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>
53 // TODO We would like to move this down, but the way gmx_nbnxn_gpu_t
54 // is currently declared means this has to be before gpu_types.h
55 #include "nbnxm_ocl_types.h"
57 // TODO Remove this comment when the above order issue is resolved
58 #include "gromacs/gpu_utils/gpu_utils.h"
59 #include "gromacs/gpu_utils/oclutils.h"
60 #include "gromacs/hardware/gpu_hw_info.h"
61 #include "gromacs/math/vectypes.h"
62 #include "gromacs/mdlib/force_flags.h"
63 #include "gromacs/mdtypes/interaction_const.h"
64 #include "gromacs/mdtypes/md_enums.h"
65 #include "gromacs/nbnxm/atomdata.h"
66 #include "gromacs/nbnxm/gpu_data_mgmt.h"
67 #include "gromacs/nbnxm/gpu_jit_support.h"
68 #include "gromacs/nbnxm/nbnxm.h"
69 #include "gromacs/nbnxm/nbnxm_gpu.h"
70 #include "gromacs/pbcutil/ishift.h"
71 #include "gromacs/timing/gpu_timing.h"
72 #include "gromacs/utility/cstringutil.h"
73 #include "gromacs/utility/fatalerror.h"
74 #include "gromacs/utility/gmxassert.h"
75 #include "gromacs/utility/real.h"
76 #include "gromacs/utility/smalloc.h"
78 #include "nbnxm_ocl_internal.h"
83 /*! \brief Copies of values from cl_driver_diagnostics_intel.h,
84 * which isn't guaranteed to be available. */
86 #define CL_CONTEXT_SHOW_DIAGNOSTICS_INTEL 0x4106
87 #define CL_CONTEXT_DIAGNOSTICS_LEVEL_GOOD_INTEL 0x1
88 #define CL_CONTEXT_DIAGNOSTICS_LEVEL_BAD_INTEL 0x2
89 #define CL_CONTEXT_DIAGNOSTICS_LEVEL_NEUTRAL_INTEL 0x4
92 /*! \brief This parameter should be determined heuristically from the
93 * kernel execution times
95 * This value is best for small systems on a single AMD Radeon R9 290X
96 * (and about 5% faster than 40, which is the default for CUDA
97 * devices). Larger simulation systems were quite insensitive to the
98 * value of this parameter.
100 static unsigned int gpu_min_ci_balanced_factor = 50;
103 /*! \brief Returns true if LJ combination rules are used in the non-bonded kernels.
105 * Full doc in nbnxn_ocl_internal.h */
106 bool useLjCombRule(int vdwType)
108 return (vdwType == evdwOclCUTCOMBGEOM ||
109 vdwType == evdwOclCUTCOMBLB);
112 /*! \brief Tabulates the Ewald Coulomb force and initializes the size/scale
113 * and the table GPU array.
115 * If called with an already allocated table, it just re-uploads the
118 static void init_ewald_coulomb_force_table(const interaction_const_t *ic,
120 const gmx_device_runtime_data_t *runData)
126 if (nbp->coulomb_tab_climg2d != nullptr)
128 freeDeviceBuffer(&(nbp->coulomb_tab_climg2d));
131 /* Switched from using textures to using buffers */
132 // TODO: decide which alternative is most efficient - textures or buffers.
134 cl_image_format array_format;
136 array_format.image_channel_data_type = CL_FLOAT;
137 array_format.image_channel_order = CL_R;
139 coul_tab = clCreateImage2D(runData->context, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
140 &array_format, tabsize, 1, 0, ftmp, &cl_error);
143 coul_tab = clCreateBuffer(runData->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
144 ic->tabq_size*sizeof(cl_float), ic->tabq_coul_F, &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 = ic->tabq_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, gmx_device_runtime_data_t *runData)
162 /* An element of the shift_vec device buffer has the same size as one element
163 of the host side shift_vec buffer. */
164 ad->shift_vec_elem_size = sizeof(*nbnxn_atomdata_t::shift_vec.data());
166 ad->shift_vec = clCreateBuffer(runData->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
167 SHIFTS * ad->shift_vec_elem_size, nullptr, &cl_error);
168 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
169 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
170 ad->bShiftVecUploaded = CL_FALSE;
172 /* An element of the fshift device buffer has the same size as one element
173 of the host side fshift buffer. */
174 ad->fshift_elem_size = sizeof(*cl_nb_staging_t::fshift);
176 ad->fshift = clCreateBuffer(runData->context, CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
177 SHIFTS * ad->fshift_elem_size, nullptr, &cl_error);
178 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
179 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
181 ad->e_lj = clCreateBuffer(runData->context, CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
182 sizeof(float), nullptr, &cl_error);
183 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
184 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
186 ad->e_el = clCreateBuffer(runData->context, CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
187 sizeof(float), nullptr, &cl_error);
188 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
189 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
191 /* initialize to nullptr pointers to data that is not allocated here and will
192 need reallocation in nbnxn_gpu_init_atomdata */
196 /* size -1 indicates that the respective array hasn't been initialized yet */
201 /*! \brief Copies all parameters related to the cut-off from ic to nbp
203 static void set_cutoff_parameters(cl_nbparam_t *nbp,
204 const interaction_const_t *ic,
205 const NbnxnListParameters *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
233 map_interaction_types_to_gpu_kernel_flavors(const interaction_const_t *ic,
238 if (ic->vdwtype == evdwCUT)
240 switch (ic->vdw_modifier)
243 case eintmodPOTSHIFT:
247 *gpu_vdwtype = evdwOclCUT;
250 *gpu_vdwtype = evdwOclCUTCOMBGEOM;
253 *gpu_vdwtype = evdwOclCUTCOMBLB;
256 gmx_incons("The requested LJ combination rule is not implemented in the OpenCL GPU accelerated kernels!");
259 case eintmodFORCESWITCH:
260 *gpu_vdwtype = evdwOclFSWITCH;
262 case eintmodPOTSWITCH:
263 *gpu_vdwtype = evdwOclPSWITCH;
266 gmx_incons("The requested VdW interaction modifier is not implemented in the GPU accelerated kernels!");
269 else if (ic->vdwtype == evdwPME)
271 if (ic->ljpme_comb_rule == ljcrGEOM)
273 *gpu_vdwtype = evdwOclEWALDGEOM;
277 *gpu_vdwtype = evdwOclEWALDLB;
282 gmx_incons("The requested VdW type is not implemented in the GPU accelerated kernels!");
285 if (ic->eeltype == eelCUT)
287 *gpu_eeltype = eelOclCUT;
289 else if (EEL_RF(ic->eeltype))
291 *gpu_eeltype = eelOclRF;
293 else if ((EEL_PME(ic->eeltype) || ic->eeltype == eelEWALD))
295 /* Initially rcoulomb == rvdw, so it's surely not twin cut-off. */
296 *gpu_eeltype = gpu_pick_ewald_kernel_type(false);
300 /* Shouldn't happen, as this is checked when choosing Verlet-scheme */
301 gmx_incons("The requested electrostatics type is not implemented in the GPU accelerated kernels!");
305 /*! \brief Initializes the nonbonded parameter data structure.
307 static void init_nbparam(cl_nbparam_t *nbp,
308 const interaction_const_t *ic,
309 const NbnxnListParameters *listParams,
310 const nbnxn_atomdata_t::Params &nbatParams,
311 const gmx_device_runtime_data_t *runData)
315 set_cutoff_parameters(nbp, ic, listParams);
317 map_interaction_types_to_gpu_kernel_flavors(ic,
318 nbatParams.comb_rule,
322 if (ic->vdwtype == evdwPME)
324 if (ic->ljpme_comb_rule == ljcrGEOM)
326 GMX_ASSERT(nbatParams.comb_rule == ljcrGEOM, "Combination rule mismatch!");
330 GMX_ASSERT(nbatParams.comb_rule == ljcrLB, "Combination rule mismatch!");
333 /* generate table for PME */
334 nbp->coulomb_tab_climg2d = nullptr;
335 if (nbp->eeltype == eelOclEWALD_TAB || nbp->eeltype == eelOclEWALD_TAB_TWIN)
337 init_ewald_coulomb_force_table(ic, nbp, runData);
340 // TODO: improvement needed.
341 // The image2d is created here even if eeltype is not eelCuEWALD_TAB or eelCuEWALD_TAB_TWIN because the OpenCL kernels
342 // don't accept nullptr values for image2D parameters.
344 /* Switched from using textures to using buffers */
345 // TODO: decide which alternative is most efficient - textures or buffers.
347 cl_image_format array_format;
349 array_format.image_channel_data_type = CL_FLOAT;
350 array_format.image_channel_order = CL_R;
352 nbp->coulomb_tab_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE,
353 &array_format, 1, 1, 0, nullptr, &cl_error);
356 nbp->coulomb_tab_climg2d = clCreateBuffer(runData->context, CL_MEM_READ_ONLY, sizeof(cl_float), nullptr, &cl_error);
357 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
358 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
361 const int nnbfp = 2*nbatParams.numTypes*nbatParams.numTypes;
362 const int nnbfp_comb = 2*nbatParams.numTypes;
365 /* Switched from using textures to using buffers */
366 // TODO: decide which alternative is most efficient - textures or buffers.
368 cl_image_format array_format;
370 array_format.image_channel_data_type = CL_FLOAT;
371 array_format.image_channel_order = CL_R;
373 nbp->nbfp_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
374 &array_format, nnbfp, 1, 0, nbat->nbfp, &cl_error);
378 clCreateBuffer(runData->context,
379 CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
380 nnbfp*sizeof(cl_float),
381 const_cast<float *>(nbatParams.nbfp.data()),
383 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
384 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
386 if (ic->vdwtype == evdwPME)
388 /* Switched from using textures to using buffers */
389 // TODO: decide which alternative is most efficient - textures or buffers.
390 /* nbp->nbfp_comb_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
391 &array_format, nnbfp_comb, 1, 0, nbat->nbfp_comb, &cl_error);*/
392 nbp->nbfp_comb_climg2d =
393 clCreateBuffer(runData->context,
394 CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
395 nnbfp_comb*sizeof(cl_float),
396 const_cast<float *>(nbatParams.nbfp_comb.data()),
398 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
399 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
403 // TODO: improvement needed.
404 // The image2d is created here even if vdwtype is not evdwPME because the OpenCL kernels
405 // don't accept nullptr values for image2D parameters.
406 /* Switched from using textures to using buffers */
407 // TODO: decide which alternative is most efficient - textures or buffers.
408 /* nbp->nbfp_comb_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE,
409 &array_format, 1, 1, 0, nullptr, &cl_error);*/
410 nbp->nbfp_comb_climg2d = clCreateBuffer(runData->context, CL_MEM_READ_ONLY, sizeof(cl_float), nullptr, &cl_error);
411 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
412 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
417 //! This function is documented in the header file
418 void gpu_pme_loadbal_update_param(const nonbonded_verlet_t *nbv,
419 const interaction_const_t *ic,
420 const NbnxnListParameters *listParams)
422 if (!nbv || !nbv->useGpu())
426 gmx_nbnxn_ocl_t *nb = nbv->gpu_nbv;
427 cl_nbparam_t *nbp = nb->nbparam;
429 set_cutoff_parameters(nbp, ic, listParams);
431 nbp->eeltype = gpu_pick_ewald_kernel_type(ic->rcoulomb != ic->rvdw);
433 init_ewald_coulomb_force_table(ic, nbp, nb->dev_rundata);
436 /*! \brief Initializes the pair list data structure.
438 static void init_plist(cl_plist_t *pl)
440 /* initialize to nullptr pointers to data that is not allocated here and will
441 need reallocation in nbnxn_gpu_init_pairlist */
447 /* size -1 indicates that the respective array hasn't been initialized yet */
454 pl->imask_nalloc = -1;
456 pl->excl_nalloc = -1;
457 pl->haveFreshList = false;
460 /*! \brief Initializes the timings data structure.
462 static void init_timings(gmx_wallclock_gpu_nbnxn_t *t)
471 for (i = 0; i < 2; i++)
473 for (j = 0; j < 2; j++)
475 t->ktime[i][j].t = 0.0;
476 t->ktime[i][j].c = 0;
481 t->pruneTime.t = 0.0;
482 t->dynamicPruneTime.c = 0;
483 t->dynamicPruneTime.t = 0.0;
487 //! OpenCL notification callback function
488 static void CL_CALLBACK
489 ocl_notify_fn( const char *pErrInfo, const void *, size_t, void *)
491 if (pErrInfo != NULL)
493 printf("%s\n", pErrInfo ); // Print error/hint
497 /*! \brief Creates context for OpenCL GPU given by \p mygpu
499 * A fatal error results if creation fails.
501 * \param[inout] runtimeData runtime data including program and context
502 * \param[in] devInfo device info struct
503 * \param[in] rank MPI rank (for error reporting)
506 nbnxn_gpu_create_context(gmx_device_runtime_data_t *runtimeData,
507 const gmx_device_info_t *devInfo,
510 cl_context_properties context_properties[5];
511 cl_platform_id platform_id;
512 cl_device_id device_id;
516 assert(runtimeData != nullptr);
517 assert(devInfo != nullptr);
519 platform_id = devInfo->ocl_gpu_id.ocl_platform_id;
520 device_id = devInfo->ocl_gpu_id.ocl_device_id;
523 context_properties[i++] = CL_CONTEXT_PLATFORM;
524 context_properties[i++] = reinterpret_cast<cl_context_properties>(platform_id);
525 if (getenv("GMX_OCL_SHOW_DIAGNOSTICS"))
527 context_properties[i++] = CL_CONTEXT_SHOW_DIAGNOSTICS_INTEL;
528 context_properties[i++] = CL_CONTEXT_DIAGNOSTICS_LEVEL_BAD_INTEL | CL_CONTEXT_DIAGNOSTICS_LEVEL_NEUTRAL_INTEL;
530 context_properties[i++] = 0; /* Terminates the list of properties */
532 context = clCreateContext(context_properties, 1, &device_id, ocl_notify_fn, nullptr, &cl_error);
533 if (CL_SUCCESS != cl_error)
535 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s:\n OpenCL error %d: %s",
537 devInfo->device_name,
538 cl_error, ocl_get_error_string(cl_error).c_str());
541 runtimeData->context = context;
544 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
545 static cl_kernel nbnxn_gpu_create_kernel(gmx_nbnxn_ocl_t *nb,
546 const char *kernel_name)
551 kernel = clCreateKernel(nb->dev_rundata->program, kernel_name, &cl_error);
552 if (CL_SUCCESS != cl_error)
554 gmx_fatal(FARGS, "Failed to create kernel '%s' for GPU #%s: OpenCL error %d",
556 nb->dev_info->device_name,
563 /*! \brief Clears nonbonded shift force output array and energy outputs on the GPU.
566 nbnxn_ocl_clear_e_fshift(gmx_nbnxn_ocl_t *nb)
570 cl_atomdata_t * adat = nb->atdat;
571 cl_command_queue ls = nb->stream[InteractionLocality::Local];
573 size_t local_work_size[3] = {1, 1, 1};
574 size_t global_work_size[3] = {1, 1, 1};
576 cl_int shifts = SHIFTS*3;
580 cl_kernel zero_e_fshift = nb->kernel_zero_e_fshift;
582 local_work_size[0] = 64;
583 // Round the total number of threads up from the array size
584 global_work_size[0] = ((shifts + local_work_size[0] - 1)/local_work_size[0])*local_work_size[0];
587 cl_error = clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->fshift));
588 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_lj));
589 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_el));
590 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_uint), &shifts);
591 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
593 cl_error = clEnqueueNDRangeKernel(ls, zero_e_fshift, 3, nullptr, global_work_size, local_work_size, 0, nullptr, nullptr);
594 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
597 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
598 static void nbnxn_gpu_init_kernels(gmx_nbnxn_ocl_t *nb)
600 /* Init to 0 main kernel arrays */
601 /* They will be later on initialized in select_nbnxn_kernel */
602 // TODO: consider always creating all variants of the kernels here so that there is no
603 // need for late call to clCreateKernel -- if that gives any advantage?
604 memset(nb->kernel_ener_noprune_ptr, 0, sizeof(nb->kernel_ener_noprune_ptr));
605 memset(nb->kernel_ener_prune_ptr, 0, sizeof(nb->kernel_ener_prune_ptr));
606 memset(nb->kernel_noener_noprune_ptr, 0, sizeof(nb->kernel_noener_noprune_ptr));
607 memset(nb->kernel_noener_prune_ptr, 0, sizeof(nb->kernel_noener_prune_ptr));
609 /* Init pruning kernels
611 * TODO: we could avoid creating kernels if dynamic pruning is turned off,
612 * but ATM that depends on force flags not passed into the initialization.
614 nb->kernel_pruneonly[epruneFirst] = nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_opencl");
615 nb->kernel_pruneonly[epruneRolling] = nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_rolling_opencl");
617 /* Init auxiliary kernels */
618 nb->kernel_zero_e_fshift = nbnxn_gpu_create_kernel(nb, "zero_e_fshift");
621 /*! \brief Initializes simulation constant data.
623 * Initializes members of the atomdata and nbparam structs and
624 * clears e/fshift output buffers.
626 static void nbnxn_ocl_init_const(gmx_nbnxn_ocl_t *nb,
627 const interaction_const_t *ic,
628 const NbnxnListParameters *listParams,
629 const nbnxn_atomdata_t::Params &nbatParams)
631 init_atomdata_first(nb->atdat, nbatParams.numTypes, nb->dev_rundata);
632 init_nbparam(nb->nbparam, ic, listParams, nbatParams, nb->dev_rundata);
636 //! This function is documented in the header file
638 gpu_init(const gmx_device_info_t *deviceInfo,
639 const interaction_const_t *ic,
640 const NbnxnListParameters *listParams,
641 const nbnxn_atomdata_t *nbat,
643 const gmx_bool bLocalAndNonlocal)
647 cl_command_queue_properties queue_properties;
653 snew(nb->nbparam, 1);
654 snew(nb->plist[InteractionLocality::Local], 1);
655 if (bLocalAndNonlocal)
657 snew(nb->plist[InteractionLocality::NonLocal], 1);
660 nb->bUseTwoStreams = static_cast<cl_bool>(bLocalAndNonlocal);
662 nb->timers = new cl_timers_t();
663 snew(nb->timings, 1);
665 /* set device info, just point it to the right GPU among the detected ones */
666 nb->dev_info = deviceInfo;
667 snew(nb->dev_rundata, 1);
670 pmalloc(reinterpret_cast<void**>(&nb->nbst.e_lj), sizeof(*nb->nbst.e_lj));
671 pmalloc(reinterpret_cast<void**>(&nb->nbst.e_el), sizeof(*nb->nbst.e_el));
672 pmalloc(reinterpret_cast<void**>(&nb->nbst.fshift), SHIFTS * sizeof(*nb->nbst.fshift));
674 init_plist(nb->plist[InteractionLocality::Local]);
676 /* OpenCL timing disabled if GMX_DISABLE_GPU_TIMING is defined. */
677 nb->bDoTime = static_cast<cl_bool>(getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
679 /* Create queues only after bDoTime has been initialized */
682 queue_properties = CL_QUEUE_PROFILING_ENABLE;
686 queue_properties = 0;
689 nbnxn_gpu_create_context(nb->dev_rundata, nb->dev_info, rank);
691 /* local/non-local GPU streams */
692 nb->stream[InteractionLocality::Local] =
693 clCreateCommandQueue(nb->dev_rundata->context, nb->dev_info->ocl_gpu_id.ocl_device_id, queue_properties, &cl_error);
694 if (CL_SUCCESS != cl_error)
696 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s: OpenCL error %d",
698 nb->dev_info->device_name,
702 if (nb->bUseTwoStreams)
704 init_plist(nb->plist[InteractionLocality::NonLocal]);
706 nb->stream[InteractionLocality::NonLocal] =
707 clCreateCommandQueue(nb->dev_rundata->context, nb->dev_info->ocl_gpu_id.ocl_device_id, queue_properties, &cl_error);
708 if (CL_SUCCESS != cl_error)
710 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s: OpenCL error %d",
712 nb->dev_info->device_name,
719 init_timings(nb->timings);
722 nbnxn_ocl_init_const(nb, ic, listParams, nbat->params());
724 /* Enable LJ param manual prefetch for AMD or Intel or if we request through env. var.
725 * TODO: decide about NVIDIA
727 nb->bPrefetchLjParam =
728 (getenv("GMX_OCL_DISABLE_I_PREFETCH") == nullptr) &&
729 ((nb->dev_info->vendor_e == OCL_VENDOR_AMD) || (nb->dev_info->vendor_e == OCL_VENDOR_INTEL)
730 || (getenv("GMX_OCL_ENABLE_I_PREFETCH") != nullptr));
732 /* NOTE: in CUDA we pick L1 cache configuration for the nbnxn kernels here,
733 * but sadly this is not supported in OpenCL (yet?). Consider adding it if
734 * it becomes supported.
736 nbnxn_gpu_compile_kernels(nb);
737 nbnxn_gpu_init_kernels(nb);
739 /* clear energy and shift force outputs */
740 nbnxn_ocl_clear_e_fshift(nb);
744 fprintf(debug, "Initialized OpenCL data structures.\n");
750 /*! \brief Clears the first natoms_clear elements of the GPU nonbonded force output array.
752 static void nbnxn_ocl_clear_f(gmx_nbnxn_ocl_t *nb, int natoms_clear)
754 if (natoms_clear == 0)
759 cl_int gmx_used_in_debug cl_error;
761 cl_atomdata_t *atomData = nb->atdat;
762 cl_command_queue ls = nb->stream[InteractionLocality::Local];
763 cl_float value = 0.0f;
765 cl_error = clEnqueueFillBuffer(ls, atomData->f, &value, sizeof(cl_float),
766 0, natoms_clear*sizeof(rvec), 0, nullptr, nullptr);
767 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS, ("clEnqueueFillBuffer failed: " +
768 ocl_get_error_string(cl_error)).c_str());
771 //! This function is documented in the header file
773 gpu_clear_outputs(gmx_nbnxn_ocl_t *nb,
776 nbnxn_ocl_clear_f(nb, nb->atdat->natoms);
777 /* clear shift force array and energies if the outputs were
778 used in the current step */
779 if (flags & GMX_FORCE_VIRIAL)
781 nbnxn_ocl_clear_e_fshift(nb);
784 /* kick off buffer clearing kernel to ensure concurrency with constraints/update */
785 cl_int gmx_unused cl_error;
786 cl_error = clFlush(nb->stream[InteractionLocality::Local]);
787 assert(CL_SUCCESS == cl_error);
790 //! This function is documented in the header file
791 void gpu_init_pairlist(gmx_nbnxn_ocl_t *nb,
792 const NbnxnPairlistGpu *h_plist,
793 const InteractionLocality iloc)
796 // Timing accumulation should happen only if there was work to do
797 // because getLastRangeTime() gets skipped with empty lists later
798 // which leads to the counter not being reset.
799 bool bDoTime = ((nb->bDoTime == CL_TRUE) && !h_plist->sci.empty());
800 cl_command_queue stream = nb->stream[iloc];
801 cl_plist_t *d_plist = nb->plist[iloc];
803 if (d_plist->na_c < 0)
805 d_plist->na_c = h_plist->na_ci;
809 if (d_plist->na_c != h_plist->na_ci)
811 sprintf(sbuf, "In cu_init_plist: the #atoms per cell has changed (from %d to %d)",
812 d_plist->na_c, h_plist->na_ci);
817 gpu_timers_t::Interaction &iTimers = nb->timers->interaction[iloc];
821 iTimers.pl_h2d.openTimingRegion(stream);
822 iTimers.didPairlistH2D = true;
825 // TODO most of this function is same in CUDA and OpenCL, move into the header
826 Context context = nb->dev_rundata->context;
828 reallocateDeviceBuffer(&d_plist->sci, h_plist->sci.size(),
829 &d_plist->nsci, &d_plist->sci_nalloc, context);
830 copyToDeviceBuffer(&d_plist->sci, h_plist->sci.data(), 0, h_plist->sci.size(),
831 stream, GpuApiCallBehavior::Async,
832 bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
834 reallocateDeviceBuffer(&d_plist->cj4, h_plist->cj4.size(),
835 &d_plist->ncj4, &d_plist->cj4_nalloc, context);
836 copyToDeviceBuffer(&d_plist->cj4, h_plist->cj4.data(), 0, h_plist->cj4.size(),
837 stream, GpuApiCallBehavior::Async,
838 bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
840 reallocateDeviceBuffer(&d_plist->imask, h_plist->cj4.size()*c_nbnxnGpuClusterpairSplit,
841 &d_plist->nimask, &d_plist->imask_nalloc, context);
843 reallocateDeviceBuffer(&d_plist->excl, h_plist->excl.size(),
844 &d_plist->nexcl, &d_plist->excl_nalloc, context);
845 copyToDeviceBuffer(&d_plist->excl, h_plist->excl.data(), 0, h_plist->excl.size(),
846 stream, GpuApiCallBehavior::Async,
847 bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
851 iTimers.pl_h2d.closeTimingRegion(stream);
854 /* need to prune the pair list during the next step */
855 d_plist->haveFreshList = true;
858 //! This function is documented in the header file
859 void gpu_upload_shiftvec(gmx_nbnxn_ocl_t *nb,
860 const nbnxn_atomdata_t *nbatom)
862 cl_atomdata_t *adat = nb->atdat;
863 cl_command_queue ls = nb->stream[InteractionLocality::Local];
865 /* only if we have a dynamic box */
866 if (nbatom->bDynamicBox || !adat->bShiftVecUploaded)
868 ocl_copy_H2D_async(adat->shift_vec, nbatom->shift_vec.data(), 0,
869 SHIFTS * adat->shift_vec_elem_size, ls, nullptr);
870 adat->bShiftVecUploaded = CL_TRUE;
874 //! This function is documented in the header file
875 void gpu_init_atomdata(gmx_nbnxn_ocl_t *nb,
876 const nbnxn_atomdata_t *nbat)
881 bool bDoTime = nb->bDoTime == CL_TRUE;
882 cl_timers_t *timers = nb->timers;
883 cl_atomdata_t *d_atdat = nb->atdat;
884 cl_command_queue ls = nb->stream[InteractionLocality::Local];
886 natoms = nbat->numAtoms();
891 /* time async copy */
892 timers->atdat.openTimingRegion(ls);
895 /* need to reallocate if we have to copy more atoms than the amount of space
896 available and only allocate if we haven't initialized yet, i.e d_atdat->natoms == -1 */
897 if (natoms > d_atdat->nalloc)
899 nalloc = over_alloc_small(natoms);
901 /* free up first if the arrays have already been initialized */
902 if (d_atdat->nalloc != -1)
904 freeDeviceBuffer(&d_atdat->f);
905 freeDeviceBuffer(&d_atdat->xq);
906 freeDeviceBuffer(&d_atdat->lj_comb);
907 freeDeviceBuffer(&d_atdat->atom_types);
910 d_atdat->f_elem_size = sizeof(rvec);
912 d_atdat->f = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
913 nalloc * d_atdat->f_elem_size, nullptr, &cl_error);
914 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
915 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
917 d_atdat->xq = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
918 nalloc * sizeof(cl_float4), nullptr, &cl_error);
919 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
920 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
922 if (useLjCombRule(nb->nbparam->vdwtype))
924 d_atdat->lj_comb = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
925 nalloc * sizeof(cl_float2), nullptr, &cl_error);
926 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
927 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
931 d_atdat->atom_types = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
932 nalloc * sizeof(int), nullptr, &cl_error);
933 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
934 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
937 d_atdat->nalloc = nalloc;
941 d_atdat->natoms = natoms;
942 d_atdat->natoms_local = nbat->natoms_local;
944 /* need to clear GPU f output if realloc happened */
947 nbnxn_ocl_clear_f(nb, nalloc);
950 if (useLjCombRule(nb->nbparam->vdwtype))
952 ocl_copy_H2D_async(d_atdat->lj_comb, nbat->params().lj_comb.data(), 0,
953 natoms*sizeof(cl_float2), ls, bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
957 ocl_copy_H2D_async(d_atdat->atom_types, nbat->params().type.data(), 0,
958 natoms*sizeof(int), ls, bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
964 timers->atdat.closeTimingRegion(ls);
967 /* kick off the tasks enqueued above to ensure concurrency with the search */
968 cl_error = clFlush(ls);
969 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
970 ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
973 /*! \brief Releases an OpenCL kernel pointer */
974 static void free_kernel(cl_kernel *kernel_ptr)
976 cl_int gmx_unused cl_error;
978 assert(nullptr != kernel_ptr);
982 cl_error = clReleaseKernel(*kernel_ptr);
983 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS, ("clReleaseKernel failed: " + ocl_get_error_string(cl_error)).c_str());
985 *kernel_ptr = nullptr;
989 /*! \brief Releases a list of OpenCL kernel pointers */
990 static void free_kernels(cl_kernel *kernels, int count)
994 for (i = 0; i < count; i++)
996 free_kernel(kernels + i);
1000 /*! \brief Free the OpenCL runtime data (context and program).
1002 * The function releases the OpenCL context and program assuciated with the
1003 * device that the calling PP rank is running on.
1005 * \param runData [in] porinter to the structure with runtime data.
1007 static void free_gpu_device_runtime_data(gmx_device_runtime_data_t *runData)
1009 if (runData == nullptr)
1014 cl_int gmx_unused cl_error;
1016 if (runData->context)
1018 cl_error = clReleaseContext(runData->context);
1019 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS, ("clReleaseContext failed: " + ocl_get_error_string(cl_error)).c_str());
1020 runData->context = nullptr;
1023 if (runData->program)
1025 cl_error = clReleaseProgram(runData->program);
1026 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS, ("clReleaseProgram failed: " + ocl_get_error_string(cl_error)).c_str());
1027 runData->program = nullptr;
1032 //! This function is documented in the header file
1033 void gpu_free(gmx_nbnxn_ocl_t *nb)
1041 int kernel_count = sizeof(nb->kernel_ener_noprune_ptr) / sizeof(nb->kernel_ener_noprune_ptr[0][0]);
1042 free_kernels(nb->kernel_ener_noprune_ptr[0], kernel_count);
1044 kernel_count = sizeof(nb->kernel_ener_prune_ptr) / sizeof(nb->kernel_ener_prune_ptr[0][0]);
1045 free_kernels(nb->kernel_ener_prune_ptr[0], kernel_count);
1047 kernel_count = sizeof(nb->kernel_noener_noprune_ptr) / sizeof(nb->kernel_noener_noprune_ptr[0][0]);
1048 free_kernels(nb->kernel_noener_noprune_ptr[0], kernel_count);
1050 kernel_count = sizeof(nb->kernel_noener_prune_ptr) / sizeof(nb->kernel_noener_prune_ptr[0][0]);
1051 free_kernels(nb->kernel_noener_prune_ptr[0], kernel_count);
1053 free_kernel(&(nb->kernel_zero_e_fshift));
1056 freeDeviceBuffer(&(nb->atdat->xq));
1057 freeDeviceBuffer(&(nb->atdat->f));
1058 freeDeviceBuffer(&(nb->atdat->e_lj));
1059 freeDeviceBuffer(&(nb->atdat->e_el));
1060 freeDeviceBuffer(&(nb->atdat->fshift));
1061 freeDeviceBuffer(&(nb->atdat->lj_comb));
1062 freeDeviceBuffer(&(nb->atdat->atom_types));
1063 freeDeviceBuffer(&(nb->atdat->shift_vec));
1067 freeDeviceBuffer(&(nb->nbparam->nbfp_climg2d));
1068 freeDeviceBuffer(&(nb->nbparam->nbfp_comb_climg2d));
1069 freeDeviceBuffer(&(nb->nbparam->coulomb_tab_climg2d));
1073 auto *plist = nb->plist[InteractionLocality::Local];
1074 freeDeviceBuffer(&plist->sci);
1075 freeDeviceBuffer(&plist->cj4);
1076 freeDeviceBuffer(&plist->imask);
1077 freeDeviceBuffer(&plist->excl);
1079 if (nb->bUseTwoStreams)
1081 auto *plist_nl = nb->plist[InteractionLocality::NonLocal];
1082 freeDeviceBuffer(&plist_nl->sci);
1083 freeDeviceBuffer(&plist_nl->cj4);
1084 freeDeviceBuffer(&plist_nl->imask);
1085 freeDeviceBuffer(&plist_nl->excl);
1090 pfree(nb->nbst.e_lj);
1091 nb->nbst.e_lj = nullptr;
1093 pfree(nb->nbst.e_el);
1094 nb->nbst.e_el = nullptr;
1096 pfree(nb->nbst.fshift);
1097 nb->nbst.fshift = nullptr;
1099 /* Free command queues */
1100 clReleaseCommandQueue(nb->stream[InteractionLocality::Local]);
1101 nb->stream[InteractionLocality::Local] = nullptr;
1102 if (nb->bUseTwoStreams)
1104 clReleaseCommandQueue(nb->stream[InteractionLocality::NonLocal]);
1105 nb->stream[InteractionLocality::NonLocal] = nullptr;
1107 /* Free other events */
1108 if (nb->nonlocal_done)
1110 clReleaseEvent(nb->nonlocal_done);
1111 nb->nonlocal_done = nullptr;
1113 if (nb->misc_ops_and_local_H2D_done)
1115 clReleaseEvent(nb->misc_ops_and_local_H2D_done);
1116 nb->misc_ops_and_local_H2D_done = nullptr;
1119 free_gpu_device_runtime_data(nb->dev_rundata);
1120 sfree(nb->dev_rundata);
1122 /* Free timers and timings */
1129 fprintf(debug, "Cleaned up OpenCL data structures.\n");
1133 //! This function is documented in the header file
1134 gmx_wallclock_gpu_nbnxn_t *gpu_get_timings(gmx_nbnxn_ocl_t *nb)
1136 return (nb != nullptr && nb->bDoTime) ? nb->timings : nullptr;
1139 //! This function is documented in the header file
1140 void gpu_reset_timings(nonbonded_verlet_t* nbv)
1142 if (nbv->gpu_nbv && nbv->gpu_nbv->bDoTime)
1144 init_timings(nbv->gpu_nbv->timings);
1148 //! This function is documented in the header file
1149 int gpu_min_ci_balanced(gmx_nbnxn_ocl_t *nb)
1151 return nb != nullptr ?
1152 gpu_min_ci_balanced_factor * nb->dev_info->compute_units : 0;
1155 //! This function is documented in the header file
1156 gmx_bool gpu_is_kernel_ewald_analytical(const gmx_nbnxn_ocl_t *nb)
1158 return ((nb->nbparam->eeltype == eelOclEWALD_ANA) ||
1159 (nb->nbparam->eeltype == eelOclEWALD_ANA_TWIN));
1162 } // namespace Nbnxm