src/gromacs/mdlib/update_constrain_gpu_impl.cu

   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 using CUDA.
  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/cudautils.cuh"
  60 #include "gromacs/gpu_utils/device_context.h"
  61 #include "gromacs/gpu_utils/device_stream.h"
  62 #include "gromacs/gpu_utils/devicebuffer.h"
  63 #include "gromacs/gpu_utils/gpueventsynchronizer.cuh"
  64 #include "gromacs/gpu_utils/gputraits.cuh"
  65 #include "gromacs/gpu_utils/vectype_ops.cuh"
  66 #include "gromacs/mdlib/leapfrog_gpu.h"
  67 #include "gromacs/mdlib/lincs_gpu.h"
  68 #include "gromacs/mdlib/settle_gpu.cuh"
  69 #include "gromacs/mdlib/update_constrain_gpu.h"
  70 #include "gromacs/mdtypes/mdatom.h"
  71 #include "gromacs/timing/wallcycle.h"
  72
  73 namespace gmx
  74 {
  75 /*!\brief Number of CUDA threads in a block
  76  *
  77  * \todo Check if using smaller block size will lead to better performance.
  78  */
  79 constexpr static int c_threadsPerBlock = 256;
  80 //! Maximum number of threads in a block (for __launch_bounds__)
  81 constexpr static int c_maxThreadsPerBlock = c_threadsPerBlock;
  82
  83 __launch_bounds__(c_maxThreadsPerBlock) __global__
  84         static void scaleCoordinates_kernel(const int numAtoms,
  85                                             float3* __restrict__ gm_x,
  86                                             const ScalingMatrix scalingMatrix)
  87 {
  88     int threadIndex = blockIdx.x * blockDim.x + threadIdx.x;
  89     if (threadIndex < numAtoms)
  90     {
  91         float3 x = gm_x[threadIndex];
  92
  93         x.x = scalingMatrix.xx * x.x + scalingMatrix.yx * x.y + scalingMatrix.zx * x.z;
  94         x.y = scalingMatrix.yy * x.y + scalingMatrix.zy * x.z;
  95         x.z = scalingMatrix.zz * x.z;
  96
  97         gm_x[threadIndex] = x;
  98     }
  99 }
 100
 101 void UpdateConstrainGpu::Impl::integrate(GpuEventSynchronizer*             fReadyOnDevice,
 102                                          const real                        dt,
 103                                          const bool                        updateVelocities,
 104                                          const bool                        computeVirial,
 105                                          tensor                            virial,
 106                                          const bool                        doTemperatureScaling,
 107                                          gmx::ArrayRef<const t_grp_tcstat> tcstat,
 108                                          const bool                        doParrinelloRahman,
 109                                          const float                       dtPressureCouple,
 110                                          const matrix                      prVelocityScalingMatrix)
 111 {
 112     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 113     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchGpuUpdateConstrain);
 114
 115     // Clearing virial matrix
 116     // TODO There is no point in having separate virial matrix for constraints
 117     clear_mat(virial);
 118
 119     // Make sure that the forces are ready on device before proceeding with the update.
 120     fReadyOnDevice->enqueueWaitEvent(deviceStream_);
 121
 122     // The integrate should save a copy of the current coordinates in d_xp_ and write updated
 123     // once into d_x_. The d_xp_ is only needed by constraints.
 124     integrator_->integrate(
 125             d_x_, d_xp_, d_v_, d_f_, dt, doTemperatureScaling, tcstat, doParrinelloRahman, dtPressureCouple, prVelocityScalingMatrix);
 126     // Constraints need both coordinates before (d_x_) and after (d_xp_) update. However, after constraints
 127     // are applied, the d_x_ can be discarded. So we intentionally swap the d_x_ and d_xp_ here to avoid the
 128     // d_xp_ -> d_x_ copy after constraints. Note that the integrate saves them in the wrong order as well.
 129     lincsGpu_->apply(d_xp_, d_x_, updateVelocities, d_v_, 1.0 / dt, computeVirial, virial, pbcAiuc_);
 130     settleGpu_->apply(d_xp_, d_x_, updateVelocities, d_v_, 1.0 / dt, computeVirial, virial, pbcAiuc_);
 131
 132     // scaledVirial -> virial (methods above returns scaled values)
 133     float scaleFactor = 0.5f / (dt * dt);
 134     for (int i = 0; i < DIM; i++)
 135     {
 136         for (int j = 0; j < DIM; j++)
 137         {
 138             virial[i][j] = scaleFactor * virial[i][j];
 139         }
 140     }
 141
 142     coordinatesReady_->markEvent(deviceStream_);
 143
 144     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchGpuUpdateConstrain);
 145     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 146
 147     return;
 148 }
 149
 150 void UpdateConstrainGpu::Impl::scaleCoordinates(const matrix scalingMatrix)
 151 {
 152     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 153     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchGpuUpdateConstrain);
 154
 155     ScalingMatrix mu(scalingMatrix);
 156
 157     const auto kernelArgs = prepareGpuKernelArguments(
 158             scaleCoordinates_kernel, coordinateScalingKernelLaunchConfig_, &numAtoms_, &d_x_, &mu);
 159     launchGpuKernel(scaleCoordinates_kernel,
 160                     coordinateScalingKernelLaunchConfig_,
 161                     deviceStream_,
 162                     nullptr,
 163                     "scaleCoordinates_kernel",
 164                     kernelArgs);
 165     // TODO: Although this only happens on the pressure coupling steps, this synchronization
 166     //       can affect the performance if nstpcouple is small.
 167     deviceStream_.synchronize();
 168
 169     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchGpuUpdateConstrain);
 170     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 171 }
 172
 173 void UpdateConstrainGpu::Impl::scaleVelocities(const matrix scalingMatrix)
 174 {
 175     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 176     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchGpuUpdateConstrain);
 177
 178     ScalingMatrix mu(scalingMatrix);
 179
 180     const auto kernelArgs = prepareGpuKernelArguments(
 181             scaleCoordinates_kernel, coordinateScalingKernelLaunchConfig_, &numAtoms_, &d_v_, &mu);
 182     launchGpuKernel(scaleCoordinates_kernel,
 183                     coordinateScalingKernelLaunchConfig_,
 184                     deviceStream_,
 185                     nullptr,
 186                     "scaleCoordinates_kernel",
 187                     kernelArgs);
 188     // TODO: Although this only happens on the pressure coupling steps, this synchronization
 189     //       can affect the performance if nstpcouple is small.
 190     deviceStream_.synchronize();
 191
 192     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchGpuUpdateConstrain);
 193     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 194 }
 195
 196 UpdateConstrainGpu::Impl::Impl(const t_inputrec&     ir,
 197                                const gmx_mtop_t&     mtop,
 198                                const int             numTempScaleValues,
 199                                const DeviceContext&  deviceContext,
 200                                const DeviceStream&   deviceStream,
 201                                GpuEventSynchronizer* xUpdatedOnDevice,
 202                                gmx_wallcycle*        wcycle) :
 203     deviceContext_(deviceContext),
 204     deviceStream_(deviceStream),
 205     coordinatesReady_(xUpdatedOnDevice),
 206     wcycle_(wcycle)
 207 {
 208     GMX_ASSERT(xUpdatedOnDevice != nullptr, "The event synchronizer can not be nullptr.");
 209
 210
 211     integrator_ = std::make_unique<LeapFrogGpu>(deviceContext_, deviceStream_, numTempScaleValues);
 212     lincsGpu_ = std::make_unique<LincsGpu>(ir.nLincsIter, ir.nProjOrder, deviceContext_, deviceStream_);
 213     settleGpu_ = std::make_unique<SettleGpu>(mtop, deviceContext_, deviceStream_);
 214
 215     coordinateScalingKernelLaunchConfig_.blockSize[0]     = c_threadsPerBlock;
 216     coordinateScalingKernelLaunchConfig_.blockSize[1]     = 1;
 217     coordinateScalingKernelLaunchConfig_.blockSize[2]     = 1;
 218     coordinateScalingKernelLaunchConfig_.sharedMemorySize = 0;
 219 }
 220
 221 UpdateConstrainGpu::Impl::~Impl() {}
 222
 223 void UpdateConstrainGpu::Impl::set(DeviceBuffer<Float3>          d_x,
 224                                    DeviceBuffer<Float3>          d_v,
 225                                    const DeviceBuffer<Float3>    d_f,
 226                                    const InteractionDefinitions& idef,
 227                                    const t_mdatoms&              md)
 228 {
 229     // TODO wallcycle
 230     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 231     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchGpuUpdateConstrain);
 232
 233     GMX_ASSERT(d_x != nullptr, "Coordinates device buffer should not be null.");
 234     GMX_ASSERT(d_v != nullptr, "Velocities device buffer should not be null.");
 235     GMX_ASSERT(d_f != nullptr, "Forces device buffer should not be null.");
 236
 237     d_x_ = d_x;
 238     d_v_ = d_v;
 239     d_f_ = d_f;
 240
 241     numAtoms_ = md.nr;
 242
 243     reallocateDeviceBuffer(&d_xp_, numAtoms_, &numXp_, &numXpAlloc_, deviceContext_);
 244
 245     reallocateDeviceBuffer(
 246             &d_inverseMasses_, numAtoms_, &numInverseMasses_, &numInverseMassesAlloc_, deviceContext_);
 247
 248     // Integrator should also update something, but it does not even have a method yet
 249     integrator_->set(numAtoms_, md.invmass, md.cTC);
 250     lincsGpu_->set(idef, numAtoms_, md.invmass);
 251     settleGpu_->set(idef);
 252
 253     coordinateScalingKernelLaunchConfig_.gridSize[0] =
 254             (numAtoms_ + c_threadsPerBlock - 1) / c_threadsPerBlock;
 255
 256     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchGpuUpdateConstrain);
 257     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 258 }
 259
 260 void UpdateConstrainGpu::Impl::setPbc(const PbcType pbcType, const matrix box)
 261 {
 262     // TODO wallcycle
 263     setPbcAiuc(numPbcDimensions(pbcType), box, &pbcAiuc_);
 264 }
 265
 266 GpuEventSynchronizer* UpdateConstrainGpu::Impl::getCoordinatesReadySync()
 267 {
 268     return coordinatesReady_;
 269 }
 270
 271 UpdateConstrainGpu::UpdateConstrainGpu(const t_inputrec&     ir,
 272                                        const gmx_mtop_t&     mtop,
 273                                        const int             numTempScaleValues,
 274                                        const DeviceContext&  deviceContext,
 275                                        const DeviceStream&   deviceStream,
 276                                        GpuEventSynchronizer* xUpdatedOnDevice,
 277                                        gmx_wallcycle*        wcycle) :
 278     impl_(new Impl(ir, mtop, numTempScaleValues, deviceContext, deviceStream, xUpdatedOnDevice, wcycle))
 279 {
 280 }
 281
 282 UpdateConstrainGpu::~UpdateConstrainGpu() = default;
 283
 284 void UpdateConstrainGpu::integrate(GpuEventSynchronizer*             fReadyOnDevice,
 285                                    const real                        dt,
 286                                    const bool                        updateVelocities,
 287                                    const bool                        computeVirial,
 288                                    tensor                            virialScaled,
 289                                    const bool                        doTemperatureScaling,
 290                                    gmx::ArrayRef<const t_grp_tcstat> tcstat,
 291                                    const bool                        doParrinelloRahman,
 292                                    const float                       dtPressureCouple,
 293                                    const matrix                      prVelocityScalingMatrix)
 294 {
 295     impl_->integrate(fReadyOnDevice,
 296                      dt,
 297                      updateVelocities,
 298                      computeVirial,
 299                      virialScaled,
 300                      doTemperatureScaling,
 301                      tcstat,
 302                      doParrinelloRahman,
 303                      dtPressureCouple,
 304                      prVelocityScalingMatrix);
 305 }
 306
 307 void UpdateConstrainGpu::scaleCoordinates(const matrix scalingMatrix)
 308 {
 309     impl_->scaleCoordinates(scalingMatrix);
 310 }
 311
 312 void UpdateConstrainGpu::scaleVelocities(const matrix scalingMatrix)
 313 {
 314     impl_->scaleVelocities(scalingMatrix);
 315 }
 316
 317 void UpdateConstrainGpu::set(DeviceBuffer<Float3>          d_x,
 318                              DeviceBuffer<Float3>          d_v,
 319                              const DeviceBuffer<Float3>    d_f,
 320                              const InteractionDefinitions& idef,
 321                              const t_mdatoms&              md)
 322 {
 323     impl_->set(d_x, d_v, d_f, idef, md);
 324 }
 325
 326 void UpdateConstrainGpu::setPbc(const PbcType pbcType, const matrix box)
 327 {
 328     impl_->setPbc(pbcType, box);
 329 }
 330
 331 GpuEventSynchronizer* UpdateConstrainGpu::getCoordinatesReadySync()
 332 {
 333     return impl_->getCoordinatesReadySync();
 334 }
 335
 336 bool UpdateConstrainGpu::isNumCoupledConstraintsSupported(const gmx_mtop_t& mtop)
 337 {
 338     return LincsGpu::isNumCoupledConstraintsSupported(mtop);
 339 }
 340
 341 } // namespace gmx