Optimizing Cell-list construction in case of reordering

src/NN/CellList/CellList_util.hpp

@@ -11,6 +11,12 @@
 #define CL_SYMMETRIC 1
 #define CL_NON_SYMMETRIC 2
 
+enum cl_construct_opt
+{
+	Full,
+	Only_reorder
+};
+
 #if defined(CUDA_GPU) && defined(__NVCC__)
 #include "util/cuda/moderngpu/kernel_mergesort.hxx"
 #endif
@@ -58,7 +64,8 @@ struct populate_cell_list_no_sym_impl
 						   openfpm::vector<prop,Memory,typename layout_base<prop>::type,layout_base > & v_prp_out,
 			   	   	   	   CellList & cli,
 			   	   	   	   mgpu::ofp_context_t & context,
-			   	   	   	   size_t g_m)
+			   	   	   	   size_t g_m,
+			   	   	   	   cl_construct_opt optc)
 	{
 		cli.clear();
 
@@ -79,13 +86,13 @@ struct populate_cell_list_no_sym_impl<true>
 						 openfpm::vector<prop,Memory,typename layout_base<prop>::type,layout_base > & v_prp_out,
 			   	   	   	   CellList & cli,
 			   	   	   	   mgpu::ofp_context_t & context,
-			   	   	   	   size_t g_m)
+			   	   	   	   size_t g_m,
+			   	   	   	   cl_construct_opt optc)
 	{
 		v_prp_out.resize(pos.size());
 		v_pos_out.resize(pos.size());
 
-		cli.template construct<decltype(pos),decltype(v_prp)>(pos,v_pos_out,v_prp,v_prp_out,context,g_m);
-
+		cli.template construct<decltype(pos),decltype(v_prp)>(pos,v_pos_out,v_prp,v_prp_out,context,g_m,optc);
 	}
 };
 
@@ -141,9 +148,10 @@ void populate_cell_list_no_sym(openfpm::vector<Point<dim,T>,Memory,typename layo
 		 	 	 	 	 	   openfpm::vector<prop,Memory,typename layout_base<prop>::type,layout_base > & v_prp_out,
 							   CellList & cli,
 							   mgpu::ofp_context_t & mgpu,
-							   size_t g_m)
+							   size_t g_m,
+							   cl_construct_opt optc)
 {
-	populate_cell_list_no_sym_impl<is_gpu_celllist<CellList>::value>::populate(pos,v_pos_out,v_prp,v_prp_out,cli,mgpu,g_m);
+	populate_cell_list_no_sym_impl<is_gpu_celllist<CellList>::value>::populate(pos,v_pos_out,v_prp,v_prp_out,cli,mgpu,g_m,optc);
 }
 
 /*! \brief populate the Cell-list with particles symmetric case
@@ -185,10 +193,11 @@ void populate_cell_list(openfpm::vector<Point<dim,T>,Memory,typename layout_base
 						CellList & cli,
 						mgpu::ofp_context_t & context,
 						size_t g_m,
-						size_t opt)
+						size_t opt,
+						cl_construct_opt optc)
 {
 	if (opt == CL_NON_SYMMETRIC)
-	{populate_cell_list_no_sym(pos,v_pos_out,v_prp,v_prp_out,cli,context,g_m);}
+	{populate_cell_list_no_sym(pos,v_pos_out,v_prp,v_prp_out,cli,context,g_m,optc);}
 	else
 	{populate_cell_list_sym(pos,cli,g_m);}
 }
@@ -212,7 +221,8 @@ void populate_cell_list(openfpm::vector<Point<dim,T>,Memory,typename layout_base
 						CellList & cli,
 						mgpu::ofp_context_t & context,
 						size_t g_m,
-						size_t opt)
+						size_t opt,
+						cl_construct_opt optc)
 {
 	typedef openfpm::vector<aggregate<int>,Memory,typename layout_base<aggregate<int>>::type,layout_base> stub_prop_type;
 
@@ -221,7 +231,7 @@ void populate_cell_list(openfpm::vector<Point<dim,T>,Memory,typename layout_base
 
 	openfpm::vector<Point<dim,T>,Memory,typename layout_base<Point<dim,T>>::type,layout_base> stub3;
 
-	populate_cell_list(pos,stub3,stub1,stub2,cli,context,g_m,opt);
+	populate_cell_list(pos,stub3,stub1,stub2,cli,context,g_m,opt,optc);
 }
 
 /*! \brief Structure that contain a reference to a vector of particles

src/NN/CellList/cuda/CellList_gpu.hpp

@@ -27,7 +27,6 @@
 constexpr int count = 0;
 constexpr int start = 1;
 
-
 template<unsigned int dim, typename T,  typename Memory, typename transform = no_transform_only<dim,T>, typename cnt_type = unsigned int, typename ids_type = int>
 class CellList_gpu : public CellDecomposer_sm<dim,T,transform>
 {
@@ -225,7 +224,14 @@ public:
 	 * \param pl Particles list
 	 *
 	 */
-	template<typename vector, typename vector_prp> void construct(vector & pl, vector & pl_out, vector_prp & pl_prp, vector_prp & pl_prp_out, mgpu::ofp_context_t & mgpuContext, size_t g_m = 0)
+	template<typename vector, typename vector_prp>
+	void construct(vector & pl,
+				   vector & pl_out,
+				   vector_prp & pl_prp,
+				   vector_prp & pl_prp_out,
+				   mgpu::ofp_context_t & mgpuContext,
+				   size_t g_m = 0,
+				   cl_construct_opt opt = cl_construct_opt::Full)
 	{
 #ifdef __NVCC__
 
@@ -277,7 +283,7 @@ public:
 
 		auto ite = pl.getGPUIterator();
 
-		// Here we test fill cell
+		// Here we reorder the particles to improve coalescing access
 		CUDA_LAUNCH((reorder_parts<decltype(pl_prp.toKernel()),
 				      decltype(pl.toKernel()),
 				      decltype(sorted_to_not_sorted.toKernel()),
@@ -291,6 +297,8 @@ public:
 				                                                           static_cast<cnt_type *>(cells.template getDeviceBuffer<0>()));
 
 
+		if (opt == cl_construct_opt::Full)
+		{
 			ite = sorted_domain_particles_ids.getGPUIterator();
 
 			CUDA_LAUNCH((mark_domain_particles),ite.wthr,ite.thr,sorted_to_not_sorted.toKernel(),sorted_domain_particles_ids.toKernel(),sorted_domain_particles_dg.toKernel(),g_m);
@@ -299,12 +307,13 @@ public:
 			// now we sort the particles
 			mergesort((int *)sorted_domain_particles_dg.template getDeviceBuffer<0>(),(int *)sorted_domain_particles_ids.template getDeviceBuffer<0>(),
 							 sorted_domain_particles_dg.size(), mgpu::template less_t<int>(), mgpuContext);
+		}
 
-#else
+	#else
 
 			std::cout << "Error: " <<  __FILE__ << ":" << __LINE__ << " you are calling CellList_gpu.construct() this function is suppose must be compiled with NVCC compiler, but it look like has been compiled by the standard system compiler" << std::endl;
 
-#endif
+	#endif
 
 	}
 

src/NN/VerletList/VerletListFast.hpp

@@ -324,9 +324,9 @@ private:
 	{
 		mgpu::ofp_context_t context(mgpu::gpu_context_opt::dummy);
 		if (opt & VL_SYMMETRIC || opt & VL_CRS_SYMMETRIC)
-		{populate_cell_list(pos,cli,context,g_m,CL_SYMMETRIC);}
+		{populate_cell_list(pos,cli,context,g_m,CL_SYMMETRIC,cl_construct_opt::Full);}
 		else
-		{populate_cell_list(pos,cli,context,g_m,CL_NON_SYMMETRIC);}
+		{populate_cell_list(pos,cli,context,g_m,CL_NON_SYMMETRIC,cl_construct_opt::Full);}
 	}
 
 	/*! \brief Create the Verlet list from a given cell-list