src/gromacs/nbnxm/opencl/nbnxm_ocl.cpp

   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) 2017,2018,2019,2020,2021, 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 /*! \internal \file
  37  *  \brief Define OpenCL implementation of nbnxm_gpu.h
  38  *
  39  *  \author Anca Hamuraru <anca@streamcomputing.eu>
  40  *  \author Teemu Virolainen <teemu@streamcomputing.eu>
  41  *  \author Dimitrios Karkoulis <dimitris.karkoulis@gmail.com>
  42  *  \author Szilárd Páll <pall.szilard@gmail.com>
  43  *  \ingroup module_nbnxm
  44  *
  45  *  TODO (psz):
  46  *  - Add a static const cl_uint c_pruneKernelWorkDim / c_nbnxnKernelWorkDim = 3;
  47  *  - Rework the copying of OCL data structures done before every invocation of both
  48  *    nb and prune kernels (using fillin_ocl_structures); also consider at the same
  49  *    time calling clSetKernelArg only on the updated parameters (if tracking changed
  50  *    parameters is feasible);
  51  *  - Consider using the event_wait_list argument to clEnqueueNDRangeKernel to mark
  52  *    dependencies on the kernel launched: e.g. the non-local nb kernel's dependency
  53  *    on the misc_ops_and_local_H2D_done event could be better expressed this way.
  54  *
  55  *  - Consider extracting common sections of the OpenCL and CUDA nbnxn logic, e.g:
  56  *    - in nbnxn_gpu_launch_kernel_pruneonly() the pre- and post-kernel launch logic
  57  *      is identical in the two implementations, so a 3-way split might allow sharing
  58  *      code;
  59  *    -
  60  *
  61  */
  62 #include "gmxpre.h"
  63
  64 #include <assert.h>
  65 #include <stdlib.h>
  66
  67 #if defined(_MSVC)
  68 #    include <limits>
  69 #endif
  70
  71 #include "gromacs/gpu_utils/device_context.h"
  72 #include "gromacs/gpu_utils/gputraits_ocl.h"
  73 #include "gromacs/gpu_utils/oclutils.h"
  74 #include "gromacs/hardware/device_information.h"
  75 #include "gromacs/hardware/hw_info.h"
  76 #include "gromacs/mdtypes/simulation_workload.h"
  77 #include "gromacs/nbnxm/atomdata.h"
  78 #include "gromacs/nbnxm/gpu_common.h"
  79 #include "gromacs/nbnxm/gpu_common_utils.h"
  80 #include "gromacs/nbnxm/gpu_data_mgmt.h"
  81 #include "gromacs/nbnxm/nbnxm.h"
  82 #include "gromacs/nbnxm/nbnxm_gpu.h"
  83 #include "gromacs/nbnxm/pairlist.h"
  84 #include "gromacs/pbcutil/ishift.h"
  85 #include "gromacs/timing/gpu_timing.h"
  86 #include "gromacs/utility/cstringutil.h"
  87 #include "gromacs/utility/fatalerror.h"
  88 #include "gromacs/utility/gmxassert.h"
  89
  90 #include "nbnxm_ocl_types.h"
  91
  92 namespace Nbnxm
  93 {
  94
  95 /*! \brief Convenience constants */
  96 //@{
  97 static constexpr int c_clSize = c_nbnxnGpuClusterSize;
  98 //@}
  99
 100
 101 /*! \brief Validates the input global work size parameter.
 102  */
 103 static inline void validate_global_work_size(const KernelLaunchConfig& config,
 104                                              int                       work_dim,
 105                                              const DeviceInformation*  dinfo)
 106 {
 107     cl_uint device_size_t_size_bits;
 108     cl_uint host_size_t_size_bits;
 109
 110     GMX_ASSERT(dinfo, "Need a valid device info object");
 111
 112     size_t global_work_size[3];
 113     GMX_ASSERT(work_dim <= 3, "Not supporting hyper-grids just yet");
 114     for (int i = 0; i < work_dim; i++)
 115     {
 116         global_work_size[i] = config.blockSize[i] * config.gridSize[i];
 117     }
 118
 119     /* Each component of a global_work_size must not exceed the range given by the
 120        sizeof(device size_t) for the device on which the kernel execution will
 121        be enqueued. See:
 122        https://www.khronos.org/registry/cl/sdk/1.0/docs/man/xhtml/clEnqueueNDRangeKernel.html
 123      */
 124     device_size_t_size_bits = dinfo->adress_bits;
 125     host_size_t_size_bits   = static_cast<cl_uint>(sizeof(size_t) * 8);
 126
 127     /* If sizeof(host size_t) <= sizeof(device size_t)
 128             => global_work_size components will always be valid
 129        else
 130             => get device limit for global work size and
 131             compare it against each component of global_work_size.
 132      */
 133     if (host_size_t_size_bits > device_size_t_size_bits)
 134     {
 135         size_t device_limit;
 136
 137         device_limit = (1ULL << device_size_t_size_bits) - 1;
 138
 139         for (int i = 0; i < work_dim; i++)
 140         {
 141             if (global_work_size[i] > device_limit)
 142             {
 143                 gmx_fatal(
 144                         FARGS,
 145                         "Watch out, the input system is too large to simulate!\n"
 146                         "The number of nonbonded work units (=number of super-clusters) exceeds the"
 147                         "device capabilities. Global work size limit exceeded (%zu > %zu)!",
 148                         global_work_size[i],
 149                         device_limit);
 150             }
 151         }
 152     }
 153 }
 154
 155 /* Constant arrays listing non-bonded kernel function names. The arrays are
 156  * organized in 2-dim arrays by: electrostatics and VDW type.
 157  *
 158  *  Note that the row- and column-order of function pointers has to match the
 159  *  order of corresponding enumerated electrostatics and vdw types, resp.,
 160  *  defined in nbnxm_ocl_types.h.
 161  */
 162
 163 /*! \brief Force-only kernel function names. */
 164 static const char* nb_kfunc_noener_noprune_ptr[c_numElecTypes][c_numVdwTypes] = {
 165     { "nbnxn_kernel_ElecCut_VdwLJ_F_opencl",
 166       "nbnxn_kernel_ElecCut_VdwLJCombGeom_F_opencl",
 167       "nbnxn_kernel_ElecCut_VdwLJCombLB_F_opencl",
 168       "nbnxn_kernel_ElecCut_VdwLJFsw_F_opencl",
 169       "nbnxn_kernel_ElecCut_VdwLJPsw_F_opencl",
 170       "nbnxn_kernel_ElecCut_VdwLJEwCombGeom_F_opencl",
 171       "nbnxn_kernel_ElecCut_VdwLJEwCombLB_F_opencl" },
 172     { "nbnxn_kernel_ElecRF_VdwLJ_F_opencl",
 173       "nbnxn_kernel_ElecRF_VdwLJCombGeom_F_opencl",
 174       "nbnxn_kernel_ElecRF_VdwLJCombLB_F_opencl",
 175       "nbnxn_kernel_ElecRF_VdwLJFsw_F_opencl",
 176       "nbnxn_kernel_ElecRF_VdwLJPsw_F_opencl",
 177       "nbnxn_kernel_ElecRF_VdwLJEwCombGeom_F_opencl",
 178       "nbnxn_kernel_ElecRF_VdwLJEwCombLB_F_opencl" },
 179     { "nbnxn_kernel_ElecEwQSTab_VdwLJ_F_opencl",
 180       "nbnxn_kernel_ElecEwQSTab_VdwLJCombGeom_F_opencl",
 181       "nbnxn_kernel_ElecEwQSTab_VdwLJCombLB_F_opencl",
 182       "nbnxn_kernel_ElecEwQSTab_VdwLJFsw_F_opencl",
 183       "nbnxn_kernel_ElecEwQSTab_VdwLJPsw_F_opencl",
 184       "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_F_opencl",
 185       "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_F_opencl" },
 186     { "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_F_opencl",
 187       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombGeom_F_opencl",
 188       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombLB_F_opencl",
 189       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_F_opencl",
 190       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_F_opencl",
 191       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_F_opencl",
 192       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_F_opencl" },
 193     { "nbnxn_kernel_ElecEw_VdwLJ_F_opencl",
 194       "nbnxn_kernel_ElecEw_VdwLJCombGeom_F_opencl",
 195       "nbnxn_kernel_ElecEw_VdwLJCombLB_F_opencl",
 196       "nbnxn_kernel_ElecEw_VdwLJFsw_F_opencl",
 197       "nbnxn_kernel_ElecEw_VdwLJPsw_F_opencl",
 198       "nbnxn_kernel_ElecEw_VdwLJEwCombGeom_F_opencl",
 199       "nbnxn_kernel_ElecEw_VdwLJEwCombLB_F_opencl" },
 200     { "nbnxn_kernel_ElecEwTwinCut_VdwLJ_F_opencl",
 201       "nbnxn_kernel_ElecEwTwinCut_VdwLJCombGeom_F_opencl",
 202       "nbnxn_kernel_ElecEwTwinCut_VdwLJCombLB_F_opencl",
 203       "nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_F_opencl",
 204       "nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_F_opencl",
 205       "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_F_opencl",
 206       "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_F_opencl" }
 207 };
 208
 209 /*! \brief Force + energy kernel function pointers. */
 210 static const char* nb_kfunc_ener_noprune_ptr[c_numElecTypes][c_numVdwTypes] = {
 211     { "nbnxn_kernel_ElecCut_VdwLJ_VF_opencl",
 212       "nbnxn_kernel_ElecCut_VdwLJCombGeom_VF_opencl",
 213       "nbnxn_kernel_ElecCut_VdwLJCombLB_VF_opencl",
 214       "nbnxn_kernel_ElecCut_VdwLJFsw_VF_opencl",
 215       "nbnxn_kernel_ElecCut_VdwLJPsw_VF_opencl",
 216       "nbnxn_kernel_ElecCut_VdwLJEwCombGeom_VF_opencl",
 217       "nbnxn_kernel_ElecCut_VdwLJEwCombLB_VF_opencl" },
 218     { "nbnxn_kernel_ElecRF_VdwLJ_VF_opencl",
 219       "nbnxn_kernel_ElecRF_VdwLJCombGeom_VF_opencl",
 220       "nbnxn_kernel_ElecRF_VdwLJCombLB_VF_opencl",
 221       "nbnxn_kernel_ElecRF_VdwLJFsw_VF_opencl",
 222       "nbnxn_kernel_ElecRF_VdwLJPsw_VF_opencl",
 223       "nbnxn_kernel_ElecRF_VdwLJEwCombGeom_VF_opencl",
 224       "nbnxn_kernel_ElecRF_VdwLJEwCombLB_VF_opencl" },
 225     { "nbnxn_kernel_ElecEwQSTab_VdwLJ_VF_opencl",
 226       "nbnxn_kernel_ElecEwQSTab_VdwLJCombGeom_VF_opencl",
 227       "nbnxn_kernel_ElecEwQSTab_VdwLJCombLB_VF_opencl",
 228       "nbnxn_kernel_ElecEwQSTab_VdwLJFsw_VF_opencl",
 229       "nbnxn_kernel_ElecEwQSTab_VdwLJPsw_VF_opencl",
 230       "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_VF_opencl",
 231       "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_VF_opencl" },
 232     { "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_VF_opencl",
 233       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombGeom_VF_opencl",
 234       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombLB_VF_opencl",
 235       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_VF_opencl",
 236       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_VF_opencl",
 237       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_VF_opencl",
 238       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_VF_opencl" },
 239     { "nbnxn_kernel_ElecEw_VdwLJ_VF_opencl",
 240       "nbnxn_kernel_ElecEw_VdwLJCombGeom_VF_opencl",
 241       "nbnxn_kernel_ElecEw_VdwLJCombLB_VF_opencl",
 242       "nbnxn_kernel_ElecEw_VdwLJFsw_VF_opencl",
 243       "nbnxn_kernel_ElecEw_VdwLJPsw_VF_opencl",
 244       "nbnxn_kernel_ElecEw_VdwLJEwCombGeom_VF_opencl",
 245       "nbnxn_kernel_ElecEw_VdwLJEwCombLB_VF_opencl" },
 246     { "nbnxn_kernel_ElecEwTwinCut_VdwLJ_VF_opencl",
 247       "nbnxn_kernel_ElecEwTwinCut_VdwLJCombGeom_VF_opencl",
 248       "nbnxn_kernel_ElecEwTwinCut_VdwLJCombLB_VF_opencl",
 249       "nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_VF_opencl",
 250       "nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_VF_opencl",
 251       "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF_opencl",
 252       "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_VF_opencl" }
 253 };
 254
 255 /*! \brief Force + pruning kernel function pointers. */
 256 static const char* nb_kfunc_noener_prune_ptr[c_numElecTypes][c_numVdwTypes] = {
 257     { "nbnxn_kernel_ElecCut_VdwLJ_F_prune_opencl",
 258       "nbnxn_kernel_ElecCut_VdwLJCombGeom_F_prune_opencl",
 259       "nbnxn_kernel_ElecCut_VdwLJCombLB_F_prune_opencl",
 260       "nbnxn_kernel_ElecCut_VdwLJFsw_F_prune_opencl",
 261       "nbnxn_kernel_ElecCut_VdwLJPsw_F_prune_opencl",
 262       "nbnxn_kernel_ElecCut_VdwLJEwCombGeom_F_prune_opencl",
 263       "nbnxn_kernel_ElecCut_VdwLJEwCombLB_F_prune_opencl" },
 264     { "nbnxn_kernel_ElecRF_VdwLJ_F_prune_opencl",
 265       "nbnxn_kernel_ElecRF_VdwLJCombGeom_F_prune_opencl",
 266       "nbnxn_kernel_ElecRF_VdwLJCombLB_F_prune_opencl",
 267       "nbnxn_kernel_ElecRF_VdwLJFsw_F_prune_opencl",
 268       "nbnxn_kernel_ElecRF_VdwLJPsw_F_prune_opencl",
 269       "nbnxn_kernel_ElecRF_VdwLJEwCombGeom_F_prune_opencl",
 270       "nbnxn_kernel_ElecRF_VdwLJEwCombLB_F_prune_opencl" },
 271     { "nbnxn_kernel_ElecEwQSTab_VdwLJ_F_prune_opencl",
 272       "nbnxn_kernel_ElecEwQSTab_VdwLJCombGeom_F_prune_opencl",
 273       "nbnxn_kernel_ElecEwQSTab_VdwLJCombLB_F_prune_opencl",
 274       "nbnxn_kernel_ElecEwQSTab_VdwLJFsw_F_prune_opencl",
 275       "nbnxn_kernel_ElecEwQSTab_VdwLJPsw_F_prune_opencl",
 276       "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_F_prune_opencl",
 277       "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_F_prune_opencl" },
 278     { "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_F_prune_opencl",
 279       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombGeom_F_prune_opencl",
 280       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombLB_F_prune_opencl",
 281       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_F_prune_opencl",
 282       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_F_prune_opencl",
 283       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_F_prune_opencl",
 284       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_F_prune_opencl" },
 285     { "nbnxn_kernel_ElecEw_VdwLJ_F_prune_opencl",
 286       "nbnxn_kernel_ElecEw_VdwLJCombGeom_F_prune_opencl",
 287       "nbnxn_kernel_ElecEw_VdwLJCombLB_F_prune_opencl",
 288       "nbnxn_kernel_ElecEw_VdwLJFsw_F_prune_opencl",
 289       "nbnxn_kernel_ElecEw_VdwLJPsw_F_prune_opencl",
 290       "nbnxn_kernel_ElecEw_VdwLJEwCombGeom_F_prune_opencl",
 291       "nbnxn_kernel_ElecEw_VdwLJEwCombLB_F_prune_opencl" },
 292     { "nbnxn_kernel_ElecEwTwinCut_VdwLJ_F_prune_opencl",
 293       "nbnxn_kernel_ElecEwTwinCut_VdwLJCombGeom_F_prune_opencl",
 294       "nbnxn_kernel_ElecEwTwinCut_VdwLJCombLB_F_prune_opencl",
 295       "nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_F_prune_opencl",
 296       "nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_F_prune_opencl",
 297       "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_F_prune_opencl",
 298       "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_F_prune_opencl" }
 299 };
 300
 301 /*! \brief Force + energy + pruning kernel function pointers. */
 302 static const char* nb_kfunc_ener_prune_ptr[c_numElecTypes][c_numVdwTypes] = {
 303     { "nbnxn_kernel_ElecCut_VdwLJ_VF_prune_opencl",
 304       "nbnxn_kernel_ElecCut_VdwLJCombGeom_VF_prune_opencl",
 305       "nbnxn_kernel_ElecCut_VdwLJCombLB_VF_prune_opencl",
 306       "nbnxn_kernel_ElecCut_VdwLJFsw_VF_prune_opencl",
 307       "nbnxn_kernel_ElecCut_VdwLJPsw_VF_prune_opencl",
 308       "nbnxn_kernel_ElecCut_VdwLJEwCombGeom_VF_prune_opencl",
 309       "nbnxn_kernel_ElecCut_VdwLJEwCombLB_VF_prune_opencl" },
 310     { "nbnxn_kernel_ElecRF_VdwLJ_VF_prune_opencl",
 311       "nbnxn_kernel_ElecRF_VdwLJCombGeom_VF_prune_opencl",
 312       "nbnxn_kernel_ElecRF_VdwLJCombLB_VF_prune_opencl",
 313       "nbnxn_kernel_ElecRF_VdwLJFsw_VF_prune_opencl",
 314       "nbnxn_kernel_ElecRF_VdwLJPsw_VF_prune_opencl",
 315       "nbnxn_kernel_ElecRF_VdwLJEwCombGeom_VF_prune_opencl",
 316       "nbnxn_kernel_ElecRF_VdwLJEwCombLB_VF_prune_opencl" },
 317     { "nbnxn_kernel_ElecEwQSTab_VdwLJ_VF_prune_opencl",
 318       "nbnxn_kernel_ElecEwQSTab_VdwLJCombGeom_VF_prune_opencl",
 319       "nbnxn_kernel_ElecEwQSTab_VdwLJCombLB_VF_prune_opencl",
 320       "nbnxn_kernel_ElecEwQSTab_VdwLJFsw_VF_prune_opencl",
 321       "nbnxn_kernel_ElecEwQSTab_VdwLJPsw_VF_prune_opencl",
 322       "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_VF_prune_opencl",
 323       "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_VF_prune_opencl" },
 324     { "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_VF_prune_opencl",
 325       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombGeom_VF_prune_opencl",
 326       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombLB_VF_prune_opencl",
 327       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_VF_prune_opencl",
 328       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_VF_prune_opencl",
 329       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_VF_prune_opencl",
 330       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_VF_prune_opencl" },
 331     { "nbnxn_kernel_ElecEw_VdwLJ_VF_prune_opencl",
 332       "nbnxn_kernel_ElecEw_VdwLJCombGeom_VF_prune_opencl",
 333       "nbnxn_kernel_ElecEw_VdwLJCombLB_VF_prune_opencl",
 334       "nbnxn_kernel_ElecEw_VdwLJFsw_VF_prune_opencl",
 335       "nbnxn_kernel_ElecEw_VdwLJPsw_VF_prune_opencl",
 336       "nbnxn_kernel_ElecEw_VdwLJEwCombGeom_VF_prune_opencl",
 337       "nbnxn_kernel_ElecEw_VdwLJEwCombLB_VF_prune_opencl" },
 338     { "nbnxn_kernel_ElecEwTwinCut_VdwLJ_VF_prune_opencl",
 339       "nbnxn_kernel_ElecEwTwinCut_VdwLJCombGeom_VF_prune_opencl",
 340       "nbnxn_kernel_ElecEwTwinCut_VdwLJCombLB_VF_prune_opencl",
 341       "nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_VF_prune_opencl",
 342       "nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_VF_prune_opencl",
 343       "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF_prune_opencl",
 344       "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_VF_prune_opencl" }
 345 };
 346
 347 /*! \brief Return a pointer to the prune kernel version to be executed at the current invocation.
 348  *
 349  * \param[in] kernel_pruneonly  array of prune kernel objects
 350  * \param[in] firstPrunePass    true if the first pruning pass is being executed
 351  */
 352 static inline cl_kernel selectPruneKernel(cl_kernel kernel_pruneonly[], bool firstPrunePass)
 353 {
 354     cl_kernel* kernelPtr;
 355
 356     if (firstPrunePass)
 357     {
 358         kernelPtr = &(kernel_pruneonly[epruneFirst]);
 359     }
 360     else
 361     {
 362         kernelPtr = &(kernel_pruneonly[epruneRolling]);
 363     }
 364     // TODO: consider creating the prune kernel object here to avoid a
 365     // clCreateKernel for the rolling prune kernel if this is not needed.
 366     return *kernelPtr;
 367 }
 368
 369 /*! \brief Return a pointer to the kernel version to be executed at the current step.
 370  *  OpenCL kernel objects are cached in nb. If the requested kernel is not
 371  *  found in the cache, it will be created and the cache will be updated.
 372  */
 373 static inline cl_kernel
 374 select_nbnxn_kernel(NbnxmGpu* nb, enum ElecType elecType, enum VdwType vdwType, bool bDoEne, bool bDoPrune)
 375 {
 376     const char* kernel_name_to_run;
 377     cl_kernel*  kernel_ptr;
 378     cl_int      cl_error;
 379
 380     const int elecTypeIdx = static_cast<int>(elecType);
 381     const int vdwTypeIdx  = static_cast<int>(vdwType);
 382
 383     GMX_ASSERT(elecTypeIdx < c_numElecTypes,
 384                "The electrostatics type requested is not implemented in the OpenCL kernels.");
 385     GMX_ASSERT(vdwTypeIdx < c_numVdwTypes,
 386                "The VdW type requested is not implemented in the OpenCL kernels.");
 387
 388     if (bDoEne)
 389     {
 390         if (bDoPrune)
 391         {
 392             kernel_name_to_run = nb_kfunc_ener_prune_ptr[elecTypeIdx][vdwTypeIdx];
 393             kernel_ptr         = &(nb->kernel_ener_prune_ptr[elecTypeIdx][vdwTypeIdx]);
 394         }
 395         else
 396         {
 397             kernel_name_to_run = nb_kfunc_ener_noprune_ptr[elecTypeIdx][vdwTypeIdx];
 398             kernel_ptr         = &(nb->kernel_ener_noprune_ptr[elecTypeIdx][vdwTypeIdx]);
 399         }
 400     }
 401     else
 402     {
 403         if (bDoPrune)
 404         {
 405             kernel_name_to_run = nb_kfunc_noener_prune_ptr[elecTypeIdx][vdwTypeIdx];
 406             kernel_ptr         = &(nb->kernel_noener_prune_ptr[elecTypeIdx][vdwTypeIdx]);
 407         }
 408         else
 409         {
 410             kernel_name_to_run = nb_kfunc_noener_noprune_ptr[elecTypeIdx][vdwTypeIdx];
 411             kernel_ptr         = &(nb->kernel_noener_noprune_ptr[elecTypeIdx][vdwTypeIdx]);
 412         }
 413     }
 414
 415     if (nullptr == kernel_ptr[0])
 416     {
 417         *kernel_ptr = clCreateKernel(nb->dev_rundata->program, kernel_name_to_run, &cl_error);
 418         GMX_ASSERT(cl_error == CL_SUCCESS,
 419                    ("clCreateKernel failed: " + ocl_get_error_string(cl_error)
 420                     + " for kernel named " + kernel_name_to_run)
 421                            .c_str());
 422     }
 423
 424     return *kernel_ptr;
 425 }
 426
 427 /*! \brief Calculates the amount of shared memory required by the nonbonded kernel in use.
 428  */
 429 static inline int calc_shmem_required_nonbonded(enum VdwType vdwType, bool bPrefetchLjParam)
 430 {
 431     int shmem;
 432
 433     /* size of shmem (force-buffers/xq/atom type preloading) */
 434     /* NOTE: with the default kernel on sm3.0 we need shmem only for pre-loading */
 435     /* i-atom x+q in shared memory */
 436     shmem = c_nbnxnGpuNumClusterPerSupercluster * c_clSize * sizeof(float) * 4; /* xqib */
 437     /* cj in shared memory, for both warps separately
 438      * TODO: in the "nowarp kernels we load cj only once  so the factor 2 is not needed.
 439      */
 440     shmem += 2 * c_nbnxnGpuJgroupSize * sizeof(int); /* cjs  */
 441     if (bPrefetchLjParam)
 442     {
 443         if (useLjCombRule(vdwType))
 444         {
 445             /* i-atom LJ combination parameters in shared memory */
 446             shmem += c_nbnxnGpuNumClusterPerSupercluster * c_clSize * 2
 447                      * sizeof(float); /* atib abused for ljcp, float2 */
 448         }
 449         else
 450         {
 451             /* i-atom types in shared memory */
 452             shmem += c_nbnxnGpuNumClusterPerSupercluster * c_clSize * sizeof(int); /* atib */
 453         }
 454     }
 455     /* force reduction buffers in shared memory */
 456     shmem += c_clSize * c_clSize * 3 * sizeof(float); /* f_buf */
 457     /* Warp vote. In fact it must be * number of warps in block.. */
 458     shmem += sizeof(cl_uint) * 2; /* warp_any */
 459     return shmem;
 460 }
 461
 462 /*! \brief Initializes data structures that are going to be sent to the OpenCL device.
 463  *
 464  *  The device can't use the same data structures as the host for two main reasons:
 465  *  - OpenCL restrictions (pointers are not accepted inside data structures)
 466  *  - some host side fields are not needed for the OpenCL kernels.
 467  *
 468  *  This function is called before the launch of both nbnxn and prune kernels.
 469  */
 470 static void fillin_ocl_structures(NBParamGpu* nbp, cl_nbparam_params_t* nbparams_params)
 471 {
 472     nbparams_params->coulomb_tab_scale = nbp->coulomb_tab_scale;
 473     nbparams_params->c_rf              = nbp->c_rf;
 474     nbparams_params->dispersion_shift  = nbp->dispersion_shift;
 475     nbparams_params->elecType          = nbp->elecType;
 476     nbparams_params->epsfac            = nbp->epsfac;
 477     nbparams_params->ewaldcoeff_lj     = nbp->ewaldcoeff_lj;
 478     nbparams_params->ewald_beta        = nbp->ewald_beta;
 479     nbparams_params->rcoulomb_sq       = nbp->rcoulomb_sq;
 480     nbparams_params->repulsion_shift   = nbp->repulsion_shift;
 481     nbparams_params->rlistOuter_sq     = nbp->rlistOuter_sq;
 482     nbparams_params->rvdw_sq           = nbp->rvdw_sq;
 483     nbparams_params->rlistInner_sq     = nbp->rlistInner_sq;
 484     nbparams_params->rvdw_switch       = nbp->rvdw_switch;
 485     nbparams_params->sh_ewald          = nbp->sh_ewald;
 486     nbparams_params->sh_lj_ewald       = nbp->sh_lj_ewald;
 487     nbparams_params->two_k_rf          = nbp->two_k_rf;
 488     nbparams_params->vdwType           = nbp->vdwType;
 489     nbparams_params->vdw_switch        = nbp->vdw_switch;
 490 }
 491
 492 void nbnxnInsertNonlocalGpuDependency(NbnxmGpu* nb, const InteractionLocality interactionLocality)
 493 {
 494     const DeviceStream& deviceStream = *nb->deviceStreams[interactionLocality];
 495
 496     /* When we get here all misc operations issued in the local stream as well as
 497        the local xq H2D are done,
 498        so we record that in the local stream and wait for it in the nonlocal one.
 499        This wait needs to precede any PP tasks, bonded or nonbonded, that may
 500        compute on interactions between local and nonlocal atoms.
 501      */
 502     if (nb->bUseTwoStreams)
 503     {
 504         if (interactionLocality == InteractionLocality::Local)
 505         {
 506             nb->misc_ops_and_local_H2D_done.markEvent(deviceStream);
 507
 508             /* Based on the v1.2 section 5.13 of the OpenCL spec, a flush is needed
 509              * in the local stream in order to be able to sync with the above event
 510              * from the non-local stream.
 511              */
 512             cl_int gmx_used_in_debug cl_error = clFlush(deviceStream.stream());
 513             GMX_ASSERT(cl_error == CL_SUCCESS,
 514                        ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
 515         }
 516         else
 517         {
 518             nb->misc_ops_and_local_H2D_done.enqueueWaitEvent(deviceStream);
 519         }
 520     }
 521 }
 522
 523 /*! \brief Launch GPU kernel
 524
 525    As we execute nonbonded workload in separate queues, before launching
 526    the kernel we need to make sure that he following operations have completed:
 527    - atomdata allocation and related H2D transfers (every nstlist step);
 528    - pair list H2D transfer (every nstlist step);
 529    - shift vector H2D transfer (every nstlist step);
 530    - force (+shift force and energy) output clearing (every step).
 531
 532    These operations are issued in the local queue at the beginning of the step
 533    and therefore always complete before the local kernel launch. The non-local
 534    kernel is launched after the local on the same device/context, so this is
 535    inherently scheduled after the operations in the local stream (including the
 536    above "misc_ops").
 537    However, for the sake of having a future-proof implementation, we use the
 538    misc_ops_done event to record the point in time when the above  operations
 539    are finished and synchronize with this event in the non-local stream.
 540  */
 541 void gpu_launch_kernel(NbnxmGpu* nb, const gmx::StepWorkload& stepWork, const Nbnxm::InteractionLocality iloc)
 542 {
 543     NBAtomData*         adat         = nb->atdat;
 544     NBParamGpu*         nbp          = nb->nbparam;
 545     gpu_plist*          plist        = nb->plist[iloc];
 546     Nbnxm::GpuTimers*   timers       = nb->timers;
 547     const DeviceStream& deviceStream = *nb->deviceStreams[iloc];
 548
 549     bool bDoTime = nb->bDoTime;
 550
 551     cl_nbparam_params_t nbparams_params;
 552
 553     /* Don't launch the non-local kernel if there is no work to do.
 554        Doing the same for the local kernel is more complicated, since the
 555        local part of the force array also depends on the non-local kernel.
 556        So to avoid complicating the code and to reduce the risk of bugs,
 557        we always call the local kernel and later (not in
 558        this function) the stream wait, local f copyback and the f buffer
 559        clearing. All these operations, except for the local interaction kernel,
 560        are needed for the non-local interactions. The skip of the local kernel
 561        call is taken care of later in this function. */
 562     if (canSkipNonbondedWork(*nb, iloc))
 563     {
 564         plist->haveFreshList = false;
 565
 566         return;
 567     }
 568
 569     if (nbp->useDynamicPruning && plist->haveFreshList)
 570     {
 571         /* Prunes for rlistOuter and rlistInner, sets plist->haveFreshList=false
 572            (that's the way the timing accounting can distinguish between
 573            separate prune kernel and combined force+prune).
 574          */
 575         Nbnxm::gpu_launch_kernel_pruneonly(nb, iloc, 1);
 576     }
 577
 578     if (plist->nsci == 0)
 579     {
 580         /* Don't launch an empty local kernel (is not allowed with OpenCL).
 581          */
 582         return;
 583     }
 584
 585     /* beginning of timed nonbonded calculation section */
 586     if (bDoTime)
 587     {
 588         timers->interaction[iloc].nb_k.openTimingRegion(deviceStream);
 589     }
 590
 591     /* kernel launch config */
 592
 593     KernelLaunchConfig config;
 594     config.sharedMemorySize = calc_shmem_required_nonbonded(nbp->vdwType, nb->bPrefetchLjParam);
 595     config.blockSize[0]     = c_clSize;
 596     config.blockSize[1]     = c_clSize;
 597     config.gridSize[0]      = plist->nsci;
 598
 599     validate_global_work_size(config, 3, &nb->deviceContext_->deviceInfo());
 600
 601     if (debug)
 602     {
 603         fprintf(debug,
 604                 "Non-bonded GPU launch configuration:\n\tLocal work size: %zux%zux%zu\n\t"
 605                 "Global work size : %zux%zu\n\t#Super-clusters/clusters: %d/%d (%d)\n",
 606                 config.blockSize[0],
 607                 config.blockSize[1],
 608                 config.blockSize[2],
 609                 config.blockSize[0] * config.gridSize[0],
 610                 config.blockSize[1] * config.gridSize[1],
 611                 plist->nsci * c_nbnxnGpuNumClusterPerSupercluster,
 612                 c_nbnxnGpuNumClusterPerSupercluster,
 613                 plist->na_c);
 614     }
 615
 616     fillin_ocl_structures(nbp, &nbparams_params);
 617
 618     auto* timingEvent = bDoTime ? timers->interaction[iloc].nb_k.fetchNextEvent() : nullptr;
 619     constexpr char kernelName[] = "k_calc_nb";
 620     const auto     kernel =
 621             select_nbnxn_kernel(nb,
 622                                 nbp->elecType,
 623                                 nbp->vdwType,
 624                                 stepWork.computeEnergy,
 625                                 (plist->haveFreshList && !nb->timers->interaction[iloc].didPrune));
 626
 627
 628     // The OpenCL kernel takes int as second to last argument because bool is
 629     // not supported as a kernel argument type (sizeof(bool) is implementation defined).
 630     const int computeFshift = static_cast<int>(stepWork.computeVirial);
 631     if (useLjCombRule(nb->nbparam->vdwType))
 632     {
 633         const auto kernelArgs = prepareGpuKernelArguments(kernel,
 634                                                           config,
 635                                                           &nbparams_params,
 636                                                           &adat->xq,
 637                                                           &adat->f,
 638                                                           &adat->eLJ,
 639                                                           &adat->eElec,
 640                                                           &adat->fShift,
 641                                                           &adat->ljComb,
 642                                                           &adat->shiftVec,
 643                                                           &nbp->nbfp,
 644                                                           &nbp->nbfp_comb,
 645                                                           &nbp->coulomb_tab,
 646                                                           &plist->sci,
 647                                                           &plist->cj4,
 648                                                           &plist->excl,
 649                                                           &computeFshift);
 650
 651         launchGpuKernel(kernel, config, deviceStream, timingEvent, kernelName, kernelArgs);
 652     }
 653     else
 654     {
 655         const auto kernelArgs = prepareGpuKernelArguments(kernel,
 656                                                           config,
 657                                                           &adat->numTypes,
 658                                                           &nbparams_params,
 659                                                           &adat->xq,
 660                                                           &adat->f,
 661                                                           &adat->eLJ,
 662                                                           &adat->eElec,
 663                                                           &adat->fShift,
 664                                                           &adat->atomTypes,
 665                                                           &adat->shiftVec,
 666                                                           &nbp->nbfp,
 667                                                           &nbp->nbfp_comb,
 668                                                           &nbp->coulomb_tab,
 669                                                           &plist->sci,
 670                                                           &plist->cj4,
 671                                                           &plist->excl,
 672                                                           &computeFshift);
 673         launchGpuKernel(kernel, config, deviceStream, timingEvent, kernelName, kernelArgs);
 674     }
 675
 676     if (bDoTime)
 677     {
 678         timers->interaction[iloc].nb_k.closeTimingRegion(deviceStream);
 679     }
 680 }
 681
 682
 683 /*! \brief Calculates the amount of shared memory required by the prune kernel.
 684  *
 685  *  Note that for the sake of simplicity we use the CUDA terminology "shared memory"
 686  *  for OpenCL local memory.
 687  *
 688  * \param[in] num_threads_z cj4 concurrency equal to the number of threads/work items in the 3-rd
 689  * dimension. \returns   the amount of local memory in bytes required by the pruning kernel
 690  */
 691 static inline int calc_shmem_required_prune(const int num_threads_z)
 692 {
 693     int shmem;
 694
 695     /* i-atom x in shared memory (for convenience we load all 4 components including q) */
 696     shmem = c_nbnxnGpuNumClusterPerSupercluster * c_clSize * sizeof(float) * 4;
 697     /* cj in shared memory, for each warp separately
 698      * Note: only need to load once per wavefront, but to keep the code simple,
 699      * for now we load twice on AMD.
 700      */
 701     shmem += num_threads_z * c_nbnxnGpuClusterpairSplit * c_nbnxnGpuJgroupSize * sizeof(int);
 702     /* Warp vote, requires one uint per warp/32 threads per block. */
 703     shmem += sizeof(cl_uint) * 2 * num_threads_z;
 704
 705     return shmem;
 706 }
 707
 708 /*! \brief
 709  * Launch the pairlist prune only kernel for the given locality.
 710  * \p numParts tells in how many parts, i.e. calls the list will be pruned.
 711  */
 712 void gpu_launch_kernel_pruneonly(NbnxmGpu* nb, const InteractionLocality iloc, const int numParts)
 713 {
 714     NBAtomData*         adat         = nb->atdat;
 715     NBParamGpu*         nbp          = nb->nbparam;
 716     gpu_plist*          plist        = nb->plist[iloc];
 717     Nbnxm::GpuTimers*   timers       = nb->timers;
 718     const DeviceStream& deviceStream = *nb->deviceStreams[iloc];
 719     bool                bDoTime      = nb->bDoTime;
 720
 721     if (plist->haveFreshList)
 722     {
 723         GMX_ASSERT(numParts == 1, "With first pruning we expect 1 part");
 724
 725         /* Set rollingPruningNumParts to signal that it is not set */
 726         plist->rollingPruningNumParts = 0;
 727         plist->rollingPruningPart     = 0;
 728     }
 729     else
 730     {
 731         if (plist->rollingPruningNumParts == 0)
 732         {
 733             plist->rollingPruningNumParts = numParts;
 734         }
 735         else
 736         {
 737             GMX_ASSERT(numParts == plist->rollingPruningNumParts,
 738                        "It is not allowed to change numParts in between list generation steps");
 739         }
 740     }
 741
 742     /* Use a local variable for part and update in plist, so we can return here
 743      * without duplicating the part increment code.
 744      */
 745     int part = plist->rollingPruningPart;
 746
 747     plist->rollingPruningPart++;
 748     if (plist->rollingPruningPart >= plist->rollingPruningNumParts)
 749     {
 750         plist->rollingPruningPart = 0;
 751     }
 752
 753     /* Compute the number of list entries to prune in this pass */
 754     int numSciInPart = (plist->nsci - part) / numParts;
 755
 756     /* Don't launch the kernel if there is no work to do. */
 757     if (numSciInPart <= 0)
 758     {
 759         plist->haveFreshList = false;
 760
 761         return;
 762     }
 763
 764     GpuRegionTimer* timer = nullptr;
 765     if (bDoTime)
 766     {
 767         timer = &(plist->haveFreshList ? timers->interaction[iloc].prune_k
 768                                        : timers->interaction[iloc].rollingPrune_k);
 769     }
 770
 771     /* beginning of timed prune calculation section */
 772     if (bDoTime)
 773     {
 774         timer->openTimingRegion(deviceStream);
 775     }
 776
 777     /* Kernel launch config:
 778      * - The thread block dimensions match the size of i-clusters, j-clusters,
 779      *   and j-cluster concurrency, in x, y, and z, respectively.
 780      * - The 1D block-grid contains as many blocks as super-clusters.
 781      */
 782     int num_threads_z = c_pruneKernelJ4Concurrency;
 783     /* kernel launch config */
 784     KernelLaunchConfig config;
 785     config.sharedMemorySize = calc_shmem_required_prune(num_threads_z);
 786     config.blockSize[0]     = c_clSize;
 787     config.blockSize[1]     = c_clSize;
 788     config.blockSize[2]     = num_threads_z;
 789     config.gridSize[0]      = numSciInPart;
 790
 791     validate_global_work_size(config, 3, &nb->deviceContext_->deviceInfo());
 792
 793     if (debug)
 794     {
 795         fprintf(debug,
 796                 "Pruning GPU kernel launch configuration:\n\tLocal work size: %zux%zux%zu\n\t"
 797                 "\tGlobal work size: %zux%zu\n\t#Super-clusters/clusters: %d/%d (%d)\n"
 798                 "\tShMem: %zu\n",
 799                 config.blockSize[0],
 800                 config.blockSize[1],
 801                 config.blockSize[2],
 802                 config.blockSize[0] * config.gridSize[0],
 803                 config.blockSize[1] * config.gridSize[1],
 804                 plist->nsci * c_nbnxnGpuNumClusterPerSupercluster,
 805                 c_nbnxnGpuNumClusterPerSupercluster,
 806                 plist->na_c,
 807                 config.sharedMemorySize);
 808     }
 809
 810     cl_nbparam_params_t nbparams_params;
 811     fillin_ocl_structures(nbp, &nbparams_params);
 812
 813     auto*          timingEvent  = bDoTime ? timer->fetchNextEvent() : nullptr;
 814     constexpr char kernelName[] = "k_pruneonly";
 815     const auto     pruneKernel  = selectPruneKernel(nb->kernel_pruneonly, plist->haveFreshList);
 816     const auto     kernelArgs   = prepareGpuKernelArguments(pruneKernel,
 817                                                       config,
 818                                                       &nbparams_params,
 819                                                       &adat->xq,
 820                                                       &adat->shiftVec,
 821                                                       &plist->sci,
 822                                                       &plist->cj4,
 823                                                       &plist->imask,
 824                                                       &numParts,
 825                                                       &part);
 826     launchGpuKernel(pruneKernel, config, deviceStream, timingEvent, kernelName, kernelArgs);
 827
 828     if (plist->haveFreshList)
 829     {
 830         plist->haveFreshList = false;
 831         /* Mark that pruning has been done */
 832         nb->timers->interaction[iloc].didPrune = true;
 833     }
 834     else
 835     {
 836         /* Mark that rolling pruning has been done */
 837         nb->timers->interaction[iloc].didRollingPrune = true;
 838     }
 839
 840     if (bDoTime)
 841     {
 842         timer->closeTimingRegion(deviceStream);
 843     }
 844 }
 845
 846 /*! \brief
 847  * Launch asynchronously the download of nonbonded forces from the GPU
 848  * (and energies/shift forces if required).
 849  */
 850 void gpu_launch_cpyback(NbnxmGpu*                nb,
 851                         struct nbnxn_atomdata_t* nbatom,
 852                         const gmx::StepWorkload& stepWork,
 853                         const AtomLocality       atomLocality)
 854 {
 855     GMX_ASSERT(nb, "Need a valid nbnxn_gpu object");
 856
 857     cl_int gmx_unused cl_error;
 858
 859     /* determine interaction locality from atom locality */
 860     const InteractionLocality iloc = gpuAtomToInteractionLocality(atomLocality);
 861     GMX_ASSERT(iloc == InteractionLocality::Local
 862                        || (iloc == InteractionLocality::NonLocal && nb->bNonLocalStreamDoneMarked == false),
 863                "Non-local stream is indicating that the copy back event is enqueued at the "
 864                "beginning of the copy back function.");
 865
 866     NBAtomData*         adat         = nb->atdat;
 867     Nbnxm::GpuTimers*   timers       = nb->timers;
 868     bool                bDoTime      = nb->bDoTime;
 869     const DeviceStream& deviceStream = *nb->deviceStreams[iloc];
 870
 871     /* don't launch non-local copy-back if there was no non-local work to do */
 872     if ((iloc == InteractionLocality::NonLocal) && !haveGpuShortRangeWork(*nb, iloc))
 873     {
 874         /* TODO An alternative way to signal that non-local work is
 875            complete is to use a clEnqueueMarker+clEnqueueBarrier
 876            pair. However, the use of bNonLocalStreamDoneMarked has the
 877            advantage of being local to the host, so probably minimizes
 878            overhead. Curiously, for NVIDIA OpenCL with an empty-domain
 879            test case, overall simulation performance was higher with
 880            the API calls, but this has not been tested on AMD OpenCL,
 881            so could be worth considering in future. */
 882         nb->bNonLocalStreamDoneMarked = false;
 883         return;
 884     }
 885
 886     /* local/nonlocal offset and length used for xq and f */
 887     auto atomsRange = getGpuAtomRange(adat, atomLocality);
 888
 889     /* beginning of timed D2H section */
 890     if (bDoTime)
 891     {
 892         timers->xf[atomLocality].nb_d2h.openTimingRegion(deviceStream);
 893     }
 894
 895     /* With DD the local D2H transfer can only start after the non-local
 896        has been launched. */
 897     if (iloc == InteractionLocality::Local && nb->bNonLocalStreamDoneMarked)
 898     {
 899         nb->nonlocal_done.enqueueWaitEvent(deviceStream);
 900         nb->bNonLocalStreamDoneMarked = false;
 901     }
 902
 903     /* DtoH f */
 904     GMX_ASSERT(sizeof(*nbatom->out[0].f.data()) == sizeof(float),
 905                "The host force buffer should be in single precision to match device data size.");
 906     copyFromDeviceBuffer(reinterpret_cast<Float3*>(nbatom->out[0].f.data()) + atomsRange.begin(),
 907                          &adat->f,
 908                          atomsRange.begin(),
 909                          atomsRange.size(),
 910                          deviceStream,
 911                          GpuApiCallBehavior::Async,
 912                          bDoTime ? timers->xf[atomLocality].nb_d2h.fetchNextEvent() : nullptr);
 913
 914     /* kick off work */
 915     cl_error = clFlush(deviceStream.stream());
 916     GMX_ASSERT(cl_error == CL_SUCCESS, ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
 917
 918     /* After the non-local D2H is launched the nonlocal_done event can be
 919        recorded which signals that the local D2H can proceed. This event is not
 920        placed after the non-local kernel because we first need the non-local
 921        data back first. */
 922     if (iloc == InteractionLocality::NonLocal)
 923     {
 924         nb->nonlocal_done.markEvent(deviceStream);
 925         nb->bNonLocalStreamDoneMarked = true;
 926     }
 927
 928     /* only transfer energies in the local stream */
 929     if (iloc == InteractionLocality::Local)
 930     {
 931         /* DtoH fshift when virial is needed */
 932         if (stepWork.computeVirial)
 933         {
 934             static_assert(
 935                     sizeof(*nb->nbst.fShift) == sizeof(Float3),
 936                     "Sizes of host- and device-side shift vector elements should be the same.");
 937             copyFromDeviceBuffer(nb->nbst.fShift,
 938                                  &adat->fShift,
 939                                  0,
 940                                  SHIFTS,
 941                                  deviceStream,
 942                                  GpuApiCallBehavior::Async,
 943                                  bDoTime ? timers->xf[atomLocality].nb_d2h.fetchNextEvent() : nullptr);
 944         }
 945
 946         /* DtoH energies */
 947         if (stepWork.computeEnergy)
 948         {
 949             static_assert(sizeof(*nb->nbst.eLJ) == sizeof(float),
 950                           "Sizes of host- and device-side LJ energy terms should be the same.");
 951             copyFromDeviceBuffer(nb->nbst.eLJ,
 952                                  &adat->eLJ,
 953                                  0,
 954                                  1,
 955                                  deviceStream,
 956                                  GpuApiCallBehavior::Async,
 957                                  bDoTime ? timers->xf[atomLocality].nb_d2h.fetchNextEvent() : nullptr);
 958             static_assert(sizeof(*nb->nbst.eElec) == sizeof(float),
 959                           "Sizes of host- and device-side electrostatic energy terms should be the "
 960                           "same.");
 961             copyFromDeviceBuffer(nb->nbst.eElec,
 962                                  &adat->eElec,
 963                                  0,
 964                                  1,
 965                                  deviceStream,
 966                                  GpuApiCallBehavior::Async,
 967                                  bDoTime ? timers->xf[atomLocality].nb_d2h.fetchNextEvent() : nullptr);
 968         }
 969     }
 970
 971     if (bDoTime)
 972     {
 973         timers->xf[atomLocality].nb_d2h.closeTimingRegion(deviceStream);
 974     }
 975 }
 976
 977 } // namespace Nbnxm