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 asynchronously the xq buffer host to device copy. */
 524 void gpu_copy_xq_to_gpu(NbnxmGpu* nb, const nbnxn_atomdata_t* nbatom, const AtomLocality atomLocality)
 525 {
 526     GMX_ASSERT(nb, "Need a valid nbnxn_gpu object");
 527
 528     const InteractionLocality iloc = gpuAtomToInteractionLocality(atomLocality);
 529
 530     NBAtomData*         adat         = nb->atdat;
 531     gpu_plist*          plist        = nb->plist[iloc];
 532     cl_timers_t*        t            = nb->timers;
 533     const DeviceStream& deviceStream = *nb->deviceStreams[iloc];
 534
 535     bool bDoTime = nb->bDoTime;
 536
 537     /* Don't launch the non-local H2D copy if there is no dependent
 538        work to do: neither non-local nor other (e.g. bonded) work
 539        to do that has as input the nbnxn coordinates.
 540        Doing the same for the local kernel is more complicated, since the
 541        local part of the force array also depends on the non-local kernel.
 542        So to avoid complicating the code and to reduce the risk of bugs,
 543        we always call the local local x+q copy (and the rest of the local
 544        work in nbnxn_gpu_launch_kernel().
 545      */
 546     if ((iloc == InteractionLocality::NonLocal) && !haveGpuShortRangeWork(*nb, iloc))
 547     {
 548         plist->haveFreshList = false;
 549
 550         // The event is marked for Local interactions unconditionally,
 551         // so it has to be released here because of the early return
 552         // for NonLocal interactions.
 553         nb->misc_ops_and_local_H2D_done.reset();
 554
 555         return;
 556     }
 557
 558     /* local/nonlocal offset and length used for xq and f */
 559     auto atomsRange = getGpuAtomRange(adat, atomLocality);
 560
 561     /* beginning of timed HtoD section */
 562     if (bDoTime)
 563     {
 564         t->xf[atomLocality].nb_h2d.openTimingRegion(deviceStream);
 565     }
 566
 567     /* HtoD x, q */
 568     static_assert(sizeof(float) == sizeof(*nbatom->x().data()),
 569                   "The size of the xyzq buffer element should be equal to the size of float4.");
 570     copyToDeviceBuffer(&adat->xq,
 571                        reinterpret_cast<const Float4*>(nbatom->x().data()) + atomsRange.begin(),
 572                        atomsRange.begin(),
 573                        atomsRange.size(),
 574                        deviceStream,
 575                        GpuApiCallBehavior::Async,
 576                        bDoTime ? t->xf[atomLocality].nb_h2d.fetchNextEvent() : nullptr);
 577
 578     if (bDoTime)
 579     {
 580         t->xf[atomLocality].nb_h2d.closeTimingRegion(deviceStream);
 581     }
 582
 583     /* When we get here all misc operations issued in the local stream as well as
 584        the local xq H2D are done,
 585        so we record that in the local stream and wait for it in the nonlocal one.
 586        This wait needs to precede any PP tasks, bonded or nonbonded, that may
 587        compute on interactions between local and nonlocal atoms.
 588      */
 589     nbnxnInsertNonlocalGpuDependency(nb, iloc);
 590 }
 591
 592
 593 /*! \brief Launch GPU kernel
 594
 595    As we execute nonbonded workload in separate queues, before launching
 596    the kernel we need to make sure that he following operations have completed:
 597    - atomdata allocation and related H2D transfers (every nstlist step);
 598    - pair list H2D transfer (every nstlist step);
 599    - shift vector H2D transfer (every nstlist step);
 600    - force (+shift force and energy) output clearing (every step).
 601
 602    These operations are issued in the local queue at the beginning of the step
 603    and therefore always complete before the local kernel launch. The non-local
 604    kernel is launched after the local on the same device/context, so this is
 605    inherently scheduled after the operations in the local stream (including the
 606    above "misc_ops").
 607    However, for the sake of having a future-proof implementation, we use the
 608    misc_ops_done event to record the point in time when the above  operations
 609    are finished and synchronize with this event in the non-local stream.
 610  */
 611 void gpu_launch_kernel(NbnxmGpu* nb, const gmx::StepWorkload& stepWork, const Nbnxm::InteractionLocality iloc)
 612 {
 613     NBAtomData*         adat         = nb->atdat;
 614     NBParamGpu*         nbp          = nb->nbparam;
 615     gpu_plist*          plist        = nb->plist[iloc];
 616     cl_timers_t*        t            = nb->timers;
 617     const DeviceStream& deviceStream = *nb->deviceStreams[iloc];
 618
 619     bool bDoTime = nb->bDoTime;
 620
 621     cl_nbparam_params_t nbparams_params;
 622
 623     /* Don't launch the non-local kernel if there is no work to do.
 624        Doing the same for the local kernel is more complicated, since the
 625        local part of the force array also depends on the non-local kernel.
 626        So to avoid complicating the code and to reduce the risk of bugs,
 627        we always call the local kernel and later (not in
 628        this function) the stream wait, local f copyback and the f buffer
 629        clearing. All these operations, except for the local interaction kernel,
 630        are needed for the non-local interactions. The skip of the local kernel
 631        call is taken care of later in this function. */
 632     if (canSkipNonbondedWork(*nb, iloc))
 633     {
 634         plist->haveFreshList = false;
 635
 636         return;
 637     }
 638
 639     if (nbp->useDynamicPruning && plist->haveFreshList)
 640     {
 641         /* Prunes for rlistOuter and rlistInner, sets plist->haveFreshList=false
 642            (that's the way the timing accounting can distinguish between
 643            separate prune kernel and combined force+prune).
 644          */
 645         Nbnxm::gpu_launch_kernel_pruneonly(nb, iloc, 1);
 646     }
 647
 648     if (plist->nsci == 0)
 649     {
 650         /* Don't launch an empty local kernel (is not allowed with OpenCL).
 651          */
 652         return;
 653     }
 654
 655     /* beginning of timed nonbonded calculation section */
 656     if (bDoTime)
 657     {
 658         t->interaction[iloc].nb_k.openTimingRegion(deviceStream);
 659     }
 660
 661     /* kernel launch config */
 662
 663     KernelLaunchConfig config;
 664     config.sharedMemorySize = calc_shmem_required_nonbonded(nbp->vdwType, nb->bPrefetchLjParam);
 665     config.blockSize[0]     = c_clSize;
 666     config.blockSize[1]     = c_clSize;
 667     config.gridSize[0]      = plist->nsci;
 668
 669     validate_global_work_size(config, 3, &nb->deviceContext_->deviceInfo());
 670
 671     if (debug)
 672     {
 673         fprintf(debug,
 674                 "Non-bonded GPU launch configuration:\n\tLocal work size: %zux%zux%zu\n\t"
 675                 "Global work size : %zux%zu\n\t#Super-clusters/clusters: %d/%d (%d)\n",
 676                 config.blockSize[0],
 677                 config.blockSize[1],
 678                 config.blockSize[2],
 679                 config.blockSize[0] * config.gridSize[0],
 680                 config.blockSize[1] * config.gridSize[1],
 681                 plist->nsci * c_nbnxnGpuNumClusterPerSupercluster,
 682                 c_nbnxnGpuNumClusterPerSupercluster,
 683                 plist->na_c);
 684     }
 685
 686     fillin_ocl_structures(nbp, &nbparams_params);
 687
 688     auto*          timingEvent  = bDoTime ? t->interaction[iloc].nb_k.fetchNextEvent() : nullptr;
 689     constexpr char kernelName[] = "k_calc_nb";
 690     const auto     kernel =
 691             select_nbnxn_kernel(nb,
 692                                 nbp->elecType,
 693                                 nbp->vdwType,
 694                                 stepWork.computeEnergy,
 695                                 (plist->haveFreshList && !nb->timers->interaction[iloc].didPrune));
 696
 697
 698     // The OpenCL kernel takes int as second to last argument because bool is
 699     // not supported as a kernel argument type (sizeof(bool) is implementation defined).
 700     const int computeFshift = static_cast<int>(stepWork.computeVirial);
 701     if (useLjCombRule(nb->nbparam->vdwType))
 702     {
 703         const auto kernelArgs = prepareGpuKernelArguments(kernel,
 704                                                           config,
 705                                                           &nbparams_params,
 706                                                           &adat->xq,
 707                                                           &adat->f,
 708                                                           &adat->eLJ,
 709                                                           &adat->eElec,
 710                                                           &adat->fShift,
 711                                                           &adat->ljComb,
 712                                                           &adat->shiftVec,
 713                                                           &nbp->nbfp,
 714                                                           &nbp->nbfp_comb,
 715                                                           &nbp->coulomb_tab,
 716                                                           &plist->sci,
 717                                                           &plist->cj4,
 718                                                           &plist->excl,
 719                                                           &computeFshift);
 720
 721         launchGpuKernel(kernel, config, deviceStream, timingEvent, kernelName, kernelArgs);
 722     }
 723     else
 724     {
 725         const auto kernelArgs = prepareGpuKernelArguments(kernel,
 726                                                           config,
 727                                                           &adat->numTypes,
 728                                                           &nbparams_params,
 729                                                           &adat->xq,
 730                                                           &adat->f,
 731                                                           &adat->eLJ,
 732                                                           &adat->eElec,
 733                                                           &adat->fShift,
 734                                                           &adat->atomTypes,
 735                                                           &adat->shiftVec,
 736                                                           &nbp->nbfp,
 737                                                           &nbp->nbfp_comb,
 738                                                           &nbp->coulomb_tab,
 739                                                           &plist->sci,
 740                                                           &plist->cj4,
 741                                                           &plist->excl,
 742                                                           &computeFshift);
 743         launchGpuKernel(kernel, config, deviceStream, timingEvent, kernelName, kernelArgs);
 744     }
 745
 746     if (bDoTime)
 747     {
 748         t->interaction[iloc].nb_k.closeTimingRegion(deviceStream);
 749     }
 750 }
 751
 752
 753 /*! \brief Calculates the amount of shared memory required by the prune kernel.
 754  *
 755  *  Note that for the sake of simplicity we use the CUDA terminology "shared memory"
 756  *  for OpenCL local memory.
 757  *
 758  * \param[in] num_threads_z cj4 concurrency equal to the number of threads/work items in the 3-rd
 759  * dimension. \returns   the amount of local memory in bytes required by the pruning kernel
 760  */
 761 static inline int calc_shmem_required_prune(const int num_threads_z)
 762 {
 763     int shmem;
 764
 765     /* i-atom x in shared memory (for convenience we load all 4 components including q) */
 766     shmem = c_nbnxnGpuNumClusterPerSupercluster * c_clSize * sizeof(float) * 4;
 767     /* cj in shared memory, for each warp separately
 768      * Note: only need to load once per wavefront, but to keep the code simple,
 769      * for now we load twice on AMD.
 770      */
 771     shmem += num_threads_z * c_nbnxnGpuClusterpairSplit * c_nbnxnGpuJgroupSize * sizeof(int);
 772     /* Warp vote, requires one uint per warp/32 threads per block. */
 773     shmem += sizeof(cl_uint) * 2 * num_threads_z;
 774
 775     return shmem;
 776 }
 777
 778 /*! \brief
 779  * Launch the pairlist prune only kernel for the given locality.
 780  * \p numParts tells in how many parts, i.e. calls the list will be pruned.
 781  */
 782 void gpu_launch_kernel_pruneonly(NbnxmGpu* nb, const InteractionLocality iloc, const int numParts)
 783 {
 784     NBAtomData*         adat         = nb->atdat;
 785     NBParamGpu*         nbp          = nb->nbparam;
 786     gpu_plist*          plist        = nb->plist[iloc];
 787     cl_timers_t*        t            = nb->timers;
 788     const DeviceStream& deviceStream = *nb->deviceStreams[iloc];
 789     bool                bDoTime      = nb->bDoTime;
 790
 791     if (plist->haveFreshList)
 792     {
 793         GMX_ASSERT(numParts == 1, "With first pruning we expect 1 part");
 794
 795         /* Set rollingPruningNumParts to signal that it is not set */
 796         plist->rollingPruningNumParts = 0;
 797         plist->rollingPruningPart     = 0;
 798     }
 799     else
 800     {
 801         if (plist->rollingPruningNumParts == 0)
 802         {
 803             plist->rollingPruningNumParts = numParts;
 804         }
 805         else
 806         {
 807             GMX_ASSERT(numParts == plist->rollingPruningNumParts,
 808                        "It is not allowed to change numParts in between list generation steps");
 809         }
 810     }
 811
 812     /* Use a local variable for part and update in plist, so we can return here
 813      * without duplicating the part increment code.
 814      */
 815     int part = plist->rollingPruningPart;
 816
 817     plist->rollingPruningPart++;
 818     if (plist->rollingPruningPart >= plist->rollingPruningNumParts)
 819     {
 820         plist->rollingPruningPart = 0;
 821     }
 822
 823     /* Compute the number of list entries to prune in this pass */
 824     int numSciInPart = (plist->nsci - part) / numParts;
 825
 826     /* Don't launch the kernel if there is no work to do. */
 827     if (numSciInPart <= 0)
 828     {
 829         plist->haveFreshList = false;
 830
 831         return;
 832     }
 833
 834     GpuRegionTimer* timer = nullptr;
 835     if (bDoTime)
 836     {
 837         timer = &(plist->haveFreshList ? t->interaction[iloc].prune_k : t->interaction[iloc].rollingPrune_k);
 838     }
 839
 840     /* beginning of timed prune calculation section */
 841     if (bDoTime)
 842     {
 843         timer->openTimingRegion(deviceStream);
 844     }
 845
 846     /* Kernel launch config:
 847      * - The thread block dimensions match the size of i-clusters, j-clusters,
 848      *   and j-cluster concurrency, in x, y, and z, respectively.
 849      * - The 1D block-grid contains as many blocks as super-clusters.
 850      */
 851     int num_threads_z = c_pruneKernelJ4Concurrency;
 852     /* kernel launch config */
 853     KernelLaunchConfig config;
 854     config.sharedMemorySize = calc_shmem_required_prune(num_threads_z);
 855     config.blockSize[0]     = c_clSize;
 856     config.blockSize[1]     = c_clSize;
 857     config.blockSize[2]     = num_threads_z;
 858     config.gridSize[0]      = numSciInPart;
 859
 860     validate_global_work_size(config, 3, &nb->deviceContext_->deviceInfo());
 861
 862     if (debug)
 863     {
 864         fprintf(debug,
 865                 "Pruning GPU kernel launch configuration:\n\tLocal work size: %zux%zux%zu\n\t"
 866                 "\tGlobal work size: %zux%zu\n\t#Super-clusters/clusters: %d/%d (%d)\n"
 867                 "\tShMem: %zu\n",
 868                 config.blockSize[0],
 869                 config.blockSize[1],
 870                 config.blockSize[2],
 871                 config.blockSize[0] * config.gridSize[0],
 872                 config.blockSize[1] * config.gridSize[1],
 873                 plist->nsci * c_nbnxnGpuNumClusterPerSupercluster,
 874                 c_nbnxnGpuNumClusterPerSupercluster,
 875                 plist->na_c,
 876                 config.sharedMemorySize);
 877     }
 878
 879     cl_nbparam_params_t nbparams_params;
 880     fillin_ocl_structures(nbp, &nbparams_params);
 881
 882     auto*          timingEvent  = bDoTime ? timer->fetchNextEvent() : nullptr;
 883     constexpr char kernelName[] = "k_pruneonly";
 884     const auto     pruneKernel  = selectPruneKernel(nb->kernel_pruneonly, plist->haveFreshList);
 885     const auto     kernelArgs   = prepareGpuKernelArguments(pruneKernel,
 886                                                       config,
 887                                                       &nbparams_params,
 888                                                       &adat->xq,
 889                                                       &adat->shiftVec,
 890                                                       &plist->sci,
 891                                                       &plist->cj4,
 892                                                       &plist->imask,
 893                                                       &numParts,
 894                                                       &part);
 895     launchGpuKernel(pruneKernel, config, deviceStream, timingEvent, kernelName, kernelArgs);
 896
 897     if (plist->haveFreshList)
 898     {
 899         plist->haveFreshList = false;
 900         /* Mark that pruning has been done */
 901         nb->timers->interaction[iloc].didPrune = true;
 902     }
 903     else
 904     {
 905         /* Mark that rolling pruning has been done */
 906         nb->timers->interaction[iloc].didRollingPrune = true;
 907     }
 908
 909     if (bDoTime)
 910     {
 911         timer->closeTimingRegion(deviceStream);
 912     }
 913 }
 914
 915 /*! \brief
 916  * Launch asynchronously the download of nonbonded forces from the GPU
 917  * (and energies/shift forces if required).
 918  */
 919 void gpu_launch_cpyback(NbnxmGpu*                nb,
 920                         struct nbnxn_atomdata_t* nbatom,
 921                         const gmx::StepWorkload& stepWork,
 922                         const AtomLocality       atomLocality)
 923 {
 924     GMX_ASSERT(nb, "Need a valid nbnxn_gpu object");
 925
 926     cl_int gmx_unused cl_error;
 927
 928     /* determine interaction locality from atom locality */
 929     const InteractionLocality iloc = gpuAtomToInteractionLocality(atomLocality);
 930     GMX_ASSERT(iloc == InteractionLocality::Local
 931                        || (iloc == InteractionLocality::NonLocal && nb->bNonLocalStreamDoneMarked == false),
 932                "Non-local stream is indicating that the copy back event is enqueued at the "
 933                "beginning of the copy back function.");
 934
 935     NBAtomData*         adat         = nb->atdat;
 936     cl_timers_t*        t            = nb->timers;
 937     bool                bDoTime      = nb->bDoTime;
 938     const DeviceStream& deviceStream = *nb->deviceStreams[iloc];
 939
 940     /* don't launch non-local copy-back if there was no non-local work to do */
 941     if ((iloc == InteractionLocality::NonLocal) && !haveGpuShortRangeWork(*nb, iloc))
 942     {
 943         /* TODO An alternative way to signal that non-local work is
 944            complete is to use a clEnqueueMarker+clEnqueueBarrier
 945            pair. However, the use of bNonLocalStreamDoneMarked has the
 946            advantage of being local to the host, so probably minimizes
 947            overhead. Curiously, for NVIDIA OpenCL with an empty-domain
 948            test case, overall simulation performance was higher with
 949            the API calls, but this has not been tested on AMD OpenCL,
 950            so could be worth considering in future. */
 951         nb->bNonLocalStreamDoneMarked = false;
 952         return;
 953     }
 954
 955     /* local/nonlocal offset and length used for xq and f */
 956     auto atomsRange = getGpuAtomRange(adat, atomLocality);
 957
 958     /* beginning of timed D2H section */
 959     if (bDoTime)
 960     {
 961         t->xf[atomLocality].nb_d2h.openTimingRegion(deviceStream);
 962     }
 963
 964     /* With DD the local D2H transfer can only start after the non-local
 965        has been launched. */
 966     if (iloc == InteractionLocality::Local && nb->bNonLocalStreamDoneMarked)
 967     {
 968         nb->nonlocal_done.enqueueWaitEvent(deviceStream);
 969         nb->bNonLocalStreamDoneMarked = false;
 970     }
 971
 972     /* DtoH f */
 973     GMX_ASSERT(sizeof(*nbatom->out[0].f.data()) == sizeof(float),
 974                "The host force buffer should be in single precision to match device data size.");
 975     copyFromDeviceBuffer(reinterpret_cast<Float3*>(nbatom->out[0].f.data()) + atomsRange.begin(),
 976                          &adat->f,
 977                          atomsRange.begin(),
 978                          atomsRange.size(),
 979                          deviceStream,
 980                          GpuApiCallBehavior::Async,
 981                          bDoTime ? t->xf[atomLocality].nb_d2h.fetchNextEvent() : nullptr);
 982
 983     /* kick off work */
 984     cl_error = clFlush(deviceStream.stream());
 985     GMX_ASSERT(cl_error == CL_SUCCESS, ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
 986
 987     /* After the non-local D2H is launched the nonlocal_done event can be
 988        recorded which signals that the local D2H can proceed. This event is not
 989        placed after the non-local kernel because we first need the non-local
 990        data back first. */
 991     if (iloc == InteractionLocality::NonLocal)
 992     {
 993         nb->nonlocal_done.markEvent(deviceStream);
 994         nb->bNonLocalStreamDoneMarked = true;
 995     }
 996
 997     /* only transfer energies in the local stream */
 998     if (iloc == InteractionLocality::Local)
 999     {
1000         /* DtoH fshift when virial is needed */
1001         if (stepWork.computeVirial)
1002         {
1003             static_assert(
1004                     sizeof(*nb->nbst.fShift) == sizeof(Float3),
1005                     "Sizes of host- and device-side shift vector elements should be the same.");
1006             copyFromDeviceBuffer(nb->nbst.fShift,
1007                                  &adat->fShift,
1008                                  0,
1009                                  SHIFTS,
1010                                  deviceStream,
1011                                  GpuApiCallBehavior::Async,
1012                                  bDoTime ? t->xf[atomLocality].nb_d2h.fetchNextEvent() : nullptr);
1013         }
1014
1015         /* DtoH energies */
1016         if (stepWork.computeEnergy)
1017         {
1018             static_assert(sizeof(*nb->nbst.eLJ) == sizeof(float),
1019                           "Sizes of host- and device-side LJ energy terms should be the same.");
1020             copyFromDeviceBuffer(nb->nbst.eLJ,
1021                                  &adat->eLJ,
1022                                  0,
1023                                  1,
1024                                  deviceStream,
1025                                  GpuApiCallBehavior::Async,
1026                                  bDoTime ? t->xf[atomLocality].nb_d2h.fetchNextEvent() : nullptr);
1027             static_assert(sizeof(*nb->nbst.eElec) == sizeof(float),
1028                           "Sizes of host- and device-side electrostatic energy terms should be the "
1029                           "same.");
1030             copyFromDeviceBuffer(nb->nbst.eElec,
1031                                  &adat->eElec,
1032                                  0,
1033                                  1,
1034                                  deviceStream,
1035                                  GpuApiCallBehavior::Async,
1036                                  bDoTime ? t->xf[atomLocality].nb_d2h.fetchNextEvent() : nullptr);
1037         }
1038     }
1039
1040     if (bDoTime)
1041     {
1042         t->xf[atomLocality].nb_d2h.closeTimingRegion(deviceStream);
1043     }
1044 }
1045
1046 } // namespace Nbnxm