2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016 by the GROMACS development team.
5 * Copyright (c) 2017,2018,2019,2020, by the GROMACS development team, led by
6 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
7 * and including many others, as listed in the AUTHORS file in the
8 * top-level source directory and at http://www.gromacs.org.
10 * GROMACS is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public License
12 * as published by the Free Software Foundation; either version 2.1
13 * of the License, or (at your option) any later version.
15 * GROMACS is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with GROMACS; if not, see
22 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
23 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
25 * If you want to redistribute modifications to GROMACS, please
26 * consider that scientific software is very special. Version
27 * control is crucial - bugs must be traceable. We will be happy to
28 * consider code for inclusion in the official distribution, but
29 * derived work must not be called official GROMACS. Details are found
30 * in the README & COPYING files - if they are missing, get the
31 * official version at http://www.gromacs.org.
33 * To help us fund GROMACS development, we humbly ask that you cite
34 * the research papers on the package. Check out http://www.gromacs.org.
37 * \brief Define OpenCL implementation of nbnxm_gpu_data_mgmt.h
39 * \author Anca Hamuraru <anca@streamcomputing.eu>
40 * \author Dimitrios Karkoulis <dimitris.karkoulis@gmail.com>
41 * \author Teemu Virolainen <teemu@streamcomputing.eu>
42 * \author Szilárd Páll <pall.szilard@gmail.com>
43 * \ingroup module_nbnxm
55 #include "gromacs/gpu_utils/device_stream_manager.h"
56 #include "gromacs/gpu_utils/oclutils.h"
57 #include "gromacs/hardware/device_information.h"
58 #include "gromacs/hardware/device_management.h"
59 #include "gromacs/math/vectypes.h"
60 #include "gromacs/mdlib/force_flags.h"
61 #include "gromacs/mdtypes/interaction_const.h"
62 #include "gromacs/mdtypes/md_enums.h"
63 #include "gromacs/nbnxm/atomdata.h"
64 #include "gromacs/nbnxm/gpu_data_mgmt.h"
65 #include "gromacs/nbnxm/gpu_jit_support.h"
66 #include "gromacs/nbnxm/nbnxm.h"
67 #include "gromacs/nbnxm/nbnxm_gpu.h"
68 #include "gromacs/nbnxm/nbnxm_gpu_data_mgmt.h"
69 #include "gromacs/nbnxm/pairlistsets.h"
70 #include "gromacs/pbcutil/ishift.h"
71 #include "gromacs/timing/gpu_timing.h"
72 #include "gromacs/utility/cstringutil.h"
73 #include "gromacs/utility/fatalerror.h"
74 #include "gromacs/utility/gmxassert.h"
75 #include "gromacs/utility/real.h"
76 #include "gromacs/utility/smalloc.h"
78 #include "nbnxm_ocl_types.h"
83 /*! \brief Copies of values from cl_driver_diagnostics_intel.h,
84 * which isn't guaranteed to be available. */
86 #define CL_CONTEXT_SHOW_DIAGNOSTICS_INTEL 0x4106
87 #define CL_CONTEXT_DIAGNOSTICS_LEVEL_GOOD_INTEL 0x1
88 #define CL_CONTEXT_DIAGNOSTICS_LEVEL_BAD_INTEL 0x2
89 #define CL_CONTEXT_DIAGNOSTICS_LEVEL_NEUTRAL_INTEL 0x4
92 /*! \brief This parameter should be determined heuristically from the
93 * kernel execution times
95 * This value is best for small systems on a single AMD Radeon R9 290X
96 * (and about 5% faster than 40, which is the default for CUDA
97 * devices). Larger simulation systems were quite insensitive to the
98 * value of this parameter.
100 static unsigned int gpu_min_ci_balanced_factor = 50;
103 /*! \brief Initializes the atomdata structure first time, it only gets filled at
106 static void init_atomdata_first(cl_atomdata_t* ad, int ntypes, const DeviceContext& deviceContext)
110 allocateDeviceBuffer(&ad->shift_vec, SHIFTS * DIM, deviceContext);
111 ad->bShiftVecUploaded = CL_FALSE;
113 allocateDeviceBuffer(&ad->fshift, SHIFTS * DIM, deviceContext);
114 allocateDeviceBuffer(&ad->e_lj, 1, deviceContext);
115 allocateDeviceBuffer(&ad->e_el, 1, deviceContext);
117 /* initialize to nullptr pointers to data that is not allocated here and will
118 need reallocation in nbnxn_gpu_init_atomdata */
122 /* size -1 indicates that the respective array hasn't been initialized yet */
128 /*! \brief Initializes the nonbonded parameter data structure.
130 static void init_nbparam(NBParamGpu* nbp,
131 const interaction_const_t* ic,
132 const PairlistParams& listParams,
133 const nbnxn_atomdata_t::Params& nbatParams,
134 const DeviceContext& deviceContext)
136 set_cutoff_parameters(nbp, ic, listParams);
138 nbp->vdwType = nbnxmGpuPickVdwKernelType(ic, nbatParams.comb_rule);
139 nbp->elecType = nbnxmGpuPickElectrostaticsKernelType(ic);
141 if (ic->vdwtype == evdwPME)
143 if (ic->ljpme_comb_rule == ljcrGEOM)
145 GMX_ASSERT(nbatParams.comb_rule == ljcrGEOM, "Combination rule mismatch!");
149 GMX_ASSERT(nbatParams.comb_rule == ljcrLB, "Combination rule mismatch!");
152 /* generate table for PME */
153 nbp->coulomb_tab = nullptr;
154 if (nbp->elecType == ElecType::EwaldTab || nbp->elecType == ElecType::EwaldTabTwin)
156 GMX_RELEASE_ASSERT(ic->coulombEwaldTables, "Need valid Coulomb Ewald correction tables");
157 init_ewald_coulomb_force_table(*ic->coulombEwaldTables, nbp, deviceContext);
161 allocateDeviceBuffer(&nbp->coulomb_tab, 1, deviceContext);
164 const int nnbfp = 2 * nbatParams.numTypes * nbatParams.numTypes;
165 const int nnbfp_comb = 2 * nbatParams.numTypes;
168 /* set up LJ parameter lookup table */
169 DeviceBuffer<real> nbfp;
170 initParamLookupTable(&nbfp, nullptr, nbatParams.nbfp.data(), nnbfp, deviceContext);
173 if (ic->vdwtype == evdwPME)
175 DeviceBuffer<float> nbfp_comb;
176 initParamLookupTable(&nbfp_comb, nullptr, nbatParams.nbfp_comb.data(), nnbfp_comb, deviceContext);
177 nbp->nbfp_comb = nbfp_comb;
182 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
183 static cl_kernel nbnxn_gpu_create_kernel(NbnxmGpu* nb, const char* kernel_name)
188 kernel = clCreateKernel(nb->dev_rundata->program, kernel_name, &cl_error);
189 if (CL_SUCCESS != cl_error)
191 gmx_fatal(FARGS, "Failed to create kernel '%s' for GPU #%s: OpenCL error %d", kernel_name,
192 nb->deviceContext_->deviceInfo().device_name, cl_error);
198 /*! \brief Clears nonbonded shift force output array and energy outputs on the GPU.
200 static void nbnxn_ocl_clear_e_fshift(NbnxmGpu* nb)
204 cl_atomdata_t* adat = nb->atdat;
205 cl_command_queue ls = nb->deviceStreams[InteractionLocality::Local]->stream();
207 size_t local_work_size[3] = { 1, 1, 1 };
208 size_t global_work_size[3] = { 1, 1, 1 };
210 cl_int shifts = SHIFTS * 3;
214 cl_kernel zero_e_fshift = nb->kernel_zero_e_fshift;
216 local_work_size[0] = 64;
217 // Round the total number of threads up from the array size
218 global_work_size[0] = ((shifts + local_work_size[0] - 1) / local_work_size[0]) * local_work_size[0];
221 cl_error = clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->fshift));
222 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_lj));
223 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_el));
224 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_uint), &shifts);
225 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
227 cl_error = clEnqueueNDRangeKernel(ls, zero_e_fshift, 3, nullptr, global_work_size,
228 local_work_size, 0, nullptr, nullptr);
229 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
232 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
233 static void nbnxn_gpu_init_kernels(NbnxmGpu* nb)
235 /* Init to 0 main kernel arrays */
236 /* They will be later on initialized in select_nbnxn_kernel */
237 // TODO: consider always creating all variants of the kernels here so that there is no
238 // need for late call to clCreateKernel -- if that gives any advantage?
239 memset(nb->kernel_ener_noprune_ptr, 0, sizeof(nb->kernel_ener_noprune_ptr));
240 memset(nb->kernel_ener_prune_ptr, 0, sizeof(nb->kernel_ener_prune_ptr));
241 memset(nb->kernel_noener_noprune_ptr, 0, sizeof(nb->kernel_noener_noprune_ptr));
242 memset(nb->kernel_noener_prune_ptr, 0, sizeof(nb->kernel_noener_prune_ptr));
244 /* Init pruning kernels
246 * TODO: we could avoid creating kernels if dynamic pruning is turned off,
247 * but ATM that depends on force flags not passed into the initialization.
249 nb->kernel_pruneonly[epruneFirst] = nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_opencl");
250 nb->kernel_pruneonly[epruneRolling] =
251 nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_rolling_opencl");
253 /* Init auxiliary kernels */
254 nb->kernel_zero_e_fshift = nbnxn_gpu_create_kernel(nb, "zero_e_fshift");
257 /*! \brief Initializes simulation constant data.
259 * Initializes members of the atomdata and nbparam structs and
260 * clears e/fshift output buffers.
262 static void nbnxn_ocl_init_const(cl_atomdata_t* atomData,
263 NBParamGpu* nbParams,
264 const interaction_const_t* ic,
265 const PairlistParams& listParams,
266 const nbnxn_atomdata_t::Params& nbatParams,
267 const DeviceContext& deviceContext)
269 init_atomdata_first(atomData, nbatParams.numTypes, deviceContext);
270 init_nbparam(nbParams, ic, listParams, nbatParams, deviceContext);
274 //! This function is documented in the header file
275 NbnxmGpu* gpu_init(const gmx::DeviceStreamManager& deviceStreamManager,
276 const interaction_const_t* ic,
277 const PairlistParams& listParams,
278 const nbnxn_atomdata_t* nbat,
279 const bool bLocalAndNonlocal)
281 GMX_ASSERT(ic, "Need a valid interaction constants object");
283 auto nb = new NbnxmGpu();
284 nb->deviceContext_ = &deviceStreamManager.context();
286 snew(nb->nbparam, 1);
287 snew(nb->plist[InteractionLocality::Local], 1);
288 if (bLocalAndNonlocal)
290 snew(nb->plist[InteractionLocality::NonLocal], 1);
293 nb->bUseTwoStreams = bLocalAndNonlocal;
295 nb->timers = new cl_timers_t();
296 snew(nb->timings, 1);
298 /* set device info, just point it to the right GPU among the detected ones */
299 nb->dev_rundata = new gmx_device_runtime_data_t();
302 pmalloc(reinterpret_cast<void**>(&nb->nbst.e_lj), sizeof(*nb->nbst.e_lj));
303 pmalloc(reinterpret_cast<void**>(&nb->nbst.e_el), sizeof(*nb->nbst.e_el));
304 pmalloc(reinterpret_cast<void**>(&nb->nbst.fshift), SHIFTS * sizeof(*nb->nbst.fshift));
306 init_plist(nb->plist[InteractionLocality::Local]);
308 /* OpenCL timing disabled if GMX_DISABLE_GPU_TIMING is defined. */
309 nb->bDoTime = (getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
311 /* local/non-local GPU streams */
312 GMX_RELEASE_ASSERT(deviceStreamManager.streamIsValid(gmx::DeviceStreamType::NonBondedLocal),
313 "Local non-bonded stream should be initialized to use GPU for non-bonded.");
314 nb->deviceStreams[InteractionLocality::Local] =
315 &deviceStreamManager.stream(gmx::DeviceStreamType::NonBondedLocal);
317 if (nb->bUseTwoStreams)
319 init_plist(nb->plist[InteractionLocality::NonLocal]);
321 GMX_RELEASE_ASSERT(deviceStreamManager.streamIsValid(gmx::DeviceStreamType::NonBondedNonLocal),
322 "Non-local non-bonded stream should be initialized to use GPU for "
323 "non-bonded with domain decomposition.");
324 nb->deviceStreams[InteractionLocality::NonLocal] =
325 &deviceStreamManager.stream(gmx::DeviceStreamType::NonBondedNonLocal);
330 init_timings(nb->timings);
333 nbnxn_ocl_init_const(nb->atdat, nb->nbparam, ic, listParams, nbat->params(), *nb->deviceContext_);
335 /* Enable LJ param manual prefetch for AMD or Intel or if we request through env. var.
336 * TODO: decide about NVIDIA
338 nb->bPrefetchLjParam = (getenv("GMX_OCL_DISABLE_I_PREFETCH") == nullptr)
339 && ((nb->deviceContext_->deviceInfo().deviceVendor == DeviceVendor::Amd)
340 || (nb->deviceContext_->deviceInfo().deviceVendor == DeviceVendor::Intel)
341 || (getenv("GMX_OCL_ENABLE_I_PREFETCH") != nullptr));
343 /* NOTE: in CUDA we pick L1 cache configuration for the nbnxn kernels here,
344 * but sadly this is not supported in OpenCL (yet?). Consider adding it if
345 * it becomes supported.
347 nbnxn_gpu_compile_kernels(nb);
348 nbnxn_gpu_init_kernels(nb);
350 /* clear energy and shift force outputs */
351 nbnxn_ocl_clear_e_fshift(nb);
355 fprintf(debug, "Initialized OpenCL data structures.\n");
361 /*! \brief Clears the first natoms_clear elements of the GPU nonbonded force output array.
363 static void nbnxn_ocl_clear_f(NbnxmGpu* nb, int natoms_clear)
365 if (natoms_clear == 0)
370 cl_atomdata_t* atomData = nb->atdat;
371 const DeviceStream& localStream = *nb->deviceStreams[InteractionLocality::Local];
373 clearDeviceBufferAsync(&atomData->f, 0, natoms_clear * DIM, localStream);
376 //! This function is documented in the header file
377 void gpu_clear_outputs(NbnxmGpu* nb, bool computeVirial)
379 nbnxn_ocl_clear_f(nb, nb->atdat->natoms);
380 /* clear shift force array and energies if the outputs were
381 used in the current step */
384 nbnxn_ocl_clear_e_fshift(nb);
387 /* kick off buffer clearing kernel to ensure concurrency with constraints/update */
388 cl_int gmx_unused cl_error;
389 cl_error = clFlush(nb->deviceStreams[InteractionLocality::Local]->stream());
390 GMX_ASSERT(cl_error == CL_SUCCESS, ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
393 //! This function is documented in the header file
394 void gpu_upload_shiftvec(NbnxmGpu* nb, const nbnxn_atomdata_t* nbatom)
396 cl_atomdata_t* adat = nb->atdat;
397 const DeviceStream& deviceStream = *nb->deviceStreams[InteractionLocality::Local];
399 /* only if we have a dynamic box */
400 if (nbatom->bDynamicBox || !adat->bShiftVecUploaded)
402 GMX_ASSERT(sizeof(float) * DIM == sizeof(*nbatom->shift_vec.data()),
403 "Sizes of host- and device-side shift vectors should be the same.");
404 copyToDeviceBuffer(&adat->shift_vec, reinterpret_cast<const float*>(nbatom->shift_vec.data()),
405 0, SHIFTS * DIM, deviceStream, GpuApiCallBehavior::Async, nullptr);
406 adat->bShiftVecUploaded = CL_TRUE;
410 //! This function is documented in the header file
411 void gpu_init_atomdata(NbnxmGpu* nb, const nbnxn_atomdata_t* nbat)
416 bool bDoTime = nb->bDoTime;
417 cl_timers_t* timers = nb->timers;
418 cl_atomdata_t* d_atdat = nb->atdat;
419 const DeviceContext& deviceContext = *nb->deviceContext_;
420 const DeviceStream& deviceStream = *nb->deviceStreams[InteractionLocality::Local];
422 natoms = nbat->numAtoms();
427 /* time async copy */
428 timers->atdat.openTimingRegion(deviceStream);
431 /* need to reallocate if we have to copy more atoms than the amount of space
432 available and only allocate if we haven't initialized yet, i.e d_atdat->natoms == -1 */
433 if (natoms > d_atdat->nalloc)
435 nalloc = over_alloc_small(natoms);
437 /* free up first if the arrays have already been initialized */
438 if (d_atdat->nalloc != -1)
440 freeDeviceBuffer(&d_atdat->f);
441 freeDeviceBuffer(&d_atdat->xq);
442 freeDeviceBuffer(&d_atdat->lj_comb);
443 freeDeviceBuffer(&d_atdat->atom_types);
447 allocateDeviceBuffer(&d_atdat->f, nalloc * DIM, deviceContext);
448 allocateDeviceBuffer(&d_atdat->xq, nalloc * (DIM + 1), deviceContext);
450 if (useLjCombRule(nb->nbparam->vdwType))
452 // Two Lennard-Jones parameters per atom
453 allocateDeviceBuffer(&d_atdat->lj_comb, nalloc * 2, deviceContext);
457 allocateDeviceBuffer(&d_atdat->atom_types, nalloc, deviceContext);
460 d_atdat->nalloc = nalloc;
464 d_atdat->natoms = natoms;
465 d_atdat->natoms_local = nbat->natoms_local;
467 /* need to clear GPU f output if realloc happened */
470 nbnxn_ocl_clear_f(nb, nalloc);
473 if (useLjCombRule(nb->nbparam->vdwType))
475 GMX_ASSERT(sizeof(float) == sizeof(*nbat->params().lj_comb.data()),
476 "Size of the LJ parameters element should be equal to the size of float2.");
477 copyToDeviceBuffer(&d_atdat->lj_comb, nbat->params().lj_comb.data(), 0, 2 * natoms,
478 deviceStream, GpuApiCallBehavior::Async,
479 bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
483 GMX_ASSERT(sizeof(int) == sizeof(*nbat->params().type.data()),
484 "Sizes of host- and device-side atom types should be the same.");
485 copyToDeviceBuffer(&d_atdat->atom_types, nbat->params().type.data(), 0, natoms, deviceStream,
486 GpuApiCallBehavior::Async, bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
491 timers->atdat.closeTimingRegion(deviceStream);
494 /* kick off the tasks enqueued above to ensure concurrency with the search */
495 cl_error = clFlush(deviceStream.stream());
496 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
497 ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
500 /*! \brief Releases an OpenCL kernel pointer */
501 static void free_kernel(cl_kernel* kernel_ptr)
503 cl_int gmx_unused cl_error;
505 GMX_ASSERT(kernel_ptr, "Need a valid kernel pointer");
509 cl_error = clReleaseKernel(*kernel_ptr);
510 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
511 ("clReleaseKernel failed: " + ocl_get_error_string(cl_error)).c_str());
513 *kernel_ptr = nullptr;
517 /*! \brief Releases a list of OpenCL kernel pointers */
518 static void free_kernels(cl_kernel* kernels, int count)
522 for (i = 0; i < count; i++)
524 free_kernel(kernels + i);
528 /*! \brief Free the OpenCL program.
530 * The function releases the OpenCL program assuciated with the
531 * device that the calling PP rank is running on.
533 * \param program [in] OpenCL program to release.
535 static void freeGpuProgram(cl_program program)
539 cl_int cl_error = clReleaseProgram(program);
540 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
541 ("clReleaseProgram failed: " + ocl_get_error_string(cl_error)).c_str());
546 //! This function is documented in the header file
547 void gpu_free(NbnxmGpu* nb)
555 // NOLINTNEXTLINE(bugprone-sizeof-expression)
556 int kernel_count = sizeof(nb->kernel_ener_noprune_ptr) / sizeof(nb->kernel_ener_noprune_ptr[0][0]);
557 free_kernels(nb->kernel_ener_noprune_ptr[0], kernel_count);
559 // NOLINTNEXTLINE(bugprone-sizeof-expression)
560 kernel_count = sizeof(nb->kernel_ener_prune_ptr) / sizeof(nb->kernel_ener_prune_ptr[0][0]);
561 free_kernels(nb->kernel_ener_prune_ptr[0], kernel_count);
563 // NOLINTNEXTLINE(bugprone-sizeof-expression)
564 kernel_count = sizeof(nb->kernel_noener_noprune_ptr) / sizeof(nb->kernel_noener_noprune_ptr[0][0]);
565 free_kernels(nb->kernel_noener_noprune_ptr[0], kernel_count);
567 // NOLINTNEXTLINE(bugprone-sizeof-expression)
568 kernel_count = sizeof(nb->kernel_noener_prune_ptr) / sizeof(nb->kernel_noener_prune_ptr[0][0]);
569 free_kernels(nb->kernel_noener_prune_ptr[0], kernel_count);
571 free_kernel(&(nb->kernel_zero_e_fshift));
574 freeDeviceBuffer(&(nb->atdat->xq));
575 freeDeviceBuffer(&(nb->atdat->f));
576 freeDeviceBuffer(&(nb->atdat->e_lj));
577 freeDeviceBuffer(&(nb->atdat->e_el));
578 freeDeviceBuffer(&(nb->atdat->fshift));
579 freeDeviceBuffer(&(nb->atdat->lj_comb));
580 freeDeviceBuffer(&(nb->atdat->atom_types));
581 freeDeviceBuffer(&(nb->atdat->shift_vec));
585 freeDeviceBuffer(&(nb->nbparam->nbfp));
586 freeDeviceBuffer(&(nb->nbparam->nbfp_comb));
587 freeDeviceBuffer(&(nb->nbparam->coulomb_tab));
591 auto* plist = nb->plist[InteractionLocality::Local];
592 freeDeviceBuffer(&plist->sci);
593 freeDeviceBuffer(&plist->cj4);
594 freeDeviceBuffer(&plist->imask);
595 freeDeviceBuffer(&plist->excl);
597 if (nb->bUseTwoStreams)
599 auto* plist_nl = nb->plist[InteractionLocality::NonLocal];
600 freeDeviceBuffer(&plist_nl->sci);
601 freeDeviceBuffer(&plist_nl->cj4);
602 freeDeviceBuffer(&plist_nl->imask);
603 freeDeviceBuffer(&plist_nl->excl);
608 pfree(nb->nbst.e_lj);
609 nb->nbst.e_lj = nullptr;
611 pfree(nb->nbst.e_el);
612 nb->nbst.e_el = nullptr;
614 pfree(nb->nbst.fshift);
615 nb->nbst.fshift = nullptr;
617 /* Free other events */
618 if (nb->nonlocal_done)
620 clReleaseEvent(nb->nonlocal_done);
621 nb->nonlocal_done = nullptr;
623 if (nb->misc_ops_and_local_H2D_done)
625 clReleaseEvent(nb->misc_ops_and_local_H2D_done);
626 nb->misc_ops_and_local_H2D_done = nullptr;
629 freeGpuProgram(nb->dev_rundata->program);
630 delete nb->dev_rundata;
632 /* Free timers and timings */
639 fprintf(debug, "Cleaned up OpenCL data structures.\n");
643 //! This function is documented in the header file
644 int gpu_min_ci_balanced(NbnxmGpu* nb)
646 return nb != nullptr ? gpu_min_ci_balanced_factor * nb->deviceContext_->deviceInfo().compute_units : 0;