src/gromacs/mdlib/update_constrain_gpu_impl.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 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 /*! \internal \file
  36  *
  37  * \brief Implements update and constraints class.
  38  *
  39  * The class combines Leap-Frog integrator with LINCS and SETTLE constraints.
  40  *
  41  * \todo The computational procedures in members should be integrated to improve
  42  *       computational performance.
  43  *
  44  * \author Artem Zhmurov <zhmurov@gmail.com>
  45  *
  46  * \ingroup module_mdlib
  47  */
  48 #include "gmxpre.h"
  49
  50 #include "update_constrain_gpu_impl.h"
  51
  52 #include <assert.h>
  53 #include <stdio.h>
  54
  55 #include <cmath>
  56
  57 #include <algorithm>
  58
  59 #include "gromacs/gpu_utils/device_context.h"
  60 #include "gromacs/gpu_utils/device_stream.h"
  61 #include "gromacs/gpu_utils/devicebuffer.h"
  62 #include "gromacs/gpu_utils/gpueventsynchronizer.h"
  63 #include "gromacs/gpu_utils/gputraits.h"
  64 #include "gromacs/mdlib/leapfrog_gpu.h"
  65 #include "gromacs/mdlib/update_constrain_gpu.h"
  66 #include "gromacs/mdlib/update_constrain_gpu_internal.h"
  67 #include "gromacs/mdtypes/mdatom.h"
  68 #include "gromacs/timing/wallcycle.h"
  69 #include "gromacs/topology/mtop_util.h"
  70
  71 static constexpr bool sc_haveGpuConstraintSupport = bool(GMX_GPU_CUDA) || bool(GMX_GPU_SYCL);
  72
  73 namespace gmx
  74 {
  75
  76 void UpdateConstrainGpu::Impl::integrate(GpuEventSynchronizer*             fReadyOnDevice,
  77                                          const real                        dt,
  78                                          const bool                        updateVelocities,
  79                                          const bool                        computeVirial,
  80                                          tensor                            virial,
  81                                          const bool                        doTemperatureScaling,
  82                                          gmx::ArrayRef<const t_grp_tcstat> tcstat,
  83                                          const bool                        doParrinelloRahman,
  84                                          const float                       dtPressureCouple,
  85                                          const matrix                      prVelocityScalingMatrix)
  86 {
  87     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
  88     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchGpuUpdateConstrain);
  89
  90     // Clearing virial matrix
  91     // TODO There is no point in having separate virial matrix for constraints
  92     clear_mat(virial);
  93
  94     // Make sure that the forces are ready on device before proceeding with the update.
  95     fReadyOnDevice->enqueueWaitEvent(deviceStream_);
  96
  97     // The integrate should save a copy of the current coordinates in d_xp_ and write updated
  98     // once into d_x_. The d_xp_ is only needed by constraints.
  99     integrator_->integrate(
 100             d_x_, d_xp_, d_v_, d_f_, dt, doTemperatureScaling, tcstat, doParrinelloRahman, dtPressureCouple, prVelocityScalingMatrix);
 101     // Constraints need both coordinates before (d_x_) and after (d_xp_) update. However, after constraints
 102     // are applied, the d_x_ can be discarded. So we intentionally swap the d_x_ and d_xp_ here to avoid the
 103     // d_xp_ -> d_x_ copy after constraints. Note that the integrate saves them in the wrong order as well.
 104     if (sc_haveGpuConstraintSupport)
 105     {
 106         lincsGpu_->apply(d_xp_, d_x_, updateVelocities, d_v_, 1.0 / dt, computeVirial, virial, pbcAiuc_);
 107         settleGpu_->apply(d_xp_, d_x_, updateVelocities, d_v_, 1.0 / dt, computeVirial, virial, pbcAiuc_);
 108     }
 109
 110     // scaledVirial -> virial (methods above returns scaled values)
 111     float scaleFactor = 0.5F / (dt * dt);
 112     for (int i = 0; i < DIM; i++)
 113     {
 114         for (int j = 0; j < DIM; j++)
 115         {
 116             virial[i][j] = scaleFactor * virial[i][j];
 117         }
 118     }
 119
 120     xUpdatedOnDeviceEvent_.markEvent(deviceStream_);
 121
 122     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchGpuUpdateConstrain);
 123     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 124 }
 125
 126 void UpdateConstrainGpu::Impl::scaleCoordinates(const matrix scalingMatrix)
 127 {
 128     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 129     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchGpuUpdateConstrain);
 130
 131     ScalingMatrix mu(scalingMatrix);
 132
 133     launchScaleCoordinatesKernel(numAtoms_, d_x_, mu, deviceStream_);
 134
 135     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchGpuUpdateConstrain);
 136     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 137 }
 138
 139 void UpdateConstrainGpu::Impl::scaleVelocities(const matrix scalingMatrix)
 140 {
 141     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 142     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchGpuUpdateConstrain);
 143
 144     ScalingMatrix mu(scalingMatrix);
 145
 146     launchScaleCoordinatesKernel(numAtoms_, d_v_, mu, deviceStream_);
 147
 148     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchGpuUpdateConstrain);
 149     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 150 }
 151
 152 UpdateConstrainGpu::Impl::Impl(const t_inputrec&    ir,
 153                                const gmx_mtop_t&    mtop,
 154                                const int            numTempScaleValues,
 155                                const DeviceContext& deviceContext,
 156                                const DeviceStream&  deviceStream,
 157                                gmx_wallcycle*       wcycle) :
 158     deviceContext_(deviceContext), deviceStream_(deviceStream), wcycle_(wcycle)
 159 {
 160     integrator_ = std::make_unique<LeapFrogGpu>(deviceContext_, deviceStream_, numTempScaleValues);
 161     if (sc_haveGpuConstraintSupport)
 162     {
 163         lincsGpu_ = std::make_unique<LincsGpu>(ir.nLincsIter, ir.nProjOrder, deviceContext_, deviceStream_);
 164         settleGpu_ = std::make_unique<SettleGpu>(mtop, deviceContext_, deviceStream_);
 165     }
 166 }
 167
 168 UpdateConstrainGpu::Impl::~Impl() {}
 169
 170 void UpdateConstrainGpu::Impl::set(DeviceBuffer<Float3>          d_x,
 171                                    DeviceBuffer<Float3>          d_v,
 172                                    const DeviceBuffer<Float3>    d_f,
 173                                    const InteractionDefinitions& idef,
 174                                    const t_mdatoms&              md)
 175 {
 176     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 177     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchGpuUpdateConstrain);
 178
 179     GMX_ASSERT(d_x, "Coordinates device buffer should not be null.");
 180     GMX_ASSERT(d_v, "Velocities device buffer should not be null.");
 181     GMX_ASSERT(d_f, "Forces device buffer should not be null.");
 182
 183     d_x_ = d_x;
 184     d_v_ = d_v;
 185     d_f_ = d_f;
 186
 187     numAtoms_ = md.nr;
 188
 189     reallocateDeviceBuffer(&d_xp_, numAtoms_, &numXp_, &numXpAlloc_, deviceContext_);
 190
 191     reallocateDeviceBuffer(
 192             &d_inverseMasses_, numAtoms_, &numInverseMasses_, &numInverseMassesAlloc_, deviceContext_);
 193
 194     // Integrator should also update something, but it does not even have a method yet
 195     integrator_->set(numAtoms_, md.invmass, md.cTC);
 196     if (sc_haveGpuConstraintSupport)
 197     {
 198         lincsGpu_->set(idef, numAtoms_, md.invmass);
 199         settleGpu_->set(idef);
 200     }
 201     else
 202     {
 203         GMX_ASSERT(idef.il[F_SETTLE].empty(), "SETTLE not supported");
 204         GMX_ASSERT(idef.il[F_CONSTR].empty(), "LINCS not supported");
 205     }
 206
 207     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchGpuUpdateConstrain);
 208     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 209 }
 210
 211 void UpdateConstrainGpu::Impl::setPbc(const PbcType pbcType, const matrix box)
 212 {
 213     // TODO wallcycle
 214     setPbcAiuc(numPbcDimensions(pbcType), box, &pbcAiuc_);
 215 }
 216
 217 GpuEventSynchronizer* UpdateConstrainGpu::Impl::xUpdatedOnDeviceEvent()
 218 {
 219     return &xUpdatedOnDeviceEvent_;
 220 }
 221
 222 UpdateConstrainGpu::UpdateConstrainGpu(const t_inputrec&    ir,
 223                                        const gmx_mtop_t&    mtop,
 224                                        const int            numTempScaleValues,
 225                                        const DeviceContext& deviceContext,
 226                                        const DeviceStream&  deviceStream,
 227                                        gmx_wallcycle*       wcycle) :
 228     impl_(new Impl(ir, mtop, numTempScaleValues, deviceContext, deviceStream, wcycle))
 229 {
 230 }
 231
 232 UpdateConstrainGpu::~UpdateConstrainGpu() = default;
 233
 234 void UpdateConstrainGpu::integrate(GpuEventSynchronizer*             fReadyOnDevice,
 235                                    const real                        dt,
 236                                    const bool                        updateVelocities,
 237                                    const bool                        computeVirial,
 238                                    tensor                            virialScaled,
 239                                    const bool                        doTemperatureScaling,
 240                                    gmx::ArrayRef<const t_grp_tcstat> tcstat,
 241                                    const bool                        doParrinelloRahman,
 242                                    const float                       dtPressureCouple,
 243                                    const matrix                      prVelocityScalingMatrix)
 244 {
 245     impl_->integrate(fReadyOnDevice,
 246                      dt,
 247                      updateVelocities,
 248                      computeVirial,
 249                      virialScaled,
 250                      doTemperatureScaling,
 251                      tcstat,
 252                      doParrinelloRahman,
 253                      dtPressureCouple,
 254                      prVelocityScalingMatrix);
 255 }
 256
 257 void UpdateConstrainGpu::scaleCoordinates(const matrix scalingMatrix)
 258 {
 259     impl_->scaleCoordinates(scalingMatrix);
 260 }
 261
 262 void UpdateConstrainGpu::scaleVelocities(const matrix scalingMatrix)
 263 {
 264     impl_->scaleVelocities(scalingMatrix);
 265 }
 266
 267 void UpdateConstrainGpu::set(DeviceBuffer<Float3>          d_x,
 268                              DeviceBuffer<Float3>          d_v,
 269                              const DeviceBuffer<Float3>    d_f,
 270                              const InteractionDefinitions& idef,
 271                              const t_mdatoms&              md)
 272 {
 273     impl_->set(d_x, d_v, d_f, idef, md);
 274 }
 275
 276 void UpdateConstrainGpu::setPbc(const PbcType pbcType, const matrix box)
 277 {
 278     impl_->setPbc(pbcType, box);
 279 }
 280
 281 GpuEventSynchronizer* UpdateConstrainGpu::xUpdatedOnDeviceEvent()
 282 {
 283     return impl_->xUpdatedOnDeviceEvent();
 284 }
 285
 286 bool UpdateConstrainGpu::isNumCoupledConstraintsSupported(const gmx_mtop_t& mtop)
 287 {
 288     return LincsGpu::isNumCoupledConstraintsSupported(mtop);
 289 }
 290
 291 bool UpdateConstrainGpu::areConstraintsSupported()
 292 {
 293     return sc_haveGpuConstraintSupport;
 294 }
 295
 296 } // namespace gmx