src/gromacs/nbnxm/opencl/nbnxm_ocl_types.h

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2016 by the GROMACS development team.
   5  * Copyright (c) 2017,2018,2019,2020, by the GROMACS development team, led by
   6  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   7  * and including many others, as listed in the AUTHORS file in the
   8  * top-level source directory and at http://www.gromacs.org.
   9  *
  10  * GROMACS is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public License
  12  * as published by the Free Software Foundation; either version 2.1
  13  * of the License, or (at your option) any later version.
  14  *
  15  * GROMACS is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * Lesser General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU Lesser General Public
  21  * License along with GROMACS; if not, see
  22  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  23  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  24  *
  25  * If you want to redistribute modifications to GROMACS, please
  26  * consider that scientific software is very special. Version
  27  * control is crucial - bugs must be traceable. We will be happy to
  28  * consider code for inclusion in the official distribution, but
  29  * derived work must not be called official GROMACS. Details are found
  30  * in the README & COPYING files - if they are missing, get the
  31  * official version at http://www.gromacs.org.
  32  *
  33  * To help us fund GROMACS development, we humbly ask that you cite
  34  * the research papers on the package. Check out http://www.gromacs.org.
  35  */
  36
  37 /*! \internal \file
  38  *  \brief
  39  *  Data types used internally in the nbnxm_ocl module.
  40  *
  41  *  \author Anca Hamuraru <anca@streamcomputing.eu>
  42  *  \author Szilárd Páll <pszilard@kth.se>
  43  *  \ingroup module_nbnxm
  44  */
  45
  46 #ifndef GMX_NBNXM_NBNXM_OPENCL_TYPES_H
  47 #define GMX_NBNXM_NBNXM_OPENCL_TYPES_H
  48
  49 #include "gromacs/gpu_utils/devicebuffer.h"
  50 #include "gromacs/gpu_utils/gmxopencl.h"
  51 #include "gromacs/gpu_utils/gputraits_ocl.h"
  52 #include "gromacs/gpu_utils/oclutils.h"
  53 #include "gromacs/mdtypes/interaction_const.h"
  54 #include "gromacs/nbnxm/gpu_types_common.h"
  55 #include "gromacs/nbnxm/nbnxm.h"
  56 #include "gromacs/nbnxm/nbnxm_gpu.h"
  57 #include "gromacs/nbnxm/pairlist.h"
  58 #include "gromacs/utility/enumerationhelpers.h"
  59 #include "gromacs/utility/fatalerror.h"
  60 #include "gromacs/utility/real.h"
  61
  62 #include "nbnxm_ocl_consts.h"
  63
  64 struct gmx_wallclock_gpu_nbnxn_t;
  65
  66 /* kernel does #include "gromacs/math/utilities.h" */
  67 /* Move the actual useful stuff here: */
  68
  69 //! Define 1/sqrt(pi)
  70 #define M_FLOAT_1_SQRTPI 0.564189583547756f
  71
  72 /*! @} */
  73 /*! \brief Constants for platform-dependent defaults for the prune kernel's j4 processing concurrency.
  74  *
  75  *  Initialized using macros that can be overridden at compile-time (using #GMX_NBNXN_PRUNE_KERNEL_J4_CONCURRENCY).
  76  */
  77 /*! @{ */
  78 const int c_oclPruneKernelJ4ConcurrencyDEFAULT = GMX_NBNXN_PRUNE_KERNEL_J4_CONCURRENCY_DEFAULT;
  79 /*! @} */
  80
  81 /*! \brief Pruning kernel flavors.
  82  *
  83  * The values correspond to the first call of the pruning post-list generation
  84  * and the rolling pruning, respectively.
  85  */
  86 enum ePruneKind
  87 {
  88     epruneFirst,
  89     epruneRolling,
  90     ePruneNR
  91 };
  92
  93 /*! \internal
  94  * \brief Staging area for temporary data downloaded from the GPU.
  95  *
  96  *  The energies/shift forces get downloaded here first, before getting added
  97  *  to the CPU-side aggregate values.
  98  */
  99 struct nb_staging_t
 100 {
 101     //! LJ energy
 102     float* e_lj = nullptr;
 103     //! electrostatic energy
 104     float* e_el = nullptr;
 105     //! float3 buffer with shift forces
 106     float (*fshift)[3] = nullptr;
 107 };
 108
 109 /*! \internal
 110  * \brief Nonbonded atom data - both inputs and outputs.
 111  */
 112 typedef struct cl_atomdata
 113 {
 114     //! number of atoms
 115     int natoms;
 116     //! number of local atoms
 117     int natoms_local;
 118     //! allocation size for the atom data (xq, f)
 119     int nalloc;
 120
 121     //! float4 buffer with atom coordinates + charges, size natoms
 122     DeviceBuffer<float> xq;
 123
 124     //! float3 buffer with force output array, size natoms
 125     DeviceBuffer<float> f;
 126
 127     //! LJ energy output, size 1
 128     DeviceBuffer<float> e_lj;
 129     //! Electrostatics energy input, size 1
 130     DeviceBuffer<float> e_el;
 131
 132     //! float3 buffer with shift forces
 133     DeviceBuffer<float> fshift;
 134
 135     //! number of atom types
 136     int ntypes;
 137     //! int buffer with atom type indices, size natoms
 138     DeviceBuffer<int> atom_types;
 139     //! float2 buffer with sqrt(c6),sqrt(c12), size natoms
 140     DeviceBuffer<float> lj_comb;
 141
 142     //! float3 buffer with shifts values
 143     DeviceBuffer<float> shift_vec;
 144
 145     //! true if the shift vector has been uploaded
 146     bool bShiftVecUploaded;
 147 } cl_atomdata_t;
 148
 149 /*! \internal
 150  * \brief Data structure shared between the OpenCL device code and OpenCL host code
 151  *
 152  * Must not contain OpenCL objects (buffers)
 153  * TODO: review, improve */
 154 typedef struct cl_nbparam_params
 155 {
 156
 157     //! type of electrostatics, takes values from #eelType
 158     int eeltype;
 159     //! type of VdW impl., takes values from #evdwType
 160     int vdwtype;
 161
 162     //! charge multiplication factor
 163     float epsfac;
 164     //! Reaction-field/plain cutoff electrostatics const.
 165     float c_rf;
 166     //! Reaction-field electrostatics constant
 167     float two_k_rf;
 168     //! Ewald/PME parameter
 169     float ewald_beta;
 170     //! Ewald/PME correction term substracted from the direct-space potential
 171     float sh_ewald;
 172     //! LJ-Ewald/PME correction term added to the correction potential
 173     float sh_lj_ewald;
 174     //! LJ-Ewald/PME coefficient
 175     float ewaldcoeff_lj;
 176
 177     //! Coulomb cut-off squared
 178     float rcoulomb_sq;
 179
 180     //! VdW cut-off squared
 181     float rvdw_sq;
 182     //! VdW switched cut-off
 183     float rvdw_switch;
 184     //! Full, outer pair-list cut-off squared
 185     float rlistOuter_sq;
 186     //! Inner, dynamic pruned pair-list cut-off squared
 187     float rlistInner_sq;
 188
 189     //! VdW shift dispersion constants
 190     shift_consts_t dispersion_shift;
 191     //! VdW shift repulsion constants
 192     shift_consts_t repulsion_shift;
 193     //! VdW switch constants
 194     switch_consts_t vdw_switch;
 195
 196     /* Ewald Coulomb force table data - accessed through texture memory */
 197     //! table scale/spacing
 198     float coulomb_tab_scale;
 199 } cl_nbparam_params_t;
 200
 201
 202 /*! \internal
 203  * \brief Pair list data.
 204  */
 205 using cl_plist_t = Nbnxm::gpu_plist;
 206
 207 /** \internal
 208  * \brief Typedef of actual timer type.
 209  */
 210 typedef struct Nbnxm::gpu_timers_t cl_timers_t;
 211
 212 /*! \internal
 213  * \brief Main data structure for OpenCL nonbonded force calculations.
 214  */
 215 struct NbnxmGpu
 216 {
 217     /* \brief OpenCL device context
 218      *
 219      * \todo Make it constant reference, once NbnxmGpu is a proper class.
 220      */
 221     const DeviceContext* deviceContext_;
 222     //! OpenCL runtime data (context, kernels)
 223     struct gmx_device_runtime_data_t* dev_rundata = nullptr;
 224
 225     /**< Pointers to non-bonded kernel functions
 226      * organized similar with nb_kfunc_xxx arrays in nbnxn_ocl.cpp */
 227     ///@{
 228     cl_kernel kernel_noener_noprune_ptr[eelTypeNR][evdwTypeNR] = { { nullptr } };
 229     cl_kernel kernel_ener_noprune_ptr[eelTypeNR][evdwTypeNR]   = { { nullptr } };
 230     cl_kernel kernel_noener_prune_ptr[eelTypeNR][evdwTypeNR]   = { { nullptr } };
 231     cl_kernel kernel_ener_prune_ptr[eelTypeNR][evdwTypeNR]     = { { nullptr } };
 232     ///@}
 233     //! prune kernels, ePruneKind defined the kernel kinds
 234     cl_kernel kernel_pruneonly[ePruneNR] = { nullptr };
 235
 236     //! true if prefetching fg i-atom LJ parameters should be used in the kernels
 237     bool bPrefetchLjParam = false;
 238
 239     /**< auxiliary kernels implementing memset-like functions */
 240     ///@{
 241     cl_kernel kernel_memset_f      = nullptr;
 242     cl_kernel kernel_memset_f2     = nullptr;
 243     cl_kernel kernel_memset_f3     = nullptr;
 244     cl_kernel kernel_zero_e_fshift = nullptr;
 245     ///@}
 246
 247     //! true if doing both local/non-local NB work on GPU
 248     bool bUseTwoStreams = false;
 249     //! true indicates that the nonlocal_done event was enqueued
 250     bool bNonLocalStreamActive = false;
 251
 252     //! atom data
 253     cl_atomdata_t* atdat = nullptr;
 254     //! parameters required for the non-bonded calc.
 255     NBParamGpu* nbparam = nullptr;
 256     //! pair-list data structures (local and non-local)
 257     gmx::EnumerationArray<Nbnxm::InteractionLocality, cl_plist_t*> plist = { nullptr };
 258     //! staging area where fshift/energies get downloaded
 259     nb_staging_t nbst;
 260
 261     //! local and non-local GPU queues
 262     gmx::EnumerationArray<Nbnxm::InteractionLocality, const DeviceStream*> deviceStreams;
 263
 264     /*! \brief Events used for synchronization */
 265     /*! \{ */
 266     /*! \brief Event triggered when the non-local non-bonded
 267      * kernel is done (and the local transfer can proceed) */
 268     cl_event nonlocal_done = nullptr;
 269     /*! \brief Event triggered when the tasks issued in the local
 270      * stream that need to precede the non-local force or buffer
 271      * operation calculations are done (e.g. f buffer 0-ing, local
 272      * x/q H2D, buffer op initialization in local stream that is
 273      * required also by nonlocal stream ) */
 274     cl_event misc_ops_and_local_H2D_done = nullptr;
 275     /*! \} */
 276
 277     //! True if there has been local/nonlocal GPU work, either bonded or nonbonded, scheduled
 278     //  to be executed in the current domain. As long as bonded work is not split up into
 279     //  local/nonlocal, if there is bonded GPU work, both flags will be true.
 280     gmx::EnumerationArray<Nbnxm::InteractionLocality, bool> haveWork;
 281
 282
 283     //! True if event-based timing is enabled.
 284     bool bDoTime = false;
 285     //! OpenCL event-based timers.
 286     cl_timers_t* timers = nullptr;
 287     //! Timing data. TODO: deprecate this and query timers for accumulated data instead
 288     gmx_wallclock_gpu_nbnxn_t* timings = nullptr;
 289 };
 290
 291 #endif /* NBNXN_OPENCL_TYPES_H */