Merge pull request #454 from kubagalecki/develop

Refactor in preparation for arbitrary grid development

Author: llaniewski

src/LatticeAccess.inc.cpp.Rt renamed to src/CartLatticeAccess.hpp.Rt

@@ -2,14 +2,17 @@
 	source("conf.R")
         c_header()
 ?>
-/** \file LatticeAccess.cu
+/** \file CartLatticeAccess.hpp
 */
 
+#ifndef LATTICEACCESS_HPP
+#define LATTICEACCESS_HPP
+
 #include "Consts.h"
 #include "Global.h"
-#include "Lattice.h"
 #include <mpi.h>
 #include "range_int.hpp"
+#include "CartLatticeContainer.h"
 
 #ifndef STORAGE_BITS
   #define storage_to_real(x__) x__
@@ -34,12 +37,6 @@
   #define real_to_storage_shift(x__,y__) real_to_storage(x__)
 #endif
 
-/// Get only type of node
-CudaDeviceFunction flag_t LatticeContainer::getType(int x, int y, int z) const
-{
-  return NodeType[(x + nx*(y + ny*z))];
-}
-
 /// Push all densities
 
 <?R
@@ -253,18 +250,22 @@ resolve.symmetries = function(D) {
 
 ?>
 
+CudaDeviceFunction flag_t fetchNodeType(const CartLatticeContainer& container, int x, int y, int z) {
+  return container.NodeType[(x + container.nx*(y + container.ny*z))];
+}
+
 template < class x_t, class y_t, class z_t >
-class LatticeAccess {
+class CartLatticeAccess {
   private:
   const x_t x;
   const y_t y;
   const z_t z;
   const flag_t nt;
   const range_int<0,1,0,1> nx, ny, nz;
   public:
-  CudaDeviceFunction LatticeAccess(const int& x_, const int& y_, const int& z_) :
+  CudaDeviceFunction CartLatticeAccess(const int& x_, const int& y_, const int& z_) :
     x(x_),y(y_),z(z_),
-    nt(constContainer.getType(x,y,z)),
+    nt(fetchNodeType(constContainer, x, y, z)),
     nx(constContainer.nx),ny(constContainer.ny),nz(constContainer.nz)
     { }
   CudaDeviceFunction real_t getX() const { return constContainer.px + x; }
@@ -295,7 +296,7 @@ class LatticeAccess {
 <?R for (f in rows(Fields)) { ?>
 template < class x_t, class y_t, class z_t >
 template <class dx_t, class dy_t, class dz_t>
-CudaDeviceFunction real_t LatticeAccess< x_t, y_t, z_t >::load_<?%s f$nicename ?> (const dx_t & dx, const dy_t & dy, const dz_t & dz) const
+CudaDeviceFunction real_t CartLatticeAccess< x_t, y_t, z_t >::load_<?%s f$nicename ?> (const dx_t & dx, const dy_t & dy, const dz_t & dz) const
 {
   storage_t ret; <?R
   con = make.context("constContainer.in");
@@ -314,7 +315,7 @@ CudaDeviceFunction real_t LatticeAccess< x_t, y_t, z_t >::load_<?%s f$nicename ?
 for (s in rows(all_stages)) { ?>
 template < class x_t, class y_t, class z_t >
 template <class N>
-CudaDeviceFunction void LatticeAccess< x_t, y_t, z_t >::pop<?%s s$suffix ?>(N & node) const
+CudaDeviceFunction void CartLatticeAccess< x_t, y_t, z_t >::pop<?%s s$suffix ?>(N & node) const
 {
 	storage_t val; <?R
   con = make.context("constContainer.in",pocket=TRUE);
@@ -328,7 +329,7 @@ CudaDeviceFunction void LatticeAccess< x_t, y_t, z_t >::pop<?%s s$suffix ?>(N &
 
 template < class x_t, class y_t, class z_t >
 template <class N>
-CudaDeviceFunction void LatticeAccess< x_t, y_t, z_t >::push<?%s s$suffix ?>(N & node) const
+CudaDeviceFunction void CartLatticeAccess< x_t, y_t, z_t >::push<?%s s$suffix ?>(N & node) const
 {
   storage_t val; <?R
   con = make.context("constContainer.out",pocket=TRUE);
@@ -341,7 +342,7 @@ CudaDeviceFunction void LatticeAccess< x_t, y_t, z_t >::push<?%s s$suffix ?>(N &
 <?R if (ADJOINT) { ?>
 template < class x_t, class y_t, class z_t >
 template <class N>
-CudaDeviceFunction void LatticeAccess< x_t, y_t, z_t >::push<?%s s$suffix ?>_adj(N & node) const
+CudaDeviceFunction void CartLatticeAccess< x_t, y_t, z_t >::push<?%s s$suffix ?>_adj(N & node) const
 { <?R
   con = make.context("constContainer.adjout",pocket=TRUE);
   for (d in rows(Density)[s$load.densities]) {
@@ -354,7 +355,7 @@ CudaDeviceFunction void LatticeAccess< x_t, y_t, z_t >::push<?%s s$suffix ?>_adj
 
 template < class x_t, class y_t, class z_t >
 template <class N>
-CudaDeviceFunction void LatticeAccess< x_t, y_t, z_t >::pop<?%s s$suffix ?>_adj(N & node) const
+CudaDeviceFunction void CartLatticeAccess< x_t, y_t, z_t >::pop<?%s s$suffix ?>_adj(N & node) const
 { <?R
   con = make.context("constContainer.adjin",pocket=TRUE);
   for (f in rows(Fields)[s$save.fields]) {
@@ -478,3 +479,4 @@ public:
 template < class ACC >
 using AccessComposite = <?%s str ?>;
 
+#endif // LATTICEACCESS_HPP

src/CartLatticeContainer.h.Rt

@@ -0,0 +1,55 @@
+<?R
+	source("conf.R");
+	c_header()
+?>
+/** \file CartLatticeContainer.h
+  File defining CartLatticeContainer and some additional CUDA functions
+*/
+
+#ifndef LATTICECONTAINER_H
+#define LATTICECONTAINER_H
+
+#include "Consts.h"
+#include "Global.h"
+
+/// Main data structure for the lattice
+/**
+    Main data structure for storing solution (Snapshots)
+    in the lattice. It stores the internal part of the fields
+    and margins.
+*/
+struct FTabs { <?R
+for (m in NonEmptyMargin) { ?>
+  storage_t* <?%s m$name ?>; ///< Data on [<?%2d m$dx ?>,<?%2d m$dy?>,<?%2d m$dz?>] margin<?R
+} ?>
+  void Clear(int,int,int);
+  void PreAlloc(int,int,int);
+  void Free();
+};
+
+/// Container class used for storing all the data needed on GPU
+/**
+  Main class for storing the data needed on GPU and running kernels.
+  CartLatticeContainer is the container class which
+  owns all the data needed by the kernels on device
+  like pointers to the global gpu memory etc.
+  It defines push/pop functions allowing access
+  to node data stored in memory.
+*/
+class CartLatticeContainer {
+public:
+  FTabs in; ///< Main FTabs used for Primal iteration as input
+  FTabs out; ///< Main FTabs used for Primal iteration as output
+#ifdef ADJOINT
+  FTabs adjout; ///< FTabs used for Adjoint iteration as output
+  FTabs adjin; ///< FTabs used for Adjoint iteration as input
+#endif
+  flag_t * NodeType; ///< Table of flags/NodeTypes of all the nodes
+  cut_t* Q; 
+  int nx, ny, nz; ///< Size of the Lattice region
+  real_t px,py,pz;
+};
+
+static_assert(std::is_trivially_copyable<CartLatticeContainer>::value, "CartLatticeContainer must be trivially copyable");
+
+#endif // LATTICECONTAINER_H

src/CartLatticeExecutor.h.Rt

@@ -0,0 +1,49 @@
+<?R
+        source("conf.R")
+c_header();
+?>
+/** \file CartLatticeExecutor.h
+*/
+
+#ifndef CARTLATTICEEXECUTOR_H
+#define CARTLATTICEEXECUTOR_H
+
+#include "CartLatticeContainer.h"
+#include "LatticeData.hpp"
+
+struct CartLatticeExecutor {
+  CartLatticeContainer container;
+  LatticeData          data;
+
+  template < eOperationType I, eCalculateGlobals G, eStage S >
+  void RunBorder(CudaStream_t stream) const;
+  template < eOperationType I, eCalculateGlobals G, eStage S >
+  void RunInterior(CudaStream_t stream) const;
+  void Color(uchar4 *out_ptr);
+
+  void CopyToConst() const;
+
+  template < eOperationType I, eCalculateGlobals G, eStage S >
+  static CudaDeviceFunction void NodeToColor(int x, int y, int z, uchar4 *optr);
+};
+
+CudaGlobalFunction void ColorKernel(uchar4 *optr, int z);
+
+<?R
+for (q in rows(Quantities)) { ifdef(q$adjoint);
+        if (q$adjoint) {
+          node = "Node_Adj"
+        } else {
+          node = "Node"
+        }
+?>
+CudaGlobalFunction void get<?%s q$name ?>(lbRegion r, <?%s q$type ?> * tab, real_t scale); <?R
+  for (tp in c("float","double"))
+  { ?>
+CudaGlobalFunction void get<?%s q$name ?>_<?%s tp ?>(lbRegion , <?%s tp ?> *, int);
+<?R
+  }
+}
+ifdef() ?>
+
+#endif // CARTLATTICEEXECUTOR_H