Decision forest random splitter (#2270)

* Random splitter without chaning high-level interfaces

* Update prototype

* Update descriptor class

* Update Descriptor and train_context classes

* Update enum class

* Working version. Unrefactored

* Update random bin index generation.

* Rename train_best_split* into train_splitter. (DPC version)

* Update splitter for small cases

* Update the code after renaming classes.

* Apply clang-format for updated files.

* Apply clang-format for train_kernel_hist*

* Update proper version of clang format.

* Update engines collestion initialization

* Update tests adding bootstrap_mode and splitter_mode

* Update docs.

* Update tests. Code formatting

* Docs with more detailed descriptions.

* Update docs for splitter_mode enum class

* Add comments forsplitter_mode enum class.

* Fix bug with large case

* Updates according change requests. In scope of #2270 pull request.

* Refactoring for PR #2270.

* Update docs regarding to PR #2270

* Working version, but prediction score is bad. Please do not consider in PR #2270

* Add bin_count as boundaries for bin selection

* Raw fix for large case

* Random splitter with debug info

* Update large_single_pass random spliiter, remove debug prints

* Add example for random splitter

* Enable distributed mode

* Add splitter by histogram

* Valid version for large, small and distr cases

* Apply clang-format

* Apply clang-format

* Default constructors/destructors for proper memory management

* Update structure for kernel_hist_impl_dpc

* Update random treshold generation

* Minor updates

* Fix minor bugs

* Minor updates for single gpu version

* Update histogram calc for single mode version

* Optimized single-GPU version

* Refactoring

* Uniform distribution for Random Splitter

* Update method description and namespaces

* Apply clang-format for common.cpp

* Fix bug for case min_bin > max_bin

* Minor updates

* Update file copyright headers.

* Update clang-format

* Fix typo, remove redundant checking, snake case.

cpp/oneapi/dal/algo/decision_forest/backend/gpu/train_kernel_hist_impl.hpp

@@ -28,8 +28,8 @@
 #include "oneapi/dal/algo/decision_forest/backend/gpu/train_service_kernels.hpp"
 #include "oneapi/dal/algo/decision_forest/backend/gpu/train_feature_type.hpp"
 #include "oneapi/dal/algo/decision_forest/backend/gpu/train_model_manager.hpp"
-#include "oneapi/dal/algo/decision_forest/backend/gpu/train_best_split_impl.hpp"
-#include "oneapi/dal/algo/decision_forest/backend/gpu/train_best_split_sp_opt_impl.hpp"
+#include "oneapi/dal/algo/decision_forest/backend/gpu/train_splitter_impl.hpp"
+#include "oneapi/dal/algo/decision_forest/backend/gpu/train_splitter_sp_opt_impl.hpp"
 
 namespace oneapi::dal::decision_forest::backend {
 
@@ -104,6 +104,12 @@ class train_kernel_hist_impl {
         const pr::ndarray<Index, 1>& node_vs_tree_map,
         rng_engine_list_t& rng_engine_list);
 
+    std::tuple<pr::ndarray<Float, 1>, sycl::event> gen_random_tresholds(
+        const train_context_t& ctx,
+        Index node_count,
+        const pr::ndarray<Index, 1>& node_vs_tree_map,
+        rng_engine_list_t& rng_engine_list);
+
     sycl::event compute_initial_imp_for_node_list(const train_context_t& ctx,
                                                   imp_data_t& imp_data_list,
                                                   pr::ndarray<Index, 1>& node_list,
@@ -168,6 +174,7 @@ class train_kernel_hist_impl {
                                    const pr::ndview<Float, 1>& response,
                                    const pr::ndarray<Index, 1>& tree_order,
                                    const pr::ndarray<Index, 1>& selected_ftr_list,
+                                   const pr::ndarray<Float, 1>& random_bins_com,
                                    const pr::ndarray<Index, 1>& bin_offset_list,
                                    const imp_data_t& imp_data_list,
                                    pr::ndarray<Index, 1>& node_list,