src/gromacs/ewald/pme_gpu_3dfft.cu

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2016,2017,2018,2019,2020,2021, 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 CUDA FFT routines for PME GPU.
  38  *
  39  *  \author Aleksei Iupinov <a.yupinov@gmail.com>
  40  *  \ingroup module_ewald
  41  */
  42
  43 #include "gmxpre.h"
  44
  45 #include "pme_gpu_3dfft.h"
  46
  47 #include <cufft.h>
  48
  49 #include "gromacs/gpu_utils/device_stream.h"
  50 #include "gromacs/utility/fatalerror.h"
  51 #include "gromacs/utility/gmxassert.h"
  52
  53 class GpuParallel3dFft::Impl
  54 {
  55 public:
  56     Impl(ivec                 realGridSize,
  57          ivec                 realGridSizePadded,
  58          ivec                 complexGridSizePadded,
  59          bool                 useDecomposition,
  60          bool                 performOutOfPlaceFFT,
  61          const DeviceContext& context,
  62          const DeviceStream&  pmeStream,
  63          DeviceBuffer<float>  realGrid,
  64          DeviceBuffer<float>  complexGrid);
  65     ~Impl();
  66
  67     cufftHandle   planR2C_;
  68     cufftHandle   planC2R_;
  69     cufftReal*    realGrid_;
  70     cufftComplex* complexGrid_;
  71 };
  72
  73 static void handleCufftError(cufftResult_t status, const char* msg)
  74 {
  75     if (status != CUFFT_SUCCESS)
  76     {
  77         gmx_fatal(FARGS, "%s (error code %d)\n", msg, status);
  78     }
  79 }
  80
  81 GpuParallel3dFft::Impl::Impl(ivec       realGridSize,
  82                              ivec       realGridSizePadded,
  83                              ivec       complexGridSizePadded,
  84                              const bool useDecomposition,
  85                              const bool /*performOutOfPlaceFFT*/,
  86                              const DeviceContext& /*context*/,
  87                              const DeviceStream& pmeStream,
  88                              DeviceBuffer<float> realGrid,
  89                              DeviceBuffer<float> complexGrid) :
  90     realGrid_(reinterpret_cast<cufftReal*>(realGrid)),
  91     complexGrid_(reinterpret_cast<cufftComplex*>(complexGrid))
  92 {
  93     GMX_RELEASE_ASSERT(!useDecomposition, "FFT decomposition not implemented");
  94
  95     const int complexGridSizePaddedTotal =
  96             complexGridSizePadded[XX] * complexGridSizePadded[YY] * complexGridSizePadded[ZZ];
  97     const int realGridSizePaddedTotal =
  98             realGridSizePadded[XX] * realGridSizePadded[YY] * realGridSizePadded[ZZ];
  99
 100     GMX_RELEASE_ASSERT(realGrid_, "Bad (null) input real-space grid");
 101     GMX_RELEASE_ASSERT(complexGrid_, "Bad (null) input complex grid");
 102
 103     cufftResult_t result;
 104     /* Commented code for a simple 3D grid with no padding */
 105     /*
 106        result = cufftPlan3d(&planR2C_, realGridSize[XX], realGridSize[YY], realGridSize[ZZ],
 107        CUFFT_R2C); handleCufftError(result, "cufftPlan3d R2C plan failure");
 108
 109        result = cufftPlan3d(&planC2R_, realGridSize[XX], realGridSize[YY], realGridSize[ZZ],
 110        CUFFT_C2R); handleCufftError(result, "cufftPlan3d C2R plan failure");
 111      */
 112
 113     const int rank = 3, batch = 1;
 114     result = cufftPlanMany(&planR2C_,
 115                            rank,
 116                            realGridSize,
 117                            realGridSizePadded,
 118                            1,
 119                            realGridSizePaddedTotal,
 120                            complexGridSizePadded,
 121                            1,
 122                            complexGridSizePaddedTotal,
 123                            CUFFT_R2C,
 124                            batch);
 125     handleCufftError(result, "cufftPlanMany R2C plan failure");
 126
 127     result = cufftPlanMany(&planC2R_,
 128                            rank,
 129                            realGridSize,
 130                            complexGridSizePadded,
 131                            1,
 132                            complexGridSizePaddedTotal,
 133                            realGridSizePadded,
 134                            1,
 135                            realGridSizePaddedTotal,
 136                            CUFFT_C2R,
 137                            batch);
 138     handleCufftError(result, "cufftPlanMany C2R plan failure");
 139
 140     cudaStream_t stream = pmeStream.stream();
 141     GMX_RELEASE_ASSERT(stream, "Can not use the default CUDA stream for PME cuFFT");
 142
 143     result = cufftSetStream(planR2C_, stream);
 144     handleCufftError(result, "cufftSetStream R2C failure");
 145
 146     result = cufftSetStream(planC2R_, stream);
 147     handleCufftError(result, "cufftSetStream C2R failure");
 148 }
 149
 150 GpuParallel3dFft::Impl::~Impl()
 151 {
 152     cufftResult_t result;
 153     result = cufftDestroy(planR2C_);
 154     handleCufftError(result, "cufftDestroy R2C failure");
 155     result = cufftDestroy(planC2R_);
 156     handleCufftError(result, "cufftDestroy C2R failure");
 157 }
 158
 159 void GpuParallel3dFft::perform3dFft(gmx_fft_direction dir, CommandEvent* /*timingEvent*/)
 160 {
 161     cufftResult_t result;
 162     if (dir == GMX_FFT_REAL_TO_COMPLEX)
 163     {
 164         result = cufftExecR2C(impl_->planR2C_, impl_->realGrid_, impl_->complexGrid_);
 165         handleCufftError(result, "cuFFT R2C execution failure");
 166     }
 167     else
 168     {
 169         result = cufftExecC2R(impl_->planC2R_, impl_->complexGrid_, impl_->realGrid_);
 170         handleCufftError(result, "cuFFT C2R execution failure");
 171     }
 172 }
 173
 174 GpuParallel3dFft::GpuParallel3dFft(ivec                 realGridSize,
 175                                    ivec                 realGridSizePadded,
 176                                    ivec                 complexGridSizePadded,
 177                                    const bool           useDecomposition,
 178                                    const bool           performOutOfPlaceFFT,
 179                                    const DeviceContext& context,
 180                                    const DeviceStream&  pmeStream,
 181                                    DeviceBuffer<float>  realGrid,
 182                                    DeviceBuffer<float>  complexGrid) :
 183     impl_(std::make_unique<Impl>(realGridSize,
 184                                  realGridSizePadded,
 185                                  complexGridSizePadded,
 186                                  useDecomposition,
 187                                  performOutOfPlaceFFT,
 188                                  context,
 189                                  pmeStream,
 190                                  realGrid,
 191                                  complexGrid))
 192 {
 193 }
 194
 195 GpuParallel3dFft::~GpuParallel3dFft() = default;