudaMemcpy(dh::Raw(nodes), h_nodes.data(),
sizeof(DevicePredictionNode) * h_nodes.size(),
cudaMemcpyHostToDevice));
tree_segments.resize(h_tree_segments.size()); tree_segments.resize(h_tree_segments.size());
thrust::copy(h_tree_segments.begin(), h_tree_segments.end(), dh::safe_cuda(cudaMemcpy(dh::Raw(tree_segments), h_tree_segments.data(),
tree_segments.begin()); sizeof(size_t) * h_tree_segments.size(),
cudaMemcpyHostToDevice));
tree_group.resize(model.tree_info.size()); tree_group.resize(model.tree_info.size());
thrust::copy(model.tree_info.begin(), model.tree_info.end(), dh::safe_cuda(cudaMemcpy(dh::Raw(tree_group), model.tree_info.data(),
tree_group.begin()); sizeof(int) * model.tree_info.size(),
cudaMemcpyHostToDevice));
device_matrix->predictions.resize(out_preds->Size()); device_matrix->predictions.resize(out_preds->Size());
auto& predictions = device_matrix->predictions; auto& predictions = device_matrix->predictions;

13
src/tree/updater_gpu_hist.cu
View File

@ -369,7 +369,7 @@ struct DeviceShard {
// find the maximum row size // find the maximum row size
thrust::device_vector<size_t> row_ptr_d( thrust::device_vector<size_t> row_ptr_d(
&row_batch.offset[row_begin_idx], &row_batch.offset[row_end_idx + 1]); row_batch.offset.data() + row_begin_idx, row_batch.offset.data() + row_end_idx + 1);
auto row_iter = row_ptr_d.begin(); auto row_iter = row_ptr_d.begin();
auto get_size = [=] __device__(size_t row) { auto get_size = [=] __device__(size_t row) {
@ -381,9 +381,8 @@ struct DeviceShard {
TransformT row_size_iter = TransformT(counting, get_size); TransformT row_size_iter = TransformT(counting, get_size);
row_stride = thrust::reduce(row_size_iter, row_size_iter + n_rows, 0, row_stride = thrust::reduce(row_size_iter, row_size_iter + n_rows, 0,
thrust::maximum<size_t>()); thrust::maximum<size_t>());
int num_symbols =
// allocate compressed bin data n_bins + 1;
int num_symbols = n_bins + 1;
size_t compressed_size_bytes = size_t compressed_size_bytes =
common::CompressedBufferWriter::CalculateBufferSize(row_stride * n_rows, common::CompressedBufferWriter::CalculateBufferSize(row_stride * n_rows,
num_symbols); num_symbols);
@ -674,8 +673,10 @@ struct DeviceShard {
CalcWeightTrainParam param_d(param); CalcWeightTrainParam param_d(param);
thrust::copy(node_sum_gradients.begin(), node_sum_gradients.end(), dh::safe_cuda(cudaMemcpy(node_sum_gradients_d.Data(),
node_sum_gradients_d.tbegin()); node_sum_gradients.data(),
sizeof(GradientPair) * node_sum_gradients.size(),
cudaMemcpyHostToDevice));
auto d_position = position.Current(); auto d_position = position.Current();
auto d_ridx = ridx.Current(); auto d_ridx = ridx.Current();
auto d_node_sum_gradients = node_sum_gradients_d.Data(); auto d_node_sum_gradients = node_sum_gradients_d.Data();