Commit 117845d5 authored by Pavel Kus's avatar Pavel Kus

fixing the non-consistent timing and thus optimum choices in autotuning

by allreducing the measured time and using average over ranks insted
parent da095029
...@@ -1035,9 +1035,14 @@ module elpa_impl ...@@ -1035,9 +1035,14 @@ module elpa_impl
#else #else
integer(kind=c_int), intent(out) :: error integer(kind=c_int), intent(out) :: error
#endif #endif
integer(kind=c_int) :: error2, error3
integer :: mpierr, mpierr2, mpi_comm_parent, mpi_string_length, np_total
logical :: unfinished logical :: unfinished
integer :: i integer :: i
real(kind=C_DOUBLE) :: time_spent real(kind=C_DOUBLE) :: time_spent, sendbuf(1), recvbuf(1)
#ifdef WITH_MPI
character(len=MPI_MAX_ERROR_STRING) :: mpierr_string
if (present(error)) then if (present(error)) then
error = ELPA_OK error = ELPA_OK
...@@ -1060,11 +1065,34 @@ module elpa_impl ...@@ -1060,11 +1065,34 @@ module elpa_impl
#else #else
print *, "Cannot do autotuning without detailed timings" print *, "Cannot do autotuning without detailed timings"
! TODO check this. Do we really want to return only if error is present? And should it be ELPA_OK?
if (present(error)) then if (present(error)) then
error = ELPA_OK error = ELPA_OK
return return
endif endif
#endif #endif
#ifdef WITH_MPI
! find the average time spent .. we need a unique value on all ranks
call self%get("mpi_comm_parent", mpi_comm_parent, error2)
call self%get("num_processes", np_total, error3)
if((error2 .ne. ELPA_OK) .or. (error3 .ne. ELPA_OK)) then
print *, "Parrent communicator is not set properly. Aborting..."
if(present(error)) &
sendbuf(1) = time_spent
call MPI_Allreduce(sendbuf, recvbuf, 1, MPI_REAL8, MPI_SUM, mpi_comm_parent, mpierr)
if (mpierr .ne. MPI_SUCCESS) then
call MPI_ERROR_STRING(mpierr,mpierr_string, mpi_string_length, mpierr2)
write(error_unit,*) "MPI ERROR occured during mpi_comm_split for row communicator: ", trim(mpierr_string)
time_spent = recvbuf(1) / np_total
if (ts_impl%min_loc == -1 .or. (time_spent < ts_impl%min_val)) then if (ts_impl%min_loc == -1 .or. (time_spent < ts_impl%min_val)) then
ts_impl%min_val = time_spent ts_impl%min_val = time_spent
ts_impl%min_loc = ts_impl%current ts_impl%min_loc = ts_impl%current
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment