2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015, 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.
37 #include "gromacs/legacyheaders/gmx_detect_hardware.h"
53 #ifdef GMX_NATIVE_WINDOWS
57 #include "thread_mpi/threads.h"
59 #include "gromacs/gmxlib/gpu_utils/gpu_utils.h"
60 #include "gromacs/legacyheaders/copyrite.h"
61 #include "gromacs/legacyheaders/gmx_cpuid.h"
62 #include "gromacs/legacyheaders/md_logging.h"
63 #include "gromacs/legacyheaders/network.h"
64 #include "gromacs/legacyheaders/types/commrec.h"
65 #include "gromacs/legacyheaders/types/enums.h"
66 #include "gromacs/legacyheaders/types/hw_info.h"
67 #include "gromacs/utility/arrayref.h"
68 #include "gromacs/utility/basenetwork.h"
69 #include "gromacs/utility/cstringutil.h"
70 #include "gromacs/utility/exceptions.h"
71 #include "gromacs/utility/fatalerror.h"
72 #include "gromacs/utility/gmxassert.h"
73 #include "gromacs/utility/gmxomp.h"
74 #include "gromacs/utility/smalloc.h"
75 #include "gromacs/utility/stringutil.h"
76 #include "gromacs/utility/sysinfo.h"
80 const gmx_bool bGPUBinary = TRUE;
82 const gmx_bool bGPUBinary = FALSE;
85 static const char * invalid_gpuid_hint =
86 "A delimiter-free sequence of valid numeric IDs of available GPUs is expected.";
88 /* The globally shared hwinfo structure. */
89 static gmx_hw_info_t *hwinfo_g;
90 /* A reference counter for the hwinfo structure */
91 static int n_hwinfo = 0;
92 /* A lock to protect the hwinfo structure */
93 static tMPI_Thread_mutex_t hw_info_lock = TMPI_THREAD_MUTEX_INITIALIZER;
97 static void set_gpu_ids(gmx_gpu_opt_t *gpu_opt, int nrank, int rank);
98 static int gmx_count_gpu_dev_unique(const gmx_gpu_info_t *gpu_info,
99 const gmx_gpu_opt_t *gpu_opt);
101 static void sprint_gpus(char *sbuf, const gmx_gpu_info_t *gpu_info)
106 ndev = gpu_info->n_dev;
109 for (i = 0; i < ndev; i++)
111 get_gpu_device_info_string(stmp, gpu_info, i);
121 static void print_gpu_detection_stats(FILE *fplog,
122 const gmx_gpu_info_t *gpu_info,
125 char onhost[266], stmp[STRLEN];
128 if (!gpu_info->bDetectGPUs)
130 /* We skipped the detection, so don't print detection stats */
134 ngpu = gpu_info->n_dev;
136 #if defined GMX_MPI && !defined GMX_THREAD_MPI
137 /* We only print the detection on one, of possibly multiple, nodes */
138 strncpy(onhost, " on host ", 10);
139 gmx_gethostname(onhost+9, 256);
141 /* We detect all relevant GPUs */
142 strncpy(onhost, "", 1);
147 sprint_gpus(stmp, gpu_info);
148 md_print_warn(cr, fplog, "%d GPU%s detected%s:\n%s\n",
149 ngpu, (ngpu > 1) ? "s" : "", onhost, stmp);
153 md_print_warn(cr, fplog, "No GPUs detected%s\n", onhost);
157 /*! \brief Helper function for reporting GPU usage information
158 * in the mdrun log file
160 * \param[in] gpu_info Pointer to per-node GPU info struct
161 * \param[in] gpu_opt Pointer to per-node GPU options struct
162 * \param[in] numPpRanks Number of PP ranks per node
163 * \return String to write to the log file
164 * \throws std::bad_alloc if out of memory */
166 makeGpuUsageReport(const gmx_gpu_info_t *gpu_info,
167 const gmx_gpu_opt_t *gpu_opt,
170 int ngpu_use = gpu_opt->n_dev_use;
171 int ngpu_comp = gpu_info->n_dev_compatible;
173 /* Issue a note if GPUs are available but not used */
174 if (ngpu_comp > 0 && ngpu_use < 1)
176 return gmx::formatString("%d compatible GPU%s detected in the system, but none will be used.\n"
177 "Consider trying GPU acceleration with the Verlet scheme!\n",
178 ngpu_comp, (ngpu_comp > 1) ? "s" : "");
182 if (!gpu_opt->bUserSet)
184 // gpu_opt->dev_compatible is only populated during auto-selection
185 std::string gpuIdsString =
186 formatAndJoin(gmx::constArrayRefFromArray(gpu_opt->dev_compatible,
187 gpu_opt->n_dev_compatible),
188 ",", gmx::StringFormatter("%d"));
189 bool bPluralGpus = gpu_opt->n_dev_compatible > 1;
190 output += gmx::formatString("%d compatible GPU%s %s present, with ID%s %s\n",
191 gpu_opt->n_dev_compatible,
192 bPluralGpus ? "s" : "",
193 bPluralGpus ? "are" : "is",
194 bPluralGpus ? "s" : "",
195 gpuIdsString.c_str());
199 std::vector<int> gpuIdsInUse;
200 for (int i = 0; i < ngpu_use; i++)
202 gpuIdsInUse.push_back(get_cuda_gpu_device_id(gpu_info, gpu_opt, i));
204 std::string gpuIdsString =
205 formatAndJoin(gpuIdsInUse, ",", gmx::StringFormatter("%d"));
206 int numGpusInUse = gmx_count_gpu_dev_unique(gpu_info, gpu_opt);
207 bool bPluralGpus = numGpusInUse > 1;
209 output += gmx::formatString("%d GPU%s %sselected for this run.\n"
210 "Mapping of GPU ID%s to the %d PP rank%s in this node: %s\n",
211 numGpusInUse, bPluralGpus ? "s" : "",
212 gpu_opt->bUserSet ? "user-" : "auto-",
213 bPluralGpus ? "s" : "",
215 (numPpRanks > 1) ? "s" : "",
216 gpuIdsString.c_str());
222 /* Give a suitable fatal error or warning if the build configuration
223 and runtime CPU do not match. */
225 check_use_of_rdtscp_on_this_cpu(FILE *fplog,
227 const gmx_hw_info_t *hwinfo)
229 gmx_bool bCpuHasRdtscp, bBinaryUsesRdtscp;
231 bBinaryUsesRdtscp = TRUE;
233 bBinaryUsesRdtscp = FALSE;
236 bCpuHasRdtscp = gmx_cpuid_feature(hwinfo->cpuid_info, GMX_CPUID_FEATURE_X86_RDTSCP);
238 if (!bCpuHasRdtscp && bBinaryUsesRdtscp)
240 gmx_fatal(FARGS, "The %s executable was compiled to use the rdtscp CPU instruction. "
241 "However, this is not supported by the current hardware and continuing would lead to a crash. "
242 "Please rebuild GROMACS with the GMX_USE_RDTSCP=OFF CMake option.",
246 if (bCpuHasRdtscp && !bBinaryUsesRdtscp)
248 md_print_warn(cr, fplog, "The current CPU can measure timings more accurately than the code in\n"
249 "%s was configured to use. This might affect your simulation\n"
250 "speed as accurate timings are needed for load-balancing.\n"
251 "Please consider rebuilding %s with the GMX_USE_RDTSCP=ON CMake option.\n",
252 ShortProgram(), ShortProgram());
256 void gmx_check_hw_runconf_consistency(FILE *fplog,
257 const gmx_hw_info_t *hwinfo,
259 const gmx_hw_opt_t *hw_opt,
263 char th_or_proc[STRLEN], th_or_proc_plural[STRLEN], pernode[STRLEN];
264 gmx_bool btMPI, bMPI, bMaxMpiThreadsSet, bNthreadsAuto, bEmulateGPU;
269 /* Below we only do consistency checks for PP and GPUs,
270 * this is irrelevant for PME only nodes, so in that case we return
273 if (!(cr->duty & DUTY_PP))
278 #if defined(GMX_THREAD_MPI)
281 bNthreadsAuto = (hw_opt->nthreads_tmpi < 1);
282 #elif defined(GMX_LIB_MPI)
285 bNthreadsAuto = FALSE;
289 bNthreadsAuto = FALSE;
292 /* GPU emulation detection is done later, but we need here as well
293 * -- uncool, but there's no elegant workaround */
294 bEmulateGPU = (getenv("GMX_EMULATE_GPU") != NULL);
295 bMaxMpiThreadsSet = (getenv("GMX_MAX_MPI_THREADS") != NULL);
297 if (hwinfo->gpu_info.n_dev_compatible > 0)
299 std::string gpuUseageReport;
302 gpuUseageReport = makeGpuUsageReport(&hwinfo->gpu_info,
304 cr->nrank_pp_intranode);
306 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
308 /* NOTE: this print is only for and on one physical node */
309 md_print_info(cr, fplog, "%s\n", gpuUseageReport.c_str());
312 /* Need to ensure that we have enough GPUs:
313 * - need one GPU per PP node
314 * - no GPU oversubscription with tMPI
316 /* number of PP processes per node */
317 npppn = cr->nrank_pp_intranode;
320 th_or_proc_plural[0] = '\0';
323 sprintf(th_or_proc, "thread-MPI thread");
326 sprintf(th_or_proc_plural, "s");
331 sprintf(th_or_proc, "MPI process");
334 sprintf(th_or_proc_plural, "es");
336 sprintf(pernode, " per node");
340 /* neither MPI nor tMPI */
341 sprintf(th_or_proc, "process");
344 if (bUseGPU && hwinfo->gpu_info.n_dev_compatible > 0 &&
347 int ngpu_comp, ngpu_use;
348 char gpu_comp_plural[2], gpu_use_plural[2];
350 ngpu_comp = hwinfo->gpu_info.n_dev_compatible;
351 ngpu_use = hw_opt->gpu_opt.n_dev_use;
353 sprintf(gpu_comp_plural, "%s", (ngpu_comp > 1) ? "s" : "");
354 sprintf(gpu_use_plural, "%s", (ngpu_use > 1) ? "s" : "");
356 /* number of tMPI threads auto-adjusted */
357 if (btMPI && bNthreadsAuto)
359 if (hw_opt->gpu_opt.bUserSet && npppn < ngpu_use)
361 /* The user manually provided more GPUs than threads we
362 could automatically start. */
364 "%d GPU%s provided, but only %d PP thread-MPI thread%s coud be started.\n"
365 "%s requires one PP tread-MPI thread per GPU; use fewer GPUs%s.",
366 ngpu_use, gpu_use_plural,
367 npppn, th_or_proc_plural,
368 ShortProgram(), bMaxMpiThreadsSet ? "\nor allow more threads to be used" : "");
371 if (!hw_opt->gpu_opt.bUserSet && npppn < ngpu_comp)
373 /* There are more GPUs than tMPI threads; we have
374 limited the number GPUs used. */
375 md_print_warn(cr, fplog,
376 "NOTE: %d GPU%s were detected, but only %d PP thread-MPI thread%s can be started.\n"
377 " %s can use one GPU per PP tread-MPI thread, so only %d GPU%s will be used.%s\n",
378 ngpu_comp, gpu_comp_plural,
379 npppn, th_or_proc_plural,
380 ShortProgram(), npppn,
381 npppn > 1 ? "s" : "",
382 bMaxMpiThreadsSet ? "\n Also, you can allow more threads to be used by increasing GMX_MAX_MPI_THREADS" : "");
386 if (hw_opt->gpu_opt.bUserSet)
388 if (ngpu_use != npppn)
391 "Incorrect launch configuration: mismatching number of PP %s%s and GPUs%s.\n"
392 "%s was started with %d PP %s%s%s, but you provided %d GPU%s.",
393 th_or_proc, btMPI ? "s" : "es", pernode,
394 ShortProgram(), npppn, th_or_proc,
395 th_or_proc_plural, pernode,
396 ngpu_use, gpu_use_plural);
401 if (ngpu_comp > npppn)
403 md_print_warn(cr, fplog,
404 "NOTE: potentially sub-optimal launch configuration, %s started with less\n"
405 " PP %s%s%s than GPU%s available.\n"
406 " Each PP %s can use only one GPU, %d GPU%s%s will be used.\n",
407 ShortProgram(), th_or_proc,
408 th_or_proc_plural, pernode, gpu_comp_plural,
409 th_or_proc, npppn, gpu_use_plural, pernode);
412 if (ngpu_use != npppn)
414 /* Avoid duplicate error messages.
415 * Unfortunately we can only do this at the physical node
416 * level, since the hardware setup and MPI process count
417 * might differ between physical nodes.
419 if (cr->rank_pp_intranode == 0)
422 "Incorrect launch configuration: mismatching number of PP %s%s and GPUs%s.\n"
423 "%s was started with %d PP %s%s%s, but only %d GPU%s were detected.",
424 th_or_proc, btMPI ? "s" : "es", pernode,
425 ShortProgram(), npppn, th_or_proc,
426 th_or_proc_plural, pernode,
427 ngpu_use, gpu_use_plural);
435 same_count = gmx_count_gpu_dev_shared(&hw_opt->gpu_opt);
439 md_print_info(cr, fplog,
440 "NOTE: You assigned %s to multiple %s%s.\n",
441 same_count > 1 ? "GPUs" : "a GPU", th_or_proc, btMPI ? "s" : "es");
449 /* Avoid other ranks to continue after
451 MPI_Barrier(cr->mpi_comm_mygroup);
457 /* Return 0 if none of the GPU (per node) are shared among PP ranks.
459 * Sharing GPUs among multiple PP ranks is possible when the user passes
460 * GPU IDs. Here we check for sharing and return a non-zero value when
461 * this is detected. Note that the return value represents the number of
462 * PP rank pairs that share a device.
464 int gmx_count_gpu_dev_shared(const gmx_gpu_opt_t *gpu_opt)
467 int ngpu = gpu_opt->n_dev_use;
469 if (gpu_opt->bUserSet)
473 for (i = 0; i < ngpu - 1; i++)
475 for (j = i + 1; j < ngpu; j++)
477 same_count += (gpu_opt->dev_use[i] ==
478 gpu_opt->dev_use[j]);
486 /* Count and return the number of unique GPUs (per node) selected.
488 * As sharing GPUs among multiple PP ranks is possible when the user passes
489 * GPU IDs, the number of GPUs user (per node) can be different from the
490 * number of GPU IDs selected.
492 static int gmx_count_gpu_dev_unique(const gmx_gpu_info_t *gpu_info,
493 const gmx_gpu_opt_t *gpu_opt)
495 int i, uniq_count, ngpu;
501 ngpu = gpu_info->n_dev;
505 snew(uniq_ids, ngpu);
507 /* Each element in uniq_ids will be set to 0 or 1. The n-th element set
508 * to 1 indicates that the respective GPU was selected to be used. */
509 for (i = 0; i < gpu_opt->n_dev_use; i++)
511 uniq_ids[get_cuda_gpu_device_id(gpu_info, gpu_opt, i)] = 1;
513 /* Count the devices used. */
514 for (i = 0; i < ngpu; i++)
516 uniq_count += uniq_ids[i];
524 static int get_ncores(gmx_cpuid_t cpuid)
526 int nprocessors, npackages, ncores_per_package, nhwthreads_per_core;
527 const int *package_id, *core_id, *hwthread_id, *locality_order;
530 rc = gmx_cpuid_topology(cpuid,
531 &nprocessors, &npackages,
532 &ncores_per_package, &nhwthreads_per_core,
533 &package_id, &core_id,
534 &hwthread_id, &locality_order);
538 return npackages*ncores_per_package;
542 /* We don't have cpuid topology info, return 0 core count */
547 /* Return the number of hardware threads supported by the current CPU.
548 * We assume that this is equal with the number of "processors"
549 * reported to be online by the OS at the time of the call. The
550 * definition of "processor" is according to an old POSIX standard.
552 * Note that the number of hardware threads is generally greater than
553 * the number of cores (e.g. x86 hyper-threading, Power). Managing the
554 * mapping of software threads to hardware threads is managed
556 static int get_nthreads_hw_avail(FILE gmx_unused *fplog, const t_commrec gmx_unused *cr)
560 #if ((defined(WIN32) || defined( _WIN32 ) || defined(WIN64) || defined( _WIN64 )) && !(defined (__CYGWIN__) || defined (__CYGWIN32__)))
563 GetSystemInfo( &sysinfo );
564 ret = sysinfo.dwNumberOfProcessors;
565 #elif defined HAVE_SYSCONF
566 /* We are probably on Unix.
567 * Now check if we have the argument to use before executing the call
569 #if defined(_SC_NPROCESSORS_ONLN)
570 ret = sysconf(_SC_NPROCESSORS_ONLN);
571 #elif defined(_SC_NPROC_ONLN)
572 ret = sysconf(_SC_NPROC_ONLN);
573 #elif defined(_SC_NPROCESSORS_CONF)
574 ret = sysconf(_SC_NPROCESSORS_CONF);
575 #elif defined(_SC_NPROC_CONF)
576 ret = sysconf(_SC_NPROC_CONF);
578 #warning "No valid sysconf argument value found. Executables will not be able to determine the number of logical cores: mdrun will use 1 thread by default!"
579 #endif /* End of check for sysconf argument values */
582 /* Neither windows nor Unix. No fscking idea how many hardware threads we have! */
588 fprintf(debug, "Detected %d hardware threads to use.\n", ret);
592 if (ret != gmx_omp_get_num_procs())
594 md_print_warn(cr, fplog,
595 "Number of logical cores detected (%d) does not match the number reported by OpenMP (%d).\n"
596 "Consider setting the launch configuration manually!",
597 ret, gmx_omp_get_num_procs());
604 static void gmx_detect_gpus(FILE *fplog, const t_commrec *cr)
608 MPI_Comm physicalnode_comm;
612 /* Under certain circumstances MPI ranks on the same physical node
613 * can not simultaneously access the same GPU(s). Therefore we run
614 * the detection only on one MPI rank per node and broadcast the info.
615 * Note that with thread-MPI only a single thread runs this code.
617 * TODO: We should also do CPU hardware detection only once on each
618 * physical node and broadcast it, instead of do it on every MPI rank.
621 /* A split of MPI_COMM_WORLD over physical nodes is only required here,
622 * so we create and destroy it locally.
624 MPI_Comm_rank(MPI_COMM_WORLD, &rank_world);
625 MPI_Comm_split(MPI_COMM_WORLD, gmx_physicalnode_id_hash(),
626 rank_world, &physicalnode_comm);
627 MPI_Comm_rank(physicalnode_comm, &rank_local);
629 /* Here there should be only one process, check this */
630 assert(cr->nnodes == 1 && cr->sim_nodeid == 0);
637 char detection_error[STRLEN] = "", sbuf[STRLEN];
639 if (detect_gpus(&hwinfo_g->gpu_info, detection_error) != 0)
641 if (detection_error[0] != '\0')
643 sprintf(sbuf, ":\n %s\n", detection_error);
649 md_print_warn(cr, fplog,
650 "NOTE: Error occurred during GPU detection%s"
651 " Can not use GPU acceleration, will fall back to CPU kernels.\n",
657 /* Broadcast the GPU info to the other ranks within this node */
658 MPI_Bcast(&hwinfo_g->gpu_info.n_dev, 1, MPI_INT, 0, physicalnode_comm);
660 if (hwinfo_g->gpu_info.n_dev > 0)
664 dev_size = hwinfo_g->gpu_info.n_dev*sizeof_gpu_dev_info();
668 hwinfo_g->gpu_info.gpu_dev =
669 (struct gmx_device_info_t *)malloc(dev_size);
671 MPI_Bcast(hwinfo_g->gpu_info.gpu_dev, dev_size, MPI_BYTE,
672 0, physicalnode_comm);
673 MPI_Bcast(&hwinfo_g->gpu_info.n_dev_compatible, 1, MPI_INT,
674 0, physicalnode_comm);
677 MPI_Comm_free(&physicalnode_comm);
681 static void gmx_collect_hardware_mpi()
685 int nrank, rank, ncore, nhwthread, ngpu, i;
689 rank_id = gmx_physicalnode_id_hash();
690 MPI_Comm_rank(MPI_COMM_WORLD, &rank);
691 MPI_Comm_size(MPI_COMM_WORLD, &nrank);
692 ncore = hwinfo_g->ncore;
693 nhwthread = hwinfo_g->nthreads_hw_avail;
694 ngpu = hwinfo_g->gpu_info.n_dev_compatible;
695 /* Create a unique hash of the GPU type(s) in this node */
697 /* Here it might be better to only loop over the compatible GPU, but we
698 * don't have that information available and it would also require
699 * removing the device ID from the device info string.
701 for (i = 0; i < hwinfo_g->gpu_info.n_dev; i++)
705 /* Since the device ID is incorporated in the hash, the order of
706 * the GPUs affects the hash. Also two identical GPUs won't give
707 * a gpu_hash of zero after XORing.
709 get_gpu_device_info_string(stmp, &hwinfo_g->gpu_info, i);
710 gpu_hash ^= gmx_string_fullhash_func(stmp, gmx_string_hash_init);
717 MPI_Allreduce(buf, all, nrank, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
720 int nnode0, ncore0, nhwthread0, ngpu0, r;
727 for (r = 0; r < nrank; r++)
729 if (all[r] == rank_id)
731 if (!bFound && r == rank)
733 /* We are the first rank in this physical node */
736 nhwthread0 = nhwthread;
746 int sum[4], maxmin[10];
751 /* Sum values from only intra-rank 0 so we get the sum over all nodes */
757 MPI_Allreduce(buf, sum, 4, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
763 /* Store + and - values for all ranks,
764 * so we can get max+min with one MPI call.
769 buf[3] = gmx_cpuid_simd_suggest(hwinfo_g->cpuid_info);
777 MPI_Allreduce(buf, maxmin, 10, MPI_INT, MPI_MAX, MPI_COMM_WORLD);
780 hwinfo_g->nphysicalnode = sum[0];
781 hwinfo_g->ncore_tot = sum[1];
782 hwinfo_g->ncore_min = -maxmin[5];
783 hwinfo_g->ncore_max = maxmin[0];
784 hwinfo_g->nhwthread_tot = sum[2];
785 hwinfo_g->nhwthread_min = -maxmin[6];
786 hwinfo_g->nhwthread_max = maxmin[1];
787 hwinfo_g->ngpu_compatible_tot = sum[3];
788 hwinfo_g->ngpu_compatible_min = -maxmin[7];
789 hwinfo_g->ngpu_compatible_max = maxmin[2];
790 hwinfo_g->simd_suggest_min = static_cast<enum gmx_cpuid_simd>(-maxmin[8]);
791 hwinfo_g->simd_suggest_max = static_cast<enum gmx_cpuid_simd>(maxmin[3]);
792 hwinfo_g->bIdenticalGPUs = (maxmin[4] == -maxmin[9]);
794 /* All ranks use the same pointer, protect it with a mutex */
795 tMPI_Thread_mutex_lock(&hw_info_lock);
796 hwinfo_g->nphysicalnode = 1;
797 hwinfo_g->ncore_tot = hwinfo_g->ncore;
798 hwinfo_g->ncore_min = hwinfo_g->ncore;
799 hwinfo_g->ncore_max = hwinfo_g->ncore;
800 hwinfo_g->nhwthread_tot = hwinfo_g->nthreads_hw_avail;
801 hwinfo_g->nhwthread_min = hwinfo_g->nthreads_hw_avail;
802 hwinfo_g->nhwthread_max = hwinfo_g->nthreads_hw_avail;
803 hwinfo_g->ngpu_compatible_tot = hwinfo_g->gpu_info.n_dev_compatible;
804 hwinfo_g->ngpu_compatible_min = hwinfo_g->gpu_info.n_dev_compatible;
805 hwinfo_g->ngpu_compatible_max = hwinfo_g->gpu_info.n_dev_compatible;
806 hwinfo_g->simd_suggest_min = gmx_cpuid_simd_suggest(hwinfo_g->cpuid_info);
807 hwinfo_g->simd_suggest_max = gmx_cpuid_simd_suggest(hwinfo_g->cpuid_info);
808 hwinfo_g->bIdenticalGPUs = TRUE;
809 tMPI_Thread_mutex_unlock(&hw_info_lock);
813 gmx_hw_info_t *gmx_detect_hardware(FILE *fplog, const t_commrec *cr,
814 gmx_bool bDetectGPUs)
818 /* make sure no one else is doing the same thing */
819 ret = tMPI_Thread_mutex_lock(&hw_info_lock);
822 gmx_fatal(FARGS, "Error locking hwinfo mutex: %s", strerror(errno));
825 /* only initialize the hwinfo structure if it is not already initalized */
830 /* detect CPUID info; no fuss, we don't detect system-wide
831 * -- sloppy, but that's it for now */
832 if (gmx_cpuid_init(&hwinfo_g->cpuid_info) != 0)
834 gmx_fatal_collective(FARGS, cr, NULL, "CPUID detection failed!");
837 /* get the number of cores, will be 0 when not detected */
838 hwinfo_g->ncore = get_ncores(hwinfo_g->cpuid_info);
840 /* detect number of hardware threads */
841 hwinfo_g->nthreads_hw_avail = get_nthreads_hw_avail(fplog, cr);
844 hwinfo_g->gpu_info.n_dev = 0;
845 hwinfo_g->gpu_info.n_dev_compatible = 0;
846 hwinfo_g->gpu_info.gpu_dev = NULL;
848 /* Run the detection if the binary was compiled with GPU support
849 * and we requested detection.
851 hwinfo_g->gpu_info.bDetectGPUs =
852 (bGPUBinary && bDetectGPUs &&
853 getenv("GMX_DISABLE_GPU_DETECTION") == NULL);
854 if (hwinfo_g->gpu_info.bDetectGPUs)
856 gmx_detect_gpus(fplog, cr);
859 /* increase the reference counter */
862 ret = tMPI_Thread_mutex_unlock(&hw_info_lock);
865 gmx_fatal(FARGS, "Error unlocking hwinfo mutex: %s", strerror(errno));
868 gmx_collect_hardware_mpi();
873 static std::string detected_hardware_string(const gmx_hw_info_t *hwinfo,
878 s = gmx::formatString("\n");
879 s += gmx::formatString("Running on %d node%s with total",
880 hwinfo->nphysicalnode,
881 hwinfo->nphysicalnode == 1 ? "" : "s");
882 if (hwinfo->ncore_tot > 0)
884 s += gmx::formatString(" %d cores,", hwinfo->ncore_tot);
886 s += gmx::formatString(" %d logical cores", hwinfo->nhwthread_tot);
887 if (hwinfo->gpu_info.bDetectGPUs)
889 s += gmx::formatString(", %d compatible GPU%s",
890 hwinfo->ngpu_compatible_tot,
891 hwinfo->ngpu_compatible_tot == 1 ? "" : "s");
895 s += gmx::formatString(" (GPU detection deactivated)");
897 s += gmx::formatString("\n");
899 if (hwinfo->nphysicalnode > 1)
901 /* Print per node hardware feature counts */
902 if (hwinfo->ncore_max > 0)
904 s += gmx::formatString(" Cores per node: %2d", hwinfo->ncore_min);
905 if (hwinfo->ncore_max > hwinfo->ncore_min)
907 s += gmx::formatString(" - %2d", hwinfo->ncore_max);
909 s += gmx::formatString("\n");
911 s += gmx::formatString(" Logical cores per node: %2d", hwinfo->nhwthread_min);
912 if (hwinfo->nhwthread_max > hwinfo->nhwthread_min)
914 s += gmx::formatString(" - %2d", hwinfo->nhwthread_max);
916 s += gmx::formatString("\n");
919 s += gmx::formatString(" Compatible GPUs per node: %2d",
920 hwinfo->ngpu_compatible_min);
921 if (hwinfo->ngpu_compatible_max > hwinfo->ngpu_compatible_min)
923 s += gmx::formatString(" - %2d", hwinfo->ngpu_compatible_max);
925 s += gmx::formatString("\n");
926 if (hwinfo->ngpu_compatible_tot > 0)
928 if (hwinfo->bIdenticalGPUs)
930 s += gmx::formatString(" All nodes have identical type(s) of GPUs\n");
934 /* This message will also appear with identical GPU types
935 * when at least one node has no GPU.
937 s += gmx::formatString(" Different nodes have different type(s) and/or order of GPUs\n");
947 gmx_gethostname(host, 255);
948 MPI_Comm_rank(MPI_COMM_WORLD, &rank);
950 s += gmx::formatString("Hardware detected on host %s (the node of MPI rank %d):\n",
953 s += gmx::formatString("Hardware detected:\n");
955 s += gmx::formatString(" CPU info:\n");
960 gmx_cpuid_formatstring(hwinfo->cpuid_info, buf, 1023);
963 s += gmx::formatString("%s", buf);
967 s += gmx::formatString(" Vendor: %s\n",
968 gmx_cpuid_vendor_string[gmx_cpuid_vendor(hwinfo->cpuid_info)]);
969 s += gmx::formatString(" Brand: %s\n",
970 gmx_cpuid_brand(hwinfo->cpuid_info));
972 s += gmx::formatString(" SIMD instructions most likely to fit this hardware: %s",
973 gmx_cpuid_simd_string[hwinfo->simd_suggest_min]);
974 if (hwinfo->simd_suggest_max > hwinfo->simd_suggest_min)
976 s += gmx::formatString(" - %s",
977 gmx_cpuid_simd_string[hwinfo->simd_suggest_max]);
979 s += gmx::formatString("\n");
980 s += gmx::formatString(" SIMD instructions selected at GROMACS compile time: %s\n",
981 gmx_cpuid_simd_string[gmx_compiled_simd()]);
982 if (bGPUBinary && (hwinfo->ngpu_compatible_tot > 0 ||
983 hwinfo->gpu_info.n_dev > 0))
985 s += gmx::formatString(" GPU info:\n");
986 s += gmx::formatString(" Number of GPUs detected: %d\n",
987 hwinfo->gpu_info.n_dev);
988 if (hwinfo->gpu_info.n_dev > 0)
992 sprint_gpus(buf, &hwinfo->gpu_info);
993 s += gmx::formatString("%s\n", buf);
1000 void gmx_print_detected_hardware(FILE *fplog, const t_commrec *cr,
1001 const gmx_hw_info_t *hwinfo)
1005 std::string detected;
1007 detected = detected_hardware_string(hwinfo, TRUE);
1009 fprintf(fplog, "%s\n", detected.c_str());
1012 if (MULTIMASTER(cr))
1014 std::string detected;
1016 detected = detected_hardware_string(hwinfo, FALSE);
1018 fprintf(stderr, "%s\n", detected.c_str());
1021 /* Check the compiled SIMD instruction set against that of the node
1022 * with the lowest SIMD level support.
1024 gmx_cpuid_simd_check(hwinfo->simd_suggest_min, fplog, MULTIMASTER(cr));
1026 /* For RDTSCP we only check on our local node and skip the MPI reduction */
1027 check_use_of_rdtscp_on_this_cpu(fplog, cr, hwinfo);
1030 void gmx_parse_gpu_ids(gmx_gpu_opt_t *gpu_opt)
1034 if (gpu_opt->gpu_id != NULL && !bGPUBinary)
1036 gmx_fatal(FARGS, "GPU ID string set, but %s was compiled without GPU support!", ShortProgram());
1039 env = getenv("GMX_GPU_ID");
1040 if (env != NULL && gpu_opt->gpu_id != NULL)
1042 gmx_fatal(FARGS, "GMX_GPU_ID and -gpu_id can not be used at the same time");
1046 env = gpu_opt->gpu_id;
1049 /* parse GPU IDs if the user passed any */
1052 /* Parse a "plain" GPU ID string which contains a sequence of
1053 * digits corresponding to GPU IDs; the order will indicate
1054 * the process/tMPI thread - GPU assignment. */
1055 parse_digits_from_plain_string(env,
1056 &gpu_opt->n_dev_use,
1059 if (gpu_opt->n_dev_use == 0)
1061 gmx_fatal(FARGS, "Empty GPU ID string encountered.\n%s\n",
1062 invalid_gpuid_hint);
1065 gpu_opt->bUserSet = TRUE;
1069 void gmx_select_gpu_ids(FILE *fplog, const t_commrec *cr,
1070 const gmx_gpu_info_t *gpu_info,
1071 gmx_bool bForceUseGPU,
1072 gmx_gpu_opt_t *gpu_opt)
1075 char sbuf[STRLEN], stmp[STRLEN];
1077 /* Bail if binary is not compiled with GPU acceleration, but this is either
1078 * explicitly (-nb gpu) or implicitly (gpu ID passed) requested. */
1079 if (bForceUseGPU && !bGPUBinary)
1081 gmx_fatal(FARGS, "GPU acceleration requested, but %s was compiled without GPU support!", ShortProgram());
1084 if (!(cr->duty & DUTY_PP))
1086 /* Our rank is not doing PP, we don't use a GPU */
1090 if (gpu_opt->bUserSet)
1092 /* Check the GPU IDs passed by the user.
1093 * (GPU IDs have been parsed by gmx_parse_gpu_ids before)
1098 snew(checkres, gpu_opt->n_dev_use);
1100 res = check_selected_gpus(checkres, gpu_info, gpu_opt);
1104 print_gpu_detection_stats(fplog, gpu_info, cr);
1106 sprintf(sbuf, "Some of the requested GPUs do not exist, behave strangely, or are not compatible:\n");
1107 for (i = 0; i < gpu_opt->n_dev_use; i++)
1109 if (checkres[i] != egpuCompatible)
1111 sprintf(stmp, " GPU #%d: %s\n",
1112 gpu_opt->dev_use[i],
1113 gpu_detect_res_str[checkres[i]]);
1117 gmx_fatal(FARGS, "%s", sbuf);
1124 pick_compatible_gpus(&hwinfo_g->gpu_info, gpu_opt);
1125 set_gpu_ids(gpu_opt, cr->nrank_pp_intranode, cr->rank_pp_intranode);
1128 /* If the user asked for a GPU, check whether we have a GPU */
1129 if (bForceUseGPU && gpu_info->n_dev_compatible == 0)
1131 gmx_fatal(FARGS, "GPU acceleration requested, but no compatible GPUs were detected.");
1135 /* Select the GPUs we will use. This is an operation local to each physical
1136 * node. If we have less MPI ranks than GPUs, we will waste some GPUs.
1137 * nrank and rank are the rank count and id for PP processes in our node.
1139 static void set_gpu_ids(gmx_gpu_opt_t *gpu_opt, int nrank, int rank)
1141 GMX_RELEASE_ASSERT(gpu_opt, "Invalid gpu_opt pointer passed");
1142 GMX_RELEASE_ASSERT(nrank >= 1,
1143 gmx::formatString("Invalid limit (%d) for the number of GPUs (detected %d compatible GPUs)",
1144 rank, gpu_opt->n_dev_compatible).c_str());
1146 if (gpu_opt->n_dev_compatible == 0)
1150 gmx_gethostname(host, 255);
1151 gmx_fatal(FARGS, "A GPU was requested on host %s, but no compatible GPUs were detected. All nodes with PP ranks need to have GPUs. If you intended to use GPU acceleration in a parallel run, you can either avoid using the nodes that don't have GPUs or place PME ranks on these nodes.", host);
1157 if (nrank > gpu_opt->n_dev_compatible)
1159 if (nrank % gpu_opt->n_dev_compatible == 0)
1161 nshare = nrank/gpu_opt->n_dev_compatible;
1167 gmx_fatal(FARGS, "The number of MPI ranks (%d) in a physical node is not a multiple of the number of GPUs (%d). Select a different number of MPI ranks or use the -gpu_id option to manually specify the GPU to be used.",
1168 nrank, gpu_opt->n_dev_compatible);
1172 /* We use a global barrier to prevent ranks from continuing with
1175 MPI_Barrier(MPI_COMM_WORLD);
1180 /* Here we will waste GPUs when nrank < gpu_opt->n_dev_compatible */
1181 gpu_opt->n_dev_use = std::min(gpu_opt->n_dev_compatible*nshare, nrank);
1182 snew(gpu_opt->dev_use, gpu_opt->n_dev_use);
1183 for (int i = 0; i != gpu_opt->n_dev_use; ++i)
1185 /* TODO: improve this implementation: either sort GPUs or remove the weakest here */
1186 gpu_opt->dev_use[i] = gpu_opt->dev_compatible[i/nshare];
1190 void gmx_hardware_info_free(gmx_hw_info_t *hwinfo)
1194 ret = tMPI_Thread_mutex_lock(&hw_info_lock);
1197 gmx_fatal(FARGS, "Error locking hwinfo mutex: %s", strerror(errno));
1200 /* decrease the reference counter */
1204 if (hwinfo != hwinfo_g)
1206 gmx_incons("hwinfo < hwinfo_g");
1211 gmx_incons("n_hwinfo < 0");
1216 gmx_cpuid_done(hwinfo_g->cpuid_info);
1217 free_gpu_info(&hwinfo_g->gpu_info);
1221 ret = tMPI_Thread_mutex_unlock(&hw_info_lock);
1224 gmx_fatal(FARGS, "Error unlocking hwinfo mutex: %s", strerror(errno));