Add STREAM benchmark application

5 files changed, 174 insertions, 0 deletions

diff --git a/.gitignore b/.gitignore
index ae10028..14f0467 100644
--- a/.gitignore
+++ b/.gitignore
@@ -7,6 +7,7 @@
 /Kconfig
 /include/config.h
 /src/app/aemr/main.cc
+/src/app/stream/log
 /src/app/prototest/nanopb.pb.h
 /src/app/prototest/nanopb.pb.cc
 /src/app/prototest/nanopbbench.pb.h
diff --git a/src/app/stream/Kconfig b/src/app/stream/Kconfig
new file mode 100644
index 0000000..ca5893c
--- /dev/null
+++ b/src/app/stream/Kconfig
@@ -0,0 +1,23 @@
+# Copyright 2020 Birte Kristina Friesel
+#
+# SPDX-License-Identifier: CC0-1.0
+
+prompt "STREAM Benchmark"
+depends on meta_driver_counter && !loop && !wakeup
+
+config app_stream_type
+string "Type"
+depends on app_stream
+default "unsigned int"
+
+config app_stream_n_elements
+int "# Elements"
+depends on app_stream
+default 128 if arch_arduino_nano
+default 8192 if arch_rm46l8lp
+default 8192 if arch_stm32f746zg_nucleo
+
+config app_stream_stride
+int "Stride"
+depends on app_stream
+default 1
diff --git a/src/app/stream/Makefile.inc b/src/app/stream/Makefile.inc
new file mode 100644
index 0000000..7f4c0cc
--- /dev/null
+++ b/src/app/stream/Makefile.inc
@@ -0,0 +1,9 @@
+# vim:ft=make
+#
+# Copyright 2020 Birte Kristina Friesel
+#
+# SPDX-License-Identifier: CC0-1.0
+
+ifdef app
+	override arch_drivers += ,counter
+endif
diff --git a/src/app/stream/main.cc b/src/app/stream/main.cc
new file mode 100644
index 0000000..6fe99bc
--- /dev/null
+++ b/src/app/stream/main.cc
@@ -0,0 +1,119 @@
+/*
+ * Copyright 2020 Birte Kristina Friesel
+ *
+ * SPDX-License-Identifier: BSD-2-Clause
+ */
+#include "arch.h"
+#include "driver/gpio.h"
+#include "driver/stdout.h"
+#include "driver/counter.h"
+#include <stdlib.h>
+
+#define XSTR(x) STR(x)
+#define STR(x) #x
+
+static const auto input_size = CONFIG_app_stream_n_elements;
+static const auto stride = CONFIG_app_stream_stride;
+typedef CONFIG_app_stream_type T;
+static T A[input_size];
+static T B[input_size];
+static T scale;
+static T X[input_size];
+static T check;
+
+int main(void)
+{
+	uint16_t i;
+	arch.setup();
+	gpio.setup();
+	kout.setup();
+
+	scale = rand();
+	for (i = 0; i < input_size; i++) {
+		A[i] = (T)rand();
+		B[i] = (T)rand();
+		X[i] = (T)0;
+	}
+
+	while (1) {
+		check = 0;
+		/*
+		 * Copy
+		 */
+		counter.start();
+		for (i = 0; i < input_size; i += stride) {
+			X[i] = A[i];
+		}
+		counter.stop();
+
+		kout << "[::] STREAM COPY";
+		kout << " | n_elements=" << input_size << " e_type=" << XSTR(CONFIG_app_stream_type) << " elem_B=" << sizeof(T) << " n_stride=" << stride;
+		kout << " | latency_us=" << counter.value << "/" << counter.overflow;
+		kout << endl;
+
+		for (i = 0; i < input_size; i++) {
+			check = (uint16_t)check ^ (uint16_t)X[i];
+		}
+
+		/*
+		 * Scale
+		 */
+		counter.start();
+		for (i = 0; i < input_size; i += stride) {
+			X[i] = scale * A[i];
+		}
+		counter.stop();
+
+		kout << "[::] STREAM SCALE";
+		kout << " | n_elements=" << input_size << " e_type=" << XSTR(CONFIG_app_stream_type) << " elem_B=" << sizeof(T) << " n_stride=" << stride;
+		kout << " | latency_us=" << counter.value << "/" << counter.overflow;
+		kout << endl;
+
+		for (i = 0; i < input_size; i++) {
+			check = (uint16_t)check ^ (uint16_t)X[i];
+		}
+
+		/*
+		 * Add
+		 */
+		counter.start();
+		for (i = 0; i < input_size; i += stride) {
+			X[i] = A[i] + B[i];
+		}
+		counter.stop();
+
+		kout << "[::] STREAM ADD";
+		kout << " | n_elements=" << input_size << " e_type=" << XSTR(CONFIG_app_stream_type) << " elem_B=" << sizeof(T) << " n_stride=" << stride;
+		kout << " | latency_us=" << counter.value << "/" << counter.overflow;
+		kout << endl;
+
+		for (i = 0; i < input_size; i++) {
+			check = (uint16_t)check ^ (uint16_t)X[i];
+		}
+
+		/*
+		 * Triad
+		 */
+		counter.start();
+		for (i = 0; i < input_size; i += stride) {
+			X[i] = A[i] + scale * B[i];
+		}
+		counter.stop();
+
+		kout << "[::] STREAM TRIAD";
+		kout << " | n_elements=" << input_size << " e_type=" << XSTR(CONFIG_app_stream_type) << " elem_B=" << sizeof(T) << " n_stride=" << stride;
+		kout << " | latency_us=" << counter.value << "/" << counter.overflow;
+		kout << endl;
+
+		for (i = 0; i < input_size; i++) {
+			check = (uint16_t)check ^ (uint16_t)X[i];
+		}
+
+		/*
+		 * Avoid optimizations
+		 */
+		kout << "// " << check << endl;
+	}
+
+	return 0;
+}
diff --git a/src/app/stream/run.sh b/src/app/stream/run.sh
new file mode 100755
index 0000000..8cc3da2
--- /dev/null
+++ b/src/app/stream/run.sh
@@ -0,0 +1,22 @@
+#!/bin/sh
+
+cd "$(git rev-parse --show-toplevel)"
+
+target="$(grep '^CONFIG_arch=' .config | cut -d '"' -f 2)"
+
+mkdir -p src/app/stream/log
+: > src/app/stream/log/${target}.txt
+
+for stride in 1 2 4 8 16 32 64; do
+	for type in uint8_t uint16_t uint32_t uint64_t float double; do
+		kconfig-tweak --keep-case --set-str app_stream_stride $stride
+		kconfig-tweak --keep-case --set-str app_stream_type $type
+
+		echo
+		echo "stride: ${stride}"
+		echo "type  : ${type}"
+		echo
+
+		./mp && make cat >> src/app/stream/log/${target}.txt
+	done
+done