Updates for GPU CI tests (#3467)

* Fail GPU CI after test failure

* Fix GPU linear tests

* Reduced number of GPU tests to speed up CI

* Remove static allocations of device memory

* Resolve illegal memory access for updater_fast_hist.cc

* Fix broken r tests dependency

* Update python install documentation for GPU

appveyor.yml

@@ -54,6 +54,8 @@ install:
           Bootstrap
           $DEPS = "c('data.table','magrittr','stringi','ggplot2','DiagrammeR','Ckmeans.1d.dp','vcd','testthat','lintr','igraph','knitr','rmarkdown')"
           cmd.exe /c "R.exe -q -e ""install.packages($DEPS, repos='$CRAN', type='both')"" 2>&1"
+          $BINARY_DEPS = "c('XML')"
+          cmd.exe /c "R.exe -q -e ""install.packages($BINARY_DEPS, repos='$CRAN', type='win.binary')"" 2>&1"
         }
 
 build_script:

doc/build.md

@@ -167,7 +167,12 @@ Other versions of Visual Studio may work but are untested.
 
 ### Building with GPU support
 
-XGBoost can be built with GPU support for both Linux and Windows using cmake. GPU support works with the Python package as well as the CLI version. See [Installing R package with GPU support](#installing-r-package-with-gpu-support) for special instructions for R.
+Linux users may simply install prebuilt python binaries:
+```bash
+pip install xgboost
+```
+
+XGBoost can be built from source with GPU support for both Linux and Windows using cmake. GPU support works with the Python package as well as the CLI version. See [Installing R package with GPU support](#installing-r-package-with-gpu-support) for special instructions for R.
 
 An up-to-date version of the CUDA toolkit is required.
 

src/c_api/c_api.cc

@@ -194,9 +194,9 @@ struct XGBAPIThreadLocalEntry {
   /*! \brief result holder for returning string pointers */
   std::vector<const char *> ret_vec_charp;
   /*! \brief returning float vector. */
-  HostDeviceVector<bst_float> ret_vec_float;
+  std::vector<bst_float> ret_vec_float;
   /*! \brief temp variable of gradient pairs. */
-  HostDeviceVector<GradientPair> tmp_gpair;
+  std::vector<GradientPair> tmp_gpair;
 };
 
 // define the threadlocal store.
@@ -861,7 +861,7 @@ XGB_DLL int XGBoosterBoostOneIter(BoosterHandle handle,
                                   bst_float *grad,
                                   bst_float *hess,
                                   xgboost::bst_ulong len) {
-  HostDeviceVector<GradientPair>& tmp_gpair = XGBAPIThreadLocalStore::Get()->tmp_gpair;
+  HostDeviceVector<GradientPair> tmp_gpair;
   API_BEGIN();
   CHECK_HANDLE();
   auto* bst = static_cast<Booster*>(handle);
@@ -908,22 +908,24 @@ XGB_DLL int XGBoosterPredict(BoosterHandle handle,
                              unsigned ntree_limit,
                              xgboost::bst_ulong *len,
                              const bst_float **out_result) {
-  HostDeviceVector<bst_float>& preds =
+  std::vector<bst_float>&preds =
     XGBAPIThreadLocalStore::Get()->ret_vec_float;
   API_BEGIN();
   CHECK_HANDLE();
   auto *bst = static_cast<Booster*>(handle);
   bst->LazyInit();
+  HostDeviceVector<bst_float> tmp_preds;
   bst->learner()->Predict(
       static_cast<std::shared_ptr<DMatrix>*>(dmat)->get(),
       (option_mask & 1) != 0,
-      &preds, ntree_limit,
+      &tmp_preds, ntree_limit,
       (option_mask & 2) != 0,
       (option_mask & 4) != 0,
       (option_mask & 8) != 0,
       (option_mask & 16) != 0);
-  *out_result = dmlc::BeginPtr(preds.HostVector());
+  preds = tmp_preds.HostVector();
-  *len = static_cast<xgboost::bst_ulong>(preds.Size());
+  *out_result = dmlc::BeginPtr(preds);
+  *len = static_cast<xgboost::bst_ulong>(preds.size());
   API_END();
 }
 

src/common/column_matrix.h

@@ -122,6 +122,7 @@ class ColumnMatrix {
       boundary_[fid].row_ind_begin = accum_row_ind_;
       if (type_[fid] == kDenseColumn) {
         accum_index_ += static_cast<size_t>(nrow);
+        accum_row_ind_ += static_cast<size_t>(nrow);
       } else {
         accum_index_ += feature_counts_[fid];
         accum_row_ind_ += feature_counts_[fid];

tests/ci_build/test_gpu.sh

@@ -1,4 +1,5 @@
 #!/usr/bin/env bash
+set -e
 
 cd python-package
 python setup.py install --user

tests/python-gpu/test_gpu_linear.py

@@ -9,6 +9,10 @@ import unittest
 class TestGPULinear(unittest.TestCase):
     def test_gpu_coordinate(self):
         tm._skip_if_no_sklearn()
-        variable_param = {'alpha': [.005, .1], 'lambda': [0.005],
+        variable_param = {'booster': ['gblinear'], 'updater': ['coord_descent'], 'eta': [0.5],
-                          'coordinate_selection': ['cyclic', 'random', 'greedy'], 'n_gpus': [-1, 1]}
+                          'top_k': [10], 'tolerance': [1e-5], 'nthread': [2], 'alpha': [.005, .1], 'lambda': [0.005],
-        test_linear.assert_updater_accuracy('gpu_coord_descent', variable_param)
+                          'coordinate_selection': ['cyclic', 'random', 'greedy'], 'n_gpus': [-1]}
+        for param in test_linear.parameter_combinations(variable_param):
+            results = test_linear.run_suite(param, 200, None, scale_features=True)
+            test_linear.assert_regression_result(results, 1e-2)
+            test_linear.assert_classification_result(results)

tests/python-gpu/test_gpu_updaters.py

@@ -29,9 +29,9 @@ class TestGPU(unittest.TestCase):
             assert_gpu_results(cpu_results, gpu_results)
 
     def test_gpu_hist(self):
-        variable_param = {'n_gpus': [1, -1], 'max_depth': [2, 6], 'max_leaves': [255, 4],
+        variable_param = {'n_gpus': [-1], 'max_depth': [2, 10], 'max_leaves': [255, 4],
-                          'max_bin': [2, 16, 1024],
+                          'max_bin': [2, 256],
-                          'grow_policy': ['depthwise', 'lossguide']}
+                          'grow_policy': ['lossguide']}
         for param in parameter_combinations(variable_param):
             param['tree_method'] = 'gpu_hist'
             gpu_results = run_suite(param, select_datasets=datasets)