--- branches/development/src/parallel/ForceMatrixDecomposition.cpp	2011/06/03 21:39:49	1575
+++ branches/development/src/parallel/ForceMatrixDecomposition.cpp	2011/06/09 20:26:29	1579
@@ -57,8 +57,9 @@ namespace OpenMD {
     storageLayout_ = sman_->getStorageLayout();
     ff_ = info_->getForceField();
     nLocal_ = snap_->getNumberOfAtoms();
-    nGroups_ = snap_->getNumberOfCutoffGroups();
 
+    nGroups_ = info_->getNLocalCutoffGroups();
+    cerr << "in dId, nGroups = " << nGroups_ << "\n";
     // gather the information for atomtype IDs (atids):
     identsLocal = info_->getIdentArray();
     AtomLocalToGlobal = info_->getGlobalAtomIndices();
@@ -104,8 +105,8 @@ namespace OpenMD {
     cgColData.resize(nGroupsInCol_);
     cgColData.setStorageLayout(DataStorage::dslPosition);
         
-    identsRow.reserve(nAtomsInRow_);
-    identsCol.reserve(nAtomsInCol_);
+    identsRow.resize(nAtomsInRow_);
+    identsCol.resize(nAtomsInCol_);
     
     AtomCommIntRow->gather(identsLocal, identsRow);
     AtomCommIntColumn->gather(identsLocal, identsCol);
@@ -120,7 +121,7 @@ namespace OpenMD {
     AtomCommRealColumn->gather(massFactorsLocal, massFactorsCol);
 
     groupListRow_.clear();
-    groupListRow_.reserve(nGroupsInRow_);
+    groupListRow_.resize(nGroupsInRow_);
     for (int i = 0; i < nGroupsInRow_; i++) {
       int gid = cgRowToGlobal[i];
       for (int j = 0; j < nAtomsInRow_; j++) {
@@ -131,7 +132,7 @@ namespace OpenMD {
     }
 
     groupListCol_.clear();
-    groupListCol_.reserve(nGroupsInCol_);
+    groupListCol_.resize(nGroupsInCol_);
     for (int i = 0; i < nGroupsInCol_; i++) {
       int gid = cgColToGlobal[i];
       for (int j = 0; j < nAtomsInCol_; j++) {
@@ -141,93 +142,273 @@ namespace OpenMD {
       }      
     }
 
-    skipsForRowAtom.clear();
-    skipsForRowAtom.reserve(nAtomsInRow_);
+    skipsForAtom.clear();
+    skipsForAtom.resize(nAtomsInRow_);
+    toposForAtom.clear();
+    toposForAtom.resize(nAtomsInRow_);
+    topoDist.clear();
+    topoDist.resize(nAtomsInRow_);
     for (int i = 0; i < nAtomsInRow_; i++) {
       int iglob = AtomRowToGlobal[i];
-      for (int j = 0; j < nAtomsInCol_; j++) {
-        int jglob = AtomColToGlobal[j];        
-        if (excludes.hasPair(iglob, jglob)) 
-          skipsForRowAtom[i].push_back(j);       
-      }      
-    }
 
-    toposForRowAtom.clear();
-    toposForRowAtom.reserve(nAtomsInRow_);
-    for (int i = 0; i < nAtomsInRow_; i++) {
-      int iglob = AtomRowToGlobal[i];
-      int nTopos = 0;
       for (int j = 0; j < nAtomsInCol_; j++) {
-        int jglob = AtomColToGlobal[j];        
+        int jglob = AtomColToGlobal[j];
+
+        if (excludes.hasPair(iglob, jglob)) 
+          skipsForAtom[i].push_back(j);       
+        
         if (oneTwo.hasPair(iglob, jglob)) {
-          toposForRowAtom[i].push_back(j);
-          topoDistRow[i][nTopos] = 1;
-          nTopos++;
+          toposForAtom[i].push_back(j);
+          topoDist[i].push_back(1);
+        } else {
+          if (oneThree.hasPair(iglob, jglob)) {
+            toposForAtom[i].push_back(j);
+            topoDist[i].push_back(2);
+          } else {
+            if (oneFour.hasPair(iglob, jglob)) {
+              toposForAtom[i].push_back(j);
+              topoDist[i].push_back(3);
+            }
+          }
         }
-        if (oneThree.hasPair(iglob, jglob)) {
-          toposForRowAtom[i].push_back(j);
-          topoDistRow[i][nTopos] = 2;
-          nTopos++;
-        }
-        if (oneFour.hasPair(iglob, jglob)) {
-          toposForRowAtom[i].push_back(j);
-          topoDistRow[i][nTopos] = 3;
-          nTopos++;
-        }
       }      
     }
 
 #endif
 
     groupList_.clear();
-    groupList_.reserve(nGroups_);
+    groupList_.resize(nGroups_);
     for (int i = 0; i < nGroups_; i++) {
       int gid = cgLocalToGlobal[i];
       for (int j = 0; j < nLocal_; j++) {
         int aid = AtomLocalToGlobal[j];
-        if (globalGroupMembership[aid] == gid)
+        if (globalGroupMembership[aid] == gid) {
           groupList_[i].push_back(j);
+        }
       }      
     }
 
-    skipsForLocalAtom.clear();
-    skipsForLocalAtom.reserve(nLocal_);
+    skipsForAtom.clear();
+    skipsForAtom.resize(nLocal_);
+    toposForAtom.clear();
+    toposForAtom.resize(nLocal_);
+    topoDist.clear();
+    topoDist.resize(nLocal_);
 
     for (int i = 0; i < nLocal_; i++) {
       int iglob = AtomLocalToGlobal[i];
+
       for (int j = 0; j < nLocal_; j++) {
-        int jglob = AtomLocalToGlobal[j];        
+        int jglob = AtomLocalToGlobal[j];
+
         if (excludes.hasPair(iglob, jglob)) 
-          skipsForLocalAtom[i].push_back(j);       
+          skipsForAtom[i].push_back(j);              
+        
+        if (oneTwo.hasPair(iglob, jglob)) {
+          toposForAtom[i].push_back(j);
+          topoDist[i].push_back(1);
+        } else {
+          if (oneThree.hasPair(iglob, jglob)) {
+            toposForAtom[i].push_back(j);
+            topoDist[i].push_back(2);
+          } else {
+            if (oneFour.hasPair(iglob, jglob)) {
+              toposForAtom[i].push_back(j);
+              topoDist[i].push_back(3);
+            }
+          }
+        }
       }      
     }
+    
+    createGtypeCutoffMap();
+  }
+   
+  void ForceMatrixDecomposition::createGtypeCutoffMap() {
 
-    toposForLocalAtom.clear();
-    toposForLocalAtom.reserve(nLocal_);
-    for (int i = 0; i < nLocal_; i++) {
-      int iglob = AtomLocalToGlobal[i];
-      int nTopos = 0;
-      for (int j = 0; j < nLocal_; j++) {
-        int jglob = AtomLocalToGlobal[j];        
-        if (oneTwo.hasPair(iglob, jglob)) {
-          toposForLocalAtom[i].push_back(j);
-          topoDistLocal[i][nTopos] = 1;
-          nTopos++;
+    RealType tol = 1e-6;
+    RealType rc;
+    int atid;
+    set<AtomType*> atypes = info_->getSimulatedAtomTypes();
+    vector<RealType> atypeCutoff;
+    atypeCutoff.resize( atypes.size() );
+
+    for (set<AtomType*>::iterator at = atypes.begin(); 
+         at != atypes.end(); ++at){
+      rc = interactionMan_->getSuggestedCutoffRadius(*at);
+      atid = (*at)->getIdent();
+      atypeCutoff[atid] = rc;
+    }
+
+    vector<RealType> gTypeCutoffs;
+
+    // first we do a single loop over the cutoff groups to find the
+    // largest cutoff for any atypes present in this group.
+#ifdef IS_MPI
+    vector<RealType> groupCutoffRow(nGroupsInRow_, 0.0);
+    groupRowToGtype.resize(nGroupsInRow_);
+    for (int cg1 = 0; cg1 < nGroupsInRow_; cg1++) {
+      vector<int> atomListRow = getAtomsInGroupRow(cg1);
+      for (vector<int>::iterator ia = atomListRow.begin(); 
+           ia != atomListRow.end(); ++ia) {            
+        int atom1 = (*ia);
+        atid = identsRow[atom1];
+        if (atypeCutoff[atid] > groupCutoffRow[cg1]) {
+          groupCutoffRow[cg1] = atypeCutoff[atid];
         }
-        if (oneThree.hasPair(iglob, jglob)) {
-          toposForLocalAtom[i].push_back(j);
-          topoDistLocal[i][nTopos] = 2;
-          nTopos++;
+      }
+
+      bool gTypeFound = false;
+      for (int gt = 0; gt < gTypeCutoffs.size(); gt++) {
+        if (abs(groupCutoffRow[cg1] - gTypeCutoffs[gt]) < tol) {
+          groupRowToGtype[cg1] = gt;
+          gTypeFound = true;
+        } 
+      }
+      if (!gTypeFound) {
+        gTypeCutoffs.push_back( groupCutoffRow[cg1] );
+        groupRowToGtype[cg1] = gTypeCutoffs.size() - 1;
+      }
+      
+    }
+    vector<RealType> groupCutoffCol(nGroupsInCol_, 0.0);
+    groupColToGtype.resize(nGroupsInCol_);
+    for (int cg2 = 0; cg2 < nGroupsInCol_; cg2++) {
+      vector<int> atomListCol = getAtomsInGroupColumn(cg2);
+      for (vector<int>::iterator jb = atomListCol.begin(); 
+           jb != atomListCol.end(); ++jb) {            
+        int atom2 = (*jb);
+        atid = identsCol[atom2];
+        if (atypeCutoff[atid] > groupCutoffCol[cg2]) {
+          groupCutoffCol[cg2] = atypeCutoff[atid];
         }
-        if (oneFour.hasPair(iglob, jglob)) {
-          toposForLocalAtom[i].push_back(j);
-          topoDistLocal[i][nTopos] = 3;
-          nTopos++;
+      }
+      bool gTypeFound = false;
+      for (int gt = 0; gt < gTypeCutoffs.size(); gt++) {
+        if (abs(groupCutoffCol[cg2] - gTypeCutoffs[gt]) < tol) {
+          groupColToGtype[cg2] = gt;
+          gTypeFound = true;
+        } 
+      }
+      if (!gTypeFound) {
+        gTypeCutoffs.push_back( groupCutoffCol[cg2] );
+        groupColToGtype[cg2] = gTypeCutoffs.size() - 1;
+      }
+    }
+#else
+
+    vector<RealType> groupCutoff(nGroups_, 0.0);
+    groupToGtype.resize(nGroups_);
+
+    cerr << "nGroups = " << nGroups_ << "\n";
+    for (int cg1 = 0; cg1 < nGroups_; cg1++) {
+
+      groupCutoff[cg1] = 0.0;
+      vector<int> atomList = getAtomsInGroupRow(cg1);
+
+      for (vector<int>::iterator ia = atomList.begin(); 
+           ia != atomList.end(); ++ia) {            
+        int atom1 = (*ia);
+        atid = identsLocal[atom1];
+        if (atypeCutoff[atid] > groupCutoff[cg1]) {
+          groupCutoff[cg1] = atypeCutoff[atid];
         }
+      }
+
+      bool gTypeFound = false;
+      for (int gt = 0; gt < gTypeCutoffs.size(); gt++) {
+        if (abs(groupCutoff[cg1] - gTypeCutoffs[gt]) < tol) {
+          groupToGtype[cg1] = gt;
+          gTypeFound = true;
+        } 
+      }
+      if (!gTypeFound) {
+        gTypeCutoffs.push_back( groupCutoff[cg1] );
+        groupToGtype[cg1] = gTypeCutoffs.size() - 1;
       }      
     }
+#endif
+
+    cerr << "gTypeCutoffs.size() = " << gTypeCutoffs.size() << "\n";
+    // Now we find the maximum group cutoff value present in the simulation
+
+    RealType groupMax = *max_element(gTypeCutoffs.begin(), gTypeCutoffs.end());
+
+#ifdef IS_MPI
+    MPI::COMM_WORLD.Allreduce(&groupMax, &groupMax, 1, MPI::REALTYPE, MPI::MAX);
+#endif
+    
+    RealType tradRcut = groupMax;
+
+    for (int i = 0; i < gTypeCutoffs.size();  i++) {
+      for (int j = 0; j < gTypeCutoffs.size();  j++) {       
+        RealType thisRcut;
+        switch(cutoffPolicy_) {
+        case TRADITIONAL:
+          thisRcut = tradRcut;
+          break;
+        case MIX:
+          thisRcut = 0.5 * (gTypeCutoffs[i] + gTypeCutoffs[j]);
+          break;
+        case MAX:
+          thisRcut = max(gTypeCutoffs[i], gTypeCutoffs[j]);
+          break;
+        default:
+          sprintf(painCave.errMsg,
+                  "ForceMatrixDecomposition::createGtypeCutoffMap " 
+                  "hit an unknown cutoff policy!\n");
+          painCave.severity = OPENMD_ERROR;
+          painCave.isFatal = 1;
+          simError();
+          break;
+        }
+
+        pair<int,int> key = make_pair(i,j);
+        gTypeCutoffMap[key].first = thisRcut;
+
+        if (thisRcut > largestRcut_) largestRcut_ = thisRcut;
+
+        gTypeCutoffMap[key].second = thisRcut*thisRcut;
+        
+        gTypeCutoffMap[key].third = pow(thisRcut + skinThickness_, 2);
+
+        // sanity check
+        
+        if (userChoseCutoff_) {
+          if (abs(gTypeCutoffMap[key].first - userCutoff_) > 0.0001) {
+            sprintf(painCave.errMsg,
+                    "ForceMatrixDecomposition::createGtypeCutoffMap " 
+                    "user-specified rCut does not match computed group Cutoff\n");
+            painCave.severity = OPENMD_ERROR;
+            painCave.isFatal = 1;
+            simError();            
+          }
+        }
+      }
+    }
   }
-   
+
+
+  groupCutoffs ForceMatrixDecomposition::getGroupCutoffs(int cg1, int cg2) {
+    int i, j;   
+#ifdef IS_MPI
+    i = groupRowToGtype[cg1];
+    j = groupColToGtype[cg2];
+#else
+    i = groupToGtype[cg1];
+    j = groupToGtype[cg2];
+#endif    
+    return gTypeCutoffMap[make_pair(i,j)];
+  }
+
+  int ForceMatrixDecomposition::getTopologicalDistance(int atom1, int atom2) {
+    for (int j = 0; j < toposForAtom[atom1].size(); j++) {
+      if (toposForAtom[atom1][j] == atom2) 
+        return topoDist[atom1][j];
+    }
+    return 0;
+  }
+
   void ForceMatrixDecomposition::zeroWorkArrays() {
 
     for (int j = 0; j < N_INTERACTION_FAMILIES; j++) {
@@ -536,12 +717,8 @@ namespace OpenMD {
     return d;    
   }
 
-  vector<int> ForceMatrixDecomposition::getSkipsForRowAtom(int atom1) {
-#ifdef IS_MPI
-    return skipsForRowAtom[atom1];
-#else
-    return skipsForLocalAtom[atom1];
-#endif
+  vector<int> ForceMatrixDecomposition::getSkipsForAtom(int atom1) {
+    return skipsForAtom[atom1];
   }
 
   /**
@@ -574,35 +751,14 @@ namespace OpenMD {
     unique_id_2 = atom2;
 #endif
     
-#ifdef IS_MPI
-    for (vector<int>::iterator i = skipsForRowAtom[atom1].begin();
-         i != skipsForRowAtom[atom1].end(); ++i) {
-      if ( (*i) == unique_id_2 ) return true;
-    }    
-#else
-    for (vector<int>::iterator i = skipsForLocalAtom[atom1].begin();
-         i != skipsForLocalAtom[atom1].end(); ++i) {
+    for (vector<int>::iterator i = skipsForAtom[atom1].begin();
+         i != skipsForAtom[atom1].end(); ++i) {
       if ( (*i) == unique_id_2 ) return true;
     }    
-#endif
-  }
 
-  int ForceMatrixDecomposition::getTopoDistance(int atom1, int atom2) {
-    
-#ifdef IS_MPI
-    for (int i = 0; i < toposForRowAtom[atom1].size(); i++) {
-      if ( toposForRowAtom[atom1][i] == atom2 ) return topoDistRow[atom1][i];
-    }
-#else
-    for (int i = 0; i < toposForLocalAtom[atom1].size(); i++) {
-      if ( toposForLocalAtom[atom1][i] == atom2 ) return topoDistLocal[atom1][i];
-    }
-#endif
-
-    // zero is default for unconnected (i.e. normal) pair interactions
-    return 0;
   }
 
+
   void ForceMatrixDecomposition::addForceToAtomRow(int atom1, Vector3d fg){
 #ifdef IS_MPI
     atomRowData.force[atom1] += fg;
@@ -765,6 +921,7 @@ namespace OpenMD {
   vector<pair<int, int> > ForceMatrixDecomposition::buildNeighborList() {
       
     vector<pair<int, int> > neighborList;
+    groupCutoffs cuts;
 #ifdef IS_MPI
     cellListRow_.clear();
     cellListCol_.clear();
@@ -772,10 +929,7 @@ namespace OpenMD {
     cellList_.clear();
 #endif
 
-    // dangerous to not do error checking.
-    RealType rCut_;
- 
-    RealType rList_ = (rCut_ + skinThickness_);
+    RealType rList_ = (largestRcut_ + skinThickness_);
     RealType rl2 = rList_ * rList_;
     Snapshot* snap_ = sman_->getCurrentSnapshot();
     Mat3x3d Hmat = snap_->getHmat();
@@ -791,8 +945,16 @@ namespace OpenMD {
     Vector3d rs, scaled, dr;
     Vector3i whichCell;
     int cellIndex;
+    int nCtot = nCells_.x() * nCells_.y() * nCells_.z();
 
 #ifdef IS_MPI
+    cellListRow_.resize(nCtot);
+    cellListCol_.resize(nCtot);
+#else
+    cellList_.resize(nCtot);
+#endif
+
+#ifdef IS_MPI
     for (int i = 0; i < nGroupsInRow_; i++) {
       rs = cgRowData.position[i];
       // scaled positions relative to the box vectors
@@ -898,7 +1060,8 @@ namespace OpenMD {
                 if (m2 != m1 || cgColToGlobal[(*j2)] < cgRowToGlobal[(*j1)]) {
                   dr = cgColData.position[(*j2)] - cgRowData.position[(*j1)];
                   snap_->wrapVector(dr);
-                  if (dr.lengthSquare() < rl2) {
+                  cuts = getGroupCutoffs( (*j1), (*j2) );
+                  if (dr.lengthSquare() < cuts.third) {
                     neighborList.push_back(make_pair((*j1), (*j2)));
                   }
                 }
@@ -917,7 +1080,8 @@ namespace OpenMD {
                 if (m2 != m1 || (*j2) < (*j1)) {
                   dr = snap_->cgData.position[(*j2)] - snap_->cgData.position[(*j1)];
                   snap_->wrapVector(dr);
-                  if (dr.lengthSquare() < rl2) {
+                  cuts = getGroupCutoffs( (*j1), (*j2) );
+                  if (dr.lengthSquare() < cuts.third) {
                     neighborList.push_back(make_pair((*j1), (*j2)));
                   }
                 }