xref: /freebsd-src/contrib/arm-optimized-routines/string/bench/memcpy.c (revision f3087bef11543b42e0d69b708f367097a4118d24)
131914882SAlex Richardson /*
231914882SAlex Richardson  * memcpy benchmark.
331914882SAlex Richardson  *
45a02ffc3SAndrew Turner  * Copyright (c) 2020-2023, Arm Limited.
5072a4ba8SAndrew Turner  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
631914882SAlex Richardson  */
731914882SAlex Richardson 
831914882SAlex Richardson #define _GNU_SOURCE
931914882SAlex Richardson #include <stdint.h>
1031914882SAlex Richardson #include <stdio.h>
1131914882SAlex Richardson #include <string.h>
1231914882SAlex Richardson #include <assert.h>
1331914882SAlex Richardson #include "stringlib.h"
1431914882SAlex Richardson #include "benchlib.h"
1531914882SAlex Richardson 
1631914882SAlex Richardson #define ITERS  5000
1731914882SAlex Richardson #define ITERS2 20000000
18d49ad206SAndrew Turner #define ITERS3 200000
19d49ad206SAndrew Turner #define NUM_TESTS 16384
20d49ad206SAndrew Turner #define MIN_SIZE 32768
21d49ad206SAndrew Turner #define MAX_SIZE (1024 * 1024)
2231914882SAlex Richardson 
23*f3087befSAndrew Turner static uint8_t a[MAX_SIZE + 4096 + 64] __attribute__((__aligned__(4096)));
24*f3087befSAndrew Turner static uint8_t b[MAX_SIZE + 4096 + 64] __attribute__((__aligned__(4096)));
2531914882SAlex Richardson 
26*f3087befSAndrew Turner #define DOTEST(STR,TESTFN)			\
27*f3087befSAndrew Turner   printf (STR);					\
28*f3087befSAndrew Turner   RUN (TESTFN, memcpy);				\
29*f3087befSAndrew Turner   RUNA64 (TESTFN, __memcpy_aarch64);		\
30*f3087befSAndrew Turner   RUNA64 (TESTFN, __memcpy_aarch64_simd);	\
31*f3087befSAndrew Turner   RUNSVE (TESTFN, __memcpy_aarch64_sve);	\
32*f3087befSAndrew Turner   RUNMOPS (TESTFN, __memcpy_aarch64_mops);	\
33*f3087befSAndrew Turner   RUNA32 (TESTFN, __memcpy_arm);		\
34*f3087befSAndrew Turner   printf ("\n");
3531914882SAlex Richardson 
3631914882SAlex Richardson typedef struct { uint16_t size; uint16_t freq; } freq_data_t;
3731914882SAlex Richardson typedef struct { uint8_t align; uint16_t freq; } align_data_t;
3831914882SAlex Richardson 
3931914882SAlex Richardson #define SIZE_NUM 65536
4031914882SAlex Richardson #define SIZE_MASK (SIZE_NUM-1)
4131914882SAlex Richardson static uint8_t size_arr[SIZE_NUM];
4231914882SAlex Richardson 
4331914882SAlex Richardson /* Frequency data for memcpy of less than 4096 bytes based on SPEC2017.  */
4431914882SAlex Richardson static freq_data_t size_freq[] =
4531914882SAlex Richardson {
4631914882SAlex Richardson {32,22320}, { 16,9554}, {  8,8915}, {152,5327}, {  4,2159}, {292,2035},
4731914882SAlex Richardson { 12,1608}, { 24,1343}, {1152,895}, {144, 813}, {884, 733}, {284, 721},
4831914882SAlex Richardson {120, 661}, {  2, 649}, {882, 550}, {  5, 475}, {  7, 461}, {108, 460},
4931914882SAlex Richardson { 10, 361}, {  9, 361}, {  6, 334}, {  3, 326}, {464, 308}, {2048,303},
5031914882SAlex Richardson {  1, 298}, { 64, 250}, { 11, 197}, {296, 194}, { 68, 187}, { 15, 185},
5131914882SAlex Richardson {192, 184}, {1764,183}, { 13, 173}, {560, 126}, {160, 115}, {288,  96},
5231914882SAlex Richardson {104,  96}, {1144, 83}, { 18,  80}, { 23,  78}, { 40,  77}, { 19,  68},
5331914882SAlex Richardson { 48,  63}, { 17,  57}, { 72,  54}, {1280, 51}, { 20,  49}, { 28,  47},
5431914882SAlex Richardson { 22,  46}, {640,  45}, { 25,  41}, { 14,  40}, { 56,  37}, { 27,  35},
5531914882SAlex Richardson { 35,  33}, {384,  33}, { 29,  32}, { 80,  30}, {4095, 22}, {232,  22},
5631914882SAlex Richardson { 36,  19}, {184,  17}, { 21,  17}, {256,  16}, { 44,  15}, { 26,  15},
5731914882SAlex Richardson { 31,  14}, { 88,  14}, {176,  13}, { 33,  12}, {1024, 12}, {208,  11},
5831914882SAlex Richardson { 62,  11}, {128,  10}, {704,  10}, {324,  10}, { 96,  10}, { 60,   9},
5931914882SAlex Richardson {136,   9}, {124,   9}, { 34,   8}, { 30,   8}, {480,   8}, {1344,  8},
6031914882SAlex Richardson {273,   7}, {520,   7}, {112,   6}, { 52,   6}, {344,   6}, {336,   6},
6131914882SAlex Richardson {504,   5}, {168,   5}, {424,   5}, {  0,   4}, { 76,   3}, {200,   3},
6231914882SAlex Richardson {512,   3}, {312,   3}, {240,   3}, {960,   3}, {264,   2}, {672,   2},
6331914882SAlex Richardson { 38,   2}, {328,   2}, { 84,   2}, { 39,   2}, {216,   2}, { 42,   2},
6431914882SAlex Richardson { 37,   2}, {1608,  2}, { 70,   2}, { 46,   2}, {536,   2}, {280,   1},
6531914882SAlex Richardson {248,   1}, { 47,   1}, {1088,  1}, {1288,  1}, {224,   1}, { 41,   1},
6631914882SAlex Richardson { 50,   1}, { 49,   1}, {808,   1}, {360,   1}, {440,   1}, { 43,   1},
6731914882SAlex Richardson { 45,   1}, { 78,   1}, {968,   1}, {392,   1}, { 54,   1}, { 53,   1},
6831914882SAlex Richardson { 59,   1}, {376,   1}, {664,   1}, { 58,   1}, {272,   1}, { 66,   1},
6931914882SAlex Richardson {2688,  1}, {472,   1}, {568,   1}, {720,   1}, { 51,   1}, { 63,   1},
7031914882SAlex Richardson { 86,   1}, {496,   1}, {776,   1}, { 57,   1}, {680,   1}, {792,   1},
7131914882SAlex Richardson {122,   1}, {760,   1}, {824,   1}, {552,   1}, { 67,   1}, {456,   1},
7231914882SAlex Richardson {984,   1}, { 74,   1}, {408,   1}, { 75,   1}, { 92,   1}, {576,   1},
7331914882SAlex Richardson {116,   1}, { 65,   1}, {117,   1}, { 82,   1}, {352,   1}, { 55,   1},
7431914882SAlex Richardson {100,   1}, { 90,   1}, {696,   1}, {111,   1}, {880,   1}, { 79,   1},
7531914882SAlex Richardson {488,   1}, { 61,   1}, {114,   1}, { 94,   1}, {1032,  1}, { 98,   1},
7631914882SAlex Richardson { 87,   1}, {584,   1}, { 85,   1}, {648,   1}, {0, 0}
7731914882SAlex Richardson };
7831914882SAlex Richardson 
7931914882SAlex Richardson #define ALIGN_NUM 1024
8031914882SAlex Richardson #define ALIGN_MASK (ALIGN_NUM-1)
8131914882SAlex Richardson static uint8_t src_align_arr[ALIGN_NUM];
8231914882SAlex Richardson static uint8_t dst_align_arr[ALIGN_NUM];
8331914882SAlex Richardson 
8431914882SAlex Richardson /* Source alignment frequency for memcpy based on SPEC2017.  */
8531914882SAlex Richardson static align_data_t src_align_freq[] =
8631914882SAlex Richardson {
8731914882SAlex Richardson   {8, 300}, {16, 292}, {32, 168}, {64, 153}, {4, 79}, {2, 14}, {1, 18}, {0, 0}
8831914882SAlex Richardson };
8931914882SAlex Richardson 
9031914882SAlex Richardson static align_data_t dst_align_freq[] =
9131914882SAlex Richardson {
9231914882SAlex Richardson   {8, 265}, {16, 263}, {64, 209}, {32, 174}, {4, 90}, {2, 10}, {1, 13}, {0, 0}
9331914882SAlex Richardson };
9431914882SAlex Richardson 
9531914882SAlex Richardson typedef struct
9631914882SAlex Richardson {
9731914882SAlex Richardson   uint64_t src : 24;
9831914882SAlex Richardson   uint64_t dst : 24;
9931914882SAlex Richardson   uint64_t len : 16;
10031914882SAlex Richardson } copy_t;
10131914882SAlex Richardson 
102d49ad206SAndrew Turner static copy_t test_arr[NUM_TESTS];
10331914882SAlex Richardson 
10431914882SAlex Richardson typedef char *(*proto_t) (char *, const char *, size_t);
10531914882SAlex Richardson 
10631914882SAlex Richardson static void
10731914882SAlex Richardson init_copy_distribution (void)
10831914882SAlex Richardson {
10931914882SAlex Richardson   int i, j, freq, size, n;
11031914882SAlex Richardson 
11131914882SAlex Richardson   for (n = i = 0; (freq = size_freq[i].freq) != 0; i++)
11231914882SAlex Richardson     for (j = 0, size = size_freq[i].size; j < freq; j++)
11331914882SAlex Richardson       size_arr[n++] = size;
11431914882SAlex Richardson   assert (n == SIZE_NUM);
11531914882SAlex Richardson 
11631914882SAlex Richardson   for (n = i = 0; (freq = src_align_freq[i].freq) != 0; i++)
11731914882SAlex Richardson     for (j = 0, size = src_align_freq[i].align; j < freq; j++)
11831914882SAlex Richardson       src_align_arr[n++] = size - 1;
11931914882SAlex Richardson   assert (n == ALIGN_NUM);
12031914882SAlex Richardson 
12131914882SAlex Richardson   for (n = i = 0; (freq = dst_align_freq[i].freq) != 0; i++)
12231914882SAlex Richardson     for (j = 0, size = dst_align_freq[i].align; j < freq; j++)
12331914882SAlex Richardson       dst_align_arr[n++] = size - 1;
12431914882SAlex Richardson   assert (n == ALIGN_NUM);
12531914882SAlex Richardson }
12631914882SAlex Richardson 
12731914882SAlex Richardson static size_t
12831914882SAlex Richardson init_copies (size_t max_size)
12931914882SAlex Richardson {
13031914882SAlex Richardson   size_t total = 0;
13131914882SAlex Richardson   /* Create a random set of copies with the given size and alignment
13231914882SAlex Richardson      distributions.  */
133d49ad206SAndrew Turner   for (int i = 0; i < NUM_TESTS; i++)
13431914882SAlex Richardson     {
135d49ad206SAndrew Turner       test_arr[i].dst = (rand32 (0) & (max_size - 1));
136d49ad206SAndrew Turner       test_arr[i].dst &= ~dst_align_arr[rand32 (0) & ALIGN_MASK];
137d49ad206SAndrew Turner       test_arr[i].src = (rand32 (0) & (max_size - 1));
138d49ad206SAndrew Turner       test_arr[i].src &= ~src_align_arr[rand32 (0) & ALIGN_MASK];
139d49ad206SAndrew Turner       test_arr[i].len = size_arr[rand32 (0) & SIZE_MASK];
140d49ad206SAndrew Turner       total += test_arr[i].len;
14131914882SAlex Richardson     }
14231914882SAlex Richardson 
14331914882SAlex Richardson   return total;
14431914882SAlex Richardson }
14531914882SAlex Richardson 
146*f3087befSAndrew Turner static void inline __attribute ((always_inline))
147*f3087befSAndrew Turner memcpy_random (const char *name, void *(*fn)(void *, const void *, size_t))
148*f3087befSAndrew Turner {
149*f3087befSAndrew Turner   printf ("%22s ", name);
150*f3087befSAndrew Turner   uint64_t total = 0, tsum = 0;
151*f3087befSAndrew Turner   for (int size = MIN_SIZE; size <= MAX_SIZE; size *= 2)
152*f3087befSAndrew Turner     {
153*f3087befSAndrew Turner       uint64_t copy_size = init_copies (size) * ITERS;
154*f3087befSAndrew Turner 
155*f3087befSAndrew Turner       for (int c = 0; c < NUM_TESTS; c++)
156*f3087befSAndrew Turner 	fn (b + test_arr[c].dst, a + test_arr[c].src, test_arr[c].len);
157*f3087befSAndrew Turner 
158*f3087befSAndrew Turner       uint64_t t = clock_get_ns ();
159*f3087befSAndrew Turner       for (int i = 0; i < ITERS; i++)
160*f3087befSAndrew Turner 	for (int c = 0; c < NUM_TESTS; c++)
161*f3087befSAndrew Turner 	  fn (b + test_arr[c].dst, a + test_arr[c].src, test_arr[c].len);
162*f3087befSAndrew Turner       t = clock_get_ns () - t;
163*f3087befSAndrew Turner       total += copy_size;
164*f3087befSAndrew Turner       tsum += t;
165*f3087befSAndrew Turner       printf ("%dK: %5.2f ", size / 1024, (double)copy_size / t);
166*f3087befSAndrew Turner     }
167*f3087befSAndrew Turner   printf( "avg %5.2f\n", (double)total / tsum);
168*f3087befSAndrew Turner }
169*f3087befSAndrew Turner 
170*f3087befSAndrew Turner static void inline __attribute ((always_inline))
171*f3087befSAndrew Turner memcpy_medium_aligned (const char *name, void *(*fn)(void *, const void *, size_t))
172*f3087befSAndrew Turner {
173*f3087befSAndrew Turner   printf ("%22s ", name);
174*f3087befSAndrew Turner 
175*f3087befSAndrew Turner   for (int size = 8; size <= 512; size *= 2)
176*f3087befSAndrew Turner     {
177*f3087befSAndrew Turner       uint64_t t = clock_get_ns ();
178*f3087befSAndrew Turner       for (int i = 0; i < ITERS2; i++)
179*f3087befSAndrew Turner 	fn (b, a, size);
180*f3087befSAndrew Turner       t = clock_get_ns () - t;
181*f3087befSAndrew Turner       printf ("%dB: %5.2f ", size, (double)size * ITERS2 / t);
182*f3087befSAndrew Turner     }
183*f3087befSAndrew Turner   printf ("\n");
184*f3087befSAndrew Turner }
185*f3087befSAndrew Turner 
186*f3087befSAndrew Turner static void inline __attribute ((always_inline))
187*f3087befSAndrew Turner memcpy_medium_unaligned (const char *name, void *(*fn)(void *, const void *, size_t))
188*f3087befSAndrew Turner {
189*f3087befSAndrew Turner   printf ("%22s ", name);
190*f3087befSAndrew Turner 
191*f3087befSAndrew Turner   for (int size = 8; size <= 512; size *= 2)
192*f3087befSAndrew Turner     {
193*f3087befSAndrew Turner       uint64_t t = clock_get_ns ();
194*f3087befSAndrew Turner       for (int i = 0; i < ITERS2; i++)
195*f3087befSAndrew Turner 	fn (b + 3, a + 1, size);
196*f3087befSAndrew Turner       t = clock_get_ns () - t;
197*f3087befSAndrew Turner       printf ("%dB: %5.2f ", size, (double)size * ITERS2 / t);
198*f3087befSAndrew Turner     }
199*f3087befSAndrew Turner   printf ("\n");
200*f3087befSAndrew Turner }
201*f3087befSAndrew Turner 
202*f3087befSAndrew Turner static void inline __attribute ((always_inline))
203*f3087befSAndrew Turner memcpy_large (const char *name, void *(*fn)(void *, const void *, size_t))
204*f3087befSAndrew Turner {
205*f3087befSAndrew Turner   printf ("%22s ", name);
206*f3087befSAndrew Turner 
207*f3087befSAndrew Turner   for (int size = 1024; size <= 65536; size *= 2)
208*f3087befSAndrew Turner     {
209*f3087befSAndrew Turner       uint64_t t = clock_get_ns ();
210*f3087befSAndrew Turner       for (int i = 0; i < ITERS3; i++)
211*f3087befSAndrew Turner 	fn (b, a, size);
212*f3087befSAndrew Turner       t = clock_get_ns () - t;
213*f3087befSAndrew Turner       printf ("%dK: %5.2f ", size / 1024, (double)size * ITERS3 / t);
214*f3087befSAndrew Turner     }
215*f3087befSAndrew Turner   printf ("\n");
216*f3087befSAndrew Turner }
217*f3087befSAndrew Turner 
218*f3087befSAndrew Turner static void inline __attribute ((always_inline))
219*f3087befSAndrew Turner memmove_forward_unaligned (const char *name, void *(*fn)(void *, const void *, size_t))
220*f3087befSAndrew Turner {
221*f3087befSAndrew Turner   printf ("%22s ", name);
222*f3087befSAndrew Turner 
223*f3087befSAndrew Turner   for (int size = 1024; size <= 65536; size *= 2)
224*f3087befSAndrew Turner     {
225*f3087befSAndrew Turner       uint64_t t = clock_get_ns ();
226*f3087befSAndrew Turner       for (int i = 0; i < ITERS3; i++)
227*f3087befSAndrew Turner         fn (a, a + 256 + (i & 31), size);
228*f3087befSAndrew Turner       t = clock_get_ns () - t;
229*f3087befSAndrew Turner       printf ("%dK: %5.2f ", size / 1024, (double)size * ITERS3 / t);
230*f3087befSAndrew Turner     }
231*f3087befSAndrew Turner 
232*f3087befSAndrew Turner   printf ("\n");
233*f3087befSAndrew Turner }
234*f3087befSAndrew Turner 
235*f3087befSAndrew Turner static void inline __attribute ((always_inline))
236*f3087befSAndrew Turner memmove_backward_unaligned (const char *name, void *(*fn)(void *, const void *, size_t))
237*f3087befSAndrew Turner {
238*f3087befSAndrew Turner   printf ("%22s ", name);
239*f3087befSAndrew Turner 
240*f3087befSAndrew Turner   for (int size = 1024; size <= 65536; size *= 2)
241*f3087befSAndrew Turner     {
242*f3087befSAndrew Turner       uint64_t t = clock_get_ns ();
243*f3087befSAndrew Turner       for (int i = 0; i < ITERS3; i++)
244*f3087befSAndrew Turner 	fn (a + 256 + (i & 31), a, size);
245*f3087befSAndrew Turner       t = clock_get_ns () - t;
246*f3087befSAndrew Turner       printf ("%dK: %5.2f ", size / 1024, (double)size * ITERS3 / t);
247*f3087befSAndrew Turner     }
248*f3087befSAndrew Turner 
249*f3087befSAndrew Turner   printf ("\n");
250*f3087befSAndrew Turner }
251*f3087befSAndrew Turner 
25231914882SAlex Richardson int main (void)
25331914882SAlex Richardson {
25431914882SAlex Richardson   init_copy_distribution ();
25531914882SAlex Richardson 
25631914882SAlex Richardson   memset (a, 1, sizeof (a));
25731914882SAlex Richardson   memset (b, 2, sizeof (b));
25831914882SAlex Richardson 
259*f3087befSAndrew Turner   DOTEST ("Random memcpy (bytes/ns):\n", memcpy_random);
260*f3087befSAndrew Turner   DOTEST ("Medium memcpy aligned (bytes/ns):\n", memcpy_medium_aligned);
261*f3087befSAndrew Turner   DOTEST ("Medium memcpy unaligned (bytes/ns):\n", memcpy_medium_unaligned);
262*f3087befSAndrew Turner   DOTEST ("Large memcpy (bytes/ns):\n", memcpy_large);
263*f3087befSAndrew Turner   DOTEST ("Forwards memmove unaligned (bytes/ns):\n", memmove_forward_unaligned);
264*f3087befSAndrew Turner   DOTEST ("Backwards memmove unaligned (bytes/ns):\n", memmove_backward_unaligned);
26531914882SAlex Richardson 
26631914882SAlex Richardson   return 0;
26731914882SAlex Richardson }
268