#!/bin/bash

#PBS -q EarlyAppAccess
#PBS -l select=2
#PBS -l walltime=00:20:00
#PBS -A LatticeQCD_aesp_CNDA

#export OMP_PROC_BIND=spread
#unset OMP_PLACES

cd $PBS_O_WORKDIR

source ../sourceme.sh
#module load pti-gpu


cp $PBS_NODEFILE nodefile

export OMP_NUM_THREADS=4
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, 2 nodes, 24 ranks
#
CMD="mpiexec -np 24 -ppn 12  -envall \
	     ./gpu_tile.sh \
	     ./Benchmark_comms_host_device --mpi 2.2.2.3 --grid 24.32.32.24 \
		--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 32" 
#$CMD | tee 2node.comms.hbm


CMD="mpiexec -np 24 -ppn 12  -envall \
	     ./Benchmark_dwf_fp32 --mpi 2.2.2.3 --grid 32.32.64.48 \
		--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap --debug-signals"

#for f in 1 2 3 4 5 6 7 8
for f in 1
do
echo $CMD
$CMD | tee 2node.32.32.64.48.dwf.hbm.$f
done

CMD="mpiexec -np 24 -ppn 12  -envall \
	     ./gpu_tile.sh \
	     ./Benchmark_dwf_fp32 --mpi 2.2.2.3 --grid 64.64.64.96 \
		--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
#$CMD | tee 2node.64.64.64.96.dwf.hbm