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)
125 if (nbp->coulomb_tab_climg2d != nullptr)
127 freeDeviceBuffer(&(nbp->coulomb_tab_climg2d));
130 /* Switched from using textures to using buffers */
131 // TODO: decide which alternative is most efficient - textures or buffers.
133 cl_image_format array_format;
135 array_format.image_channel_data_type = CL_FLOAT;
136 array_format.image_channel_order = CL_R;
138 coul_tab = clCreateImage2D(deviceContext.context(), CL_MEM_READ_WRITE |
139 CL_MEM_COPY_HOST_PTR, &array_format, tabsize, 1, 0, ftmp, &cl_error);
142 coul_tab = clCreateBuffer(deviceContext.context(),
143 CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
144 tables.tableF.size() * sizeof(cl_float),
145 const_cast<real*>(tables.tableF.data()), &cl_error);
146 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
147 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
149 nbp->coulomb_tab_climg2d = coul_tab;
150 nbp->coulomb_tab_scale = tables.scale;
154 /*! \brief Initializes the atomdata structure first time, it only gets filled at
157 static void init_atomdata_first(cl_atomdata_t* ad, int ntypes, const DeviceContext& deviceContext)
163 ad->shift_vec = clCreateBuffer(deviceContext.context(), CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
164 SHIFTS * sizeof(nbnxn_atomdata_t::shift_vec[0]), nullptr, &cl_error);
165 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
166 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
167 ad->bShiftVecUploaded = CL_FALSE;
169 ad->fshift = clCreateBuffer(deviceContext.context(), CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
170 SHIFTS * sizeof(nb_staging_t::fshift[0]), nullptr, &cl_error);
171 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
172 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
174 ad->e_lj = clCreateBuffer(deviceContext.context(), CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
175 sizeof(float), nullptr, &cl_error);
176 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
177 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
179 ad->e_el = clCreateBuffer(deviceContext.context(), CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
180 sizeof(float), nullptr, &cl_error);
181 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
182 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
184 /* initialize to nullptr pointers to data that is not allocated here and will
185 need reallocation in nbnxn_gpu_init_atomdata */
189 /* size -1 indicates that the respective array hasn't been initialized yet */
194 /*! \brief Copies all parameters related to the cut-off from ic to nbp
196 static void set_cutoff_parameters(cl_nbparam_t* nbp, const interaction_const_t* ic, const PairlistParams& listParams)
198 nbp->ewald_beta = ic->ewaldcoeff_q;
199 nbp->sh_ewald = ic->sh_ewald;
200 nbp->epsfac = ic->epsfac;
201 nbp->two_k_rf = 2.0 * ic->k_rf;
202 nbp->c_rf = ic->c_rf;
203 nbp->rvdw_sq = ic->rvdw * ic->rvdw;
204 nbp->rcoulomb_sq = ic->rcoulomb * ic->rcoulomb;
205 nbp->rlistOuter_sq = listParams.rlistOuter * listParams.rlistOuter;
206 nbp->rlistInner_sq = listParams.rlistInner * listParams.rlistInner;
207 nbp->useDynamicPruning = listParams.useDynamicPruning;
209 nbp->sh_lj_ewald = ic->sh_lj_ewald;
210 nbp->ewaldcoeff_lj = ic->ewaldcoeff_lj;
212 nbp->rvdw_switch = ic->rvdw_switch;
213 nbp->dispersion_shift = ic->dispersion_shift;
214 nbp->repulsion_shift = ic->repulsion_shift;
215 nbp->vdw_switch = ic->vdw_switch;
218 /*! \brief Returns the kinds of electrostatics and Vdw OpenCL
219 * kernels that will be used.
221 * Respectively, these values are from enum eelOcl and enum
223 static void map_interaction_types_to_gpu_kernel_flavors(const interaction_const_t* ic,
228 if (ic->vdwtype == evdwCUT)
230 switch (ic->vdw_modifier)
233 case eintmodPOTSHIFT:
236 case ljcrNONE: *gpu_vdwtype = evdwOclCUT; break;
237 case ljcrGEOM: *gpu_vdwtype = evdwOclCUTCOMBGEOM; break;
238 case ljcrLB: *gpu_vdwtype = evdwOclCUTCOMBLB; break;
241 "The requested LJ combination rule is not implemented in the "
242 "OpenCL GPU accelerated kernels!");
245 case eintmodFORCESWITCH: *gpu_vdwtype = evdwOclFSWITCH; break;
246 case eintmodPOTSWITCH: *gpu_vdwtype = evdwOclPSWITCH; break;
249 "The requested VdW interaction modifier is not implemented in the GPU "
250 "accelerated kernels!");
253 else if (ic->vdwtype == evdwPME)
255 if (ic->ljpme_comb_rule == ljcrGEOM)
257 *gpu_vdwtype = evdwOclEWALDGEOM;
261 *gpu_vdwtype = evdwOclEWALDLB;
266 gmx_incons("The requested VdW type is not implemented in the GPU accelerated kernels!");
269 if (ic->eeltype == eelCUT)
271 *gpu_eeltype = eelOclCUT;
273 else if (EEL_RF(ic->eeltype))
275 *gpu_eeltype = eelOclRF;
277 else if ((EEL_PME(ic->eeltype) || ic->eeltype == eelEWALD))
279 *gpu_eeltype = nbnxn_gpu_pick_ewald_kernel_type(*ic);
283 /* Shouldn't happen, as this is checked when choosing Verlet-scheme */
285 "The requested electrostatics type is not implemented in the GPU accelerated "
290 /*! \brief Initializes the nonbonded parameter data structure.
292 static void init_nbparam(cl_nbparam_t* nbp,
293 const interaction_const_t* ic,
294 const PairlistParams& listParams,
295 const nbnxn_atomdata_t::Params& nbatParams,
296 const DeviceContext& deviceContext)
300 set_cutoff_parameters(nbp, ic, listParams);
302 map_interaction_types_to_gpu_kernel_flavors(ic, nbatParams.comb_rule, &(nbp->eeltype), &(nbp->vdwtype));
304 if (ic->vdwtype == evdwPME)
306 if (ic->ljpme_comb_rule == ljcrGEOM)
308 GMX_ASSERT(nbatParams.comb_rule == ljcrGEOM, "Combination rule mismatch!");
312 GMX_ASSERT(nbatParams.comb_rule == ljcrLB, "Combination rule mismatch!");
315 /* generate table for PME */
316 nbp->coulomb_tab_climg2d = nullptr;
317 if (nbp->eeltype == eelOclEWALD_TAB || nbp->eeltype == eelOclEWALD_TAB_TWIN)
319 GMX_RELEASE_ASSERT(ic->coulombEwaldTables, "Need valid Coulomb Ewald correction tables");
320 init_ewald_coulomb_force_table(*ic->coulombEwaldTables, nbp, deviceContext);
323 // TODO: improvement needed.
324 // The image2d is created here even if eeltype is not eelCuEWALD_TAB or eelCuEWALD_TAB_TWIN
325 // because the OpenCL kernels don't accept nullptr values for image2D parameters.
327 /* Switched from using textures to using buffers */
328 // TODO: decide which alternative is most efficient - textures or buffers.
330 cl_image_format array_format;
332 array_format.image_channel_data_type = CL_FLOAT;
333 array_format.image_channel_order = CL_R;
335 nbp->coulomb_tab_climg2d = clCreateImage2D(deviceContext.context(),
336 CL_MEM_READ_WRITE, &array_format, 1, 1, 0, nullptr, &cl_error);
339 nbp->coulomb_tab_climg2d = clCreateBuffer(deviceContext.context(), CL_MEM_READ_ONLY,
340 sizeof(cl_float), nullptr, &cl_error);
341 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
342 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
345 const int nnbfp = 2 * nbatParams.numTypes * nbatParams.numTypes;
346 const int nnbfp_comb = 2 * nbatParams.numTypes;
349 /* Switched from using textures to using buffers */
350 // TODO: decide which alternative is most efficient - textures or buffers.
352 cl_image_format array_format;
354 array_format.image_channel_data_type = CL_FLOAT;
355 array_format.image_channel_order = CL_R;
357 nbp->nbfp_climg2d = clCreateImage2D(deviceContext.context(), CL_MEM_READ_ONLY |
358 CL_MEM_COPY_HOST_PTR, &array_format, nnbfp, 1, 0, nbat->nbfp, &cl_error);
361 nbp->nbfp_climg2d = clCreateBuffer(
362 deviceContext.context(), CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
363 nnbfp * sizeof(cl_float), const_cast<float*>(nbatParams.nbfp.data()), &cl_error);
364 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
365 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
367 if (ic->vdwtype == evdwPME)
369 /* Switched from using textures to using buffers */
370 // TODO: decide which alternative is most efficient - textures or buffers.
371 /* nbp->nbfp_comb_climg2d = clCreateImage2D(deviceContext.context(), CL_MEM_READ_WRITE |
372 CL_MEM_COPY_HOST_PTR, &array_format, nnbfp_comb, 1, 0, nbat->nbfp_comb, &cl_error);*/
373 nbp->nbfp_comb_climg2d =
374 clCreateBuffer(deviceContext.context(),
375 CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
376 nnbfp_comb * sizeof(cl_float),
377 const_cast<float*>(nbatParams.nbfp_comb.data()), &cl_error);
378 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
379 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
383 // TODO: improvement needed.
384 // The image2d is created here even if vdwtype is not evdwPME because the OpenCL kernels
385 // don't accept nullptr values for image2D parameters.
386 /* Switched from using textures to using buffers */
387 // TODO: decide which alternative is most efficient - textures or buffers.
388 /* nbp->nbfp_comb_climg2d = clCreateImage2D(deviceContext.context(),
389 CL_MEM_READ_WRITE, &array_format, 1, 1, 0, nullptr, &cl_error);*/
390 nbp->nbfp_comb_climg2d = clCreateBuffer(deviceContext.context(), CL_MEM_READ_ONLY,
391 sizeof(cl_float), nullptr, &cl_error);
392 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
393 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
398 //! This function is documented in the header file
399 void gpu_pme_loadbal_update_param(const nonbonded_verlet_t* nbv, const interaction_const_t* ic)
401 if (!nbv || !nbv->useGpu())
405 NbnxmGpu* nb = nbv->gpu_nbv;
406 cl_nbparam_t* nbp = nb->nbparam;
408 set_cutoff_parameters(nbp, ic, nbv->pairlistSets().params());
410 nbp->eeltype = nbnxn_gpu_pick_ewald_kernel_type(*ic);
412 GMX_RELEASE_ASSERT(ic->coulombEwaldTables, "Need valid Coulomb Ewald correction tables");
413 init_ewald_coulomb_force_table(*ic->coulombEwaldTables, nbp, *nb->deviceContext_);
416 /*! \brief Initializes the pair list data structure.
418 static void init_plist(cl_plist_t* pl)
420 /* initialize to nullptr pointers to data that is not allocated here and will
421 need reallocation in nbnxn_gpu_init_pairlist */
427 /* size -1 indicates that the respective array hasn't been initialized yet */
434 pl->imask_nalloc = -1;
436 pl->excl_nalloc = -1;
437 pl->haveFreshList = false;
440 /*! \brief Initializes the timings data structure.
442 static void init_timings(gmx_wallclock_gpu_nbnxn_t* t)
451 for (i = 0; i < 2; i++)
453 for (j = 0; j < 2; j++)
455 t->ktime[i][j].t = 0.0;
456 t->ktime[i][j].c = 0;
461 t->pruneTime.t = 0.0;
462 t->dynamicPruneTime.c = 0;
463 t->dynamicPruneTime.t = 0.0;
466 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
467 static cl_kernel nbnxn_gpu_create_kernel(NbnxmGpu* nb, const char* kernel_name)
472 kernel = clCreateKernel(nb->dev_rundata->program, kernel_name, &cl_error);
473 if (CL_SUCCESS != cl_error)
475 gmx_fatal(FARGS, "Failed to create kernel '%s' for GPU #%s: OpenCL error %d", kernel_name,
476 nb->deviceContext_->deviceInfo().device_name, cl_error);
482 /*! \brief Clears nonbonded shift force output array and energy outputs on the GPU.
484 static void nbnxn_ocl_clear_e_fshift(NbnxmGpu* nb)
488 cl_atomdata_t* adat = nb->atdat;
489 cl_command_queue ls = nb->deviceStreams[InteractionLocality::Local]->stream();
491 size_t local_work_size[3] = { 1, 1, 1 };
492 size_t global_work_size[3] = { 1, 1, 1 };
494 cl_int shifts = SHIFTS * 3;
498 cl_kernel zero_e_fshift = nb->kernel_zero_e_fshift;
500 local_work_size[0] = 64;
501 // Round the total number of threads up from the array size
502 global_work_size[0] = ((shifts + local_work_size[0] - 1) / local_work_size[0]) * local_work_size[0];
505 cl_error = clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->fshift));
506 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_lj));
507 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_el));
508 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_uint), &shifts);
509 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
511 cl_error = clEnqueueNDRangeKernel(ls, zero_e_fshift, 3, nullptr, global_work_size,
512 local_work_size, 0, nullptr, nullptr);
513 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
516 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
517 static void nbnxn_gpu_init_kernels(NbnxmGpu* nb)
519 /* Init to 0 main kernel arrays */
520 /* They will be later on initialized in select_nbnxn_kernel */
521 // TODO: consider always creating all variants of the kernels here so that there is no
522 // need for late call to clCreateKernel -- if that gives any advantage?
523 memset(nb->kernel_ener_noprune_ptr, 0, sizeof(nb->kernel_ener_noprune_ptr));
524 memset(nb->kernel_ener_prune_ptr, 0, sizeof(nb->kernel_ener_prune_ptr));
525 memset(nb->kernel_noener_noprune_ptr, 0, sizeof(nb->kernel_noener_noprune_ptr));
526 memset(nb->kernel_noener_prune_ptr, 0, sizeof(nb->kernel_noener_prune_ptr));
528 /* Init pruning kernels
530 * TODO: we could avoid creating kernels if dynamic pruning is turned off,
531 * but ATM that depends on force flags not passed into the initialization.
533 nb->kernel_pruneonly[epruneFirst] = nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_opencl");
534 nb->kernel_pruneonly[epruneRolling] =
535 nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_rolling_opencl");
537 /* Init auxiliary kernels */
538 nb->kernel_zero_e_fshift = nbnxn_gpu_create_kernel(nb, "zero_e_fshift");
541 /*! \brief Initializes simulation constant data.
543 * Initializes members of the atomdata and nbparam structs and
544 * clears e/fshift output buffers.
546 static void nbnxn_ocl_init_const(cl_atomdata_t* atomData,
547 cl_nbparam_t* nbParams,
548 const interaction_const_t* ic,
549 const PairlistParams& listParams,
550 const nbnxn_atomdata_t::Params& nbatParams,
551 const DeviceContext& deviceContext)
553 init_atomdata_first(atomData, nbatParams.numTypes, deviceContext);
554 init_nbparam(nbParams, ic, listParams, nbatParams, deviceContext);
558 //! This function is documented in the header file
559 NbnxmGpu* gpu_init(const gmx::DeviceStreamManager& deviceStreamManager,
560 const interaction_const_t* ic,
561 const PairlistParams& listParams,
562 const nbnxn_atomdata_t* nbat,
563 const bool bLocalAndNonlocal)
565 GMX_ASSERT(ic, "Need a valid interaction constants object");
567 auto nb = new NbnxmGpu();
568 nb->deviceContext_ = &deviceStreamManager.context();
570 snew(nb->nbparam, 1);
571 snew(nb->plist[InteractionLocality::Local], 1);
572 if (bLocalAndNonlocal)
574 snew(nb->plist[InteractionLocality::NonLocal], 1);
577 nb->bUseTwoStreams = bLocalAndNonlocal;
579 nb->timers = new cl_timers_t();
580 snew(nb->timings, 1);
582 /* set device info, just point it to the right GPU among the detected ones */
583 nb->dev_rundata = new gmx_device_runtime_data_t();
586 pmalloc(reinterpret_cast<void**>(&nb->nbst.e_lj), sizeof(*nb->nbst.e_lj));
587 pmalloc(reinterpret_cast<void**>(&nb->nbst.e_el), sizeof(*nb->nbst.e_el));
588 pmalloc(reinterpret_cast<void**>(&nb->nbst.fshift), SHIFTS * sizeof(*nb->nbst.fshift));
590 init_plist(nb->plist[InteractionLocality::Local]);
592 /* OpenCL timing disabled if GMX_DISABLE_GPU_TIMING is defined. */
593 nb->bDoTime = (getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
595 /* local/non-local GPU streams */
596 GMX_RELEASE_ASSERT(deviceStreamManager.streamIsValid(gmx::DeviceStreamType::NonBondedLocal),
597 "Local non-bonded stream should be initialized to use GPU for non-bonded.");
598 nb->deviceStreams[InteractionLocality::Local] =
599 &deviceStreamManager.stream(gmx::DeviceStreamType::NonBondedLocal);
601 if (nb->bUseTwoStreams)
603 init_plist(nb->plist[InteractionLocality::NonLocal]);
605 GMX_RELEASE_ASSERT(deviceStreamManager.streamIsValid(gmx::DeviceStreamType::NonBondedNonLocal),
606 "Non-local non-bonded stream should be initialized to use GPU for "
607 "non-bonded with domain decomposition.");
608 nb->deviceStreams[InteractionLocality::NonLocal] =
609 &deviceStreamManager.stream(gmx::DeviceStreamType::NonBondedNonLocal);
614 init_timings(nb->timings);
617 nbnxn_ocl_init_const(nb->atdat, nb->nbparam, ic, listParams, nbat->params(), *nb->deviceContext_);
619 /* Enable LJ param manual prefetch for AMD or Intel or if we request through env. var.
620 * TODO: decide about NVIDIA
622 nb->bPrefetchLjParam = (getenv("GMX_OCL_DISABLE_I_PREFETCH") == nullptr)
623 && ((nb->deviceContext_->deviceInfo().deviceVendor == DeviceVendor::Amd)
624 || (nb->deviceContext_->deviceInfo().deviceVendor == DeviceVendor::Intel)
625 || (getenv("GMX_OCL_ENABLE_I_PREFETCH") != nullptr));
627 /* NOTE: in CUDA we pick L1 cache configuration for the nbnxn kernels here,
628 * but sadly this is not supported in OpenCL (yet?). Consider adding it if
629 * it becomes supported.
631 nbnxn_gpu_compile_kernels(nb);
632 nbnxn_gpu_init_kernels(nb);
634 /* clear energy and shift force outputs */
635 nbnxn_ocl_clear_e_fshift(nb);
639 fprintf(debug, "Initialized OpenCL data structures.\n");
645 /*! \brief Clears the first natoms_clear elements of the GPU nonbonded force output array.
647 static void nbnxn_ocl_clear_f(NbnxmGpu* nb, int natoms_clear)
649 if (natoms_clear == 0)
654 cl_int gmx_used_in_debug cl_error;
656 cl_atomdata_t* atomData = nb->atdat;
657 cl_command_queue ls = nb->deviceStreams[InteractionLocality::Local]->stream();
658 cl_float value = 0.0F;
660 cl_error = clEnqueueFillBuffer(ls, atomData->f, &value, sizeof(cl_float), 0,
661 natoms_clear * sizeof(rvec), 0, nullptr, nullptr);
662 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
663 ("clEnqueueFillBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
666 //! This function is documented in the header file
667 void gpu_clear_outputs(NbnxmGpu* nb, bool computeVirial)
669 nbnxn_ocl_clear_f(nb, nb->atdat->natoms);
670 /* clear shift force array and energies if the outputs were
671 used in the current step */
674 nbnxn_ocl_clear_e_fshift(nb);
677 /* kick off buffer clearing kernel to ensure concurrency with constraints/update */
678 cl_int gmx_unused cl_error;
679 cl_error = clFlush(nb->deviceStreams[InteractionLocality::Local]->stream());
680 GMX_ASSERT(cl_error == CL_SUCCESS, ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
683 //! This function is documented in the header file
684 void gpu_init_pairlist(NbnxmGpu* nb, const NbnxnPairlistGpu* h_plist, const InteractionLocality iloc)
687 // Timing accumulation should happen only if there was work to do
688 // because getLastRangeTime() gets skipped with empty lists later
689 // which leads to the counter not being reset.
690 bool bDoTime = (nb->bDoTime && !h_plist->sci.empty());
691 const DeviceStream& deviceStream = *nb->deviceStreams[iloc];
692 cl_plist_t* d_plist = nb->plist[iloc];
694 if (d_plist->na_c < 0)
696 d_plist->na_c = h_plist->na_ci;
700 if (d_plist->na_c != h_plist->na_ci)
702 sprintf(sbuf, "In cu_init_plist: the #atoms per cell has changed (from %d to %d)",
703 d_plist->na_c, h_plist->na_ci);
708 gpu_timers_t::Interaction& iTimers = nb->timers->interaction[iloc];
712 iTimers.pl_h2d.openTimingRegion(deviceStream);
713 iTimers.didPairlistH2D = true;
716 // TODO most of this function is same in CUDA and OpenCL, move into the header
717 const DeviceContext& deviceContext = *nb->deviceContext_;
719 reallocateDeviceBuffer(&d_plist->sci, h_plist->sci.size(), &d_plist->nsci, &d_plist->sci_nalloc,
721 copyToDeviceBuffer(&d_plist->sci, h_plist->sci.data(), 0, h_plist->sci.size(), deviceStream,
722 GpuApiCallBehavior::Async, bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
724 reallocateDeviceBuffer(&d_plist->cj4, h_plist->cj4.size(), &d_plist->ncj4, &d_plist->cj4_nalloc,
726 copyToDeviceBuffer(&d_plist->cj4, h_plist->cj4.data(), 0, h_plist->cj4.size(), deviceStream,
727 GpuApiCallBehavior::Async, bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
729 reallocateDeviceBuffer(&d_plist->imask, h_plist->cj4.size() * c_nbnxnGpuClusterpairSplit,
730 &d_plist->nimask, &d_plist->imask_nalloc, deviceContext);
732 reallocateDeviceBuffer(&d_plist->excl, h_plist->excl.size(), &d_plist->nexcl,
733 &d_plist->excl_nalloc, deviceContext);
734 copyToDeviceBuffer(&d_plist->excl, h_plist->excl.data(), 0, h_plist->excl.size(), deviceStream,
735 GpuApiCallBehavior::Async, bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
739 iTimers.pl_h2d.closeTimingRegion(deviceStream);
742 /* need to prune the pair list during the next step */
743 d_plist->haveFreshList = true;
746 //! This function is documented in the header file
747 void gpu_upload_shiftvec(NbnxmGpu* nb, const nbnxn_atomdata_t* nbatom)
749 cl_atomdata_t* adat = nb->atdat;
750 cl_command_queue ls = nb->deviceStreams[InteractionLocality::Local]->stream();
752 /* only if we have a dynamic box */
753 if (nbatom->bDynamicBox || !adat->bShiftVecUploaded)
755 ocl_copy_H2D_async(adat->shift_vec, nbatom->shift_vec.data(), 0,
756 SHIFTS * sizeof(nbatom->shift_vec[0]), ls, nullptr);
757 adat->bShiftVecUploaded = CL_TRUE;
761 //! This function is documented in the header file
762 void gpu_init_atomdata(NbnxmGpu* nb, const nbnxn_atomdata_t* nbat)
767 bool bDoTime = nb->bDoTime;
768 cl_timers_t* timers = nb->timers;
769 cl_atomdata_t* d_atdat = nb->atdat;
770 const DeviceStream& deviceStream = *nb->deviceStreams[InteractionLocality::Local];
772 natoms = nbat->numAtoms();
777 /* time async copy */
778 timers->atdat.openTimingRegion(deviceStream);
781 /* need to reallocate if we have to copy more atoms than the amount of space
782 available and only allocate if we haven't initialized yet, i.e d_atdat->natoms == -1 */
783 if (natoms > d_atdat->nalloc)
785 nalloc = over_alloc_small(natoms);
787 /* free up first if the arrays have already been initialized */
788 if (d_atdat->nalloc != -1)
790 freeDeviceBuffer(&d_atdat->f);
791 freeDeviceBuffer(&d_atdat->xq);
792 freeDeviceBuffer(&d_atdat->lj_comb);
793 freeDeviceBuffer(&d_atdat->atom_types);
796 d_atdat->f = clCreateBuffer(nb->deviceContext_->context(), CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
797 nalloc * DIM * sizeof(nbat->out[0].f[0]), nullptr, &cl_error);
798 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
799 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
801 d_atdat->xq = clCreateBuffer(nb->deviceContext_->context(), CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
802 nalloc * sizeof(cl_float4), nullptr, &cl_error);
803 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
804 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
806 if (useLjCombRule(nb->nbparam->vdwtype))
808 d_atdat->lj_comb = clCreateBuffer(nb->deviceContext_->context(),
809 CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
810 nalloc * sizeof(cl_float2), nullptr, &cl_error);
811 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
812 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
816 d_atdat->atom_types = clCreateBuffer(nb->deviceContext_->context(),
817 CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
818 nalloc * sizeof(int), nullptr, &cl_error);
819 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
820 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
823 d_atdat->nalloc = nalloc;
827 d_atdat->natoms = natoms;
828 d_atdat->natoms_local = nbat->natoms_local;
830 /* need to clear GPU f output if realloc happened */
833 nbnxn_ocl_clear_f(nb, nalloc);
836 if (useLjCombRule(nb->nbparam->vdwtype))
838 ocl_copy_H2D_async(d_atdat->lj_comb, nbat->params().lj_comb.data(), 0, natoms * sizeof(cl_float2),
839 deviceStream.stream(), bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
843 ocl_copy_H2D_async(d_atdat->atom_types, nbat->params().type.data(), 0, natoms * sizeof(int),
844 deviceStream.stream(), bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
849 timers->atdat.closeTimingRegion(deviceStream);
852 /* kick off the tasks enqueued above to ensure concurrency with the search */
853 cl_error = clFlush(deviceStream.stream());
854 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
855 ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
858 /*! \brief Releases an OpenCL kernel pointer */
859 static void free_kernel(cl_kernel* kernel_ptr)
861 cl_int gmx_unused cl_error;
863 GMX_ASSERT(kernel_ptr, "Need a valid kernel pointer");
867 cl_error = clReleaseKernel(*kernel_ptr);
868 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
869 ("clReleaseKernel failed: " + ocl_get_error_string(cl_error)).c_str());
871 *kernel_ptr = nullptr;
875 /*! \brief Releases a list of OpenCL kernel pointers */
876 static void free_kernels(cl_kernel* kernels, int count)
880 for (i = 0; i < count; i++)
882 free_kernel(kernels + i);
886 /*! \brief Free the OpenCL program.
888 * The function releases the OpenCL program assuciated with the
889 * device that the calling PP rank is running on.
891 * \param program [in] OpenCL program to release.
893 static void freeGpuProgram(cl_program program)
897 cl_int cl_error = clReleaseProgram(program);
898 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
899 ("clReleaseProgram failed: " + ocl_get_error_string(cl_error)).c_str());
904 //! This function is documented in the header file
905 void gpu_free(NbnxmGpu* nb)
913 int kernel_count = sizeof(nb->kernel_ener_noprune_ptr) / sizeof(nb->kernel_ener_noprune_ptr[0][0]);
914 free_kernels(nb->kernel_ener_noprune_ptr[0], kernel_count);
916 kernel_count = sizeof(nb->kernel_ener_prune_ptr) / sizeof(nb->kernel_ener_prune_ptr[0][0]);
917 free_kernels(nb->kernel_ener_prune_ptr[0], kernel_count);
919 kernel_count = sizeof(nb->kernel_noener_noprune_ptr) / sizeof(nb->kernel_noener_noprune_ptr[0][0]);
920 free_kernels(nb->kernel_noener_noprune_ptr[0], kernel_count);
922 kernel_count = sizeof(nb->kernel_noener_prune_ptr) / sizeof(nb->kernel_noener_prune_ptr[0][0]);
923 free_kernels(nb->kernel_noener_prune_ptr[0], kernel_count);
925 free_kernel(&(nb->kernel_zero_e_fshift));
928 freeDeviceBuffer(&(nb->atdat->xq));
929 freeDeviceBuffer(&(nb->atdat->f));
930 freeDeviceBuffer(&(nb->atdat->e_lj));
931 freeDeviceBuffer(&(nb->atdat->e_el));
932 freeDeviceBuffer(&(nb->atdat->fshift));
933 freeDeviceBuffer(&(nb->atdat->lj_comb));
934 freeDeviceBuffer(&(nb->atdat->atom_types));
935 freeDeviceBuffer(&(nb->atdat->shift_vec));
939 freeDeviceBuffer(&(nb->nbparam->nbfp_climg2d));
940 freeDeviceBuffer(&(nb->nbparam->nbfp_comb_climg2d));
941 freeDeviceBuffer(&(nb->nbparam->coulomb_tab_climg2d));
945 auto* plist = nb->plist[InteractionLocality::Local];
946 freeDeviceBuffer(&plist->sci);
947 freeDeviceBuffer(&plist->cj4);
948 freeDeviceBuffer(&plist->imask);
949 freeDeviceBuffer(&plist->excl);
951 if (nb->bUseTwoStreams)
953 auto* plist_nl = nb->plist[InteractionLocality::NonLocal];
954 freeDeviceBuffer(&plist_nl->sci);
955 freeDeviceBuffer(&plist_nl->cj4);
956 freeDeviceBuffer(&plist_nl->imask);
957 freeDeviceBuffer(&plist_nl->excl);
962 pfree(nb->nbst.e_lj);
963 nb->nbst.e_lj = nullptr;
965 pfree(nb->nbst.e_el);
966 nb->nbst.e_el = nullptr;
968 pfree(nb->nbst.fshift);
969 nb->nbst.fshift = nullptr;
971 /* Free other events */
972 if (nb->nonlocal_done)
974 clReleaseEvent(nb->nonlocal_done);
975 nb->nonlocal_done = nullptr;
977 if (nb->misc_ops_and_local_H2D_done)
979 clReleaseEvent(nb->misc_ops_and_local_H2D_done);
980 nb->misc_ops_and_local_H2D_done = nullptr;
983 freeGpuProgram(nb->dev_rundata->program);
984 delete nb->dev_rundata;
986 /* Free timers and timings */
993 fprintf(debug, "Cleaned up OpenCL data structures.\n");
997 //! This function is documented in the header file
998 gmx_wallclock_gpu_nbnxn_t* gpu_get_timings(NbnxmGpu* nb)
1000 return (nb != nullptr && nb->bDoTime) ? nb->timings : nullptr;
1003 //! This function is documented in the header file
1004 void gpu_reset_timings(nonbonded_verlet_t* nbv)
1006 if (nbv->gpu_nbv && nbv->gpu_nbv->bDoTime)
1008 init_timings(nbv->gpu_nbv->timings);
1012 //! This function is documented in the header file
1013 int gpu_min_ci_balanced(NbnxmGpu* nb)
1015 return nb != nullptr ? gpu_min_ci_balanced_factor * nb->deviceContext_->deviceInfo().compute_units : 0;
1018 //! This function is documented in the header file
1019 gmx_bool gpu_is_kernel_ewald_analytical(const NbnxmGpu* nb)
1021 return ((nb->nbparam->eeltype == eelOclEWALD_ANA) || (nb->nbparam->eeltype == eelOclEWALD_ANA_TWIN));
1024 } // namespace Nbnxm