src/gromacs/fft/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 GPU 3D FFT routines for OpenCL.
  38  *
  39  *  \author Aleksei Iupinov <a.yupinov@gmail.com>
  40  *  \author Mark Abraham <mark.j.abraham@gmail.com>
  41  *  \ingroup module_fft
  42  */
  43
  44 #include "gmxpre.h"
  45
  46 #include "gpu_3dfft_ocl.h"
  47
  48 #include <array>
  49 #include <vector>
  50
  51 #include <clFFT.h>
  52
  53 #include "gromacs/gpu_utils/device_context.h"
  54 #include "gromacs/gpu_utils/device_stream.h"
  55 #include "gromacs/gpu_utils/gmxopencl.h"
  56 #include "gromacs/utility/exceptions.h"
  57 #include "gromacs/utility/gmxassert.h"
  58 #include "gromacs/utility/stringutil.h"
  59
  60 namespace gmx
  61 {
  62 //! Throws the exception on clFFT error
  63 static void handleClfftError(clfftStatus status, const char* msg)
  64 {
  65     // Supposedly it's just a superset of standard OpenCL errors
  66     if (status != CLFFT_SUCCESS)
  67     {
  68         GMX_THROW(InternalError(formatString("%s: %d", msg, status)));
  69     }
  70 }
  71
  72 Gpu3dFft::ImplOcl::ImplOcl(bool allocateGrids,
  73                            MPI_Comm /*comm*/,
  74                            ArrayRef<const int> gridSizesInXForEachRank,
  75                            ArrayRef<const int> gridSizesInYForEachRank,
  76                            const int /*nz*/,
  77                            bool                 performOutOfPlaceFFT,
  78                            const DeviceContext& context,
  79                            const DeviceStream&  pmeStream,
  80                            ivec                 realGridSize,
  81                            ivec                 realGridSizePadded,
  82                            ivec                 complexGridSizePadded,
  83                            DeviceBuffer<float>* realGrid,
  84                            DeviceBuffer<float>* complexGrid) :
  85     realGrid_(*realGrid), complexGrid_(*complexGrid)
  86 {
  87     GMX_RELEASE_ASSERT(allocateGrids == false, "Grids needs to be pre-allocated");
  88     GMX_RELEASE_ASSERT(gridSizesInXForEachRank.size() == 1 && gridSizesInYForEachRank.size() == 1,
  89                        "FFT decomposition not implemented with OpenCL backend");
  90
  91     cl_context clContext = context.context();
  92     commandStreams_.push_back(pmeStream.stream());
  93
  94     // clFFT expects row-major, so dimensions/strides are reversed (ZYX instead of XYZ)
  95     std::array<size_t, DIM> realGridDimensions = { size_t(realGridSize[ZZ]),
  96                                                    size_t(realGridSize[YY]),
  97                                                    size_t(realGridSize[XX]) };
  98     std::array<size_t, DIM> realGridStrides    = {
  99         1, size_t(realGridSizePadded[ZZ]), size_t(realGridSizePadded[YY] * realGridSizePadded[ZZ])
 100     };
 101     std::array<size_t, DIM> complexGridStrides = {
 102         1, size_t(complexGridSizePadded[ZZ]), size_t(complexGridSizePadded[YY] * complexGridSizePadded[ZZ])
 103     };
 104
 105     constexpr clfftDim dims = CLFFT_3D;
 106     handleClfftError(clfftCreateDefaultPlan(&planR2C_, clContext, dims, realGridDimensions.data()),
 107                      "clFFT planning failure");
 108     handleClfftError(clfftSetResultLocation(planR2C_, performOutOfPlaceFFT ? CLFFT_OUTOFPLACE : CLFFT_INPLACE),
 109                      "clFFT planning failure");
 110     handleClfftError(clfftSetPlanPrecision(planR2C_, CLFFT_SINGLE), "clFFT planning failure");
 111     constexpr cl_float scale = 1.0;
 112     handleClfftError(clfftSetPlanScale(planR2C_, CLFFT_FORWARD, scale),
 113                      "clFFT coefficient setup failure");
 114     handleClfftError(clfftSetPlanScale(planR2C_, CLFFT_BACKWARD, scale),
 115                      "clFFT coefficient setup failure");
 116
 117     // The only difference between 2 plans is direction
 118     handleClfftError(clfftCopyPlan(&planC2R_, clContext, planR2C_), "clFFT plan copying failure");
 119
 120     handleClfftError(clfftSetLayout(planR2C_, CLFFT_REAL, CLFFT_HERMITIAN_INTERLEAVED),
 121                      "clFFT R2C layout failure");
 122     handleClfftError(clfftSetLayout(planC2R_, CLFFT_HERMITIAN_INTERLEAVED, CLFFT_REAL),
 123                      "clFFT C2R layout failure");
 124
 125     handleClfftError(clfftSetPlanInStride(planR2C_, dims, realGridStrides.data()),
 126                      "clFFT stride setting failure");
 127     handleClfftError(clfftSetPlanOutStride(planR2C_, dims, complexGridStrides.data()),
 128                      "clFFT stride setting failure");
 129
 130     handleClfftError(clfftSetPlanInStride(planC2R_, dims, complexGridStrides.data()),
 131                      "clFFT stride setting failure");
 132     handleClfftError(clfftSetPlanOutStride(planC2R_, dims, realGridStrides.data()),
 133                      "clFFT stride setting failure");
 134
 135     handleClfftError(clfftBakePlan(planR2C_, commandStreams_.size(), commandStreams_.data(), nullptr, nullptr),
 136                      "clFFT precompiling failure");
 137     handleClfftError(clfftBakePlan(planC2R_, commandStreams_.size(), commandStreams_.data(), nullptr, nullptr),
 138                      "clFFT precompiling failure");
 139
 140     // TODO: implement solve kernel as R2C FFT callback
 141     // TODO: disable last transpose (clfftSetPlanTransposeResult)
 142 }
 143
 144 Gpu3dFft::ImplOcl::~ImplOcl()
 145 {
 146     clfftDestroyPlan(&planR2C_);
 147     clfftDestroyPlan(&planC2R_);
 148 }
 149
 150 void Gpu3dFft::ImplOcl::perform3dFft(gmx_fft_direction dir, CommandEvent* timingEvent)
 151 {
 152     cl_mem                            tempBuffer = nullptr;
 153     constexpr std::array<cl_event, 0> waitEvents{ {} };
 154
 155     clfftPlanHandle plan;
 156     clfftDirection  direction;
 157     cl_mem *        inputGrids, *outputGrids;
 158
 159     switch (dir)
 160     {
 161         case GMX_FFT_REAL_TO_COMPLEX:
 162             plan        = planR2C_;
 163             direction   = CLFFT_FORWARD;
 164             inputGrids  = &realGrid_;
 165             outputGrids = &complexGrid_;
 166             break;
 167         case GMX_FFT_COMPLEX_TO_REAL:
 168             plan        = planC2R_;
 169             direction   = CLFFT_BACKWARD;
 170             inputGrids  = &complexGrid_;
 171             outputGrids = &realGrid_;
 172             break;
 173         default:
 174             GMX_THROW(NotImplementedError("The chosen 3D-FFT case is not implemented on GPUs"));
 175     }
 176     handleClfftError(clfftEnqueueTransform(plan,
 177                                            direction,
 178                                            commandStreams_.size(),
 179                                            commandStreams_.data(),
 180                                            waitEvents.size(),
 181                                            waitEvents.data(),
 182                                            timingEvent,
 183                                            inputGrids,
 184                                            outputGrids,
 185                                            tempBuffer),
 186                      "clFFT execution failure");
 187 }
 188
 189 } // namespace gmx