1 /* 2 * memcpy benchmark. 3 * 4 * Copyright (c) 2020, Arm Limited. 5 * SPDX-License-Identifier: MIT 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 500000 19 #define MAX_COPIES 8192 20 #define SIZE (256*1024) 21 22 static uint8_t a[SIZE + 4096] __attribute__((__aligned__(64))); 23 static uint8_t b[SIZE + 4096] __attribute__((__aligned__(64))); 24 25 #define F(x) {#x, x}, 26 27 static const struct fun 28 { 29 const char *name; 30 void *(*fun)(void *, const void *, size_t); 31 } funtab[] = 32 { 33 F(memcpy) 34 #if __aarch64__ 35 F(__memcpy_aarch64) 36 # if __ARM_NEON 37 F(__memcpy_aarch64_simd) 38 # endif 39 #elif __arm__ 40 F(__memcpy_arm) 41 #endif 42 #undef F 43 {0, 0} 44 }; 45 46 typedef struct { uint16_t size; uint16_t freq; } freq_data_t; 47 typedef struct { uint8_t align; uint16_t freq; } align_data_t; 48 49 #define SIZE_NUM 65536 50 #define SIZE_MASK (SIZE_NUM-1) 51 static uint8_t size_arr[SIZE_NUM]; 52 53 /* Frequency data for memcpy of less than 4096 bytes based on SPEC2017. */ 54 static freq_data_t size_freq[] = 55 { 56 {32,22320}, { 16,9554}, { 8,8915}, {152,5327}, { 4,2159}, {292,2035}, 57 { 12,1608}, { 24,1343}, {1152,895}, {144, 813}, {884, 733}, {284, 721}, 58 {120, 661}, { 2, 649}, {882, 550}, { 5, 475}, { 7, 461}, {108, 460}, 59 { 10, 361}, { 9, 361}, { 6, 334}, { 3, 326}, {464, 308}, {2048,303}, 60 { 1, 298}, { 64, 250}, { 11, 197}, {296, 194}, { 68, 187}, { 15, 185}, 61 {192, 184}, {1764,183}, { 13, 173}, {560, 126}, {160, 115}, {288, 96}, 62 {104, 96}, {1144, 83}, { 18, 80}, { 23, 78}, { 40, 77}, { 19, 68}, 63 { 48, 63}, { 17, 57}, { 72, 54}, {1280, 51}, { 20, 49}, { 28, 47}, 64 { 22, 46}, {640, 45}, { 25, 41}, { 14, 40}, { 56, 37}, { 27, 35}, 65 { 35, 33}, {384, 33}, { 29, 32}, { 80, 30}, {4095, 22}, {232, 22}, 66 { 36, 19}, {184, 17}, { 21, 17}, {256, 16}, { 44, 15}, { 26, 15}, 67 { 31, 14}, { 88, 14}, {176, 13}, { 33, 12}, {1024, 12}, {208, 11}, 68 { 62, 11}, {128, 10}, {704, 10}, {324, 10}, { 96, 10}, { 60, 9}, 69 {136, 9}, {124, 9}, { 34, 8}, { 30, 8}, {480, 8}, {1344, 8}, 70 {273, 7}, {520, 7}, {112, 6}, { 52, 6}, {344, 6}, {336, 6}, 71 {504, 5}, {168, 5}, {424, 5}, { 0, 4}, { 76, 3}, {200, 3}, 72 {512, 3}, {312, 3}, {240, 3}, {960, 3}, {264, 2}, {672, 2}, 73 { 38, 2}, {328, 2}, { 84, 2}, { 39, 2}, {216, 2}, { 42, 2}, 74 { 37, 2}, {1608, 2}, { 70, 2}, { 46, 2}, {536, 2}, {280, 1}, 75 {248, 1}, { 47, 1}, {1088, 1}, {1288, 1}, {224, 1}, { 41, 1}, 76 { 50, 1}, { 49, 1}, {808, 1}, {360, 1}, {440, 1}, { 43, 1}, 77 { 45, 1}, { 78, 1}, {968, 1}, {392, 1}, { 54, 1}, { 53, 1}, 78 { 59, 1}, {376, 1}, {664, 1}, { 58, 1}, {272, 1}, { 66, 1}, 79 {2688, 1}, {472, 1}, {568, 1}, {720, 1}, { 51, 1}, { 63, 1}, 80 { 86, 1}, {496, 1}, {776, 1}, { 57, 1}, {680, 1}, {792, 1}, 81 {122, 1}, {760, 1}, {824, 1}, {552, 1}, { 67, 1}, {456, 1}, 82 {984, 1}, { 74, 1}, {408, 1}, { 75, 1}, { 92, 1}, {576, 1}, 83 {116, 1}, { 65, 1}, {117, 1}, { 82, 1}, {352, 1}, { 55, 1}, 84 {100, 1}, { 90, 1}, {696, 1}, {111, 1}, {880, 1}, { 79, 1}, 85 {488, 1}, { 61, 1}, {114, 1}, { 94, 1}, {1032, 1}, { 98, 1}, 86 { 87, 1}, {584, 1}, { 85, 1}, {648, 1}, {0, 0} 87 }; 88 89 #define ALIGN_NUM 1024 90 #define ALIGN_MASK (ALIGN_NUM-1) 91 static uint8_t src_align_arr[ALIGN_NUM]; 92 static uint8_t dst_align_arr[ALIGN_NUM]; 93 94 /* Source alignment frequency for memcpy based on SPEC2017. */ 95 static align_data_t src_align_freq[] = 96 { 97 {8, 300}, {16, 292}, {32, 168}, {64, 153}, {4, 79}, {2, 14}, {1, 18}, {0, 0} 98 }; 99 100 static align_data_t dst_align_freq[] = 101 { 102 {8, 265}, {16, 263}, {64, 209}, {32, 174}, {4, 90}, {2, 10}, {1, 13}, {0, 0} 103 }; 104 105 typedef struct 106 { 107 uint64_t src : 24; 108 uint64_t dst : 24; 109 uint64_t len : 16; 110 } copy_t; 111 112 static copy_t copy[MAX_COPIES]; 113 114 typedef char *(*proto_t) (char *, const char *, size_t); 115 116 static void 117 init_copy_distribution (void) 118 { 119 int i, j, freq, size, n; 120 121 for (n = i = 0; (freq = size_freq[i].freq) != 0; i++) 122 for (j = 0, size = size_freq[i].size; j < freq; j++) 123 size_arr[n++] = size; 124 assert (n == SIZE_NUM); 125 126 for (n = i = 0; (freq = src_align_freq[i].freq) != 0; i++) 127 for (j = 0, size = src_align_freq[i].align; j < freq; j++) 128 src_align_arr[n++] = size - 1; 129 assert (n == ALIGN_NUM); 130 131 for (n = i = 0; (freq = dst_align_freq[i].freq) != 0; i++) 132 for (j = 0, size = dst_align_freq[i].align; j < freq; j++) 133 dst_align_arr[n++] = size - 1; 134 assert (n == ALIGN_NUM); 135 } 136 137 static size_t 138 init_copies (size_t max_size) 139 { 140 size_t total = 0; 141 /* Create a random set of copies with the given size and alignment 142 distributions. */ 143 for (int i = 0; i < MAX_COPIES; i++) 144 { 145 copy[i].dst = (rand32 (0) & (max_size - 1)); 146 copy[i].dst &= ~dst_align_arr[rand32 (0) & ALIGN_MASK]; 147 copy[i].src = (rand32 (0) & (max_size - 1)); 148 copy[i].src &= ~src_align_arr[rand32 (0) & ALIGN_MASK]; 149 copy[i].len = size_arr[rand32 (0) & SIZE_MASK]; 150 total += copy[i].len; 151 } 152 153 return total; 154 } 155 156 int main (void) 157 { 158 init_copy_distribution (); 159 160 memset (a, 1, sizeof (a)); 161 memset (b, 2, sizeof (b)); 162 163 printf("Random memcpy:\n"); 164 for (int f = 0; funtab[f].name != 0; f++) 165 { 166 size_t total = 0; 167 uint64_t tsum = 0; 168 printf ("%22s (B/ns) ", funtab[f].name); 169 rand32 (0x12345678); 170 171 for (int size = 16384; size <= SIZE; size *= 2) 172 { 173 size_t copy_size = init_copies (size) * ITERS; 174 175 for (int c = 0; c < MAX_COPIES; c++) 176 funtab[f].fun (b + copy[c].dst, a + copy[c].src, copy[c].len); 177 178 uint64_t t = clock_get_ns (); 179 for (int i = 0; i < ITERS; i++) 180 for (int c = 0; c < MAX_COPIES; c++) 181 funtab[f].fun (b + copy[c].dst, a + copy[c].src, copy[c].len); 182 t = clock_get_ns () - t; 183 total += copy_size; 184 tsum += t; 185 printf ("%dK: %.2f ", size / 1024, (double)copy_size / t); 186 } 187 printf( "avg %.2f\n", (double)total / tsum); 188 } 189 190 printf ("\nMedium memcpy:\n"); 191 for (int f = 0; funtab[f].name != 0; f++) 192 { 193 printf ("%22s (B/ns) ", funtab[f].name); 194 195 for (int size = 16; size <= 512; size *= 2) 196 { 197 uint64_t t = clock_get_ns (); 198 for (int i = 0; i < ITERS2; i++) 199 funtab[f].fun (b, a, size); 200 t = clock_get_ns () - t; 201 printf ("%d%c: %.2f ", size < 1024 ? size : size / 1024, 202 size < 1024 ? 'B' : 'K', (double)size * ITERS2 / t); 203 } 204 printf ("\n"); 205 } 206 207 printf ("\nLarge memcpy:\n"); 208 for (int f = 0; funtab[f].name != 0; f++) 209 { 210 printf ("%22s (B/ns) ", funtab[f].name); 211 212 for (int size = 1024; size <= 32768; size *= 2) 213 { 214 uint64_t t = clock_get_ns (); 215 for (int i = 0; i < ITERS3; i++) 216 funtab[f].fun (b, a, size); 217 t = clock_get_ns () - t; 218 printf ("%d%c: %.2f ", size < 1024 ? size : size / 1024, 219 size < 1024 ? 'B' : 'K', (double)size * ITERS3 / t); 220 } 221 printf ("\n"); 222 } 223 224 printf ("\nUnaligned forwards memmove:\n"); 225 for (int f = 0; funtab[f].name != 0; f++) 226 { 227 printf ("%22s (B/ns) ", funtab[f].name); 228 229 for (int size = 1024; size <= 32768; size *= 2) 230 { 231 uint64_t t = clock_get_ns (); 232 for (int i = 0; i < ITERS3; i++) 233 funtab[f].fun (a, a + 256 + (i & 31), size); 234 t = clock_get_ns () - t; 235 printf ("%d%c: %.2f ", size < 1024 ? size : size / 1024, 236 size < 1024 ? 'B' : 'K', (double)size * ITERS3 / t); 237 } 238 printf ("\n"); 239 } 240 241 242 printf ("\nUnaligned backwards memmove:\n"); 243 for (int f = 0; funtab[f].name != 0; f++) 244 { 245 printf ("%22s (B/ns) ", funtab[f].name); 246 247 for (int size = 1024; size <= 32768; size *= 2) 248 { 249 uint64_t t = clock_get_ns (); 250 for (int i = 0; i < ITERS3; i++) 251 funtab[f].fun (a + 256 + (i & 31), a, size); 252 t = clock_get_ns () - t; 253 printf ("%d%c: %.2f ", size < 1024 ? size : size / 1024, 254 size < 1024 ? 'B' : 'K', (double)size * ITERS3 / t); 255 } 256 printf ("\n"); 257 } 258 259 return 0; 260 } 261