From 5f047804e8489cc59ea0be7ad9e1a4aca60a1fb1 Mon Sep 17 00:00:00 2001
From: Juan ESCOBAR <juan.escobar@aero.obs-mip.fr>
Date: Mon, 21 Mar 2022 18:16:06 +0100
Subject: [PATCH] Juan 21/03/2022: Cray GPU opt in tensor source , add "!dirs$
concurrent" & rewrite none optimum array syntax -> do concurrent
---
.../discretisation.f90 | 20 ++++++++++++++-----
1 file changed, 15 insertions(+), 5 deletions(-)
diff --git a/src/ZSOLVER/tensorproductmultigrid_Source/discretisation.f90 b/src/ZSOLVER/tensorproductmultigrid_Source/discretisation.f90
index a81de382a..3998269cf 100644
--- a/src/ZSOLVER/tensorproductmultigrid_Source/discretisation.f90
+++ b/src/ZSOLVER/tensorproductmultigrid_Source/discretisation.f90
@@ -797,7 +797,9 @@ end subroutine construct_vertical_coeff
#ifdef MNH_COMPILER_NVHPC
!$acc loop independent collapse(2)
#endif
+ !dir$ concurrent
do ij=ijb,ije
+ !dir$ concurrent
do ii=iib,iie
zv_st(ii,ij,iz) = zd_k(iz)* ( (-zb_k(iz)-zc_k(iz))*Tij * zu_st(ii,ij,iz ) &
+zb_k(iz) *Tij * zu_st(ii,ij,iz+1) )
@@ -808,7 +810,9 @@ end subroutine construct_vertical_coeff
#ifdef MNH_COMPILER_NVHPC
!$acc loop independent collapse(2)
#endif
+ !dir$ concurrent
do ij=ijb,ije
+ !dir$ concurrent
do ii=iib,iie
zv_st(ii,ij,iz) = zd_k(iz)* ( ((-zb_k(iz)-zc_k(iz))*Tij - 4.0_rl ) * zu_st(ii,ij,iz) &
+zb_k(iz) *Tij * zu_st(ii,ij,iz+1) &
@@ -826,7 +830,9 @@ end subroutine construct_vertical_coeff
#ifdef MNH_COMPILER_NVHPC
!$acc loop independent collapse(2)
#endif
+ !dir$ concurrent
do ij=ijb,ije
+ !dir$ concurrent
do ii=iib,iie
zv_st(ii,ij,iz) = zd_k(iz)* ( (-zb_k(iz)-zc_k(iz))*Tij * zu_st(ii,ij,iz) &
+zc_k(iz) *Tij * zu_st(ii,ij,iz-1) )
@@ -2097,11 +2103,13 @@ end subroutine line_Jacobi_mnh
iz=1
zSr_st(iib:iie,ijb:ije,iz) = zb_st(iib:iie,ijb:ije,iz)
do iz=2,nz-1
- zSr_st(iib:iie,ijb:ije,iz) = zb_st(iib:iie,ijb:ije,iz) - zd_k(iz) * ( &
- zSu_in_st(iib+1:iie+1,ijb:ije,iz) + &
- zSu_in_st(iib-1:iie-1,ijb:ije,iz) + &
- zSu_in_st(iib:iie,ijb+1:ije+1,iz) + &
- zSu_in_st(iib:iie,ijb-1:ije-1,iz) )
+ do concurrent(ii=iib:iie,ij=ijb:ije)
+ zSr_st(ii,ij,iz) = zb_st(ii,ij,iz) - zd_k(iz) * ( &
+ zSu_in_st(ii+1,ij,iz) + &
+ zSu_in_st(ii-1,ij,iz) + &
+ zSu_in_st(ii,ij+1,iz) + &
+ zSu_in_st(ii,ij-1,iz) )
+ end do
end do
iz=nz
zSr_st(iib:iie,ijb:ije,iz) = zb_st(iib:iie,ijb:ije,iz)
@@ -2116,7 +2124,9 @@ end subroutine line_Jacobi_mnh
#ifdef MNH_COMPILER_NVHPC
!$acc loop independent collapse(2)
#endif
+ !dir$ concurrent
do ij=ijb,ije
+ !dir$ concurrent
do ii=iib,iie
zSu_out_st(ii,ij,iz) = (zSr_st(ii,ij,iz) / (Tijs*zd_k(iz)) &
- zSu_out_st(ii,ij,iz-1)*zc_k(iz)) / tmp_k(iz)
--
GitLab