2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016,2017,2018,2019, 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 Define OpenCL implementation of nbnxm_gpu_data_mgmt.h
38 * \author Anca Hamuraru <anca@streamcomputing.eu>
39 * \author Dimitrios Karkoulis <dimitris.karkoulis@gmail.com>
40 * \author Teemu Virolainen <teemu@streamcomputing.eu>
41 * \author Szilárd Páll <pall.szilard@gmail.com>
53 #include "gromacs/gpu_utils/gpu_utils.h"
54 #include "gromacs/gpu_utils/oclutils.h"
55 #include "gromacs/hardware/gpu_hw_info.h"
56 #include "gromacs/math/vectypes.h"
57 #include "gromacs/mdlib/force_flags.h"
58 #include "gromacs/mdtypes/interaction_const.h"
59 #include "gromacs/mdtypes/md_enums.h"
60 #include "gromacs/nbnxm/gpu_data_mgmt.h"
61 #include "gromacs/nbnxm/gpu_jit_support.h"
62 #include "gromacs/nbnxm/nbnxm.h"
63 #include "gromacs/nbnxm/nbnxm_gpu.h"
64 #include "gromacs/pbcutil/ishift.h"
65 #include "gromacs/timing/gpu_timing.h"
66 #include "gromacs/utility/cstringutil.h"
67 #include "gromacs/utility/fatalerror.h"
68 #include "gromacs/utility/gmxassert.h"
69 #include "gromacs/utility/real.h"
70 #include "gromacs/utility/smalloc.h"
72 #include "nbnxm_ocl_internal.h"
73 #include "nbnxm_ocl_types.h"
75 /*! \brief This parameter should be determined heuristically from the
76 * kernel execution times
78 * This value is best for small systems on a single AMD Radeon R9 290X
79 * (and about 5% faster than 40, which is the default for CUDA
80 * devices). Larger simulation systems were quite insensitive to the
81 * value of this parameter.
83 static unsigned int gpu_min_ci_balanced_factor = 50;
86 /*! \brief Returns true if LJ combination rules are used in the non-bonded kernels.
88 * Full doc in nbnxn_ocl_internal.h */
89 bool useLjCombRule(int vdwType)
91 return (vdwType == evdwOclCUTCOMBGEOM ||
92 vdwType == evdwOclCUTCOMBLB);
95 /*! \brief Tabulates the Ewald Coulomb force and initializes the size/scale
96 * and the table GPU array.
98 * If called with an already allocated table, it just re-uploads the
101 static void init_ewald_coulomb_force_table(const interaction_const_t *ic,
103 const gmx_device_runtime_data_t *runData)
109 if (nbp->coulomb_tab_climg2d != nullptr)
111 freeDeviceBuffer(&(nbp->coulomb_tab_climg2d));
114 /* Switched from using textures to using buffers */
115 // TODO: decide which alternative is most efficient - textures or buffers.
117 cl_image_format array_format;
119 array_format.image_channel_data_type = CL_FLOAT;
120 array_format.image_channel_order = CL_R;
122 coul_tab = clCreateImage2D(runData->context, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
123 &array_format, tabsize, 1, 0, ftmp, &cl_error);
126 coul_tab = clCreateBuffer(runData->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
127 ic->tabq_size*sizeof(cl_float), ic->tabq_coul_F, &cl_error);
128 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
129 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
131 nbp->coulomb_tab_climg2d = coul_tab;
132 nbp->coulomb_tab_scale = ic->tabq_scale;
136 /*! \brief Initializes the atomdata structure first time, it only gets filled at
139 static void init_atomdata_first(cl_atomdata_t *ad, int ntypes, gmx_device_runtime_data_t *runData)
145 /* An element of the shift_vec device buffer has the same size as one element
146 of the host side shift_vec buffer. */
147 ad->shift_vec_elem_size = sizeof(*nbnxn_atomdata_t::shift_vec.data());
149 ad->shift_vec = clCreateBuffer(runData->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
150 SHIFTS * ad->shift_vec_elem_size, nullptr, &cl_error);
151 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
152 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
153 ad->bShiftVecUploaded = CL_FALSE;
155 /* An element of the fshift device buffer has the same size as one element
156 of the host side fshift buffer. */
157 ad->fshift_elem_size = sizeof(*cl_nb_staging_t::fshift);
159 ad->fshift = clCreateBuffer(runData->context, CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
160 SHIFTS * ad->fshift_elem_size, nullptr, &cl_error);
161 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
162 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
164 ad->e_lj = clCreateBuffer(runData->context, CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
165 sizeof(float), nullptr, &cl_error);
166 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
167 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
169 ad->e_el = clCreateBuffer(runData->context, CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
170 sizeof(float), nullptr, &cl_error);
171 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
172 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
174 /* initialize to nullptr pointers to data that is not allocated here and will
175 need reallocation in nbnxn_gpu_init_atomdata */
179 /* size -1 indicates that the respective array hasn't been initialized yet */
184 /*! \brief Copies all parameters related to the cut-off from ic to nbp
186 static void set_cutoff_parameters(cl_nbparam_t *nbp,
187 const interaction_const_t *ic,
188 const NbnxnListParameters *listParams)
190 nbp->ewald_beta = ic->ewaldcoeff_q;
191 nbp->sh_ewald = ic->sh_ewald;
192 nbp->epsfac = ic->epsfac;
193 nbp->two_k_rf = 2.0 * ic->k_rf;
194 nbp->c_rf = ic->c_rf;
195 nbp->rvdw_sq = ic->rvdw * ic->rvdw;
196 nbp->rcoulomb_sq = ic->rcoulomb * ic->rcoulomb;
197 nbp->rlistOuter_sq = listParams->rlistOuter * listParams->rlistOuter;
198 nbp->rlistInner_sq = listParams->rlistInner * listParams->rlistInner;
199 nbp->useDynamicPruning = listParams->useDynamicPruning;
201 nbp->sh_lj_ewald = ic->sh_lj_ewald;
202 nbp->ewaldcoeff_lj = ic->ewaldcoeff_lj;
204 nbp->rvdw_switch = ic->rvdw_switch;
205 nbp->dispersion_shift = ic->dispersion_shift;
206 nbp->repulsion_shift = ic->repulsion_shift;
207 nbp->vdw_switch = ic->vdw_switch;
210 /*! \brief Returns the kinds of electrostatics and Vdw OpenCL
211 * kernels that will be used.
213 * Respectively, these values are from enum eelOcl and enum
216 map_interaction_types_to_gpu_kernel_flavors(const interaction_const_t *ic,
221 if (ic->vdwtype == evdwCUT)
223 switch (ic->vdw_modifier)
226 case eintmodPOTSHIFT:
230 *gpu_vdwtype = evdwOclCUT;
233 *gpu_vdwtype = evdwOclCUTCOMBGEOM;
236 *gpu_vdwtype = evdwOclCUTCOMBLB;
239 gmx_incons("The requested LJ combination rule is not implemented in the OpenCL GPU accelerated kernels!");
242 case eintmodFORCESWITCH:
243 *gpu_vdwtype = evdwOclFSWITCH;
245 case eintmodPOTSWITCH:
246 *gpu_vdwtype = evdwOclPSWITCH;
249 gmx_incons("The requested VdW interaction modifier is not implemented in the GPU accelerated kernels!");
252 else if (ic->vdwtype == evdwPME)
254 if (ic->ljpme_comb_rule == ljcrGEOM)
256 *gpu_vdwtype = evdwOclEWALDGEOM;
260 *gpu_vdwtype = evdwOclEWALDLB;
265 gmx_incons("The requested VdW type is not implemented in the GPU accelerated kernels!");
268 if (ic->eeltype == eelCUT)
270 *gpu_eeltype = eelOclCUT;
272 else if (EEL_RF(ic->eeltype))
274 *gpu_eeltype = eelOclRF;
276 else if ((EEL_PME(ic->eeltype) || ic->eeltype == eelEWALD))
278 /* Initially rcoulomb == rvdw, so it's surely not twin cut-off. */
279 *gpu_eeltype = nbnxn_gpu_pick_ewald_kernel_type(false);
283 /* Shouldn't happen, as this is checked when choosing Verlet-scheme */
284 gmx_incons("The requested electrostatics type is not implemented in the GPU accelerated kernels!");
288 /*! \brief Initializes the nonbonded parameter data structure.
290 static void init_nbparam(cl_nbparam_t *nbp,
291 const interaction_const_t *ic,
292 const NbnxnListParameters *listParams,
293 const nbnxn_atomdata_t::Params &nbatParams,
294 const gmx_device_runtime_data_t *runData)
298 set_cutoff_parameters(nbp, ic, listParams);
300 map_interaction_types_to_gpu_kernel_flavors(ic,
301 nbatParams.comb_rule,
305 if (ic->vdwtype == evdwPME)
307 if (ic->ljpme_comb_rule == ljcrGEOM)
309 GMX_ASSERT(nbatParams.comb_rule == ljcrGEOM, "Combination rule mismatch!");
313 GMX_ASSERT(nbatParams.comb_rule == ljcrLB, "Combination rule mismatch!");
316 /* generate table for PME */
317 nbp->coulomb_tab_climg2d = nullptr;
318 if (nbp->eeltype == eelOclEWALD_TAB || nbp->eeltype == eelOclEWALD_TAB_TWIN)
320 init_ewald_coulomb_force_table(ic, nbp, runData);
323 // TODO: improvement needed.
324 // The image2d is created here even if eeltype is not eelCuEWALD_TAB or eelCuEWALD_TAB_TWIN because the OpenCL kernels
325 // 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(runData->context, CL_MEM_READ_WRITE,
336 &array_format, 1, 1, 0, nullptr, &cl_error);
339 nbp->coulomb_tab_climg2d = clCreateBuffer(runData->context, CL_MEM_READ_ONLY, sizeof(cl_float), nullptr, &cl_error);
340 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
341 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
344 const int nnbfp = 2*nbatParams.numTypes*nbatParams.numTypes;
345 const int nnbfp_comb = 2*nbatParams.numTypes;
348 /* Switched from using textures to using buffers */
349 // TODO: decide which alternative is most efficient - textures or buffers.
351 cl_image_format array_format;
353 array_format.image_channel_data_type = CL_FLOAT;
354 array_format.image_channel_order = CL_R;
356 nbp->nbfp_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
357 &array_format, nnbfp, 1, 0, nbat->nbfp, &cl_error);
361 clCreateBuffer(runData->context,
362 CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
363 nnbfp*sizeof(cl_float),
364 const_cast<float *>(nbatParams.nbfp.data()),
366 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
367 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
369 if (ic->vdwtype == evdwPME)
371 /* Switched from using textures to using buffers */
372 // TODO: decide which alternative is most efficient - textures or buffers.
373 /* nbp->nbfp_comb_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
374 &array_format, nnbfp_comb, 1, 0, nbat->nbfp_comb, &cl_error);*/
375 nbp->nbfp_comb_climg2d =
376 clCreateBuffer(runData->context,
377 CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
378 nnbfp_comb*sizeof(cl_float),
379 const_cast<float *>(nbatParams.nbfp_comb.data()),
381 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
382 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
386 // TODO: improvement needed.
387 // The image2d is created here even if vdwtype is not evdwPME because the OpenCL kernels
388 // don't accept nullptr values for image2D parameters.
389 /* Switched from using textures to using buffers */
390 // TODO: decide which alternative is most efficient - textures or buffers.
391 /* nbp->nbfp_comb_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE,
392 &array_format, 1, 1, 0, nullptr, &cl_error);*/
393 nbp->nbfp_comb_climg2d = clCreateBuffer(runData->context, CL_MEM_READ_ONLY, sizeof(cl_float), nullptr, &cl_error);
394 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
395 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
400 //! This function is documented in the header file
401 void nbnxn_gpu_pme_loadbal_update_param(const nonbonded_verlet_t *nbv,
402 const interaction_const_t *ic,
403 const NbnxnListParameters *listParams)
405 if (!nbv || nbv->grp[0].kernel_type != nbnxnk8x8x8_GPU)
409 gmx_nbnxn_ocl_t *nb = nbv->gpu_nbv;
410 cl_nbparam_t *nbp = nb->nbparam;
412 set_cutoff_parameters(nbp, ic, listParams);
414 nbp->eeltype = nbnxn_gpu_pick_ewald_kernel_type(ic->rcoulomb != ic->rvdw);
416 init_ewald_coulomb_force_table(ic, nb->nbparam, nb->dev_rundata);
419 /*! \brief Initializes the pair list data structure.
421 static void init_plist(cl_plist_t *pl)
423 /* initialize to nullptr pointers to data that is not allocated here and will
424 need reallocation in nbnxn_gpu_init_pairlist */
430 /* size -1 indicates that the respective array hasn't been initialized yet */
437 pl->imask_nalloc = -1;
439 pl->excl_nalloc = -1;
440 pl->haveFreshList = false;
443 /*! \brief Initializes the timer data structure.
445 static void init_timers(cl_timers_t *t,
448 for (int i = 0; i <= (bUseTwoStreams ? 1 : 0); i++)
450 t->didPairlistH2D[i] = false;
451 t->didPrune[i] = false;
452 t->didRollingPrune[i] = false;
456 /*! \brief Initializes the timings data structure.
458 static void init_timings(gmx_wallclock_gpu_nbnxn_t *t)
467 for (i = 0; i < 2; i++)
469 for (j = 0; j < 2; j++)
471 t->ktime[i][j].t = 0.0;
472 t->ktime[i][j].c = 0;
477 t->pruneTime.t = 0.0;
478 t->dynamicPruneTime.c = 0;
479 t->dynamicPruneTime.t = 0.0;
482 /*! \brief Creates context for OpenCL GPU given by \p mygpu
484 * A fatal error results if creation fails.
486 * \param[inout] runtimeData runtime data including program and context
487 * \param[in] devInfo device info struct
488 * \param[in] rank MPI rank (for error reporting)
491 nbnxn_gpu_create_context(gmx_device_runtime_data_t *runtimeData,
492 const gmx_device_info_t *devInfo,
495 cl_context_properties context_properties[3];
496 cl_platform_id platform_id;
497 cl_device_id device_id;
501 assert(runtimeData != nullptr);
502 assert(devInfo != nullptr);
504 platform_id = devInfo->ocl_gpu_id.ocl_platform_id;
505 device_id = devInfo->ocl_gpu_id.ocl_device_id;
507 context_properties[0] = CL_CONTEXT_PLATFORM;
508 context_properties[1] = reinterpret_cast<cl_context_properties>(platform_id);
509 context_properties[2] = 0; /* Terminates the list of properties */
511 context = clCreateContext(context_properties, 1, &device_id, nullptr, nullptr, &cl_error);
512 if (CL_SUCCESS != cl_error)
514 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s:\n OpenCL error %d: %s",
516 devInfo->device_name,
517 cl_error, ocl_get_error_string(cl_error).c_str());
520 runtimeData->context = context;
523 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
524 static cl_kernel nbnxn_gpu_create_kernel(gmx_nbnxn_ocl_t *nb,
525 const char *kernel_name)
530 kernel = clCreateKernel(nb->dev_rundata->program, kernel_name, &cl_error);
531 if (CL_SUCCESS != cl_error)
533 gmx_fatal(FARGS, "Failed to create kernel '%s' for GPU #%s: OpenCL error %d",
535 nb->dev_info->device_name,
542 /*! \brief Clears nonbonded shift force output array and energy outputs on the GPU.
545 nbnxn_ocl_clear_e_fshift(gmx_nbnxn_ocl_t *nb)
549 cl_atomdata_t * adat = nb->atdat;
550 cl_command_queue ls = nb->stream[eintLocal];
552 size_t local_work_size[3] = {1, 1, 1};
553 size_t global_work_size[3] = {1, 1, 1};
555 cl_int shifts = SHIFTS*3;
559 cl_kernel zero_e_fshift = nb->kernel_zero_e_fshift;
561 local_work_size[0] = 64;
562 // Round the total number of threads up from the array size
563 global_work_size[0] = ((shifts + local_work_size[0] - 1)/local_work_size[0])*local_work_size[0];
566 cl_error = clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->fshift));
567 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_lj));
568 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_el));
569 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_uint), &shifts);
570 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
572 cl_error = clEnqueueNDRangeKernel(ls, zero_e_fshift, 3, nullptr, global_work_size, local_work_size, 0, nullptr, nullptr);
573 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
576 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
577 static void nbnxn_gpu_init_kernels(gmx_nbnxn_ocl_t *nb)
579 /* Init to 0 main kernel arrays */
580 /* They will be later on initialized in select_nbnxn_kernel */
581 // TODO: consider always creating all variants of the kernels here so that there is no
582 // need for late call to clCreateKernel -- if that gives any advantage?
583 memset(nb->kernel_ener_noprune_ptr, 0, sizeof(nb->kernel_ener_noprune_ptr));
584 memset(nb->kernel_ener_prune_ptr, 0, sizeof(nb->kernel_ener_prune_ptr));
585 memset(nb->kernel_noener_noprune_ptr, 0, sizeof(nb->kernel_noener_noprune_ptr));
586 memset(nb->kernel_noener_prune_ptr, 0, sizeof(nb->kernel_noener_prune_ptr));
588 /* Init pruning kernels
590 * TODO: we could avoid creating kernels if dynamic pruning is turned off,
591 * but ATM that depends on force flags not passed into the initialization.
593 nb->kernel_pruneonly[epruneFirst] = nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_opencl");
594 nb->kernel_pruneonly[epruneRolling] = nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_rolling_opencl");
596 /* Init auxiliary kernels */
597 nb->kernel_zero_e_fshift = nbnxn_gpu_create_kernel(nb, "zero_e_fshift");
600 /*! \brief Initializes simulation constant data.
602 * Initializes members of the atomdata and nbparam structs and
603 * clears e/fshift output buffers.
605 static void nbnxn_ocl_init_const(gmx_nbnxn_ocl_t *nb,
606 const interaction_const_t *ic,
607 const NbnxnListParameters *listParams,
608 const nbnxn_atomdata_t::Params &nbatParams)
610 init_atomdata_first(nb->atdat, nbatParams.numTypes, nb->dev_rundata);
611 init_nbparam(nb->nbparam, ic, listParams, nbatParams, nb->dev_rundata);
615 //! This function is documented in the header file
616 void nbnxn_gpu_init(gmx_nbnxn_ocl_t **p_nb,
617 const gmx_device_info_t *deviceInfo,
618 const interaction_const_t *ic,
619 const NbnxnListParameters *listParams,
620 const nbnxn_atomdata_t *nbat,
622 gmx_bool bLocalAndNonlocal)
626 cl_command_queue_properties queue_properties;
637 snew(nb->nbparam, 1);
638 snew(nb->plist[eintLocal], 1);
639 if (bLocalAndNonlocal)
641 snew(nb->plist[eintNonlocal], 1);
644 nb->bUseTwoStreams = static_cast<cl_bool>(bLocalAndNonlocal);
646 nb->timers = new cl_timers_t();
647 snew(nb->timings, 1);
649 /* set device info, just point it to the right GPU among the detected ones */
650 nb->dev_info = deviceInfo;
651 snew(nb->dev_rundata, 1);
654 pmalloc(reinterpret_cast<void**>(&nb->nbst.e_lj), sizeof(*nb->nbst.e_lj));
655 pmalloc(reinterpret_cast<void**>(&nb->nbst.e_el), sizeof(*nb->nbst.e_el));
656 pmalloc(reinterpret_cast<void**>(&nb->nbst.fshift), SHIFTS * sizeof(*nb->nbst.fshift));
658 init_plist(nb->plist[eintLocal]);
660 /* OpenCL timing disabled if GMX_DISABLE_GPU_TIMING is defined. */
661 nb->bDoTime = static_cast<cl_bool>(getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
663 /* Create queues only after bDoTime has been initialized */
666 queue_properties = CL_QUEUE_PROFILING_ENABLE;
670 queue_properties = 0;
673 nbnxn_gpu_create_context(nb->dev_rundata, nb->dev_info, rank);
675 /* local/non-local GPU streams */
676 nb->stream[eintLocal] = clCreateCommandQueue(nb->dev_rundata->context, nb->dev_info->ocl_gpu_id.ocl_device_id, queue_properties, &cl_error);
677 if (CL_SUCCESS != cl_error)
679 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s: OpenCL error %d",
681 nb->dev_info->device_name,
685 if (nb->bUseTwoStreams)
687 init_plist(nb->plist[eintNonlocal]);
689 nb->stream[eintNonlocal] = clCreateCommandQueue(nb->dev_rundata->context, nb->dev_info->ocl_gpu_id.ocl_device_id, queue_properties, &cl_error);
690 if (CL_SUCCESS != cl_error)
692 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s: OpenCL error %d",
694 nb->dev_info->device_name,
701 init_timers(nb->timers, nb->bUseTwoStreams == CL_TRUE);
702 init_timings(nb->timings);
705 nbnxn_ocl_init_const(nb, ic, listParams, nbat->params());
707 /* Enable LJ param manual prefetch for AMD or Intel or if we request through env. var.
708 * TODO: decide about NVIDIA
710 nb->bPrefetchLjParam =
711 (getenv("GMX_OCL_DISABLE_I_PREFETCH") == nullptr) &&
712 ((nb->dev_info->vendor_e == OCL_VENDOR_AMD) || (nb->dev_info->vendor_e == OCL_VENDOR_INTEL)
713 || (getenv("GMX_OCL_ENABLE_I_PREFETCH") != nullptr));
715 /* NOTE: in CUDA we pick L1 cache configuration for the nbnxn kernels here,
716 * but sadly this is not supported in OpenCL (yet?). Consider adding it if
717 * it becomes supported.
719 nbnxn_gpu_compile_kernels(nb);
720 nbnxn_gpu_init_kernels(nb);
722 /* clear energy and shift force outputs */
723 nbnxn_ocl_clear_e_fshift(nb);
729 fprintf(debug, "Initialized OpenCL data structures.\n");
733 /*! \brief Clears the first natoms_clear elements of the GPU nonbonded force output array.
735 static void nbnxn_ocl_clear_f(gmx_nbnxn_ocl_t *nb, int natoms_clear)
737 if (natoms_clear == 0)
742 cl_int gmx_used_in_debug cl_error;
744 cl_atomdata_t *atomData = nb->atdat;
745 cl_command_queue ls = nb->stream[eintLocal];
746 cl_float value = 0.0f;
748 cl_error = clEnqueueFillBuffer(ls, atomData->f, &value, sizeof(cl_float),
749 0, natoms_clear*sizeof(rvec), 0, nullptr, nullptr);
750 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS, ("clEnqueueFillBuffer failed: " +
751 ocl_get_error_string(cl_error)).c_str());
754 //! This function is documented in the header file
756 nbnxn_gpu_clear_outputs(gmx_nbnxn_ocl_t *nb,
759 nbnxn_ocl_clear_f(nb, nb->atdat->natoms);
760 /* clear shift force array and energies if the outputs were
761 used in the current step */
762 if (flags & GMX_FORCE_VIRIAL)
764 nbnxn_ocl_clear_e_fshift(nb);
767 /* kick off buffer clearing kernel to ensure concurrency with constraints/update */
768 cl_int gmx_unused cl_error;
769 cl_error = clFlush(nb->stream[eintLocal]);
770 assert(CL_SUCCESS == cl_error);
773 //! This function is documented in the header file
774 void nbnxn_gpu_init_pairlist(gmx_nbnxn_ocl_t *nb,
775 const NbnxnPairlistGpu *h_plist,
779 // Timing accumulation should happen only if there was work to do
780 // because getLastRangeTime() gets skipped with empty lists later
781 // which leads to the counter not being reset.
782 bool bDoTime = ((nb->bDoTime == CL_TRUE) && !h_plist->sci.empty());
783 cl_command_queue stream = nb->stream[iloc];
784 cl_plist_t *d_plist = nb->plist[iloc];
786 if (d_plist->na_c < 0)
788 d_plist->na_c = h_plist->na_ci;
792 if (d_plist->na_c != h_plist->na_ci)
794 sprintf(sbuf, "In cu_init_plist: the #atoms per cell has changed (from %d to %d)",
795 d_plist->na_c, h_plist->na_ci);
802 nb->timers->pl_h2d[iloc].openTimingRegion(stream);
803 nb->timers->didPairlistH2D[iloc] = true;
806 // TODO most of this function is same in CUDA and OpenCL, move into the header
807 Context context = nb->dev_rundata->context;
809 reallocateDeviceBuffer(&d_plist->sci, h_plist->sci.size(),
810 &d_plist->nsci, &d_plist->sci_nalloc, context);
811 copyToDeviceBuffer(&d_plist->sci, h_plist->sci.data(), 0, h_plist->sci.size(),
812 stream, GpuApiCallBehavior::Async,
813 bDoTime ? nb->timers->pl_h2d[iloc].fetchNextEvent() : nullptr);
815 reallocateDeviceBuffer(&d_plist->cj4, h_plist->cj4.size(),
816 &d_plist->ncj4, &d_plist->cj4_nalloc, context);
817 copyToDeviceBuffer(&d_plist->cj4, h_plist->cj4.data(), 0, h_plist->cj4.size(),
818 stream, GpuApiCallBehavior::Async,
819 bDoTime ? nb->timers->pl_h2d[iloc].fetchNextEvent() : nullptr);
821 reallocateDeviceBuffer(&d_plist->imask, h_plist->cj4.size()*c_nbnxnGpuClusterpairSplit,
822 &d_plist->nimask, &d_plist->imask_nalloc, context);
824 reallocateDeviceBuffer(&d_plist->excl, h_plist->excl.size(),
825 &d_plist->nexcl, &d_plist->excl_nalloc, context);
826 copyToDeviceBuffer(&d_plist->excl, h_plist->excl.data(), 0, h_plist->excl.size(),
827 stream, GpuApiCallBehavior::Async,
828 bDoTime ? nb->timers->pl_h2d[iloc].fetchNextEvent() : nullptr);
832 nb->timers->pl_h2d[iloc].closeTimingRegion(stream);
835 /* need to prune the pair list during the next step */
836 d_plist->haveFreshList = true;
839 //! This function is documented in the header file
840 void nbnxn_gpu_upload_shiftvec(gmx_nbnxn_ocl_t *nb,
841 const nbnxn_atomdata_t *nbatom)
843 cl_atomdata_t *adat = nb->atdat;
844 cl_command_queue ls = nb->stream[eintLocal];
846 /* only if we have a dynamic box */
847 if (nbatom->bDynamicBox || !adat->bShiftVecUploaded)
849 ocl_copy_H2D_async(adat->shift_vec, nbatom->shift_vec.data(), 0,
850 SHIFTS * adat->shift_vec_elem_size, ls, nullptr);
851 adat->bShiftVecUploaded = CL_TRUE;
855 //! This function is documented in the header file
856 void nbnxn_gpu_init_atomdata(gmx_nbnxn_ocl_t *nb,
857 const nbnxn_atomdata_t *nbat)
862 bool bDoTime = nb->bDoTime == CL_TRUE;
863 cl_timers_t *timers = nb->timers;
864 cl_atomdata_t *d_atdat = nb->atdat;
865 cl_command_queue ls = nb->stream[eintLocal];
867 natoms = nbat->numAtoms();
872 /* time async copy */
873 timers->atdat.openTimingRegion(ls);
876 /* need to reallocate if we have to copy more atoms than the amount of space
877 available and only allocate if we haven't initialized yet, i.e d_atdat->natoms == -1 */
878 if (natoms > d_atdat->nalloc)
880 nalloc = over_alloc_small(natoms);
882 /* free up first if the arrays have already been initialized */
883 if (d_atdat->nalloc != -1)
885 freeDeviceBuffer(&d_atdat->f);
886 freeDeviceBuffer(&d_atdat->xq);
887 freeDeviceBuffer(&d_atdat->lj_comb);
888 freeDeviceBuffer(&d_atdat->atom_types);
891 d_atdat->f_elem_size = sizeof(rvec);
893 d_atdat->f = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
894 nalloc * d_atdat->f_elem_size, nullptr, &cl_error);
895 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
896 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
898 d_atdat->xq = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
899 nalloc * sizeof(cl_float4), nullptr, &cl_error);
900 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
901 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
903 if (useLjCombRule(nb->nbparam->vdwtype))
905 d_atdat->lj_comb = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
906 nalloc * sizeof(cl_float2), nullptr, &cl_error);
907 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
908 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
912 d_atdat->atom_types = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
913 nalloc * sizeof(int), nullptr, &cl_error);
914 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
915 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
918 d_atdat->nalloc = nalloc;
922 d_atdat->natoms = natoms;
923 d_atdat->natoms_local = nbat->natoms_local;
925 /* need to clear GPU f output if realloc happened */
928 nbnxn_ocl_clear_f(nb, nalloc);
931 if (useLjCombRule(nb->nbparam->vdwtype))
933 ocl_copy_H2D_async(d_atdat->lj_comb, nbat->params().lj_comb.data(), 0,
934 natoms*sizeof(cl_float2), ls, bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
938 ocl_copy_H2D_async(d_atdat->atom_types, nbat->params().type.data(), 0,
939 natoms*sizeof(int), ls, bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
945 timers->atdat.closeTimingRegion(ls);
948 /* kick off the tasks enqueued above to ensure concurrency with the search */
949 cl_error = clFlush(ls);
950 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
951 ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
954 /*! \brief Releases an OpenCL kernel pointer */
955 static void free_kernel(cl_kernel *kernel_ptr)
957 cl_int gmx_unused cl_error;
959 assert(nullptr != kernel_ptr);
963 cl_error = clReleaseKernel(*kernel_ptr);
964 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS, ("clReleaseKernel failed: " + ocl_get_error_string(cl_error)).c_str());
966 *kernel_ptr = nullptr;
970 /*! \brief Releases a list of OpenCL kernel pointers */
971 static void free_kernels(cl_kernel *kernels, int count)
975 for (i = 0; i < count; i++)
977 free_kernel(kernels + i);
981 /*! \brief Free the OpenCL runtime data (context and program).
983 * The function releases the OpenCL context and program assuciated with the
984 * device that the calling PP rank is running on.
986 * \param runData [in] porinter to the structure with runtime data.
988 static void free_gpu_device_runtime_data(gmx_device_runtime_data_t *runData)
990 if (runData == nullptr)
995 cl_int gmx_unused cl_error;
997 if (runData->context)
999 cl_error = clReleaseContext(runData->context);
1000 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS, ("clReleaseContext failed: " + ocl_get_error_string(cl_error)).c_str());
1001 runData->context = nullptr;
1004 if (runData->program)
1006 cl_error = clReleaseProgram(runData->program);
1007 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS, ("clReleaseProgram failed: " + ocl_get_error_string(cl_error)).c_str());
1008 runData->program = nullptr;
1013 //! This function is documented in the header file
1014 void nbnxn_gpu_free(gmx_nbnxn_ocl_t *nb)
1022 int kernel_count = sizeof(nb->kernel_ener_noprune_ptr) / sizeof(nb->kernel_ener_noprune_ptr[0][0]);
1023 free_kernels(nb->kernel_ener_noprune_ptr[0], kernel_count);
1025 kernel_count = sizeof(nb->kernel_ener_prune_ptr) / sizeof(nb->kernel_ener_prune_ptr[0][0]);
1026 free_kernels(nb->kernel_ener_prune_ptr[0], kernel_count);
1028 kernel_count = sizeof(nb->kernel_noener_noprune_ptr) / sizeof(nb->kernel_noener_noprune_ptr[0][0]);
1029 free_kernels(nb->kernel_noener_noprune_ptr[0], kernel_count);
1031 kernel_count = sizeof(nb->kernel_noener_prune_ptr) / sizeof(nb->kernel_noener_prune_ptr[0][0]);
1032 free_kernels(nb->kernel_noener_prune_ptr[0], kernel_count);
1034 free_kernel(&(nb->kernel_zero_e_fshift));
1037 freeDeviceBuffer(&(nb->atdat->xq));
1038 freeDeviceBuffer(&(nb->atdat->f));
1039 freeDeviceBuffer(&(nb->atdat->e_lj));
1040 freeDeviceBuffer(&(nb->atdat->e_el));
1041 freeDeviceBuffer(&(nb->atdat->fshift));
1042 freeDeviceBuffer(&(nb->atdat->lj_comb));
1043 freeDeviceBuffer(&(nb->atdat->atom_types));
1044 freeDeviceBuffer(&(nb->atdat->shift_vec));
1048 freeDeviceBuffer(&(nb->nbparam->nbfp_climg2d));
1049 freeDeviceBuffer(&(nb->nbparam->nbfp_comb_climg2d));
1050 freeDeviceBuffer(&(nb->nbparam->coulomb_tab_climg2d));
1054 auto *plist = nb->plist[eintLocal];
1055 freeDeviceBuffer(&plist->sci);
1056 freeDeviceBuffer(&plist->cj4);
1057 freeDeviceBuffer(&plist->imask);
1058 freeDeviceBuffer(&plist->excl);
1060 if (nb->bUseTwoStreams)
1062 auto *plist_nl = nb->plist[eintNonlocal];
1063 freeDeviceBuffer(&plist_nl->sci);
1064 freeDeviceBuffer(&plist_nl->cj4);
1065 freeDeviceBuffer(&plist_nl->imask);
1066 freeDeviceBuffer(&plist_nl->excl);
1071 pfree(nb->nbst.e_lj);
1072 nb->nbst.e_lj = nullptr;
1074 pfree(nb->nbst.e_el);
1075 nb->nbst.e_el = nullptr;
1077 pfree(nb->nbst.fshift);
1078 nb->nbst.fshift = nullptr;
1080 /* Free command queues */
1081 clReleaseCommandQueue(nb->stream[eintLocal]);
1082 nb->stream[eintLocal] = nullptr;
1083 if (nb->bUseTwoStreams)
1085 clReleaseCommandQueue(nb->stream[eintNonlocal]);
1086 nb->stream[eintNonlocal] = nullptr;
1088 /* Free other events */
1089 if (nb->nonlocal_done)
1091 clReleaseEvent(nb->nonlocal_done);
1092 nb->nonlocal_done = nullptr;
1094 if (nb->misc_ops_and_local_H2D_done)
1096 clReleaseEvent(nb->misc_ops_and_local_H2D_done);
1097 nb->misc_ops_and_local_H2D_done = nullptr;
1100 free_gpu_device_runtime_data(nb->dev_rundata);
1101 sfree(nb->dev_rundata);
1103 /* Free timers and timings */
1110 fprintf(debug, "Cleaned up OpenCL data structures.\n");
1114 //! This function is documented in the header file
1115 gmx_wallclock_gpu_nbnxn_t *nbnxn_gpu_get_timings(gmx_nbnxn_ocl_t *nb)
1117 return (nb != nullptr && nb->bDoTime) ? nb->timings : nullptr;
1120 //! This function is documented in the header file
1121 void nbnxn_gpu_reset_timings(nonbonded_verlet_t* nbv)
1123 if (nbv->gpu_nbv && nbv->gpu_nbv->bDoTime)
1125 init_timings(nbv->gpu_nbv->timings);
1129 //! This function is documented in the header file
1130 int nbnxn_gpu_min_ci_balanced(gmx_nbnxn_ocl_t *nb)
1132 return nb != nullptr ?
1133 gpu_min_ci_balanced_factor * nb->dev_info->compute_units : 0;
1136 //! This function is documented in the header file
1137 gmx_bool nbnxn_gpu_is_kernel_ewald_analytical(const gmx_nbnxn_ocl_t *nb)
1139 return ((nb->nbparam->eeltype == eelOclEWALD_ANA) ||
1140 (nb->nbparam->eeltype == eelOclEWALD_ANA_TWIN));