app/test/test_memcpy_perf.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2014 Intel Corporation
   3  */
   4
   5 #include <stdint.h>
   6 #include <stdio.h>
   7 #include <string.h>
   8 #include <stdlib.h>
   9 #include <sys/time.h>
  10
  11 #include <rte_common.h>
  12 #include <rte_cycles.h>
  13 #include <rte_random.h>
  14 #include <rte_malloc.h>
  15
  16 #include <rte_memcpy.h>
  17
  18 #include "test.h"
  19
  20 /*
  21  * Set this to the maximum buffer size you want to test. If it is 0, then the
  22  * values in the buf_sizes[] array below will be used.
  23  */
  24 #define TEST_VALUE_RANGE        0
  25
  26 /* List of buffer sizes to test */
  27 #if TEST_VALUE_RANGE == 0
  28 static size_t buf_sizes[] = {
  29         1, 2, 3, 4, 5, 6, 7, 8, 9, 12, 15, 16, 17, 31, 32, 33, 63, 64, 65, 127, 128,
  30         129, 191, 192, 193, 255, 256, 257, 319, 320, 321, 383, 384, 385, 447, 448,
  31         449, 511, 512, 513, 767, 768, 769, 1023, 1024, 1025, 1518, 1522, 1536, 1600,
  32         2048, 2560, 3072, 3584, 4096, 4608, 5120, 5632, 6144, 6656, 7168, 7680, 8192
  33 };
  34 /* MUST be as large as largest packet size above */
  35 #define SMALL_BUFFER_SIZE       8192
  36 #else /* TEST_VALUE_RANGE != 0 */
  37 static size_t buf_sizes[TEST_VALUE_RANGE];
  38 #define SMALL_BUFFER_SIZE       TEST_VALUE_RANGE
  39 #endif /* TEST_VALUE_RANGE == 0 */
  40
  41
  42 /*
  43  * Arrays of this size are used for measuring uncached memory accesses by
  44  * picking a random location within the buffer. Make this smaller if there are
  45  * memory allocation errors.
  46  */
  47 #define LARGE_BUFFER_SIZE       (100 * 1024 * 1024)
  48
  49 /* How many times to run timing loop for performance tests */
  50 #define TEST_ITERATIONS         1000000
  51 #define TEST_BATCH_SIZE         100
  52
  53 /* Data is aligned on this many bytes (power of 2) */
  54 #ifdef RTE_MACHINE_CPUFLAG_AVX512F
  55 #define ALIGNMENT_UNIT          64
  56 #elif defined RTE_MACHINE_CPUFLAG_AVX2
  57 #define ALIGNMENT_UNIT          32
  58 #else /* RTE_MACHINE_CPUFLAG */
  59 #define ALIGNMENT_UNIT          16
  60 #endif /* RTE_MACHINE_CPUFLAG */
  61
  62 /*
  63  * Pointers used in performance tests. The two large buffers are for uncached
  64  * access where random addresses within the buffer are used for each
  65  * memcpy. The two small buffers are for cached access.
  66  */
  67 static uint8_t *large_buf_read, *large_buf_write;
  68 static uint8_t *small_buf_read, *small_buf_write;
  69
  70 /* Initialise data buffers. */
  71 static int
  72 init_buffers(void)
  73 {
  74         unsigned i;
  75
  76         large_buf_read = rte_malloc("memcpy", LARGE_BUFFER_SIZE + ALIGNMENT_UNIT, ALIGNMENT_UNIT);
  77         if (large_buf_read == NULL)
  78                 goto error_large_buf_read;
  79
  80         large_buf_write = rte_malloc("memcpy", LARGE_BUFFER_SIZE + ALIGNMENT_UNIT, ALIGNMENT_UNIT);
  81         if (large_buf_write == NULL)
  82                 goto error_large_buf_write;
  83
  84         small_buf_read = rte_malloc("memcpy", SMALL_BUFFER_SIZE + ALIGNMENT_UNIT, ALIGNMENT_UNIT);
  85         if (small_buf_read == NULL)
  86                 goto error_small_buf_read;
  87
  88         small_buf_write = rte_malloc("memcpy", SMALL_BUFFER_SIZE + ALIGNMENT_UNIT, ALIGNMENT_UNIT);
  89         if (small_buf_write == NULL)
  90                 goto error_small_buf_write;
  91
  92         for (i = 0; i < LARGE_BUFFER_SIZE; i++)
  93                 large_buf_read[i] = rte_rand();
  94         for (i = 0; i < SMALL_BUFFER_SIZE; i++)
  95                 small_buf_read[i] = rte_rand();
  96
  97         return 0;
  98
  99 error_small_buf_write:
 100         rte_free(small_buf_read);
 101 error_small_buf_read:
 102         rte_free(large_buf_write);
 103 error_large_buf_write:
 104         rte_free(large_buf_read);
 105 error_large_buf_read:
 106         printf("ERROR: not enough memory\n");
 107         return -1;
 108 }
 109
 110 /* Cleanup data buffers */
 111 static void
 112 free_buffers(void)
 113 {
 114         rte_free(large_buf_read);
 115         rte_free(large_buf_write);
 116         rte_free(small_buf_read);
 117         rte_free(small_buf_write);
 118 }
 119
 120 /*
 121  * Get a random offset into large array, with enough space needed to perform
 122  * max copy size. Offset is aligned, uoffset is used for unalignment setting.
 123  */
 124 static inline size_t
 125 get_rand_offset(size_t uoffset)
 126 {
 127         return ((rte_rand() % (LARGE_BUFFER_SIZE - SMALL_BUFFER_SIZE)) &
 128                         ~(ALIGNMENT_UNIT - 1)) + uoffset;
 129 }
 130
 131 /* Fill in source and destination addresses. */
 132 static inline void
 133 fill_addr_arrays(size_t *dst_addr, int is_dst_cached, size_t dst_uoffset,
 134                                  size_t *src_addr, int is_src_cached, size_t src_uoffset)
 135 {
 136         unsigned int i;
 137
 138         for (i = 0; i < TEST_BATCH_SIZE; i++) {
 139                 dst_addr[i] = (is_dst_cached) ? dst_uoffset : get_rand_offset(dst_uoffset);
 140                 src_addr[i] = (is_src_cached) ? src_uoffset : get_rand_offset(src_uoffset);
 141         }
 142 }
 143
 144 /*
 145  * WORKAROUND: For some reason the first test doing an uncached write
 146  * takes a very long time (~25 times longer than is expected). So we do
 147  * it once without timing.
 148  */
 149 static void
 150 do_uncached_write(uint8_t *dst, int is_dst_cached,
 151                                   const uint8_t *src, int is_src_cached, size_t size)
 152 {
 153         unsigned i, j;
 154         size_t dst_addrs[TEST_BATCH_SIZE], src_addrs[TEST_BATCH_SIZE];
 155
 156         for (i = 0; i < (TEST_ITERATIONS / TEST_BATCH_SIZE); i++) {
 157                 fill_addr_arrays(dst_addrs, is_dst_cached, 0,
 158                                                  src_addrs, is_src_cached, 0);
 159                 for (j = 0; j < TEST_BATCH_SIZE; j++) {
 160                         rte_memcpy(dst+dst_addrs[j], src+src_addrs[j], size);
 161                 }
 162         }
 163 }
 164
 165 /*
 166  * Run a single memcpy performance test. This is a macro to ensure that if
 167  * the "size" parameter is a constant it won't be converted to a variable.
 168  */
 169 #define SINGLE_PERF_TEST(dst, is_dst_cached, dst_uoffset,                   \
 170                          src, is_src_cached, src_uoffset, size)             \
 171 do {                                                                        \
 172     unsigned int iter, t;                                                   \
 173     size_t dst_addrs[TEST_BATCH_SIZE], src_addrs[TEST_BATCH_SIZE];          \
 174     uint64_t start_time, total_time = 0;                                    \
 175     uint64_t total_time2 = 0;                                               \
 176     for (iter = 0; iter < (TEST_ITERATIONS / TEST_BATCH_SIZE); iter++) {    \
 177         fill_addr_arrays(dst_addrs, is_dst_cached, dst_uoffset,             \
 178                          src_addrs, is_src_cached, src_uoffset);            \
 179         start_time = rte_rdtsc();                                           \
 180         for (t = 0; t < TEST_BATCH_SIZE; t++)                               \
 181             rte_memcpy(dst+dst_addrs[t], src+src_addrs[t], size);           \
 182         total_time += rte_rdtsc() - start_time;                             \
 183     }                                                                       \
 184     for (iter = 0; iter < (TEST_ITERATIONS / TEST_BATCH_SIZE); iter++) {    \
 185         fill_addr_arrays(dst_addrs, is_dst_cached, dst_uoffset,             \
 186                          src_addrs, is_src_cached, src_uoffset);            \
 187         start_time = rte_rdtsc();                                           \
 188         for (t = 0; t < TEST_BATCH_SIZE; t++)                               \
 189             memcpy(dst+dst_addrs[t], src+src_addrs[t], size);               \
 190         total_time2 += rte_rdtsc() - start_time;                            \
 191     }                                                                       \
 192     printf("%3.0f -", (double)total_time  / TEST_ITERATIONS);                 \
 193     printf("%3.0f",   (double)total_time2 / TEST_ITERATIONS);                 \
 194     printf("(%6.2f%%) ", ((double)total_time - total_time2)*100/total_time2); \
 195 } while (0)
 196
 197 /* Run aligned memcpy tests for each cached/uncached permutation */
 198 #define ALL_PERF_TESTS_FOR_SIZE(n)                                       \
 199 do {                                                                     \
 200     if (__builtin_constant_p(n))                                         \
 201         printf("\nC%6u", (unsigned)n);                                   \
 202     else                                                                 \
 203         printf("\n%7u", (unsigned)n);                                    \
 204     SINGLE_PERF_TEST(small_buf_write, 1, 0, small_buf_read, 1, 0, n);    \
 205     SINGLE_PERF_TEST(large_buf_write, 0, 0, small_buf_read, 1, 0, n);    \
 206     SINGLE_PERF_TEST(small_buf_write, 1, 0, large_buf_read, 0, 0, n);    \
 207     SINGLE_PERF_TEST(large_buf_write, 0, 0, large_buf_read, 0, 0, n);    \
 208 } while (0)
 209
 210 /* Run unaligned memcpy tests for each cached/uncached permutation */
 211 #define ALL_PERF_TESTS_FOR_SIZE_UNALIGNED(n)                             \
 212 do {                                                                     \
 213     if (__builtin_constant_p(n))                                         \
 214         printf("\nC%6u", (unsigned)n);                                   \
 215     else                                                                 \
 216         printf("\n%7u", (unsigned)n);                                    \
 217     SINGLE_PERF_TEST(small_buf_write, 1, 1, small_buf_read, 1, 5, n);    \
 218     SINGLE_PERF_TEST(large_buf_write, 0, 1, small_buf_read, 1, 5, n);    \
 219     SINGLE_PERF_TEST(small_buf_write, 1, 1, large_buf_read, 0, 5, n);    \
 220     SINGLE_PERF_TEST(large_buf_write, 0, 1, large_buf_read, 0, 5, n);    \
 221 } while (0)
 222
 223 /* Run memcpy tests for constant length */
 224 #define ALL_PERF_TEST_FOR_CONSTANT                                      \
 225 do {                                                                    \
 226     TEST_CONSTANT(6U); TEST_CONSTANT(64U); TEST_CONSTANT(128U);         \
 227     TEST_CONSTANT(192U); TEST_CONSTANT(256U); TEST_CONSTANT(512U);      \
 228     TEST_CONSTANT(768U); TEST_CONSTANT(1024U); TEST_CONSTANT(1536U);    \
 229 } while (0)
 230
 231 /* Run all memcpy tests for aligned constant cases */
 232 static inline void
 233 perf_test_constant_aligned(void)
 234 {
 235 #define TEST_CONSTANT ALL_PERF_TESTS_FOR_SIZE
 236         ALL_PERF_TEST_FOR_CONSTANT;
 237 #undef TEST_CONSTANT
 238 }
 239
 240 /* Run all memcpy tests for unaligned constant cases */
 241 static inline void
 242 perf_test_constant_unaligned(void)
 243 {
 244 #define TEST_CONSTANT ALL_PERF_TESTS_FOR_SIZE_UNALIGNED
 245         ALL_PERF_TEST_FOR_CONSTANT;
 246 #undef TEST_CONSTANT
 247 }
 248
 249 /* Run all memcpy tests for aligned variable cases */
 250 static inline void
 251 perf_test_variable_aligned(void)
 252 {
 253         unsigned i;
 254         for (i = 0; i < RTE_DIM(buf_sizes); i++) {
 255                 ALL_PERF_TESTS_FOR_SIZE((size_t)buf_sizes[i]);
 256         }
 257 }
 258
 259 /* Run all memcpy tests for unaligned variable cases */
 260 static inline void
 261 perf_test_variable_unaligned(void)
 262 {
 263         unsigned i;
 264         for (i = 0; i < RTE_DIM(buf_sizes); i++) {
 265                 ALL_PERF_TESTS_FOR_SIZE_UNALIGNED((size_t)buf_sizes[i]);
 266         }
 267 }
 268
 269 /* Run all memcpy tests */
 270 static int
 271 perf_test(void)
 272 {
 273         int ret;
 274         struct timeval tv_begin, tv_end;
 275         double time_aligned, time_unaligned;
 276         double time_aligned_const, time_unaligned_const;
 277
 278         ret = init_buffers();
 279         if (ret != 0)
 280                 return ret;
 281
 282 #if TEST_VALUE_RANGE != 0
 283         /* Set up buf_sizes array, if required */
 284         unsigned i;
 285         for (i = 0; i < TEST_VALUE_RANGE; i++)
 286                 buf_sizes[i] = i;
 287 #endif
 288
 289         /* See function comment */
 290         do_uncached_write(large_buf_write, 0, small_buf_read, 1, SMALL_BUFFER_SIZE);
 291
 292         printf("\n** rte_memcpy() - memcpy perf. tests (C = compile-time constant) **\n"
 293                    "======= ================= ================= ================= =================\n"
 294                    "   Size   Cache to cache     Cache to mem      Mem to cache        Mem to mem\n"
 295                    "(bytes)          (ticks)          (ticks)           (ticks)           (ticks)\n"
 296                    "------- ----------------- ----------------- ----------------- -----------------");
 297
 298         printf("\n================================= %2dB aligned =================================",
 299                 ALIGNMENT_UNIT);
 300         /* Do aligned tests where size is a variable */
 301         gettimeofday(&tv_begin, NULL);
 302         perf_test_variable_aligned();
 303         gettimeofday(&tv_end, NULL);
 304         time_aligned = (double)(tv_end.tv_sec - tv_begin.tv_sec)
 305                 + ((double)tv_end.tv_usec - tv_begin.tv_usec)/1000000;
 306         printf("\n------- ----------------- ----------------- ----------------- -----------------");
 307         /* Do aligned tests where size is a compile-time constant */
 308         gettimeofday(&tv_begin, NULL);
 309         perf_test_constant_aligned();
 310         gettimeofday(&tv_end, NULL);
 311         time_aligned_const = (double)(tv_end.tv_sec - tv_begin.tv_sec)
 312                 + ((double)tv_end.tv_usec - tv_begin.tv_usec)/1000000;
 313         printf("\n================================== Unaligned ==================================");
 314         /* Do unaligned tests where size is a variable */
 315         gettimeofday(&tv_begin, NULL);
 316         perf_test_variable_unaligned();
 317         gettimeofday(&tv_end, NULL);
 318         time_unaligned = (double)(tv_end.tv_sec - tv_begin.tv_sec)
 319                 + ((double)tv_end.tv_usec - tv_begin.tv_usec)/1000000;
 320         printf("\n------- ----------------- ----------------- ----------------- -----------------");
 321         /* Do unaligned tests where size is a compile-time constant */
 322         gettimeofday(&tv_begin, NULL);
 323         perf_test_constant_unaligned();
 324         gettimeofday(&tv_end, NULL);
 325         time_unaligned_const = (double)(tv_end.tv_sec - tv_begin.tv_sec)
 326                 + ((double)tv_end.tv_usec - tv_begin.tv_usec)/1000000;
 327         printf("\n======= ================= ================= ================= =================\n\n");
 328
 329         printf("Test Execution Time (seconds):\n");
 330         printf("Aligned variable copy size   = %8.3f\n", time_aligned);
 331         printf("Aligned constant copy size   = %8.3f\n", time_aligned_const);
 332         printf("Unaligned variable copy size = %8.3f\n", time_unaligned);
 333         printf("Unaligned constant copy size = %8.3f\n", time_unaligned_const);
 334         free_buffers();
 335
 336         return 0;
 337 }
 338
 339 static int
 340 test_memcpy_perf(void)
 341 {
 342         int ret;
 343
 344         ret = perf_test();
 345         if (ret != 0)
 346                 return -1;
 347         return 0;
 348 }
 349
 350 REGISTER_TEST_COMMAND(memcpy_perf_autotest, test_memcpy_perf);