2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2019,2020, 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
45 #include "gromacs/domdec/domdec.h"
46 #include "gromacs/domdec/domdec_struct.h"
47 #include "gromacs/hardware/hw_info.h"
48 #include "gromacs/mdlib/gmx_omp_nthreads.h"
49 #include "gromacs/mdtypes/commrec.h"
50 #include "gromacs/mdtypes/forcerec.h"
51 #include "gromacs/mdtypes/inputrec.h"
52 #include "gromacs/mdtypes/interaction_const.h"
53 #include "gromacs/nbnxm/atomdata.h"
54 #include "gromacs/nbnxm/gpu_data_mgmt.h"
55 #include "gromacs/nbnxm/nbnxm.h"
56 #include "gromacs/nbnxm/pairlist_tuning.h"
57 #include "gromacs/simd/simd.h"
58 #include "gromacs/topology/mtop_util.h"
59 #include "gromacs/utility/fatalerror.h"
60 #include "gromacs/utility/logger.h"
63 #include "nbnxm_geometry.h"
64 #include "nbnxm_simd.h"
66 #include "pairlistset.h"
67 #include "pairlistsets.h"
68 #include "pairsearch.h"
73 /*! \brief Resources that can be used to execute non-bonded kernels on */
74 enum class NonbondedResource : int
81 /*! \brief Returns whether CPU SIMD support exists for the given inputrec
83 * If the return value is FALSE and fplog/cr != NULL, prints a fallback
84 * message to fplog/stderr.
86 static gmx_bool nbnxn_simd_supported(const gmx::MDLogger& mdlog, const t_inputrec* ir)
88 if (ir->vdwtype == evdwPME && ir->ljpme_combination_rule == eljpmeLB)
90 /* LJ PME with LB combination rule does 7 mesh operations.
91 * This so slow that we don't compile SIMD non-bonded kernels
93 GMX_LOG(mdlog.warning)
96 "LJ-PME with Lorentz-Berthelot is not supported with SIMD kernels, falling "
97 "back to plain C kernels");
104 /*! \brief Returns the most suitable CPU kernel type and Ewald handling */
105 static KernelSetup pick_nbnxn_kernel_cpu(const t_inputrec gmx_unused* ir,
106 const gmx_hw_info_t gmx_unused& hardwareInfo)
108 KernelSetup kernelSetup;
112 kernelSetup.kernelType = KernelType::Cpu4x4_PlainC;
113 kernelSetup.ewaldExclusionType = EwaldExclusionType::Table;
117 #ifdef GMX_NBNXN_SIMD_4XN
118 kernelSetup.kernelType = KernelType::Cpu4xN_Simd_4xN;
120 #ifdef GMX_NBNXN_SIMD_2XNN
121 kernelSetup.kernelType = KernelType::Cpu4xN_Simd_2xNN;
124 #if defined GMX_NBNXN_SIMD_2XNN && defined GMX_NBNXN_SIMD_4XN
125 /* We need to choose if we want 2x(N+N) or 4xN kernels.
126 * This is based on the SIMD acceleration choice and CPU information
127 * detected at runtime.
129 * 4xN calculates more (zero) interactions, but has less pair-search
130 * work and much better kernel instruction scheduling.
132 * Up till now we have only seen that on Intel Sandy/Ivy Bridge,
133 * which doesn't have FMA, both the analytical and tabulated Ewald
134 * kernels have similar pair rates for 4x8 and 2x(4+4), so we choose
135 * 2x(4+4) because it results in significantly fewer pairs.
136 * For RF, the raw pair rate of the 4x8 kernel is higher than 2x(4+4),
137 * 10% with HT, 50% without HT. As we currently don't detect the actual
138 * use of HT, use 4x8 to avoid a potential performance hit.
139 * On Intel Haswell 4x8 is always faster.
141 kernelSetup.kernelType = KernelType::Cpu4xN_Simd_4xN;
143 if (!GMX_SIMD_HAVE_FMA && (EEL_PME_EWALD(ir->coulombtype) || EVDW_PME(ir->vdwtype)))
145 /* We have Ewald kernels without FMA (Intel Sandy/Ivy Bridge).
146 * There are enough instructions to make 2x(4+4) efficient.
148 kernelSetup.kernelType = KernelType::Cpu4xN_Simd_2xNN;
151 if (hardwareInfo.haveAmdZen1Cpu)
153 /* One 256-bit FMA per cycle makes 2xNN faster */
154 kernelSetup.kernelType = KernelType::Cpu4xN_Simd_2xNN;
156 #endif /* GMX_NBNXN_SIMD_2XNN && GMX_NBNXN_SIMD_4XN */
159 if (getenv("GMX_NBNXN_SIMD_4XN") != nullptr)
161 #ifdef GMX_NBNXN_SIMD_4XN
162 kernelSetup.kernelType = KernelType::Cpu4xN_Simd_4xN;
165 "SIMD 4xN kernels requested, but GROMACS has been compiled without support "
166 "for these kernels");
169 if (getenv("GMX_NBNXN_SIMD_2XNN") != nullptr)
171 #ifdef GMX_NBNXN_SIMD_2XNN
172 kernelSetup.kernelType = KernelType::Cpu4xN_Simd_2xNN;
175 "SIMD 2x(N+N) kernels requested, but GROMACS has been compiled without "
176 "support for these kernels");
180 /* Analytical Ewald exclusion correction is only an option in
182 * Since table lookup's don't parallelize with SIMD, analytical
183 * will probably always be faster for a SIMD width of 8 or more.
184 * With FMA analytical is sometimes faster for a width if 4 as well.
185 * In single precision, this is faster on Bulldozer.
186 * On AMD Zen, tabulated Ewald kernels are faster on all 4 combinations
187 * of single or double precision and 128 or 256-bit AVX2.
191 (GMX_SIMD_REAL_WIDTH >= 8 || (GMX_SIMD_REAL_WIDTH >= 4 && GMX_SIMD_HAVE_FMA && !GMX_DOUBLE)) &&
193 !hardwareInfo.haveAmdZen1Cpu)
195 kernelSetup.ewaldExclusionType = EwaldExclusionType::Analytical;
199 kernelSetup.ewaldExclusionType = EwaldExclusionType::Table;
201 if (getenv("GMX_NBNXN_EWALD_TABLE") != nullptr)
203 kernelSetup.ewaldExclusionType = EwaldExclusionType::Table;
205 if (getenv("GMX_NBNXN_EWALD_ANALYTICAL") != nullptr)
207 kernelSetup.ewaldExclusionType = EwaldExclusionType::Analytical;
214 const char* lookup_kernel_name(const KernelType kernelType)
216 const char* returnvalue = nullptr;
219 case KernelType::NotSet: returnvalue = "not set"; break;
220 case KernelType::Cpu4x4_PlainC: returnvalue = "plain C"; break;
221 case KernelType::Cpu4xN_Simd_4xN:
222 case KernelType::Cpu4xN_Simd_2xNN:
224 returnvalue = "SIMD";
226 returnvalue = "not available";
229 case KernelType::Gpu8x8x8: returnvalue = "GPU"; break;
230 case KernelType::Cpu8x8x8_PlainC: returnvalue = "plain C"; break;
232 default: gmx_fatal(FARGS, "Illegal kernel type selected");
237 /*! \brief Returns the most suitable kernel type and Ewald handling */
238 static KernelSetup pick_nbnxn_kernel(const gmx::MDLogger& mdlog,
239 gmx_bool use_simd_kernels,
240 const gmx_hw_info_t& hardwareInfo,
241 const NonbondedResource& nonbondedResource,
242 const t_inputrec* ir,
243 gmx_bool bDoNonbonded)
245 KernelSetup kernelSetup;
247 if (nonbondedResource == NonbondedResource::EmulateGpu)
249 kernelSetup.kernelType = KernelType::Cpu8x8x8_PlainC;
250 kernelSetup.ewaldExclusionType = EwaldExclusionType::DecidedByGpuModule;
254 GMX_LOG(mdlog.warning)
256 .appendText("Emulating a GPU run on the CPU (slow)");
259 else if (nonbondedResource == NonbondedResource::Gpu)
261 kernelSetup.kernelType = KernelType::Gpu8x8x8;
262 kernelSetup.ewaldExclusionType = EwaldExclusionType::DecidedByGpuModule;
266 if (use_simd_kernels && nbnxn_simd_supported(mdlog, ir))
268 kernelSetup = pick_nbnxn_kernel_cpu(ir, hardwareInfo);
272 kernelSetup.kernelType = KernelType::Cpu4x4_PlainC;
273 kernelSetup.ewaldExclusionType = EwaldExclusionType::Analytical;
281 .appendTextFormatted("Using %s %dx%d nonbonded short-range kernels",
282 lookup_kernel_name(kernelSetup.kernelType),
283 IClusterSizePerKernelType[kernelSetup.kernelType],
284 JClusterSizePerKernelType[kernelSetup.kernelType]);
286 if (KernelType::Cpu4x4_PlainC == kernelSetup.kernelType
287 || KernelType::Cpu8x8x8_PlainC == kernelSetup.kernelType)
289 GMX_LOG(mdlog.warning)
291 .appendTextFormatted(
292 "WARNING: Using the slow %s kernels. This should\n"
293 "not happen during routine usage on supported platforms.",
294 lookup_kernel_name(kernelSetup.kernelType));
298 GMX_RELEASE_ASSERT(kernelSetup.kernelType != KernelType::NotSet
299 && kernelSetup.ewaldExclusionType != EwaldExclusionType::NotSet,
300 "All kernel setup parameters should be set here");
307 PairlistSets::PairlistSets(const PairlistParams& pairlistParams,
308 const bool haveMultipleDomains,
309 const int minimumIlistCountForGpuBalancing) :
310 params_(pairlistParams),
311 minimumIlistCountForGpuBalancing_(minimumIlistCountForGpuBalancing)
313 localSet_ = std::make_unique<PairlistSet>(gmx::InteractionLocality::Local, params_);
315 if (haveMultipleDomains)
317 nonlocalSet_ = std::make_unique<PairlistSet>(gmx::InteractionLocality::NonLocal, params_);
324 /*! \brief Gets and returns the minimum i-list count for balacing based on the GPU used or env.var. when set */
325 static int getMinimumIlistCountForGpuBalancing(NbnxmGpu* nbnxmGpu)
327 int minimumIlistCount;
329 if (const char* env = getenv("GMX_NB_MIN_CI"))
333 minimumIlistCount = strtol(env, &end, 10);
334 if (!end || (*end != 0) || minimumIlistCount < 0)
337 "Invalid value passed in GMX_NB_MIN_CI=%s, non-negative integer required", env);
342 fprintf(debug, "Neighbor-list balancing parameter: %d (passed as env. var.)\n",
348 minimumIlistCount = gpu_min_ci_balanced(nbnxmGpu);
352 "Neighbor-list balancing parameter: %d (auto-adjusted to the number of GPU "
353 "multi-processors)\n",
358 return minimumIlistCount;
361 std::unique_ptr<nonbonded_verlet_t> init_nb_verlet(const gmx::MDLogger& mdlog,
362 const t_inputrec* ir,
363 const t_forcerec* fr,
365 const gmx_hw_info_t& hardwareInfo,
366 const DeviceInformation* deviceInfo,
367 const DeviceContext* deviceContext,
368 const gmx_mtop_t* mtop,
370 gmx_wallcycle* wcycle)
372 const bool emulateGpu = (getenv("GMX_EMULATE_GPU") != nullptr);
373 const bool useGpu = deviceInfo != nullptr;
375 GMX_RELEASE_ASSERT(!(emulateGpu && useGpu),
376 "When GPU emulation is active, there cannot be a GPU assignment");
378 NonbondedResource nonbondedResource;
381 nonbondedResource = NonbondedResource::Gpu;
385 nonbondedResource = NonbondedResource::EmulateGpu;
389 nonbondedResource = NonbondedResource::Cpu;
392 Nbnxm::KernelSetup kernelSetup = pick_nbnxn_kernel(mdlog, fr->use_simd_kernels, hardwareInfo,
393 nonbondedResource, ir, fr->bNonbonded);
395 const bool haveMultipleDomains = havePPDomainDecomposition(cr);
397 bool bFEP_NonBonded = (fr->efep != efepNO) && haveFepPerturbedNBInteractions(mtop);
398 PairlistParams pairlistParams(kernelSetup.kernelType, bFEP_NonBonded, ir->rlist, haveMultipleDomains);
400 setupDynamicPairlistPruning(mdlog, ir, mtop, box, fr->ic, &pairlistParams);
402 int enbnxninitcombrule;
403 if (fr->ic->vdwtype == evdwCUT
404 && (fr->ic->vdw_modifier == eintmodNONE || fr->ic->vdw_modifier == eintmodPOTSHIFT)
405 && getenv("GMX_NO_LJ_COMB_RULE") == nullptr)
407 /* Plain LJ cut-off: we can optimize with combination rules */
408 enbnxninitcombrule = enbnxninitcombruleDETECT;
410 else if (fr->ic->vdwtype == evdwPME)
412 /* LJ-PME: we need to use a combination rule for the grid */
413 if (fr->ljpme_combination_rule == eljpmeGEOM)
415 enbnxninitcombrule = enbnxninitcombruleGEOM;
419 enbnxninitcombrule = enbnxninitcombruleLB;
424 /* We use a full combination matrix: no rule required */
425 enbnxninitcombrule = enbnxninitcombruleNONE;
428 auto pinPolicy = (useGpu ? gmx::PinningPolicy::PinnedIfSupported : gmx::PinningPolicy::CannotBePinned);
430 auto nbat = std::make_unique<nbnxn_atomdata_t>(pinPolicy);
432 int mimimumNumEnergyGroupNonbonded = ir->opts.ngener;
433 if (ir->opts.ngener - ir->nwall == 1)
435 /* We have only one non-wall energy group, we do not need energy group
436 * support in the non-bondeds kernels, since all non-bonded energy
437 * contributions go to the first element of the energy group matrix.
439 mimimumNumEnergyGroupNonbonded = 1;
441 nbnxn_atomdata_init(mdlog, nbat.get(), kernelSetup.kernelType, enbnxninitcombrule, fr->ntype,
442 fr->nbfp, mimimumNumEnergyGroupNonbonded,
443 (useGpu || emulateGpu) ? 1 : gmx_omp_nthreads_get(emntNonbonded));
445 NbnxmGpu* gpu_nbv = nullptr;
446 int minimumIlistCountForGpuBalancing = 0;
450 deviceContext != nullptr,
451 "Device context can not be nullptr when to use GPU for non-bonded forces.");
452 /* init the NxN GPU data; the last argument tells whether we'll have
453 * both local and non-local NB calculation on GPU */
454 gpu_nbv = gpu_init(*deviceContext, fr->ic, pairlistParams, nbat.get(), haveMultipleDomains);
456 minimumIlistCountForGpuBalancing = getMinimumIlistCountForGpuBalancing(gpu_nbv);
459 auto pairlistSets = std::make_unique<PairlistSets>(pairlistParams, haveMultipleDomains,
460 minimumIlistCountForGpuBalancing);
462 auto pairSearch = std::make_unique<PairSearch>(
463 ir->pbcType, EI_TPI(ir->eI), DOMAINDECOMP(cr) ? &cr->dd->numCells : nullptr,
464 DOMAINDECOMP(cr) ? domdec_zones(cr->dd) : nullptr, pairlistParams.pairlistType,
465 bFEP_NonBonded, gmx_omp_nthreads_get(emntPairsearch), pinPolicy);
467 return std::make_unique<nonbonded_verlet_t>(std::move(pairlistSets), std::move(pairSearch),
468 std::move(nbat), kernelSetup, gpu_nbv, wcycle);
473 nonbonded_verlet_t::nonbonded_verlet_t(std::unique_ptr<PairlistSets> pairlistSets,
474 std::unique_ptr<PairSearch> pairSearch,
475 std::unique_ptr<nbnxn_atomdata_t> nbat_in,
476 const Nbnxm::KernelSetup& kernelSetup,
477 NbnxmGpu* gpu_nbv_ptr,
478 gmx_wallcycle* wcycle) :
479 pairlistSets_(std::move(pairlistSets)),
480 pairSearch_(std::move(pairSearch)),
481 nbat(std::move(nbat_in)),
482 kernelSetup_(kernelSetup),
486 GMX_RELEASE_ASSERT(pairlistSets_, "Need valid pairlistSets");
487 GMX_RELEASE_ASSERT(pairSearch_, "Need valid search object");
488 GMX_RELEASE_ASSERT(nbat, "Need valid atomdata object");
491 nonbonded_verlet_t::~nonbonded_verlet_t()
493 Nbnxm::gpu_free(gpu_nbv);