diff options
| author | Birte Kristina Friesel <birte.friesel@uos.de> | 2024-07-17 15:00:09 +0200 |
|---|---|---|
| committer | Birte Kristina Friesel <birte.friesel@uos.de> | 2024-07-17 15:00:09 +0200 |
| commit | 9612a7d5e3b685f9d1d9af559999999ce2b3880c (patch) | |
| tree | 26a11e230a5baa344d6d30bd25ee7df6ad99e6d9 /VA/dimes-hetsim-nmc.sh | |
| parent | 1259f4c85f45ccf72b434d35b3bd1af841c9ae62 (diff) | |
VA memcpy: HBM support
Diffstat (limited to 'VA/dimes-hetsim-nmc.sh')
| -rwxr-xr-x | VA/dimes-hetsim-nmc.sh | 23 |
1 files changed, 13 insertions, 10 deletions
diff --git a/VA/dimes-hetsim-nmc.sh b/VA/dimes-hetsim-nmc.sh index fa9e975..1e33c09 100755 --- a/VA/dimes-hetsim-nmc.sh +++ b/VA/dimes-hetsim-nmc.sh @@ -40,20 +40,22 @@ make -B NUMA=1 ( -echo "CPU single-node operation (1/2)" >&2 +echo "CPU single-node operation (1/4)" >&2 parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \ - ./va -i {input_size} -a {ram} -b {ram} -c {cpu} -t {nr_threads} -w 0 -e 40 \ - ::: ram 0 1 \ - ::: cpu 0 1 \ + ./va -i {input_size} -a {ram_in} -b {ram_out} -c {cpu} -t {nr_threads} -w 0 -e 40 \ + ::: ram_in 0 1 \ + ::: cpu 0 1 \ + :::+ ram_out 0 1 \ ::: nr_threads 1 2 4 8 12 16 \ ::: input_size 167772160 -echo "CPU multi-node operation (1/2)" >&2 +echo "CPU multi-node operation (2/4)" >&2 parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \ - ./va -i {input_size} -a {ram} -b {ram} -c {cpu} -t {nr_threads} -w 0 -e 40 \ - ::: ram 0 1 \ + ./va -i {input_size} -a {ram_in} -b {ram_out} -c {cpu} -t {nr_threads} -w 0 -e 40 \ + ::: ram_in 0 1 \ + ::: ram_out 0 1 \ ::: cpu -1 \ ::: nr_threads 24 32 \ ::: input_size 167772160 @@ -66,10 +68,11 @@ make -B NUMA=1 NUMA_MEMCPY=1 echo "CPU single-node operation with setup cost" >&2 -parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \ - ./va -i {input_size} -a {ram} -b {cpu} -c {cpu} -t {nr_threads} -w 0 -e 40 \ +parallel -j1 --eta --joblog ${fn}.3.joblog --resume --header : \ + ./va -i {input_size} -a {ram} -b {ram_out} -c {cpu} -C {cpu} -t {nr_threads} -w 0 -e 40 \ ::: ram 0 1 \ - ::: cpu 0 1 \ + ::: cpu 0 1 \ + :::+ ram_out 0 1 \ ::: nr_threads 1 2 4 8 12 16 \ ::: input_size 167772160 |