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 <time.h> 10 11 #include <rte_common.h> 12 #include <rte_cycles.h> 13 #include <rte_os_shim.h> 14 #include <rte_random.h> 15 #include <rte_malloc.h> 16 17 #include <rte_memcpy.h> 18 19 #include "test.h" 20 21 /* 22 * Set this to the maximum buffer size you want to test. If it is 0, then the 23 * values in the buf_sizes[] array below will be used. 24 */ 25 #define TEST_VALUE_RANGE 0 26 27 /* List of buffer sizes to test */ 28 #if TEST_VALUE_RANGE == 0 29 static size_t buf_sizes[] = { 30 1, 2, 3, 4, 5, 6, 7, 8, 9, 12, 15, 16, 17, 31, 32, 33, 63, 64, 65, 127, 128, 31 129, 191, 192, 193, 255, 256, 257, 319, 320, 321, 383, 384, 385, 447, 448, 32 449, 511, 512, 513, 767, 768, 769, 1023, 1024, 1025, 1518, 1522, 1536, 1600, 33 2048, 2560, 3072, 3584, 4096, 4608, 5120, 5632, 6144, 6656, 7168, 7680, 8192 34 }; 35 /* MUST be as large as largest packet size above */ 36 #define SMALL_BUFFER_SIZE 8192 37 #else /* TEST_VALUE_RANGE != 0 */ 38 static size_t buf_sizes[TEST_VALUE_RANGE]; 39 #define SMALL_BUFFER_SIZE TEST_VALUE_RANGE 40 #endif /* TEST_VALUE_RANGE == 0 */ 41 42 43 /* 44 * Arrays of this size are used for measuring uncached memory accesses by 45 * picking a random location within the buffer. Make this smaller if there are 46 * memory allocation errors. 47 */ 48 #define LARGE_BUFFER_SIZE (100 * 1024 * 1024) 49 50 /* How many times to run timing loop for performance tests */ 51 #define TEST_ITERATIONS 1000000 52 #define TEST_BATCH_SIZE 100 53 54 /* Data is aligned on this many bytes (power of 2) */ 55 #ifdef __AVX512F__ 56 #define ALIGNMENT_UNIT 64 57 #elif defined __AVX2__ 58 #define ALIGNMENT_UNIT 32 59 #else 60 #define ALIGNMENT_UNIT 16 61 #endif 62 63 /* 64 * Pointers used in performance tests. The two large buffers are for uncached 65 * access where random addresses within the buffer are used for each 66 * memcpy. The two small buffers are for cached access. 67 */ 68 static uint8_t *large_buf_read, *large_buf_write; 69 static uint8_t *small_buf_read, *small_buf_write; 70 71 /* Initialise data buffers. */ 72 static int 73 init_buffers(void) 74 { 75 unsigned i; 76 77 large_buf_read = rte_malloc("memcpy", LARGE_BUFFER_SIZE + ALIGNMENT_UNIT, ALIGNMENT_UNIT); 78 if (large_buf_read == NULL) 79 goto error_large_buf_read; 80 81 large_buf_write = rte_malloc("memcpy", LARGE_BUFFER_SIZE + ALIGNMENT_UNIT, ALIGNMENT_UNIT); 82 if (large_buf_write == NULL) 83 goto error_large_buf_write; 84 85 small_buf_read = rte_malloc("memcpy", SMALL_BUFFER_SIZE + ALIGNMENT_UNIT, ALIGNMENT_UNIT); 86 if (small_buf_read == NULL) 87 goto error_small_buf_read; 88 89 small_buf_write = rte_malloc("memcpy", SMALL_BUFFER_SIZE + ALIGNMENT_UNIT, ALIGNMENT_UNIT); 90 if (small_buf_write == NULL) 91 goto error_small_buf_write; 92 93 for (i = 0; i < LARGE_BUFFER_SIZE; i++) 94 large_buf_read[i] = rte_rand(); 95 for (i = 0; i < SMALL_BUFFER_SIZE; i++) 96 small_buf_read[i] = rte_rand(); 97 98 return 0; 99 100 error_small_buf_write: 101 rte_free(small_buf_read); 102 error_small_buf_read: 103 rte_free(large_buf_write); 104 error_large_buf_write: 105 rte_free(large_buf_read); 106 error_large_buf_read: 107 printf("ERROR: not enough memory\n"); 108 return -1; 109 } 110 111 /* Cleanup data buffers */ 112 static void 113 free_buffers(void) 114 { 115 rte_free(large_buf_read); 116 rte_free(large_buf_write); 117 rte_free(small_buf_read); 118 rte_free(small_buf_write); 119 } 120 121 /* 122 * Get a random offset into large array, with enough space needed to perform 123 * max copy size. Offset is aligned, uoffset is used for unalignment setting. 124 */ 125 static inline size_t 126 get_rand_offset(size_t uoffset) 127 { 128 return ((rte_rand() % (LARGE_BUFFER_SIZE - SMALL_BUFFER_SIZE)) & 129 ~(ALIGNMENT_UNIT - 1)) + uoffset; 130 } 131 132 /* Fill in source and destination addresses. */ 133 static inline void 134 fill_addr_arrays(size_t *dst_addr, int is_dst_cached, size_t dst_uoffset, 135 size_t *src_addr, int is_src_cached, size_t src_uoffset) 136 { 137 unsigned int i; 138 139 for (i = 0; i < TEST_BATCH_SIZE; i++) { 140 dst_addr[i] = (is_dst_cached) ? dst_uoffset : get_rand_offset(dst_uoffset); 141 src_addr[i] = (is_src_cached) ? src_uoffset : get_rand_offset(src_uoffset); 142 } 143 } 144 145 /* 146 * WORKAROUND: For some reason the first test doing an uncached write 147 * takes a very long time (~25 times longer than is expected). So we do 148 * it once without timing. 149 */ 150 static void 151 do_uncached_write(uint8_t *dst, int is_dst_cached, 152 const uint8_t *src, int is_src_cached, size_t size) 153 { 154 unsigned i, j; 155 size_t dst_addrs[TEST_BATCH_SIZE], src_addrs[TEST_BATCH_SIZE]; 156 157 for (i = 0; i < (TEST_ITERATIONS / TEST_BATCH_SIZE); i++) { 158 fill_addr_arrays(dst_addrs, is_dst_cached, 0, 159 src_addrs, is_src_cached, 0); 160 for (j = 0; j < TEST_BATCH_SIZE; j++) { 161 rte_memcpy(dst+dst_addrs[j], src+src_addrs[j], size); 162 } 163 } 164 } 165 166 /* 167 * Run a single memcpy performance test. This is a macro to ensure that if 168 * the "size" parameter is a constant it won't be converted to a variable. 169 */ 170 #define SINGLE_PERF_TEST(dst, is_dst_cached, dst_uoffset, \ 171 src, is_src_cached, src_uoffset, size) \ 172 do { \ 173 unsigned int iter, t; \ 174 size_t dst_addrs[TEST_BATCH_SIZE], src_addrs[TEST_BATCH_SIZE]; \ 175 uint64_t start_time, total_time = 0; \ 176 uint64_t total_time2 = 0; \ 177 for (iter = 0; iter < (TEST_ITERATIONS / TEST_BATCH_SIZE); iter++) { \ 178 fill_addr_arrays(dst_addrs, is_dst_cached, dst_uoffset, \ 179 src_addrs, is_src_cached, src_uoffset); \ 180 start_time = rte_rdtsc(); \ 181 for (t = 0; t < TEST_BATCH_SIZE; t++) \ 182 rte_memcpy(dst+dst_addrs[t], src+src_addrs[t], size); \ 183 total_time += rte_rdtsc() - start_time; \ 184 } \ 185 for (iter = 0; iter < (TEST_ITERATIONS / TEST_BATCH_SIZE); iter++) { \ 186 fill_addr_arrays(dst_addrs, is_dst_cached, dst_uoffset, \ 187 src_addrs, is_src_cached, src_uoffset); \ 188 start_time = rte_rdtsc(); \ 189 for (t = 0; t < TEST_BATCH_SIZE; t++) \ 190 memcpy(dst+dst_addrs[t], src+src_addrs[t], size); \ 191 total_time2 += rte_rdtsc() - start_time; \ 192 } \ 193 printf("%3.0f -", (double)total_time / TEST_ITERATIONS); \ 194 printf("%3.0f", (double)total_time2 / TEST_ITERATIONS); \ 195 printf("(%6.2f%%) ", ((double)total_time - total_time2)*100/total_time2); \ 196 } while (0) 197 198 /* Run aligned memcpy tests for each cached/uncached permutation */ 199 #define ALL_PERF_TESTS_FOR_SIZE(n) \ 200 do { \ 201 if (__builtin_constant_p(n)) \ 202 printf("\nC%6u", (unsigned)n); \ 203 else \ 204 printf("\n%7u", (unsigned)n); \ 205 SINGLE_PERF_TEST(small_buf_write, 1, 0, small_buf_read, 1, 0, n); \ 206 SINGLE_PERF_TEST(large_buf_write, 0, 0, small_buf_read, 1, 0, n); \ 207 SINGLE_PERF_TEST(small_buf_write, 1, 0, large_buf_read, 0, 0, n); \ 208 SINGLE_PERF_TEST(large_buf_write, 0, 0, large_buf_read, 0, 0, n); \ 209 } while (0) 210 211 /* Run unaligned memcpy tests for each cached/uncached permutation */ 212 #define ALL_PERF_TESTS_FOR_SIZE_UNALIGNED(n) \ 213 do { \ 214 if (__builtin_constant_p(n)) \ 215 printf("\nC%6u", (unsigned)n); \ 216 else \ 217 printf("\n%7u", (unsigned)n); \ 218 SINGLE_PERF_TEST(small_buf_write, 1, 1, small_buf_read, 1, 5, n); \ 219 SINGLE_PERF_TEST(large_buf_write, 0, 1, small_buf_read, 1, 5, n); \ 220 SINGLE_PERF_TEST(small_buf_write, 1, 1, large_buf_read, 0, 5, n); \ 221 SINGLE_PERF_TEST(large_buf_write, 0, 1, large_buf_read, 0, 5, n); \ 222 } while (0) 223 224 /* Run memcpy tests for constant length */ 225 #define ALL_PERF_TEST_FOR_CONSTANT \ 226 do { \ 227 TEST_CONSTANT(6U); TEST_CONSTANT(64U); TEST_CONSTANT(128U); \ 228 TEST_CONSTANT(192U); TEST_CONSTANT(256U); TEST_CONSTANT(512U); \ 229 TEST_CONSTANT(768U); TEST_CONSTANT(1024U); TEST_CONSTANT(1536U); \ 230 } while (0) 231 232 /* Run all memcpy tests for aligned constant cases */ 233 static inline void 234 perf_test_constant_aligned(void) 235 { 236 #define TEST_CONSTANT ALL_PERF_TESTS_FOR_SIZE 237 ALL_PERF_TEST_FOR_CONSTANT; 238 #undef TEST_CONSTANT 239 } 240 241 /* Run all memcpy tests for unaligned constant cases */ 242 static inline void 243 perf_test_constant_unaligned(void) 244 { 245 #define TEST_CONSTANT ALL_PERF_TESTS_FOR_SIZE_UNALIGNED 246 ALL_PERF_TEST_FOR_CONSTANT; 247 #undef TEST_CONSTANT 248 } 249 250 /* Run all memcpy tests for aligned variable cases */ 251 static inline void 252 perf_test_variable_aligned(void) 253 { 254 unsigned i; 255 for (i = 0; i < RTE_DIM(buf_sizes); i++) { 256 ALL_PERF_TESTS_FOR_SIZE((size_t)buf_sizes[i]); 257 } 258 } 259 260 /* Run all memcpy tests for unaligned variable cases */ 261 static inline void 262 perf_test_variable_unaligned(void) 263 { 264 unsigned i; 265 for (i = 0; i < RTE_DIM(buf_sizes); i++) { 266 ALL_PERF_TESTS_FOR_SIZE_UNALIGNED((size_t)buf_sizes[i]); 267 } 268 } 269 270 /* Run all memcpy tests */ 271 static int 272 perf_test(void) 273 { 274 int ret; 275 struct timespec tv_begin, tv_end; 276 double time_aligned, time_unaligned; 277 double time_aligned_const, time_unaligned_const; 278 279 ret = init_buffers(); 280 if (ret != 0) 281 return ret; 282 283 #if TEST_VALUE_RANGE != 0 284 /* Set up buf_sizes array, if required */ 285 unsigned i; 286 for (i = 0; i < TEST_VALUE_RANGE; i++) 287 buf_sizes[i] = i; 288 #endif 289 290 /* See function comment */ 291 do_uncached_write(large_buf_write, 0, small_buf_read, 1, SMALL_BUFFER_SIZE); 292 293 printf("\n** rte_memcpy() - memcpy perf. tests (C = compile-time constant) **\n" 294 "======= ================= ================= ================= =================\n" 295 " Size Cache to cache Cache to mem Mem to cache Mem to mem\n" 296 "(bytes) (ticks) (ticks) (ticks) (ticks)\n" 297 "------- ----------------- ----------------- ----------------- -----------------"); 298 299 printf("\n================================= %2dB aligned =================================", 300 ALIGNMENT_UNIT); 301 /* Do aligned tests where size is a variable */ 302 timespec_get(&tv_begin, TIME_UTC); 303 perf_test_variable_aligned(); 304 timespec_get(&tv_end, TIME_UTC); 305 time_aligned = (double)(tv_end.tv_sec - tv_begin.tv_sec) 306 + ((double)tv_end.tv_nsec - tv_begin.tv_nsec) / NS_PER_S; 307 printf("\n------- ----------------- ----------------- ----------------- -----------------"); 308 /* Do aligned tests where size is a compile-time constant */ 309 timespec_get(&tv_begin, TIME_UTC); 310 perf_test_constant_aligned(); 311 timespec_get(&tv_end, TIME_UTC); 312 time_aligned_const = (double)(tv_end.tv_sec - tv_begin.tv_sec) 313 + ((double)tv_end.tv_nsec - tv_begin.tv_nsec) / NS_PER_S; 314 printf("\n================================== Unaligned =================================="); 315 /* Do unaligned tests where size is a variable */ 316 timespec_get(&tv_begin, TIME_UTC); 317 perf_test_variable_unaligned(); 318 timespec_get(&tv_end, TIME_UTC); 319 time_unaligned = (double)(tv_end.tv_sec - tv_begin.tv_sec) 320 + ((double)tv_end.tv_nsec - tv_begin.tv_nsec) / NS_PER_S; 321 printf("\n------- ----------------- ----------------- ----------------- -----------------"); 322 /* Do unaligned tests where size is a compile-time constant */ 323 timespec_get(&tv_begin, TIME_UTC); 324 perf_test_constant_unaligned(); 325 timespec_get(&tv_end, TIME_UTC); 326 time_unaligned_const = (double)(tv_end.tv_sec - tv_begin.tv_sec) 327 + ((double)tv_end.tv_nsec - tv_begin.tv_nsec) / NS_PER_S; 328 printf("\n======= ================= ================= ================= =================\n\n"); 329 330 printf("Test Execution Time (seconds):\n"); 331 printf("Aligned variable copy size = %8.3f\n", time_aligned); 332 printf("Aligned constant copy size = %8.3f\n", time_aligned_const); 333 printf("Unaligned variable copy size = %8.3f\n", time_unaligned); 334 printf("Unaligned constant copy size = %8.3f\n", time_unaligned_const); 335 free_buffers(); 336 337 return 0; 338 } 339 340 static int 341 test_memcpy_perf(void) 342 { 343 int ret; 344 345 ret = perf_test(); 346 if (ret != 0) 347 return -1; 348 return 0; 349 } 350 351 REGISTER_PERF_TEST(memcpy_perf_autotest, test_memcpy_perf); 352