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.
49 #include "types/enums.h"
50 #include "types/hw_info.h"
51 #include "types/commrec.h"
53 #include "md_logging.h"
54 #include "gmx_cpuid.h"
55 #include "gpu_utils.h"
57 #include "gmx_detect_hardware.h"
58 #include "md_logging.h"
60 #include "gromacs/utility/basenetwork.h"
61 #include "gromacs/utility/cstringutil.h"
62 #include "gromacs/utility/fatalerror.h"
63 #include "gromacs/utility/gmxomp.h"
64 #include "gromacs/utility/smalloc.h"
66 #include "thread_mpi/threads.h"
68 #ifdef GMX_NATIVE_WINDOWS
73 const gmx_bool bGPUBinary = TRUE;
75 const gmx_bool bGPUBinary = FALSE;
78 static const char * invalid_gpuid_hint =
79 "A delimiter-free sequence of valid numeric IDs of available GPUs is expected.";
81 /* The globally shared hwinfo structure. */
82 static gmx_hw_info_t *hwinfo_g;
83 /* A reference counter for the hwinfo structure */
84 static int n_hwinfo = 0;
85 /* A lock to protect the hwinfo structure */
86 static tMPI_Thread_mutex_t hw_info_lock = TMPI_THREAD_MUTEX_INITIALIZER;
90 static void limit_num_gpus_used(gmx_gpu_opt_t *gpu_opt, int count);
91 static int gmx_count_gpu_dev_unique(const gmx_gpu_info_t *gpu_info,
92 const gmx_gpu_opt_t *gpu_opt);
94 static void sprint_gpus(char *sbuf, const gmx_gpu_info_t *gpu_info)
99 ndev = gpu_info->ncuda_dev;
102 for (i = 0; i < ndev; i++)
104 get_gpu_device_info_string(stmp, gpu_info, i);
114 static void print_gpu_detection_stats(FILE *fplog,
115 const gmx_gpu_info_t *gpu_info,
118 char onhost[266], stmp[STRLEN];
121 if (!gpu_info->bDetectGPUs)
123 /* We skipped the detection, so don't print detection stats */
127 ngpu = gpu_info->ncuda_dev;
129 #if defined GMX_MPI && !defined GMX_THREAD_MPI
130 /* We only print the detection on one, of possibly multiple, nodes */
131 strncpy(onhost, " on host ", 10);
132 gmx_gethostname(onhost+9, 256);
134 /* We detect all relevant GPUs */
135 strncpy(onhost, "", 1);
140 sprint_gpus(stmp, gpu_info);
141 md_print_warn(cr, fplog, "%d GPU%s detected%s:\n%s\n",
142 ngpu, (ngpu > 1) ? "s" : "", onhost, stmp);
146 md_print_warn(cr, fplog, "No GPUs detected%s\n", onhost);
150 static void print_gpu_use_stats(FILE *fplog,
151 const gmx_gpu_info_t *gpu_info,
152 const gmx_gpu_opt_t *gpu_opt,
155 char sbuf[STRLEN], stmp[STRLEN];
156 int i, ngpu_comp, ngpu_use;
158 ngpu_comp = gpu_info->ncuda_dev_compatible;
159 ngpu_use = gpu_opt->ncuda_dev_use;
161 /* Issue a note if GPUs are available but not used */
162 if (ngpu_comp > 0 && ngpu_use < 1)
165 "%d compatible GPU%s detected in the system, but none will be used.\n"
166 "Consider trying GPU acceleration with the Verlet scheme!",
167 ngpu_comp, (ngpu_comp > 1) ? "s" : "");
173 ngpu_use_uniq = gmx_count_gpu_dev_unique(gpu_info, gpu_opt);
175 sprintf(sbuf, "%d GPU%s %sselected for this run.\n"
176 "Mapping of GPU%s to the %d PP rank%s in this node: ",
177 ngpu_use_uniq, (ngpu_use_uniq > 1) ? "s" : "",
178 gpu_opt->bUserSet ? "user-" : "auto-",
179 (ngpu_use > 1) ? "s" : "",
180 cr->nrank_pp_intranode,
181 (cr->nrank_pp_intranode > 1) ? "s" : "");
183 for (i = 0; i < ngpu_use; i++)
185 sprintf(stmp, "#%d", get_gpu_device_id(gpu_info, gpu_opt, i));
186 if (i < ngpu_use - 1)
193 md_print_info(cr, fplog, "%s\n\n", sbuf);
196 /* Give a suitable fatal error or warning if the build configuration
197 and runtime CPU do not match. */
199 check_use_of_rdtscp_on_this_cpu(FILE *fplog,
201 const gmx_hw_info_t *hwinfo)
203 gmx_bool bCpuHasRdtscp, bBinaryUsesRdtscp;
205 bBinaryUsesRdtscp = TRUE;
207 bBinaryUsesRdtscp = FALSE;
210 bCpuHasRdtscp = gmx_cpuid_feature(hwinfo->cpuid_info, GMX_CPUID_FEATURE_X86_RDTSCP);
212 if (!bCpuHasRdtscp && bBinaryUsesRdtscp)
214 gmx_fatal(FARGS, "The %s executable was compiled to use the rdtscp CPU instruction. "
215 "However, this is not supported by the current hardware and continuing would lead to a crash. "
216 "Please rebuild GROMACS with the GMX_USE_RDTSCP=OFF CMake option.",
220 if (bCpuHasRdtscp && !bBinaryUsesRdtscp)
222 md_print_warn(cr, fplog, "The current CPU can measure timings more accurately than the code in\n"
223 "%s was configured to use. This might affect your simulation\n"
224 "speed as accurate timings are needed for load-balancing.\n"
225 "Please consider rebuilding %s with the GMX_USE_RDTSCP=OFF CMake option.\n",
226 ShortProgram(), ShortProgram());
230 void gmx_check_hw_runconf_consistency(FILE *fplog,
231 const gmx_hw_info_t *hwinfo,
233 const gmx_hw_opt_t *hw_opt,
237 char th_or_proc[STRLEN], th_or_proc_plural[STRLEN], pernode[STRLEN];
238 gmx_bool btMPI, bMPI, bMaxMpiThreadsSet, bNthreadsAuto, bEmulateGPU;
243 /* Below we only do consistency checks for PP and GPUs,
244 * this is irrelevant for PME only nodes, so in that case we return
247 if (!(cr->duty & DUTY_PP))
252 #if defined(GMX_THREAD_MPI)
255 bNthreadsAuto = (hw_opt->nthreads_tmpi < 1);
256 #elif defined(GMX_LIB_MPI)
259 bNthreadsAuto = FALSE;
263 bNthreadsAuto = FALSE;
266 /* GPU emulation detection is done later, but we need here as well
267 * -- uncool, but there's no elegant workaround */
268 bEmulateGPU = (getenv("GMX_EMULATE_GPU") != NULL);
269 bMaxMpiThreadsSet = (getenv("GMX_MAX_MPI_THREADS") != NULL);
271 /* check the SIMD level mdrun is compiled with against hardware
273 /* TODO: Here we assume homogeneous hardware which is not necessarily
274 the case! Might not hurt to add an extra check over MPI. */
275 gmx_cpuid_simd_check(hwinfo->cpuid_info, fplog, SIMMASTER(cr));
277 check_use_of_rdtscp_on_this_cpu(fplog, cr, hwinfo);
279 /* NOTE: this print is only for and on one physical node */
280 print_gpu_detection_stats(fplog, &hwinfo->gpu_info, cr);
282 if (hwinfo->gpu_info.ncuda_dev_compatible > 0)
284 /* NOTE: this print is only for and on one physical node */
285 print_gpu_use_stats(fplog, &hwinfo->gpu_info, &hw_opt->gpu_opt, cr);
288 /* Need to ensure that we have enough GPUs:
289 * - need one GPU per PP node
290 * - no GPU oversubscription with tMPI
292 /* number of PP processes per node */
293 npppn = cr->nrank_pp_intranode;
296 th_or_proc_plural[0] = '\0';
299 sprintf(th_or_proc, "thread-MPI thread");
302 sprintf(th_or_proc_plural, "s");
307 sprintf(th_or_proc, "MPI process");
310 sprintf(th_or_proc_plural, "es");
312 sprintf(pernode, " per node");
316 /* neither MPI nor tMPI */
317 sprintf(th_or_proc, "process");
320 if (bUseGPU && hwinfo->gpu_info.ncuda_dev_compatible > 0 &&
323 int ngpu_comp, ngpu_use;
324 char gpu_comp_plural[2], gpu_use_plural[2];
326 ngpu_comp = hwinfo->gpu_info.ncuda_dev_compatible;
327 ngpu_use = hw_opt->gpu_opt.ncuda_dev_use;
329 sprintf(gpu_comp_plural, "%s", (ngpu_comp > 1) ? "s" : "");
330 sprintf(gpu_use_plural, "%s", (ngpu_use > 1) ? "s" : "");
332 /* number of tMPI threads auto-adjusted */
333 if (btMPI && bNthreadsAuto)
335 if (hw_opt->gpu_opt.bUserSet && npppn < ngpu_use)
337 /* The user manually provided more GPUs than threads we
338 could automatically start. */
340 "%d GPU%s provided, but only %d PP thread-MPI thread%s coud be started.\n"
341 "%s requires one PP tread-MPI thread per GPU; use fewer GPUs%s.",
342 ngpu_use, gpu_use_plural,
343 npppn, th_or_proc_plural,
344 ShortProgram(), bMaxMpiThreadsSet ? "\nor allow more threads to be used" : "");
347 if (!hw_opt->gpu_opt.bUserSet && npppn < ngpu_comp)
349 /* There are more GPUs than tMPI threads; we have
350 limited the number GPUs used. */
351 md_print_warn(cr, fplog,
352 "NOTE: %d GPU%s were detected, but only %d PP thread-MPI thread%s can be started.\n"
353 " %s can use one GPU per PP tread-MPI thread, so only %d GPU%s will be used.%s\n",
354 ngpu_comp, gpu_comp_plural,
355 npppn, th_or_proc_plural,
356 ShortProgram(), npppn,
357 npppn > 1 ? "s" : "",
358 bMaxMpiThreadsSet ? "\n Also, you can allow more threads to be used by increasing GMX_MAX_MPI_THREADS" : "");
362 if (hw_opt->gpu_opt.bUserSet)
364 if (ngpu_use != npppn)
367 "Incorrect launch configuration: mismatching number of PP %s%s and GPUs%s.\n"
368 "%s was started with %d PP %s%s%s, but you provided %d GPU%s.",
369 th_or_proc, btMPI ? "s" : "es", pernode,
370 ShortProgram(), npppn, th_or_proc,
371 th_or_proc_plural, pernode,
372 ngpu_use, gpu_use_plural);
377 if (ngpu_comp > npppn)
379 md_print_warn(cr, fplog,
380 "NOTE: potentially sub-optimal launch configuration, %s started with less\n"
381 " PP %s%s%s than GPU%s available.\n"
382 " Each PP %s can use only one GPU, %d GPU%s%s will be used.\n",
383 ShortProgram(), th_or_proc,
384 th_or_proc_plural, pernode, gpu_comp_plural,
385 th_or_proc, npppn, gpu_use_plural, pernode);
388 if (ngpu_use != npppn)
390 /* Avoid duplicate error messages.
391 * Unfortunately we can only do this at the physical node
392 * level, since the hardware setup and MPI process count
393 * might differ between physical nodes.
395 if (cr->rank_pp_intranode == 0)
398 "Incorrect launch configuration: mismatching number of PP %s%s and GPUs%s.\n"
399 "%s was started with %d PP %s%s%s, but only %d GPU%s were detected.",
400 th_or_proc, btMPI ? "s" : "es", pernode,
401 ShortProgram(), npppn, th_or_proc,
402 th_or_proc_plural, pernode,
403 ngpu_use, gpu_use_plural);
411 same_count = gmx_count_gpu_dev_shared(&hw_opt->gpu_opt);
415 md_print_info(cr, fplog,
416 "NOTE: You assigned %s to multiple %s%s.\n",
417 same_count > 1 ? "GPUs" : "a GPU", th_or_proc, btMPI ? "s" : "es");
425 /* Avoid other ranks to continue after
427 MPI_Barrier(cr->mpi_comm_mygroup);
433 /* Return 0 if none of the GPU (per node) are shared among PP ranks.
435 * Sharing GPUs among multiple PP ranks is possible when the user passes
436 * GPU IDs. Here we check for sharing and return a non-zero value when
437 * this is detected. Note that the return value represents the number of
438 * PP rank pairs that share a device.
440 int gmx_count_gpu_dev_shared(const gmx_gpu_opt_t *gpu_opt)
443 int ngpu = gpu_opt->ncuda_dev_use;
445 if (gpu_opt->bUserSet)
449 for (i = 0; i < ngpu - 1; i++)
451 for (j = i + 1; j < ngpu; j++)
453 same_count += (gpu_opt->cuda_dev_use[i] ==
454 gpu_opt->cuda_dev_use[j]);
462 /* Count and return the number of unique GPUs (per node) selected.
464 * As sharing GPUs among multiple PP ranks is possible when the user passes
465 * GPU IDs, the number of GPUs user (per node) can be different from the
466 * number of GPU IDs selected.
468 static int gmx_count_gpu_dev_unique(const gmx_gpu_info_t *gpu_info,
469 const gmx_gpu_opt_t *gpu_opt)
471 int i, uniq_count, ngpu;
477 ngpu = gpu_info->ncuda_dev;
480 snew(uniq_ids, ngpu);
482 /* Each element in uniq_ids will be set to 0 or 1. The n-th element set
483 * to 1 indicates that the respective GPU was selected to be used. */
484 for (i = 0; i < gpu_opt->ncuda_dev_use; i++)
486 uniq_ids[get_gpu_device_id(gpu_info, gpu_opt, i)] = 1;
488 /* Count the devices used. */
489 for (i = 0; i < ngpu; i++)
491 uniq_count += uniq_ids[i];
500 /* Return the number of hardware threads supported by the current CPU.
501 * We assume that this is equal with the number of CPUs reported to be
502 * online by the OS at the time of the call.
504 static int get_nthreads_hw_avail(FILE gmx_unused *fplog, const t_commrec gmx_unused *cr)
508 #if ((defined(WIN32) || defined( _WIN32 ) || defined(WIN64) || defined( _WIN64 )) && !(defined (__CYGWIN__) || defined (__CYGWIN32__)))
511 GetSystemInfo( &sysinfo );
512 ret = sysinfo.dwNumberOfProcessors;
513 #elif defined HAVE_SYSCONF
514 /* We are probably on Unix.
515 * Now check if we have the argument to use before executing the call
517 #if defined(_SC_NPROCESSORS_ONLN)
518 ret = sysconf(_SC_NPROCESSORS_ONLN);
519 #elif defined(_SC_NPROC_ONLN)
520 ret = sysconf(_SC_NPROC_ONLN);
521 #elif defined(_SC_NPROCESSORS_CONF)
522 ret = sysconf(_SC_NPROCESSORS_CONF);
523 #elif defined(_SC_NPROC_CONF)
524 ret = sysconf(_SC_NPROC_CONF);
526 #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!"
527 #endif /* End of check for sysconf argument values */
530 /* Neither windows nor Unix. No fscking idea how many CPUs we have! */
536 fprintf(debug, "Detected %d processors, will use this as the number "
537 "of supported hardware threads.\n", ret);
541 if (ret != gmx_omp_get_num_procs())
543 md_print_warn(cr, fplog,
544 "Number of CPUs detected (%d) does not match the number reported by OpenMP (%d).\n"
545 "Consider setting the launch configuration manually!",
546 ret, gmx_omp_get_num_procs());
553 static void gmx_detect_gpus(FILE *fplog, const t_commrec *cr)
557 MPI_Comm physicalnode_comm;
561 /* Under certain circumstances MPI ranks on the same physical node
562 * can not simultaneously access the same GPU(s). Therefore we run
563 * the detection only on one MPI rank per node and broadcast the info.
564 * Note that with thread-MPI only a single thread runs this code.
566 * TODO: We should also do CPU hardware detection only once on each
567 * physical node and broadcast it, instead of do it on every MPI rank.
570 /* A split of MPI_COMM_WORLD over physical nodes is only required here,
571 * so we create and destroy it locally.
573 MPI_Comm_rank(MPI_COMM_WORLD, &rank_world);
574 MPI_Comm_split(MPI_COMM_WORLD, gmx_physicalnode_id_hash(),
575 rank_world, &physicalnode_comm);
576 MPI_Comm_rank(physicalnode_comm, &rank_local);
578 /* Here there should be only one process, check this */
579 assert(cr->nnodes == 1 && cr->sim_nodeid == 0);
586 char detection_error[STRLEN] = "", sbuf[STRLEN];
588 if (detect_cuda_gpus(&hwinfo_g->gpu_info, detection_error) != 0)
590 if (detection_error[0] != '\0')
592 sprintf(sbuf, ":\n %s\n", detection_error);
598 md_print_warn(cr, fplog,
599 "NOTE: Error occurred during GPU detection%s"
600 " Can not use GPU acceleration, will fall back to CPU kernels.\n",
606 /* Broadcast the GPU info to the other ranks within this node */
607 MPI_Bcast(&hwinfo_g->gpu_info.ncuda_dev, 1, MPI_INT, 0, physicalnode_comm);
609 if (hwinfo_g->gpu_info.ncuda_dev > 0)
613 cuda_dev_size = hwinfo_g->gpu_info.ncuda_dev*sizeof_cuda_dev_info();
617 hwinfo_g->gpu_info.cuda_dev =
618 (cuda_dev_info_ptr_t)malloc(cuda_dev_size);
620 MPI_Bcast(hwinfo_g->gpu_info.cuda_dev, cuda_dev_size, MPI_BYTE,
621 0, physicalnode_comm);
622 MPI_Bcast(&hwinfo_g->gpu_info.ncuda_dev_compatible, 1, MPI_INT,
623 0, physicalnode_comm);
626 MPI_Comm_free(&physicalnode_comm);
630 gmx_hw_info_t *gmx_detect_hardware(FILE *fplog, const t_commrec *cr,
631 gmx_bool bDetectGPUs)
635 /* make sure no one else is doing the same thing */
636 ret = tMPI_Thread_mutex_lock(&hw_info_lock);
639 gmx_fatal(FARGS, "Error locking hwinfo mutex: %s", strerror(errno));
642 /* only initialize the hwinfo structure if it is not already initalized */
647 /* detect CPUID info; no fuss, we don't detect system-wide
648 * -- sloppy, but that's it for now */
649 if (gmx_cpuid_init(&hwinfo_g->cpuid_info) != 0)
651 gmx_fatal_collective(FARGS, cr, NULL, "CPUID detection failed!");
654 /* detect number of hardware threads */
655 hwinfo_g->nthreads_hw_avail = get_nthreads_hw_avail(fplog, cr);
658 hwinfo_g->gpu_info.ncuda_dev = 0;
659 hwinfo_g->gpu_info.cuda_dev = NULL;
660 hwinfo_g->gpu_info.ncuda_dev_compatible = 0;
662 /* Run the detection if the binary was compiled with GPU support
663 * and we requested detection.
665 hwinfo_g->gpu_info.bDetectGPUs =
666 (bGPUBinary && bDetectGPUs &&
667 getenv("GMX_DISABLE_GPU_DETECTION") == NULL);
668 if (hwinfo_g->gpu_info.bDetectGPUs)
670 gmx_detect_gpus(fplog, cr);
673 /* increase the reference counter */
676 ret = tMPI_Thread_mutex_unlock(&hw_info_lock);
679 gmx_fatal(FARGS, "Error unlocking hwinfo mutex: %s", strerror(errno));
685 void gmx_parse_gpu_ids(gmx_gpu_opt_t *gpu_opt)
689 if (gpu_opt->gpu_id != NULL && !bGPUBinary)
691 gmx_fatal(FARGS, "GPU ID string set, but %s was compiled without GPU support!", ShortProgram());
694 env = getenv("GMX_GPU_ID");
695 if (env != NULL && gpu_opt->gpu_id != NULL)
697 gmx_fatal(FARGS, "GMX_GPU_ID and -gpu_id can not be used at the same time");
701 env = gpu_opt->gpu_id;
704 /* parse GPU IDs if the user passed any */
707 /* Parse a "plain" GPU ID string which contains a sequence of
708 * digits corresponding to GPU IDs; the order will indicate
709 * the process/tMPI thread - GPU assignment. */
710 parse_digits_from_plain_string(env,
711 &gpu_opt->ncuda_dev_use,
712 &gpu_opt->cuda_dev_use);
714 if (gpu_opt->ncuda_dev_use == 0)
716 gmx_fatal(FARGS, "Empty GPU ID string encountered.\n%s\n",
720 gpu_opt->bUserSet = TRUE;
724 void gmx_select_gpu_ids(FILE *fplog, const t_commrec *cr,
725 const gmx_gpu_info_t *gpu_info,
726 gmx_bool bForceUseGPU,
727 gmx_gpu_opt_t *gpu_opt)
730 char sbuf[STRLEN], stmp[STRLEN];
732 /* Bail if binary is not compiled with GPU acceleration, but this is either
733 * explicitly (-nb gpu) or implicitly (gpu ID passed) requested. */
734 if (bForceUseGPU && !bGPUBinary)
736 gmx_fatal(FARGS, "GPU acceleration requested, but %s was compiled without GPU support!", ShortProgram());
739 if (gpu_opt->bUserSet)
741 /* Check the GPU IDs passed by the user.
742 * (GPU IDs have been parsed by gmx_parse_gpu_ids before)
747 snew(checkres, gpu_opt->ncuda_dev_use);
749 res = check_selected_cuda_gpus(checkres, gpu_info, gpu_opt);
753 print_gpu_detection_stats(fplog, gpu_info, cr);
755 sprintf(sbuf, "Some of the requested GPUs do not exist, behave strangely, or are not compatible:\n");
756 for (i = 0; i < gpu_opt->ncuda_dev_use; i++)
758 if (checkres[i] != egpuCompatible)
760 sprintf(stmp, " GPU #%d: %s\n",
761 gpu_opt->cuda_dev_use[i],
762 gpu_detect_res_str[checkres[i]]);
766 gmx_fatal(FARGS, "%s", sbuf);
773 pick_compatible_gpus(&hwinfo_g->gpu_info, gpu_opt);
775 if (gpu_opt->ncuda_dev_use > cr->nrank_pp_intranode)
777 /* We picked more GPUs than we can use: limit the number.
778 * We print detailed messages about this later in
779 * gmx_check_hw_runconf_consistency.
781 limit_num_gpus_used(gpu_opt, cr->nrank_pp_intranode);
784 gpu_opt->bUserSet = FALSE;
787 /* If the user asked for a GPU, check whether we have a GPU */
788 if (bForceUseGPU && gpu_info->ncuda_dev_compatible == 0)
790 gmx_fatal(FARGS, "GPU acceleration requested, but no compatible GPUs were detected.");
794 static void limit_num_gpus_used(gmx_gpu_opt_t *gpu_opt, int count)
800 ndev_use = gpu_opt->ncuda_dev_use;
802 if (count > ndev_use)
804 /* won't increase the # of GPUs */
811 sprintf(sbuf, "Limiting the number of GPUs to <1 doesn't make sense (detected %d, %d requested)!",
816 /* TODO: improve this implementation: either sort GPUs or remove the weakest here */
817 gpu_opt->ncuda_dev_use = count;
820 void gmx_hardware_info_free(gmx_hw_info_t *hwinfo)
824 ret = tMPI_Thread_mutex_lock(&hw_info_lock);
827 gmx_fatal(FARGS, "Error locking hwinfo mutex: %s", strerror(errno));
830 /* decrease the reference counter */
834 if (hwinfo != hwinfo_g)
836 gmx_incons("hwinfo < hwinfo_g");
841 gmx_incons("n_hwinfo < 0");
846 gmx_cpuid_done(hwinfo_g->cpuid_info);
847 free_gpu_info(&hwinfo_g->gpu_info);
851 ret = tMPI_Thread_mutex_unlock(&hw_info_lock);
854 gmx_fatal(FARGS, "Error unlocking hwinfo mutex: %s", strerror(errno));