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, 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 "thread_mpi/atomic.h"
  72
  73 #include "gromacs/gpu_utils/device_context.h"
  74 #include "gromacs/gpu_utils/gputraits_ocl.h"
  75 #include "gromacs/gpu_utils/oclutils.h"
  76 #include "gromacs/hardware/device_information.h"
  77 #include "gromacs/hardware/hw_info.h"
  78 #include "gromacs/mdtypes/simulation_workload.h"
  79 #include "gromacs/nbnxm/atomdata.h"
  80 #include "gromacs/nbnxm/gpu_common.h"
  81 #include "gromacs/nbnxm/gpu_common_utils.h"
  82 #include "gromacs/nbnxm/gpu_data_mgmt.h"
  83 #include "gromacs/nbnxm/nbnxm.h"
  84 #include "gromacs/nbnxm/nbnxm_gpu.h"
  85 #include "gromacs/nbnxm/pairlist.h"
  86 #include "gromacs/pbcutil/ishift.h"
  87 #include "gromacs/timing/gpu_timing.h"
  88 #include "gromacs/utility/cstringutil.h"
  89 #include "gromacs/utility/fatalerror.h"
  90 #include "gromacs/utility/gmxassert.h"
  91
  92 #include "nbnxm_ocl_types.h"
  93
  94 namespace Nbnxm
  95 {
  96
  97 /*! \brief Convenience constants */
  98 //@{
  99 static constexpr int c_clSize = c_nbnxnGpuClusterSize;
 100 //@}
 101
 102
 103 /*! \brief Validates the input global work size parameter.
 104  */
 105 static inline void validate_global_work_size(const KernelLaunchConfig& config,
 106                                              int                       work_dim,
 107                                              const DeviceInformation*  dinfo)
 108 {
 109     cl_uint device_size_t_size_bits;
 110     cl_uint host_size_t_size_bits;
 111
 112     GMX_ASSERT(dinfo, "Need a valid device info object");
 113
 114     size_t global_work_size[3];
 115     GMX_ASSERT(work_dim <= 3, "Not supporting hyper-grids just yet");
 116     for (int i = 0; i < work_dim; i++)
 117     {
 118         global_work_size[i] = config.blockSize[i] * config.gridSize[i];
 119     }
 120
 121     /* Each component of a global_work_size must not exceed the range given by the
 122        sizeof(device size_t) for the device on which the kernel execution will
 123        be enqueued. See:
 124        https://www.khronos.org/registry/cl/sdk/1.0/docs/man/xhtml/clEnqueueNDRangeKernel.html
 125      */
 126     device_size_t_size_bits = dinfo->adress_bits;
 127     host_size_t_size_bits   = static_cast<cl_uint>(sizeof(size_t) * 8);
 128
 129     /* If sizeof(host size_t) <= sizeof(device size_t)
 130             => global_work_size components will always be valid
 131        else
 132             => get device limit for global work size and
 133             compare it against each component of global_work_size.
 134      */
 135     if (host_size_t_size_bits > device_size_t_size_bits)
 136     {
 137         size_t device_limit;
 138
 139         device_limit = (1ULL << device_size_t_size_bits) - 1;
 140
 141         for (int i = 0; i < work_dim; i++)
 142         {
 143             if (global_work_size[i] > device_limit)
 144             {
 145                 gmx_fatal(
 146                         FARGS,
 147                         "Watch out, the input system is too large to simulate!\n"
 148                         "The number of nonbonded work units (=number of super-clusters) exceeds the"
 149                         "device capabilities. Global work size limit exceeded (%zu > %zu)!",
 150                         global_work_size[i], device_limit);
 151             }
 152         }
 153     }
 154 }
 155
 156 /* Constant arrays listing non-bonded kernel function names. The arrays are
 157  * organized in 2-dim arrays by: electrostatics and VDW type.
 158  *
 159  *  Note that the row- and column-order of function pointers has to match the
 160  *  order of corresponding enumerated electrostatics and vdw types, resp.,
 161  *  defined in nbnxm_ocl_types.h.
 162  */
 163
 164 /*! \brief Force-only kernel function names. */
 165 static const char* nb_kfunc_noener_noprune_ptr[c_numElecTypes][c_numVdwTypes] = {
 166     { "nbnxn_kernel_ElecCut_VdwLJ_F_opencl", "nbnxn_kernel_ElecCut_VdwLJCombGeom_F_opencl",
 167       "nbnxn_kernel_ElecCut_VdwLJCombLB_F_opencl", "nbnxn_kernel_ElecCut_VdwLJFsw_F_opencl",
 168       "nbnxn_kernel_ElecCut_VdwLJPsw_F_opencl", "nbnxn_kernel_ElecCut_VdwLJEwCombGeom_F_opencl",
 169       "nbnxn_kernel_ElecCut_VdwLJEwCombLB_F_opencl" },
 170     { "nbnxn_kernel_ElecRF_VdwLJ_F_opencl", "nbnxn_kernel_ElecRF_VdwLJCombGeom_F_opencl",
 171       "nbnxn_kernel_ElecRF_VdwLJCombLB_F_opencl", "nbnxn_kernel_ElecRF_VdwLJFsw_F_opencl",
 172       "nbnxn_kernel_ElecRF_VdwLJPsw_F_opencl", "nbnxn_kernel_ElecRF_VdwLJEwCombGeom_F_opencl",
 173       "nbnxn_kernel_ElecRF_VdwLJEwCombLB_F_opencl" },
 174     { "nbnxn_kernel_ElecEwQSTab_VdwLJ_F_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJCombGeom_F_opencl",
 175       "nbnxn_kernel_ElecEwQSTab_VdwLJCombLB_F_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJFsw_F_opencl",
 176       "nbnxn_kernel_ElecEwQSTab_VdwLJPsw_F_opencl",
 177       "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_F_opencl",
 178       "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_F_opencl" },
 179     { "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_F_opencl",
 180       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombGeom_F_opencl",
 181       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombLB_F_opencl",
 182       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_F_opencl",
 183       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_F_opencl",
 184       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_F_opencl",
 185       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_F_opencl" },
 186     { "nbnxn_kernel_ElecEw_VdwLJ_F_opencl", "nbnxn_kernel_ElecEw_VdwLJCombGeom_F_opencl",
 187       "nbnxn_kernel_ElecEw_VdwLJCombLB_F_opencl", "nbnxn_kernel_ElecEw_VdwLJFsw_F_opencl",
 188       "nbnxn_kernel_ElecEw_VdwLJPsw_F_opencl", "nbnxn_kernel_ElecEw_VdwLJEwCombGeom_F_opencl",
 189       "nbnxn_kernel_ElecEw_VdwLJEwCombLB_F_opencl" },
 190     { "nbnxn_kernel_ElecEwTwinCut_VdwLJ_F_opencl",
 191       "nbnxn_kernel_ElecEwTwinCut_VdwLJCombGeom_F_opencl",
 192       "nbnxn_kernel_ElecEwTwinCut_VdwLJCombLB_F_opencl",
 193       "nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_F_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_F_opencl",
 194       "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_F_opencl",
 195       "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_F_opencl" }
 196 };
 197
 198 /*! \brief Force + energy kernel function pointers. */
 199 static const char* nb_kfunc_ener_noprune_ptr[c_numElecTypes][c_numVdwTypes] = {
 200     { "nbnxn_kernel_ElecCut_VdwLJ_VF_opencl", "nbnxn_kernel_ElecCut_VdwLJCombGeom_VF_opencl",
 201       "nbnxn_kernel_ElecCut_VdwLJCombLB_VF_opencl", "nbnxn_kernel_ElecCut_VdwLJFsw_VF_opencl",
 202       "nbnxn_kernel_ElecCut_VdwLJPsw_VF_opencl", "nbnxn_kernel_ElecCut_VdwLJEwCombGeom_VF_opencl",
 203       "nbnxn_kernel_ElecCut_VdwLJEwCombLB_VF_opencl" },
 204     { "nbnxn_kernel_ElecRF_VdwLJ_VF_opencl", "nbnxn_kernel_ElecRF_VdwLJCombGeom_VF_opencl",
 205       "nbnxn_kernel_ElecRF_VdwLJCombLB_VF_opencl", "nbnxn_kernel_ElecRF_VdwLJFsw_VF_opencl",
 206       "nbnxn_kernel_ElecRF_VdwLJPsw_VF_opencl", "nbnxn_kernel_ElecRF_VdwLJEwCombGeom_VF_opencl",
 207       "nbnxn_kernel_ElecRF_VdwLJEwCombLB_VF_opencl" },
 208     { "nbnxn_kernel_ElecEwQSTab_VdwLJ_VF_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJCombGeom_VF_opencl",
 209       "nbnxn_kernel_ElecEwQSTab_VdwLJCombLB_VF_opencl",
 210       "nbnxn_kernel_ElecEwQSTab_VdwLJFsw_VF_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJPsw_VF_opencl",
 211       "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_VF_opencl",
 212       "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_VF_opencl" },
 213     { "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_VF_opencl",
 214       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombGeom_VF_opencl",
 215       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombLB_VF_opencl",
 216       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_VF_opencl",
 217       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_VF_opencl",
 218       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_VF_opencl",
 219       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_VF_opencl" },
 220     { "nbnxn_kernel_ElecEw_VdwLJ_VF_opencl", "nbnxn_kernel_ElecEw_VdwLJCombGeom_VF_opencl",
 221       "nbnxn_kernel_ElecEw_VdwLJCombLB_VF_opencl", "nbnxn_kernel_ElecEw_VdwLJFsw_VF_opencl",
 222       "nbnxn_kernel_ElecEw_VdwLJPsw_VF_opencl", "nbnxn_kernel_ElecEw_VdwLJEwCombGeom_VF_opencl",
 223       "nbnxn_kernel_ElecEw_VdwLJEwCombLB_VF_opencl" },
 224     { "nbnxn_kernel_ElecEwTwinCut_VdwLJ_VF_opencl",
 225       "nbnxn_kernel_ElecEwTwinCut_VdwLJCombGeom_VF_opencl",
 226       "nbnxn_kernel_ElecEwTwinCut_VdwLJCombLB_VF_opencl",
 227       "nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_VF_opencl",
 228       "nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_VF_opencl",
 229       "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF_opencl",
 230       "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_VF_opencl" }
 231 };
 232
 233 /*! \brief Force + pruning kernel function pointers. */
 234 static const char* nb_kfunc_noener_prune_ptr[c_numElecTypes][c_numVdwTypes] = {
 235     { "nbnxn_kernel_ElecCut_VdwLJ_F_prune_opencl",
 236       "nbnxn_kernel_ElecCut_VdwLJCombGeom_F_prune_opencl",
 237       "nbnxn_kernel_ElecCut_VdwLJCombLB_F_prune_opencl",
 238       "nbnxn_kernel_ElecCut_VdwLJFsw_F_prune_opencl", "nbnxn_kernel_ElecCut_VdwLJPsw_F_prune_opencl",
 239       "nbnxn_kernel_ElecCut_VdwLJEwCombGeom_F_prune_opencl",
 240       "nbnxn_kernel_ElecCut_VdwLJEwCombLB_F_prune_opencl" },
 241     { "nbnxn_kernel_ElecRF_VdwLJ_F_prune_opencl", "nbnxn_kernel_ElecRF_VdwLJCombGeom_F_prune_opencl",
 242       "nbnxn_kernel_ElecRF_VdwLJCombLB_F_prune_opencl",
 243       "nbnxn_kernel_ElecRF_VdwLJFsw_F_prune_opencl", "nbnxn_kernel_ElecRF_VdwLJPsw_F_prune_opencl",
 244       "nbnxn_kernel_ElecRF_VdwLJEwCombGeom_F_prune_opencl",
 245       "nbnxn_kernel_ElecRF_VdwLJEwCombLB_F_prune_opencl" },
 246     { "nbnxn_kernel_ElecEwQSTab_VdwLJ_F_prune_opencl",
 247       "nbnxn_kernel_ElecEwQSTab_VdwLJCombGeom_F_prune_opencl",
 248       "nbnxn_kernel_ElecEwQSTab_VdwLJCombLB_F_prune_opencl",
 249       "nbnxn_kernel_ElecEwQSTab_VdwLJFsw_F_prune_opencl",
 250       "nbnxn_kernel_ElecEwQSTab_VdwLJPsw_F_prune_opencl",
 251       "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_F_prune_opencl",
 252       "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_F_prune_opencl" },
 253     { "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_F_prune_opencl",
 254       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombGeom_F_prune_opencl",
 255       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombLB_F_prune_opencl",
 256       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_F_prune_opencl",
 257       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_F_prune_opencl",
 258       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_F_prune_opencl",
 259       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_F_prune_opencl" },
 260     { "nbnxn_kernel_ElecEw_VdwLJ_F_prune_opencl", "nbnxn_kernel_ElecEw_VdwLJCombGeom_F_prune_opencl",
 261       "nbnxn_kernel_ElecEw_VdwLJCombLB_F_prune_opencl",
 262       "nbnxn_kernel_ElecEw_VdwLJFsw_F_prune_opencl", "nbnxn_kernel_ElecEw_VdwLJPsw_F_prune_opencl",
 263       "nbnxn_kernel_ElecEw_VdwLJEwCombGeom_F_prune_opencl",
 264       "nbnxn_kernel_ElecEw_VdwLJEwCombLB_F_prune_opencl" },
 265     { "nbnxn_kernel_ElecEwTwinCut_VdwLJ_F_prune_opencl",
 266       "nbnxn_kernel_ElecEwTwinCut_VdwLJCombGeom_F_prune_opencl",
 267       "nbnxn_kernel_ElecEwTwinCut_VdwLJCombLB_F_prune_opencl",
 268       "nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_F_prune_opencl",
 269       "nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_F_prune_opencl",
 270       "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_F_prune_opencl",
 271       "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_F_prune_opencl" }
 272 };
 273
 274 /*! \brief Force + energy + pruning kernel function pointers. */
 275 static const char* nb_kfunc_ener_prune_ptr[c_numElecTypes][c_numVdwTypes] = {
 276     { "nbnxn_kernel_ElecCut_VdwLJ_VF_prune_opencl",
 277       "nbnxn_kernel_ElecCut_VdwLJCombGeom_VF_prune_opencl",
 278       "nbnxn_kernel_ElecCut_VdwLJCombLB_VF_prune_opencl",
 279       "nbnxn_kernel_ElecCut_VdwLJFsw_VF_prune_opencl",
 280       "nbnxn_kernel_ElecCut_VdwLJPsw_VF_prune_opencl",
 281       "nbnxn_kernel_ElecCut_VdwLJEwCombGeom_VF_prune_opencl",
 282       "nbnxn_kernel_ElecCut_VdwLJEwCombLB_VF_prune_opencl" },
 283     { "nbnxn_kernel_ElecRF_VdwLJ_VF_prune_opencl",
 284       "nbnxn_kernel_ElecRF_VdwLJCombGeom_VF_prune_opencl",
 285       "nbnxn_kernel_ElecRF_VdwLJCombLB_VF_prune_opencl",
 286       "nbnxn_kernel_ElecRF_VdwLJFsw_VF_prune_opencl", "nbnxn_kernel_ElecRF_VdwLJPsw_VF_prune_opencl",
 287       "nbnxn_kernel_ElecRF_VdwLJEwCombGeom_VF_prune_opencl",
 288       "nbnxn_kernel_ElecRF_VdwLJEwCombLB_VF_prune_opencl" },
 289     { "nbnxn_kernel_ElecEwQSTab_VdwLJ_VF_prune_opencl",
 290       "nbnxn_kernel_ElecEwQSTab_VdwLJCombGeom_VF_prune_opencl",
 291       "nbnxn_kernel_ElecEwQSTab_VdwLJCombLB_VF_prune_opencl",
 292       "nbnxn_kernel_ElecEwQSTab_VdwLJFsw_VF_prune_opencl",
 293       "nbnxn_kernel_ElecEwQSTab_VdwLJPsw_VF_prune_opencl",
 294       "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_VF_prune_opencl",
 295       "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_VF_prune_opencl" },
 296     { "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_VF_prune_opencl",
 297       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombGeom_VF_prune_opencl",
 298       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombLB_VF_prune_opencl",
 299       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_VF_prune_opencl",
 300       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_VF_prune_opencl",
 301       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_VF_prune_opencl",
 302       "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_VF_prune_opencl" },
 303     { "nbnxn_kernel_ElecEw_VdwLJ_VF_prune_opencl",
 304       "nbnxn_kernel_ElecEw_VdwLJCombGeom_VF_prune_opencl",
 305       "nbnxn_kernel_ElecEw_VdwLJCombLB_VF_prune_opencl",
 306       "nbnxn_kernel_ElecEw_VdwLJFsw_VF_prune_opencl", "nbnxn_kernel_ElecEw_VdwLJPsw_VF_prune_opencl",
 307       "nbnxn_kernel_ElecEw_VdwLJEwCombGeom_VF_prune_opencl",
 308       "nbnxn_kernel_ElecEw_VdwLJEwCombLB_VF_prune_opencl" },
 309     { "nbnxn_kernel_ElecEwTwinCut_VdwLJ_VF_prune_opencl",
 310       "nbnxn_kernel_ElecEwTwinCut_VdwLJCombGeom_VF_prune_opencl",
 311       "nbnxn_kernel_ElecEwTwinCut_VdwLJCombLB_VF_prune_opencl",
 312       "nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_VF_prune_opencl",
 313       "nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_VF_prune_opencl",
 314       "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF_prune_opencl",
 315       "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_VF_prune_opencl" }
 316 };
 317
 318 /*! \brief Return a pointer to the prune kernel version to be executed at the current invocation.
 319  *
 320  * \param[in] kernel_pruneonly  array of prune kernel objects
 321  * \param[in] firstPrunePass    true if the first pruning pass is being executed
 322  */
 323 static inline cl_kernel selectPruneKernel(cl_kernel kernel_pruneonly[], bool firstPrunePass)
 324 {
 325     cl_kernel* kernelPtr;
 326
 327     if (firstPrunePass)
 328     {
 329         kernelPtr = &(kernel_pruneonly[epruneFirst]);
 330     }
 331     else
 332     {
 333         kernelPtr = &(kernel_pruneonly[epruneRolling]);
 334     }
 335     // TODO: consider creating the prune kernel object here to avoid a
 336     // clCreateKernel for the rolling prune kernel if this is not needed.
 337     return *kernelPtr;
 338 }
 339
 340 /*! \brief Return a pointer to the kernel version to be executed at the current step.
 341  *  OpenCL kernel objects are cached in nb. If the requested kernel is not
 342  *  found in the cache, it will be created and the cache will be updated.
 343  */
 344 static inline cl_kernel
 345 select_nbnxn_kernel(NbnxmGpu* nb, enum ElecType elecType, enum VdwType vdwType, bool bDoEne, bool bDoPrune)
 346 {
 347     const char* kernel_name_to_run;
 348     cl_kernel*  kernel_ptr;
 349     cl_int      cl_error;
 350
 351     const int elecTypeIdx = static_cast<int>(elecType);
 352     const int vdwTypeIdx  = static_cast<int>(vdwType);
 353
 354     GMX_ASSERT(elecTypeIdx < c_numElecTypes,
 355                "The electrostatics type requested is not implemented in the OpenCL kernels.");
 356     GMX_ASSERT(vdwTypeIdx < c_numVdwTypes,
 357                "The VdW type requested is not implemented in the OpenCL kernels.");
 358
 359     if (bDoEne)
 360     {
 361         if (bDoPrune)
 362         {
 363             kernel_name_to_run = nb_kfunc_ener_prune_ptr[elecTypeIdx][vdwTypeIdx];
 364             kernel_ptr         = &(nb->kernel_ener_prune_ptr[elecTypeIdx][vdwTypeIdx]);
 365         }
 366         else
 367         {
 368             kernel_name_to_run = nb_kfunc_ener_noprune_ptr[elecTypeIdx][vdwTypeIdx];
 369             kernel_ptr         = &(nb->kernel_ener_noprune_ptr[elecTypeIdx][vdwTypeIdx]);
 370         }
 371     }
 372     else
 373     {
 374         if (bDoPrune)
 375         {
 376             kernel_name_to_run = nb_kfunc_noener_prune_ptr[elecTypeIdx][vdwTypeIdx];
 377             kernel_ptr         = &(nb->kernel_noener_prune_ptr[elecTypeIdx][vdwTypeIdx]);
 378         }
 379         else
 380         {
 381             kernel_name_to_run = nb_kfunc_noener_noprune_ptr[elecTypeIdx][vdwTypeIdx];
 382             kernel_ptr         = &(nb->kernel_noener_noprune_ptr[elecTypeIdx][vdwTypeIdx]);
 383         }
 384     }
 385
 386     if (nullptr == kernel_ptr[0])
 387     {
 388         *kernel_ptr = clCreateKernel(nb->dev_rundata->program, kernel_name_to_run, &cl_error);
 389         GMX_ASSERT(cl_error == CL_SUCCESS, ("clCreateKernel failed: " + ocl_get_error_string(cl_error)
 390                                             + " for kernel named " + kernel_name_to_run)
 391                                                    .c_str());
 392     }
 393
 394     return *kernel_ptr;
 395 }
 396
 397 /*! \brief Calculates the amount of shared memory required by the nonbonded kernel in use.
 398  */
 399 static inline int calc_shmem_required_nonbonded(enum VdwType vdwType, bool bPrefetchLjParam)
 400 {
 401     int shmem;
 402
 403     /* size of shmem (force-buffers/xq/atom type preloading) */
 404     /* NOTE: with the default kernel on sm3.0 we need shmem only for pre-loading */
 405     /* i-atom x+q in shared memory */
 406     shmem = c_nbnxnGpuNumClusterPerSupercluster * c_clSize * sizeof(float) * 4; /* xqib */
 407     /* cj in shared memory, for both warps separately
 408      * TODO: in the "nowarp kernels we load cj only once  so the factor 2 is not needed.
 409      */
 410     shmem += 2 * c_nbnxnGpuJgroupSize * sizeof(int); /* cjs  */
 411     if (bPrefetchLjParam)
 412     {
 413         if (useLjCombRule(vdwType))
 414         {
 415             /* i-atom LJ combination parameters in shared memory */
 416             shmem += c_nbnxnGpuNumClusterPerSupercluster * c_clSize * 2
 417                      * sizeof(float); /* atib abused for ljcp, float2 */
 418         }
 419         else
 420         {
 421             /* i-atom types in shared memory */
 422             shmem += c_nbnxnGpuNumClusterPerSupercluster * c_clSize * sizeof(int); /* atib */
 423         }
 424     }
 425     /* force reduction buffers in shared memory */
 426     shmem += c_clSize * c_clSize * 3 * sizeof(float); /* f_buf */
 427     /* Warp vote. In fact it must be * number of warps in block.. */
 428     shmem += sizeof(cl_uint) * 2; /* warp_any */
 429     return shmem;
 430 }
 431
 432 /*! \brief Initializes data structures that are going to be sent to the OpenCL device.
 433  *
 434  *  The device can't use the same data structures as the host for two main reasons:
 435  *  - OpenCL restrictions (pointers are not accepted inside data structures)
 436  *  - some host side fields are not needed for the OpenCL kernels.
 437  *
 438  *  This function is called before the launch of both nbnxn and prune kernels.
 439  */
 440 static void fillin_ocl_structures(NBParamGpu* nbp, cl_nbparam_params_t* nbparams_params)
 441 {
 442     nbparams_params->coulomb_tab_scale = nbp->coulomb_tab_scale;
 443     nbparams_params->c_rf              = nbp->c_rf;
 444     nbparams_params->dispersion_shift  = nbp->dispersion_shift;
 445     nbparams_params->elecType          = nbp->elecType;
 446     nbparams_params->epsfac            = nbp->epsfac;
 447     nbparams_params->ewaldcoeff_lj     = nbp->ewaldcoeff_lj;
 448     nbparams_params->ewald_beta        = nbp->ewald_beta;
 449     nbparams_params->rcoulomb_sq       = nbp->rcoulomb_sq;
 450     nbparams_params->repulsion_shift   = nbp->repulsion_shift;
 451     nbparams_params->rlistOuter_sq     = nbp->rlistOuter_sq;
 452     nbparams_params->rvdw_sq           = nbp->rvdw_sq;
 453     nbparams_params->rlistInner_sq     = nbp->rlistInner_sq;
 454     nbparams_params->rvdw_switch       = nbp->rvdw_switch;
 455     nbparams_params->sh_ewald          = nbp->sh_ewald;
 456     nbparams_params->sh_lj_ewald       = nbp->sh_lj_ewald;
 457     nbparams_params->two_k_rf          = nbp->two_k_rf;
 458     nbparams_params->vdwType           = nbp->vdwType;
 459     nbparams_params->vdw_switch        = nbp->vdw_switch;
 460 }
 461
 462 /*! \brief Enqueues a wait for event completion.
 463  *
 464  * Then it releases the event and sets it to 0.
 465  * Don't use this function when more than one wait will be issued for the event.
 466  * Equivalent to Cuda Stream Sync. */
 467 static void sync_ocl_event(cl_command_queue stream, cl_event* ocl_event)
 468 {
 469     cl_int gmx_unused cl_error;
 470
 471     /* Enqueue wait */
 472     cl_error = clEnqueueBarrierWithWaitList(stream, 1, ocl_event, nullptr);
 473     GMX_RELEASE_ASSERT(CL_SUCCESS == cl_error, ocl_get_error_string(cl_error).c_str());
 474
 475     /* Release event and reset it to 0. It is ok to release it as enqueuewaitforevents performs implicit retain for events. */
 476     cl_error = clReleaseEvent(*ocl_event);
 477     GMX_ASSERT(cl_error == CL_SUCCESS,
 478                ("clReleaseEvent failed: " + ocl_get_error_string(cl_error)).c_str());
 479     *ocl_event = nullptr;
 480 }
 481
 482 /*! \brief Launch asynchronously the xq buffer host to device copy. */
 483 void gpu_copy_xq_to_gpu(NbnxmGpu* nb, const nbnxn_atomdata_t* nbatom, const AtomLocality atomLocality)
 484 {
 485     GMX_ASSERT(nb, "Need a valid nbnxn_gpu object");
 486
 487     const InteractionLocality iloc = gpuAtomToInteractionLocality(atomLocality);
 488
 489     /* local/nonlocal offset and length used for xq and f */
 490     int adat_begin, adat_len;
 491
 492     cl_atomdata_t*      adat         = nb->atdat;
 493     gpu_plist*          plist        = nb->plist[iloc];
 494     cl_timers_t*        t            = nb->timers;
 495     const DeviceStream& deviceStream = *nb->deviceStreams[iloc];
 496
 497     bool bDoTime = nb->bDoTime;
 498
 499     /* Don't launch the non-local H2D copy if there is no dependent
 500        work to do: neither non-local nor other (e.g. bonded) work
 501        to do that has as input the nbnxn coordinates.
 502        Doing the same for the local kernel is more complicated, since the
 503        local part of the force array also depends on the non-local kernel.
 504        So to avoid complicating the code and to reduce the risk of bugs,
 505        we always call the local local x+q copy (and the rest of the local
 506        work in nbnxn_gpu_launch_kernel().
 507      */
 508     if ((iloc == InteractionLocality::NonLocal) && !haveGpuShortRangeWork(*nb, iloc))
 509     {
 510         plist->haveFreshList = false;
 511
 512         return;
 513     }
 514
 515     /* calculate the atom data index range based on locality */
 516     if (atomLocality == AtomLocality::Local)
 517     {
 518         adat_begin = 0;
 519         adat_len   = adat->natoms_local;
 520     }
 521     else
 522     {
 523         adat_begin = adat->natoms_local;
 524         adat_len   = adat->natoms - adat->natoms_local;
 525     }
 526
 527     /* beginning of timed HtoD section */
 528     if (bDoTime)
 529     {
 530         t->xf[atomLocality].nb_h2d.openTimingRegion(deviceStream);
 531     }
 532
 533     /* HtoD x, q */
 534     GMX_ASSERT(sizeof(float) == sizeof(*nbatom->x().data()),
 535                "The size of the xyzq buffer element should be equal to the size of float4.");
 536     copyToDeviceBuffer(&adat->xq, nbatom->x().data() + adat_begin * 4, adat_begin * 4, adat_len * 4,
 537                        deviceStream, GpuApiCallBehavior::Async,
 538                        bDoTime ? t->xf[atomLocality].nb_h2d.fetchNextEvent() : nullptr);
 539
 540     if (bDoTime)
 541     {
 542         t->xf[atomLocality].nb_h2d.closeTimingRegion(deviceStream);
 543     }
 544
 545     /* When we get here all misc operations issues in the local stream as well as
 546        the local xq H2D are done,
 547        so we record that in the local stream and wait for it in the nonlocal one. */
 548     if (nb->bUseTwoStreams)
 549     {
 550         if (iloc == InteractionLocality::Local)
 551         {
 552             cl_int gmx_used_in_debug cl_error = clEnqueueMarkerWithWaitList(
 553                     deviceStream.stream(), 0, nullptr, &(nb->misc_ops_and_local_H2D_done));
 554             GMX_ASSERT(cl_error == CL_SUCCESS,
 555                        ("clEnqueueMarkerWithWaitList failed: " + ocl_get_error_string(cl_error)).c_str());
 556
 557             /* Based on the v1.2 section 5.13 of the OpenCL spec, a flush is needed
 558              * in the local stream in order to be able to sync with the above event
 559              * from the non-local stream.
 560              */
 561             cl_error = clFlush(deviceStream.stream());
 562             GMX_ASSERT(cl_error == CL_SUCCESS,
 563                        ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
 564         }
 565         else
 566         {
 567             sync_ocl_event(deviceStream.stream(), &(nb->misc_ops_and_local_H2D_done));
 568         }
 569     }
 570 }
 571
 572
 573 /*! \brief Launch GPU kernel
 574
 575    As we execute nonbonded workload in separate queues, before launching
 576    the kernel we need to make sure that he following operations have completed:
 577    - atomdata allocation and related H2D transfers (every nstlist step);
 578    - pair list H2D transfer (every nstlist step);
 579    - shift vector H2D transfer (every nstlist step);
 580    - force (+shift force and energy) output clearing (every step).
 581
 582    These operations are issued in the local queue at the beginning of the step
 583    and therefore always complete before the local kernel launch. The non-local
 584    kernel is launched after the local on the same device/context, so this is
 585    inherently scheduled after the operations in the local stream (including the
 586    above "misc_ops").
 587    However, for the sake of having a future-proof implementation, we use the
 588    misc_ops_done event to record the point in time when the above  operations
 589    are finished and synchronize with this event in the non-local stream.
 590  */
 591 void gpu_launch_kernel(NbnxmGpu* nb, const gmx::StepWorkload& stepWork, const Nbnxm::InteractionLocality iloc)
 592 {
 593     cl_atomdata_t*      adat         = nb->atdat;
 594     NBParamGpu*         nbp          = nb->nbparam;
 595     gpu_plist*          plist        = nb->plist[iloc];
 596     cl_timers_t*        t            = nb->timers;
 597     const DeviceStream& deviceStream = *nb->deviceStreams[iloc];
 598
 599     bool bDoTime = nb->bDoTime;
 600
 601     cl_nbparam_params_t nbparams_params;
 602
 603     /* Don't launch the non-local kernel if there is no work to do.
 604        Doing the same for the local kernel is more complicated, since the
 605        local part of the force array also depends on the non-local kernel.
 606        So to avoid complicating the code and to reduce the risk of bugs,
 607        we always call the local kernel and later (not in
 608        this function) the stream wait, local f copyback and the f buffer
 609        clearing. All these operations, except for the local interaction kernel,
 610        are needed for the non-local interactions. The skip of the local kernel
 611        call is taken care of later in this function. */
 612     if (canSkipNonbondedWork(*nb, iloc))
 613     {
 614         plist->haveFreshList = false;
 615
 616         return;
 617     }
 618
 619     if (nbp->useDynamicPruning && plist->haveFreshList)
 620     {
 621         /* Prunes for rlistOuter and rlistInner, sets plist->haveFreshList=false
 622            (that's the way the timing accounting can distinguish between
 623            separate prune kernel and combined force+prune).
 624          */
 625         Nbnxm::gpu_launch_kernel_pruneonly(nb, iloc, 1);
 626     }
 627
 628     if (plist->nsci == 0)
 629     {
 630         /* Don't launch an empty local kernel (is not allowed with OpenCL).
 631          */
 632         return;
 633     }
 634
 635     /* beginning of timed nonbonded calculation section */
 636     if (bDoTime)
 637     {
 638         t->interaction[iloc].nb_k.openTimingRegion(deviceStream);
 639     }
 640
 641     /* kernel launch config */
 642
 643     KernelLaunchConfig config;
 644     config.sharedMemorySize = calc_shmem_required_nonbonded(nbp->vdwType, nb->bPrefetchLjParam);
 645     config.blockSize[0]     = c_clSize;
 646     config.blockSize[1]     = c_clSize;
 647     config.gridSize[0]      = plist->nsci;
 648
 649     validate_global_work_size(config, 3, &nb->deviceContext_->deviceInfo());
 650
 651     if (debug)
 652     {
 653         fprintf(debug,
 654                 "Non-bonded GPU launch configuration:\n\tLocal work size: %zux%zux%zu\n\t"
 655                 "Global work size : %zux%zu\n\t#Super-clusters/clusters: %d/%d (%d)\n",
 656                 config.blockSize[0], config.blockSize[1], config.blockSize[2],
 657                 config.blockSize[0] * config.gridSize[0], config.blockSize[1] * config.gridSize[1],
 658                 plist->nsci * c_nbnxnGpuNumClusterPerSupercluster,
 659                 c_nbnxnGpuNumClusterPerSupercluster, plist->na_c);
 660     }
 661
 662     fillin_ocl_structures(nbp, &nbparams_params);
 663
 664     auto*          timingEvent  = bDoTime ? t->interaction[iloc].nb_k.fetchNextEvent() : nullptr;
 665     constexpr char kernelName[] = "k_calc_nb";
 666     const auto     kernel =
 667             select_nbnxn_kernel(nb, nbp->elecType, nbp->vdwType, stepWork.computeEnergy,
 668                                 (plist->haveFreshList && !nb->timers->interaction[iloc].didPrune));
 669
 670
 671     // The OpenCL kernel takes int as second to last argument because bool is
 672     // not supported as a kernel argument type (sizeof(bool) is implementation defined).
 673     const int computeFshift = static_cast<int>(stepWork.computeVirial);
 674     if (useLjCombRule(nb->nbparam->vdwType))
 675     {
 676         const auto kernelArgs = prepareGpuKernelArguments(
 677                 kernel, config, &nbparams_params, &adat->xq, &adat->f, &adat->e_lj, &adat->e_el,
 678                 &adat->fshift, &adat->lj_comb, &adat->shift_vec, &nbp->nbfp, &nbp->nbfp_comb,
 679                 &nbp->coulomb_tab, &plist->sci, &plist->cj4, &plist->excl, &computeFshift);
 680
 681         launchGpuKernel(kernel, config, deviceStream, timingEvent, kernelName, kernelArgs);
 682     }
 683     else
 684     {
 685         const auto kernelArgs = prepareGpuKernelArguments(
 686                 kernel, config, &adat->ntypes, &nbparams_params, &adat->xq, &adat->f, &adat->e_lj,
 687                 &adat->e_el, &adat->fshift, &adat->atom_types, &adat->shift_vec, &nbp->nbfp, &nbp->nbfp_comb,
 688                 &nbp->coulomb_tab, &plist->sci, &plist->cj4, &plist->excl, &computeFshift);
 689         launchGpuKernel(kernel, config, deviceStream, timingEvent, kernelName, kernelArgs);
 690     }
 691
 692     if (bDoTime)
 693     {
 694         t->interaction[iloc].nb_k.closeTimingRegion(deviceStream);
 695     }
 696 }
 697
 698
 699 /*! \brief Calculates the amount of shared memory required by the prune kernel.
 700  *
 701  *  Note that for the sake of simplicity we use the CUDA terminology "shared memory"
 702  *  for OpenCL local memory.
 703  *
 704  * \param[in] num_threads_z cj4 concurrency equal to the number of threads/work items in the 3-rd
 705  * dimension. \returns   the amount of local memory in bytes required by the pruning kernel
 706  */
 707 static inline int calc_shmem_required_prune(const int num_threads_z)
 708 {
 709     int shmem;
 710
 711     /* i-atom x in shared memory (for convenience we load all 4 components including q) */
 712     shmem = c_nbnxnGpuNumClusterPerSupercluster * c_clSize * sizeof(float) * 4;
 713     /* cj in shared memory, for each warp separately
 714      * Note: only need to load once per wavefront, but to keep the code simple,
 715      * for now we load twice on AMD.
 716      */
 717     shmem += num_threads_z * c_nbnxnGpuClusterpairSplit * c_nbnxnGpuJgroupSize * sizeof(int);
 718     /* Warp vote, requires one uint per warp/32 threads per block. */
 719     shmem += sizeof(cl_uint) * 2 * num_threads_z;
 720
 721     return shmem;
 722 }
 723
 724 /*! \brief
 725  * Launch the pairlist prune only kernel for the given locality.
 726  * \p numParts tells in how many parts, i.e. calls the list will be pruned.
 727  */
 728 void gpu_launch_kernel_pruneonly(NbnxmGpu* nb, const InteractionLocality iloc, const int numParts)
 729 {
 730     cl_atomdata_t*      adat         = nb->atdat;
 731     NBParamGpu*         nbp          = nb->nbparam;
 732     gpu_plist*          plist        = nb->plist[iloc];
 733     cl_timers_t*        t            = nb->timers;
 734     const DeviceStream& deviceStream = *nb->deviceStreams[iloc];
 735     bool                bDoTime      = nb->bDoTime;
 736
 737     if (plist->haveFreshList)
 738     {
 739         GMX_ASSERT(numParts == 1, "With first pruning we expect 1 part");
 740
 741         /* Set rollingPruningNumParts to signal that it is not set */
 742         plist->rollingPruningNumParts = 0;
 743         plist->rollingPruningPart     = 0;
 744     }
 745     else
 746     {
 747         if (plist->rollingPruningNumParts == 0)
 748         {
 749             plist->rollingPruningNumParts = numParts;
 750         }
 751         else
 752         {
 753             GMX_ASSERT(numParts == plist->rollingPruningNumParts,
 754                        "It is not allowed to change numParts in between list generation steps");
 755         }
 756     }
 757
 758     /* Use a local variable for part and update in plist, so we can return here
 759      * without duplicating the part increment code.
 760      */
 761     int part = plist->rollingPruningPart;
 762
 763     plist->rollingPruningPart++;
 764     if (plist->rollingPruningPart >= plist->rollingPruningNumParts)
 765     {
 766         plist->rollingPruningPart = 0;
 767     }
 768
 769     /* Compute the number of list entries to prune in this pass */
 770     int numSciInPart = (plist->nsci - part) / numParts;
 771
 772     /* Don't launch the kernel if there is no work to do. */
 773     if (numSciInPart <= 0)
 774     {
 775         plist->haveFreshList = false;
 776
 777         return;
 778     }
 779
 780     GpuRegionTimer* timer = nullptr;
 781     if (bDoTime)
 782     {
 783         timer = &(plist->haveFreshList ? t->interaction[iloc].prune_k : t->interaction[iloc].rollingPrune_k);
 784     }
 785
 786     /* beginning of timed prune calculation section */
 787     if (bDoTime)
 788     {
 789         timer->openTimingRegion(deviceStream);
 790     }
 791
 792     /* Kernel launch config:
 793      * - The thread block dimensions match the size of i-clusters, j-clusters,
 794      *   and j-cluster concurrency, in x, y, and z, respectively.
 795      * - The 1D block-grid contains as many blocks as super-clusters.
 796      */
 797     int num_threads_z = c_oclPruneKernelJ4ConcurrencyDEFAULT;
 798
 799
 800     /* kernel launch config */
 801     KernelLaunchConfig config;
 802     config.sharedMemorySize = calc_shmem_required_prune(num_threads_z);
 803     config.blockSize[0]     = c_clSize;
 804     config.blockSize[1]     = c_clSize;
 805     config.blockSize[2]     = num_threads_z;
 806     config.gridSize[0]      = numSciInPart;
 807
 808     validate_global_work_size(config, 3, &nb->deviceContext_->deviceInfo());
 809
 810     if (debug)
 811     {
 812         fprintf(debug,
 813                 "Pruning GPU kernel launch configuration:\n\tLocal work size: %zux%zux%zu\n\t"
 814                 "\tGlobal work size: %zux%zu\n\t#Super-clusters/clusters: %d/%d (%d)\n"
 815                 "\tShMem: %zu\n",
 816                 config.blockSize[0], config.blockSize[1], config.blockSize[2],
 817                 config.blockSize[0] * config.gridSize[0], config.blockSize[1] * config.gridSize[1],
 818                 plist->nsci * c_nbnxnGpuNumClusterPerSupercluster,
 819                 c_nbnxnGpuNumClusterPerSupercluster, plist->na_c, config.sharedMemorySize);
 820     }
 821
 822     cl_nbparam_params_t nbparams_params;
 823     fillin_ocl_structures(nbp, &nbparams_params);
 824
 825     auto*          timingEvent  = bDoTime ? timer->fetchNextEvent() : nullptr;
 826     constexpr char kernelName[] = "k_pruneonly";
 827     const auto     pruneKernel  = selectPruneKernel(nb->kernel_pruneonly, plist->haveFreshList);
 828     const auto     kernelArgs   = prepareGpuKernelArguments(pruneKernel, config, &nbparams_params,
 829                                                       &adat->xq, &adat->shift_vec, &plist->sci,
 830                                                       &plist->cj4, &plist->imask, &numParts, &part);
 831     launchGpuKernel(pruneKernel, config, deviceStream, timingEvent, kernelName, kernelArgs);
 832
 833     if (plist->haveFreshList)
 834     {
 835         plist->haveFreshList = false;
 836         /* Mark that pruning has been done */
 837         nb->timers->interaction[iloc].didPrune = true;
 838     }
 839     else
 840     {
 841         /* Mark that rolling pruning has been done */
 842         nb->timers->interaction[iloc].didRollingPrune = true;
 843     }
 844
 845     if (bDoTime)
 846     {
 847         timer->closeTimingRegion(deviceStream);
 848     }
 849 }
 850
 851 /*! \brief
 852  * Launch asynchronously the download of nonbonded forces from the GPU
 853  * (and energies/shift forces if required).
 854  */
 855 void gpu_launch_cpyback(NbnxmGpu*                nb,
 856                         struct nbnxn_atomdata_t* nbatom,
 857                         const gmx::StepWorkload& stepWork,
 858                         const AtomLocality       aloc)
 859 {
 860     GMX_ASSERT(nb, "Need a valid nbnxn_gpu object");
 861
 862     cl_int gmx_unused cl_error;
 863     int               adat_begin, adat_len; /* local/nonlocal offset and length used for xq and f */
 864
 865     /* determine interaction locality from atom locality */
 866     const InteractionLocality iloc = gpuAtomToInteractionLocality(aloc);
 867
 868     cl_atomdata_t*      adat         = nb->atdat;
 869     cl_timers_t*        t            = nb->timers;
 870     bool                bDoTime      = nb->bDoTime;
 871     const DeviceStream& deviceStream = *nb->deviceStreams[iloc];
 872
 873     /* don't launch non-local copy-back if there was no non-local work to do */
 874     if ((iloc == InteractionLocality::NonLocal) && !haveGpuShortRangeWork(*nb, iloc))
 875     {
 876         /* TODO An alternative way to signal that non-local work is
 877            complete is to use a clEnqueueMarker+clEnqueueBarrier
 878            pair. However, the use of bNonLocalStreamActive has the
 879            advantage of being local to the host, so probably minimizes
 880            overhead. Curiously, for NVIDIA OpenCL with an empty-domain
 881            test case, overall simulation performance was higher with
 882            the API calls, but this has not been tested on AMD OpenCL,
 883            so could be worth considering in future. */
 884         nb->bNonLocalStreamActive = CL_FALSE;
 885         return;
 886     }
 887
 888     getGpuAtomRange(adat, aloc, &adat_begin, &adat_len);
 889
 890     /* beginning of timed D2H section */
 891     if (bDoTime)
 892     {
 893         t->xf[aloc].nb_d2h.openTimingRegion(deviceStream);
 894     }
 895
 896     /* With DD the local D2H transfer can only start after the non-local
 897        has been launched. */
 898     if (iloc == InteractionLocality::Local && nb->bNonLocalStreamActive)
 899     {
 900         sync_ocl_event(deviceStream.stream(), &(nb->nonlocal_done));
 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(&nbatom->out[0].f[adat_begin * DIM], &adat->f, adat_begin * DIM,
 907                          adat_len * DIM, deviceStream, GpuApiCallBehavior::Async,
 908                          bDoTime ? t->xf[aloc].nb_d2h.fetchNextEvent() : nullptr);
 909
 910     /* kick off work */
 911     cl_error = clFlush(deviceStream.stream());
 912     GMX_ASSERT(cl_error == CL_SUCCESS, ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
 913
 914     /* After the non-local D2H is launched the nonlocal_done event can be
 915        recorded which signals that the local D2H can proceed. This event is not
 916        placed after the non-local kernel because we first need the non-local
 917        data back first. */
 918     if (iloc == InteractionLocality::NonLocal)
 919     {
 920         cl_error = clEnqueueMarkerWithWaitList(deviceStream.stream(), 0, nullptr, &(nb->nonlocal_done));
 921         GMX_ASSERT(cl_error == CL_SUCCESS,
 922                    ("clEnqueueMarkerWithWaitList failed: " + ocl_get_error_string(cl_error)).c_str());
 923         nb->bNonLocalStreamActive = CL_TRUE;
 924     }
 925
 926     /* only transfer energies in the local stream */
 927     if (iloc == InteractionLocality::Local)
 928     {
 929         /* DtoH fshift when virial is needed */
 930         if (stepWork.computeVirial)
 931         {
 932             GMX_ASSERT(sizeof(*nb->nbst.fshift) == DIM * sizeof(float),
 933                        "Sizes of host- and device-side shift vector elements should be the same.");
 934             copyFromDeviceBuffer(reinterpret_cast<float*>(nb->nbst.fshift), &adat->fshift, 0,
 935                                  SHIFTS * DIM, deviceStream, GpuApiCallBehavior::Async,
 936                                  bDoTime ? t->xf[aloc].nb_d2h.fetchNextEvent() : nullptr);
 937         }
 938
 939         /* DtoH energies */
 940         if (stepWork.computeEnergy)
 941         {
 942             GMX_ASSERT(sizeof(*nb->nbst.e_lj) == sizeof(float),
 943                        "Sizes of host- and device-side LJ energy terms should be the same.");
 944             copyFromDeviceBuffer(nb->nbst.e_lj, &adat->e_lj, 0, 1, deviceStream, GpuApiCallBehavior::Async,
 945                                  bDoTime ? t->xf[aloc].nb_d2h.fetchNextEvent() : nullptr);
 946             GMX_ASSERT(sizeof(*nb->nbst.e_el) == sizeof(float),
 947                        "Sizes of host- and device-side electrostatic energy terms should be the "
 948                        "same.");
 949             copyFromDeviceBuffer(nb->nbst.e_el, &adat->e_el, 0, 1, deviceStream, GpuApiCallBehavior::Async,
 950                                  bDoTime ? t->xf[aloc].nb_d2h.fetchNextEvent() : nullptr);
 951         }
 952     }
 953
 954     if (bDoTime)
 955     {
 956         t->xf[aloc].nb_d2h.closeTimingRegion(deviceStream);
 957     }
 958 }
 959
 960 } // namespace Nbnxm