xref: /freebsd-src/contrib/arm-optimized-routines/string/bench/memset.c (revision f3087bef11543b42e0d69b708f367097a4118d24)
1d49ad206SAndrew Turner /*
2d49ad206SAndrew Turner  * memset benchmark.
3d49ad206SAndrew Turner  *
4d49ad206SAndrew Turner  * Copyright (c) 2021, Arm Limited.
5072a4ba8SAndrew Turner  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6d49ad206SAndrew Turner  */
7d49ad206SAndrew Turner 
8d49ad206SAndrew Turner #define _GNU_SOURCE
9d49ad206SAndrew Turner #include <stdint.h>
10d49ad206SAndrew Turner #include <stdio.h>
11d49ad206SAndrew Turner #include <string.h>
12d49ad206SAndrew Turner #include <assert.h>
13d49ad206SAndrew Turner #include "stringlib.h"
14d49ad206SAndrew Turner #include "benchlib.h"
15d49ad206SAndrew Turner 
16d49ad206SAndrew Turner #define ITERS  5000
17d49ad206SAndrew Turner #define ITERS2 20000000
18d49ad206SAndrew Turner #define ITERS3 1000000
19d49ad206SAndrew Turner #define NUM_TESTS 16384
20d49ad206SAndrew Turner #define MIN_SIZE 32768
21d49ad206SAndrew Turner #define MAX_SIZE (1024 * 1024)
22d49ad206SAndrew Turner 
23*f3087befSAndrew Turner static uint8_t a[MAX_SIZE + 4096] __attribute__((__aligned__(4096)));
24d49ad206SAndrew Turner 
25*f3087befSAndrew Turner #define DOTEST(STR,TESTFN)			\
26*f3087befSAndrew Turner   printf (STR);					\
27*f3087befSAndrew Turner   RUN (TESTFN, memset);				\
28*f3087befSAndrew Turner   RUNA64 (TESTFN, __memset_aarch64);		\
29*f3087befSAndrew Turner   RUNSVE (TESTFN, __memset_aarch64_sve);	\
30*f3087befSAndrew Turner   RUNMOPS (TESTFN, __memset_mops);		\
31*f3087befSAndrew Turner   RUNA32 (TESTFN, __memset_arm);		\
32*f3087befSAndrew Turner   printf ("\n");
33d49ad206SAndrew Turner 
34d49ad206SAndrew Turner typedef struct { uint32_t offset : 20, len : 12; } memset_test_t;
35d49ad206SAndrew Turner static memset_test_t test_arr[NUM_TESTS];
36d49ad206SAndrew Turner 
37d49ad206SAndrew Turner typedef struct { uint16_t size; uint16_t freq; } freq_data_t;
38d49ad206SAndrew Turner typedef struct { uint8_t align; uint16_t freq; } align_data_t;
39d49ad206SAndrew Turner 
40d49ad206SAndrew Turner #define SIZE_NUM 65536
41d49ad206SAndrew Turner #define SIZE_MASK (SIZE_NUM-1)
42d49ad206SAndrew Turner static uint8_t len_arr[SIZE_NUM];
43d49ad206SAndrew Turner 
44d49ad206SAndrew Turner /* Frequency data for memset sizes up to 4096 based on SPEC2017.  */
45d49ad206SAndrew Turner static freq_data_t memset_len_freq[] =
46d49ad206SAndrew Turner {
47d49ad206SAndrew Turner {40,28817}, {32,15336}, { 16,3823}, {296,3545}, { 24,3454}, {  8,1412},
48d49ad206SAndrew Turner {292,1202}, { 48, 927}, { 12, 613}, { 11, 539}, {284, 493}, {108, 414},
49d49ad206SAndrew Turner { 88, 380}, { 20, 295}, {312, 271}, { 72, 233}, {  2, 200}, {  4, 192},
50d49ad206SAndrew Turner { 15, 180}, { 14, 174}, { 13, 160}, { 56, 151}, { 36, 144}, { 64, 140},
51d49ad206SAndrew Turner {4095,133}, { 10, 130}, {  9, 124}, {  3, 124}, { 28, 120}, {  0, 118},
52d49ad206SAndrew Turner {288, 110}, {1152, 96}, {104,  90}, {  1,  86}, {832,  76}, {248,  74},
53d49ad206SAndrew Turner {1024, 69}, {120,  64}, {512,  63}, {384,  60}, {  6,  59}, { 80,  54},
54d49ad206SAndrew Turner { 17,  50}, {  7,  49}, {520,  47}, {2048, 39}, {256,  37}, {864,  33},
55d49ad206SAndrew Turner {1440, 28}, { 22,  27}, {2056, 24}, {260,  23}, { 68,  23}, {  5,  22},
56d49ad206SAndrew Turner { 18,  21}, {200,  18}, {2120, 18}, { 60,  17}, { 52,  16}, {336,  15},
57d49ad206SAndrew Turner { 44,  13}, {192,  13}, {160,  12}, {2064, 12}, {128,  12}, { 76,  11},
58d49ad206SAndrew Turner {164,  11}, {152,  10}, {136,   9}, {488,   7}, { 96,   6}, {560,   6},
59d49ad206SAndrew Turner {1016,  6}, {112,   5}, {232,   5}, {168,   5}, {952,   5}, {184,   5},
60d49ad206SAndrew Turner {144,   4}, {252,   4}, { 84,   3}, {960,   3}, {3808,  3}, {244,   3},
61d49ad206SAndrew Turner {280,   3}, {224,   3}, {156,   3}, {1088,  3}, {440,   3}, {216,   2},
62d49ad206SAndrew Turner {304,   2}, { 23,   2}, { 25,   2}, { 26,   2}, {264,   2}, {328,   2},
63d49ad206SAndrew Turner {1096,  2}, {240,   2}, {1104,  2}, {704,   2}, {1664,  2}, {360,   2},
64d49ad206SAndrew Turner {808,   1}, {544,   1}, {236,   1}, {720,   1}, {368,   1}, {424,   1},
65d49ad206SAndrew Turner {640,   1}, {1112,  1}, {552,   1}, {272,   1}, {776,   1}, {376,   1},
66d49ad206SAndrew Turner { 92,   1}, {536,   1}, {824,   1}, {496,   1}, {760,   1}, {792,   1},
67d49ad206SAndrew Turner {504,   1}, {344,   1}, {1816,  1}, {880,   1}, {176,   1}, {320,   1},
68d49ad206SAndrew Turner {352,   1}, {2008,  1}, {208,   1}, {408,   1}, {228,   1}, {2072,  1},
69d49ad206SAndrew Turner {568,   1}, {220,   1}, {616,   1}, {600,   1}, {392,   1}, {696,   1},
70d49ad206SAndrew Turner {2144,  1}, {1280,  1}, {2136,  1}, {632,   1}, {584,   1}, {456,   1},
71d49ad206SAndrew Turner {472,   1}, {3440,  1}, {2088,  1}, {680,   1}, {2928,  1}, {212,   1},
72d49ad206SAndrew Turner {648,   1}, {1752,  1}, {664,   1}, {3512,  1}, {1032,  1}, {528,   1},
73d49ad206SAndrew Turner {4072,  1}, {204,   1}, {2880,  1}, {3392,  1}, {712,   1}, { 59,   1},
74d49ad206SAndrew Turner {736,   1}, {592,   1}, {2520,  1}, {744,   1}, {196,   1}, {172,   1},
75d49ad206SAndrew Turner {728,   1}, {2040,  1}, {1192,  1}, {3600,  1}, {0, 0}
76d49ad206SAndrew Turner };
77d49ad206SAndrew Turner 
78d49ad206SAndrew Turner #define ALIGN_NUM 1024
79d49ad206SAndrew Turner #define ALIGN_MASK (ALIGN_NUM-1)
80d49ad206SAndrew Turner static uint8_t align_arr[ALIGN_NUM];
81d49ad206SAndrew Turner 
82d49ad206SAndrew Turner /* Alignment data for memset based on SPEC2017.  */
83d49ad206SAndrew Turner static align_data_t memset_align_freq[] =
84d49ad206SAndrew Turner {
85d49ad206SAndrew Turner  {16, 338}, {8, 307}, {32, 148}, {64, 131}, {4, 72}, {1, 23}, {2, 5}, {0, 0}
86d49ad206SAndrew Turner };
87d49ad206SAndrew Turner 
88d49ad206SAndrew Turner static void
89d49ad206SAndrew Turner init_memset_distribution (void)
90d49ad206SAndrew Turner {
91d49ad206SAndrew Turner   int i, j, freq, size, n;
92d49ad206SAndrew Turner 
93d49ad206SAndrew Turner   for (n = i = 0; (freq = memset_len_freq[i].freq) != 0; i++)
94d49ad206SAndrew Turner     for (j = 0, size = memset_len_freq[i].size; j < freq; j++)
95d49ad206SAndrew Turner       len_arr[n++] = size;
96d49ad206SAndrew Turner   assert (n == SIZE_NUM);
97d49ad206SAndrew Turner 
98d49ad206SAndrew Turner   for (n = i = 0; (freq = memset_align_freq[i].freq) != 0; i++)
99d49ad206SAndrew Turner     for (j = 0, size = memset_align_freq[i].align; j < freq; j++)
100d49ad206SAndrew Turner       align_arr[n++] = size - 1;
101d49ad206SAndrew Turner   assert (n == ALIGN_NUM);
102d49ad206SAndrew Turner }
103d49ad206SAndrew Turner 
104d49ad206SAndrew Turner static size_t
105d49ad206SAndrew Turner init_memset (size_t max_size)
106d49ad206SAndrew Turner {
107d49ad206SAndrew Turner   size_t total = 0;
108d49ad206SAndrew Turner   /* Create a random set of memsets with the given size and alignment
109d49ad206SAndrew Turner      distributions.  */
110d49ad206SAndrew Turner   for (int i = 0; i < NUM_TESTS; i++)
111d49ad206SAndrew Turner     {
112d49ad206SAndrew Turner       test_arr[i].offset = (rand32 (0) & (max_size - 1));
113d49ad206SAndrew Turner       test_arr[i].offset &= ~align_arr[rand32 (0) & ALIGN_MASK];
114d49ad206SAndrew Turner       test_arr[i].len = len_arr[rand32 (0) & SIZE_MASK];
115d49ad206SAndrew Turner       total += test_arr[i].len;
116d49ad206SAndrew Turner     }
117d49ad206SAndrew Turner 
118d49ad206SAndrew Turner   return total;
119d49ad206SAndrew Turner }
120d49ad206SAndrew Turner 
121*f3087befSAndrew Turner static void inline __attribute ((always_inline))
122*f3087befSAndrew Turner memset_random (const char *name, void *(*set)(void *, int, size_t))
123*f3087befSAndrew Turner {
124*f3087befSAndrew Turner   uint64_t total_size = 0;
125*f3087befSAndrew Turner   uint64_t tsum = 0;
126*f3087befSAndrew Turner   printf ("%22s ", name);
127*f3087befSAndrew Turner   rand32 (0x12345678);
128*f3087befSAndrew Turner 
129*f3087befSAndrew Turner   for (int size = MIN_SIZE; size <= MAX_SIZE; size *= 2)
130*f3087befSAndrew Turner     {
131*f3087befSAndrew Turner       uint64_t memset_size = init_memset (size) * ITERS;
132*f3087befSAndrew Turner 
133*f3087befSAndrew Turner       for (int c = 0; c < NUM_TESTS; c++)
134*f3087befSAndrew Turner 	set (a + test_arr[c].offset, 0, test_arr[c].len);
135*f3087befSAndrew Turner 
136*f3087befSAndrew Turner       uint64_t t = clock_get_ns ();
137*f3087befSAndrew Turner       for (int i = 0; i < ITERS; i++)
138*f3087befSAndrew Turner 	for (int c = 0; c < NUM_TESTS; c++)
139*f3087befSAndrew Turner 	  set (a + test_arr[c].offset, 0, test_arr[c].len);
140*f3087befSAndrew Turner       t = clock_get_ns () - t;
141*f3087befSAndrew Turner       total_size += memset_size;
142*f3087befSAndrew Turner       tsum += t;
143*f3087befSAndrew Turner       printf ("%dK: %5.2f ", size / 1024, (double)memset_size / t);
144*f3087befSAndrew Turner     }
145*f3087befSAndrew Turner   printf( "avg %5.2f\n", (double)total_size / tsum);
146*f3087befSAndrew Turner }
147*f3087befSAndrew Turner 
148*f3087befSAndrew Turner static void inline __attribute ((always_inline))
149*f3087befSAndrew Turner memset_medium (const char *name, void *(*set)(void *, int, size_t))
150*f3087befSAndrew Turner {
151*f3087befSAndrew Turner   printf ("%22s ", name);
152*f3087befSAndrew Turner 
153*f3087befSAndrew Turner   for (int size = 8; size <= 512; size *= 2)
154*f3087befSAndrew Turner     {
155*f3087befSAndrew Turner       uint64_t t = clock_get_ns ();
156*f3087befSAndrew Turner       for (int i = 0; i < ITERS2; i++)
157*f3087befSAndrew Turner 	set (a, 0, size);
158*f3087befSAndrew Turner       t = clock_get_ns () - t;
159*f3087befSAndrew Turner       printf ("%dB: %5.2f ", size, (double)size * ITERS2 / t);
160*f3087befSAndrew Turner     }
161*f3087befSAndrew Turner   printf ("\n");
162*f3087befSAndrew Turner }
163*f3087befSAndrew Turner 
164*f3087befSAndrew Turner static void inline __attribute ((always_inline))
165*f3087befSAndrew Turner memset_large (const char *name, void *(*set)(void *, int, size_t))
166*f3087befSAndrew Turner {
167*f3087befSAndrew Turner   printf ("%22s ", name);
168*f3087befSAndrew Turner 
169*f3087befSAndrew Turner   for (int size = 1024; size <= 65536; size *= 2)
170*f3087befSAndrew Turner     {
171*f3087befSAndrew Turner       uint64_t t = clock_get_ns ();
172*f3087befSAndrew Turner       for (int i = 0; i < ITERS3; i++)
173*f3087befSAndrew Turner 	set (a, 0, size);
174*f3087befSAndrew Turner       t = clock_get_ns () - t;
175*f3087befSAndrew Turner       printf ("%dKB: %6.2f ", size / 1024, (double)size * ITERS3 / t);
176*f3087befSAndrew Turner     }
177*f3087befSAndrew Turner   printf ("\n");
178*f3087befSAndrew Turner }
179d49ad206SAndrew Turner 
180d49ad206SAndrew Turner int main (void)
181d49ad206SAndrew Turner {
182d49ad206SAndrew Turner   init_memset_distribution ();
183d49ad206SAndrew Turner 
184d49ad206SAndrew Turner   memset (a, 1, sizeof (a));
185d49ad206SAndrew Turner 
186*f3087befSAndrew Turner   DOTEST ("Random memset (bytes/ns):\n", memset_random);
187*f3087befSAndrew Turner   DOTEST ("Medium memset (bytes/ns):\n", memset_medium);
188*f3087befSAndrew Turner   DOTEST ("Large memset (bytes/ns):\n", memset_large);
189d49ad206SAndrew Turner   return 0;
190d49ad206SAndrew Turner }
191