Combination of Jacobi and gauss-siedal
authoryanhua <yanhua@yanhua-Precision-T1600.(none)>
Tue, 9 Oct 2012 05:09:18 +0000 (00:09 -0500)
committeryanhua <yanhua@yanhua-Precision-T1600.(none)>
Tue, 9 Oct 2012 05:09:18 +0000 (00:09 -0500)
tests/charm++/jacobi3d-gausssiedal/Makefile [new file with mode: 0644]
tests/charm++/jacobi3d-gausssiedal/jacobi3d.C [new file with mode: 0644]
tests/charm++/jacobi3d-gausssiedal/jacobi3d.ci [new file with mode: 0644]

diff --git a/tests/charm++/jacobi3d-gausssiedal/Makefile b/tests/charm++/jacobi3d-gausssiedal/Makefile
new file mode 100644 (file)
index 0000000..61e38c3
--- /dev/null
@@ -0,0 +1,43 @@
+CHARMC=../../../bin/charmc $(OPTS) $(MOPTS)
+
+OBJS = jacobi3d.o
+
+OBJS_OMP = jacobi3d_omp.o
+
+OMP_FLAG = -DJACOBI_OPENMP -fopenmp
+
+all: jacobi3d
+
+jacobi3d: $(OBJS)
+       $(CHARMC) -language charm++ -module DummyLB -o jacobi3d $(OBJS)
+#      $(CHARMC) -language charm++ -module EveryLB -memory paranoid -o jacobi3d $(OBJS)
+
+jacobi3d_omp: $(OBJS_OMP)
+       $(CHARMC) -language charm++ -module DummyLB $(OMP_FLAG) -o jacobi3d_omp $(OBJS_OMP)
+
+projections: $(OBJS)
+       $(CHARMC) -language charm++ -tracemode projections -lz -o jacobi3d.prj $(OBJS)
+
+summary: $(OBJS)
+       $(CHARMC) -language charm++ -tracemode summary -lz -o jacobi3d.sum $(OBJS)
+
+jacobi3d.decl.h: jacobi3d.ci
+       $(CHARMC)  jacobi3d.ci
+
+syncfttest causalfttest: fttest
+fttest: jacobi3d
+       ./charmrun ./jacobi3d 256 128 +p8 +balancer DummyLB $(TESTOPTS)
+       ./charmrun ./jacobi3d 256 128 +p8 +balancer DummyLB +killFile kill_01.txt $(TESTOPTS)
+       ./charmrun ./jacobi3d 256 256 256 64 64 32 +p8 +balancer DummyLB +killFile kill_02.txt $(TESTOPTS)
+       ./charmrun ./jacobi3d 256 256 256 64 64 32 +p8 +balancer DummyLB +killFile kill_03.txt $(TESTOPTS)
+
+
+clean:
+       rm -f *.decl.h *.def.h conv-host *.o jacobi3d jacobi3d.prj charmrun *~
+
+jacobi3d.o: jacobi3d.C jacobi3d.decl.h
+       $(CHARMC) -c jacobi3d.C
+
+jacobi3d_omp.o: jacobi3d.C jacobi3d.decl.h
+       $(CHARMC) $(OMP_FLAG) -o $(OBJS_OMP) -c jacobi3d.C
+
diff --git a/tests/charm++/jacobi3d-gausssiedal/jacobi3d.C b/tests/charm++/jacobi3d-gausssiedal/jacobi3d.C
new file mode 100644 (file)
index 0000000..4ed659b
--- /dev/null
@@ -0,0 +1,583 @@
+/*****************************************************************************
+ * $Source$
+ * $Author$
+ * $Date$
+ * $Revision$
+ *****************************************************************************/
+
+/** \file jacobi3d.C
+ *  Author: Abhinav S Bhatele
+ *  Date Created: June 01st, 2009
+ *
+ *  This does a topological placement for a 3d jacobi.
+ *
+ *     
+ *           *****************
+ *        *               *  *
+ *   ^ *****************     *
+ *   | *               *     *
+ *   | *               *     *
+ *   | *               *     *
+ *   Y *               *     *
+ *   | *               *     *
+ *   | *               *     *
+ *   | *               *  * 
+ *   ~ *****************    Z
+ *     <------ X ------> 
+ *
+ *   X: left, right --> wrap_x
+ *   Y: top, bottom --> wrap_y
+ *   Z: front, back --> wrap_z
+ */
+
+#include "jacobi3d.decl.h"
+#include "TopoManager.h"
+#ifdef JACOBI_OPENMP
+#include <omp.h>
+#endif
+
+// See README for documentation
+
+/*readonly*/ CProxy_Main mainProxy;
+/*readonly*/ int arrayDimX;
+/*readonly*/ int arrayDimY;
+/*readonly*/ int arrayDimZ;
+/*readonly*/ int blockDimX;
+/*readonly*/ int blockDimY;
+/*readonly*/ int blockDimZ;
+
+// specify the number of worker chares in each dimension
+/*readonly*/ int num_chare_x;
+/*readonly*/ int num_chare_y;
+/*readonly*/ int num_chare_z;
+
+/*readonly*/ int globalBarrier;
+
+static unsigned long next = 1;
+
+int myrand(int numpes) {
+  next = next * 1103515245 + 12345;
+  return((unsigned)(next/65536) % numpes);
+}
+
+// We want to wrap entries around, and because mod operator % 
+// sometimes misbehaves on negative values. -1 maps to the highest value.
+#define wrap_x(a)      (((a)+num_chare_x)%num_chare_x)
+#define wrap_y(a)      (((a)+num_chare_y)%num_chare_y)
+#define wrap_z(a)      (((a)+num_chare_z)%num_chare_z)
+
+#define index(a, b, c) ( (a)*(blockDimY+2)*(blockDimZ+2) + (b)*(blockDimZ+2) + (c) )
+
+#define PRINT_FREQ      100
+#define CKP_FREQ               100
+#define MAX_ITER        400    
+#define WARM_ITER              5
+#define LEFT                   1
+#define RIGHT                  2
+#define TOP                    3
+#define BOTTOM                 4
+#define FRONT                  5
+#define BACK                   6
+#define DIVIDEBY7              0.14285714285714285714
+
+double startTime;
+double endTime;
+
+/** \class ghostMsg
+ *
+ */
+class ghostMsg: public CMessage_ghostMsg {
+  public:
+    int dir;
+    int height;
+    int width;
+    double* gh;
+
+    ghostMsg(int _d, int _h, int _w) : dir(_d), height(_h), width(_w) {
+    }
+};
+
+/** \class Main
+ *
+ */
+class Main : public CBase_Main {
+  public:
+    CProxy_Jacobi array;
+    int iterations;
+
+    Main(CkArgMsg* m) {
+      if ( (m->argc != 3) && (m->argc != 7) ) {
+        CkPrintf("%s [array_size] [block_size]\n", m->argv[0]);
+        CkPrintf("OR %s [array_size_X] [array_size_Y] [array_size_Z] [block_size_X] [block_size_Y] [block_size_Z]\n", m->argv[0]);
+        CkAbort("Abort");
+      }
+
+      // set iteration counter to zero
+      iterations = 0;
+
+      // store the main proxy
+      mainProxy = thisProxy;
+       
+      if(m->argc == 3) {
+       arrayDimX = arrayDimY = arrayDimZ = atoi(m->argv[1]);
+        blockDimX = blockDimY = blockDimZ = atoi(m->argv[2]); 
+      }
+      else if (m->argc == 7) {
+        arrayDimX = atoi(m->argv[1]);
+       arrayDimY = atoi(m->argv[2]);
+       arrayDimZ = atoi(m->argv[3]);
+        blockDimX = atoi(m->argv[4]); 
+       blockDimY = atoi(m->argv[5]); 
+       blockDimZ = atoi(m->argv[6]);
+      }
+
+      if (arrayDimX < blockDimX || arrayDimX % blockDimX != 0)
+        CkAbort("array_size_X % block_size_X != 0!");
+      if (arrayDimY < blockDimY || arrayDimY % blockDimY != 0)
+        CkAbort("array_size_Y % block_size_Y != 0!");
+      if (arrayDimZ < blockDimZ || arrayDimZ % blockDimZ != 0)
+        CkAbort("array_size_Z % block_size_Z != 0!");
+
+      num_chare_x = arrayDimX / blockDimX;
+      num_chare_y = arrayDimY / blockDimY;
+      num_chare_z = arrayDimZ / blockDimZ;
+
+      // print info
+      CkPrintf("\nSTENCIL COMPUTATION WITH NO BARRIERS\n");
+      CkPrintf("Running Jacobi on %d processors with (%d, %d, %d) chares\n", CkNumPes(), num_chare_x, num_chare_y, num_chare_z);
+      CkPrintf("Array Dimensions: %d %d %d\n", arrayDimX, arrayDimY, arrayDimZ);
+      CkPrintf("Block Dimensions: %d %d %d\n", blockDimX, blockDimY, blockDimZ);
+
+      // Create new array of worker chares
+#if USE_TOPOMAP
+      CProxy_JacobiMap map = CProxy_JacobiMap::ckNew(num_chare_x, num_chare_y, num_chare_z);
+      CkPrintf("Topology Mapping is being done ... \n");
+      CkArrayOptions opts(num_chare_x, num_chare_y, num_chare_z);
+      opts.setMap(map);
+      array = CProxy_Jacobi::ckNew(opts);
+#else
+      array = CProxy_Jacobi::ckNew(num_chare_x, num_chare_y, num_chare_z);
+#endif
+
+      TopoManager tmgr;
+      CkArray *jarr = array.ckLocalBranch();
+      int jmap[num_chare_x][num_chare_y][num_chare_z];
+
+      int hops=0, p;
+      for(int i=0; i<num_chare_x; i++)
+       for(int j=0; j<num_chare_y; j++)
+         for(int k=0; k<num_chare_z; k++) {
+           jmap[i][j][k] = jarr->procNum(CkArrayIndex3D(i, j, k));
+         }
+
+      for(int i=0; i<num_chare_x; i++)
+       for(int j=0; j<num_chare_y; j++)
+         for(int k=0; k<num_chare_z; k++) {
+           p = jmap[i][j][k];
+           hops += tmgr.getHopsBetweenRanks(p, jmap[wrap_x(i+1)][j][k]);
+           hops += tmgr.getHopsBetweenRanks(p, jmap[wrap_x(i-1)][j][k]);
+           hops += tmgr.getHopsBetweenRanks(p, jmap[i][wrap_y(j+1)][k]);
+           hops += tmgr.getHopsBetweenRanks(p, jmap[i][wrap_y(j-1)][k]);
+           hops += tmgr.getHopsBetweenRanks(p, jmap[i][j][wrap_z(k+1)]);
+           hops += tmgr.getHopsBetweenRanks(p, jmap[i][j][wrap_z(k-1)]);
+         }
+      CkPrintf("Total Hops: %d\n", hops);
+       
+#ifdef JACOBI_OPENMP 
+      CProxy_OmpInitializer ompInit = 
+        CProxy_OmpInitializer::ckNew(4);
+#else    
+      //Start the computation
+      start();
+#endif
+    }
+
+    void start() {
+      startTime = CmiWallTimer();                
+      array.doStep();
+    }
+
+    // Each worker reports back to here when it completes an iteration
+       void report() {
+        iterations += CKP_FREQ;
+        if (iterations < MAX_ITER) {
+#ifdef CMK_MEM_CHECKPOINT
+            CkCallback cb (CkIndex_Jacobi::doStep(), array);
+            CkStartMemCheckpoint(cb);
+#else
+                       array.doStep();
+#endif
+        } else {
+            CkPrintf("Completed %d iterations\n", MAX_ITER-1);
+            endTime = CmiWallTimer();
+            CkPrintf("Time elapsed per iteration: %f\n", (endTime - startTime)/(MAX_ITER-1-WARM_ITER));
+            CkExit();
+        }
+    }
+
+};
+
+/** \class Jacobi
+ *
+ */
+
+class Jacobi: public CBase_Jacobi {
+  Jacobi_SDAG_CODE
+
+  public:
+    int iterations;
+    int imsg;
+
+    double *temperature;
+       double timing,average;
+
+    // Constructor, initialize values
+    Jacobi() {
+      // This call is an anachronism - the need to call __sdag_init() has been
+      // removed. We still call it here to test backward compatibility.
+      __sdag_init();
+
+      int i, j, k;
+      // allocate a three dimensional array
+      temperature = new double[(blockDimX+2) * (blockDimY+2) * (blockDimZ+2)];
+
+      for(i=0; i<blockDimX+2; ++i) {
+          for(j=0; j<blockDimY+2; ++j) {
+              for(k=0; k<blockDimZ+2; ++k) {
+                  temperature[index(i, j, k)] = 0.0;
+              }
+          } 
+      }
+
+      iterations = 0;
+      imsg = 0;
+      constrainBC();
+
+      usesAtSync = CmiTrue;
+
+    }
+
+    Jacobi(CkMigrateMessage* m): CBase_Jacobi(m) { }
+
+    ~Jacobi() { 
+        delete [] temperature; 
+    }
+
+       // Pupping function for migration and fault tolerance
+       // Condition: assuming the 3D Chare Arrays are NOT used
+       void pup(PUP::er &p){
+               
+               // calling parent's pup
+               CBase_Jacobi::pup(p);
+       
+               __sdag_pup(p);
+       
+               // pupping properties of this class
+               p | iterations;
+               p | imsg;
+
+               // if unpacking, allocate the memory space
+               if(p.isUnpacking()){
+                       temperature = new double[(blockDimX+2) * (blockDimY+2) * (blockDimZ+2)];
+       
+               }
+
+               // pupping the arrays
+               p((char *)temperature, (blockDimX+2) * (blockDimY+2) * (blockDimZ+2) * sizeof(double));
+
+       }
+
+
+    // Send ghost faces to the six neighbors
+    void begin_iteration(void) {
+      if (thisIndex.x == 0 && thisIndex.y == 0 && thisIndex.z == 0) {
+//          CkPrintf("Start of iteration %d\n", iterations);
+                       if(iterations % PRINT_FREQ == 0){
+                               average = timing;
+                               timing = CmiWallTimer();
+                               average = (timing - average)/(double)PRINT_FREQ;
+                               CkPrintf("time=%.2f it=%d avg=%.4f\n",timing,iterations,average);
+                       }
+      }
+      iterations++;
+
+      // Copy different faces into messages
+      ghostMsg *leftMsg = new (blockDimY*blockDimZ) ghostMsg(RIGHT, blockDimY, blockDimZ);
+      ghostMsg *rightMsg = new (blockDimY*blockDimZ) ghostMsg(LEFT, blockDimY, blockDimZ);
+      ghostMsg *topMsg = new (blockDimX*blockDimZ) ghostMsg(BOTTOM, blockDimX, blockDimZ);
+      ghostMsg *bottomMsg = new (blockDimX*blockDimZ) ghostMsg(TOP, blockDimX, blockDimZ);
+      ghostMsg *frontMsg = new (blockDimX*blockDimY) ghostMsg(BACK, blockDimX, blockDimY);
+      ghostMsg *backMsg = new (blockDimX*blockDimY) ghostMsg(FRONT, blockDimX, blockDimY);
+
+      CkSetRefNum(leftMsg, iterations);
+      CkSetRefNum(rightMsg, iterations);
+      CkSetRefNum(topMsg, iterations);
+      CkSetRefNum(bottomMsg, iterations);
+      CkSetRefNum(frontMsg, iterations);
+      CkSetRefNum(backMsg, iterations);
+
+      for(int j=0; j<blockDimY; ++j) 
+          for(int k=0; k<blockDimZ; ++k) {
+              leftMsg->gh[k*blockDimY+j] = temperature[index(1, j+1, k+1)];
+              rightMsg->gh[k*blockDimY+j] = temperature[index(blockDimX, j+1, k+1)];
+          }
+
+      for(int i=0; i<blockDimX; ++i) 
+          for(int k=0; k<blockDimZ; ++k) {
+              topMsg->gh[k*blockDimX+i] = temperature[index(i+1, 1, k+1)];
+              bottomMsg->gh[k*blockDimX+i] = temperature[index(i+1, blockDimY, k+1)];
+          }
+
+      for(int i=0; i<blockDimX; ++i) 
+          for(int j=0; j<blockDimY; ++j) {
+              frontMsg->gh[j*blockDimX+i] = temperature[index(i+1, j+1, 1)];
+              backMsg->gh[j*blockDimX+i] = temperature[index(i+1, j+1, blockDimZ)];
+          }
+
+      // Send my left face
+       thisProxy(wrap_x(thisIndex.x-1), thisIndex.y, thisIndex.z).receiveGhosts(leftMsg);
+      // Send my right face
+      thisProxy(wrap_x(thisIndex.x+1), thisIndex.y, thisIndex.z).receiveGhosts(rightMsg);
+      // Send my top face
+      thisProxy(thisIndex.x, wrap_y(thisIndex.y-1), thisIndex.z).receiveGhosts(topMsg);
+      // Send my bottom face
+      thisProxy(thisIndex.x, wrap_y(thisIndex.y+1), thisIndex.z).receiveGhosts(bottomMsg);
+      // Send my front face
+      thisProxy(thisIndex.x, thisIndex.y, wrap_z(thisIndex.z-1)).receiveGhosts(frontMsg);
+      // Send my back face
+      thisProxy(thisIndex.x, thisIndex.y, wrap_z(thisIndex.z+1)).receiveGhosts(backMsg);
+    }
+
+    void processGhosts(ghostMsg *gmsg) {
+      int height = gmsg->height;
+      int width = gmsg->width;
+
+      switch(gmsg->dir) {
+      case LEFT:
+          for(int j=0; j<height; ++j) 
+              for(int k=0; k<width; ++k) {
+                  temperature[index(0, j+1, k+1)] = gmsg->gh[k*height+j];
+              }
+          break;
+      case RIGHT:
+          for(int j=0; j<height; ++j) 
+              for(int k=0; k<width; ++k) {
+                  temperature[index(blockDimX+1, j+1, k+1)] = gmsg->gh[k*height+j];
+              }
+          break;
+      case TOP:
+          for(int i=0; i<height; ++i) 
+              for(int k=0; k<width; ++k) {
+                  temperature[index(i+1, 0, k+1)] = gmsg->gh[k*height+i];
+              }
+          break;
+      case BOTTOM:
+          for(int i=0; i<height; ++i) 
+              for(int k=0; k<width; ++k) {
+                  temperature[index(i+1, blockDimY+1, k+1)] = gmsg->gh[k*height+i];
+              }
+          break;
+      case FRONT:
+          for(int i=0; i<height; ++i) 
+              for(int j=0; j<width; ++j) {
+                  temperature[index(i+1, j+1, blockDimZ+1)] = gmsg->gh[j*height+i];
+              }
+          break;
+      case BACK:
+          for(int i=0; i<height; ++i) 
+              for(int j=0; j<width; ++j) {
+                  temperature[index(i+1, j+1, 0)] = gmsg->gh[j*height+i];
+              }
+          break;
+      default:
+          CkAbort("ERROR\n");
+      }
+
+      delete gmsg;
+    }
+
+
+       void check_and_compute() {
+               compute_kernel();
+
+               // calculate error
+               // not being done right now since we are doing a fixed no. of iterations
+
+               constrainBC();
+
+               if (iterations % CKP_FREQ == 0 || iterations > MAX_ITER){
+#ifdef CMK_MEM_CHECKPOINT
+                       contribute(0, 0, CkReduction::concat, CkCallback(CkIndex_Main::report(), mainProxy));
+#elif CMK_MESSAGE_LOGGING
+                       if(iterations > MAX_ITER)
+                               contribute(0, 0, CkReduction::concat, CkCallback(CkIndex_Main::report(), mainProxy));
+                       else
+                               AtSync();
+#else
+            contribute(0, 0, CkReduction::concat, CkCallback(CkIndex_Main::report(), mainProxy));
+#endif
+        } else {
+                       doStep();
+        }
+    }
+
+       void ResumeFromSync(){
+               doStep();
+       }
+
+
+    // Check to see if we have received all neighbor values yet
+    // If all neighbor values have been received, we update our values and proceed
+    void compute_kernel() {     //Gauss-Siedal compute
+        int i;
+  #pragma omp parallel for schedule(dynamic) 
+        for(i=1; i<blockDimX+1; ++i) {
+          //printf("[%d] did  %d iteration out of %d \n", omp_get_thread_num(), i, blockDimX+1); 
+          for(int j=1; j<blockDimY+1; ++j) {
+              for(int k=1; k<blockDimZ+1; ++k) {
+                  // update my value based on the surrounding values
+                   temperature[index(i, j, k)] = (temperature[index(i-1, j, k)] 
+                                           +  temperature[index(i+1, j, k)]
+                                           +  temperature[index(i, j-1, k)]
+                                           +  temperature[index(i, j+1, k)]
+                                           +  temperature[index(i, j, k-1)]
+                                           +  temperature[index(i, j, k+1)]
+                                           +  temperature[index(i, j, k)] ) * DIVIDEBY7;
+              }
+          }
+      }
+    }
+
+    // Enforce some boundary conditions
+    void constrainBC() {
+      // Heat left, top and front faces of each chare's block
+      for(int i=1; i<blockDimX+1; ++i)
+      for(int k=1; k<blockDimZ+1; ++k)
+              temperature[index(i, 1, k)] = 255.0;
+      for(int j=1; j<blockDimY+1; ++j)
+          for(int k=1; k<blockDimZ+1; ++k)
+              temperature[index(1, j, k)] = 255.0;
+      for(int i=1; i<blockDimX+1; ++i)
+          for(int j=1; j<blockDimY+1; ++j)
+              temperature[index(i, j, 1)] = 255.0;
+    }
+
+};
+
+/** \class JacobiMap
+ *
+ */
+
+class JacobiMap : public CkArrayMap {
+  public:
+    int X, Y, Z;
+    int *mapping;
+
+    JacobiMap(int x, int y, int z) {
+      X = x; Y = y; Z = z;
+      mapping = new int[X*Y*Z];
+
+      // we are assuming that the no. of chares in each dimension is a 
+      // multiple of the torus dimension
+
+      TopoManager tmgr;
+      int dimNX, dimNY, dimNZ, dimNT;
+
+      dimNX = tmgr.getDimNX();
+      dimNY = tmgr.getDimNY();
+      dimNZ = tmgr.getDimNZ();
+      dimNT = tmgr.getDimNT();
+
+      // we are assuming that the no. of chares in each dimension is a 
+      // multiple of the torus dimension
+      int numCharesPerPe = X*Y*Z/CkNumPes();
+
+      int numCharesPerPeX = X / dimNX;
+      int numCharesPerPeY = Y / dimNY;
+      int numCharesPerPeZ = Z / dimNZ;
+      int pe = 0, pes = CkNumPes();
+
+#if USE_BLOCK_RNDMAP
+      int used[pes];
+      for(int i=0; i<pes; i++)
+       used[i] = 0;
+#endif
+
+      if(dimNT < 2) {  // one core per node
+       if(CkMyPe()==0) CkPrintf("%d %d %d %d : %d %d %d \n", dimNX, dimNY, dimNZ, dimNT, numCharesPerPeX, numCharesPerPeY, numCharesPerPeZ); 
+       for(int i=0; i<dimNX; i++)
+         for(int j=0; j<dimNY; j++)
+           for(int k=0; k<dimNZ; k++)
+           {
+#if USE_BLOCK_RNDMAP
+             pe = myrand(pes); 
+             while(used[pe]!=0) {
+               pe = myrand(pes); 
+             }
+             used[pe] = 1;
+#endif
+
+             for(int ci=i*numCharesPerPeX; ci<(i+1)*numCharesPerPeX; ci++)
+               for(int cj=j*numCharesPerPeY; cj<(j+1)*numCharesPerPeY; cj++)
+                 for(int ck=k*numCharesPerPeZ; ck<(k+1)*numCharesPerPeZ; ck++) {
+#if USE_TOPOMAP
+                   mapping[ci*Y*Z + cj*Z + ck] = tmgr.coordinatesToRank(i, j, k);
+#elif USE_BLOCK_RNDMAP
+                   mapping[ci*Y*Z + cj*Z + ck] = pe;
+#endif
+                 }
+           }
+      } else {         // multiple cores per node
+       // In this case, we split the chares in the X dimension among the
+       // cores on the same node. The strange thing I figured out is that
+       // doing this in the Z dimension is not as good.
+       numCharesPerPeX /= dimNT;
+       if(CkMyPe()==0) CkPrintf("%d %d %d %d : %d %d %d \n", dimNX, dimNY, dimNZ, dimNT, numCharesPerPeX, numCharesPerPeY, numCharesPerPeZ);
+
+       for(int i=0; i<dimNX; i++)
+         for(int j=0; j<dimNY; j++)
+           for(int k=0; k<dimNZ; k++)
+             for(int l=0; l<dimNT; l++)
+               for(int ci=(dimNT*i+l)*numCharesPerPeX; ci<(dimNT*i+l+1)*numCharesPerPeX; ci++)
+                 for(int cj=j*numCharesPerPeY; cj<(j+1)*numCharesPerPeY; cj++)
+                   for(int ck=k*numCharesPerPeZ; ck<(k+1)*numCharesPerPeZ; ck++) {
+                     mapping[ci*Y*Z + cj*Z + ck] = tmgr.coordinatesToRank(i, j, k, l);
+                   }
+      } // end of if
+
+      if(CkMyPe() == 0) CkPrintf("Map generated ... \n");
+    }
+
+    ~JacobiMap() { 
+      delete [] mapping;
+    }
+
+    int procNum(int, const CkArrayIndex &idx) {
+      int *index = (int *)idx.data();
+      return mapping[index[0]*Y*Z + index[1]*Z + index[2]]; 
+    }
+};
+
+class OmpInitializer : public CBase_OmpInitializer{
+public:
+       OmpInitializer(int numThreads) {
+#ifdef JACOBI_OPENMP
+          if (numThreads < 1) {
+            numThreads = 1; 
+          }
+          //omp_set_num_threads(numThreads);
+          if(CkMyPe() == 0)
+          {
+#pragma omp parallel
+              { if(omp_get_thread_num() == 0 )
+                  CkPrintf("Computation loop will be parallelized"
+               " with %d OpenMP threads\n", omp_get_num_threads());
+              }
+          }
+#endif
+          mainProxy.start();
+       }
+};
+
+
+#include "jacobi3d.def.h"
diff --git a/tests/charm++/jacobi3d-gausssiedal/jacobi3d.ci b/tests/charm++/jacobi3d-gausssiedal/jacobi3d.ci
new file mode 100644 (file)
index 0000000..62729d6
--- /dev/null
@@ -0,0 +1,57 @@
+mainmodule jacobi3d {
+
+  readonly CProxy_Main mainProxy;
+  readonly int arrayDimX;
+  readonly int arrayDimY;
+  readonly int arrayDimZ;
+  readonly int blockDimX;
+  readonly int blockDimY;
+  readonly int blockDimZ;
+
+  readonly int num_chare_x;
+  readonly int num_chare_y;
+  readonly int num_chare_z;
+
+  readonly int globalBarrier;
+
+  message ghostMsg {
+    double gh[];
+  };
+
+  mainchare Main {
+    entry Main(CkArgMsg *m);
+    entry void start();
+    entry void report();
+  };
+
+  array [3D] Jacobi {
+    // Normal Charm++ entry methods
+    entry Jacobi(void);
+    entry void begin_iteration(void);
+    entry void receiveGhosts(ghostMsg *gmsg);
+    entry void processGhosts(ghostMsg *gmsg);
+
+    entry void doStep() {
+      serial "begin_iteration" {
+       begin_iteration();
+      }
+      for(imsg = 0; imsg < 6; imsg++) {
+       // "iterations" keeps track of messages across steps
+       when receiveGhosts[iterations] (ghostMsg *gmsg)
+         serial "process ghosts" { processGhosts(gmsg); }
+      }
+      serial "doWork" {
+       check_and_compute();
+      }
+    };
+  };
+
+  group JacobiMap : CkArrayMap {
+    entry JacobiMap(int x, int y, int z);
+  };
+
+  nodegroup OmpInitializer {
+    entry OmpInitializer(int numThreads); 
+  };
+
+};