#!/bin/bash

##PBS -q LatticeQCD_aesp_CNDA
#PBS -q debug-scaling
##PBS -q prod
#PBS -l select=16
#PBS -l walltime=00:20:00
#PBS -A LatticeQCD_aesp_CNDA

cd $PBS_O_WORKDIR

source ../sourceme.sh

cp $PBS_NODEFILE nodefile

export OMP_NUM_THREADS=4
export MPICH_OFI_NIC_POLICY=GPU

#export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
#unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE
#unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE
#unset MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST
#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE=0
#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE=0
#export MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST=1
#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_BUFFER_SZ=1048576
#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_THRESHOLD=131072
#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_NUM_BUFFERS_PER_CHUNK=16
#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_MAX_NUM_BUFFERS=16

#
# Local vol 16.16.16.32
#

LX=16
LY=16
LZ=16
LT=32

NX=2
NY=2
NZ=4
NT=1

GX=2
GY=2
GZ=1
GT=3

PX=$((NX * GX ))
PY=$((NY * GY ))
PZ=$((NZ * GZ ))
PT=$((NT * GT ))

VX=$((PX * LX ))
VY=$((PY * LY ))
VZ=$((PZ * LZ ))
VT=$((PT * LT ))

NP=$((PX*PY*PZ*PT))
VOL=${VX}.${VY}.${VZ}.${VT}
AT=8
MPI=${PX}.${PY}.${PZ}.${PT}

CMD="mpiexec -np $NP -ppn 12  -envall \
	     ./gpu_tile.sh ./Benchmark_dwf_fp32 --mpi $MPI --grid $VOL \
		--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads $AT --comms-overlap "

echo VOL $VOL
echo MPI $MPI
echo NPROC $NP
echo $CMD
$CMD