2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2019,2020,2021, 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 Common functions for the different NBNXN GPU implementations.
38 * \author Berk Hess <hess@kth.se>
40 * \ingroup module_nbnxm
47 #include "gromacs/domdec/domdec.h"
48 #include "gromacs/domdec/domdec_struct.h"
49 #include "gromacs/hardware/hw_info.h"
50 #include "gromacs/mdlib/gmx_omp_nthreads.h"
51 #include "gromacs/mdtypes/commrec.h"
52 #include "gromacs/mdtypes/enerdata.h"
53 #include "gromacs/mdtypes/forcerec.h"
54 #include "gromacs/mdtypes/inputrec.h"
55 #include "gromacs/mdtypes/interaction_const.h"
56 #include "gromacs/nbnxm/atomdata.h"
57 #include "gromacs/nbnxm/gpu_data_mgmt.h"
58 #include "gromacs/nbnxm/nbnxm.h"
59 #include "gromacs/nbnxm/pairlist_tuning.h"
60 #include "gromacs/simd/simd.h"
61 #include "gromacs/topology/mtop_util.h"
62 #include "gromacs/utility/fatalerror.h"
63 #include "gromacs/utility/logger.h"
66 #include "nbnxm_geometry.h"
67 #include "nbnxm_simd.h"
69 #include "pairlistset.h"
70 #include "pairlistsets.h"
71 #include "pairsearch.h"
76 /*! \brief Resources that can be used to execute non-bonded kernels on */
77 enum class NonbondedResource : int
84 /*! \brief Returns whether CPU SIMD support exists for the given inputrec
86 * If the return value is FALSE and fplog/cr != NULL, prints a fallback
87 * message to fplog/stderr.
89 static bool nbnxn_simd_supported(const gmx::MDLogger& mdlog, const t_inputrec& inputrec)
91 if (inputrec.vdwtype == VanDerWaalsType::Pme && inputrec.ljpme_combination_rule == LongRangeVdW::LB)
93 /* LJ PME with LB combination rule does 7 mesh operations.
94 * This so slow that we don't compile SIMD non-bonded kernels
96 GMX_LOG(mdlog.warning)
99 "LJ-PME with Lorentz-Berthelot is not supported with SIMD kernels, falling "
100 "back to plain C kernels");
107 /*! \brief Returns the most suitable CPU kernel type and Ewald handling */
108 static KernelSetup pick_nbnxn_kernel_cpu(const t_inputrec gmx_unused& inputrec,
109 const gmx_hw_info_t gmx_unused& hardwareInfo)
111 KernelSetup kernelSetup;
115 kernelSetup.kernelType = KernelType::Cpu4x4_PlainC;
116 kernelSetup.ewaldExclusionType = EwaldExclusionType::Table;
120 #ifdef GMX_NBNXN_SIMD_4XN
121 kernelSetup.kernelType = KernelType::Cpu4xN_Simd_4xN;
123 #ifdef GMX_NBNXN_SIMD_2XNN
124 kernelSetup.kernelType = KernelType::Cpu4xN_Simd_2xNN;
127 #if defined GMX_NBNXN_SIMD_2XNN && defined GMX_NBNXN_SIMD_4XN
128 /* We need to choose if we want 2x(N+N) or 4xN kernels.
129 * This is based on the SIMD acceleration choice and CPU information
130 * detected at runtime.
132 * 4xN calculates more (zero) interactions, but has less pair-search
133 * work and much better kernel instruction scheduling.
135 * Up till now we have only seen that on Intel Sandy/Ivy Bridge,
136 * which doesn't have FMA, both the analytical and tabulated Ewald
137 * kernels have similar pair rates for 4x8 and 2x(4+4), so we choose
138 * 2x(4+4) because it results in significantly fewer pairs.
139 * For RF, the raw pair rate of the 4x8 kernel is higher than 2x(4+4),
140 * 10% with HT, 50% without HT. As we currently don't detect the actual
141 * use of HT, use 4x8 to avoid a potential performance hit.
142 * On Intel Haswell 4x8 is always faster.
144 kernelSetup.kernelType = KernelType::Cpu4xN_Simd_4xN;
146 if (!GMX_SIMD_HAVE_FMA && (EEL_PME_EWALD(inputrec.coulombtype) || EVDW_PME(inputrec.vdwtype)))
148 /* We have Ewald kernels without FMA (Intel Sandy/Ivy Bridge).
149 * There are enough instructions to make 2x(4+4) efficient.
151 kernelSetup.kernelType = KernelType::Cpu4xN_Simd_2xNN;
154 if (hardwareInfo.haveAmdZen1Cpu)
156 /* One 256-bit FMA per cycle makes 2xNN faster */
157 kernelSetup.kernelType = KernelType::Cpu4xN_Simd_2xNN;
159 #endif /* GMX_NBNXN_SIMD_2XNN && GMX_NBNXN_SIMD_4XN */
162 if (getenv("GMX_NBNXN_SIMD_4XN") != nullptr)
164 #ifdef GMX_NBNXN_SIMD_4XN
165 kernelSetup.kernelType = KernelType::Cpu4xN_Simd_4xN;
168 "SIMD 4xN kernels requested, but GROMACS has been compiled without support "
169 "for these kernels");
172 if (getenv("GMX_NBNXN_SIMD_2XNN") != nullptr)
174 #ifdef GMX_NBNXN_SIMD_2XNN
175 kernelSetup.kernelType = KernelType::Cpu4xN_Simd_2xNN;
178 "SIMD 2x(N+N) kernels requested, but GROMACS has been compiled without "
179 "support for these kernels");
183 /* Analytical Ewald exclusion correction is only an option in
185 * Since table lookup's don't parallelize with SIMD, analytical
186 * will probably always be faster for a SIMD width of 8 or more.
187 * With FMA analytical is sometimes faster for a width if 4 as well.
188 * In single precision, this is faster on Bulldozer.
189 * On AMD Zen, tabulated Ewald kernels are faster on all 4 combinations
190 * of single or double precision and 128 or 256-bit AVX2.
192 MSVC_DIAGNOSTIC_IGNORE(6285) // Always zero because compile time constant
195 (GMX_SIMD_REAL_WIDTH >= 8 || (GMX_SIMD_REAL_WIDTH >= 4 && GMX_SIMD_HAVE_FMA && !GMX_DOUBLE)) &&
197 !hardwareInfo.haveAmdZen1Cpu)
199 kernelSetup.ewaldExclusionType = EwaldExclusionType::Analytical;
201 MSVC_DIAGNOSTIC_RESET
202 else { kernelSetup.ewaldExclusionType = EwaldExclusionType::Table; }
203 if (getenv("GMX_NBNXN_EWALD_TABLE") != nullptr)
205 kernelSetup.ewaldExclusionType = EwaldExclusionType::Table;
207 if (getenv("GMX_NBNXN_EWALD_ANALYTICAL") != nullptr)
209 kernelSetup.ewaldExclusionType = EwaldExclusionType::Analytical;
216 const char* lookup_kernel_name(const KernelType kernelType)
220 case KernelType::NotSet: return "not set";
221 case KernelType::Cpu4x4_PlainC: return "plain C";
222 case KernelType::Cpu4xN_Simd_4xN:
223 case KernelType::Cpu4xN_Simd_2xNN:
227 return "not available";
229 case KernelType::Gpu8x8x8: return "GPU";
230 case KernelType::Cpu8x8x8_PlainC: return "plain C";
232 default: gmx_fatal(FARGS, "Illegal kernel type selected");
236 /*! \brief Returns the most suitable kernel type and Ewald handling */
237 static KernelSetup pick_nbnxn_kernel(const gmx::MDLogger& mdlog,
238 gmx_bool use_simd_kernels,
239 const gmx_hw_info_t& hardwareInfo,
240 const NonbondedResource& nonbondedResource,
241 const t_inputrec& inputrec)
243 KernelSetup kernelSetup;
245 if (nonbondedResource == NonbondedResource::EmulateGpu)
247 kernelSetup.kernelType = KernelType::Cpu8x8x8_PlainC;
248 kernelSetup.ewaldExclusionType = EwaldExclusionType::DecidedByGpuModule;
250 GMX_LOG(mdlog.warning).asParagraph().appendText("Emulating a GPU run on the CPU (slow)");
252 else if (nonbondedResource == NonbondedResource::Gpu)
254 kernelSetup.kernelType = KernelType::Gpu8x8x8;
255 kernelSetup.ewaldExclusionType = EwaldExclusionType::DecidedByGpuModule;
259 if (use_simd_kernels && nbnxn_simd_supported(mdlog, inputrec))
261 kernelSetup = pick_nbnxn_kernel_cpu(inputrec, hardwareInfo);
265 kernelSetup.kernelType = KernelType::Cpu4x4_PlainC;
266 kernelSetup.ewaldExclusionType = EwaldExclusionType::Analytical;
272 .appendTextFormatted("Using %s %dx%d nonbonded short-range kernels",
273 lookup_kernel_name(kernelSetup.kernelType),
274 IClusterSizePerKernelType[kernelSetup.kernelType],
275 JClusterSizePerKernelType[kernelSetup.kernelType]);
277 if (KernelType::Cpu4x4_PlainC == kernelSetup.kernelType
278 || KernelType::Cpu8x8x8_PlainC == kernelSetup.kernelType)
280 GMX_LOG(mdlog.warning)
282 .appendTextFormatted(
283 "WARNING: Using the slow %s kernels. This should\n"
284 "not happen during routine usage on supported platforms.",
285 lookup_kernel_name(kernelSetup.kernelType));
288 GMX_RELEASE_ASSERT(kernelSetup.kernelType != KernelType::NotSet
289 && kernelSetup.ewaldExclusionType != EwaldExclusionType::NotSet,
290 "All kernel setup parameters should be set here");
297 PairlistSets::PairlistSets(const PairlistParams& pairlistParams,
298 const bool haveMultipleDomains,
299 const int minimumIlistCountForGpuBalancing) :
300 params_(pairlistParams), minimumIlistCountForGpuBalancing_(minimumIlistCountForGpuBalancing)
302 localSet_ = std::make_unique<PairlistSet>(params_);
304 if (haveMultipleDomains)
306 nonlocalSet_ = std::make_unique<PairlistSet>(params_);
313 /*! \brief Gets and returns the minimum i-list count for balacing based on the GPU used or env.var. when set */
314 static int getMinimumIlistCountForGpuBalancing(NbnxmGpu* nbnxmGpu)
316 if (const char* env = getenv("GMX_NB_MIN_CI"))
320 int minimumIlistCount = strtol(env, &end, 10);
321 if (!end || (*end != 0) || minimumIlistCount < 0)
324 FARGS, "Invalid value passed in GMX_NB_MIN_CI=%s, non-negative integer required", env);
329 fprintf(debug, "Neighbor-list balancing parameter: %d (passed as env. var.)\n", minimumIlistCount);
331 return minimumIlistCount;
335 int minimumIlistCount = gpu_min_ci_balanced(nbnxmGpu);
339 "Neighbor-list balancing parameter: %d (auto-adjusted to the number of GPU "
340 "multi-processors)\n",
343 return minimumIlistCount;
347 static int getENbnxnInitCombRule(const t_forcerec& forcerec)
349 if (forcerec.ic->vdwtype == VanDerWaalsType::Cut
350 && (forcerec.ic->vdw_modifier == InteractionModifiers::None
351 || forcerec.ic->vdw_modifier == InteractionModifiers::PotShift)
352 && getenv("GMX_NO_LJ_COMB_RULE") == nullptr)
354 /* Plain LJ cut-off: we can optimize with combination rules */
355 return enbnxninitcombruleDETECT;
357 else if (forcerec.ic->vdwtype == VanDerWaalsType::Pme)
359 /* LJ-PME: we need to use a combination rule for the grid */
360 if (forcerec.ljpme_combination_rule == LongRangeVdW::Geom)
362 return enbnxninitcombruleGEOM;
366 return enbnxninitcombruleLB;
371 /* We use a full combination matrix: no rule required */
372 return enbnxninitcombruleNONE;
376 std::unique_ptr<nonbonded_verlet_t> init_nb_verlet(const gmx::MDLogger& mdlog,
377 const t_inputrec& inputrec,
378 const t_forcerec& forcerec,
379 const t_commrec* commrec,
380 const gmx_hw_info_t& hardwareInfo,
381 bool useGpuForNonbonded,
382 const gmx::DeviceStreamManager* deviceStreamManager,
383 const gmx_mtop_t& mtop,
385 gmx_wallcycle* wcycle)
387 const bool emulateGpu = (getenv("GMX_EMULATE_GPU") != nullptr);
389 GMX_RELEASE_ASSERT(!(emulateGpu && useGpuForNonbonded),
390 "When GPU emulation is active, there cannot be a GPU assignment");
392 NonbondedResource nonbondedResource;
393 if (useGpuForNonbonded)
395 nonbondedResource = NonbondedResource::Gpu;
399 nonbondedResource = NonbondedResource::EmulateGpu;
403 nonbondedResource = NonbondedResource::Cpu;
406 Nbnxm::KernelSetup kernelSetup = pick_nbnxn_kernel(
407 mdlog, forcerec.use_simd_kernels, hardwareInfo, nonbondedResource, inputrec);
409 const bool haveMultipleDomains = havePPDomainDecomposition(commrec);
411 bool bFEP_NonBonded = (forcerec.efep != FreeEnergyPerturbationType::No)
412 && haveFepPerturbedNBInteractions(mtop);
413 PairlistParams pairlistParams(
414 kernelSetup.kernelType, bFEP_NonBonded, inputrec.rlist, haveMultipleDomains);
416 setupDynamicPairlistPruning(mdlog, inputrec, mtop, box, *forcerec.ic, &pairlistParams);
418 const int enbnxninitcombrule = getENbnxnInitCombRule(forcerec);
420 auto pinPolicy = (useGpuForNonbonded ? gmx::PinningPolicy::PinnedIfSupported
421 : gmx::PinningPolicy::CannotBePinned);
423 int mimimumNumEnergyGroupNonbonded = inputrec.opts.ngener;
424 if (inputrec.opts.ngener - inputrec.nwall == 1)
426 /* We have only one non-wall energy group, we do not need energy group
427 * support in the non-bondeds kernels, since all non-bonded energy
428 * contributions go to the first element of the energy group matrix.
430 mimimumNumEnergyGroupNonbonded = 1;
433 auto nbat = std::make_unique<nbnxn_atomdata_t>(
436 kernelSetup.kernelType,
440 mimimumNumEnergyGroupNonbonded,
441 (useGpuForNonbonded || emulateGpu) ? 1 : gmx_omp_nthreads_get(ModuleMultiThread::Nonbonded));
443 NbnxmGpu* gpu_nbv = nullptr;
444 int minimumIlistCountForGpuBalancing = 0;
445 if (useGpuForNonbonded)
447 /* init the NxN GPU data; the last argument tells whether we'll have
448 * both local and non-local NB calculation on GPU */
450 (deviceStreamManager != nullptr),
451 "Device stream manager should be initialized in order to use GPU for non-bonded.");
453 *deviceStreamManager, forcerec.ic.get(), pairlistParams, nbat.get(), haveMultipleDomains);
455 minimumIlistCountForGpuBalancing = getMinimumIlistCountForGpuBalancing(gpu_nbv);
458 auto pairlistSets = std::make_unique<PairlistSets>(
459 pairlistParams, haveMultipleDomains, minimumIlistCountForGpuBalancing);
462 std::make_unique<PairSearch>(inputrec.pbcType,
464 DOMAINDECOMP(commrec) ? &commrec->dd->numCells : nullptr,
465 DOMAINDECOMP(commrec) ? domdec_zones(commrec->dd) : nullptr,
466 pairlistParams.pairlistType,
468 gmx_omp_nthreads_get(ModuleMultiThread::Pairsearch),
471 return std::make_unique<nonbonded_verlet_t>(
472 std::move(pairlistSets), std::move(pairSearch), std::move(nbat), kernelSetup, gpu_nbv, wcycle);
477 nonbonded_verlet_t::nonbonded_verlet_t(std::unique_ptr<PairlistSets> pairlistSets,
478 std::unique_ptr<PairSearch> pairSearch,
479 std::unique_ptr<nbnxn_atomdata_t> nbat_in,
480 const Nbnxm::KernelSetup& kernelSetup,
481 NbnxmGpu* gpu_nbv_ptr,
482 gmx_wallcycle* wcycle) :
483 pairlistSets_(std::move(pairlistSets)),
484 pairSearch_(std::move(pairSearch)),
485 nbat(std::move(nbat_in)),
486 kernelSetup_(kernelSetup),
488 foreignEnergyGroups_(std::make_unique<gmx_grppairener_t>(nbat->params().nenergrp)),
491 GMX_RELEASE_ASSERT(pairlistSets_, "Need valid pairlistSets");
492 GMX_RELEASE_ASSERT(pairSearch_, "Need valid search object");
493 GMX_RELEASE_ASSERT(nbat, "Need valid atomdata object");
496 nonbonded_verlet_t::~nonbonded_verlet_t()
498 Nbnxm::gpu_free(gpu_nbv);