#!/bin/bash

## qsub -q EarlyAppAccess -A Aurora_Deployment -I -l select=1 -l walltime=60:00

#PBS -l select=16
#PBS -q EarlyAppAccess
#PBS -A LatticeQCD_aesp_CNDA
#PBS -l walltime=01:00:00
#PBS -N dwf
#PBS -k doe

#export OMP_PROC_BIND=spread
#unset OMP_PLACES

cd $PBS_O_WORKDIR

source ../sourceme.sh

cat $PBS_NODEFILE

export OMP_NUM_THREADS=3
export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1

#module load mpich/51.2/icc-all-deterministic-pmix-gpu
#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_ALLREDUCE_DEVICE_COLLECTIVE=0
export MPIR_CVAR_REDUCE_DEVICE_COLLECTIVE=0
export MPIR_CVAR_ALLREDUCE_INTRA_ALGORITHM=recursive_doubling
unset MPIR_CVAR_CH4_COLL_SELECTION_TUNING_JSON_FILE
unset MPIR_CVAR_COLL_SELECTION_TUNING_JSON_FILE
unset MPIR_CVAR_CH4_POSIX_COLL_SELECTION_TUNING_JSON_FILE

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
export MPICH_OFI_NIC_POLICY=GPU

DIR=repro.$PBS_JOBID
mkdir $DIR
cd $DIR

CMD="mpiexec -np 192 -ppn 12  -envall \
	     ../gpu_tile_compact.sh \
	     ../Test_dwf_mixedcg_prec --mpi 2.4.4.6 --grid 64.128.128.192 \
		--shm-mpi 1 --shm 4096 --device-mem 32000 --accelerator-threads 32 --seconds 3000 --debug-stdout --log Message,Iterative"
#--comms-overlap
$CMD 

grep Oops Grid.stderr.* > failures.$PBS_JOBID
rm core.*