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/gpu_utils.h"
57 #include "gromacs/gpu_utils/oclutils.h"
58 #include "gromacs/hardware/gpu_hw_info.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/pairlistsets.h"
69 #include "gromacs/pbcutil/ishift.h"
70 #include "gromacs/timing/gpu_timing.h"
71 #include "gromacs/utility/cstringutil.h"
72 #include "gromacs/utility/fatalerror.h"
73 #include "gromacs/utility/gmxassert.h"
74 #include "gromacs/utility/real.h"
75 #include "gromacs/utility/smalloc.h"
77 #include "nbnxm_ocl_internal.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 Returns true if LJ combination rules are used in the non-bonded kernels.
105 * Full doc in nbnxn_ocl_internal.h */
106 bool useLjCombRule(int vdwType)
108 return (vdwType == evdwOclCUTCOMBGEOM || vdwType == evdwOclCUTCOMBLB);
111 /*! \brief Tabulates the Ewald Coulomb force and initializes the size/scale
112 * and the table GPU array.
114 * If called with an already allocated table, it just re-uploads the
117 static void init_ewald_coulomb_force_table(const EwaldCorrectionTables& tables,
119 const DeviceContext& deviceContext)
121 if (nbp->coulomb_tab_climg2d != nullptr)
123 freeDeviceBuffer(&(nbp->coulomb_tab_climg2d));
126 DeviceBuffer<real> coulomb_tab;
128 initParamLookupTable(&coulomb_tab, nullptr, tables.tableF.data(), tables.tableF.size(), deviceContext);
130 nbp->coulomb_tab_climg2d = coulomb_tab;
131 nbp->coulomb_tab_scale = tables.scale;
135 /*! \brief Initializes the atomdata structure first time, it only gets filled at
138 static void init_atomdata_first(cl_atomdata_t* ad, int ntypes, const DeviceContext& deviceContext)
144 ad->shift_vec = clCreateBuffer(deviceContext.context(), CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
145 SHIFTS * sizeof(nbnxn_atomdata_t::shift_vec[0]), nullptr, &cl_error);
146 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
147 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
148 ad->bShiftVecUploaded = CL_FALSE;
150 ad->fshift = clCreateBuffer(deviceContext.context(), CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
151 SHIFTS * sizeof(nb_staging_t::fshift[0]), nullptr, &cl_error);
152 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
153 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
155 ad->e_lj = clCreateBuffer(deviceContext.context(), CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
156 sizeof(float), nullptr, &cl_error);
157 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
158 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
160 ad->e_el = clCreateBuffer(deviceContext.context(), CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
161 sizeof(float), nullptr, &cl_error);
162 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
163 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
165 /* initialize to nullptr pointers to data that is not allocated here and will
166 need reallocation in nbnxn_gpu_init_atomdata */
170 /* size -1 indicates that the respective array hasn't been initialized yet */
175 /*! \brief Copies all parameters related to the cut-off from ic to nbp
177 static void set_cutoff_parameters(cl_nbparam_t* nbp, const interaction_const_t* ic, const PairlistParams& listParams)
179 nbp->ewald_beta = ic->ewaldcoeff_q;
180 nbp->sh_ewald = ic->sh_ewald;
181 nbp->epsfac = ic->epsfac;
182 nbp->two_k_rf = 2.0 * ic->k_rf;
183 nbp->c_rf = ic->c_rf;
184 nbp->rvdw_sq = ic->rvdw * ic->rvdw;
185 nbp->rcoulomb_sq = ic->rcoulomb * ic->rcoulomb;
186 nbp->rlistOuter_sq = listParams.rlistOuter * listParams.rlistOuter;
187 nbp->rlistInner_sq = listParams.rlistInner * listParams.rlistInner;
188 nbp->useDynamicPruning = listParams.useDynamicPruning;
190 nbp->sh_lj_ewald = ic->sh_lj_ewald;
191 nbp->ewaldcoeff_lj = ic->ewaldcoeff_lj;
193 nbp->rvdw_switch = ic->rvdw_switch;
194 nbp->dispersion_shift = ic->dispersion_shift;
195 nbp->repulsion_shift = ic->repulsion_shift;
196 nbp->vdw_switch = ic->vdw_switch;
199 /*! \brief Returns the kinds of electrostatics and Vdw OpenCL
200 * kernels that will be used.
202 * Respectively, these values are from enum eelOcl and enum
204 static void map_interaction_types_to_gpu_kernel_flavors(const interaction_const_t* ic,
209 if (ic->vdwtype == evdwCUT)
211 switch (ic->vdw_modifier)
214 case eintmodPOTSHIFT:
217 case ljcrNONE: *gpu_vdwtype = evdwOclCUT; break;
218 case ljcrGEOM: *gpu_vdwtype = evdwOclCUTCOMBGEOM; break;
219 case ljcrLB: *gpu_vdwtype = evdwOclCUTCOMBLB; break;
222 "The requested LJ combination rule is not implemented in the "
223 "OpenCL GPU accelerated kernels!");
226 case eintmodFORCESWITCH: *gpu_vdwtype = evdwOclFSWITCH; break;
227 case eintmodPOTSWITCH: *gpu_vdwtype = evdwOclPSWITCH; break;
230 "The requested VdW interaction modifier is not implemented in the GPU "
231 "accelerated kernels!");
234 else if (ic->vdwtype == evdwPME)
236 if (ic->ljpme_comb_rule == ljcrGEOM)
238 *gpu_vdwtype = evdwOclEWALDGEOM;
242 *gpu_vdwtype = evdwOclEWALDLB;
247 gmx_incons("The requested VdW type is not implemented in the GPU accelerated kernels!");
250 if (ic->eeltype == eelCUT)
252 *gpu_eeltype = eelOclCUT;
254 else if (EEL_RF(ic->eeltype))
256 *gpu_eeltype = eelOclRF;
258 else if ((EEL_PME(ic->eeltype) || ic->eeltype == eelEWALD))
260 *gpu_eeltype = nbnxn_gpu_pick_ewald_kernel_type(*ic);
264 /* Shouldn't happen, as this is checked when choosing Verlet-scheme */
266 "The requested electrostatics type is not implemented in the GPU accelerated "
271 /*! \brief Initializes the nonbonded parameter data structure.
273 static void init_nbparam(cl_nbparam_t* nbp,
274 const interaction_const_t* ic,
275 const PairlistParams& listParams,
276 const nbnxn_atomdata_t::Params& nbatParams,
277 const DeviceContext& deviceContext)
281 set_cutoff_parameters(nbp, ic, listParams);
283 map_interaction_types_to_gpu_kernel_flavors(ic, nbatParams.comb_rule, &(nbp->eeltype), &(nbp->vdwtype));
285 if (ic->vdwtype == evdwPME)
287 if (ic->ljpme_comb_rule == ljcrGEOM)
289 GMX_ASSERT(nbatParams.comb_rule == ljcrGEOM, "Combination rule mismatch!");
293 GMX_ASSERT(nbatParams.comb_rule == ljcrLB, "Combination rule mismatch!");
296 /* generate table for PME */
297 nbp->coulomb_tab_climg2d = nullptr;
298 if (nbp->eeltype == eelOclEWALD_TAB || nbp->eeltype == eelOclEWALD_TAB_TWIN)
300 GMX_RELEASE_ASSERT(ic->coulombEwaldTables, "Need valid Coulomb Ewald correction tables");
301 init_ewald_coulomb_force_table(*ic->coulombEwaldTables, nbp, deviceContext);
305 nbp->coulomb_tab_climg2d = clCreateBuffer(deviceContext.context(), CL_MEM_READ_ONLY,
306 sizeof(cl_float), nullptr, &cl_error);
307 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
308 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
311 const int nnbfp = 2 * nbatParams.numTypes * nbatParams.numTypes;
312 const int nnbfp_comb = 2 * nbatParams.numTypes;
315 /* set up LJ parameter lookup table */
316 DeviceBuffer<real> nbfp;
317 initParamLookupTable(&nbfp, nullptr, nbatParams.nbfp.data(), nnbfp, deviceContext);
318 nbp->nbfp_climg2d = nbfp;
320 if (ic->vdwtype == evdwPME)
322 DeviceBuffer<float> nbfp_comb;
323 initParamLookupTable(&nbfp_comb, nullptr, nbatParams.nbfp_comb.data(), nnbfp_comb, deviceContext);
324 nbp->nbfp_comb_climg2d = nbfp_comb;
329 //! This function is documented in the header file
330 void gpu_pme_loadbal_update_param(const nonbonded_verlet_t* nbv, const interaction_const_t* ic)
332 if (!nbv || !nbv->useGpu())
336 NbnxmGpu* nb = nbv->gpu_nbv;
337 cl_nbparam_t* nbp = nb->nbparam;
339 set_cutoff_parameters(nbp, ic, nbv->pairlistSets().params());
341 nbp->eeltype = nbnxn_gpu_pick_ewald_kernel_type(*ic);
343 GMX_RELEASE_ASSERT(ic->coulombEwaldTables, "Need valid Coulomb Ewald correction tables");
344 init_ewald_coulomb_force_table(*ic->coulombEwaldTables, nbp, *nb->deviceContext_);
347 /*! \brief Initializes the pair list data structure.
349 static void init_plist(cl_plist_t* pl)
351 /* initialize to nullptr pointers to data that is not allocated here and will
352 need reallocation in nbnxn_gpu_init_pairlist */
358 /* size -1 indicates that the respective array hasn't been initialized yet */
365 pl->imask_nalloc = -1;
367 pl->excl_nalloc = -1;
368 pl->haveFreshList = false;
371 /*! \brief Initializes the timings data structure.
373 static void init_timings(gmx_wallclock_gpu_nbnxn_t* t)
382 for (i = 0; i < 2; i++)
384 for (j = 0; j < 2; j++)
386 t->ktime[i][j].t = 0.0;
387 t->ktime[i][j].c = 0;
392 t->pruneTime.t = 0.0;
393 t->dynamicPruneTime.c = 0;
394 t->dynamicPruneTime.t = 0.0;
397 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
398 static cl_kernel nbnxn_gpu_create_kernel(NbnxmGpu* nb, const char* kernel_name)
403 kernel = clCreateKernel(nb->dev_rundata->program, kernel_name, &cl_error);
404 if (CL_SUCCESS != cl_error)
406 gmx_fatal(FARGS, "Failed to create kernel '%s' for GPU #%s: OpenCL error %d", kernel_name,
407 nb->deviceContext_->deviceInfo().device_name, cl_error);
413 /*! \brief Clears nonbonded shift force output array and energy outputs on the GPU.
415 static void nbnxn_ocl_clear_e_fshift(NbnxmGpu* nb)
419 cl_atomdata_t* adat = nb->atdat;
420 cl_command_queue ls = nb->deviceStreams[InteractionLocality::Local]->stream();
422 size_t local_work_size[3] = { 1, 1, 1 };
423 size_t global_work_size[3] = { 1, 1, 1 };
425 cl_int shifts = SHIFTS * 3;
429 cl_kernel zero_e_fshift = nb->kernel_zero_e_fshift;
431 local_work_size[0] = 64;
432 // Round the total number of threads up from the array size
433 global_work_size[0] = ((shifts + local_work_size[0] - 1) / local_work_size[0]) * local_work_size[0];
436 cl_error = clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->fshift));
437 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_lj));
438 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_el));
439 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_uint), &shifts);
440 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
442 cl_error = clEnqueueNDRangeKernel(ls, zero_e_fshift, 3, nullptr, global_work_size,
443 local_work_size, 0, nullptr, nullptr);
444 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
447 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
448 static void nbnxn_gpu_init_kernels(NbnxmGpu* nb)
450 /* Init to 0 main kernel arrays */
451 /* They will be later on initialized in select_nbnxn_kernel */
452 // TODO: consider always creating all variants of the kernels here so that there is no
453 // need for late call to clCreateKernel -- if that gives any advantage?
454 memset(nb->kernel_ener_noprune_ptr, 0, sizeof(nb->kernel_ener_noprune_ptr));
455 memset(nb->kernel_ener_prune_ptr, 0, sizeof(nb->kernel_ener_prune_ptr));
456 memset(nb->kernel_noener_noprune_ptr, 0, sizeof(nb->kernel_noener_noprune_ptr));
457 memset(nb->kernel_noener_prune_ptr, 0, sizeof(nb->kernel_noener_prune_ptr));
459 /* Init pruning kernels
461 * TODO: we could avoid creating kernels if dynamic pruning is turned off,
462 * but ATM that depends on force flags not passed into the initialization.
464 nb->kernel_pruneonly[epruneFirst] = nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_opencl");
465 nb->kernel_pruneonly[epruneRolling] =
466 nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_rolling_opencl");
468 /* Init auxiliary kernels */
469 nb->kernel_zero_e_fshift = nbnxn_gpu_create_kernel(nb, "zero_e_fshift");
472 /*! \brief Initializes simulation constant data.
474 * Initializes members of the atomdata and nbparam structs and
475 * clears e/fshift output buffers.
477 static void nbnxn_ocl_init_const(cl_atomdata_t* atomData,
478 cl_nbparam_t* nbParams,
479 const interaction_const_t* ic,
480 const PairlistParams& listParams,
481 const nbnxn_atomdata_t::Params& nbatParams,
482 const DeviceContext& deviceContext)
484 init_atomdata_first(atomData, nbatParams.numTypes, deviceContext);
485 init_nbparam(nbParams, ic, listParams, nbatParams, deviceContext);
489 //! This function is documented in the header file
490 NbnxmGpu* gpu_init(const gmx::DeviceStreamManager& deviceStreamManager,
491 const interaction_const_t* ic,
492 const PairlistParams& listParams,
493 const nbnxn_atomdata_t* nbat,
494 const bool bLocalAndNonlocal)
496 GMX_ASSERT(ic, "Need a valid interaction constants object");
498 auto nb = new NbnxmGpu();
499 nb->deviceContext_ = &deviceStreamManager.context();
501 snew(nb->nbparam, 1);
502 snew(nb->plist[InteractionLocality::Local], 1);
503 if (bLocalAndNonlocal)
505 snew(nb->plist[InteractionLocality::NonLocal], 1);
508 nb->bUseTwoStreams = bLocalAndNonlocal;
510 nb->timers = new cl_timers_t();
511 snew(nb->timings, 1);
513 /* set device info, just point it to the right GPU among the detected ones */
514 nb->dev_rundata = new gmx_device_runtime_data_t();
517 pmalloc(reinterpret_cast<void**>(&nb->nbst.e_lj), sizeof(*nb->nbst.e_lj));
518 pmalloc(reinterpret_cast<void**>(&nb->nbst.e_el), sizeof(*nb->nbst.e_el));
519 pmalloc(reinterpret_cast<void**>(&nb->nbst.fshift), SHIFTS * sizeof(*nb->nbst.fshift));
521 init_plist(nb->plist[InteractionLocality::Local]);
523 /* OpenCL timing disabled if GMX_DISABLE_GPU_TIMING is defined. */
524 nb->bDoTime = (getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
526 /* local/non-local GPU streams */
527 GMX_RELEASE_ASSERT(deviceStreamManager.streamIsValid(gmx::DeviceStreamType::NonBondedLocal),
528 "Local non-bonded stream should be initialized to use GPU for non-bonded.");
529 nb->deviceStreams[InteractionLocality::Local] =
530 &deviceStreamManager.stream(gmx::DeviceStreamType::NonBondedLocal);
532 if (nb->bUseTwoStreams)
534 init_plist(nb->plist[InteractionLocality::NonLocal]);
536 GMX_RELEASE_ASSERT(deviceStreamManager.streamIsValid(gmx::DeviceStreamType::NonBondedNonLocal),
537 "Non-local non-bonded stream should be initialized to use GPU for "
538 "non-bonded with domain decomposition.");
539 nb->deviceStreams[InteractionLocality::NonLocal] =
540 &deviceStreamManager.stream(gmx::DeviceStreamType::NonBondedNonLocal);
545 init_timings(nb->timings);
548 nbnxn_ocl_init_const(nb->atdat, nb->nbparam, ic, listParams, nbat->params(), *nb->deviceContext_);
550 /* Enable LJ param manual prefetch for AMD or Intel or if we request through env. var.
551 * TODO: decide about NVIDIA
553 nb->bPrefetchLjParam = (getenv("GMX_OCL_DISABLE_I_PREFETCH") == nullptr)
554 && ((nb->deviceContext_->deviceInfo().deviceVendor == DeviceVendor::Amd)
555 || (nb->deviceContext_->deviceInfo().deviceVendor == DeviceVendor::Intel)
556 || (getenv("GMX_OCL_ENABLE_I_PREFETCH") != nullptr));
558 /* NOTE: in CUDA we pick L1 cache configuration for the nbnxn kernels here,
559 * but sadly this is not supported in OpenCL (yet?). Consider adding it if
560 * it becomes supported.
562 nbnxn_gpu_compile_kernels(nb);
563 nbnxn_gpu_init_kernels(nb);
565 /* clear energy and shift force outputs */
566 nbnxn_ocl_clear_e_fshift(nb);
570 fprintf(debug, "Initialized OpenCL data structures.\n");
576 /*! \brief Clears the first natoms_clear elements of the GPU nonbonded force output array.
578 static void nbnxn_ocl_clear_f(NbnxmGpu* nb, int natoms_clear)
580 if (natoms_clear == 0)
585 cl_int gmx_used_in_debug cl_error;
587 cl_atomdata_t* atomData = nb->atdat;
588 cl_command_queue ls = nb->deviceStreams[InteractionLocality::Local]->stream();
589 cl_float value = 0.0F;
591 cl_error = clEnqueueFillBuffer(ls, atomData->f, &value, sizeof(cl_float), 0,
592 natoms_clear * sizeof(rvec), 0, nullptr, nullptr);
593 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
594 ("clEnqueueFillBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
597 //! This function is documented in the header file
598 void gpu_clear_outputs(NbnxmGpu* nb, bool computeVirial)
600 nbnxn_ocl_clear_f(nb, nb->atdat->natoms);
601 /* clear shift force array and energies if the outputs were
602 used in the current step */
605 nbnxn_ocl_clear_e_fshift(nb);
608 /* kick off buffer clearing kernel to ensure concurrency with constraints/update */
609 cl_int gmx_unused cl_error;
610 cl_error = clFlush(nb->deviceStreams[InteractionLocality::Local]->stream());
611 GMX_ASSERT(cl_error == CL_SUCCESS, ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
614 //! This function is documented in the header file
615 void gpu_init_pairlist(NbnxmGpu* nb, const NbnxnPairlistGpu* h_plist, const InteractionLocality iloc)
618 // Timing accumulation should happen only if there was work to do
619 // because getLastRangeTime() gets skipped with empty lists later
620 // which leads to the counter not being reset.
621 bool bDoTime = (nb->bDoTime && !h_plist->sci.empty());
622 const DeviceStream& deviceStream = *nb->deviceStreams[iloc];
623 cl_plist_t* d_plist = nb->plist[iloc];
625 if (d_plist->na_c < 0)
627 d_plist->na_c = h_plist->na_ci;
631 if (d_plist->na_c != h_plist->na_ci)
633 sprintf(sbuf, "In cu_init_plist: the #atoms per cell has changed (from %d to %d)",
634 d_plist->na_c, h_plist->na_ci);
639 gpu_timers_t::Interaction& iTimers = nb->timers->interaction[iloc];
643 iTimers.pl_h2d.openTimingRegion(deviceStream);
644 iTimers.didPairlistH2D = true;
647 // TODO most of this function is same in CUDA and OpenCL, move into the header
648 const DeviceContext& deviceContext = *nb->deviceContext_;
650 reallocateDeviceBuffer(&d_plist->sci, h_plist->sci.size(), &d_plist->nsci, &d_plist->sci_nalloc,
652 copyToDeviceBuffer(&d_plist->sci, h_plist->sci.data(), 0, h_plist->sci.size(), deviceStream,
653 GpuApiCallBehavior::Async, bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
655 reallocateDeviceBuffer(&d_plist->cj4, h_plist->cj4.size(), &d_plist->ncj4, &d_plist->cj4_nalloc,
657 copyToDeviceBuffer(&d_plist->cj4, h_plist->cj4.data(), 0, h_plist->cj4.size(), deviceStream,
658 GpuApiCallBehavior::Async, bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
660 reallocateDeviceBuffer(&d_plist->imask, h_plist->cj4.size() * c_nbnxnGpuClusterpairSplit,
661 &d_plist->nimask, &d_plist->imask_nalloc, deviceContext);
663 reallocateDeviceBuffer(&d_plist->excl, h_plist->excl.size(), &d_plist->nexcl,
664 &d_plist->excl_nalloc, deviceContext);
665 copyToDeviceBuffer(&d_plist->excl, h_plist->excl.data(), 0, h_plist->excl.size(), deviceStream,
666 GpuApiCallBehavior::Async, bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
670 iTimers.pl_h2d.closeTimingRegion(deviceStream);
673 /* need to prune the pair list during the next step */
674 d_plist->haveFreshList = true;
677 //! This function is documented in the header file
678 void gpu_upload_shiftvec(NbnxmGpu* nb, const nbnxn_atomdata_t* nbatom)
680 cl_atomdata_t* adat = nb->atdat;
681 cl_command_queue ls = nb->deviceStreams[InteractionLocality::Local]->stream();
683 /* only if we have a dynamic box */
684 if (nbatom->bDynamicBox || !adat->bShiftVecUploaded)
686 ocl_copy_H2D_async(adat->shift_vec, nbatom->shift_vec.data(), 0,
687 SHIFTS * sizeof(nbatom->shift_vec[0]), ls, nullptr);
688 adat->bShiftVecUploaded = CL_TRUE;
692 //! This function is documented in the header file
693 void gpu_init_atomdata(NbnxmGpu* nb, const nbnxn_atomdata_t* nbat)
698 bool bDoTime = nb->bDoTime;
699 cl_timers_t* timers = nb->timers;
700 cl_atomdata_t* d_atdat = nb->atdat;
701 const DeviceStream& deviceStream = *nb->deviceStreams[InteractionLocality::Local];
703 natoms = nbat->numAtoms();
708 /* time async copy */
709 timers->atdat.openTimingRegion(deviceStream);
712 /* need to reallocate if we have to copy more atoms than the amount of space
713 available and only allocate if we haven't initialized yet, i.e d_atdat->natoms == -1 */
714 if (natoms > d_atdat->nalloc)
716 nalloc = over_alloc_small(natoms);
718 /* free up first if the arrays have already been initialized */
719 if (d_atdat->nalloc != -1)
721 freeDeviceBuffer(&d_atdat->f);
722 freeDeviceBuffer(&d_atdat->xq);
723 freeDeviceBuffer(&d_atdat->lj_comb);
724 freeDeviceBuffer(&d_atdat->atom_types);
727 d_atdat->f = clCreateBuffer(nb->deviceContext_->context(), CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
728 nalloc * DIM * sizeof(nbat->out[0].f[0]), nullptr, &cl_error);
729 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
730 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
732 d_atdat->xq = clCreateBuffer(nb->deviceContext_->context(), CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
733 nalloc * sizeof(cl_float4), nullptr, &cl_error);
734 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
735 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
737 if (useLjCombRule(nb->nbparam->vdwtype))
739 d_atdat->lj_comb = clCreateBuffer(nb->deviceContext_->context(),
740 CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
741 nalloc * sizeof(cl_float2), nullptr, &cl_error);
742 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
743 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
747 d_atdat->atom_types = clCreateBuffer(nb->deviceContext_->context(),
748 CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
749 nalloc * sizeof(int), nullptr, &cl_error);
750 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
751 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
754 d_atdat->nalloc = nalloc;
758 d_atdat->natoms = natoms;
759 d_atdat->natoms_local = nbat->natoms_local;
761 /* need to clear GPU f output if realloc happened */
764 nbnxn_ocl_clear_f(nb, nalloc);
767 if (useLjCombRule(nb->nbparam->vdwtype))
769 ocl_copy_H2D_async(d_atdat->lj_comb, nbat->params().lj_comb.data(), 0, natoms * sizeof(cl_float2),
770 deviceStream.stream(), bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
774 ocl_copy_H2D_async(d_atdat->atom_types, nbat->params().type.data(), 0, natoms * sizeof(int),
775 deviceStream.stream(), bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
780 timers->atdat.closeTimingRegion(deviceStream);
783 /* kick off the tasks enqueued above to ensure concurrency with the search */
784 cl_error = clFlush(deviceStream.stream());
785 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
786 ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
789 /*! \brief Releases an OpenCL kernel pointer */
790 static void free_kernel(cl_kernel* kernel_ptr)
792 cl_int gmx_unused cl_error;
794 GMX_ASSERT(kernel_ptr, "Need a valid kernel pointer");
798 cl_error = clReleaseKernel(*kernel_ptr);
799 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
800 ("clReleaseKernel failed: " + ocl_get_error_string(cl_error)).c_str());
802 *kernel_ptr = nullptr;
806 /*! \brief Releases a list of OpenCL kernel pointers */
807 static void free_kernels(cl_kernel* kernels, int count)
811 for (i = 0; i < count; i++)
813 free_kernel(kernels + i);
817 /*! \brief Free the OpenCL program.
819 * The function releases the OpenCL program assuciated with the
820 * device that the calling PP rank is running on.
822 * \param program [in] OpenCL program to release.
824 static void freeGpuProgram(cl_program program)
828 cl_int cl_error = clReleaseProgram(program);
829 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
830 ("clReleaseProgram failed: " + ocl_get_error_string(cl_error)).c_str());
835 //! This function is documented in the header file
836 void gpu_free(NbnxmGpu* nb)
844 int kernel_count = sizeof(nb->kernel_ener_noprune_ptr) / sizeof(nb->kernel_ener_noprune_ptr[0][0]);
845 free_kernels(nb->kernel_ener_noprune_ptr[0], kernel_count);
847 kernel_count = sizeof(nb->kernel_ener_prune_ptr) / sizeof(nb->kernel_ener_prune_ptr[0][0]);
848 free_kernels(nb->kernel_ener_prune_ptr[0], kernel_count);
850 kernel_count = sizeof(nb->kernel_noener_noprune_ptr) / sizeof(nb->kernel_noener_noprune_ptr[0][0]);
851 free_kernels(nb->kernel_noener_noprune_ptr[0], kernel_count);
853 kernel_count = sizeof(nb->kernel_noener_prune_ptr) / sizeof(nb->kernel_noener_prune_ptr[0][0]);
854 free_kernels(nb->kernel_noener_prune_ptr[0], kernel_count);
856 free_kernel(&(nb->kernel_zero_e_fshift));
859 freeDeviceBuffer(&(nb->atdat->xq));
860 freeDeviceBuffer(&(nb->atdat->f));
861 freeDeviceBuffer(&(nb->atdat->e_lj));
862 freeDeviceBuffer(&(nb->atdat->e_el));
863 freeDeviceBuffer(&(nb->atdat->fshift));
864 freeDeviceBuffer(&(nb->atdat->lj_comb));
865 freeDeviceBuffer(&(nb->atdat->atom_types));
866 freeDeviceBuffer(&(nb->atdat->shift_vec));
870 freeDeviceBuffer(&(nb->nbparam->nbfp_climg2d));
871 freeDeviceBuffer(&(nb->nbparam->nbfp_comb_climg2d));
872 freeDeviceBuffer(&(nb->nbparam->coulomb_tab_climg2d));
876 auto* plist = nb->plist[InteractionLocality::Local];
877 freeDeviceBuffer(&plist->sci);
878 freeDeviceBuffer(&plist->cj4);
879 freeDeviceBuffer(&plist->imask);
880 freeDeviceBuffer(&plist->excl);
882 if (nb->bUseTwoStreams)
884 auto* plist_nl = nb->plist[InteractionLocality::NonLocal];
885 freeDeviceBuffer(&plist_nl->sci);
886 freeDeviceBuffer(&plist_nl->cj4);
887 freeDeviceBuffer(&plist_nl->imask);
888 freeDeviceBuffer(&plist_nl->excl);
893 pfree(nb->nbst.e_lj);
894 nb->nbst.e_lj = nullptr;
896 pfree(nb->nbst.e_el);
897 nb->nbst.e_el = nullptr;
899 pfree(nb->nbst.fshift);
900 nb->nbst.fshift = nullptr;
902 /* Free other events */
903 if (nb->nonlocal_done)
905 clReleaseEvent(nb->nonlocal_done);
906 nb->nonlocal_done = nullptr;
908 if (nb->misc_ops_and_local_H2D_done)
910 clReleaseEvent(nb->misc_ops_and_local_H2D_done);
911 nb->misc_ops_and_local_H2D_done = nullptr;
914 freeGpuProgram(nb->dev_rundata->program);
915 delete nb->dev_rundata;
917 /* Free timers and timings */
924 fprintf(debug, "Cleaned up OpenCL data structures.\n");
928 //! This function is documented in the header file
929 gmx_wallclock_gpu_nbnxn_t* gpu_get_timings(NbnxmGpu* nb)
931 return (nb != nullptr && nb->bDoTime) ? nb->timings : nullptr;
934 //! This function is documented in the header file
935 void gpu_reset_timings(nonbonded_verlet_t* nbv)
937 if (nbv->gpu_nbv && nbv->gpu_nbv->bDoTime)
939 init_timings(nbv->gpu_nbv->timings);
943 //! This function is documented in the header file
944 int gpu_min_ci_balanced(NbnxmGpu* nb)
946 return nb != nullptr ? gpu_min_ci_balanced_factor * nb->deviceContext_->deviceInfo().compute_units : 0;
949 //! This function is documented in the header file
950 gmx_bool gpu_is_kernel_ewald_analytical(const NbnxmGpu* nb)
952 return ((nb->nbparam->eeltype == eelOclEWALD_ANA) || (nb->nbparam->eeltype == eelOclEWALD_ANA_TWIN));