From 19483947cd6f85a7589a363dd91eb16d147f591a Mon Sep 17 00:00:00 2001
From: ESCOBAR Juan <escj@nuwa>
Date: Mon, 14 Mar 2022 19:05:34 +0100
Subject: [PATCH] Juan 14/03/2022:tensor/communication.f90,some GPU
optimisation & replace floor -> nint for pb with Cray Compiler
---
.../communication.f90 | 26 ++++++++++++++++---
1 file changed, 23 insertions(+), 3 deletions(-)
diff --git a/src/ZSOLVER/tensorproductmultigrid_Source/communication.f90 b/src/ZSOLVER/tensorproductmultigrid_Source/communication.f90
index 6b86c4adf..6ec3b197d 100644
--- a/src/ZSOLVER/tensorproductmultigrid_Source/communication.f90
+++ b/src/ZSOLVER/tensorproductmultigrid_Source/communication.f90
@@ -215,8 +215,10 @@ contains
call mpi_comm_rank(MPI_COMM_WORLD, rank, ierr)
i_am_master_mpi = (rank == master_rank)
! Check that nproc = 2^(2*p)
- pproc = floor(log(1.0d0*nproc)/log(4.0d0))
+ !pproc = floor(log(1.0d0*nproc)/log(4.0d0))
+ pproc = nint(log(1.0d0*nproc)/log(4.0d0))
if ( (nproc - 4**pproc) .ne. 0) then
+ print*,"Number of processors has to be 2^(2*pproc) with integer nproc,pproc=",nproc,pproc
call fatalerror("Number of processors has to be 2^(2*pproc) with integer pproc.")
end if
if (i_am_master_mpi) then
@@ -499,6 +501,9 @@ contains
! Scalar product of two fields
!==================================================================
subroutine scalarprod_mnh(m, a, b, s)
+#ifdef MNH
+ USE MODE_MPPDB
+#endif
implicit none
integer, intent(in) :: m
type(scalar3d), intent(in) :: a
@@ -516,6 +521,7 @@ contains
real(kind=rl) :: ddot
integer :: iy_min,iy_max, ix_min,ix_max
+ integer :: icompy_min,icompy_max, icompx_min,icompx_max
real , dimension(:,:,:) , pointer , contiguous :: za_st,zb_st
nlocal = a%ix_max-a%ix_min+1
@@ -526,6 +532,11 @@ contains
ix_min = a%ix_min
ix_max = a%ix_max
+ icompy_min = a%icompy_min
+ icompy_max = a%icompy_max
+ icompx_min = a%icompx_min
+ icompx_max = a%icompx_max
+
! Work out coordinates of processor
call mpi_comm_size(MPI_COMM_HORIZ,nprocs,ierr)
call mpi_comm_rank(MPI_COMM_HORIZ,rank,ierr)
@@ -552,8 +563,8 @@ contains
!$acc kernels
!$acc loop collapse(3)
do iz=0,nz+1
- do iy=a%icompy_min,a%icompy_max
- do ix=a%icompx_min,a%icompx_max
+ do iy=icompy_min,icompy_max
+ do ix=icompx_min,icompx_max
local_sumt = local_sumt &
+ za_st(ix,iy,iz)*zb_st(ix,iy,iz)
end do
@@ -598,6 +609,10 @@ contains
else
s = global_sumt
end if
+
+#ifdef MNH
+!!$ CALL MPPDB_CHECK0D_REAL_MG(s,"scalarprod_mnh")
+#endif
end subroutine scalarprod_mnh
!-------------------------------------------------------------------------------
@@ -1105,6 +1120,11 @@ contains
Gneighbour_e = (neighbour_e_rank >= 0)
Gneighbour_w = (neighbour_w_rank >= 0)
!
+ requests_ns(:) = MPI_REQUEST_NULL
+ requests_ew(:) = MPI_REQUEST_NULL
+ requests_nsT(:) = MPI_REQUEST_NULL
+ requests_ewT(:) = MPI_REQUEST_NULL
+ !
#ifdef MNH_GPUDIRECT
if (LUseT) then
!
--
GitLab