PARALLEL_PROCESS_EXPOSED_VERTEX_B is actually not parallelizable. Atleast not as I was doing.

amgprec/impl/aggregator/algoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP.cpp

@@ -424,22 +424,15 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
             //Compute the Initial Matching Set:
 
             S = numGhostVertices; //Initialize S with number of Ghost Vertices
-        } // end of single region
 
             /*
              * OMP PARALLEL_COMPUTE_CANDIDATE_MATE_B
-         * The next portion of code has been splitted
-         * to make it 100% parallelized
-         *
-         * TODO: I think it diminish the cache update, does it?
-         *
-         * TODO: would it make any sense to parallelize also the
-         *       inner for?
+             * It is actually not possible to parallelize this cycle
+             * as it is.
              *
-         * TODO: we have a false sharing on candidateMate
+             * TODO think how it could be parallelizable
              */
 
-#pragma omp for
             for ( v=0; v < NLVer; v++ ) {
 #ifdef PRINT_DEBUG_INFO_
                 cout<<"\n("<<myRank<<")Processing: "<<v+StartIndex<<endl; fflush(stdout);
@@ -468,30 +461,12 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
                 candidateMate[v] = w;
                 //End: PARALLEL_COMPUTE_CANDIDATE_MATE_B(v)
 
-    }
-
-    /*
-        TODO this cycle has a lot of margin of improvement!!!!
-             This current version introduce some errors.
-             1 - ollback to the previous verison and check if it is
-                100% stable
-            2 - if the previous verison was stable all right, if not
-                that's a big deal
-            3 - reimplement step by step to check from where the instability
-                comes from
-    */
-
-#pragma omp for reduction(+: msgInd, NumMessagesBundled, myCard, PCounter[:numProcs])
-        for ( v=0; v < NLVer; v++ ) {
-
                 //Start: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
 
 #ifdef PRINT_DEBUG_INFO_
                 cout<<"\n("<<myRank<<")Processing: "<<v+StartIndex<<endl; fflush(stdout);
 #endif
 
-            w = candidateMate[v];
-
 #ifdef PRINT_DEBUG_INFO_
                 cout<<"\n("<<myRank<<")"<<v+StartIndex<<" Points to: "<<w; fflush(stdout);
 #endif
@@ -514,9 +489,8 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
                         msgInd++;
                         NumMessagesBundled++;
                         ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
-                    PCounter[ghostOwner]++; //TODO maybe reduction?
-#pragma omp critical
-                    {
+                        PCounter[ghostOwner]++;
+
                         QLocalVtx.push_back(v + StartIndex);
                         QGhostVtx.push_back(w);
                         QMsgType.push_back(REQUEST);
@@ -550,13 +524,11 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
                             } //End of if Counter[w] > 0
                             //End: PARALLEL_PROCESS_CROSS_EDGE_B(v)
                         } //End of if CandidateMate[w] = v
-                        } // end of critical region
                     } //End of if a Ghost Vertex
                     else { // w is a local vertex
 
                         if (candidateMate[w - StartIndex] == (v + StartIndex)) {
-#pragma omp critical
-                            {
+
                             Mate[v] = w;  //v is local
                             Mate[w - StartIndex] = v + StartIndex; //w is local
                             //Q.push_back(u);
@@ -566,7 +538,7 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
 #ifdef PRINT_DEBUG_INFO_
                             cout<<"\n("<<myRank<<")MATCH: ("<<v+StartIndex<<","<<w<<") "; fflush(stdout);
 #endif
-                            }
+
                         } //End of if ( candidateMate[w-StartIndex] == (v+StartIndex) )
                     } //End of Else
                 } //End of if(w >=0)
@@ -607,12 +579,12 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
                     } //End of for loop
                 } // End of Else: w == -1
                 //End:   PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
-        //} // end of critical
             } //End of for ( v=0; v < NLVer; v++ )
+
+        } // end of single region
     } // end of parallel region
 
     tempCounter.clear(); //Do not need this any more
-    //} // end of parallel region
 
 
 #ifdef PRINT_DEBUG_INFO_