src/gromacs/fft/gpu_3dfft.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2021, 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 stub GPU 3D FFT routines for CPU-only builds
  38  *
  39  *  \author Mark Abraham <mark.j.abraham@gmail.com>
  40  *  \author Gaurav Garg <gaugarg@nvidia.com>
  41  *  \ingroup module_fft
  42  */
  43
  44 #include "gmxpre.h"
  45
  46 #include "gpu_3dfft.h"
  47 #include "gpu_3dfft_impl.h"
  48
  49 #if GMX_GPU_CUDA
  50 #    include "gpu_3dfft_cufft.h"
  51 #elif GMX_GPU_OPENCL
  52 #    include "gpu_3dfft_ocl.h"
  53 #elif GMX_GPU_SYCL
  54 #    include "gpu_3dfft_sycl.h"
  55 #    if GMX_SYCL_DPCPP && GMX_FFT_MKL
  56 #        include "gpu_3dfft_sycl_mkl.h"
  57 #    endif
  58 #    if GMX_SYCL_HIPSYCL && GMX_HIPSYCL_HAVE_HIP_TARGET
  59 #        include "gpu_3dfft_sycl_rocfft.h"
  60 #    endif
  61 #endif
  62
  63 #if Heffte_FOUND
  64 #    include "gpu_3dfft_heffte.h"
  65 #endif
  66
  67 #include "gromacs/utility/arrayref.h"
  68 #include "gromacs/utility/exceptions.h"
  69
  70 namespace gmx
  71 {
  72
  73 // [[noreturn]] attributes must be added in the common headers, so it's easier to silence the warning here
  74 #ifdef __clang__
  75 #    pragma clang diagnostic push
  76 #    pragma clang diagnostic ignored "-Wmissing-noreturn"
  77 #endif
  78
  79 #if (GMX_GPU_CUDA || GMX_GPU_OPENCL || GMX_GPU_SYCL)
  80
  81 Gpu3dFft::Gpu3dFft(FftBackend           backend,
  82                    bool                 allocateGrids,
  83                    MPI_Comm             comm,
  84                    ArrayRef<const int>  gridSizesInXForEachRank,
  85                    ArrayRef<const int>  gridSizesInYForEachRank,
  86                    const int            nz,
  87                    bool                 performOutOfPlaceFFT,
  88                    const DeviceContext& context,
  89                    const DeviceStream&  pmeStream,
  90                    ivec                 realGridSize,
  91                    ivec                 realGridSizePadded,
  92                    ivec                 complexGridSizePadded,
  93                    DeviceBuffer<float>* realGrid,
  94                    DeviceBuffer<float>* complexGrid)
  95 {
  96 #    if GMX_GPU_CUDA
  97     switch (backend)
  98     {
  99         case FftBackend::Cufft:
 100             impl_ = std::make_unique<Gpu3dFft::ImplCuFft>(allocateGrids,
 101                                                           comm,
 102                                                           gridSizesInXForEachRank,
 103                                                           gridSizesInYForEachRank,
 104                                                           nz,
 105                                                           performOutOfPlaceFFT,
 106                                                           context,
 107                                                           pmeStream,
 108                                                           realGridSize,
 109                                                           realGridSizePadded,
 110                                                           complexGridSizePadded,
 111                                                           realGrid,
 112                                                           complexGrid);
 113             break;
 114         default:
 115             GMX_RELEASE_ASSERT(backend == FftBackend::HeFFTe_CUDA,
 116                                "Unsupported FFT backend requested");
 117     }
 118 #    elif GMX_GPU_OPENCL
 119     switch (backend)
 120     {
 121         case FftBackend::Ocl:
 122             impl_ = std::make_unique<Gpu3dFft::ImplOcl>(allocateGrids,
 123                                                         comm,
 124                                                         gridSizesInXForEachRank,
 125                                                         gridSizesInYForEachRank,
 126                                                         nz,
 127                                                         performOutOfPlaceFFT,
 128                                                         context,
 129                                                         pmeStream,
 130                                                         realGridSize,
 131                                                         realGridSizePadded,
 132                                                         complexGridSizePadded,
 133                                                         realGrid,
 134                                                         complexGrid);
 135             break;
 136         default: GMX_THROW(InternalError("Unsupported FFT backend requested"));
 137     }
 138 #    elif GMX_GPU_SYCL
 139     switch (backend)
 140     {
 141 #        if GMX_SYCL_DPCPP && GMX_FFT_MKL
 142         case FftBackend::SyclMkl:
 143             impl_ = std::make_unique<Gpu3dFft::ImplSyclMkl>(allocateGrids,
 144                                                             comm,
 145                                                             gridSizesInXForEachRank,
 146                                                             gridSizesInYForEachRank,
 147                                                             nz,
 148                                                             performOutOfPlaceFFT,
 149                                                             context,
 150                                                             pmeStream,
 151                                                             realGridSize,
 152                                                             realGridSizePadded,
 153                                                             complexGridSizePadded,
 154                                                             realGrid,
 155                                                             complexGrid);
 156             break;
 157 #        endif
 158 #        if GMX_SYCL_HIPSYCL && GMX_HIPSYCL_HAVE_HIP_TARGET
 159         case FftBackend::SyclRocfft:
 160             impl_ = std::make_unique<Gpu3dFft::ImplSyclRocfft>(allocateGrids,
 161                                                                comm,
 162                                                                gridSizesInXForEachRank,
 163                                                                gridSizesInYForEachRank,
 164                                                                nz,
 165                                                                performOutOfPlaceFFT,
 166                                                                context,
 167                                                                pmeStream,
 168                                                                realGridSize,
 169                                                                realGridSizePadded,
 170                                                                complexGridSizePadded,
 171                                                                realGrid,
 172                                                                complexGrid);
 173             break;
 174 #        endif
 175         case FftBackend::Sycl:
 176             impl_ = std::make_unique<Gpu3dFft::ImplSycl>(allocateGrids,
 177                                                          comm,
 178                                                          gridSizesInXForEachRank,
 179                                                          gridSizesInYForEachRank,
 180                                                          nz,
 181                                                          performOutOfPlaceFFT,
 182                                                          context,
 183                                                          pmeStream,
 184                                                          realGridSize,
 185                                                          realGridSizePadded,
 186                                                          complexGridSizePadded,
 187                                                          realGrid,
 188                                                          complexGrid);
 189             break;
 190         default: GMX_THROW(InternalError("Unsupported FFT backend requested"));
 191     }
 192 #    endif
 193
 194 #    if Heffte_FOUND
 195     switch (backend)
 196     {
 197         case FftBackend::HeFFTe_CUDA:
 198             GMX_RELEASE_ASSERT(
 199                     GMX_GPU_CUDA,
 200                     "HeFFTe_CUDA FFT backend is supported only with GROMACS compiled with CUDA");
 201             GMX_RELEASE_ASSERT(heffte::backend::is_enabled<heffte::backend::cufft>::value,
 202                                "HeFFTe not compiled with CUDA support");
 203             impl_ = std::make_unique<Gpu3dFft::ImplHeFfte<heffte::backend::cufft>>(
 204                     allocateGrids,
 205                     comm,
 206                     gridSizesInXForEachRank,
 207                     gridSizesInYForEachRank,
 208                     nz,
 209                     performOutOfPlaceFFT,
 210                     context,
 211                     pmeStream,
 212                     realGridSize,
 213                     realGridSizePadded,
 214                     complexGridSizePadded,
 215                     realGrid,
 216                     complexGrid);
 217
 218             break;
 219         default: GMX_RELEASE_ASSERT(impl_ != nullptr, "Unsupported FFT backend requested");
 220     }
 221 #    endif
 222 }
 223
 224 #else
 225
 226 Gpu3dFft::Gpu3dFft(FftBackend /*backend */,
 227                    bool /*allocateGrids*/,
 228                    MPI_Comm /*comm*/,
 229                    ArrayRef<const int> /*gridSizesInXForEachRank*/,
 230                    ArrayRef<const int> /*gridSizesInYForEachRank*/,
 231                    const int /*nz*/,
 232                    bool /*performOutOfPlaceFFT*/,
 233                    const DeviceContext& /*context*/,
 234                    const DeviceStream& /*pmeStream*/,
 235                    ivec /*realGridSize*/,
 236                    ivec /*realGridSizePadded*/,
 237                    ivec /*complexGridSizePadded*/,
 238                    DeviceBuffer<float>* /*realGrid*/,
 239                    DeviceBuffer<float>* /*complexGrid*/)
 240 {
 241     GMX_THROW(InternalError("Cannot run GPU routines in a CPU-only configuration"));
 242 }
 243
 244 #endif
 245
 246 Gpu3dFft::~Gpu3dFft() = default;
 247
 248 void Gpu3dFft::perform3dFft(gmx_fft_direction dir, CommandEvent* timingEvent)
 249 {
 250     GMX_RELEASE_ASSERT(impl_ != nullptr, "Cannot run GPU routines in a CPU-only configuration");
 251     impl_->perform3dFft(dir, timingEvent);
 252 }
 253
 254 #ifdef __clang__
 255 #    pragma clang diagnostic pop
 256 #endif
 257
 258 } // namespace gmx