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

Optimizations for quantisation on device (#4572)

Browse Source 
* - do not create device vectors for the entire sparse page while computing histograms...
   - while creating the compressed histogram indices, the row vector is created for the entire
     sparse page batch. this is needless as we only process chunks at a time based on a slice
     of the total gpu memory
   - this pr will allocate only as much as required to store the ppropriate row indices and the entries

* - do not dereference row_ptrs once the device_vector has been created to elide host copies of those counts
   - instead, grab the entry counts directly from the sparsepage
This commit is contained in:
sriramch
2019-06-18 15:50:25 -07:00
committed by Rory Mitchell
parent ba1d848767
commit 6757654337
2 changed files with 46 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files

40
tests/cpp/tree/test_gpu_hist.cu
View File

@@ -477,7 +477,7 @@ TEST(GpuHist, SortPosition) {
TestSortPosition({1, 2, 1, 2, 3}, 1, 2);
}
TEST(GpuHist, TestHistogramIndex) {
void TestHistogramIndexImpl(int n_gpus) {
// Test if the compressed histogram index matches when using a sparse
// dmatrix with and without using external memory
@@ -491,31 +491,47 @@ TEST(GpuHist, TestHistogramIndex) {
CreateSparsePageDMatrixWithRC(kNRows, kNCols, 128UL, true));
std::vector<std::pair<std::string, std::string>> training_params = {
{"max_depth", "1"},
{"max_depth", "10"},
{"max_leaves", "0"}
};
LearnerTrainParam learner_param(CreateEmptyGenericParam(0, 1));
LearnerTrainParam learner_param(CreateEmptyGenericParam(0, n_gpus));
hist_maker.Init(training_params, &learner_param);
hist_maker.InitDataOnce(hist_maker_dmat.get());
hist_maker_ext.Init(training_params, &learner_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
// histogram index
const auto &dev_shard = hist_maker.shards_[0];
std::vector<common::CompressedByteT> h_gidx_buffer(dev_shard->gidx_buffer.size());
dh::CopyDeviceSpanToVector(&h_gidx_buffer, dev_shard->gidx_buffer);
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_ext = hist_maker_ext.shards_[0];
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.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);
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) {
TestHistogramIndexImpl(1);
}
#if defined(XGBOOST_USE_NCCL)
TEST(GpuHist, MGPU_TestHistogramIndex) {
auto devices = GPUSet::AllVisible();
CHECK_GT(devices.Size(), 1);
TestHistogramIndexImpl(-1);
}
#endif
} // namespace tree
} // namespace xgboost