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.
46 #include "gmx_fatal.h"
47 #include "gromacs/utility/smalloc.h"
50 #include "types/enums.h"
51 #include "types/nb_verlet.h"
52 #include "types/interaction_const.h"
53 #include "types/force_flags.h"
54 #include "../nbnxn_consts.h"
55 #include "gmx_detect_hardware.h"
57 #include "nbnxn_cuda_types.h"
58 #include "../../gmxlib/cuda_tools/cudautils.cuh"
59 #include "nbnxn_cuda_data_mgmt.h"
60 #include "pmalloc_cuda.h"
61 #include "gpu_utils.h"
63 #include "gromacs/utility/common.h"
65 static bool bUseCudaEventBlockingSync = false; /* makes the CPU thread block */
67 /* This is a heuristically determined parameter for the Fermi architecture for
68 * the minimum size of ci lists by multiplying this constant with the # of
69 * multiprocessors on the current device.
71 static unsigned int gpu_min_ci_balanced_factor = 40;
73 /* Functions from nbnxn_cuda.cu */
74 extern void nbnxn_cuda_set_cacheconfig(cuda_dev_info_t *devinfo);
75 extern const struct texture<float, 1, cudaReadModeElementType> &nbnxn_cuda_get_nbfp_texref();
76 extern const struct texture<float, 1, cudaReadModeElementType> &nbnxn_cuda_get_nbfp_comb_texref();
77 extern const struct texture<float, 1, cudaReadModeElementType> &nbnxn_cuda_get_coulomb_tab_texref();
79 /* We should actually be using md_print_warn in md_logging.c,
80 * but we can't include mpi.h in CUDA code.
82 static void md_print_warn(FILE *fplog,
89 /* We should only print to stderr on the master node,
90 * in most cases fplog is only set on the master node, so this works.
93 fprintf(stderr, "\n");
94 vfprintf(stderr, fmt, ap);
95 fprintf(stderr, "\n");
100 vfprintf(fplog, fmt, ap);
101 fprintf(fplog, "\n");
108 static void nbnxn_cuda_clear_e_fshift(nbnxn_cuda_ptr_t cu_nb);
111 /*! Tabulates the Ewald Coulomb force and initializes the size/scale
112 and the table GPU array. If called with an already allocated table,
113 it just re-uploads the table.
115 static void init_ewald_coulomb_force_table(cu_nbparam_t *nbp,
116 const cuda_dev_info_t *dev_info)
118 float *ftmp, *coul_tab;
123 tabsize = GPU_EWALD_COULOMB_FORCE_TABLE_SIZE;
124 /* Subtract 2 iso 1 to avoid access out of range due to rounding */
125 tabscale = (tabsize - 2) / sqrt(nbp->rcoulomb_sq);
127 pmalloc((void**)&ftmp, tabsize*sizeof(*ftmp));
129 table_spline3_fill_ewald_lr(ftmp, NULL, NULL, tabsize,
130 1/tabscale, nbp->ewald_beta, v_q_ewald_lr);
132 /* If the table pointer == NULL the table is generated the first time =>
133 the array pointer will be saved to nbparam and the texture is bound.
135 coul_tab = nbp->coulomb_tab;
136 if (coul_tab == NULL)
138 stat = cudaMalloc((void **)&coul_tab, tabsize*sizeof(*coul_tab));
139 CU_RET_ERR(stat, "cudaMalloc failed on coul_tab");
141 nbp->coulomb_tab = coul_tab;
143 #ifdef TEXOBJ_SUPPORTED
144 /* Only device CC >= 3.0 (Kepler and later) support texture objects */
145 if (dev_info->prop.major >= 3)
148 memset(&rd, 0, sizeof(rd));
149 rd.resType = cudaResourceTypeLinear;
150 rd.res.linear.devPtr = nbp->coulomb_tab;
151 rd.res.linear.desc.f = cudaChannelFormatKindFloat;
152 rd.res.linear.desc.x = 32;
153 rd.res.linear.sizeInBytes = tabsize*sizeof(*coul_tab);
156 memset(&td, 0, sizeof(td));
157 td.readMode = cudaReadModeElementType;
158 stat = cudaCreateTextureObject(&nbp->coulomb_tab_texobj, &rd, &td, NULL);
159 CU_RET_ERR(stat, "cudaCreateTextureObject on coulomb_tab_texobj failed");
164 GMX_UNUSED_VALUE(dev_info);
165 cudaChannelFormatDesc cd = cudaCreateChannelDesc<float>();
166 stat = cudaBindTexture(NULL, &nbnxn_cuda_get_coulomb_tab_texref(),
167 coul_tab, &cd, tabsize*sizeof(*coul_tab));
168 CU_RET_ERR(stat, "cudaBindTexture on coulomb_tab_texref failed");
172 cu_copy_H2D(coul_tab, ftmp, tabsize*sizeof(*coul_tab));
174 nbp->coulomb_tab_size = tabsize;
175 nbp->coulomb_tab_scale = tabscale;
181 /*! Initializes the atomdata structure first time, it only gets filled at
183 static void init_atomdata_first(cu_atomdata_t *ad, int ntypes)
188 stat = cudaMalloc((void**)&ad->shift_vec, SHIFTS*sizeof(*ad->shift_vec));
189 CU_RET_ERR(stat, "cudaMalloc failed on ad->shift_vec");
190 ad->bShiftVecUploaded = false;
192 stat = cudaMalloc((void**)&ad->fshift, SHIFTS*sizeof(*ad->fshift));
193 CU_RET_ERR(stat, "cudaMalloc failed on ad->fshift");
195 stat = cudaMalloc((void**)&ad->e_lj, sizeof(*ad->e_lj));
196 CU_RET_ERR(stat, "cudaMalloc failed on ad->e_lj");
197 stat = cudaMalloc((void**)&ad->e_el, sizeof(*ad->e_el));
198 CU_RET_ERR(stat, "cudaMalloc failed on ad->e_el");
200 /* initialize to NULL poiters to data that is not allocated here and will
201 need reallocation in nbnxn_cuda_init_atomdata */
205 /* size -1 indicates that the respective array hasn't been initialized yet */
210 /*! Selects the Ewald kernel type, analytical on SM 3.0 and later, tabulated on
211 earlier GPUs, single or twin cut-off. */
212 static int pick_ewald_kernel_type(bool bTwinCut,
213 const cuda_dev_info_t *dev_info)
215 bool bUseAnalyticalEwald, bForceAnalyticalEwald, bForceTabulatedEwald;
218 /* Benchmarking/development environment variables to force the use of
219 analytical or tabulated Ewald kernel. */
220 bForceAnalyticalEwald = (getenv("GMX_CUDA_NB_ANA_EWALD") != NULL);
221 bForceTabulatedEwald = (getenv("GMX_CUDA_NB_TAB_EWALD") != NULL);
223 if (bForceAnalyticalEwald && bForceTabulatedEwald)
225 gmx_incons("Both analytical and tabulated Ewald CUDA non-bonded kernels "
226 "requested through environment variables.");
229 /* By default, on SM 3.0 and later use analytical Ewald, on earlier tabulated. */
230 if ((dev_info->prop.major >= 3 || bForceAnalyticalEwald) && !bForceTabulatedEwald)
232 bUseAnalyticalEwald = true;
236 fprintf(debug, "Using analytical Ewald CUDA kernels\n");
241 bUseAnalyticalEwald = false;
245 fprintf(debug, "Using tabulated Ewald CUDA kernels\n");
249 /* Use twin cut-off kernels if requested by bTwinCut or the env. var.
250 forces it (use it for debugging/benchmarking only). */
251 if (!bTwinCut && (getenv("GMX_CUDA_NB_EWALD_TWINCUT") == NULL))
253 kernel_type = bUseAnalyticalEwald ? eelCuEWALD_ANA : eelCuEWALD_TAB;
257 kernel_type = bUseAnalyticalEwald ? eelCuEWALD_ANA_TWIN : eelCuEWALD_TAB_TWIN;
263 /*! Copies all parameters related to the cut-off from ic to nbp */
264 static void set_cutoff_parameters(cu_nbparam_t *nbp,
265 const interaction_const_t *ic)
267 nbp->ewald_beta = ic->ewaldcoeff_q;
268 nbp->sh_ewald = ic->sh_ewald;
269 nbp->epsfac = ic->epsfac;
270 nbp->two_k_rf = 2.0 * ic->k_rf;
271 nbp->c_rf = ic->c_rf;
272 nbp->rvdw_sq = ic->rvdw * ic->rvdw;
273 nbp->rcoulomb_sq = ic->rcoulomb * ic->rcoulomb;
274 nbp->rlist_sq = ic->rlist * ic->rlist;
276 nbp->sh_lj_ewald = ic->sh_lj_ewald;
277 nbp->ewaldcoeff_lj = ic->ewaldcoeff_lj;
279 nbp->rvdw_switch = ic->rvdw_switch;
280 nbp->dispersion_shift = ic->dispersion_shift;
281 nbp->repulsion_shift = ic->repulsion_shift;
282 nbp->vdw_switch = ic->vdw_switch;
285 /*! Initializes the nonbonded parameter data structure. */
286 static void init_nbparam(cu_nbparam_t *nbp,
287 const interaction_const_t *ic,
288 const nbnxn_atomdata_t *nbat,
289 const cuda_dev_info_t *dev_info)
292 int ntypes, nnbfp, nnbfp_comb;
294 ntypes = nbat->ntype;
296 set_cutoff_parameters(nbp, ic);
298 if (ic->vdwtype == evdwCUT)
300 switch (ic->vdw_modifier)
303 case eintmodPOTSHIFT:
304 nbp->vdwtype = evdwCuCUT;
306 case eintmodFORCESWITCH:
307 nbp->vdwtype = evdwCuFSWITCH;
309 case eintmodPOTSWITCH:
310 nbp->vdwtype = evdwCuPSWITCH;
313 gmx_incons("The requested VdW interaction modifier is not implemented in the CUDA GPU accelerated kernels!");
317 else if (ic->vdwtype == evdwPME)
319 if (ic->ljpme_comb_rule == ljcrGEOM)
321 assert(nbat->comb_rule == ljcrGEOM);
322 nbp->vdwtype = evdwCuEWALDGEOM;
326 assert(nbat->comb_rule == ljcrLB);
327 nbp->vdwtype = evdwCuEWALDLB;
332 gmx_incons("The requested VdW type is not implemented in the CUDA GPU accelerated kernels!");
335 if (ic->eeltype == eelCUT)
337 nbp->eeltype = eelCuCUT;
339 else if (EEL_RF(ic->eeltype))
341 nbp->eeltype = eelCuRF;
343 else if ((EEL_PME(ic->eeltype) || ic->eeltype == eelEWALD))
345 /* Initially rcoulomb == rvdw, so it's surely not twin cut-off. */
346 nbp->eeltype = pick_ewald_kernel_type(false, dev_info);
350 /* Shouldn't happen, as this is checked when choosing Verlet-scheme */
351 gmx_incons("The requested electrostatics type is not implemented in the CUDA GPU accelerated kernels!");
354 /* generate table for PME */
355 nbp->coulomb_tab = NULL;
356 if (nbp->eeltype == eelCuEWALD_TAB || nbp->eeltype == eelCuEWALD_TAB_TWIN)
358 init_ewald_coulomb_force_table(nbp, dev_info);
361 nnbfp = 2*ntypes*ntypes;
362 nnbfp_comb = 2*ntypes;
364 stat = cudaMalloc((void **)&nbp->nbfp, nnbfp*sizeof(*nbp->nbfp));
365 CU_RET_ERR(stat, "cudaMalloc failed on nbp->nbfp");
366 cu_copy_H2D(nbp->nbfp, nbat->nbfp, nnbfp*sizeof(*nbp->nbfp));
369 if (ic->vdwtype == evdwPME)
371 stat = cudaMalloc((void **)&nbp->nbfp_comb, nnbfp_comb*sizeof(*nbp->nbfp_comb));
372 CU_RET_ERR(stat, "cudaMalloc failed on nbp->nbfp_comb");
373 cu_copy_H2D(nbp->nbfp_comb, nbat->nbfp_comb, nnbfp_comb*sizeof(*nbp->nbfp_comb));
376 #ifdef TEXOBJ_SUPPORTED
377 /* Only device CC >= 3.0 (Kepler and later) support texture objects */
378 if (dev_info->prop.major >= 3)
383 memset(&rd, 0, sizeof(rd));
384 rd.resType = cudaResourceTypeLinear;
385 rd.res.linear.devPtr = nbp->nbfp;
386 rd.res.linear.desc.f = cudaChannelFormatKindFloat;
387 rd.res.linear.desc.x = 32;
388 rd.res.linear.sizeInBytes = nnbfp*sizeof(*nbp->nbfp);
390 memset(&td, 0, sizeof(td));
391 td.readMode = cudaReadModeElementType;
392 stat = cudaCreateTextureObject(&nbp->nbfp_texobj, &rd, &td, NULL);
393 CU_RET_ERR(stat, "cudaCreateTextureObject on nbfp_texobj failed");
395 if (ic->vdwtype == evdwPME)
397 memset(&rd, 0, sizeof(rd));
398 rd.resType = cudaResourceTypeLinear;
399 rd.res.linear.devPtr = nbp->nbfp_comb;
400 rd.res.linear.desc.f = cudaChannelFormatKindFloat;
401 rd.res.linear.desc.x = 32;
402 rd.res.linear.sizeInBytes = nnbfp_comb*sizeof(*nbp->nbfp_comb);
404 memset(&td, 0, sizeof(td));
405 td.readMode = cudaReadModeElementType;
406 stat = cudaCreateTextureObject(&nbp->nbfp_comb_texobj, &rd, &td, NULL);
407 CU_RET_ERR(stat, "cudaCreateTextureObject on nbfp_comb_texobj failed");
413 cudaChannelFormatDesc cd = cudaCreateChannelDesc<float>();
414 stat = cudaBindTexture(NULL, &nbnxn_cuda_get_nbfp_texref(),
415 nbp->nbfp, &cd, nnbfp*sizeof(*nbp->nbfp));
416 CU_RET_ERR(stat, "cudaBindTexture on nbfp_texref failed");
418 if (ic->vdwtype == evdwPME)
420 stat = cudaBindTexture(NULL, &nbnxn_cuda_get_nbfp_comb_texref(),
421 nbp->nbfp_comb, &cd, nnbfp_comb*sizeof(*nbp->nbfp_comb));
422 CU_RET_ERR(stat, "cudaBindTexture on nbfp_comb_texref failed");
427 /*! Re-generate the GPU Ewald force table, resets rlist, and update the
428 * electrostatic type switching to twin cut-off (or back) if needed. */
429 void nbnxn_cuda_pme_loadbal_update_param(nbnxn_cuda_ptr_t cu_nb,
430 const interaction_const_t *ic)
432 cu_nbparam_t *nbp = cu_nb->nbparam;
434 set_cutoff_parameters(nbp, ic);
436 nbp->eeltype = pick_ewald_kernel_type(ic->rcoulomb != ic->rvdw,
439 init_ewald_coulomb_force_table(cu_nb->nbparam, cu_nb->dev_info);
442 /*! Initializes the pair list data structure. */
443 static void init_plist(cu_plist_t *pl)
445 /* initialize to NULL pointers to data that is not allocated here and will
446 need reallocation in nbnxn_cuda_init_pairlist */
451 /* size -1 indicates that the respective array hasn't been initialized yet */
458 pl->excl_nalloc = -1;
459 pl->bDoPrune = false;
462 /*! Initializes the timer data structure. */
463 static void init_timers(cu_timers_t *t, bool bUseTwoStreams)
466 int eventflags = ( bUseCudaEventBlockingSync ? cudaEventBlockingSync : cudaEventDefault );
468 stat = cudaEventCreateWithFlags(&(t->start_atdat), eventflags);
469 CU_RET_ERR(stat, "cudaEventCreate on start_atdat failed");
470 stat = cudaEventCreateWithFlags(&(t->stop_atdat), eventflags);
471 CU_RET_ERR(stat, "cudaEventCreate on stop_atdat failed");
473 /* The non-local counters/stream (second in the array) are needed only with DD. */
474 for (int i = 0; i <= (bUseTwoStreams ? 1 : 0); i++)
476 stat = cudaEventCreateWithFlags(&(t->start_nb_k[i]), eventflags);
477 CU_RET_ERR(stat, "cudaEventCreate on start_nb_k failed");
478 stat = cudaEventCreateWithFlags(&(t->stop_nb_k[i]), eventflags);
479 CU_RET_ERR(stat, "cudaEventCreate on stop_nb_k failed");
482 stat = cudaEventCreateWithFlags(&(t->start_pl_h2d[i]), eventflags);
483 CU_RET_ERR(stat, "cudaEventCreate on start_pl_h2d failed");
484 stat = cudaEventCreateWithFlags(&(t->stop_pl_h2d[i]), eventflags);
485 CU_RET_ERR(stat, "cudaEventCreate on stop_pl_h2d failed");
487 stat = cudaEventCreateWithFlags(&(t->start_nb_h2d[i]), eventflags);
488 CU_RET_ERR(stat, "cudaEventCreate on start_nb_h2d failed");
489 stat = cudaEventCreateWithFlags(&(t->stop_nb_h2d[i]), eventflags);
490 CU_RET_ERR(stat, "cudaEventCreate on stop_nb_h2d failed");
492 stat = cudaEventCreateWithFlags(&(t->start_nb_d2h[i]), eventflags);
493 CU_RET_ERR(stat, "cudaEventCreate on start_nb_d2h failed");
494 stat = cudaEventCreateWithFlags(&(t->stop_nb_d2h[i]), eventflags);
495 CU_RET_ERR(stat, "cudaEventCreate on stop_nb_d2h failed");
499 /*! Initializes the timings data structure. */
500 static void init_timings(wallclock_gpu_t *t)
509 for (i = 0; i < 2; i++)
511 for (j = 0; j < 2; j++)
513 t->ktime[i][j].t = 0.0;
514 t->ktime[i][j].c = 0;
519 void nbnxn_cuda_init(FILE *fplog,
520 nbnxn_cuda_ptr_t *p_cu_nb,
521 const gmx_gpu_info_t *gpu_info,
522 const gmx_gpu_opt_t *gpu_opt,
524 gmx_bool bLocalAndNonlocal)
529 bool bStreamSync, bNoStreamSync, bTMPIAtomics, bX86, bOldDriver;
541 snew(nb->nbparam, 1);
542 snew(nb->plist[eintLocal], 1);
543 if (bLocalAndNonlocal)
545 snew(nb->plist[eintNonlocal], 1);
548 nb->bUseTwoStreams = bLocalAndNonlocal;
551 snew(nb->timings, 1);
554 pmalloc((void**)&nb->nbst.e_lj, sizeof(*nb->nbst.e_lj));
555 pmalloc((void**)&nb->nbst.e_el, sizeof(*nb->nbst.e_el));
556 pmalloc((void**)&nb->nbst.fshift, SHIFTS * sizeof(*nb->nbst.fshift));
558 init_plist(nb->plist[eintLocal]);
560 /* set device info, just point it to the right GPU among the detected ones */
561 nb->dev_info = &gpu_info->cuda_dev[get_gpu_device_id(gpu_info, gpu_opt, my_gpu_index)];
563 /* local/non-local GPU streams */
564 stat = cudaStreamCreate(&nb->stream[eintLocal]);
565 CU_RET_ERR(stat, "cudaStreamCreate on stream[eintLocal] failed");
566 if (nb->bUseTwoStreams)
568 init_plist(nb->plist[eintNonlocal]);
570 /* CUDA stream priority available in the CUDA RT 5.5 API.
571 * Note that the device we're running on does not have to support
572 * priorities, because we are querying the priority range which in this
573 * case will be a single value.
575 #if CUDA_VERSION >= 5050
577 int highest_priority;
578 stat = cudaDeviceGetStreamPriorityRange(NULL, &highest_priority);
579 CU_RET_ERR(stat, "cudaDeviceGetStreamPriorityRange failed");
581 stat = cudaStreamCreateWithPriority(&nb->stream[eintNonlocal],
584 CU_RET_ERR(stat, "cudaStreamCreateWithPriority on stream[eintNonlocal] failed");
587 stat = cudaStreamCreate(&nb->stream[eintNonlocal]);
588 CU_RET_ERR(stat, "cudaStreamCreate on stream[eintNonlocal] failed");
592 /* init events for sychronization (timing disabled for performance reasons!) */
593 stat = cudaEventCreateWithFlags(&nb->nonlocal_done, cudaEventDisableTiming);
594 CU_RET_ERR(stat, "cudaEventCreate on nonlocal_done failed");
595 stat = cudaEventCreateWithFlags(&nb->misc_ops_done, cudaEventDisableTiming);
596 CU_RET_ERR(stat, "cudaEventCreate on misc_ops_one failed");
598 /* On GPUs with ECC enabled, cudaStreamSynchronize shows a large overhead
599 * (which increases with shorter time/step) caused by a known CUDA driver bug.
600 * To work around the issue we'll use an (admittedly fragile) memory polling
601 * waiting to preserve performance. This requires support for atomic
602 * operations and only works on x86/x86_64.
603 * With polling wait event-timing also needs to be disabled.
605 * The overhead is greatly reduced in API v5.0 drivers and the improvement
606 * is independent of runtime version. Hence, with API v5.0 drivers and later
607 * we won't switch to polling.
609 * NOTE: Unfortunately, this is known to fail when GPUs are shared by (t)MPI,
610 * ranks so we will also disable it in that case.
613 bStreamSync = getenv("GMX_CUDA_STREAMSYNC") != NULL;
614 bNoStreamSync = getenv("GMX_NO_CUDA_STREAMSYNC") != NULL;
619 bTMPIAtomics = false;
622 #ifdef GMX_TARGET_X86
628 if (bStreamSync && bNoStreamSync)
630 gmx_fatal(FARGS, "Conflicting environment variables: both GMX_CUDA_STREAMSYNC and GMX_NO_CUDA_STREAMSYNC defined");
633 stat = cudaDriverGetVersion(&cuda_drv_ver);
634 CU_RET_ERR(stat, "cudaDriverGetVersion failed");
636 bOldDriver = (cuda_drv_ver < 5000);
638 if ((nb->dev_info->prop.ECCEnabled == 1) && bOldDriver)
640 /* Polling wait should be used instead of cudaStreamSynchronize only if:
641 * - ECC is ON & driver is old (checked above),
642 * - we're on x86/x86_64,
643 * - atomics are available, and
644 * - GPUs are not being shared.
646 bool bShouldUsePollSync = (bX86 && bTMPIAtomics &&
647 (gmx_count_gpu_dev_shared(gpu_opt) < 1));
651 nb->bUseStreamSync = true;
653 /* only warn if polling should be used */
654 if (bShouldUsePollSync)
657 "NOTE: Using a GPU with ECC enabled and CUDA driver API version <5.0, but\n"
658 " cudaStreamSynchronize waiting is forced by the GMX_CUDA_STREAMSYNC env. var.\n");
663 nb->bUseStreamSync = !bShouldUsePollSync;
665 if (bShouldUsePollSync)
668 "NOTE: Using a GPU with ECC enabled and CUDA driver API version <5.0, known to\n"
669 " cause performance loss. Switching to the alternative polling GPU wait.\n"
670 " If you encounter issues, switch back to standard GPU waiting by setting\n"
671 " the GMX_CUDA_STREAMSYNC environment variable.\n");
675 /* Tell the user that the ECC+old driver combination can be bad */
677 "NOTE: Using a GPU with ECC enabled and CUDA driver API version <5.0.\n"
678 " A known bug in this driver version can cause performance loss.\n"
679 " However, the polling wait workaround can not be used because\n%s\n"
680 " Consider updating the driver or turning ECC off.",
681 (bX86 && bTMPIAtomics) ?
682 " GPU(s) are being oversubscribed." :
683 " atomic operations are not supported by the platform/CPU+compiler.");
684 md_print_warn(fplog, sbuf);
692 nb->bUseStreamSync = false;
695 "NOTE: Polling wait for GPU synchronization requested by GMX_NO_CUDA_STREAMSYNC\n");
699 /* no/off ECC, cudaStreamSynchronize not turned off by env. var. */
700 nb->bUseStreamSync = true;
704 /* CUDA timing disabled as event timers don't work:
705 - with multiple streams = domain-decomposition;
706 - with the polling waiting hack (without cudaStreamSynchronize);
707 - when turned off by GMX_DISABLE_CUDA_TIMING.
709 nb->bDoTime = (!nb->bUseTwoStreams && nb->bUseStreamSync &&
710 (getenv("GMX_DISABLE_CUDA_TIMING") == NULL));
714 init_timers(nb->timers, nb->bUseTwoStreams);
715 init_timings(nb->timings);
718 /* set the kernel type for the current GPU */
719 /* pick L1 cache configuration */
720 nbnxn_cuda_set_cacheconfig(nb->dev_info);
726 fprintf(debug, "Initialized CUDA data structures.\n");
730 void nbnxn_cuda_init_const(nbnxn_cuda_ptr_t cu_nb,
731 const interaction_const_t *ic,
732 const nonbonded_verlet_group_t *nbv_group)
734 init_atomdata_first(cu_nb->atdat, nbv_group[0].nbat->ntype);
735 init_nbparam(cu_nb->nbparam, ic, nbv_group[0].nbat, cu_nb->dev_info);
737 /* clear energy and shift force outputs */
738 nbnxn_cuda_clear_e_fshift(cu_nb);
741 void nbnxn_cuda_init_pairlist(nbnxn_cuda_ptr_t cu_nb,
742 const nbnxn_pairlist_t *h_plist,
747 bool bDoTime = cu_nb->bDoTime;
748 cudaStream_t stream = cu_nb->stream[iloc];
749 cu_plist_t *d_plist = cu_nb->plist[iloc];
751 if (d_plist->na_c < 0)
753 d_plist->na_c = h_plist->na_ci;
757 if (d_plist->na_c != h_plist->na_ci)
759 sprintf(sbuf, "In cu_init_plist: the #atoms per cell has changed (from %d to %d)",
760 d_plist->na_c, h_plist->na_ci);
767 stat = cudaEventRecord(cu_nb->timers->start_pl_h2d[iloc], stream);
768 CU_RET_ERR(stat, "cudaEventRecord failed");
771 cu_realloc_buffered((void **)&d_plist->sci, h_plist->sci, sizeof(*d_plist->sci),
772 &d_plist->nsci, &d_plist->sci_nalloc,
776 cu_realloc_buffered((void **)&d_plist->cj4, h_plist->cj4, sizeof(*d_plist->cj4),
777 &d_plist->ncj4, &d_plist->cj4_nalloc,
781 cu_realloc_buffered((void **)&d_plist->excl, h_plist->excl, sizeof(*d_plist->excl),
782 &d_plist->nexcl, &d_plist->excl_nalloc,
788 stat = cudaEventRecord(cu_nb->timers->stop_pl_h2d[iloc], stream);
789 CU_RET_ERR(stat, "cudaEventRecord failed");
792 /* need to prune the pair list during the next step */
793 d_plist->bDoPrune = true;
796 void nbnxn_cuda_upload_shiftvec(nbnxn_cuda_ptr_t cu_nb,
797 const nbnxn_atomdata_t *nbatom)
799 cu_atomdata_t *adat = cu_nb->atdat;
800 cudaStream_t ls = cu_nb->stream[eintLocal];
802 /* only if we have a dynamic box */
803 if (nbatom->bDynamicBox || !adat->bShiftVecUploaded)
805 cu_copy_H2D_async(adat->shift_vec, nbatom->shift_vec,
806 SHIFTS * sizeof(*adat->shift_vec), ls);
807 adat->bShiftVecUploaded = true;
811 /*! Clears the first natoms_clear elements of the GPU nonbonded force output array. */
812 static void nbnxn_cuda_clear_f(nbnxn_cuda_ptr_t cu_nb, int natoms_clear)
815 cu_atomdata_t *adat = cu_nb->atdat;
816 cudaStream_t ls = cu_nb->stream[eintLocal];
818 stat = cudaMemsetAsync(adat->f, 0, natoms_clear * sizeof(*adat->f), ls);
819 CU_RET_ERR(stat, "cudaMemsetAsync on f falied");
822 /*! Clears nonbonded shift force output array and energy outputs on the GPU. */
823 static void nbnxn_cuda_clear_e_fshift(nbnxn_cuda_ptr_t cu_nb)
826 cu_atomdata_t *adat = cu_nb->atdat;
827 cudaStream_t ls = cu_nb->stream[eintLocal];
829 stat = cudaMemsetAsync(adat->fshift, 0, SHIFTS * sizeof(*adat->fshift), ls);
830 CU_RET_ERR(stat, "cudaMemsetAsync on fshift falied");
831 stat = cudaMemsetAsync(adat->e_lj, 0, sizeof(*adat->e_lj), ls);
832 CU_RET_ERR(stat, "cudaMemsetAsync on e_lj falied");
833 stat = cudaMemsetAsync(adat->e_el, 0, sizeof(*adat->e_el), ls);
834 CU_RET_ERR(stat, "cudaMemsetAsync on e_el falied");
837 void nbnxn_cuda_clear_outputs(nbnxn_cuda_ptr_t cu_nb, int flags)
839 nbnxn_cuda_clear_f(cu_nb, cu_nb->atdat->natoms);
840 /* clear shift force array and energies if the outputs were
841 used in the current step */
842 if (flags & GMX_FORCE_VIRIAL)
844 nbnxn_cuda_clear_e_fshift(cu_nb);
848 void nbnxn_cuda_init_atomdata(nbnxn_cuda_ptr_t cu_nb,
849 const nbnxn_atomdata_t *nbat)
854 bool bDoTime = cu_nb->bDoTime;
855 cu_timers_t *timers = cu_nb->timers;
856 cu_atomdata_t *d_atdat = cu_nb->atdat;
857 cudaStream_t ls = cu_nb->stream[eintLocal];
859 natoms = nbat->natoms;
864 /* time async copy */
865 stat = cudaEventRecord(timers->start_atdat, ls);
866 CU_RET_ERR(stat, "cudaEventRecord failed");
869 /* need to reallocate if we have to copy more atoms than the amount of space
870 available and only allocate if we haven't initialized yet, i.e d_atdat->natoms == -1 */
871 if (natoms > d_atdat->nalloc)
873 nalloc = over_alloc_small(natoms);
875 /* free up first if the arrays have already been initialized */
876 if (d_atdat->nalloc != -1)
878 cu_free_buffered(d_atdat->f, &d_atdat->natoms, &d_atdat->nalloc);
879 cu_free_buffered(d_atdat->xq);
880 cu_free_buffered(d_atdat->atom_types);
883 stat = cudaMalloc((void **)&d_atdat->f, nalloc*sizeof(*d_atdat->f));
884 CU_RET_ERR(stat, "cudaMalloc failed on d_atdat->f");
885 stat = cudaMalloc((void **)&d_atdat->xq, nalloc*sizeof(*d_atdat->xq));
886 CU_RET_ERR(stat, "cudaMalloc failed on d_atdat->xq");
888 stat = cudaMalloc((void **)&d_atdat->atom_types, nalloc*sizeof(*d_atdat->atom_types));
889 CU_RET_ERR(stat, "cudaMalloc failed on d_atdat->atom_types");
891 d_atdat->nalloc = nalloc;
895 d_atdat->natoms = natoms;
896 d_atdat->natoms_local = nbat->natoms_local;
898 /* need to clear GPU f output if realloc happened */
901 nbnxn_cuda_clear_f(cu_nb, nalloc);
904 cu_copy_H2D_async(d_atdat->atom_types, nbat->type,
905 natoms*sizeof(*d_atdat->atom_types), ls);
909 stat = cudaEventRecord(timers->stop_atdat, ls);
910 CU_RET_ERR(stat, "cudaEventRecord failed");
914 void nbnxn_cuda_free(nbnxn_cuda_ptr_t cu_nb)
917 cu_atomdata_t *atdat;
918 cu_nbparam_t *nbparam;
919 cu_plist_t *plist, *plist_nl;
927 atdat = cu_nb->atdat;
928 nbparam = cu_nb->nbparam;
929 plist = cu_nb->plist[eintLocal];
930 plist_nl = cu_nb->plist[eintNonlocal];
931 timers = cu_nb->timers;
933 if (nbparam->eeltype == eelCuEWALD_TAB || nbparam->eeltype == eelCuEWALD_TAB_TWIN)
936 #ifdef TEXOBJ_SUPPORTED
937 /* Only device CC >= 3.0 (Kepler and later) support texture objects */
938 if (cu_nb->dev_info->prop.major >= 3)
940 stat = cudaDestroyTextureObject(nbparam->coulomb_tab_texobj);
941 CU_RET_ERR(stat, "cudaDestroyTextureObject on coulomb_tab_texobj failed");
946 stat = cudaUnbindTexture(nbnxn_cuda_get_coulomb_tab_texref());
947 CU_RET_ERR(stat, "cudaUnbindTexture on coulomb_tab_texref failed");
949 cu_free_buffered(nbparam->coulomb_tab, &nbparam->coulomb_tab_size);
952 stat = cudaEventDestroy(cu_nb->nonlocal_done);
953 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->nonlocal_done");
954 stat = cudaEventDestroy(cu_nb->misc_ops_done);
955 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->misc_ops_done");
959 stat = cudaEventDestroy(timers->start_atdat);
960 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_atdat");
961 stat = cudaEventDestroy(timers->stop_atdat);
962 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_atdat");
964 /* The non-local counters/stream (second in the array) are needed only with DD. */
965 for (int i = 0; i <= (cu_nb->bUseTwoStreams ? 1 : 0); i++)
967 stat = cudaEventDestroy(timers->start_nb_k[i]);
968 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_nb_k");
969 stat = cudaEventDestroy(timers->stop_nb_k[i]);
970 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_nb_k");
972 stat = cudaEventDestroy(timers->start_pl_h2d[i]);
973 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_pl_h2d");
974 stat = cudaEventDestroy(timers->stop_pl_h2d[i]);
975 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_pl_h2d");
977 stat = cudaStreamDestroy(cu_nb->stream[i]);
978 CU_RET_ERR(stat, "cudaStreamDestroy failed on stream");
980 stat = cudaEventDestroy(timers->start_nb_h2d[i]);
981 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_nb_h2d");
982 stat = cudaEventDestroy(timers->stop_nb_h2d[i]);
983 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_nb_h2d");
985 stat = cudaEventDestroy(timers->start_nb_d2h[i]);
986 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->start_nb_d2h");
987 stat = cudaEventDestroy(timers->stop_nb_d2h[i]);
988 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->stop_nb_d2h");
992 #ifdef TEXOBJ_SUPPORTED
993 /* Only device CC >= 3.0 (Kepler and later) support texture objects */
994 if (cu_nb->dev_info->prop.major >= 3)
996 stat = cudaDestroyTextureObject(nbparam->nbfp_texobj);
997 CU_RET_ERR(stat, "cudaDestroyTextureObject on nbfp_texobj failed");
1002 stat = cudaUnbindTexture(nbnxn_cuda_get_nbfp_texref());
1003 CU_RET_ERR(stat, "cudaUnbindTexture on nbfp_texref failed");
1005 cu_free_buffered(nbparam->nbfp);
1007 if (nbparam->vdwtype == evdwCuEWALDGEOM || nbparam->vdwtype == evdwCuEWALDLB)
1009 #ifdef TEXOBJ_SUPPORTED
1010 /* Only device CC >= 3.0 (Kepler and later) support texture objects */
1011 if (cu_nb->dev_info->prop.major >= 3)
1013 stat = cudaDestroyTextureObject(nbparam->nbfp_comb_texobj);
1014 CU_RET_ERR(stat, "cudaDestroyTextureObject on nbfp_comb_texobj failed");
1019 stat = cudaUnbindTexture(nbnxn_cuda_get_nbfp_comb_texref());
1020 CU_RET_ERR(stat, "cudaUnbindTexture on nbfp_comb_texref failed");
1022 cu_free_buffered(nbparam->nbfp_comb);
1025 stat = cudaFree(atdat->shift_vec);
1026 CU_RET_ERR(stat, "cudaFree failed on atdat->shift_vec");
1027 stat = cudaFree(atdat->fshift);
1028 CU_RET_ERR(stat, "cudaFree failed on atdat->fshift");
1030 stat = cudaFree(atdat->e_lj);
1031 CU_RET_ERR(stat, "cudaFree failed on atdat->e_lj");
1032 stat = cudaFree(atdat->e_el);
1033 CU_RET_ERR(stat, "cudaFree failed on atdat->e_el");
1035 cu_free_buffered(atdat->f, &atdat->natoms, &atdat->nalloc);
1036 cu_free_buffered(atdat->xq);
1037 cu_free_buffered(atdat->atom_types, &atdat->ntypes);
1039 cu_free_buffered(plist->sci, &plist->nsci, &plist->sci_nalloc);
1040 cu_free_buffered(plist->cj4, &plist->ncj4, &plist->cj4_nalloc);
1041 cu_free_buffered(plist->excl, &plist->nexcl, &plist->excl_nalloc);
1042 if (cu_nb->bUseTwoStreams)
1044 cu_free_buffered(plist_nl->sci, &plist_nl->nsci, &plist_nl->sci_nalloc);
1045 cu_free_buffered(plist_nl->cj4, &plist_nl->ncj4, &plist_nl->cj4_nalloc);
1046 cu_free_buffered(plist_nl->excl, &plist_nl->nexcl, &plist->excl_nalloc);
1052 if (cu_nb->bUseTwoStreams)
1057 sfree(cu_nb->timings);
1062 fprintf(debug, "Cleaned up CUDA data structures.\n");
1066 void cu_synchstream_atdat(nbnxn_cuda_ptr_t cu_nb, int iloc)
1069 cudaStream_t stream = cu_nb->stream[iloc];
1071 stat = cudaStreamWaitEvent(stream, cu_nb->timers->stop_atdat, 0);
1072 CU_RET_ERR(stat, "cudaStreamWaitEvent failed");
1075 wallclock_gpu_t * nbnxn_cuda_get_timings(nbnxn_cuda_ptr_t cu_nb)
1077 return (cu_nb != NULL && cu_nb->bDoTime) ? cu_nb->timings : NULL;
1080 void nbnxn_cuda_reset_timings(nbnxn_cuda_ptr_t cu_nb)
1084 init_timings(cu_nb->timings);
1088 int nbnxn_cuda_min_ci_balanced(nbnxn_cuda_ptr_t cu_nb)
1090 return cu_nb != NULL ?
1091 gpu_min_ci_balanced_factor*cu_nb->dev_info->prop.multiProcessorCount : 0;
1095 gmx_bool nbnxn_cuda_is_kernel_ewald_analytical(const nbnxn_cuda_ptr_t cu_nb)
1097 return ((cu_nb->nbparam->eeltype == eelCuEWALD_ANA) ||
1098 (cu_nb->nbparam->eeltype == eelCuEWALD_ANA_TWIN));