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 DeviceContext& deviceContext)
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(deviceContext.context(), CL_MEM_READ_WRITE |
138 CL_MEM_COPY_HOST_PTR, &array_format, tabsize, 1, 0, ftmp, &cl_error);
141 coul_tab = clCreateBuffer(deviceContext.context(),
142 CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
143 tables.tableF.size() * sizeof(cl_float),
144 const_cast<real*>(tables.tableF.data()), &cl_error);
145 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
146 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
148 nbp->coulomb_tab_climg2d = coul_tab;
149 nbp->coulomb_tab_scale = tables.scale;
153 /*! \brief Initializes the atomdata structure first time, it only gets filled at
156 static void init_atomdata_first(cl_atomdata_t* ad, int ntypes, const DeviceContext& deviceContext)
162 ad->shift_vec = clCreateBuffer(deviceContext.context(), CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
163 SHIFTS * sizeof(nbnxn_atomdata_t::shift_vec[0]), nullptr, &cl_error);
164 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
165 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
166 ad->bShiftVecUploaded = CL_FALSE;
168 ad->fshift = clCreateBuffer(deviceContext.context(), CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
169 SHIFTS * sizeof(nb_staging_t::fshift[0]), nullptr, &cl_error);
170 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
171 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
173 ad->e_lj = clCreateBuffer(deviceContext.context(), CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
174 sizeof(float), nullptr, &cl_error);
175 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
176 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
178 ad->e_el = clCreateBuffer(deviceContext.context(), CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
179 sizeof(float), nullptr, &cl_error);
180 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
181 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
183 /* initialize to nullptr pointers to data that is not allocated here and will
184 need reallocation in nbnxn_gpu_init_atomdata */
188 /* size -1 indicates that the respective array hasn't been initialized yet */
193 /*! \brief Copies all parameters related to the cut-off from ic to nbp
195 static void set_cutoff_parameters(cl_nbparam_t* nbp, const interaction_const_t* ic, const PairlistParams& listParams)
197 nbp->ewald_beta = ic->ewaldcoeff_q;
198 nbp->sh_ewald = ic->sh_ewald;
199 nbp->epsfac = ic->epsfac;
200 nbp->two_k_rf = 2.0 * ic->k_rf;
201 nbp->c_rf = ic->c_rf;
202 nbp->rvdw_sq = ic->rvdw * ic->rvdw;
203 nbp->rcoulomb_sq = ic->rcoulomb * ic->rcoulomb;
204 nbp->rlistOuter_sq = listParams.rlistOuter * listParams.rlistOuter;
205 nbp->rlistInner_sq = listParams.rlistInner * listParams.rlistInner;
206 nbp->useDynamicPruning = listParams.useDynamicPruning;
208 nbp->sh_lj_ewald = ic->sh_lj_ewald;
209 nbp->ewaldcoeff_lj = ic->ewaldcoeff_lj;
211 nbp->rvdw_switch = ic->rvdw_switch;
212 nbp->dispersion_shift = ic->dispersion_shift;
213 nbp->repulsion_shift = ic->repulsion_shift;
214 nbp->vdw_switch = ic->vdw_switch;
217 /*! \brief Returns the kinds of electrostatics and Vdw OpenCL
218 * kernels that will be used.
220 * Respectively, these values are from enum eelOcl and enum
222 static void map_interaction_types_to_gpu_kernel_flavors(const interaction_const_t* ic,
227 if (ic->vdwtype == evdwCUT)
229 switch (ic->vdw_modifier)
232 case eintmodPOTSHIFT:
235 case ljcrNONE: *gpu_vdwtype = evdwOclCUT; break;
236 case ljcrGEOM: *gpu_vdwtype = evdwOclCUTCOMBGEOM; break;
237 case ljcrLB: *gpu_vdwtype = evdwOclCUTCOMBLB; break;
240 "The requested LJ combination rule is not implemented in the "
241 "OpenCL GPU accelerated kernels!");
244 case eintmodFORCESWITCH: *gpu_vdwtype = evdwOclFSWITCH; break;
245 case eintmodPOTSWITCH: *gpu_vdwtype = evdwOclPSWITCH; break;
248 "The requested VdW interaction modifier is not implemented in the GPU "
249 "accelerated kernels!");
252 else if (ic->vdwtype == evdwPME)
254 if (ic->ljpme_comb_rule == ljcrGEOM)
256 *gpu_vdwtype = evdwOclEWALDGEOM;
260 *gpu_vdwtype = evdwOclEWALDLB;
265 gmx_incons("The requested VdW type is not implemented in the GPU accelerated kernels!");
268 if (ic->eeltype == eelCUT)
270 *gpu_eeltype = eelOclCUT;
272 else if (EEL_RF(ic->eeltype))
274 *gpu_eeltype = eelOclRF;
276 else if ((EEL_PME(ic->eeltype) || ic->eeltype == eelEWALD))
278 *gpu_eeltype = nbnxn_gpu_pick_ewald_kernel_type(*ic);
282 /* Shouldn't happen, as this is checked when choosing Verlet-scheme */
284 "The requested electrostatics type is not implemented in the GPU accelerated "
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 PairlistParams& listParams,
294 const nbnxn_atomdata_t::Params& nbatParams,
295 const DeviceContext& deviceContext)
299 set_cutoff_parameters(nbp, ic, listParams);
301 map_interaction_types_to_gpu_kernel_flavors(ic, nbatParams.comb_rule, &(nbp->eeltype), &(nbp->vdwtype));
303 if (ic->vdwtype == evdwPME)
305 if (ic->ljpme_comb_rule == ljcrGEOM)
307 GMX_ASSERT(nbatParams.comb_rule == ljcrGEOM, "Combination rule mismatch!");
311 GMX_ASSERT(nbatParams.comb_rule == ljcrLB, "Combination rule mismatch!");
314 /* generate table for PME */
315 nbp->coulomb_tab_climg2d = nullptr;
316 if (nbp->eeltype == eelOclEWALD_TAB || nbp->eeltype == eelOclEWALD_TAB_TWIN)
318 GMX_RELEASE_ASSERT(ic->coulombEwaldTables, "Need valid Coulomb Ewald correction tables");
319 init_ewald_coulomb_force_table(*ic->coulombEwaldTables, nbp, deviceContext);
322 // TODO: improvement needed.
323 // The image2d is created here even if eeltype is not eelCuEWALD_TAB or eelCuEWALD_TAB_TWIN
324 // because the OpenCL kernels don't accept nullptr values for image2D parameters.
326 /* Switched from using textures to using buffers */
327 // TODO: decide which alternative is most efficient - textures or buffers.
329 cl_image_format array_format;
331 array_format.image_channel_data_type = CL_FLOAT;
332 array_format.image_channel_order = CL_R;
334 nbp->coulomb_tab_climg2d = clCreateImage2D(deviceContext.context(),
335 CL_MEM_READ_WRITE, &array_format, 1, 1, 0, nullptr, &cl_error);
338 nbp->coulomb_tab_climg2d = clCreateBuffer(deviceContext.context(), CL_MEM_READ_ONLY,
339 sizeof(cl_float), nullptr, &cl_error);
340 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
341 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
344 const int nnbfp = 2 * nbatParams.numTypes * nbatParams.numTypes;
345 const int nnbfp_comb = 2 * nbatParams.numTypes;
348 /* Switched from using textures to using buffers */
349 // TODO: decide which alternative is most efficient - textures or buffers.
351 cl_image_format array_format;
353 array_format.image_channel_data_type = CL_FLOAT;
354 array_format.image_channel_order = CL_R;
356 nbp->nbfp_climg2d = clCreateImage2D(deviceContext.context(), CL_MEM_READ_ONLY |
357 CL_MEM_COPY_HOST_PTR, &array_format, nnbfp, 1, 0, nbat->nbfp, &cl_error);
360 nbp->nbfp_climg2d = clCreateBuffer(
361 deviceContext.context(), CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
362 nnbfp * sizeof(cl_float), const_cast<float*>(nbatParams.nbfp.data()), &cl_error);
363 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
364 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
366 if (ic->vdwtype == evdwPME)
368 /* Switched from using textures to using buffers */
369 // TODO: decide which alternative is most efficient - textures or buffers.
370 /* nbp->nbfp_comb_climg2d = clCreateImage2D(deviceContext.context(), CL_MEM_READ_WRITE |
371 CL_MEM_COPY_HOST_PTR, &array_format, nnbfp_comb, 1, 0, nbat->nbfp_comb, &cl_error);*/
372 nbp->nbfp_comb_climg2d =
373 clCreateBuffer(deviceContext.context(),
374 CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
375 nnbfp_comb * sizeof(cl_float),
376 const_cast<float*>(nbatParams.nbfp_comb.data()), &cl_error);
377 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
378 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
382 // TODO: improvement needed.
383 // The image2d is created here even if vdwtype is not evdwPME because the OpenCL kernels
384 // don't accept nullptr values for image2D parameters.
385 /* Switched from using textures to using buffers */
386 // TODO: decide which alternative is most efficient - textures or buffers.
387 /* nbp->nbfp_comb_climg2d = clCreateImage2D(deviceContext.context(),
388 CL_MEM_READ_WRITE, &array_format, 1, 1, 0, nullptr, &cl_error);*/
389 nbp->nbfp_comb_climg2d = clCreateBuffer(deviceContext.context(), CL_MEM_READ_ONLY,
390 sizeof(cl_float), nullptr, &cl_error);
391 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
392 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
397 //! This function is documented in the header file
398 void gpu_pme_loadbal_update_param(const nonbonded_verlet_t* nbv, const interaction_const_t* ic)
400 if (!nbv || !nbv->useGpu())
404 NbnxmGpu* nb = nbv->gpu_nbv;
405 cl_nbparam_t* nbp = nb->nbparam;
407 set_cutoff_parameters(nbp, ic, nbv->pairlistSets().params());
409 nbp->eeltype = nbnxn_gpu_pick_ewald_kernel_type(*ic);
411 GMX_RELEASE_ASSERT(ic->coulombEwaldTables, "Need valid Coulomb Ewald correction tables");
412 init_ewald_coulomb_force_table(*ic->coulombEwaldTables, nbp, nb->dev_rundata->deviceContext_);
415 /*! \brief Initializes the pair list data structure.
417 static void init_plist(cl_plist_t* pl)
419 /* initialize to nullptr pointers to data that is not allocated here and will
420 need reallocation in nbnxn_gpu_init_pairlist */
426 /* size -1 indicates that the respective array hasn't been initialized yet */
433 pl->imask_nalloc = -1;
435 pl->excl_nalloc = -1;
436 pl->haveFreshList = false;
439 /*! \brief Initializes the timings data structure.
441 static void init_timings(gmx_wallclock_gpu_nbnxn_t* t)
450 for (i = 0; i < 2; i++)
452 for (j = 0; j < 2; j++)
454 t->ktime[i][j].t = 0.0;
455 t->ktime[i][j].c = 0;
460 t->pruneTime.t = 0.0;
461 t->dynamicPruneTime.c = 0;
462 t->dynamicPruneTime.t = 0.0;
465 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
466 static cl_kernel nbnxn_gpu_create_kernel(NbnxmGpu* nb, const char* kernel_name)
471 kernel = clCreateKernel(nb->dev_rundata->program, kernel_name, &cl_error);
472 if (CL_SUCCESS != cl_error)
474 gmx_fatal(FARGS, "Failed to create kernel '%s' for GPU #%s: OpenCL error %d", kernel_name,
475 nb->deviceInfo->device_name, cl_error);
481 /*! \brief Clears nonbonded shift force output array and energy outputs on the GPU.
483 static void nbnxn_ocl_clear_e_fshift(NbnxmGpu* nb)
487 cl_atomdata_t* adat = nb->atdat;
488 cl_command_queue ls = nb->deviceStreams[InteractionLocality::Local].stream();
490 size_t local_work_size[3] = { 1, 1, 1 };
491 size_t global_work_size[3] = { 1, 1, 1 };
493 cl_int shifts = SHIFTS * 3;
497 cl_kernel zero_e_fshift = nb->kernel_zero_e_fshift;
499 local_work_size[0] = 64;
500 // Round the total number of threads up from the array size
501 global_work_size[0] = ((shifts + local_work_size[0] - 1) / local_work_size[0]) * local_work_size[0];
504 cl_error = clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->fshift));
505 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_lj));
506 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_el));
507 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_uint), &shifts);
508 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
510 cl_error = clEnqueueNDRangeKernel(ls, zero_e_fshift, 3, nullptr, global_work_size,
511 local_work_size, 0, nullptr, nullptr);
512 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
515 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
516 static void nbnxn_gpu_init_kernels(NbnxmGpu* nb)
518 /* Init to 0 main kernel arrays */
519 /* They will be later on initialized in select_nbnxn_kernel */
520 // TODO: consider always creating all variants of the kernels here so that there is no
521 // need for late call to clCreateKernel -- if that gives any advantage?
522 memset(nb->kernel_ener_noprune_ptr, 0, sizeof(nb->kernel_ener_noprune_ptr));
523 memset(nb->kernel_ener_prune_ptr, 0, sizeof(nb->kernel_ener_prune_ptr));
524 memset(nb->kernel_noener_noprune_ptr, 0, sizeof(nb->kernel_noener_noprune_ptr));
525 memset(nb->kernel_noener_prune_ptr, 0, sizeof(nb->kernel_noener_prune_ptr));
527 /* Init pruning kernels
529 * TODO: we could avoid creating kernels if dynamic pruning is turned off,
530 * but ATM that depends on force flags not passed into the initialization.
532 nb->kernel_pruneonly[epruneFirst] = nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_opencl");
533 nb->kernel_pruneonly[epruneRolling] =
534 nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_rolling_opencl");
536 /* Init auxiliary kernels */
537 nb->kernel_zero_e_fshift = nbnxn_gpu_create_kernel(nb, "zero_e_fshift");
540 /*! \brief Initializes simulation constant data.
542 * Initializes members of the atomdata and nbparam structs and
543 * clears e/fshift output buffers.
545 static void nbnxn_ocl_init_const(cl_atomdata_t* atomData,
546 cl_nbparam_t* nbParams,
547 const interaction_const_t* ic,
548 const PairlistParams& listParams,
549 const nbnxn_atomdata_t::Params& nbatParams,
550 const DeviceContext& deviceContext)
552 init_atomdata_first(atomData, nbatParams.numTypes, deviceContext);
553 init_nbparam(nbParams, ic, listParams, nbatParams, deviceContext);
557 //! This function is documented in the header file
558 NbnxmGpu* gpu_init(const DeviceInformation* deviceInfo,
559 const DeviceContext& deviceContext,
560 const interaction_const_t* ic,
561 const PairlistParams& listParams,
562 const nbnxn_atomdata_t* nbat,
563 const bool bLocalAndNonlocal)
565 GMX_ASSERT(ic, "Need a valid interaction constants object");
567 auto nb = new NbnxmGpu;
569 snew(nb->nbparam, 1);
570 snew(nb->plist[InteractionLocality::Local], 1);
571 if (bLocalAndNonlocal)
573 snew(nb->plist[InteractionLocality::NonLocal], 1);
576 nb->bUseTwoStreams = bLocalAndNonlocal;
578 nb->timers = new cl_timers_t();
579 snew(nb->timings, 1);
581 /* set device info, just point it to the right GPU among the detected ones */
582 nb->deviceInfo = deviceInfo;
583 nb->dev_rundata = new gmx_device_runtime_data_t(deviceContext);
586 pmalloc(reinterpret_cast<void**>(&nb->nbst.e_lj), sizeof(*nb->nbst.e_lj));
587 pmalloc(reinterpret_cast<void**>(&nb->nbst.e_el), sizeof(*nb->nbst.e_el));
588 pmalloc(reinterpret_cast<void**>(&nb->nbst.fshift), SHIFTS * sizeof(*nb->nbst.fshift));
590 init_plist(nb->plist[InteractionLocality::Local]);
592 /* OpenCL timing disabled if GMX_DISABLE_GPU_TIMING is defined. */
593 nb->bDoTime = (getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
595 /* local/non-local GPU streams */
596 nb->deviceStreams[InteractionLocality::Local].init(*nb->deviceInfo, nb->dev_rundata->deviceContext_,
597 DeviceStreamPriority::Normal, nb->bDoTime);
599 if (nb->bUseTwoStreams)
601 init_plist(nb->plist[InteractionLocality::NonLocal]);
603 nb->deviceStreams[InteractionLocality::NonLocal].init(
604 *nb->deviceInfo, nb->dev_rundata->deviceContext_, DeviceStreamPriority::High, nb->bDoTime);
609 init_timings(nb->timings);
612 nbnxn_ocl_init_const(nb->atdat, nb->nbparam, ic, listParams, nbat->params(),
613 nb->dev_rundata->deviceContext_);
615 /* Enable LJ param manual prefetch for AMD or Intel or if we request through env. var.
616 * TODO: decide about NVIDIA
618 nb->bPrefetchLjParam = (getenv("GMX_OCL_DISABLE_I_PREFETCH") == nullptr)
619 && ((nb->deviceInfo->deviceVendor == DeviceVendor::Amd)
620 || (nb->deviceInfo->deviceVendor == DeviceVendor::Intel)
621 || (getenv("GMX_OCL_ENABLE_I_PREFETCH") != nullptr));
623 /* NOTE: in CUDA we pick L1 cache configuration for the nbnxn kernels here,
624 * but sadly this is not supported in OpenCL (yet?). Consider adding it if
625 * it becomes supported.
627 nbnxn_gpu_compile_kernels(nb);
628 nbnxn_gpu_init_kernels(nb);
630 /* clear energy and shift force outputs */
631 nbnxn_ocl_clear_e_fshift(nb);
635 fprintf(debug, "Initialized OpenCL data structures.\n");
641 /*! \brief Clears the first natoms_clear elements of the GPU nonbonded force output array.
643 static void nbnxn_ocl_clear_f(NbnxmGpu* nb, int natoms_clear)
645 if (natoms_clear == 0)
650 cl_int gmx_used_in_debug cl_error;
652 cl_atomdata_t* atomData = nb->atdat;
653 cl_command_queue ls = nb->deviceStreams[InteractionLocality::Local].stream();
654 cl_float value = 0.0F;
656 cl_error = clEnqueueFillBuffer(ls, atomData->f, &value, sizeof(cl_float), 0,
657 natoms_clear * sizeof(rvec), 0, nullptr, nullptr);
658 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
659 ("clEnqueueFillBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
662 //! This function is documented in the header file
663 void gpu_clear_outputs(NbnxmGpu* nb, bool computeVirial)
665 nbnxn_ocl_clear_f(nb, nb->atdat->natoms);
666 /* clear shift force array and energies if the outputs were
667 used in the current step */
670 nbnxn_ocl_clear_e_fshift(nb);
673 /* kick off buffer clearing kernel to ensure concurrency with constraints/update */
674 cl_int gmx_unused cl_error;
675 cl_error = clFlush(nb->deviceStreams[InteractionLocality::Local].stream());
676 GMX_ASSERT(cl_error == CL_SUCCESS, ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
679 //! This function is documented in the header file
680 void gpu_init_pairlist(NbnxmGpu* nb, const NbnxnPairlistGpu* h_plist, const InteractionLocality iloc)
683 // Timing accumulation should happen only if there was work to do
684 // because getLastRangeTime() gets skipped with empty lists later
685 // which leads to the counter not being reset.
686 bool bDoTime = (nb->bDoTime && !h_plist->sci.empty());
687 const DeviceStream& deviceStream = nb->deviceStreams[iloc];
688 cl_plist_t* d_plist = nb->plist[iloc];
690 if (d_plist->na_c < 0)
692 d_plist->na_c = h_plist->na_ci;
696 if (d_plist->na_c != h_plist->na_ci)
698 sprintf(sbuf, "In cu_init_plist: the #atoms per cell has changed (from %d to %d)",
699 d_plist->na_c, h_plist->na_ci);
704 gpu_timers_t::Interaction& iTimers = nb->timers->interaction[iloc];
708 iTimers.pl_h2d.openTimingRegion(deviceStream);
709 iTimers.didPairlistH2D = true;
712 // TODO most of this function is same in CUDA and OpenCL, move into the header
713 const DeviceContext& deviceContext = nb->dev_rundata->deviceContext_;
715 reallocateDeviceBuffer(&d_plist->sci, h_plist->sci.size(), &d_plist->nsci, &d_plist->sci_nalloc,
717 copyToDeviceBuffer(&d_plist->sci, h_plist->sci.data(), 0, h_plist->sci.size(), deviceStream,
718 GpuApiCallBehavior::Async, bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
720 reallocateDeviceBuffer(&d_plist->cj4, h_plist->cj4.size(), &d_plist->ncj4, &d_plist->cj4_nalloc,
722 copyToDeviceBuffer(&d_plist->cj4, h_plist->cj4.data(), 0, h_plist->cj4.size(), deviceStream,
723 GpuApiCallBehavior::Async, bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
725 reallocateDeviceBuffer(&d_plist->imask, h_plist->cj4.size() * c_nbnxnGpuClusterpairSplit,
726 &d_plist->nimask, &d_plist->imask_nalloc, deviceContext);
728 reallocateDeviceBuffer(&d_plist->excl, h_plist->excl.size(), &d_plist->nexcl,
729 &d_plist->excl_nalloc, deviceContext);
730 copyToDeviceBuffer(&d_plist->excl, h_plist->excl.data(), 0, h_plist->excl.size(), deviceStream,
731 GpuApiCallBehavior::Async, bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
735 iTimers.pl_h2d.closeTimingRegion(deviceStream);
738 /* need to prune the pair list during the next step */
739 d_plist->haveFreshList = true;
742 //! This function is documented in the header file
743 void gpu_upload_shiftvec(NbnxmGpu* nb, const nbnxn_atomdata_t* nbatom)
745 cl_atomdata_t* adat = nb->atdat;
746 cl_command_queue ls = nb->deviceStreams[InteractionLocality::Local].stream();
748 /* only if we have a dynamic box */
749 if (nbatom->bDynamicBox || !adat->bShiftVecUploaded)
751 ocl_copy_H2D_async(adat->shift_vec, nbatom->shift_vec.data(), 0,
752 SHIFTS * sizeof(nbatom->shift_vec[0]), ls, nullptr);
753 adat->bShiftVecUploaded = CL_TRUE;
757 //! This function is documented in the header file
758 void gpu_init_atomdata(NbnxmGpu* nb, const nbnxn_atomdata_t* nbat)
763 bool bDoTime = nb->bDoTime;
764 cl_timers_t* timers = nb->timers;
765 cl_atomdata_t* d_atdat = nb->atdat;
766 const DeviceStream& deviceStream = nb->deviceStreams[InteractionLocality::Local];
768 natoms = nbat->numAtoms();
773 /* time async copy */
774 timers->atdat.openTimingRegion(deviceStream);
777 /* need to reallocate if we have to copy more atoms than the amount of space
778 available and only allocate if we haven't initialized yet, i.e d_atdat->natoms == -1 */
779 if (natoms > d_atdat->nalloc)
781 nalloc = over_alloc_small(natoms);
783 /* free up first if the arrays have already been initialized */
784 if (d_atdat->nalloc != -1)
786 freeDeviceBuffer(&d_atdat->f);
787 freeDeviceBuffer(&d_atdat->xq);
788 freeDeviceBuffer(&d_atdat->lj_comb);
789 freeDeviceBuffer(&d_atdat->atom_types);
792 d_atdat->f = clCreateBuffer(nb->dev_rundata->deviceContext_.context(),
793 CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
794 nalloc * DIM * sizeof(nbat->out[0].f[0]), nullptr, &cl_error);
795 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
796 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
798 d_atdat->xq = clCreateBuffer(nb->dev_rundata->deviceContext_.context(),
799 CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
800 nalloc * sizeof(cl_float4), nullptr, &cl_error);
801 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
802 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
804 if (useLjCombRule(nb->nbparam->vdwtype))
806 d_atdat->lj_comb = clCreateBuffer(nb->dev_rundata->deviceContext_.context(),
807 CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
808 nalloc * sizeof(cl_float2), nullptr, &cl_error);
809 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
810 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
814 d_atdat->atom_types = clCreateBuffer(nb->dev_rundata->deviceContext_.context(),
815 CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
816 nalloc * sizeof(int), nullptr, &cl_error);
817 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
818 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
821 d_atdat->nalloc = nalloc;
825 d_atdat->natoms = natoms;
826 d_atdat->natoms_local = nbat->natoms_local;
828 /* need to clear GPU f output if realloc happened */
831 nbnxn_ocl_clear_f(nb, nalloc);
834 if (useLjCombRule(nb->nbparam->vdwtype))
836 ocl_copy_H2D_async(d_atdat->lj_comb, nbat->params().lj_comb.data(), 0, natoms * sizeof(cl_float2),
837 deviceStream.stream(), bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
841 ocl_copy_H2D_async(d_atdat->atom_types, nbat->params().type.data(), 0, natoms * sizeof(int),
842 deviceStream.stream(), bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
847 timers->atdat.closeTimingRegion(deviceStream);
850 /* kick off the tasks enqueued above to ensure concurrency with the search */
851 cl_error = clFlush(deviceStream.stream());
852 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
853 ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
856 /*! \brief Releases an OpenCL kernel pointer */
857 static void free_kernel(cl_kernel* kernel_ptr)
859 cl_int gmx_unused cl_error;
861 GMX_ASSERT(kernel_ptr, "Need a valid kernel pointer");
865 cl_error = clReleaseKernel(*kernel_ptr);
866 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
867 ("clReleaseKernel failed: " + ocl_get_error_string(cl_error)).c_str());
869 *kernel_ptr = nullptr;
873 /*! \brief Releases a list of OpenCL kernel pointers */
874 static void free_kernels(cl_kernel* kernels, int count)
878 for (i = 0; i < count; i++)
880 free_kernel(kernels + i);
884 /*! \brief Free the OpenCL program.
886 * The function releases the OpenCL program assuciated with the
887 * device that the calling PP rank is running on.
889 * \param program [in] OpenCL program to release.
891 static void freeGpuProgram(cl_program program)
895 cl_int cl_error = clReleaseProgram(program);
896 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
897 ("clReleaseProgram failed: " + ocl_get_error_string(cl_error)).c_str());
902 //! This function is documented in the header file
903 void gpu_free(NbnxmGpu* nb)
911 int kernel_count = sizeof(nb->kernel_ener_noprune_ptr) / sizeof(nb->kernel_ener_noprune_ptr[0][0]);
912 free_kernels(nb->kernel_ener_noprune_ptr[0], kernel_count);
914 kernel_count = sizeof(nb->kernel_ener_prune_ptr) / sizeof(nb->kernel_ener_prune_ptr[0][0]);
915 free_kernels(nb->kernel_ener_prune_ptr[0], kernel_count);
917 kernel_count = sizeof(nb->kernel_noener_noprune_ptr) / sizeof(nb->kernel_noener_noprune_ptr[0][0]);
918 free_kernels(nb->kernel_noener_noprune_ptr[0], kernel_count);
920 kernel_count = sizeof(nb->kernel_noener_prune_ptr) / sizeof(nb->kernel_noener_prune_ptr[0][0]);
921 free_kernels(nb->kernel_noener_prune_ptr[0], kernel_count);
923 free_kernel(&(nb->kernel_zero_e_fshift));
926 freeDeviceBuffer(&(nb->atdat->xq));
927 freeDeviceBuffer(&(nb->atdat->f));
928 freeDeviceBuffer(&(nb->atdat->e_lj));
929 freeDeviceBuffer(&(nb->atdat->e_el));
930 freeDeviceBuffer(&(nb->atdat->fshift));
931 freeDeviceBuffer(&(nb->atdat->lj_comb));
932 freeDeviceBuffer(&(nb->atdat->atom_types));
933 freeDeviceBuffer(&(nb->atdat->shift_vec));
937 freeDeviceBuffer(&(nb->nbparam->nbfp_climg2d));
938 freeDeviceBuffer(&(nb->nbparam->nbfp_comb_climg2d));
939 freeDeviceBuffer(&(nb->nbparam->coulomb_tab_climg2d));
943 auto* plist = nb->plist[InteractionLocality::Local];
944 freeDeviceBuffer(&plist->sci);
945 freeDeviceBuffer(&plist->cj4);
946 freeDeviceBuffer(&plist->imask);
947 freeDeviceBuffer(&plist->excl);
949 if (nb->bUseTwoStreams)
951 auto* plist_nl = nb->plist[InteractionLocality::NonLocal];
952 freeDeviceBuffer(&plist_nl->sci);
953 freeDeviceBuffer(&plist_nl->cj4);
954 freeDeviceBuffer(&plist_nl->imask);
955 freeDeviceBuffer(&plist_nl->excl);
960 pfree(nb->nbst.e_lj);
961 nb->nbst.e_lj = nullptr;
963 pfree(nb->nbst.e_el);
964 nb->nbst.e_el = nullptr;
966 pfree(nb->nbst.fshift);
967 nb->nbst.fshift = nullptr;
969 /* Free other events */
970 if (nb->nonlocal_done)
972 clReleaseEvent(nb->nonlocal_done);
973 nb->nonlocal_done = nullptr;
975 if (nb->misc_ops_and_local_H2D_done)
977 clReleaseEvent(nb->misc_ops_and_local_H2D_done);
978 nb->misc_ops_and_local_H2D_done = nullptr;
981 freeGpuProgram(nb->dev_rundata->program);
982 delete nb->dev_rundata;
984 /* Free timers and timings */
991 fprintf(debug, "Cleaned up OpenCL data structures.\n");
995 //! This function is documented in the header file
996 gmx_wallclock_gpu_nbnxn_t* gpu_get_timings(NbnxmGpu* nb)
998 return (nb != nullptr && nb->bDoTime) ? nb->timings : nullptr;
1001 //! This function is documented in the header file
1002 void gpu_reset_timings(nonbonded_verlet_t* nbv)
1004 if (nbv->gpu_nbv && nbv->gpu_nbv->bDoTime)
1006 init_timings(nbv->gpu_nbv->timings);
1010 //! This function is documented in the header file
1011 int gpu_min_ci_balanced(NbnxmGpu* nb)
1013 return nb != nullptr ? gpu_min_ci_balanced_factor * nb->deviceInfo->compute_units : 0;
1016 //! This function is documented in the header file
1017 gmx_bool gpu_is_kernel_ewald_analytical(const NbnxmGpu* nb)
1019 return ((nb->nbparam->eeltype == eelOclEWALD_ANA) || (nb->nbparam->eeltype == eelOclEWALD_ANA_TWIN));
1022 } // namespace Nbnxm