2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014, 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.
52 #include "types/enums.h"
53 #include "types/hw_info.h"
54 #include "types/commrec.h"
56 #include "md_logging.h"
57 #include "gmx_cpuid.h"
58 #include "gpu_utils.h"
60 #include "gmx_detect_hardware.h"
61 #include "md_logging.h"
63 #include "gromacs/utility/basenetwork.h"
64 #include "gromacs/utility/cstringutil.h"
65 #include "gromacs/utility/exceptions.h"
66 #include "gromacs/utility/fatalerror.h"
67 #include "gromacs/utility/gmxomp.h"
68 #include "gromacs/utility/smalloc.h"
69 #include "gromacs/utility/stringutil.h"
71 #include "thread_mpi/threads.h"
73 #ifdef GMX_NATIVE_WINDOWS
78 const gmx_bool bGPUBinary = TRUE;
80 const gmx_bool bGPUBinary = FALSE;
83 static const char * invalid_gpuid_hint =
84 "A delimiter-free sequence of valid numeric IDs of available GPUs is expected.";
86 /* The globally shared hwinfo structure. */
87 static gmx_hw_info_t *hwinfo_g;
88 /* A reference counter for the hwinfo structure */
89 static int n_hwinfo = 0;
90 /* A lock to protect the hwinfo structure */
91 static tMPI_Thread_mutex_t hw_info_lock = TMPI_THREAD_MUTEX_INITIALIZER;
95 static void limit_num_gpus_used(gmx_gpu_opt_t *gpu_opt, int count);
96 static int gmx_count_gpu_dev_unique(const gmx_gpu_info_t *gpu_info,
97 const gmx_gpu_opt_t *gpu_opt);
99 static void sprint_gpus(char *sbuf, const gmx_gpu_info_t *gpu_info)
104 ndev = gpu_info->ncuda_dev;
107 for (i = 0; i < ndev; i++)
109 get_gpu_device_info_string(stmp, gpu_info, i);
119 static void print_gpu_detection_stats(FILE *fplog,
120 const gmx_gpu_info_t *gpu_info,
123 char onhost[266], stmp[STRLEN];
126 if (!gpu_info->bDetectGPUs)
128 /* We skipped the detection, so don't print detection stats */
132 ngpu = gpu_info->ncuda_dev;
134 #if defined GMX_MPI && !defined GMX_THREAD_MPI
135 /* We only print the detection on one, of possibly multiple, nodes */
136 strncpy(onhost, " on host ", 10);
137 gmx_gethostname(onhost+9, 256);
139 /* We detect all relevant GPUs */
140 strncpy(onhost, "", 1);
145 sprint_gpus(stmp, gpu_info);
146 md_print_warn(cr, fplog, "%d GPU%s detected%s:\n%s\n",
147 ngpu, (ngpu > 1) ? "s" : "", onhost, stmp);
151 md_print_warn(cr, fplog, "No GPUs detected%s\n", onhost);
155 /*! \brief Helper function for writing comma-separated GPU IDs.
157 * \param[in] ids A container of integer GPU IDs
158 * \return A comma-separated string of GPU IDs */
159 template <typename Container>
160 static std::string makeGpuIdsString(const Container &ids)
166 typename Container::const_iterator it = ids.begin();
167 output += gmx::formatString("%d", *it);
168 for (++it; it != ids.end(); ++it)
170 output += gmx::formatString(",%d", *it);
176 /*! \brief Helper function for reporting GPU usage information
177 * in the mdrun log file
179 * \param[in] gpu_info Pointer to per-node GPU info struct
180 * \param[in] gpu_opt Pointer to per-node GPU options struct
181 * \param[in] numPpRanks Number of PP ranks per node
182 * \return String to write to the log file
183 * \throws std::bad_alloc if out of memory */
185 makeGpuUsageReport(const gmx_gpu_info_t *gpu_info,
186 const gmx_gpu_opt_t *gpu_opt,
189 int ngpu_use = gpu_opt->ncuda_dev_use;
190 int ngpu_comp = gpu_info->ncuda_dev_compatible;
192 /* Issue a note if GPUs are available but not used */
193 if (ngpu_comp > 0 && ngpu_use < 1)
195 return gmx::formatString("%d compatible GPU%s detected in the system, but none will be used.\n"
196 "Consider trying GPU acceleration with the Verlet scheme!\n",
197 ngpu_comp, (ngpu_comp > 1) ? "s" : "");
203 std::vector<int> gpuIdsInUse;
204 for (int i = 0; i < ngpu_use; i++)
206 gpuIdsInUse.push_back(get_gpu_device_id(gpu_info, gpu_opt, i));
208 std::string gpuIdsString = makeGpuIdsString(gpuIdsInUse);
209 int numGpusInUse = gmx_count_gpu_dev_unique(gpu_info, gpu_opt);
210 bool bPluralGpus = numGpusInUse > 1;
212 output += gmx::formatString("%d GPU%s %sselected for this run.\n"
213 "Mapping of GPU ID%s to the %d PP rank%s in this node: %s\n",
214 numGpusInUse, bPluralGpus ? "s" : "",
215 gpu_opt->bUserSet ? "user-" : "auto-",
216 bPluralGpus ? "s" : "",
218 (numPpRanks > 1) ? "s" : "",
219 gpuIdsString.c_str());
225 /* Give a suitable fatal error or warning if the build configuration
226 and runtime CPU do not match. */
228 check_use_of_rdtscp_on_this_cpu(FILE *fplog,
230 const gmx_hw_info_t *hwinfo)
232 gmx_bool bCpuHasRdtscp, bBinaryUsesRdtscp;
234 bBinaryUsesRdtscp = TRUE;
236 bBinaryUsesRdtscp = FALSE;
239 bCpuHasRdtscp = gmx_cpuid_feature(hwinfo->cpuid_info, GMX_CPUID_FEATURE_X86_RDTSCP);
241 if (!bCpuHasRdtscp && bBinaryUsesRdtscp)
243 gmx_fatal(FARGS, "The %s executable was compiled to use the rdtscp CPU instruction. "
244 "However, this is not supported by the current hardware and continuing would lead to a crash. "
245 "Please rebuild GROMACS with the GMX_USE_RDTSCP=OFF CMake option.",
249 if (bCpuHasRdtscp && !bBinaryUsesRdtscp)
251 md_print_warn(cr, fplog, "The current CPU can measure timings more accurately than the code in\n"
252 "%s was configured to use. This might affect your simulation\n"
253 "speed as accurate timings are needed for load-balancing.\n"
254 "Please consider rebuilding %s with the GMX_USE_RDTSCP=OFF CMake option.\n",
255 ShortProgram(), ShortProgram());
259 void gmx_check_hw_runconf_consistency(FILE *fplog,
260 const gmx_hw_info_t *hwinfo,
262 const gmx_hw_opt_t *hw_opt,
266 char th_or_proc[STRLEN], th_or_proc_plural[STRLEN], pernode[STRLEN];
267 gmx_bool btMPI, bMPI, bMaxMpiThreadsSet, bNthreadsAuto, bEmulateGPU;
272 /* Below we only do consistency checks for PP and GPUs,
273 * this is irrelevant for PME only nodes, so in that case we return
276 if (!(cr->duty & DUTY_PP))
281 #if defined(GMX_THREAD_MPI)
284 bNthreadsAuto = (hw_opt->nthreads_tmpi < 1);
285 #elif defined(GMX_LIB_MPI)
288 bNthreadsAuto = FALSE;
292 bNthreadsAuto = FALSE;
295 /* GPU emulation detection is done later, but we need here as well
296 * -- uncool, but there's no elegant workaround */
297 bEmulateGPU = (getenv("GMX_EMULATE_GPU") != NULL);
298 bMaxMpiThreadsSet = (getenv("GMX_MAX_MPI_THREADS") != NULL);
300 /* check the SIMD level mdrun is compiled with against hardware
302 /* TODO: Here we assume homogeneous hardware which is not necessarily
303 the case! Might not hurt to add an extra check over MPI. */
304 gmx_cpuid_simd_check(hwinfo->cpuid_info, fplog, SIMMASTER(cr));
306 check_use_of_rdtscp_on_this_cpu(fplog, cr, hwinfo);
308 /* NOTE: this print is only for and on one physical node */
309 print_gpu_detection_stats(fplog, &hwinfo->gpu_info, cr);
311 if (hwinfo->gpu_info.ncuda_dev_compatible > 0)
313 std::string gpuUseageReport;
316 gpuUseageReport = makeGpuUsageReport(&hwinfo->gpu_info,
318 cr->nrank_pp_intranode);
320 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
322 /* NOTE: this print is only for and on one physical node */
323 md_print_info(cr, fplog, gpuUseageReport.c_str());
326 /* Need to ensure that we have enough GPUs:
327 * - need one GPU per PP node
328 * - no GPU oversubscription with tMPI
330 /* number of PP processes per node */
331 npppn = cr->nrank_pp_intranode;
334 th_or_proc_plural[0] = '\0';
337 sprintf(th_or_proc, "thread-MPI thread");
340 sprintf(th_or_proc_plural, "s");
345 sprintf(th_or_proc, "MPI process");
348 sprintf(th_or_proc_plural, "es");
350 sprintf(pernode, " per node");
354 /* neither MPI nor tMPI */
355 sprintf(th_or_proc, "process");
358 if (bUseGPU && hwinfo->gpu_info.ncuda_dev_compatible > 0 &&
361 int ngpu_comp, ngpu_use;
362 char gpu_comp_plural[2], gpu_use_plural[2];
364 ngpu_comp = hwinfo->gpu_info.ncuda_dev_compatible;
365 ngpu_use = hw_opt->gpu_opt.ncuda_dev_use;
367 sprintf(gpu_comp_plural, "%s", (ngpu_comp > 1) ? "s" : "");
368 sprintf(gpu_use_plural, "%s", (ngpu_use > 1) ? "s" : "");
370 /* number of tMPI threads auto-adjusted */
371 if (btMPI && bNthreadsAuto)
373 if (hw_opt->gpu_opt.bUserSet && npppn < ngpu_use)
375 /* The user manually provided more GPUs than threads we
376 could automatically start. */
378 "%d GPU%s provided, but only %d PP thread-MPI thread%s coud be started.\n"
379 "%s requires one PP tread-MPI thread per GPU; use fewer GPUs%s.",
380 ngpu_use, gpu_use_plural,
381 npppn, th_or_proc_plural,
382 ShortProgram(), bMaxMpiThreadsSet ? "\nor allow more threads to be used" : "");
385 if (!hw_opt->gpu_opt.bUserSet && npppn < ngpu_comp)
387 /* There are more GPUs than tMPI threads; we have
388 limited the number GPUs used. */
389 md_print_warn(cr, fplog,
390 "NOTE: %d GPU%s were detected, but only %d PP thread-MPI thread%s can be started.\n"
391 " %s can use one GPU per PP tread-MPI thread, so only %d GPU%s will be used.%s\n",
392 ngpu_comp, gpu_comp_plural,
393 npppn, th_or_proc_plural,
394 ShortProgram(), npppn,
395 npppn > 1 ? "s" : "",
396 bMaxMpiThreadsSet ? "\n Also, you can allow more threads to be used by increasing GMX_MAX_MPI_THREADS" : "");
400 if (hw_opt->gpu_opt.bUserSet)
402 if (ngpu_use != npppn)
405 "Incorrect launch configuration: mismatching number of PP %s%s and GPUs%s.\n"
406 "%s was started with %d PP %s%s%s, but you provided %d GPU%s.",
407 th_or_proc, btMPI ? "s" : "es", pernode,
408 ShortProgram(), npppn, th_or_proc,
409 th_or_proc_plural, pernode,
410 ngpu_use, gpu_use_plural);
415 if (ngpu_comp > npppn)
417 md_print_warn(cr, fplog,
418 "NOTE: potentially sub-optimal launch configuration, %s started with less\n"
419 " PP %s%s%s than GPU%s available.\n"
420 " Each PP %s can use only one GPU, %d GPU%s%s will be used.\n",
421 ShortProgram(), th_or_proc,
422 th_or_proc_plural, pernode, gpu_comp_plural,
423 th_or_proc, npppn, gpu_use_plural, pernode);
426 if (ngpu_use != npppn)
428 /* Avoid duplicate error messages.
429 * Unfortunately we can only do this at the physical node
430 * level, since the hardware setup and MPI process count
431 * might differ between physical nodes.
433 if (cr->rank_pp_intranode == 0)
436 "Incorrect launch configuration: mismatching number of PP %s%s and GPUs%s.\n"
437 "%s was started with %d PP %s%s%s, but only %d GPU%s were detected.",
438 th_or_proc, btMPI ? "s" : "es", pernode,
439 ShortProgram(), npppn, th_or_proc,
440 th_or_proc_plural, pernode,
441 ngpu_use, gpu_use_plural);
449 same_count = gmx_count_gpu_dev_shared(&hw_opt->gpu_opt);
453 md_print_info(cr, fplog,
454 "NOTE: You assigned %s to multiple %s%s.\n",
455 same_count > 1 ? "GPUs" : "a GPU", th_or_proc, btMPI ? "s" : "es");
463 /* Avoid other ranks to continue after
465 MPI_Barrier(cr->mpi_comm_mygroup);
471 /* Return 0 if none of the GPU (per node) are shared among PP ranks.
473 * Sharing GPUs among multiple PP ranks is possible when the user passes
474 * GPU IDs. Here we check for sharing and return a non-zero value when
475 * this is detected. Note that the return value represents the number of
476 * PP rank pairs that share a device.
478 int gmx_count_gpu_dev_shared(const gmx_gpu_opt_t *gpu_opt)
481 int ngpu = gpu_opt->ncuda_dev_use;
483 if (gpu_opt->bUserSet)
487 for (i = 0; i < ngpu - 1; i++)
489 for (j = i + 1; j < ngpu; j++)
491 same_count += (gpu_opt->cuda_dev_use[i] ==
492 gpu_opt->cuda_dev_use[j]);
500 /* Count and return the number of unique GPUs (per node) selected.
502 * As sharing GPUs among multiple PP ranks is possible when the user passes
503 * GPU IDs, the number of GPUs user (per node) can be different from the
504 * number of GPU IDs selected.
506 static int gmx_count_gpu_dev_unique(const gmx_gpu_info_t *gpu_info,
507 const gmx_gpu_opt_t *gpu_opt)
509 int i, uniq_count, ngpu;
515 ngpu = gpu_info->ncuda_dev;
518 snew(uniq_ids, ngpu);
520 /* Each element in uniq_ids will be set to 0 or 1. The n-th element set
521 * to 1 indicates that the respective GPU was selected to be used. */
522 for (i = 0; i < gpu_opt->ncuda_dev_use; i++)
524 uniq_ids[get_gpu_device_id(gpu_info, gpu_opt, i)] = 1;
526 /* Count the devices used. */
527 for (i = 0; i < ngpu; i++)
529 uniq_count += uniq_ids[i];
538 /* Return the number of hardware threads supported by the current CPU.
539 * We assume that this is equal with the number of CPUs reported to be
540 * online by the OS at the time of the call.
542 static int get_nthreads_hw_avail(FILE gmx_unused *fplog, const t_commrec gmx_unused *cr)
546 #if ((defined(WIN32) || defined( _WIN32 ) || defined(WIN64) || defined( _WIN64 )) && !(defined (__CYGWIN__) || defined (__CYGWIN32__)))
549 GetSystemInfo( &sysinfo );
550 ret = sysinfo.dwNumberOfProcessors;
551 #elif defined HAVE_SYSCONF
552 /* We are probably on Unix.
553 * Now check if we have the argument to use before executing the call
555 #if defined(_SC_NPROCESSORS_ONLN)
556 ret = sysconf(_SC_NPROCESSORS_ONLN);
557 #elif defined(_SC_NPROC_ONLN)
558 ret = sysconf(_SC_NPROC_ONLN);
559 #elif defined(_SC_NPROCESSORS_CONF)
560 ret = sysconf(_SC_NPROCESSORS_CONF);
561 #elif defined(_SC_NPROC_CONF)
562 ret = sysconf(_SC_NPROC_CONF);
564 #warning "No valid sysconf argument value found. Executables will not be able to determine the number of hardware threads: mdrun will use 1 thread by default!"
565 #endif /* End of check for sysconf argument values */
568 /* Neither windows nor Unix. No fscking idea how many CPUs we have! */
574 fprintf(debug, "Detected %d processors, will use this as the number "
575 "of supported hardware threads.\n", ret);
579 if (ret != gmx_omp_get_num_procs())
581 md_print_warn(cr, fplog,
582 "Number of CPUs detected (%d) does not match the number reported by OpenMP (%d).\n"
583 "Consider setting the launch configuration manually!",
584 ret, gmx_omp_get_num_procs());
591 static void gmx_detect_gpus(FILE *fplog, const t_commrec *cr)
595 MPI_Comm physicalnode_comm;
599 /* Under certain circumstances MPI ranks on the same physical node
600 * can not simultaneously access the same GPU(s). Therefore we run
601 * the detection only on one MPI rank per node and broadcast the info.
602 * Note that with thread-MPI only a single thread runs this code.
604 * TODO: We should also do CPU hardware detection only once on each
605 * physical node and broadcast it, instead of do it on every MPI rank.
608 /* A split of MPI_COMM_WORLD over physical nodes is only required here,
609 * so we create and destroy it locally.
611 MPI_Comm_rank(MPI_COMM_WORLD, &rank_world);
612 MPI_Comm_split(MPI_COMM_WORLD, gmx_physicalnode_id_hash(),
613 rank_world, &physicalnode_comm);
614 MPI_Comm_rank(physicalnode_comm, &rank_local);
616 /* Here there should be only one process, check this */
617 assert(cr->nnodes == 1 && cr->sim_nodeid == 0);
624 char detection_error[STRLEN] = "", sbuf[STRLEN];
626 if (detect_cuda_gpus(&hwinfo_g->gpu_info, detection_error) != 0)
628 if (detection_error[0] != '\0')
630 sprintf(sbuf, ":\n %s\n", detection_error);
636 md_print_warn(cr, fplog,
637 "NOTE: Error occurred during GPU detection%s"
638 " Can not use GPU acceleration, will fall back to CPU kernels.\n",
644 /* Broadcast the GPU info to the other ranks within this node */
645 MPI_Bcast(&hwinfo_g->gpu_info.ncuda_dev, 1, MPI_INT, 0, physicalnode_comm);
647 if (hwinfo_g->gpu_info.ncuda_dev > 0)
651 cuda_dev_size = hwinfo_g->gpu_info.ncuda_dev*sizeof_cuda_dev_info();
655 hwinfo_g->gpu_info.cuda_dev =
656 (cuda_dev_info_ptr_t)malloc(cuda_dev_size);
658 MPI_Bcast(hwinfo_g->gpu_info.cuda_dev, cuda_dev_size, MPI_BYTE,
659 0, physicalnode_comm);
660 MPI_Bcast(&hwinfo_g->gpu_info.ncuda_dev_compatible, 1, MPI_INT,
661 0, physicalnode_comm);
664 MPI_Comm_free(&physicalnode_comm);
668 gmx_hw_info_t *gmx_detect_hardware(FILE *fplog, const t_commrec *cr,
669 gmx_bool bDetectGPUs)
673 /* make sure no one else is doing the same thing */
674 ret = tMPI_Thread_mutex_lock(&hw_info_lock);
677 gmx_fatal(FARGS, "Error locking hwinfo mutex: %s", strerror(errno));
680 /* only initialize the hwinfo structure if it is not already initalized */
685 /* detect CPUID info; no fuss, we don't detect system-wide
686 * -- sloppy, but that's it for now */
687 if (gmx_cpuid_init(&hwinfo_g->cpuid_info) != 0)
689 gmx_fatal_collective(FARGS, cr, NULL, "CPUID detection failed!");
692 /* detect number of hardware threads */
693 hwinfo_g->nthreads_hw_avail = get_nthreads_hw_avail(fplog, cr);
696 hwinfo_g->gpu_info.ncuda_dev = 0;
697 hwinfo_g->gpu_info.cuda_dev = NULL;
698 hwinfo_g->gpu_info.ncuda_dev_compatible = 0;
700 /* Run the detection if the binary was compiled with GPU support
701 * and we requested detection.
703 hwinfo_g->gpu_info.bDetectGPUs =
704 (bGPUBinary && bDetectGPUs &&
705 getenv("GMX_DISABLE_GPU_DETECTION") == NULL);
706 if (hwinfo_g->gpu_info.bDetectGPUs)
708 gmx_detect_gpus(fplog, cr);
711 /* increase the reference counter */
714 ret = tMPI_Thread_mutex_unlock(&hw_info_lock);
717 gmx_fatal(FARGS, "Error unlocking hwinfo mutex: %s", strerror(errno));
723 void gmx_parse_gpu_ids(gmx_gpu_opt_t *gpu_opt)
727 if (gpu_opt->gpu_id != NULL && !bGPUBinary)
729 gmx_fatal(FARGS, "GPU ID string set, but %s was compiled without GPU support!", ShortProgram());
732 env = getenv("GMX_GPU_ID");
733 if (env != NULL && gpu_opt->gpu_id != NULL)
735 gmx_fatal(FARGS, "GMX_GPU_ID and -gpu_id can not be used at the same time");
739 env = gpu_opt->gpu_id;
742 /* parse GPU IDs if the user passed any */
745 /* Parse a "plain" GPU ID string which contains a sequence of
746 * digits corresponding to GPU IDs; the order will indicate
747 * the process/tMPI thread - GPU assignment. */
748 parse_digits_from_plain_string(env,
749 &gpu_opt->ncuda_dev_use,
750 &gpu_opt->cuda_dev_use);
752 if (gpu_opt->ncuda_dev_use == 0)
754 gmx_fatal(FARGS, "Empty GPU ID string encountered.\n%s\n",
758 gpu_opt->bUserSet = TRUE;
762 void gmx_select_gpu_ids(FILE *fplog, const t_commrec *cr,
763 const gmx_gpu_info_t *gpu_info,
764 gmx_bool bForceUseGPU,
765 gmx_gpu_opt_t *gpu_opt)
768 char sbuf[STRLEN], stmp[STRLEN];
770 /* Bail if binary is not compiled with GPU acceleration, but this is either
771 * explicitly (-nb gpu) or implicitly (gpu ID passed) requested. */
772 if (bForceUseGPU && !bGPUBinary)
774 gmx_fatal(FARGS, "GPU acceleration requested, but %s was compiled without GPU support!", ShortProgram());
777 if (gpu_opt->bUserSet)
779 /* Check the GPU IDs passed by the user.
780 * (GPU IDs have been parsed by gmx_parse_gpu_ids before)
785 snew(checkres, gpu_opt->ncuda_dev_use);
787 res = check_selected_cuda_gpus(checkres, gpu_info, gpu_opt);
791 print_gpu_detection_stats(fplog, gpu_info, cr);
793 sprintf(sbuf, "Some of the requested GPUs do not exist, behave strangely, or are not compatible:\n");
794 for (i = 0; i < gpu_opt->ncuda_dev_use; i++)
796 if (checkres[i] != egpuCompatible)
798 sprintf(stmp, " GPU #%d: %s\n",
799 gpu_opt->cuda_dev_use[i],
800 gpu_detect_res_str[checkres[i]]);
804 gmx_fatal(FARGS, "%s", sbuf);
811 pick_compatible_gpus(&hwinfo_g->gpu_info, gpu_opt);
813 if (gpu_opt->ncuda_dev_use > cr->nrank_pp_intranode)
815 /* We picked more GPUs than we can use: limit the number.
816 * We print detailed messages about this later in
817 * gmx_check_hw_runconf_consistency.
819 limit_num_gpus_used(gpu_opt, cr->nrank_pp_intranode);
822 gpu_opt->bUserSet = FALSE;
825 /* If the user asked for a GPU, check whether we have a GPU */
826 if (bForceUseGPU && gpu_info->ncuda_dev_compatible == 0)
828 gmx_fatal(FARGS, "GPU acceleration requested, but no compatible GPUs were detected.");
832 static void limit_num_gpus_used(gmx_gpu_opt_t *gpu_opt, int count)
838 ndev_use = gpu_opt->ncuda_dev_use;
840 if (count > ndev_use)
842 /* won't increase the # of GPUs */
849 sprintf(sbuf, "Limiting the number of GPUs to <1 doesn't make sense (detected %d, %d requested)!",
854 /* TODO: improve this implementation: either sort GPUs or remove the weakest here */
855 gpu_opt->ncuda_dev_use = count;
858 void gmx_hardware_info_free(gmx_hw_info_t *hwinfo)
862 ret = tMPI_Thread_mutex_lock(&hw_info_lock);
865 gmx_fatal(FARGS, "Error locking hwinfo mutex: %s", strerror(errno));
868 /* decrease the reference counter */
872 if (hwinfo != hwinfo_g)
874 gmx_incons("hwinfo < hwinfo_g");
879 gmx_incons("n_hwinfo < 0");
884 gmx_cpuid_done(hwinfo_g->cpuid_info);
885 free_gpu_info(&hwinfo_g->gpu_info);
889 ret = tMPI_Thread_mutex_unlock(&hw_info_lock);
892 gmx_fatal(FARGS, "Error unlocking hwinfo mutex: %s", strerror(errno));