From 5976e5c372ce0eded29b628c6b233fca77762975 Mon Sep 17 00:00:00 2001
From: Birte Kristina Friesel <birte.friesel@uos.de>
Date: Mon, 15 Jul 2024 07:48:36 +0200
Subject: VA, GEMV: Add roofline benchmarks

---
 GEMV/dimes-hetsim-hbm-roofline.sh | 36 ++++++++++++++++++++++++++++++++++++
 VA/dimes-hetsim-hbm-roofline.sh   | 25 +++++++++++++++++++++++++
 2 files changed, 61 insertions(+)
 create mode 100755 GEMV/dimes-hetsim-hbm-roofline.sh
 create mode 100755 VA/dimes-hetsim-hbm-roofline.sh

diff --git a/GEMV/dimes-hetsim-hbm-roofline.sh b/GEMV/dimes-hetsim-hbm-roofline.sh
new file mode 100755
index 0000000..74e4051
--- /dev/null
+++ b/GEMV/dimes-hetsim-hbm-roofline.sh
@@ -0,0 +1,36 @@
+#!/bin/bash
+
+cd baselines/cpu
+make -B NUMA=1
+
+mkdir -p log/$(hostname)
+fn=log/$(hostname)/$(date +%Y%m%d)-roofline
+
+# upstream DPU version uses -m 163840 -n 4096
+# → (163840 * 4096 + 4096) * uint32 ≈ 2.5 GiB
+
+# upstream baseline uses 20480 rows and 8192 cols, allocating 20480 * double + 8192 * double + 20480 * double + 20480 * 8192 * double
+# → ≈ 1.25 GiB
+
+run_benchmark() {
+	local "$@"
+	OMP_NUM_THREADS=$nr_threads ./gemv $ram $ram $cpu
+	return $?
+}
+
+export -f run_benchmark
+
+(
+
+for i in $(seq 0 7); do
+	echo "single-node execution ($i/8)" >&2
+	parallel -j1 --eta --joblog ${fn}.${i}.joblog --header : \
+	run_benchmark nr_threads={nr_threads} ram={ram} cpu={cpu} \
+	::: cpu $(seq 0 7) \
+	::: ram $(seq 0 15) \
+	::: nr_threads 1 2 4 8 12 16
+done
+
+) > ${fn}.txt
+
+xz -f -v -9 -M 800M ${fn}.txt
diff --git a/VA/dimes-hetsim-hbm-roofline.sh b/VA/dimes-hetsim-hbm-roofline.sh
new file mode 100755
index 0000000..f987f5a
--- /dev/null
+++ b/VA/dimes-hetsim-hbm-roofline.sh
@@ -0,0 +1,25 @@
+#!/bin/sh
+
+cd baselines/cpu
+make -B NUMA=1
+
+mkdir -p log/$(hostname)
+fn=log/$(hostname)/$(date +%Y%m%d)-roofline
+
+# upstream uses 167772160 * 2 * int32 == 2.5 GiB input and 1.25 GiB output for DPU version
+
+(
+
+for i in $(seq 0 7); do
+	echo "single-node execution ($i/8)" >&2
+	parallel -j1 --eta --joblog ${fn}.${i}.joblog --header : \
+		./va -i {input_size} -a {ram} -b {ram} -c {cpu} -t {nr_threads} -w 0 -e 5 \
+		::: nr_threads $(seq 1 16) \
+		::: cpu $i \
+		::: ram $i $((i+8)) \
+		::: input_size 167772160
+done
+
+) > ${fn}.txt
+
+xz -f -v -9 -M 800M ${fn}.txt
-- 
cgit v1.2.3