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.
48 #include "gmx_fatal.h"
52 #include "types/nb_verlet.h"
53 #include "types/interaction_const.h"
54 #include "types/force_flags.h"
55 #include "../nbnxn_consts.h"
56 #include "gmx_detect_hardware.h"
58 #include "nbnxn_cuda_types.h"
59 #include "../../gmxlib/cuda_tools/cudautils.cuh"
60 #include "nbnxn_cuda_data_mgmt.h"
61 #include "pmalloc_cuda.h"
62 #include "gpu_utils.h"
64 static bool bUseCudaEventBlockingSync = false; /* makes the CPU thread block */
66 /* This is a heuristically determined parameter for the Fermi architecture for
67 * the minimum size of ci lists by multiplying this constant with the # of
68 * multiprocessors on the current device.
70 static unsigned int gpu_min_ci_balanced_factor = 40;
72 /* Functions from nbnxn_cuda.cu */
73 extern void nbnxn_cuda_set_cacheconfig(cuda_dev_info_t *devinfo);
74 extern const struct texture<float, 1, cudaReadModeElementType>& nbnxn_cuda_get_nbfp_texref();
75 extern const struct texture<float, 1, cudaReadModeElementType>& nbnxn_cuda_get_coulomb_tab_texref();
77 /* We should actually be using md_print_warn in md_logging.c,
78 * but we can't include mpi.h in CUDA code.
80 static void md_print_warn(FILE *fplog,
87 /* We should only print to stderr on the master node,
88 * in most cases fplog is only set on the master node, so this works.
91 fprintf(stderr, "\n");
92 vfprintf(stderr, fmt, ap);
93 fprintf(stderr, "\n");
98 vfprintf(fplog, fmt, ap);
106 static void nbnxn_cuda_clear_e_fshift(nbnxn_cuda_ptr_t cu_nb);
109 /*! Tabulates the Ewald Coulomb force and initializes the size/scale
110 and the table GPU array. If called with an already allocated table,
111 it just re-uploads the table.
113 static void init_ewald_coulomb_force_table(cu_nbparam_t *nbp,
114 const cuda_dev_info_t *dev_info)
116 float *ftmp, *coul_tab;
121 tabsize = GPU_EWALD_COULOMB_FORCE_TABLE_SIZE;
122 /* Subtract 2 iso 1 to avoid access out of range due to rounding */
123 tabscale = (tabsize - 2) / sqrt(nbp->rcoulomb_sq);
125 pmalloc((void**)&ftmp, tabsize*sizeof(*ftmp));
127 table_spline3_fill_ewald_lr(ftmp, NULL, NULL, tabsize,
128 1/tabscale, nbp->ewald_beta);
130 /* If the table pointer == NULL the table is generated the first time =>
131 the array pointer will be saved to nbparam and the texture is bound.
133 coul_tab = nbp->coulomb_tab;
134 if (coul_tab == NULL)
136 stat = cudaMalloc((void **)&coul_tab, tabsize*sizeof(*coul_tab));
137 CU_RET_ERR(stat, "cudaMalloc failed on coul_tab");
139 nbp->coulomb_tab = coul_tab;
141 #ifdef TEXOBJ_SUPPORTED
142 /* Only device CC >= 3.0 (Kepler and later) support texture objects */
143 if (dev_info->prop.major >= 3)
146 memset(&rd, 0, sizeof(rd));
147 rd.resType = cudaResourceTypeLinear;
148 rd.res.linear.devPtr = nbp->coulomb_tab;
149 rd.res.linear.desc.f = cudaChannelFormatKindFloat;
150 rd.res.linear.desc.x = 32;
151 rd.res.linear.sizeInBytes = tabsize*sizeof(*coul_tab);
154 memset(&td, 0, sizeof(td));
155 td.readMode = cudaReadModeElementType;
156 stat = cudaCreateTextureObject(&nbp->coulomb_tab_texobj, &rd, &td, NULL);
157 CU_RET_ERR(stat, "cudaCreateTextureObject on coulomb_tab_texobj failed");
162 cudaChannelFormatDesc cd = cudaCreateChannelDesc<float>();
163 stat = cudaBindTexture(NULL, &nbnxn_cuda_get_coulomb_tab_texref(),
164 coul_tab, &cd, tabsize*sizeof(*coul_tab));
165 CU_RET_ERR(stat, "cudaBindTexture on coulomb_tab_texref failed");
169 cu_copy_H2D(coul_tab, ftmp, tabsize*sizeof(*coul_tab));
171 nbp->coulomb_tab_size = tabsize;
172 nbp->coulomb_tab_scale = tabscale;
178 /*! Initializes the atomdata structure first time, it only gets filled at
180 static void init_atomdata_first(cu_atomdata_t *ad, int ntypes)
185 stat = cudaMalloc((void**)&ad->shift_vec, SHIFTS*sizeof(*ad->shift_vec));
186 CU_RET_ERR(stat, "cudaMalloc failed on ad->shift_vec");
187 ad->bShiftVecUploaded = false;
189 stat = cudaMalloc((void**)&ad->fshift, SHIFTS*sizeof(*ad->fshift));
190 CU_RET_ERR(stat, "cudaMalloc failed on ad->fshift");
192 stat = cudaMalloc((void**)&ad->e_lj, sizeof(*ad->e_lj));
193 CU_RET_ERR(stat, "cudaMalloc failed on ad->e_lj");
194 stat = cudaMalloc((void**)&ad->e_el, sizeof(*ad->e_el));
195 CU_RET_ERR(stat, "cudaMalloc failed on ad->e_el");
197 /* initialize to NULL poiters to data that is not allocated here and will
198 need reallocation in nbnxn_cuda_init_atomdata */
202 /* size -1 indicates that the respective array hasn't been initialized yet */
207 /*! Selects the Ewald kernel type, analytical on SM 3.0 and later, tabulated on
208 earlier GPUs, single or twin cut-off. */
209 static int pick_ewald_kernel_type(bool bTwinCut,
210 const cuda_dev_info_t *dev_info)
212 bool bUseAnalyticalEwald, bForceAnalyticalEwald, bForceTabulatedEwald;
215 /* Benchmarking/development environment variables to force the use of
216 analytical or tabulated Ewald kernel. */
217 bForceAnalyticalEwald = (getenv("GMX_CUDA_NB_ANA_EWALD") != NULL);
218 bForceTabulatedEwald = (getenv("GMX_CUDA_NB_TAB_EWALD") != NULL);
220 if (bForceAnalyticalEwald && bForceTabulatedEwald)
222 gmx_incons("Both analytical and tabulated Ewald CUDA non-bonded kernels "
223 "requested through environment variables.");
226 /* By default, on SM 3.0 and later use analytical Ewald, on earlier tabulated. */
227 if ((dev_info->prop.major >= 3 || bForceAnalyticalEwald) && !bForceTabulatedEwald)
229 bUseAnalyticalEwald = true;
233 fprintf(debug, "Using analytical Ewald CUDA kernels\n");
238 bUseAnalyticalEwald = false;
242 fprintf(debug, "Using tabulated Ewald CUDA kernels\n");
246 /* Use twin cut-off kernels if requested by bTwinCut or the env. var.
247 forces it (use it for debugging/benchmarking only). */
248 if (!bTwinCut && (getenv("GMX_CUDA_NB_EWALD_TWINCUT") == NULL))
250 kernel_type = bUseAnalyticalEwald ? eelCuEWALD_ANA : eelCuEWALD_TAB;
254 kernel_type = bUseAnalyticalEwald ? eelCuEWALD_ANA_TWIN : eelCuEWALD_TAB_TWIN;
261 /*! Initializes the nonbonded parameter data structure. */
262 static void init_nbparam(cu_nbparam_t *nbp,
263 const interaction_const_t *ic,
264 const nbnxn_atomdata_t *nbat,
265 const cuda_dev_info_t *dev_info)
270 ntypes = nbat->ntype;
272 nbp->ewald_beta = ic->ewaldcoeff;
273 nbp->sh_ewald = ic->sh_ewald;
274 nbp->epsfac = ic->epsfac;
275 nbp->two_k_rf = 2.0 * ic->k_rf;
276 nbp->c_rf = ic->c_rf;
277 nbp->rvdw_sq = ic->rvdw * ic->rvdw;
278 nbp->rcoulomb_sq= ic->rcoulomb * ic->rcoulomb;
279 nbp->rlist_sq = ic->rlist * ic->rlist;
280 nbp->sh_invrc6 = ic->sh_invrc6;
282 if (ic->eeltype == eelCUT)
284 nbp->eeltype = eelCuCUT;
286 else if (EEL_RF(ic->eeltype))
288 nbp->eeltype = eelCuRF;
290 else if ((EEL_PME(ic->eeltype) || ic->eeltype==eelEWALD))
292 /* Initially rcoulomb == rvdw, so it's surely not twin cut-off. */
293 nbp->eeltype = pick_ewald_kernel_type(false, dev_info);
297 /* Shouldn't happen, as this is checked when choosing Verlet-scheme */
298 gmx_incons("The requested electrostatics type is not implemented in the CUDA GPU accelerated kernels!");
301 /* generate table for PME */
302 nbp->coulomb_tab = NULL;
303 if (nbp->eeltype == eelCuEWALD_TAB || nbp->eeltype == eelCuEWALD_TAB_TWIN)
305 init_ewald_coulomb_force_table(nbp, dev_info);
308 nnbfp = 2*ntypes*ntypes;
309 stat = cudaMalloc((void **)&nbp->nbfp, nnbfp*sizeof(*nbp->nbfp));
310 CU_RET_ERR(stat, "cudaMalloc failed on nbp->nbfp");
311 cu_copy_H2D(nbp->nbfp, nbat->nbfp, nnbfp*sizeof(*nbp->nbfp));
313 #ifdef TEXOBJ_SUPPORTED
314 /* Only device CC >= 3.0 (Kepler and later) support texture objects */
315 if (dev_info->prop.major >= 3)
318 memset(&rd, 0, sizeof(rd));
319 rd.resType = cudaResourceTypeLinear;
320 rd.res.linear.devPtr = nbp->nbfp;
321 rd.res.linear.desc.f = cudaChannelFormatKindFloat;
322 rd.res.linear.desc.x = 32;
323 rd.res.linear.sizeInBytes = nnbfp*sizeof(*nbp->nbfp);
326 memset(&td, 0, sizeof(td));
327 td.readMode = cudaReadModeElementType;
328 stat = cudaCreateTextureObject(&nbp->nbfp_texobj, &rd, &td, NULL);
329 CU_RET_ERR(stat, "cudaCreateTextureObject on nbfp_texobj failed");
334 cudaChannelFormatDesc cd = cudaCreateChannelDesc<float>();
335 stat = cudaBindTexture(NULL, &nbnxn_cuda_get_nbfp_texref(),
336 nbp->nbfp, &cd, nnbfp*sizeof(*nbp->nbfp));
337 CU_RET_ERR(stat, "cudaBindTexture on nbfp_texref failed");
341 /*! Re-generate the GPU Ewald force table, resets rlist, and update the
342 * electrostatic type switching to twin cut-off (or back) if needed. */
343 void nbnxn_cuda_pme_loadbal_update_param(nbnxn_cuda_ptr_t cu_nb,
344 const interaction_const_t *ic)
346 cu_nbparam_t *nbp = cu_nb->nbparam;
348 nbp->rlist_sq = ic->rlist * ic->rlist;
349 nbp->rcoulomb_sq = ic->rcoulomb * ic->rcoulomb;
350 nbp->ewald_beta = ic->ewaldcoeff;
352 nbp->eeltype = pick_ewald_kernel_type(ic->rcoulomb != ic->rvdw,
355 init_ewald_coulomb_force_table(cu_nb->nbparam, cu_nb->dev_info);
358 /*! Initializes the pair list data structure. */
359 static void init_plist(cu_plist_t *pl)
361 /* initialize to NULL pointers to data that is not allocated here and will
362 need reallocation in nbnxn_cuda_init_pairlist */
367 /* size -1 indicates that the respective array hasn't been initialized yet */
374 pl->excl_nalloc = -1;
375 pl->bDoPrune = false;
378 /*! Initializes the timer data structure. */
379 static void init_timers(cu_timers_t *t, bool bUseTwoStreams)
382 int eventflags = ( bUseCudaEventBlockingSync ? cudaEventBlockingSync: cudaEventDefault );
384 stat = cudaEventCreateWithFlags(&(t->start_atdat), eventflags);
385 CU_RET_ERR(stat, "cudaEventCreate on start_atdat failed");
386 stat = cudaEventCreateWithFlags(&(t->stop_atdat), eventflags);
387 CU_RET_ERR(stat, "cudaEventCreate on stop_atdat failed");
389 /* The non-local counters/stream (second in the array) are needed only with DD. */
390 for (int i = 0; i <= (bUseTwoStreams ? 1 : 0); i++)
392 stat = cudaEventCreateWithFlags(&(t->start_nb_k[i]), eventflags);
393 CU_RET_ERR(stat, "cudaEventCreate on start_nb_k failed");
394 stat = cudaEventCreateWithFlags(&(t->stop_nb_k[i]), eventflags);
395 CU_RET_ERR(stat, "cudaEventCreate on stop_nb_k failed");
398 stat = cudaEventCreateWithFlags(&(t->start_pl_h2d[i]), eventflags);
399 CU_RET_ERR(stat, "cudaEventCreate on start_pl_h2d failed");
400 stat = cudaEventCreateWithFlags(&(t->stop_pl_h2d[i]), eventflags);
401 CU_RET_ERR(stat, "cudaEventCreate on stop_pl_h2d failed");
403 stat = cudaEventCreateWithFlags(&(t->start_nb_h2d[i]), eventflags);
404 CU_RET_ERR(stat, "cudaEventCreate on start_nb_h2d failed");
405 stat = cudaEventCreateWithFlags(&(t->stop_nb_h2d[i]), eventflags);
406 CU_RET_ERR(stat, "cudaEventCreate on stop_nb_h2d failed");
408 stat = cudaEventCreateWithFlags(&(t->start_nb_d2h[i]), eventflags);
409 CU_RET_ERR(stat, "cudaEventCreate on start_nb_d2h failed");
410 stat = cudaEventCreateWithFlags(&(t->stop_nb_d2h[i]), eventflags);
411 CU_RET_ERR(stat, "cudaEventCreate on stop_nb_d2h failed");
415 /*! Initializes the timings data structure. */
416 static void init_timings(wallclock_gpu_t *t)
425 for (i = 0; i < 2; i++)
427 for(j = 0; j < 2; j++)
429 t->ktime[i][j].t = 0.0;
430 t->ktime[i][j].c = 0;
435 /* Decide which kernel version to use (default or legacy) based on:
436 * - CUDA version used for compilation
437 * - non-bonded kernel selector environment variables
438 * - GPU architecture version
440 static int pick_nbnxn_kernel_version(FILE *fplog,
441 cuda_dev_info_t *devinfo)
443 bool bForceLegacyKernel, bForceDefaultKernel, bCUDA40, bCUDA32;
447 /* Legacy kernel (former k2), kept for backward compatibility as it is
448 faster than the default with CUDA 3.2/4.0 on Fermi (not on Kepler). */
449 bForceLegacyKernel = (getenv("GMX_CUDA_NB_LEGACY") != NULL);
450 /* default kernel (former k3). */
451 bForceDefaultKernel = (getenv("GMX_CUDA_NB_DEFAULT") != NULL);
453 if ((unsigned)(bForceLegacyKernel + bForceDefaultKernel) > 1)
455 gmx_fatal(FARGS, "Multiple CUDA non-bonded kernels requested; to manually pick a kernel set only one \n"
456 "of the following environment variables: \n"
457 "GMX_CUDA_NB_DEFAULT, GMX_CUDA_NB_LEGACY");
460 bCUDA32 = bCUDA40 = false;
461 #if CUDA_VERSION == 3200
463 sprintf(sbuf, "3.2");
464 #elif CUDA_VERSION == 4000
466 sprintf(sbuf, "4.0");
469 /* default is default ;) */
470 kver = eNbnxnCuKDefault;
472 /* Consider switching to legacy kernels only on Fermi */
473 if (devinfo->prop.major < 3 && (bCUDA32 || bCUDA40))
475 /* use legacy kernel unless something else is forced by an env. var */
476 if (bForceDefaultKernel)
479 "NOTE: CUDA %s compilation detected; with this compiler version the legacy\n"
480 " non-bonded kernels perform best. However, the default kernels were\n"
481 " selected by the GMX_CUDA_NB_DEFAULT environment variable.\n"
482 " For best performance upgrade your CUDA toolkit.\n",
487 kver = eNbnxnCuKLegacy;
492 /* issue note if the non-default kernel is forced by an env. var */
493 if (bForceLegacyKernel)
496 "NOTE: Legacy non-bonded CUDA kernels selected by the GMX_CUDA_NB_LEGACY\n"
497 " env. var. Consider using using the default kernels which should be faster!\n");
499 kver = eNbnxnCuKLegacy;
506 void nbnxn_cuda_init(FILE *fplog,
507 nbnxn_cuda_ptr_t *p_cu_nb,
508 const gmx_gpu_info_t *gpu_info, int my_gpu_index,
509 gmx_bool bLocalAndNonlocal)
514 bool bStreamSync, bNoStreamSync, bTMPIAtomics, bX86, bOldDriver;
519 if (p_cu_nb == NULL) return;
523 snew(nb->nbparam, 1);
524 snew(nb->plist[eintLocal], 1);
525 if (bLocalAndNonlocal)
527 snew(nb->plist[eintNonlocal], 1);
530 nb->bUseTwoStreams = bLocalAndNonlocal;
533 snew(nb->timings, 1);
536 pmalloc((void**)&nb->nbst.e_lj, sizeof(*nb->nbst.e_lj));
537 pmalloc((void**)&nb->nbst.e_el, sizeof(*nb->nbst.e_el));
538 pmalloc((void**)&nb->nbst.fshift, SHIFTS * sizeof(*nb->nbst.fshift));
540 init_plist(nb->plist[eintLocal]);
542 /* local/non-local GPU streams */
543 stat = cudaStreamCreate(&nb->stream[eintLocal]);
544 CU_RET_ERR(stat, "cudaStreamCreate on stream[eintLocal] failed");
545 if (nb->bUseTwoStreams)
547 init_plist(nb->plist[eintNonlocal]);
548 stat = cudaStreamCreate(&nb->stream[eintNonlocal]);
549 CU_RET_ERR(stat, "cudaStreamCreate on stream[eintNonlocal] failed");
552 /* init events for sychronization (timing disabled for performance reasons!) */
553 stat = cudaEventCreateWithFlags(&nb->nonlocal_done, cudaEventDisableTiming);
554 CU_RET_ERR(stat, "cudaEventCreate on nonlocal_done failed");
555 stat = cudaEventCreateWithFlags(&nb->misc_ops_done, cudaEventDisableTiming);
556 CU_RET_ERR(stat, "cudaEventCreate on misc_ops_one failed");
558 /* set device info, just point it to the right GPU among the detected ones */
559 nb->dev_info = &gpu_info->cuda_dev[get_gpu_device_id(gpu_info, my_gpu_index)];
561 /* On GPUs with ECC enabled, cudaStreamSynchronize shows a large overhead
562 * (which increases with shorter time/step) caused by a known CUDA driver bug.
563 * To work around the issue we'll use an (admittedly fragile) memory polling
564 * waiting to preserve performance. This requires support for atomic
565 * operations and only works on x86/x86_64.
566 * With polling wait event-timing also needs to be disabled.
568 * The overhead is greatly reduced in API v5.0 drivers and the improvement
569 $ is independent of runtime version. Hence, with API v5.0 drivers and later
570 * we won't switch to polling.
572 * NOTE: Unfortunately, this is known to fail when GPUs are shared by (t)MPI,
573 * ranks so we will also disable it in that case.
576 bStreamSync = getenv("GMX_CUDA_STREAMSYNC") != NULL;
577 bNoStreamSync = getenv("GMX_NO_CUDA_STREAMSYNC") != NULL;
582 bTMPIAtomics = false;
585 #if defined(i386) || defined(__x86_64__)
591 if (bStreamSync && bNoStreamSync)
593 gmx_fatal(FARGS, "Conflicting environment variables: both GMX_CUDA_STREAMSYNC and GMX_NO_CUDA_STREAMSYNC defined");
596 stat = cudaDriverGetVersion(&cuda_drv_ver);
597 CU_RET_ERR(stat, "cudaDriverGetVersion failed");
599 bOldDriver = (cuda_drv_ver < 5000);
601 if ((nb->dev_info->prop.ECCEnabled == 1) && bOldDriver)
603 /* Polling wait should be used instead of cudaStreamSynchronize only if:
604 * - ECC is ON & driver is old (checked above),
605 * - we're on x86/x86_64,
606 * - atomics are available, and
607 * - GPUs are not being shared.
609 bool bShouldUsePollSync = (bX86 && bTMPIAtomics &&
610 (gmx_count_gpu_dev_shared(gpu_info) < 1));
614 nb->bUseStreamSync = true;
616 /* only warn if polling should be used */
617 if (bShouldUsePollSync)
620 "NOTE: Using a GPU with ECC enabled and CUDA driver API version <5.0, but\n"
621 " cudaStreamSynchronize waiting is forced by the GMX_CUDA_STREAMSYNC env. var.\n");
626 nb->bUseStreamSync = !bShouldUsePollSync;
628 if (bShouldUsePollSync)
631 "NOTE: Using a GPU with ECC enabled and CUDA driver API version <5.0, known to\n"
632 " cause performance loss. Switching to the alternative polling GPU wait.\n"
633 " If you encounter issues, switch back to standard GPU waiting by setting\n"
634 " the GMX_CUDA_STREAMSYNC environment variable.\n");
638 /* Tell the user that the ECC+old driver combination can be bad */
640 "NOTE: Using a GPU with ECC enabled and CUDA driver API version <5.0.\n"
641 " A known bug in this driver version can cause performance loss.\n"
642 " However, the polling wait workaround can not be used because\n%s\n"
643 " Consider updating the driver or turning ECC off.",
644 (bX86 && bTMPIAtomics) ?
645 " GPU(s) are being oversubscribed." :
646 " atomic operations are not supported by the platform/CPU+compiler.");
647 md_print_warn(fplog, sbuf);
655 nb->bUseStreamSync = false;
658 "NOTE: Polling wait for GPU synchronization requested by GMX_NO_CUDA_STREAMSYNC\n");
662 /* no/off ECC, cudaStreamSynchronize not turned off by env. var. */
663 nb->bUseStreamSync = true;
667 /* CUDA timing disabled as event timers don't work:
668 - with multiple streams = domain-decomposition;
669 - with the polling waiting hack (without cudaStreamSynchronize);
670 - when turned off by GMX_DISABLE_CUDA_TIMING.
672 nb->bDoTime = (!nb->bUseTwoStreams && nb->bUseStreamSync &&
673 (getenv("GMX_DISABLE_CUDA_TIMING") == NULL));
677 init_timers(nb->timers, nb->bUseTwoStreams);
678 init_timings(nb->timings);
681 /* set the kernel type for the current GPU */
682 nb->kernel_ver = pick_nbnxn_kernel_version(fplog, nb->dev_info);
683 /* pick L1 cache configuration */
684 nbnxn_cuda_set_cacheconfig(nb->dev_info);
690 fprintf(debug, "Initialized CUDA data structures.\n");
694 void nbnxn_cuda_init_const(nbnxn_cuda_ptr_t cu_nb,
695 const interaction_const_t *ic,
696 const nonbonded_verlet_group_t *nbv_group)
698 init_atomdata_first(cu_nb->atdat, nbv_group[0].nbat->ntype);
699 init_nbparam(cu_nb->nbparam, ic, nbv_group[0].nbat, cu_nb->dev_info);
701 /* clear energy and shift force outputs */
702 nbnxn_cuda_clear_e_fshift(cu_nb);
705 void nbnxn_cuda_init_pairlist(nbnxn_cuda_ptr_t cu_nb,
706 const nbnxn_pairlist_t *h_plist,
711 bool bDoTime = cu_nb->bDoTime;
712 cudaStream_t stream = cu_nb->stream[iloc];
713 cu_plist_t *d_plist = cu_nb->plist[iloc];
715 if (d_plist->na_c < 0)
717 d_plist->na_c = h_plist->na_ci;
721 if (d_plist->na_c != h_plist->na_ci)
723 sprintf(sbuf, "In cu_init_plist: the #atoms per cell has changed (from %d to %d)",
724 d_plist->na_c, h_plist->na_ci);
731 stat = cudaEventRecord(cu_nb->timers->start_pl_h2d[iloc], stream);
732 CU_RET_ERR(stat, "cudaEventRecord failed");
735 cu_realloc_buffered((void **)&d_plist->sci, h_plist->sci, sizeof(*d_plist->sci),
736 &d_plist->nsci, &d_plist->sci_nalloc,
740 cu_realloc_buffered((void **)&d_plist->cj4, h_plist->cj4, sizeof(*d_plist->cj4),
741 &d_plist->ncj4, &d_plist->cj4_nalloc,
745 cu_realloc_buffered((void **)&d_plist->excl, h_plist->excl, sizeof(*d_plist->excl),
746 &d_plist->nexcl, &d_plist->excl_nalloc,
752 stat = cudaEventRecord(cu_nb->timers->stop_pl_h2d[iloc], stream);
753 CU_RET_ERR(stat, "cudaEventRecord failed");
756 /* need to prune the pair list during the next step */
757 d_plist->bDoPrune = true;
760 void nbnxn_cuda_upload_shiftvec(nbnxn_cuda_ptr_t cu_nb,
761 const nbnxn_atomdata_t *nbatom)
763 cu_atomdata_t *adat = cu_nb->atdat;
764 cudaStream_t ls = cu_nb->stream[eintLocal];
766 /* only if we have a dynamic box */
767 if (nbatom->bDynamicBox || !adat->bShiftVecUploaded)
769 cu_copy_H2D_async(adat->shift_vec, nbatom->shift_vec,
770 SHIFTS * sizeof(*adat->shift_vec), ls);
771 adat->bShiftVecUploaded = true;
775 /*! Clears the first natoms_clear elements of the GPU nonbonded force output array. */
776 static void nbnxn_cuda_clear_f(nbnxn_cuda_ptr_t cu_nb, int natoms_clear)
779 cu_atomdata_t *adat = cu_nb->atdat;
780 cudaStream_t ls = cu_nb->stream[eintLocal];
782 stat = cudaMemsetAsync(adat->f, 0, natoms_clear * sizeof(*adat->f), ls);
783 CU_RET_ERR(stat, "cudaMemsetAsync on f falied");
786 /*! Clears nonbonded shift force output array and energy outputs on the GPU. */
787 static void nbnxn_cuda_clear_e_fshift(nbnxn_cuda_ptr_t cu_nb)
790 cu_atomdata_t *adat = cu_nb->atdat;
791 cudaStream_t ls = cu_nb->stream[eintLocal];
793 stat = cudaMemsetAsync(adat->fshift, 0, SHIFTS * sizeof(*adat->fshift), ls);
794 CU_RET_ERR(stat, "cudaMemsetAsync on fshift falied");
795 stat = cudaMemsetAsync(adat->e_lj, 0, sizeof(*adat->e_lj), ls);
796 CU_RET_ERR(stat, "cudaMemsetAsync on e_lj falied");
797 stat = cudaMemsetAsync(adat->e_el, 0, sizeof(*adat->e_el), ls);
798 CU_RET_ERR(stat, "cudaMemsetAsync on e_el falied");
801 void nbnxn_cuda_clear_outputs(nbnxn_cuda_ptr_t cu_nb, int flags)
803 nbnxn_cuda_clear_f(cu_nb, cu_nb->atdat->natoms);
804 /* clear shift force array and energies if the outputs were
805 used in the current step */
806 if (flags & GMX_FORCE_VIRIAL)
808 nbnxn_cuda_clear_e_fshift(cu_nb);
812 void nbnxn_cuda_init_atomdata(nbnxn_cuda_ptr_t cu_nb,
813 const nbnxn_atomdata_t *nbat)
818 bool bDoTime = cu_nb->bDoTime;
819 cu_timers_t *timers = cu_nb->timers;
820 cu_atomdata_t *d_atdat = cu_nb->atdat;
821 cudaStream_t ls = cu_nb->stream[eintLocal];
823 natoms = nbat->natoms;
828 /* time async copy */
829 stat = cudaEventRecord(timers->start_atdat, ls);
830 CU_RET_ERR(stat, "cudaEventRecord failed");
833 /* need to reallocate if we have to copy more atoms than the amount of space
834 available and only allocate if we haven't initialized yet, i.e d_atdat->natoms == -1 */
835 if (natoms > d_atdat->nalloc)
837 nalloc = over_alloc_small(natoms);
839 /* free up first if the arrays have already been initialized */
840 if (d_atdat->nalloc != -1)
842 cu_free_buffered(d_atdat->f, &d_atdat->natoms, &d_atdat->nalloc);
843 cu_free_buffered(d_atdat->xq);
844 cu_free_buffered(d_atdat->atom_types);
847 stat = cudaMalloc((void **)&d_atdat->f, nalloc*sizeof(*d_atdat->f));
848 CU_RET_ERR(stat, "cudaMalloc failed on d_atdat->f");
849 stat = cudaMalloc((void **)&d_atdat->xq, nalloc*sizeof(*d_atdat->xq));
850 CU_RET_ERR(stat, "cudaMalloc failed on d_atdat->xq");
852 stat = cudaMalloc((void **)&d_atdat->atom_types, nalloc*sizeof(*d_atdat->atom_types));
853 CU_RET_ERR(stat, "cudaMalloc failed on d_atdat->atom_types");
855 d_atdat->nalloc = nalloc;
859 d_atdat->natoms = natoms;
860 d_atdat->natoms_local = nbat->natoms_local;
862 /* need to clear GPU f output if realloc happened */
865 nbnxn_cuda_clear_f(cu_nb, nalloc);
868 cu_copy_H2D_async(d_atdat->atom_types, nbat->type,
869 natoms*sizeof(*d_atdat->atom_types), ls);
873 stat = cudaEventRecord(timers->stop_atdat, ls);
874 CU_RET_ERR(stat, "cudaEventRecord failed");
878 void nbnxn_cuda_free(FILE *fplog, nbnxn_cuda_ptr_t cu_nb)
881 cu_atomdata_t *atdat;
882 cu_nbparam_t *nbparam;
883 cu_plist_t *plist, *plist_nl;
886 if (cu_nb == NULL) return;
888 atdat = cu_nb->atdat;
889 nbparam = cu_nb->nbparam;
890 plist = cu_nb->plist[eintLocal];
891 plist_nl = cu_nb->plist[eintNonlocal];
892 timers = cu_nb->timers;
894 if (nbparam->eeltype == eelCuEWALD_TAB || nbparam->eeltype == eelCuEWALD_TAB_TWIN)
897 #ifdef TEXOBJ_SUPPORTED
898 /* Only device CC >= 3.0 (Kepler and later) support texture objects */
899 if (cu_nb->dev_info->prop.major >= 3)
901 stat = cudaDestroyTextureObject(nbparam->coulomb_tab_texobj);
902 CU_RET_ERR(stat, "cudaDestroyTextureObject on coulomb_tab_texobj failed");
907 stat = cudaUnbindTexture(nbnxn_cuda_get_coulomb_tab_texref());
908 CU_RET_ERR(stat, "cudaUnbindTexture on coulomb_tab_texref failed");
910 cu_free_buffered(nbparam->coulomb_tab, &nbparam->coulomb_tab_size);
913 stat = cudaEventDestroy(cu_nb->nonlocal_done);
914 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->nonlocal_done");
915 stat = cudaEventDestroy(cu_nb->misc_ops_done);
916 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->misc_ops_done");
920 stat = cudaEventDestroy(timers->start_atdat);
921 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_atdat");
922 stat = cudaEventDestroy(timers->stop_atdat);
923 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_atdat");
925 /* The non-local counters/stream (second in the array) are needed only with DD. */
926 for (int i = 0; i <= (cu_nb->bUseTwoStreams ? 1 : 0); i++)
928 stat = cudaEventDestroy(timers->start_nb_k[i]);
929 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_nb_k");
930 stat = cudaEventDestroy(timers->stop_nb_k[i]);
931 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_nb_k");
933 stat = cudaEventDestroy(timers->start_pl_h2d[i]);
934 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_pl_h2d");
935 stat = cudaEventDestroy(timers->stop_pl_h2d[i]);
936 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_pl_h2d");
938 stat = cudaStreamDestroy(cu_nb->stream[i]);
939 CU_RET_ERR(stat, "cudaStreamDestroy failed on stream");
941 stat = cudaEventDestroy(timers->start_nb_h2d[i]);
942 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_nb_h2d");
943 stat = cudaEventDestroy(timers->stop_nb_h2d[i]);
944 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_nb_h2d");
946 stat = cudaEventDestroy(timers->start_nb_d2h[i]);
947 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_nb_d2h");
948 stat = cudaEventDestroy(timers->stop_nb_d2h[i]);
949 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_nb_d2h");
953 #ifdef TEXOBJ_SUPPORTED
954 /* Only device CC >= 3.0 (Kepler and later) support texture objects */
955 if (cu_nb->dev_info->prop.major >= 3)
957 stat = cudaDestroyTextureObject(nbparam->nbfp_texobj);
958 CU_RET_ERR(stat, "cudaDestroyTextureObject on nbfp_texobj failed");
963 stat = cudaUnbindTexture(nbnxn_cuda_get_nbfp_texref());
964 CU_RET_ERR(stat, "cudaUnbindTexture on nbfp_texref failed");
966 cu_free_buffered(nbparam->nbfp);
968 stat = cudaFree(atdat->shift_vec);
969 CU_RET_ERR(stat, "cudaFree failed on atdat->shift_vec");
970 stat = cudaFree(atdat->fshift);
971 CU_RET_ERR(stat, "cudaFree failed on atdat->fshift");
973 stat = cudaFree(atdat->e_lj);
974 CU_RET_ERR(stat, "cudaFree failed on atdat->e_lj");
975 stat = cudaFree(atdat->e_el);
976 CU_RET_ERR(stat, "cudaFree failed on atdat->e_el");
978 cu_free_buffered(atdat->f, &atdat->natoms, &atdat->nalloc);
979 cu_free_buffered(atdat->xq);
980 cu_free_buffered(atdat->atom_types, &atdat->ntypes);
982 cu_free_buffered(plist->sci, &plist->nsci, &plist->sci_nalloc);
983 cu_free_buffered(plist->cj4, &plist->ncj4, &plist->cj4_nalloc);
984 cu_free_buffered(plist->excl, &plist->nexcl, &plist->excl_nalloc);
985 if (cu_nb->bUseTwoStreams)
987 cu_free_buffered(plist_nl->sci, &plist_nl->nsci, &plist_nl->sci_nalloc);
988 cu_free_buffered(plist_nl->cj4, &plist_nl->ncj4, &plist_nl->cj4_nalloc);
989 cu_free_buffered(plist_nl->excl, &plist_nl->nexcl, &plist->excl_nalloc);
995 if (cu_nb->bUseTwoStreams)
1000 sfree(cu_nb->timings);
1005 fprintf(debug, "Cleaned up CUDA data structures.\n");
1009 void cu_synchstream_atdat(nbnxn_cuda_ptr_t cu_nb, int iloc)
1012 cudaStream_t stream = cu_nb->stream[iloc];
1014 stat = cudaStreamWaitEvent(stream, cu_nb->timers->stop_atdat, 0);
1015 CU_RET_ERR(stat, "cudaStreamWaitEvent failed");
1018 wallclock_gpu_t * nbnxn_cuda_get_timings(nbnxn_cuda_ptr_t cu_nb)
1020 return (cu_nb != NULL && cu_nb->bDoTime) ? cu_nb->timings : NULL;
1023 void nbnxn_cuda_reset_timings(nbnxn_cuda_ptr_t cu_nb)
1027 init_timings(cu_nb->timings);
1031 int nbnxn_cuda_min_ci_balanced(nbnxn_cuda_ptr_t cu_nb)
1033 return cu_nb != NULL ?
1034 gpu_min_ci_balanced_factor*cu_nb->dev_info->prop.multiProcessorCount : 0;
1038 gmx_bool nbnxn_cuda_is_kernel_ewald_analytical(const nbnxn_cuda_ptr_t cu_nb)
1040 return ((cu_nb->nbparam->eeltype == eelCuEWALD_ANA) ||
1041 (cu_nb->nbparam->eeltype == eelCuEWALD_ANA_TWIN));