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 CUDA implementation of nbnxn_gpu.h
38 * \author Szilard Pall <pall.szilard@gmail.com>
47 #include "gromacs/nbnxm/nbnxm_gpu.h"
54 #include "nbnxm_cuda.h"
56 #include "gromacs/gpu_utils/cudautils.cuh"
57 #include "gromacs/gpu_utils/gpueventsynchronizer.cuh"
58 #include "gromacs/gpu_utils/vectype_ops.cuh"
59 #include "gromacs/mdtypes/simulation_workload.h"
60 #include "gromacs/nbnxm/atomdata.h"
61 #include "gromacs/nbnxm/gpu_common.h"
62 #include "gromacs/nbnxm/gpu_common_utils.h"
63 #include "gromacs/nbnxm/gpu_data_mgmt.h"
64 #include "gromacs/nbnxm/grid.h"
65 #include "gromacs/nbnxm/nbnxm.h"
66 #include "gromacs/nbnxm/pairlist.h"
67 #include "gromacs/timing/gpu_timing.h"
68 #include "gromacs/utility/cstringutil.h"
69 #include "gromacs/utility/gmxassert.h"
71 #include "nbnxm_buffer_ops_kernels.cuh"
72 #include "nbnxm_cuda_types.h"
74 /***** The kernel declarations/definitions come here *****/
76 /* Top-level kernel declaration generation: will generate through multiple
77 * inclusion the following flavors for all kernel declarations:
78 * - force-only output;
79 * - force and energy output;
80 * - force-only with pair list pruning;
81 * - force and energy output with pair list pruning.
83 #define FUNCTION_DECLARATION_ONLY
85 #include "nbnxm_cuda_kernels.cuh"
86 /** Force & energy **/
88 #include "nbnxm_cuda_kernels.cuh"
91 /*** Pair-list pruning kernels ***/
94 #include "nbnxm_cuda_kernels.cuh"
95 /** Force & energy **/
97 #include "nbnxm_cuda_kernels.cuh"
101 /* Prune-only kernels */
102 #include "nbnxm_cuda_kernel_pruneonly.cuh"
103 #undef FUNCTION_DECLARATION_ONLY
105 /* Now generate the function definitions if we are using a single compilation unit. */
106 #if GMX_CUDA_NB_SINGLE_COMPILATION_UNIT
107 #include "nbnxm_cuda_kernel_F_noprune.cu"
108 #include "nbnxm_cuda_kernel_F_prune.cu"
109 #include "nbnxm_cuda_kernel_VF_noprune.cu"
110 #include "nbnxm_cuda_kernel_VF_prune.cu"
111 #include "nbnxm_cuda_kernel_pruneonly.cu"
112 #endif /* GMX_CUDA_NB_SINGLE_COMPILATION_UNIT */
117 //! Number of CUDA threads in a block
118 //TODO Optimize this through experimentation
119 constexpr static int c_bufOpsThreadsPerBlock = 128;
121 /*! Nonbonded kernel function pointer type */
122 typedef void (*nbnxn_cu_kfunc_ptr_t)(const cu_atomdata_t,
127 /*********************************/
129 /*! Returns the number of blocks to be used for the nonbonded GPU kernel. */
130 static inline int calc_nb_kernel_nblock(int nwork_units, const gmx_device_info_t *dinfo)
135 /* CUDA does not accept grid dimension of 0 (which can happen e.g. with an
136 empty domain) and that case should be handled before this point. */
137 assert(nwork_units > 0);
139 max_grid_x_size = dinfo->prop.maxGridSize[0];
141 /* do we exceed the grid x dimension limit? */
142 if (nwork_units > max_grid_x_size)
144 gmx_fatal(FARGS, "Watch out, the input system is too large to simulate!\n"
145 "The number of nonbonded work units (=number of super-clusters) exceeds the"
146 "maximum grid size in x dimension (%d > %d)!", nwork_units, max_grid_x_size);
153 /* Constant arrays listing all kernel function pointers and enabling selection
154 of a kernel in an elegant manner. */
156 /*! Pointers to the non-bonded kernels organized in 2-dim arrays by:
157 * electrostatics and VDW type.
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 nbnxn_cuda_types.h.
164 /*! Force-only kernel function pointers. */
165 static const nbnxn_cu_kfunc_ptr_t nb_kfunc_noener_noprune_ptr[eelCuNR][evdwCuNR] =
167 { nbnxn_kernel_ElecCut_VdwLJ_F_cuda, nbnxn_kernel_ElecCut_VdwLJCombGeom_F_cuda, nbnxn_kernel_ElecCut_VdwLJCombLB_F_cuda, nbnxn_kernel_ElecCut_VdwLJFsw_F_cuda, nbnxn_kernel_ElecCut_VdwLJPsw_F_cuda, nbnxn_kernel_ElecCut_VdwLJEwCombGeom_F_cuda, nbnxn_kernel_ElecCut_VdwLJEwCombLB_F_cuda },
168 { nbnxn_kernel_ElecRF_VdwLJ_F_cuda, nbnxn_kernel_ElecRF_VdwLJCombGeom_F_cuda, nbnxn_kernel_ElecRF_VdwLJCombLB_F_cuda, nbnxn_kernel_ElecRF_VdwLJFsw_F_cuda, nbnxn_kernel_ElecRF_VdwLJPsw_F_cuda, nbnxn_kernel_ElecRF_VdwLJEwCombGeom_F_cuda, nbnxn_kernel_ElecRF_VdwLJEwCombLB_F_cuda },
169 { nbnxn_kernel_ElecEwQSTab_VdwLJ_F_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJCombGeom_F_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJCombLB_F_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJFsw_F_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJPsw_F_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_F_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_F_cuda },
170 { nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_F_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombGeom_F_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombLB_F_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_F_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_F_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_F_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_F_cuda },
171 { nbnxn_kernel_ElecEw_VdwLJ_F_cuda, nbnxn_kernel_ElecEw_VdwLJCombGeom_F_cuda, nbnxn_kernel_ElecEw_VdwLJCombLB_F_cuda, nbnxn_kernel_ElecEw_VdwLJFsw_F_cuda, nbnxn_kernel_ElecEw_VdwLJPsw_F_cuda, nbnxn_kernel_ElecEw_VdwLJEwCombGeom_F_cuda, nbnxn_kernel_ElecEw_VdwLJEwCombLB_F_cuda },
172 { nbnxn_kernel_ElecEwTwinCut_VdwLJ_F_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJCombGeom_F_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJCombLB_F_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_F_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_F_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_F_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_F_cuda }
175 /*! Force + energy kernel function pointers. */
176 static const nbnxn_cu_kfunc_ptr_t nb_kfunc_ener_noprune_ptr[eelCuNR][evdwCuNR] =
178 { nbnxn_kernel_ElecCut_VdwLJ_VF_cuda, nbnxn_kernel_ElecCut_VdwLJCombGeom_VF_cuda, nbnxn_kernel_ElecCut_VdwLJCombLB_VF_cuda, nbnxn_kernel_ElecCut_VdwLJFsw_VF_cuda, nbnxn_kernel_ElecCut_VdwLJPsw_VF_cuda, nbnxn_kernel_ElecCut_VdwLJEwCombGeom_VF_cuda, nbnxn_kernel_ElecCut_VdwLJEwCombLB_VF_cuda },
179 { nbnxn_kernel_ElecRF_VdwLJ_VF_cuda, nbnxn_kernel_ElecRF_VdwLJCombGeom_VF_cuda, nbnxn_kernel_ElecRF_VdwLJCombLB_VF_cuda, nbnxn_kernel_ElecRF_VdwLJFsw_VF_cuda, nbnxn_kernel_ElecRF_VdwLJPsw_VF_cuda, nbnxn_kernel_ElecRF_VdwLJEwCombGeom_VF_cuda, nbnxn_kernel_ElecRF_VdwLJEwCombLB_VF_cuda },
180 { nbnxn_kernel_ElecEwQSTab_VdwLJ_VF_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJCombGeom_VF_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJCombLB_VF_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJFsw_VF_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJPsw_VF_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_VF_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_VF_cuda },
181 { nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_VF_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombGeom_VF_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombLB_VF_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_VF_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_VF_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_VF_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_VF_cuda },
182 { nbnxn_kernel_ElecEw_VdwLJ_VF_cuda, nbnxn_kernel_ElecEw_VdwLJCombGeom_VF_cuda, nbnxn_kernel_ElecEw_VdwLJCombLB_VF_cuda, nbnxn_kernel_ElecEw_VdwLJFsw_VF_cuda, nbnxn_kernel_ElecEw_VdwLJPsw_VF_cuda, nbnxn_kernel_ElecEw_VdwLJEwCombGeom_VF_cuda, nbnxn_kernel_ElecEw_VdwLJEwCombLB_VF_cuda },
183 { nbnxn_kernel_ElecEwTwinCut_VdwLJ_VF_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJCombGeom_VF_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJCombLB_VF_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_VF_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_VF_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_VF_cuda }
186 /*! Force + pruning kernel function pointers. */
187 static const nbnxn_cu_kfunc_ptr_t nb_kfunc_noener_prune_ptr[eelCuNR][evdwCuNR] =
189 { nbnxn_kernel_ElecCut_VdwLJ_F_prune_cuda, nbnxn_kernel_ElecCut_VdwLJCombGeom_F_prune_cuda, nbnxn_kernel_ElecCut_VdwLJCombLB_F_prune_cuda, nbnxn_kernel_ElecCut_VdwLJFsw_F_prune_cuda, nbnxn_kernel_ElecCut_VdwLJPsw_F_prune_cuda, nbnxn_kernel_ElecCut_VdwLJEwCombGeom_F_prune_cuda, nbnxn_kernel_ElecCut_VdwLJEwCombLB_F_prune_cuda },
190 { nbnxn_kernel_ElecRF_VdwLJ_F_prune_cuda, nbnxn_kernel_ElecRF_VdwLJCombGeom_F_prune_cuda, nbnxn_kernel_ElecRF_VdwLJCombLB_F_prune_cuda, nbnxn_kernel_ElecRF_VdwLJFsw_F_prune_cuda, nbnxn_kernel_ElecRF_VdwLJPsw_F_prune_cuda, nbnxn_kernel_ElecRF_VdwLJEwCombGeom_F_prune_cuda, nbnxn_kernel_ElecRF_VdwLJEwCombLB_F_prune_cuda },
191 { nbnxn_kernel_ElecEwQSTab_VdwLJ_F_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJCombGeom_F_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJCombLB_F_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJFsw_F_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJPsw_F_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_F_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_F_prune_cuda },
192 { nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_F_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombGeom_F_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombLB_F_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_F_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_F_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_F_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_F_prune_cuda },
193 { nbnxn_kernel_ElecEw_VdwLJ_F_prune_cuda, nbnxn_kernel_ElecEw_VdwLJCombGeom_F_prune_cuda, nbnxn_kernel_ElecEw_VdwLJCombLB_F_prune_cuda, nbnxn_kernel_ElecEw_VdwLJFsw_F_prune_cuda, nbnxn_kernel_ElecEw_VdwLJPsw_F_prune_cuda, nbnxn_kernel_ElecEw_VdwLJEwCombGeom_F_prune_cuda, nbnxn_kernel_ElecEw_VdwLJEwCombLB_F_prune_cuda },
194 { nbnxn_kernel_ElecEwTwinCut_VdwLJ_F_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJCombGeom_F_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJCombLB_F_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_F_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_F_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_F_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_F_prune_cuda }
197 /*! Force + energy + pruning kernel function pointers. */
198 static const nbnxn_cu_kfunc_ptr_t nb_kfunc_ener_prune_ptr[eelCuNR][evdwCuNR] =
200 { nbnxn_kernel_ElecCut_VdwLJ_VF_prune_cuda, nbnxn_kernel_ElecCut_VdwLJCombGeom_VF_prune_cuda, nbnxn_kernel_ElecCut_VdwLJCombLB_VF_prune_cuda, nbnxn_kernel_ElecCut_VdwLJFsw_VF_prune_cuda, nbnxn_kernel_ElecCut_VdwLJPsw_VF_prune_cuda, nbnxn_kernel_ElecCut_VdwLJEwCombGeom_VF_prune_cuda, nbnxn_kernel_ElecCut_VdwLJEwCombLB_VF_prune_cuda },
201 { nbnxn_kernel_ElecRF_VdwLJ_VF_prune_cuda, nbnxn_kernel_ElecRF_VdwLJCombGeom_VF_prune_cuda, nbnxn_kernel_ElecRF_VdwLJCombLB_VF_prune_cuda, nbnxn_kernel_ElecRF_VdwLJFsw_VF_prune_cuda, nbnxn_kernel_ElecRF_VdwLJPsw_VF_prune_cuda, nbnxn_kernel_ElecRF_VdwLJEwCombGeom_VF_prune_cuda, nbnxn_kernel_ElecRF_VdwLJEwCombLB_VF_prune_cuda },
202 { nbnxn_kernel_ElecEwQSTab_VdwLJ_VF_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJCombGeom_VF_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJCombLB_VF_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJFsw_VF_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJPsw_VF_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_VF_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_VF_prune_cuda },
203 { nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_VF_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombGeom_VF_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJCombLB_VF_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_VF_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_VF_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_VF_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_VF_prune_cuda },
204 { nbnxn_kernel_ElecEw_VdwLJ_VF_prune_cuda, nbnxn_kernel_ElecEw_VdwLJCombGeom_VF_prune_cuda, nbnxn_kernel_ElecEw_VdwLJCombLB_VF_prune_cuda, nbnxn_kernel_ElecEw_VdwLJFsw_VF_prune_cuda, nbnxn_kernel_ElecEw_VdwLJPsw_VF_prune_cuda, nbnxn_kernel_ElecEw_VdwLJEwCombGeom_VF_prune_cuda, nbnxn_kernel_ElecEw_VdwLJEwCombLB_VF_prune_cuda },
205 { nbnxn_kernel_ElecEwTwinCut_VdwLJ_VF_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJCombGeom_VF_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJCombLB_VF_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_VF_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_VF_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_VF_prune_cuda }
208 /*! Return a pointer to the kernel version to be executed at the current step. */
209 static inline nbnxn_cu_kfunc_ptr_t select_nbnxn_kernel(int eeltype,
213 const gmx_device_info_t gmx_unused *devInfo)
215 nbnxn_cu_kfunc_ptr_t res;
217 GMX_ASSERT(eeltype < eelCuNR,
218 "The electrostatics type requested is not implemented in the CUDA kernels.");
219 GMX_ASSERT(evdwtype < evdwCuNR,
220 "The VdW type requested is not implemented in the CUDA kernels.");
222 /* assert assumptions made by the kernels */
223 GMX_ASSERT(c_nbnxnGpuClusterSize*c_nbnxnGpuClusterSize/c_nbnxnGpuClusterpairSplit == devInfo->prop.warpSize,
224 "The CUDA kernels require the cluster_size_i*cluster_size_j/nbnxn_gpu_clusterpair_split to match the warp size of the architecture targeted.");
230 res = nb_kfunc_ener_prune_ptr[eeltype][evdwtype];
234 res = nb_kfunc_ener_noprune_ptr[eeltype][evdwtype];
241 res = nb_kfunc_noener_prune_ptr[eeltype][evdwtype];
245 res = nb_kfunc_noener_noprune_ptr[eeltype][evdwtype];
252 /*! \brief Calculates the amount of shared memory required by the nonbonded kernel in use. */
253 static inline int calc_shmem_required_nonbonded(const int num_threads_z, const gmx_device_info_t gmx_unused *dinfo, const cu_nbparam_t *nbp)
259 /* size of shmem (force-buffers/xq/atom type preloading) */
260 /* NOTE: with the default kernel on sm3.0 we need shmem only for pre-loading */
261 /* i-atom x+q in shared memory */
262 shmem = c_numClPerSupercl * c_clSize * sizeof(float4);
263 /* cj in shared memory, for each warp separately */
264 shmem += num_threads_z * c_nbnxnGpuClusterpairSplit * c_nbnxnGpuJgroupSize * sizeof(int);
266 if (nbp->vdwtype == evdwCuCUTCOMBGEOM ||
267 nbp->vdwtype == evdwCuCUTCOMBLB)
269 /* i-atom LJ combination parameters in shared memory */
270 shmem += c_numClPerSupercl * c_clSize * sizeof(float2);
274 /* i-atom types in shared memory */
275 shmem += c_numClPerSupercl * c_clSize * sizeof(int);
281 /*! \brief Sync the nonlocal stream with dependent tasks in the local queue.
283 * As the point where the local stream tasks can be considered complete happens
284 * at the same call point where the nonlocal stream should be synced with the
285 * the local, this function records the event if called with the local stream as
286 * argument and inserts in the GPU stream a wait on the event on the nonlocal.
288 void nbnxnInsertNonlocalGpuDependency(const gmx_nbnxn_cuda_t *nb,
289 const InteractionLocality interactionLocality)
291 cudaStream_t stream = nb->stream[interactionLocality];
293 /* When we get here all misc operations issued in the local stream as well as
294 the local xq H2D are done,
295 so we record that in the local stream and wait for it in the nonlocal one.
296 This wait needs to precede any PP tasks, bonded or nonbonded, that may
297 compute on interactions between local and nonlocal atoms.
299 if (nb->bUseTwoStreams)
301 if (interactionLocality == InteractionLocality::Local)
303 cudaError_t stat = cudaEventRecord(nb->misc_ops_and_local_H2D_done, stream);
304 CU_RET_ERR(stat, "cudaEventRecord on misc_ops_and_local_H2D_done failed");
308 cudaError_t stat = cudaStreamWaitEvent(stream, nb->misc_ops_and_local_H2D_done, 0);
309 CU_RET_ERR(stat, "cudaStreamWaitEvent on misc_ops_and_local_H2D_done failed");
314 /*! \brief Launch asynchronously the xq buffer host to device copy. */
315 void gpu_copy_xq_to_gpu(gmx_nbnxn_cuda_t *nb,
316 const nbnxn_atomdata_t *nbatom,
317 const AtomLocality atomLocality)
319 GMX_ASSERT(nb, "Need a valid nbnxn_gpu object");
321 GMX_ASSERT(atomLocality == AtomLocality::Local || atomLocality == AtomLocality::NonLocal,
322 "Only local and non-local xq transfers are supported");
324 const InteractionLocality iloc = gpuAtomToInteractionLocality(atomLocality);
326 int adat_begin, adat_len; /* local/nonlocal offset and length used for xq and f */
328 cu_atomdata_t *adat = nb->atdat;
329 cu_plist_t *plist = nb->plist[iloc];
330 cu_timers_t *t = nb->timers;
331 cudaStream_t stream = nb->stream[iloc];
333 bool bDoTime = nb->bDoTime;
335 /* Don't launch the non-local H2D copy if there is no dependent
336 work to do: neither non-local nor other (e.g. bonded) work
337 to do that has as input the nbnxn coordaintes.
338 Doing the same for the local kernel is more complicated, since the
339 local part of the force array also depends on the non-local kernel.
340 So to avoid complicating the code and to reduce the risk of bugs,
341 we always call the local local x+q copy (and the rest of the local
342 work in nbnxn_gpu_launch_kernel().
344 if ((iloc == InteractionLocality::NonLocal) && !haveGpuShortRangeWork(*nb, iloc))
346 plist->haveFreshList = false;
351 /* calculate the atom data index range based on locality */
352 if (atomLocality == AtomLocality::Local)
355 adat_len = adat->natoms_local;
359 adat_begin = adat->natoms_local;
360 adat_len = adat->natoms - adat->natoms_local;
364 /* beginning of timed HtoD section */
367 t->xf[atomLocality].nb_h2d.openTimingRegion(stream);
370 cu_copy_H2D_async(adat->xq + adat_begin, static_cast<const void *>(nbatom->x().data() + adat_begin * 4),
371 adat_len * sizeof(*adat->xq), stream);
375 t->xf[atomLocality].nb_h2d.closeTimingRegion(stream);
378 /* When we get here all misc operations issued in the local stream as well as
379 the local xq H2D are done,
380 so we record that in the local stream and wait for it in the nonlocal one.
381 This wait needs to precede any PP tasks, bonded or nonbonded, that may
382 compute on interactions between local and nonlocal atoms.
384 nbnxnInsertNonlocalGpuDependency(nb, iloc);
387 /*! As we execute nonbonded workload in separate streams, before launching
388 the kernel we need to make sure that he following operations have completed:
389 - atomdata allocation and related H2D transfers (every nstlist step);
390 - pair list H2D transfer (every nstlist step);
391 - shift vector H2D transfer (every nstlist step);
392 - force (+shift force and energy) output clearing (every step).
394 These operations are issued in the local stream at the beginning of the step
395 and therefore always complete before the local kernel launch. The non-local
396 kernel is launched after the local on the same device/context hence it is
397 inherently scheduled after the operations in the local stream (including the
398 above "misc_ops") on pre-GK110 devices with single hardware queue, but on later
399 devices with multiple hardware queues the dependency needs to be enforced.
400 We use the misc_ops_and_local_H2D_done event to record the point where
401 the local x+q H2D (and all preceding) tasks are complete and synchronize
402 with this event in the non-local stream before launching the non-bonded kernel.
404 void gpu_launch_kernel(gmx_nbnxn_cuda_t *nb,
405 const gmx::StepWorkload &stepWork,
406 const InteractionLocality iloc)
408 cu_atomdata_t *adat = nb->atdat;
409 cu_nbparam_t *nbp = nb->nbparam;
410 cu_plist_t *plist = nb->plist[iloc];
411 cu_timers_t *t = nb->timers;
412 cudaStream_t stream = nb->stream[iloc];
414 bool bDoTime = nb->bDoTime;
416 /* Don't launch the non-local kernel if there is no work to do.
417 Doing the same for the local kernel is more complicated, since the
418 local part of the force array also depends on the non-local kernel.
419 So to avoid complicating the code and to reduce the risk of bugs,
420 we always call the local kernel, and later (not in
421 this function) the stream wait, local f copyback and the f buffer
422 clearing. All these operations, except for the local interaction kernel,
423 are needed for the non-local interactions. The skip of the local kernel
424 call is taken care of later in this function. */
425 if (canSkipNonbondedWork(*nb, iloc))
427 plist->haveFreshList = false;
432 if (nbp->useDynamicPruning && plist->haveFreshList)
434 /* Prunes for rlistOuter and rlistInner, sets plist->haveFreshList=false
435 (TODO: ATM that's the way the timing accounting can distinguish between
436 separate prune kernel and combined force+prune, maybe we need a better way?).
438 gpu_launch_kernel_pruneonly(nb, iloc, 1);
441 if (plist->nsci == 0)
443 /* Don't launch an empty local kernel (not allowed with CUDA) */
447 /* beginning of timed nonbonded calculation section */
450 t->interaction[iloc].nb_k.openTimingRegion(stream);
453 /* Kernel launch config:
454 * - The thread block dimensions match the size of i-clusters, j-clusters,
455 * and j-cluster concurrency, in x, y, and z, respectively.
456 * - The 1D block-grid contains as many blocks as super-clusters.
458 int num_threads_z = 1;
459 if (nb->dev_info->prop.major == 3 && nb->dev_info->prop.minor == 7)
463 int nblock = calc_nb_kernel_nblock(plist->nsci, nb->dev_info);
466 KernelLaunchConfig config;
467 config.blockSize[0] = c_clSize;
468 config.blockSize[1] = c_clSize;
469 config.blockSize[2] = num_threads_z;
470 config.gridSize[0] = nblock;
471 config.sharedMemorySize = calc_shmem_required_nonbonded(num_threads_z, nb->dev_info, nbp);
472 config.stream = stream;
476 fprintf(debug, "Non-bonded GPU launch configuration:\n\tThread block: %zux%zux%zu\n\t"
477 "\tGrid: %zux%zu\n\t#Super-clusters/clusters: %d/%d (%d)\n"
479 config.blockSize[0], config.blockSize[1], config.blockSize[2],
480 config.gridSize[0], config.gridSize[1], plist->nsci*c_numClPerSupercl,
481 c_numClPerSupercl, plist->na_c,
482 config.sharedMemorySize);
485 auto *timingEvent = bDoTime ? t->interaction[iloc].nb_k.fetchNextEvent() : nullptr;
486 const auto kernel = select_nbnxn_kernel(nbp->eeltype,
488 stepWork.computeEnergy,
489 (plist->haveFreshList && !nb->timers->interaction[iloc].didPrune),
491 const auto kernelArgs = prepareGpuKernelArguments(kernel, config, adat, nbp, plist, &stepWork.computeVirial);
492 launchGpuKernel(kernel, config, timingEvent, "k_calc_nb", kernelArgs);
496 t->interaction[iloc].nb_k.closeTimingRegion(stream);
499 if (GMX_NATIVE_WINDOWS)
501 /* Windows: force flushing WDDM queue */
502 cudaStreamQuery(stream);
506 /*! Calculates the amount of shared memory required by the CUDA kernel in use. */
507 static inline int calc_shmem_required_prune(const int num_threads_z)
511 /* i-atom x in shared memory */
512 shmem = c_numClPerSupercl * c_clSize * sizeof(float4);
513 /* cj in shared memory, for each warp separately */
514 shmem += num_threads_z * c_nbnxnGpuClusterpairSplit * c_nbnxnGpuJgroupSize * sizeof(int);
519 void gpu_launch_kernel_pruneonly(gmx_nbnxn_cuda_t *nb,
520 const InteractionLocality iloc,
523 cu_atomdata_t *adat = nb->atdat;
524 cu_nbparam_t *nbp = nb->nbparam;
525 cu_plist_t *plist = nb->plist[iloc];
526 cu_timers_t *t = nb->timers;
527 cudaStream_t stream = nb->stream[iloc];
529 bool bDoTime = nb->bDoTime;
531 if (plist->haveFreshList)
533 GMX_ASSERT(numParts == 1, "With first pruning we expect 1 part");
535 /* Set rollingPruningNumParts to signal that it is not set */
536 plist->rollingPruningNumParts = 0;
537 plist->rollingPruningPart = 0;
541 if (plist->rollingPruningNumParts == 0)
543 plist->rollingPruningNumParts = numParts;
547 GMX_ASSERT(numParts == plist->rollingPruningNumParts, "It is not allowed to change numParts in between list generation steps");
551 /* Use a local variable for part and update in plist, so we can return here
552 * without duplicating the part increment code.
554 int part = plist->rollingPruningPart;
556 plist->rollingPruningPart++;
557 if (plist->rollingPruningPart >= plist->rollingPruningNumParts)
559 plist->rollingPruningPart = 0;
562 /* Compute the number of list entries to prune in this pass */
563 int numSciInPart = (plist->nsci - part)/numParts;
565 /* Don't launch the kernel if there is no work to do (not allowed with CUDA) */
566 if (numSciInPart <= 0)
568 plist->haveFreshList = false;
573 GpuRegionTimer *timer = nullptr;
576 timer = &(plist->haveFreshList ? t->interaction[iloc].prune_k : t->interaction[iloc].rollingPrune_k);
579 /* beginning of timed prune calculation section */
582 timer->openTimingRegion(stream);
585 /* Kernel launch config:
586 * - The thread block dimensions match the size of i-clusters, j-clusters,
587 * and j-cluster concurrency, in x, y, and z, respectively.
588 * - The 1D block-grid contains as many blocks as super-clusters.
590 int num_threads_z = c_cudaPruneKernelJ4Concurrency;
591 int nblock = calc_nb_kernel_nblock(numSciInPart, nb->dev_info);
592 KernelLaunchConfig config;
593 config.blockSize[0] = c_clSize;
594 config.blockSize[1] = c_clSize;
595 config.blockSize[2] = num_threads_z;
596 config.gridSize[0] = nblock;
597 config.sharedMemorySize = calc_shmem_required_prune(num_threads_z);
598 config.stream = stream;
602 fprintf(debug, "Pruning GPU kernel launch configuration:\n\tThread block: %zux%zux%zu\n\t"
603 "\tGrid: %zux%zu\n\t#Super-clusters/clusters: %d/%d (%d)\n"
605 config.blockSize[0], config.blockSize[1], config.blockSize[2],
606 config.gridSize[0], config.gridSize[1], numSciInPart*c_numClPerSupercl,
607 c_numClPerSupercl, plist->na_c,
608 config.sharedMemorySize);
611 auto *timingEvent = bDoTime ? timer->fetchNextEvent() : nullptr;
612 constexpr char kernelName[] = "k_pruneonly";
613 const auto kernel = plist->haveFreshList ? nbnxn_kernel_prune_cuda<true> : nbnxn_kernel_prune_cuda<false>;
614 const auto kernelArgs = prepareGpuKernelArguments(kernel, config, adat, nbp, plist, &numParts, &part);
615 launchGpuKernel(kernel, config, timingEvent, kernelName, kernelArgs);
617 /* TODO: consider a more elegant way to track which kernel has been called
618 (combined or separate 1st pass prune, rolling prune). */
619 if (plist->haveFreshList)
621 plist->haveFreshList = false;
622 /* Mark that pruning has been done */
623 nb->timers->interaction[iloc].didPrune = true;
627 /* Mark that rolling pruning has been done */
628 nb->timers->interaction[iloc].didRollingPrune = true;
633 timer->closeTimingRegion(stream);
636 if (GMX_NATIVE_WINDOWS)
638 /* Windows: force flushing WDDM queue */
639 cudaStreamQuery(stream);
643 void gpu_launch_cpyback(gmx_nbnxn_cuda_t *nb,
644 nbnxn_atomdata_t *nbatom,
645 const gmx::StepWorkload &stepWork,
646 const AtomLocality atomLocality)
648 GMX_ASSERT(nb, "Need a valid nbnxn_gpu object");
651 int adat_begin, adat_len; /* local/nonlocal offset and length used for xq and f */
653 /* determine interaction locality from atom locality */
654 const InteractionLocality iloc = gpuAtomToInteractionLocality(atomLocality);
656 /* extract the data */
657 cu_atomdata_t *adat = nb->atdat;
658 cu_timers_t *t = nb->timers;
659 bool bDoTime = nb->bDoTime;
660 cudaStream_t stream = nb->stream[iloc];
662 /* don't launch non-local copy-back if there was no non-local work to do */
663 if ((iloc == InteractionLocality::NonLocal) && !haveGpuShortRangeWork(*nb, iloc))
668 getGpuAtomRange(adat, atomLocality, &adat_begin, &adat_len);
670 /* beginning of timed D2H section */
673 t->xf[atomLocality].nb_d2h.openTimingRegion(stream);
676 /* With DD the local D2H transfer can only start after the non-local
677 kernel has finished. */
678 if (iloc == InteractionLocality::Local && nb->bUseTwoStreams)
680 stat = cudaStreamWaitEvent(stream, nb->nonlocal_done, 0);
681 CU_RET_ERR(stat, "cudaStreamWaitEvent on nonlocal_done failed");
685 * Skip if buffer ops / reduction is offloaded to the GPU.
687 if (!stepWork.useGpuFBufferOps)
689 cu_copy_D2H_async(nbatom->out[0].f.data() + adat_begin * 3, adat->f + adat_begin,
690 (adat_len)*sizeof(*adat->f), stream);
693 /* After the non-local D2H is launched the nonlocal_done event can be
694 recorded which signals that the local D2H can proceed. This event is not
695 placed after the non-local kernel because we want the non-local data
697 if (iloc == InteractionLocality::NonLocal)
699 stat = cudaEventRecord(nb->nonlocal_done, stream);
700 CU_RET_ERR(stat, "cudaEventRecord on nonlocal_done failed");
703 /* only transfer energies in the local stream */
704 if (iloc == InteractionLocality::Local)
706 /* DtoH fshift when virial is needed */
707 if (stepWork.computeVirial)
709 cu_copy_D2H_async(nb->nbst.fshift, adat->fshift,
710 SHIFTS * sizeof(*nb->nbst.fshift), stream);
714 if (stepWork.computeEnergy)
716 cu_copy_D2H_async(nb->nbst.e_lj, adat->e_lj,
717 sizeof(*nb->nbst.e_lj), stream);
718 cu_copy_D2H_async(nb->nbst.e_el, adat->e_el,
719 sizeof(*nb->nbst.e_el), stream);
725 t->xf[atomLocality].nb_d2h.closeTimingRegion(stream);
729 void cuda_set_cacheconfig()
733 for (int i = 0; i < eelCuNR; i++)
735 for (int j = 0; j < evdwCuNR; j++)
737 /* Default kernel 32/32 kB Shared/L1 */
738 cudaFuncSetCacheConfig(nb_kfunc_ener_prune_ptr[i][j], cudaFuncCachePreferEqual);
739 cudaFuncSetCacheConfig(nb_kfunc_ener_noprune_ptr[i][j], cudaFuncCachePreferEqual);
740 cudaFuncSetCacheConfig(nb_kfunc_noener_prune_ptr[i][j], cudaFuncCachePreferEqual);
741 stat = cudaFuncSetCacheConfig(nb_kfunc_noener_noprune_ptr[i][j], cudaFuncCachePreferEqual);
742 CU_RET_ERR(stat, "cudaFuncSetCacheConfig failed");
747 /* X buffer operations on GPU: performs conversion from rvec to nb format. */
748 void nbnxn_gpu_x_to_nbat_x(const Nbnxm::Grid &grid,
749 bool setFillerCoords,
751 DeviceBuffer<float> d_x,
752 GpuEventSynchronizer *xReadyOnDevice,
753 const Nbnxm::AtomLocality locality,
757 GMX_ASSERT(nb, "Need a valid nbnxn_gpu object");
759 cu_atomdata_t *adat = nb->atdat;
761 const int numColumns = grid.numColumns();
762 const int cellOffset = grid.cellOffset();
763 const int numAtomsPerCell = grid.numAtomsPerCell();
764 Nbnxm::InteractionLocality interactionLoc = gpuAtomToInteractionLocality(locality);
766 cudaStream_t stream = nb->stream[interactionLoc];
768 int numAtoms = grid.srcAtomEnd() - grid.srcAtomBegin();
769 // avoid empty kernel launch, skip to inserting stream dependency
772 // TODO: This will only work with CUDA
773 GMX_ASSERT(d_x, "Need a valid device pointer");
775 // ensure that coordinates are ready on the device before launching the kernel
776 GMX_ASSERT(xReadyOnDevice, "Need a valid GpuEventSynchronizer object");
777 xReadyOnDevice->enqueueWaitEvent(stream);
779 KernelLaunchConfig config;
780 config.blockSize[0] = c_bufOpsThreadsPerBlock;
781 config.blockSize[1] = 1;
782 config.blockSize[2] = 1;
783 config.gridSize[0] = (grid.numCellsColumnMax()*numAtomsPerCell + c_bufOpsThreadsPerBlock - 1)/c_bufOpsThreadsPerBlock;
784 config.gridSize[1] = numColumns;
785 config.gridSize[2] = 1;
786 GMX_ASSERT(config.gridSize[0] > 0, "Can not have empty grid, early return above avoids this");
787 config.sharedMemorySize = 0;
788 config.stream = stream;
790 auto kernelFn = nbnxn_gpu_x_to_nbat_x_kernel;
791 float *xqPtr = &(adat->xq->x);
792 const int *d_atomIndices = nb->atomIndices;
793 const int *d_cxy_na = &nb->cxy_na[numColumnsMax*gridId];
794 const int *d_cxy_ind = &nb->cxy_ind[numColumnsMax*gridId];
795 const auto kernelArgs = prepareGpuKernelArguments(kernelFn, config,
805 launchGpuKernel(kernelFn, config, nullptr, "XbufferOps", kernelArgs);
808 // TODO: note that this is not necessary when there are no local atoms, that is:
809 // (numAtoms == 0 && interactionLoc == InteractionLocality::Local)
810 // but for now we avoid that optimization
811 nbnxnInsertNonlocalGpuDependency(nb, interactionLoc);
814 /* F buffer operations on GPU: performs force summations and conversion from nb to rvec format. */
815 void nbnxn_gpu_add_nbat_f_to_f(const AtomLocality atomLocality,
816 DeviceBuffer<float> totalForcesDevice,
818 void *pmeForcesDevice,
819 gmx::ArrayRef<GpuEventSynchronizer* const> dependencyList,
822 bool useGpuFPmeReduction,
823 bool accumulateForce)
825 GMX_ASSERT(nb, "Need a valid nbnxn_gpu object");
826 GMX_ASSERT(numAtoms != 0, "Cannot call function with no atoms");
827 GMX_ASSERT(totalForcesDevice, "Need a valid totalForcesDevice pointer");
829 const InteractionLocality iLocality = gpuAtomToInteractionLocality(atomLocality);
830 cudaStream_t stream = nb->stream[iLocality];
831 cu_atomdata_t *adat = nb->atdat;
833 size_t gmx_used_in_debug numDependency =
834 static_cast<size_t>((useGpuFPmeReduction == true)) +
835 static_cast<size_t>((accumulateForce == true));
836 GMX_ASSERT(numDependency >= dependencyList.size(), "Mismatching number of dependencies and call signature");
838 // Enqueue wait on all dependencies passed
839 for (auto const synchronizer : dependencyList)
841 synchronizer->enqueueWaitEvent(stream);
846 KernelLaunchConfig config;
847 config.blockSize[0] = c_bufOpsThreadsPerBlock;
848 config.blockSize[1] = 1;
849 config.blockSize[2] = 1;
850 config.gridSize[0] = ((numAtoms+1)+c_bufOpsThreadsPerBlock-1)/c_bufOpsThreadsPerBlock;
851 config.gridSize[1] = 1;
852 config.gridSize[2] = 1;
853 config.sharedMemorySize = 0;
854 config.stream = stream;
856 auto kernelFn = accumulateForce ?
857 nbnxn_gpu_add_nbat_f_to_f_kernel<true, false> :
858 nbnxn_gpu_add_nbat_f_to_f_kernel<false, false>;
860 if (useGpuFPmeReduction)
862 GMX_ASSERT(pmeForcesDevice, "Need a valid pmeForcesDevice pointer");
863 kernelFn = accumulateForce ?
864 nbnxn_gpu_add_nbat_f_to_f_kernel<true, true> :
865 nbnxn_gpu_add_nbat_f_to_f_kernel<false, true>;
868 const float3 *d_fNB = adat->f;
869 const float3 *d_fPme = (float3*) pmeForcesDevice;
870 float3 *d_fTotal = (float3*) totalForcesDevice;
871 const int *d_cell = nb->cell;
873 const auto kernelArgs = prepareGpuKernelArguments(kernelFn, config,
881 launchGpuKernel(kernelFn, config, nullptr, "FbufferOps", kernelArgs);
883 if (atomLocality == AtomLocality::Local)
885 GMX_ASSERT(nb->localFReductionDone != nullptr, "localFReductionDone has to be a valid pointer");
886 nb->localFReductionDone->markEvent(stream);
890 void* nbnxn_get_x_on_device_event(const gmx_nbnxn_cuda_t *nb)
892 return static_cast<void*> (nb->xAvailableOnDevice);
895 void nbnxn_wait_nonlocal_x_copy_D2H_done(gmx_nbnxn_cuda_t *nb)
897 nb->xNonLocalCopyD2HDone->waitForEvent();
900 void nbnxn_stream_local_wait_for_nonlocal(gmx_nbnxn_cuda_t *nb)
902 cudaStream_t localStream = nb->stream[InteractionLocality::Local];
903 cudaStream_t nonLocalStream = nb->stream[InteractionLocality::NonLocal];
905 GpuEventSynchronizer event;
906 event.markEvent(nonLocalStream);
907 event.enqueueWaitEvent(localStream);