src/gromacs/mdlib/nbnxn_gpu.h

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, 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  *  \brief Declare interface for GPU execution for NBNXN module
  37  *
  38  *  \author Szilard Pall <pall.szilard@gmail.com>
  39  *  \author Mark Abraham <mark.j.abraham@gmail.com>
  40  *  \ingroup module_mdlib
  41  */
  42
  43 #ifndef GMX_MDLIB_NBNXN_GPU_H
  44 #define GMX_MDLIB_NBNXN_GPU_H
  45
  46 #include "gromacs/gpu_utils/gpu_macros.h"
  47 #include "gromacs/math/vectypes.h"
  48 #include "gromacs/mdlib/nbnxn_gpu_types.h"
  49 #include "gromacs/utility/basedefinitions.h"
  50 #include "gromacs/utility/real.h"
  51
  52 struct nbnxn_atomdata_t;
  53 enum class GpuTaskCompletion;
  54
  55 /*! \brief
  56  * Launch asynchronously the nonbonded force calculations.
  57  *
  58  *  This consists of the following (async) steps launched:
  59  *  - upload x and q;
  60  *  - upload shift vector;
  61  *  - launch kernel;
  62  *  The local and non-local interaction calculations are launched in two
  63  *  separate streams.
  64  */
  65 GPU_FUNC_QUALIFIER
  66 void nbnxn_gpu_launch_kernel(gmx_nbnxn_gpu_t gmx_unused               *nb,
  67                              const struct nbnxn_atomdata_t gmx_unused *nbdata,
  68                              int gmx_unused                            flags,
  69                              int gmx_unused                            iloc) GPU_FUNC_TERM
  70
  71 /*! \brief
  72  * Launch asynchronously the nonbonded prune-only kernel.
  73  *
  74  *  The local and non-local list pruning are launched in their separate streams.
  75  *
  76  *  Notes for future scheduling tuning:
  77  *  Currently we schedule the dynamic pruning between two MD steps *after* both local and
  78  *  nonlocal force D2H transfers completed. We could launch already after the cpyback
  79  *  is launched, but we want to avoid prune kernels (especially in the non-local
  80  *  high prio-stream) competing with nonbonded work.
  81  *
  82  *  However, this is not ideal as this schedule does not expose the available
  83  *  concurrency. The dynamic pruning kernel:
  84  *    - should be allowed to overlap with any task other than force compute, including
  85  *      transfers (F D2H and the next step's x H2D as well as force clearing).
  86  *    - we'd prefer to avoid competition with non-bonded force kernels belonging
  87  *      to the same rank and ideally other ranks too.
  88  *
  89  *  In the most general case, the former would require scheduling pruning in a separate
  90  *  stream and adding additional event sync points to ensure that force kernels read
  91  *  consistent pair list data. This would lead to some overhead (due to extra
  92  *  cudaStreamWaitEvent calls, 3-5 us/call) which we might be able to live with.
  93  *  The gains from additional overlap might not be significant as long as
  94  *  update+constraints anyway takes longer than pruning, but there will still
  95  *  be use-cases where more overlap may help (e.g. multiple ranks per GPU,
  96  *  no/hbonds only constraints).
  97  *  The above second point is harder to address given that multiple ranks will often
  98  *  share a GPU. Ranks that complete their nonbondeds sooner can schedule pruning earlier
  99  *  and without a third priority level it is difficult to avoid some interference of
 100  *  prune kernels with force tasks (in particular preemption of low-prio local force task).
 101  *
 102  * \param [inout] nb        GPU nonbonded data.
 103  * \param [in]    iloc      Interaction locality flag.
 104  * \param [in]    numParts  Number of parts the pair list is split into in the rolling kernel.
 105  */
 106 GPU_FUNC_QUALIFIER
 107 void nbnxn_gpu_launch_kernel_pruneonly(gmx_nbnxn_gpu_t gmx_unused *nb,
 108                                        int gmx_unused              iloc,
 109                                        int gmx_unused              numParts) GPU_FUNC_TERM
 110
 111 /*! \brief
 112  * Launch asynchronously the download of nonbonded forces from the GPU
 113  * (and energies/shift forces if required).
 114  */
 115 GPU_FUNC_QUALIFIER
 116 void nbnxn_gpu_launch_cpyback(gmx_nbnxn_gpu_t  gmx_unused              *nb,
 117                               const struct nbnxn_atomdata_t gmx_unused *nbatom,
 118                               int                    gmx_unused         flags,
 119                               int                    gmx_unused         aloc) GPU_FUNC_TERM
 120
 121 /*! \brief Attempts to complete nonbonded GPU task.
 122  *
 123  *  This function attempts to complete the nonbonded task (both GPU and CPU auxiliary work).
 124  *  Success, i.e. that the tasks completed and results are ready to be consumed, is signaled
 125  *  by the return value (always true if blocking wait mode requested).
 126  *
 127  *  The \p completionKind parameter controls whether the behavior is non-blocking
 128  *  (achieved by passing GpuTaskCompletion::Check) or blocking wait until the results
 129  *  are ready (when GpuTaskCompletion::Wait is passed).
 130  *  As the "Check" mode the function will return immediately if the GPU stream
 131  *  still contain tasks that have not completed, it allows more flexible overlapping
 132  *  of work on the CPU with GPU execution.
 133  *
 134  *  Note that it is only safe to use the results, and to continue to the next MD
 135  *  step when this function has returned true which indicates successful completion of
 136  *  - All nonbonded GPU tasks: both compute and device transfer(s)
 137  *  - auxiliary tasks: updating the internal module state (timing accumulation, list pruning states) and
 138  *  - internal staging reduction of (\p fshift, \p e_el, \p e_lj).
 139  *
 140  *  TODO: improve the handling of outputs e.g. by ensuring that this function explcitly returns the
 141  *  force buffer (instead of that being passed only to nbnxn_gpu_launch_cpyback()) and by returning
 142  *  the energy and Fshift contributions for some external/centralized reduction.
 143  *
 144  * \param[in]  nb     The nonbonded data GPU structure
 145  * \param[in]  flags  Force flags
 146  * \param[in]  aloc   Atom locality identifier
 147  * \param[out] e_lj   Pointer to the LJ energy output to accumulate into
 148  * \param[out] e_el   Pointer to the electrostatics energy output to accumulate into
 149  * \param[out] fshift Pointer to the shift force buffer to accumulate into
 150  * \param[in]  completionKind Indicates whether nnbonded task completion should only be checked rather than waited for
 151  * \returns              True if the nonbonded tasks associated with \p aloc locality have completed
 152  */
 153 GPU_FUNC_QUALIFIER
 154 bool nbnxn_gpu_try_finish_task(gmx_nbnxn_gpu_t gmx_unused  *nb,
 155                                int             gmx_unused   flags,
 156                                int             gmx_unused   aloc,
 157                                real            gmx_unused  *e_lj,
 158                                real            gmx_unused  *e_el,
 159                                rvec            gmx_unused  *fshift,
 160                                GpuTaskCompletion gmx_unused completionKind) GPU_FUNC_TERM_WITH_RETURN(false)
 161
 162 /*! \brief  Completes the nonbonded GPU task blocking until GPU tasks and data
 163  * transfers to finish.
 164  *
 165  * Also does timing accounting and reduction of the internal staging buffers.
 166  * As this is called at the end of the step, it also resets the pair list and
 167  * pruning flags.
 168  *
 169  * \param[in] nb The nonbonded data GPU structure
 170  * \param[in] flags Force flags
 171  * \param[in] aloc Atom locality identifier
 172  * \param[out] e_lj Pointer to the LJ energy output to accumulate into
 173  * \param[out] e_el Pointer to the electrostatics energy output to accumulate into
 174  * \param[out] fshift Pointer to the shift force buffer to accumulate into
 175  */
 176 GPU_FUNC_QUALIFIER
 177 void nbnxn_gpu_wait_finish_task(gmx_nbnxn_gpu_t gmx_unused *nb,
 178                                 int             gmx_unused  flags,
 179                                 int             gmx_unused  aloc,
 180                                 real            gmx_unused *e_lj,
 181                                 real            gmx_unused *e_el,
 182                                 rvec            gmx_unused *fshift) GPU_FUNC_TERM
 183
 184 /*! \brief Selects the Ewald kernel type, analytical or tabulated, single or twin cut-off. */
 185 GPU_FUNC_QUALIFIER
 186 int nbnxn_gpu_pick_ewald_kernel_type(bool gmx_unused bTwinCut) GPU_FUNC_TERM_WITH_RETURN(-1)
 187
 188 #endif