1*d49ad206SAndrew Turner /* 2*d49ad206SAndrew Turner * memset benchmark. 3*d49ad206SAndrew Turner * 4*d49ad206SAndrew Turner * Copyright (c) 2021, Arm Limited. 5*d49ad206SAndrew Turner * SPDX-License-Identifier: MIT 6*d49ad206SAndrew Turner */ 7*d49ad206SAndrew Turner 8*d49ad206SAndrew Turner #define _GNU_SOURCE 9*d49ad206SAndrew Turner #include <stdint.h> 10*d49ad206SAndrew Turner #include <stdio.h> 11*d49ad206SAndrew Turner #include <string.h> 12*d49ad206SAndrew Turner #include <assert.h> 13*d49ad206SAndrew Turner #include "stringlib.h" 14*d49ad206SAndrew Turner #include "benchlib.h" 15*d49ad206SAndrew Turner 16*d49ad206SAndrew Turner #define ITERS 5000 17*d49ad206SAndrew Turner #define ITERS2 20000000 18*d49ad206SAndrew Turner #define ITERS3 1000000 19*d49ad206SAndrew Turner #define NUM_TESTS 16384 20*d49ad206SAndrew Turner #define MIN_SIZE 32768 21*d49ad206SAndrew Turner #define MAX_SIZE (1024 * 1024) 22*d49ad206SAndrew Turner 23*d49ad206SAndrew Turner static uint8_t a[MAX_SIZE + 4096] __attribute__((__aligned__(64))); 24*d49ad206SAndrew Turner 25*d49ad206SAndrew Turner #define F(x) {#x, x}, 26*d49ad206SAndrew Turner 27*d49ad206SAndrew Turner static const struct fun 28*d49ad206SAndrew Turner { 29*d49ad206SAndrew Turner const char *name; 30*d49ad206SAndrew Turner void *(*fun)(void *, int, size_t); 31*d49ad206SAndrew Turner } funtab[] = 32*d49ad206SAndrew Turner { 33*d49ad206SAndrew Turner #if __aarch64__ 34*d49ad206SAndrew Turner F(__memset_aarch64) 35*d49ad206SAndrew Turner #elif __arm__ 36*d49ad206SAndrew Turner F(__memset_arm) 37*d49ad206SAndrew Turner #endif 38*d49ad206SAndrew Turner F(memset) 39*d49ad206SAndrew Turner #undef F 40*d49ad206SAndrew Turner {0, 0} 41*d49ad206SAndrew Turner }; 42*d49ad206SAndrew Turner 43*d49ad206SAndrew Turner typedef struct { uint32_t offset : 20, len : 12; } memset_test_t; 44*d49ad206SAndrew Turner static memset_test_t test_arr[NUM_TESTS]; 45*d49ad206SAndrew Turner 46*d49ad206SAndrew Turner typedef struct { uint16_t size; uint16_t freq; } freq_data_t; 47*d49ad206SAndrew Turner typedef struct { uint8_t align; uint16_t freq; } align_data_t; 48*d49ad206SAndrew Turner 49*d49ad206SAndrew Turner #define SIZE_NUM 65536 50*d49ad206SAndrew Turner #define SIZE_MASK (SIZE_NUM-1) 51*d49ad206SAndrew Turner static uint8_t len_arr[SIZE_NUM]; 52*d49ad206SAndrew Turner 53*d49ad206SAndrew Turner /* Frequency data for memset sizes up to 4096 based on SPEC2017. */ 54*d49ad206SAndrew Turner static freq_data_t memset_len_freq[] = 55*d49ad206SAndrew Turner { 56*d49ad206SAndrew Turner {40,28817}, {32,15336}, { 16,3823}, {296,3545}, { 24,3454}, { 8,1412}, 57*d49ad206SAndrew Turner {292,1202}, { 48, 927}, { 12, 613}, { 11, 539}, {284, 493}, {108, 414}, 58*d49ad206SAndrew Turner { 88, 380}, { 20, 295}, {312, 271}, { 72, 233}, { 2, 200}, { 4, 192}, 59*d49ad206SAndrew Turner { 15, 180}, { 14, 174}, { 13, 160}, { 56, 151}, { 36, 144}, { 64, 140}, 60*d49ad206SAndrew Turner {4095,133}, { 10, 130}, { 9, 124}, { 3, 124}, { 28, 120}, { 0, 118}, 61*d49ad206SAndrew Turner {288, 110}, {1152, 96}, {104, 90}, { 1, 86}, {832, 76}, {248, 74}, 62*d49ad206SAndrew Turner {1024, 69}, {120, 64}, {512, 63}, {384, 60}, { 6, 59}, { 80, 54}, 63*d49ad206SAndrew Turner { 17, 50}, { 7, 49}, {520, 47}, {2048, 39}, {256, 37}, {864, 33}, 64*d49ad206SAndrew Turner {1440, 28}, { 22, 27}, {2056, 24}, {260, 23}, { 68, 23}, { 5, 22}, 65*d49ad206SAndrew Turner { 18, 21}, {200, 18}, {2120, 18}, { 60, 17}, { 52, 16}, {336, 15}, 66*d49ad206SAndrew Turner { 44, 13}, {192, 13}, {160, 12}, {2064, 12}, {128, 12}, { 76, 11}, 67*d49ad206SAndrew Turner {164, 11}, {152, 10}, {136, 9}, {488, 7}, { 96, 6}, {560, 6}, 68*d49ad206SAndrew Turner {1016, 6}, {112, 5}, {232, 5}, {168, 5}, {952, 5}, {184, 5}, 69*d49ad206SAndrew Turner {144, 4}, {252, 4}, { 84, 3}, {960, 3}, {3808, 3}, {244, 3}, 70*d49ad206SAndrew Turner {280, 3}, {224, 3}, {156, 3}, {1088, 3}, {440, 3}, {216, 2}, 71*d49ad206SAndrew Turner {304, 2}, { 23, 2}, { 25, 2}, { 26, 2}, {264, 2}, {328, 2}, 72*d49ad206SAndrew Turner {1096, 2}, {240, 2}, {1104, 2}, {704, 2}, {1664, 2}, {360, 2}, 73*d49ad206SAndrew Turner {808, 1}, {544, 1}, {236, 1}, {720, 1}, {368, 1}, {424, 1}, 74*d49ad206SAndrew Turner {640, 1}, {1112, 1}, {552, 1}, {272, 1}, {776, 1}, {376, 1}, 75*d49ad206SAndrew Turner { 92, 1}, {536, 1}, {824, 1}, {496, 1}, {760, 1}, {792, 1}, 76*d49ad206SAndrew Turner {504, 1}, {344, 1}, {1816, 1}, {880, 1}, {176, 1}, {320, 1}, 77*d49ad206SAndrew Turner {352, 1}, {2008, 1}, {208, 1}, {408, 1}, {228, 1}, {2072, 1}, 78*d49ad206SAndrew Turner {568, 1}, {220, 1}, {616, 1}, {600, 1}, {392, 1}, {696, 1}, 79*d49ad206SAndrew Turner {2144, 1}, {1280, 1}, {2136, 1}, {632, 1}, {584, 1}, {456, 1}, 80*d49ad206SAndrew Turner {472, 1}, {3440, 1}, {2088, 1}, {680, 1}, {2928, 1}, {212, 1}, 81*d49ad206SAndrew Turner {648, 1}, {1752, 1}, {664, 1}, {3512, 1}, {1032, 1}, {528, 1}, 82*d49ad206SAndrew Turner {4072, 1}, {204, 1}, {2880, 1}, {3392, 1}, {712, 1}, { 59, 1}, 83*d49ad206SAndrew Turner {736, 1}, {592, 1}, {2520, 1}, {744, 1}, {196, 1}, {172, 1}, 84*d49ad206SAndrew Turner {728, 1}, {2040, 1}, {1192, 1}, {3600, 1}, {0, 0} 85*d49ad206SAndrew Turner }; 86*d49ad206SAndrew Turner 87*d49ad206SAndrew Turner #define ALIGN_NUM 1024 88*d49ad206SAndrew Turner #define ALIGN_MASK (ALIGN_NUM-1) 89*d49ad206SAndrew Turner static uint8_t align_arr[ALIGN_NUM]; 90*d49ad206SAndrew Turner 91*d49ad206SAndrew Turner /* Alignment data for memset based on SPEC2017. */ 92*d49ad206SAndrew Turner static align_data_t memset_align_freq[] = 93*d49ad206SAndrew Turner { 94*d49ad206SAndrew Turner {16, 338}, {8, 307}, {32, 148}, {64, 131}, {4, 72}, {1, 23}, {2, 5}, {0, 0} 95*d49ad206SAndrew Turner }; 96*d49ad206SAndrew Turner 97*d49ad206SAndrew Turner static void 98*d49ad206SAndrew Turner init_memset_distribution (void) 99*d49ad206SAndrew Turner { 100*d49ad206SAndrew Turner int i, j, freq, size, n; 101*d49ad206SAndrew Turner 102*d49ad206SAndrew Turner for (n = i = 0; (freq = memset_len_freq[i].freq) != 0; i++) 103*d49ad206SAndrew Turner for (j = 0, size = memset_len_freq[i].size; j < freq; j++) 104*d49ad206SAndrew Turner len_arr[n++] = size; 105*d49ad206SAndrew Turner assert (n == SIZE_NUM); 106*d49ad206SAndrew Turner 107*d49ad206SAndrew Turner for (n = i = 0; (freq = memset_align_freq[i].freq) != 0; i++) 108*d49ad206SAndrew Turner for (j = 0, size = memset_align_freq[i].align; j < freq; j++) 109*d49ad206SAndrew Turner align_arr[n++] = size - 1; 110*d49ad206SAndrew Turner assert (n == ALIGN_NUM); 111*d49ad206SAndrew Turner } 112*d49ad206SAndrew Turner 113*d49ad206SAndrew Turner static size_t 114*d49ad206SAndrew Turner init_memset (size_t max_size) 115*d49ad206SAndrew Turner { 116*d49ad206SAndrew Turner size_t total = 0; 117*d49ad206SAndrew Turner /* Create a random set of memsets with the given size and alignment 118*d49ad206SAndrew Turner distributions. */ 119*d49ad206SAndrew Turner for (int i = 0; i < NUM_TESTS; i++) 120*d49ad206SAndrew Turner { 121*d49ad206SAndrew Turner test_arr[i].offset = (rand32 (0) & (max_size - 1)); 122*d49ad206SAndrew Turner test_arr[i].offset &= ~align_arr[rand32 (0) & ALIGN_MASK]; 123*d49ad206SAndrew Turner test_arr[i].len = len_arr[rand32 (0) & SIZE_MASK]; 124*d49ad206SAndrew Turner total += test_arr[i].len; 125*d49ad206SAndrew Turner } 126*d49ad206SAndrew Turner 127*d49ad206SAndrew Turner return total; 128*d49ad206SAndrew Turner } 129*d49ad206SAndrew Turner 130*d49ad206SAndrew Turner 131*d49ad206SAndrew Turner int main (void) 132*d49ad206SAndrew Turner { 133*d49ad206SAndrew Turner init_memset_distribution (); 134*d49ad206SAndrew Turner 135*d49ad206SAndrew Turner memset (a, 1, sizeof (a)); 136*d49ad206SAndrew Turner 137*d49ad206SAndrew Turner printf("Random memset (bytes/ns):\n"); 138*d49ad206SAndrew Turner for (int f = 0; funtab[f].name != 0; f++) 139*d49ad206SAndrew Turner { 140*d49ad206SAndrew Turner size_t total_size = 0; 141*d49ad206SAndrew Turner uint64_t tsum = 0; 142*d49ad206SAndrew Turner printf ("%22s ", funtab[f].name); 143*d49ad206SAndrew Turner rand32 (0x12345678); 144*d49ad206SAndrew Turner 145*d49ad206SAndrew Turner for (int size = MIN_SIZE; size <= MAX_SIZE; size *= 2) 146*d49ad206SAndrew Turner { 147*d49ad206SAndrew Turner size_t memset_size = init_memset (size) * ITERS; 148*d49ad206SAndrew Turner 149*d49ad206SAndrew Turner for (int c = 0; c < NUM_TESTS; c++) 150*d49ad206SAndrew Turner funtab[f].fun (a + test_arr[c].offset, 0, test_arr[c].len); 151*d49ad206SAndrew Turner 152*d49ad206SAndrew Turner uint64_t t = clock_get_ns (); 153*d49ad206SAndrew Turner for (int i = 0; i < ITERS; i++) 154*d49ad206SAndrew Turner for (int c = 0; c < NUM_TESTS; c++) 155*d49ad206SAndrew Turner funtab[f].fun (a + test_arr[c].offset, 0, test_arr[c].len); 156*d49ad206SAndrew Turner t = clock_get_ns () - t; 157*d49ad206SAndrew Turner total_size += memset_size; 158*d49ad206SAndrew Turner tsum += t; 159*d49ad206SAndrew Turner printf ("%dK: %.2f ", size / 1024, (double)memset_size / t); 160*d49ad206SAndrew Turner } 161*d49ad206SAndrew Turner printf( "avg %.2f\n", (double)total_size / tsum); 162*d49ad206SAndrew Turner } 163*d49ad206SAndrew Turner 164*d49ad206SAndrew Turner size_t total_size = 0; 165*d49ad206SAndrew Turner uint64_t tsum = 0; 166*d49ad206SAndrew Turner printf ("%22s ", "memset_call"); 167*d49ad206SAndrew Turner rand32 (0x12345678); 168*d49ad206SAndrew Turner 169*d49ad206SAndrew Turner for (int size = MIN_SIZE; size <= MAX_SIZE; size *= 2) 170*d49ad206SAndrew Turner { 171*d49ad206SAndrew Turner size_t memset_size = init_memset (size) * ITERS; 172*d49ad206SAndrew Turner 173*d49ad206SAndrew Turner for (int c = 0; c < NUM_TESTS; c++) 174*d49ad206SAndrew Turner memset (a + test_arr[c].offset, 0, test_arr[c].len); 175*d49ad206SAndrew Turner 176*d49ad206SAndrew Turner uint64_t t = clock_get_ns (); 177*d49ad206SAndrew Turner for (int i = 0; i < ITERS; i++) 178*d49ad206SAndrew Turner for (int c = 0; c < NUM_TESTS; c++) 179*d49ad206SAndrew Turner memset (a + test_arr[c].offset, 0, test_arr[c].len); 180*d49ad206SAndrew Turner t = clock_get_ns () - t; 181*d49ad206SAndrew Turner total_size += memset_size; 182*d49ad206SAndrew Turner tsum += t; 183*d49ad206SAndrew Turner printf ("%dK: %.2f ", size / 1024, (double)memset_size / t); 184*d49ad206SAndrew Turner } 185*d49ad206SAndrew Turner printf( "avg %.2f\n", (double)total_size / tsum); 186*d49ad206SAndrew Turner 187*d49ad206SAndrew Turner 188*d49ad206SAndrew Turner printf ("\nMedium memset (bytes/ns):\n"); 189*d49ad206SAndrew Turner for (int f = 0; funtab[f].name != 0; f++) 190*d49ad206SAndrew Turner { 191*d49ad206SAndrew Turner printf ("%22s ", funtab[f].name); 192*d49ad206SAndrew Turner 193*d49ad206SAndrew Turner for (int size = 8; size <= 512; size *= 2) 194*d49ad206SAndrew Turner { 195*d49ad206SAndrew Turner uint64_t t = clock_get_ns (); 196*d49ad206SAndrew Turner for (int i = 0; i < ITERS2; i++) 197*d49ad206SAndrew Turner funtab[f].fun (a, 0, size); 198*d49ad206SAndrew Turner t = clock_get_ns () - t; 199*d49ad206SAndrew Turner printf ("%dB: %.2f ", size, (double)size * ITERS2 / t); 200*d49ad206SAndrew Turner } 201*d49ad206SAndrew Turner printf ("\n"); 202*d49ad206SAndrew Turner } 203*d49ad206SAndrew Turner 204*d49ad206SAndrew Turner printf ("%22s ", "memset_call"); 205*d49ad206SAndrew Turner for (int size = 8; size <= 512; size *= 2) 206*d49ad206SAndrew Turner { 207*d49ad206SAndrew Turner uint64_t t = clock_get_ns (); 208*d49ad206SAndrew Turner for (int i = 0; i < ITERS2; i++) 209*d49ad206SAndrew Turner memset (a, 0, size); 210*d49ad206SAndrew Turner t = clock_get_ns () - t; 211*d49ad206SAndrew Turner printf ("%dB: %.2f ", size, (double)size * ITERS2 / t); 212*d49ad206SAndrew Turner } 213*d49ad206SAndrew Turner 214*d49ad206SAndrew Turner 215*d49ad206SAndrew Turner printf ("\nLarge memset (bytes/ns):\n"); 216*d49ad206SAndrew Turner for (int f = 0; funtab[f].name != 0; f++) 217*d49ad206SAndrew Turner { 218*d49ad206SAndrew Turner printf ("%22s ", funtab[f].name); 219*d49ad206SAndrew Turner 220*d49ad206SAndrew Turner for (int size = 1024; size <= 65536; size *= 2) 221*d49ad206SAndrew Turner { 222*d49ad206SAndrew Turner uint64_t t = clock_get_ns (); 223*d49ad206SAndrew Turner for (int i = 0; i < ITERS3; i++) 224*d49ad206SAndrew Turner funtab[f].fun (a, 0, size); 225*d49ad206SAndrew Turner t = clock_get_ns () - t; 226*d49ad206SAndrew Turner printf ("%dK: %.2f ", size / 1024, (double)size * ITERS3 / t); 227*d49ad206SAndrew Turner } 228*d49ad206SAndrew Turner printf ("\n"); 229*d49ad206SAndrew Turner } 230*d49ad206SAndrew Turner 231*d49ad206SAndrew Turner printf ("%22s ", "memset_call"); 232*d49ad206SAndrew Turner for (int size = 1024; size <= 65536; size *= 2) 233*d49ad206SAndrew Turner { 234*d49ad206SAndrew Turner uint64_t t = clock_get_ns (); 235*d49ad206SAndrew Turner for (int i = 0; i < ITERS3; i++) 236*d49ad206SAndrew Turner memset (a, 0, size); 237*d49ad206SAndrew Turner t = clock_get_ns () - t; 238*d49ad206SAndrew Turner printf ("%dK: %.2f ", size / 1024, (double)size * ITERS3 / t); 239*d49ad206SAndrew Turner } 240*d49ad206SAndrew Turner printf ("\n\n"); 241*d49ad206SAndrew Turner 242*d49ad206SAndrew Turner return 0; 243*d49ad206SAndrew Turner } 244