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) 2017,2018,2019,2020,2021, 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(NBAtomData* ad, int ntypes, const DeviceContext& deviceContext)
108 ad->numTypes = ntypes;
110 allocateDeviceBuffer(&ad->shiftVec, SHIFTS * DIM, deviceContext);
111 ad->shiftVecUploaded = false;
113 allocateDeviceBuffer(&ad->fShift, SHIFTS * DIM, deviceContext);
114 allocateDeviceBuffer(&ad->eLJ, 1, deviceContext);
115 allocateDeviceBuffer(&ad->eElec, 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 */
124 ad->numAtomsAlloc = -1;
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.ljCombinationRule);
139 nbp->elecType = nbnxmGpuPickElectrostaticsKernelType(ic, deviceContext.deviceInfo());
141 if (ic->vdwtype == VanDerWaalsType::Pme)
143 if (ic->ljpme_comb_rule == LongRangeVdW::Geom)
145 GMX_ASSERT(nbatParams.ljCombinationRule == LJCombinationRule::Geometric,
146 "Combination rule mismatch!");
150 GMX_ASSERT(nbatParams.ljCombinationRule == LJCombinationRule::LorentzBerthelot,
151 "Combination rule mismatch!");
154 /* generate table for PME */
155 nbp->coulomb_tab = nullptr;
156 if (nbp->elecType == ElecType::EwaldTab || nbp->elecType == ElecType::EwaldTabTwin)
158 GMX_RELEASE_ASSERT(ic->coulombEwaldTables, "Need valid Coulomb Ewald correction tables");
159 init_ewald_coulomb_force_table(*ic->coulombEwaldTables, nbp, deviceContext);
163 allocateDeviceBuffer(&nbp->coulomb_tab, 1, deviceContext);
166 const int nnbfp = 2 * nbatParams.numTypes * nbatParams.numTypes;
167 const int nnbfp_comb = 2 * nbatParams.numTypes;
170 /* set up LJ parameter lookup table */
171 DeviceBuffer<real> nbfp;
172 initParamLookupTable(&nbfp, nullptr, nbatParams.nbfp.data(), nnbfp, deviceContext);
175 if (ic->vdwtype == VanDerWaalsType::Pme)
177 DeviceBuffer<float> nbfp_comb;
178 initParamLookupTable(&nbfp_comb, nullptr, nbatParams.nbfp_comb.data(), nnbfp_comb, deviceContext);
179 nbp->nbfp_comb = nbfp_comb;
184 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
185 static cl_kernel nbnxn_gpu_create_kernel(NbnxmGpu* nb, const char* kernel_name)
190 kernel = clCreateKernel(nb->dev_rundata->program, kernel_name, &cl_error);
191 if (CL_SUCCESS != cl_error)
194 "Failed to create kernel '%s' for GPU #%s: OpenCL error %d",
196 nb->deviceContext_->deviceInfo().device_name,
203 /*! \brief Clears nonbonded shift force output array and energy outputs on the GPU.
205 static void nbnxn_ocl_clear_e_fshift(NbnxmGpu* nb)
209 NBAtomData* adat = nb->atdat;
210 cl_command_queue ls = nb->deviceStreams[InteractionLocality::Local]->stream();
212 size_t local_work_size[3] = { 1, 1, 1 };
213 size_t global_work_size[3] = { 1, 1, 1 };
215 cl_int shifts = SHIFTS * 3;
219 cl_kernel zero_e_fshift = nb->kernel_zero_e_fshift;
221 local_work_size[0] = 64;
222 // Round the total number of threads up from the array size
223 global_work_size[0] = ((shifts + local_work_size[0] - 1) / local_work_size[0]) * local_work_size[0];
226 cl_error = clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->fShift));
227 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->eLJ));
228 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->eElec));
229 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_uint), &shifts);
230 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
232 cl_error = clEnqueueNDRangeKernel(
233 ls, zero_e_fshift, 3, nullptr, global_work_size, local_work_size, 0, nullptr, nullptr);
234 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
237 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
238 static void nbnxn_gpu_init_kernels(NbnxmGpu* nb)
240 /* Init to 0 main kernel arrays */
241 /* They will be later on initialized in select_nbnxn_kernel */
242 // TODO: consider always creating all variants of the kernels here so that there is no
243 // need for late call to clCreateKernel -- if that gives any advantage?
244 memset(nb->kernel_ener_noprune_ptr, 0, sizeof(nb->kernel_ener_noprune_ptr));
245 memset(nb->kernel_ener_prune_ptr, 0, sizeof(nb->kernel_ener_prune_ptr));
246 memset(nb->kernel_noener_noprune_ptr, 0, sizeof(nb->kernel_noener_noprune_ptr));
247 memset(nb->kernel_noener_prune_ptr, 0, sizeof(nb->kernel_noener_prune_ptr));
249 /* Init pruning kernels
251 * TODO: we could avoid creating kernels if dynamic pruning is turned off,
252 * but ATM that depends on force flags not passed into the initialization.
254 nb->kernel_pruneonly[epruneFirst] = nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_opencl");
255 nb->kernel_pruneonly[epruneRolling] =
256 nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_rolling_opencl");
258 /* Init auxiliary kernels */
259 nb->kernel_zero_e_fshift = nbnxn_gpu_create_kernel(nb, "zero_e_fshift");
262 /*! \brief Initializes simulation constant data.
264 * Initializes members of the atomdata and nbparam structs and
265 * clears e/fshift output buffers.
267 static void nbnxn_ocl_init_const(NBAtomData* atomData,
268 NBParamGpu* nbParams,
269 const interaction_const_t* ic,
270 const PairlistParams& listParams,
271 const nbnxn_atomdata_t::Params& nbatParams,
272 const DeviceContext& deviceContext)
274 init_atomdata_first(atomData, nbatParams.numTypes, deviceContext);
275 init_nbparam(nbParams, ic, listParams, nbatParams, deviceContext);
279 //! This function is documented in the header file
280 NbnxmGpu* gpu_init(const gmx::DeviceStreamManager& deviceStreamManager,
281 const interaction_const_t* ic,
282 const PairlistParams& listParams,
283 const nbnxn_atomdata_t* nbat,
284 const bool bLocalAndNonlocal)
286 GMX_ASSERT(ic, "Need a valid interaction constants object");
288 auto nb = new NbnxmGpu();
289 nb->deviceContext_ = &deviceStreamManager.context();
291 snew(nb->nbparam, 1);
292 snew(nb->plist[InteractionLocality::Local], 1);
293 if (bLocalAndNonlocal)
295 snew(nb->plist[InteractionLocality::NonLocal], 1);
298 nb->bUseTwoStreams = bLocalAndNonlocal;
300 nb->timers = new cl_timers_t();
301 snew(nb->timings, 1);
303 /* set device info, just point it to the right GPU among the detected ones */
304 nb->dev_rundata = new gmx_device_runtime_data_t();
307 pmalloc(reinterpret_cast<void**>(&nb->nbst.eLJ), sizeof(*nb->nbst.eLJ));
308 pmalloc(reinterpret_cast<void**>(&nb->nbst.eElec), sizeof(*nb->nbst.eElec));
309 pmalloc(reinterpret_cast<void**>(&nb->nbst.fShift), SHIFTS * sizeof(*nb->nbst.fShift));
311 init_plist(nb->plist[InteractionLocality::Local]);
313 /* OpenCL timing disabled if GMX_DISABLE_GPU_TIMING is defined. */
314 nb->bDoTime = (getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
316 /* local/non-local GPU streams */
317 GMX_RELEASE_ASSERT(deviceStreamManager.streamIsValid(gmx::DeviceStreamType::NonBondedLocal),
318 "Local non-bonded stream should be initialized to use GPU for non-bonded.");
319 nb->deviceStreams[InteractionLocality::Local] =
320 &deviceStreamManager.stream(gmx::DeviceStreamType::NonBondedLocal);
322 if (nb->bUseTwoStreams)
324 init_plist(nb->plist[InteractionLocality::NonLocal]);
326 GMX_RELEASE_ASSERT(deviceStreamManager.streamIsValid(gmx::DeviceStreamType::NonBondedNonLocal),
327 "Non-local non-bonded stream should be initialized to use GPU for "
328 "non-bonded with domain decomposition.");
329 nb->deviceStreams[InteractionLocality::NonLocal] =
330 &deviceStreamManager.stream(gmx::DeviceStreamType::NonBondedNonLocal);
335 init_timings(nb->timings);
338 nbnxn_ocl_init_const(nb->atdat, nb->nbparam, ic, listParams, nbat->params(), *nb->deviceContext_);
340 /* Enable LJ param manual prefetch for AMD or Intel or if we request through env. var.
341 * TODO: decide about NVIDIA
343 nb->bPrefetchLjParam = (getenv("GMX_OCL_DISABLE_I_PREFETCH") == nullptr)
344 && ((nb->deviceContext_->deviceInfo().deviceVendor == DeviceVendor::Amd)
345 || (nb->deviceContext_->deviceInfo().deviceVendor == DeviceVendor::Intel)
346 || (getenv("GMX_OCL_ENABLE_I_PREFETCH") != nullptr));
348 /* NOTE: in CUDA we pick L1 cache configuration for the nbnxn kernels here,
349 * but sadly this is not supported in OpenCL (yet?). Consider adding it if
350 * it becomes supported.
352 nbnxn_gpu_compile_kernels(nb);
353 nbnxn_gpu_init_kernels(nb);
355 /* clear energy and shift force outputs */
356 nbnxn_ocl_clear_e_fshift(nb);
360 fprintf(debug, "Initialized OpenCL data structures.\n");
366 /*! \brief Clears the first natoms_clear elements of the GPU nonbonded force output array.
368 static void nbnxn_ocl_clear_f(NbnxmGpu* nb, int natoms_clear)
370 if (natoms_clear == 0)
375 NBAtomData* atomData = nb->atdat;
376 const DeviceStream& localStream = *nb->deviceStreams[InteractionLocality::Local];
378 clearDeviceBufferAsync(&atomData->f, 0, natoms_clear, localStream);
381 //! This function is documented in the header file
382 void gpu_clear_outputs(NbnxmGpu* nb, bool computeVirial)
384 nbnxn_ocl_clear_f(nb, nb->atdat->numAtoms);
385 /* clear shift force array and energies if the outputs were
386 used in the current step */
389 nbnxn_ocl_clear_e_fshift(nb);
392 /* kick off buffer clearing kernel to ensure concurrency with constraints/update */
393 cl_int gmx_unused cl_error;
394 cl_error = clFlush(nb->deviceStreams[InteractionLocality::Local]->stream());
395 GMX_ASSERT(cl_error == CL_SUCCESS, ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
398 //! This function is documented in the header file
399 void gpu_upload_shiftvec(NbnxmGpu* nb, const nbnxn_atomdata_t* nbatom)
401 NBAtomData* adat = nb->atdat;
402 const DeviceStream& localStream = *nb->deviceStreams[InteractionLocality::Local];
404 /* only if we have a dynamic box */
405 if (nbatom->bDynamicBox || !adat->shiftVecUploaded)
407 static_assert(sizeof(Float3) == sizeof(nbatom->shift_vec[0]),
408 "Sizes of host- and device-side shift vectors should be the same.");
409 copyToDeviceBuffer(&adat->shiftVec,
410 reinterpret_cast<const Float3*>(nbatom->shift_vec.data()),
414 GpuApiCallBehavior::Async,
416 adat->shiftVecUploaded = true;
420 //! This function is documented in the header file
421 void gpu_init_atomdata(NbnxmGpu* nb, const nbnxn_atomdata_t* nbat)
426 bool bDoTime = nb->bDoTime;
427 cl_timers_t* timers = nb->timers;
428 NBAtomData* d_atdat = nb->atdat;
429 const DeviceContext& deviceContext = *nb->deviceContext_;
430 const DeviceStream& localStream = *nb->deviceStreams[InteractionLocality::Local];
432 natoms = nbat->numAtoms();
437 /* time async copy */
438 timers->atdat.openTimingRegion(localStream);
441 /* need to reallocate if we have to copy more atoms than the amount of space
442 available and only allocate if we haven't initialized yet, i.e d_atdat->natoms == -1 */
443 if (natoms > d_atdat->numAtomsAlloc)
445 nalloc = over_alloc_small(natoms);
447 /* free up first if the arrays have already been initialized */
448 if (d_atdat->numAtomsAlloc != -1)
450 freeDeviceBuffer(&d_atdat->f);
451 freeDeviceBuffer(&d_atdat->xq);
452 freeDeviceBuffer(&d_atdat->ljComb);
453 freeDeviceBuffer(&d_atdat->atomTypes);
457 allocateDeviceBuffer(&d_atdat->f, nalloc, deviceContext);
458 allocateDeviceBuffer(&d_atdat->xq, nalloc, deviceContext);
460 if (useLjCombRule(nb->nbparam->vdwType))
462 // Two Lennard-Jones parameters per atom
463 allocateDeviceBuffer(&d_atdat->ljComb, nalloc, deviceContext);
467 allocateDeviceBuffer(&d_atdat->atomTypes, nalloc, deviceContext);
470 d_atdat->numAtomsAlloc = nalloc;
474 d_atdat->numAtoms = natoms;
475 d_atdat->numAtomsLocal = nbat->natoms_local;
477 /* need to clear GPU f output if realloc happened */
480 nbnxn_ocl_clear_f(nb, nalloc);
483 if (useLjCombRule(nb->nbparam->vdwType))
485 static_assert(sizeof(float) == sizeof(*nbat->params().lj_comb.data()),
486 "Size of the LJ parameters element should be equal to the size of float2.");
487 copyToDeviceBuffer(&d_atdat->ljComb,
488 reinterpret_cast<const Float2*>(nbat->params().lj_comb.data()),
492 GpuApiCallBehavior::Async,
493 bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
497 static_assert(sizeof(int) == sizeof(*nbat->params().type.data()),
498 "Sizes of host- and device-side atom types should be the same.");
499 copyToDeviceBuffer(&d_atdat->atomTypes,
500 nbat->params().type.data(),
504 GpuApiCallBehavior::Async,
505 bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
510 timers->atdat.closeTimingRegion(localStream);
513 /* kick off the tasks enqueued above to ensure concurrency with the search */
514 cl_error = clFlush(localStream.stream());
515 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
516 ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
519 /*! \brief Releases an OpenCL kernel pointer */
520 static void free_kernel(cl_kernel* kernel_ptr)
522 cl_int gmx_unused cl_error;
524 GMX_ASSERT(kernel_ptr, "Need a valid kernel pointer");
528 cl_error = clReleaseKernel(*kernel_ptr);
529 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
530 ("clReleaseKernel failed: " + ocl_get_error_string(cl_error)).c_str());
532 *kernel_ptr = nullptr;
536 /*! \brief Releases a list of OpenCL kernel pointers */
537 static void free_kernels(cl_kernel* kernels, int count)
541 for (i = 0; i < count; i++)
543 free_kernel(kernels + i);
547 /*! \brief Free the OpenCL program.
549 * The function releases the OpenCL program assuciated with the
550 * device that the calling PP rank is running on.
552 * \param program [in] OpenCL program to release.
554 static void freeGpuProgram(cl_program program)
558 cl_int cl_error = clReleaseProgram(program);
559 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
560 ("clReleaseProgram failed: " + ocl_get_error_string(cl_error)).c_str());
565 //! This function is documented in the header file
566 void gpu_free(NbnxmGpu* nb)
574 // NOLINTNEXTLINE(bugprone-sizeof-expression)
575 int kernel_count = sizeof(nb->kernel_ener_noprune_ptr) / sizeof(nb->kernel_ener_noprune_ptr[0][0]);
576 free_kernels(nb->kernel_ener_noprune_ptr[0], kernel_count);
578 // NOLINTNEXTLINE(bugprone-sizeof-expression)
579 kernel_count = sizeof(nb->kernel_ener_prune_ptr) / sizeof(nb->kernel_ener_prune_ptr[0][0]);
580 free_kernels(nb->kernel_ener_prune_ptr[0], kernel_count);
582 // NOLINTNEXTLINE(bugprone-sizeof-expression)
583 kernel_count = sizeof(nb->kernel_noener_noprune_ptr) / sizeof(nb->kernel_noener_noprune_ptr[0][0]);
584 free_kernels(nb->kernel_noener_noprune_ptr[0], kernel_count);
586 // NOLINTNEXTLINE(bugprone-sizeof-expression)
587 kernel_count = sizeof(nb->kernel_noener_prune_ptr) / sizeof(nb->kernel_noener_prune_ptr[0][0]);
588 free_kernels(nb->kernel_noener_prune_ptr[0], kernel_count);
590 free_kernel(&(nb->kernel_zero_e_fshift));
593 freeDeviceBuffer(&(nb->atdat->xq));
594 freeDeviceBuffer(&(nb->atdat->f));
595 freeDeviceBuffer(&(nb->atdat->eLJ));
596 freeDeviceBuffer(&(nb->atdat->eElec));
597 freeDeviceBuffer(&(nb->atdat->fShift));
598 freeDeviceBuffer(&(nb->atdat->ljComb));
599 freeDeviceBuffer(&(nb->atdat->atomTypes));
600 freeDeviceBuffer(&(nb->atdat->shiftVec));
604 freeDeviceBuffer(&(nb->nbparam->nbfp));
605 freeDeviceBuffer(&(nb->nbparam->nbfp_comb));
606 freeDeviceBuffer(&(nb->nbparam->coulomb_tab));
610 auto* plist = nb->plist[InteractionLocality::Local];
611 freeDeviceBuffer(&plist->sci);
612 freeDeviceBuffer(&plist->cj4);
613 freeDeviceBuffer(&plist->imask);
614 freeDeviceBuffer(&plist->excl);
616 if (nb->bUseTwoStreams)
618 auto* plist_nl = nb->plist[InteractionLocality::NonLocal];
619 freeDeviceBuffer(&plist_nl->sci);
620 freeDeviceBuffer(&plist_nl->cj4);
621 freeDeviceBuffer(&plist_nl->imask);
622 freeDeviceBuffer(&plist_nl->excl);
628 nb->nbst.eLJ = nullptr;
630 pfree(nb->nbst.eElec);
631 nb->nbst.eElec = nullptr;
633 pfree(nb->nbst.fShift);
634 nb->nbst.fShift = nullptr;
636 freeGpuProgram(nb->dev_rundata->program);
637 delete nb->dev_rundata;
639 /* Free timers and timings */
646 fprintf(debug, "Cleaned up OpenCL data structures.\n");
650 //! This function is documented in the header file
651 int gpu_min_ci_balanced(NbnxmGpu* nb)
653 return nb != nullptr ? gpu_min_ci_balanced_factor * nb->deviceContext_->deviceInfo().compute_units : 0;