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, 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);
314 if (hwinfo->gpu_info.n_dev_compatible > 0)
316 std::string gpuUseageReport;
319 gpuUseageReport = makeGpuUsageReport(&hwinfo->gpu_info,
321 cr->nrank_pp_intranode,
322 bMPI && cr->nnodes > 1);
324 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
326 /* NOTE: this print is only for and on one physical node */
327 md_print_info(cr, fplog, "%s\n", gpuUseageReport.c_str());
330 /* Need to ensure that we have enough GPUs:
331 * - need one GPU per PP node
332 * - no GPU oversubscription with tMPI
334 /* number of PP processes per node */
335 npppn = cr->nrank_pp_intranode;
338 th_or_proc_plural[0] = '\0';
341 sprintf(th_or_proc, "thread-MPI thread");
344 sprintf(th_or_proc_plural, "s");
349 sprintf(th_or_proc, "MPI process");
352 sprintf(th_or_proc_plural, "es");
354 sprintf(pernode, " per node");
358 /* neither MPI nor tMPI */
359 sprintf(th_or_proc, "process");
362 if (bUseGPU && hwinfo->gpu_info.n_dev_compatible > 0 &&
365 int ngpu_comp, ngpu_use;
366 char gpu_comp_plural[2], gpu_use_plural[2];
368 ngpu_comp = hwinfo->gpu_info.n_dev_compatible;
369 ngpu_use = hw_opt->gpu_opt.n_dev_use;
371 sprintf(gpu_comp_plural, "%s", (ngpu_comp > 1) ? "s" : "");
372 sprintf(gpu_use_plural, "%s", (ngpu_use > 1) ? "s" : "");
374 /* number of tMPI threads auto-adjusted */
375 if (btMPI && bNthreadsAuto)
377 if (hw_opt->gpu_opt.bUserSet && npppn < ngpu_use)
379 /* The user manually provided more GPUs than threads we
380 could automatically start. */
382 "%d GPU%s provided, but only %d PP thread-MPI thread%s coud be started.\n"
383 "%s requires one PP tread-MPI thread per GPU; use fewer GPUs.",
384 ngpu_use, gpu_use_plural,
385 npppn, th_or_proc_plural,
389 if (!hw_opt->gpu_opt.bUserSet && npppn < ngpu_comp)
391 /* There are more GPUs than tMPI threads; we have
392 limited the number GPUs used. */
393 md_print_warn(cr, fplog,
394 "NOTE: %d GPU%s were detected, but only %d PP thread-MPI thread%s can be started.\n"
395 " %s can use one GPU per PP tread-MPI thread, so only %d GPU%s will be used.\n",
396 ngpu_comp, gpu_comp_plural,
397 npppn, th_or_proc_plural,
398 ShortProgram(), npppn,
399 npppn > 1 ? "s" : "");
403 if (hw_opt->gpu_opt.bUserSet)
405 if (ngpu_use != npppn)
408 "Incorrect launch configuration: mismatching number of PP %s%s and GPUs%s.\n"
409 "%s was started with %d PP %s%s%s, but you provided %d GPU%s.",
410 th_or_proc, btMPI ? "s" : "es", pernode,
411 ShortProgram(), npppn, th_or_proc,
412 th_or_proc_plural, pernode,
413 ngpu_use, gpu_use_plural);
418 if (ngpu_comp > npppn)
420 md_print_warn(cr, fplog,
421 "NOTE: potentially sub-optimal launch configuration, %s started with less\n"
422 " PP %s%s%s than GPU%s available.\n"
423 " Each PP %s can use only one GPU, %d GPU%s%s will be used.\n",
424 ShortProgram(), th_or_proc,
425 th_or_proc_plural, pernode, gpu_comp_plural,
426 th_or_proc, npppn, gpu_use_plural, pernode);
429 if (ngpu_use != npppn)
431 /* Avoid duplicate error messages.
432 * Unfortunately we can only do this at the physical node
433 * level, since the hardware setup and MPI process count
434 * might differ between physical nodes.
436 if (cr->rank_pp_intranode == 0)
439 "Incorrect launch configuration: mismatching number of PP %s%s and GPUs%s.\n"
440 "%s was started with %d PP %s%s%s, but only %d GPU%s were detected.",
441 th_or_proc, btMPI ? "s" : "es", pernode,
442 ShortProgram(), npppn, th_or_proc,
443 th_or_proc_plural, pernode,
444 ngpu_use, gpu_use_plural);
452 same_count = gmx_count_gpu_dev_shared(&hw_opt->gpu_opt);
456 md_print_info(cr, fplog,
457 "NOTE: You assigned %s to multiple %s%s.\n",
458 same_count > 1 ? "GPUs" : "a GPU", th_or_proc, btMPI ? "s" : "es");
466 /* Avoid other ranks to continue after
468 MPI_Barrier(cr->mpi_comm_mygroup);
474 /* Return 0 if none of the GPU (per node) are shared among PP ranks.
476 * Sharing GPUs among multiple PP ranks is possible when the user passes
477 * GPU IDs. Here we check for sharing and return a non-zero value when
478 * this is detected. Note that the return value represents the number of
479 * PP rank pairs that share a device.
481 int gmx_count_gpu_dev_shared(const gmx_gpu_opt_t *gpu_opt)
484 int ngpu = gpu_opt->n_dev_use;
486 if (gpu_opt->bUserSet)
490 for (i = 0; i < ngpu - 1; i++)
492 for (j = i + 1; j < ngpu; j++)
494 same_count += (gpu_opt->dev_use[i] ==
495 gpu_opt->dev_use[j]);
503 /* Count and return the number of unique GPUs (per node) selected.
505 * As sharing GPUs among multiple PP ranks is possible when the user passes
506 * GPU IDs, the number of GPUs user (per node) can be different from the
507 * number of GPU IDs selected.
509 static int gmx_count_gpu_dev_unique(const gmx_gpu_info_t *gpu_info,
510 const gmx_gpu_opt_t *gpu_opt)
512 int i, uniq_count, ngpu;
518 ngpu = gpu_info->n_dev;
522 snew(uniq_ids, ngpu);
524 /* Each element in uniq_ids will be set to 0 or 1. The n-th element set
525 * to 1 indicates that the respective GPU was selected to be used. */
526 for (i = 0; i < gpu_opt->n_dev_use; i++)
528 uniq_ids[get_cuda_gpu_device_id(gpu_info, gpu_opt, i)] = 1;
530 /* Count the devices used. */
531 for (i = 0; i < ngpu; i++)
533 uniq_count += uniq_ids[i];
541 static int get_ncores(gmx_cpuid_t cpuid)
543 int nprocessors, npackages, ncores_per_package, nhwthreads_per_core;
544 const int *package_id, *core_id, *hwthread_id, *locality_order;
547 rc = gmx_cpuid_topology(cpuid,
548 &nprocessors, &npackages,
549 &ncores_per_package, &nhwthreads_per_core,
550 &package_id, &core_id,
551 &hwthread_id, &locality_order);
555 return npackages*ncores_per_package;
559 /* We don't have cpuid topology info, return 0 core count */
564 /* Return the number of hardware threads supported by the current CPU.
565 * We assume that this is equal with the number of "processors"
566 * reported to be online by the OS at the time of the call. The
567 * definition of "processor" is according to an old POSIX standard.
569 * Note that the number of hardware threads is generally greater than
570 * the number of cores (e.g. x86 hyper-threading, Power). Managing the
571 * mapping of software threads to hardware threads is managed
573 static int get_nthreads_hw_avail(FILE gmx_unused *fplog, const t_commrec gmx_unused *cr)
577 #if ((defined(WIN32) || defined( _WIN32 ) || defined(WIN64) || defined( _WIN64 )) && !(defined (__CYGWIN__) || defined (__CYGWIN32__)))
580 GetSystemInfo( &sysinfo );
581 ret = sysinfo.dwNumberOfProcessors;
582 #elif defined HAVE_SYSCONF
583 /* We are probably on Unix.
584 * Now check if we have the argument to use before executing the call
586 #if defined(_SC_NPROCESSORS_ONLN)
587 ret = sysconf(_SC_NPROCESSORS_ONLN);
588 #elif defined(_SC_NPROC_ONLN)
589 ret = sysconf(_SC_NPROC_ONLN);
590 #elif defined(_SC_NPROCESSORS_CONF)
591 ret = sysconf(_SC_NPROCESSORS_CONF);
592 #elif defined(_SC_NPROC_CONF)
593 ret = sysconf(_SC_NPROC_CONF);
595 #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!"
596 #endif /* End of check for sysconf argument values */
599 /* Neither windows nor Unix. No fscking idea how many hardware threads we have! */
605 fprintf(debug, "Detected %d hardware threads to use.\n", ret);
609 if (ret != gmx_omp_get_num_procs())
611 md_print_warn(cr, fplog,
612 "Number of logical cores detected (%d) does not match the number reported by OpenMP (%d).\n"
613 "Consider setting the launch configuration manually!",
614 ret, gmx_omp_get_num_procs());
621 static void gmx_detect_gpus(FILE *fplog, const t_commrec *cr)
625 MPI_Comm physicalnode_comm;
629 /* Under certain circumstances MPI ranks on the same physical node
630 * can not simultaneously access the same GPU(s). Therefore we run
631 * the detection only on one MPI rank per node and broadcast the info.
632 * Note that with thread-MPI only a single thread runs this code.
634 * TODO: We should also do CPU hardware detection only once on each
635 * physical node and broadcast it, instead of do it on every MPI rank.
638 /* A split of MPI_COMM_WORLD over physical nodes is only required here,
639 * so we create and destroy it locally.
641 MPI_Comm_rank(MPI_COMM_WORLD, &rank_world);
642 MPI_Comm_split(MPI_COMM_WORLD, gmx_physicalnode_id_hash(),
643 rank_world, &physicalnode_comm);
644 MPI_Comm_rank(physicalnode_comm, &rank_local);
646 /* Here there should be only one process, check this */
647 assert(cr->nnodes == 1 && cr->sim_nodeid == 0);
654 char detection_error[STRLEN] = "", sbuf[STRLEN];
656 if (detect_gpus(&hwinfo_g->gpu_info, detection_error) != 0)
658 if (detection_error[0] != '\0')
660 sprintf(sbuf, ":\n %s\n", detection_error);
666 md_print_warn(cr, fplog,
667 "NOTE: Error occurred during GPU detection%s"
668 " Can not use GPU acceleration, will fall back to CPU kernels.\n",
674 /* Broadcast the GPU info to the other ranks within this node */
675 MPI_Bcast(&hwinfo_g->gpu_info.n_dev, 1, MPI_INT, 0, physicalnode_comm);
677 if (hwinfo_g->gpu_info.n_dev > 0)
681 dev_size = hwinfo_g->gpu_info.n_dev*sizeof_gpu_dev_info();
685 hwinfo_g->gpu_info.gpu_dev =
686 (struct gmx_device_info_t *)malloc(dev_size);
688 MPI_Bcast(hwinfo_g->gpu_info.gpu_dev, dev_size, MPI_BYTE,
689 0, physicalnode_comm);
690 MPI_Bcast(&hwinfo_g->gpu_info.n_dev_compatible, 1, MPI_INT,
691 0, physicalnode_comm);
694 MPI_Comm_free(&physicalnode_comm);
698 static void gmx_collect_hardware_mpi()
702 int nrank, rank, ncore, nhwthread, ngpu, i;
706 rank_id = gmx_physicalnode_id_hash();
707 MPI_Comm_rank(MPI_COMM_WORLD, &rank);
708 MPI_Comm_size(MPI_COMM_WORLD, &nrank);
709 ncore = hwinfo_g->ncore;
710 nhwthread = hwinfo_g->nthreads_hw_avail;
711 ngpu = hwinfo_g->gpu_info.n_dev_compatible;
712 /* Create a unique hash of the GPU type(s) in this node */
714 /* Here it might be better to only loop over the compatible GPU, but we
715 * don't have that information available and it would also require
716 * removing the device ID from the device info string.
718 for (i = 0; i < hwinfo_g->gpu_info.n_dev; i++)
722 /* Since the device ID is incorporated in the hash, the order of
723 * the GPUs affects the hash. Also two identical GPUs won't give
724 * a gpu_hash of zero after XORing.
726 get_gpu_device_info_string(stmp, &hwinfo_g->gpu_info, i);
727 gpu_hash ^= gmx_string_fullhash_func(stmp, gmx_string_hash_init);
734 MPI_Allreduce(buf, all, nrank, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
737 int nnode0, ncore0, nhwthread0, ngpu0, r;
744 for (r = 0; r < nrank; r++)
746 if (all[r] == rank_id)
748 if (!bFound && r == rank)
750 /* We are the first rank in this physical node */
753 nhwthread0 = nhwthread;
763 int sum[4], maxmin[10];
768 /* Sum values from only intra-rank 0 so we get the sum over all nodes */
774 MPI_Allreduce(buf, sum, 4, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
780 /* Store + and - values for all ranks,
781 * so we can get max+min with one MPI call.
786 buf[3] = gmx_cpuid_simd_suggest(hwinfo_g->cpuid_info);
794 MPI_Allreduce(buf, maxmin, 10, MPI_INT, MPI_MAX, MPI_COMM_WORLD);
797 hwinfo_g->nphysicalnode = sum[0];
798 hwinfo_g->ncore_tot = sum[1];
799 hwinfo_g->ncore_min = -maxmin[5];
800 hwinfo_g->ncore_max = maxmin[0];
801 hwinfo_g->nhwthread_tot = sum[2];
802 hwinfo_g->nhwthread_min = -maxmin[6];
803 hwinfo_g->nhwthread_max = maxmin[1];
804 hwinfo_g->ngpu_compatible_tot = sum[3];
805 hwinfo_g->ngpu_compatible_min = -maxmin[7];
806 hwinfo_g->ngpu_compatible_max = maxmin[2];
807 hwinfo_g->simd_suggest_min = static_cast<enum gmx_cpuid_simd>(-maxmin[8]);
808 hwinfo_g->simd_suggest_max = static_cast<enum gmx_cpuid_simd>(maxmin[3]);
809 hwinfo_g->bIdenticalGPUs = (maxmin[4] == -maxmin[9]);
811 /* All ranks use the same pointer, protect it with a mutex */
812 tMPI_Thread_mutex_lock(&hw_info_lock);
813 hwinfo_g->nphysicalnode = 1;
814 hwinfo_g->ncore_tot = hwinfo_g->ncore;
815 hwinfo_g->ncore_min = hwinfo_g->ncore;
816 hwinfo_g->ncore_max = hwinfo_g->ncore;
817 hwinfo_g->nhwthread_tot = hwinfo_g->nthreads_hw_avail;
818 hwinfo_g->nhwthread_min = hwinfo_g->nthreads_hw_avail;
819 hwinfo_g->nhwthread_max = hwinfo_g->nthreads_hw_avail;
820 hwinfo_g->ngpu_compatible_tot = hwinfo_g->gpu_info.n_dev_compatible;
821 hwinfo_g->ngpu_compatible_min = hwinfo_g->gpu_info.n_dev_compatible;
822 hwinfo_g->ngpu_compatible_max = hwinfo_g->gpu_info.n_dev_compatible;
823 hwinfo_g->simd_suggest_min = gmx_cpuid_simd_suggest(hwinfo_g->cpuid_info);
824 hwinfo_g->simd_suggest_max = gmx_cpuid_simd_suggest(hwinfo_g->cpuid_info);
825 hwinfo_g->bIdenticalGPUs = TRUE;
826 tMPI_Thread_mutex_unlock(&hw_info_lock);
830 gmx_hw_info_t *gmx_detect_hardware(FILE *fplog, const t_commrec *cr,
831 gmx_bool bDetectGPUs)
835 /* make sure no one else is doing the same thing */
836 ret = tMPI_Thread_mutex_lock(&hw_info_lock);
839 gmx_fatal(FARGS, "Error locking hwinfo mutex: %s", strerror(errno));
842 /* only initialize the hwinfo structure if it is not already initalized */
847 /* detect CPUID info; no fuss, we don't detect system-wide
848 * -- sloppy, but that's it for now */
849 if (gmx_cpuid_init(&hwinfo_g->cpuid_info) != 0)
851 gmx_fatal_collective(FARGS, cr, NULL, "CPUID detection failed!");
854 /* get the number of cores, will be 0 when not detected */
855 hwinfo_g->ncore = get_ncores(hwinfo_g->cpuid_info);
857 /* detect number of hardware threads */
858 hwinfo_g->nthreads_hw_avail = get_nthreads_hw_avail(fplog, cr);
861 hwinfo_g->gpu_info.n_dev = 0;
862 hwinfo_g->gpu_info.n_dev_compatible = 0;
863 hwinfo_g->gpu_info.gpu_dev = NULL;
865 /* Run the detection if the binary was compiled with GPU support
866 * and we requested detection.
868 hwinfo_g->gpu_info.bDetectGPUs =
869 (bGPUBinary && bDetectGPUs &&
870 getenv("GMX_DISABLE_GPU_DETECTION") == NULL);
871 if (hwinfo_g->gpu_info.bDetectGPUs)
873 gmx_detect_gpus(fplog, cr);
876 /* increase the reference counter */
879 ret = tMPI_Thread_mutex_unlock(&hw_info_lock);
882 gmx_fatal(FARGS, "Error unlocking hwinfo mutex: %s", strerror(errno));
885 gmx_collect_hardware_mpi();
890 static std::string detected_hardware_string(const gmx_hw_info_t *hwinfo,
895 s = gmx::formatString("\n");
896 s += gmx::formatString("Running on %d node%s with total",
897 hwinfo->nphysicalnode,
898 hwinfo->nphysicalnode == 1 ? "" : "s");
899 if (hwinfo->ncore_tot > 0)
901 s += gmx::formatString(" %d cores,", hwinfo->ncore_tot);
903 s += gmx::formatString(" %d logical cores", hwinfo->nhwthread_tot);
904 if (hwinfo->gpu_info.bDetectGPUs)
906 s += gmx::formatString(", %d compatible GPU%s",
907 hwinfo->ngpu_compatible_tot,
908 hwinfo->ngpu_compatible_tot == 1 ? "" : "s");
912 s += gmx::formatString(" (GPU detection deactivated)");
914 s += gmx::formatString("\n");
916 if (hwinfo->nphysicalnode > 1)
918 /* Print per node hardware feature counts */
919 if (hwinfo->ncore_max > 0)
921 s += gmx::formatString(" Cores per node: %2d", hwinfo->ncore_min);
922 if (hwinfo->ncore_max > hwinfo->ncore_min)
924 s += gmx::formatString(" - %2d", hwinfo->ncore_max);
926 s += gmx::formatString("\n");
928 s += gmx::formatString(" Logical cores per node: %2d", hwinfo->nhwthread_min);
929 if (hwinfo->nhwthread_max > hwinfo->nhwthread_min)
931 s += gmx::formatString(" - %2d", hwinfo->nhwthread_max);
933 s += gmx::formatString("\n");
936 s += gmx::formatString(" Compatible GPUs per node: %2d",
937 hwinfo->ngpu_compatible_min);
938 if (hwinfo->ngpu_compatible_max > hwinfo->ngpu_compatible_min)
940 s += gmx::formatString(" - %2d", hwinfo->ngpu_compatible_max);
942 s += gmx::formatString("\n");
943 if (hwinfo->ngpu_compatible_tot > 0)
945 if (hwinfo->bIdenticalGPUs)
947 s += gmx::formatString(" All nodes have identical type(s) of GPUs\n");
951 /* This message will also appear with identical GPU types
952 * when at least one node has no GPU.
954 s += gmx::formatString(" Different nodes have different type(s) and/or order of GPUs\n");
961 char host[HOSTNAMELEN];
964 gmx_gethostname(host, HOSTNAMELEN);
965 MPI_Comm_rank(MPI_COMM_WORLD, &rank);
967 s += gmx::formatString("Hardware detected on host %s (the node of MPI rank %d):\n",
970 s += gmx::formatString("Hardware detected:\n");
972 s += gmx::formatString(" CPU info:\n");
977 gmx_cpuid_formatstring(hwinfo->cpuid_info, buf, 1023);
980 s += gmx::formatString("%s", buf);
984 s += gmx::formatString(" Vendor: %s\n",
985 gmx_cpuid_vendor_string[gmx_cpuid_vendor(hwinfo->cpuid_info)]);
986 s += gmx::formatString(" Brand: %s\n",
987 gmx_cpuid_brand(hwinfo->cpuid_info));
989 s += gmx::formatString(" SIMD instructions most likely to fit this hardware: %s",
990 gmx_cpuid_simd_string[hwinfo->simd_suggest_min]);
991 if (hwinfo->simd_suggest_max > hwinfo->simd_suggest_min)
993 s += gmx::formatString(" - %s",
994 gmx_cpuid_simd_string[hwinfo->simd_suggest_max]);
996 s += gmx::formatString("\n");
997 s += gmx::formatString(" SIMD instructions selected at GROMACS compile time: %s\n",
998 gmx_cpuid_simd_string[gmx_compiled_simd()]);
999 if (bGPUBinary && (hwinfo->ngpu_compatible_tot > 0 ||
1000 hwinfo->gpu_info.n_dev > 0))
1002 s += gmx::formatString(" GPU info:\n");
1003 s += gmx::formatString(" Number of GPUs detected: %d\n",
1004 hwinfo->gpu_info.n_dev);
1005 if (hwinfo->gpu_info.n_dev > 0)
1009 sprint_gpus(buf, &hwinfo->gpu_info);
1010 s += gmx::formatString("%s\n", buf);
1017 void gmx_print_detected_hardware(FILE *fplog, const t_commrec *cr,
1018 const gmx_hw_info_t *hwinfo)
1022 std::string detected;
1024 detected = detected_hardware_string(hwinfo, TRUE);
1026 fprintf(fplog, "%s\n", detected.c_str());
1029 if (MULTIMASTER(cr))
1031 std::string detected;
1033 detected = detected_hardware_string(hwinfo, FALSE);
1035 fprintf(stderr, "%s\n", detected.c_str());
1038 /* Check the compiled SIMD instruction set against that of the node
1039 * with the lowest SIMD level support.
1041 gmx_cpuid_simd_check(hwinfo->simd_suggest_min, fplog, MULTIMASTER(cr));
1043 /* For RDTSCP we only check on our local node and skip the MPI reduction */
1044 check_use_of_rdtscp_on_this_cpu(fplog, cr, hwinfo);
1047 void gmx_parse_gpu_ids(gmx_gpu_opt_t *gpu_opt)
1051 if (gpu_opt->gpu_id != NULL && !bGPUBinary)
1053 gmx_fatal(FARGS, "GPU ID string set, but %s was compiled without GPU support!", ShortProgram());
1056 env = getenv("GMX_GPU_ID");
1057 if (env != NULL && gpu_opt->gpu_id != NULL)
1059 gmx_fatal(FARGS, "GMX_GPU_ID and -gpu_id can not be used at the same time");
1063 env = gpu_opt->gpu_id;
1066 /* parse GPU IDs if the user passed any */
1069 /* Parse a "plain" GPU ID string which contains a sequence of
1070 * digits corresponding to GPU IDs; the order will indicate
1071 * the process/tMPI thread - GPU assignment. */
1072 parse_digits_from_plain_string(env,
1073 &gpu_opt->n_dev_use,
1076 if (gpu_opt->n_dev_use == 0)
1078 gmx_fatal(FARGS, "Empty GPU ID string encountered.\n%s\n",
1079 invalid_gpuid_hint);
1082 gpu_opt->bUserSet = TRUE;
1086 void gmx_select_gpu_ids(FILE *fplog, const t_commrec *cr,
1087 const gmx_gpu_info_t *gpu_info,
1088 gmx_bool bForceUseGPU,
1089 gmx_gpu_opt_t *gpu_opt)
1092 char sbuf[STRLEN], stmp[STRLEN];
1094 /* Bail if binary is not compiled with GPU acceleration, but this is either
1095 * explicitly (-nb gpu) or implicitly (gpu ID passed) requested. */
1096 if (bForceUseGPU && !bGPUBinary)
1098 gmx_fatal(FARGS, "GPU acceleration requested, but %s was compiled without GPU support!", ShortProgram());
1101 if (!(cr->duty & DUTY_PP))
1103 /* Our rank is not doing PP, we don't use a GPU */
1107 if (gpu_opt->bUserSet)
1109 /* Check the GPU IDs passed by the user.
1110 * (GPU IDs have been parsed by gmx_parse_gpu_ids before)
1115 snew(checkres, gpu_opt->n_dev_use);
1117 res = check_selected_gpus(checkres, gpu_info, gpu_opt);
1121 print_gpu_detection_stats(fplog, gpu_info, cr);
1123 sprintf(sbuf, "Some of the requested GPUs do not exist, behave strangely, or are not compatible:\n");
1124 for (i = 0; i < gpu_opt->n_dev_use; i++)
1126 if (checkres[i] != egpuCompatible)
1128 sprintf(stmp, " GPU #%d: %s\n",
1129 gpu_opt->dev_use[i],
1130 gpu_detect_res_str[checkres[i]]);
1134 gmx_fatal(FARGS, "%s", sbuf);
1139 else if (getenv("GMX_EMULATE_GPU") == NULL)
1141 pick_compatible_gpus(&hwinfo_g->gpu_info, gpu_opt);
1142 set_gpu_ids(gpu_opt, cr->nrank_pp_intranode, cr->rank_pp_intranode);
1145 /* If the user asked for a GPU, check whether we have a GPU */
1146 if (bForceUseGPU && gpu_info->n_dev_compatible == 0)
1148 gmx_fatal(FARGS, "GPU acceleration requested, but no compatible GPUs were detected.");
1152 /* Select the GPUs we will use. This is an operation local to each physical
1153 * node. If we have less MPI ranks than GPUs, we will waste some GPUs.
1154 * nrank and rank are the rank count and id for PP processes in our node.
1156 static void set_gpu_ids(gmx_gpu_opt_t *gpu_opt, int nrank, int rank)
1158 GMX_RELEASE_ASSERT(gpu_opt, "Invalid gpu_opt pointer passed");
1159 GMX_RELEASE_ASSERT(nrank >= 1,
1160 gmx::formatString("Invalid limit (%d) for the number of GPUs (detected %d compatible GPUs)",
1161 rank, gpu_opt->n_dev_compatible).c_str());
1163 if (gpu_opt->n_dev_compatible == 0)
1165 char host[HOSTNAMELEN];
1167 gmx_gethostname(host, HOSTNAMELEN);
1168 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);
1174 if (nrank > gpu_opt->n_dev_compatible)
1176 if (nrank % gpu_opt->n_dev_compatible == 0)
1178 nshare = nrank/gpu_opt->n_dev_compatible;
1184 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.",
1185 nrank, gpu_opt->n_dev_compatible);
1189 /* We use a global barrier to prevent ranks from continuing with
1192 MPI_Barrier(MPI_COMM_WORLD);
1197 /* Here we will waste GPUs when nrank < gpu_opt->n_dev_compatible */
1198 gpu_opt->n_dev_use = std::min(gpu_opt->n_dev_compatible*nshare, nrank);
1199 snew(gpu_opt->dev_use, gpu_opt->n_dev_use);
1200 for (int i = 0; i != gpu_opt->n_dev_use; ++i)
1202 /* TODO: improve this implementation: either sort GPUs or remove the weakest here */
1203 gpu_opt->dev_use[i] = gpu_opt->dev_compatible[i/nshare];
1207 void gmx_hardware_info_free(gmx_hw_info_t *hwinfo)
1211 ret = tMPI_Thread_mutex_lock(&hw_info_lock);
1214 gmx_fatal(FARGS, "Error locking hwinfo mutex: %s", strerror(errno));
1217 /* decrease the reference counter */
1221 if (hwinfo != hwinfo_g)
1223 gmx_incons("hwinfo < hwinfo_g");
1228 gmx_incons("n_hwinfo < 0");
1233 gmx_cpuid_done(hwinfo_g->cpuid_info);
1234 free_gpu_info(&hwinfo_g->gpu_info);
1238 ret = tMPI_Thread_mutex_unlock(&hw_info_lock);
1241 gmx_fatal(FARGS, "Error unlocking hwinfo mutex: %s", strerror(errno));