Update Kokkos library in LAMMPS to v3.5.0

lib/kokkos/example/query_device/query_device.cpp

@@ -47,7 +47,8 @@
 
 #include <Kokkos_Macros.hpp>
 
-#if defined(KOKKOS_ENABLE_MPI)
+//#define USE_MPI
+#if defined(USE_MPI)
 #include <mpi.h>
 #endif
 
@@ -61,7 +62,7 @@ int main(int argc, char** argv) {
 
   (void)argc;
   (void)argv;
-#if defined(KOKKOS_ENABLE_MPI)
+#if defined(USE_MPI)
 
   MPI_Init(&argc, &argv);
 
@@ -72,7 +73,7 @@ int main(int argc, char** argv) {
   msg << "MPI rank(" << mpi_rank << ") ";
 
 #endif
-
+  Kokkos::initialize(argc, argv);
   msg << "{" << std::endl;
 
   if (Kokkos::hwloc::available()) {
@@ -82,15 +83,13 @@ int main(int argc, char** argv) {
         << std::endl;
   }
 
-#if defined(KOKKOS_ENABLE_CUDA)
-  Kokkos::Cuda::print_configuration(msg);
-#endif
+  Kokkos::print_configuration(msg);
 
   msg << "}" << std::endl;
 
   std::cout << msg.str();
-
-#if defined(KOKKOS_ENABLE_MPI)
+  Kokkos::finalize();
+#if defined(USE_MPI)
 
   MPI_Finalize();
 

lib/kokkos/example/tutorial/06_simple_mdrangepolicy/simple_mdrangepolicy.cpp

@@ -107,8 +107,8 @@ int main(int argc, char* argv[]) {
 
   // ViewType aliases for Rank<2>, Rank<3> for example usage
   using ScalarType  = double;
-  using ViewType_2D = typename Kokkos::View<ScalarType**>;
-  using ViewType_3D = typename Kokkos::View<ScalarType***>;
+  using ViewType_2D = Kokkos::View<ScalarType**>;
+  using ViewType_3D = Kokkos::View<ScalarType***>;
 
   /////////////////////////////////////////////////////////////////////////////
   // Explanation of MDRangePolicy usage, template parameters, constructor
@@ -160,8 +160,7 @@ int main(int argc, char* argv[]) {
   long incorrect_count_2d = 0;
   {
     // Rank<2> Case: Rank is provided, all other parameters are default
-    using MDPolicyType_2D = typename Kokkos::Experimental::MDRangePolicy<
-        Kokkos::Experimental::Rank<2> >;
+    using MDPolicyType_2D = Kokkos::MDRangePolicy<Kokkos::Rank<2> >;
 
     // Construct 2D MDRangePolicy: lower and upper bounds provided, tile dims
     // defaulted
@@ -185,9 +184,8 @@ int main(int argc, char* argv[]) {
   long incorrect_count_3d = 0;
   {
     // Rank<3> Case: Rank, inner iterate pattern, outer iterate pattern provided
-    using MDPolicyType_3D = typename Kokkos::Experimental::MDRangePolicy<
-        Kokkos::Experimental::Rank<3, Kokkos::Experimental::Iterate::Left,
-                                   Kokkos::Experimental::Iterate::Left> >;
+    using MDPolicyType_3D = Kokkos::MDRangePolicy<
+        Kokkos::Rank<3, Kokkos::Iterate::Left, Kokkos::Iterate::Left> >;
 
     // Construct 3D MDRangePolicy: lower, upper bounds, tile dims provided
     MDPolicyType_3D mdpolicy_3d({{0, 0, 0}}, {{n, n, n}}, {{4, 4, 4}});

lib/kokkos/example/tutorial/Advanced_Views/01_data_layouts/data_layouts.cpp

@@ -43,7 +43,7 @@
 */
 
 #include <Kokkos_Core.hpp>
-#include <impl/Kokkos_Timer.hpp>
+#include <Kokkos_Timer.hpp>
 #include <cstdio>
 
 // These two View types are both 2-D arrays of double.  However, they

lib/kokkos/example/tutorial/Advanced_Views/02_memory_traits/memory_traits.cpp

@@ -43,7 +43,7 @@
 */
 
 #include <Kokkos_Core.hpp>
-#include <impl/Kokkos_Timer.hpp>
+#include <Kokkos_Timer.hpp>
 #include <cstdio>
 #include <cstdlib>
 

lib/kokkos/example/tutorial/Advanced_Views/03_subviews/subviews.cpp

@@ -49,7 +49,7 @@
 // the mesh.
 
 #include <Kokkos_Core.hpp>
-#include <impl/Kokkos_Timer.hpp>
+#include <Kokkos_Timer.hpp>
 #include <cstdio>
 
 using mesh_type = Kokkos::View<double***, Kokkos::LayoutRight>;

lib/kokkos/example/tutorial/Advanced_Views/04_dualviews/dual_view.cpp

@@ -44,7 +44,7 @@
 
 #include <Kokkos_Core.hpp>
 #include <Kokkos_DualView.hpp>
-#include <impl/Kokkos_Timer.hpp>
+#include <Kokkos_Timer.hpp>
 #include <cstdio>
 #include <cstdlib>
 

lib/kokkos/example/tutorial/Advanced_Views/05_NVIDIA_UVM/uvm_example.cpp

@@ -44,7 +44,7 @@
 
 #include <Kokkos_Core.hpp>
 #include <Kokkos_DualView.hpp>
-#include <impl/Kokkos_Timer.hpp>
+#include <Kokkos_Timer.hpp>
 #include <cstdio>
 #include <cstdlib>
 

lib/kokkos/example/tutorial/Advanced_Views/07_Overlapping_DeepCopy/overlapping_deepcopy.cpp

@@ -46,7 +46,7 @@
 #include <cstdio>
 #include <typeinfo>
 #include <cmath>
-#include <impl/Kokkos_Timer.hpp>
+#include <Kokkos_Timer.hpp>
 
 struct FillDevice {
   double value;

lib/kokkos/example/tutorial/Algorithms/01_random_numbers/random_numbers.cpp

@@ -45,7 +45,7 @@
 #include <Kokkos_Core.hpp>
 #include <Kokkos_Random.hpp>
 #include <Kokkos_DualView.hpp>
-#include <impl/Kokkos_Timer.hpp>
+#include <Kokkos_Timer.hpp>
 #include <cstdlib>
 
 using DefaultHostType = Kokkos::HostSpace::execution_space;
@@ -74,7 +74,7 @@ using DefaultHostType = Kokkos::HostSpace::execution_space;
 template <class GeneratorPool>
 struct generate_random {
   // Output View for the random numbers
-  Kokkos::View<uint64_t*> vals;
+  Kokkos::View<uint64_t**> vals;
 
   // The GeneratorPool
   GeneratorPool rand_pool;
@@ -82,7 +82,7 @@ struct generate_random {
   int samples;
 
   // Initialize all members
-  generate_random(Kokkos::View<uint64_t*> vals_, GeneratorPool rand_pool_,
+  generate_random(Kokkos::View<uint64_t**> vals_, GeneratorPool rand_pool_,
                   int samples_)
       : vals(vals_), rand_pool(rand_pool_), samples(samples_) {}
 
@@ -94,8 +94,7 @@ struct generate_random {
     // Draw samples numbers from the pool as urand64 between 0 and
     // rand_pool.MAX_URAND64 Note there are function calls to get other type of
     // scalars, and also to specify Ranges or get a normal distributed float.
-    for (int k = 0; k < samples; k++)
-      vals(i * samples + k) = rand_gen.urand64();
+    for (int k = 0; k < samples; k++) vals(i, k) = rand_gen.urand64();
 
     // Give the state back, which will allow another thread to acquire it
     rand_pool.free_state(rand_gen);
@@ -103,11 +102,11 @@ struct generate_random {
 };
 
 int main(int argc, char* args[]) {
+  Kokkos::initialize(argc, args);
   if (argc != 3) {
     printf("Please pass two integers on the command line\n");
   } else {
     // Initialize Kokkos
-    Kokkos::initialize(argc, args);
     int size    = std::stoi(args[1]);
     int samples = std::stoi(args[2]);
 
@@ -117,7 +116,7 @@ int main(int argc, char* args[]) {
     // pool.
     Kokkos::Random_XorShift64_Pool<> rand_pool64(5374857);
     Kokkos::Random_XorShift1024_Pool<> rand_pool1024(5374857);
-    Kokkos::DualView<uint64_t*> vals("Vals", size * samples);
+    Kokkos::DualView<uint64_t**> vals("Vals", size, samples);
 
     // Run some performance comparisons
     Kokkos::Timer timer;
@@ -151,8 +150,7 @@ int main(int argc, char* args[]) {
            1.0e-9 * samples * size / time_1024);
 
     Kokkos::deep_copy(vals.h_view, vals.d_view);
-
-    Kokkos::finalize();
   }
+  Kokkos::finalize();
   return 0;
 }

lib/kokkos/example/tutorial/Hierarchical_Parallelism/04_team_scan/team_scan.cpp

@@ -44,7 +44,7 @@
 
 #include <Kokkos_Core.hpp>
 #include <Kokkos_DualView.hpp>
-#include <impl/Kokkos_Timer.hpp>
+#include <Kokkos_Timer.hpp>
 #include <cstdio>
 #include <cstdlib>