Add RIKYU support for MHDTurbulence

programs/MHDTurbulence/build.sh

@@ -36,6 +36,23 @@ case "$system" in
 	echo "Executable is "${BIN}" and copied to "${artdir}
 	cp ../exe/$BIN ../../${artdir}
 	;;
+    RIKYU)
+	module load nvhpc-hpcx-cuda13/26.5
+	# Portability fixes; skipped when already upstream
+	grep -q "integer,dimension(8) :: seed" src_f90_omp_host/main.f90 || \
+	    patch -p1 < ../programs/${code}/patches/random_seed_put_size.patch
+	grep -q "acc_init(acc_device_nvidia)" src_f90_acc_device/main.f90 || \
+	    patch -p1 < ../programs/${code}/patches/acc_init_before_mpi.patch
+	# ntiles is compile-time: one binary per list.csv rank count
+	cd src_f90_acc_device
+	for tiles in 1 4 8; do
+	    make clean >/dev/null 2>&1 || true
+	    sed -i "s/ntiles(3) = \[.*,.*,.*\]/ntiles(3) = [ ${tiles},1,1 ]/" config.f90
+	    make > /dev/null
+	    echo "Executable is "${BIN}".t"${tiles}" and copied to "${artdir}
+	    cp ../exe/$BIN ../../${artdir}/${BIN}.t${tiles}
+	done
+	;;
 # in the future, we may add this
 #    MiyabiG/OpenMP)
 #	cd src_f90_omp_device

programs/MHDTurbulence/list.csv

@@ -2,3 +2,6 @@ system,enable,nodes,numproc_node,nthreads,elapse
 Fugaku,no,1,4,12,0:10:00
 MiyabiC,yes,1,96,1,0:10:00
 MiyabiG,yes,1,1,1,0:10:00
+RIKYU,yes,1,1,1,0:10:00
+RIKYU,yes,1,4,1,0:10:00
+RIKYU,yes,2,4,1,0:10:00

programs/MHDTurbulence/patches/README.md

@@ -0,0 +1,22 @@
+# Patches
+
+Applied by `build.sh` only when the fix is not already present in the
+fetched source (grep-guarded), so they become no-ops once upstream
+merges them. Both originate from `william-dawson/MHDTurbulence`
+branch `gfortran-port` (candidate upstream PRs):
+
+- `random_seed_put_size.patch` — commit `d9a19d9`: `random_seed(PUT=)`
+  requires the array to match the compiler's seed size (8 for gfortran);
+  the shipped size-2 array breaks strict compilers. Physically inert
+  (the seeded RNG feeds a disabled perturbation, `rrv=0`).
+- `acc_init_before_mpi.patch` — commit `4db3adb`: initialize the CUDA
+  context before `MPI_Init`. Without it, multi-node runs hang silently
+  in the first `BoundaryCondition` (UCX "cannot find remote protocol"
+  on the device-buffer `MPI_ISEND`, then `MPI_WAITALL` on a dead
+  request; the app's `MPI_ERRORS_RETURN` handler means the failed ISEND
+  returns an error code that is never checked).
+
+Launch side (in `run.sh`, RIKYU case): each rank pins one GPU via
+`CUDA_VISIBLE_DEVICES=$OMPI_COMM_WORLD_LOCAL_RANK`; combined with the
+patch this is required for multi-node GPU runs. Validated 2026-09-03 on
+Rikyu (8 GPUs / 2 nodes, results bit-consistent with 1-GPU runs).

programs/MHDTurbulence/patches/acc_init_before_mpi.patch

@@ -0,0 +1,19 @@
+diff --git a/src_f90_acc_device/main.f90 b/src_f90_acc_device/main.f90
+index c808716..3356a5e 100644
+--- a/src_f90_acc_device/main.f90
++++ b/src_f90_acc_device/main.f90
+@@ -1,5 +1,6 @@
+ program main
+   use omp_lib
++  use openacc
+   use basicmod
+   use mpimod
+   use boundarymod
+@@ -9,6 +10,7 @@ program main
+   logical::is_final
+   logical,parameter:: forceoutput=.true., usualoutput=.false.
+   data is_final /.false./
++  call acc_init(acc_device_nvidia)
+   call InitializeMPI
+   if(myid_w == 0) print *, "setup grids and fields"
+   if(myid_w == 0) print *, "grid size for x y z",ngrid1*ntiles(1),ngrid2*ntiles(2),ngrid3*ntiles(3)

programs/MHDTurbulence/patches/random_seed_put_size.patch

@@ -0,0 +1,50 @@
+diff --git a/src_f90_acc_device/main.f90 b/src_f90_acc_device/main.f90
+index 669188c..c808716 100644
+--- a/src_f90_acc_device/main.f90
++++ b/src_f90_acc_device/main.f90
+@@ -123,7 +123,7 @@ subroutine GenerateProblem
+   real(8):: pi
+   real(8):: den, B0, rho1, rho2, dv, wid, sig
+
+-  integer,dimension(2) :: seed
++  integer,dimension(8) :: seed
+   real(8),dimension(1) :: rnum
+   real(8),parameter :: rrv =0.0d-2
+   real(8):: dv_harm
+@@ -169,9 +169,10 @@ subroutine GenerateProblem
+
+   if(myid_w == 0) write(6,*) rrv*100.0d0 &
+        & , "% of Randam Perturbation imposed on velocity"
++  seed = 0
+   seed(1) = 1
+   seed(2) = 1 + myid_w*in*jn*kn
+-  call random_seed(PUT=seed(1:2))
++  call random_seed(PUT=seed)
+
+ ! pert     
+   do k=ks,ke 
+diff --git a/src_f90_omp_host/main.f90 b/src_f90_omp_host/main.f90
+index 62cc30b..d27d395 100644
+--- a/src_f90_omp_host/main.f90
++++ b/src_f90_omp_host/main.f90
+@@ -120,7 +120,7 @@ subroutine GenerateProblem
+   real(8):: pi
+   real(8):: den, B0, rho1, rho2, dv, wid, sig
+
+-  integer,dimension(2) :: seed
++  integer,dimension(8) :: seed
+   real(8),dimension(1) :: rnum
+   real(8),parameter :: rrv =0.0d-2
+
+@@ -163,9 +163,10 @@ subroutine GenerateProblem
+
+   if(myid_w == 0) write(6,*) rrv*100.0d0 &
+        & , "% of Randam Perturbation imposed on velocity"
++  seed = 0
+   seed(1) = 1
+   seed(2) = 1 + myid_w*in*jn*kn
+-  call random_seed(PUT=seed(1:2))
++  call random_seed(PUT=seed)
+
+ ! pert     
+   do k=ks,ke 

programs/MHDTurbulence/run.sh

@@ -40,6 +40,20 @@ case "$system" in
 	elapsed=$(grep "sim time \[s\]:" ${LOG} | awk '{print $4}' | tail -n 1)
 	tcc=$(grep "time/count/cell" ${LOG} | awk '{print $2}' | tail -n 1)
 	;;
+    RIKYU)
+	module load nvhpc-hpcx-cuda13/26.5
+	exedir=exe
+	mkdir -p $code/$exedir/
+	NP=$((nodes * numproc_node))
+	cp $artdir/${BIN}.t${NP} $code/$exedir/${BIN}
+	cd $code/$exedir/
+	echo "code is executed in "$code/$exedir/
+	# CUDA context must exist before MPI_Init and each rank must pin one
+	# GPU, or the inter-node device exchange hangs (UCX protocol failure)
+	mpirun -np "$NP" bash -c 'export CUDA_VISIBLE_DEVICES=$((OMPI_COMM_WORLD_LOCAL_RANK % 4)); exec ./Simulation.x' > ${LOG} 2>&1
+	elapsed=$(grep "sim time \[s\]:" ${LOG} | awk '{print $4}' | tail -n 1)
+	tcc=$(grep "time/count/cell" ${LOG} | awk '{print $2}' | tail -n 1)
+	;;
     *)
 	echo "Unknown system: $system"
 	exit 1