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

Multi-GPU support in GPUPredictor. (#3738)

Browse Source 
* Multi-GPU support in GPUPredictor.

- GPUPredictor is multi-GPU
- removed DeviceMatrix, as it has been made obsolete by using HostDeviceVector in DMatrix

* Replaced pointers with spans in GPUPredictor.

* Added a multi-GPU predictor test.

* Fix multi-gpu test.

* Fix n_rows < n_gpus.

* Reinitialize shards when GPUSet is changed.
* Tests range of data.

* Remove commented code.

* Remove commented code.
This commit is contained in:
Andy Adinets
2018-10-24 07:59:11 +02:00
committed by Philip Hyunsu Cho
parent 32de54fdee
commit 2a59ff2f9b
7 changed files with 198 additions and 160 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/objective/hinge.cu
View File

@@ -22,7 +22,7 @@ struct HingeObjParam : public dmlc::Parameter<HingeObjParam> {
int n_gpus;
int gpu_id;
DMLC_DECLARE_PARAMETER(HingeObjParam) {
DMLC_DECLARE_FIELD(n_gpus).set_default(0).set_lower_bound(0)
DMLC_DECLARE_FIELD(n_gpus).set_default(1).set_lower_bound(-1)
.describe("Number of GPUs to use for multi-gpu algorithms.");
DMLC_DECLARE_FIELD(gpu_id)
.set_lower_bound(0)

14
src/objective/multiclass_obj.cu
View File

@@ -31,7 +31,7 @@ struct SoftmaxMultiClassParam : public dmlc::Parameter<SoftmaxMultiClassParam> {
DMLC_DECLARE_PARAMETER(SoftmaxMultiClassParam) {
DMLC_DECLARE_FIELD(num_class).set_lower_bound(1)
.describe("Number of output class in the multi-class classification.");
DMLC_DECLARE_FIELD(n_gpus).set_default(-1).set_lower_bound(-1)
DMLC_DECLARE_FIELD(n_gpus).set_default(1).set_lower_bound(-1)
.describe("Number of GPUs to use for multi-gpu algorithms.");
DMLC_DECLARE_FIELD(gpu_id)
.set_lower_bound(0)
@@ -64,10 +64,6 @@ class SoftmaxMultiClassObj : public ObjFunction {
const int nclass = param_.num_class;
const auto ndata = static_cast<int64_t>(preds.Size() / nclass);
// clear out device memory;
out_gpair->Reshard(GPUSet::Empty());
preds.Reshard(GPUSet::Empty());
out_gpair->Reshard(GPUDistribution::Granular(devices_, nclass));
info.labels_.Reshard(GPUDistribution::Block(devices_));
info.weights_.Reshard(GPUDistribution::Block(devices_));
@@ -109,11 +105,6 @@ class SoftmaxMultiClassObj : public ObjFunction {
}, common::Range{0, ndata}, devices_, false)
.Eval(out_gpair, &info.labels_, &preds, &info.weights_, &label_correct_);
out_gpair->Reshard(GPUSet::Empty());
out_gpair->Reshard(GPUDistribution::Block(devices_));
preds.Reshard(GPUSet::Empty());
preds.Reshard(GPUDistribution::Block(devices_));
std::vector<int>& label_correct_h = label_correct_.HostVector();
for (auto const flag : label_correct_h) {
if (flag != 1) {
@@ -136,7 +127,6 @@ class SoftmaxMultiClassObj : public ObjFunction {
const auto ndata = static_cast<int64_t>(io_preds->Size() / nclass);
max_preds_.Resize(ndata);
io_preds->Reshard(GPUSet::Empty()); // clear out device memory
if (prob) {
common::Transform<>::Init(
[=] XGBOOST_DEVICE(size_t _idx, common::Span<bst_float> _preds) {
@@ -166,8 +156,6 @@ class SoftmaxMultiClassObj : public ObjFunction {
io_preds->Resize(max_preds_.Size());
io_preds->Copy(max_preds_);
}
io_preds->Reshard(GPUSet::Empty()); // clear out device memory
io_preds->Reshard(GPUDistribution::Block(devices_));
}
private:

2
src/objective/regression_obj.cu
View File

@@ -34,7 +34,7 @@ struct RegLossParam : public dmlc::Parameter<RegLossParam> {
DMLC_DECLARE_PARAMETER(RegLossParam) {
DMLC_DECLARE_FIELD(scale_pos_weight).set_default(1.0f).set_lower_bound(0.0f)
.describe("Scale the weight of positive examples by this factor");
DMLC_DECLARE_FIELD(n_gpus).set_default(-1).set_lower_bound(-1)
DMLC_DECLARE_FIELD(n_gpus).set_default(1).set_lower_bound(-1)
.describe("Number of GPUs to use for multi-gpu algorithms.");
DMLC_DECLARE_FIELD(gpu_id)
.set_lower_bound(0)