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;
95 #define HOSTNAMELEN 80
98 static void set_gpu_ids(gmx_gpu_opt_t *gpu_opt, int nrank, int rank);
99 static int gmx_count_gpu_dev_unique(const gmx_gpu_info_t *gpu_info,
100 const gmx_gpu_opt_t *gpu_opt);
102 static void sprint_gpus(char *sbuf, const gmx_gpu_info_t *gpu_info)
107 ndev = gpu_info->n_dev;
110 for (i = 0; i < ndev; i++)
112 get_gpu_device_info_string(stmp, gpu_info, i);
122 static void print_gpu_detection_stats(FILE *fplog,
123 const gmx_gpu_info_t *gpu_info,
126 char onhost[HOSTNAMELEN+10], stmp[STRLEN];
129 if (!gpu_info->bDetectGPUs)
131 /* We skipped the detection, so don't print detection stats */
135 ngpu = gpu_info->n_dev;
137 #if defined GMX_MPI && !defined GMX_THREAD_MPI
138 /* We only print the detection on one, of possibly multiple, nodes */
139 strncpy(onhost, " on host ", 10);
140 gmx_gethostname(onhost + 9, HOSTNAMELEN);
142 /* We detect all relevant GPUs */
143 strncpy(onhost, "", 1);
148 sprint_gpus(stmp, gpu_info);
149 md_print_warn(cr, fplog, "%d GPU%s detected%s:\n%s\n",
150 ngpu, (ngpu > 1) ? "s" : "", onhost, stmp);
154 md_print_warn(cr, fplog, "No GPUs detected%s\n", onhost);
158 /*! \brief Helper function for reporting GPU usage information
159 * in the mdrun log file
161 * \param[in] gpu_info Pointer to per-node GPU info struct
162 * \param[in] gpu_opt Pointer to per-node GPU options struct
163 * \param[in] numPpRanks Number of PP ranks per node
164 * \param[in] bPrintHostName Print the hostname in the usage information
165 * \return String to write to the log file
166 * \throws std::bad_alloc if out of memory */
168 makeGpuUsageReport(const gmx_gpu_info_t *gpu_info,
169 const gmx_gpu_opt_t *gpu_opt,
173 int ngpu_use = gpu_opt->n_dev_use;
174 int ngpu_comp = gpu_info->n_dev_compatible;
175 char host[HOSTNAMELEN];
179 gmx_gethostname(host, HOSTNAMELEN);
182 /* Issue a note if GPUs are available but not used */
183 if (ngpu_comp > 0 && ngpu_use < 1)
185 return gmx::formatString("%d compatible GPU%s detected in the system, but none will be used.\n"
186 "Consider trying GPU acceleration with the Verlet scheme!\n",
187 ngpu_comp, (ngpu_comp > 1) ? "s" : "");
191 if (!gpu_opt->bUserSet)
193 // gpu_opt->dev_compatible is only populated during auto-selection
194 std::string gpuIdsString =
195 formatAndJoin(gmx::constArrayRefFromArray(gpu_opt->dev_compatible,
196 gpu_opt->n_dev_compatible),
197 ",", gmx::StringFormatter("%d"));
198 bool bPluralGpus = gpu_opt->n_dev_compatible > 1;
202 output += gmx::formatString("On host %s ", host);
204 output += gmx::formatString("%d compatible GPU%s %s present, with ID%s %s\n",
205 gpu_opt->n_dev_compatible,
206 bPluralGpus ? "s" : "",
207 bPluralGpus ? "are" : "is",
208 bPluralGpus ? "s" : "",
209 gpuIdsString.c_str());
213 std::vector<int> gpuIdsInUse;
214 for (int i = 0; i < ngpu_use; i++)
216 gpuIdsInUse.push_back(get_cuda_gpu_device_id(gpu_info, gpu_opt, i));
218 std::string gpuIdsString =
219 formatAndJoin(gpuIdsInUse, ",", gmx::StringFormatter("%d"));
220 int numGpusInUse = gmx_count_gpu_dev_unique(gpu_info, gpu_opt);
221 bool bPluralGpus = numGpusInUse > 1;
225 output += gmx::formatString("On host %s ", host);
227 output += gmx::formatString("%d GPU%s %sselected for this run.\n"
228 "Mapping of GPU ID%s to the %d PP rank%s in this node: %s\n",
229 numGpusInUse, bPluralGpus ? "s" : "",
230 gpu_opt->bUserSet ? "user-" : "auto-",
231 bPluralGpus ? "s" : "",
233 (numPpRanks > 1) ? "s" : "",
234 gpuIdsString.c_str());
240 /* Give a suitable fatal error or warning if the build configuration
241 and runtime CPU do not match. */
243 check_use_of_rdtscp_on_this_cpu(FILE *fplog,
245 const gmx_hw_info_t *hwinfo)
247 gmx_bool bCpuHasRdtscp, bBinaryUsesRdtscp;
249 bBinaryUsesRdtscp = TRUE;
251 bBinaryUsesRdtscp = FALSE;
254 bCpuHasRdtscp = gmx_cpuid_feature(hwinfo->cpuid_info, GMX_CPUID_FEATURE_X86_RDTSCP);
256 if (!bCpuHasRdtscp && bBinaryUsesRdtscp)
258 gmx_fatal(FARGS, "The %s executable was compiled to use the rdtscp CPU instruction. "
259 "However, this is not supported by the current hardware and continuing would lead to a crash. "
260 "Please rebuild GROMACS with the GMX_USE_RDTSCP=OFF CMake option.",
264 if (bCpuHasRdtscp && !bBinaryUsesRdtscp)
266 md_print_warn(cr, fplog, "The current CPU can measure timings more accurately than the code in\n"
267 "%s was configured to use. This might affect your simulation\n"
268 "speed as accurate timings are needed for load-balancing.\n"
269 "Please consider rebuilding %s with the GMX_USE_RDTSCP=ON CMake option.\n",
270 ShortProgram(), ShortProgram());
274 void gmx_check_hw_runconf_consistency(FILE *fplog,
275 const gmx_hw_info_t *hwinfo,
277 const gmx_hw_opt_t *hw_opt,
281 char th_or_proc[STRLEN], th_or_proc_plural[STRLEN], pernode[STRLEN];
282 gmx_bool btMPI, bMPI, bMaxMpiThreadsSet, bNthreadsAuto, bEmulateGPU;
287 /* Below we only do consistency checks for PP and GPUs,
288 * this is irrelevant for PME only nodes, so in that case we return
291 if (!(cr->duty & DUTY_PP))
296 #if defined(GMX_THREAD_MPI)
299 bNthreadsAuto = (hw_opt->nthreads_tmpi < 1);
300 #elif defined(GMX_LIB_MPI)
303 bNthreadsAuto = FALSE;
307 bNthreadsAuto = FALSE;
310 /* GPU emulation detection is done later, but we need here as well
311 * -- uncool, but there's no elegant workaround */
312 bEmulateGPU = (getenv("GMX_EMULATE_GPU") != NULL);
313 bMaxMpiThreadsSet = (getenv("GMX_MAX_MPI_THREADS") != NULL);
315 if (hwinfo->gpu_info.n_dev_compatible > 0)
317 std::string gpuUseageReport;
320 gpuUseageReport = makeGpuUsageReport(&hwinfo->gpu_info,
322 cr->nrank_pp_intranode,
323 bMPI && cr->nnodes > 1);
325 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
327 /* NOTE: this print is only for and on one physical node */
328 md_print_info(cr, fplog, "%s\n", gpuUseageReport.c_str());
331 /* Need to ensure that we have enough GPUs:
332 * - need one GPU per PP node
333 * - no GPU oversubscription with tMPI
335 /* number of PP processes per node */
336 npppn = cr->nrank_pp_intranode;
339 th_or_proc_plural[0] = '\0';
342 sprintf(th_or_proc, "thread-MPI thread");
345 sprintf(th_or_proc_plural, "s");
350 sprintf(th_or_proc, "MPI process");
353 sprintf(th_or_proc_plural, "es");
355 sprintf(pernode, " per node");
359 /* neither MPI nor tMPI */
360 sprintf(th_or_proc, "process");
363 if (bUseGPU && hwinfo->gpu_info.n_dev_compatible > 0 &&
366 int ngpu_comp, ngpu_use;
367 char gpu_comp_plural[2], gpu_use_plural[2];
369 ngpu_comp = hwinfo->gpu_info.n_dev_compatible;
370 ngpu_use = hw_opt->gpu_opt.n_dev_use;
372 sprintf(gpu_comp_plural, "%s", (ngpu_comp > 1) ? "s" : "");
373 sprintf(gpu_use_plural, "%s", (ngpu_use > 1) ? "s" : "");
375 /* number of tMPI threads auto-adjusted */
376 if (btMPI && bNthreadsAuto)
378 if (hw_opt->gpu_opt.bUserSet && npppn < ngpu_use)
380 /* The user manually provided more GPUs than threads we
381 could automatically start. */
383 "%d GPU%s provided, but only %d PP thread-MPI thread%s coud be started.\n"
384 "%s requires one PP tread-MPI thread per GPU; use fewer GPUs%s.",
385 ngpu_use, gpu_use_plural,
386 npppn, th_or_proc_plural,
387 ShortProgram(), bMaxMpiThreadsSet ? "\nor allow more threads to be used" : "");
390 if (!hw_opt->gpu_opt.bUserSet && npppn < ngpu_comp)
392 /* There are more GPUs than tMPI threads; we have
393 limited the number GPUs used. */
394 md_print_warn(cr, fplog,
395 "NOTE: %d GPU%s were detected, but only %d PP thread-MPI thread%s can be started.\n"
396 " %s can use one GPU per PP tread-MPI thread, so only %d GPU%s will be used.%s\n",
397 ngpu_comp, gpu_comp_plural,
398 npppn, th_or_proc_plural,
399 ShortProgram(), npppn,
400 npppn > 1 ? "s" : "",
401 bMaxMpiThreadsSet ? "\n Also, you can allow more threads to be used by increasing GMX_MAX_MPI_THREADS" : "");
405 if (hw_opt->gpu_opt.bUserSet)
407 if (ngpu_use != npppn)
410 "Incorrect launch configuration: mismatching number of PP %s%s and GPUs%s.\n"
411 "%s was started with %d PP %s%s%s, but you provided %d GPU%s.",
412 th_or_proc, btMPI ? "s" : "es", pernode,
413 ShortProgram(), npppn, th_or_proc,
414 th_or_proc_plural, pernode,
415 ngpu_use, gpu_use_plural);
420 if (ngpu_comp > npppn)
422 md_print_warn(cr, fplog,
423 "NOTE: potentially sub-optimal launch configuration, %s started with less\n"
424 " PP %s%s%s than GPU%s available.\n"
425 " Each PP %s can use only one GPU, %d GPU%s%s will be used.\n",
426 ShortProgram(), th_or_proc,
427 th_or_proc_plural, pernode, gpu_comp_plural,
428 th_or_proc, npppn, gpu_use_plural, pernode);
431 if (ngpu_use != npppn)
433 /* Avoid duplicate error messages.
434 * Unfortunately we can only do this at the physical node
435 * level, since the hardware setup and MPI process count
436 * might differ between physical nodes.
438 if (cr->rank_pp_intranode == 0)
441 "Incorrect launch configuration: mismatching number of PP %s%s and GPUs%s.\n"
442 "%s was started with %d PP %s%s%s, but only %d GPU%s were detected.",
443 th_or_proc, btMPI ? "s" : "es", pernode,
444 ShortProgram(), npppn, th_or_proc,
445 th_or_proc_plural, pernode,
446 ngpu_use, gpu_use_plural);
454 same_count = gmx_count_gpu_dev_shared(&hw_opt->gpu_opt);
458 md_print_info(cr, fplog,
459 "NOTE: You assigned %s to multiple %s%s.\n",
460 same_count > 1 ? "GPUs" : "a GPU", th_or_proc, btMPI ? "s" : "es");
468 /* Avoid other ranks to continue after
470 MPI_Barrier(cr->mpi_comm_mygroup);
476 /* Return 0 if none of the GPU (per node) are shared among PP ranks.
478 * Sharing GPUs among multiple PP ranks is possible when the user passes
479 * GPU IDs. Here we check for sharing and return a non-zero value when
480 * this is detected. Note that the return value represents the number of
481 * PP rank pairs that share a device.
483 int gmx_count_gpu_dev_shared(const gmx_gpu_opt_t *gpu_opt)
486 int ngpu = gpu_opt->n_dev_use;
488 if (gpu_opt->bUserSet)
492 for (i = 0; i < ngpu - 1; i++)
494 for (j = i + 1; j < ngpu; j++)
496 same_count += (gpu_opt->dev_use[i] ==
497 gpu_opt->dev_use[j]);
505 /* Count and return the number of unique GPUs (per node) selected.
507 * As sharing GPUs among multiple PP ranks is possible when the user passes
508 * GPU IDs, the number of GPUs user (per node) can be different from the
509 * number of GPU IDs selected.
511 static int gmx_count_gpu_dev_unique(const gmx_gpu_info_t *gpu_info,
512 const gmx_gpu_opt_t *gpu_opt)
514 int i, uniq_count, ngpu;
520 ngpu = gpu_info->n_dev;
524 snew(uniq_ids, ngpu);
526 /* Each element in uniq_ids will be set to 0 or 1. The n-th element set
527 * to 1 indicates that the respective GPU was selected to be used. */
528 for (i = 0; i < gpu_opt->n_dev_use; i++)
530 uniq_ids[get_cuda_gpu_device_id(gpu_info, gpu_opt, i)] = 1;
532 /* Count the devices used. */
533 for (i = 0; i < ngpu; i++)
535 uniq_count += uniq_ids[i];
543 static int get_ncores(gmx_cpuid_t cpuid)
545 int nprocessors, npackages, ncores_per_package, nhwthreads_per_core;
546 const int *package_id, *core_id, *hwthread_id, *locality_order;
549 rc = gmx_cpuid_topology(cpuid,
550 &nprocessors, &npackages,
551 &ncores_per_package, &nhwthreads_per_core,
552 &package_id, &core_id,
553 &hwthread_id, &locality_order);
557 return npackages*ncores_per_package;
561 /* We don't have cpuid topology info, return 0 core count */
566 /* Return the number of hardware threads supported by the current CPU.
567 * We assume that this is equal with the number of "processors"
568 * reported to be online by the OS at the time of the call. The
569 * definition of "processor" is according to an old POSIX standard.
571 * Note that the number of hardware threads is generally greater than
572 * the number of cores (e.g. x86 hyper-threading, Power). Managing the
573 * mapping of software threads to hardware threads is managed
575 static int get_nthreads_hw_avail(FILE gmx_unused *fplog, const t_commrec gmx_unused *cr)
579 #if ((defined(WIN32) || defined( _WIN32 ) || defined(WIN64) || defined( _WIN64 )) && !(defined (__CYGWIN__) || defined (__CYGWIN32__)))
582 GetSystemInfo( &sysinfo );
583 ret = sysinfo.dwNumberOfProcessors;
584 #elif defined HAVE_SYSCONF
585 /* We are probably on Unix.
586 * Now check if we have the argument to use before executing the call
588 #if defined(_SC_NPROCESSORS_ONLN)
589 ret = sysconf(_SC_NPROCESSORS_ONLN);
590 #elif defined(_SC_NPROC_ONLN)
591 ret = sysconf(_SC_NPROC_ONLN);
592 #elif defined(_SC_NPROCESSORS_CONF)
593 ret = sysconf(_SC_NPROCESSORS_CONF);
594 #elif defined(_SC_NPROC_CONF)
595 ret = sysconf(_SC_NPROC_CONF);
597 #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!"
598 #endif /* End of check for sysconf argument values */
601 /* Neither windows nor Unix. No fscking idea how many hardware threads we have! */
607 fprintf(debug, "Detected %d hardware threads to use.\n", ret);
611 if (ret != gmx_omp_get_num_procs())
613 md_print_warn(cr, fplog,
614 "Number of logical cores detected (%d) does not match the number reported by OpenMP (%d).\n"
615 "Consider setting the launch configuration manually!",
616 ret, gmx_omp_get_num_procs());
623 static void gmx_detect_gpus(FILE *fplog, const t_commrec *cr)
627 MPI_Comm physicalnode_comm;
631 /* Under certain circumstances MPI ranks on the same physical node
632 * can not simultaneously access the same GPU(s). Therefore we run
633 * the detection only on one MPI rank per node and broadcast the info.
634 * Note that with thread-MPI only a single thread runs this code.
636 * TODO: We should also do CPU hardware detection only once on each
637 * physical node and broadcast it, instead of do it on every MPI rank.
640 /* A split of MPI_COMM_WORLD over physical nodes is only required here,
641 * so we create and destroy it locally.
643 MPI_Comm_rank(MPI_COMM_WORLD, &rank_world);
644 MPI_Comm_split(MPI_COMM_WORLD, gmx_physicalnode_id_hash(),
645 rank_world, &physicalnode_comm);
646 MPI_Comm_rank(physicalnode_comm, &rank_local);
648 /* Here there should be only one process, check this */
649 assert(cr->nnodes == 1 && cr->sim_nodeid == 0);
656 char detection_error[STRLEN] = "", sbuf[STRLEN];
658 if (detect_gpus(&hwinfo_g->gpu_info, detection_error) != 0)
660 if (detection_error[0] != '\0')
662 sprintf(sbuf, ":\n %s\n", detection_error);
668 md_print_warn(cr, fplog,
669 "NOTE: Error occurred during GPU detection%s"
670 " Can not use GPU acceleration, will fall back to CPU kernels.\n",
676 /* Broadcast the GPU info to the other ranks within this node */
677 MPI_Bcast(&hwinfo_g->gpu_info.n_dev, 1, MPI_INT, 0, physicalnode_comm);
679 if (hwinfo_g->gpu_info.n_dev > 0)
683 dev_size = hwinfo_g->gpu_info.n_dev*sizeof_gpu_dev_info();
687 hwinfo_g->gpu_info.gpu_dev =
688 (struct gmx_device_info_t *)malloc(dev_size);
690 MPI_Bcast(hwinfo_g->gpu_info.gpu_dev, dev_size, MPI_BYTE,
691 0, physicalnode_comm);
692 MPI_Bcast(&hwinfo_g->gpu_info.n_dev_compatible, 1, MPI_INT,
693 0, physicalnode_comm);
696 MPI_Comm_free(&physicalnode_comm);
700 static void gmx_collect_hardware_mpi()
704 int nrank, rank, ncore, nhwthread, ngpu, i;
708 rank_id = gmx_physicalnode_id_hash();
709 MPI_Comm_rank(MPI_COMM_WORLD, &rank);
710 MPI_Comm_size(MPI_COMM_WORLD, &nrank);
711 ncore = hwinfo_g->ncore;
712 nhwthread = hwinfo_g->nthreads_hw_avail;
713 ngpu = hwinfo_g->gpu_info.n_dev_compatible;
714 /* Create a unique hash of the GPU type(s) in this node */
716 /* Here it might be better to only loop over the compatible GPU, but we
717 * don't have that information available and it would also require
718 * removing the device ID from the device info string.
720 for (i = 0; i < hwinfo_g->gpu_info.n_dev; i++)
724 /* Since the device ID is incorporated in the hash, the order of
725 * the GPUs affects the hash. Also two identical GPUs won't give
726 * a gpu_hash of zero after XORing.
728 get_gpu_device_info_string(stmp, &hwinfo_g->gpu_info, i);
729 gpu_hash ^= gmx_string_fullhash_func(stmp, gmx_string_hash_init);
736 MPI_Allreduce(buf, all, nrank, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
739 int nnode0, ncore0, nhwthread0, ngpu0, r;
746 for (r = 0; r < nrank; r++)
748 if (all[r] == rank_id)
750 if (!bFound && r == rank)
752 /* We are the first rank in this physical node */
755 nhwthread0 = nhwthread;
765 int sum[4], maxmin[10];
770 /* Sum values from only intra-rank 0 so we get the sum over all nodes */
776 MPI_Allreduce(buf, sum, 4, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
782 /* Store + and - values for all ranks,
783 * so we can get max+min with one MPI call.
788 buf[3] = gmx_cpuid_simd_suggest(hwinfo_g->cpuid_info);
796 MPI_Allreduce(buf, maxmin, 10, MPI_INT, MPI_MAX, MPI_COMM_WORLD);
799 hwinfo_g->nphysicalnode = sum[0];
800 hwinfo_g->ncore_tot = sum[1];
801 hwinfo_g->ncore_min = -maxmin[5];
802 hwinfo_g->ncore_max = maxmin[0];
803 hwinfo_g->nhwthread_tot = sum[2];
804 hwinfo_g->nhwthread_min = -maxmin[6];
805 hwinfo_g->nhwthread_max = maxmin[1];
806 hwinfo_g->ngpu_compatible_tot = sum[3];
807 hwinfo_g->ngpu_compatible_min = -maxmin[7];
808 hwinfo_g->ngpu_compatible_max = maxmin[2];
809 hwinfo_g->simd_suggest_min = static_cast<enum gmx_cpuid_simd>(-maxmin[8]);
810 hwinfo_g->simd_suggest_max = static_cast<enum gmx_cpuid_simd>(maxmin[3]);
811 hwinfo_g->bIdenticalGPUs = (maxmin[4] == -maxmin[9]);
813 /* All ranks use the same pointer, protect it with a mutex */
814 tMPI_Thread_mutex_lock(&hw_info_lock);
815 hwinfo_g->nphysicalnode = 1;
816 hwinfo_g->ncore_tot = hwinfo_g->ncore;
817 hwinfo_g->ncore_min = hwinfo_g->ncore;
818 hwinfo_g->ncore_max = hwinfo_g->ncore;
819 hwinfo_g->nhwthread_tot = hwinfo_g->nthreads_hw_avail;
820 hwinfo_g->nhwthread_min = hwinfo_g->nthreads_hw_avail;
821 hwinfo_g->nhwthread_max = hwinfo_g->nthreads_hw_avail;
822 hwinfo_g->ngpu_compatible_tot = hwinfo_g->gpu_info.n_dev_compatible;
823 hwinfo_g->ngpu_compatible_min = hwinfo_g->gpu_info.n_dev_compatible;
824 hwinfo_g->ngpu_compatible_max = hwinfo_g->gpu_info.n_dev_compatible;
825 hwinfo_g->simd_suggest_min = gmx_cpuid_simd_suggest(hwinfo_g->cpuid_info);
826 hwinfo_g->simd_suggest_max = gmx_cpuid_simd_suggest(hwinfo_g->cpuid_info);
827 hwinfo_g->bIdenticalGPUs = TRUE;
828 tMPI_Thread_mutex_unlock(&hw_info_lock);
832 gmx_hw_info_t *gmx_detect_hardware(FILE *fplog, const t_commrec *cr,
833 gmx_bool bDetectGPUs)
837 /* make sure no one else is doing the same thing */
838 ret = tMPI_Thread_mutex_lock(&hw_info_lock);
841 gmx_fatal(FARGS, "Error locking hwinfo mutex: %s", strerror(errno));
844 /* only initialize the hwinfo structure if it is not already initalized */
849 /* detect CPUID info; no fuss, we don't detect system-wide
850 * -- sloppy, but that's it for now */
851 if (gmx_cpuid_init(&hwinfo_g->cpuid_info) != 0)
853 gmx_fatal_collective(FARGS, cr, NULL, "CPUID detection failed!");
856 /* get the number of cores, will be 0 when not detected */
857 hwinfo_g->ncore = get_ncores(hwinfo_g->cpuid_info);
859 /* detect number of hardware threads */
860 hwinfo_g->nthreads_hw_avail = get_nthreads_hw_avail(fplog, cr);
863 hwinfo_g->gpu_info.n_dev = 0;
864 hwinfo_g->gpu_info.n_dev_compatible = 0;
865 hwinfo_g->gpu_info.gpu_dev = NULL;
867 /* Run the detection if the binary was compiled with GPU support
868 * and we requested detection.
870 hwinfo_g->gpu_info.bDetectGPUs =
871 (bGPUBinary && bDetectGPUs &&
872 getenv("GMX_DISABLE_GPU_DETECTION") == NULL);
873 if (hwinfo_g->gpu_info.bDetectGPUs)
875 gmx_detect_gpus(fplog, cr);
878 /* increase the reference counter */
881 ret = tMPI_Thread_mutex_unlock(&hw_info_lock);
884 gmx_fatal(FARGS, "Error unlocking hwinfo mutex: %s", strerror(errno));
887 gmx_collect_hardware_mpi();
892 static std::string detected_hardware_string(const gmx_hw_info_t *hwinfo,
897 s = gmx::formatString("\n");
898 s += gmx::formatString("Running on %d node%s with total",
899 hwinfo->nphysicalnode,
900 hwinfo->nphysicalnode == 1 ? "" : "s");
901 if (hwinfo->ncore_tot > 0)
903 s += gmx::formatString(" %d cores,", hwinfo->ncore_tot);
905 s += gmx::formatString(" %d logical cores", hwinfo->nhwthread_tot);
906 if (hwinfo->gpu_info.bDetectGPUs)
908 s += gmx::formatString(", %d compatible GPU%s",
909 hwinfo->ngpu_compatible_tot,
910 hwinfo->ngpu_compatible_tot == 1 ? "" : "s");
914 s += gmx::formatString(" (GPU detection deactivated)");
916 s += gmx::formatString("\n");
918 if (hwinfo->nphysicalnode > 1)
920 /* Print per node hardware feature counts */
921 if (hwinfo->ncore_max > 0)
923 s += gmx::formatString(" Cores per node: %2d", hwinfo->ncore_min);
924 if (hwinfo->ncore_max > hwinfo->ncore_min)
926 s += gmx::formatString(" - %2d", hwinfo->ncore_max);
928 s += gmx::formatString("\n");
930 s += gmx::formatString(" Logical cores per node: %2d", hwinfo->nhwthread_min);
931 if (hwinfo->nhwthread_max > hwinfo->nhwthread_min)
933 s += gmx::formatString(" - %2d", hwinfo->nhwthread_max);
935 s += gmx::formatString("\n");
938 s += gmx::formatString(" Compatible GPUs per node: %2d",
939 hwinfo->ngpu_compatible_min);
940 if (hwinfo->ngpu_compatible_max > hwinfo->ngpu_compatible_min)
942 s += gmx::formatString(" - %2d", hwinfo->ngpu_compatible_max);
944 s += gmx::formatString("\n");
945 if (hwinfo->ngpu_compatible_tot > 0)
947 if (hwinfo->bIdenticalGPUs)
949 s += gmx::formatString(" All nodes have identical type(s) of GPUs\n");
953 /* This message will also appear with identical GPU types
954 * when at least one node has no GPU.
956 s += gmx::formatString(" Different nodes have different type(s) and/or order of GPUs\n");
963 char host[HOSTNAMELEN];
966 gmx_gethostname(host, HOSTNAMELEN);
967 MPI_Comm_rank(MPI_COMM_WORLD, &rank);
969 s += gmx::formatString("Hardware detected on host %s (the node of MPI rank %d):\n",
972 s += gmx::formatString("Hardware detected:\n");
974 s += gmx::formatString(" CPU info:\n");
979 gmx_cpuid_formatstring(hwinfo->cpuid_info, buf, 1023);
982 s += gmx::formatString("%s", buf);
986 s += gmx::formatString(" Vendor: %s\n",
987 gmx_cpuid_vendor_string[gmx_cpuid_vendor(hwinfo->cpuid_info)]);
988 s += gmx::formatString(" Brand: %s\n",
989 gmx_cpuid_brand(hwinfo->cpuid_info));
991 s += gmx::formatString(" SIMD instructions most likely to fit this hardware: %s",
992 gmx_cpuid_simd_string[hwinfo->simd_suggest_min]);
993 if (hwinfo->simd_suggest_max > hwinfo->simd_suggest_min)
995 s += gmx::formatString(" - %s",
996 gmx_cpuid_simd_string[hwinfo->simd_suggest_max]);
998 s += gmx::formatString("\n");
999 s += gmx::formatString(" SIMD instructions selected at GROMACS compile time: %s\n",
1000 gmx_cpuid_simd_string[gmx_compiled_simd()]);
1001 if (bGPUBinary && (hwinfo->ngpu_compatible_tot > 0 ||
1002 hwinfo->gpu_info.n_dev > 0))
1004 s += gmx::formatString(" GPU info:\n");
1005 s += gmx::formatString(" Number of GPUs detected: %d\n",
1006 hwinfo->gpu_info.n_dev);
1007 if (hwinfo->gpu_info.n_dev > 0)
1011 sprint_gpus(buf, &hwinfo->gpu_info);
1012 s += gmx::formatString("%s\n", buf);
1019 void gmx_print_detected_hardware(FILE *fplog, const t_commrec *cr,
1020 const gmx_hw_info_t *hwinfo)
1024 std::string detected;
1026 detected = detected_hardware_string(hwinfo, TRUE);
1028 fprintf(fplog, "%s\n", detected.c_str());
1031 if (MULTIMASTER(cr))
1033 std::string detected;
1035 detected = detected_hardware_string(hwinfo, FALSE);
1037 fprintf(stderr, "%s\n", detected.c_str());
1040 /* Check the compiled SIMD instruction set against that of the node
1041 * with the lowest SIMD level support.
1043 gmx_cpuid_simd_check(hwinfo->simd_suggest_min, fplog, MULTIMASTER(cr));
1045 /* For RDTSCP we only check on our local node and skip the MPI reduction */
1046 check_use_of_rdtscp_on_this_cpu(fplog, cr, hwinfo);
1049 void gmx_parse_gpu_ids(gmx_gpu_opt_t *gpu_opt)
1053 if (gpu_opt->gpu_id != NULL && !bGPUBinary)
1055 gmx_fatal(FARGS, "GPU ID string set, but %s was compiled without GPU support!", ShortProgram());
1058 env = getenv("GMX_GPU_ID");
1059 if (env != NULL && gpu_opt->gpu_id != NULL)
1061 gmx_fatal(FARGS, "GMX_GPU_ID and -gpu_id can not be used at the same time");
1065 env = gpu_opt->gpu_id;
1068 /* parse GPU IDs if the user passed any */
1071 /* Parse a "plain" GPU ID string which contains a sequence of
1072 * digits corresponding to GPU IDs; the order will indicate
1073 * the process/tMPI thread - GPU assignment. */
1074 parse_digits_from_plain_string(env,
1075 &gpu_opt->n_dev_use,
1078 if (gpu_opt->n_dev_use == 0)
1080 gmx_fatal(FARGS, "Empty GPU ID string encountered.\n%s\n",
1081 invalid_gpuid_hint);
1084 gpu_opt->bUserSet = TRUE;
1088 void gmx_select_gpu_ids(FILE *fplog, const t_commrec *cr,
1089 const gmx_gpu_info_t *gpu_info,
1090 gmx_bool bForceUseGPU,
1091 gmx_gpu_opt_t *gpu_opt)
1094 char sbuf[STRLEN], stmp[STRLEN];
1096 /* Bail if binary is not compiled with GPU acceleration, but this is either
1097 * explicitly (-nb gpu) or implicitly (gpu ID passed) requested. */
1098 if (bForceUseGPU && !bGPUBinary)
1100 gmx_fatal(FARGS, "GPU acceleration requested, but %s was compiled without GPU support!", ShortProgram());
1103 if (!(cr->duty & DUTY_PP))
1105 /* Our rank is not doing PP, we don't use a GPU */
1109 if (gpu_opt->bUserSet)
1111 /* Check the GPU IDs passed by the user.
1112 * (GPU IDs have been parsed by gmx_parse_gpu_ids before)
1117 snew(checkres, gpu_opt->n_dev_use);
1119 res = check_selected_gpus(checkres, gpu_info, gpu_opt);
1123 print_gpu_detection_stats(fplog, gpu_info, cr);
1125 sprintf(sbuf, "Some of the requested GPUs do not exist, behave strangely, or are not compatible:\n");
1126 for (i = 0; i < gpu_opt->n_dev_use; i++)
1128 if (checkres[i] != egpuCompatible)
1130 sprintf(stmp, " GPU #%d: %s\n",
1131 gpu_opt->dev_use[i],
1132 gpu_detect_res_str[checkres[i]]);
1136 gmx_fatal(FARGS, "%s", sbuf);
1143 pick_compatible_gpus(&hwinfo_g->gpu_info, gpu_opt);
1144 set_gpu_ids(gpu_opt, cr->nrank_pp_intranode, cr->rank_pp_intranode);
1147 /* If the user asked for a GPU, check whether we have a GPU */
1148 if (bForceUseGPU && gpu_info->n_dev_compatible == 0)
1150 gmx_fatal(FARGS, "GPU acceleration requested, but no compatible GPUs were detected.");
1154 /* Select the GPUs we will use. This is an operation local to each physical
1155 * node. If we have less MPI ranks than GPUs, we will waste some GPUs.
1156 * nrank and rank are the rank count and id for PP processes in our node.
1158 static void set_gpu_ids(gmx_gpu_opt_t *gpu_opt, int nrank, int rank)
1160 GMX_RELEASE_ASSERT(gpu_opt, "Invalid gpu_opt pointer passed");
1161 GMX_RELEASE_ASSERT(nrank >= 1,
1162 gmx::formatString("Invalid limit (%d) for the number of GPUs (detected %d compatible GPUs)",
1163 rank, gpu_opt->n_dev_compatible).c_str());
1165 if (gpu_opt->n_dev_compatible == 0)
1167 char host[HOSTNAMELEN];
1169 gmx_gethostname(host, HOSTNAMELEN);
1170 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);
1176 if (nrank > gpu_opt->n_dev_compatible)
1178 if (nrank % gpu_opt->n_dev_compatible == 0)
1180 nshare = nrank/gpu_opt->n_dev_compatible;
1186 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.",
1187 nrank, gpu_opt->n_dev_compatible);
1191 /* We use a global barrier to prevent ranks from continuing with
1194 MPI_Barrier(MPI_COMM_WORLD);
1199 /* Here we will waste GPUs when nrank < gpu_opt->n_dev_compatible */
1200 gpu_opt->n_dev_use = std::min(gpu_opt->n_dev_compatible*nshare, nrank);
1201 snew(gpu_opt->dev_use, gpu_opt->n_dev_use);
1202 for (int i = 0; i != gpu_opt->n_dev_use; ++i)
1204 /* TODO: improve this implementation: either sort GPUs or remove the weakest here */
1205 gpu_opt->dev_use[i] = gpu_opt->dev_compatible[i/nshare];
1209 void gmx_hardware_info_free(gmx_hw_info_t *hwinfo)
1213 ret = tMPI_Thread_mutex_lock(&hw_info_lock);
1216 gmx_fatal(FARGS, "Error locking hwinfo mutex: %s", strerror(errno));
1219 /* decrease the reference counter */
1223 if (hwinfo != hwinfo_g)
1225 gmx_incons("hwinfo < hwinfo_g");
1230 gmx_incons("n_hwinfo < 0");
1235 gmx_cpuid_done(hwinfo_g->cpuid_info);
1236 free_gpu_info(&hwinfo_g->gpu_info);
1240 ret = tMPI_Thread_mutex_unlock(&hw_info_lock);
1243 gmx_fatal(FARGS, "Error unlocking hwinfo mutex: %s", strerror(errno));