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