/// <summary> /// converts the temporary CSR format into the final matrix format /// </summary> abstract protected FormatBase AssembleFinalFormat(TempCSR tmp);
/// <summary>
/// matrix assembly; must be called by each implementation,
/// </summary>
/// <param name="M"></param>
protected void PackMatrix(IMutableMatrixEx M)
{
ilPSP.MPICollectiveWatchDog.Watch();
IPartitioning rp = M.RowPartitioning;
IPartitioning cp = m_ColPart;
// define Comm List
// ================
SortedDictionary <int, List <int> > CommLists = new SortedDictionary <int, List <int> >();
// keys: processor rank p
// values: List of global indices, which processor p needs to send to this processor
int Lr;
double[] val = null;
int[] col = null;
int L = rp.LocalLength;
int i0 = (int)rp.i0;
for (int iLoc = 0; iLoc < L; iLoc++) // loop over all matrix rows...
{
int iGlob = i0 + iLoc;
//MsrMatrix.MatrixEntry[] row = (asMsr==null) ? M.GetRow(iGlob) : asMsr.GetRowShallow(iGlob);
Lr = M.GetOccupiedColumnIndices(iGlob, ref col);
for (int j = 0; j < Lr; j++) // loop over all nonzero entries in row 'iGlob'
{
int jGlob = col[j];
if (cp.i0 <= jGlob && jGlob < (cp.i0 + cp.LocalLength))
{
// Entry on current processor
}
else
{
int proc = cp.FindProcess(jGlob);
// Entry on Processor proc
if (!CommLists.ContainsKey(proc))
{
CommLists.Add(proc, new List <int>());
}
List <int> CommList_proc = CommLists[proc];
if (!CommList_proc.Contains(jGlob)) // a lot of room for optimization
{
CommList_proc.Add(jGlob);
}
}
}
}
// sort com list
// =============
{
foreach (List <int> cl in CommLists.Values)
{
cl.Sort();
}
}
// define matrix
// =============
{
TempCSR intTmp = new TempCSR();
SortedDictionary <int, ExternalTmp> extTmp = new SortedDictionary <int, ExternalTmp>();
foreach (int proc in CommLists.Keys)
{
extTmp.Add(proc, new ExternalTmp());
}
for (int iLoc = 0; iLoc < L; iLoc++)
{
int iGlob = i0 + iLoc;
Lr = M.GetRow(iGlob, ref col, ref val);
for (int j = 0; j < Lr; j++)
{
int jGlob = col[j];
double Value = val[j];
bool bIsDiag = (iGlob == jGlob);
if (cp.i0 <= jGlob && jGlob < (cp.i0 + cp.LocalLength))
{
// Entry on current processor
intTmp.AddEntry(jGlob - (int)cp.i0, Value, bIsDiag);
}
else
{
int proc = cp.FindProcess(jGlob);
// Entry on Processor proc
List <int> CommList_proc = CommLists[proc];
int jloc = CommList_proc.IndexOf(jGlob);
ExternalTmp et = extTmp[proc];
et.AddEntry(jloc, jGlob, Value);
}
}
intTmp.NextRow();
foreach (ExternalTmp et in extTmp.Values)
{
et.NextRow();
}
}
m_LocalMtx = AssembleFinalFormat(intTmp);
ExtMatrix = new Dictionary <int, External>();
foreach (int proc in extTmp.Keys)
{
ExtMatrix.Add(proc, extTmp[proc].GetFinalObj());
}
}
// send/receive & transform Comm lists
// ====================================
{
SerialisationMessenger sms = new SerialisationMessenger(csMPI.Raw._COMM.WORLD);
SortedDictionary <int, int[]> CommListsTo = new SortedDictionary <int, int[]>();
foreach (int proc in CommLists.Keys)
{
sms.SetCommPath(proc);
}
sms.CommitCommPaths();
foreach (int proc in CommLists.Keys)
{
sms.Transmitt(proc, CommLists[proc].ToArray());
}
int _proc;
int[] CommListReceived;
sms.GetNext(out _proc, out CommListReceived);
int Lcol = m_ColPart.LocalLength;
int i0col = (int)m_ColPart.i0;
while (CommListReceived != null)
{
// convert indices to local coordinates
for (int i = 0; i < CommListReceived.Length; i++)
{
CommListReceived[i] -= i0col;
// check:
if (CommListReceived[i] < 0 || CommListReceived[i] >= Lcol)
{
throw new ApplicationException("internal error: something wrong with received Comm List.");
}
}
CommListsTo.Add(_proc, CommListReceived);
sms.GetNext(out _proc, out CommListReceived);
}
sms.Dispose();
m_SpmvCommPattern = new SpmvCommPattern();
m_SpmvCommPattern.ComLists = CommListsTo;
}
// record the number of elements which we receive
// ==============================================
{
m_SpmvCommPattern.NoOfReceivedEntries = new Dictionary <int, int>();
foreach (int p in CommLists.Keys)
{
m_SpmvCommPattern.NoOfReceivedEntries.Add(p, CommLists[p].Count);
}
}
}
