portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-09 23:45:36 +00:00
Code Releases Activity

Merge branch 'feature/sliceSum_gpu' of https://github.com/dbollweg/Grid into feature/sliceSum_gpu

Browse Source
This commit is contained in:
dbollweg 2024-02-22 21:43:06 -05:00
commit 0a816b5509
11 changed files with 318 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
Grid/lattice/Lattice_slicesum_core.h
View File

@ -125,7 +125,7 @@ template<class vobj> inline void sliceSumReduction_cub_large(const vobj *Data, V
typedef typename vobj::vector_type vector;
const int words = sizeof(vobj)/sizeof(vector);
const int osites = rd*e1*e2;
Vector<vector>buffer(osites);
commVector<vector>buffer(osites);
vector *dat = (vector *)Data;
vector *buf = &buffer[0];
Vector<vector> lvSum_small(rd);

2
Grid/qcd/smearing/StoutSmearing.h
View File

@ -69,7 +69,7 @@ public:
/*! Construct stout smearing object from explicitly specified rho matrix */
Smear_Stout(const std::vector<double>& rho_)
: OwnedBase{new Smear_APE<Gimpl>(rho_)}, SmearBase{OwnedBase.get()} {
std::cout << GridLogDebug << "Stout smearing constructor : Smear_Stout(const std::vector<double>& " << rho_ << " )" << std::endl
std::cout << GridLogDebug << "Stout smearing constructor : Smear_Stout(const std::vector<double>& " << rho_ << " )" << std::endl;
assert(Nc == 3 && "Stout smearing currently implemented only for Nc==3");
}

19
Grid/threads/Accelerator.cc
View File

@ -7,6 +7,8 @@ uint32_t accelerator_threads=2;
uint32_t acceleratorThreads(void) {return accelerator_threads;};
void acceleratorThreads(uint32_t t) {accelerator_threads = t;};
#define ENV_LOCAL_RANK_PALS "PALS_LOCAL_RANKID"
#define ENV_RANK_PALS "PALS_RANKID"
#define ENV_LOCAL_RANK_OMPI "OMPI_COMM_WORLD_LOCAL_RANK"
#define ENV_RANK_OMPI "OMPI_COMM_WORLD_RANK"
#define ENV_LOCAL_RANK_SLURM "SLURM_LOCALID"
@ -228,8 +230,17 @@ void acceleratorInit(void)
{
rank = atoi(localRankStr);
}
if ((localRankStr = getenv(ENV_LOCAL_RANK_PALS)) != NULL)
{
rank = atoi(localRankStr);
}
if ((localRankStr = getenv(ENV_RANK_OMPI )) != NULL) { world_rank = atoi(localRankStr);}
if ((localRankStr = getenv(ENV_RANK_MVAPICH)) != NULL) { world_rank = atoi(localRankStr);}
if ((localRankStr = getenv(ENV_RANK_PALS )) != NULL) { world_rank = atoi(localRankStr);}
char hostname[HOST_NAME_MAX+1];
gethostname(hostname, HOST_NAME_MAX+1);
if ( rank==0 ) printf(" acceleratorInit world_rank %d is host %s \n",world_rank,hostname);
auto devices = cl::sycl::device::get_devices();
for(int d = 0;d<devices.size();d++){
@ -241,9 +252,10 @@ void acceleratorInit(void)
printf("AcceleratorSyclInit: " #prop ": " FMT" \n",devices[d].get_info<cl::sycl::info::device::prop>());
#define GPU_PROP(prop) GPU_PROP_FMT(prop,"%ld");
if ( world_rank == 0) {
GPU_PROP_STR(vendor);
GPU_PROP_STR(version);
GPU_PROP_STR(vendor);
GPU_PROP_STR(version);
// GPU_PROP_STR(device_type);
/*
GPU_PROP(max_compute_units);
@ -259,7 +271,8 @@ void acceleratorInit(void)
GPU_PROP(single_fp_config);
*/
// GPU_PROP(double_fp_config);
GPU_PROP(global_mem_size);
GPU_PROP(global_mem_size);
}
}
if ( world_rank == 0 ) {

20
benchmarks/Benchmark_dwf_fp32.cc
View File

@ -90,11 +90,11 @@ int main (int argc, char ** argv)
std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
for(int d=0;d<Nd;d++) CommDim[d]= (mpi[d]/shm[d])>1 ? 1 : 0;
Dirichlet[0] = 0;
Dirichlet[1] = CommDim[0]*latt4[0]/mpi[0] * shm[0];
Dirichlet[2] = CommDim[1]*latt4[1]/mpi[1] * shm[1];
Dirichlet[3] = CommDim[2]*latt4[2]/mpi[2] * shm[2];
Dirichlet[4] = CommDim[3]*latt4[3]/mpi[3] * shm[3];
// Dirichlet[0] = 0;
// Dirichlet[1] = CommDim[0]*latt4[0]/mpi[0] * shm[0];
// Dirichlet[2] = CommDim[1]*latt4[1]/mpi[1] * shm[1];
// Dirichlet[3] = CommDim[2]*latt4[2]/mpi[2] * shm[2];
// Dirichlet[4] = CommDim[3]*latt4[3]/mpi[3] * shm[3];
Benchmark(Ls,Dirichlet);
@ -105,11 +105,11 @@ int main (int argc, char ** argv)
std::cout << GridLogMessage<< "++++++++++++++++++++++++++++++++++++++++++++++++" <<std::endl;
for(int d=0;d<Nd;d++) CommDim[d]= mpi[d]>1 ? 1 : 0;
Dirichlet[0] = 0;
Dirichlet[1] = CommDim[0]*latt4[0]/mpi[0];
Dirichlet[2] = CommDim[1]*latt4[1]/mpi[1];
Dirichlet[3] = CommDim[2]*latt4[2]/mpi[2];
Dirichlet[4] = CommDim[3]*latt4[3]/mpi[3];
// Dirichlet[0] = 0;
// Dirichlet[1] = CommDim[0]*latt4[0]/mpi[0];
// Dirichlet[2] = CommDim[1]*latt4[1]/mpi[1];
// Dirichlet[3] = CommDim[2]*latt4[2]/mpi[2];
// Dirichlet[4] = CommDim[3]*latt4[3]/mpi[3];
Benchmark(Ls,Dirichlet);

56
systems/Aurora/benchmarks/bench1024.pbs Normal file
View File

@ -0,0 +1,56 @@
#!/bin/bash
## qsub -q EarlyAppAccess -A Aurora_Deployment -I -l select=1 -l walltime=60:00
#PBS -q EarlyAppAccess
#PBS -l select=1024
#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 12288 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_comms_host_device --mpi 8.6.16.16 --grid 64.48.64.284 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32"
$CMD
CMD="mpiexec -np 12288 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_dwf_fp32 --mpi 8.8.8.24 --grid 128.128.128.384 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
$CMD | tee 1024node.dwf.small
CMD="mpiexec -np 12288 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_dwf_fp32 --mpi 16.8.8.12 --grid 256.256.256.384 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
$CMD | tee 1024node.dwf

56
systems/Aurora/benchmarks/bench2048.pbs Normal file
View File

@ -0,0 +1,56 @@
#!/bin/bash
## qsub -q EarlyAppAccess -A Aurora_Deployment -I -l select=1 -l walltime=60:00
#PBS -q EarlyAppAccess
#PBS -l select=2048
#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 24576 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_comms_host_device --mpi 8.12.16.16 --grid 64.48.64.284 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32"
$CMD
CMD="mpiexec -np 24576 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_dwf_fp32 --mpi 16.8.8.24 --grid 128.128.128.384 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
$CMD | tee 2048node.dwf.small
CMD="mpiexec -np 24576 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_dwf_fp32 --mpi 16.8.8.24 --grid 256.256.256.768 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
$CMD | tee 2048node.dwf

48
systems/Aurora/benchmarks/bench256.pbs Normal file
View File

@ -0,0 +1,48 @@
#!/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

48
systems/Aurora/benchmarks/bench512.pbs Normal file
View File

@ -0,0 +1,48 @@
#!/bin/bash
## qsub -q EarlyAppAccess -A Aurora_Deployment -I -l select=1 -l walltime=60:00
#PBS -q EarlyAppAccess
#PBS -l select=512
#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 6144 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_comms_host_device --mpi 8.6.8.16 --grid 32.24.32.192 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32"
$CMD
CMD="mpiexec -np 6144 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_dwf_fp32 --mpi 8.8.8.12 --grid 256.128.128.768 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
$CMD | tee 512node.dwf.large

80
systems/Aurora/benchmarks/bench_scaling.pbs Normal file
View File

@ -0,0 +1,80 @@
#!/bin/bash
## qsub -q EarlyAppAccess -A Aurora_Deployment -I -l select=1 -l walltime=60:00
#PBS -q EarlyAppAccess
#PBS -l select=32
#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 384 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_comms_host_device --mpi 4.6.4.4 --grid 32.24.32.192 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32"
$CMD
CMD="mpiexec -np 12 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_dwf_fp32 --mpi 1.2.2.3 --grid 16.64.64.96 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
$CMD | tee 1node.dwf
CMD="mpiexec -np 24 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_dwf_fp32 --mpi 2.2.2.3 --grid 32.64.64.96 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
$CMD | tee 2node.dwf
CMD="mpiexec -np 48 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_dwf_fp32 --mpi 2.2.2.6 --grid 32.64.64.192 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
$CMD | tee 4node.dwf
CMD="mpiexec -np 96 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_dwf_fp32 --mpi 2.2.4.6 --grid 32.64.128.192 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
$CMD | tee 8node.dwf
CMD="mpiexec -np 192 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_dwf_fp32 --mpi 2.4.4.6 --grid 32.128.128.192 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
$CMD | tee 16node.dwf
CMD="mpiexec -np 384 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_dwf_fp32 --mpi 4.4.4.6 --grid 64.128.128.192 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
$CMD | tee 32node.dwf

2
systems/Aurora/benchmarks/gpu_tile_compact.sh
View File

@ -28,6 +28,6 @@ export ONEAPI_DEVICE_FILTER=gpu,level_zero
#export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE_FOR_D2D_COPY=1
#export SYCL_PI_LEVEL_ZERO_USM_RESIDENT=1
echo "rank $PALS_RANKID ; local rank $PALS_LOCAL_RANKID ; ZE_AFFINITY_MASK=$ZE_AFFINITY_MASK ; NUMA $NUMA "
#echo "rank $PALS_RANKID ; local rank $PALS_LOCAL_RANKID ; ZE_AFFINITY_MASK=$ZE_AFFINITY_MASK ; NUMA $NUMA "
numactl -m $NUMA -N $NUMAP "$@"

2
tests/core/Test_sliceSum.cc
View File

@ -260,4 +260,4 @@ int main (int argc, char ** argv) {
Grid_finalize();
return 0;
}
}