test/test/test_memcpy_perf.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <stdint.h>
  35 #include <stdio.h>
  36 #include <string.h>
  37 #include <stdlib.h>
  38
  39 #include <rte_common.h>
  40 #include <rte_cycles.h>
  41 #include <rte_random.h>
  42 #include <rte_malloc.h>
  43
  44 #include <rte_memcpy.h>
  45 #include <rte_cpuflags.h>
  46
  47 #include "test.h"
  48
  49 /*
  50  * Set this to the maximum buffer size you want to test. If it is 0, then the
  51  * values in the buf_sizes[] array below will be used.
  52  */
  53 #define TEST_VALUE_RANGE        0
  54
  55 /* List of buffer sizes to test */
  56 #if TEST_VALUE_RANGE == 0
  57 static size_t buf_sizes[] = {
  58         1, 2, 3, 4, 5, 6, 7, 8, 9, 12, 15, 16, 17, 31, 32, 33, 63, 64, 65, 127, 128,
  59         129, 191, 192, 193, 255, 256, 257, 319, 320, 321, 383, 384, 385, 447, 448,
  60         449, 511, 512, 513, 767, 768, 769, 1023, 1024, 1025, 1518, 1522, 1536, 1600,
  61         2048, 2560, 3072, 3584, 4096, 4608, 5120, 5632, 6144, 6656, 7168, 7680, 8192
  62 };
  63 /* MUST be as large as largest packet size above */
  64 #define SMALL_BUFFER_SIZE       8192
  65 #else /* TEST_VALUE_RANGE != 0 */
  66 static size_t buf_sizes[TEST_VALUE_RANGE];
  67 #define SMALL_BUFFER_SIZE       TEST_VALUE_RANGE
  68 #endif /* TEST_VALUE_RANGE == 0 */
  69
  70
  71 /*
  72  * Arrays of this size are used for measuring uncached memory accesses by
  73  * picking a random location within the buffer. Make this smaller if there are
  74  * memory allocation errors.
  75  */
  76 #define LARGE_BUFFER_SIZE       (100 * 1024 * 1024)
  77
  78 /* How many times to run timing loop for performance tests */
  79 #define TEST_ITERATIONS         1000000
  80 #define TEST_BATCH_SIZE         100
  81
  82 /* Data is aligned on this many bytes (power of 2) */
  83 static uint8_t alignment_unit = 16;
  84
  85 /*
  86  * Pointers used in performance tests. The two large buffers are for uncached
  87  * access where random addresses within the buffer are used for each
  88  * memcpy. The two small buffers are for cached access.
  89  */
  90 static uint8_t *large_buf_read, *large_buf_write;
  91 static uint8_t *small_buf_read, *small_buf_write;
  92
  93 /* Initialise alignment_unit based on machine at run-time. */
  94 static void
  95 init_alignment_unit(void)
  96 {
  97 #ifdef CC_SUPPORT_AVX512
  98         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F)) {
  99                 alignment_unit = 64;
 100                 return;
 101         }
 102 #endif
 103 #ifdef CC_SUPPORT_AVX2
 104         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2)) {
 105                 alignment_unit = 32;
 106                 return;
 107         }
 108 #endif
 109         alignment_unit = 16;
 110 }
 111
 112 /* Initialise data buffers. */
 113 static int
 114 init_buffers(void)
 115 {
 116         unsigned i;
 117
 118         init_alignment_unit();
 119
 120         large_buf_read = rte_malloc("memcpy",
 121                                     LARGE_BUFFER_SIZE + alignment_unit,
 122                                     alignment_unit);
 123         if (large_buf_read == NULL)
 124                 goto error_large_buf_read;
 125
 126         large_buf_write = rte_malloc("memcpy",
 127                                      LARGE_BUFFER_SIZE + alignment_unit,
 128                                      alignment_unit);
 129         if (large_buf_write == NULL)
 130                 goto error_large_buf_write;
 131
 132         small_buf_read = rte_malloc("memcpy",
 133                                     SMALL_BUFFER_SIZE + alignment_unit,
 134                                     alignment_unit);
 135         if (small_buf_read == NULL)
 136                 goto error_small_buf_read;
 137
 138         small_buf_write = rte_malloc("memcpy",
 139                                      SMALL_BUFFER_SIZE + alignment_unit,
 140                                      alignment_unit);
 141         if (small_buf_write == NULL)
 142                 goto error_small_buf_write;
 143
 144         for (i = 0; i < LARGE_BUFFER_SIZE; i++)
 145                 large_buf_read[i] = rte_rand();
 146         for (i = 0; i < SMALL_BUFFER_SIZE; i++)
 147                 small_buf_read[i] = rte_rand();
 148
 149         return 0;
 150
 151 error_small_buf_write:
 152         rte_free(small_buf_read);
 153 error_small_buf_read:
 154         rte_free(large_buf_write);
 155 error_large_buf_write:
 156         rte_free(large_buf_read);
 157 error_large_buf_read:
 158         printf("ERROR: not enough memory\n");
 159         return -1;
 160 }
 161
 162 /* Cleanup data buffers */
 163 static void
 164 free_buffers(void)
 165 {
 166         rte_free(large_buf_read);
 167         rte_free(large_buf_write);
 168         rte_free(small_buf_read);
 169         rte_free(small_buf_write);
 170 }
 171
 172 /*
 173  * Get a random offset into large array, with enough space needed to perform
 174  * max copy size. Offset is aligned, uoffset is used for unalignment setting.
 175  */
 176 static inline size_t
 177 get_rand_offset(size_t uoffset)
 178 {
 179         return ((rte_rand() % (LARGE_BUFFER_SIZE - SMALL_BUFFER_SIZE)) &
 180                         ~(alignment_unit - 1)) + uoffset;
 181 }
 182
 183 /* Fill in source and destination addresses. */
 184 static inline void
 185 fill_addr_arrays(size_t *dst_addr, int is_dst_cached, size_t dst_uoffset,
 186                                  size_t *src_addr, int is_src_cached, size_t src_uoffset)
 187 {
 188         unsigned int i;
 189
 190         for (i = 0; i < TEST_BATCH_SIZE; i++) {
 191                 dst_addr[i] = (is_dst_cached) ? dst_uoffset : get_rand_offset(dst_uoffset);
 192                 src_addr[i] = (is_src_cached) ? src_uoffset : get_rand_offset(src_uoffset);
 193         }
 194 }
 195
 196 /*
 197  * WORKAROUND: For some reason the first test doing an uncached write
 198  * takes a very long time (~25 times longer than is expected). So we do
 199  * it once without timing.
 200  */
 201 static void
 202 do_uncached_write(uint8_t *dst, int is_dst_cached,
 203                                   const uint8_t *src, int is_src_cached, size_t size)
 204 {
 205         unsigned i, j;
 206         size_t dst_addrs[TEST_BATCH_SIZE], src_addrs[TEST_BATCH_SIZE];
 207
 208         for (i = 0; i < (TEST_ITERATIONS / TEST_BATCH_SIZE); i++) {
 209                 fill_addr_arrays(dst_addrs, is_dst_cached, 0,
 210                                                  src_addrs, is_src_cached, 0);
 211                 for (j = 0; j < TEST_BATCH_SIZE; j++) {
 212                         rte_memcpy(dst+dst_addrs[j], src+src_addrs[j], size);
 213                 }
 214         }
 215 }
 216
 217 /*
 218  * Run a single memcpy performance test. This is a macro to ensure that if
 219  * the "size" parameter is a constant it won't be converted to a variable.
 220  */
 221 #define SINGLE_PERF_TEST(dst, is_dst_cached, dst_uoffset,                   \
 222                          src, is_src_cached, src_uoffset, size)             \
 223 do {                                                                        \
 224     unsigned int iter, t;                                                   \
 225     size_t dst_addrs[TEST_BATCH_SIZE], src_addrs[TEST_BATCH_SIZE];          \
 226     uint64_t start_time, total_time = 0;                                    \
 227     uint64_t total_time2 = 0;                                               \
 228     for (iter = 0; iter < (TEST_ITERATIONS / TEST_BATCH_SIZE); iter++) {    \
 229         fill_addr_arrays(dst_addrs, is_dst_cached, dst_uoffset,             \
 230                          src_addrs, is_src_cached, src_uoffset);            \
 231         start_time = rte_rdtsc();                                           \
 232         for (t = 0; t < TEST_BATCH_SIZE; t++)                               \
 233             rte_memcpy(dst+dst_addrs[t], src+src_addrs[t], size);           \
 234         total_time += rte_rdtsc() - start_time;                             \
 235     }                                                                       \
 236     for (iter = 0; iter < (TEST_ITERATIONS / TEST_BATCH_SIZE); iter++) {    \
 237         fill_addr_arrays(dst_addrs, is_dst_cached, dst_uoffset,             \
 238                          src_addrs, is_src_cached, src_uoffset);            \
 239         start_time = rte_rdtsc();                                           \
 240         for (t = 0; t < TEST_BATCH_SIZE; t++)                               \
 241             memcpy(dst+dst_addrs[t], src+src_addrs[t], size);               \
 242         total_time2 += rte_rdtsc() - start_time;                            \
 243     }                                                                       \
 244     printf("%8.0f -",  (double)total_time /TEST_ITERATIONS);                \
 245     printf("%5.0f",  (double)total_time2 / TEST_ITERATIONS);                \
 246 } while (0)
 247
 248 /* Run aligned memcpy tests for each cached/uncached permutation */
 249 #define ALL_PERF_TESTS_FOR_SIZE(n)                                       \
 250 do {                                                                     \
 251     if (__builtin_constant_p(n))                                         \
 252         printf("\nC%6u", (unsigned)n);                                   \
 253     else                                                                 \
 254         printf("\n%7u", (unsigned)n);                                    \
 255     SINGLE_PERF_TEST(small_buf_write, 1, 0, small_buf_read, 1, 0, n);    \
 256     SINGLE_PERF_TEST(large_buf_write, 0, 0, small_buf_read, 1, 0, n);    \
 257     SINGLE_PERF_TEST(small_buf_write, 1, 0, large_buf_read, 0, 0, n);    \
 258     SINGLE_PERF_TEST(large_buf_write, 0, 0, large_buf_read, 0, 0, n);    \
 259 } while (0)
 260
 261 /* Run unaligned memcpy tests for each cached/uncached permutation */
 262 #define ALL_PERF_TESTS_FOR_SIZE_UNALIGNED(n)                             \
 263 do {                                                                     \
 264     if (__builtin_constant_p(n))                                         \
 265         printf("\nC%6u", (unsigned)n);                                   \
 266     else                                                                 \
 267         printf("\n%7u", (unsigned)n);                                    \
 268     SINGLE_PERF_TEST(small_buf_write, 1, 1, small_buf_read, 1, 5, n);    \
 269     SINGLE_PERF_TEST(large_buf_write, 0, 1, small_buf_read, 1, 5, n);    \
 270     SINGLE_PERF_TEST(small_buf_write, 1, 1, large_buf_read, 0, 5, n);    \
 271     SINGLE_PERF_TEST(large_buf_write, 0, 1, large_buf_read, 0, 5, n);    \
 272 } while (0)
 273
 274 /* Run memcpy tests for constant length */
 275 #define ALL_PERF_TEST_FOR_CONSTANT                                      \
 276 do {                                                                    \
 277     TEST_CONSTANT(6U); TEST_CONSTANT(64U); TEST_CONSTANT(128U);         \
 278     TEST_CONSTANT(192U); TEST_CONSTANT(256U); TEST_CONSTANT(512U);      \
 279     TEST_CONSTANT(768U); TEST_CONSTANT(1024U); TEST_CONSTANT(1536U);    \
 280 } while (0)
 281
 282 /* Run all memcpy tests for aligned constant cases */
 283 static inline void
 284 perf_test_constant_aligned(void)
 285 {
 286 #define TEST_CONSTANT ALL_PERF_TESTS_FOR_SIZE
 287         ALL_PERF_TEST_FOR_CONSTANT;
 288 #undef TEST_CONSTANT
 289 }
 290
 291 /* Run all memcpy tests for unaligned constant cases */
 292 static inline void
 293 perf_test_constant_unaligned(void)
 294 {
 295 #define TEST_CONSTANT ALL_PERF_TESTS_FOR_SIZE_UNALIGNED
 296         ALL_PERF_TEST_FOR_CONSTANT;
 297 #undef TEST_CONSTANT
 298 }
 299
 300 /* Run all memcpy tests for aligned variable cases */
 301 static inline void
 302 perf_test_variable_aligned(void)
 303 {
 304         unsigned n = sizeof(buf_sizes) / sizeof(buf_sizes[0]);
 305         unsigned i;
 306         for (i = 0; i < n; i++) {
 307                 ALL_PERF_TESTS_FOR_SIZE((size_t)buf_sizes[i]);
 308         }
 309 }
 310
 311 /* Run all memcpy tests for unaligned variable cases */
 312 static inline void
 313 perf_test_variable_unaligned(void)
 314 {
 315         unsigned n = sizeof(buf_sizes) / sizeof(buf_sizes[0]);
 316         unsigned i;
 317         for (i = 0; i < n; i++) {
 318                 ALL_PERF_TESTS_FOR_SIZE_UNALIGNED((size_t)buf_sizes[i]);
 319         }
 320 }
 321
 322 /* Run all memcpy tests */
 323 static int
 324 perf_test(void)
 325 {
 326         int ret;
 327
 328         ret = init_buffers();
 329         if (ret != 0)
 330                 return ret;
 331
 332 #if TEST_VALUE_RANGE != 0
 333         /* Set up buf_sizes array, if required */
 334         unsigned i;
 335         for (i = 0; i < TEST_VALUE_RANGE; i++)
 336                 buf_sizes[i] = i;
 337 #endif
 338
 339         /* See function comment */
 340         do_uncached_write(large_buf_write, 0, small_buf_read, 1, SMALL_BUFFER_SIZE);
 341
 342         printf("\n** rte_memcpy() - memcpy perf. tests (C = compile-time constant) **\n"
 343                    "======= ============== ============== ============== ==============\n"
 344                    "   Size Cache to cache   Cache to mem   Mem to cache     Mem to mem\n"
 345                    "(bytes)        (ticks)        (ticks)        (ticks)        (ticks)\n"
 346                    "------- -------------- -------------- -------------- --------------");
 347
 348         printf("\n========================= %2dB aligned ============================",
 349                 alignment_unit);
 350         /* Do aligned tests where size is a variable */
 351         perf_test_variable_aligned();
 352         printf("\n------- -------------- -------------- -------------- --------------");
 353         /* Do aligned tests where size is a compile-time constant */
 354         perf_test_constant_aligned();
 355         printf("\n=========================== Unaligned =============================");
 356         /* Do unaligned tests where size is a variable */
 357         perf_test_variable_unaligned();
 358         printf("\n------- -------------- -------------- -------------- --------------");
 359         /* Do unaligned tests where size is a compile-time constant */
 360         perf_test_constant_unaligned();
 361         printf("\n======= ============== ============== ============== ==============\n\n");
 362
 363         free_buffers();
 364
 365         return 0;
 366 }
 367
 368 static int
 369 test_memcpy_perf(void)
 370 {
 371         int ret;
 372
 373         ret = perf_test();
 374         if (ret != 0)
 375                 return -1;
 376         return 0;
 377 }
 378
 379 REGISTER_TEST_COMMAND(memcpy_perf_autotest, test_memcpy_perf);