1d49ad206SAndrew Turner /* 2d49ad206SAndrew Turner * memset benchmark. 3d49ad206SAndrew Turner * 4d49ad206SAndrew Turner * Copyright (c) 2021, Arm Limited. 5*072a4ba8SAndrew 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 23d49ad206SAndrew Turner static uint8_t a[MAX_SIZE + 4096] __attribute__((__aligned__(64))); 24d49ad206SAndrew Turner 25d49ad206SAndrew Turner #define F(x) {#x, x}, 26d49ad206SAndrew Turner 27d49ad206SAndrew Turner static const struct fun 28d49ad206SAndrew Turner { 29d49ad206SAndrew Turner const char *name; 30d49ad206SAndrew Turner void *(*fun)(void *, int, size_t); 31d49ad206SAndrew Turner } funtab[] = 32d49ad206SAndrew Turner { 33d49ad206SAndrew Turner #if __aarch64__ 34d49ad206SAndrew Turner F(__memset_aarch64) 35d49ad206SAndrew Turner #elif __arm__ 36d49ad206SAndrew Turner F(__memset_arm) 37d49ad206SAndrew Turner #endif 38d49ad206SAndrew Turner F(memset) 39d49ad206SAndrew Turner #undef F 40d49ad206SAndrew Turner {0, 0} 41d49ad206SAndrew Turner }; 42d49ad206SAndrew Turner 43d49ad206SAndrew Turner typedef struct { uint32_t offset : 20, len : 12; } memset_test_t; 44d49ad206SAndrew Turner static memset_test_t test_arr[NUM_TESTS]; 45d49ad206SAndrew Turner 46d49ad206SAndrew Turner typedef struct { uint16_t size; uint16_t freq; } freq_data_t; 47d49ad206SAndrew Turner typedef struct { uint8_t align; uint16_t freq; } align_data_t; 48d49ad206SAndrew Turner 49d49ad206SAndrew Turner #define SIZE_NUM 65536 50d49ad206SAndrew Turner #define SIZE_MASK (SIZE_NUM-1) 51d49ad206SAndrew Turner static uint8_t len_arr[SIZE_NUM]; 52d49ad206SAndrew Turner 53d49ad206SAndrew Turner /* Frequency data for memset sizes up to 4096 based on SPEC2017. */ 54d49ad206SAndrew Turner static freq_data_t memset_len_freq[] = 55d49ad206SAndrew Turner { 56d49ad206SAndrew Turner {40,28817}, {32,15336}, { 16,3823}, {296,3545}, { 24,3454}, { 8,1412}, 57d49ad206SAndrew Turner {292,1202}, { 48, 927}, { 12, 613}, { 11, 539}, {284, 493}, {108, 414}, 58d49ad206SAndrew Turner { 88, 380}, { 20, 295}, {312, 271}, { 72, 233}, { 2, 200}, { 4, 192}, 59d49ad206SAndrew Turner { 15, 180}, { 14, 174}, { 13, 160}, { 56, 151}, { 36, 144}, { 64, 140}, 60d49ad206SAndrew Turner {4095,133}, { 10, 130}, { 9, 124}, { 3, 124}, { 28, 120}, { 0, 118}, 61d49ad206SAndrew Turner {288, 110}, {1152, 96}, {104, 90}, { 1, 86}, {832, 76}, {248, 74}, 62d49ad206SAndrew Turner {1024, 69}, {120, 64}, {512, 63}, {384, 60}, { 6, 59}, { 80, 54}, 63d49ad206SAndrew Turner { 17, 50}, { 7, 49}, {520, 47}, {2048, 39}, {256, 37}, {864, 33}, 64d49ad206SAndrew Turner {1440, 28}, { 22, 27}, {2056, 24}, {260, 23}, { 68, 23}, { 5, 22}, 65d49ad206SAndrew Turner { 18, 21}, {200, 18}, {2120, 18}, { 60, 17}, { 52, 16}, {336, 15}, 66d49ad206SAndrew Turner { 44, 13}, {192, 13}, {160, 12}, {2064, 12}, {128, 12}, { 76, 11}, 67d49ad206SAndrew Turner {164, 11}, {152, 10}, {136, 9}, {488, 7}, { 96, 6}, {560, 6}, 68d49ad206SAndrew Turner {1016, 6}, {112, 5}, {232, 5}, {168, 5}, {952, 5}, {184, 5}, 69d49ad206SAndrew Turner {144, 4}, {252, 4}, { 84, 3}, {960, 3}, {3808, 3}, {244, 3}, 70d49ad206SAndrew Turner {280, 3}, {224, 3}, {156, 3}, {1088, 3}, {440, 3}, {216, 2}, 71d49ad206SAndrew Turner {304, 2}, { 23, 2}, { 25, 2}, { 26, 2}, {264, 2}, {328, 2}, 72d49ad206SAndrew Turner {1096, 2}, {240, 2}, {1104, 2}, {704, 2}, {1664, 2}, {360, 2}, 73d49ad206SAndrew Turner {808, 1}, {544, 1}, {236, 1}, {720, 1}, {368, 1}, {424, 1}, 74d49ad206SAndrew Turner {640, 1}, {1112, 1}, {552, 1}, {272, 1}, {776, 1}, {376, 1}, 75d49ad206SAndrew Turner { 92, 1}, {536, 1}, {824, 1}, {496, 1}, {760, 1}, {792, 1}, 76d49ad206SAndrew Turner {504, 1}, {344, 1}, {1816, 1}, {880, 1}, {176, 1}, {320, 1}, 77d49ad206SAndrew Turner {352, 1}, {2008, 1}, {208, 1}, {408, 1}, {228, 1}, {2072, 1}, 78d49ad206SAndrew Turner {568, 1}, {220, 1}, {616, 1}, {600, 1}, {392, 1}, {696, 1}, 79d49ad206SAndrew Turner {2144, 1}, {1280, 1}, {2136, 1}, {632, 1}, {584, 1}, {456, 1}, 80d49ad206SAndrew Turner {472, 1}, {3440, 1}, {2088, 1}, {680, 1}, {2928, 1}, {212, 1}, 81d49ad206SAndrew Turner {648, 1}, {1752, 1}, {664, 1}, {3512, 1}, {1032, 1}, {528, 1}, 82d49ad206SAndrew Turner {4072, 1}, {204, 1}, {2880, 1}, {3392, 1}, {712, 1}, { 59, 1}, 83d49ad206SAndrew Turner {736, 1}, {592, 1}, {2520, 1}, {744, 1}, {196, 1}, {172, 1}, 84d49ad206SAndrew Turner {728, 1}, {2040, 1}, {1192, 1}, {3600, 1}, {0, 0} 85d49ad206SAndrew Turner }; 86d49ad206SAndrew Turner 87d49ad206SAndrew Turner #define ALIGN_NUM 1024 88d49ad206SAndrew Turner #define ALIGN_MASK (ALIGN_NUM-1) 89d49ad206SAndrew Turner static uint8_t align_arr[ALIGN_NUM]; 90d49ad206SAndrew Turner 91d49ad206SAndrew Turner /* Alignment data for memset based on SPEC2017. */ 92d49ad206SAndrew Turner static align_data_t memset_align_freq[] = 93d49ad206SAndrew Turner { 94d49ad206SAndrew Turner {16, 338}, {8, 307}, {32, 148}, {64, 131}, {4, 72}, {1, 23}, {2, 5}, {0, 0} 95d49ad206SAndrew Turner }; 96d49ad206SAndrew Turner 97d49ad206SAndrew Turner static void 98d49ad206SAndrew Turner init_memset_distribution (void) 99d49ad206SAndrew Turner { 100d49ad206SAndrew Turner int i, j, freq, size, n; 101d49ad206SAndrew Turner 102d49ad206SAndrew Turner for (n = i = 0; (freq = memset_len_freq[i].freq) != 0; i++) 103d49ad206SAndrew Turner for (j = 0, size = memset_len_freq[i].size; j < freq; j++) 104d49ad206SAndrew Turner len_arr[n++] = size; 105d49ad206SAndrew Turner assert (n == SIZE_NUM); 106d49ad206SAndrew Turner 107d49ad206SAndrew Turner for (n = i = 0; (freq = memset_align_freq[i].freq) != 0; i++) 108d49ad206SAndrew Turner for (j = 0, size = memset_align_freq[i].align; j < freq; j++) 109d49ad206SAndrew Turner align_arr[n++] = size - 1; 110d49ad206SAndrew Turner assert (n == ALIGN_NUM); 111d49ad206SAndrew Turner } 112d49ad206SAndrew Turner 113d49ad206SAndrew Turner static size_t 114d49ad206SAndrew Turner init_memset (size_t max_size) 115d49ad206SAndrew Turner { 116d49ad206SAndrew Turner size_t total = 0; 117d49ad206SAndrew Turner /* Create a random set of memsets with the given size and alignment 118d49ad206SAndrew Turner distributions. */ 119d49ad206SAndrew Turner for (int i = 0; i < NUM_TESTS; i++) 120d49ad206SAndrew Turner { 121d49ad206SAndrew Turner test_arr[i].offset = (rand32 (0) & (max_size - 1)); 122d49ad206SAndrew Turner test_arr[i].offset &= ~align_arr[rand32 (0) & ALIGN_MASK]; 123d49ad206SAndrew Turner test_arr[i].len = len_arr[rand32 (0) & SIZE_MASK]; 124d49ad206SAndrew Turner total += test_arr[i].len; 125d49ad206SAndrew Turner } 126d49ad206SAndrew Turner 127d49ad206SAndrew Turner return total; 128d49ad206SAndrew Turner } 129d49ad206SAndrew Turner 130d49ad206SAndrew Turner 131d49ad206SAndrew Turner int main (void) 132d49ad206SAndrew Turner { 133d49ad206SAndrew Turner init_memset_distribution (); 134d49ad206SAndrew Turner 135d49ad206SAndrew Turner memset (a, 1, sizeof (a)); 136d49ad206SAndrew Turner 137d49ad206SAndrew Turner printf("Random memset (bytes/ns):\n"); 138d49ad206SAndrew Turner for (int f = 0; funtab[f].name != 0; f++) 139d49ad206SAndrew Turner { 140d49ad206SAndrew Turner size_t total_size = 0; 141d49ad206SAndrew Turner uint64_t tsum = 0; 142d49ad206SAndrew Turner printf ("%22s ", funtab[f].name); 143d49ad206SAndrew Turner rand32 (0x12345678); 144d49ad206SAndrew Turner 145d49ad206SAndrew Turner for (int size = MIN_SIZE; size <= MAX_SIZE; size *= 2) 146d49ad206SAndrew Turner { 147d49ad206SAndrew Turner size_t memset_size = init_memset (size) * ITERS; 148d49ad206SAndrew Turner 149d49ad206SAndrew Turner for (int c = 0; c < NUM_TESTS; c++) 150d49ad206SAndrew Turner funtab[f].fun (a + test_arr[c].offset, 0, test_arr[c].len); 151d49ad206SAndrew Turner 152d49ad206SAndrew Turner uint64_t t = clock_get_ns (); 153d49ad206SAndrew Turner for (int i = 0; i < ITERS; i++) 154d49ad206SAndrew Turner for (int c = 0; c < NUM_TESTS; c++) 155d49ad206SAndrew Turner funtab[f].fun (a + test_arr[c].offset, 0, test_arr[c].len); 156d49ad206SAndrew Turner t = clock_get_ns () - t; 157d49ad206SAndrew Turner total_size += memset_size; 158d49ad206SAndrew Turner tsum += t; 159d49ad206SAndrew Turner printf ("%dK: %.2f ", size / 1024, (double)memset_size / t); 160d49ad206SAndrew Turner } 161d49ad206SAndrew Turner printf( "avg %.2f\n", (double)total_size / tsum); 162d49ad206SAndrew Turner } 163d49ad206SAndrew Turner 164d49ad206SAndrew Turner size_t total_size = 0; 165d49ad206SAndrew Turner uint64_t tsum = 0; 166d49ad206SAndrew Turner printf ("%22s ", "memset_call"); 167d49ad206SAndrew Turner rand32 (0x12345678); 168d49ad206SAndrew Turner 169d49ad206SAndrew Turner for (int size = MIN_SIZE; size <= MAX_SIZE; size *= 2) 170d49ad206SAndrew Turner { 171d49ad206SAndrew Turner size_t memset_size = init_memset (size) * ITERS; 172d49ad206SAndrew Turner 173d49ad206SAndrew Turner for (int c = 0; c < NUM_TESTS; c++) 174d49ad206SAndrew Turner memset (a + test_arr[c].offset, 0, test_arr[c].len); 175d49ad206SAndrew Turner 176d49ad206SAndrew Turner uint64_t t = clock_get_ns (); 177d49ad206SAndrew Turner for (int i = 0; i < ITERS; i++) 178d49ad206SAndrew Turner for (int c = 0; c < NUM_TESTS; c++) 179d49ad206SAndrew Turner memset (a + test_arr[c].offset, 0, test_arr[c].len); 180d49ad206SAndrew Turner t = clock_get_ns () - t; 181d49ad206SAndrew Turner total_size += memset_size; 182d49ad206SAndrew Turner tsum += t; 183d49ad206SAndrew Turner printf ("%dK: %.2f ", size / 1024, (double)memset_size / t); 184d49ad206SAndrew Turner } 185d49ad206SAndrew Turner printf( "avg %.2f\n", (double)total_size / tsum); 186d49ad206SAndrew Turner 187d49ad206SAndrew Turner 188d49ad206SAndrew Turner printf ("\nMedium memset (bytes/ns):\n"); 189d49ad206SAndrew Turner for (int f = 0; funtab[f].name != 0; f++) 190d49ad206SAndrew Turner { 191d49ad206SAndrew Turner printf ("%22s ", funtab[f].name); 192d49ad206SAndrew Turner 193d49ad206SAndrew Turner for (int size = 8; size <= 512; size *= 2) 194d49ad206SAndrew Turner { 195d49ad206SAndrew Turner uint64_t t = clock_get_ns (); 196d49ad206SAndrew Turner for (int i = 0; i < ITERS2; i++) 197d49ad206SAndrew Turner funtab[f].fun (a, 0, size); 198d49ad206SAndrew Turner t = clock_get_ns () - t; 199d49ad206SAndrew Turner printf ("%dB: %.2f ", size, (double)size * ITERS2 / t); 200d49ad206SAndrew Turner } 201d49ad206SAndrew Turner printf ("\n"); 202d49ad206SAndrew Turner } 203d49ad206SAndrew Turner 204d49ad206SAndrew Turner printf ("%22s ", "memset_call"); 205d49ad206SAndrew Turner for (int size = 8; size <= 512; size *= 2) 206d49ad206SAndrew Turner { 207d49ad206SAndrew Turner uint64_t t = clock_get_ns (); 208d49ad206SAndrew Turner for (int i = 0; i < ITERS2; i++) 209d49ad206SAndrew Turner memset (a, 0, size); 210d49ad206SAndrew Turner t = clock_get_ns () - t; 211d49ad206SAndrew Turner printf ("%dB: %.2f ", size, (double)size * ITERS2 / t); 212d49ad206SAndrew Turner } 213d49ad206SAndrew Turner 214d49ad206SAndrew Turner 215d49ad206SAndrew Turner printf ("\nLarge memset (bytes/ns):\n"); 216d49ad206SAndrew Turner for (int f = 0; funtab[f].name != 0; f++) 217d49ad206SAndrew Turner { 218d49ad206SAndrew Turner printf ("%22s ", funtab[f].name); 219d49ad206SAndrew Turner 220d49ad206SAndrew Turner for (int size = 1024; size <= 65536; size *= 2) 221d49ad206SAndrew Turner { 222d49ad206SAndrew Turner uint64_t t = clock_get_ns (); 223d49ad206SAndrew Turner for (int i = 0; i < ITERS3; i++) 224d49ad206SAndrew Turner funtab[f].fun (a, 0, size); 225d49ad206SAndrew Turner t = clock_get_ns () - t; 226d49ad206SAndrew Turner printf ("%dK: %.2f ", size / 1024, (double)size * ITERS3 / t); 227d49ad206SAndrew Turner } 228d49ad206SAndrew Turner printf ("\n"); 229d49ad206SAndrew Turner } 230d49ad206SAndrew Turner 231d49ad206SAndrew Turner printf ("%22s ", "memset_call"); 232d49ad206SAndrew Turner for (int size = 1024; size <= 65536; size *= 2) 233d49ad206SAndrew Turner { 234d49ad206SAndrew Turner uint64_t t = clock_get_ns (); 235d49ad206SAndrew Turner for (int i = 0; i < ITERS3; i++) 236d49ad206SAndrew Turner memset (a, 0, size); 237d49ad206SAndrew Turner t = clock_get_ns () - t; 238d49ad206SAndrew Turner printf ("%dK: %.2f ", size / 1024, (double)size * ITERS3 / t); 239d49ad206SAndrew Turner } 240d49ad206SAndrew Turner printf ("\n\n"); 241d49ad206SAndrew Turner 242d49ad206SAndrew Turner return 0; 243d49ad206SAndrew Turner } 244