1 /* 2 * strlen benchmark. 3 * 4 * Copyright (c) 2020-2021, 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 2000000 19 #define NUM_TESTS 16384 20 21 #define MAX_ALIGN 32 22 #define MAX_STRLEN 256 23 24 static char a[(MAX_STRLEN + 1) * MAX_ALIGN] __attribute__((__aligned__(4096))); 25 26 #define F(x, mte) {#x, x, mte}, 27 28 static const struct fun 29 { 30 const char *name; 31 size_t (*fun) (const char *s); 32 int test_mte; 33 } funtab[] = { 34 // clang-format off 35 F(strlen, 0) 36 #if __aarch64__ 37 F(__strlen_aarch64, 0) 38 F(__strlen_aarch64_mte, 1) 39 # if __ARM_FEATURE_SVE 40 F(__strlen_aarch64_sve, 1) 41 # endif 42 #elif __arm__ 43 # if __ARM_ARCH >= 6 && __ARM_ARCH_ISA_THUMB == 2 44 F(__strlen_armv6t2, 0) 45 # endif 46 #endif 47 {0, 0, 0} 48 // clang-format on 49 }; 50 #undef F 51 52 static uint16_t strlen_tests[NUM_TESTS]; 53 54 typedef struct { uint16_t size; uint16_t freq; } freq_data_t; 55 typedef struct { uint8_t align; uint16_t freq; } align_data_t; 56 57 #define SIZE_NUM 65536 58 #define SIZE_MASK (SIZE_NUM - 1) 59 static uint8_t strlen_len_arr[SIZE_NUM]; 60 61 /* Frequency data for strlen sizes up to 128 based on SPEC2017. */ 62 static freq_data_t strlen_len_freq[] = 63 { 64 { 12,22671}, { 18,12834}, { 13, 9555}, { 6, 6348}, { 17, 6095}, { 11, 2115}, 65 { 10, 1335}, { 7, 814}, { 2, 646}, { 9, 483}, { 8, 471}, { 16, 418}, 66 { 4, 390}, { 1, 388}, { 5, 233}, { 3, 204}, { 0, 79}, { 14, 79}, 67 { 15, 69}, { 26, 36}, { 22, 35}, { 31, 24}, { 32, 24}, { 19, 21}, 68 { 25, 17}, { 28, 15}, { 21, 14}, { 33, 14}, { 20, 13}, { 24, 9}, 69 { 29, 9}, { 30, 9}, { 23, 7}, { 34, 7}, { 27, 6}, { 44, 5}, 70 { 42, 4}, { 45, 3}, { 47, 3}, { 40, 2}, { 41, 2}, { 43, 2}, 71 { 58, 2}, { 78, 2}, { 36, 2}, { 48, 1}, { 52, 1}, { 60, 1}, 72 { 64, 1}, { 56, 1}, { 76, 1}, { 68, 1}, { 80, 1}, { 84, 1}, 73 { 72, 1}, { 86, 1}, { 35, 1}, { 39, 1}, { 50, 1}, { 38, 1}, 74 { 37, 1}, { 46, 1}, { 98, 1}, {102, 1}, {128, 1}, { 51, 1}, 75 {107, 1}, { 0, 0} 76 }; 77 78 #define ALIGN_NUM 1024 79 #define ALIGN_MASK (ALIGN_NUM - 1) 80 static uint8_t strlen_align_arr[ALIGN_NUM]; 81 82 /* Alignment data for strlen based on SPEC2017. */ 83 static align_data_t string_align_freq[] = 84 { 85 {8, 470}, {32, 427}, {16, 99}, {1, 19}, {2, 6}, {4, 3}, {0, 0} 86 }; 87 88 static void 89 init_strlen_distribution (void) 90 { 91 int i, j, freq, size, n; 92 93 for (n = i = 0; (freq = strlen_len_freq[i].freq) != 0; i++) 94 for (j = 0, size = strlen_len_freq[i].size; j < freq; j++) 95 strlen_len_arr[n++] = size; 96 assert (n == SIZE_NUM); 97 98 for (n = i = 0; (freq = string_align_freq[i].freq) != 0; i++) 99 for (j = 0, size = string_align_freq[i].align; j < freq; j++) 100 strlen_align_arr[n++] = size; 101 assert (n == ALIGN_NUM); 102 } 103 104 static void 105 init_strlen_tests (void) 106 { 107 uint16_t index[MAX_ALIGN]; 108 109 memset (a, 'x', sizeof (a)); 110 111 /* Create indices for strings at all alignments. */ 112 for (int i = 0; i < MAX_ALIGN; i++) 113 { 114 index[i] = i * (MAX_STRLEN + 1); 115 a[index[i] + MAX_STRLEN] = 0; 116 } 117 118 /* Create a random set of strlen input strings using the string length 119 and alignment distributions. */ 120 for (int n = 0; n < NUM_TESTS; n++) 121 { 122 int align = strlen_align_arr[rand32 (0) & ALIGN_MASK]; 123 int exp_len = strlen_len_arr[rand32 (0) & SIZE_MASK]; 124 125 strlen_tests[n] = 126 index[(align + exp_len) & (MAX_ALIGN - 1)] + MAX_STRLEN - exp_len; 127 } 128 } 129 130 static volatile size_t maskv = 0; 131 132 int main (void) 133 { 134 rand32 (0x12345678); 135 init_strlen_distribution (); 136 init_strlen_tests (); 137 138 printf ("\nRandom strlen (bytes/ns):\n"); 139 for (int f = 0; funtab[f].name != 0; f++) 140 { 141 size_t res = 0, strlen_size = 0, mask = maskv; 142 printf ("%22s ", funtab[f].name); 143 144 for (int c = 0; c < NUM_TESTS; c++) 145 strlen_size += funtab[f].fun (a + strlen_tests[c]); 146 strlen_size *= ITERS; 147 148 /* Measure latency of strlen result with (res & mask). */ 149 uint64_t t = clock_get_ns (); 150 for (int i = 0; i < ITERS; i++) 151 for (int c = 0; c < NUM_TESTS; c++) 152 res = funtab[f].fun (a + strlen_tests[c] + (res & mask)); 153 t = clock_get_ns () - t; 154 printf ("%.2f\n", (double)strlen_size / t); 155 } 156 157 printf ("\nSmall aligned strlen (bytes/ns):\n"); 158 for (int f = 0; funtab[f].name != 0; f++) 159 { 160 printf ("%22s ", funtab[f].name); 161 162 for (int size = 1; size <= 64; size *= 2) 163 { 164 memset (a, 'x', size); 165 a[size - 1] = 0; 166 167 uint64_t t = clock_get_ns (); 168 for (int i = 0; i < ITERS2; i++) 169 funtab[f].fun (a); 170 t = clock_get_ns () - t; 171 printf ("%d%c: %.2f ", size < 1024 ? size : size / 1024, 172 size < 1024 ? 'B' : 'K', (double)size * ITERS2 / t); 173 } 174 printf ("\n"); 175 } 176 177 printf ("\nSmall unaligned strlen (bytes/ns):\n"); 178 for (int f = 0; funtab[f].name != 0; f++) 179 { 180 printf ("%22s ", funtab[f].name); 181 182 int align = 9; 183 for (int size = 1; size <= 64; size *= 2) 184 { 185 memset (a + align, 'x', size); 186 a[align + size - 1] = 0; 187 188 uint64_t t = clock_get_ns (); 189 for (int i = 0; i < ITERS2; i++) 190 funtab[f].fun (a + align); 191 t = clock_get_ns () - t; 192 printf ("%d%c: %.2f ", size < 1024 ? size : size / 1024, 193 size < 1024 ? 'B' : 'K', (double)size * ITERS2 / t); 194 } 195 printf ("\n"); 196 } 197 198 printf ("\nMedium strlen (bytes/ns):\n"); 199 for (int f = 0; funtab[f].name != 0; f++) 200 { 201 printf ("%22s ", funtab[f].name); 202 203 for (int size = 128; size <= 4096; size *= 2) 204 { 205 memset (a, 'x', size); 206 a[size - 1] = 0; 207 208 uint64_t t = clock_get_ns (); 209 for (int i = 0; i < ITERS3; i++) 210 funtab[f].fun (a); 211 t = clock_get_ns () - t; 212 printf ("%d%c: %.2f ", size < 1024 ? size : size / 1024, 213 size < 1024 ? 'B' : 'K', (double)size * ITERS3 / t); 214 } 215 printf ("\n"); 216 } 217 218 printf ("\n"); 219 220 return 0; 221 } 222