Improved gpu_hist_experimental algorithm (#2866)

- Implement colsampling, subsampling for gpu_hist_experimental

 - Optimised multi-GPU implementation for gpu_hist_experimental

 - Make nccl optional

 - Add Volta architecture flag

 - Optimise RegLossObj

 - Add timing utilities for debug verbose mode

 - Bump required cuda version to 8.0

tests/cpp/common/test_device_helpers.cu

@@ -6,6 +6,7 @@
 #include <xgboost/base.h>
 #include "../../../src/common/device_helpers.cuh"
 #include "gtest/gtest.h"
+#include "../../../src/common/timer.h"
 
 void CreateTestData(xgboost::bst_uint num_rows, int max_row_size,
                     thrust::host_vector<int> *row_ptr,
@@ -35,7 +36,7 @@ void SpeedTest() {
   thrust::device_vector<int> output_row(h_rows.size());
   auto d_output_row = output_row.data();
 
-  dh::Timer t;
+  xgboost::common::Timer t;
   dh::TransformLbs(
       0, &temp_memory, h_rows.size(), dh::raw(row_ptr), row_ptr.size() - 1, false,
       [=] __device__(size_t idx, size_t ridx) { d_output_row[idx] = ridx; });

tests/cpp/tree/test_gpu_hist_experimental.cu

@@ -7,8 +7,8 @@
 #include "../helpers.h"
 #include "gtest/gtest.h"
 
-#include "../../../src/tree/updater_gpu_hist_experimental.cu"
 #include "../../../src/gbm/gbtree_model.h"
+#include "../../../src/tree/updater_gpu_hist_experimental.cu"
 
 namespace xgboost {
 namespace tree {
@@ -22,7 +22,9 @@ TEST(gpu_hist_experimental, TestSparseShard) {
   hmat.Init(dmat.get(), max_bins);
   gmat.cut = &hmat;
   gmat.Init(dmat.get());
-  DeviceShard shard(0, 0, gmat, 0, rows, hmat.row_ptr.back(), TrainParam());
+  ncclComm_t comm;
+  DeviceShard shard(0, 0, gmat, 0, rows, hmat.row_ptr.back(),
+                    TrainParam());
 
   ASSERT_LT(shard.row_stride, columns);
 
@@ -54,7 +56,9 @@ TEST(gpu_hist_experimental, TestDenseShard) {
   hmat.Init(dmat.get(), max_bins);
   gmat.cut = &hmat;
   gmat.Init(dmat.get());
-  DeviceShard shard(0, 0, gmat, 0, rows, hmat.row_ptr.back(), TrainParam());
+  ncclComm_t comm;
+  DeviceShard shard(0, 0, gmat, 0, rows, hmat.row_ptr.back(),
+                    TrainParam());
 
   ASSERT_EQ(shard.row_stride, columns);