2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016,2017,2018, 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 nbnxn_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/mdlib/nb_verlet.h"
59 #include "gromacs/mdlib/nbnxn_consts.h"
60 #include "gromacs/mdlib/nbnxn_gpu.h"
61 #include "gromacs/mdlib/nbnxn_gpu_common_utils.h"
62 #include "gromacs/mdlib/nbnxn_gpu_data_mgmt.h"
63 #include "gromacs/mdlib/nbnxn_gpu_jit_support.h"
64 #include "gromacs/mdtypes/interaction_const.h"
65 #include "gromacs/mdtypes/md_enums.h"
66 #include "gromacs/pbcutil/ishift.h"
67 #include "gromacs/timing/gpu_timing.h"
68 #include "gromacs/utility/cstringutil.h"
69 #include "gromacs/utility/fatalerror.h"
70 #include "gromacs/utility/gmxassert.h"
71 #include "gromacs/utility/real.h"
72 #include "gromacs/utility/smalloc.h"
74 #include "nbnxn_ocl_internal.h"
75 #include "nbnxn_ocl_types.h"
77 /*! \brief This parameter should be determined heuristically from the
78 * kernel execution times
80 * This value is best for small systems on a single AMD Radeon R9 290X
81 * (and about 5% faster than 40, which is the default for CUDA
82 * devices). Larger simulation systems were quite insensitive to the
83 * value of this parameter.
85 static unsigned int gpu_min_ci_balanced_factor = 50;
88 /*! \brief Returns true if LJ combination rules are used in the non-bonded kernels.
90 * Full doc in nbnxn_ocl_internal.h */
91 bool useLjCombRule(int vdwType)
93 return (vdwType == evdwOclCUTCOMBGEOM ||
94 vdwType == evdwOclCUTCOMBLB);
97 /*! \brief Tabulates the Ewald Coulomb force and initializes the size/scale
98 * and the table GPU array.
100 * If called with an already allocated table, it just re-uploads the
103 static void init_ewald_coulomb_force_table(const interaction_const_t *ic,
105 const gmx_device_runtime_data_t *runData)
111 if (nbp->coulomb_tab_climg2d != nullptr)
113 freeDeviceBuffer(&(nbp->coulomb_tab_climg2d));
116 /* Switched from using textures to using buffers */
117 // TODO: decide which alternative is most efficient - textures or buffers.
119 cl_image_format array_format;
121 array_format.image_channel_data_type = CL_FLOAT;
122 array_format.image_channel_order = CL_R;
124 coul_tab = clCreateImage2D(runData->context, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
125 &array_format, tabsize, 1, 0, ftmp, &cl_error);
128 coul_tab = clCreateBuffer(runData->context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, ic->tabq_size*sizeof(cl_float), ic->tabq_coul_F, &cl_error);
129 assert(cl_error == CL_SUCCESS);
130 // TODO: handle errors, check clCreateBuffer flags
132 nbp->coulomb_tab_climg2d = coul_tab;
133 nbp->coulomb_tab_scale = ic->tabq_scale;
137 /*! \brief Initializes the atomdata structure first time, it only gets filled at
140 static void init_atomdata_first(cl_atomdata_t *ad, int ntypes, gmx_device_runtime_data_t *runData)
146 /* An element of the shift_vec device buffer has the same size as one element
147 of the host side shift_vec buffer. */
148 ad->shift_vec_elem_size = sizeof(*nbnxn_atomdata_t::shift_vec);
150 // TODO: handle errors, check clCreateBuffer flags
151 ad->shift_vec = clCreateBuffer(runData->context, CL_MEM_READ_WRITE, SHIFTS * ad->shift_vec_elem_size, nullptr, &cl_error);
152 assert(cl_error == CL_SUCCESS);
153 ad->bShiftVecUploaded = 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, SHIFTS * ad->fshift_elem_size, nullptr, &cl_error);
160 assert(cl_error == CL_SUCCESS);
161 // TODO: handle errors, check clCreateBuffer flags
163 ad->e_lj = clCreateBuffer(runData->context, CL_MEM_READ_WRITE, sizeof(float), nullptr, &cl_error);
164 assert(cl_error == CL_SUCCESS);
165 // TODO: handle errors, check clCreateBuffer flags
167 ad->e_el = clCreateBuffer(runData->context, CL_MEM_READ_WRITE, sizeof(float), nullptr, &cl_error);
168 assert(cl_error == CL_SUCCESS);
169 // TODO: handle errors, check clCreateBuffer flags
171 /* initialize to nullptr pointers to data that is not allocated here and will
172 need reallocation in nbnxn_gpu_init_atomdata */
176 /* size -1 indicates that the respective array hasn't been initialized yet */
181 /*! \brief Copies all parameters related to the cut-off from ic to nbp
183 static void set_cutoff_parameters(cl_nbparam_t *nbp,
184 const interaction_const_t *ic,
185 const NbnxnListParameters *listParams)
187 nbp->ewald_beta = ic->ewaldcoeff_q;
188 nbp->sh_ewald = ic->sh_ewald;
189 nbp->epsfac = ic->epsfac;
190 nbp->two_k_rf = 2.0 * ic->k_rf;
191 nbp->c_rf = ic->c_rf;
192 nbp->rvdw_sq = ic->rvdw * ic->rvdw;
193 nbp->rcoulomb_sq = ic->rcoulomb * ic->rcoulomb;
194 nbp->rlistOuter_sq = listParams->rlistOuter * listParams->rlistOuter;
195 nbp->rlistInner_sq = listParams->rlistInner * listParams->rlistInner;
196 nbp->useDynamicPruning = listParams->useDynamicPruning;
198 nbp->sh_lj_ewald = ic->sh_lj_ewald;
199 nbp->ewaldcoeff_lj = ic->ewaldcoeff_lj;
201 nbp->rvdw_switch = ic->rvdw_switch;
202 nbp->dispersion_shift = ic->dispersion_shift;
203 nbp->repulsion_shift = ic->repulsion_shift;
204 nbp->vdw_switch = ic->vdw_switch;
207 /*! \brief Returns the kinds of electrostatics and Vdw OpenCL
208 * kernels that will be used.
210 * Respectively, these values are from enum eelOcl and enum
213 map_interaction_types_to_gpu_kernel_flavors(const interaction_const_t *ic,
218 if (ic->vdwtype == evdwCUT)
220 switch (ic->vdw_modifier)
223 case eintmodPOTSHIFT:
227 *gpu_vdwtype = evdwOclCUT;
230 *gpu_vdwtype = evdwOclCUTCOMBGEOM;
233 *gpu_vdwtype = evdwOclCUTCOMBLB;
236 gmx_incons("The requested LJ combination rule is not implemented in the OpenCL GPU accelerated kernels!");
240 case eintmodFORCESWITCH:
241 *gpu_vdwtype = evdwOclFSWITCH;
243 case eintmodPOTSWITCH:
244 *gpu_vdwtype = evdwOclPSWITCH;
247 gmx_incons("The requested VdW interaction modifier is not implemented in the GPU accelerated kernels!");
251 else if (ic->vdwtype == evdwPME)
253 if (ic->ljpme_comb_rule == ljcrGEOM)
255 *gpu_vdwtype = evdwOclEWALDGEOM;
259 *gpu_vdwtype = evdwOclEWALDLB;
264 gmx_incons("The requested VdW type is not implemented in the GPU accelerated kernels!");
267 if (ic->eeltype == eelCUT)
269 *gpu_eeltype = eelOclCUT;
271 else if (EEL_RF(ic->eeltype))
273 *gpu_eeltype = eelOclRF;
275 else if ((EEL_PME(ic->eeltype) || ic->eeltype == eelEWALD))
277 /* Initially rcoulomb == rvdw, so it's surely not twin cut-off. */
278 *gpu_eeltype = nbnxn_gpu_pick_ewald_kernel_type(false);
282 /* Shouldn't happen, as this is checked when choosing Verlet-scheme */
283 gmx_incons("The requested electrostatics type is not implemented in the GPU accelerated kernels!");
287 /*! \brief Initializes the nonbonded parameter data structure.
289 static void init_nbparam(cl_nbparam_t *nbp,
290 const interaction_const_t *ic,
291 const NbnxnListParameters *listParams,
292 const nbnxn_atomdata_t *nbat,
293 const gmx_device_runtime_data_t *runData)
295 int ntypes, nnbfp, nnbfp_comb;
299 ntypes = nbat->ntype;
301 set_cutoff_parameters(nbp, ic, listParams);
303 map_interaction_types_to_gpu_kernel_flavors(ic,
308 if (ic->vdwtype == evdwPME)
310 if (ic->ljpme_comb_rule == ljcrGEOM)
312 assert(nbat->comb_rule == ljcrGEOM);
316 assert(nbat->comb_rule == ljcrLB);
319 /* generate table for PME */
320 nbp->coulomb_tab_climg2d = nullptr;
321 if (nbp->eeltype == eelOclEWALD_TAB || nbp->eeltype == eelOclEWALD_TAB_TWIN)
323 init_ewald_coulomb_force_table(ic, nbp, runData);
326 // TODO: improvement needed.
327 // The image2d is created here even if eeltype is not eelCuEWALD_TAB or eelCuEWALD_TAB_TWIN because the OpenCL kernels
328 // don't accept nullptr values for image2D parameters.
330 /* Switched from using textures to using buffers */
331 // TODO: decide which alternative is most efficient - textures or buffers.
333 cl_image_format array_format;
335 array_format.image_channel_data_type = CL_FLOAT;
336 array_format.image_channel_order = CL_R;
338 nbp->coulomb_tab_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE,
339 &array_format, 1, 1, 0, nullptr, &cl_error);
342 nbp->coulomb_tab_climg2d = clCreateBuffer(runData->context, CL_MEM_READ_ONLY, sizeof(cl_float), nullptr, &cl_error);
343 // TODO: handle errors
346 nnbfp = 2*ntypes*ntypes;
347 nnbfp_comb = 2*ntypes;
350 /* Switched from using textures to using buffers */
351 // TODO: decide which alternative is most efficient - textures or buffers.
353 cl_image_format array_format;
355 array_format.image_channel_data_type = CL_FLOAT;
356 array_format.image_channel_order = CL_R;
358 nbp->nbfp_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
359 &array_format, nnbfp, 1, 0, nbat->nbfp, &cl_error);
362 nbp->nbfp_climg2d = clCreateBuffer(runData->context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, nnbfp*sizeof(cl_float), nbat->nbfp, &cl_error);
363 assert(cl_error == CL_SUCCESS);
364 // TODO: handle errors
366 if (ic->vdwtype == evdwPME)
368 /* Switched from using textures to using buffers */
369 // TODO: decide which alternative is most efficient - textures or buffers.
370 /* nbp->nbfp_comb_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
371 &array_format, nnbfp_comb, 1, 0, nbat->nbfp_comb, &cl_error);*/
372 nbp->nbfp_comb_climg2d = clCreateBuffer(runData->context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, nnbfp_comb*sizeof(cl_float), nbat->nbfp_comb, &cl_error);
375 assert(cl_error == CL_SUCCESS);
376 // TODO: handle errors
380 // TODO: improvement needed.
381 // The image2d is created here even if vdwtype is not evdwPME because the OpenCL kernels
382 // don't accept nullptr values for image2D parameters.
383 /* Switched from using textures to using buffers */
384 // TODO: decide which alternative is most efficient - textures or buffers.
385 /* nbp->nbfp_comb_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE,
386 &array_format, 1, 1, 0, nullptr, &cl_error);*/
387 nbp->nbfp_comb_climg2d = clCreateBuffer(runData->context, CL_MEM_READ_ONLY, sizeof(cl_float), nullptr, &cl_error);
390 assert(cl_error == CL_SUCCESS);
391 // TODO: handle errors
396 //! This function is documented in the header file
397 void nbnxn_gpu_pme_loadbal_update_param(const nonbonded_verlet_t *nbv,
398 const interaction_const_t *ic,
399 const NbnxnListParameters *listParams)
401 if (!nbv || nbv->grp[0].kernel_type != nbnxnk8x8x8_GPU)
405 gmx_nbnxn_ocl_t *nb = nbv->gpu_nbv;
406 cl_nbparam_t *nbp = nb->nbparam;
408 set_cutoff_parameters(nbp, ic, listParams);
410 nbp->eeltype = nbnxn_gpu_pick_ewald_kernel_type(ic->rcoulomb != ic->rvdw);
412 init_ewald_coulomb_force_table(ic, nb->nbparam, nb->dev_rundata);
415 /*! \brief Initializes the pair list data structure.
417 static void init_plist(cl_plist_t *pl)
419 /* initialize to nullptr pointers to data that is not allocated here and will
420 need reallocation in nbnxn_gpu_init_pairlist */
426 /* size -1 indicates that the respective array hasn't been initialized yet */
433 pl->imask_nalloc = -1;
435 pl->excl_nalloc = -1;
436 pl->haveFreshList = false;
439 /*! \brief Initializes the timer data structure.
441 static void init_timers(cl_timers_t *t,
444 for (int i = 0; i <= (bUseTwoStreams ? 1 : 0); i++)
446 t->didPairlistH2D[i] = false;
447 t->didPrune[i] = false;
448 t->didRollingPrune[i] = false;
452 /*! \brief Initializes the timings data structure.
454 static void init_timings(gmx_wallclock_gpu_nbnxn_t *t)
463 for (i = 0; i < 2; i++)
465 for (j = 0; j < 2; j++)
467 t->ktime[i][j].t = 0.0;
468 t->ktime[i][j].c = 0;
473 t->pruneTime.t = 0.0;
474 t->dynamicPruneTime.c = 0;
475 t->dynamicPruneTime.t = 0.0;
478 /*! \brief Creates context for OpenCL GPU given by \p mygpu
480 * A fatal error results if creation fails.
482 * \param[inout] runtimeData runtime data including program and context
483 * \param[in] devInfo device info struct
484 * \param[in] rank MPI rank (for error reporting)
487 nbnxn_gpu_create_context(gmx_device_runtime_data_t *runtimeData,
488 const gmx_device_info_t *devInfo,
491 cl_context_properties context_properties[3];
492 cl_platform_id platform_id;
493 cl_device_id device_id;
497 assert(runtimeData != nullptr);
498 assert(devInfo != nullptr);
500 platform_id = devInfo->ocl_gpu_id.ocl_platform_id;
501 device_id = devInfo->ocl_gpu_id.ocl_device_id;
503 context_properties[0] = CL_CONTEXT_PLATFORM;
504 context_properties[1] = (cl_context_properties) platform_id;
505 context_properties[2] = 0; /* Terminates the list of properties */
507 context = clCreateContext(context_properties, 1, &device_id, nullptr, nullptr, &cl_error);
508 if (CL_SUCCESS != cl_error)
510 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s:\n OpenCL error %d: %s",
512 devInfo->device_name,
513 cl_error, ocl_get_error_string(cl_error).c_str());
517 runtimeData->context = context;
520 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
521 static cl_kernel nbnxn_gpu_create_kernel(gmx_nbnxn_ocl_t *nb,
522 const char *kernel_name)
527 kernel = clCreateKernel(nb->dev_rundata->program, kernel_name, &cl_error);
528 if (CL_SUCCESS != cl_error)
530 gmx_fatal(FARGS, "Failed to create kernel '%s' for GPU #%s: OpenCL error %d",
532 nb->dev_info->device_name,
539 /*! \brief Clears nonbonded shift force output array and energy outputs on the GPU.
542 nbnxn_ocl_clear_e_fshift(gmx_nbnxn_ocl_t *nb)
546 cl_atomdata_t * adat = nb->atdat;
547 cl_command_queue ls = nb->stream[eintLocal];
549 size_t local_work_size[3] = {1, 1, 1};
550 size_t global_work_size[3] = {1, 1, 1};
552 cl_int shifts = SHIFTS*3;
556 cl_kernel zero_e_fshift = nb->kernel_zero_e_fshift;
558 local_work_size[0] = 64;
559 // Round the total number of threads up from the array size
560 global_work_size[0] = ((shifts + local_work_size[0] - 1)/local_work_size[0])*local_work_size[0];
563 cl_error = clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->fshift));
564 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_lj));
565 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_el));
566 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_uint), &shifts);
567 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
569 cl_error = clEnqueueNDRangeKernel(ls, zero_e_fshift, 3, nullptr, global_work_size, local_work_size, 0, nullptr, nullptr);
570 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
573 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
574 static void nbnxn_gpu_init_kernels(gmx_nbnxn_ocl_t *nb)
576 /* Init to 0 main kernel arrays */
577 /* They will be later on initialized in select_nbnxn_kernel */
578 // TODO: consider always creating all variants of the kernels here so that there is no
579 // need for late call to clCreateKernel -- if that gives any advantage?
580 memset(nb->kernel_ener_noprune_ptr, 0, sizeof(nb->kernel_ener_noprune_ptr));
581 memset(nb->kernel_ener_prune_ptr, 0, sizeof(nb->kernel_ener_prune_ptr));
582 memset(nb->kernel_noener_noprune_ptr, 0, sizeof(nb->kernel_noener_noprune_ptr));
583 memset(nb->kernel_noener_prune_ptr, 0, sizeof(nb->kernel_noener_prune_ptr));
585 /* Init pruning kernels
587 * TODO: we could avoid creating kernels if dynamic pruning is turned off,
588 * but ATM that depends on force flags not passed into the initialization.
590 nb->kernel_pruneonly[epruneFirst] = nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_opencl");
591 nb->kernel_pruneonly[epruneRolling] = nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_rolling_opencl");
593 /* Init auxiliary kernels */
594 nb->kernel_memset_f = nbnxn_gpu_create_kernel(nb, "memset_f");
595 nb->kernel_memset_f2 = nbnxn_gpu_create_kernel(nb, "memset_f2");
596 nb->kernel_memset_f3 = nbnxn_gpu_create_kernel(nb, "memset_f3");
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 *nbat)
610 init_atomdata_first(nb->atdat, nbat->ntype, nb->dev_rundata);
611 init_nbparam(nb->nbparam, ic, listParams, nbat, 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 = 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((void**)&nb->nbst.e_lj, sizeof(*nb->nbst.e_lj));
655 pmalloc((void**)&nb->nbst.e_el, sizeof(*nb->nbst.e_el));
656 pmalloc((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 = (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,
686 if (nb->bUseTwoStreams)
688 init_plist(nb->plist[eintNonlocal]);
690 nb->stream[eintNonlocal] = clCreateCommandQueue(nb->dev_rundata->context, nb->dev_info->ocl_gpu_id.ocl_device_id, queue_properties, &cl_error);
691 if (CL_SUCCESS != cl_error)
693 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s: OpenCL error %d",
695 nb->dev_info->device_name,
703 init_timers(nb->timers, nb->bUseTwoStreams);
704 init_timings(nb->timings);
707 nbnxn_ocl_init_const(nb, ic, listParams, nbat);
709 /* Enable LJ param manual prefetch for AMD or if we request through env. var.
710 * TODO: decide about NVIDIA
712 nb->bPrefetchLjParam =
713 (getenv("GMX_OCL_DISABLE_I_PREFETCH") == nullptr) &&
714 ((nb->dev_info->vendor_e == OCL_VENDOR_AMD) || (getenv("GMX_OCL_ENABLE_I_PREFETCH") != nullptr));
716 /* NOTE: in CUDA we pick L1 cache configuration for the nbnxn kernels here,
717 * but sadly this is not supported in OpenCL (yet?). Consider adding it if
718 * it becomes supported.
720 nbnxn_gpu_compile_kernels(nb);
721 nbnxn_gpu_init_kernels(nb);
723 /* clear energy and shift force outputs */
724 nbnxn_ocl_clear_e_fshift(nb);
730 fprintf(debug, "Initialized OpenCL data structures.\n");
734 /*! \brief Clears the first natoms_clear elements of the GPU nonbonded force output array.
736 static void nbnxn_ocl_clear_f(gmx_nbnxn_ocl_t *nb, int natoms_clear)
738 if (natoms_clear == 0)
743 cl_atomdata_t * adat = nb->atdat;
744 cl_command_queue ls = nb->stream[eintLocal];
745 cl_float value = 0.0f;
747 size_t local_work_size[3] = {1, 1, 1};
748 size_t global_work_size[3] = {1, 1, 1};
752 cl_kernel memset_f = nb->kernel_memset_f;
754 cl_uint natoms_flat = natoms_clear * (sizeof(rvec)/sizeof(real));
756 local_work_size[0] = 64;
757 // Round the total number of threads up from the array size
758 global_work_size[0] = ((natoms_flat + local_work_size[0] - 1)/local_work_size[0])*local_work_size[0];
761 cl_int gmx_used_in_debug cl_error;
763 cl_error = clSetKernelArg(memset_f, arg_no++, sizeof(cl_mem), &(adat->f));
764 cl_error |= clSetKernelArg(memset_f, arg_no++, sizeof(cl_float), &value);
765 cl_error |= clSetKernelArg(memset_f, arg_no++, sizeof(cl_uint), &natoms_flat);
766 assert(cl_error == CL_SUCCESS);
768 cl_error = clEnqueueNDRangeKernel(ls, memset_f, 3, nullptr, global_work_size, local_work_size, 0, nullptr, nullptr);
769 assert(cl_error == CL_SUCCESS);
772 //! This function is documented in the header file
774 nbnxn_gpu_clear_outputs(gmx_nbnxn_ocl_t *nb,
777 nbnxn_ocl_clear_f(nb, nb->atdat->natoms);
778 /* clear shift force array and energies if the outputs were
779 used in the current step */
780 if (flags & GMX_FORCE_VIRIAL)
782 nbnxn_ocl_clear_e_fshift(nb);
785 /* kick off buffer clearing kernel to ensure concurrency with constraints/update */
786 cl_int gmx_unused cl_error;
787 cl_error = clFlush(nb->stream[eintLocal]);
788 assert(CL_SUCCESS == cl_error);
791 //! This function is documented in the header file
792 void nbnxn_gpu_init_pairlist(gmx_nbnxn_ocl_t *nb,
793 const nbnxn_pairlist_t *h_plist,
796 if (canSkipWork(nb, iloc))
802 bool bDoTime = nb->bDoTime;
803 cl_command_queue stream = nb->stream[iloc];
804 cl_plist_t *d_plist = nb->plist[iloc];
806 if (d_plist->na_c < 0)
808 d_plist->na_c = h_plist->na_ci;
812 if (d_plist->na_c != h_plist->na_ci)
814 sprintf(sbuf, "In cu_init_plist: the #atoms per cell has changed (from %d to %d)",
815 d_plist->na_c, h_plist->na_ci);
822 nb->timers->pl_h2d[iloc].openTimingRegion(stream);
823 nb->timers->didPairlistH2D[iloc] = true;
826 // TODO most of this function is same in CUDA and OpenCL, move into the header
827 Context context = nb->dev_rundata->context;
829 reallocateDeviceBuffer(&d_plist->sci, h_plist->nsci,
830 &d_plist->nsci, &d_plist->sci_nalloc, context);
831 copyToDeviceBuffer(&d_plist->sci, h_plist->sci, 0, h_plist->nsci,
832 stream, GpuApiCallBehavior::Async,
833 bDoTime ? nb->timers->pl_h2d[iloc].fetchNextEvent() : nullptr);
835 reallocateDeviceBuffer(&d_plist->cj4, h_plist->ncj4,
836 &d_plist->ncj4, &d_plist->cj4_nalloc, context);
837 copyToDeviceBuffer(&d_plist->cj4, h_plist->cj4, 0, h_plist->ncj4,
838 stream, GpuApiCallBehavior::Async,
839 bDoTime ? nb->timers->pl_h2d[iloc].fetchNextEvent() : nullptr);
841 reallocateDeviceBuffer(&d_plist->imask, h_plist->ncj4*c_nbnxnGpuClusterpairSplit,
842 &d_plist->nimask, &d_plist->imask_nalloc, context);
844 reallocateDeviceBuffer(&d_plist->excl, h_plist->nexcl,
845 &d_plist->nexcl, &d_plist->excl_nalloc, context);
846 copyToDeviceBuffer(&d_plist->excl, h_plist->excl, 0, h_plist->nexcl,
847 stream, GpuApiCallBehavior::Async,
848 bDoTime ? nb->timers->pl_h2d[iloc].fetchNextEvent() : nullptr);
852 nb->timers->pl_h2d[iloc].closeTimingRegion(stream);
855 /* need to prune the pair list during the next step */
856 d_plist->haveFreshList = true;
859 //! This function is documented in the header file
860 void nbnxn_gpu_upload_shiftvec(gmx_nbnxn_ocl_t *nb,
861 const nbnxn_atomdata_t *nbatom)
863 cl_atomdata_t *adat = nb->atdat;
864 cl_command_queue ls = nb->stream[eintLocal];
866 /* only if we have a dynamic box */
867 if (nbatom->bDynamicBox || !adat->bShiftVecUploaded)
869 ocl_copy_H2D_async(adat->shift_vec, nbatom->shift_vec, 0,
870 SHIFTS * adat->shift_vec_elem_size, ls, nullptr);
871 adat->bShiftVecUploaded = true;
875 //! This function is documented in the header file
876 void nbnxn_gpu_init_atomdata(gmx_nbnxn_ocl_t *nb,
877 const nbnxn_atomdata_t *nbat)
882 bool bDoTime = nb->bDoTime;
883 cl_timers_t *timers = nb->timers;
884 cl_atomdata_t *d_atdat = nb->atdat;
885 cl_command_queue ls = nb->stream[eintLocal];
887 natoms = nbat->natoms;
892 /* time async copy */
893 timers->atdat.openTimingRegion(ls);
896 /* need to reallocate if we have to copy more atoms than the amount of space
897 available and only allocate if we haven't initialized yet, i.e d_atdat->natoms == -1 */
898 if (natoms > d_atdat->nalloc)
900 nalloc = over_alloc_small(natoms);
902 /* free up first if the arrays have already been initialized */
903 if (d_atdat->nalloc != -1)
905 freeDeviceBuffer(&d_atdat->f);
906 freeDeviceBuffer(&d_atdat->xq);
907 freeDeviceBuffer(&d_atdat->lj_comb);
908 freeDeviceBuffer(&d_atdat->atom_types);
911 d_atdat->f_elem_size = sizeof(rvec);
913 // TODO: handle errors, check clCreateBuffer flags
914 d_atdat->f = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_WRITE, nalloc * d_atdat->f_elem_size, nullptr, &cl_error);
915 assert(CL_SUCCESS == cl_error);
917 // TODO: change the flag to read-only
918 d_atdat->xq = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_WRITE, nalloc * sizeof(cl_float4), nullptr, &cl_error);
919 assert(CL_SUCCESS == cl_error);
920 // TODO: handle errors, check clCreateBuffer flags
922 if (useLjCombRule(nb->nbparam->vdwtype))
924 // TODO: change the flag to read-only
925 d_atdat->lj_comb = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_WRITE, nalloc * sizeof(cl_float2), nullptr, &cl_error);
926 assert(CL_SUCCESS == cl_error);
927 // TODO: handle errors, check clCreateBuffer flags
931 // TODO: change the flag to read-only
932 d_atdat->atom_types = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_WRITE, nalloc * sizeof(int), nullptr, &cl_error);
933 assert(CL_SUCCESS == cl_error);
934 // TODO: handle errors, check clCreateBuffer flags
937 d_atdat->nalloc = nalloc;
941 d_atdat->natoms = natoms;
942 d_atdat->natoms_local = nbat->natoms_local;
944 /* need to clear GPU f output if realloc happened */
947 nbnxn_ocl_clear_f(nb, nalloc);
950 if (useLjCombRule(nb->nbparam->vdwtype))
952 ocl_copy_H2D_async(d_atdat->lj_comb, nbat->lj_comb, 0,
953 natoms*sizeof(cl_float2), ls, bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
957 ocl_copy_H2D_async(d_atdat->atom_types, nbat->type, 0,
958 natoms*sizeof(int), ls, bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
964 timers->atdat.closeTimingRegion(ls);
967 /* kick off the tasks enqueued above to ensure concurrency with the search */
968 cl_error = clFlush(ls);
969 assert(CL_SUCCESS == cl_error);
972 /*! \brief Releases an OpenCL kernel pointer */
973 static void free_kernel(cl_kernel *kernel_ptr)
975 cl_int gmx_unused cl_error;
977 assert(nullptr != kernel_ptr);
981 cl_error = clReleaseKernel(*kernel_ptr);
982 assert(cl_error == CL_SUCCESS);
984 *kernel_ptr = nullptr;
988 /*! \brief Releases a list of OpenCL kernel pointers */
989 static void free_kernels(cl_kernel *kernels, int count)
993 for (i = 0; i < count; i++)
995 free_kernel(kernels + i);
999 /*! \brief Free the OpenCL runtime data (context and program).
1001 * The function releases the OpenCL context and program assuciated with the
1002 * device that the calling PP rank is running on.
1004 * \param runData [in] porinter to the structure with runtime data.
1006 static void free_gpu_device_runtime_data(gmx_device_runtime_data_t *runData)
1008 if (runData == nullptr)
1013 cl_int gmx_unused cl_error;
1015 if (runData->context)
1017 cl_error = clReleaseContext(runData->context);
1018 runData->context = nullptr;
1019 assert(CL_SUCCESS == cl_error);
1022 if (runData->program)
1024 cl_error = clReleaseProgram(runData->program);
1025 runData->program = nullptr;
1026 assert(CL_SUCCESS == cl_error);
1031 //! This function is documented in the header file
1032 void nbnxn_gpu_free(gmx_nbnxn_ocl_t *nb)
1040 int kernel_count = sizeof(nb->kernel_ener_noprune_ptr) / sizeof(nb->kernel_ener_noprune_ptr[0][0]);
1041 free_kernels((cl_kernel*)nb->kernel_ener_noprune_ptr, kernel_count);
1043 kernel_count = sizeof(nb->kernel_ener_prune_ptr) / sizeof(nb->kernel_ener_prune_ptr[0][0]);
1044 free_kernels((cl_kernel*)nb->kernel_ener_prune_ptr, kernel_count);
1046 kernel_count = sizeof(nb->kernel_noener_noprune_ptr) / sizeof(nb->kernel_noener_noprune_ptr[0][0]);
1047 free_kernels((cl_kernel*)nb->kernel_noener_noprune_ptr, kernel_count);
1049 kernel_count = sizeof(nb->kernel_noener_prune_ptr) / sizeof(nb->kernel_noener_prune_ptr[0][0]);
1050 free_kernels((cl_kernel*)nb->kernel_noener_prune_ptr, kernel_count);
1052 free_kernel(&(nb->kernel_memset_f));
1053 free_kernel(&(nb->kernel_memset_f2));
1054 free_kernel(&(nb->kernel_memset_f3));
1055 free_kernel(&(nb->kernel_zero_e_fshift));
1058 freeDeviceBuffer(&(nb->atdat->xq));
1059 freeDeviceBuffer(&(nb->atdat->f));
1060 freeDeviceBuffer(&(nb->atdat->e_lj));
1061 freeDeviceBuffer(&(nb->atdat->e_el));
1062 freeDeviceBuffer(&(nb->atdat->fshift));
1063 freeDeviceBuffer(&(nb->atdat->lj_comb));
1064 freeDeviceBuffer(&(nb->atdat->atom_types));
1065 freeDeviceBuffer(&(nb->atdat->shift_vec));
1069 freeDeviceBuffer(&(nb->nbparam->nbfp_climg2d));
1070 freeDeviceBuffer(&(nb->nbparam->nbfp_comb_climg2d));
1071 freeDeviceBuffer(&(nb->nbparam->coulomb_tab_climg2d));
1075 auto *plist = nb->plist[eintLocal];
1076 freeDeviceBuffer(&plist->sci);
1077 freeDeviceBuffer(&plist->cj4);
1078 freeDeviceBuffer(&plist->imask);
1079 freeDeviceBuffer(&plist->excl);
1081 if (nb->bUseTwoStreams)
1083 auto *plist_nl = nb->plist[eintNonlocal];
1084 freeDeviceBuffer(&plist_nl->sci);
1085 freeDeviceBuffer(&plist_nl->cj4);
1086 freeDeviceBuffer(&plist_nl->imask);
1087 freeDeviceBuffer(&plist_nl->excl);
1092 pfree(nb->nbst.e_lj);
1093 nb->nbst.e_lj = nullptr;
1095 pfree(nb->nbst.e_el);
1096 nb->nbst.e_el = nullptr;
1098 pfree(nb->nbst.fshift);
1099 nb->nbst.fshift = nullptr;
1101 /* Free command queues */
1102 clReleaseCommandQueue(nb->stream[eintLocal]);
1103 nb->stream[eintLocal] = nullptr;
1104 if (nb->bUseTwoStreams)
1106 clReleaseCommandQueue(nb->stream[eintNonlocal]);
1107 nb->stream[eintNonlocal] = nullptr;
1109 /* Free other events */
1110 if (nb->nonlocal_done)
1112 clReleaseEvent(nb->nonlocal_done);
1113 nb->nonlocal_done = nullptr;
1115 if (nb->misc_ops_and_local_H2D_done)
1117 clReleaseEvent(nb->misc_ops_and_local_H2D_done);
1118 nb->misc_ops_and_local_H2D_done = nullptr;
1121 free_gpu_device_runtime_data(nb->dev_rundata);
1122 sfree(nb->dev_rundata);
1124 /* Free timers and timings */
1131 fprintf(debug, "Cleaned up OpenCL data structures.\n");
1135 //! This function is documented in the header file
1136 gmx_wallclock_gpu_nbnxn_t *nbnxn_gpu_get_timings(gmx_nbnxn_ocl_t *nb)
1138 return (nb != nullptr && nb->bDoTime) ? nb->timings : nullptr;
1141 //! This function is documented in the header file
1142 void nbnxn_gpu_reset_timings(nonbonded_verlet_t* nbv)
1144 if (nbv->gpu_nbv && nbv->gpu_nbv->bDoTime)
1146 init_timings(nbv->gpu_nbv->timings);
1150 //! This function is documented in the header file
1151 int nbnxn_gpu_min_ci_balanced(gmx_nbnxn_ocl_t *nb)
1153 return nb != nullptr ?
1154 gpu_min_ci_balanced_factor * nb->dev_info->compute_units : 0;
1157 //! This function is documented in the header file
1158 gmx_bool nbnxn_gpu_is_kernel_ewald_analytical(const gmx_nbnxn_ocl_t *nb)
1160 return ((nb->nbparam->eeltype == eelOclEWALD_ANA) ||
1161 (nb->nbparam->eeltype == eelOclEWALD_ANA_TWIN));