src/gmxlib/gmx_cpuid.c

   1 /* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
   2  *
   3  *
   4  * This file is part of GROMACS.
   5  * Copyright (c) 2012-
   6  *
   7  * Written by the Gromacs development team under coordination of
   8  * David van der Spoel, Berk Hess, and Erik Lindahl.
   9  *
  10  * This library is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public License
  12  * as published by the Free Software Foundation; either version 2
  13  * of the License, or (at your option) any later version.
  14  *
  15  * To help us fund GROMACS development, we humbly ask that you cite
  16  * the research papers on the package. Check out http://www.gromacs.org
  17  *
  18  * And Hey:
  19  * Gnomes, ROck Monsters And Chili Sauce
  20  */
  21 #ifdef HAVE_CONFIG_H
  22 #include <config.h>
  23 #endif
  24
  25 #ifdef HAVE_SCHED_H
  26 #define _GNU_SOURCE
  27 #include <sched.h>
  28 #endif
  29
  30 #include <stdio.h>
  31 #include <stdlib.h>
  32 #include <string.h>
  33 #include <ctype.h>
  34 #ifdef _MSC_VER
  35 /* MSVC definition for __cpuid() */
  36 #include <intrin.h>
  37 #endif
  38 #ifdef HAVE_UNISTD_H
  39 /* sysconf() definition */
  40 #include <unistd.h>
  41 #endif
  42
  43
  44
  45
  46 #include "gmx_cpuid.h"
  47
  48
  49 /* Global constant character strings corresponding to our enumerated types */
  50 const char *
  51 gmx_cpuid_vendor_string[GMX_CPUID_NVENDORS] =
  52 {
  53     "CannotDetect",
  54     "Unknown",
  55     "GenuineIntel",
  56     "AuthenticAMD"
  57 };
  58
  59 const char *
  60 gmx_cpuid_feature_string[GMX_CPUID_NFEATURES] =
  61 {
  62     "CannotDetect",
  63     "aes",
  64     "apic",
  65     "avx",
  66     "avx2",
  67     "clfsh",
  68     "cmov",
  69     "cx8",
  70     "cx16",
  71     "f16c",
  72     "fma",
  73     "fma4",
  74     "htt",
  75     "lahf_lm",
  76     "misalignsse",
  77     "mmx",
  78     "msr",
  79     "nonstop_tsc",
  80     "pcid",
  81     "pclmuldq",
  82     "pdcm",
  83     "pdpe1gb",
  84     "popcnt",
  85     "pse",
  86     "rdrnd",
  87     "rdtscp",
  88     "sse2",
  89     "sse3",
  90     "sse4a",
  91     "sse4.1",
  92     "sse4.2",
  93     "ssse3",
  94     "tdt",
  95     "x2apic",
  96     "xop"
  97 };
  98
  99 const char *
 100 gmx_cpuid_acceleration_string[GMX_CPUID_NACCELERATIONS] =
 101 {
 102     "CannotDetect",
 103     "None",
 104     "SSE2",
 105     "SSE4.1",
 106     "AVX_128_FMA",
 107     "AVX_256"
 108 };
 109
 110 /* Max length of brand string */
 111 #define GMX_CPUID_BRAND_MAXLEN 256
 112
 113
 114 /* Contents of the abstract datatype */
 115 struct gmx_cpuid
 116 {
 117     enum gmx_cpuid_vendor      vendor;
 118     char                       brand[GMX_CPUID_BRAND_MAXLEN];
 119     int                        family;
 120     int                        model;
 121     int                        stepping;
 122     /* Not using gmx_bool here, since this file must be possible to compile without simple.h */
 123     char                       feature[GMX_CPUID_NFEATURES];
 124 };
 125
 126
 127 /* Simple routines to access the data structure. The initialization routine is
 128  * further down since that needs to call other static routines in this file.
 129  */
 130 enum gmx_cpuid_vendor
 131 gmx_cpuid_vendor            (gmx_cpuid_t                cpuid)
 132 {
 133     return cpuid->vendor;
 134 }
 135
 136
 137 const char *
 138 gmx_cpuid_brand             (gmx_cpuid_t                cpuid)
 139 {
 140     return cpuid->brand;
 141 }
 142
 143 int
 144 gmx_cpuid_family            (gmx_cpuid_t                cpuid)
 145 {
 146     return cpuid->family;
 147 }
 148
 149 int
 150 gmx_cpuid_model             (gmx_cpuid_t                cpuid)
 151 {
 152     return cpuid->model;
 153 }
 154
 155 int
 156 gmx_cpuid_stepping          (gmx_cpuid_t                cpuid)
 157 {
 158     return cpuid->stepping;
 159 }
 160
 161 int
 162 gmx_cpuid_feature           (gmx_cpuid_t                cpuid,
 163                              enum gmx_cpuid_feature     feature)
 164 {
 165     return (cpuid->feature[feature]!=0);
 166 }
 167
 168
 169
 170
 171 /* What type of acceleration was compiled in, if any?
 172  * This is set from Cmake. Note that the SSE2 and SSE4_1 macros are set for
 173  * AVX too, so it is important that they appear last in the list.
 174  */
 175 #ifdef GMX_X86_AVX_256
 176 static const
 177 enum gmx_cpuid_acceleration
 178 compiled_acc = GMX_CPUID_ACCELERATION_X86_AVX_256;
 179 #elif defined GMX_X86_AVX_128_FMA
 180 static const
 181 enum gmx_cpuid_acceleration
 182 compiled_acc = GMX_CPUID_ACCELERATION_X86_AVX_128_FMA;
 183 #elif defined GMX_X86_SSE4_1
 184 static const
 185 enum gmx_cpuid_acceleration
 186 compiled_acc = GMX_CPUID_ACCELERATION_X86_SSE4_1;
 187 #elif defined GMX_X86_SSE2
 188 static const
 189 enum gmx_cpuid_acceleration
 190 compiled_acc = GMX_CPUID_ACCELERATION_X86_SSE2;
 191 #else
 192 static const
 193 enum gmx_cpuid_acceleration
 194 compiled_acc = GMX_CPUID_ACCELERATION_NONE;
 195 #endif
 196
 197
 198 /* Currently CPUID is only supported (1) if we can use an instruction on MSVC, or (2)
 199  * if the compiler handles GNU-style inline assembly.
 200  */
 201 #if defined (__i386__) || defined (__x86_64__) || defined (_M_IX86) || defined (_M_X64)
 202
 203 /* Execute CPUID on x86 class CPUs. level sets function to exec, and the
 204  * contents of register output is returned. See Intel/AMD docs for details.
 205  *
 206  * This version supports extended information where we can also have an input
 207  * value in the ecx register. This is ignored for most levels, but some of them
 208  * (e.g. level 0xB on Intel) use it.
 209  */
 210 static int
 211 execute_x86cpuid(unsigned int   level,
 212                  unsigned int   ecxval,
 213                  unsigned int * eax,
 214                  unsigned int * ebx,
 215                  unsigned int * ecx,
 216                  unsigned int * edx)
 217 {
 218     int rc = 0;
 219
 220 #if (defined _MSC_VER)
 221     int CPUInfo[4];
 222
 223 #if (_MSC_VER > 1500) || (_MSC_VER==1500 & _MSC_FULL_VER >= 150030729)
 224     /* MSVC 9.0 SP1 or later */
 225     __cpuidex(CPUInfo,level,ecxval);
 226     rc = 0;
 227 #else
 228     __cpuid(CPUInfo,level);
 229     /* Set an error code if the user wanted a non-zero ecxval, since we did not have cpuidex */
 230     rc = (ecxval>0) ? -1 : 0;
 231 #endif
 232     *eax=CPUInfo[0];
 233     *ebx=CPUInfo[1];
 234     *ecx=CPUInfo[2];
 235     *edx=CPUInfo[3];
 236
 237 #elif (defined GMX_X86_GCC_INLINE_ASM)
 238     /* for now this means GMX_X86_GCC_INLINE_ASM should be defined,
 239      * but there might be more options added in the future.
 240      */
 241     *eax = level;
 242     *ecx = ecxval;
 243     *ebx = 0;
 244     *edx = 0;
 245 #if defined(__i386__) && defined(__PIC__)
 246     /* Avoid clobbering the global offset table in 32-bit pic code (ebx register) */
 247     __asm__ __volatile__ ("xchgl %%ebx, %1  \n\t"
 248                           "cpuid            \n\t"
 249                           "xchgl %%ebx, %1  \n\t"
 250                           : "+a"(*eax), "+r"(*ebx), "+c"(*ecx), "+d"(*edx));
 251 #else
 252     /* i386 without PIC, or x86-64. Things are easy and we can clobber any reg we want :-) */
 253     __asm__ __volatile__ ("cpuid            \n\t"
 254                           : "+a"(*eax), "+b"(*ebx), "+c"(*ecx), "+d"(*edx));
 255 #endif
 256     rc = 0;
 257 #else
 258     /* Death and horror!
 259      * Apparently this is an x86 platform where we don't know how to call cpuid.
 260      *
 261      * This is REALLY bad, since we will lose all Gromacs acceleration.
 262      */
 263     *eax = 0;
 264     *ebx = 0;
 265     *ecx = 0;
 266     *edx = 0;
 267
 268     rc = -1;
 269 #endif
 270     return rc;
 271 }
 272 #endif /* architecture is x86 */
 273
 274
 275 /* Identify CPU features common to Intel & AMD - mainly brand string,
 276  * version and some features. Vendor has already been detected outside this.
 277  */
 278 static int
 279 cpuid_check_common_x86(gmx_cpuid_t                cpuid)
 280 {
 281     int                       fn,max_stdfn,max_extfn;
 282     unsigned int              eax,ebx,ecx,edx;
 283     char                      str[GMX_CPUID_BRAND_MAXLEN];
 284     char *                    p;
 285
 286     /* Find largest standard/extended function input value */
 287     execute_x86cpuid(0x0,0,&eax,&ebx,&ecx,&edx);
 288     max_stdfn = eax;
 289     execute_x86cpuid(0x80000000,0,&eax,&ebx,&ecx,&edx);
 290     max_extfn = eax;
 291
 292     p = str;
 293     if(max_extfn>=0x80000005)
 294     {
 295         /* Get CPU brand string */
 296         for(fn=0x80000002;fn<0x80000005;fn++)
 297         {
 298             execute_x86cpuid(fn,0,&eax,&ebx,&ecx,&edx);
 299             memcpy(p,&eax,4);
 300             memcpy(p+4,&ebx,4);
 301             memcpy(p+8,&ecx,4);
 302             memcpy(p+12,&edx,4);
 303             p+=16;
 304         }
 305         *p='\0';
 306
 307         /* Remove empty initial space */
 308         p = str;
 309         while(isspace(*(p)))
 310         {
 311             p++;
 312         }
 313         strncpy(cpuid->brand,p,GMX_CPUID_BRAND_MAXLEN);
 314     }
 315     else
 316     {
 317         strncpy(cpuid->brand,"Unknown CPU brand",GMX_CPUID_BRAND_MAXLEN);
 318     }
 319
 320     /* Find basic CPU properties */
 321     if(max_stdfn>=1)
 322     {
 323         execute_x86cpuid(0x1,0,&eax,&ebx,&ecx,&edx);
 324
 325         cpuid->family   = ((eax & 0x0FF00000) >> 20) + ((eax & 0x00000F00) >> 8);
 326         /* Note that extended model should be shifted left 4, so only shift right 12 iso 16. */
 327         cpuid->model    = ((eax & 0x000F0000) >> 12) + ((eax & 0x000000F0) >> 4);
 328         cpuid->stepping = (eax & 0x0000000F);
 329
 330         /* Feature flags common to AMD and intel */
 331         cpuid->feature[GMX_CPUID_FEATURE_X86_SSE3]     = (ecx & (1 << 0))  != 0;
 332         cpuid->feature[GMX_CPUID_FEATURE_X86_PCLMULDQ] = (ecx & (1 << 1))  != 0;
 333         cpuid->feature[GMX_CPUID_FEATURE_X86_SSSE3]    = (ecx & (1 << 9))  != 0;
 334         cpuid->feature[GMX_CPUID_FEATURE_X86_FMA]      = (ecx & (1 << 12)) != 0;
 335         cpuid->feature[GMX_CPUID_FEATURE_X86_CX16]     = (ecx & (1 << 13)) != 0;
 336         cpuid->feature[GMX_CPUID_FEATURE_X86_SSE4_1]   = (ecx & (1 << 19)) != 0;
 337         cpuid->feature[GMX_CPUID_FEATURE_X86_SSE4_2]   = (ecx & (1 << 20)) != 0;
 338         cpuid->feature[GMX_CPUID_FEATURE_X86_POPCNT]   = (ecx & (1 << 23)) != 0;
 339         cpuid->feature[GMX_CPUID_FEATURE_X86_AES]      = (ecx & (1 << 25)) != 0;
 340         cpuid->feature[GMX_CPUID_FEATURE_X86_AVX]      = (ecx & (1 << 28)) != 0;
 341         cpuid->feature[GMX_CPUID_FEATURE_X86_F16C]     = (ecx & (1 << 29)) != 0;
 342         cpuid->feature[GMX_CPUID_FEATURE_X86_RDRND]    = (ecx & (1 << 30)) != 0;
 343
 344         cpuid->feature[GMX_CPUID_FEATURE_X86_PSE]      = (edx & (1 << 3))  != 0;
 345         cpuid->feature[GMX_CPUID_FEATURE_X86_MSR]      = (edx & (1 << 5))  != 0;
 346         cpuid->feature[GMX_CPUID_FEATURE_X86_CX8]      = (edx & (1 << 8))  != 0;
 347         cpuid->feature[GMX_CPUID_FEATURE_X86_APIC]     = (edx & (1 << 9))  != 0;
 348         cpuid->feature[GMX_CPUID_FEATURE_X86_CMOV]     = (edx & (1 << 15)) != 0;
 349         cpuid->feature[GMX_CPUID_FEATURE_X86_CLFSH]    = (edx & (1 << 19)) != 0;
 350         cpuid->feature[GMX_CPUID_FEATURE_X86_MMX]      = (edx & (1 << 23)) != 0;
 351         cpuid->feature[GMX_CPUID_FEATURE_X86_SSE2]     = (edx & (1 << 26)) != 0;
 352         cpuid->feature[GMX_CPUID_FEATURE_X86_HTT]      = (edx & (1 << 28)) != 0;
 353     }
 354     else
 355     {
 356         cpuid->family   = -1;
 357         cpuid->model    = -1;
 358         cpuid->stepping = -1;
 359     }
 360
 361     if(max_extfn>=0x80000001)
 362     {
 363         execute_x86cpuid(0x80000001,0,&eax,&ebx,&ecx,&edx);
 364         cpuid->feature[GMX_CPUID_FEATURE_X86_LAHF_LM] = (ecx & (1 << 0))  != 0;
 365         cpuid->feature[GMX_CPUID_FEATURE_X86_PDPE1GB] = (edx & (1 << 26)) != 0;
 366         cpuid->feature[GMX_CPUID_FEATURE_X86_RDTSCP]  = (edx & (1 << 27)) != 0;
 367     }
 368
 369     if(max_extfn>=0x80000007)
 370     {
 371         execute_x86cpuid(0x80000007,0,&eax,&ebx,&ecx,&edx);
 372         cpuid->feature[GMX_CPUID_FEATURE_X86_NONSTOP_TSC]  = (edx & (1 << 8))  != 0;
 373     }
 374
 375     return 0;
 376 }
 377
 378 /* Detection of AMD-specific CPU features */
 379 static int
 380 cpuid_check_amd_x86(gmx_cpuid_t                cpuid)
 381 {
 382     int                       max_stdfn,max_extfn;
 383     unsigned int              eax,ebx,ecx,edx;
 384
 385     cpuid_check_common_x86(cpuid);
 386
 387     execute_x86cpuid(0x0,0,&eax,&ebx,&ecx,&edx);
 388     max_stdfn = eax;
 389
 390     execute_x86cpuid(0x80000000,0,&eax,&ebx,&ecx,&edx);
 391     max_extfn = eax;
 392
 393     if(max_extfn>=0x80000001)
 394     {
 395         execute_x86cpuid(0x80000001,0,&eax,&ebx,&ecx,&edx);
 396
 397         cpuid->feature[GMX_CPUID_FEATURE_X86_SSE4A]       = (ecx & (1 << 6))  != 0;
 398         cpuid->feature[GMX_CPUID_FEATURE_X86_MISALIGNSSE] = (ecx & (1 << 7))  != 0;
 399         cpuid->feature[GMX_CPUID_FEATURE_X86_XOP]         = (ecx & (1 << 11)) != 0;
 400         cpuid->feature[GMX_CPUID_FEATURE_X86_FMA4]        = (ecx & (1 << 16)) != 0;
 401     }
 402
 403     return 0;
 404 }
 405
 406 /* Detection of Intel-specific CPU features */
 407 static int
 408 cpuid_check_intel_x86(gmx_cpuid_t                cpuid)
 409 {
 410     unsigned int              max_stdfn,max_extfn;
 411     unsigned int              eax,ebx,ecx,edx;
 412     unsigned int              i;
 413     unsigned int              max_logical_cores,max_physical_cores;
 414
 415     cpuid_check_common_x86(cpuid);
 416
 417     execute_x86cpuid(0x0,0,&eax,&ebx,&ecx,&edx);
 418     max_stdfn = eax;
 419
 420     execute_x86cpuid(0x80000000,0,&eax,&ebx,&ecx,&edx);
 421     max_extfn = eax;
 422
 423     if(max_stdfn>=1)
 424     {
 425         execute_x86cpuid(0x1,0,&eax,&ebx,&ecx,&edx);
 426         cpuid->feature[GMX_CPUID_FEATURE_X86_PDCM]    = (ecx & (1 << 15)) != 0;
 427         cpuid->feature[GMX_CPUID_FEATURE_X86_PCID]    = (ecx & (1 << 17)) != 0;
 428         cpuid->feature[GMX_CPUID_FEATURE_X86_X2APIC]  = (ecx & (1 << 21)) != 0;
 429         cpuid->feature[GMX_CPUID_FEATURE_X86_TDT]     = (ecx & (1 << 24)) != 0;
 430     }
 431
 432     if(max_stdfn>=7)
 433     {
 434         execute_x86cpuid(0x7,0,&eax,&ebx,&ecx,&edx);
 435         cpuid->feature[GMX_CPUID_FEATURE_X86_AVX2]    = (ebx & (1 << 5))  != 0;
 436     }
 437
 438     /* Check whether Hyper-Threading is enabled, not only supported */
 439     if(cpuid->feature[GMX_CPUID_FEATURE_X86_HTT] && max_stdfn>=4)
 440     {
 441         execute_x86cpuid(0x1,0,&eax,&ebx,&ecx,&edx);
 442         max_logical_cores  = (ebx >> 16) & 0x0FF;
 443         execute_x86cpuid(0x4,0,&eax,&ebx,&ecx,&edx);
 444         max_physical_cores = ((eax >> 26) & 0x3F) + 1;
 445
 446         /* Clear HTT flag if we only have 1 logical core per physical */
 447         if(max_logical_cores/max_physical_cores < 2)
 448         {
 449             cpuid->feature[GMX_CPUID_FEATURE_X86_HTT] = 0;
 450         }
 451     }
 452     return 0;
 453 }
 454
 455 /* Try to find the vendor of the current CPU, so we know what specific
 456  * detection routine to call.
 457  */
 458 static enum gmx_cpuid_vendor
 459 cpuid_check_vendor(void)
 460 {
 461     enum gmx_cpuid_vendor      i,vendor;
 462     /* Register data used on x86 */
 463     unsigned int               eax,ebx,ecx,edx;
 464     char                       vendorstring[13];
 465
 466     /* Set default first */
 467     vendor = GMX_CPUID_VENDOR_UNKNOWN;
 468
 469     execute_x86cpuid(0x0,0,&eax,&ebx,&ecx,&edx);
 470
 471     memcpy(vendorstring,&ebx,4);
 472     memcpy(vendorstring+4,&edx,4);
 473     memcpy(vendorstring+8,&ecx,4);
 474
 475     vendorstring[12]='\0';
 476
 477     for(i=GMX_CPUID_VENDOR_UNKNOWN;i<GMX_CPUID_NVENDORS;i++)
 478     {
 479         if(!strncmp(vendorstring,gmx_cpuid_vendor_string[i],12))
 480         {
 481             vendor = i;
 482         }
 483     }
 484
 485     return vendor;
 486 }
 487
 488
 489
 490
 491 int
 492 gmx_cpuid_init               (gmx_cpuid_t *              pcpuid)
 493 {
 494     gmx_cpuid_t cpuid;
 495     int i;
 496
 497     cpuid = malloc(sizeof(*cpuid));
 498
 499     *pcpuid = cpuid;
 500
 501     for(i=0;i<GMX_CPUID_NFEATURES;i++)
 502     {
 503         cpuid->feature[i]=0;
 504     }
 505
 506     cpuid->vendor = cpuid_check_vendor();
 507
 508     switch(cpuid->vendor)
 509     {
 510         case GMX_CPUID_VENDOR_INTEL:
 511             cpuid_check_intel_x86(cpuid);
 512             break;
 513         case GMX_CPUID_VENDOR_AMD:
 514             cpuid_check_amd_x86(cpuid);
 515             break;
 516         default:
 517             /* Could not find vendor */
 518             strncpy(cpuid->brand,"Unknown CPU brand",GMX_CPUID_BRAND_MAXLEN);
 519             cpuid->family         = 0;
 520             cpuid->model          = 0;
 521             cpuid->stepping       = 0;
 522
 523             for(i=0;i<GMX_CPUID_NFEATURES;i++)
 524             {
 525                 cpuid->feature[i]=0;
 526             }
 527             cpuid->feature[GMX_CPUID_FEATURE_CANNOTDETECT] = 1;
 528             break;
 529     }
 530
 531     return 0;
 532 }
 533
 534
 535
 536 void
 537 gmx_cpuid_done               (gmx_cpuid_t              cpuid)
 538 {
 539     free(cpuid);
 540 }
 541
 542
 543 int
 544 gmx_cpuid_formatstring       (gmx_cpuid_t              cpuid,
 545                               char *                   str,
 546                               int                      n)
 547 {
 548     int c;
 549     int i;
 550     enum gmx_cpuid_feature  feature;
 551
 552 #ifdef _MSC_VER
 553     _snprintf(str,n,
 554               "Vendor: %s\n"
 555               "Brand:  %s\n"
 556               "Family: %2d  Model: %2d  Stepping: %2d\n"
 557               "Features:",
 558               gmx_cpuid_vendor_string[gmx_cpuid_vendor(cpuid)],
 559               gmx_cpuid_brand(cpuid),
 560               gmx_cpuid_family(cpuid),gmx_cpuid_model(cpuid),gmx_cpuid_stepping(cpuid));
 561 #else
 562     snprintf(str,n,
 563              "Vendor: %s\n"
 564              "Brand:  %s\n"
 565              "Family: %2d  Model: %2d  Stepping: %2d\n"
 566              "Features:",
 567              gmx_cpuid_vendor_string[gmx_cpuid_vendor(cpuid)],
 568              gmx_cpuid_brand(cpuid),
 569              gmx_cpuid_family(cpuid),gmx_cpuid_model(cpuid),gmx_cpuid_stepping(cpuid));
 570 #endif
 571
 572     str[n-1] = '\0';
 573     c = strlen(str);
 574     n   -= c;
 575     str += c;
 576
 577     for(feature=GMX_CPUID_FEATURE_CANNOTDETECT;feature<GMX_CPUID_NFEATURES;feature++)
 578     {
 579         if(gmx_cpuid_feature(cpuid,feature)==1)
 580         {
 581 #ifdef _MSC_VER
 582             _snprintf(str,n," %s",gmx_cpuid_feature_string[feature]);
 583 #else
 584             snprintf(str,n," %s",gmx_cpuid_feature_string[feature]);
 585 #endif
 586             str[n-1] = '\0';
 587             c = strlen(str);
 588             n   -= c;
 589             str += c;
 590         }
 591     }
 592 #ifdef _MSC_VER
 593     _snprintf(str,n,"\n");
 594 #else
 595     snprintf(str,n,"\n");
 596 #endif
 597     str[n-1] = '\0';
 598
 599     return 0;
 600 }
 601
 602
 603
 604 enum gmx_cpuid_acceleration
 605 gmx_cpuid_acceleration_suggest  (gmx_cpuid_t                 cpuid)
 606 {
 607     enum gmx_cpuid_acceleration  tmpacc;
 608
 609     tmpacc = GMX_CPUID_ACCELERATION_NONE;
 610
 611     if(gmx_cpuid_vendor(cpuid)==GMX_CPUID_VENDOR_INTEL)
 612     {
 613         if(gmx_cpuid_feature(cpuid,GMX_CPUID_FEATURE_X86_AVX))
 614         {
 615             tmpacc = GMX_CPUID_ACCELERATION_X86_AVX_256;
 616         }
 617         else if(gmx_cpuid_feature(cpuid,GMX_CPUID_FEATURE_X86_SSE4_1))
 618         {
 619             tmpacc = GMX_CPUID_ACCELERATION_X86_SSE4_1;
 620         }
 621         else if(gmx_cpuid_feature(cpuid,GMX_CPUID_FEATURE_X86_SSE2))
 622         {
 623             tmpacc = GMX_CPUID_ACCELERATION_X86_SSE2;
 624         }
 625     }
 626     else if(gmx_cpuid_vendor(cpuid)==GMX_CPUID_VENDOR_AMD)
 627     {
 628         if(gmx_cpuid_feature(cpuid,GMX_CPUID_FEATURE_X86_AVX))
 629         {
 630             tmpacc = GMX_CPUID_ACCELERATION_X86_AVX_128_FMA;
 631         }
 632         else if(gmx_cpuid_feature(cpuid,GMX_CPUID_FEATURE_X86_SSE4_1))
 633         {
 634             tmpacc = GMX_CPUID_ACCELERATION_X86_SSE4_1;
 635         }
 636         else if(gmx_cpuid_feature(cpuid,GMX_CPUID_FEATURE_X86_SSE2))
 637         {
 638             tmpacc = GMX_CPUID_ACCELERATION_X86_SSE2;
 639         }
 640     }
 641
 642     return tmpacc;
 643 }
 644
 645
 646
 647 int
 648 gmx_cpuid_acceleration_check(gmx_cpuid_t   cpuid,
 649                              FILE *        log)
 650 {
 651     int                           rc;
 652     char                          str[1024];
 653     enum gmx_cpuid_acceleration   acc;
 654
 655     acc = gmx_cpuid_acceleration_suggest(cpuid);
 656
 657     rc = (acc != compiled_acc);
 658
 659     gmx_cpuid_formatstring(cpuid,str,1023);
 660     str[1023] = '\0';
 661
 662     if(log!=NULL)
 663     {
 664         fprintf(log,
 665                 "\nDetecting CPU-specific acceleration.\nPresent hardware specification:\n"
 666                 "%s"
 667                 "Acceleration most likely to fit this hardware: %s\n"
 668                 "Acceleration selected at GROMACS compile time: %s\n\n",
 669                 str,
 670                 gmx_cpuid_acceleration_string[acc],
 671                 gmx_cpuid_acceleration_string[compiled_acc]);
 672     }
 673
 674     if(rc!=0)
 675     {
 676         if(log!=NULL)
 677         {
 678             fprintf(log,"WARNING! Binary not matching hardware - you are likely losing performance.\n\n");
 679         }
 680         printf("\nWARNING! Binary not matching hardware - you are likely losing performance.\n"
 681                "Acceleration most likely to fit this hardware: %s\n"
 682                "Acceleration selected at GROMACS compile time: %s\n\n",
 683                gmx_cpuid_acceleration_string[acc],
 684                gmx_cpuid_acceleration_string[compiled_acc]);
 685     }
 686
 687     return rc;
 688 }
 689
 690
 691 enum gmx_cpuid_x86_smt
 692 gmx_cpuid_x86_smt(gmx_cpuid_t cpuid)
 693 {
 694
 695 #if (defined HAVE_SCHED_H && defined HAVE_SCHED_SETAFFINITY && defined HAVE_SYSCONF && defined __linux__)
 696     int            i;
 697     int            nproc;
 698     cpu_set_t      cpuset,save_cpuset;
 699     int *          apic_id;
 700     unsigned int   eax,ebx,ecx,edx;
 701     int            core_shift_bits;
 702     int            smt_found;
 703
 704     if( gmx_cpuid_vendor(cpuid)!=GMX_CPUID_VENDOR_INTEL ||
 705        gmx_cpuid_feature(cpuid,GMX_CPUID_FEATURE_X86_HTT)==0)
 706     {
 707         return GMX_CPUID_X86_SMT_DISABLED;
 708     }
 709
 710     /* Check cpuid max standard function */
 711     execute_x86cpuid(0x0,0,&eax,&ebx,&ecx,&edx);
 712
 713     /* Early CPUs that do not support function 11 do not support SMT either */
 714     if(eax<0xB)
 715     {
 716         return GMX_CPUID_X86_SMT_DISABLED;
 717     }
 718
 719     /* If we got here, it is a modern Intel CPU that supports detection, as does our OS */
 720
 721     /* How many processors? */
 722     nproc = sysconf(_SC_NPROCESSORS_ONLN);
 723
 724     apic_id      = malloc(sizeof(int)*nproc);
 725
 726     sched_getaffinity(0,sizeof(cpu_set_t),&save_cpuset);
 727
 728     /* Get x2APIC ID from each hardware thread */
 729     CPU_ZERO(&cpuset);
 730     for(i=0;i<nproc;i++)
 731     {
 732         CPU_SET(i,&cpuset);
 733         sched_setaffinity(0,sizeof(cpu_set_t),&cpuset);
 734         execute_x86cpuid(0xB,0,&eax,&ebx,&ecx,&edx);
 735         apic_id[i]=edx;
 736         CPU_CLR(i,&cpuset);
 737     }
 738     /* Reset affinity to the value it had when calling this routine */
 739     sched_setaffinity(0,sizeof(cpu_set_t),&save_cpuset);
 740
 741     core_shift_bits = eax & 0x1F;
 742
 743     /* Check if there is any other APIC id that is identical to [0], apart from
 744      * the hardware thread bit.
 745      */
 746     smt_found  = 0;
 747     for(i=1;i<nproc && smt_found==0;i++)
 748     {
 749         smt_found = (apic_id[i]>>core_shift_bits == apic_id[0] >> core_shift_bits);
 750     }
 751
 752     free(apic_id);
 753
 754     if(smt_found==1)
 755     {
 756         return GMX_CPUID_X86_SMT_ENABLED;
 757     }
 758     else
 759     {
 760         return GMX_CPUID_X86_SMT_DISABLED;
 761     }
 762 #else
 763     /* Do the trivial stuff first. If Hyper-Threading isn't even supported it
 764      * cannot be enabled, no matter what OS detection we use!
 765      */
 766     if(0==gmx_cpuid_feature(cpuid,GMX_CPUID_FEATURE_X86_HTT))
 767     {
 768         return GMX_CPUID_X86_SMT_DISABLED;
 769     }
 770     else
 771     {
 772         return GMX_CPUID_X86_SMT_CANNOTDETECT;
 773     }
 774 #endif
 775 }
 776
 777
 778
 779
 780 #ifdef GMX_CPUID_STANDALONE
 781 /* Stand-alone program to enable queries of CPU features from Cmake.
 782  * Note that you need to check inline ASM capabilities before compiling and set
 783  * -DGMX_X86_GCC_INLINE_ASM for the cpuid instruction to work...
 784  */
 785 int
 786 main(int argc, char **argv)
 787 {
 788     gmx_cpuid_t                   cpuid;
 789     enum gmx_cpuid_acceleration   acc;
 790     int                           i,cnt;
 791
 792     if(argc<2)
 793     {
 794         fprintf(stdout,
 795                 "Usage:\n\n%s [flags]\n\n"
 796                 "Available flags:\n"
 797                 "-vendor        Print CPU vendor.\n"
 798                 "-brand         Print CPU brand string.\n"
 799                 "-family        Print CPU family version.\n"
 800                 "-model         Print CPU model version.\n"
 801                 "-stepping      Print CPU stepping version.\n"
 802                 "-features      Print CPU feature flags.\n"
 803                 "-acceleration  Print suggested GROMACS acceleration.\n"
 804                 ,argv[0]);
 805         exit(0);
 806     }
 807
 808     gmx_cpuid_init(&cpuid);
 809
 810     if(!strncmp(argv[1],"-vendor",3))
 811     {
 812         printf("%s\n",gmx_cpuid_vendor_string[cpuid->vendor]);
 813     }
 814     else if(!strncmp(argv[1],"-brand",3))
 815     {
 816         printf("%s\n",cpuid->brand);
 817     }
 818     else if(!strncmp(argv[1],"-family",3))
 819     {
 820         printf("%d\n",cpuid->family);
 821     }
 822     else if(!strncmp(argv[1],"-model",3))
 823     {
 824         printf("%d\n",cpuid->model);
 825     }
 826     else if(!strncmp(argv[1],"-stepping",3))
 827     {
 828         printf("%d\n",cpuid->stepping);
 829     }
 830     else if(!strncmp(argv[1],"-features",3))
 831     {
 832         cnt = 0;
 833         for(i=0;i<GMX_CPUID_NFEATURES;i++)
 834         {
 835             if(cpuid->feature[i]==1)
 836             {
 837                 if(cnt++ > 0)
 838                 {
 839                     printf(" ");
 840                 }
 841                 printf("%s",gmx_cpuid_feature_string[i]);
 842             }
 843         }
 844         printf("\n");
 845     }
 846     else if(!strncmp(argv[1],"-acceleration",3))
 847     {
 848         acc = gmx_cpuid_acceleration_suggest(cpuid);
 849         fprintf(stdout,"%s\n",gmx_cpuid_acceleration_string[acc]);
 850     }
 851
 852     gmx_cpuid_done(cpuid);
 853
 854
 855     return 0;
 856 }
 857
 858 #endif