src/gromacs/hardware/identifyavx512fmaunits.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 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  * \brief Implements a routine to check the number of AVX512 fma units
  38  *
  39  * Just as the CpuInfo code, we need to be able to compile this file in stand-alone mode
  40  * to set the SIMD acceleration and similar things during CMake configuration.
  41  */
  42
  43 #ifndef GMX_IDENTIFY_AVX512_FMA_UNITS_STANDALONE
  44 #    include "gmxpre.h"
  45 #endif
  46
  47 #include "identifyavx512fmaunits.h"
  48
  49 #ifndef GMX_IDENTIFY_AVX512_FMA_UNITS_STANDALONE
  50 #    include "config.h"
  51 #endif
  52
  53 #include <cstdint>
  54 #include <cstdio>
  55
  56 #include <algorithm>
  57 #include <mutex>
  58
  59 #ifndef GMX_IDENTIFY_AVX512_FMA_UNITS_STANDALONE
  60 #    include "gromacs/hardware/cpuinfo.h"
  61 #endif
  62
  63 namespace gmx
  64 {
  65
  66 namespace
  67 {
  68
  69 #if GMX_X86_GCC_INLINE_ASM && SIMD_AVX_512_CXX_SUPPORTED
  70 /*\ brief Loop over mixed FMA and shuffle AVX512 instructions
  71  *
  72  * This function executes a meaningless loop that includes both
  73  * FMA and shuffle instructions from the AVX512 instruction set.
  74  * We need a bit of complex logic to make sure it cannot be
  75  * optimized away by the compiler.
  76  *
  77  * \param loopCount  Number of iterations. Each iteration will
  78  *                   execute 12 FMA and 12 shuffle instructions.
  79  * \return Number of cycles used for the loop.
  80  */
  81 uint64_t timeFmaAndShuffleLoop(uint64_t loopCount)
  82 {
  83     uint64_t cycles;
  84     // Unfortunately we need to resort to inline ASM since we are
  85     // making a choice based on timing, and without efficient optimization
  86     // (e.g. when doing debugging) the usual intrinsics are often implemented
  87     // as independent load/store operations, which completely screws up timing.
  88     __asm__ __volatile__(
  89             "\tvpxord %%zmm0, %%zmm0, %%zmm0\n"
  90             "\tvmovaps %%zmm0, %%zmm1\n"
  91             "\tvmovaps %%zmm0, %%zmm2\n"
  92             "\tvmovaps %%zmm0, %%zmm3\n"
  93             "\tvmovaps %%zmm0, %%zmm4\n"
  94             "\tvmovaps %%zmm0, %%zmm5\n"
  95             "\tvmovaps %%zmm0, %%zmm6\n"
  96             "\tvmovaps %%zmm0, %%zmm7\n"
  97             "\tvmovaps %%zmm0, %%zmm8\n"
  98             "\tvmovaps %%zmm0, %%zmm9\n"
  99             "\tvmovaps %%zmm0, %%zmm10\n"
 100             "\tvmovaps %%zmm0, %%zmm11\n"
 101             "\tvpxord %%zmm12, %%zmm12, %%zmm12\n"
 102             "\tvmovaps %%zmm12, %%zmm13\n"
 103             "\tvmovaps %%zmm12, %%zmm14\n"
 104             "\tvmovaps %%zmm12, %%zmm15\n"
 105             "\tvmovaps %%zmm12, %%zmm16\n"
 106             "\tvmovaps %%zmm12, %%zmm17\n"
 107             "\tvmovaps %%zmm12, %%zmm18\n"
 108             "\tvmovaps %%zmm12, %%zmm19\n"
 109             "\tvmovaps %%zmm12, %%zmm20\n"
 110             "\tvmovaps %%zmm12, %%zmm21\n"
 111             "\tvmovaps %%zmm12, %%zmm22\n"
 112             "\tvmovaps %%zmm12, %%zmm23\n"
 113             "\tvmovaps %%zmm12, %%zmm30\n"
 114             "\trdtscp\n"
 115             "\tsalq $32, %%rdx\n"
 116             "\tmovl %%eax, %%eax\n"
 117             "\tmovq %%rdx, %%rbx\n"
 118             "\torq %%rax, %%rbx\n"
 119             "\tmovq %1, %%rdx\n"
 120             "1:\n"
 121             "\tvfmadd231pd %%zmm0, %%zmm0, %%zmm0\n"
 122             "\tvfmadd231pd %%zmm1, %%zmm1, %%zmm1\n"
 123             "\tvfmadd231pd %%zmm2, %%zmm2, %%zmm2\n"
 124             "\tvfmadd231pd %%zmm3, %%zmm3, %%zmm3\n"
 125             "\tvfmadd231pd %%zmm4, %%zmm4, %%zmm4\n"
 126             "\tvfmadd231pd %%zmm5, %%zmm5, %%zmm5\n"
 127             "\tvfmadd231pd %%zmm6, %%zmm6, %%zmm6\n"
 128             "\tvfmadd231pd %%zmm7, %%zmm7, %%zmm7\n"
 129             "\tvfmadd231pd %%zmm8, %%zmm8, %%zmm8\n"
 130             "\tvfmadd231pd %%zmm9, %%zmm9, %%zmm9\n"
 131             "\tvfmadd231pd %%zmm10, %%zmm10, %%zmm10\n"
 132             "\tvfmadd231pd %%zmm11, %%zmm11, %%zmm11\n"
 133             "\tvpermd %%zmm30, %%zmm30, %%zmm12\n"
 134             "\tvpermd %%zmm30, %%zmm30, %%zmm13\n"
 135             "\tvpermd %%zmm30, %%zmm30, %%zmm14\n"
 136             "\tvpermd %%zmm30, %%zmm30, %%zmm15\n"
 137             "\tvpermd %%zmm30, %%zmm30, %%zmm16\n"
 138             "\tvpermd %%zmm30, %%zmm30, %%zmm17\n"
 139             "\tvpermd %%zmm30, %%zmm30, %%zmm18\n"
 140             "\tvpermd %%zmm30, %%zmm30, %%zmm19\n"
 141             "\tvpermd %%zmm30, %%zmm30, %%zmm20\n"
 142             "\tvpermd %%zmm30, %%zmm30, %%zmm21\n"
 143             "\tvpermd %%zmm30, %%zmm30, %%zmm22\n"
 144             "\tvpermd %%zmm30, %%zmm30, %%zmm23\n"
 145             "\tdec %%rdx\n"
 146             "\tjg 1b\n"
 147             "\trdtscp\n"
 148             "\tsalq $32, %%rdx\n"
 149             "\tmovl %%eax, %%eax\n"
 150             "\torq %%rax, %%rdx\n"
 151             "\tsubq %%rbx, %%rdx\n"
 152             "\tmovq %%rdx, %0\n"
 153             : "=r"(cycles)
 154             : "r"(loopCount)
 155             : "rax", "rbx", "rcx", "rdx", "zmm0", "zmm1", "zmm2", "zmm3", "zmm4", "zmm5", "zmm6",
 156               "zmm7", "zmm8", "zmm9", "zmm10", "zmm11", "zmm12", "zmm13", "zmm14", "zmm15", "zmm16",
 157               "zmm17", "zmm18", "zmm19", "zmm20", "zmm21", "zmm22", "zmm23", "zmm30");
 158
 159     return cycles;
 160 }
 161
 162 /*\ brief Loop over FMA AVX512 instructions
 163  *
 164  * This function executes a meaningless loop that includes only
 165  * FMA instructions from the AVX512 instruction set.
 166  * We need a bit of complex logic to make sure it cannot be
 167  * optimized away by the compiler.
 168  *
 169  * \param loopCount  Number of iterations. Each iteration will
 170  *                   execute 12 FMA instructions.
 171  * \return Number of cycles used for the loop.
 172  */
 173 uint64_t timeFmaOnlyLoop(uint64_t loopCount)
 174 {
 175     uint64_t cycles;
 176     // Unfortunately we need to resort to inline ASM since we are
 177     // making a choice based on timing, and without efficient optimization
 178     // (e.g. when doing debugging) the usual intrinsics are often implemented
 179     // as independent load/store operations, which completely screws up timing.
 180     __asm__ __volatile__(
 181             "\tvpxord %%zmm0, %%zmm0, %%zmm0\n"
 182             "\tvmovaps %%zmm0, %%zmm1\n"
 183             "\tvmovaps %%zmm0, %%zmm2\n"
 184             "\tvmovaps %%zmm0, %%zmm3\n"
 185             "\tvmovaps %%zmm0, %%zmm4\n"
 186             "\tvmovaps %%zmm0, %%zmm5\n"
 187             "\tvmovaps %%zmm0, %%zmm6\n"
 188             "\tvmovaps %%zmm0, %%zmm7\n"
 189             "\tvmovaps %%zmm0, %%zmm8\n"
 190             "\tvmovaps %%zmm0, %%zmm9\n"
 191             "\tvmovaps %%zmm0, %%zmm10\n"
 192             "\tvmovaps %%zmm0, %%zmm11\n"
 193             "\trdtscp\n"
 194             "\tsalq $32, %%rdx\n"
 195             "\tmovl %%eax, %%eax\n"
 196             "\tmovq %%rdx, %%rbx\n"
 197             "\torq %%rax, %%rbx\n"
 198             "\tmovq %1, %%rdx\n"
 199             "1:\n"
 200             "\tvfmadd231pd %%zmm0, %%zmm0, %%zmm0\n"
 201             "\tvfmadd231pd %%zmm1, %%zmm1, %%zmm1\n"
 202             "\tvfmadd231pd %%zmm2, %%zmm2, %%zmm2\n"
 203             "\tvfmadd231pd %%zmm3, %%zmm3, %%zmm3\n"
 204             "\tvfmadd231pd %%zmm4, %%zmm4, %%zmm4\n"
 205             "\tvfmadd231pd %%zmm5, %%zmm5, %%zmm5\n"
 206             "\tvfmadd231pd %%zmm6, %%zmm6, %%zmm6\n"
 207             "\tvfmadd231pd %%zmm7, %%zmm7, %%zmm7\n"
 208             "\tvfmadd231pd %%zmm8, %%zmm8, %%zmm8\n"
 209             "\tvfmadd231pd %%zmm9, %%zmm9, %%zmm9\n"
 210             "\tvfmadd231pd %%zmm10, %%zmm10, %%zmm10\n"
 211             "\tvfmadd231pd %%zmm11, %%zmm11, %%zmm11\n"
 212             "\tdec %%rdx\n"
 213             "\tjg 1b\n"
 214             "\trdtscp\n"
 215             "\tsalq $32, %%rdx\n"
 216             "\tmovl %%eax, %%eax\n"
 217             "\torq %%rax, %%rdx\n"
 218             "\tsubq %%rbx, %%rdx\n"
 219             "\tmovq %%rdx, %0\n"
 220             : "=r"(cycles)
 221             : "r"(loopCount)
 222             : "rax", "rbx", "rcx", "rdx", "zmm0", "zmm1", "zmm2", "zmm3", "zmm4", "zmm5", "zmm6",
 223               "zmm7", "zmm8", "zmm9", "zmm10", "zmm11");
 224
 225     return cycles;
 226 }
 227
 228 bool checkDualAvx512FmaUnits()
 229 {
 230     uint64_t timeFmaAndShuf = static_cast<uint64_t>(1e9); // Large value
 231
 232     // Make sure the CPU is in AVX512 mode by executing a fairly long loop.
 233     // Use the return value to make sure it is not optimized away. Later invocations
 234     // use fewer iterations, so they should always be faster.
 235     uint64_t timeFmaOnly = timeFmaOnlyLoop(100000);
 236
 237     // Execute the loops three times
 238     for (int i = 0; i < 3; i++)
 239     {
 240         timeFmaAndShuf = std::min(timeFmaAndShuf, timeFmaAndShuffleLoop(1000));
 241         timeFmaOnly    = std::min(timeFmaOnly, timeFmaOnlyLoop(1000));
 242     }
 243
 244     return timeFmaAndShuf > 1.5 * timeFmaOnly;
 245 }
 246
 247 #endif // GMX_X86_GCC_INLINE_ASM && SIMD_AVX_512_CXX_SUPPORTED
 248
 249
 250 /*! \brief Mutex to guard the execution of the timing test
 251  *
 252  * We only execute the test once, and return the saved result
 253  * on subsequent calls.
 254  */
 255 std::mutex initMutex;
 256
 257 } // namespace
 258
 259 int identifyAvx512FmaUnits()
 260 {
 261     static bool initialized = false;
 262     static int  result      = 0;
 263
 264     if (!initialized)
 265     {
 266         std::lock_guard<std::mutex> lock(initMutex);
 267
 268         if (!initialized)
 269         {
 270             // For the standalone test binary we assume it will
 271             // only be executed on AVX512 hardware, but for the
 272             // library version we check the hardware support.
 273 #ifdef GMX_IDENTIFY_AVX512_FMA_UNITS_STANDALONE
 274             bool haveAvx512Hardware = true;
 275 #else
 276             bool haveAvx512Hardware = CpuInfo::detect().feature(CpuInfo::Feature::X86_Avx512F);
 277 #endif
 278
 279             if (haveAvx512Hardware)
 280             {
 281 #if GMX_X86_GCC_INLINE_ASM && SIMD_AVX_512_CXX_SUPPORTED
 282                 result = checkDualAvx512FmaUnits() ? 2 : 1;
 283 #else
 284                 result = -1; // Cannot run the tests
 285 #endif
 286             }
 287             else
 288             {
 289                 result = 0; // Not AVX-512 hardware
 290             }
 291             initialized = true;
 292         }
 293     }
 294     return result;
 295 }
 296
 297 } // namespace gmx
 298
 299 #ifdef GMX_IDENTIFY_AVX512_FMA_UNITS_STANDALONE
 300 int main()
 301 {
 302     printf("%d\n", gmx::identifyAvx512FmaUnits());
 303     return 0;
 304 }
 305 #endif