src/gromacs/ewald/pme_gpu_3dfft_ocl.cpp

   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 OpenCL 3D 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 <array>
  48 #include <vector>
  49
  50 #include <clFFT.h>
  51
  52 #include "gromacs/gpu_utils/device_context.h"
  53 #include "gromacs/gpu_utils/device_stream.h"
  54 #include "gromacs/gpu_utils/gmxopencl.h"
  55 #include "gromacs/utility/exceptions.h"
  56 #include "gromacs/utility/gmxassert.h"
  57 #include "gromacs/utility/stringutil.h"
  58
  59 class GpuParallel3dFft::Impl
  60 {
  61 public:
  62     Impl(ivec                 realGridSize,
  63          ivec                 realGridSizePadded,
  64          ivec                 complexGridSizePadded,
  65          bool                 useDecomposition,
  66          bool                 performOutOfPlaceFFT,
  67          const DeviceContext& context,
  68          const DeviceStream&  pmeStream,
  69          DeviceBuffer<float>  realGrid,
  70          DeviceBuffer<float>  complexGrid);
  71     ~Impl();
  72
  73     clfftPlanHandle               planR2C_;
  74     clfftPlanHandle               planC2R_;
  75     std::vector<cl_command_queue> commandStreams_;
  76     cl_mem                        realGrid_;
  77     cl_mem                        complexGrid_;
  78 };
  79
  80 //! Throws the exception on clFFT error
  81 static void handleClfftError(clfftStatus status, const char* msg)
  82 {
  83     // Supposedly it's just a superset of standard OpenCL errors
  84     if (status != CLFFT_SUCCESS)
  85     {
  86         GMX_THROW(gmx::InternalError(gmx::formatString("%s: %d", msg, status)));
  87     }
  88 }
  89
  90 GpuParallel3dFft::Impl::Impl(ivec                 realGridSize,
  91                              ivec                 realGridSizePadded,
  92                              ivec                 complexGridSizePadded,
  93                              const bool           useDecomposition,
  94                              const bool           performOutOfPlaceFFT,
  95                              const DeviceContext& context,
  96                              const DeviceStream&  pmeStream,
  97                              DeviceBuffer<float>  realGrid,
  98                              DeviceBuffer<float>  complexGrid) :
  99     realGrid_(realGrid), complexGrid_(complexGrid)
 100 {
 101     GMX_RELEASE_ASSERT(!useDecomposition, "FFT decomposition not implemented");
 102
 103     cl_context clContext = context.context();
 104     commandStreams_.push_back(pmeStream.stream());
 105
 106     // clFFT expects row-major, so dimensions/strides are reversed (ZYX instead of XYZ)
 107     std::array<size_t, DIM> realGridDimensions = { size_t(realGridSize[ZZ]),
 108                                                    size_t(realGridSize[YY]),
 109                                                    size_t(realGridSize[XX]) };
 110     std::array<size_t, DIM> realGridStrides    = {
 111         1, size_t(realGridSizePadded[ZZ]), size_t(realGridSizePadded[YY] * realGridSizePadded[ZZ])
 112     };
 113     std::array<size_t, DIM> complexGridStrides = {
 114         1, size_t(complexGridSizePadded[ZZ]), size_t(complexGridSizePadded[YY] * complexGridSizePadded[ZZ])
 115     };
 116
 117     constexpr clfftDim dims = CLFFT_3D;
 118     handleClfftError(clfftCreateDefaultPlan(&planR2C_, clContext, dims, realGridDimensions.data()),
 119                      "clFFT planning failure");
 120     handleClfftError(clfftSetResultLocation(planR2C_, performOutOfPlaceFFT ? CLFFT_OUTOFPLACE : CLFFT_INPLACE),
 121                      "clFFT planning failure");
 122     handleClfftError(clfftSetPlanPrecision(planR2C_, CLFFT_SINGLE), "clFFT planning failure");
 123     constexpr cl_float scale = 1.0;
 124     handleClfftError(clfftSetPlanScale(planR2C_, CLFFT_FORWARD, scale),
 125                      "clFFT coefficient setup failure");
 126     handleClfftError(clfftSetPlanScale(planR2C_, CLFFT_BACKWARD, scale),
 127                      "clFFT coefficient setup failure");
 128
 129     // The only difference between 2 plans is direction
 130     handleClfftError(clfftCopyPlan(&planC2R_, clContext, planR2C_), "clFFT plan copying failure");
 131
 132     handleClfftError(clfftSetLayout(planR2C_, CLFFT_REAL, CLFFT_HERMITIAN_INTERLEAVED),
 133                      "clFFT R2C layout failure");
 134     handleClfftError(clfftSetLayout(planC2R_, CLFFT_HERMITIAN_INTERLEAVED, CLFFT_REAL),
 135                      "clFFT C2R layout failure");
 136
 137     handleClfftError(clfftSetPlanInStride(planR2C_, dims, realGridStrides.data()),
 138                      "clFFT stride setting failure");
 139     handleClfftError(clfftSetPlanOutStride(planR2C_, dims, complexGridStrides.data()),
 140                      "clFFT stride setting failure");
 141
 142     handleClfftError(clfftSetPlanInStride(planC2R_, dims, complexGridStrides.data()),
 143                      "clFFT stride setting failure");
 144     handleClfftError(clfftSetPlanOutStride(planC2R_, dims, realGridStrides.data()),
 145                      "clFFT stride setting failure");
 146
 147     handleClfftError(clfftBakePlan(planR2C_, commandStreams_.size(), commandStreams_.data(), nullptr, nullptr),
 148                      "clFFT precompiling failure");
 149     handleClfftError(clfftBakePlan(planC2R_, commandStreams_.size(), commandStreams_.data(), nullptr, nullptr),
 150                      "clFFT precompiling failure");
 151
 152     // TODO: implement solve kernel as R2C FFT callback
 153     // TODO: disable last transpose (clfftSetPlanTransposeResult)
 154 }
 155
 156 GpuParallel3dFft::Impl::~Impl()
 157 {
 158     clfftDestroyPlan(&planR2C_);
 159     clfftDestroyPlan(&planC2R_);
 160 }
 161
 162 void GpuParallel3dFft::perform3dFft(gmx_fft_direction dir, CommandEvent* timingEvent)
 163 {
 164     cl_mem                            tempBuffer = nullptr;
 165     constexpr std::array<cl_event, 0> waitEvents{ {} };
 166
 167     clfftPlanHandle plan;
 168     clfftDirection  direction;
 169     cl_mem *        inputGrids, *outputGrids;
 170
 171     switch (dir)
 172     {
 173         case GMX_FFT_REAL_TO_COMPLEX:
 174             plan        = impl_->planR2C_;
 175             direction   = CLFFT_FORWARD;
 176             inputGrids  = &impl_->realGrid_;
 177             outputGrids = &impl_->complexGrid_;
 178             break;
 179         case GMX_FFT_COMPLEX_TO_REAL:
 180             plan        = impl_->planC2R_;
 181             direction   = CLFFT_BACKWARD;
 182             inputGrids  = &impl_->complexGrid_;
 183             outputGrids = &impl_->realGrid_;
 184             break;
 185         default:
 186             GMX_THROW(
 187                     gmx::NotImplementedError("The chosen 3D-FFT case is not implemented on GPUs"));
 188     }
 189     handleClfftError(clfftEnqueueTransform(plan,
 190                                            direction,
 191                                            impl_->commandStreams_.size(),
 192                                            impl_->commandStreams_.data(),
 193                                            waitEvents.size(),
 194                                            waitEvents.data(),
 195                                            timingEvent,
 196                                            inputGrids,
 197                                            outputGrids,
 198                                            tempBuffer),
 199                      "clFFT execution failure");
 200 }
 201
 202 GpuParallel3dFft::GpuParallel3dFft(ivec                 realGridSize,
 203                                    ivec                 realGridSizePadded,
 204                                    ivec                 complexGridSizePadded,
 205                                    const bool           useDecomposition,
 206                                    const bool           performOutOfPlaceFFT,
 207                                    const DeviceContext& context,
 208                                    const DeviceStream&  pmeStream,
 209                                    DeviceBuffer<float>  realGrid,
 210                                    DeviceBuffer<float>  complexGrid) :
 211     impl_(std::make_unique<Impl>(realGridSize,
 212                                  realGridSizePadded,
 213                                  complexGridSizePadded,
 214                                  useDecomposition,
 215                                  performOutOfPlaceFFT,
 216                                  context,
 217                                  pmeStream,
 218                                  realGrid,
 219                                  complexGrid))
 220 {
 221 }
 222
 223 GpuParallel3dFft::~GpuParallel3dFft() = default;