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_HIPSYCL
  56 #        include "gpu_3dfft_sycl_rocfft.h"
  57 #    endif
  58 #endif
  59
  60 #if Heffte_FOUND
  61 #    include "gpu_3dfft_heffte.h"
  62 #endif
  63
  64 #include "gromacs/utility/arrayref.h"
  65 #include "gromacs/utility/exceptions.h"
  66
  67 namespace gmx
  68 {
  69
  70 // [[noreturn]] attributes must be added in the common headers, so it's easier to silence the warning here
  71 #ifdef __clang__
  72 #    pragma clang diagnostic push
  73 #    pragma clang diagnostic ignored "-Wmissing-noreturn"
  74 #endif
  75
  76 #if (GMX_GPU_CUDA || GMX_GPU_OPENCL || GMX_GPU_SYCL)
  77
  78 Gpu3dFft::Gpu3dFft(FftBackend           backend,
  79                    bool                 allocateGrids,
  80                    MPI_Comm             comm,
  81                    ArrayRef<const int>  gridSizesInXForEachRank,
  82                    ArrayRef<const int>  gridSizesInYForEachRank,
  83                    const int            nz,
  84                    bool                 performOutOfPlaceFFT,
  85                    const DeviceContext& context,
  86                    const DeviceStream&  pmeStream,
  87                    ivec                 realGridSize,
  88                    ivec                 realGridSizePadded,
  89                    ivec                 complexGridSizePadded,
  90                    DeviceBuffer<float>* realGrid,
  91                    DeviceBuffer<float>* complexGrid)
  92 {
  93 #    if GMX_GPU_CUDA
  94     switch (backend)
  95     {
  96         case FftBackend::Cufft:
  97             impl_ = std::make_unique<Gpu3dFft::ImplCuFft>(allocateGrids,
  98                                                           comm,
  99                                                           gridSizesInXForEachRank,
 100                                                           gridSizesInYForEachRank,
 101                                                           nz,
 102                                                           performOutOfPlaceFFT,
 103                                                           context,
 104                                                           pmeStream,
 105                                                           realGridSize,
 106                                                           realGridSizePadded,
 107                                                           complexGridSizePadded,
 108                                                           realGrid,
 109                                                           complexGrid);
 110             break;
 111         default:
 112             GMX_RELEASE_ASSERT(backend == FftBackend::HeFFTe_CUDA,
 113                                "Unsupported FFT backend requested");
 114     }
 115 #    elif GMX_GPU_OPENCL
 116     switch (backend)
 117     {
 118         case FftBackend::Ocl:
 119             impl_ = std::make_unique<Gpu3dFft::ImplOcl>(allocateGrids,
 120                                                         comm,
 121                                                         gridSizesInXForEachRank,
 122                                                         gridSizesInYForEachRank,
 123                                                         nz,
 124                                                         performOutOfPlaceFFT,
 125                                                         context,
 126                                                         pmeStream,
 127                                                         realGridSize,
 128                                                         realGridSizePadded,
 129                                                         complexGridSizePadded,
 130                                                         realGrid,
 131                                                         complexGrid);
 132             break;
 133         default: GMX_THROW(InternalError("Unsupported FFT backend requested"));
 134     }
 135 #    elif GMX_GPU_SYCL
 136     switch (backend)
 137     {
 138 #        if GMX_SYCL_HIPSYCL
 139         case FftBackend::SyclRocfft:
 140             impl_ = std::make_unique<Gpu3dFft::ImplSyclRocfft>(allocateGrids,
 141                                                                comm,
 142                                                                gridSizesInXForEachRank,
 143                                                                gridSizesInYForEachRank,
 144                                                                nz,
 145                                                                performOutOfPlaceFFT,
 146                                                                context,
 147                                                                pmeStream,
 148                                                                realGridSize,
 149                                                                realGridSizePadded,
 150                                                                complexGridSizePadded,
 151                                                                realGrid,
 152                                                                complexGrid);
 153             break;
 154 #        endif
 155         case FftBackend::Sycl:
 156             impl_ = std::make_unique<Gpu3dFft::ImplSycl>(allocateGrids,
 157                                                          comm,
 158                                                          gridSizesInXForEachRank,
 159                                                          gridSizesInYForEachRank,
 160                                                          nz,
 161                                                          performOutOfPlaceFFT,
 162                                                          context,
 163                                                          pmeStream,
 164                                                          realGridSize,
 165                                                          realGridSizePadded,
 166                                                          complexGridSizePadded,
 167                                                          realGrid,
 168                                                          complexGrid);
 169             break;
 170         default: GMX_THROW(InternalError("Unsupported FFT backend requested"));
 171     }
 172 #    endif
 173
 174 #    if Heffte_FOUND
 175     switch (backend)
 176     {
 177         case FftBackend::HeFFTe_CUDA:
 178             GMX_RELEASE_ASSERT(
 179                     GMX_GPU_CUDA,
 180                     "HeFFTe_CUDA FFT backend is supported only with GROMACS compiled with CUDA");
 181             GMX_RELEASE_ASSERT(heffte::backend::is_enabled<heffte::backend::cufft>::value,
 182                                "HeFFTe not compiled with CUDA support");
 183             impl_ = std::make_unique<Gpu3dFft::ImplHeFfte<heffte::backend::cufft>>(
 184                     allocateGrids,
 185                     comm,
 186                     gridSizesInXForEachRank,
 187                     gridSizesInYForEachRank,
 188                     nz,
 189                     performOutOfPlaceFFT,
 190                     context,
 191                     pmeStream,
 192                     realGridSize,
 193                     realGridSizePadded,
 194                     complexGridSizePadded,
 195                     realGrid,
 196                     complexGrid);
 197
 198             break;
 199         default: GMX_RELEASE_ASSERT(impl_ != nullptr, "Unsupported FFT backend requested");
 200     }
 201 #    endif
 202 }
 203
 204 #else
 205
 206 Gpu3dFft::Gpu3dFft(FftBackend /*backend */,
 207                    bool /*allocateGrids*/,
 208                    MPI_Comm /*comm*/,
 209                    ArrayRef<const int> /*gridSizesInXForEachRank*/,
 210                    ArrayRef<const int> /*gridSizesInYForEachRank*/,
 211                    const int /*nz*/,
 212                    bool /*performOutOfPlaceFFT*/,
 213                    const DeviceContext& /*context*/,
 214                    const DeviceStream& /*pmeStream*/,
 215                    ivec /*realGridSize*/,
 216                    ivec /*realGridSizePadded*/,
 217                    ivec /*complexGridSizePadded*/,
 218                    DeviceBuffer<float>* /*realGrid*/,
 219                    DeviceBuffer<float>* /*complexGrid*/)
 220 {
 221     GMX_THROW(InternalError("Cannot run GPU routines in a CPU-only configuration"));
 222 }
 223
 224 #endif
 225
 226 Gpu3dFft::~Gpu3dFft() = default;
 227
 228 void Gpu3dFft::perform3dFft(gmx_fft_direction dir, CommandEvent* timingEvent)
 229 {
 230     GMX_RELEASE_ASSERT(impl_ != nullptr, "Cannot run GPU routines in a CPU-only configuration");
 231     impl_->perform3dFft(dir, timingEvent);
 232 }
 233
 234 #ifdef __clang__
 235 #    pragma clang diagnostic pop
 236 #endif
 237
 238 } // namespace gmx