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/mdlib/force_flags.h"
58 #include "gromacs/nbnxm/atomdata.h"
59 #include "gromacs/nbnxm/gpu_common.h"
60 #include "gromacs/nbnxm/gpu_common_utils.h"
61 #include "gromacs/nbnxm/gpu_data_mgmt.h"
62 #include "gromacs/nbnxm/nbnxm.h"
63 #include "gromacs/nbnxm/pairlist.h"
64 #include "gromacs/timing/gpu_timing.h"
65 #include "gromacs/utility/cstringutil.h"
66 #include "gromacs/utility/gmxassert.h"
68 #include "nbnxm_cuda_types.h"
70 /***** The kernel declarations/definitions come here *****/
72 /* Top-level kernel declaration generation: will generate through multiple
73 * inclusion the following flavors for all kernel declarations:
74 * - force-only output;
75 * - force and energy output;
76 * - force-only with pair list pruning;
77 * - force and energy output with pair list pruning.
79 #define FUNCTION_DECLARATION_ONLY
81 #include "gromacs/nbnxm/cuda/nbnxm_cuda_kernels.cuh"
82 /** Force & energy **/
84 #include "gromacs/nbnxm/cuda/nbnxm_cuda_kernels.cuh"
87 /*** Pair-list pruning kernels ***/
90 #include "gromacs/nbnxm/cuda/nbnxm_cuda_kernels.cuh"
91 /** Force & energy **/
93 #include "gromacs/nbnxm/cuda/nbnxm_cuda_kernels.cuh"
97 /* Prune-only kernels */
98 #include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel_pruneonly.cuh"
99 #undef FUNCTION_DECLARATION_ONLY
101 /* Now generate the function definitions if we are using a single compilation unit. */
102 #if GMX_CUDA_NB_SINGLE_COMPILATION_UNIT
103 #include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel_F_noprune.cu"
104 #include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel_F_prune.cu"
105 #include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel_VF_noprune.cu"
106 #include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel_VF_prune.cu"
107 #include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel_pruneonly.cu"
108 #endif /* GMX_CUDA_NB_SINGLE_COMPILATION_UNIT */
114 /*! Nonbonded kernel function pointer type */
115 typedef void (*nbnxn_cu_kfunc_ptr_t)(const cu_atomdata_t,
120 /*********************************/
122 /*! Returns the number of blocks to be used for the nonbonded GPU kernel. */
123 static inline int calc_nb_kernel_nblock(int nwork_units, const gmx_device_info_t *dinfo)
128 /* CUDA does not accept grid dimension of 0 (which can happen e.g. with an
129 empty domain) and that case should be handled before this point. */
130 assert(nwork_units > 0);
132 max_grid_x_size = dinfo->prop.maxGridSize[0];
134 /* do we exceed the grid x dimension limit? */
135 if (nwork_units > max_grid_x_size)
137 gmx_fatal(FARGS, "Watch out, the input system is too large to simulate!\n"
138 "The number of nonbonded work units (=number of super-clusters) exceeds the"
139 "maximum grid size in x dimension (%d > %d)!", nwork_units, max_grid_x_size);
146 /* Constant arrays listing all kernel function pointers and enabling selection
147 of a kernel in an elegant manner. */
149 /*! Pointers to the non-bonded kernels organized in 2-dim arrays by:
150 * 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 /*! Force-only kernel function pointers. */
158 static const nbnxn_cu_kfunc_ptr_t nb_kfunc_noener_noprune_ptr[eelCuNR][evdwCuNR] =
160 { 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 },
161 { 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 },
162 { 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 },
163 { 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 },
164 { 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 },
165 { 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 }
168 /*! Force + energy kernel function pointers. */
169 static const nbnxn_cu_kfunc_ptr_t nb_kfunc_ener_noprune_ptr[eelCuNR][evdwCuNR] =
171 { 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 },
172 { 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 },
173 { 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 },
174 { 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 },
175 { 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 },
176 { 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 }
179 /*! Force + pruning kernel function pointers. */
180 static const nbnxn_cu_kfunc_ptr_t nb_kfunc_noener_prune_ptr[eelCuNR][evdwCuNR] =
182 { 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 },
183 { 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 },
184 { 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 },
185 { 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 },
186 { 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 },
187 { 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 }
190 /*! Force + energy + pruning kernel function pointers. */
191 static const nbnxn_cu_kfunc_ptr_t nb_kfunc_ener_prune_ptr[eelCuNR][evdwCuNR] =
193 { 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 },
194 { 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 },
195 { 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 },
196 { 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 },
197 { 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 },
198 { 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 }
201 /*! Return a pointer to the kernel version to be executed at the current step. */
202 static inline nbnxn_cu_kfunc_ptr_t select_nbnxn_kernel(int eeltype,
206 const gmx_device_info_t gmx_unused *devInfo)
208 nbnxn_cu_kfunc_ptr_t res;
210 GMX_ASSERT(eeltype < eelCuNR,
211 "The electrostatics type requested is not implemented in the CUDA kernels.");
212 GMX_ASSERT(evdwtype < evdwCuNR,
213 "The VdW type requested is not implemented in the CUDA kernels.");
215 /* assert assumptions made by the kernels */
216 GMX_ASSERT(c_nbnxnGpuClusterSize*c_nbnxnGpuClusterSize/c_nbnxnGpuClusterpairSplit == devInfo->prop.warpSize,
217 "The CUDA kernels require the cluster_size_i*cluster_size_j/nbnxn_gpu_clusterpair_split to match the warp size of the architecture targeted.");
223 res = nb_kfunc_ener_prune_ptr[eeltype][evdwtype];
227 res = nb_kfunc_ener_noprune_ptr[eeltype][evdwtype];
234 res = nb_kfunc_noener_prune_ptr[eeltype][evdwtype];
238 res = nb_kfunc_noener_noprune_ptr[eeltype][evdwtype];
245 /*! \brief Calculates the amount of shared memory required by the nonbonded kernel in use. */
246 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)
252 /* size of shmem (force-buffers/xq/atom type preloading) */
253 /* NOTE: with the default kernel on sm3.0 we need shmem only for pre-loading */
254 /* i-atom x+q in shared memory */
255 shmem = c_numClPerSupercl * c_clSize * sizeof(float4);
256 /* cj in shared memory, for each warp separately */
257 shmem += num_threads_z * c_nbnxnGpuClusterpairSplit * c_nbnxnGpuJgroupSize * sizeof(int);
259 if (nbp->vdwtype == evdwCuCUTCOMBGEOM ||
260 nbp->vdwtype == evdwCuCUTCOMBLB)
262 /* i-atom LJ combination parameters in shared memory */
263 shmem += c_numClPerSupercl * c_clSize * sizeof(float2);
267 /* i-atom types in shared memory */
268 shmem += c_numClPerSupercl * c_clSize * sizeof(int);
274 /*! \brief Launch asynchronously the xq buffer host to device copy. */
275 void gpu_copy_xq_to_gpu(gmx_nbnxn_cuda_t *nb,
276 const nbnxn_atomdata_t *nbatom,
277 const AtomLocality atomLocality,
278 const bool haveOtherWork)
280 GMX_ASSERT(atomLocality == AtomLocality::Local || atomLocality == AtomLocality::NonLocal,
281 "Only local and non-local xq transfers are supported");
283 const InteractionLocality iloc = gpuAtomToInteractionLocality(atomLocality);
285 int adat_begin, adat_len; /* local/nonlocal offset and length used for xq and f */
287 cu_atomdata_t *adat = nb->atdat;
288 cu_plist_t *plist = nb->plist[iloc];
289 cu_timers_t *t = nb->timers;
290 cudaStream_t stream = nb->stream[iloc];
292 bool bDoTime = nb->bDoTime;
294 /* Don't launch the non-local H2D copy if there is no dependent
295 work to do: neither non-local nor other (e.g. bonded) work
296 to do that has as input the nbnxn coordaintes.
297 Doing the same for the local kernel is more complicated, since the
298 local part of the force array also depends on the non-local kernel.
299 So to avoid complicating the code and to reduce the risk of bugs,
300 we always call the local local x+q copy (and the rest of the local
301 work in nbnxn_gpu_launch_kernel().
303 if (!haveOtherWork && canSkipWork(*nb, iloc))
305 plist->haveFreshList = false;
310 /* calculate the atom data index range based on locality */
311 if (atomLocality == AtomLocality::Local)
314 adat_len = adat->natoms_local;
318 adat_begin = adat->natoms_local;
319 adat_len = adat->natoms - adat->natoms_local;
322 /* beginning of timed HtoD section */
325 t->xf[atomLocality].nb_h2d.openTimingRegion(stream);
329 cu_copy_H2D_async(adat->xq + adat_begin,
330 static_cast<const void *>(nbatom->x().data() + adat_begin * 4),
331 adat_len * sizeof(*adat->xq), stream);
335 t->xf[atomLocality].nb_h2d.closeTimingRegion(stream);
338 /* When we get here all misc operations issued in the local stream as well as
339 the local xq H2D are done,
340 so we record that in the local stream and wait for it in the nonlocal one.
341 This wait needs to precede any PP tasks, bonded or nonbonded, that may
342 compute on interactions between local and nonlocal atoms.
344 if (nb->bUseTwoStreams)
346 if (iloc == InteractionLocality::Local)
348 cudaError_t stat = cudaEventRecord(nb->misc_ops_and_local_H2D_done, stream);
349 CU_RET_ERR(stat, "cudaEventRecord on misc_ops_and_local_H2D_done failed");
353 cudaError_t stat = cudaStreamWaitEvent(stream, nb->misc_ops_and_local_H2D_done, 0);
354 CU_RET_ERR(stat, "cudaStreamWaitEvent on misc_ops_and_local_H2D_done failed");
359 /*! As we execute nonbonded workload in separate streams, before launching
360 the kernel we need to make sure that he following operations have completed:
361 - atomdata allocation and related H2D transfers (every nstlist step);
362 - pair list H2D transfer (every nstlist step);
363 - shift vector H2D transfer (every nstlist step);
364 - force (+shift force and energy) output clearing (every step).
366 These operations are issued in the local stream at the beginning of the step
367 and therefore always complete before the local kernel launch. The non-local
368 kernel is launched after the local on the same device/context hence it is
369 inherently scheduled after the operations in the local stream (including the
370 above "misc_ops") on pre-GK110 devices with single hardware queue, but on later
371 devices with multiple hardware queues the dependency needs to be enforced.
372 We use the misc_ops_and_local_H2D_done event to record the point where
373 the local x+q H2D (and all preceding) tasks are complete and synchronize
374 with this event in the non-local stream before launching the non-bonded kernel.
376 void gpu_launch_kernel(gmx_nbnxn_cuda_t *nb,
378 const InteractionLocality iloc)
380 cu_atomdata_t *adat = nb->atdat;
381 cu_nbparam_t *nbp = nb->nbparam;
382 cu_plist_t *plist = nb->plist[iloc];
383 cu_timers_t *t = nb->timers;
384 cudaStream_t stream = nb->stream[iloc];
386 bool bCalcEner = flags & GMX_FORCE_ENERGY;
387 bool bCalcFshift = flags & GMX_FORCE_VIRIAL;
388 bool bDoTime = nb->bDoTime;
390 /* Don't launch the non-local kernel if there is no work to do.
391 Doing the same for the local kernel is more complicated, since the
392 local part of the force array also depends on the non-local kernel.
393 So to avoid complicating the code and to reduce the risk of bugs,
394 we always call the local kernel, and later (not in
395 this function) the stream wait, local f copyback and the f buffer
396 clearing. All these operations, except for the local interaction kernel,
397 are needed for the non-local interactions. The skip of the local kernel
398 call is taken care of later in this function. */
399 if (canSkipWork(*nb, iloc))
401 plist->haveFreshList = false;
406 if (nbp->useDynamicPruning && plist->haveFreshList)
408 /* Prunes for rlistOuter and rlistInner, sets plist->haveFreshList=false
409 (TODO: ATM that's the way the timing accounting can distinguish between
410 separate prune kernel and combined force+prune, maybe we need a better way?).
412 gpu_launch_kernel_pruneonly(nb, iloc, 1);
415 if (plist->nsci == 0)
417 /* Don't launch an empty local kernel (not allowed with CUDA) */
421 /* beginning of timed nonbonded calculation section */
424 t->interaction[iloc].nb_k.openTimingRegion(stream);
427 /* Kernel launch config:
428 * - The thread block dimensions match the size of i-clusters, j-clusters,
429 * and j-cluster concurrency, in x, y, and z, respectively.
430 * - The 1D block-grid contains as many blocks as super-clusters.
432 int num_threads_z = 1;
433 if (nb->dev_info->prop.major == 3 && nb->dev_info->prop.minor == 7)
437 int nblock = calc_nb_kernel_nblock(plist->nsci, nb->dev_info);
440 KernelLaunchConfig config;
441 config.blockSize[0] = c_clSize;
442 config.blockSize[1] = c_clSize;
443 config.blockSize[2] = num_threads_z;
444 config.gridSize[0] = nblock;
445 config.sharedMemorySize = calc_shmem_required_nonbonded(num_threads_z, nb->dev_info, nbp);
446 config.stream = stream;
450 fprintf(debug, "Non-bonded GPU launch configuration:\n\tThread block: %zux%zux%zu\n\t"
451 "\tGrid: %zux%zu\n\t#Super-clusters/clusters: %d/%d (%d)\n"
453 config.blockSize[0], config.blockSize[1], config.blockSize[2],
454 config.gridSize[0], config.gridSize[1], plist->nsci*c_numClPerSupercl,
455 c_numClPerSupercl, plist->na_c,
456 config.sharedMemorySize);
459 auto *timingEvent = bDoTime ? t->interaction[iloc].nb_k.fetchNextEvent() : nullptr;
460 const auto kernel = select_nbnxn_kernel(nbp->eeltype,
463 (plist->haveFreshList && !nb->timers->interaction[iloc].didPrune),
465 const auto kernelArgs = prepareGpuKernelArguments(kernel, config, adat, nbp, plist, &bCalcFshift);
466 launchGpuKernel(kernel, config, timingEvent, "k_calc_nb", kernelArgs);
470 t->interaction[iloc].nb_k.closeTimingRegion(stream);
473 if (GMX_NATIVE_WINDOWS)
475 /* Windows: force flushing WDDM queue */
476 cudaStreamQuery(stream);
480 /*! Calculates the amount of shared memory required by the CUDA kernel in use. */
481 static inline int calc_shmem_required_prune(const int num_threads_z)
485 /* i-atom x in shared memory */
486 shmem = c_numClPerSupercl * c_clSize * sizeof(float4);
487 /* cj in shared memory, for each warp separately */
488 shmem += num_threads_z * c_nbnxnGpuClusterpairSplit * c_nbnxnGpuJgroupSize * sizeof(int);
493 void gpu_launch_kernel_pruneonly(gmx_nbnxn_cuda_t *nb,
494 const InteractionLocality iloc,
497 cu_atomdata_t *adat = nb->atdat;
498 cu_nbparam_t *nbp = nb->nbparam;
499 cu_plist_t *plist = nb->plist[iloc];
500 cu_timers_t *t = nb->timers;
501 cudaStream_t stream = nb->stream[iloc];
503 bool bDoTime = nb->bDoTime;
505 if (plist->haveFreshList)
507 GMX_ASSERT(numParts == 1, "With first pruning we expect 1 part");
509 /* Set rollingPruningNumParts to signal that it is not set */
510 plist->rollingPruningNumParts = 0;
511 plist->rollingPruningPart = 0;
515 if (plist->rollingPruningNumParts == 0)
517 plist->rollingPruningNumParts = numParts;
521 GMX_ASSERT(numParts == plist->rollingPruningNumParts, "It is not allowed to change numParts in between list generation steps");
525 /* Use a local variable for part and update in plist, so we can return here
526 * without duplicating the part increment code.
528 int part = plist->rollingPruningPart;
530 plist->rollingPruningPart++;
531 if (plist->rollingPruningPart >= plist->rollingPruningNumParts)
533 plist->rollingPruningPart = 0;
536 /* Compute the number of list entries to prune in this pass */
537 int numSciInPart = (plist->nsci - part)/numParts;
539 /* Don't launch the kernel if there is no work to do (not allowed with CUDA) */
540 if (numSciInPart <= 0)
542 plist->haveFreshList = false;
547 GpuRegionTimer *timer = nullptr;
550 timer = &(plist->haveFreshList ? t->interaction[iloc].prune_k : t->interaction[iloc].rollingPrune_k);
553 /* beginning of timed prune calculation section */
556 timer->openTimingRegion(stream);
559 /* Kernel launch config:
560 * - The thread block dimensions match the size of i-clusters, j-clusters,
561 * and j-cluster concurrency, in x, y, and z, respectively.
562 * - The 1D block-grid contains as many blocks as super-clusters.
564 int num_threads_z = c_cudaPruneKernelJ4Concurrency;
565 int nblock = calc_nb_kernel_nblock(numSciInPart, nb->dev_info);
566 KernelLaunchConfig config;
567 config.blockSize[0] = c_clSize;
568 config.blockSize[1] = c_clSize;
569 config.blockSize[2] = num_threads_z;
570 config.gridSize[0] = nblock;
571 config.sharedMemorySize = calc_shmem_required_prune(num_threads_z);
572 config.stream = stream;
576 fprintf(debug, "Pruning GPU kernel launch configuration:\n\tThread block: %zux%zux%zu\n\t"
577 "\tGrid: %zux%zu\n\t#Super-clusters/clusters: %d/%d (%d)\n"
579 config.blockSize[0], config.blockSize[1], config.blockSize[2],
580 config.gridSize[0], config.gridSize[1], numSciInPart*c_numClPerSupercl,
581 c_numClPerSupercl, plist->na_c,
582 config.sharedMemorySize);
585 auto *timingEvent = bDoTime ? timer->fetchNextEvent() : nullptr;
586 constexpr char kernelName[] = "k_pruneonly";
587 const auto kernel = plist->haveFreshList ? nbnxn_kernel_prune_cuda<true> : nbnxn_kernel_prune_cuda<false>;
588 const auto kernelArgs = prepareGpuKernelArguments(kernel, config, adat, nbp, plist, &numParts, &part);
589 launchGpuKernel(kernel, config, timingEvent, kernelName, kernelArgs);
591 /* TODO: consider a more elegant way to track which kernel has been called
592 (combined or separate 1st pass prune, rolling prune). */
593 if (plist->haveFreshList)
595 plist->haveFreshList = false;
596 /* Mark that pruning has been done */
597 nb->timers->interaction[iloc].didPrune = true;
601 /* Mark that rolling pruning has been done */
602 nb->timers->interaction[iloc].didRollingPrune = true;
607 timer->closeTimingRegion(stream);
610 if (GMX_NATIVE_WINDOWS)
612 /* Windows: force flushing WDDM queue */
613 cudaStreamQuery(stream);
617 void gpu_launch_cpyback(gmx_nbnxn_cuda_t *nb,
618 nbnxn_atomdata_t *nbatom,
620 const AtomLocality atomLocality,
621 const bool haveOtherWork)
624 int adat_begin, adat_len; /* local/nonlocal offset and length used for xq and f */
626 /* determine interaction locality from atom locality */
627 const InteractionLocality iloc = gpuAtomToInteractionLocality(atomLocality);
629 /* extract the data */
630 cu_atomdata_t *adat = nb->atdat;
631 cu_timers_t *t = nb->timers;
632 bool bDoTime = nb->bDoTime;
633 cudaStream_t stream = nb->stream[iloc];
635 bool bCalcEner = flags & GMX_FORCE_ENERGY;
636 bool bCalcFshift = flags & GMX_FORCE_VIRIAL;
638 /* don't launch non-local copy-back if there was no non-local work to do */
639 if (!haveOtherWork && canSkipWork(*nb, iloc))
644 getGpuAtomRange(adat, atomLocality, &adat_begin, &adat_len);
646 /* beginning of timed D2H section */
649 t->xf[atomLocality].nb_d2h.openTimingRegion(stream);
652 /* With DD the local D2H transfer can only start after the non-local
653 kernel has finished. */
654 if (iloc == InteractionLocality::Local && nb->bUseTwoStreams)
656 stat = cudaStreamWaitEvent(stream, nb->nonlocal_done, 0);
657 CU_RET_ERR(stat, "cudaStreamWaitEvent on nonlocal_done failed");
661 cu_copy_D2H_async(nbatom->out[0].f.data() + adat_begin * 3, adat->f + adat_begin,
662 (adat_len)*sizeof(*adat->f), stream);
664 /* After the non-local D2H is launched the nonlocal_done event can be
665 recorded which signals that the local D2H can proceed. This event is not
666 placed after the non-local kernel because we want the non-local data
668 if (iloc == InteractionLocality::NonLocal)
670 stat = cudaEventRecord(nb->nonlocal_done, stream);
671 CU_RET_ERR(stat, "cudaEventRecord on nonlocal_done failed");
674 /* only transfer energies in the local stream */
675 if (iloc == InteractionLocality::Local)
680 cu_copy_D2H_async(nb->nbst.fshift, adat->fshift,
681 SHIFTS * sizeof(*nb->nbst.fshift), stream);
687 cu_copy_D2H_async(nb->nbst.e_lj, adat->e_lj,
688 sizeof(*nb->nbst.e_lj), stream);
689 cu_copy_D2H_async(nb->nbst.e_el, adat->e_el,
690 sizeof(*nb->nbst.e_el), stream);
696 t->xf[atomLocality].nb_d2h.closeTimingRegion(stream);
700 void cuda_set_cacheconfig()
704 for (int i = 0; i < eelCuNR; i++)
706 for (int j = 0; j < evdwCuNR; j++)
708 /* Default kernel 32/32 kB Shared/L1 */
709 cudaFuncSetCacheConfig(nb_kfunc_ener_prune_ptr[i][j], cudaFuncCachePreferEqual);
710 cudaFuncSetCacheConfig(nb_kfunc_ener_noprune_ptr[i][j], cudaFuncCachePreferEqual);
711 cudaFuncSetCacheConfig(nb_kfunc_noener_prune_ptr[i][j], cudaFuncCachePreferEqual);
712 stat = cudaFuncSetCacheConfig(nb_kfunc_noener_noprune_ptr[i][j], cudaFuncCachePreferEqual);
713 CU_RET_ERR(stat, "cudaFuncSetCacheConfig failed");