2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016 by the GROMACS development team.
5 * Copyright (c) 2017,2018,2019,2020, by the GROMACS development team, led by
6 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
7 * and including many others, as listed in the AUTHORS file in the
8 * top-level source directory and at http://www.gromacs.org.
10 * GROMACS is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public License
12 * as published by the Free Software Foundation; either version 2.1
13 * of the License, or (at your option) any later version.
15 * GROMACS is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with GROMACS; if not, see
22 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
23 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
25 * If you want to redistribute modifications to GROMACS, please
26 * consider that scientific software is very special. Version
27 * control is crucial - bugs must be traceable. We will be happy to
28 * consider code for inclusion in the official distribution, but
29 * derived work must not be called official GROMACS. Details are found
30 * in the README & COPYING files - if they are missing, get the
31 * official version at http://www.gromacs.org.
33 * To help us fund GROMACS development, we humbly ask that you cite
34 * the research papers on the package. Check out http://www.gromacs.org.
37 * \brief Define OpenCL implementation of nbnxm_gpu_data_mgmt.h
39 * \author Anca Hamuraru <anca@streamcomputing.eu>
40 * \author Dimitrios Karkoulis <dimitris.karkoulis@gmail.com>
41 * \author Teemu Virolainen <teemu@streamcomputing.eu>
42 * \author Szilárd Páll <pall.szilard@gmail.com>
43 * \ingroup module_nbnxm
55 #include "gromacs/gpu_utils/gpu_utils.h"
56 #include "gromacs/gpu_utils/oclutils.h"
57 #include "gromacs/hardware/gpu_hw_info.h"
58 #include "gromacs/math/vectypes.h"
59 #include "gromacs/mdlib/force_flags.h"
60 #include "gromacs/mdtypes/interaction_const.h"
61 #include "gromacs/mdtypes/md_enums.h"
62 #include "gromacs/nbnxm/atomdata.h"
63 #include "gromacs/nbnxm/gpu_data_mgmt.h"
64 #include "gromacs/nbnxm/gpu_jit_support.h"
65 #include "gromacs/nbnxm/nbnxm.h"
66 #include "gromacs/nbnxm/nbnxm_gpu.h"
67 #include "gromacs/nbnxm/pairlistsets.h"
68 #include "gromacs/pbcutil/ishift.h"
69 #include "gromacs/timing/gpu_timing.h"
70 #include "gromacs/utility/cstringutil.h"
71 #include "gromacs/utility/fatalerror.h"
72 #include "gromacs/utility/gmxassert.h"
73 #include "gromacs/utility/real.h"
74 #include "gromacs/utility/smalloc.h"
76 #include "nbnxm_ocl_internal.h"
77 #include "nbnxm_ocl_types.h"
82 /*! \brief Copies of values from cl_driver_diagnostics_intel.h,
83 * which isn't guaranteed to be available. */
85 #define CL_CONTEXT_SHOW_DIAGNOSTICS_INTEL 0x4106
86 #define CL_CONTEXT_DIAGNOSTICS_LEVEL_GOOD_INTEL 0x1
87 #define CL_CONTEXT_DIAGNOSTICS_LEVEL_BAD_INTEL 0x2
88 #define CL_CONTEXT_DIAGNOSTICS_LEVEL_NEUTRAL_INTEL 0x4
91 /*! \brief This parameter should be determined heuristically from the
92 * kernel execution times
94 * This value is best for small systems on a single AMD Radeon R9 290X
95 * (and about 5% faster than 40, which is the default for CUDA
96 * devices). Larger simulation systems were quite insensitive to the
97 * value of this parameter.
99 static unsigned int gpu_min_ci_balanced_factor = 50;
102 /*! \brief Returns true if LJ combination rules are used in the non-bonded kernels.
104 * Full doc in nbnxn_ocl_internal.h */
105 bool useLjCombRule(int vdwType)
107 return (vdwType == evdwOclCUTCOMBGEOM || vdwType == evdwOclCUTCOMBLB);
110 /*! \brief Tabulates the Ewald Coulomb force and initializes the size/scale
111 * and the table GPU array.
113 * If called with an already allocated table, it just re-uploads the
116 static void init_ewald_coulomb_force_table(const EwaldCorrectionTables& tables,
118 const gmx_device_runtime_data_t* runData)
124 if (nbp->coulomb_tab_climg2d != nullptr)
126 freeDeviceBuffer(&(nbp->coulomb_tab_climg2d));
129 /* Switched from using textures to using buffers */
130 // TODO: decide which alternative is most efficient - textures or buffers.
132 cl_image_format array_format;
134 array_format.image_channel_data_type = CL_FLOAT;
135 array_format.image_channel_order = CL_R;
137 coul_tab = clCreateImage2D(runData->context, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
138 &array_format, tabsize, 1, 0, ftmp, &cl_error);
141 coul_tab = clCreateBuffer(
142 runData->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
143 tables.tableF.size() * sizeof(cl_float), const_cast<real*>(tables.tableF.data()), &cl_error);
144 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
145 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
147 nbp->coulomb_tab_climg2d = coul_tab;
148 nbp->coulomb_tab_scale = tables.scale;
152 /*! \brief Initializes the atomdata structure first time, it only gets filled at
155 static void init_atomdata_first(cl_atomdata_t* ad, int ntypes, gmx_device_runtime_data_t* runData)
161 /* An element of the shift_vec device buffer has the same size as one element
162 of the host side shift_vec buffer. */
163 ad->shift_vec_elem_size = sizeof(*nbnxn_atomdata_t::shift_vec.data());
165 ad->shift_vec = clCreateBuffer(runData->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
166 SHIFTS * ad->shift_vec_elem_size, nullptr, &cl_error);
167 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
168 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
169 ad->bShiftVecUploaded = CL_FALSE;
171 /* An element of the fshift device buffer has the same size as one element
172 of the host side fshift buffer. */
173 ad->fshift_elem_size = sizeof(*nb_staging_t::fshift);
175 ad->fshift = clCreateBuffer(runData->context, CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
176 SHIFTS * ad->fshift_elem_size, nullptr, &cl_error);
177 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
178 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
180 ad->e_lj = clCreateBuffer(runData->context, CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
181 sizeof(float), nullptr, &cl_error);
182 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
183 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
185 ad->e_el = clCreateBuffer(runData->context, CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
186 sizeof(float), nullptr, &cl_error);
187 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
188 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
190 /* initialize to nullptr pointers to data that is not allocated here and will
191 need reallocation in nbnxn_gpu_init_atomdata */
195 /* size -1 indicates that the respective array hasn't been initialized yet */
200 /*! \brief Copies all parameters related to the cut-off from ic to nbp
202 static void set_cutoff_parameters(cl_nbparam_t* nbp, const interaction_const_t* ic, const PairlistParams& listParams)
204 nbp->ewald_beta = ic->ewaldcoeff_q;
205 nbp->sh_ewald = ic->sh_ewald;
206 nbp->epsfac = ic->epsfac;
207 nbp->two_k_rf = 2.0 * ic->k_rf;
208 nbp->c_rf = ic->c_rf;
209 nbp->rvdw_sq = ic->rvdw * ic->rvdw;
210 nbp->rcoulomb_sq = ic->rcoulomb * ic->rcoulomb;
211 nbp->rlistOuter_sq = listParams.rlistOuter * listParams.rlistOuter;
212 nbp->rlistInner_sq = listParams.rlistInner * listParams.rlistInner;
213 nbp->useDynamicPruning = listParams.useDynamicPruning;
215 nbp->sh_lj_ewald = ic->sh_lj_ewald;
216 nbp->ewaldcoeff_lj = ic->ewaldcoeff_lj;
218 nbp->rvdw_switch = ic->rvdw_switch;
219 nbp->dispersion_shift = ic->dispersion_shift;
220 nbp->repulsion_shift = ic->repulsion_shift;
221 nbp->vdw_switch = ic->vdw_switch;
224 /*! \brief Returns the kinds of electrostatics and Vdw OpenCL
225 * kernels that will be used.
227 * Respectively, these values are from enum eelOcl and enum
229 static void map_interaction_types_to_gpu_kernel_flavors(const interaction_const_t* ic,
234 if (ic->vdwtype == evdwCUT)
236 switch (ic->vdw_modifier)
239 case eintmodPOTSHIFT:
242 case ljcrNONE: *gpu_vdwtype = evdwOclCUT; break;
243 case ljcrGEOM: *gpu_vdwtype = evdwOclCUTCOMBGEOM; break;
244 case ljcrLB: *gpu_vdwtype = evdwOclCUTCOMBLB; break;
247 "The requested LJ combination rule is not implemented in the "
248 "OpenCL GPU accelerated kernels!");
251 case eintmodFORCESWITCH: *gpu_vdwtype = evdwOclFSWITCH; break;
252 case eintmodPOTSWITCH: *gpu_vdwtype = evdwOclPSWITCH; break;
255 "The requested VdW interaction modifier is not implemented in the GPU "
256 "accelerated kernels!");
259 else if (ic->vdwtype == evdwPME)
261 if (ic->ljpme_comb_rule == ljcrGEOM)
263 *gpu_vdwtype = evdwOclEWALDGEOM;
267 *gpu_vdwtype = evdwOclEWALDLB;
272 gmx_incons("The requested VdW type is not implemented in the GPU accelerated kernels!");
275 if (ic->eeltype == eelCUT)
277 *gpu_eeltype = eelOclCUT;
279 else if (EEL_RF(ic->eeltype))
281 *gpu_eeltype = eelOclRF;
283 else if ((EEL_PME(ic->eeltype) || ic->eeltype == eelEWALD))
285 *gpu_eeltype = nbnxn_gpu_pick_ewald_kernel_type(*ic);
289 /* Shouldn't happen, as this is checked when choosing Verlet-scheme */
291 "The requested electrostatics type is not implemented in the GPU accelerated "
296 /*! \brief Initializes the nonbonded parameter data structure.
298 static void init_nbparam(cl_nbparam_t* nbp,
299 const interaction_const_t* ic,
300 const PairlistParams& listParams,
301 const nbnxn_atomdata_t::Params& nbatParams,
302 const gmx_device_runtime_data_t* runData)
306 set_cutoff_parameters(nbp, ic, listParams);
308 map_interaction_types_to_gpu_kernel_flavors(ic, nbatParams.comb_rule, &(nbp->eeltype), &(nbp->vdwtype));
310 if (ic->vdwtype == evdwPME)
312 if (ic->ljpme_comb_rule == ljcrGEOM)
314 GMX_ASSERT(nbatParams.comb_rule == ljcrGEOM, "Combination rule mismatch!");
318 GMX_ASSERT(nbatParams.comb_rule == ljcrLB, "Combination rule mismatch!");
321 /* generate table for PME */
322 nbp->coulomb_tab_climg2d = nullptr;
323 if (nbp->eeltype == eelOclEWALD_TAB || nbp->eeltype == eelOclEWALD_TAB_TWIN)
325 GMX_RELEASE_ASSERT(ic->coulombEwaldTables, "Need valid Coulomb Ewald correction tables");
326 init_ewald_coulomb_force_table(*ic->coulombEwaldTables, nbp, runData);
329 // TODO: improvement needed.
330 // The image2d is created here even if eeltype is not eelCuEWALD_TAB or eelCuEWALD_TAB_TWIN
331 // because the OpenCL kernels don't accept nullptr values for image2D parameters.
333 /* Switched from using textures to using buffers */
334 // TODO: decide which alternative is most efficient - textures or buffers.
336 cl_image_format array_format;
338 array_format.image_channel_data_type = CL_FLOAT;
339 array_format.image_channel_order = CL_R;
341 nbp->coulomb_tab_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE,
342 &array_format, 1, 1, 0, nullptr, &cl_error);
345 nbp->coulomb_tab_climg2d = clCreateBuffer(runData->context, CL_MEM_READ_ONLY,
346 sizeof(cl_float), nullptr, &cl_error);
347 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
348 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
351 const int nnbfp = 2 * nbatParams.numTypes * nbatParams.numTypes;
352 const int nnbfp_comb = 2 * nbatParams.numTypes;
355 /* Switched from using textures to using buffers */
356 // TODO: decide which alternative is most efficient - textures or buffers.
358 cl_image_format array_format;
360 array_format.image_channel_data_type = CL_FLOAT;
361 array_format.image_channel_order = CL_R;
363 nbp->nbfp_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_ONLY |
364 CL_MEM_COPY_HOST_PTR, &array_format, nnbfp, 1, 0, nbat->nbfp, &cl_error);
367 nbp->nbfp_climg2d = clCreateBuffer(
368 runData->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
369 nnbfp * sizeof(cl_float), const_cast<float*>(nbatParams.nbfp.data()), &cl_error);
370 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
371 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
373 if (ic->vdwtype == evdwPME)
375 /* Switched from using textures to using buffers */
376 // TODO: decide which alternative is most efficient - textures or buffers.
377 /* nbp->nbfp_comb_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE |
378 CL_MEM_COPY_HOST_PTR, &array_format, nnbfp_comb, 1, 0, nbat->nbfp_comb, &cl_error);*/
379 nbp->nbfp_comb_climg2d = clCreateBuffer(
380 runData->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
381 nnbfp_comb * sizeof(cl_float), const_cast<float*>(nbatParams.nbfp_comb.data()),
383 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
384 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
388 // TODO: improvement needed.
389 // The image2d is created here even if vdwtype is not evdwPME because the OpenCL kernels
390 // don't accept nullptr values for image2D parameters.
391 /* Switched from using textures to using buffers */
392 // TODO: decide which alternative is most efficient - textures or buffers.
393 /* nbp->nbfp_comb_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE,
394 &array_format, 1, 1, 0, nullptr, &cl_error);*/
395 nbp->nbfp_comb_climg2d = clCreateBuffer(runData->context, CL_MEM_READ_ONLY,
396 sizeof(cl_float), nullptr, &cl_error);
397 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
398 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
403 //! This function is documented in the header file
404 void gpu_pme_loadbal_update_param(const nonbonded_verlet_t* nbv, const interaction_const_t* ic)
406 if (!nbv || !nbv->useGpu())
410 NbnxmGpu* nb = nbv->gpu_nbv;
411 cl_nbparam_t* nbp = nb->nbparam;
413 set_cutoff_parameters(nbp, ic, nbv->pairlistSets().params());
415 nbp->eeltype = nbnxn_gpu_pick_ewald_kernel_type(*ic);
417 GMX_RELEASE_ASSERT(ic->coulombEwaldTables, "Need valid Coulomb Ewald correction tables");
418 init_ewald_coulomb_force_table(*ic->coulombEwaldTables, nbp, nb->dev_rundata);
421 /*! \brief Initializes the pair list data structure.
423 static void init_plist(cl_plist_t* pl)
425 /* initialize to nullptr pointers to data that is not allocated here and will
426 need reallocation in nbnxn_gpu_init_pairlist */
432 /* size -1 indicates that the respective array hasn't been initialized yet */
439 pl->imask_nalloc = -1;
441 pl->excl_nalloc = -1;
442 pl->haveFreshList = false;
445 /*! \brief Initializes the timings data structure.
447 static void init_timings(gmx_wallclock_gpu_nbnxn_t* t)
456 for (i = 0; i < 2; i++)
458 for (j = 0; j < 2; j++)
460 t->ktime[i][j].t = 0.0;
461 t->ktime[i][j].c = 0;
466 t->pruneTime.t = 0.0;
467 t->dynamicPruneTime.c = 0;
468 t->dynamicPruneTime.t = 0.0;
472 //! OpenCL notification callback function
473 static void CL_CALLBACK ocl_notify_fn(const char* pErrInfo,
474 const void gmx_unused* private_info,
475 size_t gmx_unused cb,
476 void gmx_unused* user_data)
478 if (pErrInfo != nullptr)
480 printf("%s\n", pErrInfo); // Print error/hint
484 /*! \brief Creates context for OpenCL GPU given by \p mygpu
486 * A fatal error results if creation fails.
488 * \param[inout] runtimeData runtime data including program and context
489 * \param[in] devInfo device info struct
490 * \param[in] rank MPI rank (for error reporting)
492 static void nbnxn_gpu_create_context(gmx_device_runtime_data_t* runtimeData,
493 const gmx_device_info_t* devInfo,
496 cl_context_properties context_properties[5];
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;
509 context_properties[i++] = CL_CONTEXT_PLATFORM;
510 context_properties[i++] = reinterpret_cast<cl_context_properties>(platform_id);
511 if (getenv("GMX_OCL_SHOW_DIAGNOSTICS"))
513 context_properties[i++] = CL_CONTEXT_SHOW_DIAGNOSTICS_INTEL;
514 context_properties[i++] =
515 CL_CONTEXT_DIAGNOSTICS_LEVEL_BAD_INTEL | CL_CONTEXT_DIAGNOSTICS_LEVEL_NEUTRAL_INTEL;
517 context_properties[i++] = 0; /* Terminates the list of properties */
519 context = clCreateContext(context_properties, 1, &device_id, ocl_notify_fn, nullptr, &cl_error);
520 if (CL_SUCCESS != cl_error)
522 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s:\n OpenCL error %d: %s",
523 rank, devInfo->device_name, cl_error, ocl_get_error_string(cl_error).c_str());
526 runtimeData->context = context;
529 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
530 static cl_kernel nbnxn_gpu_create_kernel(NbnxmGpu* nb, const char* kernel_name)
535 kernel = clCreateKernel(nb->dev_rundata->program, kernel_name, &cl_error);
536 if (CL_SUCCESS != cl_error)
538 gmx_fatal(FARGS, "Failed to create kernel '%s' for GPU #%s: OpenCL error %d", kernel_name,
539 nb->dev_info->device_name, cl_error);
545 /*! \brief Clears nonbonded shift force output array and energy outputs on the GPU.
547 static void nbnxn_ocl_clear_e_fshift(NbnxmGpu* nb)
551 cl_atomdata_t* adat = nb->atdat;
552 cl_command_queue ls = nb->stream[InteractionLocality::Local];
554 size_t local_work_size[3] = { 1, 1, 1 };
555 size_t global_work_size[3] = { 1, 1, 1 };
557 cl_int shifts = SHIFTS * 3;
561 cl_kernel zero_e_fshift = nb->kernel_zero_e_fshift;
563 local_work_size[0] = 64;
564 // Round the total number of threads up from the array size
565 global_work_size[0] = ((shifts + local_work_size[0] - 1) / local_work_size[0]) * local_work_size[0];
568 cl_error = clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->fshift));
569 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_lj));
570 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_el));
571 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_uint), &shifts);
572 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
574 cl_error = clEnqueueNDRangeKernel(ls, zero_e_fshift, 3, nullptr, global_work_size,
575 local_work_size, 0, nullptr, nullptr);
576 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
579 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
580 static void nbnxn_gpu_init_kernels(NbnxmGpu* nb)
582 /* Init to 0 main kernel arrays */
583 /* They will be later on initialized in select_nbnxn_kernel */
584 // TODO: consider always creating all variants of the kernels here so that there is no
585 // need for late call to clCreateKernel -- if that gives any advantage?
586 memset(nb->kernel_ener_noprune_ptr, 0, sizeof(nb->kernel_ener_noprune_ptr));
587 memset(nb->kernel_ener_prune_ptr, 0, sizeof(nb->kernel_ener_prune_ptr));
588 memset(nb->kernel_noener_noprune_ptr, 0, sizeof(nb->kernel_noener_noprune_ptr));
589 memset(nb->kernel_noener_prune_ptr, 0, sizeof(nb->kernel_noener_prune_ptr));
591 /* Init pruning kernels
593 * TODO: we could avoid creating kernels if dynamic pruning is turned off,
594 * but ATM that depends on force flags not passed into the initialization.
596 nb->kernel_pruneonly[epruneFirst] = nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_opencl");
597 nb->kernel_pruneonly[epruneRolling] =
598 nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_rolling_opencl");
600 /* Init auxiliary kernels */
601 nb->kernel_zero_e_fshift = nbnxn_gpu_create_kernel(nb, "zero_e_fshift");
604 /*! \brief Initializes simulation constant data.
606 * Initializes members of the atomdata and nbparam structs and
607 * clears e/fshift output buffers.
609 static void nbnxn_ocl_init_const(NbnxmGpu* nb,
610 const interaction_const_t* ic,
611 const PairlistParams& listParams,
612 const nbnxn_atomdata_t::Params& nbatParams)
614 init_atomdata_first(nb->atdat, nbatParams.numTypes, nb->dev_rundata);
615 init_nbparam(nb->nbparam, ic, listParams, nbatParams, nb->dev_rundata);
619 //! This function is documented in the header file
620 NbnxmGpu* gpu_init(const gmx_device_info_t* deviceInfo,
621 const interaction_const_t* ic,
622 const PairlistParams& listParams,
623 const nbnxn_atomdata_t* nbat,
625 const gmx_bool bLocalAndNonlocal)
628 cl_command_queue_properties queue_properties;
632 auto nb = new NbnxmGpu;
634 snew(nb->nbparam, 1);
635 snew(nb->plist[InteractionLocality::Local], 1);
636 if (bLocalAndNonlocal)
638 snew(nb->plist[InteractionLocality::NonLocal], 1);
641 nb->bUseTwoStreams = static_cast<cl_bool>(bLocalAndNonlocal);
643 nb->timers = new cl_timers_t();
644 snew(nb->timings, 1);
646 /* set device info, just point it to the right GPU among the detected ones */
647 nb->dev_info = deviceInfo;
648 snew(nb->dev_rundata, 1);
651 pmalloc(reinterpret_cast<void**>(&nb->nbst.e_lj), sizeof(*nb->nbst.e_lj));
652 pmalloc(reinterpret_cast<void**>(&nb->nbst.e_el), sizeof(*nb->nbst.e_el));
653 pmalloc(reinterpret_cast<void**>(&nb->nbst.fshift), SHIFTS * sizeof(*nb->nbst.fshift));
655 init_plist(nb->plist[InteractionLocality::Local]);
657 /* OpenCL timing disabled if GMX_DISABLE_GPU_TIMING is defined. */
658 nb->bDoTime = static_cast<cl_bool>(getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
660 /* Create queues only after bDoTime has been initialized */
663 queue_properties = CL_QUEUE_PROFILING_ENABLE;
667 queue_properties = 0;
670 nbnxn_gpu_create_context(nb->dev_rundata, nb->dev_info, rank);
672 /* local/non-local GPU streams */
673 nb->stream[InteractionLocality::Local] = clCreateCommandQueue(
674 nb->dev_rundata->context, nb->dev_info->ocl_gpu_id.ocl_device_id, queue_properties, &cl_error);
675 if (CL_SUCCESS != cl_error)
677 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s: OpenCL error %d", rank,
678 nb->dev_info->device_name, cl_error);
681 if (nb->bUseTwoStreams)
683 init_plist(nb->plist[InteractionLocality::NonLocal]);
685 nb->stream[InteractionLocality::NonLocal] =
686 clCreateCommandQueue(nb->dev_rundata->context, nb->dev_info->ocl_gpu_id.ocl_device_id,
687 queue_properties, &cl_error);
688 if (CL_SUCCESS != cl_error)
690 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s: OpenCL error %d",
691 rank, nb->dev_info->device_name, cl_error);
697 init_timings(nb->timings);
700 nbnxn_ocl_init_const(nb, ic, listParams, nbat->params());
702 /* Enable LJ param manual prefetch for AMD or Intel or if we request through env. var.
703 * TODO: decide about NVIDIA
705 nb->bPrefetchLjParam = (getenv("GMX_OCL_DISABLE_I_PREFETCH") == nullptr)
706 && ((nb->dev_info->vendor_e == OCL_VENDOR_AMD)
707 || (nb->dev_info->vendor_e == OCL_VENDOR_INTEL)
708 || (getenv("GMX_OCL_ENABLE_I_PREFETCH") != nullptr));
710 /* NOTE: in CUDA we pick L1 cache configuration for the nbnxn kernels here,
711 * but sadly this is not supported in OpenCL (yet?). Consider adding it if
712 * it becomes supported.
714 nbnxn_gpu_compile_kernels(nb);
715 nbnxn_gpu_init_kernels(nb);
717 /* clear energy and shift force outputs */
718 nbnxn_ocl_clear_e_fshift(nb);
722 fprintf(debug, "Initialized OpenCL data structures.\n");
728 /*! \brief Clears the first natoms_clear elements of the GPU nonbonded force output array.
730 static void nbnxn_ocl_clear_f(NbnxmGpu* nb, int natoms_clear)
732 if (natoms_clear == 0)
737 cl_int gmx_used_in_debug cl_error;
739 cl_atomdata_t* atomData = nb->atdat;
740 cl_command_queue ls = nb->stream[InteractionLocality::Local];
741 cl_float value = 0.0F;
743 cl_error = clEnqueueFillBuffer(ls, atomData->f, &value, sizeof(cl_float), 0,
744 natoms_clear * sizeof(rvec), 0, nullptr, nullptr);
745 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
746 ("clEnqueueFillBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
749 //! This function is documented in the header file
750 void gpu_clear_outputs(NbnxmGpu* nb, bool computeVirial)
752 nbnxn_ocl_clear_f(nb, nb->atdat->natoms);
753 /* clear shift force array and energies if the outputs were
754 used in the current step */
757 nbnxn_ocl_clear_e_fshift(nb);
760 /* kick off buffer clearing kernel to ensure concurrency with constraints/update */
761 cl_int gmx_unused cl_error;
762 cl_error = clFlush(nb->stream[InteractionLocality::Local]);
763 assert(CL_SUCCESS == cl_error);
766 //! This function is documented in the header file
767 void gpu_init_pairlist(NbnxmGpu* nb, const NbnxnPairlistGpu* h_plist, const InteractionLocality iloc)
770 // Timing accumulation should happen only if there was work to do
771 // because getLastRangeTime() gets skipped with empty lists later
772 // which leads to the counter not being reset.
773 bool bDoTime = ((nb->bDoTime == CL_TRUE) && !h_plist->sci.empty());
774 cl_command_queue stream = nb->stream[iloc];
775 cl_plist_t* d_plist = nb->plist[iloc];
777 if (d_plist->na_c < 0)
779 d_plist->na_c = h_plist->na_ci;
783 if (d_plist->na_c != h_plist->na_ci)
785 sprintf(sbuf, "In cu_init_plist: the #atoms per cell has changed (from %d to %d)",
786 d_plist->na_c, h_plist->na_ci);
791 gpu_timers_t::Interaction& iTimers = nb->timers->interaction[iloc];
795 iTimers.pl_h2d.openTimingRegion(stream);
796 iTimers.didPairlistH2D = true;
799 // TODO most of this function is same in CUDA and OpenCL, move into the header
800 DeviceContext context = nb->dev_rundata->context;
802 reallocateDeviceBuffer(&d_plist->sci, h_plist->sci.size(), &d_plist->nsci, &d_plist->sci_nalloc, context);
803 copyToDeviceBuffer(&d_plist->sci, h_plist->sci.data(), 0, h_plist->sci.size(), stream,
804 GpuApiCallBehavior::Async, bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
806 reallocateDeviceBuffer(&d_plist->cj4, h_plist->cj4.size(), &d_plist->ncj4, &d_plist->cj4_nalloc, context);
807 copyToDeviceBuffer(&d_plist->cj4, h_plist->cj4.data(), 0, h_plist->cj4.size(), stream,
808 GpuApiCallBehavior::Async, bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
810 reallocateDeviceBuffer(&d_plist->imask, h_plist->cj4.size() * c_nbnxnGpuClusterpairSplit,
811 &d_plist->nimask, &d_plist->imask_nalloc, context);
813 reallocateDeviceBuffer(&d_plist->excl, h_plist->excl.size(), &d_plist->nexcl,
814 &d_plist->excl_nalloc, context);
815 copyToDeviceBuffer(&d_plist->excl, h_plist->excl.data(), 0, h_plist->excl.size(), stream,
816 GpuApiCallBehavior::Async, bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
820 iTimers.pl_h2d.closeTimingRegion(stream);
823 /* need to prune the pair list during the next step */
824 d_plist->haveFreshList = true;
827 //! This function is documented in the header file
828 void gpu_upload_shiftvec(NbnxmGpu* nb, const nbnxn_atomdata_t* nbatom)
830 cl_atomdata_t* adat = nb->atdat;
831 cl_command_queue ls = nb->stream[InteractionLocality::Local];
833 /* only if we have a dynamic box */
834 if (nbatom->bDynamicBox || !adat->bShiftVecUploaded)
836 ocl_copy_H2D_async(adat->shift_vec, nbatom->shift_vec.data(), 0,
837 SHIFTS * adat->shift_vec_elem_size, ls, nullptr);
838 adat->bShiftVecUploaded = CL_TRUE;
842 //! This function is documented in the header file
843 void gpu_init_atomdata(NbnxmGpu* nb, const nbnxn_atomdata_t* nbat)
848 bool bDoTime = nb->bDoTime == CL_TRUE;
849 cl_timers_t* timers = nb->timers;
850 cl_atomdata_t* d_atdat = nb->atdat;
851 cl_command_queue ls = nb->stream[InteractionLocality::Local];
853 natoms = nbat->numAtoms();
858 /* time async copy */
859 timers->atdat.openTimingRegion(ls);
862 /* need to reallocate if we have to copy more atoms than the amount of space
863 available and only allocate if we haven't initialized yet, i.e d_atdat->natoms == -1 */
864 if (natoms > d_atdat->nalloc)
866 nalloc = over_alloc_small(natoms);
868 /* free up first if the arrays have already been initialized */
869 if (d_atdat->nalloc != -1)
871 freeDeviceBuffer(&d_atdat->f);
872 freeDeviceBuffer(&d_atdat->xq);
873 freeDeviceBuffer(&d_atdat->lj_comb);
874 freeDeviceBuffer(&d_atdat->atom_types);
877 d_atdat->f_elem_size = sizeof(rvec);
879 d_atdat->f = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
880 nalloc * d_atdat->f_elem_size, nullptr, &cl_error);
881 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
882 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
884 d_atdat->xq = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
885 nalloc * sizeof(cl_float4), nullptr, &cl_error);
886 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
887 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
889 if (useLjCombRule(nb->nbparam->vdwtype))
891 d_atdat->lj_comb = clCreateBuffer(nb->dev_rundata->context,
892 CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
893 nalloc * sizeof(cl_float2), nullptr, &cl_error);
894 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
895 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
899 d_atdat->atom_types = clCreateBuffer(nb->dev_rundata->context,
900 CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
901 nalloc * sizeof(int), nullptr, &cl_error);
902 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
903 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
906 d_atdat->nalloc = nalloc;
910 d_atdat->natoms = natoms;
911 d_atdat->natoms_local = nbat->natoms_local;
913 /* need to clear GPU f output if realloc happened */
916 nbnxn_ocl_clear_f(nb, nalloc);
919 if (useLjCombRule(nb->nbparam->vdwtype))
921 ocl_copy_H2D_async(d_atdat->lj_comb, nbat->params().lj_comb.data(), 0, natoms * sizeof(cl_float2),
922 ls, bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
926 ocl_copy_H2D_async(d_atdat->atom_types, nbat->params().type.data(), 0, natoms * sizeof(int),
927 ls, bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
932 timers->atdat.closeTimingRegion(ls);
935 /* kick off the tasks enqueued above to ensure concurrency with the search */
936 cl_error = clFlush(ls);
937 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
938 ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
941 /*! \brief Releases an OpenCL kernel pointer */
942 static void free_kernel(cl_kernel* kernel_ptr)
944 cl_int gmx_unused cl_error;
946 assert(nullptr != kernel_ptr);
950 cl_error = clReleaseKernel(*kernel_ptr);
951 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
952 ("clReleaseKernel failed: " + ocl_get_error_string(cl_error)).c_str());
954 *kernel_ptr = nullptr;
958 /*! \brief Releases a list of OpenCL kernel pointers */
959 static void free_kernels(cl_kernel* kernels, int count)
963 for (i = 0; i < count; i++)
965 free_kernel(kernels + i);
969 /*! \brief Free the OpenCL runtime data (context and program).
971 * The function releases the OpenCL context and program assuciated with the
972 * device that the calling PP rank is running on.
974 * \param runData [in] porinter to the structure with runtime data.
976 static void free_gpu_device_runtime_data(gmx_device_runtime_data_t* runData)
978 if (runData == nullptr)
983 cl_int gmx_unused cl_error;
985 if (runData->context)
987 cl_error = clReleaseContext(runData->context);
988 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
989 ("clReleaseContext failed: " + ocl_get_error_string(cl_error)).c_str());
990 runData->context = nullptr;
993 if (runData->program)
995 cl_error = clReleaseProgram(runData->program);
996 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
997 ("clReleaseProgram failed: " + ocl_get_error_string(cl_error)).c_str());
998 runData->program = nullptr;
1002 //! This function is documented in the header file
1003 void gpu_free(NbnxmGpu* nb)
1011 int kernel_count = sizeof(nb->kernel_ener_noprune_ptr) / sizeof(nb->kernel_ener_noprune_ptr[0][0]);
1012 free_kernels(nb->kernel_ener_noprune_ptr[0], kernel_count);
1014 kernel_count = sizeof(nb->kernel_ener_prune_ptr) / sizeof(nb->kernel_ener_prune_ptr[0][0]);
1015 free_kernels(nb->kernel_ener_prune_ptr[0], kernel_count);
1017 kernel_count = sizeof(nb->kernel_noener_noprune_ptr) / sizeof(nb->kernel_noener_noprune_ptr[0][0]);
1018 free_kernels(nb->kernel_noener_noprune_ptr[0], kernel_count);
1020 kernel_count = sizeof(nb->kernel_noener_prune_ptr) / sizeof(nb->kernel_noener_prune_ptr[0][0]);
1021 free_kernels(nb->kernel_noener_prune_ptr[0], kernel_count);
1023 free_kernel(&(nb->kernel_zero_e_fshift));
1026 freeDeviceBuffer(&(nb->atdat->xq));
1027 freeDeviceBuffer(&(nb->atdat->f));
1028 freeDeviceBuffer(&(nb->atdat->e_lj));
1029 freeDeviceBuffer(&(nb->atdat->e_el));
1030 freeDeviceBuffer(&(nb->atdat->fshift));
1031 freeDeviceBuffer(&(nb->atdat->lj_comb));
1032 freeDeviceBuffer(&(nb->atdat->atom_types));
1033 freeDeviceBuffer(&(nb->atdat->shift_vec));
1037 freeDeviceBuffer(&(nb->nbparam->nbfp_climg2d));
1038 freeDeviceBuffer(&(nb->nbparam->nbfp_comb_climg2d));
1039 freeDeviceBuffer(&(nb->nbparam->coulomb_tab_climg2d));
1043 auto* plist = nb->plist[InteractionLocality::Local];
1044 freeDeviceBuffer(&plist->sci);
1045 freeDeviceBuffer(&plist->cj4);
1046 freeDeviceBuffer(&plist->imask);
1047 freeDeviceBuffer(&plist->excl);
1049 if (nb->bUseTwoStreams)
1051 auto* plist_nl = nb->plist[InteractionLocality::NonLocal];
1052 freeDeviceBuffer(&plist_nl->sci);
1053 freeDeviceBuffer(&plist_nl->cj4);
1054 freeDeviceBuffer(&plist_nl->imask);
1055 freeDeviceBuffer(&plist_nl->excl);
1060 pfree(nb->nbst.e_lj);
1061 nb->nbst.e_lj = nullptr;
1063 pfree(nb->nbst.e_el);
1064 nb->nbst.e_el = nullptr;
1066 pfree(nb->nbst.fshift);
1067 nb->nbst.fshift = nullptr;
1069 /* Free command queues */
1070 clReleaseCommandQueue(nb->stream[InteractionLocality::Local]);
1071 nb->stream[InteractionLocality::Local] = nullptr;
1072 if (nb->bUseTwoStreams)
1074 clReleaseCommandQueue(nb->stream[InteractionLocality::NonLocal]);
1075 nb->stream[InteractionLocality::NonLocal] = nullptr;
1077 /* Free other events */
1078 if (nb->nonlocal_done)
1080 clReleaseEvent(nb->nonlocal_done);
1081 nb->nonlocal_done = nullptr;
1083 if (nb->misc_ops_and_local_H2D_done)
1085 clReleaseEvent(nb->misc_ops_and_local_H2D_done);
1086 nb->misc_ops_and_local_H2D_done = nullptr;
1089 free_gpu_device_runtime_data(nb->dev_rundata);
1090 sfree(nb->dev_rundata);
1092 /* Free timers and timings */
1099 fprintf(debug, "Cleaned up OpenCL data structures.\n");
1103 //! This function is documented in the header file
1104 gmx_wallclock_gpu_nbnxn_t* gpu_get_timings(NbnxmGpu* nb)
1106 return (nb != nullptr && nb->bDoTime) ? nb->timings : nullptr;
1109 //! This function is documented in the header file
1110 void gpu_reset_timings(nonbonded_verlet_t* nbv)
1112 if (nbv->gpu_nbv && nbv->gpu_nbv->bDoTime)
1114 init_timings(nbv->gpu_nbv->timings);
1118 //! This function is documented in the header file
1119 int gpu_min_ci_balanced(NbnxmGpu* nb)
1121 return nb != nullptr ? gpu_min_ci_balanced_factor * nb->dev_info->compute_units : 0;
1124 //! This function is documented in the header file
1125 gmx_bool gpu_is_kernel_ewald_analytical(const NbnxmGpu* nb)
1127 return ((nb->nbparam->eeltype == eelOclEWALD_ANA) || (nb->nbparam->eeltype == eelOclEWALD_ANA_TWIN));
1130 } // namespace Nbnxm