From 6d46c9de8e637db533e3ab211d33962cf9ad1be2 Mon Sep 17 00:00:00 2001
From: Philippe WAUTELET <philippe.wautelet@aero.obs-mip.fr>
Date: Tue, 18 Jan 2022 13:28:54 +0100
Subject: [PATCH] Philippe 18/01/2022: OpenACC: Cray compiler does not support
directives with kernels and loop in the same one
---
src/ZSOLVER/pressurez.f90 | 20 +++++++++---
.../communication.f90 | 4 ++-
.../datatypes.f90 | 3 +-
.../discretisation.f90 | 7 ++++-
.../multigrid.f90 | 6 ++--
.../profiles.f90 | 15 ++++++---
src/ZSOLVER/zsolver_inv.f90 | 31 ++++++++++---------
7 files changed, 58 insertions(+), 28 deletions(-)
diff --git a/src/ZSOLVER/pressurez.f90 b/src/ZSOLVER/pressurez.f90
index 4be2ba2ad..2d9b879ea 100644
--- a/src/ZSOLVER/pressurez.f90
+++ b/src/ZSOLVER/pressurez.f90
@@ -798,8 +798,11 @@ CALL GX_M_U_DEVICE(1,IKU,1,ZPHIT,PDXX,PDZZ,PDZX,ZDV_SOURCE)
IF ( GWEST ) THEN
!!!!!!!!!!!!!!!! FUJI compiler directive !!!!!!!!!!
!!!!!!!!!!!!!!!! FUJI compiler directive !!!!!!!!!!
+ !$acc kernels async
#ifdef MNH_COMPILER_NVHPC
- !$acc kernels loop independent collapse(2) async
+ !$acc loop independent collapse(2)
+#else
+ !$acc loop independent
#endif
DO CONCURRENT (JJ=1:IJU , JK=2:IKU-1)
ZDV_SOURCE(IIB,JJ,JK)= &
@@ -813,8 +816,11 @@ IF ( GWEST ) THEN
ENDIF
!
IF( GEAST ) THEN
+ !$acc kernels async
#ifdef MNH_COMPILER_NVHPC
- !$acc kernels loop independent collapse(2) async
+ !$acc loop independent collapse(2)
+#else
+ !$acc loop independent
#endif
DO CONCURRENT (JJ=1:IJU , JK=2:IKU-1)
ZDV_SOURCE(IIE+1,JJ,JK)= &
@@ -863,8 +869,11 @@ IF(.NOT. L2D) THEN
IF ( GSOUTH ) THEN
!!!!!!!!!!!!!!!! FUJI compiler directive !!!!!!!!!!
!!!!!!!!!!!!!!!! FUJI compiler directive !!!!!!!!!!
+ !$acc kernels async
#ifdef MNH_COMPILER_NVHPC
- !$acc kernels loop independent collapse(2) async
+ !$acc loop independent collapse(2)
+#else
+ !$acc loop independent
#endif
DO CONCURRENT (JI=1:IIU , JK=2:IKU-1)
ZDV_SOURCE(JI,IJB,JK)= &
@@ -878,8 +887,11 @@ IF(.NOT. L2D) THEN
END IF
!
IF ( GNORTH ) THEN
+ !$acc kernels async
#ifdef MNH_COMPILER_NVHPC
- !$acc kernels loop independent collapse(2) async
+ !$acc loop independent collapse(2)
+#else
+ !$acc loop independent
#endif
DO CONCURRENT (JI=1:IIU , JK=2:IKU-1)
ZDV_SOURCE(JI,IJE+1,JK)= &
diff --git a/src/ZSOLVER/tensorproductmultigrid_Source/communication.f90 b/src/ZSOLVER/tensorproductmultigrid_Source/communication.f90
index d95cde8d4..6b86c4adf 100644
--- a/src/ZSOLVER/tensorproductmultigrid_Source/communication.f90
+++ b/src/ZSOLVER/tensorproductmultigrid_Source/communication.f90
@@ -549,7 +549,8 @@ contains
local_sumt = 0.0_rl
za_st => a%st
zb_st => b%st
- !$acc kernels loop collapse(3)
+ !$acc kernels
+ !$acc loop collapse(3)
do iz=0,nz+1
do iy=a%icompy_min,a%icompy_max
do ix=a%icompx_min,a%icompx_max
@@ -558,6 +559,7 @@ contains
end do
end do
end do
+ !$acc end kernels
end if
else
if (LUseO) local_sum = 0.0_rl
diff --git a/src/ZSOLVER/tensorproductmultigrid_Source/datatypes.f90 b/src/ZSOLVER/tensorproductmultigrid_Source/datatypes.f90
index e4c2a3703..f99505180 100644
--- a/src/ZSOLVER/tensorproductmultigrid_Source/datatypes.f90
+++ b/src/ZSOLVER/tensorproductmultigrid_Source/datatypes.f90
@@ -360,7 +360,8 @@ private
if (LUseT) then
zphi_st => phi%st
- !$acc kernels loop collapse(3)
+ !$acc kernels
+ !$acc loop collapse(3)
do iz=1,nz
do iy=1,nlocaly
do ix=1,nlocalx
diff --git a/src/ZSOLVER/tensorproductmultigrid_Source/discretisation.f90 b/src/ZSOLVER/tensorproductmultigrid_Source/discretisation.f90
index b81b9dc95..fa4092f54 100644
--- a/src/ZSOLVER/tensorproductmultigrid_Source/discretisation.f90
+++ b/src/ZSOLVER/tensorproductmultigrid_Source/discretisation.f90
@@ -1578,7 +1578,10 @@ end subroutine construct_vertical_coeff
if (LUseO) u0(:,:,:) = u%s(:,:,:)
if (LUseT) then
zu_st => u%st
- !$acc kernels loop independent collapse(3)
+ !$acc kernels
+#ifdef MNH_COMPILER_NVHPC
+ !$acc loop independent collapse(3)
+#endif
do concurrent (ii=nib:nie,ij=njb:nje,ik=nkb:nke)
ut0(ii,ij,ik) = zu_st(ii,ij,ik)
end do
@@ -1674,7 +1677,9 @@ end subroutine construct_vertical_coeff
Sut0, &
Sutmp,level )
!$acc kernels
+#ifdef MNH_COMPILER_NVHPC
!$acc loop independent collapse(3)
+#endif
do concurrent ( ix=iib:iie,iy=ijb:ije,iz=1:nz )
zu_st(ix,iy,iz) = &
rho*zSutmp_st(ix,iy,iz) &
diff --git a/src/ZSOLVER/tensorproductmultigrid_Source/multigrid.f90 b/src/ZSOLVER/tensorproductmultigrid_Source/multigrid.f90
index b43fa68dd..6b55f0c13 100644
--- a/src/ZSOLVER/tensorproductmultigrid_Source/multigrid.f90
+++ b/src/ZSOLVER/tensorproductmultigrid_Source/multigrid.f90
@@ -503,7 +503,8 @@ contains
if (LUseT) then
zphifine_st => phifine%st
zphicoarse_st => phicoarse%st
- !$acc kernels loop independent collapse(3)
+ !$acc kernels
+ !$acc loop independent collapse(3)
do iz=1,nz
do iy=iy_min,iy_max
do ix=ix_min,ix_max
@@ -835,7 +836,8 @@ contains
zphifine_st => phifine%st
zphicoarse_st => phicoarse%st
- !$acc kernels loop independent collapse(5)
+ !$acc kernels
+ !$acc loop independent collapse(5)
do iz=1,nz
do diy = -1,0
do dix = -1,0
diff --git a/src/ZSOLVER/tensorproductmultigrid_Source/profiles.f90 b/src/ZSOLVER/tensorproductmultigrid_Source/profiles.f90
index 70e7508d5..3fb50577b 100644
--- a/src/ZSOLVER/tensorproductmultigrid_Source/profiles.f90
+++ b/src/ZSOLVER/tensorproductmultigrid_Source/profiles.f90
@@ -88,7 +88,8 @@ private
end if
if (LUseT) then
zu_st => u%st
- !$acc kernels loop independent collapse(3)
+ !$acc kernels
+ !$acc loop independent collapse(3)
do iz=1,nz
do iy=iy_min, iy_max
do ix=ix_min, ix_max
@@ -113,7 +114,8 @@ private
end if
if (LUseT) then
zu_st => u%st
- !$acc kernels loop independent collapse(3)
+ !$acc kernels
+ !$acc loop independent collapse(3)
do iz=1,nz
do iy=iy_min, iy_max
do ix=ix_min, ix_max
@@ -162,7 +164,8 @@ private
end do
else
zu_st => u%st
- !$acc kernels loop independent collapse(3)
+ !$acc kernels
+ !$acc loop independent collapse(3)
do iz=1,nz
do iy=iy_min, iy_max
do ix=ix_min, ix_max
@@ -225,7 +228,8 @@ private
endif
if (LUseT) then
zb_st => b%st
- !$acc kernels loop independent collapse(3)
+ !$acc kernels
+ !$acc loop independent collapse(3)
do iz=1,nz
do iy=iy_min, iy_max
do ix=ix_min, ix_max
@@ -262,7 +266,8 @@ private
end if
if (LUseT) then
zb_st => b%st
- !$acc kernels loop independent collapse(3)
+ !$acc kernels
+ !$acc loop independent collapse(3)
do iz=1,nz
do iy=iy_min, iy_max
do ix=ix_min, ix_max
diff --git a/src/ZSOLVER/zsolver_inv.f90 b/src/ZSOLVER/zsolver_inv.f90
index d15dcc1ca..ceaa670be 100644
--- a/src/ZSOLVER/zsolver_inv.f90
+++ b/src/ZSOLVER/zsolver_inv.f90
@@ -1,13 +1,8 @@
-!MNH_LIC Copyright 1994-2014 CNRS, Meteo-France and Universite Paul Sabatier
+!MNH_LIC Copyright 1994-2022 CNRS, Meteo-France and Universite Paul Sabatier
!MNH_LIC This is part of the Meso-NH software governed by the CeCILL-C licence
!MNH_LIC version 1. See LICENSE, CeCILL-C_V1-en.txt and CeCILL-C_V1-fr.txt
!MNH_LIC for details. version 1.
!-----------------------------------------------------------------
-!--------------- special set of characters for RCS information
-!-----------------------------------------------------------------
-! $Source$ $Revision$
-! MASDEV4_7 solver 2006/05/18 13:07:25
-!-----------------------------------------------------------------
! ####################
MODULE MODI_ZSOLVER_INV
! ####################
@@ -350,7 +345,8 @@ INTEGER :: ZDXM2, ZDYM2
ZDXM2 = PDXHATM*PDXHATM
!
IF (LWEST_ll(HSPLITTING='B')) THEN
- !$acc kernels loop independent collapse(2) async
+ !$acc kernels async
+ !$acc loop independent collapse(2)
DO JK=IKB,IKE
DO JJ = IJB,IJE
PF_1_Y(IIB-1,JJ,JK) = PF_1_Y(IIB,JJ,JK) - PY(IIB-1,JJ,JK)*ZDXM2/PRHOM(JK)
@@ -360,7 +356,8 @@ INTEGER :: ZDXM2, ZDYM2
END IF
!
IF (LEAST_ll(HSPLITTING='B')) THEN
- !$acc kernels loop independent collapse(2) async
+ !$acc kernels async
+ !$acc loop independent collapse(2)
DO JK=IKB,IKE
DO JJ = IJB,IJE
PF_1_Y(IIE+1,JJ,JK) = PF_1_Y(IIE,JJ,JK) + PY(IIE+1,JJ,JK)*ZDXM2/PRHOM(JK)
@@ -373,7 +370,8 @@ INTEGER :: ZDXM2, ZDYM2
! we set the solution at the corner point by the condition:
! dxm ( P ) = 0
IF (LWEST_ll(HSPLITTING='B')) THEN
- !$acc kernels loop independent async
+ !$acc kernels async
+ !$acc loop independent
DO JJ = IJB,IJE
PF_1_Y(IIB-1,JJ,IKB-1) = PF_1_Y(IIB,JJ,IKB-1)
PF_1_Y(IIB-1,JJ,IKE+1) = PF_1_Y(IIB,JJ,IKE+1)
@@ -381,7 +379,8 @@ INTEGER :: ZDXM2, ZDYM2
!$acc end kernels
END IF
IF (LEAST_ll(HSPLITTING='B')) THEN
- !$acc kernels loop independent async
+ !$acc kernels async
+ !$acc loop independent
DO JJ = IJB,IJE
PF_1_Y(IIE+1,JJ,IKB-1) = PF_1_Y(IIE,JJ,IKB-1)
PF_1_Y(IIE+1,JJ,IKE+1) = PF_1_Y(IIE,JJ,IKE+1)
@@ -407,7 +406,8 @@ INTEGER :: ZDXM2, ZDYM2
ZDYM2 = PDYHATM*PDYHATM
!
IF (LSOUTH_ll(HSPLITTING='B')) THEN
- !$acc kernels loop independent collapse(2) async
+ !$acc kernels async
+ !$acc loop independent collapse(2)
DO JK=IKB,IKE
DO JI = IIB,IIE
PF_1_Y(JI,IJB-1,JK) = PF_1_Y(JI,IJB,JK) - PY(JI,IJB-1,JK)*ZDYM2/PRHOM(JK)
@@ -417,7 +417,8 @@ INTEGER :: ZDXM2, ZDYM2
END IF
!
IF (LNORTH_ll(HSPLITTING='B')) THEN
- !$acc kernels loop independent collapse(2) async
+ !$acc kernels async
+ !$acc loop independent collapse(2)
DO JK=IKB,IKE
DO JI = IIB,IIE
PF_1_Y(JI,IJE+1,JK) = PF_1_Y(JI,IJE,JK) + PY(JI,IJE+1,JK)*ZDYM2/PRHOM(JK)
@@ -430,7 +431,8 @@ INTEGER :: ZDXM2, ZDYM2
! dym ( P ) = 0
!
IF (LSOUTH_ll(HSPLITTING='B')) THEN
- !$acc kernels loop independent async
+ !$acc kernels async
+ !$acc loop independent
DO JI = IIB,IIE
PF_1_Y(JI,IJB-1,IKB-1) = PF_1_Y(JI,IJB,IKB-1)
PF_1_Y(JI,IJB-1,IKE+1) = PF_1_Y(JI,IJB,IKE+1)
@@ -439,7 +441,8 @@ INTEGER :: ZDXM2, ZDYM2
END IF
!
IF (LNORTH_ll(HSPLITTING='B')) THEN
- !$acc kernels loop independent async
+ !$acc kernels async
+ !$acc loop independent
DO JI = IIB,IIE
PF_1_Y(JI,IJE+1,IKB-1) = PF_1_Y(JI,IJE,IKB-1)
PF_1_Y(JI,IJE+1,IKE+1) = PF_1_Y(JI,IJE,IKE+1)
--
GitLab