2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
5 * Copyright (c) 2001-2012, The GROMACS development team,
6 * check out http://www.gromacs.org for more information.
7 * Copyright (c) 2012, by the GROMACS development team, led by
8 * David van der Spoel, Berk Hess, Erik Lindahl, and including many
9 * others, as listed in the AUTHORS file in the top-level source
10 * directory and at http://www.gromacs.org.
12 * GROMACS is free software; you can redistribute it and/or
13 * modify it under the terms of the GNU Lesser General Public License
14 * as published by the Free Software Foundation; either version 2.1
15 * of the License, or (at your option) any later version.
17 * GROMACS is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20 * Lesser General Public License for more details.
22 * You should have received a copy of the GNU Lesser General Public
23 * License along with GROMACS; if not, see
24 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
25 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
27 * If you want to redistribute modifications to GROMACS, please
28 * consider that scientific software is very special. Version
29 * control is crucial - bugs must be traceable. We will be happy to
30 * consider code for inclusion in the official distribution, but
31 * derived work must not be called official GROMACS. Details are found
32 * in the README & COPYING files - if they are missing, get the
33 * official version at http://www.gromacs.org.
35 * To help us fund GROMACS development, we humbly ask that you cite
36 * the research papers on the package. Check out http://www.gromacs.org.
46 #include "gmx_fatal.h"
50 #include "types/nb_verlet.h"
51 #include "types/interaction_const.h"
52 #include "types/force_flags.h"
53 #include "../nbnxn_consts.h"
55 #include "nbnxn_cuda_types.h"
56 #include "../../gmxlib/cuda_tools/cudautils.cuh"
57 #include "nbnxn_cuda_data_mgmt.h"
58 #include "pmalloc_cuda.h"
59 #include "gpu_utils.h"
61 static bool bUseCudaEventBlockingSync = false; /* makes the CPU thread block */
63 /* This is a heuristically determined parameter for the Fermi architecture for
64 * the minimum size of ci lists by multiplying this constant with the # of
65 * multiprocessors on the current device.
67 static unsigned int gpu_min_ci_balanced_factor = 40;
69 /* Functions from nbnxn_cuda.cu */
70 extern void nbnxn_cuda_set_cacheconfig(cuda_dev_info_t *devinfo);
71 extern const struct texture<float, 1, cudaReadModeElementType>& nbnxn_cuda_get_nbfp_texref();
72 extern const struct texture<float, 1, cudaReadModeElementType>& nbnxn_cuda_get_coulomb_tab_texref();
74 /* We should actually be using md_print_warn in md_logging.c,
75 * but we can't include mpi.h in CUDA code.
77 static void md_print_warn(FILE *fplog, const char *buf)
81 /* We should only print to stderr on the master node,
82 * in most cases fplog is only set on the master node, so this works.
84 fprintf(stderr, "\n%s\n", buf);
85 fprintf(fplog, "\n%s\n", buf);
90 static void nbnxn_cuda_clear_e_fshift(nbnxn_cuda_ptr_t cu_nb);
93 /*! Tabulates the Ewald Coulomb force and initializes the size/scale
94 and the table GPU array. If called with an already allocated table,
95 it just re-uploads the table.
97 static void init_ewald_coulomb_force_table(cu_nbparam_t *nbp)
99 float *ftmp, *coul_tab;
104 tabsize = GPU_EWALD_COULOMB_FORCE_TABLE_SIZE;
105 /* Subtract 2 iso 1 to avoid access out of range due to rounding */
106 tabscale = (tabsize - 2) / sqrt(nbp->rcoulomb_sq);
108 pmalloc((void**)&ftmp, tabsize*sizeof(*ftmp));
110 table_spline3_fill_ewald_lr(ftmp, NULL, NULL, tabsize,
111 1/tabscale, nbp->ewald_beta);
113 /* If the table pointer == NULL the table is generated the first time =>
114 the array pointer will be saved to nbparam and the texture is bound.
116 coul_tab = nbp->coulomb_tab;
117 if (coul_tab == NULL)
119 stat = cudaMalloc((void **)&coul_tab, tabsize*sizeof(*coul_tab));
120 CU_RET_ERR(stat, "cudaMalloc failed on coul_tab");
122 nbp->coulomb_tab = coul_tab;
124 cudaChannelFormatDesc cd = cudaCreateChannelDesc<float>();
125 stat = cudaBindTexture(NULL, &nbnxn_cuda_get_coulomb_tab_texref(),
126 coul_tab, &cd, tabsize*sizeof(*coul_tab));
127 CU_RET_ERR(stat, "cudaBindTexture on coul_tab failed");
130 cu_copy_H2D(coul_tab, ftmp, tabsize*sizeof(*coul_tab));
132 nbp->coulomb_tab_size = tabsize;
133 nbp->coulomb_tab_scale = tabscale;
139 /*! Initializes the atomdata structure first time, it only gets filled at
141 static void init_atomdata_first(cu_atomdata_t *ad, int ntypes)
146 stat = cudaMalloc((void**)&ad->shift_vec, SHIFTS*sizeof(*ad->shift_vec));
147 CU_RET_ERR(stat, "cudaMalloc failed on ad->shift_vec");
148 ad->bShiftVecUploaded = false;
150 stat = cudaMalloc((void**)&ad->fshift, SHIFTS*sizeof(*ad->fshift));
151 CU_RET_ERR(stat, "cudaMalloc failed on ad->fshift");
153 stat = cudaMalloc((void**)&ad->e_lj, sizeof(*ad->e_lj));
154 CU_RET_ERR(stat, "cudaMalloc failed on ad->e_lj");
155 stat = cudaMalloc((void**)&ad->e_el, sizeof(*ad->e_el));
156 CU_RET_ERR(stat, "cudaMalloc failed on ad->e_el");
158 /* initialize to NULL poiters to data that is not allocated here and will
159 need reallocation in nbnxn_cuda_init_atomdata */
163 /* size -1 indicates that the respective array hasn't been initialized yet */
168 /*! Initializes the nonbonded parameter data structure. */
169 static void init_nbparam(cu_nbparam_t *nbp,
170 const interaction_const_t *ic,
171 const nonbonded_verlet_t *nbv)
176 ntypes = nbv->grp[0].nbat->ntype;
178 nbp->ewald_beta = ic->ewaldcoeff;
179 nbp->sh_ewald = ic->sh_ewald;
180 nbp->epsfac = ic->epsfac;
181 nbp->two_k_rf = 2.0 * ic->k_rf;
182 nbp->c_rf = ic->c_rf;
183 nbp->rvdw_sq = ic->rvdw * ic->rvdw;
184 nbp->rcoulomb_sq= ic->rcoulomb * ic->rcoulomb;
185 nbp->rlist_sq = ic->rlist * ic->rlist;
186 nbp->sh_invrc6 = ic->sh_invrc6;
188 if (ic->eeltype == eelCUT)
190 nbp->eeltype = eelCuCUT;
192 else if (EEL_RF(ic->eeltype))
194 nbp->eeltype = eelCuRF;
196 else if ((EEL_PME(ic->eeltype) || ic->eeltype==eelEWALD))
198 /* Initially rcoulomb == rvdw, so it's surely not twin cut-off, unless
199 forced by the env. var. (used only for benchmarking). */
200 if (getenv("GMX_CUDA_NB_EWALD_TWINCUT") == NULL)
202 nbp->eeltype = eelCuEWALD;
206 nbp->eeltype = eelCuEWALD_TWIN;
211 /* Shouldn't happen, as this is checked when choosing Verlet-scheme */
212 gmx_incons("The requested electrostatics type is not implemented in the CUDA GPU accelerated kernels!");
215 /* generate table for PME */
216 if (nbp->eeltype == eelCuEWALD)
218 nbp->coulomb_tab = NULL;
219 init_ewald_coulomb_force_table(nbp);
222 nnbfp = 2*ntypes*ntypes;
223 stat = cudaMalloc((void **)&nbp->nbfp, nnbfp*sizeof(*nbp->nbfp));
224 CU_RET_ERR(stat, "cudaMalloc failed on nbp->nbfp");
225 cu_copy_H2D(nbp->nbfp, nbv->grp[0].nbat->nbfp, nnbfp*sizeof(*nbp->nbfp));
227 cudaChannelFormatDesc cd = cudaCreateChannelDesc<float>();
228 stat = cudaBindTexture(NULL, &nbnxn_cuda_get_nbfp_texref(),
229 nbp->nbfp, &cd, nnbfp*sizeof(*nbp->nbfp));
230 CU_RET_ERR(stat, "cudaBindTexture on nbfp failed");
233 /*! Re-generate the GPU Ewald force table, resets rlist, and update the
234 * electrostatic type switching to twin cut-off (or back) if needed. */
235 void nbnxn_cuda_pme_loadbal_update_param(nbnxn_cuda_ptr_t cu_nb,
236 const interaction_const_t *ic)
238 cu_nbparam_t *nbp = cu_nb->nbparam;
240 nbp->rlist_sq = ic->rlist * ic->rlist;
241 nbp->rcoulomb_sq = ic->rcoulomb * ic->rcoulomb;
242 nbp->ewald_beta = ic->ewaldcoeff;
244 /* When switching to/from twin cut-off, the electrostatics type needs updating.
245 (The env. var. that forces twin cut-off is for benchmarking only!) */
246 if (ic->rcoulomb == ic->rvdw &&
247 getenv("GMX_CUDA_NB_EWALD_TWINCUT") == NULL)
249 nbp->eeltype = eelCuEWALD;
253 nbp->eeltype = eelCuEWALD_TWIN;
256 init_ewald_coulomb_force_table(cu_nb->nbparam);
259 /*! Initializes the pair list data structure. */
260 static void init_plist(cu_plist_t *pl)
262 /* initialize to NULL pointers to data that is not allocated here and will
263 need reallocation in nbnxn_cuda_init_pairlist */
268 /* size -1 indicates that the respective array hasn't been initialized yet */
275 pl->excl_nalloc = -1;
276 pl->bDoPrune = false;
279 /*! Initializes the timer data structure. */
280 static void init_timers(cu_timers_t *t, bool bUseTwoStreams)
283 int eventflags = ( bUseCudaEventBlockingSync ? cudaEventBlockingSync: cudaEventDefault );
285 stat = cudaEventCreateWithFlags(&(t->start_atdat), eventflags);
286 CU_RET_ERR(stat, "cudaEventCreate on start_atdat failed");
287 stat = cudaEventCreateWithFlags(&(t->stop_atdat), eventflags);
288 CU_RET_ERR(stat, "cudaEventCreate on stop_atdat failed");
290 /* The non-local counters/stream (second in the array) are needed only with DD. */
291 for (int i = 0; i <= (bUseTwoStreams ? 1 : 0); i++)
293 stat = cudaEventCreateWithFlags(&(t->start_nb_k[i]), eventflags);
294 CU_RET_ERR(stat, "cudaEventCreate on start_nb_k failed");
295 stat = cudaEventCreateWithFlags(&(t->stop_nb_k[i]), eventflags);
296 CU_RET_ERR(stat, "cudaEventCreate on stop_nb_k failed");
299 stat = cudaEventCreateWithFlags(&(t->start_pl_h2d[i]), eventflags);
300 CU_RET_ERR(stat, "cudaEventCreate on start_pl_h2d failed");
301 stat = cudaEventCreateWithFlags(&(t->stop_pl_h2d[i]), eventflags);
302 CU_RET_ERR(stat, "cudaEventCreate on stop_pl_h2d failed");
304 stat = cudaEventCreateWithFlags(&(t->start_nb_h2d[i]), eventflags);
305 CU_RET_ERR(stat, "cudaEventCreate on start_nb_h2d failed");
306 stat = cudaEventCreateWithFlags(&(t->stop_nb_h2d[i]), eventflags);
307 CU_RET_ERR(stat, "cudaEventCreate on stop_nb_h2d failed");
309 stat = cudaEventCreateWithFlags(&(t->start_nb_d2h[i]), eventflags);
310 CU_RET_ERR(stat, "cudaEventCreate on start_nb_d2h failed");
311 stat = cudaEventCreateWithFlags(&(t->stop_nb_d2h[i]), eventflags);
312 CU_RET_ERR(stat, "cudaEventCreate on stop_nb_d2h failed");
316 /*! Initializes the timings data structure. */
317 static void init_timings(wallclock_gpu_t *t)
326 for (i = 0; i < 2; i++)
328 for(j = 0; j < 2; j++)
330 t->ktime[i][j].t = 0.0;
331 t->ktime[i][j].c = 0;
336 /* Decide which kernel version to use (default or legacy) based on:
338 * - non-bonded kernel selector environment variables
339 * - GPU SM version TODO ???
341 static int pick_nbnxn_kernel_version()
343 bool bLegacyKernel, bDefaultKernel, bCUDA40, bCUDA32;
347 /* legacy kernel (former k2), kept for now for backward compatibility,
348 faster than the default with CUDA 3.2/4.0 (TODO: on Kepler?). */
349 bLegacyKernel = (getenv("GMX_CUDA_NB_LEGACY") != NULL);
350 /* default kernel (former k3). */
351 bDefaultKernel = (getenv("GMX_CUDA_NB_DEFAULT") != NULL);
353 if ((unsigned)(bLegacyKernel + bDefaultKernel) > 1)
355 gmx_fatal(FARGS, "Multiple CUDA non-bonded kernels requested; to manually pick a kernel set only one \n"
356 "of the following environment variables: \n"
357 "GMX_CUDA_NB_DEFAULT, GMX_CUDA_NB_LEGACY");
360 bCUDA32 = bCUDA40 = false;
361 #if CUDA_VERSION == 3200
363 sprintf(sbuf, "3.2");
364 #elif CUDA_VERSION == 4000
366 sprintf(sbuf, "4.0");
369 /* default is default ;) */
370 kver = eNbnxnCuKDefault;
372 if (bCUDA32 || bCUDA40)
374 /* use legacy kernel unless something else is forced by an env. var */
378 "\nNOTE: CUDA %s compilation detected; with this compiler version the legacy\n"
379 " non-bonded kernels perform best. However, the default kernels were\n"
380 " selected by the GMX_CUDA_NB_DEFAULT environment variable.\n"
381 " For best performance upgrade your CUDA toolkit.",
386 kver = eNbnxnCuKLegacy;
391 /* issue not if the non-default kernel is forced by an env. var */
395 "\nNOTE: Legacy non-bonded CUDA kernels were selected by the GMX_CUDA_NB_LEGACY\n"
396 " env. var. Consider using using the default kernels which should be faster!\n");
398 kver = eNbnxnCuKLegacy;
405 void nbnxn_cuda_init(FILE *fplog,
406 nbnxn_cuda_ptr_t *p_cu_nb,
407 gmx_gpu_info_t *gpu_info, int my_gpu_index,
408 gmx_bool bLocalAndNonlocal)
413 bool bStreamSync, bNoStreamSync, bTMPIAtomics, bX86, bOldDriver;
418 if (p_cu_nb == NULL) return;
422 snew(nb->nbparam, 1);
423 snew(nb->plist[eintLocal], 1);
424 if (bLocalAndNonlocal)
426 snew(nb->plist[eintNonlocal], 1);
429 nb->bUseTwoStreams = bLocalAndNonlocal;
432 snew(nb->timings, 1);
435 pmalloc((void**)&nb->nbst.e_lj, sizeof(*nb->nbst.e_lj));
436 pmalloc((void**)&nb->nbst.e_el, sizeof(*nb->nbst.e_el));
437 pmalloc((void**)&nb->nbst.fshift, SHIFTS * sizeof(*nb->nbst.fshift));
439 init_plist(nb->plist[eintLocal]);
441 /* local/non-local GPU streams */
442 stat = cudaStreamCreate(&nb->stream[eintLocal]);
443 CU_RET_ERR(stat, "cudaStreamCreate on stream[eintLocal] failed");
444 if (nb->bUseTwoStreams)
446 init_plist(nb->plist[eintNonlocal]);
447 stat = cudaStreamCreate(&nb->stream[eintNonlocal]);
448 CU_RET_ERR(stat, "cudaStreamCreate on stream[eintNonlocal] failed");
451 /* init events for sychronization (timing disabled for performance reasons!) */
452 stat = cudaEventCreateWithFlags(&nb->nonlocal_done, cudaEventDisableTiming);
453 CU_RET_ERR(stat, "cudaEventCreate on nonlocal_done failed");
454 stat = cudaEventCreateWithFlags(&nb->misc_ops_done, cudaEventDisableTiming);
455 CU_RET_ERR(stat, "cudaEventCreate on misc_ops_one failed");
457 /* set device info, just point it to the right GPU among the detected ones */
458 nb->dev_info = &gpu_info->cuda_dev[get_gpu_device_id(gpu_info, my_gpu_index)];
460 /* On GPUs with ECC enabled, cudaStreamSynchronize shows a large overhead
461 * (which increases with shorter time/step) caused by a known CUDA driver bug.
462 * To work around the issue we'll use an (admittedly fragile) memory polling
463 * waiting to preserve performance. This requires support for atomic
464 * operations and only works on x86/x86_64.
465 * With polling wait event-timing also needs to be disabled.
467 * The overhead is greatly reduced in API v5.0 drivers and the improvement
468 $ is independent of runtime version. Hence, with API v5.0 drivers and later
469 * we won't switch to polling.
471 * NOTE: Unfortunately, this is known to fail when GPUs are shared by (t)MPI,
472 * ranks so we will also disable it in that case.
475 bStreamSync = getenv("GMX_CUDA_STREAMSYNC") != NULL;
476 bNoStreamSync = getenv("GMX_NO_CUDA_STREAMSYNC") != NULL;
481 bTMPIAtomics = false;
484 #if defined(i386) || defined(__x86_64__)
490 if (bStreamSync && bNoStreamSync)
492 gmx_fatal(FARGS, "Conflicting environment variables: both GMX_CUDA_STREAMSYNC and GMX_NO_CUDA_STREAMSYNC defined");
495 stat = cudaDriverGetVersion(&cuda_drv_ver);
496 CU_RET_ERR(stat, "cudaDriverGetVersion failed");
497 bOldDriver = (cuda_drv_ver < 5000);
499 if (nb->dev_info->prop.ECCEnabled == 1)
503 nb->bUseStreamSync = true;
505 /* only warn if polling should be used */
506 if (bOldDriver && !gpu_info->bDevShare)
509 "NOTE: Using a GPU with ECC enabled and CUDA driver API version <5.0, but\n"
510 " cudaStreamSynchronize waiting is forced by the GMX_CUDA_STREAMSYNC env. var.\n");
515 /* Can/should turn of cudaStreamSynchronize wait only if
516 * - we're on x86/x86_64
517 * - atomics are available
518 * - GPUs are not being shared
519 * - and driver is old. */
521 (bX86 && bTMPIAtomics && !gpu_info->bDevShare && bOldDriver) ?
524 if (nb->bUseStreamSync)
527 "NOTE: Using a GPU with ECC enabled and CUDA driver API version <5.0, known to\n"
528 " cause performance loss. Switching to the alternative polling GPU waiting.\n"
529 " If you encounter issues, switch back to standard GPU waiting by setting\n"
530 " the GMX_CUDA_STREAMSYNC environment variable.\n");
534 /* Tell the user that the ECC+old driver combination can be bad */
536 "NOTE: Using a GPU with ECC enabled and CUDA driver API version <5.0. A bug in this\n"
537 " driver can cause performance loss.\n"
538 " However, the polling waiting workaround can not be used because\n%s\n"
539 " Consider updating the driver or turning ECC off.",
540 (!bX86 || !bTMPIAtomics) ?
541 " atomic operations are not supported by the platform/CPU+compiler." :
542 " GPU(s) are being oversubscribed.");
543 md_print_warn(fplog, sbuf);
551 nb->bUseStreamSync = false;
554 "NOTE: Polling wait for GPU synchronization requested by GMX_NO_CUDA_STREAMSYNC\n");
558 /* no/off ECC, cudaStreamSynchronize not turned off by env. var. */
559 nb->bUseStreamSync = true;
563 /* CUDA timing disabled as event timers don't work:
564 - with multiple streams = domain-decomposition;
565 - with the polling waiting hack (without cudaStreamSynchronize);
566 - when turned off by GMX_DISABLE_CUDA_TIMING.
568 nb->bDoTime = (!nb->bUseTwoStreams && nb->bUseStreamSync &&
569 (getenv("GMX_DISABLE_CUDA_TIMING") == NULL));
573 init_timers(nb->timers, nb->bUseTwoStreams);
574 init_timings(nb->timings);
577 /* set the kernel type for the current GPU */
578 nb->kernel_ver = pick_nbnxn_kernel_version();
579 /* pick L1 cache configuration */
580 nbnxn_cuda_set_cacheconfig(nb->dev_info);
586 fprintf(debug, "Initialized CUDA data structures.\n");
590 void nbnxn_cuda_init_const(nbnxn_cuda_ptr_t cu_nb,
591 const interaction_const_t *ic,
592 const nonbonded_verlet_t *nbv)
594 init_atomdata_first(cu_nb->atdat, nbv->grp[0].nbat->ntype);
595 init_nbparam(cu_nb->nbparam, ic, nbv);
597 /* clear energy and shift force outputs */
598 nbnxn_cuda_clear_e_fshift(cu_nb);
601 void nbnxn_cuda_init_pairlist(nbnxn_cuda_ptr_t cu_nb,
602 const nbnxn_pairlist_t *h_plist,
607 bool bDoTime = cu_nb->bDoTime;
608 cudaStream_t stream = cu_nb->stream[iloc];
609 cu_plist_t *d_plist = cu_nb->plist[iloc];
611 if (d_plist->na_c < 0)
613 d_plist->na_c = h_plist->na_ci;
617 if (d_plist->na_c != h_plist->na_ci)
619 sprintf(sbuf, "In cu_init_plist: the #atoms per cell has changed (from %d to %d)",
620 d_plist->na_c, h_plist->na_ci);
627 stat = cudaEventRecord(cu_nb->timers->start_pl_h2d[iloc], stream);
628 CU_RET_ERR(stat, "cudaEventRecord failed");
631 cu_realloc_buffered((void **)&d_plist->sci, h_plist->sci, sizeof(*d_plist->sci),
632 &d_plist->nsci, &d_plist->sci_nalloc,
636 cu_realloc_buffered((void **)&d_plist->cj4, h_plist->cj4, sizeof(*d_plist->cj4),
637 &d_plist->ncj4, &d_plist->cj4_nalloc,
641 cu_realloc_buffered((void **)&d_plist->excl, h_plist->excl, sizeof(*d_plist->excl),
642 &d_plist->nexcl, &d_plist->excl_nalloc,
648 stat = cudaEventRecord(cu_nb->timers->stop_pl_h2d[iloc], stream);
649 CU_RET_ERR(stat, "cudaEventRecord failed");
652 /* need to prune the pair list during the next step */
653 d_plist->bDoPrune = true;
656 void nbnxn_cuda_upload_shiftvec(nbnxn_cuda_ptr_t cu_nb,
657 const nbnxn_atomdata_t *nbatom)
659 cu_atomdata_t *adat = cu_nb->atdat;
660 cudaStream_t ls = cu_nb->stream[eintLocal];
662 /* only if we have a dynamic box */
663 if (nbatom->bDynamicBox || !adat->bShiftVecUploaded)
665 cu_copy_H2D_async(adat->shift_vec, nbatom->shift_vec,
666 SHIFTS * sizeof(*adat->shift_vec), ls);
667 adat->bShiftVecUploaded = true;
671 /*! Clears the first natoms_clear elements of the GPU nonbonded force output array. */
672 static void nbnxn_cuda_clear_f(nbnxn_cuda_ptr_t cu_nb, int natoms_clear)
675 cu_atomdata_t *adat = cu_nb->atdat;
676 cudaStream_t ls = cu_nb->stream[eintLocal];
678 stat = cudaMemsetAsync(adat->f, 0, natoms_clear * sizeof(*adat->f), ls);
679 CU_RET_ERR(stat, "cudaMemsetAsync on f falied");
682 /*! Clears nonbonded shift force output array and energy outputs on the GPU. */
683 static void nbnxn_cuda_clear_e_fshift(nbnxn_cuda_ptr_t cu_nb)
686 cu_atomdata_t *adat = cu_nb->atdat;
687 cudaStream_t ls = cu_nb->stream[eintLocal];
689 stat = cudaMemsetAsync(adat->fshift, 0, SHIFTS * sizeof(*adat->fshift), ls);
690 CU_RET_ERR(stat, "cudaMemsetAsync on fshift falied");
691 stat = cudaMemsetAsync(adat->e_lj, 0, sizeof(*adat->e_lj), ls);
692 CU_RET_ERR(stat, "cudaMemsetAsync on e_lj falied");
693 stat = cudaMemsetAsync(adat->e_el, 0, sizeof(*adat->e_el), ls);
694 CU_RET_ERR(stat, "cudaMemsetAsync on e_el falied");
697 void nbnxn_cuda_clear_outputs(nbnxn_cuda_ptr_t cu_nb, int flags)
699 nbnxn_cuda_clear_f(cu_nb, cu_nb->atdat->natoms);
700 /* clear shift force array and energies if the outputs were
701 used in the current step */
702 if (flags & GMX_FORCE_VIRIAL)
704 nbnxn_cuda_clear_e_fshift(cu_nb);
708 void nbnxn_cuda_init_atomdata(nbnxn_cuda_ptr_t cu_nb,
709 const nbnxn_atomdata_t *nbat)
714 bool bDoTime = cu_nb->bDoTime;
715 cu_timers_t *timers = cu_nb->timers;
716 cu_atomdata_t *d_atdat = cu_nb->atdat;
717 cudaStream_t ls = cu_nb->stream[eintLocal];
719 natoms = nbat->natoms;
724 /* time async copy */
725 stat = cudaEventRecord(timers->start_atdat, ls);
726 CU_RET_ERR(stat, "cudaEventRecord failed");
729 /* need to reallocate if we have to copy more atoms than the amount of space
730 available and only allocate if we haven't initialized yet, i.e d_atdat->natoms == -1 */
731 if (natoms > d_atdat->nalloc)
733 nalloc = over_alloc_small(natoms);
735 /* free up first if the arrays have already been initialized */
736 if (d_atdat->nalloc != -1)
738 cu_free_buffered(d_atdat->f, &d_atdat->natoms, &d_atdat->nalloc);
739 cu_free_buffered(d_atdat->xq);
740 cu_free_buffered(d_atdat->atom_types);
743 stat = cudaMalloc((void **)&d_atdat->f, nalloc*sizeof(*d_atdat->f));
744 CU_RET_ERR(stat, "cudaMalloc failed on d_atdat->f");
745 stat = cudaMalloc((void **)&d_atdat->xq, nalloc*sizeof(*d_atdat->xq));
746 CU_RET_ERR(stat, "cudaMalloc failed on d_atdat->xq");
748 stat = cudaMalloc((void **)&d_atdat->atom_types, nalloc*sizeof(*d_atdat->atom_types));
749 CU_RET_ERR(stat, "cudaMalloc failed on d_atdat->atom_types");
751 d_atdat->nalloc = nalloc;
755 d_atdat->natoms = natoms;
756 d_atdat->natoms_local = nbat->natoms_local;
758 /* need to clear GPU f output if realloc happened */
761 nbnxn_cuda_clear_f(cu_nb, nalloc);
764 cu_copy_H2D_async(d_atdat->atom_types, nbat->type,
765 natoms*sizeof(*d_atdat->atom_types), ls);
769 stat = cudaEventRecord(timers->stop_atdat, ls);
770 CU_RET_ERR(stat, "cudaEventRecord failed");
774 void nbnxn_cuda_free(FILE *fplog, nbnxn_cuda_ptr_t cu_nb)
777 cu_atomdata_t *atdat;
778 cu_nbparam_t *nbparam;
779 cu_plist_t *plist, *plist_nl;
782 if (cu_nb == NULL) return;
784 atdat = cu_nb->atdat;
785 nbparam = cu_nb->nbparam;
786 plist = cu_nb->plist[eintLocal];
787 plist_nl = cu_nb->plist[eintNonlocal];
788 timers = cu_nb->timers;
790 if (nbparam->eeltype == eelCuEWALD || nbparam->eeltype == eelCuEWALD_TWIN)
792 stat = cudaUnbindTexture(nbnxn_cuda_get_coulomb_tab_texref());
793 CU_RET_ERR(stat, "cudaUnbindTexture on coulomb_tab failed");
794 cu_free_buffered(nbparam->coulomb_tab, &nbparam->coulomb_tab_size);
797 stat = cudaEventDestroy(cu_nb->nonlocal_done);
798 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->nonlocal_done");
799 stat = cudaEventDestroy(cu_nb->misc_ops_done);
800 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->misc_ops_done");
804 stat = cudaEventDestroy(timers->start_atdat);
805 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_atdat");
806 stat = cudaEventDestroy(timers->stop_atdat);
807 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_atdat");
809 /* The non-local counters/stream (second in the array) are needed only with DD. */
810 for (int i = 0; i <= (cu_nb->bUseTwoStreams ? 1 : 0); i++)
812 stat = cudaEventDestroy(timers->start_nb_k[i]);
813 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_nb_k");
814 stat = cudaEventDestroy(timers->stop_nb_k[i]);
815 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_nb_k");
817 stat = cudaEventDestroy(timers->start_pl_h2d[i]);
818 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_pl_h2d");
819 stat = cudaEventDestroy(timers->stop_pl_h2d[i]);
820 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_pl_h2d");
822 stat = cudaStreamDestroy(cu_nb->stream[i]);
823 CU_RET_ERR(stat, "cudaStreamDestroy failed on stream");
825 stat = cudaEventDestroy(timers->start_nb_h2d[i]);
826 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_nb_h2d");
827 stat = cudaEventDestroy(timers->stop_nb_h2d[i]);
828 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_nb_h2d");
830 stat = cudaEventDestroy(timers->start_nb_d2h[i]);
831 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_nb_d2h");
832 stat = cudaEventDestroy(timers->stop_nb_d2h[i]);
833 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_nb_d2h");
837 stat = cudaUnbindTexture(nbnxn_cuda_get_nbfp_texref());
838 CU_RET_ERR(stat, "cudaUnbindTexture on coulomb_tab failed");
839 cu_free_buffered(nbparam->nbfp);
841 stat = cudaFree(atdat->shift_vec);
842 CU_RET_ERR(stat, "cudaFree failed on atdat->shift_vec");
843 stat = cudaFree(atdat->fshift);
844 CU_RET_ERR(stat, "cudaFree failed on atdat->fshift");
846 stat = cudaFree(atdat->e_lj);
847 CU_RET_ERR(stat, "cudaFree failed on atdat->e_lj");
848 stat = cudaFree(atdat->e_el);
849 CU_RET_ERR(stat, "cudaFree failed on atdat->e_el");
851 cu_free_buffered(atdat->f, &atdat->natoms, &atdat->nalloc);
852 cu_free_buffered(atdat->xq);
853 cu_free_buffered(atdat->atom_types, &atdat->ntypes);
855 cu_free_buffered(plist->sci, &plist->nsci, &plist->sci_nalloc);
856 cu_free_buffered(plist->cj4, &plist->ncj4, &plist->cj4_nalloc);
857 cu_free_buffered(plist->excl, &plist->nexcl, &plist->excl_nalloc);
858 if (cu_nb->bUseTwoStreams)
860 cu_free_buffered(plist_nl->sci, &plist_nl->nsci, &plist_nl->sci_nalloc);
861 cu_free_buffered(plist_nl->cj4, &plist_nl->ncj4, &plist_nl->cj4_nalloc);
862 cu_free_buffered(plist_nl->excl, &plist_nl->nexcl, &plist->excl_nalloc);
867 fprintf(debug, "Cleaned up CUDA data structures.\n");
871 void cu_synchstream_atdat(nbnxn_cuda_ptr_t cu_nb, int iloc)
874 cudaStream_t stream = cu_nb->stream[iloc];
876 stat = cudaStreamWaitEvent(stream, cu_nb->timers->stop_atdat, 0);
877 CU_RET_ERR(stat, "cudaStreamWaitEvent failed");
880 wallclock_gpu_t * nbnxn_cuda_get_timings(nbnxn_cuda_ptr_t cu_nb)
882 return (cu_nb != NULL && cu_nb->bDoTime) ? cu_nb->timings : NULL;
885 void nbnxn_cuda_reset_timings(nbnxn_cuda_ptr_t cu_nb)
889 init_timings(cu_nb->timings);
893 int nbnxn_cuda_min_ci_balanced(nbnxn_cuda_ptr_t cu_nb)
895 return cu_nb != NULL ?
896 gpu_min_ci_balanced_factor*cu_nb->dev_info->prop.multiProcessorCount : 0;