2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015, 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 * \ingroup module_mdlib
54 #include "gromacs/gpu_utils/oclutils.h"
55 #include "gromacs/hardware/hw_info.h"
56 #include "gromacs/mdlib/force_flags.h"
57 #include "gromacs/mdlib/nb_verlet.h"
58 #include "gromacs/mdlib/nbnxn_consts.h"
59 #include "gromacs/mdlib/nbnxn_pairlist.h"
60 #include "gromacs/timing/gpu_timing.h"
63 #include "thread_mpi/atomic.h"
66 #include "gromacs/mdlib/nbnxn_gpu.h"
67 #include "gromacs/mdlib/nbnxn_gpu_data_mgmt.h"
68 #include "gromacs/pbcutil/ishift.h"
69 #include "gromacs/utility/cstringutil.h"
70 #include "gromacs/utility/fatalerror.h"
71 #include "gromacs/utility/gmxassert.h"
73 #include "nbnxn_ocl_types.h"
75 #if defined TEXOBJ_SUPPORTED && __CUDA_ARCH__ >= 300
79 /*! \brief Convenience defines */
81 #define NCL_PER_SUPERCL (NBNXN_GPU_NCLUSTER_PER_SUPERCLUSTER)
82 #define CL_SIZE (NBNXN_GPU_CLUSTER_SIZE)
85 /*! \brief Always/never run the energy/pruning kernels -- only for benchmarking purposes */
87 static bool always_ener = (getenv("GMX_GPU_ALWAYS_ENER") != NULL);
88 static bool never_ener = (getenv("GMX_GPU_NEVER_ENER") != NULL);
89 static bool always_prune = (getenv("GMX_GPU_ALWAYS_PRUNE") != NULL);
92 /* Uncomment this define to enable kernel debugging */
95 /*! \brief Specifies which kernel run to debug */
96 #define DEBUG_RUN_STEP 2
98 /*! \brief Validates the input global work size parameter.
100 static inline void validate_global_work_size(size_t *global_work_size, int work_dim, gmx_device_info_t *dinfo)
102 cl_uint device_size_t_size_bits;
103 cl_uint host_size_t_size_bits;
107 /* Each component of a global_work_size must not exceed the range given by the
108 sizeof(device size_t) for the device on which the kernel execution will
110 https://www.khronos.org/registry/cl/sdk/1.0/docs/man/xhtml/clEnqueueNDRangeKernel.html
112 device_size_t_size_bits = dinfo->adress_bits;
113 host_size_t_size_bits = (cl_uint)(sizeof(size_t) * 8);
115 /* If sizeof(host size_t) <= sizeof(device size_t)
116 => global_work_size components will always be valid
118 => get device limit for global work size and
119 compare it against each component of global_work_size.
121 if (host_size_t_size_bits > device_size_t_size_bits)
125 device_limit = (((size_t)1) << device_size_t_size_bits) - 1;
127 for (int i = 0; i < work_dim; i++)
129 if (global_work_size[i] > device_limit)
131 gmx_fatal(FARGS, "Watch out, the input system is too large to simulate!\n"
132 "The number of nonbonded work units (=number of super-clusters) exceeds the"
133 "device capabilities. Global work size limit exceeded (%d > %d)!",
134 global_work_size[i], device_limit);
140 /* Constant arrays listing non-bonded kernel function names. The arrays are
141 * organized in 2-dim arrays by: electrostatics and VDW type.
143 * Note that the row- and column-order of function pointers has to match the
144 * order of corresponding enumerated electrostatics and vdw types, resp.,
145 * defined in nbnxn_cuda_types.h.
148 /*! \brief Force-only kernel function names. */
149 static const char* nb_kfunc_noener_noprune_ptr[eelOclNR][evdwOclNR] =
151 { "nbnxn_kernel_ElecCut_VdwLJ_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" },
152 { "nbnxn_kernel_ElecRF_VdwLJ_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" },
153 { "nbnxn_kernel_ElecEwQSTab_VdwLJ_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" },
154 { "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_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" },
155 { "nbnxn_kernel_ElecEw_VdwLJ_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" },
156 { "nbnxn_kernel_ElecEwTwinCut_VdwLJ_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" }
159 /*! \brief Force + energy kernel function pointers. */
160 static const char* nb_kfunc_ener_noprune_ptr[eelOclNR][evdwOclNR] =
162 { "nbnxn_kernel_ElecCut_VdwLJ_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" },
163 { "nbnxn_kernel_ElecRF_VdwLJ_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" },
164 { "nbnxn_kernel_ElecEwQSTab_VdwLJ_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" },
165 { "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_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" },
166 { "nbnxn_kernel_ElecEw_VdwLJ_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" },
167 { "nbnxn_kernel_ElecEwTwinCut_VdwLJ_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" }
170 /*! \brief Force + pruning kernel function pointers. */
171 static const char* nb_kfunc_noener_prune_ptr[eelOclNR][evdwOclNR] =
173 { "nbnxn_kernel_ElecCut_VdwLJ_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" },
174 { "nbnxn_kernel_ElecRF_VdwLJ_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" },
175 { "nbnxn_kernel_ElecEwQSTab_VdwLJ_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" },
176 { "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_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" },
177 { "nbnxn_kernel_ElecEw_VdwLJ_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" },
178 { "nbnxn_kernel_ElecEwTwinCut_VdwLJ_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" }
181 /*! \brief Force + energy + pruning kernel function pointers. */
182 static const char* nb_kfunc_ener_prune_ptr[eelOclNR][evdwOclNR] =
184 { "nbnxn_kernel_ElecCut_VdwLJ_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" },
185 { "nbnxn_kernel_ElecRF_VdwLJ_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" },
186 { "nbnxn_kernel_ElecEwQSTab_VdwLJ_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" },
187 { "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_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" },
188 { "nbnxn_kernel_ElecEw_VdwLJ_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" },
189 { "nbnxn_kernel_ElecEwTwinCut_VdwLJ_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" }
192 /*! \brief Return a pointer to the kernel version to be executed at the current step.
193 * OpenCL kernel objects are cached in nb. If the requested kernel is not
194 * found in the cache, it will be created and the cache will be updated.
196 static inline cl_kernel select_nbnxn_kernel(gmx_nbnxn_ocl_t *nb,
202 const char* kernel_name_to_run;
203 cl_kernel *kernel_ptr;
206 assert(eeltype < eelOclNR);
207 assert(evdwtype < eelOclNR);
213 kernel_name_to_run = nb_kfunc_ener_prune_ptr[eeltype][evdwtype];
214 kernel_ptr = &(nb->kernel_ener_prune_ptr[eeltype][evdwtype]);
218 kernel_name_to_run = nb_kfunc_ener_noprune_ptr[eeltype][evdwtype];
219 kernel_ptr = &(nb->kernel_ener_noprune_ptr[eeltype][evdwtype]);
226 kernel_name_to_run = nb_kfunc_noener_prune_ptr[eeltype][evdwtype];
227 kernel_ptr = &(nb->kernel_noener_prune_ptr[eeltype][evdwtype]);
231 kernel_name_to_run = nb_kfunc_noener_noprune_ptr[eeltype][evdwtype];
232 kernel_ptr = &(nb->kernel_noener_noprune_ptr[eeltype][evdwtype]);
236 if (NULL == kernel_ptr[0])
238 *kernel_ptr = clCreateKernel(nb->dev_info->program, kernel_name_to_run, &cl_error);
239 assert(cl_error == CL_SUCCESS);
241 // TODO: handle errors
246 /*! \brief Calculates the amount of shared memory required by the OpenCL kernel in use.
248 static inline int calc_shmem_required()
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 = NCL_PER_SUPERCL * CL_SIZE * sizeof(float) * 4; /* xqib */
256 /* cj in shared memory, for both warps separately */
257 shmem += 2 * NBNXN_GPU_JGROUP_SIZE * sizeof(int); /* cjs */
259 /* FIXME: this should not be compile-time decided but rather at runtime.
260 * This issue propagated from the CUDA code where due to the source to source
261 * compilation there was confusion the way to set up arch-dependent launch parameters.
262 * Here too this should be converted to a hardware/arch/generation dependent
263 * conditional when re-evaluating the need for i atom type preloading.
265 /* i-atom types in shared memory */
266 #pragma error "Should not be defined"
267 shmem += NCL_PER_SUPERCL * CL_SIZE * sizeof(int); /* atib */
269 /* force reduction buffers in shared memory */
270 shmem += CL_SIZE * CL_SIZE * 3 * sizeof(float); /* f_buf */
271 /* Warp vote. In fact it must be * number of warps in block.. */
272 shmem += sizeof(cl_uint) * 2; /* warp_any */
276 /*! \brief Initializes data structures that are going to be sent to the OpenCL device.
278 * The device can't use the same data structures as the host for two main reasons:
279 * - OpenCL restrictions (pointers are not accepted inside data structures)
280 * - some host side fields are not needed for the OpenCL kernels.
282 static void fillin_ocl_structures(cl_nbparam_t *nbp,
283 cl_nbparam_params_t *nbparams_params)
285 nbparams_params->coulomb_tab_scale = nbp->coulomb_tab_scale;
286 nbparams_params->coulomb_tab_size = nbp->coulomb_tab_size;
287 nbparams_params->c_rf = nbp->c_rf;
288 nbparams_params->dispersion_shift = nbp->dispersion_shift;
289 nbparams_params->eeltype = nbp->eeltype;
290 nbparams_params->epsfac = nbp->epsfac;
291 nbparams_params->ewaldcoeff_lj = nbp->ewaldcoeff_lj;
292 nbparams_params->ewald_beta = nbp->ewald_beta;
293 nbparams_params->rcoulomb_sq = nbp->rcoulomb_sq;
294 nbparams_params->repulsion_shift = nbp->repulsion_shift;
295 nbparams_params->rlist_sq = nbp->rlist_sq;
296 nbparams_params->rvdw_sq = nbp->rvdw_sq;
297 nbparams_params->rvdw_switch = nbp->rvdw_switch;
298 nbparams_params->sh_ewald = nbp->sh_ewald;
299 nbparams_params->sh_lj_ewald = nbp->sh_lj_ewald;
300 nbparams_params->two_k_rf = nbp->two_k_rf;
301 nbparams_params->vdwtype = nbp->vdwtype;
302 nbparams_params->vdw_switch = nbp->vdw_switch;
305 /*! \brief Waits for the commands associated with the input event to finish.
306 * Then it releases the event and sets it to 0.
307 * Don't use this function when more than one wait will be issued for the event.
309 void wait_ocl_event(cl_event *ocl_event)
311 cl_int gmx_unused cl_error;
313 /* Blocking wait for the event */
314 cl_error = clWaitForEvents(1, ocl_event);
315 assert(CL_SUCCESS == cl_error);
317 /* Release event and reset it to 0 */
318 cl_error = clReleaseEvent(*ocl_event);
319 assert(CL_SUCCESS == cl_error);
323 /*! \brief Enqueues a wait for event completion.
325 * Then it releases the event and sets it to 0.
326 * Don't use this function when more than one wait will be issued for the event.
327 * Equivalent to Cuda Stream Sync. */
328 void sync_ocl_event(cl_command_queue stream, cl_event *ocl_event)
330 cl_int gmx_unused cl_error;
333 #ifdef CL_VERSION_1_2
334 cl_error = clEnqueueBarrierWithWaitList(stream, 1, ocl_event, NULL);
336 cl_error = clEnqueueWaitForEvents(stream, 1, ocl_event);
339 GMX_RELEASE_ASSERT(CL_SUCCESS == cl_error, ocl_get_error_string(cl_error));
341 /* Release event and reset it to 0. It is ok to release it as enqueuewaitforevents performs implicit retain for events. */
342 cl_error = clReleaseEvent(*ocl_event);
343 assert(CL_SUCCESS == cl_error);
347 /*! \brief Returns the duration in milliseconds for the command associated with the event.
349 * It then releases the event and sets it to 0.
350 * Before calling this function, make sure the command has finished either by
351 * calling clFinish or clWaitForEvents.
352 * The function returns 0.0 if the input event, *ocl_event, is 0.
353 * Don't use this function when more than one wait will be issued for the event.
355 double ocl_event_elapsed_ms(cl_event *ocl_event)
357 cl_int gmx_unused cl_error;
358 cl_ulong start_ns, end_ns;
362 assert(NULL != ocl_event);
366 cl_error = clGetEventProfilingInfo(*ocl_event, CL_PROFILING_COMMAND_START,
367 sizeof(cl_ulong), &start_ns, NULL);
368 assert(CL_SUCCESS == cl_error);
370 cl_error = clGetEventProfilingInfo(*ocl_event, CL_PROFILING_COMMAND_END,
371 sizeof(cl_ulong), &end_ns, NULL);
372 assert(CL_SUCCESS == cl_error);
374 clReleaseEvent(*ocl_event);
377 elapsed_ms = (end_ns - start_ns) / 1000000.0;
383 /*! \brief Launch GPU kernel
385 As we execute nonbonded workload in separate queues, before launching
386 the kernel we need to make sure that he following operations have completed:
387 - atomdata allocation and related H2D transfers (every nstlist step);
388 - pair list H2D transfer (every nstlist step);
389 - shift vector H2D transfer (every nstlist step);
390 - force (+shift force and energy) output clearing (every step).
392 These operations are issued in the local queue at the beginning of the step
393 and therefore always complete before the local kernel launch. The non-local
394 kernel is launched after the local on the same device/context, so this is
395 inherently scheduled after the operations in the local stream (including the
397 However, for the sake of having a future-proof implementation, we use the
398 misc_ops_done event to record the point in time when the above operations
399 are finished and synchronize with this event in the non-local stream.
401 void nbnxn_gpu_launch_kernel(gmx_nbnxn_ocl_t *nb,
402 const struct nbnxn_atomdata_t *nbatom,
407 int adat_begin, adat_len; /* local/nonlocal offset and length used for xq and f */
408 /* OpenCL kernel launch-related stuff */
410 size_t local_work_size[3], global_work_size[3];
411 cl_kernel nb_kernel = NULL; /* fn pointer to the nonbonded kernel */
413 cl_atomdata_t *adat = nb->atdat;
414 cl_nbparam_t *nbp = nb->nbparam;
415 cl_plist_t *plist = nb->plist[iloc];
416 cl_timers_t *t = nb->timers;
417 cl_command_queue stream = nb->stream[iloc];
419 bool bCalcEner = flags & GMX_FORCE_ENERGY;
420 int bCalcFshift = flags & GMX_FORCE_VIRIAL;
421 bool bDoTime = nb->bDoTime;
424 cl_nbparam_params_t nbparams_params;
426 float * debug_buffer_h;
427 size_t debug_buffer_size;
430 /* turn energy calculation always on/off (for debugging/testing only) */
431 bCalcEner = (bCalcEner || always_ener) && !never_ener;
433 /* Don't launch the non-local kernel if there is no work to do.
434 Doing the same for the local kernel is more complicated, since the
435 local part of the force array also depends on the non-local kernel.
436 So to avoid complicating the code and to reduce the risk of bugs,
437 we always call the local kernel, the local x+q copy and later (not in
438 this function) the stream wait, local f copyback and the f buffer
439 clearing. All these operations, except for the local interaction kernel,
440 are needed for the non-local interactions. The skip of the local kernel
441 call is taken care of later in this function. */
442 if (iloc == eintNonlocal && plist->nsci == 0)
447 /* calculate the atom data index range based on locality */
451 adat_len = adat->natoms_local;
455 adat_begin = adat->natoms_local;
456 adat_len = adat->natoms - adat->natoms_local;
459 /* beginning of timed HtoD section */
462 ocl_copy_H2D_async(adat->xq, nbatom->x + adat_begin * 4, adat_begin*sizeof(float)*4,
463 adat_len * sizeof(float) * 4, stream, bDoTime ? (&(t->nb_h2d[iloc])) : NULL);
465 /* When we get here all misc operations issues in the local stream as well as
466 the local xq H2D are done,
467 so we record that in the local stream and wait for it in the nonlocal one. */
468 if (nb->bUseTwoStreams)
470 if (iloc == eintLocal)
472 #ifdef CL_VERSION_1_2
473 cl_error = clEnqueueMarkerWithWaitList(stream, 0, NULL, &(nb->misc_ops_and_local_H2D_done));
475 cl_error = clEnqueueMarker(stream, &(nb->misc_ops_and_local_H2D_done));
477 assert(CL_SUCCESS == cl_error);
479 /* Based on the v1.2 section 5.13 of the OpenCL spec, a flush is needed
480 * in the local stream in order to be able to sync with the above event
481 * from the non-local stream.
483 cl_error = clFlush(stream);
484 assert(CL_SUCCESS == cl_error);
488 sync_ocl_event(stream, &(nb->misc_ops_and_local_H2D_done));
492 if (plist->nsci == 0)
494 /* Don't launch an empty local kernel (is not allowed with OpenCL).
495 * TODO: Separate H2D and kernel launch into separate functions.
500 /* beginning of timed nonbonded calculation section */
502 /* get the pointer to the kernel flavor we need to use */
503 nb_kernel = select_nbnxn_kernel(nb,
507 plist->bDoPrune || always_prune);
509 /* kernel launch config */
510 local_work_size[0] = CL_SIZE;
511 local_work_size[1] = CL_SIZE;
512 local_work_size[2] = 1;
514 global_work_size[0] = plist->nsci * local_work_size[0];
515 global_work_size[1] = 1 * local_work_size[1];
516 global_work_size[2] = 1 * local_work_size[2];
518 validate_global_work_size(global_work_size, 3, nb->dev_info);
520 shmem = calc_shmem_required();
524 static int run_step = 1;
526 if (DEBUG_RUN_STEP == run_step)
528 debug_buffer_size = global_work_size[0] * global_work_size[1] * global_work_size[2] * sizeof(float);
529 debug_buffer_h = (float*)calloc(1, debug_buffer_size);
530 assert(NULL != debug_buffer_h);
532 if (NULL == nb->debug_buffer)
534 nb->debug_buffer = clCreateBuffer(nb->dev_info->context, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
535 debug_buffer_size, debug_buffer_h, &cl_error);
537 assert(CL_SUCCESS == cl_error);
546 fprintf(debug, "GPU launch configuration:\n\tLocal work size: %dx%dx%d\n\t"
547 "Global work size : %dx%d\n\t#Super-clusters/clusters: %d/%d (%d)\n",
548 (int)(local_work_size[0]), (int)(local_work_size[1]), (int)(local_work_size[2]),
549 (int)(global_work_size[0]), (int)(global_work_size[1]), plist->nsci*NCL_PER_SUPERCL,
550 NCL_PER_SUPERCL, plist->na_c);
553 fillin_ocl_structures(nbp, &nbparams_params);
556 cl_error = clSetKernelArg(nb_kernel, arg_no++, sizeof(int), &(adat->ntypes));
557 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(nbparams_params), &(nbparams_params));
558 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(adat->xq));
559 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(adat->f));
560 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(adat->e_lj));
561 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(adat->e_el));
562 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(adat->fshift));
563 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(adat->atom_types));
564 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(adat->shift_vec));
565 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(nbp->nbfp_climg2d));
566 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(nbp->nbfp_comb_climg2d));
567 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(nbp->coulomb_tab_climg2d));
568 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(plist->sci));
569 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(plist->cj4));
570 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(plist->excl));
571 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(int), &bCalcFshift);
572 cl_error |= clSetKernelArg(nb_kernel, arg_no++, shmem, NULL);
573 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(nb->debug_buffer));
575 assert(cl_error == CL_SUCCESS);
579 printf("ClERROR! %d\n", cl_error);
581 cl_error = clEnqueueNDRangeKernel(stream, nb_kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, bDoTime ? &(t->nb_k[iloc]) : NULL);
582 assert(cl_error == CL_SUCCESS);
586 static int run_step = 1;
588 if (DEBUG_RUN_STEP == run_step)
591 char file_name[256] = {0};
593 ocl_copy_D2H_async(debug_buffer_h, nb->debug_buffer, 0,
594 debug_buffer_size, stream, NULL);
596 // Make sure all data has been transfered back from device
599 printf("\nWriting debug_buffer to debug_buffer_ocl.txt...");
601 sprintf(file_name, "debug_buffer_ocl_%d.txt", DEBUG_RUN_STEP);
602 pf = fopen(file_name, "wt");
605 fprintf(pf, "%20s", "");
606 for (int j = 0; j < global_work_size[0]; j++)
609 sprintf(label, "(wIdx=%2d thIdx=%2d)", j / local_work_size[0], j % local_work_size[0]);
610 fprintf(pf, "%20s", label);
613 for (int i = 0; i < global_work_size[1]; i++)
616 sprintf(label, "(wIdy=%2d thIdy=%2d)", i / local_work_size[1], i % local_work_size[1]);
617 fprintf(pf, "\n%20s", label);
619 for (int j = 0; j < global_work_size[0]; j++)
621 fprintf(pf, "%20.5f", debug_buffer_h[i * global_work_size[0] + j]);
632 free(debug_buffer_h);
633 debug_buffer_h = NULL;
641 /*! \brief Debugging helper function */
642 void dump_compare_results_cj4(nbnxn_cj4_t* results, int cnt, char* out_file, char* ref_file)
646 pf = fopen(out_file, "wt");
649 fprintf(pf, "%20s%20s%20s%20s%20s%20s%20s%20s\n",
650 "cj[0]", "cj[1]", "cj[2]", "cj[3]",
651 "imei[0].excl_ind", "imei[0].imask",
652 "imei[1].excl_ind", "imei[1].imask");
654 for (int index = 0; index < cnt; index++)
656 fprintf(pf, "%20d%20d%20d%20d%20d%20u%20d%20u\n",
657 results[index].cj[0], results[index].cj[1], results[index].cj[2], results[index].cj[3],
658 results[index].imei[0].excl_ind, results[index].imei[0].imask,
659 results[index].imei[1].excl_ind, results[index].imei[1].imask);
664 printf("\nWrote results to %s", out_file);
666 pf = fopen(ref_file, "rt");
671 printf("\n%s file found. Comparing results...", ref_file);
673 /* Skip the first line */
677 if (1 != fscanf(pf, "%c", &c))
683 for (int index = 0; index < cnt; index++)
686 unsigned int u_ref_val;
688 for (int j = 0; j < 4; j++)
690 if (1 != fscanf(pf, "%20d", &ref_val))
695 if (ref_val != results[index].cj[j])
697 printf("\nDifference for cj[%d] at index %d computed value = %d reference value = %d",
698 j, index, results[index].cj[j], ref_val);
704 for (int j = 0; j < 2; j++)
706 if (1 != fscanf(pf, "%20d", &ref_val))
711 if (ref_val != results[index].imei[j].excl_ind)
713 printf("\nDifference for imei[%d].excl_ind at index %d computed value = %d reference value = %d",
714 j, index, results[index].imei[j].excl_ind, ref_val);
719 if (1 != fscanf(pf, "%20u", &u_ref_val))
724 if (u_ref_val != results[index].imei[j].imask)
726 printf("\nDifference for imei[%d].imask at index %d computed value = %u reference value = %u",
727 j, index, results[index].imei[j].imask, u_ref_val);
735 printf("\nFinished comparing results. Total number of differences: %d", diff);
740 printf("\n%s file not found. No comparison performed.", ref_file);
744 /*! \brief Debugging helper function */
745 void dump_compare_results_f(float* results, int cnt, char* out_file, char* ref_file)
748 float cmp_eps = 0.001f;
750 pf = fopen(out_file, "wt");
753 for (int index = 0; index < cnt; index++)
755 fprintf(pf, "%15.5f\n", results[index]);
760 printf("\nWrote results to %s", out_file);
762 pf = fopen(ref_file, "rt");
766 printf("\n%s file found. Comparing results...", ref_file);
767 for (int index = 0; index < cnt; index++)
770 if (1 != fscanf(pf, "%20f", &ref_val))
775 if (((ref_val - results[index]) > cmp_eps) ||
776 ((ref_val - results[index]) < -cmp_eps))
778 printf("\nDifference at index %d computed value = %15.5f reference value = %15.5f",
779 index, results[index], ref_val);
785 printf("\nFinished comparing results. Total number of differences: %d", diff);
790 printf("\n%s file not found. No comparison performed.", ref_file);
795 * Debug function for dumping cj4, f and fshift buffers.
796 * By default this function does nothing. To enable debugging for any of these
797 * buffers, uncomment the corresponding definition inside the function:
798 * DEBUG_DUMP_CJ4_OCL, DEBUG_DUMP_F_OCL, DEBUG_DUMP_FSHIFT_OCL.
801 void debug_dump_cj4_f_fshift(gmx_nbnxn_ocl_t gmx_unused *nb,
802 const struct nbnxn_atomdata_t gmx_unused *nbatom,
803 cl_command_queue gmx_unused stream,
804 int gmx_unused adat_begin,
805 int gmx_unused adat_len)
807 /* Uncomment this define to enable cj4 debugging for the first kernel run */
808 //#define DEBUG_DUMP_CJ4_OCL
809 #ifdef DEBUG_DUMP_CJ4_OCL
811 static int run_step = 1;
813 if (DEBUG_RUN_STEP == run_step)
815 nbnxn_cj4_t *temp_cj4;
818 char ocl_file_name[256] = {0};
819 char cuda_file_name[256] = {0};
821 cnt = nb->plist[0]->ncj4;
822 size = cnt * sizeof(nbnxn_cj4_t);
823 temp_cj4 = (nbnxn_cj4_t*)malloc(size);
825 ocl_copy_D2H_async(temp_cj4, nb->plist[0]->cj4, 0,
828 // Make sure all data has been transfered back from device
831 sprintf(ocl_file_name, "ocl_cj4_%d.txt", DEBUG_RUN_STEP);
832 sprintf(cuda_file_name, "cuda_cj4_%d.txt", DEBUG_RUN_STEP);
833 dump_compare_results_cj4(temp_cj4, cnt, ocl_file_name, cuda_file_name);
842 /* Uncomment this define to enable f debugging for the first kernel run */
843 //#define DEBUG_DUMP_F_OCL
844 #ifdef DEBUG_DUMP_F_OCL
846 static int run_step = 1;
848 if (DEBUG_RUN_STEP == run_step)
850 char ocl_file_name[256] = {0};
851 char cuda_file_name[256] = {0};
853 // Make sure all data has been transfered back from device
856 sprintf(ocl_file_name, "ocl_f_%d.txt", DEBUG_RUN_STEP);
857 sprintf(cuda_file_name, "cuda_f_%d.txt", DEBUG_RUN_STEP);
859 dump_compare_results_f(nbatom->out[0].f + adat_begin * 3, (adat_len) * 3,
860 ocl_file_name, cuda_file_name);
867 /* Uncomment this define to enable fshift debugging for the first kernel run */
868 //#define DEBUG_DUMP_FSHIFT_OCL
869 #ifdef DEBUG_DUMP_FSHIFT_OCL
871 static int run_step = 1;
873 if (DEBUG_RUN_STEP == run_step)
875 char ocl_file_name[256] = {0};
876 char cuda_file_name[256] = {0};
878 // Make sure all data has been transfered back from device
881 sprintf(ocl_file_name, "ocl_fshift_%d.txt", DEBUG_RUN_STEP);
882 sprintf(cuda_file_name, "cuda_fshift_%d.txt", DEBUG_RUN_STEP);
884 dump_compare_results_f((float*)(nb->nbst.fshift), SHIFTS * 3,
885 ocl_file_name, cuda_file_name);
894 * Launch asynchronously the download of nonbonded forces from the GPU
895 * (and energies/shift forces if required).
897 void nbnxn_gpu_launch_cpyback(gmx_nbnxn_ocl_t *nb,
898 const struct nbnxn_atomdata_t *nbatom,
902 cl_int gmx_unused cl_error;
903 int adat_begin, adat_len; /* local/nonlocal offset and length used for xq and f */
906 /* determine interaction locality from atom locality */
911 else if (NONLOCAL_A(aloc))
918 sprintf(stmp, "Invalid atom locality passed (%d); valid here is only "
919 "local (%d) or nonlocal (%d)", aloc, eatLocal, eatNonlocal);
924 cl_atomdata_t *adat = nb->atdat;
925 cl_timers_t *t = nb->timers;
926 bool bDoTime = nb->bDoTime;
927 cl_command_queue stream = nb->stream[iloc];
929 bool bCalcEner = flags & GMX_FORCE_ENERGY;
930 int bCalcFshift = flags & GMX_FORCE_VIRIAL;
933 /* don't launch non-local copy-back if there was no non-local work to do */
934 if (iloc == eintNonlocal && nb->plist[iloc]->nsci == 0)
936 /* TODO An alternative way to signal that non-local work is
937 complete is to use a clEnqueueMarker+clEnqueueBarrier
938 pair. However, the use of bNonLocalStreamActive has the
939 advantage of being local to the host, so probably minimizes
940 overhead. Curiously, for NVIDIA OpenCL with an empty-domain
941 test case, overall simulation performance was higher with
942 the API calls, but this has not been tested on AMD OpenCL,
943 so could be worth considering in future. */
944 nb->bNonLocalStreamActive = false;
948 /* calculate the atom data index range based on locality */
952 adat_len = adat->natoms_local;
956 adat_begin = adat->natoms_local;
957 adat_len = adat->natoms - adat->natoms_local;
960 /* beginning of timed D2H section */
962 /* With DD the local D2H transfer can only start after the non-local
963 has been launched. */
964 if (iloc == eintLocal && nb->bNonLocalStreamActive)
966 sync_ocl_event(stream, &(nb->nonlocal_done));
970 ocl_copy_D2H_async(nbatom->out[0].f + adat_begin * 3, adat->f, adat_begin*3*sizeof(float),
971 (adat_len)* adat->f_elem_size, stream, bDoTime ? &(t->nb_d2h_f[iloc]) : NULL);
974 cl_error = clFlush(stream);
975 assert(CL_SUCCESS == cl_error);
977 /* After the non-local D2H is launched the nonlocal_done event can be
978 recorded which signals that the local D2H can proceed. This event is not
979 placed after the non-local kernel because we first need the non-local
981 if (iloc == eintNonlocal)
983 #ifdef CL_VERSION_1_2
984 cl_error = clEnqueueMarkerWithWaitList(stream, 0, NULL, &(nb->nonlocal_done));
986 cl_error = clEnqueueMarker(stream, &(nb->nonlocal_done));
988 assert(CL_SUCCESS == cl_error);
989 nb->bNonLocalStreamActive = true;
992 /* only transfer energies in the local stream */
998 ocl_copy_D2H_async(nb->nbst.fshift, adat->fshift, 0,
999 SHIFTS * adat->fshift_elem_size, stream, bDoTime ? &(t->nb_d2h_fshift[iloc]) : NULL);
1005 ocl_copy_D2H_async(nb->nbst.e_lj, adat->e_lj, 0,
1006 sizeof(float), stream, bDoTime ? &(t->nb_d2h_e_lj[iloc]) : NULL);
1008 ocl_copy_D2H_async(nb->nbst.e_el, adat->e_el, 0,
1009 sizeof(float), stream, bDoTime ? &(t->nb_d2h_e_el[iloc]) : NULL);
1013 debug_dump_cj4_f_fshift(nb, nbatom, stream, adat_begin, adat_len);
1017 * Wait for the asynchronously launched nonbonded calculations and data
1018 * transfers to finish.
1020 void nbnxn_gpu_wait_for_gpu(gmx_nbnxn_ocl_t *nb,
1021 int flags, int aloc,
1022 real *e_lj, real *e_el, rvec *fshift)
1024 /* NOTE: only implemented for single-precision at this time */
1025 cl_int gmx_unused cl_error;
1028 /* determine interaction locality from atom locality */
1033 else if (NONLOCAL_A(aloc))
1035 iloc = eintNonlocal;
1040 sprintf(stmp, "Invalid atom locality passed (%d); valid here is only "
1041 "local (%d) or nonlocal (%d)", aloc, eatLocal, eatNonlocal);
1045 cl_plist_t *plist = nb->plist[iloc];
1046 cl_timers_t *timers = nb->timers;
1047 struct gmx_wallclock_gpu_t *timings = nb->timings;
1048 cl_nb_staging nbst = nb->nbst;
1050 bool bCalcEner = flags & GMX_FORCE_ENERGY;
1051 int bCalcFshift = flags & GMX_FORCE_VIRIAL;
1053 /* turn energy calculation always on/off (for debugging/testing only) */
1054 bCalcEner = (bCalcEner || always_ener) && !never_ener;
1056 /* Launch wait/update timers & counters, unless doing the non-local phase
1057 when there is not actually work to do. This is consistent with
1058 nbnxn_gpu_launch_kernel.
1060 NOTE: if timing with multiple GPUs (streams) becomes possible, the
1061 counters could end up being inconsistent due to not being incremented
1062 on some of the nodes! */
1063 if (iloc == eintNonlocal && nb->plist[iloc]->nsci == 0)
1068 /* Actual sync point. Waits for everything to be finished in the command queue. TODO: Find out if a more fine grained solution is needed */
1069 cl_error = clFinish(nb->stream[iloc]);
1070 assert(CL_SUCCESS == cl_error);
1072 /* timing data accumulation */
1075 /* only increase counter once (at local F wait) */
1079 timings->ktime[plist->bDoPrune ? 1 : 0][bCalcEner ? 1 : 0].c += 1;
1082 /* kernel timings */
1084 timings->ktime[plist->bDoPrune ? 1 : 0][bCalcEner ? 1 : 0].t +=
1085 ocl_event_elapsed_ms(timers->nb_k + iloc);
1087 /* X/q H2D and F D2H timings */
1088 timings->nb_h2d_t += ocl_event_elapsed_ms(timers->nb_h2d + iloc);
1089 timings->nb_d2h_t += ocl_event_elapsed_ms(timers->nb_d2h_f + iloc);
1090 timings->nb_d2h_t += ocl_event_elapsed_ms(timers->nb_d2h_fshift + iloc);
1091 timings->nb_d2h_t += ocl_event_elapsed_ms(timers->nb_d2h_e_el + iloc);
1092 timings->nb_d2h_t += ocl_event_elapsed_ms(timers->nb_d2h_e_lj + iloc);
1094 /* only count atdat and pair-list H2D at pair-search step */
1095 if (plist->bDoPrune)
1097 /* atdat transfer timing (add only once, at local F wait) */
1100 timings->pl_h2d_c++;
1101 timings->pl_h2d_t += ocl_event_elapsed_ms(&(timers->atdat));
1104 timings->pl_h2d_t +=
1105 ocl_event_elapsed_ms(timers->pl_h2d_sci + iloc) +
1106 ocl_event_elapsed_ms(timers->pl_h2d_cj4 + iloc) +
1107 ocl_event_elapsed_ms(timers->pl_h2d_excl + iloc);
1112 /* add up energies and shift forces (only once at local F wait) */
1117 *e_lj += *nbst.e_lj;
1118 *e_el += *nbst.e_el;
1123 for (i = 0; i < SHIFTS; i++)
1125 fshift[i][0] += (nbst.fshift)[i][0];
1126 fshift[i][1] += (nbst.fshift)[i][1];
1127 fshift[i][2] += (nbst.fshift)[i][2];
1132 /* turn off pruning (doesn't matter if this is pair-search step or not) */
1133 plist->bDoPrune = false;
1137 /*! \brief Selects the Ewald kernel type, analytical or tabulated, single or twin cut-off. */
1138 int nbnxn_gpu_pick_ewald_kernel_type(bool bTwinCut)
1140 bool bUseAnalyticalEwald, bForceAnalyticalEwald, bForceTabulatedEwald;
1143 /* Benchmarking/development environment variables to force the use of
1144 analytical or tabulated Ewald kernel. */
1145 bForceAnalyticalEwald = (getenv("GMX_OCL_NB_ANA_EWALD") != NULL);
1146 bForceTabulatedEwald = (getenv("GMX_OCL_NB_TAB_EWALD") != NULL);
1148 if (bForceAnalyticalEwald && bForceTabulatedEwald)
1150 gmx_incons("Both analytical and tabulated Ewald OpenCL non-bonded kernels "
1151 "requested through environment variables.");
1154 /* OpenCL: By default, use analytical Ewald
1155 * TODO: tabulated does not work, it needs fixing, see init_nbparam() in nbnxn_ocl_data_mgmt.cpp
1157 * TODO: decide if dev_info parameter should be added to recognize NVIDIA CC>=3.0 devices.
1160 //if ((dev_info->prop.major >= 3 || bForceAnalyticalEwald) && !bForceTabulatedEwald)
1161 if ((1 || bForceAnalyticalEwald) && !bForceTabulatedEwald)
1163 bUseAnalyticalEwald = true;
1167 fprintf(debug, "Using analytical Ewald OpenCL kernels\n");
1172 bUseAnalyticalEwald = false;
1176 fprintf(debug, "Using tabulated Ewald OpenCL kernels\n");
1180 /* Use twin cut-off kernels if requested by bTwinCut or the env. var.
1181 forces it (use it for debugging/benchmarking only). */
1182 if (!bTwinCut && (getenv("GMX_OCL_NB_EWALD_TWINCUT") == NULL))
1184 kernel_type = bUseAnalyticalEwald ? eelOclEWALD_ANA : eelOclEWALD_TAB;
1188 kernel_type = bUseAnalyticalEwald ? eelOclEWALD_ANA_TWIN : eelOclEWALD_TAB_TWIN;