2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016,2017,2018,2019, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
36 * \brief Define OpenCL implementation of nbnxn_gpu.h
38 * \author Anca Hamuraru <anca@streamcomputing.eu>
39 * \author Teemu Virolainen <teemu@streamcomputing.eu>
40 * \author Dimitrios Karkoulis <dimitris.karkoulis@gmail.com>
41 * \author Szilárd Páll <pall.szilard@gmail.com>
42 * \ingroup module_mdlib
45 * - Add a static const cl_uint c_pruneKernelWorkDim / c_nbnxnKernelWorkDim = 3;
46 * - Rework the copying of OCL data structures done before every invocation of both
47 * nb and prune kernels (using fillin_ocl_structures); also consider at the same
48 * time calling clSetKernelArg only on the updated parameters (if tracking changed
49 * parameters is feasible);
50 * - Consider using the event_wait_list argument to clEnqueueNDRangeKernel to mark
51 * dependencies on the kernel launched: e.g. the non-local nb kernel's dependency
52 * on the misc_ops_and_local_H2D_done event could be better expressed this way.
54 * - Consider extracting common sections of the OpenCL and CUDA nbnxn logic, e.g:
55 * - in nbnxn_gpu_launch_kernel_pruneonly() the pre- and post-kernel launch logic
56 * is identical in the two implementations, so a 3-way split might allow sharing
70 #include "thread_mpi/atomic.h"
72 #include "gromacs/gpu_utils/gputraits_ocl.h"
73 #include "gromacs/gpu_utils/oclutils.h"
74 #include "gromacs/hardware/hw_info.h"
75 #include "gromacs/mdlib/force_flags.h"
76 #include "gromacs/mdlib/nb_verlet.h"
77 #include "gromacs/mdlib/nbnxn_consts.h"
78 #include "gromacs/mdlib/nbnxn_gpu.h"
79 #include "gromacs/mdlib/nbnxn_gpu_common.h"
80 #include "gromacs/mdlib/nbnxn_gpu_common_utils.h"
81 #include "gromacs/mdlib/nbnxn_gpu_data_mgmt.h"
82 #include "gromacs/mdlib/nbnxn_pairlist.h"
83 #include "gromacs/pbcutil/ishift.h"
84 #include "gromacs/timing/gpu_timing.h"
85 #include "gromacs/utility/cstringutil.h"
86 #include "gromacs/utility/fatalerror.h"
87 #include "gromacs/utility/gmxassert.h"
89 #include "nbnxn_ocl_internal.h"
90 #include "nbnxn_ocl_types.h"
93 /*! \brief Convenience constants */
95 static const int c_numClPerSupercl = c_nbnxnGpuNumClusterPerSupercluster;
96 static const int c_clSize = c_nbnxnGpuClusterSize;
100 /*! \brief Validates the input global work size parameter.
102 static inline void validate_global_work_size(const KernelLaunchConfig &config, int work_dim, const gmx_device_info_t *dinfo)
104 cl_uint device_size_t_size_bits;
105 cl_uint host_size_t_size_bits;
109 size_t global_work_size[3];
110 GMX_ASSERT(work_dim <= 3, "Not supporting hyper-grids just yet");
111 for (int i = 0; i < work_dim; i++)
113 global_work_size[i] = config.blockSize[i] * config.gridSize[i];
116 /* Each component of a global_work_size must not exceed the range given by the
117 sizeof(device size_t) for the device on which the kernel execution will
119 https://www.khronos.org/registry/cl/sdk/1.0/docs/man/xhtml/clEnqueueNDRangeKernel.html
121 device_size_t_size_bits = dinfo->adress_bits;
122 host_size_t_size_bits = static_cast<cl_uint>(sizeof(size_t) * 8);
124 /* If sizeof(host size_t) <= sizeof(device size_t)
125 => global_work_size components will always be valid
127 => get device limit for global work size and
128 compare it against each component of global_work_size.
130 if (host_size_t_size_bits > device_size_t_size_bits)
134 device_limit = (1ull << device_size_t_size_bits) - 1;
136 for (int i = 0; i < work_dim; i++)
138 if (global_work_size[i] > device_limit)
140 gmx_fatal(FARGS, "Watch out, the input system is too large to simulate!\n"
141 "The number of nonbonded work units (=number of super-clusters) exceeds the"
142 "device capabilities. Global work size limit exceeded (%zu > %zu)!",
143 global_work_size[i], device_limit);
149 /* Constant arrays listing non-bonded kernel function names. The arrays are
150 * organized in 2-dim arrays by: electrostatics and VDW type.
152 * Note that the row- and column-order of function pointers has to match the
153 * order of corresponding enumerated electrostatics and vdw types, resp.,
154 * defined in nbnxn_cuda_types.h.
157 /*! \brief Force-only kernel function names. */
158 static const char* nb_kfunc_noener_noprune_ptr[eelOclNR][evdwOclNR] =
160 { "nbnxn_kernel_ElecCut_VdwLJ_F_opencl", "nbnxn_kernel_ElecCut_VdwLJCombGeom_F_opencl", "nbnxn_kernel_ElecCut_VdwLJCombLB_F_opencl", "nbnxn_kernel_ElecCut_VdwLJFsw_F_opencl", "nbnxn_kernel_ElecCut_VdwLJPsw_F_opencl", "nbnxn_kernel_ElecCut_VdwLJEwCombGeom_F_opencl", "nbnxn_kernel_ElecCut_VdwLJEwCombLB_F_opencl" },
161 { "nbnxn_kernel_ElecRF_VdwLJ_F_opencl", "nbnxn_kernel_ElecRF_VdwLJCombGeom_F_opencl", "nbnxn_kernel_ElecRF_VdwLJCombLB_F_opencl", "nbnxn_kernel_ElecRF_VdwLJFsw_F_opencl", "nbnxn_kernel_ElecRF_VdwLJPsw_F_opencl", "nbnxn_kernel_ElecRF_VdwLJEwCombGeom_F_opencl", "nbnxn_kernel_ElecRF_VdwLJEwCombLB_F_opencl" },
162 { "nbnxn_kernel_ElecEwQSTab_VdwLJ_F_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJCombGeom_F_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJCombLB_F_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJFsw_F_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJPsw_F_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_F_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_F_opencl" },
163 { "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_F_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombGeom_F_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombLB_F_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_F_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_F_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_F_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_F_opencl" },
164 { "nbnxn_kernel_ElecEw_VdwLJ_F_opencl", "nbnxn_kernel_ElecEw_VdwLJCombGeom_F_opencl", "nbnxn_kernel_ElecEw_VdwLJCombLB_F_opencl", "nbnxn_kernel_ElecEw_VdwLJFsw_F_opencl", "nbnxn_kernel_ElecEw_VdwLJPsw_F_opencl", "nbnxn_kernel_ElecEw_VdwLJEwCombGeom_F_opencl", "nbnxn_kernel_ElecEw_VdwLJEwCombLB_F_opencl" },
165 { "nbnxn_kernel_ElecEwTwinCut_VdwLJ_F_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJCombGeom_F_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJCombLB_F_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_F_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_F_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_F_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_F_opencl" }
168 /*! \brief Force + energy kernel function pointers. */
169 static const char* nb_kfunc_ener_noprune_ptr[eelOclNR][evdwOclNR] =
171 { "nbnxn_kernel_ElecCut_VdwLJ_VF_opencl", "nbnxn_kernel_ElecCut_VdwLJCombGeom_VF_opencl", "nbnxn_kernel_ElecCut_VdwLJCombLB_VF_opencl", "nbnxn_kernel_ElecCut_VdwLJFsw_VF_opencl", "nbnxn_kernel_ElecCut_VdwLJPsw_VF_opencl", "nbnxn_kernel_ElecCut_VdwLJEwCombGeom_VF_opencl", "nbnxn_kernel_ElecCut_VdwLJEwCombLB_VF_opencl" },
172 { "nbnxn_kernel_ElecRF_VdwLJ_VF_opencl", "nbnxn_kernel_ElecRF_VdwLJCombGeom_VF_opencl", "nbnxn_kernel_ElecRF_VdwLJCombLB_VF_opencl", "nbnxn_kernel_ElecRF_VdwLJFsw_VF_opencl", "nbnxn_kernel_ElecRF_VdwLJPsw_VF_opencl", "nbnxn_kernel_ElecRF_VdwLJEwCombGeom_VF_opencl", "nbnxn_kernel_ElecRF_VdwLJEwCombLB_VF_opencl" },
173 { "nbnxn_kernel_ElecEwQSTab_VdwLJ_VF_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJCombGeom_VF_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJCombLB_VF_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJFsw_VF_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJPsw_VF_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_VF_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_VF_opencl" },
174 { "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_VF_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombGeom_VF_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombLB_VF_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_VF_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_VF_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_VF_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_VF_opencl" },
175 { "nbnxn_kernel_ElecEw_VdwLJ_VF_opencl", "nbnxn_kernel_ElecEw_VdwLJCombGeom_VF_opencl", "nbnxn_kernel_ElecEw_VdwLJCombLB_VF_opencl", "nbnxn_kernel_ElecEw_VdwLJFsw_VF_opencl", "nbnxn_kernel_ElecEw_VdwLJPsw_VF_opencl", "nbnxn_kernel_ElecEw_VdwLJEwCombGeom_VF_opencl", "nbnxn_kernel_ElecEw_VdwLJEwCombLB_VF_opencl" },
176 { "nbnxn_kernel_ElecEwTwinCut_VdwLJ_VF_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJCombGeom_VF_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJCombLB_VF_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_VF_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_VF_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_VF_opencl" }
179 /*! \brief Force + pruning kernel function pointers. */
180 static const char* nb_kfunc_noener_prune_ptr[eelOclNR][evdwOclNR] =
182 { "nbnxn_kernel_ElecCut_VdwLJ_F_prune_opencl", "nbnxn_kernel_ElecCut_VdwLJCombGeom_F_prune_opencl", "nbnxn_kernel_ElecCut_VdwLJCombLB_F_prune_opencl", "nbnxn_kernel_ElecCut_VdwLJFsw_F_prune_opencl", "nbnxn_kernel_ElecCut_VdwLJPsw_F_prune_opencl", "nbnxn_kernel_ElecCut_VdwLJEwCombGeom_F_prune_opencl", "nbnxn_kernel_ElecCut_VdwLJEwCombLB_F_prune_opencl" },
183 { "nbnxn_kernel_ElecRF_VdwLJ_F_prune_opencl", "nbnxn_kernel_ElecRF_VdwLJCombGeom_F_prune_opencl", "nbnxn_kernel_ElecRF_VdwLJCombLB_F_prune_opencl", "nbnxn_kernel_ElecRF_VdwLJFsw_F_prune_opencl", "nbnxn_kernel_ElecRF_VdwLJPsw_F_prune_opencl", "nbnxn_kernel_ElecRF_VdwLJEwCombGeom_F_prune_opencl", "nbnxn_kernel_ElecRF_VdwLJEwCombLB_F_prune_opencl" },
184 { "nbnxn_kernel_ElecEwQSTab_VdwLJ_F_prune_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJCombGeom_F_prune_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJCombLB_F_prune_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJFsw_F_prune_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJPsw_F_prune_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_F_prune_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_F_prune_opencl" },
185 { "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_F_prune_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombGeom_F_prune_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombLB_F_prune_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_F_prune_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_F_prune_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_F_prune_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_F_prune_opencl" },
186 { "nbnxn_kernel_ElecEw_VdwLJ_F_prune_opencl", "nbnxn_kernel_ElecEw_VdwLJCombGeom_F_prune_opencl", "nbnxn_kernel_ElecEw_VdwLJCombLB_F_prune_opencl", "nbnxn_kernel_ElecEw_VdwLJFsw_F_prune_opencl", "nbnxn_kernel_ElecEw_VdwLJPsw_F_prune_opencl", "nbnxn_kernel_ElecEw_VdwLJEwCombGeom_F_prune_opencl", "nbnxn_kernel_ElecEw_VdwLJEwCombLB_F_prune_opencl" },
187 { "nbnxn_kernel_ElecEwTwinCut_VdwLJ_F_prune_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJCombGeom_F_prune_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJCombLB_F_prune_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_F_prune_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_F_prune_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_F_prune_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_F_prune_opencl" }
190 /*! \brief Force + energy + pruning kernel function pointers. */
191 static const char* nb_kfunc_ener_prune_ptr[eelOclNR][evdwOclNR] =
193 { "nbnxn_kernel_ElecCut_VdwLJ_VF_prune_opencl", "nbnxn_kernel_ElecCut_VdwLJCombGeom_VF_prune_opencl", "nbnxn_kernel_ElecCut_VdwLJCombLB_VF_prune_opencl", "nbnxn_kernel_ElecCut_VdwLJFsw_VF_prune_opencl", "nbnxn_kernel_ElecCut_VdwLJPsw_VF_prune_opencl", "nbnxn_kernel_ElecCut_VdwLJEwCombGeom_VF_prune_opencl", "nbnxn_kernel_ElecCut_VdwLJEwCombLB_VF_prune_opencl" },
194 { "nbnxn_kernel_ElecRF_VdwLJ_VF_prune_opencl", "nbnxn_kernel_ElecRF_VdwLJCombGeom_VF_prune_opencl", "nbnxn_kernel_ElecRF_VdwLJCombLB_VF_prune_opencl", "nbnxn_kernel_ElecRF_VdwLJFsw_VF_prune_opencl", "nbnxn_kernel_ElecRF_VdwLJPsw_VF_prune_opencl", "nbnxn_kernel_ElecRF_VdwLJEwCombGeom_VF_prune_opencl", "nbnxn_kernel_ElecRF_VdwLJEwCombLB_VF_prune_opencl" },
195 { "nbnxn_kernel_ElecEwQSTab_VdwLJ_VF_prune_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJCombGeom_VF_prune_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJCombLB_VF_prune_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJFsw_VF_prune_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJPsw_VF_prune_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_VF_prune_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_VF_prune_opencl" },
196 { "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_VF_prune_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombGeom_VF_prune_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombLB_VF_prune_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_VF_prune_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_VF_prune_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_VF_prune_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_VF_prune_opencl" },
197 { "nbnxn_kernel_ElecEw_VdwLJ_VF_prune_opencl", "nbnxn_kernel_ElecEw_VdwLJCombGeom_VF_prune_opencl", "nbnxn_kernel_ElecEw_VdwLJCombLB_VF_prune_opencl", "nbnxn_kernel_ElecEw_VdwLJFsw_VF_prune_opencl", "nbnxn_kernel_ElecEw_VdwLJPsw_VF_prune_opencl", "nbnxn_kernel_ElecEw_VdwLJEwCombGeom_VF_prune_opencl", "nbnxn_kernel_ElecEw_VdwLJEwCombLB_VF_prune_opencl" },
198 { "nbnxn_kernel_ElecEwTwinCut_VdwLJ_VF_prune_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJCombGeom_VF_prune_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJCombLB_VF_prune_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_VF_prune_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_VF_prune_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF_prune_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_VF_prune_opencl" }
201 /*! \brief Return a pointer to the prune kernel version to be executed at the current invocation.
203 * \param[in] kernel_pruneonly array of prune kernel objects
204 * \param[in] firstPrunePass true if the first pruning pass is being executed
206 static inline cl_kernel selectPruneKernel(cl_kernel kernel_pruneonly[],
209 cl_kernel *kernelPtr;
213 kernelPtr = &(kernel_pruneonly[epruneFirst]);
217 kernelPtr = &(kernel_pruneonly[epruneRolling]);
219 // TODO: consider creating the prune kernel object here to avoid a
220 // clCreateKernel for the rolling prune kernel if this is not needed.
224 /*! \brief Return a pointer to the kernel version to be executed at the current step.
225 * OpenCL kernel objects are cached in nb. If the requested kernel is not
226 * found in the cache, it will be created and the cache will be updated.
228 static inline cl_kernel select_nbnxn_kernel(gmx_nbnxn_ocl_t *nb,
234 const char* kernel_name_to_run;
235 cl_kernel *kernel_ptr;
238 assert(eeltype < eelOclNR);
239 assert(evdwtype < evdwOclNR);
245 kernel_name_to_run = nb_kfunc_ener_prune_ptr[eeltype][evdwtype];
246 kernel_ptr = &(nb->kernel_ener_prune_ptr[eeltype][evdwtype]);
250 kernel_name_to_run = nb_kfunc_ener_noprune_ptr[eeltype][evdwtype];
251 kernel_ptr = &(nb->kernel_ener_noprune_ptr[eeltype][evdwtype]);
258 kernel_name_to_run = nb_kfunc_noener_prune_ptr[eeltype][evdwtype];
259 kernel_ptr = &(nb->kernel_noener_prune_ptr[eeltype][evdwtype]);
263 kernel_name_to_run = nb_kfunc_noener_noprune_ptr[eeltype][evdwtype];
264 kernel_ptr = &(nb->kernel_noener_noprune_ptr[eeltype][evdwtype]);
268 if (nullptr == kernel_ptr[0])
270 *kernel_ptr = clCreateKernel(nb->dev_rundata->program, kernel_name_to_run, &cl_error);
271 assert(cl_error == CL_SUCCESS);
273 // TODO: handle errors
278 /*! \brief Calculates the amount of shared memory required by the nonbonded kernel in use.
280 static inline int calc_shmem_required_nonbonded(int vdwType,
281 bool bPrefetchLjParam)
285 /* size of shmem (force-buffers/xq/atom type preloading) */
286 /* NOTE: with the default kernel on sm3.0 we need shmem only for pre-loading */
287 /* i-atom x+q in shared memory */
288 shmem = c_numClPerSupercl * c_clSize * sizeof(float) * 4; /* xqib */
289 /* cj in shared memory, for both warps separately
290 * TODO: in the "nowarp kernels we load cj only once so the factor 2 is not needed.
292 shmem += 2 * c_nbnxnGpuJgroupSize * sizeof(int); /* cjs */
293 if (bPrefetchLjParam)
295 if (useLjCombRule(vdwType))
297 /* i-atom LJ combination parameters in shared memory */
298 shmem += c_numClPerSupercl * c_clSize * 2*sizeof(float); /* atib abused for ljcp, float2 */
302 /* i-atom types in shared memory */
303 shmem += c_numClPerSupercl * c_clSize * sizeof(int); /* atib */
306 /* force reduction buffers in shared memory */
307 shmem += c_clSize * c_clSize * 3 * sizeof(float); /* f_buf */
308 /* Warp vote. In fact it must be * number of warps in block.. */
309 shmem += sizeof(cl_uint) * 2; /* warp_any */
313 /*! \brief Initializes data structures that are going to be sent to the OpenCL device.
315 * The device can't use the same data structures as the host for two main reasons:
316 * - OpenCL restrictions (pointers are not accepted inside data structures)
317 * - some host side fields are not needed for the OpenCL kernels.
319 * This function is called before the launch of both nbnxn and prune kernels.
321 static void fillin_ocl_structures(cl_nbparam_t *nbp,
322 cl_nbparam_params_t *nbparams_params)
324 nbparams_params->coulomb_tab_scale = nbp->coulomb_tab_scale;
325 nbparams_params->c_rf = nbp->c_rf;
326 nbparams_params->dispersion_shift = nbp->dispersion_shift;
327 nbparams_params->eeltype = nbp->eeltype;
328 nbparams_params->epsfac = nbp->epsfac;
329 nbparams_params->ewaldcoeff_lj = nbp->ewaldcoeff_lj;
330 nbparams_params->ewald_beta = nbp->ewald_beta;
331 nbparams_params->rcoulomb_sq = nbp->rcoulomb_sq;
332 nbparams_params->repulsion_shift = nbp->repulsion_shift;
333 nbparams_params->rlistOuter_sq = nbp->rlistOuter_sq;
334 nbparams_params->rvdw_sq = nbp->rvdw_sq;
335 nbparams_params->rlistInner_sq = nbp->rlistInner_sq;
336 nbparams_params->rvdw_switch = nbp->rvdw_switch;
337 nbparams_params->sh_ewald = nbp->sh_ewald;
338 nbparams_params->sh_lj_ewald = nbp->sh_lj_ewald;
339 nbparams_params->two_k_rf = nbp->two_k_rf;
340 nbparams_params->vdwtype = nbp->vdwtype;
341 nbparams_params->vdw_switch = nbp->vdw_switch;
344 /*! \brief Enqueues a wait for event completion.
346 * Then it releases the event and sets it to 0.
347 * Don't use this function when more than one wait will be issued for the event.
348 * Equivalent to Cuda Stream Sync. */
349 static void sync_ocl_event(cl_command_queue stream, cl_event *ocl_event)
351 cl_int gmx_unused cl_error;
354 cl_error = clEnqueueBarrierWithWaitList(stream, 1, ocl_event, nullptr);
355 GMX_RELEASE_ASSERT(CL_SUCCESS == cl_error, ocl_get_error_string(cl_error).c_str());
357 /* Release event and reset it to 0. It is ok to release it as enqueuewaitforevents performs implicit retain for events. */
358 cl_error = clReleaseEvent(*ocl_event);
359 assert(CL_SUCCESS == cl_error);
360 *ocl_event = nullptr;
363 /*! \brief Launch asynchronously the xq buffer host to device copy. */
364 void nbnxn_gpu_copy_xq_to_gpu(gmx_nbnxn_ocl_t *nb,
365 const nbnxn_atomdata_t *nbatom,
369 int adat_begin, adat_len; /* local/nonlocal offset and length used for xq and f */
371 cl_atomdata_t *adat = nb->atdat;
372 cl_plist_t *plist = nb->plist[iloc];
373 cl_timers_t *t = nb->timers;
374 cl_command_queue stream = nb->stream[iloc];
376 bool bDoTime = (nb->bDoTime) != 0;
378 /* Don't launch the non-local H2D copy if there is no dependent
379 work to do: neither non-local nor other (e.g. bonded) work
380 to do that has as input the nbnxn coordinates.
381 Doing the same for the local kernel is more complicated, since the
382 local part of the force array also depends on the non-local kernel.
383 So to avoid complicating the code and to reduce the risk of bugs,
384 we always call the local local x+q copy (and the rest of the local
385 work in nbnxn_gpu_launch_kernel().
387 if (!haveOtherWork && canSkipWork(nb, iloc))
389 plist->haveFreshList = false;
394 /* calculate the atom data index range based on locality */
398 adat_len = adat->natoms_local;
402 adat_begin = adat->natoms_local;
403 adat_len = adat->natoms - adat->natoms_local;
406 /* beginning of timed HtoD section */
409 t->nb_h2d[iloc].openTimingRegion(stream);
413 ocl_copy_H2D_async(adat->xq, nbatom->x().data() + adat_begin * 4, adat_begin*sizeof(float)*4,
414 adat_len * sizeof(float) * 4, stream, bDoTime ? t->nb_h2d[iloc].fetchNextEvent() : nullptr);
418 t->nb_h2d[iloc].closeTimingRegion(stream);
421 /* When we get here all misc operations issues in the local stream as well as
422 the local xq H2D are done,
423 so we record that in the local stream and wait for it in the nonlocal one. */
424 if (nb->bUseTwoStreams)
426 if (iloc == eintLocal)
428 cl_int gmx_used_in_debug cl_error = clEnqueueMarkerWithWaitList(stream, 0, nullptr, &(nb->misc_ops_and_local_H2D_done));
429 assert(CL_SUCCESS == cl_error);
431 /* Based on the v1.2 section 5.13 of the OpenCL spec, a flush is needed
432 * in the local stream in order to be able to sync with the above event
433 * from the non-local stream.
435 cl_error = clFlush(stream);
436 assert(CL_SUCCESS == cl_error);
440 sync_ocl_event(stream, &(nb->misc_ops_and_local_H2D_done));
446 /*! \brief Launch GPU kernel
448 As we execute nonbonded workload in separate queues, before launching
449 the kernel we need to make sure that he following operations have completed:
450 - atomdata allocation and related H2D transfers (every nstlist step);
451 - pair list H2D transfer (every nstlist step);
452 - shift vector H2D transfer (every nstlist step);
453 - force (+shift force and energy) output clearing (every step).
455 These operations are issued in the local queue at the beginning of the step
456 and therefore always complete before the local kernel launch. The non-local
457 kernel is launched after the local on the same device/context, so this is
458 inherently scheduled after the operations in the local stream (including the
460 However, for the sake of having a future-proof implementation, we use the
461 misc_ops_done event to record the point in time when the above operations
462 are finished and synchronize with this event in the non-local stream.
464 void nbnxn_gpu_launch_kernel(gmx_nbnxn_ocl_t *nb,
468 /* OpenCL kernel launch-related stuff */
469 cl_kernel nb_kernel = nullptr; /* fn pointer to the nonbonded kernel */
471 cl_atomdata_t *adat = nb->atdat;
472 cl_nbparam_t *nbp = nb->nbparam;
473 cl_plist_t *plist = nb->plist[iloc];
474 cl_timers_t *t = nb->timers;
475 cl_command_queue stream = nb->stream[iloc];
477 bool bCalcEner = (flags & GMX_FORCE_ENERGY) != 0;
478 int bCalcFshift = flags & GMX_FORCE_VIRIAL;
479 bool bDoTime = (nb->bDoTime) != 0;
481 cl_nbparam_params_t nbparams_params;
483 /* Don't launch the non-local kernel if there is no work to do.
484 Doing the same for the local kernel is more complicated, since the
485 local part of the force array also depends on the non-local kernel.
486 So to avoid complicating the code and to reduce the risk of bugs,
487 we always call the local kernel and later (not in
488 this function) the stream wait, local f copyback and the f buffer
489 clearing. All these operations, except for the local interaction kernel,
490 are needed for the non-local interactions. The skip of the local kernel
491 call is taken care of later in this function. */
492 if (canSkipWork(nb, iloc))
494 plist->haveFreshList = false;
499 if (nbp->useDynamicPruning && plist->haveFreshList)
501 /* Prunes for rlistOuter and rlistInner, sets plist->haveFreshList=false
502 (that's the way the timing accounting can distinguish between
503 separate prune kernel and combined force+prune).
505 nbnxn_gpu_launch_kernel_pruneonly(nb, iloc, 1);
508 if (plist->nsci == 0)
510 /* Don't launch an empty local kernel (is not allowed with OpenCL).
515 /* beginning of timed nonbonded calculation section */
518 t->nb_k[iloc].openTimingRegion(stream);
521 /* get the pointer to the kernel flavor we need to use */
522 nb_kernel = select_nbnxn_kernel(nb,
526 (plist->haveFreshList && !nb->timers->didPrune[iloc]));
528 /* kernel launch config */
530 KernelLaunchConfig config;
531 config.sharedMemorySize = calc_shmem_required_nonbonded(nbp->vdwtype, nb->bPrefetchLjParam);
532 config.stream = stream;
533 config.blockSize[0] = c_clSize;
534 config.blockSize[1] = c_clSize;
535 config.gridSize[0] = plist->nsci;
537 validate_global_work_size(config, 3, nb->dev_info);
541 fprintf(debug, "Non-bonded GPU launch configuration:\n\tLocal work size: %zux%zux%zu\n\t"
542 "Global work size : %zux%zu\n\t#Super-clusters/clusters: %d/%d (%d)\n",
543 config.blockSize[0], config.blockSize[1], config.blockSize[2],
544 config.blockSize[0] * config.gridSize[0], config.blockSize[1] * config.gridSize[1], plist->nsci*c_numClPerSupercl,
545 c_numClPerSupercl, plist->na_c);
548 fillin_ocl_structures(nbp, &nbparams_params);
550 auto *timingEvent = bDoTime ? t->nb_k[iloc].fetchNextEvent() : nullptr;
551 constexpr char kernelName[] = "k_calc_nb";
552 if (useLjCombRule(nb->nbparam->vdwtype))
554 const auto kernelArgs = prepareGpuKernelArguments(nb_kernel, config,
555 &nbparams_params, &adat->xq, &adat->f, &adat->e_lj, &adat->e_el, &adat->fshift,
557 &adat->shift_vec, &nbp->nbfp_climg2d, &nbp->nbfp_comb_climg2d, &nbp->coulomb_tab_climg2d,
558 &plist->sci, &plist->cj4, &plist->excl, &bCalcFshift);
560 launchGpuKernel(nb_kernel, config, timingEvent, kernelName, kernelArgs);
564 const auto kernelArgs = prepareGpuKernelArguments(nb_kernel, config,
566 &nbparams_params, &adat->xq, &adat->f, &adat->e_lj, &adat->e_el, &adat->fshift,
568 &adat->shift_vec, &nbp->nbfp_climg2d, &nbp->nbfp_comb_climg2d, &nbp->coulomb_tab_climg2d,
569 &plist->sci, &plist->cj4, &plist->excl, &bCalcFshift);
570 launchGpuKernel(nb_kernel, config, timingEvent, kernelName, kernelArgs);
575 t->nb_k[iloc].closeTimingRegion(stream);
580 /*! \brief Calculates the amount of shared memory required by the prune kernel.
582 * Note that for the sake of simplicity we use the CUDA terminology "shared memory"
583 * for OpenCL local memory.
585 * \param[in] num_threads_z cj4 concurrency equal to the number of threads/work items in the 3-rd dimension.
586 * \returns the amount of local memory in bytes required by the pruning kernel
588 static inline int calc_shmem_required_prune(const int num_threads_z)
592 /* i-atom x in shared memory (for convenience we load all 4 components including q) */
593 shmem = c_numClPerSupercl * c_clSize * sizeof(float)*4;
594 /* cj in shared memory, for each warp separately
595 * Note: only need to load once per wavefront, but to keep the code simple,
596 * for now we load twice on AMD.
598 shmem += num_threads_z * c_nbnxnGpuClusterpairSplit * c_nbnxnGpuJgroupSize * sizeof(int);
599 /* Warp vote, requires one uint per warp/32 threads per block. */
600 shmem += sizeof(cl_uint) * 2*num_threads_z;
605 void nbnxn_gpu_launch_kernel_pruneonly(gmx_nbnxn_gpu_t *nb,
609 cl_atomdata_t *adat = nb->atdat;
610 cl_nbparam_t *nbp = nb->nbparam;
611 cl_plist_t *plist = nb->plist[iloc];
612 cl_timers_t *t = nb->timers;
613 cl_command_queue stream = nb->stream[iloc];
614 bool bDoTime = nb->bDoTime == CL_TRUE;
616 if (plist->haveFreshList)
618 GMX_ASSERT(numParts == 1, "With first pruning we expect 1 part");
620 /* Set rollingPruningNumParts to signal that it is not set */
621 plist->rollingPruningNumParts = 0;
622 plist->rollingPruningPart = 0;
626 if (plist->rollingPruningNumParts == 0)
628 plist->rollingPruningNumParts = numParts;
632 GMX_ASSERT(numParts == plist->rollingPruningNumParts, "It is not allowed to change numParts in between list generation steps");
636 /* Use a local variable for part and update in plist, so we can return here
637 * without duplicating the part increment code.
639 int part = plist->rollingPruningPart;
641 plist->rollingPruningPart++;
642 if (plist->rollingPruningPart >= plist->rollingPruningNumParts)
644 plist->rollingPruningPart = 0;
647 /* Compute the number of list entries to prune in this pass */
648 int numSciInPart = (plist->nsci - part)/numParts;
650 /* Don't launch the kernel if there is no work to do. */
651 if (numSciInPart <= 0)
653 plist->haveFreshList = false;
658 GpuRegionTimer *timer = nullptr;
661 timer = &(plist->haveFreshList ? t->prune_k[iloc] : t->rollingPrune_k[iloc]);
664 /* beginning of timed prune calculation section */
667 timer->openTimingRegion(stream);
670 /* Kernel launch config:
671 * - The thread block dimensions match the size of i-clusters, j-clusters,
672 * and j-cluster concurrency, in x, y, and z, respectively.
673 * - The 1D block-grid contains as many blocks as super-clusters.
675 int num_threads_z = getOclPruneKernelJ4Concurrency(nb->dev_info->vendor_e);
676 cl_kernel pruneKernel = selectPruneKernel(nb->kernel_pruneonly, plist->haveFreshList);
678 /* kernel launch config */
679 KernelLaunchConfig config;
680 config.sharedMemorySize = calc_shmem_required_prune(num_threads_z);
681 config.stream = stream;
682 config.blockSize[0] = c_clSize;
683 config.blockSize[1] = c_clSize;
684 config.blockSize[2] = num_threads_z;
685 config.gridSize[0] = numSciInPart;
687 validate_global_work_size(config, 3, nb->dev_info);
691 fprintf(debug, "Pruning GPU kernel launch configuration:\n\tLocal work size: %zux%zux%zu\n\t"
692 "\tGlobal work size: %zux%zu\n\t#Super-clusters/clusters: %d/%d (%d)\n"
694 config.blockSize[0], config.blockSize[1], config.blockSize[2],
695 config.blockSize[0] * config.gridSize[0], config.blockSize[1] * config.gridSize[1], plist->nsci*c_numClPerSupercl,
696 c_numClPerSupercl, plist->na_c, config.sharedMemorySize);
699 cl_nbparam_params_t nbparams_params;
700 fillin_ocl_structures(nbp, &nbparams_params);
702 auto *timingEvent = bDoTime ? timer->fetchNextEvent() : nullptr;
703 constexpr char kernelName[] = "k_pruneonly";
704 const auto kernelArgs = prepareGpuKernelArguments(pruneKernel, config,
705 &nbparams_params, &adat->xq, &adat->shift_vec,
706 &plist->sci, &plist->cj4, &plist->imask, &numParts, &part);
707 launchGpuKernel(pruneKernel, config, timingEvent, kernelName, kernelArgs);
709 if (plist->haveFreshList)
711 plist->haveFreshList = false;
712 /* Mark that pruning has been done */
713 nb->timers->didPrune[iloc] = true;
717 /* Mark that rolling pruning has been done */
718 nb->timers->didRollingPrune[iloc] = true;
723 timer->closeTimingRegion(stream);
728 * Launch asynchronously the download of nonbonded forces from the GPU
729 * (and energies/shift forces if required).
731 void nbnxn_gpu_launch_cpyback(gmx_nbnxn_ocl_t *nb,
732 struct nbnxn_atomdata_t *nbatom,
737 cl_int gmx_unused cl_error;
738 int adat_begin, adat_len; /* local/nonlocal offset and length used for xq and f */
740 /* determine interaction locality from atom locality */
741 int iloc = gpuAtomToInteractionLocality(aloc);
743 cl_atomdata_t *adat = nb->atdat;
744 cl_timers_t *t = nb->timers;
745 bool bDoTime = nb->bDoTime == CL_TRUE;
746 cl_command_queue stream = nb->stream[iloc];
748 bool bCalcEner = (flags & GMX_FORCE_ENERGY) != 0;
749 int bCalcFshift = flags & GMX_FORCE_VIRIAL;
752 /* don't launch non-local copy-back if there was no non-local work to do */
753 if (!haveOtherWork && canSkipWork(nb, iloc))
755 /* TODO An alternative way to signal that non-local work is
756 complete is to use a clEnqueueMarker+clEnqueueBarrier
757 pair. However, the use of bNonLocalStreamActive has the
758 advantage of being local to the host, so probably minimizes
759 overhead. Curiously, for NVIDIA OpenCL with an empty-domain
760 test case, overall simulation performance was higher with
761 the API calls, but this has not been tested on AMD OpenCL,
762 so could be worth considering in future. */
763 nb->bNonLocalStreamActive = CL_FALSE;
767 getGpuAtomRange(adat, aloc, &adat_begin, &adat_len);
769 /* beginning of timed D2H section */
772 t->nb_d2h[iloc].openTimingRegion(stream);
775 /* With DD the local D2H transfer can only start after the non-local
776 has been launched. */
777 if (iloc == eintLocal && nb->bNonLocalStreamActive)
779 sync_ocl_event(stream, &(nb->nonlocal_done));
783 ocl_copy_D2H_async(nbatom->out[0].f.data() + adat_begin * 3, adat->f, adat_begin*3*sizeof(float),
784 (adat_len)* adat->f_elem_size, stream, bDoTime ? t->nb_d2h[iloc].fetchNextEvent() : nullptr);
787 cl_error = clFlush(stream);
788 assert(CL_SUCCESS == cl_error);
790 /* After the non-local D2H is launched the nonlocal_done event can be
791 recorded which signals that the local D2H can proceed. This event is not
792 placed after the non-local kernel because we first need the non-local
794 if (iloc == eintNonlocal)
796 cl_error = clEnqueueMarkerWithWaitList(stream, 0, nullptr, &(nb->nonlocal_done));
797 assert(CL_SUCCESS == cl_error);
798 nb->bNonLocalStreamActive = CL_TRUE;
801 /* only transfer energies in the local stream */
807 ocl_copy_D2H_async(nb->nbst.fshift, adat->fshift, 0,
808 SHIFTS * adat->fshift_elem_size, stream, bDoTime ? t->nb_d2h[iloc].fetchNextEvent() : nullptr);
814 ocl_copy_D2H_async(nb->nbst.e_lj, adat->e_lj, 0,
815 sizeof(float), stream, bDoTime ? t->nb_d2h[iloc].fetchNextEvent() : nullptr);
817 ocl_copy_D2H_async(nb->nbst.e_el, adat->e_el, 0,
818 sizeof(float), stream, bDoTime ? t->nb_d2h[iloc].fetchNextEvent() : nullptr);
824 t->nb_d2h[iloc].closeTimingRegion(stream);
829 /*! \brief Selects the Ewald kernel type, analytical or tabulated, single or twin cut-off. */
830 int nbnxn_gpu_pick_ewald_kernel_type(bool bTwinCut)
832 bool bUseAnalyticalEwald, bForceAnalyticalEwald, bForceTabulatedEwald;
835 /* Benchmarking/development environment variables to force the use of
836 analytical or tabulated Ewald kernel. */
837 bForceAnalyticalEwald = (getenv("GMX_OCL_NB_ANA_EWALD") != nullptr);
838 bForceTabulatedEwald = (getenv("GMX_OCL_NB_TAB_EWALD") != nullptr);
840 if (bForceAnalyticalEwald && bForceTabulatedEwald)
842 gmx_incons("Both analytical and tabulated Ewald OpenCL non-bonded kernels "
843 "requested through environment variables.");
846 /* OpenCL: By default, use analytical Ewald
847 * TODO: tabulated does not work, it needs fixing, see init_nbparam() in nbnxn_ocl_data_mgmt.cpp
849 * TODO: decide if dev_info parameter should be added to recognize NVIDIA CC>=3.0 devices.
852 /* By default use analytical Ewald. */
853 bUseAnalyticalEwald = true;
854 if (bForceAnalyticalEwald)
858 fprintf(debug, "Using analytical Ewald OpenCL kernels\n");
861 else if (bForceTabulatedEwald)
863 bUseAnalyticalEwald = false;
867 fprintf(debug, "Using tabulated Ewald OpenCL kernels\n");
871 /* Use twin cut-off kernels if requested by bTwinCut or the env. var.
872 forces it (use it for debugging/benchmarking only). */
873 if (!bTwinCut && (getenv("GMX_OCL_NB_EWALD_TWINCUT") == nullptr))
875 kernel_type = bUseAnalyticalEwald ? eelOclEWALD_ANA : eelOclEWALD_TAB;
879 kernel_type = bUseAnalyticalEwald ? eelOclEWALD_ANA_TWIN : eelOclEWALD_TAB_TWIN;