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/pmalloc.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, deviceContext);
111 ad->shiftVecUploaded = false;
113 allocateDeviceBuffer(&ad->fShift, SHIFTS, 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);
167 /* set up LJ parameter lookup table */
168 static_assert(sizeof(Float2) == 2 * sizeof(decltype(*nbatParams.nbfp.data())),
169 "Mismatch in the size of host / device data types");
170 DeviceBuffer<Float2> nbfp;
171 initParamLookupTable(&nbfp,
173 reinterpret_cast<const Float2*>(nbatParams.nbfp.data()),
174 nbatParams.numTypes * nbatParams.numTypes,
178 if (ic->vdwtype == VanDerWaalsType::Pme)
180 static_assert(sizeof(Float2) == 2 * sizeof(decltype(*nbatParams.nbfp_comb.data())),
181 "Mismatch in the size of host / device data types");
182 DeviceBuffer<Float2> nbfp_comb;
183 initParamLookupTable(&nbfp_comb,
185 reinterpret_cast<const Float2*>(nbatParams.nbfp_comb.data()),
188 nbp->nbfp_comb = nbfp_comb;
193 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
194 static cl_kernel nbnxn_gpu_create_kernel(NbnxmGpu* nb, const char* kernel_name)
199 kernel = clCreateKernel(nb->dev_rundata->program, kernel_name, &cl_error);
200 if (CL_SUCCESS != cl_error)
203 "Failed to create kernel '%s' for GPU #%s: OpenCL error %d",
205 nb->deviceContext_->deviceInfo().device_name,
212 /*! \brief Clears nonbonded shift force output array and energy outputs on the GPU.
214 static void nbnxn_ocl_clear_e_fshift(NbnxmGpu* nb)
218 NBAtomData* adat = nb->atdat;
219 cl_command_queue ls = nb->deviceStreams[InteractionLocality::Local]->stream();
221 size_t local_work_size[3] = { 1, 1, 1 };
222 size_t global_work_size[3] = { 1, 1, 1 };
224 cl_int shifts = SHIFTS * 3;
228 cl_kernel zero_e_fshift = nb->kernel_zero_e_fshift;
230 local_work_size[0] = 64;
231 // Round the total number of threads up from the array size
232 global_work_size[0] = ((shifts + local_work_size[0] - 1) / local_work_size[0]) * local_work_size[0];
235 cl_error = clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->fShift));
236 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->eLJ));
237 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->eElec));
238 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_uint), &shifts);
239 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
241 cl_error = clEnqueueNDRangeKernel(
242 ls, zero_e_fshift, 3, nullptr, global_work_size, local_work_size, 0, nullptr, nullptr);
243 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
246 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
247 static void nbnxn_gpu_init_kernels(NbnxmGpu* nb)
249 /* Init to 0 main kernel arrays */
250 /* They will be later on initialized in select_nbnxn_kernel */
251 // TODO: consider always creating all variants of the kernels here so that there is no
252 // need for late call to clCreateKernel -- if that gives any advantage?
253 memset(nb->kernel_ener_noprune_ptr, 0, sizeof(nb->kernel_ener_noprune_ptr));
254 memset(nb->kernel_ener_prune_ptr, 0, sizeof(nb->kernel_ener_prune_ptr));
255 memset(nb->kernel_noener_noprune_ptr, 0, sizeof(nb->kernel_noener_noprune_ptr));
256 memset(nb->kernel_noener_prune_ptr, 0, sizeof(nb->kernel_noener_prune_ptr));
258 /* Init pruning kernels
260 * TODO: we could avoid creating kernels if dynamic pruning is turned off,
261 * but ATM that depends on force flags not passed into the initialization.
263 nb->kernel_pruneonly[epruneFirst] = nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_opencl");
264 nb->kernel_pruneonly[epruneRolling] =
265 nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_rolling_opencl");
267 /* Init auxiliary kernels */
268 nb->kernel_zero_e_fshift = nbnxn_gpu_create_kernel(nb, "zero_e_fshift");
271 /*! \brief Initializes simulation constant data.
273 * Initializes members of the atomdata and nbparam structs and
274 * clears e/fshift output buffers.
276 static void nbnxn_ocl_init_const(NBAtomData* atomData,
277 NBParamGpu* nbParams,
278 const interaction_const_t* ic,
279 const PairlistParams& listParams,
280 const nbnxn_atomdata_t::Params& nbatParams,
281 const DeviceContext& deviceContext)
283 init_atomdata_first(atomData, nbatParams.numTypes, deviceContext);
284 init_nbparam(nbParams, ic, listParams, nbatParams, deviceContext);
288 //! This function is documented in the header file
289 NbnxmGpu* gpu_init(const gmx::DeviceStreamManager& deviceStreamManager,
290 const interaction_const_t* ic,
291 const PairlistParams& listParams,
292 const nbnxn_atomdata_t* nbat,
293 const bool bLocalAndNonlocal)
295 GMX_ASSERT(ic, "Need a valid interaction constants object");
297 auto nb = new NbnxmGpu();
298 nb->deviceContext_ = &deviceStreamManager.context();
300 snew(nb->nbparam, 1);
301 snew(nb->plist[InteractionLocality::Local], 1);
302 if (bLocalAndNonlocal)
304 snew(nb->plist[InteractionLocality::NonLocal], 1);
307 nb->bUseTwoStreams = bLocalAndNonlocal;
309 nb->timers = new Nbnxm::GpuTimers();
310 snew(nb->timings, 1);
312 /* set device info, just point it to the right GPU among the detected ones */
313 nb->dev_rundata = new gmx_device_runtime_data_t();
316 pmalloc(reinterpret_cast<void**>(&nb->nbst.eLJ), sizeof(*nb->nbst.eLJ));
317 pmalloc(reinterpret_cast<void**>(&nb->nbst.eElec), sizeof(*nb->nbst.eElec));
318 pmalloc(reinterpret_cast<void**>(&nb->nbst.fShift), SHIFTS * sizeof(*nb->nbst.fShift));
320 init_plist(nb->plist[InteractionLocality::Local]);
322 /* OpenCL timing disabled if GMX_DISABLE_GPU_TIMING is defined. */
323 nb->bDoTime = (getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
325 /* local/non-local GPU streams */
326 GMX_RELEASE_ASSERT(deviceStreamManager.streamIsValid(gmx::DeviceStreamType::NonBondedLocal),
327 "Local non-bonded stream should be initialized to use GPU for non-bonded.");
328 nb->deviceStreams[InteractionLocality::Local] =
329 &deviceStreamManager.stream(gmx::DeviceStreamType::NonBondedLocal);
331 if (nb->bUseTwoStreams)
333 init_plist(nb->plist[InteractionLocality::NonLocal]);
335 GMX_RELEASE_ASSERT(deviceStreamManager.streamIsValid(gmx::DeviceStreamType::NonBondedNonLocal),
336 "Non-local non-bonded stream should be initialized to use GPU for "
337 "non-bonded with domain decomposition.");
338 nb->deviceStreams[InteractionLocality::NonLocal] =
339 &deviceStreamManager.stream(gmx::DeviceStreamType::NonBondedNonLocal);
344 init_timings(nb->timings);
347 nbnxn_ocl_init_const(nb->atdat, nb->nbparam, ic, listParams, nbat->params(), *nb->deviceContext_);
349 /* Enable LJ param manual prefetch for AMD or Intel or if we request through env. var.
350 * TODO: decide about NVIDIA
352 nb->bPrefetchLjParam = (getenv("GMX_OCL_DISABLE_I_PREFETCH") == nullptr)
353 && ((nb->deviceContext_->deviceInfo().deviceVendor == DeviceVendor::Amd)
354 || (nb->deviceContext_->deviceInfo().deviceVendor == DeviceVendor::Intel)
355 || (getenv("GMX_OCL_ENABLE_I_PREFETCH") != nullptr));
357 /* NOTE: in CUDA we pick L1 cache configuration for the nbnxn kernels here,
358 * but sadly this is not supported in OpenCL (yet?). Consider adding it if
359 * it becomes supported.
361 nbnxn_gpu_compile_kernels(nb);
362 nbnxn_gpu_init_kernels(nb);
364 /* clear energy and shift force outputs */
365 nbnxn_ocl_clear_e_fshift(nb);
369 fprintf(debug, "Initialized OpenCL data structures.\n");
375 /*! \brief Clears the first natoms_clear elements of the GPU nonbonded force output array.
377 static void nbnxn_ocl_clear_f(NbnxmGpu* nb, int natoms_clear)
379 if (natoms_clear == 0)
384 NBAtomData* atomData = nb->atdat;
385 const DeviceStream& localStream = *nb->deviceStreams[InteractionLocality::Local];
387 clearDeviceBufferAsync(&atomData->f, 0, natoms_clear, localStream);
390 //! This function is documented in the header file
391 void gpu_clear_outputs(NbnxmGpu* nb, bool computeVirial)
393 nbnxn_ocl_clear_f(nb, nb->atdat->numAtoms);
394 /* clear shift force array and energies if the outputs were
395 used in the current step */
398 nbnxn_ocl_clear_e_fshift(nb);
401 /* kick off buffer clearing kernel to ensure concurrency with constraints/update */
402 cl_int gmx_unused cl_error;
403 cl_error = clFlush(nb->deviceStreams[InteractionLocality::Local]->stream());
404 GMX_ASSERT(cl_error == CL_SUCCESS, ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
407 //! This function is documented in the header file
408 void gpu_upload_shiftvec(NbnxmGpu* nb, const nbnxn_atomdata_t* nbatom)
410 NBAtomData* adat = nb->atdat;
411 const DeviceStream& localStream = *nb->deviceStreams[InteractionLocality::Local];
413 /* only if we have a dynamic box */
414 if (nbatom->bDynamicBox || !adat->shiftVecUploaded)
416 static_assert(sizeof(Float3) == sizeof(nbatom->shift_vec[0]),
417 "Sizes of host- and device-side shift vectors should be the same.");
418 copyToDeviceBuffer(&adat->shiftVec,
419 reinterpret_cast<const Float3*>(nbatom->shift_vec.data()),
423 GpuApiCallBehavior::Async,
425 adat->shiftVecUploaded = true;
429 /*! \brief Releases an OpenCL kernel pointer */
430 static void free_kernel(cl_kernel* kernel_ptr)
432 cl_int gmx_unused cl_error;
434 GMX_ASSERT(kernel_ptr, "Need a valid kernel pointer");
438 cl_error = clReleaseKernel(*kernel_ptr);
439 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
440 ("clReleaseKernel failed: " + ocl_get_error_string(cl_error)).c_str());
442 *kernel_ptr = nullptr;
446 /*! \brief Releases a list of OpenCL kernel pointers */
447 static void free_kernels(cl_kernel* kernels, int count)
451 for (i = 0; i < count; i++)
453 free_kernel(kernels + i);
457 /*! \brief Free the OpenCL program.
459 * The function releases the OpenCL program assuciated with the
460 * device that the calling PP rank is running on.
462 * \param program [in] OpenCL program to release.
464 static void freeGpuProgram(cl_program program)
468 cl_int cl_error = clReleaseProgram(program);
469 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
470 ("clReleaseProgram failed: " + ocl_get_error_string(cl_error)).c_str());
475 //! This function is documented in the header file
476 void gpu_free(NbnxmGpu* nb)
484 // NOLINTNEXTLINE(bugprone-sizeof-expression)
485 int kernel_count = sizeof(nb->kernel_ener_noprune_ptr) / sizeof(nb->kernel_ener_noprune_ptr[0][0]);
486 free_kernels(nb->kernel_ener_noprune_ptr[0], kernel_count);
488 // NOLINTNEXTLINE(bugprone-sizeof-expression)
489 kernel_count = sizeof(nb->kernel_ener_prune_ptr) / sizeof(nb->kernel_ener_prune_ptr[0][0]);
490 free_kernels(nb->kernel_ener_prune_ptr[0], kernel_count);
492 // NOLINTNEXTLINE(bugprone-sizeof-expression)
493 kernel_count = sizeof(nb->kernel_noener_noprune_ptr) / sizeof(nb->kernel_noener_noprune_ptr[0][0]);
494 free_kernels(nb->kernel_noener_noprune_ptr[0], kernel_count);
496 // NOLINTNEXTLINE(bugprone-sizeof-expression)
497 kernel_count = sizeof(nb->kernel_noener_prune_ptr) / sizeof(nb->kernel_noener_prune_ptr[0][0]);
498 free_kernels(nb->kernel_noener_prune_ptr[0], kernel_count);
500 free_kernel(&(nb->kernel_zero_e_fshift));
503 freeDeviceBuffer(&(nb->atdat->xq));
504 freeDeviceBuffer(&(nb->atdat->f));
505 freeDeviceBuffer(&(nb->atdat->eLJ));
506 freeDeviceBuffer(&(nb->atdat->eElec));
507 freeDeviceBuffer(&(nb->atdat->fShift));
508 freeDeviceBuffer(&(nb->atdat->ljComb));
509 freeDeviceBuffer(&(nb->atdat->atomTypes));
510 freeDeviceBuffer(&(nb->atdat->shiftVec));
514 freeDeviceBuffer(&(nb->nbparam->nbfp));
515 freeDeviceBuffer(&(nb->nbparam->nbfp_comb));
516 freeDeviceBuffer(&(nb->nbparam->coulomb_tab));
520 auto* plist = nb->plist[InteractionLocality::Local];
521 freeDeviceBuffer(&plist->sci);
522 freeDeviceBuffer(&plist->cj4);
523 freeDeviceBuffer(&plist->imask);
524 freeDeviceBuffer(&plist->excl);
526 if (nb->bUseTwoStreams)
528 auto* plist_nl = nb->plist[InteractionLocality::NonLocal];
529 freeDeviceBuffer(&plist_nl->sci);
530 freeDeviceBuffer(&plist_nl->cj4);
531 freeDeviceBuffer(&plist_nl->imask);
532 freeDeviceBuffer(&plist_nl->excl);
538 nb->nbst.eLJ = nullptr;
540 pfree(nb->nbst.eElec);
541 nb->nbst.eElec = nullptr;
543 pfree(nb->nbst.fShift);
544 nb->nbst.fShift = nullptr;
546 freeGpuProgram(nb->dev_rundata->program);
547 delete nb->dev_rundata;
549 /* Free timers and timings */
556 fprintf(debug, "Cleaned up OpenCL data structures.\n");
560 //! This function is documented in the header file
561 int gpu_min_ci_balanced(NbnxmGpu* nb)
563 return nb != nullptr ? gpu_min_ci_balanced_factor * nb->deviceContext_->deviceInfo().compute_units : 0;