#!/bin/bash

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

#PBS -q EarlyAppAccess
#PBS -l select=256
#PBS -l walltime=01:00:00
#PBS -A LatticeQCD_aesp_CNDA

#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

#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
export MPICH_OFI_NIC_POLICY=GPU

# 12 ppn, 32 nodes, 384 ranks
#
CMD="mpiexec -np 3072 -ppn 12  -envall \
	     ./gpu_tile_compact.sh \
	     ./Benchmark_comms_host_device --mpi 8.6.8.8 --grid 32.24.32.192 \
		--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32"

$CMD 

CMD="mpiexec -np 3072 -ppn 12  -envall \
	     ./gpu_tile_compact.sh \
	     ./Benchmark_dwf_fp32 --mpi 8.8.4.12 --grid 128.128.128.768 \
		--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
$CMD | tee 256node.dwf.large