initialize fix

amgprec/impl/aggregator/initialize.cpp

@@ -8,7 +8,7 @@
 #include "dataStrStaticQueue.h"
 #include "omp.h"
 
-#define NUM_THREAD 12
+#define NUM_THREAD 4
 
 inline void initialize(MilanLongInt NLVer, MilanLongInt NLEdge,
                        MilanLongInt StartIndex, MilanLongInt EndIndex,
@@ -50,7 +50,6 @@ inline void initialize(MilanLongInt NLVer, MilanLongInt NLEdge,
 
 #pragma omp single
         {
-
             // Initialize the locks
 #pragma omp taskloop num_tasks(NUM_THREAD)
             for (i = 0; i < NLVer; i++)
@@ -71,9 +70,13 @@ inline void initialize(MilanLongInt NLVer, MilanLongInt NLEdge,
              * only when a ghost edge is found and ghost edges are a minority,
              * circa 3.5% during the tests.
              */
+
+#pragma omp task depend(out \
+                        : numGhostEdges, Counter, Ghost2LocalMap, insertMe, storedAlready, numGhostVertices)
+            {
+
 #pragma omp taskloop num_tasks(NUM_THREAD) reduction(+ \
-                                                     : numGhostEdges) depend(out \
-                                                                             : numGhostEdges, Counter, Ghost2LocalMap)
+                                                     : numGhostEdges)
                 for (i = 0; i < NLEdge; i++)
                 { // O(m) - Each edge stored twice
                     insertMe = verLocInd[i];
@@ -96,7 +99,9 @@ inline void initialize(MilanLongInt NLVer, MilanLongInt NLEdge,
                         }
                     } // End of if ( (insertMe < StartIndex) || (insertMe > EndIndex) )
                 }     // End of for(ghost vertices)
+            }         // end of task depend
 
+            // numGhostEdges = atomicNumGhostEdges;
 #ifdef TIME_TRACKER
             Ghost2LocalInitialization = MPI_Wtime() - Ghost2LocalInitialization;
             fprintf(stderr, "Ghost2LocalInitialization time: %f\n", Ghost2LocalInitialization);
@@ -121,7 +126,7 @@ inline void initialize(MilanLongInt NLVer, MilanLongInt NLEdge,
 
 #pragma omp task depend(out                                                       \
                         : verGhostPtr, tempCounter, verGhostInd, GMate) depend(in \
-                                                                               : numGhostVertices)
+                                                                               : numGhostVertices, numGhostEdges)
             {
 
                 // Initialize adjacency Lists for Ghost Vertices:
@@ -151,7 +156,7 @@ inline void initialize(MilanLongInt NLVer, MilanLongInt NLEdge,
 
             } // End of task
 
-#pragma omp task depent(out                      \
+#pragma omp task depend(out                      \
                         : verGhostPtr) depend(in \
                                               : Counter, numGhostVertices)
             {
@@ -198,9 +203,11 @@ inline void initialize(MilanLongInt NLVer, MilanLongInt NLEdge,
              * are a minority hence the critical region is executed
              * few times, circa 3.5% of the times in the tests.
              */
-#pragma omp taskloop num_tasks(NUM_THREAD) depend(in                                                  \
-                                                  : insertMe, Ghost2LocalMap, tempCounter) depend(out \
+#pragma omp task depend(in                                                               \
+                        : insertMe, Ghost2LocalMap, tempCounter, verGhostPtr) depend(out \
                                                                                      : verGhostInd)
+            {
+#pragma omp taskloop num_tasks(NUM_THREAD)
                 for (v = 0; v < NLVer; v++)
                 {
                     adj1 = verLocPtr[v]; // Vertex Pointer
@@ -219,8 +226,7 @@ inline void initialize(MilanLongInt NLVer, MilanLongInt NLEdge,
                         }                                           // End of if((w < StartIndex) || (w > EndIndex))
                     }                                               // End of for(k)
                 }                                                   // End of for (v)
-
-        } // End of parallel region
+            }                                                       // end of tasklopp
 
 #ifdef TIME_TRACKER
             verGhostIndInitialization = MPI_Wtime() - verGhostIndInitialization;
@@ -252,7 +258,7 @@ inline void initialize(MilanLongInt NLVer, MilanLongInt NLEdge,
                     cout << "Not enough memory to allocate the internal variables \n";
                     exit(1);
                 }
-        }
+            } // end of task
 
 #ifdef PRINT_DEBUG_INFO_
             cout << "\n(" << myRank << ")Allocating CandidateMate.. ";
@@ -270,6 +276,7 @@ inline void initialize(MilanLongInt NLVer, MilanLongInt NLEdge,
             fflush(stdout);
             fflush(stdout);
 #endif
+
 #ifdef DEBUG_HANG_
             if (myRank == 0)
                 cout << "\n(" << myRank << ") Setup Time :" << *ph0_time << endl;
@@ -281,7 +288,6 @@ inline void initialize(MilanLongInt NLVer, MilanLongInt NLEdge,
                                                                   : candidateMate, S, U, privateU, privateQLocalVtx, privateQGhostVtx, privateQMsgType, privateQOwner)
             {
 
-            
                 // The values calculated in this function are sent back to the calling function
                 *numGhostEdgesPtr = numGhostEdges;
                 *numGhostVerticesPtr = numGhostVertices;
@@ -315,6 +321,8 @@ inline void initialize(MilanLongInt NLVer, MilanLongInt NLEdge,
                 new (&privateQGhostVtx) staticQueue(size);
                 new (&privateQMsgType) staticQueue(size);
                 new (&privateQOwner) staticQueue(size);
-        }
-    } // End of single
+            } // end of task
+
+        } // End of single region
+    }     // End of parallel region
 }

exec.sh

@@ -1,3 +1,4 @@
+rm amgprec/impl/aggregator/algoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP.o
 make all
 cd samples/advanced/pdegen
 make amg_d_pde3d