[Dune-devel] dune-istl communication question

[Christian Engwer]

Dear all,

why does the mask object in the istl communication:

novlpschwarz.hh:141:            mask[i] = 1;
novlpschwarz.hh:144:              mask[i->local().local()] = 0;
novlpschwarz.hh:146:              mask[i->local().local()] = 2;
novlpschwarz.hh:158:          if (mask[i.index()] == 0)
novlpschwarz.hh:174:          if (mask[i.index()] == 0) {
novlpschwarz.hh:179:              if (mask[j.index()] == 0) {
novlpschwarz.hh:212:        if (mask[i.index()] == 0) {
novlpschwarz.hh:215:            if (mask[j.index()] == 1) //j is owner => then sum entries
novlpschwarz.hh:217:            else if (mask[j.index()] == 0) {
novlpschwarz.hh:226:        else if (mask[i.index()] == 1) {
novlpschwarz.hh:228:            if (mask[j.index()] != 2)
owneroverlapcopy.hh:410:          mask[i] = 1;
owneroverlapcopy.hh:413:            mask[i->local().local()] = 0;
owneroverlapcopy.hh:418:        result += x[i]*(y[i])*mask[i];
owneroverlapcopy.hh:437:          mask[i] = 1;
owneroverlapcopy.hh:440:            mask[i->local().local()] = 0;
owneroverlapcopy.hh:444:        result += x[i].two_norm2()*mask[i];

It only stores entries from the set [0,1,2]

Currently it is implemented as std::vector<double>. Why?

(A) why double?
 Is it faster to use double here? shouldn't we at least
 restrict ourselfs to float? Why not use int or size_t or what ever?
 Although char would be ric enough, I assume it might lead to
 performance penalties due to alignment issues.
 The mask is as large as the original data so storage is not
 negligible.
(B) why std::vector?
 I have the impression it will fail with hard to detect errors if we
 types further nested than BlockVector<FieldVector<...>>.

Ciao
Christian