1 /*
2 * memset benchmark.
3 *
4 * Copyright (c) 2021, Arm Limited.
5 * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6 */
7
8 #define _GNU_SOURCE
9 #include <stdint.h>
10 #include <stdio.h>
11 #include <string.h>
12 #include <assert.h>
13 #include "stringlib.h"
14 #include "benchlib.h"
15
16 #define ITERS 5000
17 #define ITERS2 20000000
18 #define ITERS3 1000000
19 #define NUM_TESTS 16384
20 #define MIN_SIZE 32768
21 #define MAX_SIZE (1024 * 1024)
22
23 static uint8_t a[MAX_SIZE + 4096] __attribute__((__aligned__(4096)));
24
25 #define DOTEST(STR,TESTFN) \
26 printf (STR); \
27 RUN (TESTFN, memset); \
28 RUNA64 (TESTFN, __memset_aarch64); \
29 RUNSVE (TESTFN, __memset_aarch64_sve); \
30 RUNMOPS (TESTFN, __memset_mops); \
31 RUNA32 (TESTFN, __memset_arm); \
32 printf ("\n");
33
34 typedef struct { uint32_t offset : 20, len : 12; } memset_test_t;
35 static memset_test_t test_arr[NUM_TESTS];
36
37 typedef struct { uint16_t size; uint16_t freq; } freq_data_t;
38 typedef struct { uint8_t align; uint16_t freq; } align_data_t;
39
40 #define SIZE_NUM 65536
41 #define SIZE_MASK (SIZE_NUM-1)
42 static uint8_t len_arr[SIZE_NUM];
43
44 /* Frequency data for memset sizes up to 4096 based on SPEC2017. */
45 static freq_data_t memset_len_freq[] =
46 {
47 {40,28817}, {32,15336}, { 16,3823}, {296,3545}, { 24,3454}, { 8,1412},
48 {292,1202}, { 48, 927}, { 12, 613}, { 11, 539}, {284, 493}, {108, 414},
49 { 88, 380}, { 20, 295}, {312, 271}, { 72, 233}, { 2, 200}, { 4, 192},
50 { 15, 180}, { 14, 174}, { 13, 160}, { 56, 151}, { 36, 144}, { 64, 140},
51 {4095,133}, { 10, 130}, { 9, 124}, { 3, 124}, { 28, 120}, { 0, 118},
52 {288, 110}, {1152, 96}, {104, 90}, { 1, 86}, {832, 76}, {248, 74},
53 {1024, 69}, {120, 64}, {512, 63}, {384, 60}, { 6, 59}, { 80, 54},
54 { 17, 50}, { 7, 49}, {520, 47}, {2048, 39}, {256, 37}, {864, 33},
55 {1440, 28}, { 22, 27}, {2056, 24}, {260, 23}, { 68, 23}, { 5, 22},
56 { 18, 21}, {200, 18}, {2120, 18}, { 60, 17}, { 52, 16}, {336, 15},
57 { 44, 13}, {192, 13}, {160, 12}, {2064, 12}, {128, 12}, { 76, 11},
58 {164, 11}, {152, 10}, {136, 9}, {488, 7}, { 96, 6}, {560, 6},
59 {1016, 6}, {112, 5}, {232, 5}, {168, 5}, {952, 5}, {184, 5},
60 {144, 4}, {252, 4}, { 84, 3}, {960, 3}, {3808, 3}, {244, 3},
61 {280, 3}, {224, 3}, {156, 3}, {1088, 3}, {440, 3}, {216, 2},
62 {304, 2}, { 23, 2}, { 25, 2}, { 26, 2}, {264, 2}, {328, 2},
63 {1096, 2}, {240, 2}, {1104, 2}, {704, 2}, {1664, 2}, {360, 2},
64 {808, 1}, {544, 1}, {236, 1}, {720, 1}, {368, 1}, {424, 1},
65 {640, 1}, {1112, 1}, {552, 1}, {272, 1}, {776, 1}, {376, 1},
66 { 92, 1}, {536, 1}, {824, 1}, {496, 1}, {760, 1}, {792, 1},
67 {504, 1}, {344, 1}, {1816, 1}, {880, 1}, {176, 1}, {320, 1},
68 {352, 1}, {2008, 1}, {208, 1}, {408, 1}, {228, 1}, {2072, 1},
69 {568, 1}, {220, 1}, {616, 1}, {600, 1}, {392, 1}, {696, 1},
70 {2144, 1}, {1280, 1}, {2136, 1}, {632, 1}, {584, 1}, {456, 1},
71 {472, 1}, {3440, 1}, {2088, 1}, {680, 1}, {2928, 1}, {212, 1},
72 {648, 1}, {1752, 1}, {664, 1}, {3512, 1}, {1032, 1}, {528, 1},
73 {4072, 1}, {204, 1}, {2880, 1}, {3392, 1}, {712, 1}, { 59, 1},
74 {736, 1}, {592, 1}, {2520, 1}, {744, 1}, {196, 1}, {172, 1},
75 {728, 1}, {2040, 1}, {1192, 1}, {3600, 1}, {0, 0}
76 };
77
78 #define ALIGN_NUM 1024
79 #define ALIGN_MASK (ALIGN_NUM-1)
80 static uint8_t align_arr[ALIGN_NUM];
81
82 /* Alignment data for memset based on SPEC2017. */
83 static align_data_t memset_align_freq[] =
84 {
85 {16, 338}, {8, 307}, {32, 148}, {64, 131}, {4, 72}, {1, 23}, {2, 5}, {0, 0}
86 };
87
88 static void
init_memset_distribution(void)89 init_memset_distribution (void)
90 {
91 int i, j, freq, size, n;
92
93 for (n = i = 0; (freq = memset_len_freq[i].freq) != 0; i++)
94 for (j = 0, size = memset_len_freq[i].size; j < freq; j++)
95 len_arr[n++] = size;
96 assert (n == SIZE_NUM);
97
98 for (n = i = 0; (freq = memset_align_freq[i].freq) != 0; i++)
99 for (j = 0, size = memset_align_freq[i].align; j < freq; j++)
100 align_arr[n++] = size - 1;
101 assert (n == ALIGN_NUM);
102 }
103
104 static size_t
init_memset(size_t max_size)105 init_memset (size_t max_size)
106 {
107 size_t total = 0;
108 /* Create a random set of memsets with the given size and alignment
109 distributions. */
110 for (int i = 0; i < NUM_TESTS; i++)
111 {
112 test_arr[i].offset = (rand32 (0) & (max_size - 1));
113 test_arr[i].offset &= ~align_arr[rand32 (0) & ALIGN_MASK];
114 test_arr[i].len = len_arr[rand32 (0) & SIZE_MASK];
115 total += test_arr[i].len;
116 }
117
118 return total;
119 }
120
121 static void inline __attribute ((always_inline))
memset_random(const char * name,void * (* set)(void *,int,size_t))122 memset_random (const char *name, void *(*set)(void *, int, size_t))
123 {
124 uint64_t total_size = 0;
125 uint64_t tsum = 0;
126 printf ("%22s ", name);
127 rand32 (0x12345678);
128
129 for (int size = MIN_SIZE; size <= MAX_SIZE; size *= 2)
130 {
131 uint64_t memset_size = init_memset (size) * ITERS;
132
133 for (int c = 0; c < NUM_TESTS; c++)
134 set (a + test_arr[c].offset, 0, test_arr[c].len);
135
136 uint64_t t = clock_get_ns ();
137 for (int i = 0; i < ITERS; i++)
138 for (int c = 0; c < NUM_TESTS; c++)
139 set (a + test_arr[c].offset, 0, test_arr[c].len);
140 t = clock_get_ns () - t;
141 total_size += memset_size;
142 tsum += t;
143 printf ("%dK: %5.2f ", size / 1024, (double)memset_size / t);
144 }
145 printf( "avg %5.2f\n", (double)total_size / tsum);
146 }
147
148 static void inline __attribute ((always_inline))
memset_medium(const char * name,void * (* set)(void *,int,size_t))149 memset_medium (const char *name, void *(*set)(void *, int, size_t))
150 {
151 printf ("%22s ", name);
152
153 for (int size = 8; size <= 512; size *= 2)
154 {
155 uint64_t t = clock_get_ns ();
156 for (int i = 0; i < ITERS2; i++)
157 set (a, 0, size);
158 t = clock_get_ns () - t;
159 printf ("%dB: %5.2f ", size, (double)size * ITERS2 / t);
160 }
161 printf ("\n");
162 }
163
164 static void inline __attribute ((always_inline))
memset_large(const char * name,void * (* set)(void *,int,size_t))165 memset_large (const char *name, void *(*set)(void *, int, size_t))
166 {
167 printf ("%22s ", name);
168
169 for (int size = 1024; size <= 65536; size *= 2)
170 {
171 uint64_t t = clock_get_ns ();
172 for (int i = 0; i < ITERS3; i++)
173 set (a, 0, size);
174 t = clock_get_ns () - t;
175 printf ("%dKB: %6.2f ", size / 1024, (double)size * ITERS3 / t);
176 }
177 printf ("\n");
178 }
179
main(void)180 int main (void)
181 {
182 init_memset_distribution ();
183
184 memset (a, 1, sizeof (a));
185
186 DOTEST ("Random memset (bytes/ns):\n", memset_random);
187 DOTEST ("Medium memset (bytes/ns):\n", memset_medium);
188 DOTEST ("Large memset (bytes/ns):\n", memset_large);
189 return 0;
190 }
191