2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016,2017,2018,2019, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
36 * \brief Define OpenCL implementation of nbnxm_gpu_data_mgmt.h
38 * \author Anca Hamuraru <anca@streamcomputing.eu>
39 * \author Dimitrios Karkoulis <dimitris.karkoulis@gmail.com>
40 * \author Teemu Virolainen <teemu@streamcomputing.eu>
41 * \author Szilárd Páll <pall.szilard@gmail.com>
42 * \ingroup module_nbnxm
54 // TODO We would like to move this down, but the way gmx_nbnxn_gpu_t
55 // is currently declared means this has to be before gpu_types.h
56 #include "nbnxm_ocl_types.h"
58 // TODO Remove this comment when the above order issue is resolved
59 #include "gromacs/gpu_utils/gpu_utils.h"
60 #include "gromacs/gpu_utils/oclutils.h"
61 #include "gromacs/hardware/gpu_hw_info.h"
62 #include "gromacs/math/vectypes.h"
63 #include "gromacs/mdlib/force_flags.h"
64 #include "gromacs/mdtypes/interaction_const.h"
65 #include "gromacs/mdtypes/md_enums.h"
66 #include "gromacs/nbnxm/atomdata.h"
67 #include "gromacs/nbnxm/gpu_data_mgmt.h"
68 #include "gromacs/nbnxm/gpu_jit_support.h"
69 #include "gromacs/nbnxm/nbnxm.h"
70 #include "gromacs/nbnxm/nbnxm_gpu.h"
71 #include "gromacs/nbnxm/pairlistsets.h"
72 #include "gromacs/pbcutil/ishift.h"
73 #include "gromacs/timing/gpu_timing.h"
74 #include "gromacs/utility/cstringutil.h"
75 #include "gromacs/utility/fatalerror.h"
76 #include "gromacs/utility/gmxassert.h"
77 #include "gromacs/utility/real.h"
78 #include "gromacs/utility/smalloc.h"
80 #include "nbnxm_ocl_internal.h"
85 /*! \brief Copies of values from cl_driver_diagnostics_intel.h,
86 * which isn't guaranteed to be available. */
88 #define CL_CONTEXT_SHOW_DIAGNOSTICS_INTEL 0x4106
89 #define CL_CONTEXT_DIAGNOSTICS_LEVEL_GOOD_INTEL 0x1
90 #define CL_CONTEXT_DIAGNOSTICS_LEVEL_BAD_INTEL 0x2
91 #define CL_CONTEXT_DIAGNOSTICS_LEVEL_NEUTRAL_INTEL 0x4
94 /*! \brief This parameter should be determined heuristically from the
95 * kernel execution times
97 * This value is best for small systems on a single AMD Radeon R9 290X
98 * (and about 5% faster than 40, which is the default for CUDA
99 * devices). Larger simulation systems were quite insensitive to the
100 * value of this parameter.
102 static unsigned int gpu_min_ci_balanced_factor = 50;
105 /*! \brief Returns true if LJ combination rules are used in the non-bonded kernels.
107 * Full doc in nbnxn_ocl_internal.h */
108 bool useLjCombRule(int vdwType)
110 return (vdwType == evdwOclCUTCOMBGEOM || vdwType == evdwOclCUTCOMBLB);
113 /*! \brief Tabulates the Ewald Coulomb force and initializes the size/scale
114 * and the table GPU array.
116 * If called with an already allocated table, it just re-uploads the
119 static void init_ewald_coulomb_force_table(const EwaldCorrectionTables& tables,
121 const gmx_device_runtime_data_t* runData)
127 if (nbp->coulomb_tab_climg2d != nullptr)
129 freeDeviceBuffer(&(nbp->coulomb_tab_climg2d));
132 /* Switched from using textures to using buffers */
133 // TODO: decide which alternative is most efficient - textures or buffers.
135 cl_image_format array_format;
137 array_format.image_channel_data_type = CL_FLOAT;
138 array_format.image_channel_order = CL_R;
140 coul_tab = clCreateImage2D(runData->context, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
141 &array_format, tabsize, 1, 0, ftmp, &cl_error);
144 coul_tab = clCreateBuffer(
145 runData->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
146 tables.tableF.size() * sizeof(cl_float), const_cast<real*>(tables.tableF.data()), &cl_error);
147 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
148 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
150 nbp->coulomb_tab_climg2d = coul_tab;
151 nbp->coulomb_tab_scale = tables.scale;
155 /*! \brief Initializes the atomdata structure first time, it only gets filled at
158 static void init_atomdata_first(cl_atomdata_t* ad, int ntypes, gmx_device_runtime_data_t* runData)
164 /* An element of the shift_vec device buffer has the same size as one element
165 of the host side shift_vec buffer. */
166 ad->shift_vec_elem_size = sizeof(*nbnxn_atomdata_t::shift_vec.data());
168 ad->shift_vec = clCreateBuffer(runData->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
169 SHIFTS * ad->shift_vec_elem_size, nullptr, &cl_error);
170 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
171 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
172 ad->bShiftVecUploaded = CL_FALSE;
174 /* An element of the fshift device buffer has the same size as one element
175 of the host side fshift buffer. */
176 ad->fshift_elem_size = sizeof(*cl_nb_staging_t::fshift);
178 ad->fshift = clCreateBuffer(runData->context, CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
179 SHIFTS * ad->fshift_elem_size, nullptr, &cl_error);
180 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
181 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
183 ad->e_lj = clCreateBuffer(runData->context, CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
184 sizeof(float), nullptr, &cl_error);
185 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
186 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
188 ad->e_el = clCreateBuffer(runData->context, CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
189 sizeof(float), nullptr, &cl_error);
190 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
191 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
193 /* initialize to nullptr pointers to data that is not allocated here and will
194 need reallocation in nbnxn_gpu_init_atomdata */
198 /* size -1 indicates that the respective array hasn't been initialized yet */
203 /*! \brief Copies all parameters related to the cut-off from ic to nbp
205 static void set_cutoff_parameters(cl_nbparam_t* nbp, const interaction_const_t* ic, const PairlistParams& listParams)
207 nbp->ewald_beta = ic->ewaldcoeff_q;
208 nbp->sh_ewald = ic->sh_ewald;
209 nbp->epsfac = ic->epsfac;
210 nbp->two_k_rf = 2.0 * ic->k_rf;
211 nbp->c_rf = ic->c_rf;
212 nbp->rvdw_sq = ic->rvdw * ic->rvdw;
213 nbp->rcoulomb_sq = ic->rcoulomb * ic->rcoulomb;
214 nbp->rlistOuter_sq = listParams.rlistOuter * listParams.rlistOuter;
215 nbp->rlistInner_sq = listParams.rlistInner * listParams.rlistInner;
216 nbp->useDynamicPruning = listParams.useDynamicPruning;
218 nbp->sh_lj_ewald = ic->sh_lj_ewald;
219 nbp->ewaldcoeff_lj = ic->ewaldcoeff_lj;
221 nbp->rvdw_switch = ic->rvdw_switch;
222 nbp->dispersion_shift = ic->dispersion_shift;
223 nbp->repulsion_shift = ic->repulsion_shift;
224 nbp->vdw_switch = ic->vdw_switch;
227 /*! \brief Returns the kinds of electrostatics and Vdw OpenCL
228 * kernels that will be used.
230 * Respectively, these values are from enum eelOcl and enum
232 static void map_interaction_types_to_gpu_kernel_flavors(const interaction_const_t* ic,
237 if (ic->vdwtype == evdwCUT)
239 switch (ic->vdw_modifier)
242 case eintmodPOTSHIFT:
245 case ljcrNONE: *gpu_vdwtype = evdwOclCUT; break;
246 case ljcrGEOM: *gpu_vdwtype = evdwOclCUTCOMBGEOM; break;
247 case ljcrLB: *gpu_vdwtype = evdwOclCUTCOMBLB; break;
250 "The requested LJ combination rule is not implemented in the "
251 "OpenCL GPU accelerated kernels!");
254 case eintmodFORCESWITCH: *gpu_vdwtype = evdwOclFSWITCH; break;
255 case eintmodPOTSWITCH: *gpu_vdwtype = evdwOclPSWITCH; break;
258 "The requested VdW interaction modifier is not implemented in the GPU "
259 "accelerated kernels!");
262 else if (ic->vdwtype == evdwPME)
264 if (ic->ljpme_comb_rule == ljcrGEOM)
266 *gpu_vdwtype = evdwOclEWALDGEOM;
270 *gpu_vdwtype = evdwOclEWALDLB;
275 gmx_incons("The requested VdW type is not implemented in the GPU accelerated kernels!");
278 if (ic->eeltype == eelCUT)
280 *gpu_eeltype = eelOclCUT;
282 else if (EEL_RF(ic->eeltype))
284 *gpu_eeltype = eelOclRF;
286 else if ((EEL_PME(ic->eeltype) || ic->eeltype == eelEWALD))
288 *gpu_eeltype = nbnxn_gpu_pick_ewald_kernel_type(*ic);
292 /* Shouldn't happen, as this is checked when choosing Verlet-scheme */
294 "The requested electrostatics type is not implemented in the GPU accelerated "
299 /*! \brief Initializes the nonbonded parameter data structure.
301 static void init_nbparam(cl_nbparam_t* nbp,
302 const interaction_const_t* ic,
303 const PairlistParams& listParams,
304 const nbnxn_atomdata_t::Params& nbatParams,
305 const gmx_device_runtime_data_t* runData)
309 set_cutoff_parameters(nbp, ic, listParams);
311 map_interaction_types_to_gpu_kernel_flavors(ic, nbatParams.comb_rule, &(nbp->eeltype), &(nbp->vdwtype));
313 if (ic->vdwtype == evdwPME)
315 if (ic->ljpme_comb_rule == ljcrGEOM)
317 GMX_ASSERT(nbatParams.comb_rule == ljcrGEOM, "Combination rule mismatch!");
321 GMX_ASSERT(nbatParams.comb_rule == ljcrLB, "Combination rule mismatch!");
324 /* generate table for PME */
325 nbp->coulomb_tab_climg2d = nullptr;
326 if (nbp->eeltype == eelOclEWALD_TAB || nbp->eeltype == eelOclEWALD_TAB_TWIN)
328 GMX_RELEASE_ASSERT(ic->coulombEwaldTables, "Need valid Coulomb Ewald correction tables");
329 init_ewald_coulomb_force_table(*ic->coulombEwaldTables, nbp, runData);
332 // TODO: improvement needed.
333 // The image2d is created here even if eeltype is not eelCuEWALD_TAB or eelCuEWALD_TAB_TWIN
334 // because the OpenCL kernels don't accept nullptr values for image2D parameters.
336 /* Switched from using textures to using buffers */
337 // TODO: decide which alternative is most efficient - textures or buffers.
339 cl_image_format array_format;
341 array_format.image_channel_data_type = CL_FLOAT;
342 array_format.image_channel_order = CL_R;
344 nbp->coulomb_tab_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE,
345 &array_format, 1, 1, 0, nullptr, &cl_error);
348 nbp->coulomb_tab_climg2d = clCreateBuffer(runData->context, CL_MEM_READ_ONLY,
349 sizeof(cl_float), nullptr, &cl_error);
350 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
351 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
354 const int nnbfp = 2 * nbatParams.numTypes * nbatParams.numTypes;
355 const int nnbfp_comb = 2 * nbatParams.numTypes;
358 /* Switched from using textures to using buffers */
359 // TODO: decide which alternative is most efficient - textures or buffers.
361 cl_image_format array_format;
363 array_format.image_channel_data_type = CL_FLOAT;
364 array_format.image_channel_order = CL_R;
366 nbp->nbfp_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_ONLY |
367 CL_MEM_COPY_HOST_PTR, &array_format, nnbfp, 1, 0, nbat->nbfp, &cl_error);
370 nbp->nbfp_climg2d = clCreateBuffer(
371 runData->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
372 nnbfp * sizeof(cl_float), const_cast<float*>(nbatParams.nbfp.data()), &cl_error);
373 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
374 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
376 if (ic->vdwtype == evdwPME)
378 /* Switched from using textures to using buffers */
379 // TODO: decide which alternative is most efficient - textures or buffers.
380 /* nbp->nbfp_comb_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE |
381 CL_MEM_COPY_HOST_PTR, &array_format, nnbfp_comb, 1, 0, nbat->nbfp_comb, &cl_error);*/
382 nbp->nbfp_comb_climg2d = clCreateBuffer(
383 runData->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
384 nnbfp_comb * sizeof(cl_float), const_cast<float*>(nbatParams.nbfp_comb.data()),
386 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
387 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
391 // TODO: improvement needed.
392 // The image2d is created here even if vdwtype is not evdwPME because the OpenCL kernels
393 // don't accept nullptr values for image2D parameters.
394 /* Switched from using textures to using buffers */
395 // TODO: decide which alternative is most efficient - textures or buffers.
396 /* nbp->nbfp_comb_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE,
397 &array_format, 1, 1, 0, nullptr, &cl_error);*/
398 nbp->nbfp_comb_climg2d = clCreateBuffer(runData->context, CL_MEM_READ_ONLY,
399 sizeof(cl_float), nullptr, &cl_error);
400 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
401 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
406 //! This function is documented in the header file
407 void gpu_pme_loadbal_update_param(const nonbonded_verlet_t* nbv, const interaction_const_t* ic)
409 if (!nbv || !nbv->useGpu())
413 gmx_nbnxn_ocl_t* nb = nbv->gpu_nbv;
414 cl_nbparam_t* nbp = nb->nbparam;
416 set_cutoff_parameters(nbp, ic, nbv->pairlistSets().params());
418 nbp->eeltype = nbnxn_gpu_pick_ewald_kernel_type(*ic);
420 GMX_RELEASE_ASSERT(ic->coulombEwaldTables, "Need valid Coulomb Ewald correction tables");
421 init_ewald_coulomb_force_table(*ic->coulombEwaldTables, nbp, nb->dev_rundata);
424 /*! \brief Initializes the pair list data structure.
426 static void init_plist(cl_plist_t* pl)
428 /* initialize to nullptr pointers to data that is not allocated here and will
429 need reallocation in nbnxn_gpu_init_pairlist */
435 /* size -1 indicates that the respective array hasn't been initialized yet */
442 pl->imask_nalloc = -1;
444 pl->excl_nalloc = -1;
445 pl->haveFreshList = false;
448 /*! \brief Initializes the timings data structure.
450 static void init_timings(gmx_wallclock_gpu_nbnxn_t* t)
459 for (i = 0; i < 2; i++)
461 for (j = 0; j < 2; j++)
463 t->ktime[i][j].t = 0.0;
464 t->ktime[i][j].c = 0;
469 t->pruneTime.t = 0.0;
470 t->dynamicPruneTime.c = 0;
471 t->dynamicPruneTime.t = 0.0;
475 //! OpenCL notification callback function
476 static void CL_CALLBACK ocl_notify_fn(const char* pErrInfo,
477 const void gmx_unused* private_info,
478 size_t gmx_unused cb,
479 void gmx_unused* user_data)
481 if (pErrInfo != nullptr)
483 printf("%s\n", pErrInfo); // Print error/hint
487 /*! \brief Creates context for OpenCL GPU given by \p mygpu
489 * A fatal error results if creation fails.
491 * \param[inout] runtimeData runtime data including program and context
492 * \param[in] devInfo device info struct
493 * \param[in] rank MPI rank (for error reporting)
495 static void nbnxn_gpu_create_context(gmx_device_runtime_data_t* runtimeData,
496 const gmx_device_info_t* devInfo,
499 cl_context_properties context_properties[5];
500 cl_platform_id platform_id;
501 cl_device_id device_id;
505 assert(runtimeData != nullptr);
506 assert(devInfo != nullptr);
508 platform_id = devInfo->ocl_gpu_id.ocl_platform_id;
509 device_id = devInfo->ocl_gpu_id.ocl_device_id;
512 context_properties[i++] = CL_CONTEXT_PLATFORM;
513 context_properties[i++] = reinterpret_cast<cl_context_properties>(platform_id);
514 if (getenv("GMX_OCL_SHOW_DIAGNOSTICS"))
516 context_properties[i++] = CL_CONTEXT_SHOW_DIAGNOSTICS_INTEL;
517 context_properties[i++] =
518 CL_CONTEXT_DIAGNOSTICS_LEVEL_BAD_INTEL | CL_CONTEXT_DIAGNOSTICS_LEVEL_NEUTRAL_INTEL;
520 context_properties[i++] = 0; /* Terminates the list of properties */
522 context = clCreateContext(context_properties, 1, &device_id, ocl_notify_fn, nullptr, &cl_error);
523 if (CL_SUCCESS != cl_error)
525 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s:\n OpenCL error %d: %s",
526 rank, devInfo->device_name, cl_error, ocl_get_error_string(cl_error).c_str());
529 runtimeData->context = context;
532 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
533 static cl_kernel nbnxn_gpu_create_kernel(gmx_nbnxn_ocl_t* nb, const char* kernel_name)
538 kernel = clCreateKernel(nb->dev_rundata->program, kernel_name, &cl_error);
539 if (CL_SUCCESS != cl_error)
541 gmx_fatal(FARGS, "Failed to create kernel '%s' for GPU #%s: OpenCL error %d", kernel_name,
542 nb->dev_info->device_name, cl_error);
548 /*! \brief Clears nonbonded shift force output array and energy outputs on the GPU.
550 static void nbnxn_ocl_clear_e_fshift(gmx_nbnxn_ocl_t* nb)
554 cl_atomdata_t* adat = nb->atdat;
555 cl_command_queue ls = nb->stream[InteractionLocality::Local];
557 size_t local_work_size[3] = { 1, 1, 1 };
558 size_t global_work_size[3] = { 1, 1, 1 };
560 cl_int shifts = SHIFTS * 3;
564 cl_kernel zero_e_fshift = nb->kernel_zero_e_fshift;
566 local_work_size[0] = 64;
567 // Round the total number of threads up from the array size
568 global_work_size[0] = ((shifts + local_work_size[0] - 1) / local_work_size[0]) * local_work_size[0];
571 cl_error = clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->fshift));
572 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_lj));
573 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_el));
574 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_uint), &shifts);
575 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
577 cl_error = clEnqueueNDRangeKernel(ls, zero_e_fshift, 3, nullptr, global_work_size,
578 local_work_size, 0, nullptr, nullptr);
579 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
582 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
583 static void nbnxn_gpu_init_kernels(gmx_nbnxn_ocl_t* nb)
585 /* Init to 0 main kernel arrays */
586 /* They will be later on initialized in select_nbnxn_kernel */
587 // TODO: consider always creating all variants of the kernels here so that there is no
588 // need for late call to clCreateKernel -- if that gives any advantage?
589 memset(nb->kernel_ener_noprune_ptr, 0, sizeof(nb->kernel_ener_noprune_ptr));
590 memset(nb->kernel_ener_prune_ptr, 0, sizeof(nb->kernel_ener_prune_ptr));
591 memset(nb->kernel_noener_noprune_ptr, 0, sizeof(nb->kernel_noener_noprune_ptr));
592 memset(nb->kernel_noener_prune_ptr, 0, sizeof(nb->kernel_noener_prune_ptr));
594 /* Init pruning kernels
596 * TODO: we could avoid creating kernels if dynamic pruning is turned off,
597 * but ATM that depends on force flags not passed into the initialization.
599 nb->kernel_pruneonly[epruneFirst] = nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_opencl");
600 nb->kernel_pruneonly[epruneRolling] =
601 nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_rolling_opencl");
603 /* Init auxiliary kernels */
604 nb->kernel_zero_e_fshift = nbnxn_gpu_create_kernel(nb, "zero_e_fshift");
607 /*! \brief Initializes simulation constant data.
609 * Initializes members of the atomdata and nbparam structs and
610 * clears e/fshift output buffers.
612 static void nbnxn_ocl_init_const(gmx_nbnxn_ocl_t* nb,
613 const interaction_const_t* ic,
614 const PairlistParams& listParams,
615 const nbnxn_atomdata_t::Params& nbatParams)
617 init_atomdata_first(nb->atdat, nbatParams.numTypes, nb->dev_rundata);
618 init_nbparam(nb->nbparam, ic, listParams, nbatParams, nb->dev_rundata);
622 //! This function is documented in the header file
623 gmx_nbnxn_ocl_t* gpu_init(const gmx_device_info_t* deviceInfo,
624 const interaction_const_t* ic,
625 const PairlistParams& listParams,
626 const nbnxn_atomdata_t* nbat,
628 const gmx_bool bLocalAndNonlocal)
632 cl_command_queue_properties queue_properties;
638 snew(nb->nbparam, 1);
639 snew(nb->plist[InteractionLocality::Local], 1);
640 if (bLocalAndNonlocal)
642 snew(nb->plist[InteractionLocality::NonLocal], 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[InteractionLocality::Local]);
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[InteractionLocality::Local] = clCreateCommandQueue(
678 nb->dev_rundata->context, nb->dev_info->ocl_gpu_id.ocl_device_id, queue_properties, &cl_error);
679 if (CL_SUCCESS != cl_error)
681 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s: OpenCL error %d", rank,
682 nb->dev_info->device_name, cl_error);
685 if (nb->bUseTwoStreams)
687 init_plist(nb->plist[InteractionLocality::NonLocal]);
689 nb->stream[InteractionLocality::NonLocal] =
690 clCreateCommandQueue(nb->dev_rundata->context, nb->dev_info->ocl_gpu_id.ocl_device_id,
691 queue_properties, &cl_error);
692 if (CL_SUCCESS != cl_error)
694 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s: OpenCL error %d",
695 rank, nb->dev_info->device_name, cl_error);
701 init_timings(nb->timings);
704 nbnxn_ocl_init_const(nb, ic, listParams, nbat->params());
706 /* Enable LJ param manual prefetch for AMD or Intel or if we request through env. var.
707 * TODO: decide about NVIDIA
709 nb->bPrefetchLjParam = (getenv("GMX_OCL_DISABLE_I_PREFETCH") == nullptr)
710 && ((nb->dev_info->vendor_e == OCL_VENDOR_AMD)
711 || (nb->dev_info->vendor_e == OCL_VENDOR_INTEL)
712 || (getenv("GMX_OCL_ENABLE_I_PREFETCH") != nullptr));
714 /* NOTE: in CUDA we pick L1 cache configuration for the nbnxn kernels here,
715 * but sadly this is not supported in OpenCL (yet?). Consider adding it if
716 * it becomes supported.
718 nbnxn_gpu_compile_kernels(nb);
719 nbnxn_gpu_init_kernels(nb);
721 /* clear energy and shift force outputs */
722 nbnxn_ocl_clear_e_fshift(nb);
726 fprintf(debug, "Initialized OpenCL data structures.\n");
732 /*! \brief Clears the first natoms_clear elements of the GPU nonbonded force output array.
734 static void nbnxn_ocl_clear_f(gmx_nbnxn_ocl_t* nb, int natoms_clear)
736 if (natoms_clear == 0)
741 cl_int gmx_used_in_debug cl_error;
743 cl_atomdata_t* atomData = nb->atdat;
744 cl_command_queue ls = nb->stream[InteractionLocality::Local];
745 cl_float value = 0.0F;
747 cl_error = clEnqueueFillBuffer(ls, atomData->f, &value, sizeof(cl_float), 0,
748 natoms_clear * sizeof(rvec), 0, nullptr, nullptr);
749 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
750 ("clEnqueueFillBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
753 //! This function is documented in the header file
754 void gpu_clear_outputs(gmx_nbnxn_ocl_t* nb, bool computeVirial)
756 nbnxn_ocl_clear_f(nb, nb->atdat->natoms);
757 /* clear shift force array and energies if the outputs were
758 used in the current step */
761 nbnxn_ocl_clear_e_fshift(nb);
764 /* kick off buffer clearing kernel to ensure concurrency with constraints/update */
765 cl_int gmx_unused cl_error;
766 cl_error = clFlush(nb->stream[InteractionLocality::Local]);
767 assert(CL_SUCCESS == cl_error);
770 //! This function is documented in the header file
771 void gpu_init_pairlist(gmx_nbnxn_ocl_t* nb, const NbnxnPairlistGpu* h_plist, const InteractionLocality iloc)
774 // Timing accumulation should happen only if there was work to do
775 // because getLastRangeTime() gets skipped with empty lists later
776 // which leads to the counter not being reset.
777 bool bDoTime = ((nb->bDoTime == CL_TRUE) && !h_plist->sci.empty());
778 cl_command_queue stream = nb->stream[iloc];
779 cl_plist_t* d_plist = nb->plist[iloc];
781 if (d_plist->na_c < 0)
783 d_plist->na_c = h_plist->na_ci;
787 if (d_plist->na_c != h_plist->na_ci)
789 sprintf(sbuf, "In cu_init_plist: the #atoms per cell has changed (from %d to %d)",
790 d_plist->na_c, h_plist->na_ci);
795 gpu_timers_t::Interaction& iTimers = nb->timers->interaction[iloc];
799 iTimers.pl_h2d.openTimingRegion(stream);
800 iTimers.didPairlistH2D = true;
803 // TODO most of this function is same in CUDA and OpenCL, move into the header
804 DeviceContext context = nb->dev_rundata->context;
806 reallocateDeviceBuffer(&d_plist->sci, h_plist->sci.size(), &d_plist->nsci, &d_plist->sci_nalloc, context);
807 copyToDeviceBuffer(&d_plist->sci, h_plist->sci.data(), 0, h_plist->sci.size(), stream,
808 GpuApiCallBehavior::Async, bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
810 reallocateDeviceBuffer(&d_plist->cj4, h_plist->cj4.size(), &d_plist->ncj4, &d_plist->cj4_nalloc, context);
811 copyToDeviceBuffer(&d_plist->cj4, h_plist->cj4.data(), 0, h_plist->cj4.size(), stream,
812 GpuApiCallBehavior::Async, bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
814 reallocateDeviceBuffer(&d_plist->imask, h_plist->cj4.size() * c_nbnxnGpuClusterpairSplit,
815 &d_plist->nimask, &d_plist->imask_nalloc, context);
817 reallocateDeviceBuffer(&d_plist->excl, h_plist->excl.size(), &d_plist->nexcl,
818 &d_plist->excl_nalloc, context);
819 copyToDeviceBuffer(&d_plist->excl, h_plist->excl.data(), 0, h_plist->excl.size(), stream,
820 GpuApiCallBehavior::Async, bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
824 iTimers.pl_h2d.closeTimingRegion(stream);
827 /* need to prune the pair list during the next step */
828 d_plist->haveFreshList = true;
831 //! This function is documented in the header file
832 void gpu_upload_shiftvec(gmx_nbnxn_ocl_t* nb, const nbnxn_atomdata_t* nbatom)
834 cl_atomdata_t* adat = nb->atdat;
835 cl_command_queue ls = nb->stream[InteractionLocality::Local];
837 /* only if we have a dynamic box */
838 if (nbatom->bDynamicBox || !adat->bShiftVecUploaded)
840 ocl_copy_H2D_async(adat->shift_vec, nbatom->shift_vec.data(), 0,
841 SHIFTS * adat->shift_vec_elem_size, ls, nullptr);
842 adat->bShiftVecUploaded = CL_TRUE;
846 //! This function is documented in the header file
847 void gpu_init_atomdata(gmx_nbnxn_ocl_t* nb, const nbnxn_atomdata_t* nbat)
852 bool bDoTime = nb->bDoTime == CL_TRUE;
853 cl_timers_t* timers = nb->timers;
854 cl_atomdata_t* d_atdat = nb->atdat;
855 cl_command_queue ls = nb->stream[InteractionLocality::Local];
857 natoms = nbat->numAtoms();
862 /* time async copy */
863 timers->atdat.openTimingRegion(ls);
866 /* need to reallocate if we have to copy more atoms than the amount of space
867 available and only allocate if we haven't initialized yet, i.e d_atdat->natoms == -1 */
868 if (natoms > d_atdat->nalloc)
870 nalloc = over_alloc_small(natoms);
872 /* free up first if the arrays have already been initialized */
873 if (d_atdat->nalloc != -1)
875 freeDeviceBuffer(&d_atdat->f);
876 freeDeviceBuffer(&d_atdat->xq);
877 freeDeviceBuffer(&d_atdat->lj_comb);
878 freeDeviceBuffer(&d_atdat->atom_types);
881 d_atdat->f_elem_size = sizeof(rvec);
883 d_atdat->f = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
884 nalloc * d_atdat->f_elem_size, nullptr, &cl_error);
885 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
886 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
888 d_atdat->xq = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
889 nalloc * sizeof(cl_float4), nullptr, &cl_error);
890 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
891 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
893 if (useLjCombRule(nb->nbparam->vdwtype))
895 d_atdat->lj_comb = clCreateBuffer(nb->dev_rundata->context,
896 CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
897 nalloc * sizeof(cl_float2), nullptr, &cl_error);
898 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
899 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
903 d_atdat->atom_types = clCreateBuffer(nb->dev_rundata->context,
904 CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
905 nalloc * sizeof(int), nullptr, &cl_error);
906 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
907 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
910 d_atdat->nalloc = nalloc;
914 d_atdat->natoms = natoms;
915 d_atdat->natoms_local = nbat->natoms_local;
917 /* need to clear GPU f output if realloc happened */
920 nbnxn_ocl_clear_f(nb, nalloc);
923 if (useLjCombRule(nb->nbparam->vdwtype))
925 ocl_copy_H2D_async(d_atdat->lj_comb, nbat->params().lj_comb.data(), 0, natoms * sizeof(cl_float2),
926 ls, bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
930 ocl_copy_H2D_async(d_atdat->atom_types, nbat->params().type.data(), 0, natoms * sizeof(int),
931 ls, bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
936 timers->atdat.closeTimingRegion(ls);
939 /* kick off the tasks enqueued above to ensure concurrency with the search */
940 cl_error = clFlush(ls);
941 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
942 ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
945 /*! \brief Releases an OpenCL kernel pointer */
946 static void free_kernel(cl_kernel* kernel_ptr)
948 cl_int gmx_unused cl_error;
950 assert(nullptr != kernel_ptr);
954 cl_error = clReleaseKernel(*kernel_ptr);
955 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
956 ("clReleaseKernel failed: " + ocl_get_error_string(cl_error)).c_str());
958 *kernel_ptr = nullptr;
962 /*! \brief Releases a list of OpenCL kernel pointers */
963 static void free_kernels(cl_kernel* kernels, int count)
967 for (i = 0; i < count; i++)
969 free_kernel(kernels + i);
973 /*! \brief Free the OpenCL runtime data (context and program).
975 * The function releases the OpenCL context and program assuciated with the
976 * device that the calling PP rank is running on.
978 * \param runData [in] porinter to the structure with runtime data.
980 static void free_gpu_device_runtime_data(gmx_device_runtime_data_t* runData)
982 if (runData == nullptr)
987 cl_int gmx_unused cl_error;
989 if (runData->context)
991 cl_error = clReleaseContext(runData->context);
992 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
993 ("clReleaseContext failed: " + ocl_get_error_string(cl_error)).c_str());
994 runData->context = nullptr;
997 if (runData->program)
999 cl_error = clReleaseProgram(runData->program);
1000 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
1001 ("clReleaseProgram failed: " + ocl_get_error_string(cl_error)).c_str());
1002 runData->program = nullptr;
1006 //! This function is documented in the header file
1007 void gpu_free(gmx_nbnxn_ocl_t* nb)
1015 int kernel_count = sizeof(nb->kernel_ener_noprune_ptr) / sizeof(nb->kernel_ener_noprune_ptr[0][0]);
1016 free_kernels(nb->kernel_ener_noprune_ptr[0], kernel_count);
1018 kernel_count = sizeof(nb->kernel_ener_prune_ptr) / sizeof(nb->kernel_ener_prune_ptr[0][0]);
1019 free_kernels(nb->kernel_ener_prune_ptr[0], kernel_count);
1021 kernel_count = sizeof(nb->kernel_noener_noprune_ptr) / sizeof(nb->kernel_noener_noprune_ptr[0][0]);
1022 free_kernels(nb->kernel_noener_noprune_ptr[0], kernel_count);
1024 kernel_count = sizeof(nb->kernel_noener_prune_ptr) / sizeof(nb->kernel_noener_prune_ptr[0][0]);
1025 free_kernels(nb->kernel_noener_prune_ptr[0], kernel_count);
1027 free_kernel(&(nb->kernel_zero_e_fshift));
1030 freeDeviceBuffer(&(nb->atdat->xq));
1031 freeDeviceBuffer(&(nb->atdat->f));
1032 freeDeviceBuffer(&(nb->atdat->e_lj));
1033 freeDeviceBuffer(&(nb->atdat->e_el));
1034 freeDeviceBuffer(&(nb->atdat->fshift));
1035 freeDeviceBuffer(&(nb->atdat->lj_comb));
1036 freeDeviceBuffer(&(nb->atdat->atom_types));
1037 freeDeviceBuffer(&(nb->atdat->shift_vec));
1041 freeDeviceBuffer(&(nb->nbparam->nbfp_climg2d));
1042 freeDeviceBuffer(&(nb->nbparam->nbfp_comb_climg2d));
1043 freeDeviceBuffer(&(nb->nbparam->coulomb_tab_climg2d));
1047 auto* plist = nb->plist[InteractionLocality::Local];
1048 freeDeviceBuffer(&plist->sci);
1049 freeDeviceBuffer(&plist->cj4);
1050 freeDeviceBuffer(&plist->imask);
1051 freeDeviceBuffer(&plist->excl);
1053 if (nb->bUseTwoStreams)
1055 auto* plist_nl = nb->plist[InteractionLocality::NonLocal];
1056 freeDeviceBuffer(&plist_nl->sci);
1057 freeDeviceBuffer(&plist_nl->cj4);
1058 freeDeviceBuffer(&plist_nl->imask);
1059 freeDeviceBuffer(&plist_nl->excl);
1064 pfree(nb->nbst.e_lj);
1065 nb->nbst.e_lj = nullptr;
1067 pfree(nb->nbst.e_el);
1068 nb->nbst.e_el = nullptr;
1070 pfree(nb->nbst.fshift);
1071 nb->nbst.fshift = nullptr;
1073 /* Free command queues */
1074 clReleaseCommandQueue(nb->stream[InteractionLocality::Local]);
1075 nb->stream[InteractionLocality::Local] = nullptr;
1076 if (nb->bUseTwoStreams)
1078 clReleaseCommandQueue(nb->stream[InteractionLocality::NonLocal]);
1079 nb->stream[InteractionLocality::NonLocal] = nullptr;
1081 /* Free other events */
1082 if (nb->nonlocal_done)
1084 clReleaseEvent(nb->nonlocal_done);
1085 nb->nonlocal_done = nullptr;
1087 if (nb->misc_ops_and_local_H2D_done)
1089 clReleaseEvent(nb->misc_ops_and_local_H2D_done);
1090 nb->misc_ops_and_local_H2D_done = nullptr;
1093 free_gpu_device_runtime_data(nb->dev_rundata);
1094 sfree(nb->dev_rundata);
1096 /* Free timers and timings */
1103 fprintf(debug, "Cleaned up OpenCL data structures.\n");
1107 //! This function is documented in the header file
1108 gmx_wallclock_gpu_nbnxn_t* gpu_get_timings(gmx_nbnxn_ocl_t* nb)
1110 return (nb != nullptr && nb->bDoTime) ? nb->timings : nullptr;
1113 //! This function is documented in the header file
1114 void gpu_reset_timings(nonbonded_verlet_t* nbv)
1116 if (nbv->gpu_nbv && nbv->gpu_nbv->bDoTime)
1118 init_timings(nbv->gpu_nbv->timings);
1122 //! This function is documented in the header file
1123 int gpu_min_ci_balanced(gmx_nbnxn_ocl_t* nb)
1125 return nb != nullptr ? gpu_min_ci_balanced_factor * nb->dev_info->compute_units : 0;
1128 //! This function is documented in the header file
1129 gmx_bool gpu_is_kernel_ewald_analytical(const gmx_nbnxn_ocl_t* nb)
1131 return ((nb->nbparam->eeltype == eelOclEWALD_ANA) || (nb->nbparam->eeltype == eelOclEWALD_ANA_TWIN));
1134 } // namespace Nbnxm