perf(kernels): split move_kernel with precomputed randon numbers

Use of 2D random number buffer for each particle (layout left)

Total time move kernel (move+random for new )
|   |OMP  | MPI | CUDA |
| --| --  | --- |--    |
old |  137.069686   |  119.201964   |  136.673589   |
new |  103.405564+14.312186   |  98.172346 +13.936711   |  52.763011+4|

Author: Benncs

apps/libs/mc/public/mc/alias.hpp

@@ -4,6 +4,7 @@
 #include <Kokkos_Core_fwd.hpp>
 #include <Kokkos_ScatterView.hpp>
 #include <common/traits.hpp>
+#include <traits/Kokkos_IterationPatternTrait.hpp>
 
 // static_assert(FloatingPointType<Kokkos::Experimental::half_t>,
 //               "Kokkos half_t ok");
@@ -65,7 +66,7 @@ namespace MC
   using ParticleAges = ParticleAgesBase<ComputeSpace>;
 
   using ParticleSamples = Kokkos::
-      View<Kokkos::Experimental::half_t**, Kokkos::LayoutLeft, ComputeSpace>;
+      View<Kokkos::Experimental::half_t**, Kokkos::LayoutRight, ComputeSpace>;
 
   template <typename MemorySpace>
   using ParticlePropertyViewType =

apps/libs/simulation/public/simulation/kernels/kernels.hpp

@@ -22,6 +22,8 @@ namespace Simulation::KernelInline
     using cycle_kernel_type = CycleFunctor<Model>;
     using move_kernel_type = MoveFunctor;
 
+    ComputeSpace move_space{};
+
     cycle_reducer_view_type<Space> cycle_reducer;
     move_reducer_view_type<Space> move_reducer;
     cycle_kernel_type cycle_kernel;
@@ -50,7 +52,8 @@ namespace Simulation::KernelInline
       cycle_kernel.update(d_t, container);
 
       // TODO: Why need to update all views (where did we lost the refcount ? )
-      move_kernel.update(d_t,
+      move_kernel.update(move_space,
+                         d_t,
                          container.n_particles(),
                          std::move(new_move),
                          container.position,

apps/libs/simulation/public/simulation/kernels/move_kernel.hpp

@@ -2,6 +2,7 @@
 #define __SIMULATION_MOVE_KERNEL_HPP__
 
 #include "Kokkos_Macros.hpp"
+#include "common/common.hpp"
 #include <Kokkos_Assert.hpp>
 #include <Kokkos_Core.hpp>
 #include <Kokkos_Printf.hpp>
@@ -67,8 +68,7 @@ namespace Simulation::KernelInline
       const std::size_t i_compartment,
       const double random_number)
   {
-    const int mask_do_serch =
-        static_cast<int>(do_serch); // Image this is from another calculation
+    const int mask_do_serch = static_cast<int>(do_serch);
 
     const int max_neighbor = static_cast<int>(neighbors.extent(1));
 
@@ -172,7 +172,8 @@ namespace Simulation::KernelInline
           random(std::move(_random)), enable_move(b_move),
           enable_leave(b_leave) {};
 
-    void update(double _d_t,
+    void update(const ComputeSpace& ex,
+                double _d_t,
                 std::size_t n_p,
                 MoveInfo<ComputeSpace>&& move_i,
                 MC::ParticlePositions _positions,
@@ -194,7 +195,7 @@ namespace Simulation::KernelInline
       this->status = std::move(_status);
       this->ages = std::move(_ages);
 
-      Kokkos::fill_random(random, random_pool, 0., 1.);
+      Kokkos::fill_random(ex, random, random_pool, 0., 1.);
     }
 
     KOKKOS_INLINE_FUNCTION void operator()(

apps/libs/simulation/public/simulation/simulation.hpp

@@ -213,13 +213,16 @@ namespace Simulation
     {
       return;
     }
+
+    pre_cycle(container, d_t, cycle_functors);
+
     if (f_reaction)
     {
       this->contribs_scatter.reset();
       cycle_functors.launch_model(n_particle);
     }
-    pre_cycle(container, d_t, cycle_functors);
 
+    cycle_functors.move_space.fence();
     if (cycle_functors.move_kernel.need_launch())
     {
 

tools/cases.xml

@@ -10,8 +10,8 @@
       <number_particle>5000</number_particle>
       <delta_time>0.001</delta_time>
       <results_file_name>debug</results_file_name>
-      <number_exported_result>10</number_exported_result>
-      <model_name>monod</model_name>
+      <number_exported_result>0</number_exported_result>
+      <model_name>none</model_name>
       <!--<initialiser_path>./cma_data/0d_4s_init.h5</initialiser_path> -->
       <!-- <initialiser_path>./cma_data/n14_init.h5</initialiser_path> -->
        <initialiser_path>./cma_data/sanofi_init.h5</initialiser_path>

tools/runner.py

@@ -41,7 +41,7 @@ def get_executable(instal: str, mpi: bool = True):
     if instal:
         return f"/opt/biomc/{appname}"
     else:
-        return f"{ROOT}/builddir/host/apps/cli/{appname}"
+        return f"{ROOT}/builddir/gpu/apps/cli/{appname}"
 
 
 def mk_parser():