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 #if GMX_GPU_CUDA
  63 #    include "gromacs/gpu_utils/gpueventsynchronizer.cuh"
  64 #elif GMX_GPU_SYCL
  65 #    include "gromacs/gpu_utils/gpueventsynchronizer_sycl.h"
  66 #endif
  67 #include "gromacs/gpu_utils/gputraits.h"
  68 #include "gromacs/mdlib/leapfrog_gpu.h"
  69 #include "gromacs/mdlib/update_constrain_gpu.h"
  70 #include "gromacs/mdlib/update_constrain_gpu_internal.h"
  71 #include "gromacs/mdtypes/mdatom.h"
  72 #include "gromacs/timing/wallcycle.h"
  73 #include "gromacs/topology/mtop_util.h"
  74
  75 static constexpr bool sc_haveGpuConstraintSupport = (GMX_GPU_CUDA);
  76
  77 namespace gmx
  78 {
  79
  80 void UpdateConstrainGpu::Impl::integrate(GpuEventSynchronizer*             fReadyOnDevice,
  81                                          const real                        dt,
  82                                          const bool                        updateVelocities,
  83                                          const bool                        computeVirial,
  84                                          tensor                            virial,
  85                                          const bool                        doTemperatureScaling,
  86                                          gmx::ArrayRef<const t_grp_tcstat> tcstat,
  87                                          const bool                        doParrinelloRahman,
  88                                          const float                       dtPressureCouple,
  89                                          const matrix                      prVelocityScalingMatrix)
  90 {
  91     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
  92     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchGpuUpdateConstrain);
  93
  94     // Clearing virial matrix
  95     // TODO There is no point in having separate virial matrix for constraints
  96     clear_mat(virial);
  97
  98     // Make sure that the forces are ready on device before proceeding with the update.
  99     fReadyOnDevice->enqueueWaitEvent(deviceStream_);
 100
 101     // The integrate should save a copy of the current coordinates in d_xp_ and write updated
 102     // once into d_x_. The d_xp_ is only needed by constraints.
 103     integrator_->integrate(
 104             d_x_, d_xp_, d_v_, d_f_, dt, doTemperatureScaling, tcstat, doParrinelloRahman, dtPressureCouple, prVelocityScalingMatrix);
 105     // Constraints need both coordinates before (d_x_) and after (d_xp_) update. However, after constraints
 106     // are applied, the d_x_ can be discarded. So we intentionally swap the d_x_ and d_xp_ here to avoid the
 107     // d_xp_ -> d_x_ copy after constraints. Note that the integrate saves them in the wrong order as well.
 108     if (sc_haveGpuConstraintSupport)
 109     {
 110         lincsGpu_->apply(d_xp_, d_x_, updateVelocities, d_v_, 1.0 / dt, computeVirial, virial, pbcAiuc_);
 111         settleGpu_->apply(d_xp_, d_x_, updateVelocities, d_v_, 1.0 / dt, computeVirial, virial, pbcAiuc_);
 112     }
 113
 114     // scaledVirial -> virial (methods above returns scaled values)
 115     float scaleFactor = 0.5F / (dt * dt);
 116     for (int i = 0; i < DIM; i++)
 117     {
 118         for (int j = 0; j < DIM; j++)
 119         {
 120             virial[i][j] = scaleFactor * virial[i][j];
 121         }
 122     }
 123
 124     coordinatesReady_->markEvent(deviceStream_);
 125
 126     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchGpuUpdateConstrain);
 127     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 128
 129     return;
 130 }
 131
 132 void UpdateConstrainGpu::Impl::scaleCoordinates(const matrix scalingMatrix)
 133 {
 134     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 135     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchGpuUpdateConstrain);
 136
 137     ScalingMatrix mu(scalingMatrix);
 138
 139     launchScaleCoordinatesKernel(numAtoms_, d_x_, mu, deviceStream_);
 140
 141     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchGpuUpdateConstrain);
 142     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 143 }
 144
 145 void UpdateConstrainGpu::Impl::scaleVelocities(const matrix scalingMatrix)
 146 {
 147     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 148     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchGpuUpdateConstrain);
 149
 150     ScalingMatrix mu(scalingMatrix);
 151
 152     launchScaleCoordinatesKernel(numAtoms_, d_v_, mu, deviceStream_);
 153
 154     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchGpuUpdateConstrain);
 155     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 156 }
 157
 158 UpdateConstrainGpu::Impl::Impl(const t_inputrec&     ir,
 159                                const gmx_mtop_t&     mtop,
 160                                const int             numTempScaleValues,
 161                                const DeviceContext&  deviceContext,
 162                                const DeviceStream&   deviceStream,
 163                                GpuEventSynchronizer* xUpdatedOnDevice,
 164                                gmx_wallcycle*        wcycle) :
 165     deviceContext_(deviceContext),
 166     deviceStream_(deviceStream),
 167     coordinatesReady_(xUpdatedOnDevice),
 168     wcycle_(wcycle)
 169 {
 170     GMX_ASSERT(xUpdatedOnDevice != nullptr, "The event synchronizer can not be nullptr.");
 171
 172     integrator_ = std::make_unique<LeapFrogGpu>(deviceContext_, deviceStream_, numTempScaleValues);
 173     if (sc_haveGpuConstraintSupport)
 174     {
 175         lincsGpu_ = std::make_unique<LincsGpu>(ir.nLincsIter, ir.nProjOrder, deviceContext_, deviceStream_);
 176         settleGpu_ = std::make_unique<SettleGpu>(mtop, deviceContext_, deviceStream_);
 177     }
 178 }
 179
 180 UpdateConstrainGpu::Impl::~Impl() {}
 181
 182 void UpdateConstrainGpu::Impl::set(DeviceBuffer<Float3>          d_x,
 183                                    DeviceBuffer<Float3>          d_v,
 184                                    const DeviceBuffer<Float3>    d_f,
 185                                    const InteractionDefinitions& idef,
 186                                    const t_mdatoms&              md)
 187 {
 188     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 189     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchGpuUpdateConstrain);
 190
 191     GMX_ASSERT(d_x, "Coordinates device buffer should not be null.");
 192     GMX_ASSERT(d_v, "Velocities device buffer should not be null.");
 193     GMX_ASSERT(d_f, "Forces device buffer should not be null.");
 194
 195     d_x_ = d_x;
 196     d_v_ = d_v;
 197     d_f_ = d_f;
 198
 199     numAtoms_ = md.nr;
 200
 201     reallocateDeviceBuffer(&d_xp_, numAtoms_, &numXp_, &numXpAlloc_, deviceContext_);
 202
 203     reallocateDeviceBuffer(
 204             &d_inverseMasses_, numAtoms_, &numInverseMasses_, &numInverseMassesAlloc_, deviceContext_);
 205
 206     // Integrator should also update something, but it does not even have a method yet
 207     integrator_->set(numAtoms_, md.invmass, md.cTC);
 208     if (sc_haveGpuConstraintSupport)
 209     {
 210         lincsGpu_->set(idef, numAtoms_, md.invmass);
 211         settleGpu_->set(idef);
 212     }
 213     else
 214     {
 215         GMX_ASSERT(idef.il[F_SETTLE].empty(), "SETTLE not supported");
 216         GMX_ASSERT(idef.il[F_CONSTR].empty(), "LINCS not supported");
 217     }
 218
 219     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchGpuUpdateConstrain);
 220     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 221 }
 222
 223 void UpdateConstrainGpu::Impl::setPbc(const PbcType pbcType, const matrix box)
 224 {
 225     // TODO wallcycle
 226     setPbcAiuc(numPbcDimensions(pbcType), box, &pbcAiuc_);
 227 }
 228
 229 GpuEventSynchronizer* UpdateConstrainGpu::Impl::getCoordinatesReadySync()
 230 {
 231     return coordinatesReady_;
 232 }
 233
 234 UpdateConstrainGpu::UpdateConstrainGpu(const t_inputrec&     ir,
 235                                        const gmx_mtop_t&     mtop,
 236                                        const int             numTempScaleValues,
 237                                        const DeviceContext&  deviceContext,
 238                                        const DeviceStream&   deviceStream,
 239                                        GpuEventSynchronizer* xUpdatedOnDevice,
 240                                        gmx_wallcycle*        wcycle) :
 241     impl_(new Impl(ir, mtop, numTempScaleValues, deviceContext, deviceStream, xUpdatedOnDevice, wcycle))
 242 {
 243 }
 244
 245 UpdateConstrainGpu::~UpdateConstrainGpu() = default;
 246
 247 void UpdateConstrainGpu::integrate(GpuEventSynchronizer*             fReadyOnDevice,
 248                                    const real                        dt,
 249                                    const bool                        updateVelocities,
 250                                    const bool                        computeVirial,
 251                                    tensor                            virialScaled,
 252                                    const bool                        doTemperatureScaling,
 253                                    gmx::ArrayRef<const t_grp_tcstat> tcstat,
 254                                    const bool                        doParrinelloRahman,
 255                                    const float                       dtPressureCouple,
 256                                    const matrix                      prVelocityScalingMatrix)
 257 {
 258     impl_->integrate(fReadyOnDevice,
 259                      dt,
 260                      updateVelocities,
 261                      computeVirial,
 262                      virialScaled,
 263                      doTemperatureScaling,
 264                      tcstat,
 265                      doParrinelloRahman,
 266                      dtPressureCouple,
 267                      prVelocityScalingMatrix);
 268 }
 269
 270 void UpdateConstrainGpu::scaleCoordinates(const matrix scalingMatrix)
 271 {
 272     impl_->scaleCoordinates(scalingMatrix);
 273 }
 274
 275 void UpdateConstrainGpu::scaleVelocities(const matrix scalingMatrix)
 276 {
 277     impl_->scaleVelocities(scalingMatrix);
 278 }
 279
 280 void UpdateConstrainGpu::set(DeviceBuffer<Float3>          d_x,
 281                              DeviceBuffer<Float3>          d_v,
 282                              const DeviceBuffer<Float3>    d_f,
 283                              const InteractionDefinitions& idef,
 284                              const t_mdatoms&              md)
 285 {
 286     impl_->set(d_x, d_v, d_f, idef, md);
 287 }
 288
 289 void UpdateConstrainGpu::setPbc(const PbcType pbcType, const matrix box)
 290 {
 291     impl_->setPbc(pbcType, box);
 292 }
 293
 294 GpuEventSynchronizer* UpdateConstrainGpu::getCoordinatesReadySync()
 295 {
 296     return impl_->getCoordinatesReadySync();
 297 }
 298
 299 bool UpdateConstrainGpu::isNumCoupledConstraintsSupported(const gmx_mtop_t& mtop)
 300 {
 301     return LincsGpu::isNumCoupledConstraintsSupported(mtop);
 302 }
 303
 304 bool UpdateConstrainGpu::areConstraintsSupported()
 305 {
 306     return sc_haveGpuConstraintSupport;
 307 }
 308
 309 } // namespace gmx