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;
81 # ifdef GMX_USE_OPENCL
82 const char *gpu_implementation = "OpenCL";
83 /* Our current OpenCL implementation only supports using exactly one
84 * GPU per PP rank, so sharing is impossible */
85 const gmx_bool bGpuSharingSupported = FALSE;
86 /* Our current OpenCL implementation is not known to handle
87 * concurrency correctly (at context creation, JIT compilation, or JIT
88 * cache-management stages). OpenCL runtimes need not support it
89 * either; library MPI segfaults when creating OpenCL contexts;
90 * thread-MPI seems to work but is not yet known to be safe. */
91 const gmx_bool bMultiGpuPerNodeSupported = FALSE;
93 const char *gpu_implementation = "CUDA";
94 const gmx_bool bGpuSharingSupported = TRUE;
95 const gmx_bool bMultiGpuPerNodeSupported = TRUE;
98 const gmx_bool bGPUBinary = FALSE;
99 const char *gpu_implementation = "non-GPU";
100 const gmx_bool bGpuSharingSupported = FALSE;
101 const gmx_bool bMultiGpuPerNodeSupported = FALSE;
104 /* Names of the GPU detection/check results (see e_gpu_detect_res_t in hw_info.h). */
105 const char * const gpu_detect_res_str[egpuNR] =
107 "compatible", "inexistent", "incompatible", "insane"
110 static const char * invalid_gpuid_hint =
111 "A delimiter-free sequence of valid numeric IDs of available GPUs is expected.";
113 /* The globally shared hwinfo structure. */
114 static gmx_hw_info_t *hwinfo_g;
115 /* A reference counter for the hwinfo structure */
116 static int n_hwinfo = 0;
117 /* A lock to protect the hwinfo structure */
118 static tMPI_Thread_mutex_t hw_info_lock = TMPI_THREAD_MUTEX_INITIALIZER;
120 #define HOSTNAMELEN 80
123 static void set_gpu_ids(gmx_gpu_opt_t *gpu_opt, int nrank, int rank);
124 static int gmx_count_gpu_dev_unique(const gmx_gpu_info_t *gpu_info,
125 const gmx_gpu_opt_t *gpu_opt);
127 static void sprint_gpus(char *sbuf, const gmx_gpu_info_t *gpu_info)
132 ndev = gpu_info->n_dev;
135 for (i = 0; i < ndev; i++)
137 get_gpu_device_info_string(stmp, gpu_info, i);
147 static void print_gpu_detection_stats(FILE *fplog,
148 const gmx_gpu_info_t *gpu_info,
151 char onhost[HOSTNAMELEN+10], stmp[STRLEN];
154 if (!gpu_info->bDetectGPUs)
156 /* We skipped the detection, so don't print detection stats */
160 ngpu = gpu_info->n_dev;
162 #if defined GMX_MPI && !defined GMX_THREAD_MPI
163 /* We only print the detection on one, of possibly multiple, nodes */
164 strncpy(onhost, " on host ", 10);
165 gmx_gethostname(onhost + 9, HOSTNAMELEN);
167 /* We detect all relevant GPUs */
168 strncpy(onhost, "", 1);
173 sprint_gpus(stmp, gpu_info);
174 md_print_warn(cr, fplog, "%d GPU%s detected%s:\n%s\n",
175 ngpu, (ngpu > 1) ? "s" : "", onhost, stmp);
179 md_print_warn(cr, fplog, "No GPUs detected%s\n", onhost);
183 /*! \brief Helper function for reporting GPU usage information
184 * in the mdrun log file
186 * \param[in] gpu_info Pointer to per-node GPU info struct
187 * \param[in] gpu_opt Pointer to per-node GPU options struct
188 * \param[in] numPpRanks Number of PP ranks per node
189 * \param[in] bPrintHostName Print the hostname in the usage information
190 * \return String to write to the log file
191 * \throws std::bad_alloc if out of memory */
193 makeGpuUsageReport(const gmx_gpu_info_t *gpu_info,
194 const gmx_gpu_opt_t *gpu_opt,
198 int ngpu_use = gpu_opt->n_dev_use;
199 int ngpu_comp = gpu_info->n_dev_compatible;
200 char host[HOSTNAMELEN];
204 gmx_gethostname(host, HOSTNAMELEN);
207 /* Issue a note if GPUs are available but not used */
208 if (ngpu_comp > 0 && ngpu_use < 1)
210 return gmx::formatString("%d compatible GPU%s detected in the system, but none will be used.\n"
211 "Consider trying GPU acceleration with the Verlet scheme!\n",
212 ngpu_comp, (ngpu_comp > 1) ? "s" : "");
216 if (!gpu_opt->bUserSet)
218 // gpu_opt->dev_compatible is only populated during auto-selection
219 std::string gpuIdsString =
220 formatAndJoin(gmx::constArrayRefFromArray(gpu_opt->dev_compatible,
221 gpu_opt->n_dev_compatible),
222 ",", gmx::StringFormatter("%d"));
223 bool bPluralGpus = gpu_opt->n_dev_compatible > 1;
227 output += gmx::formatString("On host %s ", host);
229 output += gmx::formatString("%d compatible GPU%s %s present, with ID%s %s\n",
230 gpu_opt->n_dev_compatible,
231 bPluralGpus ? "s" : "",
232 bPluralGpus ? "are" : "is",
233 bPluralGpus ? "s" : "",
234 gpuIdsString.c_str());
238 std::vector<int> gpuIdsInUse;
239 for (int i = 0; i < ngpu_use; i++)
241 gpuIdsInUse.push_back(get_gpu_device_id(gpu_info, gpu_opt, i));
243 std::string gpuIdsString =
244 formatAndJoin(gpuIdsInUse, ",", gmx::StringFormatter("%d"));
245 int numGpusInUse = gmx_count_gpu_dev_unique(gpu_info, gpu_opt);
246 bool bPluralGpus = numGpusInUse > 1;
250 output += gmx::formatString("On host %s ", host);
252 output += gmx::formatString("%d GPU%s %sselected for this run.\n"
253 "Mapping of GPU ID%s to the %d PP rank%s in this node: %s\n",
254 numGpusInUse, bPluralGpus ? "s" : "",
255 gpu_opt->bUserSet ? "user-" : "auto-",
256 bPluralGpus ? "s" : "",
258 (numPpRanks > 1) ? "s" : "",
259 gpuIdsString.c_str());
265 /* Give a suitable fatal error or warning if the build configuration
266 and runtime CPU do not match. */
268 check_use_of_rdtscp_on_this_cpu(FILE *fplog,
270 const gmx_hw_info_t *hwinfo)
272 gmx_bool bCpuHasRdtscp, bBinaryUsesRdtscp;
274 bBinaryUsesRdtscp = TRUE;
276 bBinaryUsesRdtscp = FALSE;
279 bCpuHasRdtscp = gmx_cpuid_feature(hwinfo->cpuid_info, GMX_CPUID_FEATURE_X86_RDTSCP);
281 if (!bCpuHasRdtscp && bBinaryUsesRdtscp)
283 gmx_fatal(FARGS, "The %s executable was compiled to use the rdtscp CPU instruction. "
284 "However, this is not supported by the current hardware and continuing would lead to a crash. "
285 "Please rebuild GROMACS with the GMX_USE_RDTSCP=OFF CMake option.",
289 if (bCpuHasRdtscp && !bBinaryUsesRdtscp)
291 md_print_warn(cr, fplog, "The current CPU can measure timings more accurately than the code in\n"
292 "%s was configured to use. This might affect your simulation\n"
293 "speed as accurate timings are needed for load-balancing.\n"
294 "Please consider rebuilding %s with the GMX_USE_RDTSCP=ON CMake option.\n",
295 ShortProgram(), ShortProgram());
299 void gmx_check_hw_runconf_consistency(FILE *fplog,
300 const gmx_hw_info_t *hwinfo,
302 const gmx_hw_opt_t *hw_opt,
306 char th_or_proc[STRLEN], th_or_proc_plural[STRLEN], pernode[STRLEN];
307 gmx_bool btMPI, bMPI, bNthreadsAuto, bEmulateGPU;
312 /* Below we only do consistency checks for PP and GPUs,
313 * this is irrelevant for PME only nodes, so in that case we return
316 if (!(cr->duty & DUTY_PP))
321 #if defined(GMX_THREAD_MPI)
324 bNthreadsAuto = (hw_opt->nthreads_tmpi < 1);
325 #elif defined(GMX_LIB_MPI)
328 bNthreadsAuto = FALSE;
332 bNthreadsAuto = FALSE;
335 /* GPU emulation detection is done later, but we need here as well
336 * -- uncool, but there's no elegant workaround */
337 bEmulateGPU = (getenv("GMX_EMULATE_GPU") != NULL);
339 if (hwinfo->gpu_info.n_dev_compatible > 0)
341 std::string gpuUseageReport;
344 gpuUseageReport = makeGpuUsageReport(&hwinfo->gpu_info,
346 cr->nrank_pp_intranode,
347 bMPI && cr->nnodes > 1);
349 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
351 /* NOTE: this print is only for and on one physical node */
352 md_print_info(cr, fplog, "%s\n", gpuUseageReport.c_str());
355 /* Need to ensure that we have enough GPUs:
356 * - need one GPU per PP node
357 * - no GPU oversubscription with tMPI
359 /* number of PP processes per node */
360 npppn = cr->nrank_pp_intranode;
363 th_or_proc_plural[0] = '\0';
366 sprintf(th_or_proc, "thread-MPI thread");
369 sprintf(th_or_proc_plural, "s");
374 sprintf(th_or_proc, "MPI process");
377 sprintf(th_or_proc_plural, "es");
379 sprintf(pernode, " per node");
383 /* neither MPI nor tMPI */
384 sprintf(th_or_proc, "process");
387 if (bUseGPU && hwinfo->gpu_info.n_dev_compatible > 0 &&
390 int ngpu_comp, ngpu_use;
391 char gpu_comp_plural[2], gpu_use_plural[2];
393 ngpu_comp = hwinfo->gpu_info.n_dev_compatible;
394 ngpu_use = hw_opt->gpu_opt.n_dev_use;
396 sprintf(gpu_comp_plural, "%s", (ngpu_comp > 1) ? "s" : "");
397 sprintf(gpu_use_plural, "%s", (ngpu_use > 1) ? "s" : "");
399 /* number of tMPI threads auto-adjusted */
400 if (btMPI && bNthreadsAuto)
402 if (hw_opt->gpu_opt.bUserSet && npppn < ngpu_use)
404 /* The user manually provided more GPUs than threads we
405 could automatically start. */
407 "%d GPU%s provided, but only %d PP thread-MPI thread%s coud be started.\n"
408 "%s requires one PP tread-MPI thread per GPU; use fewer GPUs.",
409 ngpu_use, gpu_use_plural,
410 npppn, th_or_proc_plural,
414 if (!hw_opt->gpu_opt.bUserSet && npppn < ngpu_comp)
416 /* There are more GPUs than tMPI threads; we have
417 limited the number GPUs used. */
418 md_print_warn(cr, fplog,
419 "NOTE: %d GPU%s were detected, but only %d PP thread-MPI thread%s can be started.\n"
420 " %s can use one GPU per PP tread-MPI thread, so only %d GPU%s will be used.\n",
421 ngpu_comp, gpu_comp_plural,
422 npppn, th_or_proc_plural,
423 ShortProgram(), npppn,
424 npppn > 1 ? "s" : "");
428 if (hw_opt->gpu_opt.bUserSet)
430 if (ngpu_use != npppn)
433 "Incorrect launch configuration: mismatching number of PP %s%s and GPUs%s.\n"
434 "%s was started with %d PP %s%s%s, but you provided %d GPU%s.",
435 th_or_proc, btMPI ? "s" : "es", pernode,
436 ShortProgram(), npppn, th_or_proc,
437 th_or_proc_plural, pernode,
438 ngpu_use, gpu_use_plural);
443 if (ngpu_comp > npppn)
445 md_print_warn(cr, fplog,
446 "NOTE: potentially sub-optimal launch configuration, %s started with less\n"
447 " PP %s%s%s than GPU%s available.\n"
448 " Each PP %s can use only one GPU, %d GPU%s%s will be used.\n",
449 ShortProgram(), th_or_proc,
450 th_or_proc_plural, pernode, gpu_comp_plural,
451 th_or_proc, npppn, gpu_use_plural, pernode);
454 if (ngpu_use != npppn)
456 /* Avoid duplicate error messages.
457 * Unfortunately we can only do this at the physical node
458 * level, since the hardware setup and MPI process count
459 * might differ between physical nodes.
461 if (cr->rank_pp_intranode == 0)
464 "Incorrect launch configuration: mismatching number of PP %s%s and GPUs%s.\n"
465 "%s was started with %d PP %s%s%s, but only %d GPU%s were detected.",
466 th_or_proc, btMPI ? "s" : "es", pernode,
467 ShortProgram(), npppn, th_or_proc,
468 th_or_proc_plural, pernode,
469 ngpu_use, gpu_use_plural);
477 same_count = gmx_count_gpu_dev_shared(&hw_opt->gpu_opt);
481 md_print_info(cr, fplog,
482 "NOTE: You assigned %s to multiple %s%s.\n",
483 same_count > 1 ? "GPUs" : "a GPU", th_or_proc, btMPI ? "s" : "es");
491 /* Avoid other ranks to continue after
493 MPI_Barrier(cr->mpi_comm_mygroup);
499 /* Return 0 if none of the GPU (per node) are shared among PP ranks.
501 * Sharing GPUs among multiple PP ranks is possible when the user passes
502 * GPU IDs. Here we check for sharing and return a non-zero value when
503 * this is detected. Note that the return value represents the number of
504 * PP rank pairs that share a device.
506 int gmx_count_gpu_dev_shared(const gmx_gpu_opt_t *gpu_opt)
509 int ngpu = gpu_opt->n_dev_use;
511 if (gpu_opt->bUserSet)
515 for (i = 0; i < ngpu - 1; i++)
517 for (j = i + 1; j < ngpu; j++)
519 same_count += (gpu_opt->dev_use[i] ==
520 gpu_opt->dev_use[j]);
528 /* Count and return the number of unique GPUs (per node) selected.
530 * As sharing GPUs among multiple PP ranks is possible when the user passes
531 * GPU IDs, the number of GPUs user (per node) can be different from the
532 * number of GPU IDs selected.
534 static int gmx_count_gpu_dev_unique(const gmx_gpu_info_t *gpu_info,
535 const gmx_gpu_opt_t *gpu_opt)
537 int i, uniq_count, ngpu;
543 ngpu = gpu_info->n_dev;
547 snew(uniq_ids, ngpu);
549 /* Each element in uniq_ids will be set to 0 or 1. The n-th element set
550 * to 1 indicates that the respective GPU was selected to be used. */
551 for (i = 0; i < gpu_opt->n_dev_use; i++)
555 device_id = bGpuSharingSupported ? get_gpu_device_id(gpu_info, gpu_opt, i) : i;
556 uniq_ids[device_id] = 1;
558 /* Count the devices used. */
559 for (i = 0; i < ngpu; i++)
561 uniq_count += uniq_ids[i];
569 static int get_ncores(gmx_cpuid_t cpuid)
571 int nprocessors, npackages, ncores_per_package, nhwthreads_per_core;
572 const int *package_id, *core_id, *hwthread_id, *locality_order;
575 rc = gmx_cpuid_topology(cpuid,
576 &nprocessors, &npackages,
577 &ncores_per_package, &nhwthreads_per_core,
578 &package_id, &core_id,
579 &hwthread_id, &locality_order);
583 return npackages*ncores_per_package;
587 /* We don't have cpuid topology info, return 0 core count */
592 /* Return the number of hardware threads supported by the current CPU.
593 * We assume that this is equal with the number of "processors"
594 * reported to be online by the OS at the time of the call. The
595 * definition of "processor" is according to an old POSIX standard.
597 * Note that the number of hardware threads is generally greater than
598 * the number of cores (e.g. x86 hyper-threading, Power). Managing the
599 * mapping of software threads to hardware threads is managed
601 static int get_nthreads_hw_avail(FILE gmx_unused *fplog, const t_commrec gmx_unused *cr)
605 #if ((defined(WIN32) || defined( _WIN32 ) || defined(WIN64) || defined( _WIN64 )) && !(defined (__CYGWIN__) || defined (__CYGWIN32__)))
608 GetSystemInfo( &sysinfo );
609 ret = sysinfo.dwNumberOfProcessors;
610 #elif defined HAVE_SYSCONF
611 /* We are probably on Unix.
612 * Now check if we have the argument to use before executing the call
614 #if defined(_SC_NPROCESSORS_ONLN)
615 ret = sysconf(_SC_NPROCESSORS_ONLN);
616 #elif defined(_SC_NPROC_ONLN)
617 ret = sysconf(_SC_NPROC_ONLN);
618 #elif defined(_SC_NPROCESSORS_CONF)
619 ret = sysconf(_SC_NPROCESSORS_CONF);
620 #elif defined(_SC_NPROC_CONF)
621 ret = sysconf(_SC_NPROC_CONF);
623 #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!"
624 #endif /* End of check for sysconf argument values */
627 /* Neither windows nor Unix. No fscking idea how many hardware threads we have! */
633 fprintf(debug, "Detected %d hardware threads to use.\n", ret);
637 if (ret != gmx_omp_get_num_procs())
639 md_print_warn(cr, fplog,
640 "Number of logical cores detected (%d) does not match the number reported by OpenMP (%d).\n"
641 "Consider setting the launch configuration manually!",
642 ret, gmx_omp_get_num_procs());
649 static void gmx_detect_gpus(FILE *fplog, const t_commrec *cr)
653 MPI_Comm physicalnode_comm;
657 /* Under certain circumstances MPI ranks on the same physical node
658 * can not simultaneously access the same GPU(s). Therefore we run
659 * the detection only on one MPI rank per node and broadcast the info.
660 * Note that with thread-MPI only a single thread runs this code.
662 * TODO: We should also do CPU hardware detection only once on each
663 * physical node and broadcast it, instead of do it on every MPI rank.
666 /* A split of MPI_COMM_WORLD over physical nodes is only required here,
667 * so we create and destroy it locally.
669 MPI_Comm_rank(MPI_COMM_WORLD, &rank_world);
670 MPI_Comm_split(MPI_COMM_WORLD, gmx_physicalnode_id_hash(),
671 rank_world, &physicalnode_comm);
672 MPI_Comm_rank(physicalnode_comm, &rank_local);
674 /* Here there should be only one process, check this */
675 assert(cr->nnodes == 1 && cr->sim_nodeid == 0);
682 char detection_error[STRLEN] = "", sbuf[STRLEN];
684 if (detect_gpus(&hwinfo_g->gpu_info, detection_error) != 0)
686 if (detection_error[0] != '\0')
688 sprintf(sbuf, ":\n %s\n", detection_error);
694 md_print_warn(cr, fplog,
695 "NOTE: Error occurred during GPU detection%s"
696 " Can not use GPU acceleration, will fall back to CPU kernels.\n",
702 /* Broadcast the GPU info to the other ranks within this node */
703 MPI_Bcast(&hwinfo_g->gpu_info.n_dev, 1, MPI_INT, 0, physicalnode_comm);
705 if (hwinfo_g->gpu_info.n_dev > 0)
709 dev_size = hwinfo_g->gpu_info.n_dev*sizeof_gpu_dev_info();
713 hwinfo_g->gpu_info.gpu_dev =
714 (struct gmx_device_info_t *)malloc(dev_size);
716 MPI_Bcast(hwinfo_g->gpu_info.gpu_dev, dev_size, MPI_BYTE,
717 0, physicalnode_comm);
718 MPI_Bcast(&hwinfo_g->gpu_info.n_dev_compatible, 1, MPI_INT,
719 0, physicalnode_comm);
722 MPI_Comm_free(&physicalnode_comm);
726 static void gmx_collect_hardware_mpi()
730 int nrank, rank, ncore, nhwthread, ngpu, i;
734 rank_id = gmx_physicalnode_id_hash();
735 MPI_Comm_rank(MPI_COMM_WORLD, &rank);
736 MPI_Comm_size(MPI_COMM_WORLD, &nrank);
737 ncore = hwinfo_g->ncore;
738 nhwthread = hwinfo_g->nthreads_hw_avail;
739 ngpu = hwinfo_g->gpu_info.n_dev_compatible;
740 /* Create a unique hash of the GPU type(s) in this node */
742 /* Here it might be better to only loop over the compatible GPU, but we
743 * don't have that information available and it would also require
744 * removing the device ID from the device info string.
746 for (i = 0; i < hwinfo_g->gpu_info.n_dev; i++)
750 /* Since the device ID is incorporated in the hash, the order of
751 * the GPUs affects the hash. Also two identical GPUs won't give
752 * a gpu_hash of zero after XORing.
754 get_gpu_device_info_string(stmp, &hwinfo_g->gpu_info, i);
755 gpu_hash ^= gmx_string_fullhash_func(stmp, gmx_string_hash_init);
762 MPI_Allreduce(buf, all, nrank, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
765 int nnode0, ncore0, nhwthread0, ngpu0, r;
772 for (r = 0; r < nrank; r++)
774 if (all[r] == rank_id)
776 if (!bFound && r == rank)
778 /* We are the first rank in this physical node */
781 nhwthread0 = nhwthread;
791 int sum[4], maxmin[10];
796 /* Sum values from only intra-rank 0 so we get the sum over all nodes */
802 MPI_Allreduce(buf, sum, 4, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
808 /* Store + and - values for all ranks,
809 * so we can get max+min with one MPI call.
814 buf[3] = gmx_cpuid_simd_suggest(hwinfo_g->cpuid_info);
822 MPI_Allreduce(buf, maxmin, 10, MPI_INT, MPI_MAX, MPI_COMM_WORLD);
825 hwinfo_g->nphysicalnode = sum[0];
826 hwinfo_g->ncore_tot = sum[1];
827 hwinfo_g->ncore_min = -maxmin[5];
828 hwinfo_g->ncore_max = maxmin[0];
829 hwinfo_g->nhwthread_tot = sum[2];
830 hwinfo_g->nhwthread_min = -maxmin[6];
831 hwinfo_g->nhwthread_max = maxmin[1];
832 hwinfo_g->ngpu_compatible_tot = sum[3];
833 hwinfo_g->ngpu_compatible_min = -maxmin[7];
834 hwinfo_g->ngpu_compatible_max = maxmin[2];
835 hwinfo_g->simd_suggest_min = static_cast<enum gmx_cpuid_simd>(-maxmin[8]);
836 hwinfo_g->simd_suggest_max = static_cast<enum gmx_cpuid_simd>(maxmin[3]);
837 hwinfo_g->bIdenticalGPUs = (maxmin[4] == -maxmin[9]);
839 /* All ranks use the same pointer, protect it with a mutex */
840 tMPI_Thread_mutex_lock(&hw_info_lock);
841 hwinfo_g->nphysicalnode = 1;
842 hwinfo_g->ncore_tot = hwinfo_g->ncore;
843 hwinfo_g->ncore_min = hwinfo_g->ncore;
844 hwinfo_g->ncore_max = hwinfo_g->ncore;
845 hwinfo_g->nhwthread_tot = hwinfo_g->nthreads_hw_avail;
846 hwinfo_g->nhwthread_min = hwinfo_g->nthreads_hw_avail;
847 hwinfo_g->nhwthread_max = hwinfo_g->nthreads_hw_avail;
848 hwinfo_g->ngpu_compatible_tot = hwinfo_g->gpu_info.n_dev_compatible;
849 hwinfo_g->ngpu_compatible_min = hwinfo_g->gpu_info.n_dev_compatible;
850 hwinfo_g->ngpu_compatible_max = hwinfo_g->gpu_info.n_dev_compatible;
851 hwinfo_g->simd_suggest_min = gmx_cpuid_simd_suggest(hwinfo_g->cpuid_info);
852 hwinfo_g->simd_suggest_max = gmx_cpuid_simd_suggest(hwinfo_g->cpuid_info);
853 hwinfo_g->bIdenticalGPUs = TRUE;
854 tMPI_Thread_mutex_unlock(&hw_info_lock);
858 gmx_hw_info_t *gmx_detect_hardware(FILE *fplog, const t_commrec *cr,
859 gmx_bool bDetectGPUs)
863 /* make sure no one else is doing the same thing */
864 ret = tMPI_Thread_mutex_lock(&hw_info_lock);
867 gmx_fatal(FARGS, "Error locking hwinfo mutex: %s", strerror(errno));
870 /* only initialize the hwinfo structure if it is not already initalized */
875 /* detect CPUID info; no fuss, we don't detect system-wide
876 * -- sloppy, but that's it for now */
877 if (gmx_cpuid_init(&hwinfo_g->cpuid_info) != 0)
879 gmx_fatal_collective(FARGS, cr, NULL, "CPUID detection failed!");
882 /* get the number of cores, will be 0 when not detected */
883 hwinfo_g->ncore = get_ncores(hwinfo_g->cpuid_info);
885 /* detect number of hardware threads */
886 hwinfo_g->nthreads_hw_avail = get_nthreads_hw_avail(fplog, cr);
889 hwinfo_g->gpu_info.n_dev = 0;
890 hwinfo_g->gpu_info.n_dev_compatible = 0;
891 hwinfo_g->gpu_info.gpu_dev = NULL;
893 /* Run the detection if the binary was compiled with GPU support
894 * and we requested detection.
896 hwinfo_g->gpu_info.bDetectGPUs =
897 (bGPUBinary && bDetectGPUs &&
898 getenv("GMX_DISABLE_GPU_DETECTION") == NULL);
899 if (hwinfo_g->gpu_info.bDetectGPUs)
901 gmx_detect_gpus(fplog, cr);
904 /* increase the reference counter */
907 ret = tMPI_Thread_mutex_unlock(&hw_info_lock);
910 gmx_fatal(FARGS, "Error unlocking hwinfo mutex: %s", strerror(errno));
913 gmx_collect_hardware_mpi();
918 static std::string detected_hardware_string(const gmx_hw_info_t *hwinfo,
923 s = gmx::formatString("\n");
924 s += gmx::formatString("Running on %d node%s with total",
925 hwinfo->nphysicalnode,
926 hwinfo->nphysicalnode == 1 ? "" : "s");
927 if (hwinfo->ncore_tot > 0)
929 s += gmx::formatString(" %d cores,", hwinfo->ncore_tot);
931 s += gmx::formatString(" %d logical cores", hwinfo->nhwthread_tot);
932 if (hwinfo->gpu_info.bDetectGPUs)
934 s += gmx::formatString(", %d compatible GPU%s",
935 hwinfo->ngpu_compatible_tot,
936 hwinfo->ngpu_compatible_tot == 1 ? "" : "s");
940 s += gmx::formatString(" (GPU detection deactivated)");
942 s += gmx::formatString("\n");
944 if (hwinfo->nphysicalnode > 1)
946 /* Print per node hardware feature counts */
947 if (hwinfo->ncore_max > 0)
949 s += gmx::formatString(" Cores per node: %2d", hwinfo->ncore_min);
950 if (hwinfo->ncore_max > hwinfo->ncore_min)
952 s += gmx::formatString(" - %2d", hwinfo->ncore_max);
954 s += gmx::formatString("\n");
956 s += gmx::formatString(" Logical cores per node: %2d", hwinfo->nhwthread_min);
957 if (hwinfo->nhwthread_max > hwinfo->nhwthread_min)
959 s += gmx::formatString(" - %2d", hwinfo->nhwthread_max);
961 s += gmx::formatString("\n");
964 s += gmx::formatString(" Compatible GPUs per node: %2d",
965 hwinfo->ngpu_compatible_min);
966 if (hwinfo->ngpu_compatible_max > hwinfo->ngpu_compatible_min)
968 s += gmx::formatString(" - %2d", hwinfo->ngpu_compatible_max);
970 s += gmx::formatString("\n");
971 if (hwinfo->ngpu_compatible_tot > 0)
973 if (hwinfo->bIdenticalGPUs)
975 s += gmx::formatString(" All nodes have identical type(s) of GPUs\n");
979 /* This message will also appear with identical GPU types
980 * when at least one node has no GPU.
982 s += gmx::formatString(" Different nodes have different type(s) and/or order of GPUs\n");
989 char host[HOSTNAMELEN];
992 gmx_gethostname(host, HOSTNAMELEN);
993 MPI_Comm_rank(MPI_COMM_WORLD, &rank);
995 s += gmx::formatString("Hardware detected on host %s (the node of MPI rank %d):\n",
998 s += gmx::formatString("Hardware detected:\n");
1000 s += gmx::formatString(" CPU info:\n");
1005 gmx_cpuid_formatstring(hwinfo->cpuid_info, buf, 1023);
1008 s += gmx::formatString("%s", buf);
1012 s += gmx::formatString(" Vendor: %s\n",
1013 gmx_cpuid_vendor_string[gmx_cpuid_vendor(hwinfo->cpuid_info)]);
1014 s += gmx::formatString(" Brand: %s\n",
1015 gmx_cpuid_brand(hwinfo->cpuid_info));
1017 s += gmx::formatString(" SIMD instructions most likely to fit this hardware: %s",
1018 gmx_cpuid_simd_string[hwinfo->simd_suggest_min]);
1019 if (hwinfo->simd_suggest_max > hwinfo->simd_suggest_min)
1021 s += gmx::formatString(" - %s",
1022 gmx_cpuid_simd_string[hwinfo->simd_suggest_max]);
1024 s += gmx::formatString("\n");
1025 s += gmx::formatString(" SIMD instructions selected at GROMACS compile time: %s\n",
1026 gmx_cpuid_simd_string[gmx_compiled_simd()]);
1027 if (bGPUBinary && (hwinfo->ngpu_compatible_tot > 0 ||
1028 hwinfo->gpu_info.n_dev > 0))
1030 s += gmx::formatString(" GPU info:\n");
1031 s += gmx::formatString(" Number of GPUs detected: %d\n",
1032 hwinfo->gpu_info.n_dev);
1033 if (hwinfo->gpu_info.n_dev > 0)
1037 sprint_gpus(buf, &hwinfo->gpu_info);
1038 s += gmx::formatString("%s\n", buf);
1045 void gmx_print_detected_hardware(FILE *fplog, const t_commrec *cr,
1046 const gmx_hw_info_t *hwinfo)
1050 std::string detected;
1052 detected = detected_hardware_string(hwinfo, TRUE);
1054 fprintf(fplog, "%s\n", detected.c_str());
1057 if (MULTIMASTER(cr))
1059 std::string detected;
1061 detected = detected_hardware_string(hwinfo, FALSE);
1063 fprintf(stderr, "%s\n", detected.c_str());
1066 /* Check the compiled SIMD instruction set against that of the node
1067 * with the lowest SIMD level support.
1069 gmx_cpuid_simd_check(hwinfo->simd_suggest_min, fplog, MULTIMASTER(cr));
1071 /* For RDTSCP we only check on our local node and skip the MPI reduction */
1072 check_use_of_rdtscp_on_this_cpu(fplog, cr, hwinfo);
1075 //! \brief Return if any GPU ID (e.g in a user-supplied string) is repeated
1076 static gmx_bool anyGpuIdIsRepeated(const gmx_gpu_opt_t *gpu_opt)
1078 /* Loop over IDs in the string */
1079 for (int i = 0; i < gpu_opt->n_dev_use - 1; ++i)
1081 /* Look for the ID in location i in the following part of the
1083 for (int j = i + 1; j < gpu_opt->n_dev_use; ++j)
1085 if (gpu_opt->dev_use[i] == gpu_opt->dev_use[j])
1087 /* Same ID found in locations i and j */
1096 void gmx_parse_gpu_ids(gmx_gpu_opt_t *gpu_opt)
1100 if (gpu_opt->gpu_id != NULL && !bGPUBinary)
1102 gmx_fatal(FARGS, "GPU ID string set, but %s was compiled without GPU support!", ShortProgram());
1105 env = getenv("GMX_GPU_ID");
1106 if (env != NULL && gpu_opt->gpu_id != NULL)
1108 gmx_fatal(FARGS, "GMX_GPU_ID and -gpu_id can not be used at the same time");
1112 env = gpu_opt->gpu_id;
1115 /* parse GPU IDs if the user passed any */
1118 /* Parse a "plain" GPU ID string which contains a sequence of
1119 * digits corresponding to GPU IDs; the order will indicate
1120 * the process/tMPI thread - GPU assignment. */
1121 parse_digits_from_plain_string(env,
1122 &gpu_opt->n_dev_use,
1124 if (!bMultiGpuPerNodeSupported && 1 < gpu_opt->n_dev_use)
1126 gmx_fatal(FARGS, "The %s implementation only supports using exactly one PP rank per node", gpu_implementation);
1128 if (!bGpuSharingSupported && anyGpuIdIsRepeated(gpu_opt))
1130 gmx_fatal(FARGS, "The %s implementation only supports using exactly one PP rank per GPU", gpu_implementation);
1132 if (gpu_opt->n_dev_use == 0)
1134 gmx_fatal(FARGS, "Empty GPU ID string encountered.\n%s\n",
1135 invalid_gpuid_hint);
1138 gpu_opt->bUserSet = TRUE;
1142 void gmx_select_gpu_ids(FILE *fplog, const t_commrec *cr,
1143 const gmx_gpu_info_t *gpu_info,
1144 gmx_bool bForceUseGPU,
1145 gmx_gpu_opt_t *gpu_opt)
1148 char sbuf[STRLEN], stmp[STRLEN];
1150 /* Bail if binary is not compiled with GPU acceleration, but this is either
1151 * explicitly (-nb gpu) or implicitly (gpu ID passed) requested. */
1152 if (bForceUseGPU && !bGPUBinary)
1154 gmx_fatal(FARGS, "GPU acceleration requested, but %s was compiled without GPU support!", ShortProgram());
1157 if (!(cr->duty & DUTY_PP))
1159 /* Our rank is not doing PP, we don't use a GPU */
1163 if (gpu_opt->bUserSet)
1165 /* Check the GPU IDs passed by the user.
1166 * (GPU IDs have been parsed by gmx_parse_gpu_ids before)
1171 snew(checkres, gpu_opt->n_dev_use);
1173 res = check_selected_gpus(checkres, gpu_info, gpu_opt);
1177 print_gpu_detection_stats(fplog, gpu_info, cr);
1179 sprintf(sbuf, "Some of the requested GPUs do not exist, behave strangely, or are not compatible:\n");
1180 for (i = 0; i < gpu_opt->n_dev_use; i++)
1182 if (checkres[i] != egpuCompatible)
1184 sprintf(stmp, " GPU #%d: %s\n",
1185 gpu_opt->dev_use[i],
1186 gpu_detect_res_str[checkres[i]]);
1190 gmx_fatal(FARGS, "%s", sbuf);
1195 else if (getenv("GMX_EMULATE_GPU") == NULL)
1197 pick_compatible_gpus(&hwinfo_g->gpu_info, gpu_opt);
1198 set_gpu_ids(gpu_opt, cr->nrank_pp_intranode, cr->rank_pp_intranode);
1201 /* If the user asked for a GPU, check whether we have a GPU */
1202 if (bForceUseGPU && gpu_info->n_dev_compatible == 0)
1204 gmx_fatal(FARGS, "GPU acceleration requested, but no compatible GPUs were detected.");
1208 /* Select the GPUs we will use. This is an operation local to each physical
1209 * node. If we have less MPI ranks than GPUs, we will waste some GPUs.
1210 * nrank and rank are the rank count and id for PP processes in our node.
1212 static void set_gpu_ids(gmx_gpu_opt_t *gpu_opt, int nrank, int rank)
1214 GMX_RELEASE_ASSERT(gpu_opt, "Invalid gpu_opt pointer passed");
1215 GMX_RELEASE_ASSERT(nrank >= 1,
1216 gmx::formatString("Invalid limit (%d) for the number of GPUs (detected %d compatible GPUs)",
1217 rank, gpu_opt->n_dev_compatible).c_str());
1219 if (gpu_opt->n_dev_compatible == 0)
1221 char host[HOSTNAMELEN];
1223 gmx_gethostname(host, HOSTNAMELEN);
1224 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);
1230 if (nrank > gpu_opt->n_dev_compatible)
1232 if (nrank % gpu_opt->n_dev_compatible == 0)
1234 nshare = bGpuSharingSupported ? nrank/gpu_opt->n_dev_compatible : 1;
1240 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.",
1241 nrank, gpu_opt->n_dev_compatible);
1245 /* We use a global barrier to prevent ranks from continuing with
1248 MPI_Barrier(MPI_COMM_WORLD);
1253 /* Here we will waste GPUs when nrank < gpu_opt->n_dev_compatible */
1254 gpu_opt->n_dev_use = std::min(gpu_opt->n_dev_compatible*nshare, nrank);
1255 if (!bMultiGpuPerNodeSupported)
1257 gpu_opt->n_dev_use = std::min(gpu_opt->n_dev_use, 1);
1259 snew(gpu_opt->dev_use, gpu_opt->n_dev_use);
1260 for (int i = 0; i != gpu_opt->n_dev_use; ++i)
1262 /* TODO: improve this implementation: either sort GPUs or remove the weakest here */
1263 gpu_opt->dev_use[i] = gpu_opt->dev_compatible[i/nshare];
1267 void gmx_hardware_info_free(gmx_hw_info_t *hwinfo)
1271 ret = tMPI_Thread_mutex_lock(&hw_info_lock);
1274 gmx_fatal(FARGS, "Error locking hwinfo mutex: %s", strerror(errno));
1277 /* decrease the reference counter */
1281 if (hwinfo != hwinfo_g)
1283 gmx_incons("hwinfo < hwinfo_g");
1288 gmx_incons("n_hwinfo < 0");
1293 gmx_cpuid_done(hwinfo_g->cpuid_info);
1294 free_gpu_info(&hwinfo_g->gpu_info);
1298 ret = tMPI_Thread_mutex_unlock(&hw_info_lock);
1301 gmx_fatal(FARGS, "Error unlocking hwinfo mutex: %s", strerror(errno));