ENH: provide MPI native bitOrOp reduce with single/multiple values

- can be used, for example, to track global states:

      // Encode as 0:empty, 1:uniform, 2:nonuniform, 3:mixed
      PackedList<2> uniformity(fields.size());

      forAll(fields, i)
      {
          uniformity.set(i, fields[i].whichUniformity());
      }

      reduce
      (
          uniformity.data(),
          uniformity.size_data(),
          bitOrOp<unsigned>()
      );
This commit is contained in:
Mark Olesen
2023-07-05 14:48:31 +02:00
parent f398d7b313
commit 8ee4b52560
4 changed files with 144 additions and 7 deletions

View File

@ -39,6 +39,7 @@ Description
#include "Tuple2.H"
#include "IOstreams.H"
#include "PstreamReduceOps.H"
#include "bitSet.H"
using namespace Foam;
@ -87,7 +88,10 @@ int main(int argc, char *argv[])
// Reductions (using MPI intrinsics)
{
label val = Pstream::myProcNo(UPstream::commWorld());
const label myRank = UPstream::myProcNo(UPstream::commWorld());
const label nProcs = UPstream::nProcs(UPstream::commWorld());
label val = myRank;
label worldVal = returnReduce
(
@ -108,6 +112,42 @@ int main(int argc, char *argv[])
Pout<< "value " << val
<< " (world) reduced " << worldVal
<< " (self) reduced " << selfVal << nl;
// Identical size on all procs
bitSet procUsed(nProcs);
if ((myRank % 4) == 0)
{
procUsed.set(myRank);
}
Pout<< "local procUsed " << procUsed << nl;
reduce(procUsed.data(), procUsed.size_data(), bitOrOp<unsigned>());
Pout<< "reduce procUsed " << procUsed << nl;
// Identical size on all procs
// encode as 0:empty, 1:uniform, 2:nonuniform, 3:mixed
PackedList<2> uniformity(10);
if ((myRank % 2) == 0)
{
// Every second is uniform
uniformity.set(2, 1);
uniformity.set(4, 1);
uniformity.set(6, 1);
uniformity.set(8, 1);
}
else if ((myRank % 3) == 0)
{
// Every third is nonuniform
uniformity.set(3, 2);
uniformity.set(6, 2);
uniformity.set(9, 2);
}
Pout<< "local uniform " << uniformity << nl;
reduce(uniformity.data(), uniformity.size_data(), bitOrOp<unsigned>());
Pout<< "reduce uniform " << uniformity << nl;
}
// Reductions (not using MPI intrinsics)