src/gromacs/simd/support.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2015,2016,2017,2018,2019, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35
  36 /*! \internal \file
  37  *
  38  * \brief Implements SIMD architecture support query routines
  39  *
  40  * \author Erik Lindahl <erik.lindahl@scilifelab.se>
  41  *
  42  * \ingroup module_simd
  43  */
  44
  45 #include "gmxpre.h"
  46
  47 #include "support.h"
  48
  49 #include "config.h"
  50
  51 #include <cstdio>
  52 #include <cstdlib>
  53
  54 #include <map>
  55 #include <string>
  56
  57 #include "gromacs/hardware/cpuinfo.h"
  58 #include "gromacs/hardware/identifyavx512fmaunits.h"
  59 #include "gromacs/utility/stringutil.h"
  60
  61 namespace gmx
  62 {
  63
  64 /*! \cond libapi */
  65
  66 const std::string &
  67 simdString(SimdType s)
  68 {
  69     static const std::map<SimdType, std::string> name =
  70     {
  71         { SimdType::None,           "None"            },
  72         { SimdType::Reference,      "Reference"       },
  73         { SimdType::Generic,        "Generic"         },
  74         { SimdType::X86_Sse2,       "SSE2"            },
  75         { SimdType::X86_Sse4_1,     "SSE4.1"          },
  76         { SimdType::X86_Avx128Fma,  "AVX_128_FMA"     },
  77         { SimdType::X86_Avx,        "AVX_256"         },
  78         { SimdType::X86_Avx2,       "AVX2_256"        },
  79         { SimdType::X86_Avx2_128,   "AVX2_128"        },
  80         { SimdType::X86_Avx512,     "AVX_512"         },
  81         { SimdType::X86_Avx512Knl,  "AVX_512_KNL"     },
  82         { SimdType::X86_Mic,        "X86_MIC"         },
  83         { SimdType::Arm_Neon,       "ARM_NEON"        },
  84         { SimdType::Arm_NeonAsimd,  "ARM_NEON_ASIMD"  },
  85         { SimdType::Ibm_Vmx,        "IBM_VMX"         },
  86         { SimdType::Ibm_Vsx,        "IBM_VSX"         },
  87         { SimdType::Fujitsu_HpcAce, "Fujitsu HPC-ACE" }
  88     };
  89
  90     return name.at(s);
  91 }
  92
  93 SimdType
  94 simdSuggested(const CpuInfo &c)
  95 {
  96     SimdType suggested = SimdType::None;
  97
  98     if (c.supportLevel() >= CpuInfo::SupportLevel::Features)
  99     {
 100         switch (c.vendor())
 101         {
 102             case CpuInfo::Vendor::Intel:
 103                 if (c.feature(CpuInfo::Feature::X86_Avx512ER))
 104                 {
 105                     suggested = SimdType::X86_Avx512Knl;
 106                 }
 107                 else if (c.feature(CpuInfo::Feature::X86_Avx512F))
 108                 {
 109                     // If we could not identify the number of AVX512 FMA units we assume 2
 110                     suggested = ( identifyAvx512FmaUnits() == 1 ) ? SimdType::X86_Avx2 : SimdType::X86_Avx512;
 111                 }
 112                 else if (c.feature(CpuInfo::Feature::X86_Avx2))
 113                 {
 114                     suggested = SimdType::X86_Avx2;
 115                 }
 116                 else if (c.feature(CpuInfo::Feature::X86_Avx))
 117                 {
 118                     suggested = SimdType::X86_Avx;
 119                 }
 120                 else if (c.feature(CpuInfo::Feature::X86_Sse4_1))
 121                 {
 122                     suggested = SimdType::X86_Sse4_1;
 123                 }
 124                 else if (c.feature(CpuInfo::Feature::X86_Sse2))
 125                 {
 126                     suggested = SimdType::X86_Sse2;
 127                 }
 128                 break;
 129             case CpuInfo::Vendor::Amd:
 130             case CpuInfo::Vendor::Hygon:
 131                 if (c.feature(CpuInfo::Feature::X86_Avx2))
 132                 {
 133                     // AMD Ryzen supports 256-bit AVX2, but performs better with 128-bit
 134                     // since it can execute two independent such instructions per cycle,
 135                     // and wider SIMD has slightly lower efficiency in GROMACS.
 136                     suggested = SimdType::X86_Avx2_128;
 137                 }
 138                 else if (c.feature(CpuInfo::Feature::X86_Avx))
 139                 {
 140                     // Use 128-bit FMA SIMD if Fma4 flag is set, otherwise plain 256-bit AVX
 141                     if (c.feature(CpuInfo::Feature::X86_Fma4))
 142                     {
 143                         suggested = SimdType::X86_Avx128Fma;
 144                     }
 145                     else
 146                     {
 147                         suggested = SimdType::X86_Avx;
 148                     }
 149                 }
 150                 else if (c.feature(CpuInfo::Feature::X86_Sse4_1))
 151                 {
 152                     suggested = SimdType::X86_Sse4_1;
 153                 }
 154                 else if (c.feature(CpuInfo::Feature::X86_Sse2))
 155                 {
 156                     suggested = SimdType::X86_Sse2;
 157                 }
 158
 159                 break;
 160             case CpuInfo::Vendor::Arm:
 161                 if (c.feature(CpuInfo::Feature::Arm_NeonAsimd))
 162                 {
 163                     suggested = SimdType::Arm_NeonAsimd;
 164                 }
 165                 else if (c.feature(CpuInfo::Feature::Arm_Neon))
 166                 {
 167                     suggested = SimdType::Arm_Neon;
 168                 }
 169                 break;
 170             case CpuInfo::Vendor::Ibm:
 171                 if (c.feature(CpuInfo::Feature::Ibm_Vsx))
 172                 {
 173                     suggested = SimdType::Ibm_Vsx;
 174                 }
 175                 else if (c.feature(CpuInfo::Feature::Ibm_Vmx))
 176                 {
 177                     suggested = SimdType::Ibm_Vmx;
 178                 }
 179                 break;
 180             case CpuInfo::Vendor::Fujitsu:
 181                 if (c.feature(CpuInfo::Feature::Fujitsu_HpcAce))
 182                 {
 183                     suggested = SimdType::Fujitsu_HpcAce;
 184                 }
 185                 break;
 186             default:
 187                 break;
 188         }
 189     }
 190     return suggested;
 191 }
 192
 193 SimdType
 194 simdCompiled()
 195 {
 196 #if GMX_SIMD_X86_AVX_512_KNL
 197     return SimdType::X86_Avx512Knl;
 198 #elif GMX_SIMD_X86_AVX_512
 199     return SimdType::X86_Avx512;
 200 #elif GMX_SIMD_X86_MIC
 201     return SimdType::X86_Mic;
 202 #elif GMX_SIMD_X86_AVX2_256
 203     return SimdType::X86_Avx2;
 204 #elif GMX_SIMD_X86_AVX2_128
 205     return SimdType::X86_Avx2_128;
 206 #elif GMX_SIMD_X86_AVX_256
 207     return SimdType::X86_Avx;
 208 #elif GMX_SIMD_X86_AVX_128_FMA
 209     return SimdType::X86_Avx128Fma;
 210 #elif GMX_SIMD_X86_SSE4_1
 211     return SimdType::X86_Sse4_1;
 212 #elif GMX_SIMD_X86_SSE2
 213     return SimdType::X86_Sse2;
 214 #elif GMX_SIMD_ARM_NEON
 215     return SimdType::Arm_Neon;
 216 #elif GMX_SIMD_ARM_NEON_ASIMD
 217     return SimdType::Arm_NeonAsimd;
 218 #elif GMX_SIMD_IBM_VMX
 219     return SimdType::Ibm_Vmx;
 220 #elif GMX_SIMD_IBM_VSX
 221     return SimdType::Ibm_Vsx;
 222 #elif GMX_SIMD_SPARC64_HPC_ACE
 223     return SimdType::Fujitsu_HpcAce;
 224 #elif GMX_SIMD_REFERENCE
 225     return SimdType::Reference;
 226 #else
 227     return SimdType::None;
 228 #endif
 229 }
 230
 231 bool
 232 simdCheck(gmx::SimdType    wanted,
 233           FILE *           log,
 234           bool             warnToStdErr)
 235 {
 236     SimdType             compiled  = simdCompiled();
 237
 238     gmx::TextLineWrapper wrapper;
 239     std::string          logMsg;
 240     std::string          warnMsg;
 241
 242     wrapper.settings().setLineLength(78);
 243
 244     if (compiled == SimdType::X86_Avx2 && wanted == SimdType::X86_Avx512)
 245     {
 246         logMsg = wrapper.wrapToString(formatString("Highest SIMD level requested by all nodes in run: %s\n"
 247                                                    "SIMD instructions selected at compile time:       %s\n"
 248                                                    "This program was compiled for different hardware than you are running on, "
 249                                                    "which could influence performance. This build might have been configured on "
 250                                                    "a login node with only a single AVX-512 FMA unit (in which case AVX2 is faster), "
 251                                                    "while the node you are running on has dual AVX-512 FMA units.",
 252                                                    simdString(wanted).c_str(), simdString(compiled).c_str()));
 253         warnMsg = wrapper.wrapToString(formatString("Compiled SIMD: %s, but for this host/run %s might be better (see log).",
 254                                                     simdString(compiled).c_str(), simdString(wanted).c_str()));
 255     }
 256     else if (compiled == SimdType::X86_Avx512 && wanted == SimdType::X86_Avx2 && identifyAvx512FmaUnits() == 1)
 257     {
 258         // The reason for explicitly checking the number of FMA units above is to avoid triggering
 259         // this conditional if the AVX2 SIMD was requested by some other node in a heterogeneous MPI run.
 260         logMsg = wrapper.wrapToString(formatString("Highest SIMD level requested by all nodes in run: %s\n"
 261                                                    "SIMD instructions selected at compile time:       %s\n"
 262                                                    "This program was compiled for different hardware than you are running on, "
 263                                                    "which could influence performance."
 264                                                    "This host supports AVX-512, but since it only has 1 AVX-512"
 265                                                    "FMA unit, it would be faster to use AVX2 instead.",
 266                                                    simdString(wanted).c_str(), simdString(compiled).c_str()));
 267         warnMsg = wrapper.wrapToString(formatString("Compiled SIMD: %s, but for this host/run %s might be better (see log).",
 268                                                     simdString(compiled).c_str(), simdString(wanted).c_str()));
 269     }
 270     else if (compiled == SimdType::X86_Avx2 && wanted == SimdType::X86_Avx2_128)
 271     {
 272         // Wanted SimdType::X86_Avx2_128 can only be the AMD Zen architecture.
 273         // AVX2_256 is only up to a few percent slower than AVX2_128
 274         // in both single and double precision. AVX2_256 is slightly
 275         // faster with nonbondeds and PME on a GPU. Don't warn the user.
 276     }
 277     else if (compiled > wanted && !(compiled == SimdType::X86_Avx && wanted == SimdType::X86_Avx128Fma))
 278     {
 279         // Normally it is close to catastrophic if the compiled SIMD type is larger than
 280         // the supported one, but AVX128Fma is an exception: AMD CPUs will (strongly) prefer
 281         // AVX128Fma, but they will work fine with AVX too. Thus, make an exception for this.
 282         logMsg = wrapper.wrapToString(formatString("Highest SIMD level requested by all nodes in run: %s\n"
 283                                                    "SIMD instructions selected at compile time:       %s\n"
 284                                                    "Compiled SIMD newer than requested; program might crash.",
 285                                                    simdString(wanted).c_str(), simdString(compiled).c_str()));
 286         warnMsg = logMsg;
 287     }
 288     else if (wanted != compiled)
 289     {
 290         // This warning will also occur if compiled is X86_Avx and wanted is X86_Avx128Fma
 291         logMsg = wrapper.wrapToString(formatString("Highest SIMD level requested by all nodes in run: %s\n"
 292                                                    "SIMD instructions selected at compile time:       %s\n"
 293                                                    "This program was compiled for different hardware than you are running on, "
 294                                                    "which could influence performance.",
 295                                                    simdString(wanted).c_str(), simdString(compiled).c_str()));
 296         warnMsg = wrapper.wrapToString(formatString("Compiled SIMD: %s, but for this host/run %s might be better (see log).",
 297                                                     simdString(compiled).c_str(), simdString(wanted).c_str()));
 298     }
 299
 300     if (!logMsg.empty() && log != nullptr)
 301     {
 302         fprintf(log, "%s\n", logMsg.c_str());
 303     }
 304     if (!warnMsg.empty() && warnToStdErr)
 305     {
 306         fprintf(stderr, "%s\n", warnMsg.c_str());
 307     }
 308
 309     return (wanted == compiled);
 310 }
 311
 312 /*! \endcond */
 313
 314 }  // namespace gmx