diff options
author | Birte Kristina Friesel <birte.friesel@uos.de> | 2024-05-13 10:01:46 +0200 |
---|---|---|
committer | Birte Kristina Friesel <birte.friesel@uos.de> | 2024-05-13 10:01:46 +0200 |
commit | 98df3d2727182f6e0d249cc22cc08a3561e52de0 (patch) | |
tree | a4fa3c3ac79ca4cabc4b47897c656720021639b8 /Microbenchmarks | |
parent | 2dc701bb9ee987b3044fefe51d2c8bc49b04088e (diff) |
past derf has been naughty and not committed changes in time
Diffstat (limited to 'Microbenchmarks')
-rwxr-xr-x | Microbenchmarks/STREAM/run-rank.sh | 22 |
1 files changed, 12 insertions, 10 deletions
diff --git a/Microbenchmarks/STREAM/run-rank.sh b/Microbenchmarks/STREAM/run-rank.sh index 89bac24..5e77002 100755 --- a/Microbenchmarks/STREAM/run-rank.sh +++ b/Microbenchmarks/STREAM/run-rank.sh @@ -10,22 +10,24 @@ echo "Revision $(git describe --always)" # Each DPU uses three buffers, each of which holds $i * sizeof($dt) bytes. # With a total MRAM capacity of 64M, this gives us ~21M per buffer, or 16M when rounding down to the next power of two. # With a maximum data type width of 8B (uint64_t, double), this limits the number of elements per DPU to 2097152. -for dt in uint64_t uint32_t ; do #uint8_t uint16_t float double; do - for i in 2097152 1048576 524288 131072 16384 4096; do - for nr_dpus in 1 4 8 16 32 48 64; do +for dt in uint64_t ; do #uint8_t uint16_t uint32_t float double; do + for i in 16384; do # 2097152; do # 524288 131072 4096; do + for nr_dpus in 64 1 4 8 16 32 48; do for nr_tasklets in 1 8 12 16; do for op in triad scale add copy copyw; do # BL: use 2^(BL) B blocks for MRAM <-> WRAM transfers on PIM module # Our largest data type holds 8B, so the minimum block size is 3. # From a performance perspective, 8 to 10 is usually best for sequential operations. for bl in 3 8 10; do - echo - if make -B OP=${op} NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=${bl} T=${dt} UNROLL=1 WITH_ALLOC_OVERHEAD=0 WITH_LOAD_OVERHEAD=0 WITH_FREE_OVERHEAD=0 \ - || make -B OP=${op} NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=${bl} T=${dt} UNROLL=0 WITH_ALLOC_OVERHEAD=0 WITH_LOAD_OVERHEAD=0 WITH_FREE_OVERHEAD=0; then - # -w: number of un-timed warmup iterations - # -e: number of timed iterations - timeout --foreground -k 1m 30m bin/host_code -w 0 -e 100 -i $i -x 0 || true - fi + for transfer in SERIAL PUSH; do + echo "Running at $(date)" + if make -B OP=${op} NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=${bl} T=${dt} TRANSFER=${transfer} UNROLL=1 WITH_ALLOC_OVERHEAD=0 WITH_LOAD_OVERHEAD=0 WITH_FREE_OVERHEAD=0; then + timeout --foreground -k 1m 30m bin/host_code -w 0 -e 100 -i $i -x 0 || true + fi + if make -B OP=${op} NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=${bl} T=${dt} TRANSFER=${transfer} UNROLL=0 WITH_ALLOC_OVERHEAD=0 WITH_LOAD_OVERHEAD=0 WITH_FREE_OVERHEAD=0; then + timeout --foreground -k 1m 30m bin/host_code -w 0 -e 100 -i $i -x 0 || true + fi + done done done done |