Merge branch PR #55 into naming_scheme

[ROCm/rocshmem commit: d8ce066adc]
This commit is contained in:
avinashkethineedi
2024-12-04 21:46:38 +00:00
180 changed files with 4436 additions and 4437 deletions
@@ -47,88 +47,88 @@ case $2 in
###########################################################################
*"serial")
echo "get_n2_w1_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -s 1048576 -a 0 > $3/get_n2_w1_z1_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -s 1048576 -a 0 > $3/get_n2_w1_z1_1MB.log
check get_n2_w1_z1_1MB
echo "getnbi_n2_w1_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -s 1048576 -a 1 > $3/getnbi_n2_w1_z1_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -s 1048576 -a 1 > $3/getnbi_n2_w1_z1_1MB.log
check getnbi_n2_w1_z1_1MB
echo "put_n2_w1_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -s 1048576 -a 2 > $3/put_n2_w1_z1_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -s 1048576 -a 2 > $3/put_n2_w1_z1_1MB.log
check put_n2_w1_z1_1MB
echo "putnbi_n2_w1_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -s 1048576 -a 3 > $3/putnbi_n2_w1_z1_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -s 1048576 -a 3 > $3/putnbi_n2_w1_z1_1MB.log
check putnbi_n2_w1_z1_1MB
echo "wg_get_n2_w1_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 1048576 -a 28 > $3/wg_get_n2_w1_z64_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 1048576 -a 28 > $3/wg_get_n2_w1_z64_1MB.log
check wg_get_n2_w1_z1_1MB
echo "wg_getnbi_n2_w1_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 1048576 -a 29 > $3/wg_getnbi_n2_w1_z64_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 1048576 -a 29 > $3/wg_getnbi_n2_w1_z64_1MB.log
check wg_getnbi_n2_w1_z1_1MB
echo "wg_put_n2_w1_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 1048576 -a 30 > $3/wg_put_n2_w1_z64_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 1048576 -a 30 > $3/wg_put_n2_w1_z64_1MB.log
check wg_put_n2_w1_z1_1MB
echo "wg_putnbi_n2_w1_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 1048576 -a 31 > $3/wg_putnbi_n2_w1_z64_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 1048576 -a 31 > $3/wg_putnbi_n2_w1_z64_1MB.log
check wg_putnbi_n2_w1_z1_1MB
echo "wg_get_tiled_n2_w1_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 64 -s 1048576 -a 28 > $3/wg_get_tiled_n2_w2_z64_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 64 -s 1048576 -a 28 > $3/wg_get_tiled_n2_w2_z64_1MB.log
check wg_get_tiled_n2_w1_z1_1MB
echo "wg_getnbi_tiled_n2_w1_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 64 -s 1048576 -a 29 > $3/wg_getnbi_tiled_n2_w2_z64_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 64 -s 1048576 -a 29 > $3/wg_getnbi_tiled_n2_w2_z64_1MB.log
check wg_getnbi_tiled_n2_w1_z1_1MB
echo "wg_put_tiled_n2_w1_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 64 -s 1048576 -a 30 > $3/wg_put_tiled_n2_w2_z64_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 64 -s 1048576 -a 30 > $3/wg_put_tiled_n2_w2_z64_1MB.log
check wg_put_tiled_n2_w1_z1_1MB
echo "wg_putnbi_tiled_n2_w1_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 64 -s 1048576 -a 31 > $3/wg_putnbi_tiled_n2_w2_z64_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 64 -s 1048576 -a 31 > $3/wg_putnbi_tiled_n2_w2_z64_1MB.log
check wg_putnbi_tiled_n2_w1_z1_1MB
echo "wave_get_n2_w1_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 1048576 -a 32 > $3/wave_get_n2_w1_z64_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 1048576 -a 32 > $3/wave_get_n2_w1_z64_1MB.log
check wave_get_n2_w1_z1_1MB
echo "wave_getnbi_n2_w1_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 1048576 -a 33 > $3/wave_getnbi_n2_w1_z64_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 1048576 -a 33 > $3/wave_getnbi_n2_w1_z64_1MB.log
check wave_getnbi_n2_w1_z1_1MB
echo "wave_put_n2_w1_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 1048576 -a 34 > $3/wave_put_n2_w1_z64_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 1048576 -a 34 > $3/wave_put_n2_w1_z64_1MB.log
check wave_put_n2_w1_z1_1MB
echo "wave_putnbi_n2_w1_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 1048576 -a 35 > $3/wave_putnbi_n2_w1_z64_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 1048576 -a 35 > $3/wave_putnbi_n2_w1_z64_1MB.log
check wave_putnbi_n2_w1_z1_1MB
echo "wave_get_tiled_n2_w1_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 128 -s 1048576 -a 32 > $3/wave_get_tiled_n2_w2_z128_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 128 -s 1048576 -a 32 > $3/wave_get_tiled_n2_w2_z128_1MB.log
check wave_get_tiled_n2_w1_z1_1MB
echo "wave_getnbi_tiled_n2_w1_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 128 -s 1048576 -a 33 > $3/wave_getnbi_tiled_n2_w2_z128_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 128 -s 1048576 -a 33 > $3/wave_getnbi_tiled_n2_w2_z128_1MB.log
check wave_getnbi_tiled_n2_w1_z1_1MB
echo "wave_put_tiled_n2_w1_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 128 -s 1048576 -a 34 > $3/wave_put_tiled_n2_w2_z128_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 128 -s 1048576 -a 34 > $3/wave_put_tiled_n2_w2_z128_1MB.log
check wave_put_tiled_n2_w1_z1_1MB
echo "wave_putnbi_tiled_n2_w1_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 128 -s 1048576 -a 35 > $3/wave_putnbi_tiled_n2_w2_z128_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 128 -s 1048576 -a 35 > $3/wave_putnbi_tiled_n2_w2_z128_1MB.log
check wave_putnbi_tiled_n2_w1_z1_1MB
echo "amofadd_n2_w1_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 6 > $3/amofadd_n2_w1_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 6 > $3/amofadd_n2_w1_z1.log
check amofadd_n2_w1_z1
echo "amofinc_n2_w1_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 7 > $3/amofinc_n2_w1_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 7 > $3/amofinc_n2_w1_z1.log
check amofinc_n2_w1_z1
echo "amofetch_n2_w1_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 8 > $3/amofetch_n2_w1_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 8 > $3/amofetch_n2_w1_z1.log
check amofetch_n2_w1_z1
echo "amofcswap_n2_w1_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 9 > $3/amofcswap_n2_w1_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 9 > $3/amofcswap_n2_w1_z1.log
check amofcswap_n2_w1_z1
echo "amoadd_n2_w1_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 10 > $3/amoadd_n2_w1_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 10 > $3/amoadd_n2_w1_z1.log
check amoadd_n2_w1_z1
echo "amoinc_n2_w1_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 11 > $3/amoinc_n2_w1_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 11 > $3/amoinc_n2_w1_z1.log
check amoinc_n2_w1_z1
# echo "pingpong_n2_w1"
# ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -a 14 > $3/pingpong_n2_w1.log
# ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -a 14 > $3/pingpong_n2_w1.log
# check pingpong_n2_w1
echo "amoset_n2_w1_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 44 > $3/amoset_n2_w1_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 44 > $3/amoset_n2_w1_z1.log
check amoset_n2_w1_z1
;;
@@ -137,94 +137,94 @@ case $2 in
###########################################################################
*"short")
echo "get_n2_w16_z128_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 128 -s 8 -a 0 > $3/get_n2_w16_z128_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 128 -s 8 -a 0 > $3/get_n2_w16_z128_8B.log
check get_n2_w16_z128_8B
echo "getnbi_n2_w16_z128_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 128 -s 8 -a 1 > $3/getnbi_n2_w16_z128_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 128 -s 8 -a 1 > $3/getnbi_n2_w16_z128_8B.log
check getnbi_n2_w16_z128_8B
echo "put_n2_w16_z128_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 128 -s 8 -a 2 > $3/put_n2_w16_z128_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 128 -s 8 -a 2 > $3/put_n2_w16_z128_8B.log
check put_n2_w16_z128_8B
echo "putnbi_n2_w16_z128_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 128 -s 8 -a 3 > $3/putnbi_n2_w16_z128_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 128 -s 8 -a 3 > $3/putnbi_n2_w16_z128_8B.log
check putnbi_n2_w16_z128_8B
echo "wg_get_n2_w1_z64_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 8 -a 28 > $3/wg_get_n2_w1_z64_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 8 -a 28 > $3/wg_get_n2_w1_z64_8B.log
check wg_get_n2_w1_z64_8B
echo "wg_getnbi_n2_w1_z64_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 8 -a 29 > $3/wg_getnbi_n2_w1_z64_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 8 -a 29 > $3/wg_getnbi_n2_w1_z64_8B.log
check wg_getnbi_n2_w1_z64_8B
echo "wg_put_n2_w1_z64_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 8 -a 30 > $3/wg_put_n2_w1_z64_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 8 -a 30 > $3/wg_put_n2_w1_z64_8B.log
check wg_put_n2_w1_z64_8B
echo "wg_putnbi_n2_w1_z64_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 8 -a 31 > $3/wg_putnbi_n2_w1_z64_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 8 -a 31 > $3/wg_putnbi_n2_w1_z64_8B.log
check wg_putnbi_n2_w1_z64_8B
echo "wg_get_tiled_n2_w16_z64_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 64 -s 8 -a 28 > $3/wg_get_tiled_n2_w16_z64_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 64 -s 8 -a 28 > $3/wg_get_tiled_n2_w16_z64_8B.log
check wg_get_tiled_n2_w16_z64_8B
echo "wg_getnbi_tiled_n2_w16_z64_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 64 -s 8 -a 29 > $3/wg_getnbi_tiled_n2_w16_z64_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 64 -s 8 -a 29 > $3/wg_getnbi_tiled_n2_w16_z64_8B.log
check wg_getnbi_tiled_n2_w16_z64_8B
echo "wg_put_tiled_n2_w16_z64_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 64 -s 8 -a 30 > $3/wg_put_tiled_n2_w16_z64_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 64 -s 8 -a 30 > $3/wg_put_tiled_n2_w16_z64_8B.log
check wg_put_tiled_n2_w16_z64_8B
echo "wg_putnbi_tiled_n2_w16_z64_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 64 -s 8 -a 31 > $3/wg_putnbi_tiled_n2_w16_z64_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 64 -s 8 -a 31 > $3/wg_putnbi_tiled_n2_w16_z64_8B.log
check wg_putnbi_tiled_n2_w16_z64_8B
echo "wave_get_n2_w1_z64_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 8 -a 32 > $3/wave_get_n2_w1_z64_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 8 -a 32 > $3/wave_get_n2_w1_z64_8B.log
check wave_get_n2_w1_z64_8B
echo "wave_getnbi_n2_w1_z64_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 8 -a 33 > $3/wave_getnbi_n2_w1_z64_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 8 -a 33 > $3/wave_getnbi_n2_w1_z64_8B.log
check wave_getnbi_n2_w1_z64_8B
echo "wave_put_n2_w1_z64_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 8 -a 34 > $3/wave_put_n2_w1_z64_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 8 -a 34 > $3/wave_put_n2_w1_z64_8B.log
check wave_put_n2_w1_z64_8B
echo "wave_putnbi_n2_w1_z64_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 8 -a 35 > $3/wave_putnbi_n2_w1_z64_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 64 -s 8 -a 35 > $3/wave_putnbi_n2_w1_z64_8B.log
check wave_putnbi_n2_w1_z64_8B
echo "wave_get_tiled_n2_w16_z128_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 128 -s 8 -a 32 > $3/wave_get_tiled_n2_w16_z128_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 128 -s 8 -a 32 > $3/wave_get_tiled_n2_w16_z128_8B.log
check wave_get_tiled_n2_w16_z128_8B
echo "wave_getnbi_tiled_n2_w16_z128_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 128 -s 8 -a 33 > $3/wave_getnbi_tiled_n2_w16_z128_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 128 -s 8 -a 33 > $3/wave_getnbi_tiled_n2_w16_z128_8B.log
check wave_getnbi_tiled_n2_w16_z128_8B
echo "wave_put_tiled_n2_w16_z128_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 128 -s 8 -a 34 > $3/wave_put_tiled_n2_w16_z128_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 128 -s 8 -a 34 > $3/wave_put_tiled_n2_w16_z128_8B.log
check wave_put_tiled_n2_w16_z128_8B
echo "wave_putnbi_tiled_n2_w16_z128_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 128 -s 8 -a 35 > $3/wave_putnbi_tiled_n2_w16_z128_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 128 -s 8 -a 35 > $3/wave_putnbi_tiled_n2_w16_z128_8B.log
check wave_putnbi_tiled_n2_w16_z128_8B
echo "amofadd_n2_w8_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 6 > $3/amofadd_n2_w8_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 6 > $3/amofadd_n2_w8_z1.log
check amofadd_n2_w8_z1
echo "amofinc_n2_w8_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 7 > $3/amofinc_n2_w8_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 7 > $3/amofinc_n2_w8_z1.log
check amofinc_n2_w8_z1
echo "amofetch_n2_w8_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 8 > $3/amofetch_n2_w8_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 8 > $3/amofetch_n2_w8_z1.log
check amofetch_n2_w8_z1
echo "amofcswap_n2_w8_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 9 > $3/amofcswap_n2_w8_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 9 > $3/amofcswap_n2_w8_z1.log
check amofcswap_n2_w8_z1
echo "amoadd_n2_w8_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 10 > $3/amoadd_n2_w8_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 10 > $3/amoadd_n2_w8_z1.log
check amoadd_n2_w8_z1
echo "amoinc_n2_w8_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 11 > $3/amoinc_n2_w8_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 11 > $3/amoinc_n2_w8_z1.log
check amoinc_n2_w8_z1
# echo "pingpong_n2_w1"
# ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -a 14 > $3/pingpong_n2_w1.log
# ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -a 14 > $3/pingpong_n2_w1.log
# check pingpong_n2_w1
echo "amoset_n2_w8_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 44 > $3/amoset_n2_w8_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 44 > $3/amoset_n2_w8_z1.log
check amoset_n2_w8_z1
echo "putsignal_n2_w1_z1"
mpirun -np 2 $1 -w 1 -z 1 -s 1048576 -a 53 > $3/putsignal_n2_w1_z1.log
check putsignal_n2_w1_z1
echo "putsignalwg_n2_w2_z32"
ROC_SHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 32 -s 1048576 -a 55 > $3/putsignalwg_n2_w2_z32.log
ROCSHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 32 -s 1048576 -a 55 > $3/putsignalwg_n2_w2_z32.log
check putsignalwg_n2_w2_z32
echo "putsignalwave_n2_w1_z32"
mpirun -np 2 $1 -w 1 -z 32 -s 1048576 -a 55 > $3/putsignalwave_n2_w1_z32.log
@@ -233,7 +233,7 @@ case $2 in
mpirun -np 2 $1 -w 1 -z 1 -s 1048576 -a 56 > $3/putsignalnbi_n2_w1_z1.log
check putsignalnbi_n2_w1_z1
echo "putsignalnbiwg"
ROC_SHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 32 -s 1048576 -a 57 > $3/putsignalnbiwg_n2_w2_z32.log
ROCSHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 32 -s 1048576 -a 57 > $3/putsignalnbiwg_n2_w2_z32.log
check putsignalnbiwg_n2_w2_z32
echo "putsignalnbiwave"
mpirun -np 2 $1 -w 1 -z 32 -s 1048576 -a 58 > $3/putsignalnbiwave_n2_w1_z32.log
@@ -242,7 +242,7 @@ case $2 in
mpirun -np 2 $1 -w 1 -z 1 -a 59 > $3/signalfetch_n2_w1_z1.log
check signalfetch_n2_w1_z1
echo "signalfetchwg"
ROC_SHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 32 -a 60 > $3/signalfetchwg_n2_w2_z32.log
ROCSHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 32 -a 60 > $3/signalfetchwg_n2_w2_z32.log
check signalfetchwg_n2_w2_z32
echo "signalfetchwave"
mpirun -np 2 $1 -w 1 -z 32 -a 60 > $3/signalfetchwave_n2_w2_z32.log
@@ -255,246 +255,246 @@ case $2 in
*"exhaustive")
############################### GET ###################################
echo "get_n2_w1_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -s 1048576 -a 0 > $3/get_n2_w1_z1_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -s 1048576 -a 0 > $3/get_n2_w1_z1_1MB.log
check get_n2_w1_z1_1MB
echo "get_n2_w1_z1024_512B"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1024 -s 512 -a 0 > $3/get_n2_w1_z1024_512B.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1024 -s 512 -a 0 > $3/get_n2_w1_z1024_512B.log
check get_n2_w1_z1024_512B
echo "get_n2_w8_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -s 1048576 -a 0 > $3/get_n2_w8_z1_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -s 1048576 -a 0 > $3/get_n2_w8_z1_1MB.log
check get_n2_w8_z1_1MB
echo "get_n2_w16_z128_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 128 -s 8 -a 0 > $3/get_n2_w16_z128_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 128 -s 8 -a 0 > $3/get_n2_w16_z128_8B.log
check get_n2_w16_z128_8B
echo "get_n2_w32_z256_512B"
ROC_SHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -z 256 -s 512 -a 0 > $3/get_n2_w32_z256_512B.log
ROCSHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -z 256 -s 512 -a 0 > $3/get_n2_w32_z256_512B.log
check get_n2_w32_z256_512B
echo "get_n2_w64_z1024_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=64 mpirun -np 2 $1 -w 64 -z 1024 -s 8 -a 0 > $3/get_n2_w64_z1024_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=64 mpirun -np 2 $1 -w 64 -z 1024 -s 8 -a 0 > $3/get_n2_w64_z1024_8B.log
check get_n2_w64_z1024_8B
############################### GETNBI ################################
echo "getnbi_n2_w1_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -s 1048576 -a 1 > $3/getnbi_n2_w1_z1_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -s 1048576 -a 1 > $3/getnbi_n2_w1_z1_1MB.log
check getnbi_n2_w1_z1_1MB
echo "getnbi_n2_w1_z1024_512B"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1024 -s 512 -a 1 > $3/getnbi_n2_w1_z1024_512B.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1024 -s 512 -a 1 > $3/getnbi_n2_w1_z1024_512B.log
check getnbi_n2_w1_z1024_512B
echo "getnbi_n2_w8_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -s 1048576 -a 1 > $3/getnbi_n2_w8_z1_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -s 1048576 -a 1 > $3/getnbi_n2_w8_z1_1MB.log
check getnbi_n2_w8_z1_1MB
echo "getnbi_n2_w16_z128_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 128 -s 8 -a 1 > $3/getnbi_n2_w16_z128_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 128 -s 8 -a 1 > $3/getnbi_n2_w16_z128_8B.log
check getnbi_n2_w16_z128_8B
echo "getnbi_n2_w32_z256_512B"
ROC_SHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -z 256 -s 512 -a 1 > $3/getnbi_n2_w32_z256_512B.log
ROCSHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -z 256 -s 512 -a 1 > $3/getnbi_n2_w32_z256_512B.log
check getnbi_n2_w32_z256_512B
echo "getnbi_n2_w64_z1024_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=64 mpirun -np 2 $1 -w 64 -z 1024 -s 8 -a 1 > $3/getnbi_n2_w64_z1024_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=64 mpirun -np 2 $1 -w 64 -z 1024 -s 8 -a 1 > $3/getnbi_n2_w64_z1024_8B.log
check getnbi_n2_w64_z1024_8B
############################### PUT ###################################
echo "put_n2_w1_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -s 1048576 -a 2 > $3/put_n2_w1_z1_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -s 1048576 -a 2 > $3/put_n2_w1_z1_1MB.log
check put_n2_w1_z1_1MB
echo "put_n2_w1_z1024_512B"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1024 -s 512 -a 2 > $3/put_n2_w1_z1024_512B.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1024 -s 512 -a 2 > $3/put_n2_w1_z1024_512B.log
check put_n2_w1_z1024_512B
echo "put_n2_w8_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -s 1048576 -a 2 > $3/put_n2_w8_z1_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -s 1048576 -a 2 > $3/put_n2_w8_z1_1MB.log
check put_n2_w8_z1_1MB
echo "put_n2_w16_z128_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 128 -s 8 -a 2 > $3/put_n2_w16_z128_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 128 -s 8 -a 2 > $3/put_n2_w16_z128_8B.log
check put_n2_w16_z128_8B
echo "put_n2_w32_z256_512B"
ROC_SHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -z 256 -s 512 -a 2 > $3/put_n2_w32_z256_512B.log
ROCSHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -z 256 -s 512 -a 2 > $3/put_n2_w32_z256_512B.log
check put_n2_w32_z256_512B
echo "put_n2_w64_z1024_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=64 mpirun -np 2 $1 -w 64 -z 1024 -s 8 -a 2 > $3/put_n2_w64_z1024_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=64 mpirun -np 2 $1 -w 64 -z 1024 -s 8 -a 2 > $3/put_n2_w64_z1024_8B.log
check put_n2_w64_z1024_8B
############################### PUTNBI ################################
echo "putnbi_n2_w1_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -s 1048576 -a 3 > $3/putnbi_n2_w1_z1_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -s 1048576 -a 3 > $3/putnbi_n2_w1_z1_1MB.log
check putnbi_n2_w1_z1_1MB
echo "putnbi_n2_w1_z1024_512B"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1024 -s 512 -a 3 > $3/putnbi_n2_w1_z1024_512B.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1024 -s 512 -a 3 > $3/putnbi_n2_w1_z1024_512B.log
check putnbi_n2_w1_z1024_512B
echo "putnbi_n2_w8_z1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -s 1048576 -a 3 > $3/putnbi_n2_w8_z1_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -s 1048576 -a 3 > $3/putnbi_n2_w8_z1_1MB.log
check putnbi_n2_w8_z1_1MB
echo "putnbi_n2_w16_z128_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 128 -s 8 -a 3 > $3/putnbi_n2_w16_z128_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=16 mpirun -np 2 $1 -w 16 -z 128 -s 8 -a 3 > $3/putnbi_n2_w16_z128_8B.log
check putnbi_n2_w16_z128_8B
echo "putnbi_n2_w32_z256_512B"
ROC_SHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -z 256 -s 512 -a 3 > $3/putnbi_n2_w32_z256_512B.log
ROCSHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -z 256 -s 512 -a 3 > $3/putnbi_n2_w32_z256_512B.log
check putnbi_n2_w32_z256_512B
echo "putnbi_n2_w64_z1024_8B"
ROC_SHMEM_MAX_NUM_CONTEXTS=64 mpirun -np 2 $1 -w 64 -z 1024 -s 8 -a 3 > $3/putnbi_n2_w64_z1024_8B.log
ROCSHMEM_MAX_NUM_CONTEXTS=64 mpirun -np 2 $1 -w 64 -z 1024 -s 8 -a 3 > $3/putnbi_n2_w64_z1024_8B.log
check putnbi_n2_w64_z1024_8B
############################# REDUCTION ##############################
echo "reduction_n2_w1_z1_32K"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -s 32768 -a 5 > $3/reduction_n2_w1_z1_32K.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -s 32768 -a 5 > $3/reduction_n2_w1_z1_32K.log
check reduction_n2_w1_z1_32K
echo "reduction_n2_w8_z1_32K"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -s 32768 -a 5 > $3/reduction_n2_w8_z1_32K.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -s 32768 -a 5 > $3/reduction_n2_w8_z1_32K.log
check reduction_n2_w8_z1_32K
echo "reduction_n2_w32_z1_32K"
ROC_SHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -z 1 -s 32768 -a 5 > $3/reduction_n2_w32_z1_32K.log
ROCSHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -z 1 -s 32768 -a 5 > $3/reduction_n2_w32_z1_32K.log
check reduction_n2_w32_z1_32K
############################## AMOFADD ###############################
echo "amofadd_n2_w1_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 6 > $3/amofadd_n2_w1_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 6 > $3/amofadd_n2_w1_z1.log
check amofadd_n2_w1_z1
echo "amofadd_n2_w1_z1024"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1024 -a 6 > $3/amofadd_n2_w1_z1024.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1024 -a 6 > $3/amofadd_n2_w1_z1024.log
check amofadd_n2_w1_z1024
echo "amofadd_n2_w8_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 6 > $3/amofadd_n2_w8_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 6 > $3/amofadd_n2_w8_z1.log
check amofadd_n2_w8_z1
echo "amofadd_n2_w32_z128"
ROC_SHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -z 128 -a 6 > $3/amofadd_n2_w32_z128.log
ROCSHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -z 128 -a 6 > $3/amofadd_n2_w32_z128.log
check amofadd_n2_w32_z128
############################## AMOFINC ###############################
echo "amofinc_n2_w1_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 7 > $3/amofinc_n2_w1_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 7 > $3/amofinc_n2_w1_z1.log
check amofinc_n2_w1_z1
echo "amofinc_n2_w1_z1024"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1024 -a 7 > $3/amofinc_n2_w1_z1024.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1024 -a 7 > $3/amofinc_n2_w1_z1024.log
check amofinc_n2_w1_z1024
echo "amofinc_n2_w8_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 7 > $3/amofinc_n2_w8_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 7 > $3/amofinc_n2_w8_z1.log
check amofinc_n2_w8_z1
echo "amofinc_n2_w32_z128"
ROC_SHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -z 128 -a 7 > $3/amofinc_n2_w32_z128.log
ROCSHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -z 128 -a 7 > $3/amofinc_n2_w32_z128.log
check amofinc_n2_w32_z128
############################ AMOFETCH ################################
echo "amofetch_n2_w1_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 8 > $3/amofetch_n2_w1_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 8 > $3/amofetch_n2_w1_z1.log
check amofetch_n2_w1_z1
echo "amofetch_n2_w1_z1024"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1024 -a 8 > $3/amofetch_n2_w1_z1024.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1024 -a 8 > $3/amofetch_n2_w1_z1024.log
check amofetch_n2_w1_z1024
echo "amofetch_n2_w8_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 8 > $3/amofetch_n2_w8_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 8 > $3/amofetch_n2_w8_z1.log
check amofetch_n2_w8_z1
echo "amofetch_n2_w32_z128"
ROC_SHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -z 128 -a 8 > $3/amofetch_n2_w32_z128.log
ROCSHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -z 128 -a 8 > $3/amofetch_n2_w32_z128.log
check amofetch_n2_w32_z128
########################### AMOFCSWAP ################################
echo "amofcswap_n2_w1_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 9 > $3/amofcswap_n2_w1_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 9 > $3/amofcswap_n2_w1_z1.log
check amofcswap_n2_w1_z1
echo "amofcswap_n2_w8_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 9 > $3/amofcswap_n2_w8_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 9 > $3/amofcswap_n2_w8_z1.log
check amofcswap_n2_w8_z1
echo "amofcswap_n2_w32_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -z 1 -a 9 > $3/amofcswap_n2_w32_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -z 1 -a 9 > $3/amofcswap_n2_w32_z1.log
check amofcswap_n2_w32_z1
############################# AMOADD ################################
echo "amoadd_n2_w1_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 10 > $3/amoadd_n2_w1_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 10 > $3/amoadd_n2_w1_z1.log
check amoadd_n2_w1_z1
echo "amoadd_n2_w1_z1024"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1024 -a 10 > $3/amoadd_n2_w1_z1024.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1024 -a 10 > $3/amoadd_n2_w1_z1024.log
check amoadd_n2_w1_z1024
echo "amoadd_n2_w8_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 10 > $3/amoadd_n2_w8_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 10 > $3/amoadd_n2_w8_z1.log
check amoadd_n2_w8_z1
echo "amoadd_n2_w32_z128"
ROC_SHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -z 128 -a 10 > $3/amoadd_n2_w32_z128.log
ROCSHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -z 128 -a 10 > $3/amoadd_n2_w32_z128.log
check amoadd_n2_w32_z128
############################# AMOINC ################################
echo "amoinc_n2_w1_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 11 > $3/amoinc_n2_w1_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 11 > $3/amoinc_n2_w1_z1.log
check amoinc_n2_w1_z1
echo "amoinc_n2_w1_z1024"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1024 -a 11 > $3/amoinc_n2_w1_z1024.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1024 -a 11 > $3/amoinc_n2_w1_z1024.log
check amoinc_n2_w1_z1024
echo "amoinc_n2_w8_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 11 > $3/amoinc_n2_w8_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 11 > $3/amoinc_n2_w8_z1.log
check amoinc_n2_w8_z1
echo "amoinc_n2_w32_z128"
ROC_SHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -z 128 -a 11 > $3/amoinc_n2_w32_z128.log
ROCSHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -z 128 -a 11 > $3/amoinc_n2_w32_z128.log
check amoinc_n2_w32_z128
############################## INIT #################################
echo "init_n2"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -a 13 > $3/init_n2.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -a 13 > $3/init_n2.log
check init_n2
########################### PINGPONG ################################
echo "pingpong_n2_w1"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -a 14 > $3/pingpong_n2_w1.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -a 14 > $3/pingpong_n2_w1.log
check pingpong_n2_w1
echo "pingpong_n2_w8"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -a 14 > $3/pingpong_n2_w8.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -a 14 > $3/pingpong_n2_w8.log
check pingpong_n2_w8
echo "pingpong_n2_w32"
ROC_SHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -a 14 > $3/pingpong_n2_w32.log
ROCSHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -a 14 > $3/pingpong_n2_w32.log
check pingpong_n2_w32
############################ BARRIER ################################
echo "barrier_n2_w1"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -a 17 > $3/barrier_n2_w1.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -a 17 > $3/barrier_n2_w1.log
check barrier_n2_w1
echo "barrier_n2_w8"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -a 17 > $3/barrier_n2_w8.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -a 17 > $3/barrier_n2_w8.log
check barrier_n2_w8
echo "barrier_n2_w32"
ROC_SHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -a 17 > $3/barrier_n2_w32.log
ROCSHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -a 17 > $3/barrier_n2_w32.log
check barrier_n2_w32
############################ SYNCALL ################################
echo "syncall_n2_w1"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -a 18 > $3/syncall_n2_w1.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -a 18 > $3/syncall_n2_w1.log
check syncall_n2_w1
echo "syncall_n2_w8"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -a 18 > $3/syncall_n2_w8.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -a 18 > $3/syncall_n2_w8.log
check syncall_n2_w8
echo "syncall_n2_w32"
ROC_SHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -a 18 > $3/syncall_n2_w32.log
ROCSHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -a 18 > $3/syncall_n2_w32.log
check syncall_n2_w32
############################# SYNC ##################################
echo "sync_n2_w1"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -a 19 > $3/sync_n2_w1.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -a 19 > $3/sync_n2_w1.log
check sync_n2_w1
echo "sync_n2_w8"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -a 19 > $3/sync_n2_w8.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -a 19 > $3/sync_n2_w8.log
check sync_n2_w8
echo "sync_n2_w32"
ROC_SHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -a 19 > $3/sync_n2_w32.log
ROCSHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -a 19 > $3/sync_n2_w32.log
check sync_n2_w32
########################### FCOLLECT ################################
echo "fcollect_n2_w1_512B"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -s 512 -a 22 > $3/fcollect_n2_w1_512B.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -s 512 -a 22 > $3/fcollect_n2_w1_512B.log
check fcollect_n2_w1_512B
echo "fcollect_n2_w8_512B"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -s 512 -a 22 > $3/fcollect_n2_w8_512B.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -s 512 -a 22 > $3/fcollect_n2_w8_512B.log
check fcollect_n2_w8_512B
echo "fcollect_n2_w32_512B"
ROC_SHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -s 512 -a 22 > $3/fcollect_n2_w32_512B.log
ROCSHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -s 512 -a 22 > $3/fcollect_n2_w32_512B.log
check fcollect_n2_w32_512B
########################### ALLTOALL ################################
echo "alltoall_n2_w1_512B"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -s 512 -a 23 > $3/alltoall_n2_w1_512B.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -s 512 -a 23 > $3/alltoall_n2_w1_512B.log
check alltoall_n2_w1_512B
echo "alltoall_n2_w8_512B"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -s 512 -a 23 > $3/alltoall_n2_w8_512B.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -s 512 -a 23 > $3/alltoall_n2_w8_512B.log
check alltoall_n2_w8_512B
echo "alltoall_n2_w32_512B"
ROC_SHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -s 512 -a 23 > $3/alltoall_n2_w32_512B.log
ROCSHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -s 512 -a 23 > $3/alltoall_n2_w32_512B.log
check alltoall_n2_w32_512B
########################## TEAMGETNBI ###############################
echo "teamgetnbi_n2_w1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -s 1048576 -a 39 > $3/teamgetnbi_n2_w1_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -s 1048576 -a 39 > $3/teamgetnbi_n2_w1_1MB.log
check teamgetnbi_n2_w1_1MB
########################## TEAMPUTNBI ###############################
echo "teamputnbi_n2_w1_1MB"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -s 1048576 -a 41 > $3/teamputnbi_n2_w1_1MB.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -s 1048576 -a 41 > $3/teamputnbi_n2_w1_1MB.log
check teamputnbi_n2_w1_1MB
############################ AMOSET #################################
echo "amoset_n2_w1_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 44 > $3/amoset_n2_w1_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=1 mpirun -np 2 $1 -w 1 -z 1 -a 44 > $3/amoset_n2_w1_z1.log
check amoset_n2_w1_z1
echo "amoset_n2_w8_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 44 > $3/amoset_n2_w8_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=8 mpirun -np 2 $1 -w 8 -z 1 -a 44 > $3/amoset_n2_w8_z1.log
check amoset_n2_w8_z1
echo "amoset_n2_w32_z1"
ROC_SHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -z 1 -a 44 > $3/amoset_n2_w32_z1.log
ROCSHMEM_MAX_NUM_CONTEXTS=32 mpirun -np 2 $1 -w 32 -z 1 -a 44 > $3/amoset_n2_w32_z1.log
check amoset_n2_w32_z1
;;
@@ -607,7 +607,7 @@ case $2 in
mpirun -np 2 $1 -w 1 -z 1 -a 53
;;
*"putsignalwg")
ROC_SHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 32 -a 55
ROCSHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 32 -a 55
;;
*"putsignalwave")
mpirun -np 2 $1 -w 1 -z 32 -a 55
@@ -616,7 +616,7 @@ case $2 in
mpirun -np 2 $1 -w 1 -z 1 -a 56
;;
*"putsignalnbiwg")
ROC_SHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 32 -a 57
ROCSHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 32 -a 57
;;
*"putsignalnbiwave")
mpirun -np 2 $1 -w 1 -z 32 -a 58
@@ -625,7 +625,7 @@ case $2 in
mpirun -np 2 $1 -w 1 -z 1 -a 59
;;
*"signalfetchwg")
ROC_SHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 32 -a 60
ROCSHMEM_MAX_NUM_CONTEXTS=2 mpirun -np 2 $1 -w 2 -z 32 -a 60
;;
*"signalfetchwave")
mpirun -np 2 $1 -w 1 -z 32 -a 60