Skip to content

Latest commit

 

History

History
155 lines (94 loc) · 4.14 KB

build-run.md

File metadata and controls

155 lines (94 loc) · 4.14 KB
Raw

Build and Run Instructions

DR Clients

Important: If using oprecord, use _apponly for applications which build with OpenMP or MPI. For the Cray compiler, set CRAYPE_LINK_TYPE=dynamic

STREAM

Use plain McCalpin STREAM: https://www.cs.virginia.edu/stream/FTP/Code/stream.c.

Compile

gcc -march=armv8.1-a+sve -O3 -ffast-math -DSTREAM_ARRAY_SIZE=4194304 -DNTIMES=2 -fno-openmp
armclang -march=armv8.1-a+sve -O3 -fsimdmath -DSTREAM_ARRAY_SIZE=4194304 -DNTIMES=2 -fno-openmp
cc -hnoomp -DSTREAM_ARRAY_SIZE=4194304 -DNTIMES=2

Run

./stream

BUDE

Use https://github.com/UoB-HPC/bude-benchmark, armv8 branch.

Compile

gcc -std=c99 -Wall -Ofast -ffast-math -mcpu=thunderx2t99+sve bude.c vec-pose-inner.c -DWGSIZE=64 -fopenmp
armclang -std=c99 -Wall -Ofast -ffast-math -mcpu=thunderx2t99+sve -fsimdmath bude.c vec-pose-inner.c -DWGSIZE=64 -fopenmp
cc -hstd=c99 -homp bude.c vec-pose-inner.c -DWGSIZE=64

Run

OMP_NUM_THREADS=1 ./bude -n 1024 -i 1

TeaLeaf

Use https://github.com/UoB-HPC/TeaLeaf, 2d folder.

Compile

make -B CC=armclang CPP=armclang++ OPTIONS='-DNO_MPI -mcpu=thunderx2t99+sve -fno-openmp'
make -B CC=gcc CPP=g++ OPTIONS='-DNO_MPI -mcpu=thunderx2t99+sve -fno-openmp'
make -B CC=cc CPP=CC COMPILER=CRAY OPTIONS='-DNO_MPI -hnoomp'

Run

Use tea_bm_1.in input.

./tealeaf

CloverLeaf

Use https://github.com/UK-MAC/CloverLeaf_ref.

Compile

make -B COMPILER=GNU MPI_COMPILER=mpifort C_MPI_COMPILER=mpicc CFLAGS_GNU="-Ofast -ffast-math -ffp-contract=fast -mcpu=thunderx2t99+sve -funroll-loops" FLAGS_GNU="-Ofast -ffast-math -ffp-contract=fast -mcpu=thunderx2t99+sve -funroll-loops"
make -B COMPILER=GNU MPI_COMPILER=mpifort C_MPI_COMPILER=mpicc CFLAGS_GNU="-Ofast -ffast-math -ffp-contract=fast -mcpu=thunderx2t99+sve -funroll-loops" FLAGS_GNU="-Ofast -ffast-math -ffp-contract=fast -mcpu=thunderx2t99+sve -funroll-loops"

make -B COMPILER=CRAY C_MPI_COMPILER=cc MPI_COMPILER=ftn
make -B COMPILER=CRAY C_MPI_COMPILER=cc MPI_COMPILER=ftn CFLAGS_CRAY="-Ofast -ffast-math -ffp-contract=fast -mcpu=thunderx2t99 -funroll-loops -fopenmp -fno-vectorize -fno-slp-vectorize" FLAGS_CRAY='-homp -hvector0 -hfp1' # for no-vec

Run

Use the small (default) clover.in, but remove profiler_on and reduce end_stepr=4.

./clover_leaf

MegaSweep

Use https://github.com/UK-MAC/mega-stream/, sweep3d folder.

Compile

 make -B COMPILER=GNU FFLAGS_GNU='-O3 -ffree-line-length-none -mcpu=thunderx2t99+sve'
 make -B COMPILER=CRAY FFLAGS_CRAY='-homp'
 make -B COMPILER=CRAY FFLAGS_CRAY='-homp -hvector0 -hfp1' # for no-vec

Run

Either set the parameters in mega-sweep3d.f90 before compiling of use CLI args (run with -h to list them):

nx = 16
ny = 16
nz = 16
ng = 16
nang = 136
nsweeps = 8
chunk = 16
ntimes = 2

For memtrace or oprecord, use:

nx = 4
ny = 4
nz = 4
chunk = 4
ntimes = 1

./mega-sweep3d
./mega-sweep3d-sve-gcc8.2 --ntimes 1 --prof --nx 4 --ny 4 --nz 4 --chunk 4

Neutral

Compile

git clone https://github.com/UoB-HPC/arch
cd arch
git checkout f19d9325d9

git clone https://github.com/UoB-HPC/neutral
cd neutral
git checkout d983598634

sed -i '/defined(__powerpc__)$/s/$/ \&\& !defined(__aarch64__)/' $SRC_DIR/Random123/features/gccfeatures.h

make COMPILER=GCC ARCH_COMPILER_CC=gcc ARCH_COMPILER_CPP=g++ CFLAGS_GCC="-std=gnu99 -Wall -fopenmp -Ofast -mcpu=thunderx2t99+sve -ffast-math -ffp-contract=fast
COMPILER=CRAY ARCH_COMPILER_CC=cc ARCH_COMPILER_CPP=CC CFLAGS_CRAY=-hfp2

Run

Edit the following in problems/csp.params (or make a copy), leaving everything else the same:

nx                400
ny                400
iterations        1

./neutral-omp3 problems/csp-small.params

MiniFMM

Use https://github.com/uob-hpc/minifmm-cpp.

Compile

make ARCH=thunderx2t99+sve
make ARCH=thunderx2t99+sve CC_GNU=armclang++

Run

./fmm.omp -n 50000 -m 1000 -e 0.5 -t 4 -c 32 # For instrace
./fmm.omp -n 10000 -m 0 -e 0.5 -t 3 -c 32 # For memtrace, smaller problem
./fmm.omp -i inputs/small.in # where the flags above go in *.in, one per line