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 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_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
128 ic->tabq_size*sizeof(cl_float), ic->tabq_coul_F, &cl_error);
129 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
130 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
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.data());
150 ad->shift_vec = clCreateBuffer(runData->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
151 SHIFTS * ad->shift_vec_elem_size, nullptr, &cl_error);
152 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
153 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
154 ad->bShiftVecUploaded = CL_FALSE;
156 /* An element of the fshift device buffer has the same size as one element
157 of the host side fshift buffer. */
158 ad->fshift_elem_size = sizeof(*cl_nb_staging_t::fshift);
160 ad->fshift = clCreateBuffer(runData->context, CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
161 SHIFTS * ad->fshift_elem_size, nullptr, &cl_error);
162 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
163 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
165 ad->e_lj = clCreateBuffer(runData->context, CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
166 sizeof(float), nullptr, &cl_error);
167 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
168 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
170 ad->e_el = clCreateBuffer(runData->context, CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
171 sizeof(float), nullptr, &cl_error);
172 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
173 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
175 /* initialize to nullptr pointers to data that is not allocated here and will
176 need reallocation in nbnxn_gpu_init_atomdata */
180 /* size -1 indicates that the respective array hasn't been initialized yet */
185 /*! \brief Copies all parameters related to the cut-off from ic to nbp
187 static void set_cutoff_parameters(cl_nbparam_t *nbp,
188 const interaction_const_t *ic,
189 const NbnxnListParameters *listParams)
191 nbp->ewald_beta = ic->ewaldcoeff_q;
192 nbp->sh_ewald = ic->sh_ewald;
193 nbp->epsfac = ic->epsfac;
194 nbp->two_k_rf = 2.0 * ic->k_rf;
195 nbp->c_rf = ic->c_rf;
196 nbp->rvdw_sq = ic->rvdw * ic->rvdw;
197 nbp->rcoulomb_sq = ic->rcoulomb * ic->rcoulomb;
198 nbp->rlistOuter_sq = listParams->rlistOuter * listParams->rlistOuter;
199 nbp->rlistInner_sq = listParams->rlistInner * listParams->rlistInner;
200 nbp->useDynamicPruning = listParams->useDynamicPruning;
202 nbp->sh_lj_ewald = ic->sh_lj_ewald;
203 nbp->ewaldcoeff_lj = ic->ewaldcoeff_lj;
205 nbp->rvdw_switch = ic->rvdw_switch;
206 nbp->dispersion_shift = ic->dispersion_shift;
207 nbp->repulsion_shift = ic->repulsion_shift;
208 nbp->vdw_switch = ic->vdw_switch;
211 /*! \brief Returns the kinds of electrostatics and Vdw OpenCL
212 * kernels that will be used.
214 * Respectively, these values are from enum eelOcl and enum
217 map_interaction_types_to_gpu_kernel_flavors(const interaction_const_t *ic,
222 if (ic->vdwtype == evdwCUT)
224 switch (ic->vdw_modifier)
227 case eintmodPOTSHIFT:
231 *gpu_vdwtype = evdwOclCUT;
234 *gpu_vdwtype = evdwOclCUTCOMBGEOM;
237 *gpu_vdwtype = evdwOclCUTCOMBLB;
240 gmx_incons("The requested LJ combination rule is not implemented in the OpenCL GPU accelerated kernels!");
243 case eintmodFORCESWITCH:
244 *gpu_vdwtype = evdwOclFSWITCH;
246 case eintmodPOTSWITCH:
247 *gpu_vdwtype = evdwOclPSWITCH;
250 gmx_incons("The requested VdW interaction modifier is not implemented in the GPU accelerated kernels!");
253 else if (ic->vdwtype == evdwPME)
255 if (ic->ljpme_comb_rule == ljcrGEOM)
257 *gpu_vdwtype = evdwOclEWALDGEOM;
261 *gpu_vdwtype = evdwOclEWALDLB;
266 gmx_incons("The requested VdW type is not implemented in the GPU accelerated kernels!");
269 if (ic->eeltype == eelCUT)
271 *gpu_eeltype = eelOclCUT;
273 else if (EEL_RF(ic->eeltype))
275 *gpu_eeltype = eelOclRF;
277 else if ((EEL_PME(ic->eeltype) || ic->eeltype == eelEWALD))
279 /* Initially rcoulomb == rvdw, so it's surely not twin cut-off. */
280 *gpu_eeltype = nbnxn_gpu_pick_ewald_kernel_type(false);
284 /* Shouldn't happen, as this is checked when choosing Verlet-scheme */
285 gmx_incons("The requested electrostatics type is not implemented in the GPU accelerated kernels!");
289 /*! \brief Initializes the nonbonded parameter data structure.
291 static void init_nbparam(cl_nbparam_t *nbp,
292 const interaction_const_t *ic,
293 const NbnxnListParameters *listParams,
294 const nbnxn_atomdata_t::Params &nbatParams,
295 const gmx_device_runtime_data_t *runData)
299 set_cutoff_parameters(nbp, ic, listParams);
301 map_interaction_types_to_gpu_kernel_flavors(ic,
302 nbatParams.comb_rule,
306 if (ic->vdwtype == evdwPME)
308 if (ic->ljpme_comb_rule == ljcrGEOM)
310 GMX_ASSERT(nbatParams.comb_rule == ljcrGEOM, "Combination rule mismatch!");
314 GMX_ASSERT(nbatParams.comb_rule == ljcrLB, "Combination rule mismatch!");
317 /* generate table for PME */
318 nbp->coulomb_tab_climg2d = nullptr;
319 if (nbp->eeltype == eelOclEWALD_TAB || nbp->eeltype == eelOclEWALD_TAB_TWIN)
321 init_ewald_coulomb_force_table(ic, nbp, runData);
324 // TODO: improvement needed.
325 // The image2d is created here even if eeltype is not eelCuEWALD_TAB or eelCuEWALD_TAB_TWIN because the OpenCL kernels
326 // don't accept nullptr values for image2D parameters.
328 /* Switched from using textures to using buffers */
329 // TODO: decide which alternative is most efficient - textures or buffers.
331 cl_image_format array_format;
333 array_format.image_channel_data_type = CL_FLOAT;
334 array_format.image_channel_order = CL_R;
336 nbp->coulomb_tab_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE,
337 &array_format, 1, 1, 0, nullptr, &cl_error);
340 nbp->coulomb_tab_climg2d = clCreateBuffer(runData->context, CL_MEM_READ_ONLY, sizeof(cl_float), nullptr, &cl_error);
341 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
342 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
345 const int nnbfp = 2*nbatParams.numTypes*nbatParams.numTypes;
346 const int nnbfp_comb = 2*nbatParams.numTypes;
349 /* Switched from using textures to using buffers */
350 // TODO: decide which alternative is most efficient - textures or buffers.
352 cl_image_format array_format;
354 array_format.image_channel_data_type = CL_FLOAT;
355 array_format.image_channel_order = CL_R;
357 nbp->nbfp_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
358 &array_format, nnbfp, 1, 0, nbat->nbfp, &cl_error);
362 clCreateBuffer(runData->context,
363 CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
364 nnbfp*sizeof(cl_float),
365 const_cast<float *>(nbatParams.nbfp.data()),
367 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
368 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
370 if (ic->vdwtype == evdwPME)
372 /* Switched from using textures to using buffers */
373 // TODO: decide which alternative is most efficient - textures or buffers.
374 /* nbp->nbfp_comb_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
375 &array_format, nnbfp_comb, 1, 0, nbat->nbfp_comb, &cl_error);*/
376 nbp->nbfp_comb_climg2d =
377 clCreateBuffer(runData->context,
378 CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
379 nnbfp_comb*sizeof(cl_float),
380 const_cast<float *>(nbatParams.nbfp_comb.data()),
382 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
383 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
387 // TODO: improvement needed.
388 // The image2d is created here even if vdwtype is not evdwPME because the OpenCL kernels
389 // don't accept nullptr values for image2D parameters.
390 /* Switched from using textures to using buffers */
391 // TODO: decide which alternative is most efficient - textures or buffers.
392 /* nbp->nbfp_comb_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE,
393 &array_format, 1, 1, 0, nullptr, &cl_error);*/
394 nbp->nbfp_comb_climg2d = clCreateBuffer(runData->context, CL_MEM_READ_ONLY, sizeof(cl_float), nullptr, &cl_error);
395 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
396 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
401 //! This function is documented in the header file
402 void nbnxn_gpu_pme_loadbal_update_param(const nonbonded_verlet_t *nbv,
403 const interaction_const_t *ic,
404 const NbnxnListParameters *listParams)
406 if (!nbv || nbv->grp[0].kernel_type != nbnxnk8x8x8_GPU)
410 gmx_nbnxn_ocl_t *nb = nbv->gpu_nbv;
411 cl_nbparam_t *nbp = nb->nbparam;
413 set_cutoff_parameters(nbp, ic, listParams);
415 nbp->eeltype = nbnxn_gpu_pick_ewald_kernel_type(ic->rcoulomb != ic->rvdw);
417 init_ewald_coulomb_force_table(ic, nb->nbparam, nb->dev_rundata);
420 /*! \brief Initializes the pair list data structure.
422 static void init_plist(cl_plist_t *pl)
424 /* initialize to nullptr pointers to data that is not allocated here and will
425 need reallocation in nbnxn_gpu_init_pairlist */
431 /* size -1 indicates that the respective array hasn't been initialized yet */
438 pl->imask_nalloc = -1;
440 pl->excl_nalloc = -1;
441 pl->haveFreshList = false;
444 /*! \brief Initializes the timer data structure.
446 static void init_timers(cl_timers_t *t,
449 for (int i = 0; i <= (bUseTwoStreams ? 1 : 0); i++)
451 t->didPairlistH2D[i] = false;
452 t->didPrune[i] = false;
453 t->didRollingPrune[i] = false;
457 /*! \brief Initializes the timings data structure.
459 static void init_timings(gmx_wallclock_gpu_nbnxn_t *t)
468 for (i = 0; i < 2; i++)
470 for (j = 0; j < 2; j++)
472 t->ktime[i][j].t = 0.0;
473 t->ktime[i][j].c = 0;
478 t->pruneTime.t = 0.0;
479 t->dynamicPruneTime.c = 0;
480 t->dynamicPruneTime.t = 0.0;
483 /*! \brief Creates context for OpenCL GPU given by \p mygpu
485 * A fatal error results if creation fails.
487 * \param[inout] runtimeData runtime data including program and context
488 * \param[in] devInfo device info struct
489 * \param[in] rank MPI rank (for error reporting)
492 nbnxn_gpu_create_context(gmx_device_runtime_data_t *runtimeData,
493 const gmx_device_info_t *devInfo,
496 cl_context_properties context_properties[3];
497 cl_platform_id platform_id;
498 cl_device_id device_id;
502 assert(runtimeData != nullptr);
503 assert(devInfo != nullptr);
505 platform_id = devInfo->ocl_gpu_id.ocl_platform_id;
506 device_id = devInfo->ocl_gpu_id.ocl_device_id;
508 context_properties[0] = CL_CONTEXT_PLATFORM;
509 context_properties[1] = reinterpret_cast<cl_context_properties>(platform_id);
510 context_properties[2] = 0; /* Terminates the list of properties */
512 context = clCreateContext(context_properties, 1, &device_id, nullptr, nullptr, &cl_error);
513 if (CL_SUCCESS != cl_error)
515 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s:\n OpenCL error %d: %s",
517 devInfo->device_name,
518 cl_error, ocl_get_error_string(cl_error).c_str());
521 runtimeData->context = context;
524 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
525 static cl_kernel nbnxn_gpu_create_kernel(gmx_nbnxn_ocl_t *nb,
526 const char *kernel_name)
531 kernel = clCreateKernel(nb->dev_rundata->program, kernel_name, &cl_error);
532 if (CL_SUCCESS != cl_error)
534 gmx_fatal(FARGS, "Failed to create kernel '%s' for GPU #%s: OpenCL error %d",
536 nb->dev_info->device_name,
543 /*! \brief Clears nonbonded shift force output array and energy outputs on the GPU.
546 nbnxn_ocl_clear_e_fshift(gmx_nbnxn_ocl_t *nb)
550 cl_atomdata_t * adat = nb->atdat;
551 cl_command_queue ls = nb->stream[eintLocal];
553 size_t local_work_size[3] = {1, 1, 1};
554 size_t global_work_size[3] = {1, 1, 1};
556 cl_int shifts = SHIFTS*3;
560 cl_kernel zero_e_fshift = nb->kernel_zero_e_fshift;
562 local_work_size[0] = 64;
563 // Round the total number of threads up from the array size
564 global_work_size[0] = ((shifts + local_work_size[0] - 1)/local_work_size[0])*local_work_size[0];
567 cl_error = clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->fshift));
568 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_lj));
569 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_el));
570 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_uint), &shifts);
571 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
573 cl_error = clEnqueueNDRangeKernel(ls, zero_e_fshift, 3, nullptr, global_work_size, local_work_size, 0, nullptr, nullptr);
574 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
577 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
578 static void nbnxn_gpu_init_kernels(gmx_nbnxn_ocl_t *nb)
580 /* Init to 0 main kernel arrays */
581 /* They will be later on initialized in select_nbnxn_kernel */
582 // TODO: consider always creating all variants of the kernels here so that there is no
583 // need for late call to clCreateKernel -- if that gives any advantage?
584 memset(nb->kernel_ener_noprune_ptr, 0, sizeof(nb->kernel_ener_noprune_ptr));
585 memset(nb->kernel_ener_prune_ptr, 0, sizeof(nb->kernel_ener_prune_ptr));
586 memset(nb->kernel_noener_noprune_ptr, 0, sizeof(nb->kernel_noener_noprune_ptr));
587 memset(nb->kernel_noener_prune_ptr, 0, sizeof(nb->kernel_noener_prune_ptr));
589 /* Init pruning kernels
591 * TODO: we could avoid creating kernels if dynamic pruning is turned off,
592 * but ATM that depends on force flags not passed into the initialization.
594 nb->kernel_pruneonly[epruneFirst] = nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_opencl");
595 nb->kernel_pruneonly[epruneRolling] = nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_rolling_opencl");
597 /* Init auxiliary kernels */
598 nb->kernel_zero_e_fshift = nbnxn_gpu_create_kernel(nb, "zero_e_fshift");
601 /*! \brief Initializes simulation constant data.
603 * Initializes members of the atomdata and nbparam structs and
604 * clears e/fshift output buffers.
606 static void nbnxn_ocl_init_const(gmx_nbnxn_ocl_t *nb,
607 const interaction_const_t *ic,
608 const NbnxnListParameters *listParams,
609 const nbnxn_atomdata_t::Params &nbatParams)
611 init_atomdata_first(nb->atdat, nbatParams.numTypes, nb->dev_rundata);
612 init_nbparam(nb->nbparam, ic, listParams, nbatParams, nb->dev_rundata);
616 //! This function is documented in the header file
617 void nbnxn_gpu_init(gmx_nbnxn_ocl_t **p_nb,
618 const gmx_device_info_t *deviceInfo,
619 const interaction_const_t *ic,
620 const NbnxnListParameters *listParams,
621 const nbnxn_atomdata_t *nbat,
623 gmx_bool bLocalAndNonlocal)
627 cl_command_queue_properties queue_properties;
638 snew(nb->nbparam, 1);
639 snew(nb->plist[eintLocal], 1);
640 if (bLocalAndNonlocal)
642 snew(nb->plist[eintNonlocal], 1);
645 nb->bUseTwoStreams = static_cast<cl_bool>(bLocalAndNonlocal);
647 nb->timers = new cl_timers_t();
648 snew(nb->timings, 1);
650 /* set device info, just point it to the right GPU among the detected ones */
651 nb->dev_info = deviceInfo;
652 snew(nb->dev_rundata, 1);
655 pmalloc(reinterpret_cast<void**>(&nb->nbst.e_lj), sizeof(*nb->nbst.e_lj));
656 pmalloc(reinterpret_cast<void**>(&nb->nbst.e_el), sizeof(*nb->nbst.e_el));
657 pmalloc(reinterpret_cast<void**>(&nb->nbst.fshift), SHIFTS * sizeof(*nb->nbst.fshift));
659 init_plist(nb->plist[eintLocal]);
661 /* OpenCL timing disabled if GMX_DISABLE_GPU_TIMING is defined. */
662 nb->bDoTime = static_cast<cl_bool>(getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
664 /* Create queues only after bDoTime has been initialized */
667 queue_properties = CL_QUEUE_PROFILING_ENABLE;
671 queue_properties = 0;
674 nbnxn_gpu_create_context(nb->dev_rundata, nb->dev_info, rank);
676 /* local/non-local GPU streams */
677 nb->stream[eintLocal] = clCreateCommandQueue(nb->dev_rundata->context, nb->dev_info->ocl_gpu_id.ocl_device_id, queue_properties, &cl_error);
678 if (CL_SUCCESS != cl_error)
680 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s: OpenCL error %d",
682 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,
702 init_timers(nb->timers, nb->bUseTwoStreams == CL_TRUE);
703 init_timings(nb->timings);
706 nbnxn_ocl_init_const(nb, ic, listParams, nbat->params());
708 /* Enable LJ param manual prefetch for AMD or Intel 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) || (nb->dev_info->vendor_e == OCL_VENDOR_INTEL)
714 || (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_int gmx_used_in_debug cl_error;
745 cl_atomdata_t *atomData = nb->atdat;
746 cl_command_queue ls = nb->stream[eintLocal];
747 cl_float value = 0.0f;
749 cl_error = clEnqueueFillBuffer(ls, atomData->f, &value, sizeof(cl_float),
750 0, natoms_clear*sizeof(rvec), 0, nullptr, nullptr);
751 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS, ("clEnqueueFillBuffer failed: " +
752 ocl_get_error_string(cl_error)).c_str());
755 //! This function is documented in the header file
757 nbnxn_gpu_clear_outputs(gmx_nbnxn_ocl_t *nb,
760 nbnxn_ocl_clear_f(nb, nb->atdat->natoms);
761 /* clear shift force array and energies if the outputs were
762 used in the current step */
763 if (flags & GMX_FORCE_VIRIAL)
765 nbnxn_ocl_clear_e_fshift(nb);
768 /* kick off buffer clearing kernel to ensure concurrency with constraints/update */
769 cl_int gmx_unused cl_error;
770 cl_error = clFlush(nb->stream[eintLocal]);
771 assert(CL_SUCCESS == cl_error);
774 //! This function is documented in the header file
775 void nbnxn_gpu_init_pairlist(gmx_nbnxn_ocl_t *nb,
776 const NbnxnPairlistGpu *h_plist,
780 // Timing accumulation should happen only if there was work to do
781 // because getLastRangeTime() gets skipped with empty lists later
782 // which leads to the counter not being reset.
783 bool bDoTime = ((nb->bDoTime == CL_TRUE) && h_plist->nsci > 0);
784 cl_command_queue stream = nb->stream[iloc];
785 cl_plist_t *d_plist = nb->plist[iloc];
787 if (d_plist->na_c < 0)
789 d_plist->na_c = h_plist->na_ci;
793 if (d_plist->na_c != h_plist->na_ci)
795 sprintf(sbuf, "In cu_init_plist: the #atoms per cell has changed (from %d to %d)",
796 d_plist->na_c, h_plist->na_ci);
803 nb->timers->pl_h2d[iloc].openTimingRegion(stream);
804 nb->timers->didPairlistH2D[iloc] = true;
807 // TODO most of this function is same in CUDA and OpenCL, move into the header
808 Context context = nb->dev_rundata->context;
810 reallocateDeviceBuffer(&d_plist->sci, h_plist->nsci,
811 &d_plist->nsci, &d_plist->sci_nalloc, context);
812 copyToDeviceBuffer(&d_plist->sci, h_plist->sci, 0, h_plist->nsci,
813 stream, GpuApiCallBehavior::Async,
814 bDoTime ? nb->timers->pl_h2d[iloc].fetchNextEvent() : nullptr);
816 reallocateDeviceBuffer(&d_plist->cj4, h_plist->ncj4,
817 &d_plist->ncj4, &d_plist->cj4_nalloc, context);
818 copyToDeviceBuffer(&d_plist->cj4, h_plist->cj4, 0, h_plist->ncj4,
819 stream, GpuApiCallBehavior::Async,
820 bDoTime ? nb->timers->pl_h2d[iloc].fetchNextEvent() : nullptr);
822 reallocateDeviceBuffer(&d_plist->imask, h_plist->ncj4*c_nbnxnGpuClusterpairSplit,
823 &d_plist->nimask, &d_plist->imask_nalloc, context);
825 reallocateDeviceBuffer(&d_plist->excl, h_plist->nexcl,
826 &d_plist->nexcl, &d_plist->excl_nalloc, context);
827 copyToDeviceBuffer(&d_plist->excl, h_plist->excl, 0, h_plist->nexcl,
828 stream, GpuApiCallBehavior::Async,
829 bDoTime ? nb->timers->pl_h2d[iloc].fetchNextEvent() : nullptr);
833 nb->timers->pl_h2d[iloc].closeTimingRegion(stream);
836 /* need to prune the pair list during the next step */
837 d_plist->haveFreshList = true;
840 //! This function is documented in the header file
841 void nbnxn_gpu_upload_shiftvec(gmx_nbnxn_ocl_t *nb,
842 const nbnxn_atomdata_t *nbatom)
844 cl_atomdata_t *adat = nb->atdat;
845 cl_command_queue ls = nb->stream[eintLocal];
847 /* only if we have a dynamic box */
848 if (nbatom->bDynamicBox || !adat->bShiftVecUploaded)
850 ocl_copy_H2D_async(adat->shift_vec, nbatom->shift_vec.data(), 0,
851 SHIFTS * adat->shift_vec_elem_size, ls, nullptr);
852 adat->bShiftVecUploaded = CL_TRUE;
856 //! This function is documented in the header file
857 void nbnxn_gpu_init_atomdata(gmx_nbnxn_ocl_t *nb,
858 const nbnxn_atomdata_t *nbat)
863 bool bDoTime = nb->bDoTime == CL_TRUE;
864 cl_timers_t *timers = nb->timers;
865 cl_atomdata_t *d_atdat = nb->atdat;
866 cl_command_queue ls = nb->stream[eintLocal];
868 natoms = nbat->numAtoms();
873 /* time async copy */
874 timers->atdat.openTimingRegion(ls);
877 /* need to reallocate if we have to copy more atoms than the amount of space
878 available and only allocate if we haven't initialized yet, i.e d_atdat->natoms == -1 */
879 if (natoms > d_atdat->nalloc)
881 nalloc = over_alloc_small(natoms);
883 /* free up first if the arrays have already been initialized */
884 if (d_atdat->nalloc != -1)
886 freeDeviceBuffer(&d_atdat->f);
887 freeDeviceBuffer(&d_atdat->xq);
888 freeDeviceBuffer(&d_atdat->lj_comb);
889 freeDeviceBuffer(&d_atdat->atom_types);
892 d_atdat->f_elem_size = sizeof(rvec);
894 d_atdat->f = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
895 nalloc * d_atdat->f_elem_size, nullptr, &cl_error);
896 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
897 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
899 d_atdat->xq = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
900 nalloc * sizeof(cl_float4), nullptr, &cl_error);
901 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
902 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
904 if (useLjCombRule(nb->nbparam->vdwtype))
906 d_atdat->lj_comb = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
907 nalloc * sizeof(cl_float2), nullptr, &cl_error);
908 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
909 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
913 d_atdat->atom_types = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
914 nalloc * sizeof(int), nullptr, &cl_error);
915 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
916 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
919 d_atdat->nalloc = nalloc;
923 d_atdat->natoms = natoms;
924 d_atdat->natoms_local = nbat->natoms_local;
926 /* need to clear GPU f output if realloc happened */
929 nbnxn_ocl_clear_f(nb, nalloc);
932 if (useLjCombRule(nb->nbparam->vdwtype))
934 ocl_copy_H2D_async(d_atdat->lj_comb, nbat->params().lj_comb.data(), 0,
935 natoms*sizeof(cl_float2), ls, bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
939 ocl_copy_H2D_async(d_atdat->atom_types, nbat->params().type.data(), 0,
940 natoms*sizeof(int), ls, bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
946 timers->atdat.closeTimingRegion(ls);
949 /* kick off the tasks enqueued above to ensure concurrency with the search */
950 cl_error = clFlush(ls);
951 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
952 ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
955 /*! \brief Releases an OpenCL kernel pointer */
956 static void free_kernel(cl_kernel *kernel_ptr)
958 cl_int gmx_unused cl_error;
960 assert(nullptr != kernel_ptr);
964 cl_error = clReleaseKernel(*kernel_ptr);
965 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS, ("clReleaseKernel failed: " + ocl_get_error_string(cl_error)).c_str());
967 *kernel_ptr = nullptr;
971 /*! \brief Releases a list of OpenCL kernel pointers */
972 static void free_kernels(cl_kernel *kernels, int count)
976 for (i = 0; i < count; i++)
978 free_kernel(kernels + i);
982 /*! \brief Free the OpenCL runtime data (context and program).
984 * The function releases the OpenCL context and program assuciated with the
985 * device that the calling PP rank is running on.
987 * \param runData [in] porinter to the structure with runtime data.
989 static void free_gpu_device_runtime_data(gmx_device_runtime_data_t *runData)
991 if (runData == nullptr)
996 cl_int gmx_unused cl_error;
998 if (runData->context)
1000 cl_error = clReleaseContext(runData->context);
1001 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS, ("clReleaseContext failed: " + ocl_get_error_string(cl_error)).c_str());
1002 runData->context = nullptr;
1005 if (runData->program)
1007 cl_error = clReleaseProgram(runData->program);
1008 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS, ("clReleaseProgram failed: " + ocl_get_error_string(cl_error)).c_str());
1009 runData->program = nullptr;
1014 //! This function is documented in the header file
1015 void nbnxn_gpu_free(gmx_nbnxn_ocl_t *nb)
1023 int kernel_count = sizeof(nb->kernel_ener_noprune_ptr) / sizeof(nb->kernel_ener_noprune_ptr[0][0]);
1024 free_kernels(nb->kernel_ener_noprune_ptr[0], kernel_count);
1026 kernel_count = sizeof(nb->kernel_ener_prune_ptr) / sizeof(nb->kernel_ener_prune_ptr[0][0]);
1027 free_kernels(nb->kernel_ener_prune_ptr[0], kernel_count);
1029 kernel_count = sizeof(nb->kernel_noener_noprune_ptr) / sizeof(nb->kernel_noener_noprune_ptr[0][0]);
1030 free_kernels(nb->kernel_noener_noprune_ptr[0], kernel_count);
1032 kernel_count = sizeof(nb->kernel_noener_prune_ptr) / sizeof(nb->kernel_noener_prune_ptr[0][0]);
1033 free_kernels(nb->kernel_noener_prune_ptr[0], kernel_count);
1035 free_kernel(&(nb->kernel_zero_e_fshift));
1038 freeDeviceBuffer(&(nb->atdat->xq));
1039 freeDeviceBuffer(&(nb->atdat->f));
1040 freeDeviceBuffer(&(nb->atdat->e_lj));
1041 freeDeviceBuffer(&(nb->atdat->e_el));
1042 freeDeviceBuffer(&(nb->atdat->fshift));
1043 freeDeviceBuffer(&(nb->atdat->lj_comb));
1044 freeDeviceBuffer(&(nb->atdat->atom_types));
1045 freeDeviceBuffer(&(nb->atdat->shift_vec));
1049 freeDeviceBuffer(&(nb->nbparam->nbfp_climg2d));
1050 freeDeviceBuffer(&(nb->nbparam->nbfp_comb_climg2d));
1051 freeDeviceBuffer(&(nb->nbparam->coulomb_tab_climg2d));
1055 auto *plist = nb->plist[eintLocal];
1056 freeDeviceBuffer(&plist->sci);
1057 freeDeviceBuffer(&plist->cj4);
1058 freeDeviceBuffer(&plist->imask);
1059 freeDeviceBuffer(&plist->excl);
1061 if (nb->bUseTwoStreams)
1063 auto *plist_nl = nb->plist[eintNonlocal];
1064 freeDeviceBuffer(&plist_nl->sci);
1065 freeDeviceBuffer(&plist_nl->cj4);
1066 freeDeviceBuffer(&plist_nl->imask);
1067 freeDeviceBuffer(&plist_nl->excl);
1072 pfree(nb->nbst.e_lj);
1073 nb->nbst.e_lj = nullptr;
1075 pfree(nb->nbst.e_el);
1076 nb->nbst.e_el = nullptr;
1078 pfree(nb->nbst.fshift);
1079 nb->nbst.fshift = nullptr;
1081 /* Free command queues */
1082 clReleaseCommandQueue(nb->stream[eintLocal]);
1083 nb->stream[eintLocal] = nullptr;
1084 if (nb->bUseTwoStreams)
1086 clReleaseCommandQueue(nb->stream[eintNonlocal]);
1087 nb->stream[eintNonlocal] = nullptr;
1089 /* Free other events */
1090 if (nb->nonlocal_done)
1092 clReleaseEvent(nb->nonlocal_done);
1093 nb->nonlocal_done = nullptr;
1095 if (nb->misc_ops_and_local_H2D_done)
1097 clReleaseEvent(nb->misc_ops_and_local_H2D_done);
1098 nb->misc_ops_and_local_H2D_done = nullptr;
1101 free_gpu_device_runtime_data(nb->dev_rundata);
1102 sfree(nb->dev_rundata);
1104 /* Free timers and timings */
1111 fprintf(debug, "Cleaned up OpenCL data structures.\n");
1115 //! This function is documented in the header file
1116 gmx_wallclock_gpu_nbnxn_t *nbnxn_gpu_get_timings(gmx_nbnxn_ocl_t *nb)
1118 return (nb != nullptr && nb->bDoTime) ? nb->timings : nullptr;
1121 //! This function is documented in the header file
1122 void nbnxn_gpu_reset_timings(nonbonded_verlet_t* nbv)
1124 if (nbv->gpu_nbv && nbv->gpu_nbv->bDoTime)
1126 init_timings(nbv->gpu_nbv->timings);
1130 //! This function is documented in the header file
1131 int nbnxn_gpu_min_ci_balanced(gmx_nbnxn_ocl_t *nb)
1133 return nb != nullptr ?
1134 gpu_min_ci_balanced_factor * nb->dev_info->compute_units : 0;
1137 //! This function is documented in the header file
1138 gmx_bool nbnxn_gpu_is_kernel_ewald_analytical(const gmx_nbnxn_ocl_t *nb)
1140 return ((nb->nbparam->eeltype == eelOclEWALD_ANA) ||
1141 (nb->nbparam->eeltype == eelOclEWALD_ANA_TWIN));