blob: ac6658511ac74257b343ae296941f3ff16f03601 (
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
|
#!/bin/bash
set -e
export LC_ALL=C
echo "prim-benchmarks STREAM microbenchmark (dfatool edition)"
echo "Started at $(date)"
echo "Revision $(git describe --always)"
echo "Host: $(uname -a) $(lsb_release -sd)"
echo "Compilers: $(make info)"
echo "---"
ITERATIONS=100
TIMEOUT=30m
# -i: input size (number of elements, not number of bytes!)
# 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 ; do #uint8_t uint16_t uint32_t float double; do
for i in 2097152 16384; do # 524288 131072 4096
for nr_dpus in 64; do # 1 4 8 16 32 48
for nr_tasklets in 1 4 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
for transfer in SERIAL PUSH; do
for unroll in 1 0; do
echo "Running at $(date)"
if make -B OP=${op} NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=${bl} T=${dt} TRANSFER=${transfer} UNROLL=${unroll} WITH_ALLOC_OVERHEAD=0 WITH_LOAD_OVERHEAD=0 WITH_FREE_OVERHEAD=0; then
timeout --foreground -k 1m ${TIMEOUT} bin/host_code -w 0 -e ${ITERATIONS} -i $i -x 0 || true
fi
done
done
done
done
done
done
done
done
echo "Completed at $(date)"
|
