From 87c311f9cd766128d23643f12b91e40ae4a1f9b8 Mon Sep 17 00:00:00 2001
From: Juan ESCOBAR <juan.escobar@aero.obs-mip.fr>
Date: Tue, 15 Mar 2022 18:13:22 +0100
Subject: [PATCH] Juan 15/03/2022:tensor/communication.f90, Cray OpenACC
optimization , disccard collapse + do concurrent
---
.../communication.f90 | 138 +++++++++++++++---
1 file changed, 114 insertions(+), 24 deletions(-)
diff --git a/src/ZSOLVER/tensorproductmultigrid_Source/communication.f90 b/src/ZSOLVER/tensorproductmultigrid_Source/communication.f90
index 6ec3b197d..6d5c71c75 100644
--- a/src/ZSOLVER/tensorproductmultigrid_Source/communication.f90
+++ b/src/ZSOLVER/tensorproductmultigrid_Source/communication.f90
@@ -561,7 +561,9 @@ contains
za_st => a%st
zb_st => b%st
!$acc kernels
+#ifdef MNH_COMPILER_NVHPC
!$acc loop collapse(3)
+#endif
do iz=0,nz+1
do iy=icompy_min,icompy_max
do ix=icompx_min,icompx_max
@@ -1133,34 +1135,50 @@ contains
! Send to south
if (Gneighbour_s) then
ztab_halo_st_haloTin => tab_halo_st(level,m)%haloTin
- !$acc parallel loop collapse(3) async(IS_SOUTH)
+ !$acc kernels async(IS_SOUTH)
+#ifdef MNH_COMPILER_NVHPC
+ !$acc loop independent collapse(3)
+#endif
do concurrent ( ii=1:a_n,ij=1:halo_size,ik=1:nz+2 )
ztab_halo_st_haloTin(ii,ij,ik) = zst(ii,ij+a_n-halo_size,ik-1)
end do
+ !$acc end kernels
end if
! Send to north
if (Gneighbour_n) then
ztab_halo_nt_haloTin => tab_halo_nt(level,m)%haloTin
- !$acc parallel loop collapse(3) async(IS_NORTH)
+ !$acc kernels async(IS_NORTH)
+#ifdef MNH_COMPILER_NVHPC
+ !$acc loop independent collapse(3)
+#endif
do concurrent ( ii=1:a_n,ij=1:halo_size,ik=1:nz+2 )
ztab_halo_nt_haloTin(ii,ij,ik) = zst(ii,ij,ik-1)
end do
+ !$acc end kernels
end if
! Send to east
if (Gneighbour_e) then
ztab_halo_et_haloTin => tab_halo_et(level,m)%haloTin
- !$acc parallel loop collapse(3) async(IS_EAST)
+ !$acc kernels async(IS_EAST)
+#ifdef MNH_COMPILER_NVHPC
+ !$acc loop independent collapse(3)
+#endif
do concurrent ( ii=1:halo_size,ij=1:a_n+2*halo_size,ik=1:nz+2 )
ztab_halo_et_haloTin(ii,ij,ik) = zst(ii+a_n-halo_size,ij-halo_size,ik-1)
end do
+ !$acc end kernels
end if
! Send to west
if (Gneighbour_w) then
ztab_halo_wt_haloTin => tab_halo_wt(level,m)%haloTin
- !$acc parallel loop collapse(3) async(IS_WEST)
+ !$acc kernels async(IS_WEST)
+#ifdef MNH_COMPILER_NVHPC
+ !$acc loop independent collapse(3)
+#endif
do concurrent ( ii=1:halo_size,ij=1:a_n+2*halo_size,ik=1:nz+2 )
ztab_halo_wt_haloTin(ii,ij,ik) = zst(ii,ij-halo_size,ik-1)
end do
+ !$acc end kernels
end if
end if
#endif
@@ -1368,31 +1386,47 @@ contains
if (LUseT) then
if (Gneighbour_n) then
! copy north halo for GPU managed
- !$acc parallel loop collapse(3) async(IS_NORTH)
+ !$acc kernels async(IS_NORTH)
+#ifdef MNH_COMPILER_NVHPC
+ !$acc loop independent collapse(3)
+#endif
do concurrent ( ii=1:a_n,ij=1:halo_size,ik=1:nz+2 )
zst(ii,ij-halo_size,ik-1) = ztab_halo_nt_haloTout(ii,ij,ik)
end do
+ !$acc end kernels
end if
if (Gneighbour_s) then
! copy south halo for GPU managed
- !$acc parallel loop collapse(3) async(IS_SOUTH)
+ !$acc kernels async(IS_SOUTH)
+#ifdef MNH_COMPILER_NVHPC
+ !$acc loop independent collapse(3)
+#endif
do concurrent ( ii=1:a_n,ij=1:halo_size,ik=1:nz+2 )
zst(ii,ij+a_n,ik-1) = ztab_halo_st_haloTout(ii,ij,ik)
- end do
+ end do
+ !$acc end kernels
end if
if (Gneighbour_w) then
! copy west halo for GPU managed
- !$acc parallel loop collapse(3) async(IS_WEST)
+ !$acc kernels async(IS_WEST)
+#ifdef MNH_COMPILER_NVHPC
+ !$acc loop independent collapse(3)
+#endif
do concurrent ( ii=1:halo_size,ij=1:a_n+2*halo_size,ik=1:nz+2 )
zst(ii-halo_size,ij-halo_size,ik-1) = ztab_halo_wt_haloTout(ii,ij,ik)
end do
+ !$acc end kernels
end if
if (Gneighbour_e) then
! copy east halo for GPU managed
- !$acc parallel loop collapse(3) async(IS_EAST)
+ !$acc kernels async(IS_EAST)
+#ifdef MNH_COMPILER_NVHPC
+ !$acc loop independent collapse(3)
+#endif
do concurrent ( ii=1:halo_size,ij=1:a_n+2*halo_size,ik=1:nz+2 )
zst(ii+a_n,ij-halo_size,ik-1) = ztab_halo_et_haloTout(ii,ij,ik)
end do
+ !$acc end kernels
end if
! wait for async copy of send buffer to GPU
call acc_wait_haloswap_mnh()
@@ -1690,10 +1724,14 @@ contains
#ifdef MNH_GPUDIRECT
zb_st => b%st
za_st => a%st
- !$acc parallel loop collapse(3)
+ !$acc kernels
+#ifdef MNH_COMPILER_NVHPC
+ !$acc loop independent collapse(3)
+#endif
do concurrent (ii=1:a_n,ij=1:a_n,ik=1:nz+2)
zb_st(ii,ij,ik-1) = za_st(ii,ij,ik-1)
end do
+ !$acc end kernels
#else
b%st(1:a_n,1:a_n,0:nz+1) = a%st(1:a_n,1:a_n,0:nz+1)
#endif
@@ -1706,20 +1744,32 @@ contains
zb_st => b%st
! copy from buffer for GPU DIRECT
! Receive from NE
- !$acc parallel loop collapse(3)
+ !$acc kernels
+#ifdef MNH_COMPILER_NVHPC
+ !$acc loop independent collapse(3)
+#endif
do concurrent (ii=1:b_n/2,ij=1:b_n/2,ik=1:nz+2)
zb_st(ii+b_n/2,ij,ik-1) = ztab_sub_interiorT_ne_m_1_haloTout(ii,ij,ik)
end do
+ !$acc end kernels
! Receive from SW
- !$acc parallel loop collapse(3)
+ !$acc kernels
+#ifdef MNH_COMPILER_NVHPC
+ !$acc loop independent collapse(3)
+#endif
do concurrent (ii=1:b_n/2,ij=1:b_n/2,ik=1:nz+2)
zb_st(ii,ij+b_n/2,ik-1) = ztab_sub_interiorT_sw_m_1_haloTout(ii,ij,ik)
end do
+ !$acc end kernels
! Receive from SE
- !$acc parallel loop collapse(3)
+ !$acc kernels
+#ifdef MNH_COMPILER_NVHPC
+ !$acc loop independent collapse(3)
+#endif
do concurrent (ii=1:b_n/2,ij=1:b_n/2,ik=1:nz+2)
zb_st(ii+b_n/2,ij+b_n/2,ik-1) = ztab_sub_interiorT_se_m_1_haloTout(ii,ij,ik)
end do
+ !$acc end kernels
end if
#endif
end if
@@ -1738,10 +1788,14 @@ contains
if (LUseT) then
#ifdef MNH_GPUDIRECT
ztab_interiorT_ne_m_haloTin => tab_interiorT_ne(level,m)%haloTin
- !$acc parallel loop collapse(3)
+ !$acc kernels
+#ifdef MNH_COMPILER_NVHPC
+ !$acc loop independent collapse(3)
+#endif
do concurrent (ii=1:a_n,ij=1:a_n,ik=1:nz+2)
ztab_interiorT_ne_m_haloTin(ii,ij,ik) = za_st(ii,ij,ik-1)
end do
+ !$acc end kernels
!$acc host_data use_device(ztab_interiorT_ne_m_haloTin)
call mpi_send(ztab_interiorT_ne_m_haloTin,size(ztab_interiorT_ne_m_haloTin), &
MPI_DOUBLE_PRECISION,dest_rank,send_tag,MPI_COMM_HORIZ,ierr)
@@ -1770,10 +1824,14 @@ contains
if (LUseT) then
#ifdef MNH_GPUDIRECT
ztab_interiorT_sw_m_haloTin => tab_interiorT_sw(level,m)%haloTin
- !$acc parallel loop collapse(3)
+ !$acc kernels
+#ifdef MNH_COMPILER_NVHPC
+ !$acc loop independent collapse(3)
+#endif
do concurrent (ii=1:a_n,ij=1:a_n,ik=1:nz+2)
ztab_interiorT_sw_m_haloTin(ii,ij,ik) = za_st(ii,ij,ik-1)
end do
+ !$acc end kernels
!$acc host_data use_device(ztab_interiorT_sw_m_haloTin)
call mpi_send(ztab_interiorT_sw_m_haloTin,size(ztab_interiorT_sw_m_haloTin), &
MPI_DOUBLE_PRECISION,dest_rank,send_tag,MPI_COMM_HORIZ,ierr)
@@ -1802,10 +1860,14 @@ contains
if (LUseT) then
#ifdef MNH_GPUDIRECT
ztab_interiorT_se_m_haloTin => tab_interiorT_se(level,m)%haloTin
- !$acc parallel loop collapse(3)
+ !$acc kernels
+#ifdef MNH_COMPILER_NVHPC
+ !$acc loop independent collapse(3)
+#endif
do concurrent (ii=1:a_n,ij=1:a_n,ik=1:nz+2)
ztab_interiorT_se_m_haloTin(ii,ij,ik) = za_st(ii,ij,ik-1)
end do
+ !$acc end kernels
!$acc host_data use_device(ztab_interiorT_se_m_haloTin)
call mpi_send(ztab_interiorT_se_m_haloTin,size(ztab_interiorT_se_m_haloTin), &
MPI_DOUBLE_PRECISION,dest_rank,send_tag,MPI_COMM_HORIZ,ierr)
@@ -1909,10 +1971,14 @@ contains
if (LUseT) then
#ifdef MNH_GPUDIRECT
ztab_sub_interiorT_ne_m_1_haloTin => tab_sub_interiorT_ne(level,m-1)%haloTin
- !$acc parallel loop collapse(3)
+ !$acc kernels
+#ifdef MNH_COMPILER_NVHPC
+ !$acc loop independent collapse(3)
+#endif
do concurrent (ii=1:a_n/2,ij=1:a_n/2,ik=1:nz+2)
ztab_sub_interiorT_ne_m_1_haloTin(ii,ij,ik) = za_st(ii+a_n/2,ij,ik-1)
end do
+ !$acc end kernels
!$acc host_data use_device(ztab_sub_interiorT_ne_m_1_haloTin)
call mpi_isend(ztab_sub_interiorT_ne_m_1_haloTin,size(ztab_sub_interiorT_ne_m_1_haloTin), &
MPI_DOUBLE_PRECISION,dest_rank, send_tag, &
@@ -1939,10 +2005,14 @@ contains
if (LUseT) then
#ifdef MNH_GPUDIRECT
ztab_sub_interiorT_sw_m_1_haloTin => tab_sub_interiorT_sw(level,m-1)%haloTin
- !$acc parallel loop collapse(3)
+ !$acc kernels
+#ifdef MNH_COMPILER_NVHPC
+ !$acc loop independent collapse(3)
+#endif
do concurrent (ii=1:a_n/2,ij=1:a_n/2,ik=1:nz+2)
ztab_sub_interiorT_sw_m_1_haloTin(ii,ij,ik) = za_st(ii,ij+a_n/2,ik-1)
end do
+ !$acc end kernels
!$acc host_data use_device(ztab_sub_interiorT_sw_m_1_haloTin)
call mpi_isend(ztab_sub_interiorT_sw_m_1_haloTin,size(ztab_sub_interiorT_sw_m_1_haloTin), &
MPI_DOUBLE_PRECISION, dest_rank, send_tag, &
@@ -1970,10 +2040,14 @@ contains
if (LUseT) then
#ifdef MNH_GPUDIRECT
ztab_sub_interiorT_se_m_1_haloTin => tab_sub_interiorT_se(level,m-1)%haloTin
- !$acc parallel loop collapse(3)
+ !$acc kernels
+#ifdef MNH_COMPILER_NVHPC
+ !$acc loop independent collapse(3)
+#endif
do concurrent (ii=1:a_n/2,ij=1:a_n/2,ik=1:nz+2)
ztab_sub_interiorT_se_m_1_haloTin(ii,ij,ik) = za_st(ii+a_n/2,ij+a_n/2,ik-1)
end do
+ !$acc end kernels
!$acc host_data use_device(ztab_sub_interiorT_se_m_1_haloTin)
call mpi_isend(ztab_sub_interiorT_se_m_1_haloTin,size(ztab_sub_interiorT_se_m_1_haloTin), &
MPI_DOUBLE_PRECISION, dest_rank, send_tag, &
@@ -1994,10 +2068,14 @@ contains
#ifdef MNH_GPUDIRECT
zb_st => b%st
za_st => a%st
- !$acc parallel loop collapse(3)
+ !$acc kernels
+#ifdef MNH_COMPILER_NVHPC
+ !$acc loop independent collapse(3)
+#endif
do concurrent (ii=1:b_n,ij=1:b_n,ik=1:nz+2)
zb_st(ii,ij,ik-1) = za_st(ii,ij,ik-1)
end do
+ !$acc end kernels
#else
b%st(1:b_n,1:b_n,0:nz+1) = a%st(1:b_n,1:b_n,0:nz+1)
#endif
@@ -2026,10 +2104,14 @@ contains
call mpi_recv(ztab_interiorT_ne_m_haloTout,size(ztab_interiorT_ne_m_haloTout), &
MPI_DOUBLE_PRECISION,source_rank,recv_tag,MPI_COMM_HORIZ,stat,ierr)
!$acc end host_data
- !$acc parallel loop collapse(3)
+ !$acc kernels
+#ifdef MNH_COMPILER_NVHPC
+ !$acc loop independent collapse(3)
+#endif
do concurrent (ii=1:b_n,ij=1:b_n,ik=1:nz+2)
zb_st(ii,ij,ik-1) = ztab_interiorT_ne_m_haloTout(ii,ij,ik)
end do
+ !$acc end kernels
#else
call mpi_recv(b%st(1,1,0),1,interiorT(level,m),source_rank,recv_tag,MPI_COMM_HORIZ,stat,ierr)
#endif
@@ -2058,10 +2140,14 @@ contains
call mpi_recv(ztab_interiorT_sw_m_haloTout,size(ztab_interiorT_sw_m_haloTout), &
MPI_DOUBLE_PRECISION,source_rank,recv_tag,MPI_COMM_HORIZ,stat,ierr)
!$acc end host_data
- !$acc parallel loop collapse(3)
+ !$acc kernels
+#ifdef MNH_COMPILER_NVHPC
+ !$acc loop independent collapse(3)
+#endif
do concurrent (ii=1:b_n,ij=1:b_n,ik=1:nz+2)
zb_st(ii,ij,ik-1) = ztab_interiorT_sw_m_haloTout(ii,ij,ik)
end do
+ !$acc end kernels
#else
call mpi_recv(b%st(1,1,0),1,interiorT(level,m),source_rank,recv_tag,MPI_COMM_HORIZ,stat,ierr)
#endif
@@ -2090,10 +2176,14 @@ contains
call mpi_recv(ztab_interiorT_se_m_haloTout,size(ztab_interiorT_se_m_haloTout), &
MPI_DOUBLE_PRECISION,source_rank,recv_tag,MPI_COMM_HORIZ,stat,ierr)
!$acc end host_data
- !$acc parallel loop collapse(3)
+ !$acc kernels
+#ifdef MNH_COMPILER_NVHPC
+ !$acc loop independent collapse(3)
+#endif
do concurrent (ii=1:b_n,ij=1:b_n,ik=1:nz+2)
zb_st(ii,ij,ik-1) = ztab_interiorT_se_m_haloTout(ii,ij,ik)
- end do
+ end do
+ !$acc end kernels
#else
call mpi_recv(b%st(1,1,0),1,interiorT(level,m),source_rank,recv_tag,MPI_COMM_HORIZ,stat,ierr)
#endif
--
GitLab