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, 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/gputraits.cuh"
  64 #include "gromacs/gpu_utils/vectype_ops.cuh"
  65 #include "gromacs/mdlib/leapfrog_gpu.h"
  66 #include "gromacs/mdlib/lincs_gpu.cuh"
  67 #include "gromacs/mdlib/settle_gpu.cuh"
  68 #include "gromacs/mdlib/update_constrain_gpu.h"
  69 #include "gromacs/mdtypes/mdatom.h"
  70 #include "gromacs/timing/wallcycle.h"
  71
  72 namespace gmx
  73 {
  74 /*!\brief Number of CUDA threads in a block
  75  *
  76  * \todo Check if using smaller block size will lead to better performance.
  77  */
  78 constexpr static int c_threadsPerBlock = 256;
  79 //! Maximum number of threads in a block (for __launch_bounds__)
  80 constexpr static int c_maxThreadsPerBlock = c_threadsPerBlock;
  81
  82 /*! \brief Scaling matrix struct.
  83  *
  84  * \todo Should be generalized.
  85  */
  86 struct ScalingMatrix
  87 {
  88     float xx, yy, zz, yx, zx, zy;
  89 };
  90
  91 __launch_bounds__(c_maxThreadsPerBlock) __global__
  92         static void scaleCoordinates_kernel(const int numAtoms,
  93                                             float3* __restrict__ gm_x,
  94                                             const ScalingMatrix scalingMatrix)
  95 {
  96     int threadIndex = blockIdx.x * blockDim.x + threadIdx.x;
  97     if (threadIndex < numAtoms)
  98     {
  99         float3 x = gm_x[threadIndex];
 100
 101         x.x = scalingMatrix.xx * x.x + scalingMatrix.yx * x.y + scalingMatrix.zx * x.z;
 102         x.y = scalingMatrix.yy * x.y + scalingMatrix.zy * x.z;
 103         x.z = scalingMatrix.zz * x.z;
 104
 105         gm_x[threadIndex] = x;
 106     }
 107 }
 108
 109 void UpdateConstrainGpu::Impl::integrate(GpuEventSynchronizer*             fReadyOnDevice,
 110                                          const real                        dt,
 111                                          const bool                        updateVelocities,
 112                                          const bool                        computeVirial,
 113                                          tensor                            virial,
 114                                          const bool                        doTemperatureScaling,
 115                                          gmx::ArrayRef<const t_grp_tcstat> tcstat,
 116                                          const bool                        doParrinelloRahman,
 117                                          const float                       dtPressureCouple,
 118                                          const matrix                      prVelocityScalingMatrix)
 119 {
 120     wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
 121     wallcycle_sub_start(wcycle_, ewcsLAUNCH_GPU_UPDATE_CONSTRAIN);
 122
 123     // Clearing virial matrix
 124     // TODO There is no point in having separate virial matrix for constraints
 125     clear_mat(virial);
 126
 127     // Make sure that the forces are ready on device before proceeding with the update.
 128     fReadyOnDevice->enqueueWaitEvent(deviceStream_);
 129
 130     // The integrate should save a copy of the current coordinates in d_xp_ and write updated
 131     // once into d_x_. The d_xp_ is only needed by constraints.
 132     integrator_->integrate(
 133             d_x_, d_xp_, d_v_, d_f_, dt, doTemperatureScaling, tcstat, doParrinelloRahman, dtPressureCouple, prVelocityScalingMatrix);
 134     // Constraints need both coordinates before (d_x_) and after (d_xp_) update. However, after constraints
 135     // are applied, the d_x_ can be discarded. So we intentionally swap the d_x_ and d_xp_ here to avoid the
 136     // d_xp_ -> d_x_ copy after constraints. Note that the integrate saves them in the wrong order as well.
 137     lincsGpu_->apply(d_xp_, d_x_, updateVelocities, d_v_, 1.0 / dt, computeVirial, virial, pbcAiuc_);
 138     settleGpu_->apply(d_xp_, d_x_, updateVelocities, d_v_, 1.0 / dt, computeVirial, virial, pbcAiuc_);
 139
 140     // scaledVirial -> virial (methods above returns scaled values)
 141     float scaleFactor = 0.5f / (dt * dt);
 142     for (int i = 0; i < DIM; i++)
 143     {
 144         for (int j = 0; j < DIM; j++)
 145         {
 146             virial[i][j] = scaleFactor * virial[i][j];
 147         }
 148     }
 149
 150     coordinatesReady_->markEvent(deviceStream_);
 151
 152     wallcycle_sub_stop(wcycle_, ewcsLAUNCH_GPU_UPDATE_CONSTRAIN);
 153     wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
 154
 155     return;
 156 }
 157
 158 void UpdateConstrainGpu::Impl::scaleCoordinates(const matrix scalingMatrix)
 159 {
 160     wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
 161     wallcycle_sub_start(wcycle_, ewcsLAUNCH_GPU_UPDATE_CONSTRAIN);
 162
 163     ScalingMatrix mu;
 164     mu.xx = scalingMatrix[XX][XX];
 165     mu.yy = scalingMatrix[YY][YY];
 166     mu.zz = scalingMatrix[ZZ][ZZ];
 167     mu.yx = scalingMatrix[YY][XX];
 168     mu.zx = scalingMatrix[ZZ][XX];
 169     mu.zy = scalingMatrix[ZZ][YY];
 170
 171     const auto kernelArgs = prepareGpuKernelArguments(
 172             scaleCoordinates_kernel, coordinateScalingKernelLaunchConfig_, &numAtoms_, &d_x_, &mu);
 173     launchGpuKernel(scaleCoordinates_kernel,
 174                     coordinateScalingKernelLaunchConfig_,
 175                     deviceStream_,
 176                     nullptr,
 177                     "scaleCoordinates_kernel",
 178                     kernelArgs);
 179     // TODO: Although this only happens on the pressure coupling steps, this synchronization
 180     //       can affect the performance if nstpcouple is small.
 181     deviceStream_.synchronize();
 182
 183     wallcycle_sub_stop(wcycle_, ewcsLAUNCH_GPU_UPDATE_CONSTRAIN);
 184     wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
 185 }
 186
 187 void UpdateConstrainGpu::Impl::scaleVelocities(const matrix scalingMatrix)
 188 {
 189     wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
 190     wallcycle_sub_start(wcycle_, ewcsLAUNCH_GPU_UPDATE_CONSTRAIN);
 191
 192     ScalingMatrix mu;
 193     mu.xx = scalingMatrix[XX][XX];
 194     mu.yy = scalingMatrix[YY][YY];
 195     mu.zz = scalingMatrix[ZZ][ZZ];
 196     mu.yx = scalingMatrix[YY][XX];
 197     mu.zx = scalingMatrix[ZZ][XX];
 198     mu.zy = scalingMatrix[ZZ][YY];
 199
 200     const auto kernelArgs = prepareGpuKernelArguments(
 201             scaleCoordinates_kernel, coordinateScalingKernelLaunchConfig_, &numAtoms_, &d_v_, &mu);
 202     launchGpuKernel(scaleCoordinates_kernel,
 203                     coordinateScalingKernelLaunchConfig_,
 204                     deviceStream_,
 205                     nullptr,
 206                     "scaleCoordinates_kernel",
 207                     kernelArgs);
 208     // TODO: Although this only happens on the pressure coupling steps, this synchronization
 209     //       can affect the performance if nstpcouple is small.
 210     deviceStream_.synchronize();
 211
 212     wallcycle_sub_stop(wcycle_, ewcsLAUNCH_GPU_UPDATE_CONSTRAIN);
 213     wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
 214 }
 215
 216 UpdateConstrainGpu::Impl::Impl(const t_inputrec&     ir,
 217                                const gmx_mtop_t&     mtop,
 218                                const DeviceContext&  deviceContext,
 219                                const DeviceStream&   deviceStream,
 220                                GpuEventSynchronizer* xUpdatedOnDevice,
 221                                gmx_wallcycle*        wcycle) :
 222     deviceContext_(deviceContext),
 223     deviceStream_(deviceStream),
 224     coordinatesReady_(xUpdatedOnDevice),
 225     wcycle_(wcycle)
 226 {
 227     GMX_ASSERT(xUpdatedOnDevice != nullptr, "The event synchronizer can not be nullptr.");
 228
 229
 230     integrator_ = std::make_unique<LeapFrogGpu>(deviceContext_, deviceStream_);
 231     lincsGpu_ = std::make_unique<LincsGpu>(ir.nLincsIter, ir.nProjOrder, deviceContext_, deviceStream_);
 232     settleGpu_ = std::make_unique<SettleGpu>(mtop, deviceContext_, deviceStream_);
 233
 234     coordinateScalingKernelLaunchConfig_.blockSize[0]     = c_threadsPerBlock;
 235     coordinateScalingKernelLaunchConfig_.blockSize[1]     = 1;
 236     coordinateScalingKernelLaunchConfig_.blockSize[2]     = 1;
 237     coordinateScalingKernelLaunchConfig_.sharedMemorySize = 0;
 238 }
 239
 240 UpdateConstrainGpu::Impl::~Impl() {}
 241
 242 void UpdateConstrainGpu::Impl::set(DeviceBuffer<RVec>            d_x,
 243                                    DeviceBuffer<RVec>            d_v,
 244                                    const DeviceBuffer<RVec>      d_f,
 245                                    const InteractionDefinitions& idef,
 246                                    const t_mdatoms&              md,
 247                                    const int                     numTempScaleValues)
 248 {
 249     // TODO wallcycle
 250     wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
 251     wallcycle_sub_start(wcycle_, ewcsLAUNCH_GPU_UPDATE_CONSTRAIN);
 252
 253     GMX_ASSERT(d_x != nullptr, "Coordinates device buffer should not be null.");
 254     GMX_ASSERT(d_v != nullptr, "Velocities device buffer should not be null.");
 255     GMX_ASSERT(d_f != nullptr, "Forces device buffer should not be null.");
 256
 257     d_x_ = reinterpret_cast<float3*>(d_x);
 258     d_v_ = reinterpret_cast<float3*>(d_v);
 259     d_f_ = reinterpret_cast<float3*>(d_f);
 260
 261     numAtoms_ = md.nr;
 262
 263     reallocateDeviceBuffer(&d_xp_, numAtoms_, &numXp_, &numXpAlloc_, deviceContext_);
 264
 265     reallocateDeviceBuffer(
 266             &d_inverseMasses_, numAtoms_, &numInverseMasses_, &numInverseMassesAlloc_, deviceContext_);
 267
 268     // Integrator should also update something, but it does not even have a method yet
 269     integrator_->set(numAtoms_, md.invmass, numTempScaleValues, md.cTC);
 270     lincsGpu_->set(idef, numAtoms_, md.invmass);
 271     settleGpu_->set(idef);
 272
 273     coordinateScalingKernelLaunchConfig_.gridSize[0] =
 274             (numAtoms_ + c_threadsPerBlock - 1) / c_threadsPerBlock;
 275
 276     wallcycle_sub_stop(wcycle_, ewcsLAUNCH_GPU_UPDATE_CONSTRAIN);
 277     wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
 278 }
 279
 280 void UpdateConstrainGpu::Impl::setPbc(const PbcType pbcType, const matrix box)
 281 {
 282     // TODO wallcycle
 283     setPbcAiuc(numPbcDimensions(pbcType), box, &pbcAiuc_);
 284 }
 285
 286 GpuEventSynchronizer* UpdateConstrainGpu::Impl::getCoordinatesReadySync()
 287 {
 288     return coordinatesReady_;
 289 }
 290
 291 UpdateConstrainGpu::UpdateConstrainGpu(const t_inputrec&     ir,
 292                                        const gmx_mtop_t&     mtop,
 293                                        const DeviceContext&  deviceContext,
 294                                        const DeviceStream&   deviceStream,
 295                                        GpuEventSynchronizer* xUpdatedOnDevice,
 296                                        gmx_wallcycle*        wcycle) :
 297     impl_(new Impl(ir, mtop, deviceContext, deviceStream, xUpdatedOnDevice, wcycle))
 298 {
 299 }
 300
 301 UpdateConstrainGpu::~UpdateConstrainGpu() = default;
 302
 303 void UpdateConstrainGpu::integrate(GpuEventSynchronizer*             fReadyOnDevice,
 304                                    const real                        dt,
 305                                    const bool                        updateVelocities,
 306                                    const bool                        computeVirial,
 307                                    tensor                            virialScaled,
 308                                    const bool                        doTemperatureScaling,
 309                                    gmx::ArrayRef<const t_grp_tcstat> tcstat,
 310                                    const bool                        doParrinelloRahman,
 311                                    const float                       dtPressureCouple,
 312                                    const matrix                      prVelocityScalingMatrix)
 313 {
 314     impl_->integrate(fReadyOnDevice,
 315                      dt,
 316                      updateVelocities,
 317                      computeVirial,
 318                      virialScaled,
 319                      doTemperatureScaling,
 320                      tcstat,
 321                      doParrinelloRahman,
 322                      dtPressureCouple,
 323                      prVelocityScalingMatrix);
 324 }
 325
 326 void UpdateConstrainGpu::scaleCoordinates(const matrix scalingMatrix)
 327 {
 328     impl_->scaleCoordinates(scalingMatrix);
 329 }
 330
 331 void UpdateConstrainGpu::scaleVelocities(const matrix scalingMatrix)
 332 {
 333     impl_->scaleVelocities(scalingMatrix);
 334 }
 335
 336 void UpdateConstrainGpu::set(DeviceBuffer<RVec>            d_x,
 337                              DeviceBuffer<RVec>            d_v,
 338                              const DeviceBuffer<RVec>      d_f,
 339                              const InteractionDefinitions& idef,
 340                              const t_mdatoms&              md,
 341                              const int                     numTempScaleValues)
 342 {
 343     impl_->set(d_x, d_v, d_f, idef, md, numTempScaleValues);
 344 }
 345
 346 void UpdateConstrainGpu::setPbc(const PbcType pbcType, const matrix box)
 347 {
 348     impl_->setPbc(pbcType, box);
 349 }
 350
 351 GpuEventSynchronizer* UpdateConstrainGpu::getCoordinatesReadySync()
 352 {
 353     return impl_->getCoordinatesReadySync();
 354 }
 355
 356 bool UpdateConstrainGpu::isNumCoupledConstraintsSupported(const gmx_mtop_t& mtop)
 357 {
 358     return LincsGpu::isNumCoupledConstraintsSupported(mtop);
 359 }
 360
 361 } // namespace gmx