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, 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 <array>
  46
  47 #include "gromacs/utility/exceptions.h"
  48 #include "gromacs/utility/gmxassert.h"
  49 #include "gromacs/utility/stringutil.h"
  50
  51 #include "pme_gpu_3dfft.h"
  52 #include "pme_gpu_internal.h"
  53 #include "pme_gpu_types.h"
  54 #include "pme_gpu_types_host_impl.h"
  55
  56 //! Throws the exception on clFFT error
  57 static void handleClfftError(clfftStatus status, const char* msg)
  58 {
  59     // Supposedly it's just a superset of standard OpenCL errors
  60     if (status != CLFFT_SUCCESS)
  61     {
  62         GMX_THROW(gmx::InternalError(gmx::formatString("%s: %d", msg, status)));
  63     }
  64 }
  65
  66 GpuParallel3dFft::GpuParallel3dFft(const PmeGpu* pmeGpu)
  67 {
  68     // Extracting all the data from PME GPU
  69     std::array<size_t, DIM> realGridSize, realGridSizePadded, complexGridSizePadded;
  70
  71     GMX_RELEASE_ASSERT(!pme_gpu_settings(pmeGpu).useDecomposition,
  72                        "FFT decomposition not implemented");
  73     PmeGpuKernelParamsBase* kernelParamsPtr = pmeGpu->kernelParams.get();
  74     for (int i = 0; i < DIM; i++)
  75     {
  76         realGridSize[i]          = kernelParamsPtr->grid.realGridSize[i];
  77         realGridSizePadded[i]    = kernelParamsPtr->grid.realGridSizePadded[i];
  78         complexGridSizePadded[i] = kernelParamsPtr->grid.complexGridSizePadded[i];
  79         GMX_ASSERT(kernelParamsPtr->grid.complexGridSizePadded[i]
  80                            == kernelParamsPtr->grid.complexGridSize[i],
  81                    "Complex padding not implemented");
  82     }
  83     cl_context context = pmeGpu->archSpecific->deviceContext_.context();
  84     deviceStreams_.push_back(pmeGpu->archSpecific->pmeStream_.stream());
  85     realGrid_                       = kernelParamsPtr->grid.d_realGrid;
  86     complexGrid_                    = kernelParamsPtr->grid.d_fourierGrid;
  87     const bool performOutOfPlaceFFT = pmeGpu->archSpecific->performOutOfPlaceFFT;
  88
  89
  90     // clFFT expects row-major, so dimensions/strides are reversed (ZYX instead of XYZ)
  91     std::array<size_t, DIM> realGridDimensions = { realGridSize[ZZ], realGridSize[YY], realGridSize[XX] };
  92     std::array<size_t, DIM> realGridStrides    = { 1, realGridSizePadded[ZZ],
  93                                                 realGridSizePadded[YY] * realGridSizePadded[ZZ] };
  94     std::array<size_t, DIM> complexGridStrides = { 1, complexGridSizePadded[ZZ],
  95                                                    complexGridSizePadded[YY] * complexGridSizePadded[ZZ] };
  96
  97     constexpr clfftDim dims = CLFFT_3D;
  98     handleClfftError(clfftCreateDefaultPlan(&planR2C_, context, dims, realGridDimensions.data()),
  99                      "clFFT planning failure");
 100     handleClfftError(clfftSetResultLocation(planR2C_, performOutOfPlaceFFT ? CLFFT_OUTOFPLACE : CLFFT_INPLACE),
 101                      "clFFT planning failure");
 102     handleClfftError(clfftSetPlanPrecision(planR2C_, CLFFT_SINGLE), "clFFT planning failure");
 103     constexpr cl_float scale = 1.0;
 104     handleClfftError(clfftSetPlanScale(planR2C_, CLFFT_FORWARD, scale),
 105                      "clFFT coefficient setup failure");
 106     handleClfftError(clfftSetPlanScale(planR2C_, CLFFT_BACKWARD, scale),
 107                      "clFFT coefficient setup failure");
 108
 109     // The only difference between 2 plans is direction
 110     handleClfftError(clfftCopyPlan(&planC2R_, context, planR2C_), "clFFT plan copying failure");
 111
 112     handleClfftError(clfftSetLayout(planR2C_, CLFFT_REAL, CLFFT_HERMITIAN_INTERLEAVED),
 113                      "clFFT R2C layout failure");
 114     handleClfftError(clfftSetLayout(planC2R_, CLFFT_HERMITIAN_INTERLEAVED, CLFFT_REAL),
 115                      "clFFT C2R layout failure");
 116
 117     handleClfftError(clfftSetPlanInStride(planR2C_, dims, realGridStrides.data()),
 118                      "clFFT stride setting failure");
 119     handleClfftError(clfftSetPlanOutStride(planR2C_, dims, complexGridStrides.data()),
 120                      "clFFT stride setting failure");
 121
 122     handleClfftError(clfftSetPlanInStride(planC2R_, dims, complexGridStrides.data()),
 123                      "clFFT stride setting failure");
 124     handleClfftError(clfftSetPlanOutStride(planC2R_, dims, realGridStrides.data()),
 125                      "clFFT stride setting failure");
 126
 127     handleClfftError(clfftBakePlan(planR2C_, deviceStreams_.size(), deviceStreams_.data(), nullptr, nullptr),
 128                      "clFFT precompiling failure");
 129     handleClfftError(clfftBakePlan(planC2R_, deviceStreams_.size(), deviceStreams_.data(), nullptr, nullptr),
 130                      "clFFT precompiling failure");
 131
 132     // TODO: implement solve kernel as R2C FFT callback
 133     // TODO: disable last transpose (clfftSetPlanTransposeResult)
 134 }
 135
 136 GpuParallel3dFft::~GpuParallel3dFft()
 137 {
 138     clfftDestroyPlan(&planR2C_);
 139     clfftDestroyPlan(&planC2R_);
 140 }
 141
 142 void GpuParallel3dFft::perform3dFft(gmx_fft_direction dir, CommandEvent* timingEvent)
 143 {
 144     cl_mem                            tempBuffer = nullptr;
 145     constexpr std::array<cl_event, 0> waitEvents{ {} };
 146
 147     clfftPlanHandle plan;
 148     clfftDirection  direction;
 149     cl_mem *        inputGrids, *outputGrids;
 150
 151     switch (dir)
 152     {
 153         case GMX_FFT_REAL_TO_COMPLEX:
 154             plan        = planR2C_;
 155             direction   = CLFFT_FORWARD;
 156             inputGrids  = &realGrid_;
 157             outputGrids = &complexGrid_;
 158             break;
 159         case GMX_FFT_COMPLEX_TO_REAL:
 160             plan        = planC2R_;
 161             direction   = CLFFT_BACKWARD;
 162             inputGrids  = &complexGrid_;
 163             outputGrids = &realGrid_;
 164             break;
 165         default:
 166             GMX_THROW(
 167                     gmx::NotImplementedError("The chosen 3D-FFT case is not implemented on GPUs"));
 168     }
 169     handleClfftError(clfftEnqueueTransform(plan, direction, deviceStreams_.size(),
 170                                            deviceStreams_.data(), waitEvents.size(), waitEvents.data(),
 171                                            timingEvent, inputGrids, outputGrids, tempBuffer),
 172                      "clFFT execution failure");
 173 }