From c20f5fb67ffe6d8011ce7d7a41dd8d8c571ca178 Mon Sep 17 00:00:00 2001
From: Howard Mao <zhehao.mao@gmail.com>
Date: Fri, 9 Sep 2016 13:53:07 -0700
Subject: add a software memcpy benchmark

---
 benchmarks/Makefile             |  1 +
 benchmarks/memcpy/bmark.mk      | 31 +++++++++++++++++++++++++++++++
 benchmarks/memcpy/memcpy.c      | 40 ++++++++++++++++++++++++++++++++++++++++
 benchmarks/memcpy/memcpy.h      |  9 +++++++++
 benchmarks/memcpy/memcpy_main.c | 28 ++++++++++++++++++++++++++++
 5 files changed, 109 insertions(+)
 create mode 100644 benchmarks/memcpy/bmark.mk
 create mode 100644 benchmarks/memcpy/memcpy.c
 create mode 100644 benchmarks/memcpy/memcpy.h
 create mode 100644 benchmarks/memcpy/memcpy_main.c

diff --git a/benchmarks/Makefile b/benchmarks/Makefile
index ce9d4f9..d246ef7 100644
--- a/benchmarks/Makefile
+++ b/benchmarks/Makefile
@@ -29,6 +29,7 @@ bmarks = \
 	spmv \
 	mt-vvadd \
 	mt-matmul \
+	memcpy \
 
 bmarks_host = \
 	median \
diff --git a/benchmarks/memcpy/bmark.mk b/benchmarks/memcpy/bmark.mk
new file mode 100644
index 0000000..2808b73
--- /dev/null
+++ b/benchmarks/memcpy/bmark.mk
@@ -0,0 +1,31 @@
+#=======================================================================
+# UCB CS250 Makefile fragment for benchmarks
+#-----------------------------------------------------------------------
+#
+# Each benchmark directory should have its own fragment which
+# essentially lists what the source files are and how to link them
+# into an riscv and/or host executable. All variables should include
+# the benchmark name as a prefix so that they are unique.
+#
+
+memcpy_c_src = \
+	memcpy_main.c \
+	memcpy.c \
+	syscalls.c \
+
+memcpy_riscv_src = \
+	crt.S \
+
+memcpy_c_objs     = $(patsubst %.c, %.o, $(memcpy_c_src))
+memcpy_riscv_objs = $(patsubst %.S, %.o, $(memcpy_riscv_src))
+
+memcpy_host_bin = memcpy.host
+$(memcpy_host_bin) : $(memcpy_c_src)
+	$(HOST_COMP) $^ -o $(memcpy_host_bin)
+
+memcpy_riscv_bin = memcpy.riscv
+$(memcpy_riscv_bin) : $(memcpy_c_objs) $(memcpy_riscv_objs)
+	$(RISCV_LINK) $(memcpy_c_objs) $(memcpy_riscv_objs) -o $(memcpy_riscv_bin) $(RISCV_LINK_OPTS)
+
+junk += $(memcpy_c_objs) $(memcpy_riscv_objs) \
+        $(memcpy_host_bin) $(memcpy_riscv_bin)
diff --git a/benchmarks/memcpy/memcpy.c b/benchmarks/memcpy/memcpy.c
new file mode 100644
index 0000000..b2511d5
--- /dev/null
+++ b/benchmarks/memcpy/memcpy.c
@@ -0,0 +1,40 @@
+#include "memcpy.h"
+#include <stdlib.h>
+#include <stdint.h>
+
+void fill(int *arr, size_t n)
+{
+	for (int i = 0; i < n; i++) {
+		arr[i] = i << 4;
+	}
+}
+
+#pragma GCC optimize ("unroll-loops")
+
+#define UNROLL 4
+
+void fast_memcpy(void *dst, void *src, size_t len)
+{
+	uintptr_t ptr_mask = (uintptr_t) dst | (uintptr_t) src | len;
+
+	if ((ptr_mask & (sizeof(uintptr_t) - 1)) == 0) {
+		uintptr_t *d = dst, *s = src;
+		size_t n = len / sizeof(uintptr_t);
+		size_t lenu = ((n / UNROLL) * UNROLL) * sizeof(uintptr_t);
+
+		while (d < (uintptr_t *)(dst + lenu)) {
+			for (int i = 0; i < UNROLL; i++)
+				d[i] = s[i];
+			d += UNROLL;
+			s += UNROLL;
+		}
+
+		while (d < (uintptr_t *)(dst + len))
+			*(d++) = *(s++);
+	} else {
+		char *d = dst, *s = src;
+
+		while (d < (char *)(dst + len))
+			*(d++) = *(s++);
+	}
+}
diff --git a/benchmarks/memcpy/memcpy.h b/benchmarks/memcpy/memcpy.h
new file mode 100644
index 0000000..ca65756
--- /dev/null
+++ b/benchmarks/memcpy/memcpy.h
@@ -0,0 +1,9 @@
+#ifndef __MEMCPY_H__
+#define __MEMCPY_H__
+
+#include <stdlib.h>
+
+void fill(int *arr, size_t n);
+void fast_memcpy(void *dst, void *src, size_t len);
+
+#endif
diff --git a/benchmarks/memcpy/memcpy_main.c b/benchmarks/memcpy/memcpy_main.c
new file mode 100644
index 0000000..0436c73
--- /dev/null
+++ b/benchmarks/memcpy/memcpy_main.c
@@ -0,0 +1,28 @@
+#include <stdio.h>
+#include "util.h"
+#include "memcpy.h"
+
+#define MAX_SIZE 4 * 4096
+#define START_SIZE 256
+
+int a[MAX_SIZE / sizeof(int)];
+int b[MAX_SIZE / sizeof(int)];
+
+void thread_entry(int cid, int nc)
+{
+
+	int copy_size = START_SIZE;
+
+	fill(a, MAX_SIZE / sizeof(uint64_t));
+	asm volatile ("fence");
+
+	while (copy_size <= MAX_SIZE) {
+		memcpy(b, a, MAX_SIZE);
+		printf("%d: ", copy_size);
+		stats(fast_memcpy(b, a, copy_size), 3);
+		verify(copy_size / sizeof(int), b, a);
+		copy_size *= 2;
+	}
+
+	exit(0);
+}
-- 
cgit v1.1