examples/nettle-benchmark.c

   1 /* nettle-benchmark.c
   2  *
   3  * Tries the performance of the various algorithms.
   4  *
   5  */
   6
   7 /* nettle, low-level cryptographics library
   8  *
   9  * Copyright (C) 2001, 2010 Niels Möller
  10  *
  11  * The nettle library is free software; you can redistribute it and/or modify
  12  * it under the terms of the GNU Lesser General Public License as published by
  13  * the Free Software Foundation; either version 2.1 of the License, or (at your
  14  * option) any later version.
  15  *
  16  * The nettle library is distributed in the hope that it will be useful, but
  17  * WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
  18  * or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public
  19  * License for more details.
  20  *
  21  * You should have received a copy of the GNU Lesser General Public License
  22  * along with the nettle library; see the file COPYING.LIB.  If not, write to
  23  * the Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston,
  24  * MA 02111-1301, USA.
  25  */
  26
  27 #if HAVE_CONFIG_H
  28 # include "config.h"
  29 #endif
  30
  31 #include <assert.h>
  32 #include <errno.h>
  33 #include <math.h>
  34 #include <stdarg.h>
  35 #include <stdio.h>
  36 #include <stdlib.h>
  37 #include <string.h>
  38
  39 #include <time.h>
  40
  41 #include "timing.h"
  42
  43 #include "aes.h"
  44 #include "arcfour.h"
  45 #include "blowfish.h"
  46 #include "cast128.h"
  47 #include "cbc.h"
  48 #include "ctr.h"
  49 #include "des.h"
  50 #include "gcm.h"
  51 #include "memxor.h"
  52 #include "salsa20.h"
  53 #include "serpent.h"
  54 #include "sha1.h"
  55 #include "sha2.h"
  56 #include "sha3.h"
  57 #include "twofish.h"
  58 #include "umac.h"
  59
  60 #include "nettle-meta.h"
  61 #include "nettle-internal.h"
  62
  63 #include "getopt.h"
  64
  65 static double frequency = 0.0;
  66
  67 /* Process BENCH_BLOCK bytes at a time, for BENCH_INTERVAL seconds. */
  68 #define BENCH_BLOCK 10240
  69 #define BENCH_INTERVAL 0.1
  70
  71 /* FIXME: Proper configure test for rdtsc? */
  72 #ifndef WITH_CYCLE_COUNTER
  73 # if defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__))
  74 #  define WITH_CYCLE_COUNTER 1
  75 # else
  76 #  define WITH_CYCLE_COUNTER 0
  77 # endif
  78 #endif
  79
  80 #if WITH_CYCLE_COUNTER
  81 # if defined(__i386__)
  82 #define GET_CYCLE_COUNTER(hi, lo)               \
  83   __asm__("xorl %%eax,%%eax\n"                  \
  84           "movl %%ebx, %%edi\n"                 \
  85           "cpuid\n"                             \
  86           "rdtsc\n"                             \
  87           "movl %%edi, %%ebx\n"                 \
  88           : "=a" (lo), "=d" (hi)                \
  89           : /* No inputs. */                    \
  90           : "%edi", "%ecx", "cc")
  91 # elif defined(__x86_64__)
  92 #define GET_CYCLE_COUNTER(hi, lo)               \
  93   __asm__("xorl %%eax,%%eax\n"                  \
  94           "mov %%rbx, %%r10\n"                  \
  95           "cpuid\n"                             \
  96           "rdtsc\n"                             \
  97           "mov %%r10, %%rbx\n"                  \
  98           : "=a" (lo), "=d" (hi)                \
  99           : /* No inputs. */                    \
 100           : "%r10", "%rcx", "cc")
 101 # endif
 102 #define BENCH_ITERATIONS 10
 103 #endif
 104
 105 static void NORETURN PRINTF_STYLE(1,2)
 106 die(const char *format, ...)
 107 {
 108   va_list args;
 109   va_start(args, format);
 110   vfprintf(stderr, format, args);
 111   va_end(args);
 112
 113   exit(EXIT_FAILURE);
 114 }
 115
 116 static double overhead = 0.0;
 117
 118 /* Returns second per function call */
 119 static double
 120 time_function(void (*f)(void *arg), void *arg)
 121 {
 122   unsigned ncalls;
 123   double elapsed;
 124
 125   for (ncalls = 10 ;;)
 126     {
 127       unsigned i;
 128
 129       time_start();
 130       for (i = 0; i < ncalls; i++)
 131         f(arg);
 132       elapsed = time_end();
 133       if (elapsed > BENCH_INTERVAL)
 134         break;
 135       else if (elapsed < BENCH_INTERVAL / 10)
 136         ncalls *= 10;
 137       else
 138         ncalls *= 2;
 139     }
 140   return elapsed / ncalls - overhead;
 141 }
 142
 143 static void
 144 bench_nothing(void *arg UNUSED)
 145 {
 146   return;
 147 }
 148
 149 struct bench_memxor_info
 150 {
 151   uint8_t *dst;
 152   const uint8_t *src;
 153   const uint8_t *other;
 154 };
 155
 156 static void
 157 bench_memxor(void *arg)
 158 {
 159   struct bench_memxor_info *info = arg;
 160   memxor (info->dst, info->src, BENCH_BLOCK);
 161 }
 162
 163 static void
 164 bench_memxor3(void *arg)
 165 {
 166   struct bench_memxor_info *info = arg;
 167   memxor3 (info->dst, info->src, info->other, BENCH_BLOCK);
 168 }
 169
 170 struct bench_hash_info
 171 {
 172   void *ctx;
 173   nettle_hash_update_func *update;
 174   const uint8_t *data;
 175 };
 176
 177 static void
 178 bench_hash(void *arg)
 179 {
 180   struct bench_hash_info *info = arg;
 181   info->update(info->ctx, BENCH_BLOCK, info->data);
 182 }
 183
 184 struct bench_cipher_info
 185 {
 186   void *ctx;
 187   nettle_crypt_func *crypt;
 188   uint8_t *data;
 189 };
 190
 191 static void
 192 bench_cipher(void *arg)
 193 {
 194   struct bench_cipher_info *info = arg;
 195   info->crypt(info->ctx, BENCH_BLOCK, info->data, info->data);
 196 }
 197
 198 struct bench_cbc_info
 199 {
 200   void *ctx;
 201   nettle_crypt_func *crypt;
 202
 203   uint8_t *data;
 204
 205   unsigned block_size;
 206   uint8_t *iv;
 207 };
 208
 209 static void
 210 bench_cbc_encrypt(void *arg)
 211 {
 212   struct bench_cbc_info *info = arg;
 213   cbc_encrypt(info->ctx, info->crypt,
 214               info->block_size, info->iv,
 215               BENCH_BLOCK, info->data, info->data);
 216 }
 217
 218 static void
 219 bench_cbc_decrypt(void *arg)
 220 {
 221   struct bench_cbc_info *info = arg;
 222   cbc_decrypt(info->ctx, info->crypt,
 223               info->block_size, info->iv,
 224               BENCH_BLOCK, info->data, info->data);
 225 }
 226
 227 static void
 228 bench_ctr(void *arg)
 229 {
 230   struct bench_cbc_info *info = arg;
 231   ctr_crypt(info->ctx, info->crypt,
 232             info->block_size, info->iv,
 233             BENCH_BLOCK, info->data, info->data);
 234 }
 235
 236 /* Set data[i] = floor(sqrt(i)) */
 237 static void
 238 init_data(uint8_t *data)
 239 {
 240   unsigned i,j;
 241   for (i = j = 0; i<BENCH_BLOCK;  i++)
 242     {
 243       if (j*j < i)
 244         j++;
 245       data[i] = j;
 246     }
 247 }
 248
 249 static void
 250 init_key(unsigned length,
 251          uint8_t *key)
 252 {
 253   unsigned i;
 254   for (i = 0; i<length; i++)
 255     key[i] = i;
 256 }
 257
 258 static void
 259 header(void)
 260 {
 261   printf("%18s %11s Mbyte/s%s\n",
 262          "Algorithm", "mode",
 263          frequency > 0.0 ? " cycles/byte cycles/block" : "");
 264 }
 265
 266 static void
 267 display(const char *name, const char *mode, unsigned block_size,
 268         double time)
 269 {
 270   printf("%18s %11s %7.2f",
 271          name, mode,
 272          BENCH_BLOCK / (time * 1048576.0));
 273   if (frequency > 0.0)
 274     {
 275       printf(" %11.2f", time * frequency / BENCH_BLOCK);
 276       if (block_size > 0)
 277         printf(" %12.2f", time * frequency * block_size / BENCH_BLOCK);
 278     }
 279   printf("\n");
 280 }
 281
 282 static void *
 283 xalloc(size_t size)
 284 {
 285   void *p = malloc(size);
 286   if (!p)
 287     die("Virtual memory exhausted.\n");
 288
 289   return p;
 290 }
 291
 292 static void
 293 time_overhead(void)
 294 {
 295   overhead = time_function(bench_nothing, NULL);
 296   printf("benchmark call overhead: %7f us", overhead * 1e6);
 297   if (frequency > 0.0)
 298     printf("%7.2f cycles\n", overhead * frequency);
 299   printf("\n");
 300 }
 301
 302
 303
 304 static void
 305 time_memxor(void)
 306 {
 307   struct bench_memxor_info info;
 308   uint8_t src[BENCH_BLOCK + sizeof(long)];
 309   uint8_t other[BENCH_BLOCK + sizeof(long)];
 310   uint8_t dst[BENCH_BLOCK];
 311
 312   info.src = src;
 313   info.dst = dst;
 314
 315   display ("memxor", "aligned", sizeof(unsigned long),
 316            time_function(bench_memxor, &info));
 317   info.src = src + 1;
 318   display ("memxor", "unaligned", sizeof(unsigned long),
 319            time_function(bench_memxor, &info));
 320
 321   info.src = src;
 322   info.other = other;
 323   display ("memxor3", "aligned", sizeof(unsigned long),
 324            time_function(bench_memxor3, &info));
 325
 326   info.other = other + 1;
 327   display ("memxor3", "unaligned01", sizeof(unsigned long),
 328            time_function(bench_memxor3, &info));
 329   info.src = src + 1;
 330   display ("memxor3", "unaligned11", sizeof(unsigned long),
 331            time_function(bench_memxor3, &info));
 332   info.other = other + 2;
 333   display ("memxor3", "unaligned12", sizeof(unsigned long),
 334            time_function(bench_memxor3, &info));
 335 }
 336
 337 static void
 338 time_hash(const struct nettle_hash *hash)
 339 {
 340   static uint8_t data[BENCH_BLOCK];
 341   struct bench_hash_info info;
 342
 343   info.ctx = xalloc(hash->context_size);
 344   info.update = hash->update;
 345   info.data = data;
 346
 347   init_data(data);
 348   hash->init(info.ctx);
 349
 350   display(hash->name, "update", hash->block_size,
 351           time_function(bench_hash, &info));
 352
 353   free(info.ctx);
 354 }
 355
 356 static void
 357 time_umac(void)
 358 {
 359   static uint8_t data[BENCH_BLOCK];
 360   struct bench_hash_info info;
 361   struct umac32_ctx ctx32;
 362   struct umac64_ctx ctx64;
 363   struct umac96_ctx ctx96;
 364   struct umac128_ctx ctx128;
 365
 366   uint8_t key[16];
 367
 368   umac32_set_key (&ctx32, key);
 369   info.ctx = &ctx32;
 370   info.update = (nettle_hash_update_func *) umac32_update;
 371   info.data = data;
 372
 373   display("umac32", "update", UMAC_DATA_SIZE,
 374           time_function(bench_hash, &info));
 375
 376   umac64_set_key (&ctx64, key);
 377   info.ctx = &ctx64;
 378   info.update = (nettle_hash_update_func *) umac64_update;
 379   info.data = data;
 380
 381   display("umac64", "update", UMAC_DATA_SIZE,
 382           time_function(bench_hash, &info));
 383
 384   umac96_set_key (&ctx96, key);
 385   info.ctx = &ctx96;
 386   info.update = (nettle_hash_update_func *) umac96_update;
 387   info.data = data;
 388
 389   display("umac96", "update", UMAC_DATA_SIZE,
 390           time_function(bench_hash, &info));
 391
 392   umac128_set_key (&ctx128, key);
 393   info.ctx = &ctx128;
 394   info.update = (nettle_hash_update_func *) umac128_update;
 395   info.data = data;
 396
 397   display("umac128", "update", UMAC_DATA_SIZE,
 398           time_function(bench_hash, &info));
 399 }
 400
 401 static void
 402 time_gcm(void)
 403 {
 404   static uint8_t data[BENCH_BLOCK];
 405   struct bench_hash_info hinfo;
 406   struct bench_cipher_info cinfo;
 407   struct gcm_aes_ctx ctx;
 408
 409   uint8_t key[16];
 410   uint8_t iv[GCM_IV_SIZE];
 411
 412   gcm_aes_set_key(&ctx, sizeof(key), key);
 413   gcm_aes_set_iv(&ctx, sizeof(iv), iv);
 414
 415   hinfo.ctx = &ctx;
 416   hinfo.update = (nettle_hash_update_func *) gcm_aes_update;
 417   hinfo.data = data;
 418
 419   display("gcm-aes", "update", GCM_BLOCK_SIZE,
 420           time_function(bench_hash, &hinfo));
 421
 422   cinfo.ctx = &ctx;
 423   cinfo.crypt = (nettle_crypt_func *) gcm_aes_encrypt;
 424   cinfo.data = data;
 425
 426   display("gcm-aes", "encrypt", GCM_BLOCK_SIZE,
 427           time_function(bench_cipher, &cinfo));
 428
 429   cinfo.crypt = (nettle_crypt_func *) gcm_aes_decrypt;
 430
 431   display("gcm-aes", "decrypt", GCM_BLOCK_SIZE,
 432           time_function(bench_cipher, &cinfo));
 433 }
 434
 435 static int
 436 prefix_p(const char *prefix, const char *s)
 437 {
 438   size_t i;
 439   for (i = 0; prefix[i]; i++)
 440     if (prefix[i] != s[i])
 441       return 0;
 442   return 1;
 443 }
 444
 445 static int
 446 block_cipher_p(const struct nettle_cipher *cipher)
 447 {
 448   /* Don't use nettle cbc and ctr for openssl ciphers. */
 449   return cipher->block_size > 0 && !prefix_p("openssl", cipher->name);
 450 }
 451
 452 static void
 453 time_cipher(const struct nettle_cipher *cipher)
 454 {
 455   void *ctx = xalloc(cipher->context_size);
 456   uint8_t *key = xalloc(cipher->key_size);
 457
 458   static uint8_t data[BENCH_BLOCK];
 459
 460   printf("\n");
 461
 462   init_data(data);
 463
 464   {
 465     /* Decent initializers are a GNU extension, so don't use it here. */
 466     struct bench_cipher_info info;
 467     info.ctx = ctx;
 468     info.crypt = cipher->encrypt;
 469     info.data = data;
 470
 471     init_key(cipher->key_size, key);
 472     cipher->set_encrypt_key(ctx, cipher->key_size, key);
 473
 474     display(cipher->name, "ECB encrypt", cipher->block_size,
 475             time_function(bench_cipher, &info));
 476   }
 477
 478   {
 479     struct bench_cipher_info info;
 480     info.ctx = ctx;
 481     info.crypt = cipher->decrypt;
 482     info.data = data;
 483
 484     init_key(cipher->key_size, key);
 485     cipher->set_decrypt_key(ctx, cipher->key_size, key);
 486
 487     display(cipher->name, "ECB decrypt", cipher->block_size,
 488             time_function(bench_cipher, &info));
 489   }
 490
 491   if (block_cipher_p(cipher))
 492     {
 493       uint8_t *iv = xalloc(cipher->block_size);
 494
 495       /* Do CBC mode */
 496       {
 497         struct bench_cbc_info info;
 498         info.ctx = ctx;
 499         info.crypt = cipher->encrypt;
 500         info.data = data;
 501         info.block_size = cipher->block_size;
 502         info.iv = iv;
 503
 504         memset(iv, 0, sizeof(iv));
 505
 506         cipher->set_encrypt_key(ctx, cipher->key_size, key);
 507
 508         display(cipher->name, "CBC encrypt", cipher->block_size,
 509                 time_function(bench_cbc_encrypt, &info));
 510       }
 511
 512       {
 513         struct bench_cbc_info info;
 514         info.ctx = ctx;
 515         info.crypt = cipher->decrypt;
 516         info.data = data;
 517         info.block_size = cipher->block_size;
 518         info.iv = iv;
 519
 520         memset(iv, 0, sizeof(iv));
 521
 522         cipher->set_decrypt_key(ctx, cipher->key_size, key);
 523
 524         display(cipher->name, "CBC decrypt", cipher->block_size,
 525                 time_function(bench_cbc_decrypt, &info));
 526       }
 527
 528       /* Do CTR mode */
 529       {
 530         struct bench_cbc_info info;
 531         info.ctx = ctx;
 532         info.crypt = cipher->encrypt;
 533         info.data = data;
 534         info.block_size = cipher->block_size;
 535         info.iv = iv;
 536
 537         memset(iv, 0, sizeof(iv));
 538
 539         cipher->set_encrypt_key(ctx, cipher->key_size, key);
 540
 541         display(cipher->name, "CTR", cipher->block_size,
 542                 time_function(bench_ctr, &info));
 543       }
 544
 545       free(iv);
 546     }
 547   free(ctx);
 548   free(key);
 549 }
 550
 551 /* Try to get accurate cycle times for assembler functions. */
 552 #if WITH_CYCLE_COUNTER
 553 static int
 554 compare_double(const void *ap, const void *bp)
 555 {
 556   double a = *(const double *) ap;
 557   double b = *(const double *) bp;
 558   if (a < b)
 559     return -1;
 560   else if (a > b)
 561     return 1;
 562   else
 563     return 0;
 564 }
 565
 566 #define TIME_CYCLES(t, code) do {                               \
 567   double tc_count[5];                                           \
 568   uint32_t tc_start_lo, tc_start_hi, tc_end_lo, tc_end_hi;      \
 569   unsigned tc_i, tc_j;                                          \
 570   for (tc_j = 0; tc_j < 5; tc_j++)                              \
 571     {                                                           \
 572       tc_i = 0;                                                 \
 573       GET_CYCLE_COUNTER(tc_start_hi, tc_start_lo);              \
 574       for (; tc_i < BENCH_ITERATIONS; tc_i++)                   \
 575         { code; }                                               \
 576                                                                 \
 577       GET_CYCLE_COUNTER(tc_end_hi, tc_end_lo);                  \
 578                                                                 \
 579       tc_end_hi -= (tc_start_hi + (tc_start_lo > tc_end_lo));   \
 580       tc_end_lo -= tc_start_lo;                                 \
 581                                                                 \
 582       tc_count[tc_j] = ldexp(tc_end_hi, 32) + tc_end_lo;        \
 583     }                                                           \
 584   qsort(tc_count, 5, sizeof(double), compare_double);           \
 585   (t) = tc_count[2] / BENCH_ITERATIONS;                         \
 586 } while (0)
 587
 588 static void
 589 bench_sha1_compress(void)
 590 {
 591   uint32_t state[_SHA1_DIGEST_LENGTH];
 592   uint8_t data[SHA1_DATA_SIZE];
 593   double t;
 594
 595   TIME_CYCLES (t, _nettle_sha1_compress(state, data));
 596
 597   printf("sha1_compress: %.2f cycles\n", t);
 598 }
 599
 600 static void
 601 bench_salsa20_core(void)
 602 {
 603   uint32_t state[_SALSA20_INPUT_LENGTH];
 604   double t;
 605
 606   TIME_CYCLES (t, _nettle_salsa20_core(state, state, 20));
 607   printf("salsa20_core: %.2f cycles\n", t);
 608 }
 609
 610 static void
 611 bench_sha3_permute(void)
 612 {
 613   struct sha3_state state;
 614   double t;
 615
 616   TIME_CYCLES (t, sha3_permute (&state));
 617   printf("sha3_permute: %.2f cycles (%.2f / round)\n", t, t / 24.0);
 618 }
 619 #else
 620 #define bench_sha1_compress()
 621 #define bench_salsa20_core()
 622 #define bench_sha3_permute()
 623 #endif
 624
 625 #if WITH_OPENSSL
 626 # define OPENSSL(x) x,
 627 #else
 628 # define OPENSSL(x)
 629 #endif
 630
 631 int
 632 main(int argc, char **argv)
 633 {
 634   unsigned i;
 635   int c;
 636   const char *alg;
 637
 638   const struct nettle_hash *hashes[] =
 639     {
 640       &nettle_md2, &nettle_md4, &nettle_md5,
 641       OPENSSL(&nettle_openssl_md5)
 642       &nettle_sha1, OPENSSL(&nettle_openssl_sha1)
 643       &nettle_sha224, &nettle_sha256,
 644       &nettle_sha384, &nettle_sha512,
 645       &nettle_sha3_224, &nettle_sha3_256,
 646       &nettle_sha3_384, &nettle_sha3_512,
 647       &nettle_ripemd160, &nettle_gosthash94,
 648       NULL
 649     };
 650
 651   const struct nettle_cipher *ciphers[] =
 652     {
 653       &nettle_aes128, &nettle_aes192, &nettle_aes256,
 654       OPENSSL(&nettle_openssl_aes128)
 655       OPENSSL(&nettle_openssl_aes192)
 656       OPENSSL(&nettle_openssl_aes256)
 657       &nettle_arcfour128, OPENSSL(&nettle_openssl_arcfour128)
 658       &nettle_blowfish128, OPENSSL(&nettle_openssl_blowfish128)
 659       &nettle_camellia128, &nettle_camellia192, &nettle_camellia256,
 660       &nettle_cast128, OPENSSL(&nettle_openssl_cast128)
 661       &nettle_des, OPENSSL(&nettle_openssl_des)
 662       &nettle_des3,
 663       &nettle_serpent256,
 664       &nettle_twofish128, &nettle_twofish192, &nettle_twofish256,
 665       &nettle_salsa20, &nettle_salsa20r12,
 666       NULL
 667     };
 668
 669   enum { OPT_HELP = 300 };
 670   static const struct option options[] =
 671     {
 672       /* Name, args, flag, val */
 673       { "help", no_argument, NULL, OPT_HELP },
 674       { "clock-frequency", required_argument, NULL, 'f' },
 675       { NULL, 0, NULL, 0 }
 676     };
 677
 678   while ( (c = getopt_long(argc, argv, "f:", options, NULL)) != -1)
 679     switch (c)
 680       {
 681       case 'f':
 682         frequency = atof(optarg);
 683         if (frequency > 0.0)
 684           break;
 685
 686       case OPT_HELP:
 687         printf("Usage: nettle-benchmark [-f clock frequency] [alg]\n");
 688         return EXIT_SUCCESS;
 689
 690       case '?':
 691         return EXIT_FAILURE;
 692
 693       default:
 694         abort();
 695     }
 696
 697   alg = argv[optind];
 698
 699   time_init();
 700   bench_sha1_compress();
 701   bench_salsa20_core();
 702   bench_sha3_permute();
 703   printf("\n");
 704   time_overhead();
 705
 706   header();
 707
 708   if (!alg || strstr ("memxor", alg))
 709     {
 710       time_memxor();
 711       printf("\n");
 712     }
 713
 714   for (i = 0; hashes[i]; i++)
 715     if (!alg || strstr(hashes[i]->name, alg))
 716       time_hash(hashes[i]);
 717
 718   if (!alg || strstr ("umac", alg))
 719     time_umac();
 720
 721   for (i = 0; ciphers[i]; i++)
 722     if (!alg || strstr(ciphers[i]->name, alg))
 723       time_cipher(ciphers[i]);
 724
 725   if (!alg || strstr ("gcm", alg))
 726     {
 727       printf("\n");
 728       time_gcm();
 729     }
 730
 731   return 0;
 732 }