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

further cleanup of single process multi-GPU code (#4810)

Browse Source 
* use subspan in gpu predictor instead of copying
* Revise `HostDeviceVector`
This commit is contained in:
Rong Ou
2019-08-30 02:27:23 -07:00
committed by Jiaming Yuan
parent 0184eb5d02
commit 733ed24dd9
12 changed files with 289 additions and 593 deletions
Download Patch File Download Diff File Expand all files Collapse all files

35
tests/cpp/tree/test_gpu_hist.cu
View File

@@ -113,7 +113,7 @@ TEST(GpuHist, BuildGidxDense) {
{"max_leaves", "0"},
};
param.Init(args);
DeviceShard<GradientPairPrecise> shard(0, 0, 0, kNRows, param, kNCols, kNCols);
DeviceShard<GradientPairPrecise> shard(0, kNRows, param, kNCols, kNCols);
BuildGidx(&shard, kNRows, kNCols);
std::vector<common::CompressedByteT> h_gidx_buffer(shard.gidx_buffer.size());
@@ -154,8 +154,7 @@ TEST(GpuHist, BuildGidxSparse) {
};
param.Init(args);
DeviceShard<GradientPairPrecise> shard(0, 0, 0, kNRows, param, kNCols,
kNCols);
DeviceShard<GradientPairPrecise> shard(0, kNRows, param, kNCols, kNCols);
BuildGidx(&shard, kNRows, kNCols, 0.9f);
std::vector<common::CompressedByteT> h_gidx_buffer(shard.gidx_buffer.size());
@@ -200,8 +199,7 @@ void TestBuildHist(bool use_shared_memory_histograms) {
{"max_leaves", "0"},
};
param.Init(args);
DeviceShard<GradientSumT> shard(0, 0, 0, kNRows, param, kNCols,
kNCols);
DeviceShard<GradientSumT> shard(0, kNRows, param, kNCols, kNCols);
BuildGidx(&shard, kNRows, kNCols);
xgboost::SimpleLCG gen;
@@ -303,8 +301,7 @@ TEST(GpuHist, EvaluateSplits) {
// Initialize DeviceShard
std::unique_ptr<DeviceShard<GradientPairPrecise>> shard{
new DeviceShard<GradientPairPrecise>(0, 0, 0, kNRows, param, kNCols,
kNCols)};
new DeviceShard<GradientPairPrecise>(0, kNRows, param, kNCols, kNCols)};
// Initialize DeviceShard::node_sum_gradients
shard->node_sum_gradients = {{6.4f, 12.8f}};
@@ -391,24 +388,20 @@ void TestHistogramIndexImpl() {
hist_maker_ext.Configure(training_params, &generic_param);
hist_maker_ext.InitDataOnce(hist_maker_ext_dmat.get());
ASSERT_EQ(hist_maker.shards_.size(), hist_maker_ext.shards_.size());
// Extract the device shards from the histogram makers and from that its compressed
// Extract the device shard from the histogram makers and from that its compressed
// histogram index
for (size_t i = 0; i < hist_maker.shards_.size(); ++i) {
const auto &dev_shard = hist_maker.shards_[i];
std::vector<common::CompressedByteT> h_gidx_buffer(dev_shard->gidx_buffer.size());
dh::CopyDeviceSpanToVector(&h_gidx_buffer, dev_shard->gidx_buffer);
const auto &dev_shard = hist_maker.shard_;
std::vector<common::CompressedByteT> h_gidx_buffer(dev_shard->gidx_buffer.size());
dh::CopyDeviceSpanToVector(&h_gidx_buffer, dev_shard->gidx_buffer);
const auto &dev_shard_ext = hist_maker_ext.shards_[i];
std::vector<common::CompressedByteT> h_gidx_buffer_ext(dev_shard_ext->gidx_buffer.size());
dh::CopyDeviceSpanToVector(&h_gidx_buffer_ext, dev_shard_ext->gidx_buffer);
const auto &dev_shard_ext = hist_maker_ext.shard_;
std::vector<common::CompressedByteT> h_gidx_buffer_ext(dev_shard_ext->gidx_buffer.size());
dh::CopyDeviceSpanToVector(&h_gidx_buffer_ext, dev_shard_ext->gidx_buffer);
ASSERT_EQ(dev_shard->n_bins, dev_shard_ext->n_bins);
ASSERT_EQ(dev_shard->gidx_buffer.size(), dev_shard_ext->gidx_buffer.size());
ASSERT_EQ(dev_shard->n_bins, dev_shard_ext->n_bins);
ASSERT_EQ(dev_shard->gidx_buffer.size(), dev_shard_ext->gidx_buffer.size());
ASSERT_EQ(h_gidx_buffer, h_gidx_buffer_ext);
}
ASSERT_EQ(h_gidx_buffer, h_gidx_buffer_ext);
}
TEST(GpuHist, TestHistogramIndex) {