Point multicolor GS: faster handling of long/bulk rows

perf_test/sparse/CMakeLists.txt

@@ -78,6 +78,7 @@ KOKKOSKERNELS_ADD_EXECUTABLE(
 KOKKOSKERNELS_ADD_EXECUTABLE(
   sparse_gs
   SOURCES KokkosSparse_gs.cpp
+  TESTONLYLIBS kokkoskernelsperf_gtest
   )
 
 KOKKOSKERNELS_ADD_EXECUTABLE(

src/common/KokkosKernels_Handle.hpp

@@ -636,9 +636,6 @@ class KokkosKernelsHandle
   }
   void destroy_gs_handle(){
     if (is_owner_of_the_gs_handle && this->gsHandle != NULL){
-      if (this->gsHandle->is_owner_of_coloring()){
-        this->destroy_graph_coloring_handle();
-      }
       delete this->gsHandle;
       this->gsHandle = NULL;
     }

src/sparse/KokkosSparse_gauss_seidel_handle.hpp

@@ -135,8 +135,6 @@ namespace KokkosSparse{
     //getters
     GSAlgorithm get_algorithm_type() const {return this->algorithm_type;}
 
-    virtual bool is_owner_of_coloring() const {return false;}
-
     nnz_lno_persistent_work_host_view_t get_color_xadj() const {
       return this->color_xadj;
     }
@@ -245,11 +243,18 @@ namespace KokkosSparse{
     scalar_persistent_work_view_t permuted_inverse_diagonal;
     nnz_lno_t block_size; //this is for block sgs
 
-    nnz_lno_t max_nnz_input_row;
-
     nnz_lno_t num_values_in_l1, num_values_in_l2, num_big_rows;
     size_t level_1_mem, level_2_mem;
-    bool owner_of_coloring;
+
+    //Option set by user: rows with at least this many nonzeros are handled by a separate kernel
+    nnz_lno_t long_row_threshold;
+    //Number of long rows per color set. They are all grouped at the end of each color set.
+    nnz_lno_persistent_work_host_view_t long_rows_per_color;
+    //Maximum row length in each color set.
+    nnz_lno_persistent_work_host_view_t max_row_length_per_color;
+    //Temporary space for matvec over long rows - size is only max num long rows in a color.
+    scalar_persistent_work_view_t long_row_x;
+
   public:
 
     /**
@@ -260,17 +265,13 @@ namespace KokkosSparse{
       permuted_xadj(), permuted_adj(), permuted_adj_vals(), old_to_new_map(),
       permuted_y_vector(), permuted_x_vector(),
       permuted_inverse_diagonal(), block_size(1),
-      max_nnz_input_row(-1),
       num_values_in_l1(-1), num_values_in_l2(-1),num_big_rows(0), level_1_mem(0), level_2_mem(0),
-      owner_of_coloring(false)
+      long_row_threshold(0)
     {
       if (gs == GS_DEFAULT)
         this->choose_default_algorithm();
     }
 
-    bool is_owner_of_coloring() const override {return this->owner_of_coloring;}
-    void set_owner_of_coloring(bool owner = true) {this->owner_of_coloring = owner;}
-
     void set_block_size(nnz_lno_t bs){this->block_size = bs; }
     nnz_lno_t get_block_size() const {return this->block_size;}
 
@@ -363,14 +364,44 @@ namespace KokkosSparse{
       return this->num_big_rows;
     }
 
-    nnz_lno_t get_max_nnz() const {
-      if(max_nnz_input_row == static_cast<nnz_lno_t>(-1))
-        throw std::runtime_error("Requested max nnz per input row, but this has not been set in the PointGS handle.");
-      return this->max_nnz_input_row;
+    nnz_lno_t get_long_row_threshold() const
+    {
+      return long_row_threshold;
     }
 
-    void set_max_nnz(nnz_lno_t num_result_nnz_) {
-      this->max_nnz_input_row = num_result_nnz_;
+    void set_long_row_threshold(nnz_lno_t lrt)
+    {
+      long_row_threshold = lrt;
+    }
+
+    nnz_lno_persistent_work_host_view_t get_long_rows_per_color() const
+    {
+      return long_rows_per_color;
+    }
+
+    void set_long_rows_per_color(const nnz_lno_persistent_work_host_view_t& long_rows_per_color_)
+    {
+      long_rows_per_color = long_rows_per_color_;
+    }
+
+    nnz_lno_persistent_work_host_view_t get_max_row_length_per_color() const
+    {
+      return max_row_length_per_color;
+    }
+
+    void set_max_row_length_per_color(const nnz_lno_persistent_work_host_view_t& max_row_length_per_color_)
+    {
+      max_row_length_per_color = max_row_length_per_color_;
+    }
+
+    scalar_persistent_work_view_t get_long_row_x() const
+    {
+      return long_row_x;
+    }
+
+    void set_long_row_x(const scalar_persistent_work_view_t& long_row_x_)
+    {
+      long_row_x = long_row_x_;
     }
 
     void allocate_x_y_vectors(nnz_lno_t num_rows, nnz_lno_t num_cols, nnz_lno_t num_vecs){