blob: b6d645adaa5c2a81bc1543700edc02ff97ecfcf5 (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
|
#!/bin/sh
cd baselines/cpu
make -B NUMA=1
mkdir -p log/$(hostname)
fn=log/$(hostname)/$(date +%Y%m%d)
# Args: -m m -n n -o M_ -p N_
#
# Input: (M_ * m) × (N_ * n) matrix
# Output: (N_* n) × (M_ * m) matrix
# Step 1: transpose (M_ * m) × N_ matrix that consits of tiles of size n
# CPU version: explicit
# DPU version: implicit (appropriate write operations to DPUs)
# Step 2: transpose m × n matrix; this happens N_ * M_ times.
# DPU version: Each tasklet transposes a single m × n matrix / tile.
# (16 × 8 tile takes up 1 KiB WRAM)
# Step 3: Transpose M_ × n matrix that consists of tiles of size m.
#
# Note for DPU version: if M_ > #DPUs, steps 1 through 3 are repeated.
# Number of repetitions == ceil(M_ / #DPUS)
# For Hetsim benchmarks, we set M_ == #DPUs to simplify the task graph (no repetitions that depend on the number of available DPUs).
# Just in case, there is also a configuration with M_ == 2048 independent of #DPUs
#
# input size: uint64(DPU)/double(CPU) * M_ * m * N_ * n
# output size: uint64(DPU)/double(CPU) * M_ * m * N_ * n -- on DPU only; CPU version operates in-place
# Upstream DPU version uses int64_t, -p 2048 -o 12288 -x 1 [implicit -m 16 -n 8]
# Upstream CPU version uses double, -p 2556 -o 4096 -m 16 -n 8 and fails with -o 12288 (allocation error)
#
# -p 2048 -o 2048 -m 16 -n 8 -> matrix size: 4 GiB
# -p [64 .. 2304] -o 2048 -m 16 -n 8 -> matrix size: 128 MiB .. 4.5 GiB
(
echo "CPU single-node operation (1/2)" >&2
parallel -j1 --eta --joblog ${fn}.1.joblog --header : \
./trns -w 0 -r 5 -p {p} -o 2048 -m 16 -n 8 -t {nr_threads} -a {ram} -b {ram} -c {cpu} \
::: p 64 128 256 512 768 1024 1536 2048 2304 \
::: ram $(seq 0 15) \
::: cpu $(seq 0 7) \
::: nr_threads 1 2 4 8 12 16
echo "CPU multi-node operation (1/2)" >&2
parallel -j1 --eta --joblog ${fn}.2.joblog --header : \
./trns -w 0 -r 40 -p {p} -o 2048 -m 16 -n 8 -t {nr_threads} -a {ram} -b {ram} -c {cpu} \
::: p 64 128 256 512 768 1024 1536 2048 2304 \
::: ram $(seq 0 15) \
::: cpu -1 \
::: nr_threads 32 48 64 96 128
) > ${fn}.txt
xz -f -v -9 -M 800M ${fn}.txt
|
