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.
48 #include "types/enums.h"
49 #include "types/nb_verlet.h"
50 #include "types/interaction_const.h"
51 #include "types/force_flags.h"
52 #include "../nbnxn_consts.h"
53 #include "gmx_detect_hardware.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 #include "gromacs/utility/common.h"
62 #include "gromacs/utility/cstringutil.h"
63 #include "gromacs/utility/fatalerror.h"
64 #include "gromacs/utility/smalloc.h"
66 static bool bUseCudaEventBlockingSync = false; /* makes the CPU thread block */
68 /* This is a heuristically determined parameter for the Fermi architecture for
69 * the minimum size of ci lists by multiplying this constant with the # of
70 * multiprocessors on the current device.
72 static unsigned int gpu_min_ci_balanced_factor = 40;
74 /* Functions from nbnxn_cuda.cu */
75 extern void nbnxn_cuda_set_cacheconfig(cuda_dev_info_t *devinfo);
76 extern const struct texture<float, 1, cudaReadModeElementType> &nbnxn_cuda_get_nbfp_texref();
77 extern const struct texture<float, 1, cudaReadModeElementType> &nbnxn_cuda_get_nbfp_comb_texref();
78 extern const struct texture<float, 1, cudaReadModeElementType> &nbnxn_cuda_get_coulomb_tab_texref();
80 /* We should actually be using md_print_warn in md_logging.c,
81 * but we can't include mpi.h in CUDA code.
83 static void md_print_warn(FILE *fplog,
90 /* We should only print to stderr on the master node,
91 * in most cases fplog is only set on the master node, so this works.
94 fprintf(stderr, "\n");
95 vfprintf(stderr, fmt, ap);
96 fprintf(stderr, "\n");
100 fprintf(fplog, "\n");
101 vfprintf(fplog, fmt, ap);
102 fprintf(fplog, "\n");
109 static void nbnxn_cuda_clear_e_fshift(nbnxn_cuda_ptr_t cu_nb);
112 /*! Tabulates the Ewald Coulomb force and initializes the size/scale
113 and the table GPU array. If called with an already allocated table,
114 it just re-uploads the table.
116 static void init_ewald_coulomb_force_table(cu_nbparam_t *nbp,
117 const cuda_dev_info_t *dev_info)
119 float *ftmp, *coul_tab;
124 tabsize = GPU_EWALD_COULOMB_FORCE_TABLE_SIZE;
125 /* Subtract 2 iso 1 to avoid access out of range due to rounding */
126 tabscale = (tabsize - 2) / sqrt(nbp->rcoulomb_sq);
128 pmalloc((void**)&ftmp, tabsize*sizeof(*ftmp));
130 table_spline3_fill_ewald_lr(ftmp, NULL, NULL, tabsize,
131 1/tabscale, nbp->ewald_beta, v_q_ewald_lr);
133 /* If the table pointer == NULL the table is generated the first time =>
134 the array pointer will be saved to nbparam and the texture is bound.
136 coul_tab = nbp->coulomb_tab;
137 if (coul_tab == NULL)
139 stat = cudaMalloc((void **)&coul_tab, tabsize*sizeof(*coul_tab));
140 CU_RET_ERR(stat, "cudaMalloc failed on coul_tab");
142 nbp->coulomb_tab = coul_tab;
144 #ifdef TEXOBJ_SUPPORTED
145 /* Only device CC >= 3.0 (Kepler and later) support texture objects */
146 if (dev_info->prop.major >= 3)
149 memset(&rd, 0, sizeof(rd));
150 rd.resType = cudaResourceTypeLinear;
151 rd.res.linear.devPtr = nbp->coulomb_tab;
152 rd.res.linear.desc.f = cudaChannelFormatKindFloat;
153 rd.res.linear.desc.x = 32;
154 rd.res.linear.sizeInBytes = tabsize*sizeof(*coul_tab);
157 memset(&td, 0, sizeof(td));
158 td.readMode = cudaReadModeElementType;
159 stat = cudaCreateTextureObject(&nbp->coulomb_tab_texobj, &rd, &td, NULL);
160 CU_RET_ERR(stat, "cudaCreateTextureObject on coulomb_tab_texobj failed");
165 GMX_UNUSED_VALUE(dev_info);
166 cudaChannelFormatDesc cd = cudaCreateChannelDesc<float>();
167 stat = cudaBindTexture(NULL, &nbnxn_cuda_get_coulomb_tab_texref(),
168 coul_tab, &cd, tabsize*sizeof(*coul_tab));
169 CU_RET_ERR(stat, "cudaBindTexture on coulomb_tab_texref failed");
173 cu_copy_H2D(coul_tab, ftmp, tabsize*sizeof(*coul_tab));
175 nbp->coulomb_tab_size = tabsize;
176 nbp->coulomb_tab_scale = tabscale;
182 /*! Initializes the atomdata structure first time, it only gets filled at
184 static void init_atomdata_first(cu_atomdata_t *ad, int ntypes)
189 stat = cudaMalloc((void**)&ad->shift_vec, SHIFTS*sizeof(*ad->shift_vec));
190 CU_RET_ERR(stat, "cudaMalloc failed on ad->shift_vec");
191 ad->bShiftVecUploaded = false;
193 stat = cudaMalloc((void**)&ad->fshift, SHIFTS*sizeof(*ad->fshift));
194 CU_RET_ERR(stat, "cudaMalloc failed on ad->fshift");
196 stat = cudaMalloc((void**)&ad->e_lj, sizeof(*ad->e_lj));
197 CU_RET_ERR(stat, "cudaMalloc failed on ad->e_lj");
198 stat = cudaMalloc((void**)&ad->e_el, sizeof(*ad->e_el));
199 CU_RET_ERR(stat, "cudaMalloc failed on ad->e_el");
201 /* initialize to NULL poiters to data that is not allocated here and will
202 need reallocation in nbnxn_cuda_init_atomdata */
206 /* size -1 indicates that the respective array hasn't been initialized yet */
211 /*! Selects the Ewald kernel type, analytical on SM 3.0 and later, tabulated on
212 earlier GPUs, single or twin cut-off. */
213 static int pick_ewald_kernel_type(bool bTwinCut,
214 const cuda_dev_info_t *dev_info)
216 bool bUseAnalyticalEwald, bForceAnalyticalEwald, bForceTabulatedEwald;
219 /* Benchmarking/development environment variables to force the use of
220 analytical or tabulated Ewald kernel. */
221 bForceAnalyticalEwald = (getenv("GMX_CUDA_NB_ANA_EWALD") != NULL);
222 bForceTabulatedEwald = (getenv("GMX_CUDA_NB_TAB_EWALD") != NULL);
224 if (bForceAnalyticalEwald && bForceTabulatedEwald)
226 gmx_incons("Both analytical and tabulated Ewald CUDA non-bonded kernels "
227 "requested through environment variables.");
230 /* By default, on SM 3.0 and later use analytical Ewald, on earlier tabulated. */
231 if ((dev_info->prop.major >= 3 || bForceAnalyticalEwald) && !bForceTabulatedEwald)
233 bUseAnalyticalEwald = true;
237 fprintf(debug, "Using analytical Ewald CUDA kernels\n");
242 bUseAnalyticalEwald = false;
246 fprintf(debug, "Using tabulated Ewald CUDA kernels\n");
250 /* Use twin cut-off kernels if requested by bTwinCut or the env. var.
251 forces it (use it for debugging/benchmarking only). */
252 if (!bTwinCut && (getenv("GMX_CUDA_NB_EWALD_TWINCUT") == NULL))
254 kernel_type = bUseAnalyticalEwald ? eelCuEWALD_ANA : eelCuEWALD_TAB;
258 kernel_type = bUseAnalyticalEwald ? eelCuEWALD_ANA_TWIN : eelCuEWALD_TAB_TWIN;
264 /*! Copies all parameters related to the cut-off from ic to nbp */
265 static void set_cutoff_parameters(cu_nbparam_t *nbp,
266 const interaction_const_t *ic)
268 nbp->ewald_beta = ic->ewaldcoeff_q;
269 nbp->sh_ewald = ic->sh_ewald;
270 nbp->epsfac = ic->epsfac;
271 nbp->two_k_rf = 2.0 * ic->k_rf;
272 nbp->c_rf = ic->c_rf;
273 nbp->rvdw_sq = ic->rvdw * ic->rvdw;
274 nbp->rcoulomb_sq = ic->rcoulomb * ic->rcoulomb;
275 nbp->rlist_sq = ic->rlist * ic->rlist;
277 nbp->sh_lj_ewald = ic->sh_lj_ewald;
278 nbp->ewaldcoeff_lj = ic->ewaldcoeff_lj;
280 nbp->rvdw_switch = ic->rvdw_switch;
281 nbp->dispersion_shift = ic->dispersion_shift;
282 nbp->repulsion_shift = ic->repulsion_shift;
283 nbp->vdw_switch = ic->vdw_switch;
286 /*! Initializes the nonbonded parameter data structure. */
287 static void init_nbparam(cu_nbparam_t *nbp,
288 const interaction_const_t *ic,
289 const nbnxn_atomdata_t *nbat,
290 const cuda_dev_info_t *dev_info)
293 int ntypes, nnbfp, nnbfp_comb;
295 ntypes = nbat->ntype;
297 set_cutoff_parameters(nbp, ic);
299 if (ic->vdwtype == evdwCUT)
301 switch (ic->vdw_modifier)
304 case eintmodPOTSHIFT:
305 nbp->vdwtype = evdwCuCUT;
307 case eintmodFORCESWITCH:
308 nbp->vdwtype = evdwCuFSWITCH;
310 case eintmodPOTSWITCH:
311 nbp->vdwtype = evdwCuPSWITCH;
314 gmx_incons("The requested VdW interaction modifier is not implemented in the CUDA GPU accelerated kernels!");
318 else if (ic->vdwtype == evdwPME)
320 if (ic->ljpme_comb_rule == ljcrGEOM)
322 assert(nbat->comb_rule == ljcrGEOM);
323 nbp->vdwtype = evdwCuEWALDGEOM;
327 assert(nbat->comb_rule == ljcrLB);
328 nbp->vdwtype = evdwCuEWALDLB;
333 gmx_incons("The requested VdW type is not implemented in the CUDA GPU accelerated kernels!");
336 if (ic->eeltype == eelCUT)
338 nbp->eeltype = eelCuCUT;
340 else if (EEL_RF(ic->eeltype))
342 nbp->eeltype = eelCuRF;
344 else if ((EEL_PME(ic->eeltype) || ic->eeltype == eelEWALD))
346 /* Initially rcoulomb == rvdw, so it's surely not twin cut-off. */
347 nbp->eeltype = pick_ewald_kernel_type(false, dev_info);
351 /* Shouldn't happen, as this is checked when choosing Verlet-scheme */
352 gmx_incons("The requested electrostatics type is not implemented in the CUDA GPU accelerated kernels!");
355 /* generate table for PME */
356 nbp->coulomb_tab = NULL;
357 if (nbp->eeltype == eelCuEWALD_TAB || nbp->eeltype == eelCuEWALD_TAB_TWIN)
359 init_ewald_coulomb_force_table(nbp, dev_info);
362 nnbfp = 2*ntypes*ntypes;
363 nnbfp_comb = 2*ntypes;
365 stat = cudaMalloc((void **)&nbp->nbfp, nnbfp*sizeof(*nbp->nbfp));
366 CU_RET_ERR(stat, "cudaMalloc failed on nbp->nbfp");
367 cu_copy_H2D(nbp->nbfp, nbat->nbfp, nnbfp*sizeof(*nbp->nbfp));
370 if (ic->vdwtype == evdwPME)
372 stat = cudaMalloc((void **)&nbp->nbfp_comb, nnbfp_comb*sizeof(*nbp->nbfp_comb));
373 CU_RET_ERR(stat, "cudaMalloc failed on nbp->nbfp_comb");
374 cu_copy_H2D(nbp->nbfp_comb, nbat->nbfp_comb, nnbfp_comb*sizeof(*nbp->nbfp_comb));
377 #ifdef TEXOBJ_SUPPORTED
378 /* Only device CC >= 3.0 (Kepler and later) support texture objects */
379 if (dev_info->prop.major >= 3)
384 memset(&rd, 0, sizeof(rd));
385 rd.resType = cudaResourceTypeLinear;
386 rd.res.linear.devPtr = nbp->nbfp;
387 rd.res.linear.desc.f = cudaChannelFormatKindFloat;
388 rd.res.linear.desc.x = 32;
389 rd.res.linear.sizeInBytes = nnbfp*sizeof(*nbp->nbfp);
391 memset(&td, 0, sizeof(td));
392 td.readMode = cudaReadModeElementType;
393 stat = cudaCreateTextureObject(&nbp->nbfp_texobj, &rd, &td, NULL);
394 CU_RET_ERR(stat, "cudaCreateTextureObject on nbfp_texobj failed");
396 if (ic->vdwtype == evdwPME)
398 memset(&rd, 0, sizeof(rd));
399 rd.resType = cudaResourceTypeLinear;
400 rd.res.linear.devPtr = nbp->nbfp_comb;
401 rd.res.linear.desc.f = cudaChannelFormatKindFloat;
402 rd.res.linear.desc.x = 32;
403 rd.res.linear.sizeInBytes = nnbfp_comb*sizeof(*nbp->nbfp_comb);
405 memset(&td, 0, sizeof(td));
406 td.readMode = cudaReadModeElementType;
407 stat = cudaCreateTextureObject(&nbp->nbfp_comb_texobj, &rd, &td, NULL);
408 CU_RET_ERR(stat, "cudaCreateTextureObject on nbfp_comb_texobj failed");
414 cudaChannelFormatDesc cd = cudaCreateChannelDesc<float>();
415 stat = cudaBindTexture(NULL, &nbnxn_cuda_get_nbfp_texref(),
416 nbp->nbfp, &cd, nnbfp*sizeof(*nbp->nbfp));
417 CU_RET_ERR(stat, "cudaBindTexture on nbfp_texref failed");
419 if (ic->vdwtype == evdwPME)
421 stat = cudaBindTexture(NULL, &nbnxn_cuda_get_nbfp_comb_texref(),
422 nbp->nbfp_comb, &cd, nnbfp_comb*sizeof(*nbp->nbfp_comb));
423 CU_RET_ERR(stat, "cudaBindTexture on nbfp_comb_texref failed");
428 /*! Re-generate the GPU Ewald force table, resets rlist, and update the
429 * electrostatic type switching to twin cut-off (or back) if needed. */
430 void nbnxn_cuda_pme_loadbal_update_param(nbnxn_cuda_ptr_t cu_nb,
431 const interaction_const_t *ic)
433 cu_nbparam_t *nbp = cu_nb->nbparam;
435 set_cutoff_parameters(nbp, ic);
437 nbp->eeltype = pick_ewald_kernel_type(ic->rcoulomb != ic->rvdw,
440 init_ewald_coulomb_force_table(cu_nb->nbparam, cu_nb->dev_info);
443 /*! Initializes the pair list data structure. */
444 static void init_plist(cu_plist_t *pl)
446 /* initialize to NULL pointers to data that is not allocated here and will
447 need reallocation in nbnxn_cuda_init_pairlist */
452 /* size -1 indicates that the respective array hasn't been initialized yet */
459 pl->excl_nalloc = -1;
460 pl->bDoPrune = false;
463 /*! Initializes the timer data structure. */
464 static void init_timers(cu_timers_t *t, bool bUseTwoStreams)
467 int eventflags = ( bUseCudaEventBlockingSync ? cudaEventBlockingSync : cudaEventDefault );
469 stat = cudaEventCreateWithFlags(&(t->start_atdat), eventflags);
470 CU_RET_ERR(stat, "cudaEventCreate on start_atdat failed");
471 stat = cudaEventCreateWithFlags(&(t->stop_atdat), eventflags);
472 CU_RET_ERR(stat, "cudaEventCreate on stop_atdat failed");
474 /* The non-local counters/stream (second in the array) are needed only with DD. */
475 for (int i = 0; i <= (bUseTwoStreams ? 1 : 0); i++)
477 stat = cudaEventCreateWithFlags(&(t->start_nb_k[i]), eventflags);
478 CU_RET_ERR(stat, "cudaEventCreate on start_nb_k failed");
479 stat = cudaEventCreateWithFlags(&(t->stop_nb_k[i]), eventflags);
480 CU_RET_ERR(stat, "cudaEventCreate on stop_nb_k failed");
483 stat = cudaEventCreateWithFlags(&(t->start_pl_h2d[i]), eventflags);
484 CU_RET_ERR(stat, "cudaEventCreate on start_pl_h2d failed");
485 stat = cudaEventCreateWithFlags(&(t->stop_pl_h2d[i]), eventflags);
486 CU_RET_ERR(stat, "cudaEventCreate on stop_pl_h2d failed");
488 stat = cudaEventCreateWithFlags(&(t->start_nb_h2d[i]), eventflags);
489 CU_RET_ERR(stat, "cudaEventCreate on start_nb_h2d failed");
490 stat = cudaEventCreateWithFlags(&(t->stop_nb_h2d[i]), eventflags);
491 CU_RET_ERR(stat, "cudaEventCreate on stop_nb_h2d failed");
493 stat = cudaEventCreateWithFlags(&(t->start_nb_d2h[i]), eventflags);
494 CU_RET_ERR(stat, "cudaEventCreate on start_nb_d2h failed");
495 stat = cudaEventCreateWithFlags(&(t->stop_nb_d2h[i]), eventflags);
496 CU_RET_ERR(stat, "cudaEventCreate on stop_nb_d2h failed");
500 /*! Initializes the timings data structure. */
501 static void init_timings(wallclock_gpu_t *t)
510 for (i = 0; i < 2; i++)
512 for (j = 0; j < 2; j++)
514 t->ktime[i][j].t = 0.0;
515 t->ktime[i][j].c = 0;
520 void nbnxn_cuda_init(FILE *fplog,
521 nbnxn_cuda_ptr_t *p_cu_nb,
522 const gmx_gpu_info_t *gpu_info,
523 const gmx_gpu_opt_t *gpu_opt,
525 gmx_bool bLocalAndNonlocal)
530 bool bStreamSync, bNoStreamSync, bTMPIAtomics, bX86, bOldDriver;
542 snew(nb->nbparam, 1);
543 snew(nb->plist[eintLocal], 1);
544 if (bLocalAndNonlocal)
546 snew(nb->plist[eintNonlocal], 1);
549 nb->bUseTwoStreams = bLocalAndNonlocal;
552 snew(nb->timings, 1);
555 pmalloc((void**)&nb->nbst.e_lj, sizeof(*nb->nbst.e_lj));
556 pmalloc((void**)&nb->nbst.e_el, sizeof(*nb->nbst.e_el));
557 pmalloc((void**)&nb->nbst.fshift, SHIFTS * sizeof(*nb->nbst.fshift));
559 init_plist(nb->plist[eintLocal]);
561 /* set device info, just point it to the right GPU among the detected ones */
562 nb->dev_info = &gpu_info->cuda_dev[get_gpu_device_id(gpu_info, gpu_opt, my_gpu_index)];
564 /* local/non-local GPU streams */
565 stat = cudaStreamCreate(&nb->stream[eintLocal]);
566 CU_RET_ERR(stat, "cudaStreamCreate on stream[eintLocal] failed");
567 if (nb->bUseTwoStreams)
569 init_plist(nb->plist[eintNonlocal]);
571 /* CUDA stream priority available in the CUDA RT 5.5 API.
572 * Note that the device we're running on does not have to support
573 * priorities, because we are querying the priority range which in this
574 * case will be a single value.
576 #if CUDA_VERSION >= 5500
578 int highest_priority;
579 stat = cudaDeviceGetStreamPriorityRange(NULL, &highest_priority);
580 CU_RET_ERR(stat, "cudaDeviceGetStreamPriorityRange failed");
582 stat = cudaStreamCreateWithPriority(&nb->stream[eintNonlocal],
585 CU_RET_ERR(stat, "cudaStreamCreateWithPriority on stream[eintNonlocal] failed");
588 stat = cudaStreamCreate(&nb->stream[eintNonlocal]);
589 CU_RET_ERR(stat, "cudaStreamCreate on stream[eintNonlocal] failed");
593 /* init events for sychronization (timing disabled for performance reasons!) */
594 stat = cudaEventCreateWithFlags(&nb->nonlocal_done, cudaEventDisableTiming);
595 CU_RET_ERR(stat, "cudaEventCreate on nonlocal_done failed");
596 stat = cudaEventCreateWithFlags(&nb->misc_ops_done, cudaEventDisableTiming);
597 CU_RET_ERR(stat, "cudaEventCreate on misc_ops_one failed");
599 /* On GPUs with ECC enabled, cudaStreamSynchronize shows a large overhead
600 * (which increases with shorter time/step) caused by a known CUDA driver bug.
601 * To work around the issue we'll use an (admittedly fragile) memory polling
602 * waiting to preserve performance. This requires support for atomic
603 * operations and only works on x86/x86_64.
604 * With polling wait event-timing also needs to be disabled.
606 * The overhead is greatly reduced in API v5.0 drivers and the improvement
607 * is independent of runtime version. Hence, with API v5.0 drivers and later
608 * we won't switch to polling.
610 * NOTE: Unfortunately, this is known to fail when GPUs are shared by (t)MPI,
611 * ranks so we will also disable it in that case.
614 bStreamSync = getenv("GMX_CUDA_STREAMSYNC") != NULL;
615 bNoStreamSync = getenv("GMX_NO_CUDA_STREAMSYNC") != NULL;
620 bTMPIAtomics = false;
623 #ifdef GMX_TARGET_X86
629 if (bStreamSync && bNoStreamSync)
631 gmx_fatal(FARGS, "Conflicting environment variables: both GMX_CUDA_STREAMSYNC and GMX_NO_CUDA_STREAMSYNC defined");
634 stat = cudaDriverGetVersion(&cuda_drv_ver);
635 CU_RET_ERR(stat, "cudaDriverGetVersion failed");
637 bOldDriver = (cuda_drv_ver < 5000);
639 if ((nb->dev_info->prop.ECCEnabled == 1) && bOldDriver)
641 /* Polling wait should be used instead of cudaStreamSynchronize only if:
642 * - ECC is ON & driver is old (checked above),
643 * - we're on x86/x86_64,
644 * - atomics are available, and
645 * - GPUs are not being shared.
647 bool bShouldUsePollSync = (bX86 && bTMPIAtomics &&
648 (gmx_count_gpu_dev_shared(gpu_opt) < 1));
652 nb->bUseStreamSync = true;
654 /* only warn if polling should be used */
655 if (bShouldUsePollSync)
658 "NOTE: Using a GPU with ECC enabled and CUDA driver API version <5.0, but\n"
659 " cudaStreamSynchronize waiting is forced by the GMX_CUDA_STREAMSYNC env. var.\n");
664 nb->bUseStreamSync = !bShouldUsePollSync;
666 if (bShouldUsePollSync)
669 "NOTE: Using a GPU with ECC enabled and CUDA driver API version <5.0, known to\n"
670 " cause performance loss. Switching to the alternative polling GPU wait.\n"
671 " If you encounter issues, switch back to standard GPU waiting by setting\n"
672 " the GMX_CUDA_STREAMSYNC environment variable.\n");
676 /* Tell the user that the ECC+old driver combination can be bad */
678 "NOTE: Using a GPU with ECC enabled and CUDA driver API version <5.0.\n"
679 " A known bug in this driver version can cause performance loss.\n"
680 " However, the polling wait workaround can not be used because\n%s\n"
681 " Consider updating the driver or turning ECC off.",
682 (bX86 && bTMPIAtomics) ?
683 " GPU(s) are being oversubscribed." :
684 " atomic operations are not supported by the platform/CPU+compiler.");
685 md_print_warn(fplog, sbuf);
693 nb->bUseStreamSync = false;
696 "NOTE: Polling wait for GPU synchronization requested by GMX_NO_CUDA_STREAMSYNC\n");
700 /* no/off ECC, cudaStreamSynchronize not turned off by env. var. */
701 nb->bUseStreamSync = true;
705 /* CUDA timing disabled as event timers don't work:
706 - with multiple streams = domain-decomposition;
707 - with the polling waiting hack (without cudaStreamSynchronize);
708 - when turned off by GMX_DISABLE_CUDA_TIMING.
710 nb->bDoTime = (!nb->bUseTwoStreams && nb->bUseStreamSync &&
711 (getenv("GMX_DISABLE_CUDA_TIMING") == NULL));
715 init_timers(nb->timers, nb->bUseTwoStreams);
716 init_timings(nb->timings);
719 /* set the kernel type for the current GPU */
720 /* pick L1 cache configuration */
721 nbnxn_cuda_set_cacheconfig(nb->dev_info);
727 fprintf(debug, "Initialized CUDA data structures.\n");
731 void nbnxn_cuda_init_const(nbnxn_cuda_ptr_t cu_nb,
732 const interaction_const_t *ic,
733 const nonbonded_verlet_group_t *nbv_group)
735 init_atomdata_first(cu_nb->atdat, nbv_group[0].nbat->ntype);
736 init_nbparam(cu_nb->nbparam, ic, nbv_group[0].nbat, cu_nb->dev_info);
738 /* clear energy and shift force outputs */
739 nbnxn_cuda_clear_e_fshift(cu_nb);
742 void nbnxn_cuda_init_pairlist(nbnxn_cuda_ptr_t cu_nb,
743 const nbnxn_pairlist_t *h_plist,
748 bool bDoTime = cu_nb->bDoTime;
749 cudaStream_t stream = cu_nb->stream[iloc];
750 cu_plist_t *d_plist = cu_nb->plist[iloc];
752 if (d_plist->na_c < 0)
754 d_plist->na_c = h_plist->na_ci;
758 if (d_plist->na_c != h_plist->na_ci)
760 sprintf(sbuf, "In cu_init_plist: the #atoms per cell has changed (from %d to %d)",
761 d_plist->na_c, h_plist->na_ci);
768 stat = cudaEventRecord(cu_nb->timers->start_pl_h2d[iloc], stream);
769 CU_RET_ERR(stat, "cudaEventRecord failed");
772 cu_realloc_buffered((void **)&d_plist->sci, h_plist->sci, sizeof(*d_plist->sci),
773 &d_plist->nsci, &d_plist->sci_nalloc,
777 cu_realloc_buffered((void **)&d_plist->cj4, h_plist->cj4, sizeof(*d_plist->cj4),
778 &d_plist->ncj4, &d_plist->cj4_nalloc,
782 cu_realloc_buffered((void **)&d_plist->excl, h_plist->excl, sizeof(*d_plist->excl),
783 &d_plist->nexcl, &d_plist->excl_nalloc,
789 stat = cudaEventRecord(cu_nb->timers->stop_pl_h2d[iloc], stream);
790 CU_RET_ERR(stat, "cudaEventRecord failed");
793 /* need to prune the pair list during the next step */
794 d_plist->bDoPrune = true;
797 void nbnxn_cuda_upload_shiftvec(nbnxn_cuda_ptr_t cu_nb,
798 const nbnxn_atomdata_t *nbatom)
800 cu_atomdata_t *adat = cu_nb->atdat;
801 cudaStream_t ls = cu_nb->stream[eintLocal];
803 /* only if we have a dynamic box */
804 if (nbatom->bDynamicBox || !adat->bShiftVecUploaded)
806 cu_copy_H2D_async(adat->shift_vec, nbatom->shift_vec,
807 SHIFTS * sizeof(*adat->shift_vec), ls);
808 adat->bShiftVecUploaded = true;
812 /*! Clears the first natoms_clear elements of the GPU nonbonded force output array. */
813 static void nbnxn_cuda_clear_f(nbnxn_cuda_ptr_t cu_nb, int natoms_clear)
816 cu_atomdata_t *adat = cu_nb->atdat;
817 cudaStream_t ls = cu_nb->stream[eintLocal];
819 stat = cudaMemsetAsync(adat->f, 0, natoms_clear * sizeof(*adat->f), ls);
820 CU_RET_ERR(stat, "cudaMemsetAsync on f falied");
823 /*! Clears nonbonded shift force output array and energy outputs on the GPU. */
824 static void nbnxn_cuda_clear_e_fshift(nbnxn_cuda_ptr_t cu_nb)
827 cu_atomdata_t *adat = cu_nb->atdat;
828 cudaStream_t ls = cu_nb->stream[eintLocal];
830 stat = cudaMemsetAsync(adat->fshift, 0, SHIFTS * sizeof(*adat->fshift), ls);
831 CU_RET_ERR(stat, "cudaMemsetAsync on fshift falied");
832 stat = cudaMemsetAsync(adat->e_lj, 0, sizeof(*adat->e_lj), ls);
833 CU_RET_ERR(stat, "cudaMemsetAsync on e_lj falied");
834 stat = cudaMemsetAsync(adat->e_el, 0, sizeof(*adat->e_el), ls);
835 CU_RET_ERR(stat, "cudaMemsetAsync on e_el falied");
838 void nbnxn_cuda_clear_outputs(nbnxn_cuda_ptr_t cu_nb, int flags)
840 nbnxn_cuda_clear_f(cu_nb, cu_nb->atdat->natoms);
841 /* clear shift force array and energies if the outputs were
842 used in the current step */
843 if (flags & GMX_FORCE_VIRIAL)
845 nbnxn_cuda_clear_e_fshift(cu_nb);
849 void nbnxn_cuda_init_atomdata(nbnxn_cuda_ptr_t cu_nb,
850 const nbnxn_atomdata_t *nbat)
855 bool bDoTime = cu_nb->bDoTime;
856 cu_timers_t *timers = cu_nb->timers;
857 cu_atomdata_t *d_atdat = cu_nb->atdat;
858 cudaStream_t ls = cu_nb->stream[eintLocal];
860 natoms = nbat->natoms;
865 /* time async copy */
866 stat = cudaEventRecord(timers->start_atdat, ls);
867 CU_RET_ERR(stat, "cudaEventRecord failed");
870 /* need to reallocate if we have to copy more atoms than the amount of space
871 available and only allocate if we haven't initialized yet, i.e d_atdat->natoms == -1 */
872 if (natoms > d_atdat->nalloc)
874 nalloc = over_alloc_small(natoms);
876 /* free up first if the arrays have already been initialized */
877 if (d_atdat->nalloc != -1)
879 cu_free_buffered(d_atdat->f, &d_atdat->natoms, &d_atdat->nalloc);
880 cu_free_buffered(d_atdat->xq);
881 cu_free_buffered(d_atdat->atom_types);
884 stat = cudaMalloc((void **)&d_atdat->f, nalloc*sizeof(*d_atdat->f));
885 CU_RET_ERR(stat, "cudaMalloc failed on d_atdat->f");
886 stat = cudaMalloc((void **)&d_atdat->xq, nalloc*sizeof(*d_atdat->xq));
887 CU_RET_ERR(stat, "cudaMalloc failed on d_atdat->xq");
889 stat = cudaMalloc((void **)&d_atdat->atom_types, nalloc*sizeof(*d_atdat->atom_types));
890 CU_RET_ERR(stat, "cudaMalloc failed on d_atdat->atom_types");
892 d_atdat->nalloc = nalloc;
896 d_atdat->natoms = natoms;
897 d_atdat->natoms_local = nbat->natoms_local;
899 /* need to clear GPU f output if realloc happened */
902 nbnxn_cuda_clear_f(cu_nb, nalloc);
905 cu_copy_H2D_async(d_atdat->atom_types, nbat->type,
906 natoms*sizeof(*d_atdat->atom_types), ls);
910 stat = cudaEventRecord(timers->stop_atdat, ls);
911 CU_RET_ERR(stat, "cudaEventRecord failed");
915 void nbnxn_cuda_free(nbnxn_cuda_ptr_t cu_nb)
918 cu_atomdata_t *atdat;
919 cu_nbparam_t *nbparam;
920 cu_plist_t *plist, *plist_nl;
928 atdat = cu_nb->atdat;
929 nbparam = cu_nb->nbparam;
930 plist = cu_nb->plist[eintLocal];
931 plist_nl = cu_nb->plist[eintNonlocal];
932 timers = cu_nb->timers;
934 if (nbparam->eeltype == eelCuEWALD_TAB || nbparam->eeltype == eelCuEWALD_TAB_TWIN)
937 #ifdef TEXOBJ_SUPPORTED
938 /* Only device CC >= 3.0 (Kepler and later) support texture objects */
939 if (cu_nb->dev_info->prop.major >= 3)
941 stat = cudaDestroyTextureObject(nbparam->coulomb_tab_texobj);
942 CU_RET_ERR(stat, "cudaDestroyTextureObject on coulomb_tab_texobj failed");
947 stat = cudaUnbindTexture(nbnxn_cuda_get_coulomb_tab_texref());
948 CU_RET_ERR(stat, "cudaUnbindTexture on coulomb_tab_texref failed");
950 cu_free_buffered(nbparam->coulomb_tab, &nbparam->coulomb_tab_size);
953 stat = cudaEventDestroy(cu_nb->nonlocal_done);
954 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->nonlocal_done");
955 stat = cudaEventDestroy(cu_nb->misc_ops_done);
956 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->misc_ops_done");
960 stat = cudaEventDestroy(timers->start_atdat);
961 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_atdat");
962 stat = cudaEventDestroy(timers->stop_atdat);
963 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_atdat");
965 /* The non-local counters/stream (second in the array) are needed only with DD. */
966 for (int i = 0; i <= (cu_nb->bUseTwoStreams ? 1 : 0); i++)
968 stat = cudaEventDestroy(timers->start_nb_k[i]);
969 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_nb_k");
970 stat = cudaEventDestroy(timers->stop_nb_k[i]);
971 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_nb_k");
973 stat = cudaEventDestroy(timers->start_pl_h2d[i]);
974 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_pl_h2d");
975 stat = cudaEventDestroy(timers->stop_pl_h2d[i]);
976 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_pl_h2d");
978 stat = cudaStreamDestroy(cu_nb->stream[i]);
979 CU_RET_ERR(stat, "cudaStreamDestroy failed on stream");
981 stat = cudaEventDestroy(timers->start_nb_h2d[i]);
982 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_nb_h2d");
983 stat = cudaEventDestroy(timers->stop_nb_h2d[i]);
984 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_nb_h2d");
986 stat = cudaEventDestroy(timers->start_nb_d2h[i]);
987 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_nb_d2h");
988 stat = cudaEventDestroy(timers->stop_nb_d2h[i]);
989 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_nb_d2h");
993 #ifdef TEXOBJ_SUPPORTED
994 /* Only device CC >= 3.0 (Kepler and later) support texture objects */
995 if (cu_nb->dev_info->prop.major >= 3)
997 stat = cudaDestroyTextureObject(nbparam->nbfp_texobj);
998 CU_RET_ERR(stat, "cudaDestroyTextureObject on nbfp_texobj failed");
1003 stat = cudaUnbindTexture(nbnxn_cuda_get_nbfp_texref());
1004 CU_RET_ERR(stat, "cudaUnbindTexture on nbfp_texref failed");
1006 cu_free_buffered(nbparam->nbfp);
1008 if (nbparam->vdwtype == evdwCuEWALDGEOM || nbparam->vdwtype == evdwCuEWALDLB)
1010 #ifdef TEXOBJ_SUPPORTED
1011 /* Only device CC >= 3.0 (Kepler and later) support texture objects */
1012 if (cu_nb->dev_info->prop.major >= 3)
1014 stat = cudaDestroyTextureObject(nbparam->nbfp_comb_texobj);
1015 CU_RET_ERR(stat, "cudaDestroyTextureObject on nbfp_comb_texobj failed");
1020 stat = cudaUnbindTexture(nbnxn_cuda_get_nbfp_comb_texref());
1021 CU_RET_ERR(stat, "cudaUnbindTexture on nbfp_comb_texref failed");
1023 cu_free_buffered(nbparam->nbfp_comb);
1026 stat = cudaFree(atdat->shift_vec);
1027 CU_RET_ERR(stat, "cudaFree failed on atdat->shift_vec");
1028 stat = cudaFree(atdat->fshift);
1029 CU_RET_ERR(stat, "cudaFree failed on atdat->fshift");
1031 stat = cudaFree(atdat->e_lj);
1032 CU_RET_ERR(stat, "cudaFree failed on atdat->e_lj");
1033 stat = cudaFree(atdat->e_el);
1034 CU_RET_ERR(stat, "cudaFree failed on atdat->e_el");
1036 cu_free_buffered(atdat->f, &atdat->natoms, &atdat->nalloc);
1037 cu_free_buffered(atdat->xq);
1038 cu_free_buffered(atdat->atom_types, &atdat->ntypes);
1040 cu_free_buffered(plist->sci, &plist->nsci, &plist->sci_nalloc);
1041 cu_free_buffered(plist->cj4, &plist->ncj4, &plist->cj4_nalloc);
1042 cu_free_buffered(plist->excl, &plist->nexcl, &plist->excl_nalloc);
1043 if (cu_nb->bUseTwoStreams)
1045 cu_free_buffered(plist_nl->sci, &plist_nl->nsci, &plist_nl->sci_nalloc);
1046 cu_free_buffered(plist_nl->cj4, &plist_nl->ncj4, &plist_nl->cj4_nalloc);
1047 cu_free_buffered(plist_nl->excl, &plist_nl->nexcl, &plist->excl_nalloc);
1053 if (cu_nb->bUseTwoStreams)
1058 sfree(cu_nb->timings);
1063 fprintf(debug, "Cleaned up CUDA data structures.\n");
1067 void cu_synchstream_atdat(nbnxn_cuda_ptr_t cu_nb, int iloc)
1070 cudaStream_t stream = cu_nb->stream[iloc];
1072 stat = cudaStreamWaitEvent(stream, cu_nb->timers->stop_atdat, 0);
1073 CU_RET_ERR(stat, "cudaStreamWaitEvent failed");
1076 wallclock_gpu_t * nbnxn_cuda_get_timings(nbnxn_cuda_ptr_t cu_nb)
1078 return (cu_nb != NULL && cu_nb->bDoTime) ? cu_nb->timings : NULL;
1081 void nbnxn_cuda_reset_timings(nbnxn_cuda_ptr_t cu_nb)
1085 init_timings(cu_nb->timings);
1089 int nbnxn_cuda_min_ci_balanced(nbnxn_cuda_ptr_t cu_nb)
1091 return cu_nb != NULL ?
1092 gpu_min_ci_balanced_factor*cu_nb->dev_info->prop.multiProcessorCount : 0;
1096 gmx_bool nbnxn_cuda_is_kernel_ewald_analytical(const nbnxn_cuda_ptr_t cu_nb)
1098 return ((cu_nb->nbparam->eeltype == eelCuEWALD_ANA) ||
1099 (cu_nb->nbparam->eeltype == eelCuEWALD_ANA_TWIN));