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_data_mgmt.h"
62 #include "gromacs/mdlib/nbnxn_gpu_jit_support.h"
63 #include "gromacs/mdtypes/interaction_const.h"
64 #include "gromacs/mdtypes/md_enums.h"
65 #include "gromacs/pbcutil/ishift.h"
66 #include "gromacs/timing/gpu_timing.h"
67 #include "gromacs/utility/cstringutil.h"
68 #include "gromacs/utility/fatalerror.h"
69 #include "gromacs/utility/gmxassert.h"
70 #include "gromacs/utility/real.h"
71 #include "gromacs/utility/smalloc.h"
73 #include "nbnxn_ocl_internal.h"
74 #include "nbnxn_ocl_types.h"
76 /*! \brief This parameter should be determined heuristically from the
77 * kernel execution times
79 * This value is best for small systems on a single AMD Radeon R9 290X
80 * (and about 5% faster than 40, which is the default for CUDA
81 * devices). Larger simulation systems were quite insensitive to the
82 * value of this parameter.
84 static unsigned int gpu_min_ci_balanced_factor = 50;
87 /*! \brief Returns true if LJ combination rules are used in the non-bonded kernels.
89 * Full doc in nbnxn_ocl_internal.h */
90 bool useLjCombRule(int vdwType)
92 return (vdwType == evdwOclCUTCOMBGEOM ||
93 vdwType == evdwOclCUTCOMBLB);
96 /*! \brief Tabulates the Ewald Coulomb force and initializes the size/scale
97 * and the table GPU array.
99 * If called with an already allocated table, it just re-uploads the
102 static void init_ewald_coulomb_force_table(const interaction_const_t *ic,
104 const gmx_device_runtime_data_t *runData)
110 if (nbp->coulomb_tab_climg2d != nullptr)
112 freeDeviceBuffer(&(nbp->coulomb_tab_climg2d));
115 /* Switched from using textures to using buffers */
116 // TODO: decide which alternative is most efficient - textures or buffers.
118 cl_image_format array_format;
120 array_format.image_channel_data_type = CL_FLOAT;
121 array_format.image_channel_order = CL_R;
123 coul_tab = clCreateImage2D(runData->context, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
124 &array_format, tabsize, 1, 0, ftmp, &cl_error);
127 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);
128 assert(cl_error == CL_SUCCESS);
129 // TODO: handle errors, check clCreateBuffer flags
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);
149 // TODO: handle errors, check clCreateBuffer flags
150 ad->shift_vec = clCreateBuffer(runData->context, CL_MEM_READ_WRITE, SHIFTS * ad->shift_vec_elem_size, nullptr, &cl_error);
151 assert(cl_error == CL_SUCCESS);
152 ad->bShiftVecUploaded = false;
154 /* An element of the fshift device buffer has the same size as one element
155 of the host side fshift buffer. */
156 ad->fshift_elem_size = sizeof(*cl_nb_staging_t::fshift);
158 ad->fshift = clCreateBuffer(runData->context, CL_MEM_READ_WRITE, SHIFTS * ad->fshift_elem_size, nullptr, &cl_error);
159 assert(cl_error == CL_SUCCESS);
160 // TODO: handle errors, check clCreateBuffer flags
162 ad->e_lj = clCreateBuffer(runData->context, CL_MEM_READ_WRITE, sizeof(float), nullptr, &cl_error);
163 assert(cl_error == CL_SUCCESS);
164 // TODO: handle errors, check clCreateBuffer flags
166 ad->e_el = clCreateBuffer(runData->context, CL_MEM_READ_WRITE, sizeof(float), nullptr, &cl_error);
167 assert(cl_error == CL_SUCCESS);
168 // TODO: handle errors, check clCreateBuffer flags
170 /* initialize to nullptr pointers to data that is not allocated here and will
171 need reallocation in nbnxn_gpu_init_atomdata */
175 /* size -1 indicates that the respective array hasn't been initialized yet */
180 /*! \brief Copies all parameters related to the cut-off from ic to nbp
182 static void set_cutoff_parameters(cl_nbparam_t *nbp,
183 const interaction_const_t *ic,
184 const NbnxnListParameters *listParams)
186 nbp->ewald_beta = ic->ewaldcoeff_q;
187 nbp->sh_ewald = ic->sh_ewald;
188 nbp->epsfac = ic->epsfac;
189 nbp->two_k_rf = 2.0 * ic->k_rf;
190 nbp->c_rf = ic->c_rf;
191 nbp->rvdw_sq = ic->rvdw * ic->rvdw;
192 nbp->rcoulomb_sq = ic->rcoulomb * ic->rcoulomb;
193 nbp->rlistOuter_sq = listParams->rlistOuter * listParams->rlistOuter;
194 nbp->rlistInner_sq = listParams->rlistInner * listParams->rlistInner;
195 nbp->useDynamicPruning = listParams->useDynamicPruning;
197 nbp->sh_lj_ewald = ic->sh_lj_ewald;
198 nbp->ewaldcoeff_lj = ic->ewaldcoeff_lj;
200 nbp->rvdw_switch = ic->rvdw_switch;
201 nbp->dispersion_shift = ic->dispersion_shift;
202 nbp->repulsion_shift = ic->repulsion_shift;
203 nbp->vdw_switch = ic->vdw_switch;
206 /*! \brief Returns the kinds of electrostatics and Vdw OpenCL
207 * kernels that will be used.
209 * Respectively, these values are from enum eelOcl and enum
212 map_interaction_types_to_gpu_kernel_flavors(const interaction_const_t *ic,
217 if (ic->vdwtype == evdwCUT)
219 switch (ic->vdw_modifier)
222 case eintmodPOTSHIFT:
226 *gpu_vdwtype = evdwOclCUT;
229 *gpu_vdwtype = evdwOclCUTCOMBGEOM;
232 *gpu_vdwtype = evdwOclCUTCOMBLB;
235 gmx_incons("The requested LJ combination rule is not implemented in the OpenCL GPU accelerated kernels!");
239 case eintmodFORCESWITCH:
240 *gpu_vdwtype = evdwOclFSWITCH;
242 case eintmodPOTSWITCH:
243 *gpu_vdwtype = evdwOclPSWITCH;
246 gmx_incons("The requested VdW interaction modifier is not implemented in the GPU accelerated kernels!");
250 else if (ic->vdwtype == evdwPME)
252 if (ic->ljpme_comb_rule == ljcrGEOM)
254 *gpu_vdwtype = evdwOclEWALDGEOM;
258 *gpu_vdwtype = evdwOclEWALDLB;
263 gmx_incons("The requested VdW type is not implemented in the GPU accelerated kernels!");
266 if (ic->eeltype == eelCUT)
268 *gpu_eeltype = eelOclCUT;
270 else if (EEL_RF(ic->eeltype))
272 *gpu_eeltype = eelOclRF;
274 else if ((EEL_PME(ic->eeltype) || ic->eeltype == eelEWALD))
276 /* Initially rcoulomb == rvdw, so it's surely not twin cut-off. */
277 *gpu_eeltype = nbnxn_gpu_pick_ewald_kernel_type(false);
281 /* Shouldn't happen, as this is checked when choosing Verlet-scheme */
282 gmx_incons("The requested electrostatics type is not implemented in the GPU accelerated kernels!");
286 /*! \brief Initializes the nonbonded parameter data structure.
288 static void init_nbparam(cl_nbparam_t *nbp,
289 const interaction_const_t *ic,
290 const NbnxnListParameters *listParams,
291 const nbnxn_atomdata_t *nbat,
292 const gmx_device_runtime_data_t *runData)
294 int ntypes, nnbfp, nnbfp_comb;
298 ntypes = nbat->ntype;
300 set_cutoff_parameters(nbp, ic, listParams);
302 map_interaction_types_to_gpu_kernel_flavors(ic,
307 if (ic->vdwtype == evdwPME)
309 if (ic->ljpme_comb_rule == ljcrGEOM)
311 assert(nbat->comb_rule == ljcrGEOM);
315 assert(nbat->comb_rule == ljcrLB);
318 /* generate table for PME */
319 nbp->coulomb_tab_climg2d = nullptr;
320 if (nbp->eeltype == eelOclEWALD_TAB || nbp->eeltype == eelOclEWALD_TAB_TWIN)
322 init_ewald_coulomb_force_table(ic, nbp, runData);
325 // TODO: improvement needed.
326 // The image2d is created here even if eeltype is not eelCuEWALD_TAB or eelCuEWALD_TAB_TWIN because the OpenCL kernels
327 // don't accept nullptr values for image2D parameters.
329 /* Switched from using textures to using buffers */
330 // TODO: decide which alternative is most efficient - textures or buffers.
332 cl_image_format array_format;
334 array_format.image_channel_data_type = CL_FLOAT;
335 array_format.image_channel_order = CL_R;
337 nbp->coulomb_tab_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE,
338 &array_format, 1, 1, 0, nullptr, &cl_error);
341 nbp->coulomb_tab_climg2d = clCreateBuffer(runData->context, CL_MEM_READ_ONLY, sizeof(cl_float), nullptr, &cl_error);
342 // TODO: handle errors
345 nnbfp = 2*ntypes*ntypes;
346 nnbfp_comb = 2*ntypes;
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(runData->context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
358 &array_format, nnbfp, 1, 0, nbat->nbfp, &cl_error);
361 nbp->nbfp_climg2d = clCreateBuffer(runData->context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, nnbfp*sizeof(cl_float), nbat->nbfp, &cl_error);
362 assert(cl_error == CL_SUCCESS);
363 // TODO: handle errors
365 if (ic->vdwtype == evdwPME)
367 /* Switched from using textures to using buffers */
368 // TODO: decide which alternative is most efficient - textures or buffers.
369 /* nbp->nbfp_comb_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
370 &array_format, nnbfp_comb, 1, 0, nbat->nbfp_comb, &cl_error);*/
371 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);
374 assert(cl_error == CL_SUCCESS);
375 // TODO: handle errors
379 // TODO: improvement needed.
380 // The image2d is created here even if vdwtype is not evdwPME because the OpenCL kernels
381 // don't accept nullptr values for image2D parameters.
382 /* Switched from using textures to using buffers */
383 // TODO: decide which alternative is most efficient - textures or buffers.
384 /* nbp->nbfp_comb_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE,
385 &array_format, 1, 1, 0, nullptr, &cl_error);*/
386 nbp->nbfp_comb_climg2d = clCreateBuffer(runData->context, CL_MEM_READ_ONLY, sizeof(cl_float), nullptr, &cl_error);
389 assert(cl_error == CL_SUCCESS);
390 // TODO: handle errors
395 //! This function is documented in the header file
396 void nbnxn_gpu_pme_loadbal_update_param(const nonbonded_verlet_t *nbv,
397 const interaction_const_t *ic,
398 const NbnxnListParameters *listParams)
400 if (!nbv || nbv->grp[0].kernel_type != nbnxnk8x8x8_GPU)
404 gmx_nbnxn_ocl_t *nb = nbv->gpu_nbv;
405 cl_nbparam_t *nbp = nb->nbparam;
407 set_cutoff_parameters(nbp, ic, listParams);
409 nbp->eeltype = nbnxn_gpu_pick_ewald_kernel_type(ic->rcoulomb != ic->rvdw);
411 init_ewald_coulomb_force_table(ic, nb->nbparam, nb->dev_rundata);
414 /*! \brief Initializes the pair list data structure.
416 static void init_plist(cl_plist_t *pl)
418 /* initialize to nullptr pointers to data that is not allocated here and will
419 need reallocation in nbnxn_gpu_init_pairlist */
425 /* size -1 indicates that the respective array hasn't been initialized yet */
432 pl->imask_nalloc = -1;
434 pl->excl_nalloc = -1;
435 pl->haveFreshList = false;
438 /*! \brief Initializes the timer data structure.
440 static void init_timers(cl_timers_t *t,
443 for (int i = 0; i <= (bUseTwoStreams ? 1 : 0); i++)
445 t->didPairlistH2D[i] = false;
446 t->didPrune[i] = false;
447 t->didRollingPrune[i] = false;
451 /*! \brief Initializes the timings data structure.
453 static void init_timings(gmx_wallclock_gpu_nbnxn_t *t)
462 for (i = 0; i < 2; i++)
464 for (j = 0; j < 2; j++)
466 t->ktime[i][j].t = 0.0;
467 t->ktime[i][j].c = 0;
472 t->pruneTime.t = 0.0;
473 t->dynamicPruneTime.c = 0;
474 t->dynamicPruneTime.t = 0.0;
477 /*! \brief Creates context for OpenCL GPU given by \p mygpu
479 * A fatal error results if creation fails.
481 * \param[inout] runtimeData runtime data including program and context
482 * \param[in] devInfo device info struct
483 * \param[in] rank MPI rank (for error reporting)
486 nbnxn_gpu_create_context(gmx_device_runtime_data_t *runtimeData,
487 const gmx_device_info_t *devInfo,
490 cl_context_properties context_properties[3];
491 cl_platform_id platform_id;
492 cl_device_id device_id;
496 assert(runtimeData != nullptr);
497 assert(devInfo != nullptr);
499 platform_id = devInfo->ocl_gpu_id.ocl_platform_id;
500 device_id = devInfo->ocl_gpu_id.ocl_device_id;
502 context_properties[0] = CL_CONTEXT_PLATFORM;
503 context_properties[1] = (cl_context_properties) platform_id;
504 context_properties[2] = 0; /* Terminates the list of properties */
506 context = clCreateContext(context_properties, 1, &device_id, nullptr, nullptr, &cl_error);
507 if (CL_SUCCESS != cl_error)
509 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s:\n OpenCL error %d: %s",
511 devInfo->device_name,
512 cl_error, ocl_get_error_string(cl_error).c_str());
516 runtimeData->context = context;
519 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
520 static cl_kernel nbnxn_gpu_create_kernel(gmx_nbnxn_ocl_t *nb,
521 const char *kernel_name)
526 kernel = clCreateKernel(nb->dev_rundata->program, kernel_name, &cl_error);
527 if (CL_SUCCESS != cl_error)
529 gmx_fatal(FARGS, "Failed to create kernel '%s' for GPU #%s: OpenCL error %d",
531 nb->dev_info->device_name,
538 /*! \brief Clears nonbonded shift force output array and energy outputs on the GPU.
541 nbnxn_ocl_clear_e_fshift(gmx_nbnxn_ocl_t *nb)
545 cl_atomdata_t * adat = nb->atdat;
546 cl_command_queue ls = nb->stream[eintLocal];
548 size_t local_work_size[3] = {1, 1, 1};
549 size_t global_work_size[3] = {1, 1, 1};
551 cl_int shifts = SHIFTS*3;
555 cl_kernel zero_e_fshift = nb->kernel_zero_e_fshift;
557 local_work_size[0] = 64;
558 // Round the total number of threads up from the array size
559 global_work_size[0] = ((shifts + local_work_size[0] - 1)/local_work_size[0])*local_work_size[0];
562 cl_error = clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->fshift));
563 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_lj));
564 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_el));
565 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_uint), &shifts);
566 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
568 cl_error = clEnqueueNDRangeKernel(ls, zero_e_fshift, 3, nullptr, global_work_size, local_work_size, 0, nullptr, nullptr);
569 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
572 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
573 static void nbnxn_gpu_init_kernels(gmx_nbnxn_ocl_t *nb)
575 /* Init to 0 main kernel arrays */
576 /* They will be later on initialized in select_nbnxn_kernel */
577 // TODO: consider always creating all variants of the kernels here so that there is no
578 // need for late call to clCreateKernel -- if that gives any advantage?
579 memset(nb->kernel_ener_noprune_ptr, 0, sizeof(nb->kernel_ener_noprune_ptr));
580 memset(nb->kernel_ener_prune_ptr, 0, sizeof(nb->kernel_ener_prune_ptr));
581 memset(nb->kernel_noener_noprune_ptr, 0, sizeof(nb->kernel_noener_noprune_ptr));
582 memset(nb->kernel_noener_prune_ptr, 0, sizeof(nb->kernel_noener_prune_ptr));
584 /* Init pruning kernels
586 * TODO: we could avoid creating kernels if dynamic pruning is turned off,
587 * but ATM that depends on force flags not passed into the initialization.
589 nb->kernel_pruneonly[epruneFirst] = nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_opencl");
590 nb->kernel_pruneonly[epruneRolling] = nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_rolling_opencl");
592 /* Init auxiliary kernels */
593 nb->kernel_memset_f = nbnxn_gpu_create_kernel(nb, "memset_f");
594 nb->kernel_memset_f2 = nbnxn_gpu_create_kernel(nb, "memset_f2");
595 nb->kernel_memset_f3 = nbnxn_gpu_create_kernel(nb, "memset_f3");
596 nb->kernel_zero_e_fshift = nbnxn_gpu_create_kernel(nb, "zero_e_fshift");
599 /*! \brief Initializes simulation constant data.
601 * Initializes members of the atomdata and nbparam structs and
602 * clears e/fshift output buffers.
604 static void nbnxn_ocl_init_const(gmx_nbnxn_ocl_t *nb,
605 const interaction_const_t *ic,
606 const NbnxnListParameters *listParams,
607 const nbnxn_atomdata_t *nbat)
609 init_atomdata_first(nb->atdat, nbat->ntype, nb->dev_rundata);
610 init_nbparam(nb->nbparam, ic, listParams, nbat, nb->dev_rundata);
614 //! This function is documented in the header file
615 void nbnxn_gpu_init(gmx_nbnxn_ocl_t **p_nb,
616 const gmx_device_info_t *deviceInfo,
617 const interaction_const_t *ic,
618 const NbnxnListParameters *listParams,
619 const nbnxn_atomdata_t *nbat,
621 gmx_bool bLocalAndNonlocal)
625 cl_command_queue_properties queue_properties;
636 snew(nb->nbparam, 1);
637 snew(nb->plist[eintLocal], 1);
638 if (bLocalAndNonlocal)
640 snew(nb->plist[eintNonlocal], 1);
643 nb->bUseTwoStreams = bLocalAndNonlocal;
645 nb->timers = new cl_timers_t();
646 snew(nb->timings, 1);
648 /* set device info, just point it to the right GPU among the detected ones */
649 nb->dev_info = deviceInfo;
650 snew(nb->dev_rundata, 1);
653 pmalloc((void**)&nb->nbst.e_lj, sizeof(*nb->nbst.e_lj));
654 pmalloc((void**)&nb->nbst.e_el, sizeof(*nb->nbst.e_el));
655 pmalloc((void**)&nb->nbst.fshift, SHIFTS * sizeof(*nb->nbst.fshift));
657 init_plist(nb->plist[eintLocal]);
659 /* OpenCL timing disabled if GMX_DISABLE_GPU_TIMING is defined. */
660 nb->bDoTime = (getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
662 /* Create queues only after bDoTime has been initialized */
665 queue_properties = CL_QUEUE_PROFILING_ENABLE;
669 queue_properties = 0;
672 nbnxn_gpu_create_context(nb->dev_rundata, nb->dev_info, rank);
674 /* local/non-local GPU streams */
675 nb->stream[eintLocal] = clCreateCommandQueue(nb->dev_rundata->context, nb->dev_info->ocl_gpu_id.ocl_device_id, queue_properties, &cl_error);
676 if (CL_SUCCESS != cl_error)
678 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s: OpenCL error %d",
680 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,
702 init_timers(nb->timers, nb->bUseTwoStreams);
703 init_timings(nb->timings);
706 nbnxn_ocl_init_const(nb, ic, listParams, nbat);
708 /* Enable LJ param manual prefetch for AMD or if we request through env. var.
709 * TODO: decide about NVIDIA
711 nb->bPrefetchLjParam =
712 (getenv("GMX_OCL_DISABLE_I_PREFETCH") == nullptr) &&
713 ((nb->dev_info->vendor_e == OCL_VENDOR_AMD) || (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_atomdata_t * adat = nb->atdat;
743 cl_command_queue ls = nb->stream[eintLocal];
744 cl_float value = 0.0f;
746 size_t local_work_size[3] = {1, 1, 1};
747 size_t global_work_size[3] = {1, 1, 1};
751 cl_kernel memset_f = nb->kernel_memset_f;
753 cl_uint natoms_flat = natoms_clear * (sizeof(rvec)/sizeof(real));
755 local_work_size[0] = 64;
756 // Round the total number of threads up from the array size
757 global_work_size[0] = ((natoms_flat + local_work_size[0] - 1)/local_work_size[0])*local_work_size[0];
760 cl_int gmx_used_in_debug cl_error;
762 cl_error = clSetKernelArg(memset_f, arg_no++, sizeof(cl_mem), &(adat->f));
763 cl_error |= clSetKernelArg(memset_f, arg_no++, sizeof(cl_float), &value);
764 cl_error |= clSetKernelArg(memset_f, arg_no++, sizeof(cl_uint), &natoms_flat);
765 assert(cl_error == CL_SUCCESS);
767 cl_error = clEnqueueNDRangeKernel(ls, memset_f, 3, nullptr, global_work_size, local_work_size, 0, nullptr, nullptr);
768 assert(cl_error == CL_SUCCESS);
771 //! This function is documented in the header file
773 nbnxn_gpu_clear_outputs(gmx_nbnxn_ocl_t *nb,
776 nbnxn_ocl_clear_f(nb, nb->atdat->natoms);
777 /* clear shift force array and energies if the outputs were
778 used in the current step */
779 if (flags & GMX_FORCE_VIRIAL)
781 nbnxn_ocl_clear_e_fshift(nb);
784 /* kick off buffer clearing kernel to ensure concurrency with constraints/update */
785 cl_int gmx_unused cl_error;
786 cl_error = clFlush(nb->stream[eintLocal]);
787 assert(CL_SUCCESS == cl_error);
790 //! This function is documented in the header file
791 void nbnxn_gpu_init_pairlist(gmx_nbnxn_ocl_t *nb,
792 const nbnxn_pairlist_t *h_plist,
796 // Timing accumulation should happen only if there was work to do
797 // because getLastRangeTime() gets skipped with empty lists later
798 // which leads to the counter not being reset.
799 bool bDoTime = (nb->bDoTime && h_plist->nsci > 0);
800 cl_command_queue stream = nb->stream[iloc];
801 cl_plist_t *d_plist = nb->plist[iloc];
803 if (d_plist->na_c < 0)
805 d_plist->na_c = h_plist->na_ci;
809 if (d_plist->na_c != h_plist->na_ci)
811 sprintf(sbuf, "In cu_init_plist: the #atoms per cell has changed (from %d to %d)",
812 d_plist->na_c, h_plist->na_ci);
819 nb->timers->pl_h2d[iloc].openTimingRegion(stream);
820 nb->timers->didPairlistH2D[iloc] = true;
823 // TODO most of this function is same in CUDA and OpenCL, move into the header
824 Context context = nb->dev_rundata->context;
826 reallocateDeviceBuffer(&d_plist->sci, h_plist->nsci,
827 &d_plist->nsci, &d_plist->sci_nalloc, context);
828 copyToDeviceBuffer(&d_plist->sci, h_plist->sci, 0, h_plist->nsci,
829 stream, GpuApiCallBehavior::Async,
830 bDoTime ? nb->timers->pl_h2d[iloc].fetchNextEvent() : nullptr);
832 reallocateDeviceBuffer(&d_plist->cj4, h_plist->ncj4,
833 &d_plist->ncj4, &d_plist->cj4_nalloc, context);
834 copyToDeviceBuffer(&d_plist->cj4, h_plist->cj4, 0, h_plist->ncj4,
835 stream, GpuApiCallBehavior::Async,
836 bDoTime ? nb->timers->pl_h2d[iloc].fetchNextEvent() : nullptr);
838 reallocateDeviceBuffer(&d_plist->imask, h_plist->ncj4*c_nbnxnGpuClusterpairSplit,
839 &d_plist->nimask, &d_plist->imask_nalloc, context);
841 reallocateDeviceBuffer(&d_plist->excl, h_plist->nexcl,
842 &d_plist->nexcl, &d_plist->excl_nalloc, context);
843 copyToDeviceBuffer(&d_plist->excl, h_plist->excl, 0, h_plist->nexcl,
844 stream, GpuApiCallBehavior::Async,
845 bDoTime ? nb->timers->pl_h2d[iloc].fetchNextEvent() : nullptr);
849 nb->timers->pl_h2d[iloc].closeTimingRegion(stream);
852 /* need to prune the pair list during the next step */
853 d_plist->haveFreshList = true;
856 //! This function is documented in the header file
857 void nbnxn_gpu_upload_shiftvec(gmx_nbnxn_ocl_t *nb,
858 const nbnxn_atomdata_t *nbatom)
860 cl_atomdata_t *adat = nb->atdat;
861 cl_command_queue ls = nb->stream[eintLocal];
863 /* only if we have a dynamic box */
864 if (nbatom->bDynamicBox || !adat->bShiftVecUploaded)
866 ocl_copy_H2D_async(adat->shift_vec, nbatom->shift_vec, 0,
867 SHIFTS * adat->shift_vec_elem_size, ls, nullptr);
868 adat->bShiftVecUploaded = true;
872 //! This function is documented in the header file
873 void nbnxn_gpu_init_atomdata(gmx_nbnxn_ocl_t *nb,
874 const nbnxn_atomdata_t *nbat)
879 bool bDoTime = nb->bDoTime;
880 cl_timers_t *timers = nb->timers;
881 cl_atomdata_t *d_atdat = nb->atdat;
882 cl_command_queue ls = nb->stream[eintLocal];
884 natoms = nbat->natoms;
889 /* time async copy */
890 timers->atdat.openTimingRegion(ls);
893 /* need to reallocate if we have to copy more atoms than the amount of space
894 available and only allocate if we haven't initialized yet, i.e d_atdat->natoms == -1 */
895 if (natoms > d_atdat->nalloc)
897 nalloc = over_alloc_small(natoms);
899 /* free up first if the arrays have already been initialized */
900 if (d_atdat->nalloc != -1)
902 freeDeviceBuffer(&d_atdat->f);
903 freeDeviceBuffer(&d_atdat->xq);
904 freeDeviceBuffer(&d_atdat->lj_comb);
905 freeDeviceBuffer(&d_atdat->atom_types);
908 d_atdat->f_elem_size = sizeof(rvec);
910 // TODO: handle errors, check clCreateBuffer flags
911 d_atdat->f = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_WRITE, nalloc * d_atdat->f_elem_size, nullptr, &cl_error);
912 assert(CL_SUCCESS == cl_error);
914 // TODO: change the flag to read-only
915 d_atdat->xq = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_WRITE, nalloc * sizeof(cl_float4), nullptr, &cl_error);
916 assert(CL_SUCCESS == cl_error);
917 // TODO: handle errors, check clCreateBuffer flags
919 if (useLjCombRule(nb->nbparam->vdwtype))
921 // TODO: change the flag to read-only
922 d_atdat->lj_comb = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_WRITE, nalloc * sizeof(cl_float2), nullptr, &cl_error);
923 assert(CL_SUCCESS == cl_error);
924 // TODO: handle errors, check clCreateBuffer flags
928 // TODO: change the flag to read-only
929 d_atdat->atom_types = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_WRITE, nalloc * sizeof(int), nullptr, &cl_error);
930 assert(CL_SUCCESS == cl_error);
931 // TODO: handle errors, check clCreateBuffer flags
934 d_atdat->nalloc = nalloc;
938 d_atdat->natoms = natoms;
939 d_atdat->natoms_local = nbat->natoms_local;
941 /* need to clear GPU f output if realloc happened */
944 nbnxn_ocl_clear_f(nb, nalloc);
947 if (useLjCombRule(nb->nbparam->vdwtype))
949 ocl_copy_H2D_async(d_atdat->lj_comb, nbat->lj_comb, 0,
950 natoms*sizeof(cl_float2), ls, bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
954 ocl_copy_H2D_async(d_atdat->atom_types, nbat->type, 0,
955 natoms*sizeof(int), ls, bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
961 timers->atdat.closeTimingRegion(ls);
964 /* kick off the tasks enqueued above to ensure concurrency with the search */
965 cl_error = clFlush(ls);
966 assert(CL_SUCCESS == cl_error);
969 /*! \brief Releases an OpenCL kernel pointer */
970 static void free_kernel(cl_kernel *kernel_ptr)
972 cl_int gmx_unused cl_error;
974 assert(nullptr != kernel_ptr);
978 cl_error = clReleaseKernel(*kernel_ptr);
979 assert(cl_error == CL_SUCCESS);
981 *kernel_ptr = nullptr;
985 /*! \brief Releases a list of OpenCL kernel pointers */
986 static void free_kernels(cl_kernel *kernels, int count)
990 for (i = 0; i < count; i++)
992 free_kernel(kernels + i);
996 /*! \brief Free the OpenCL runtime data (context and program).
998 * The function releases the OpenCL context and program assuciated with the
999 * device that the calling PP rank is running on.
1001 * \param runData [in] porinter to the structure with runtime data.
1003 static void free_gpu_device_runtime_data(gmx_device_runtime_data_t *runData)
1005 if (runData == nullptr)
1010 cl_int gmx_unused cl_error;
1012 if (runData->context)
1014 cl_error = clReleaseContext(runData->context);
1015 runData->context = nullptr;
1016 assert(CL_SUCCESS == cl_error);
1019 if (runData->program)
1021 cl_error = clReleaseProgram(runData->program);
1022 runData->program = nullptr;
1023 assert(CL_SUCCESS == cl_error);
1028 //! This function is documented in the header file
1029 void nbnxn_gpu_free(gmx_nbnxn_ocl_t *nb)
1037 int kernel_count = sizeof(nb->kernel_ener_noprune_ptr) / sizeof(nb->kernel_ener_noprune_ptr[0][0]);
1038 free_kernels((cl_kernel*)nb->kernel_ener_noprune_ptr, kernel_count);
1040 kernel_count = sizeof(nb->kernel_ener_prune_ptr) / sizeof(nb->kernel_ener_prune_ptr[0][0]);
1041 free_kernels((cl_kernel*)nb->kernel_ener_prune_ptr, kernel_count);
1043 kernel_count = sizeof(nb->kernel_noener_noprune_ptr) / sizeof(nb->kernel_noener_noprune_ptr[0][0]);
1044 free_kernels((cl_kernel*)nb->kernel_noener_noprune_ptr, kernel_count);
1046 kernel_count = sizeof(nb->kernel_noener_prune_ptr) / sizeof(nb->kernel_noener_prune_ptr[0][0]);
1047 free_kernels((cl_kernel*)nb->kernel_noener_prune_ptr, kernel_count);
1049 free_kernel(&(nb->kernel_memset_f));
1050 free_kernel(&(nb->kernel_memset_f2));
1051 free_kernel(&(nb->kernel_memset_f3));
1052 free_kernel(&(nb->kernel_zero_e_fshift));
1055 freeDeviceBuffer(&(nb->atdat->xq));
1056 freeDeviceBuffer(&(nb->atdat->f));
1057 freeDeviceBuffer(&(nb->atdat->e_lj));
1058 freeDeviceBuffer(&(nb->atdat->e_el));
1059 freeDeviceBuffer(&(nb->atdat->fshift));
1060 freeDeviceBuffer(&(nb->atdat->lj_comb));
1061 freeDeviceBuffer(&(nb->atdat->atom_types));
1062 freeDeviceBuffer(&(nb->atdat->shift_vec));
1066 freeDeviceBuffer(&(nb->nbparam->nbfp_climg2d));
1067 freeDeviceBuffer(&(nb->nbparam->nbfp_comb_climg2d));
1068 freeDeviceBuffer(&(nb->nbparam->coulomb_tab_climg2d));
1072 auto *plist = nb->plist[eintLocal];
1073 freeDeviceBuffer(&plist->sci);
1074 freeDeviceBuffer(&plist->cj4);
1075 freeDeviceBuffer(&plist->imask);
1076 freeDeviceBuffer(&plist->excl);
1078 if (nb->bUseTwoStreams)
1080 auto *plist_nl = nb->plist[eintNonlocal];
1081 freeDeviceBuffer(&plist_nl->sci);
1082 freeDeviceBuffer(&plist_nl->cj4);
1083 freeDeviceBuffer(&plist_nl->imask);
1084 freeDeviceBuffer(&plist_nl->excl);
1089 pfree(nb->nbst.e_lj);
1090 nb->nbst.e_lj = nullptr;
1092 pfree(nb->nbst.e_el);
1093 nb->nbst.e_el = nullptr;
1095 pfree(nb->nbst.fshift);
1096 nb->nbst.fshift = nullptr;
1098 /* Free command queues */
1099 clReleaseCommandQueue(nb->stream[eintLocal]);
1100 nb->stream[eintLocal] = nullptr;
1101 if (nb->bUseTwoStreams)
1103 clReleaseCommandQueue(nb->stream[eintNonlocal]);
1104 nb->stream[eintNonlocal] = nullptr;
1106 /* Free other events */
1107 if (nb->nonlocal_done)
1109 clReleaseEvent(nb->nonlocal_done);
1110 nb->nonlocal_done = nullptr;
1112 if (nb->misc_ops_and_local_H2D_done)
1114 clReleaseEvent(nb->misc_ops_and_local_H2D_done);
1115 nb->misc_ops_and_local_H2D_done = nullptr;
1118 free_gpu_device_runtime_data(nb->dev_rundata);
1119 sfree(nb->dev_rundata);
1121 /* Free timers and timings */
1128 fprintf(debug, "Cleaned up OpenCL data structures.\n");
1132 //! This function is documented in the header file
1133 gmx_wallclock_gpu_nbnxn_t *nbnxn_gpu_get_timings(gmx_nbnxn_ocl_t *nb)
1135 return (nb != nullptr && nb->bDoTime) ? nb->timings : nullptr;
1138 //! This function is documented in the header file
1139 void nbnxn_gpu_reset_timings(nonbonded_verlet_t* nbv)
1141 if (nbv->gpu_nbv && nbv->gpu_nbv->bDoTime)
1143 init_timings(nbv->gpu_nbv->timings);
1147 //! This function is documented in the header file
1148 int nbnxn_gpu_min_ci_balanced(gmx_nbnxn_ocl_t *nb)
1150 return nb != nullptr ?
1151 gpu_min_ci_balanced_factor * nb->dev_info->compute_units : 0;
1154 //! This function is documented in the header file
1155 gmx_bool nbnxn_gpu_is_kernel_ewald_analytical(const gmx_nbnxn_ocl_t *nb)
1157 return ((nb->nbparam->eeltype == eelOclEWALD_ANA) ||
1158 (nb->nbparam->eeltype == eelOclEWALD_ANA_TWIN));