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 /* Names of the GPU detection/check results (see e_gpu_detect_res_t in hw_info.h). */
86 const char * const gpu_detect_res_str[egpuNR] =
88 "compatible", "inexistent", "incompatible", "insane"
91 static const char * invalid_gpuid_hint =
92 "A delimiter-free sequence of valid numeric IDs of available GPUs is expected.";
94 /* The globally shared hwinfo structure. */
95 static gmx_hw_info_t *hwinfo_g;
96 /* A reference counter for the hwinfo structure */
97 static int n_hwinfo = 0;
98 /* A lock to protect the hwinfo structure */
99 static tMPI_Thread_mutex_t hw_info_lock = TMPI_THREAD_MUTEX_INITIALIZER;
101 #define HOSTNAMELEN 80
104 static void set_gpu_ids(gmx_gpu_opt_t *gpu_opt, int nrank, int rank);
105 static int gmx_count_gpu_dev_unique(const gmx_gpu_info_t *gpu_info,
106 const gmx_gpu_opt_t *gpu_opt);
108 static void sprint_gpus(char *sbuf, const gmx_gpu_info_t *gpu_info)
113 ndev = gpu_info->n_dev;
116 for (i = 0; i < ndev; i++)
118 get_gpu_device_info_string(stmp, gpu_info, i);
128 static void print_gpu_detection_stats(FILE *fplog,
129 const gmx_gpu_info_t *gpu_info,
132 char onhost[HOSTNAMELEN+10], stmp[STRLEN];
135 if (!gpu_info->bDetectGPUs)
137 /* We skipped the detection, so don't print detection stats */
141 ngpu = gpu_info->n_dev;
143 #if defined GMX_MPI && !defined GMX_THREAD_MPI
144 /* We only print the detection on one, of possibly multiple, nodes */
145 strncpy(onhost, " on host ", 10);
146 gmx_gethostname(onhost + 9, HOSTNAMELEN);
148 /* We detect all relevant GPUs */
149 strncpy(onhost, "", 1);
154 sprint_gpus(stmp, gpu_info);
155 md_print_warn(cr, fplog, "%d GPU%s detected%s:\n%s\n",
156 ngpu, (ngpu > 1) ? "s" : "", onhost, stmp);
160 md_print_warn(cr, fplog, "No GPUs detected%s\n", onhost);
164 /*! \brief Helper function for reporting GPU usage information
165 * in the mdrun log file
167 * \param[in] gpu_info Pointer to per-node GPU info struct
168 * \param[in] gpu_opt Pointer to per-node GPU options struct
169 * \param[in] numPpRanks Number of PP ranks per node
170 * \param[in] bPrintHostName Print the hostname in the usage information
171 * \return String to write to the log file
172 * \throws std::bad_alloc if out of memory */
174 makeGpuUsageReport(const gmx_gpu_info_t *gpu_info,
175 const gmx_gpu_opt_t *gpu_opt,
179 int ngpu_use = gpu_opt->n_dev_use;
180 int ngpu_comp = gpu_info->n_dev_compatible;
181 char host[HOSTNAMELEN];
185 gmx_gethostname(host, HOSTNAMELEN);
188 /* Issue a note if GPUs are available but not used */
189 if (ngpu_comp > 0 && ngpu_use < 1)
191 return gmx::formatString("%d compatible GPU%s detected in the system, but none will be used.\n"
192 "Consider trying GPU acceleration with the Verlet scheme!\n",
193 ngpu_comp, (ngpu_comp > 1) ? "s" : "");
197 if (!gpu_opt->bUserSet)
199 // gpu_opt->dev_compatible is only populated during auto-selection
200 std::string gpuIdsString =
201 formatAndJoin(gmx::constArrayRefFromArray(gpu_opt->dev_compatible,
202 gpu_opt->n_dev_compatible),
203 ",", gmx::StringFormatter("%d"));
204 bool bPluralGpus = gpu_opt->n_dev_compatible > 1;
208 output += gmx::formatString("On host %s ", host);
210 output += gmx::formatString("%d compatible GPU%s %s present, with ID%s %s\n",
211 gpu_opt->n_dev_compatible,
212 bPluralGpus ? "s" : "",
213 bPluralGpus ? "are" : "is",
214 bPluralGpus ? "s" : "",
215 gpuIdsString.c_str());
219 std::vector<int> gpuIdsInUse;
220 for (int i = 0; i < ngpu_use; i++)
222 gpuIdsInUse.push_back(get_cuda_gpu_device_id(gpu_info, gpu_opt, i));
224 std::string gpuIdsString =
225 formatAndJoin(gpuIdsInUse, ",", gmx::StringFormatter("%d"));
226 int numGpusInUse = gmx_count_gpu_dev_unique(gpu_info, gpu_opt);
227 bool bPluralGpus = numGpusInUse > 1;
231 output += gmx::formatString("On host %s ", host);
233 output += gmx::formatString("%d GPU%s %sselected for this run.\n"
234 "Mapping of GPU ID%s to the %d PP rank%s in this node: %s\n",
235 numGpusInUse, bPluralGpus ? "s" : "",
236 gpu_opt->bUserSet ? "user-" : "auto-",
237 bPluralGpus ? "s" : "",
239 (numPpRanks > 1) ? "s" : "",
240 gpuIdsString.c_str());
246 /* Give a suitable fatal error or warning if the build configuration
247 and runtime CPU do not match. */
249 check_use_of_rdtscp_on_this_cpu(FILE *fplog,
251 const gmx_hw_info_t *hwinfo)
253 gmx_bool bCpuHasRdtscp, bBinaryUsesRdtscp;
255 bBinaryUsesRdtscp = TRUE;
257 bBinaryUsesRdtscp = FALSE;
260 bCpuHasRdtscp = gmx_cpuid_feature(hwinfo->cpuid_info, GMX_CPUID_FEATURE_X86_RDTSCP);
262 if (!bCpuHasRdtscp && bBinaryUsesRdtscp)
264 gmx_fatal(FARGS, "The %s executable was compiled to use the rdtscp CPU instruction. "
265 "However, this is not supported by the current hardware and continuing would lead to a crash. "
266 "Please rebuild GROMACS with the GMX_USE_RDTSCP=OFF CMake option.",
270 if (bCpuHasRdtscp && !bBinaryUsesRdtscp)
272 md_print_warn(cr, fplog, "The current CPU can measure timings more accurately than the code in\n"
273 "%s was configured to use. This might affect your simulation\n"
274 "speed as accurate timings are needed for load-balancing.\n"
275 "Please consider rebuilding %s with the GMX_USE_RDTSCP=ON CMake option.\n",
276 ShortProgram(), ShortProgram());
280 void gmx_check_hw_runconf_consistency(FILE *fplog,
281 const gmx_hw_info_t *hwinfo,
283 const gmx_hw_opt_t *hw_opt,
287 char th_or_proc[STRLEN], th_or_proc_plural[STRLEN], pernode[STRLEN];
288 gmx_bool btMPI, bMPI, bNthreadsAuto, bEmulateGPU;
293 /* Below we only do consistency checks for PP and GPUs,
294 * this is irrelevant for PME only nodes, so in that case we return
297 if (!(cr->duty & DUTY_PP))
302 #if defined(GMX_THREAD_MPI)
305 bNthreadsAuto = (hw_opt->nthreads_tmpi < 1);
306 #elif defined(GMX_LIB_MPI)
309 bNthreadsAuto = FALSE;
313 bNthreadsAuto = FALSE;
316 /* GPU emulation detection is done later, but we need here as well
317 * -- uncool, but there's no elegant workaround */
318 bEmulateGPU = (getenv("GMX_EMULATE_GPU") != NULL);
320 if (hwinfo->gpu_info.n_dev_compatible > 0)
322 std::string gpuUseageReport;
325 gpuUseageReport = makeGpuUsageReport(&hwinfo->gpu_info,
327 cr->nrank_pp_intranode,
328 bMPI && cr->nnodes > 1);
330 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
332 /* NOTE: this print is only for and on one physical node */
333 md_print_info(cr, fplog, "%s\n", gpuUseageReport.c_str());
336 /* Need to ensure that we have enough GPUs:
337 * - need one GPU per PP node
338 * - no GPU oversubscription with tMPI
340 /* number of PP processes per node */
341 npppn = cr->nrank_pp_intranode;
344 th_or_proc_plural[0] = '\0';
347 sprintf(th_or_proc, "thread-MPI thread");
350 sprintf(th_or_proc_plural, "s");
355 sprintf(th_or_proc, "MPI process");
358 sprintf(th_or_proc_plural, "es");
360 sprintf(pernode, " per node");
364 /* neither MPI nor tMPI */
365 sprintf(th_or_proc, "process");
368 if (bUseGPU && hwinfo->gpu_info.n_dev_compatible > 0 &&
371 int ngpu_comp, ngpu_use;
372 char gpu_comp_plural[2], gpu_use_plural[2];
374 ngpu_comp = hwinfo->gpu_info.n_dev_compatible;
375 ngpu_use = hw_opt->gpu_opt.n_dev_use;
377 sprintf(gpu_comp_plural, "%s", (ngpu_comp > 1) ? "s" : "");
378 sprintf(gpu_use_plural, "%s", (ngpu_use > 1) ? "s" : "");
380 /* number of tMPI threads auto-adjusted */
381 if (btMPI && bNthreadsAuto)
383 if (hw_opt->gpu_opt.bUserSet && npppn < ngpu_use)
385 /* The user manually provided more GPUs than threads we
386 could automatically start. */
388 "%d GPU%s provided, but only %d PP thread-MPI thread%s coud be started.\n"
389 "%s requires one PP tread-MPI thread per GPU; use fewer GPUs.",
390 ngpu_use, gpu_use_plural,
391 npppn, th_or_proc_plural,
395 if (!hw_opt->gpu_opt.bUserSet && npppn < ngpu_comp)
397 /* There are more GPUs than tMPI threads; we have
398 limited the number GPUs used. */
399 md_print_warn(cr, fplog,
400 "NOTE: %d GPU%s were detected, but only %d PP thread-MPI thread%s can be started.\n"
401 " %s can use one GPU per PP tread-MPI thread, so only %d GPU%s will be used.\n",
402 ngpu_comp, gpu_comp_plural,
403 npppn, th_or_proc_plural,
404 ShortProgram(), npppn,
405 npppn > 1 ? "s" : "");
409 if (hw_opt->gpu_opt.bUserSet)
411 if (ngpu_use != npppn)
414 "Incorrect launch configuration: mismatching number of PP %s%s and GPUs%s.\n"
415 "%s was started with %d PP %s%s%s, but you provided %d GPU%s.",
416 th_or_proc, btMPI ? "s" : "es", pernode,
417 ShortProgram(), npppn, th_or_proc,
418 th_or_proc_plural, pernode,
419 ngpu_use, gpu_use_plural);
424 if (ngpu_comp > npppn)
426 md_print_warn(cr, fplog,
427 "NOTE: potentially sub-optimal launch configuration, %s started with less\n"
428 " PP %s%s%s than GPU%s available.\n"
429 " Each PP %s can use only one GPU, %d GPU%s%s will be used.\n",
430 ShortProgram(), th_or_proc,
431 th_or_proc_plural, pernode, gpu_comp_plural,
432 th_or_proc, npppn, gpu_use_plural, pernode);
435 if (ngpu_use != npppn)
437 /* Avoid duplicate error messages.
438 * Unfortunately we can only do this at the physical node
439 * level, since the hardware setup and MPI process count
440 * might differ between physical nodes.
442 if (cr->rank_pp_intranode == 0)
445 "Incorrect launch configuration: mismatching number of PP %s%s and GPUs%s.\n"
446 "%s was started with %d PP %s%s%s, but only %d GPU%s were detected.",
447 th_or_proc, btMPI ? "s" : "es", pernode,
448 ShortProgram(), npppn, th_or_proc,
449 th_or_proc_plural, pernode,
450 ngpu_use, gpu_use_plural);
458 same_count = gmx_count_gpu_dev_shared(&hw_opt->gpu_opt);
462 md_print_info(cr, fplog,
463 "NOTE: You assigned %s to multiple %s%s.\n",
464 same_count > 1 ? "GPUs" : "a GPU", th_or_proc, btMPI ? "s" : "es");
472 /* Avoid other ranks to continue after
474 MPI_Barrier(cr->mpi_comm_mygroup);
480 /* Return 0 if none of the GPU (per node) are shared among PP ranks.
482 * Sharing GPUs among multiple PP ranks is possible when the user passes
483 * GPU IDs. Here we check for sharing and return a non-zero value when
484 * this is detected. Note that the return value represents the number of
485 * PP rank pairs that share a device.
487 int gmx_count_gpu_dev_shared(const gmx_gpu_opt_t *gpu_opt)
490 int ngpu = gpu_opt->n_dev_use;
492 if (gpu_opt->bUserSet)
496 for (i = 0; i < ngpu - 1; i++)
498 for (j = i + 1; j < ngpu; j++)
500 same_count += (gpu_opt->dev_use[i] ==
501 gpu_opt->dev_use[j]);
509 /* Count and return the number of unique GPUs (per node) selected.
511 * As sharing GPUs among multiple PP ranks is possible when the user passes
512 * GPU IDs, the number of GPUs user (per node) can be different from the
513 * number of GPU IDs selected.
515 static int gmx_count_gpu_dev_unique(const gmx_gpu_info_t *gpu_info,
516 const gmx_gpu_opt_t *gpu_opt)
518 int i, uniq_count, ngpu;
524 ngpu = gpu_info->n_dev;
528 snew(uniq_ids, ngpu);
530 /* Each element in uniq_ids will be set to 0 or 1. The n-th element set
531 * to 1 indicates that the respective GPU was selected to be used. */
532 for (i = 0; i < gpu_opt->n_dev_use; i++)
534 uniq_ids[get_cuda_gpu_device_id(gpu_info, gpu_opt, i)] = 1;
536 /* Count the devices used. */
537 for (i = 0; i < ngpu; i++)
539 uniq_count += uniq_ids[i];
547 static int get_ncores(gmx_cpuid_t cpuid)
549 int nprocessors, npackages, ncores_per_package, nhwthreads_per_core;
550 const int *package_id, *core_id, *hwthread_id, *locality_order;
553 rc = gmx_cpuid_topology(cpuid,
554 &nprocessors, &npackages,
555 &ncores_per_package, &nhwthreads_per_core,
556 &package_id, &core_id,
557 &hwthread_id, &locality_order);
561 return npackages*ncores_per_package;
565 /* We don't have cpuid topology info, return 0 core count */
570 /* Return the number of hardware threads supported by the current CPU.
571 * We assume that this is equal with the number of "processors"
572 * reported to be online by the OS at the time of the call. The
573 * definition of "processor" is according to an old POSIX standard.
575 * Note that the number of hardware threads is generally greater than
576 * the number of cores (e.g. x86 hyper-threading, Power). Managing the
577 * mapping of software threads to hardware threads is managed
579 static int get_nthreads_hw_avail(FILE gmx_unused *fplog, const t_commrec gmx_unused *cr)
583 #if ((defined(WIN32) || defined( _WIN32 ) || defined(WIN64) || defined( _WIN64 )) && !(defined (__CYGWIN__) || defined (__CYGWIN32__)))
586 GetSystemInfo( &sysinfo );
587 ret = sysinfo.dwNumberOfProcessors;
588 #elif defined HAVE_SYSCONF
589 /* We are probably on Unix.
590 * Now check if we have the argument to use before executing the call
592 #if defined(_SC_NPROCESSORS_ONLN)
593 ret = sysconf(_SC_NPROCESSORS_ONLN);
594 #elif defined(_SC_NPROC_ONLN)
595 ret = sysconf(_SC_NPROC_ONLN);
596 #elif defined(_SC_NPROCESSORS_CONF)
597 ret = sysconf(_SC_NPROCESSORS_CONF);
598 #elif defined(_SC_NPROC_CONF)
599 ret = sysconf(_SC_NPROC_CONF);
601 #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!"
602 #endif /* End of check for sysconf argument values */
605 /* Neither windows nor Unix. No fscking idea how many hardware threads we have! */
611 fprintf(debug, "Detected %d hardware threads to use.\n", ret);
615 if (ret != gmx_omp_get_num_procs())
617 md_print_warn(cr, fplog,
618 "Number of logical cores detected (%d) does not match the number reported by OpenMP (%d).\n"
619 "Consider setting the launch configuration manually!",
620 ret, gmx_omp_get_num_procs());
627 static void gmx_detect_gpus(FILE *fplog, const t_commrec *cr)
631 MPI_Comm physicalnode_comm;
635 /* Under certain circumstances MPI ranks on the same physical node
636 * can not simultaneously access the same GPU(s). Therefore we run
637 * the detection only on one MPI rank per node and broadcast the info.
638 * Note that with thread-MPI only a single thread runs this code.
640 * TODO: We should also do CPU hardware detection only once on each
641 * physical node and broadcast it, instead of do it on every MPI rank.
644 /* A split of MPI_COMM_WORLD over physical nodes is only required here,
645 * so we create and destroy it locally.
647 MPI_Comm_rank(MPI_COMM_WORLD, &rank_world);
648 MPI_Comm_split(MPI_COMM_WORLD, gmx_physicalnode_id_hash(),
649 rank_world, &physicalnode_comm);
650 MPI_Comm_rank(physicalnode_comm, &rank_local);
652 /* Here there should be only one process, check this */
653 assert(cr->nnodes == 1 && cr->sim_nodeid == 0);
660 char detection_error[STRLEN] = "", sbuf[STRLEN];
662 if (detect_gpus(&hwinfo_g->gpu_info, detection_error) != 0)
664 if (detection_error[0] != '\0')
666 sprintf(sbuf, ":\n %s\n", detection_error);
672 md_print_warn(cr, fplog,
673 "NOTE: Error occurred during GPU detection%s"
674 " Can not use GPU acceleration, will fall back to CPU kernels.\n",
680 /* Broadcast the GPU info to the other ranks within this node */
681 MPI_Bcast(&hwinfo_g->gpu_info.n_dev, 1, MPI_INT, 0, physicalnode_comm);
683 if (hwinfo_g->gpu_info.n_dev > 0)
687 dev_size = hwinfo_g->gpu_info.n_dev*sizeof_gpu_dev_info();
691 hwinfo_g->gpu_info.gpu_dev =
692 (struct gmx_device_info_t *)malloc(dev_size);
694 MPI_Bcast(hwinfo_g->gpu_info.gpu_dev, dev_size, MPI_BYTE,
695 0, physicalnode_comm);
696 MPI_Bcast(&hwinfo_g->gpu_info.n_dev_compatible, 1, MPI_INT,
697 0, physicalnode_comm);
700 MPI_Comm_free(&physicalnode_comm);
704 static void gmx_collect_hardware_mpi()
708 int nrank, rank, ncore, nhwthread, ngpu, i;
712 rank_id = gmx_physicalnode_id_hash();
713 MPI_Comm_rank(MPI_COMM_WORLD, &rank);
714 MPI_Comm_size(MPI_COMM_WORLD, &nrank);
715 ncore = hwinfo_g->ncore;
716 nhwthread = hwinfo_g->nthreads_hw_avail;
717 ngpu = hwinfo_g->gpu_info.n_dev_compatible;
718 /* Create a unique hash of the GPU type(s) in this node */
720 /* Here it might be better to only loop over the compatible GPU, but we
721 * don't have that information available and it would also require
722 * removing the device ID from the device info string.
724 for (i = 0; i < hwinfo_g->gpu_info.n_dev; i++)
728 /* Since the device ID is incorporated in the hash, the order of
729 * the GPUs affects the hash. Also two identical GPUs won't give
730 * a gpu_hash of zero after XORing.
732 get_gpu_device_info_string(stmp, &hwinfo_g->gpu_info, i);
733 gpu_hash ^= gmx_string_fullhash_func(stmp, gmx_string_hash_init);
740 MPI_Allreduce(buf, all, nrank, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
743 int nnode0, ncore0, nhwthread0, ngpu0, r;
750 for (r = 0; r < nrank; r++)
752 if (all[r] == rank_id)
754 if (!bFound && r == rank)
756 /* We are the first rank in this physical node */
759 nhwthread0 = nhwthread;
769 int sum[4], maxmin[10];
774 /* Sum values from only intra-rank 0 so we get the sum over all nodes */
780 MPI_Allreduce(buf, sum, 4, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
786 /* Store + and - values for all ranks,
787 * so we can get max+min with one MPI call.
792 buf[3] = gmx_cpuid_simd_suggest(hwinfo_g->cpuid_info);
800 MPI_Allreduce(buf, maxmin, 10, MPI_INT, MPI_MAX, MPI_COMM_WORLD);
803 hwinfo_g->nphysicalnode = sum[0];
804 hwinfo_g->ncore_tot = sum[1];
805 hwinfo_g->ncore_min = -maxmin[5];
806 hwinfo_g->ncore_max = maxmin[0];
807 hwinfo_g->nhwthread_tot = sum[2];
808 hwinfo_g->nhwthread_min = -maxmin[6];
809 hwinfo_g->nhwthread_max = maxmin[1];
810 hwinfo_g->ngpu_compatible_tot = sum[3];
811 hwinfo_g->ngpu_compatible_min = -maxmin[7];
812 hwinfo_g->ngpu_compatible_max = maxmin[2];
813 hwinfo_g->simd_suggest_min = static_cast<enum gmx_cpuid_simd>(-maxmin[8]);
814 hwinfo_g->simd_suggest_max = static_cast<enum gmx_cpuid_simd>(maxmin[3]);
815 hwinfo_g->bIdenticalGPUs = (maxmin[4] == -maxmin[9]);
817 /* All ranks use the same pointer, protect it with a mutex */
818 tMPI_Thread_mutex_lock(&hw_info_lock);
819 hwinfo_g->nphysicalnode = 1;
820 hwinfo_g->ncore_tot = hwinfo_g->ncore;
821 hwinfo_g->ncore_min = hwinfo_g->ncore;
822 hwinfo_g->ncore_max = hwinfo_g->ncore;
823 hwinfo_g->nhwthread_tot = hwinfo_g->nthreads_hw_avail;
824 hwinfo_g->nhwthread_min = hwinfo_g->nthreads_hw_avail;
825 hwinfo_g->nhwthread_max = hwinfo_g->nthreads_hw_avail;
826 hwinfo_g->ngpu_compatible_tot = hwinfo_g->gpu_info.n_dev_compatible;
827 hwinfo_g->ngpu_compatible_min = hwinfo_g->gpu_info.n_dev_compatible;
828 hwinfo_g->ngpu_compatible_max = hwinfo_g->gpu_info.n_dev_compatible;
829 hwinfo_g->simd_suggest_min = gmx_cpuid_simd_suggest(hwinfo_g->cpuid_info);
830 hwinfo_g->simd_suggest_max = gmx_cpuid_simd_suggest(hwinfo_g->cpuid_info);
831 hwinfo_g->bIdenticalGPUs = TRUE;
832 tMPI_Thread_mutex_unlock(&hw_info_lock);
836 gmx_hw_info_t *gmx_detect_hardware(FILE *fplog, const t_commrec *cr,
837 gmx_bool bDetectGPUs)
841 /* make sure no one else is doing the same thing */
842 ret = tMPI_Thread_mutex_lock(&hw_info_lock);
845 gmx_fatal(FARGS, "Error locking hwinfo mutex: %s", strerror(errno));
848 /* only initialize the hwinfo structure if it is not already initalized */
853 /* detect CPUID info; no fuss, we don't detect system-wide
854 * -- sloppy, but that's it for now */
855 if (gmx_cpuid_init(&hwinfo_g->cpuid_info) != 0)
857 gmx_fatal_collective(FARGS, cr, NULL, "CPUID detection failed!");
860 /* get the number of cores, will be 0 when not detected */
861 hwinfo_g->ncore = get_ncores(hwinfo_g->cpuid_info);
863 /* detect number of hardware threads */
864 hwinfo_g->nthreads_hw_avail = get_nthreads_hw_avail(fplog, cr);
867 hwinfo_g->gpu_info.n_dev = 0;
868 hwinfo_g->gpu_info.n_dev_compatible = 0;
869 hwinfo_g->gpu_info.gpu_dev = NULL;
871 /* Run the detection if the binary was compiled with GPU support
872 * and we requested detection.
874 hwinfo_g->gpu_info.bDetectGPUs =
875 (bGPUBinary && bDetectGPUs &&
876 getenv("GMX_DISABLE_GPU_DETECTION") == NULL);
877 if (hwinfo_g->gpu_info.bDetectGPUs)
879 gmx_detect_gpus(fplog, cr);
882 /* increase the reference counter */
885 ret = tMPI_Thread_mutex_unlock(&hw_info_lock);
888 gmx_fatal(FARGS, "Error unlocking hwinfo mutex: %s", strerror(errno));
891 gmx_collect_hardware_mpi();
896 static std::string detected_hardware_string(const gmx_hw_info_t *hwinfo,
901 s = gmx::formatString("\n");
902 s += gmx::formatString("Running on %d node%s with total",
903 hwinfo->nphysicalnode,
904 hwinfo->nphysicalnode == 1 ? "" : "s");
905 if (hwinfo->ncore_tot > 0)
907 s += gmx::formatString(" %d cores,", hwinfo->ncore_tot);
909 s += gmx::formatString(" %d logical cores", hwinfo->nhwthread_tot);
910 if (hwinfo->gpu_info.bDetectGPUs)
912 s += gmx::formatString(", %d compatible GPU%s",
913 hwinfo->ngpu_compatible_tot,
914 hwinfo->ngpu_compatible_tot == 1 ? "" : "s");
918 s += gmx::formatString(" (GPU detection deactivated)");
920 s += gmx::formatString("\n");
922 if (hwinfo->nphysicalnode > 1)
924 /* Print per node hardware feature counts */
925 if (hwinfo->ncore_max > 0)
927 s += gmx::formatString(" Cores per node: %2d", hwinfo->ncore_min);
928 if (hwinfo->ncore_max > hwinfo->ncore_min)
930 s += gmx::formatString(" - %2d", hwinfo->ncore_max);
932 s += gmx::formatString("\n");
934 s += gmx::formatString(" Logical cores per node: %2d", hwinfo->nhwthread_min);
935 if (hwinfo->nhwthread_max > hwinfo->nhwthread_min)
937 s += gmx::formatString(" - %2d", hwinfo->nhwthread_max);
939 s += gmx::formatString("\n");
942 s += gmx::formatString(" Compatible GPUs per node: %2d",
943 hwinfo->ngpu_compatible_min);
944 if (hwinfo->ngpu_compatible_max > hwinfo->ngpu_compatible_min)
946 s += gmx::formatString(" - %2d", hwinfo->ngpu_compatible_max);
948 s += gmx::formatString("\n");
949 if (hwinfo->ngpu_compatible_tot > 0)
951 if (hwinfo->bIdenticalGPUs)
953 s += gmx::formatString(" All nodes have identical type(s) of GPUs\n");
957 /* This message will also appear with identical GPU types
958 * when at least one node has no GPU.
960 s += gmx::formatString(" Different nodes have different type(s) and/or order of GPUs\n");
967 char host[HOSTNAMELEN];
970 gmx_gethostname(host, HOSTNAMELEN);
971 MPI_Comm_rank(MPI_COMM_WORLD, &rank);
973 s += gmx::formatString("Hardware detected on host %s (the node of MPI rank %d):\n",
976 s += gmx::formatString("Hardware detected:\n");
978 s += gmx::formatString(" CPU info:\n");
983 gmx_cpuid_formatstring(hwinfo->cpuid_info, buf, 1023);
986 s += gmx::formatString("%s", buf);
990 s += gmx::formatString(" Vendor: %s\n",
991 gmx_cpuid_vendor_string[gmx_cpuid_vendor(hwinfo->cpuid_info)]);
992 s += gmx::formatString(" Brand: %s\n",
993 gmx_cpuid_brand(hwinfo->cpuid_info));
995 s += gmx::formatString(" SIMD instructions most likely to fit this hardware: %s",
996 gmx_cpuid_simd_string[hwinfo->simd_suggest_min]);
997 if (hwinfo->simd_suggest_max > hwinfo->simd_suggest_min)
999 s += gmx::formatString(" - %s",
1000 gmx_cpuid_simd_string[hwinfo->simd_suggest_max]);
1002 s += gmx::formatString("\n");
1003 s += gmx::formatString(" SIMD instructions selected at GROMACS compile time: %s\n",
1004 gmx_cpuid_simd_string[gmx_compiled_simd()]);
1005 if (bGPUBinary && (hwinfo->ngpu_compatible_tot > 0 ||
1006 hwinfo->gpu_info.n_dev > 0))
1008 s += gmx::formatString(" GPU info:\n");
1009 s += gmx::formatString(" Number of GPUs detected: %d\n",
1010 hwinfo->gpu_info.n_dev);
1011 if (hwinfo->gpu_info.n_dev > 0)
1015 sprint_gpus(buf, &hwinfo->gpu_info);
1016 s += gmx::formatString("%s\n", buf);
1023 void gmx_print_detected_hardware(FILE *fplog, const t_commrec *cr,
1024 const gmx_hw_info_t *hwinfo)
1028 std::string detected;
1030 detected = detected_hardware_string(hwinfo, TRUE);
1032 fprintf(fplog, "%s\n", detected.c_str());
1035 if (MULTIMASTER(cr))
1037 std::string detected;
1039 detected = detected_hardware_string(hwinfo, FALSE);
1041 fprintf(stderr, "%s\n", detected.c_str());
1044 /* Check the compiled SIMD instruction set against that of the node
1045 * with the lowest SIMD level support.
1047 gmx_cpuid_simd_check(hwinfo->simd_suggest_min, fplog, MULTIMASTER(cr));
1049 /* For RDTSCP we only check on our local node and skip the MPI reduction */
1050 check_use_of_rdtscp_on_this_cpu(fplog, cr, hwinfo);
1053 void gmx_parse_gpu_ids(gmx_gpu_opt_t *gpu_opt)
1057 if (gpu_opt->gpu_id != NULL && !bGPUBinary)
1059 gmx_fatal(FARGS, "GPU ID string set, but %s was compiled without GPU support!", ShortProgram());
1062 env = getenv("GMX_GPU_ID");
1063 if (env != NULL && gpu_opt->gpu_id != NULL)
1065 gmx_fatal(FARGS, "GMX_GPU_ID and -gpu_id can not be used at the same time");
1069 env = gpu_opt->gpu_id;
1072 /* parse GPU IDs if the user passed any */
1075 /* Parse a "plain" GPU ID string which contains a sequence of
1076 * digits corresponding to GPU IDs; the order will indicate
1077 * the process/tMPI thread - GPU assignment. */
1078 parse_digits_from_plain_string(env,
1079 &gpu_opt->n_dev_use,
1082 if (gpu_opt->n_dev_use == 0)
1084 gmx_fatal(FARGS, "Empty GPU ID string encountered.\n%s\n",
1085 invalid_gpuid_hint);
1088 gpu_opt->bUserSet = TRUE;
1092 void gmx_select_gpu_ids(FILE *fplog, const t_commrec *cr,
1093 const gmx_gpu_info_t *gpu_info,
1094 gmx_bool bForceUseGPU,
1095 gmx_gpu_opt_t *gpu_opt)
1098 char sbuf[STRLEN], stmp[STRLEN];
1100 /* Bail if binary is not compiled with GPU acceleration, but this is either
1101 * explicitly (-nb gpu) or implicitly (gpu ID passed) requested. */
1102 if (bForceUseGPU && !bGPUBinary)
1104 gmx_fatal(FARGS, "GPU acceleration requested, but %s was compiled without GPU support!", ShortProgram());
1107 if (!(cr->duty & DUTY_PP))
1109 /* Our rank is not doing PP, we don't use a GPU */
1113 if (gpu_opt->bUserSet)
1115 /* Check the GPU IDs passed by the user.
1116 * (GPU IDs have been parsed by gmx_parse_gpu_ids before)
1121 snew(checkres, gpu_opt->n_dev_use);
1123 res = check_selected_gpus(checkres, gpu_info, gpu_opt);
1127 print_gpu_detection_stats(fplog, gpu_info, cr);
1129 sprintf(sbuf, "Some of the requested GPUs do not exist, behave strangely, or are not compatible:\n");
1130 for (i = 0; i < gpu_opt->n_dev_use; i++)
1132 if (checkres[i] != egpuCompatible)
1134 sprintf(stmp, " GPU #%d: %s\n",
1135 gpu_opt->dev_use[i],
1136 gpu_detect_res_str[checkres[i]]);
1140 gmx_fatal(FARGS, "%s", sbuf);
1145 else if (getenv("GMX_EMULATE_GPU") == NULL)
1147 pick_compatible_gpus(&hwinfo_g->gpu_info, gpu_opt);
1148 set_gpu_ids(gpu_opt, cr->nrank_pp_intranode, cr->rank_pp_intranode);
1151 /* If the user asked for a GPU, check whether we have a GPU */
1152 if (bForceUseGPU && gpu_info->n_dev_compatible == 0)
1154 gmx_fatal(FARGS, "GPU acceleration requested, but no compatible GPUs were detected.");
1158 /* Select the GPUs we will use. This is an operation local to each physical
1159 * node. If we have less MPI ranks than GPUs, we will waste some GPUs.
1160 * nrank and rank are the rank count and id for PP processes in our node.
1162 static void set_gpu_ids(gmx_gpu_opt_t *gpu_opt, int nrank, int rank)
1164 GMX_RELEASE_ASSERT(gpu_opt, "Invalid gpu_opt pointer passed");
1165 GMX_RELEASE_ASSERT(nrank >= 1,
1166 gmx::formatString("Invalid limit (%d) for the number of GPUs (detected %d compatible GPUs)",
1167 rank, gpu_opt->n_dev_compatible).c_str());
1169 if (gpu_opt->n_dev_compatible == 0)
1171 char host[HOSTNAMELEN];
1173 gmx_gethostname(host, HOSTNAMELEN);
1174 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);
1180 if (nrank > gpu_opt->n_dev_compatible)
1182 if (nrank % gpu_opt->n_dev_compatible == 0)
1184 nshare = nrank/gpu_opt->n_dev_compatible;
1190 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.",
1191 nrank, gpu_opt->n_dev_compatible);
1195 /* We use a global barrier to prevent ranks from continuing with
1198 MPI_Barrier(MPI_COMM_WORLD);
1203 /* Here we will waste GPUs when nrank < gpu_opt->n_dev_compatible */
1204 gpu_opt->n_dev_use = std::min(gpu_opt->n_dev_compatible*nshare, nrank);
1205 snew(gpu_opt->dev_use, gpu_opt->n_dev_use);
1206 for (int i = 0; i != gpu_opt->n_dev_use; ++i)
1208 /* TODO: improve this implementation: either sort GPUs or remove the weakest here */
1209 gpu_opt->dev_use[i] = gpu_opt->dev_compatible[i/nshare];
1213 void gmx_hardware_info_free(gmx_hw_info_t *hwinfo)
1217 ret = tMPI_Thread_mutex_lock(&hw_info_lock);
1220 gmx_fatal(FARGS, "Error locking hwinfo mutex: %s", strerror(errno));
1223 /* decrease the reference counter */
1227 if (hwinfo != hwinfo_g)
1229 gmx_incons("hwinfo < hwinfo_g");
1234 gmx_incons("n_hwinfo < 0");
1239 gmx_cpuid_done(hwinfo_g->cpuid_info);
1240 free_gpu_info(&hwinfo_g->gpu_info);
1244 ret = tMPI_Thread_mutex_unlock(&hw_info_lock);
1247 gmx_fatal(FARGS, "Error unlocking hwinfo mutex: %s", strerror(errno));