cwpearson/astaroth
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

MPI_Allreduce -> MPI_Reduce for MPI reductions + benchmark batch script

Browse Source 
Slightly ugly because this changes the benchmark behaviour slightly
However we now have a way to run batch benchmarks from one script, no need to generate new ones
This commit is contained in:
Oskar Lappi
2020-06-06 22:53:08 +03:00
parent eb05e02793
commit 53b48bb8ce
2 changed files with 53 additions and 39 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
src/core/device.cc
View File

@@ -1645,23 +1645,22 @@ acMPIReduceScal(const AcReal local_result, const ReductionType rtype, AcReal* re
MPI_Datatype datatype = MPI_FLOAT;
#endif
/*
int rank;
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
*/
int world_size;
MPI_Comm_size(MPI_COMM_WORLD, &world_size);
AcReal mpi_res;
MPI_Allreduce(&local_result, &mpi_res, 1, datatype, op, MPI_COMM_WORLD);
if (rtype == RTYPE_RMS || rtype == RTYPE_RMS_EXP) {
const AcReal inv_n = AcReal(1.) / (grid.nn.x * grid.decomposition.x * grid.nn.y *
MPI_Reduce(&local_result, &mpi_res, 1, datatype, op, 0, MPI_COMM_WORLD);
if (rank == 0){
if (rtype == RTYPE_RMS || rtype == RTYPE_RMS_EXP) {
const AcReal inv_n = AcReal(1.) / (grid.nn.x * grid.decomposition.x * grid.nn.y *
grid.decomposition.y * grid.nn.z * grid.decomposition.z);
mpi_res = sqrt(inv_n * mpi_res);
mpi_res = sqrt(inv_n * mpi_res);
}
*result = mpi_res;
}
*result = mpi_res;
return AC_SUCCESS;
}