2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014, 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.
45 #include "gmx_fatal.h"
49 #include "types/enums.h"
50 #include "types/nb_verlet.h"
51 #include "types/interaction_const.h"
52 #include "types/force_flags.h"
53 #include "../nbnxn_consts.h"
54 #include "gmx_detect_hardware.h"
56 #include "nbnxn_cuda_types.h"
57 #include "../../gmxlib/cuda_tools/cudautils.cuh"
58 #include "nbnxn_cuda_data_mgmt.h"
59 #include "pmalloc_cuda.h"
60 #include "gpu_utils.h"
62 #include "gromacs/utility/common.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_nbfp_comb_texref();
76 extern const struct texture<float, 1, cudaReadModeElementType> &nbnxn_cuda_get_coulomb_tab_texref();
78 /* We should actually be using md_print_warn in md_logging.c,
79 * but we can't include mpi.h in CUDA code.
81 static void md_print_warn(FILE *fplog,
88 /* We should only print to stderr on the master node,
89 * in most cases fplog is only set on the master node, so this works.
92 fprintf(stderr, "\n");
93 vfprintf(stderr, fmt, ap);
94 fprintf(stderr, "\n");
99 vfprintf(fplog, fmt, ap);
100 fprintf(fplog, "\n");
107 static void nbnxn_cuda_clear_e_fshift(nbnxn_cuda_ptr_t cu_nb);
110 /*! Tabulates the Ewald Coulomb force and initializes the size/scale
111 and the table GPU array. If called with an already allocated table,
112 it just re-uploads the table.
114 static void init_ewald_coulomb_force_table(cu_nbparam_t *nbp,
115 const cuda_dev_info_t *dev_info)
117 float *ftmp, *coul_tab;
122 tabsize = GPU_EWALD_COULOMB_FORCE_TABLE_SIZE;
123 /* Subtract 2 iso 1 to avoid access out of range due to rounding */
124 tabscale = (tabsize - 2) / sqrt(nbp->rcoulomb_sq);
126 pmalloc((void**)&ftmp, tabsize*sizeof(*ftmp));
128 table_spline3_fill_ewald_lr(ftmp, NULL, NULL, tabsize,
129 1/tabscale, nbp->ewald_beta);
131 /* If the table pointer == NULL the table is generated the first time =>
132 the array pointer will be saved to nbparam and the texture is bound.
134 coul_tab = nbp->coulomb_tab;
135 if (coul_tab == NULL)
137 stat = cudaMalloc((void **)&coul_tab, tabsize*sizeof(*coul_tab));
138 CU_RET_ERR(stat, "cudaMalloc failed on coul_tab");
140 nbp->coulomb_tab = coul_tab;
142 #ifdef TEXOBJ_SUPPORTED
143 /* Only device CC >= 3.0 (Kepler and later) support texture objects */
144 if (dev_info->prop.major >= 3)
147 memset(&rd, 0, sizeof(rd));
148 rd.resType = cudaResourceTypeLinear;
149 rd.res.linear.devPtr = nbp->coulomb_tab;
150 rd.res.linear.desc.f = cudaChannelFormatKindFloat;
151 rd.res.linear.desc.x = 32;
152 rd.res.linear.sizeInBytes = tabsize*sizeof(*coul_tab);
155 memset(&td, 0, sizeof(td));
156 td.readMode = cudaReadModeElementType;
157 stat = cudaCreateTextureObject(&nbp->coulomb_tab_texobj, &rd, &td, NULL);
158 CU_RET_ERR(stat, "cudaCreateTextureObject on coulomb_tab_texobj failed");
163 GMX_UNUSED_VALUE(dev_info);
164 cudaChannelFormatDesc cd = cudaCreateChannelDesc<float>();
165 stat = cudaBindTexture(NULL, &nbnxn_cuda_get_coulomb_tab_texref(),
166 coul_tab, &cd, tabsize*sizeof(*coul_tab));
167 CU_RET_ERR(stat, "cudaBindTexture on coulomb_tab_texref failed");
171 cu_copy_H2D(coul_tab, ftmp, tabsize*sizeof(*coul_tab));
173 nbp->coulomb_tab_size = tabsize;
174 nbp->coulomb_tab_scale = tabscale;
180 /*! Initializes the atomdata structure first time, it only gets filled at
182 static void init_atomdata_first(cu_atomdata_t *ad, int ntypes)
187 stat = cudaMalloc((void**)&ad->shift_vec, SHIFTS*sizeof(*ad->shift_vec));
188 CU_RET_ERR(stat, "cudaMalloc failed on ad->shift_vec");
189 ad->bShiftVecUploaded = false;
191 stat = cudaMalloc((void**)&ad->fshift, SHIFTS*sizeof(*ad->fshift));
192 CU_RET_ERR(stat, "cudaMalloc failed on ad->fshift");
194 stat = cudaMalloc((void**)&ad->e_lj, sizeof(*ad->e_lj));
195 CU_RET_ERR(stat, "cudaMalloc failed on ad->e_lj");
196 stat = cudaMalloc((void**)&ad->e_el, sizeof(*ad->e_el));
197 CU_RET_ERR(stat, "cudaMalloc failed on ad->e_el");
199 /* initialize to NULL poiters to data that is not allocated here and will
200 need reallocation in nbnxn_cuda_init_atomdata */
204 /* size -1 indicates that the respective array hasn't been initialized yet */
209 /*! Selects the Ewald kernel type, analytical on SM 3.0 and later, tabulated on
210 earlier GPUs, single or twin cut-off. */
211 static int pick_ewald_kernel_type(bool bTwinCut,
212 const cuda_dev_info_t *dev_info)
214 bool bUseAnalyticalEwald, bForceAnalyticalEwald, bForceTabulatedEwald;
217 /* Benchmarking/development environment variables to force the use of
218 analytical or tabulated Ewald kernel. */
219 bForceAnalyticalEwald = (getenv("GMX_CUDA_NB_ANA_EWALD") != NULL);
220 bForceTabulatedEwald = (getenv("GMX_CUDA_NB_TAB_EWALD") != NULL);
222 if (bForceAnalyticalEwald && bForceTabulatedEwald)
224 gmx_incons("Both analytical and tabulated Ewald CUDA non-bonded kernels "
225 "requested through environment variables.");
228 /* By default, on SM 3.0 and later use analytical Ewald, on earlier tabulated. */
229 if ((dev_info->prop.major >= 3 || bForceAnalyticalEwald) && !bForceTabulatedEwald)
231 bUseAnalyticalEwald = true;
235 fprintf(debug, "Using analytical Ewald CUDA kernels\n");
240 bUseAnalyticalEwald = false;
244 fprintf(debug, "Using tabulated Ewald CUDA kernels\n");
248 /* Use twin cut-off kernels if requested by bTwinCut or the env. var.
249 forces it (use it for debugging/benchmarking only). */
250 if (!bTwinCut && (getenv("GMX_CUDA_NB_EWALD_TWINCUT") == NULL))
252 kernel_type = bUseAnalyticalEwald ? eelCuEWALD_ANA : eelCuEWALD_TAB;
256 kernel_type = bUseAnalyticalEwald ? eelCuEWALD_ANA_TWIN : eelCuEWALD_TAB_TWIN;
262 /*! Copies all parameters related to the cut-off from ic to nbp */
263 static void set_cutoff_parameters(cu_nbparam_t *nbp,
264 const interaction_const_t *ic)
266 nbp->ewald_beta = ic->ewaldcoeff_q;
267 nbp->sh_ewald = ic->sh_ewald;
268 nbp->epsfac = ic->epsfac;
269 nbp->two_k_rf = 2.0 * ic->k_rf;
270 nbp->c_rf = ic->c_rf;
271 nbp->rvdw_sq = ic->rvdw * ic->rvdw;
272 nbp->rcoulomb_sq = ic->rcoulomb * ic->rcoulomb;
273 nbp->rlist_sq = ic->rlist * ic->rlist;
275 nbp->sh_lj_ewald = ic->sh_lj_ewald;
276 nbp->ewaldcoeff_lj = ic->ewaldcoeff_lj;
278 nbp->rvdw_switch = ic->rvdw_switch;
279 nbp->dispersion_shift = ic->dispersion_shift;
280 nbp->repulsion_shift = ic->repulsion_shift;
281 nbp->vdw_switch = ic->vdw_switch;
284 /*! Initializes the nonbonded parameter data structure. */
285 static void init_nbparam(cu_nbparam_t *nbp,
286 const interaction_const_t *ic,
287 const nbnxn_atomdata_t *nbat,
288 const cuda_dev_info_t *dev_info)
291 int ntypes, nnbfp, nnbfp_comb;
293 ntypes = nbat->ntype;
295 set_cutoff_parameters(nbp, ic);
297 if (ic->vdwtype == evdwCUT)
299 switch (ic->vdw_modifier)
302 case eintmodPOTSHIFT:
303 nbp->vdwtype = evdwCuCUT;
305 case eintmodFORCESWITCH:
306 nbp->vdwtype = evdwCuFSWITCH;
308 case eintmodPOTSWITCH:
309 nbp->vdwtype = evdwCuPSWITCH;
312 gmx_incons("The requested VdW interaction modifier is not implemented in the CUDA GPU accelerated kernels!");
316 else if (ic->vdwtype == evdwPME)
318 if (ic->ljpme_comb_rule == ljcrGEOM)
320 assert(nbat->comb_rule == ljcrGEOM);
321 nbp->vdwtype = evdwCuEWALDGEOM;
325 assert(nbat->comb_rule == ljcrLB);
326 nbp->vdwtype = evdwCuEWALDLB;
331 gmx_incons("The requested VdW type is not implemented in the CUDA GPU accelerated kernels!");
334 if (ic->eeltype == eelCUT)
336 nbp->eeltype = eelCuCUT;
338 else if (EEL_RF(ic->eeltype))
340 nbp->eeltype = eelCuRF;
342 else if ((EEL_PME(ic->eeltype) || ic->eeltype == eelEWALD))
344 /* Initially rcoulomb == rvdw, so it's surely not twin cut-off. */
345 nbp->eeltype = pick_ewald_kernel_type(false, dev_info);
349 /* Shouldn't happen, as this is checked when choosing Verlet-scheme */
350 gmx_incons("The requested electrostatics type is not implemented in the CUDA GPU accelerated kernels!");
353 /* generate table for PME */
354 nbp->coulomb_tab = NULL;
355 if (nbp->eeltype == eelCuEWALD_TAB || nbp->eeltype == eelCuEWALD_TAB_TWIN)
357 init_ewald_coulomb_force_table(nbp, dev_info);
360 nnbfp = 2*ntypes*ntypes;
361 nnbfp_comb = 2*ntypes;
363 stat = cudaMalloc((void **)&nbp->nbfp, nnbfp*sizeof(*nbp->nbfp));
364 CU_RET_ERR(stat, "cudaMalloc failed on nbp->nbfp");
365 cu_copy_H2D(nbp->nbfp, nbat->nbfp, nnbfp*sizeof(*nbp->nbfp));
368 if (ic->vdwtype == evdwPME)
370 stat = cudaMalloc((void **)&nbp->nbfp_comb, nnbfp_comb*sizeof(*nbp->nbfp_comb));
371 CU_RET_ERR(stat, "cudaMalloc failed on nbp->nbfp_comb");
372 cu_copy_H2D(nbp->nbfp_comb, nbat->nbfp_comb, nnbfp_comb*sizeof(*nbp->nbfp_comb));
375 #ifdef TEXOBJ_SUPPORTED
376 /* Only device CC >= 3.0 (Kepler and later) support texture objects */
377 if (dev_info->prop.major >= 3)
382 memset(&rd, 0, sizeof(rd));
383 rd.resType = cudaResourceTypeLinear;
384 rd.res.linear.devPtr = nbp->nbfp;
385 rd.res.linear.desc.f = cudaChannelFormatKindFloat;
386 rd.res.linear.desc.x = 32;
387 rd.res.linear.sizeInBytes = nnbfp*sizeof(*nbp->nbfp);
389 memset(&td, 0, sizeof(td));
390 td.readMode = cudaReadModeElementType;
391 stat = cudaCreateTextureObject(&nbp->nbfp_texobj, &rd, &td, NULL);
392 CU_RET_ERR(stat, "cudaCreateTextureObject on nbfp_texobj failed");
394 if (ic->vdwtype == evdwPME)
396 memset(&rd, 0, sizeof(rd));
397 rd.resType = cudaResourceTypeLinear;
398 rd.res.linear.devPtr = nbp->nbfp_comb;
399 rd.res.linear.desc.f = cudaChannelFormatKindFloat;
400 rd.res.linear.desc.x = 32;
401 rd.res.linear.sizeInBytes = nnbfp_comb*sizeof(*nbp->nbfp_comb);
403 memset(&td, 0, sizeof(td));
404 td.readMode = cudaReadModeElementType;
405 stat = cudaCreateTextureObject(&nbp->nbfp_comb_texobj, &rd, &td, NULL);
406 CU_RET_ERR(stat, "cudaCreateTextureObject on nbfp_comb_texobj failed");
412 cudaChannelFormatDesc cd = cudaCreateChannelDesc<float>();
413 stat = cudaBindTexture(NULL, &nbnxn_cuda_get_nbfp_texref(),
414 nbp->nbfp, &cd, nnbfp*sizeof(*nbp->nbfp));
415 CU_RET_ERR(stat, "cudaBindTexture on nbfp_texref failed");
417 if (ic->vdwtype == evdwPME)
419 stat = cudaBindTexture(NULL, &nbnxn_cuda_get_nbfp_comb_texref(),
420 nbp->nbfp_comb, &cd, nnbfp_comb*sizeof(*nbp->nbfp_comb));
421 CU_RET_ERR(stat, "cudaBindTexture on nbfp_comb_texref failed");
426 /*! Re-generate the GPU Ewald force table, resets rlist, and update the
427 * electrostatic type switching to twin cut-off (or back) if needed. */
428 void nbnxn_cuda_pme_loadbal_update_param(nbnxn_cuda_ptr_t cu_nb,
429 const interaction_const_t *ic)
431 cu_nbparam_t *nbp = cu_nb->nbparam;
433 set_cutoff_parameters(nbp, ic);
435 nbp->eeltype = pick_ewald_kernel_type(ic->rcoulomb != ic->rvdw,
438 init_ewald_coulomb_force_table(cu_nb->nbparam, cu_nb->dev_info);
441 /*! Initializes the pair list data structure. */
442 static void init_plist(cu_plist_t *pl)
444 /* initialize to NULL pointers to data that is not allocated here and will
445 need reallocation in nbnxn_cuda_init_pairlist */
450 /* size -1 indicates that the respective array hasn't been initialized yet */
457 pl->excl_nalloc = -1;
458 pl->bDoPrune = false;
461 /*! Initializes the timer data structure. */
462 static void init_timers(cu_timers_t *t, bool bUseTwoStreams)
465 int eventflags = ( bUseCudaEventBlockingSync ? cudaEventBlockingSync : cudaEventDefault );
467 stat = cudaEventCreateWithFlags(&(t->start_atdat), eventflags);
468 CU_RET_ERR(stat, "cudaEventCreate on start_atdat failed");
469 stat = cudaEventCreateWithFlags(&(t->stop_atdat), eventflags);
470 CU_RET_ERR(stat, "cudaEventCreate on stop_atdat failed");
472 /* The non-local counters/stream (second in the array) are needed only with DD. */
473 for (int i = 0; i <= (bUseTwoStreams ? 1 : 0); i++)
475 stat = cudaEventCreateWithFlags(&(t->start_nb_k[i]), eventflags);
476 CU_RET_ERR(stat, "cudaEventCreate on start_nb_k failed");
477 stat = cudaEventCreateWithFlags(&(t->stop_nb_k[i]), eventflags);
478 CU_RET_ERR(stat, "cudaEventCreate on stop_nb_k failed");
481 stat = cudaEventCreateWithFlags(&(t->start_pl_h2d[i]), eventflags);
482 CU_RET_ERR(stat, "cudaEventCreate on start_pl_h2d failed");
483 stat = cudaEventCreateWithFlags(&(t->stop_pl_h2d[i]), eventflags);
484 CU_RET_ERR(stat, "cudaEventCreate on stop_pl_h2d failed");
486 stat = cudaEventCreateWithFlags(&(t->start_nb_h2d[i]), eventflags);
487 CU_RET_ERR(stat, "cudaEventCreate on start_nb_h2d failed");
488 stat = cudaEventCreateWithFlags(&(t->stop_nb_h2d[i]), eventflags);
489 CU_RET_ERR(stat, "cudaEventCreate on stop_nb_h2d failed");
491 stat = cudaEventCreateWithFlags(&(t->start_nb_d2h[i]), eventflags);
492 CU_RET_ERR(stat, "cudaEventCreate on start_nb_d2h failed");
493 stat = cudaEventCreateWithFlags(&(t->stop_nb_d2h[i]), eventflags);
494 CU_RET_ERR(stat, "cudaEventCreate on stop_nb_d2h failed");
498 /*! Initializes the timings data structure. */
499 static void init_timings(wallclock_gpu_t *t)
508 for (i = 0; i < 2; i++)
510 for (j = 0; j < 2; j++)
512 t->ktime[i][j].t = 0.0;
513 t->ktime[i][j].c = 0;
518 void nbnxn_cuda_init(FILE *fplog,
519 nbnxn_cuda_ptr_t *p_cu_nb,
520 const gmx_gpu_info_t *gpu_info,
521 const gmx_gpu_opt_t *gpu_opt,
523 gmx_bool bLocalAndNonlocal)
528 bool bStreamSync, bNoStreamSync, bTMPIAtomics, bX86, bOldDriver;
540 snew(nb->nbparam, 1);
541 snew(nb->plist[eintLocal], 1);
542 if (bLocalAndNonlocal)
544 snew(nb->plist[eintNonlocal], 1);
547 nb->bUseTwoStreams = bLocalAndNonlocal;
550 snew(nb->timings, 1);
553 pmalloc((void**)&nb->nbst.e_lj, sizeof(*nb->nbst.e_lj));
554 pmalloc((void**)&nb->nbst.e_el, sizeof(*nb->nbst.e_el));
555 pmalloc((void**)&nb->nbst.fshift, SHIFTS * sizeof(*nb->nbst.fshift));
557 init_plist(nb->plist[eintLocal]);
559 /* set device info, just point it to the right GPU among the detected ones */
560 nb->dev_info = &gpu_info->cuda_dev[get_gpu_device_id(gpu_info, gpu_opt, my_gpu_index)];
562 /* local/non-local GPU streams */
563 stat = cudaStreamCreate(&nb->stream[eintLocal]);
564 CU_RET_ERR(stat, "cudaStreamCreate on stream[eintLocal] failed");
565 if (nb->bUseTwoStreams)
567 init_plist(nb->plist[eintNonlocal]);
569 /* CUDA stream priority available in the CUDA RT 5.5 API.
570 * Note that the device we're running on does not have to support
571 * priorities, because we are querying the priority range which in this
572 * case will be a single value.
574 #if CUDA_VERSION >= 5500
576 int highest_priority;
577 stat = cudaDeviceGetStreamPriorityRange(NULL, &highest_priority);
578 CU_RET_ERR(stat, "cudaDeviceGetStreamPriorityRange failed");
580 stat = cudaStreamCreateWithPriority(&nb->stream[eintNonlocal],
583 CU_RET_ERR(stat, "cudaStreamCreateWithPriority on stream[eintNonlocal] failed");
586 stat = cudaStreamCreate(&nb->stream[eintNonlocal]);
587 CU_RET_ERR(stat, "cudaStreamCreate on stream[eintNonlocal] failed");
591 /* init events for sychronization (timing disabled for performance reasons!) */
592 stat = cudaEventCreateWithFlags(&nb->nonlocal_done, cudaEventDisableTiming);
593 CU_RET_ERR(stat, "cudaEventCreate on nonlocal_done failed");
594 stat = cudaEventCreateWithFlags(&nb->misc_ops_done, cudaEventDisableTiming);
595 CU_RET_ERR(stat, "cudaEventCreate on misc_ops_one failed");
597 /* On GPUs with ECC enabled, cudaStreamSynchronize shows a large overhead
598 * (which increases with shorter time/step) caused by a known CUDA driver bug.
599 * To work around the issue we'll use an (admittedly fragile) memory polling
600 * waiting to preserve performance. This requires support for atomic
601 * operations and only works on x86/x86_64.
602 * With polling wait event-timing also needs to be disabled.
604 * The overhead is greatly reduced in API v5.0 drivers and the improvement
605 * is independent of runtime version. Hence, with API v5.0 drivers and later
606 * we won't switch to polling.
608 * NOTE: Unfortunately, this is known to fail when GPUs are shared by (t)MPI,
609 * ranks so we will also disable it in that case.
612 bStreamSync = getenv("GMX_CUDA_STREAMSYNC") != NULL;
613 bNoStreamSync = getenv("GMX_NO_CUDA_STREAMSYNC") != NULL;
618 bTMPIAtomics = false;
621 #ifdef GMX_TARGET_X86
627 if (bStreamSync && bNoStreamSync)
629 gmx_fatal(FARGS, "Conflicting environment variables: both GMX_CUDA_STREAMSYNC and GMX_NO_CUDA_STREAMSYNC defined");
632 stat = cudaDriverGetVersion(&cuda_drv_ver);
633 CU_RET_ERR(stat, "cudaDriverGetVersion failed");
635 bOldDriver = (cuda_drv_ver < 5000);
637 if ((nb->dev_info->prop.ECCEnabled == 1) && bOldDriver)
639 /* Polling wait should be used instead of cudaStreamSynchronize only if:
640 * - ECC is ON & driver is old (checked above),
641 * - we're on x86/x86_64,
642 * - atomics are available, and
643 * - GPUs are not being shared.
645 bool bShouldUsePollSync = (bX86 && bTMPIAtomics &&
646 (gmx_count_gpu_dev_shared(gpu_opt) < 1));
650 nb->bUseStreamSync = true;
652 /* only warn if polling should be used */
653 if (bShouldUsePollSync)
656 "NOTE: Using a GPU with ECC enabled and CUDA driver API version <5.0, but\n"
657 " cudaStreamSynchronize waiting is forced by the GMX_CUDA_STREAMSYNC env. var.\n");
662 nb->bUseStreamSync = !bShouldUsePollSync;
664 if (bShouldUsePollSync)
667 "NOTE: Using a GPU with ECC enabled and CUDA driver API version <5.0, known to\n"
668 " cause performance loss. Switching to the alternative polling GPU wait.\n"
669 " If you encounter issues, switch back to standard GPU waiting by setting\n"
670 " the GMX_CUDA_STREAMSYNC environment variable.\n");
674 /* Tell the user that the ECC+old driver combination can be bad */
676 "NOTE: Using a GPU with ECC enabled and CUDA driver API version <5.0.\n"
677 " A known bug in this driver version can cause performance loss.\n"
678 " However, the polling wait workaround can not be used because\n%s\n"
679 " Consider updating the driver or turning ECC off.",
680 (bX86 && bTMPIAtomics) ?
681 " GPU(s) are being oversubscribed." :
682 " atomic operations are not supported by the platform/CPU+compiler.");
683 md_print_warn(fplog, sbuf);
691 nb->bUseStreamSync = false;
694 "NOTE: Polling wait for GPU synchronization requested by GMX_NO_CUDA_STREAMSYNC\n");
698 /* no/off ECC, cudaStreamSynchronize not turned off by env. var. */
699 nb->bUseStreamSync = true;
703 /* CUDA timing disabled as event timers don't work:
704 - with multiple streams = domain-decomposition;
705 - with the polling waiting hack (without cudaStreamSynchronize);
706 - when turned off by GMX_DISABLE_CUDA_TIMING.
708 nb->bDoTime = (!nb->bUseTwoStreams && nb->bUseStreamSync &&
709 (getenv("GMX_DISABLE_CUDA_TIMING") == NULL));
713 init_timers(nb->timers, nb->bUseTwoStreams);
714 init_timings(nb->timings);
717 /* set the kernel type for the current GPU */
718 /* pick L1 cache configuration */
719 nbnxn_cuda_set_cacheconfig(nb->dev_info);
725 fprintf(debug, "Initialized CUDA data structures.\n");
729 void nbnxn_cuda_init_const(nbnxn_cuda_ptr_t cu_nb,
730 const interaction_const_t *ic,
731 const nonbonded_verlet_group_t *nbv_group)
733 init_atomdata_first(cu_nb->atdat, nbv_group[0].nbat->ntype);
734 init_nbparam(cu_nb->nbparam, ic, nbv_group[0].nbat, cu_nb->dev_info);
736 /* clear energy and shift force outputs */
737 nbnxn_cuda_clear_e_fshift(cu_nb);
740 void nbnxn_cuda_init_pairlist(nbnxn_cuda_ptr_t cu_nb,
741 const nbnxn_pairlist_t *h_plist,
746 bool bDoTime = cu_nb->bDoTime;
747 cudaStream_t stream = cu_nb->stream[iloc];
748 cu_plist_t *d_plist = cu_nb->plist[iloc];
750 if (d_plist->na_c < 0)
752 d_plist->na_c = h_plist->na_ci;
756 if (d_plist->na_c != h_plist->na_ci)
758 sprintf(sbuf, "In cu_init_plist: the #atoms per cell has changed (from %d to %d)",
759 d_plist->na_c, h_plist->na_ci);
766 stat = cudaEventRecord(cu_nb->timers->start_pl_h2d[iloc], stream);
767 CU_RET_ERR(stat, "cudaEventRecord failed");
770 cu_realloc_buffered((void **)&d_plist->sci, h_plist->sci, sizeof(*d_plist->sci),
771 &d_plist->nsci, &d_plist->sci_nalloc,
775 cu_realloc_buffered((void **)&d_plist->cj4, h_plist->cj4, sizeof(*d_plist->cj4),
776 &d_plist->ncj4, &d_plist->cj4_nalloc,
780 cu_realloc_buffered((void **)&d_plist->excl, h_plist->excl, sizeof(*d_plist->excl),
781 &d_plist->nexcl, &d_plist->excl_nalloc,
787 stat = cudaEventRecord(cu_nb->timers->stop_pl_h2d[iloc], stream);
788 CU_RET_ERR(stat, "cudaEventRecord failed");
791 /* need to prune the pair list during the next step */
792 d_plist->bDoPrune = true;
795 void nbnxn_cuda_upload_shiftvec(nbnxn_cuda_ptr_t cu_nb,
796 const nbnxn_atomdata_t *nbatom)
798 cu_atomdata_t *adat = cu_nb->atdat;
799 cudaStream_t ls = cu_nb->stream[eintLocal];
801 /* only if we have a dynamic box */
802 if (nbatom->bDynamicBox || !adat->bShiftVecUploaded)
804 cu_copy_H2D_async(adat->shift_vec, nbatom->shift_vec,
805 SHIFTS * sizeof(*adat->shift_vec), ls);
806 adat->bShiftVecUploaded = true;
810 /*! Clears the first natoms_clear elements of the GPU nonbonded force output array. */
811 static void nbnxn_cuda_clear_f(nbnxn_cuda_ptr_t cu_nb, int natoms_clear)
814 cu_atomdata_t *adat = cu_nb->atdat;
815 cudaStream_t ls = cu_nb->stream[eintLocal];
817 stat = cudaMemsetAsync(adat->f, 0, natoms_clear * sizeof(*adat->f), ls);
818 CU_RET_ERR(stat, "cudaMemsetAsync on f falied");
821 /*! Clears nonbonded shift force output array and energy outputs on the GPU. */
822 static void nbnxn_cuda_clear_e_fshift(nbnxn_cuda_ptr_t cu_nb)
825 cu_atomdata_t *adat = cu_nb->atdat;
826 cudaStream_t ls = cu_nb->stream[eintLocal];
828 stat = cudaMemsetAsync(adat->fshift, 0, SHIFTS * sizeof(*adat->fshift), ls);
829 CU_RET_ERR(stat, "cudaMemsetAsync on fshift falied");
830 stat = cudaMemsetAsync(adat->e_lj, 0, sizeof(*adat->e_lj), ls);
831 CU_RET_ERR(stat, "cudaMemsetAsync on e_lj falied");
832 stat = cudaMemsetAsync(adat->e_el, 0, sizeof(*adat->e_el), ls);
833 CU_RET_ERR(stat, "cudaMemsetAsync on e_el falied");
836 void nbnxn_cuda_clear_outputs(nbnxn_cuda_ptr_t cu_nb, int flags)
838 nbnxn_cuda_clear_f(cu_nb, cu_nb->atdat->natoms);
839 /* clear shift force array and energies if the outputs were
840 used in the current step */
841 if (flags & GMX_FORCE_VIRIAL)
843 nbnxn_cuda_clear_e_fshift(cu_nb);
847 void nbnxn_cuda_init_atomdata(nbnxn_cuda_ptr_t cu_nb,
848 const nbnxn_atomdata_t *nbat)
853 bool bDoTime = cu_nb->bDoTime;
854 cu_timers_t *timers = cu_nb->timers;
855 cu_atomdata_t *d_atdat = cu_nb->atdat;
856 cudaStream_t ls = cu_nb->stream[eintLocal];
858 natoms = nbat->natoms;
863 /* time async copy */
864 stat = cudaEventRecord(timers->start_atdat, ls);
865 CU_RET_ERR(stat, "cudaEventRecord failed");
868 /* need to reallocate if we have to copy more atoms than the amount of space
869 available and only allocate if we haven't initialized yet, i.e d_atdat->natoms == -1 */
870 if (natoms > d_atdat->nalloc)
872 nalloc = over_alloc_small(natoms);
874 /* free up first if the arrays have already been initialized */
875 if (d_atdat->nalloc != -1)
877 cu_free_buffered(d_atdat->f, &d_atdat->natoms, &d_atdat->nalloc);
878 cu_free_buffered(d_atdat->xq);
879 cu_free_buffered(d_atdat->atom_types);
882 stat = cudaMalloc((void **)&d_atdat->f, nalloc*sizeof(*d_atdat->f));
883 CU_RET_ERR(stat, "cudaMalloc failed on d_atdat->f");
884 stat = cudaMalloc((void **)&d_atdat->xq, nalloc*sizeof(*d_atdat->xq));
885 CU_RET_ERR(stat, "cudaMalloc failed on d_atdat->xq");
887 stat = cudaMalloc((void **)&d_atdat->atom_types, nalloc*sizeof(*d_atdat->atom_types));
888 CU_RET_ERR(stat, "cudaMalloc failed on d_atdat->atom_types");
890 d_atdat->nalloc = nalloc;
894 d_atdat->natoms = natoms;
895 d_atdat->natoms_local = nbat->natoms_local;
897 /* need to clear GPU f output if realloc happened */
900 nbnxn_cuda_clear_f(cu_nb, nalloc);
903 cu_copy_H2D_async(d_atdat->atom_types, nbat->type,
904 natoms*sizeof(*d_atdat->atom_types), ls);
908 stat = cudaEventRecord(timers->stop_atdat, ls);
909 CU_RET_ERR(stat, "cudaEventRecord failed");
913 void nbnxn_cuda_free(nbnxn_cuda_ptr_t cu_nb)
916 cu_atomdata_t *atdat;
917 cu_nbparam_t *nbparam;
918 cu_plist_t *plist, *plist_nl;
926 atdat = cu_nb->atdat;
927 nbparam = cu_nb->nbparam;
928 plist = cu_nb->plist[eintLocal];
929 plist_nl = cu_nb->plist[eintNonlocal];
930 timers = cu_nb->timers;
932 if (nbparam->eeltype == eelCuEWALD_TAB || nbparam->eeltype == eelCuEWALD_TAB_TWIN)
935 #ifdef TEXOBJ_SUPPORTED
936 /* Only device CC >= 3.0 (Kepler and later) support texture objects */
937 if (cu_nb->dev_info->prop.major >= 3)
939 stat = cudaDestroyTextureObject(nbparam->coulomb_tab_texobj);
940 CU_RET_ERR(stat, "cudaDestroyTextureObject on coulomb_tab_texobj failed");
945 stat = cudaUnbindTexture(nbnxn_cuda_get_coulomb_tab_texref());
946 CU_RET_ERR(stat, "cudaUnbindTexture on coulomb_tab_texref failed");
948 cu_free_buffered(nbparam->coulomb_tab, &nbparam->coulomb_tab_size);
951 stat = cudaEventDestroy(cu_nb->nonlocal_done);
952 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->nonlocal_done");
953 stat = cudaEventDestroy(cu_nb->misc_ops_done);
954 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->misc_ops_done");
958 stat = cudaEventDestroy(timers->start_atdat);
959 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_atdat");
960 stat = cudaEventDestroy(timers->stop_atdat);
961 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_atdat");
963 /* The non-local counters/stream (second in the array) are needed only with DD. */
964 for (int i = 0; i <= (cu_nb->bUseTwoStreams ? 1 : 0); i++)
966 stat = cudaEventDestroy(timers->start_nb_k[i]);
967 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_nb_k");
968 stat = cudaEventDestroy(timers->stop_nb_k[i]);
969 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_nb_k");
971 stat = cudaEventDestroy(timers->start_pl_h2d[i]);
972 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_pl_h2d");
973 stat = cudaEventDestroy(timers->stop_pl_h2d[i]);
974 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_pl_h2d");
976 stat = cudaStreamDestroy(cu_nb->stream[i]);
977 CU_RET_ERR(stat, "cudaStreamDestroy failed on stream");
979 stat = cudaEventDestroy(timers->start_nb_h2d[i]);
980 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_nb_h2d");
981 stat = cudaEventDestroy(timers->stop_nb_h2d[i]);
982 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_nb_h2d");
984 stat = cudaEventDestroy(timers->start_nb_d2h[i]);
985 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_nb_d2h");
986 stat = cudaEventDestroy(timers->stop_nb_d2h[i]);
987 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_nb_d2h");
991 #ifdef TEXOBJ_SUPPORTED
992 /* Only device CC >= 3.0 (Kepler and later) support texture objects */
993 if (cu_nb->dev_info->prop.major >= 3)
995 stat = cudaDestroyTextureObject(nbparam->nbfp_texobj);
996 CU_RET_ERR(stat, "cudaDestroyTextureObject on nbfp_texobj failed");
1001 stat = cudaUnbindTexture(nbnxn_cuda_get_nbfp_texref());
1002 CU_RET_ERR(stat, "cudaUnbindTexture on nbfp_texref failed");
1004 cu_free_buffered(nbparam->nbfp);
1006 if (nbparam->vdwtype == evdwCuEWALDGEOM || nbparam->vdwtype == evdwCuEWALDLB)
1008 #ifdef TEXOBJ_SUPPORTED
1009 /* Only device CC >= 3.0 (Kepler and later) support texture objects */
1010 if (cu_nb->dev_info->prop.major >= 3)
1012 stat = cudaDestroyTextureObject(nbparam->nbfp_comb_texobj);
1013 CU_RET_ERR(stat, "cudaDestroyTextureObject on nbfp_comb_texobj failed");
1018 stat = cudaUnbindTexture(nbnxn_cuda_get_nbfp_comb_texref());
1019 CU_RET_ERR(stat, "cudaUnbindTexture on nbfp_comb_texref failed");
1021 cu_free_buffered(nbparam->nbfp_comb);
1024 stat = cudaFree(atdat->shift_vec);
1025 CU_RET_ERR(stat, "cudaFree failed on atdat->shift_vec");
1026 stat = cudaFree(atdat->fshift);
1027 CU_RET_ERR(stat, "cudaFree failed on atdat->fshift");
1029 stat = cudaFree(atdat->e_lj);
1030 CU_RET_ERR(stat, "cudaFree failed on atdat->e_lj");
1031 stat = cudaFree(atdat->e_el);
1032 CU_RET_ERR(stat, "cudaFree failed on atdat->e_el");
1034 cu_free_buffered(atdat->f, &atdat->natoms, &atdat->nalloc);
1035 cu_free_buffered(atdat->xq);
1036 cu_free_buffered(atdat->atom_types, &atdat->ntypes);
1038 cu_free_buffered(plist->sci, &plist->nsci, &plist->sci_nalloc);
1039 cu_free_buffered(plist->cj4, &plist->ncj4, &plist->cj4_nalloc);
1040 cu_free_buffered(plist->excl, &plist->nexcl, &plist->excl_nalloc);
1041 if (cu_nb->bUseTwoStreams)
1043 cu_free_buffered(plist_nl->sci, &plist_nl->nsci, &plist_nl->sci_nalloc);
1044 cu_free_buffered(plist_nl->cj4, &plist_nl->ncj4, &plist_nl->cj4_nalloc);
1045 cu_free_buffered(plist_nl->excl, &plist_nl->nexcl, &plist->excl_nalloc);
1051 if (cu_nb->bUseTwoStreams)
1056 sfree(cu_nb->timings);
1061 fprintf(debug, "Cleaned up CUDA data structures.\n");
1065 void cu_synchstream_atdat(nbnxn_cuda_ptr_t cu_nb, int iloc)
1068 cudaStream_t stream = cu_nb->stream[iloc];
1070 stat = cudaStreamWaitEvent(stream, cu_nb->timers->stop_atdat, 0);
1071 CU_RET_ERR(stat, "cudaStreamWaitEvent failed");
1074 wallclock_gpu_t * nbnxn_cuda_get_timings(nbnxn_cuda_ptr_t cu_nb)
1076 return (cu_nb != NULL && cu_nb->bDoTime) ? cu_nb->timings : NULL;
1079 void nbnxn_cuda_reset_timings(nbnxn_cuda_ptr_t cu_nb)
1083 init_timings(cu_nb->timings);
1087 int nbnxn_cuda_min_ci_balanced(nbnxn_cuda_ptr_t cu_nb)
1089 return cu_nb != NULL ?
1090 gpu_min_ci_balanced_factor*cu_nb->dev_info->prop.multiProcessorCount : 0;
1094 gmx_bool nbnxn_cuda_is_kernel_ewald_analytical(const nbnxn_cuda_ptr_t cu_nb)
1096 return ((cu_nb->nbparam->eeltype == eelCuEWALD_ANA) ||
1097 (cu_nb->nbparam->eeltype == eelCuEWALD_ANA_TWIN));