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 #endif
  56
  57 #if Heffte_FOUND
  58 #    include "gpu_3dfft_heffte.h"
  59 #endif
  60
  61 #include "gromacs/utility/arrayref.h"
  62 #include "gromacs/utility/exceptions.h"
  63
  64 namespace gmx
  65 {
  66
  67 // [[noreturn]] attributes must be added in the common headers, so it's easier to silence the warning here
  68 #ifdef __clang__
  69 #    pragma clang diagnostic push
  70 #    pragma clang diagnostic ignored "-Wmissing-noreturn"
  71 #endif
  72
  73 #if (GMX_GPU_CUDA || GMX_GPU_OPENCL || GMX_GPU_SYCL)
  74
  75 Gpu3dFft::Gpu3dFft(FftBackend           backend,
  76                    bool                 allocateGrids,
  77                    MPI_Comm             comm,
  78                    ArrayRef<const int>  gridSizesInXForEachRank,
  79                    ArrayRef<const int>  gridSizesInYForEachRank,
  80                    const int            nz,
  81                    bool                 performOutOfPlaceFFT,
  82                    const DeviceContext& context,
  83                    const DeviceStream&  pmeStream,
  84                    ivec                 realGridSize,
  85                    ivec                 realGridSizePadded,
  86                    ivec                 complexGridSizePadded,
  87                    DeviceBuffer<float>* realGrid,
  88                    DeviceBuffer<float>* complexGrid)
  89 {
  90 #    if GMX_GPU_CUDA
  91     switch (backend)
  92     {
  93         case FftBackend::Cufft:
  94             impl_ = std::make_unique<Gpu3dFft::ImplCuFft>(allocateGrids,
  95                                                           comm,
  96                                                           gridSizesInXForEachRank,
  97                                                           gridSizesInYForEachRank,
  98                                                           nz,
  99                                                           performOutOfPlaceFFT,
 100                                                           context,
 101                                                           pmeStream,
 102                                                           realGridSize,
 103                                                           realGridSizePadded,
 104                                                           complexGridSizePadded,
 105                                                           realGrid,
 106                                                           complexGrid);
 107             break;
 108         default:
 109             GMX_RELEASE_ASSERT(backend == FftBackend::HeFFTe_CUDA,
 110                                "Unsupported FFT backend requested");
 111     }
 112 #    elif GMX_GPU_OPENCL
 113     switch (backend)
 114     {
 115         case FftBackend::Ocl:
 116             impl_ = std::make_unique<Gpu3dFft::ImplOcl>(allocateGrids,
 117                                                         comm,
 118                                                         gridSizesInXForEachRank,
 119                                                         gridSizesInYForEachRank,
 120                                                         nz,
 121                                                         performOutOfPlaceFFT,
 122                                                         context,
 123                                                         pmeStream,
 124                                                         realGridSize,
 125                                                         realGridSizePadded,
 126                                                         complexGridSizePadded,
 127                                                         realGrid,
 128                                                         complexGrid);
 129             break;
 130         default: GMX_THROW(InternalError("Unsupported FFT backend requested"));
 131     }
 132 #    elif GMX_GPU_SYCL
 133     switch (backend)
 134     {
 135         case FftBackend::Sycl:
 136             impl_ = std::make_unique<Gpu3dFft::ImplSycl>(allocateGrids,
 137                                                          comm,
 138                                                          gridSizesInXForEachRank,
 139                                                          gridSizesInYForEachRank,
 140                                                          nz,
 141                                                          performOutOfPlaceFFT,
 142                                                          context,
 143                                                          pmeStream,
 144                                                          realGridSize,
 145                                                          realGridSizePadded,
 146                                                          complexGridSizePadded,
 147                                                          realGrid,
 148                                                          complexGrid);
 149             break;
 150         default: GMX_THROW(InternalError("Unsupported FFT backend requested"));
 151     }
 152 #    endif
 153
 154 #    if Heffte_FOUND
 155     switch (backend)
 156     {
 157         case FftBackend::HeFFTe_CUDA:
 158             GMX_RELEASE_ASSERT(
 159                     GMX_GPU_CUDA,
 160                     "HeFFTe_CUDA FFT backend is supported only with GROMACS compiled with CUDA");
 161             GMX_RELEASE_ASSERT(heffte::backend::is_enabled<heffte::backend::cufft>::value,
 162                                "HeFFTe not compiled with CUDA support");
 163             impl_ = std::make_unique<Gpu3dFft::ImplHeFfte<heffte::backend::cufft>>(
 164                     allocateGrids,
 165                     comm,
 166                     gridSizesInXForEachRank,
 167                     gridSizesInYForEachRank,
 168                     nz,
 169                     performOutOfPlaceFFT,
 170                     context,
 171                     pmeStream,
 172                     realGridSize,
 173                     realGridSizePadded,
 174                     complexGridSizePadded,
 175                     realGrid,
 176                     complexGrid);
 177
 178             break;
 179         default: GMX_RELEASE_ASSERT(impl_ != nullptr, "Unsupported FFT backend requested");
 180     }
 181 #    endif
 182 }
 183
 184 #else
 185
 186 Gpu3dFft::Gpu3dFft(FftBackend /*backend */,
 187                    bool /*allocateGrids*/,
 188                    MPI_Comm /*comm*/,
 189                    ArrayRef<const int> /*gridSizesInXForEachRank*/,
 190                    ArrayRef<const int> /*gridSizesInYForEachRank*/,
 191                    const int /*nz*/,
 192                    bool /*performOutOfPlaceFFT*/,
 193                    const DeviceContext& /*context*/,
 194                    const DeviceStream& /*pmeStream*/,
 195                    ivec /*realGridSize*/,
 196                    ivec /*realGridSizePadded*/,
 197                    ivec /*complexGridSizePadded*/,
 198                    DeviceBuffer<float>* /*realGrid*/,
 199                    DeviceBuffer<float>* /*complexGrid*/)
 200 {
 201     GMX_THROW(InternalError("Cannot run GPU routines in a CPU-only configuration"));
 202 }
 203
 204 #endif
 205
 206 Gpu3dFft::~Gpu3dFft() = default;
 207
 208 void Gpu3dFft::perform3dFft(gmx_fft_direction dir, CommandEvent* timingEvent)
 209 {
 210     GMX_RELEASE_ASSERT(impl_ != nullptr, "Cannot run GPU routines in a CPU-only configuration");
 211     impl_->perform3dFft(dir, timingEvent);
 212 }
 213
 214 #ifdef __clang__
 215 #    pragma clang diagnostic pop
 216 #endif
 217
 218 } // namespace gmx