|
| 1 | +set -e |
| 2 | +set -x |
| 3 | + |
| 4 | +PRK_TARGET="$1" |
| 5 | + |
| 6 | +# make this runtime configurable later |
| 7 | +COMPILER=gcc |
| 8 | + |
| 9 | +# Needed for Charm++ and AMPI below |
| 10 | +os=`uname` |
| 11 | +case "$os" in |
| 12 | + Darwin) |
| 13 | + export MY_CHARM_TOP=$HOME/charm/netlrts-darwin-x86_64-smp |
| 14 | + ;; |
| 15 | + Linux) |
| 16 | + #export MY_CHARM_TOP=$HOME/charm/netlrts-linux-x86_64 |
| 17 | + #export MY_CHARM_TOP=$HOME/charm/netlrts-linux-x86_64-smp |
| 18 | + export MY_CHARM_TOP=$HOME/charm/multicore-linux64 |
| 19 | + ;; |
| 20 | +esac |
| 21 | + |
| 22 | +# Needed for FG-MPI below |
| 23 | +os=`uname` |
| 24 | +case "$os" in |
| 25 | + Darwin) |
| 26 | + export MY_FGMPI_TOP=$HOME/fgmpi |
| 27 | + ;; |
| 28 | + Linux) |
| 29 | + export MY_FGMPI_TOP=/usr |
| 30 | + ;; |
| 31 | +esac |
| 32 | + |
| 33 | +case "$PRK_TARGET" in |
| 34 | + allserial) |
| 35 | + echo "Serial" |
| 36 | + sh ./travis/create-make-defs.sh $COMPILER |
| 37 | + make $PRK_TARGET |
| 38 | + export PRK_TARGET_PATH=SERIAL |
| 39 | + # widely supported |
| 40 | + $PRK_TARGET_PATH/Synch_p2p/p2p 10 1024 1024 |
| 41 | + $PRK_TARGET_PATH/Stencil/stencil 10 1024 32 |
| 42 | + $PRK_TARGET_PATH/Transpose/transpose 10 1024 32 |
| 43 | + # less support |
| 44 | + $PRK_TARGET_PATH/Reduce/reduce 10 1024 |
| 45 | + $PRK_TARGET_PATH/Random/random 64 10 16384 |
| 46 | + $PRK_TARGET_PATH/Nstream/nstream 10 16777216 32 |
| 47 | + $PRK_TARGET_PATH/Sparse/sparse 10 10 5 |
| 48 | + $PRK_TARGET_PATH/DGEMM/dgemm 10 1024 32 |
| 49 | + ;; |
| 50 | + allopenmp) |
| 51 | + echo "OpenMP" |
| 52 | + sh ./travis/create-make-defs.sh $COMPILER |
| 53 | + make $PRK_TARGET |
| 54 | + export PRK_TARGET_PATH=OPENMP |
| 55 | + export OMP_NUM_THREADS=4 |
| 56 | + # widely supported |
| 57 | + $PRK_TARGET_PATH/Synch_p2p/p2p $OMP_NUM_THREADS 10 1024 1024 |
| 58 | + $PRK_TARGET_PATH/Stencil/stencil $OMP_NUM_THREADS 10 1024 |
| 59 | + $PRK_TARGET_PATH/Transpose/transpose $OMP_NUM_THREADS 10 1024 32 |
| 60 | + # less support |
| 61 | + $PRK_TARGET_PATH/Reduce/reduce $OMP_NUM_THREADS 10 16777216 |
| 62 | + $PRK_TARGET_PATH/Nstream/nstream $OMP_NUM_THREADS 10 16777216 32 |
| 63 | + $PRK_TARGET_PATH/Sparse/sparse $OMP_NUM_THREADS 10 10 5 |
| 64 | + $PRK_TARGET_PATH/DGEMM/dgemm $OMP_NUM_THREADS 10 1024 32 |
| 65 | + # random is broken right now it seems |
| 66 | + #$PRK_TARGET_PATH/Random/random $OMP_NUM_THREADS 10 16384 32 |
| 67 | + # no serial equivalent |
| 68 | + $PRK_TARGET_PATH/Synch_global/global $OMP_NUM_THREADS 10 16384 |
| 69 | + $PRK_TARGET_PATH/RefCount_private/private $OMP_NUM_THREADS 16777216 |
| 70 | + $PRK_TARGET_PATH/RefCount_shared/shared $OMP_NUM_THREADS 16777216 1024 |
| 71 | + ;; |
| 72 | + allmpi1) |
| 73 | + echo "MPI-1" |
| 74 | + echo "MPICC=mpicc" > common/make.defs |
| 75 | + make $PRK_TARGET |
| 76 | + export PRK_TARGET_PATH=MPI1 |
| 77 | + export PRK_MPI_PROCS=4 |
| 78 | + # widely supported |
| 79 | + mpirun -n $PRK_MPI_PROCS $PRK_TARGET_PATH/Synch_p2p/p2p 10 1024 1024 |
| 80 | + mpirun -n $PRK_MPI_PROCS $PRK_TARGET_PATH/Stencil/stencil 10 1024 |
| 81 | + mpirun -n $PRK_MPI_PROCS $PRK_TARGET_PATH/Transpose/transpose 10 1024 32 |
| 82 | + # less support |
| 83 | + mpirun -n $PRK_MPI_PROCS $PRK_TARGET_PATH/Reduce/reduce 10 16777216 |
| 84 | + mpirun -n $PRK_MPI_PROCS $PRK_TARGET_PATH/Nstream/nstream 10 16777216 32 |
| 85 | + mpirun -n $PRK_MPI_PROCS $PRK_TARGET_PATH/Sparse/sparse 10 10 5 |
| 86 | + mpirun -n $PRK_MPI_PROCS $PRK_TARGET_PATH/DGEMM/dgemm 10 1024 32 1 |
| 87 | + mpirun -n $PRK_MPI_PROCS $PRK_TARGET_PATH/Random/random 32 20 |
| 88 | + # no serial equivalent |
| 89 | + mpirun -n $PRK_MPI_PROCS $PRK_TARGET_PATH/Synch_global/global 10 16384 |
| 90 | + ;; |
| 91 | + allmpiomp) |
| 92 | + echo "MPI+OpenMP" |
| 93 | + echo "MPICC=mpicc" > common/make.defs |
| 94 | + make $PRK_TARGET |
| 95 | + export PRK_TARGET_PATH=MPIOMP |
| 96 | + export PRK_MPI_PROCS=4 |
| 97 | + # widely supported |
| 98 | + mpirun -n $PRK_MPI_PROCS $PRK_TARGET_PATH/Synch_p2p/p2p 10 1024 1024 |
| 99 | + mpirun -n $PRK_MPI_PROCS $PRK_TARGET_PATH/Stencil/stencil 10 1024 |
| 100 | + mpirun -n $PRK_MPI_PROCS $PRK_TARGET_PATH/Transpose/transpose 10 1024 32 |
| 101 | + # less support |
| 102 | + mpirun -n $PRK_MPI_PROCS $PRK_TARGET_PATH/Nstream/nstream 10 16777216 32 |
| 103 | + ;; |
| 104 | + allmpirma) |
| 105 | + echo "MPI-RMA" |
| 106 | + echo "MPICC=mpicc" > common/make.defs |
| 107 | + make $PRK_TARGET |
| 108 | + export PRK_TARGET_PATH=MPIRMA |
| 109 | + export PRK_MPI_PROCS=4 |
| 110 | + # widely supported |
| 111 | + mpirun -n $PRK_MPI_PROCS $PRK_TARGET_PATH/Synch_p2p/p2p 10 1024 1024 |
| 112 | + mpirun -n $PRK_MPI_PROCS $PRK_TARGET_PATH/Stencil/stencil 10 1024 |
| 113 | + mpirun -n $PRK_MPI_PROCS $PRK_TARGET_PATH/Transpose/transpose 10 1024 32 |
| 114 | + ;; |
| 115 | + allmpishm) |
| 116 | + echo "MPI+MPI" |
| 117 | + echo "MPICC=mpicc" > common/make.defs |
| 118 | + make $PRK_TARGET |
| 119 | + export PRK_TARGET_PATH=MPISHM |
| 120 | + export PRK_MPI_PROCS=4 |
| 121 | + export PRK_MPISHM_RANKS=$(($PRK_MPI_PROCS/2)) |
| 122 | + # widely supported |
| 123 | + mpirun -n $PRK_MPI_PROCS $PRK_TARGET_PATH/Synch_p2p/p2p 10 1024 1024 |
| 124 | + mpirun -n $PRK_MPI_PROCS $PRK_TARGET_PATH/Stencil/stencil $PRK_MPISHM_RANKS 10 1024 |
| 125 | + mpirun -n $PRK_MPI_PROCS $PRK_TARGET_PATH/Transpose/transpose $PRK_MPISHM_RANKS 10 1024 32 |
| 126 | + ;; |
| 127 | + allshmem) |
| 128 | + echo "SHMEM" |
| 129 | + echo "SHMEMTOP=$HOME\nSHMEMCC=oshcc" > common/make.defs |
| 130 | + make $PRK_TARGET |
| 131 | + export PRK_TARGET_PATH=SHMEM |
| 132 | + export PRK_SHMEM_PROCS=4 |
| 133 | + # widely supported |
| 134 | + mpirun -n $PRK_SHMEM_PROCS $PRK_TARGET_PATH/Synch_p2p/p2p 10 1024 1024 |
| 135 | + mpirun -n $PRK_SHMEM_PROCS $PRK_TARGET_PATH/Stencil/stencil 10 1024 |
| 136 | + mpirun -n $PRK_SHMEM_PROCS $PRK_TARGET_PATH/Transpose/transpose 10 1024 32 |
| 137 | + ;; |
| 138 | + allupc) |
| 139 | + echo "UPC" |
| 140 | + # compiler for static thread execution, so set this prior to build |
| 141 | + echo "UPCC=/usr/local/gupc/bin/upc" > common/make.defs |
| 142 | + make $PRK_TARGET |
| 143 | + export PRK_TARGET_PATH=UPC |
| 144 | + export PRK_UPC_PROCS=4 |
| 145 | + # widely supported |
| 146 | + $PRK_TARGET_PATH/Synch_p2p/p2p -n $PRK_UPC_PROCS 10 1024 1024 |
| 147 | + $PRK_TARGET_PATH/Stencil/stencil -n $PRK_UPC_PROCS 10 1024 |
| 148 | + $PRK_TARGET_PATH/Transpose/transpose -n $PRK_UPC_PROCS 10 1024 32 |
| 149 | + ;; |
| 150 | + allcharm++) |
| 151 | + echo "Charm++" |
| 152 | + echo "CHARMTOP=$MY_CHARM_TOP" > common/make.defs |
| 153 | + make $PRK_TARGET |
| 154 | + export PRK_TARGET_PATH=CHARM++ |
| 155 | + export PRK_CHARM_PROCS=4 |
| 156 | + # widely supported | |
| 157 | + # For Charm++, the last argument is the overdecomposition factor --> \|/ |
| 158 | + $MY_CHARM_TOP/bin/charmrun $PRK_TARGET_PATH/Synch_p2p/p2p +p$PRK_CHARM_PROCS 10 1024 1024 1 |
| 159 | + $MY_CHARM_TOP/bin/charmrun $PRK_TARGET_PATH/Stencil/stencil +p$PRK_CHARM_PROCS 10 1024 1 |
| 160 | + $MY_CHARM_TOP/bin/charmrun $PRK_TARGET_PATH/Transpose/transpose +p$PRK_CHARM_PROCS 10 1024 32 1 |
| 161 | + ;; |
| 162 | + allampi) |
| 163 | + echo "Adaptive MPI (AMPI)" |
| 164 | + echo "CHARMTOP=$MY_CHARM_TOP" > common/make.defs |
| 165 | + make $PRK_TARGET |
| 166 | + export PRK_TARGET_PATH=AMPI |
| 167 | + export PRK_CHARM_PROCS=4 |
| 168 | + # widely supported |
| 169 | + $MY_CHARM_TOP/bin/charmrun $PRK_TARGET_PATH/Synch_p2p/p2p +p$PRK_CHARM_PROCS +vp$PRK_CHARM_PROCS 10 1024 1024 |
| 170 | + $MY_CHARM_TOP/bin/charmrun $PRK_TARGET_PATH/Stencil/stencil +p$PRK_CHARM_PROCS +vp$PRK_CHARM_PROCS 10 1024 |
| 171 | + $MY_CHARM_TOP/bin/charmrun $PRK_TARGET_PATH/Transpose/transpose +p$PRK_CHARM_PROCS +vp$PRK_CHARM_PROCS 10 1024 32 |
| 172 | + # less support |
| 173 | + $MY_CHARM_TOP/bin/charmrun $PRK_TARGET_PATH/Reduce/reduce +p$PRK_CHARM_PROCS +vp$PRK_CHARM_PROCS 10 16777216 |
| 174 | + $MY_CHARM_TOP/bin/charmrun $PRK_TARGET_PATH/Nstream/nstream +p$PRK_CHARM_PROCS +vp$PRK_CHARM_PROCS 10 16777216 32 |
| 175 | + $MY_CHARM_TOP/bin/charmrun $PRK_TARGET_PATH/Sparse/sparse +p$PRK_CHARM_PROCS +vp$PRK_CHARM_PROCS 10 10 5 |
| 176 | + $MY_CHARM_TOP/bin/charmrun $PRK_TARGET_PATH/DGEMM/dgemm +p$PRK_CHARM_PROCS +vp$PRK_CHARM_PROCS 10 1024 32 1 |
| 177 | + $MY_CHARM_TOP/bin/charmrun $PRK_TARGET_PATH/Random/random +p$PRK_CHARM_PROCS +vp$PRK_CHARM_PROCS 32 20 |
| 178 | + # no serial equivalent |
| 179 | + $MY_CHARM_TOP/bin/charmrun $PRK_TARGET_PATH/Synch_global/global +p$PRK_CHARM_PROCS +vp$PRK_CHARM_PROCS 10 16384 |
| 180 | + ;; |
| 181 | + allfgmpi) |
| 182 | + echo "Fine-Grain MPI (FG-MPI)" |
| 183 | + # |
| 184 | + echo "FGMPITOP=$MY_FGMPI_TOP\nFGMPICC=$MY_FGMPI_TOP/bin/mpicc" > common/make.defs |
| 185 | + make $PRK_TARGET |
| 186 | + export PRK_TARGET_PATH=FG_MPI |
| 187 | + export PRK_MPI_PROCS=2 |
| 188 | + export PRK_FGMPI_THREADS=2 |
| 189 | + # widely supported |
| 190 | + # widely supported |
| 191 | + $MY_FGMPI_TOP/bin/mpiexec -np $PRK_MPI_PROCS -nfg $PRK_FGMPI_THREADS $PRK_TARGET_PATH/Synch_p2p/p2p 10 1024 1024 |
| 192 | + $MY_FGMPI_TOP/bin/mpiexec -np $PRK_MPI_PROCS -nfg $PRK_FGMPI_THREADS $PRK_TARGET_PATH/Stencil/stencil 10 1024 |
| 193 | + $MY_FGMPI_TOP/bin/mpiexec -np $PRK_MPI_PROCS -nfg $PRK_FGMPI_THREADS $PRK_TARGET_PATH/Transpose/transpose 10 1024 32 |
| 194 | + # less support |
| 195 | + $MY_FGMPI_TOP/bin/mpiexec -np $PRK_MPI_PROCS -nfg $PRK_FGMPI_THREADS $PRK_TARGET_PATH/Reduce/reduce 10 16777216 |
| 196 | + $MY_FGMPI_TOP/bin/mpiexec -np $PRK_MPI_PROCS -nfg $PRK_FGMPI_THREADS $PRK_TARGET_PATH/Nstream/nstream 10 16777216 32 |
| 197 | + $MY_FGMPI_TOP/bin/mpiexec -np $PRK_MPI_PROCS -nfg $PRK_FGMPI_THREADS $PRK_TARGET_PATH/Sparse/sparse 10 10 5 |
| 198 | + $MY_FGMPI_TOP/bin/mpiexec -np $PRK_MPI_PROCS -nfg $PRK_FGMPI_THREADS $PRK_TARGET_PATH/DGEMM/dgemm 10 1024 32 1 |
| 199 | + $MY_FGMPI_TOP/bin/mpiexec -np $PRK_MPI_PROCS -nfg $PRK_FGMPI_THREADS $PRK_TARGET_PATH/Random/random 32 20 |
| 200 | + # no serial equivalent |
| 201 | + $MY_FGMPI_TOP/bin/mpiexec -np $PRK_MPI_PROCS -nfg $PRK_FGMPI_THREADS $PRK_TARGET_PATH/Synch_global/global 10 16384 |
| 202 | + ;; |
| 203 | + allgrappa) |
| 204 | + echo "Grappa" |
| 205 | + # compiler for static thread execution, so set this prior to build |
| 206 | + echo "GRAPPATOP=$HOME/grappa" > common/make.defs |
| 207 | + make $PRK_TARGET |
| 208 | + export PRK_TARGET_PATH=GRAPPA |
| 209 | + export PRK_MPI_PROCS=4 |
| 210 | + # widely supported |
| 211 | + mpirun -n $PRK_MPI_PROCS $PRK_TARGET_PATH/Synch_p2p/p2p 10 1024 1024 |
| 212 | + mpirun -n $PRK_MPI_PROCS $PRK_TARGET_PATH/Stencil/stencil 10 1024 |
| 213 | + mpirun -n $PRK_MPI_PROCS $PRK_TARGET_PATH/Transpose/transpose 10 1024 32 |
| 214 | + ;; |
| 215 | +esac |
0 commit comments