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.
50 #include "types/enums.h"
51 #include "types/hw_info.h"
52 #include "types/commrec.h"
54 #include "md_logging.h"
55 #include "gmx_cpuid.h"
56 #include "gpu_utils.h"
58 #include "gmx_detect_hardware.h"
59 #include "md_logging.h"
61 #include "gromacs/utility/basenetwork.h"
62 #include "gromacs/utility/cstringutil.h"
63 #include "gromacs/utility/exceptions.h"
64 #include "gromacs/utility/fatalerror.h"
65 #include "gromacs/utility/gmxomp.h"
66 #include "gromacs/utility/smalloc.h"
67 #include "gromacs/utility/stringutil.h"
69 #include "thread_mpi/threads.h"
71 #ifdef GMX_NATIVE_WINDOWS
76 const gmx_bool bGPUBinary = TRUE;
78 const gmx_bool bGPUBinary = FALSE;
81 static const char * invalid_gpuid_hint =
82 "A delimiter-free sequence of valid numeric IDs of available GPUs is expected.";
84 /* The globally shared hwinfo structure. */
85 static gmx_hw_info_t *hwinfo_g;
86 /* A reference counter for the hwinfo structure */
87 static int n_hwinfo = 0;
88 /* A lock to protect the hwinfo structure */
89 static tMPI_Thread_mutex_t hw_info_lock = TMPI_THREAD_MUTEX_INITIALIZER;
93 static void limit_num_gpus_used(gmx_gpu_opt_t *gpu_opt, int count);
94 static int gmx_count_gpu_dev_unique(const gmx_gpu_info_t *gpu_info,
95 const gmx_gpu_opt_t *gpu_opt);
97 static void sprint_gpus(char *sbuf, const gmx_gpu_info_t *gpu_info)
102 ndev = gpu_info->ncuda_dev;
105 for (i = 0; i < ndev; i++)
107 get_gpu_device_info_string(stmp, gpu_info, i);
117 static void print_gpu_detection_stats(FILE *fplog,
118 const gmx_gpu_info_t *gpu_info,
121 char onhost[266], stmp[STRLEN];
124 if (!gpu_info->bDetectGPUs)
126 /* We skipped the detection, so don't print detection stats */
130 ngpu = gpu_info->ncuda_dev;
132 #if defined GMX_MPI && !defined GMX_THREAD_MPI
133 /* We only print the detection on one, of possibly multiple, nodes */
134 strncpy(onhost, " on host ", 10);
135 gmx_gethostname(onhost+9, 256);
137 /* We detect all relevant GPUs */
138 strncpy(onhost, "", 1);
143 sprint_gpus(stmp, gpu_info);
144 md_print_warn(cr, fplog, "%d GPU%s detected%s:\n%s\n",
145 ngpu, (ngpu > 1) ? "s" : "", onhost, stmp);
149 md_print_warn(cr, fplog, "No GPUs detected%s\n", onhost);
153 /*! \brief Helper function for writing comma-separated GPU IDs.
155 * \param[in] ids A container of integer GPU IDs
156 * \return A comma-separated string of GPU IDs */
157 template <typename Container>
158 static std::string makeGpuIdsString(const Container &ids)
164 typename Container::const_iterator it = ids.begin();
165 output += gmx::formatString("%d", *it);
166 for (++it; it != ids.end(); ++it)
168 output += gmx::formatString(",%d", *it);
174 /*! \brief Helper function for reporting GPU usage information
175 * in the mdrun log file
177 * \param[in] gpu_info Pointer to per-node GPU info struct
178 * \param[in] gpu_opt Pointer to per-node GPU options struct
179 * \param[in] numPpRanks Number of PP ranks per node
180 * \return String to write to the log file
181 * \throws std::bad_alloc if out of memory */
183 makeGpuUsageReport(const gmx_gpu_info_t *gpu_info,
184 const gmx_gpu_opt_t *gpu_opt,
187 int ngpu_use = gpu_opt->ncuda_dev_use;
188 int ngpu_comp = gpu_info->ncuda_dev_compatible;
190 /* Issue a note if GPUs are available but not used */
191 if (ngpu_comp > 0 && ngpu_use < 1)
193 return gmx::formatString("%d compatible GPU%s detected in the system, but none will be used.\n"
194 "Consider trying GPU acceleration with the Verlet scheme!\n",
195 ngpu_comp, (ngpu_comp > 1) ? "s" : "");
201 std::vector<int> gpuIdsInUse;
202 for (int i = 0; i < ngpu_use; i++)
204 gpuIdsInUse.push_back(get_gpu_device_id(gpu_info, gpu_opt, i));
206 std::string gpuIdsString = makeGpuIdsString(gpuIdsInUse);
207 int numGpusInUse = gmx_count_gpu_dev_unique(gpu_info, gpu_opt);
208 bool bPluralGpus = numGpusInUse > 1;
210 output += gmx::formatString("%d GPU%s %sselected for this run.\n"
211 "Mapping of GPU ID%s to the %d PP rank%s in this node: %s\n",
212 numGpusInUse, bPluralGpus ? "s" : "",
213 gpu_opt->bUserSet ? "user-" : "auto-",
214 bPluralGpus ? "s" : "",
216 (numPpRanks > 1) ? "s" : "",
217 gpuIdsString.c_str());
223 /* Give a suitable fatal error or warning if the build configuration
224 and runtime CPU do not match. */
226 check_use_of_rdtscp_on_this_cpu(FILE *fplog,
228 const gmx_hw_info_t *hwinfo)
230 gmx_bool bCpuHasRdtscp, bBinaryUsesRdtscp;
232 bBinaryUsesRdtscp = TRUE;
234 bBinaryUsesRdtscp = FALSE;
237 bCpuHasRdtscp = gmx_cpuid_feature(hwinfo->cpuid_info, GMX_CPUID_FEATURE_X86_RDTSCP);
239 if (!bCpuHasRdtscp && bBinaryUsesRdtscp)
241 gmx_fatal(FARGS, "The %s executable was compiled to use the rdtscp CPU instruction. "
242 "However, this is not supported by the current hardware and continuing would lead to a crash. "
243 "Please rebuild GROMACS with the GMX_USE_RDTSCP=OFF CMake option.",
247 if (bCpuHasRdtscp && !bBinaryUsesRdtscp)
249 md_print_warn(cr, fplog, "The current CPU can measure timings more accurately than the code in\n"
250 "%s was configured to use. This might affect your simulation\n"
251 "speed as accurate timings are needed for load-balancing.\n"
252 "Please consider rebuilding %s with the GMX_USE_RDTSCP=OFF CMake option.\n",
253 ShortProgram(), ShortProgram());
257 void gmx_check_hw_runconf_consistency(FILE *fplog,
258 const gmx_hw_info_t *hwinfo,
260 const gmx_hw_opt_t *hw_opt,
264 char th_or_proc[STRLEN], th_or_proc_plural[STRLEN], pernode[STRLEN];
265 gmx_bool btMPI, bMPI, bMaxMpiThreadsSet, bNthreadsAuto, bEmulateGPU;
270 /* Below we only do consistency checks for PP and GPUs,
271 * this is irrelevant for PME only nodes, so in that case we return
274 if (!(cr->duty & DUTY_PP))
279 #if defined(GMX_THREAD_MPI)
282 bNthreadsAuto = (hw_opt->nthreads_tmpi < 1);
283 #elif defined(GMX_LIB_MPI)
286 bNthreadsAuto = FALSE;
290 bNthreadsAuto = FALSE;
293 /* GPU emulation detection is done later, but we need here as well
294 * -- uncool, but there's no elegant workaround */
295 bEmulateGPU = (getenv("GMX_EMULATE_GPU") != NULL);
296 bMaxMpiThreadsSet = (getenv("GMX_MAX_MPI_THREADS") != NULL);
298 /* check the SIMD level mdrun is compiled with against hardware
300 /* TODO: Here we assume homogeneous hardware which is not necessarily
301 the case! Might not hurt to add an extra check over MPI. */
302 gmx_cpuid_simd_check(hwinfo->cpuid_info, fplog, SIMMASTER(cr));
304 check_use_of_rdtscp_on_this_cpu(fplog, cr, hwinfo);
306 /* NOTE: this print is only for and on one physical node */
307 print_gpu_detection_stats(fplog, &hwinfo->gpu_info, cr);
309 if (hwinfo->gpu_info.ncuda_dev_compatible > 0)
311 std::string gpuUseageReport;
314 gpuUseageReport = makeGpuUsageReport(&hwinfo->gpu_info,
316 cr->nrank_pp_intranode);
318 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
320 /* NOTE: this print is only for and on one physical node */
321 md_print_info(cr, fplog, gpuUseageReport.c_str());
324 /* Need to ensure that we have enough GPUs:
325 * - need one GPU per PP node
326 * - no GPU oversubscription with tMPI
328 /* number of PP processes per node */
329 npppn = cr->nrank_pp_intranode;
332 th_or_proc_plural[0] = '\0';
335 sprintf(th_or_proc, "thread-MPI thread");
338 sprintf(th_or_proc_plural, "s");
343 sprintf(th_or_proc, "MPI process");
346 sprintf(th_or_proc_plural, "es");
348 sprintf(pernode, " per node");
352 /* neither MPI nor tMPI */
353 sprintf(th_or_proc, "process");
356 if (bUseGPU && hwinfo->gpu_info.ncuda_dev_compatible > 0 &&
359 int ngpu_comp, ngpu_use;
360 char gpu_comp_plural[2], gpu_use_plural[2];
362 ngpu_comp = hwinfo->gpu_info.ncuda_dev_compatible;
363 ngpu_use = hw_opt->gpu_opt.ncuda_dev_use;
365 sprintf(gpu_comp_plural, "%s", (ngpu_comp > 1) ? "s" : "");
366 sprintf(gpu_use_plural, "%s", (ngpu_use > 1) ? "s" : "");
368 /* number of tMPI threads auto-adjusted */
369 if (btMPI && bNthreadsAuto)
371 if (hw_opt->gpu_opt.bUserSet && npppn < ngpu_use)
373 /* The user manually provided more GPUs than threads we
374 could automatically start. */
376 "%d GPU%s provided, but only %d PP thread-MPI thread%s coud be started.\n"
377 "%s requires one PP tread-MPI thread per GPU; use fewer GPUs%s.",
378 ngpu_use, gpu_use_plural,
379 npppn, th_or_proc_plural,
380 ShortProgram(), bMaxMpiThreadsSet ? "\nor allow more threads to be used" : "");
383 if (!hw_opt->gpu_opt.bUserSet && npppn < ngpu_comp)
385 /* There are more GPUs than tMPI threads; we have
386 limited the number GPUs used. */
387 md_print_warn(cr, fplog,
388 "NOTE: %d GPU%s were detected, but only %d PP thread-MPI thread%s can be started.\n"
389 " %s can use one GPU per PP tread-MPI thread, so only %d GPU%s will be used.%s\n",
390 ngpu_comp, gpu_comp_plural,
391 npppn, th_or_proc_plural,
392 ShortProgram(), npppn,
393 npppn > 1 ? "s" : "",
394 bMaxMpiThreadsSet ? "\n Also, you can allow more threads to be used by increasing GMX_MAX_MPI_THREADS" : "");
398 if (hw_opt->gpu_opt.bUserSet)
400 if (ngpu_use != npppn)
403 "Incorrect launch configuration: mismatching number of PP %s%s and GPUs%s.\n"
404 "%s was started with %d PP %s%s%s, but you provided %d GPU%s.",
405 th_or_proc, btMPI ? "s" : "es", pernode,
406 ShortProgram(), npppn, th_or_proc,
407 th_or_proc_plural, pernode,
408 ngpu_use, gpu_use_plural);
413 if (ngpu_comp > npppn)
415 md_print_warn(cr, fplog,
416 "NOTE: potentially sub-optimal launch configuration, %s started with less\n"
417 " PP %s%s%s than GPU%s available.\n"
418 " Each PP %s can use only one GPU, %d GPU%s%s will be used.\n",
419 ShortProgram(), th_or_proc,
420 th_or_proc_plural, pernode, gpu_comp_plural,
421 th_or_proc, npppn, gpu_use_plural, pernode);
424 if (ngpu_use != npppn)
426 /* Avoid duplicate error messages.
427 * Unfortunately we can only do this at the physical node
428 * level, since the hardware setup and MPI process count
429 * might differ between physical nodes.
431 if (cr->rank_pp_intranode == 0)
434 "Incorrect launch configuration: mismatching number of PP %s%s and GPUs%s.\n"
435 "%s was started with %d PP %s%s%s, but only %d GPU%s were detected.",
436 th_or_proc, btMPI ? "s" : "es", pernode,
437 ShortProgram(), npppn, th_or_proc,
438 th_or_proc_plural, pernode,
439 ngpu_use, gpu_use_plural);
447 same_count = gmx_count_gpu_dev_shared(&hw_opt->gpu_opt);
451 md_print_info(cr, fplog,
452 "NOTE: You assigned %s to multiple %s%s.\n",
453 same_count > 1 ? "GPUs" : "a GPU", th_or_proc, btMPI ? "s" : "es");
461 /* Avoid other ranks to continue after
463 MPI_Barrier(cr->mpi_comm_mygroup);
469 /* Return 0 if none of the GPU (per node) are shared among PP ranks.
471 * Sharing GPUs among multiple PP ranks is possible when the user passes
472 * GPU IDs. Here we check for sharing and return a non-zero value when
473 * this is detected. Note that the return value represents the number of
474 * PP rank pairs that share a device.
476 int gmx_count_gpu_dev_shared(const gmx_gpu_opt_t *gpu_opt)
479 int ngpu = gpu_opt->ncuda_dev_use;
481 if (gpu_opt->bUserSet)
485 for (i = 0; i < ngpu - 1; i++)
487 for (j = i + 1; j < ngpu; j++)
489 same_count += (gpu_opt->cuda_dev_use[i] ==
490 gpu_opt->cuda_dev_use[j]);
498 /* Count and return the number of unique GPUs (per node) selected.
500 * As sharing GPUs among multiple PP ranks is possible when the user passes
501 * GPU IDs, the number of GPUs user (per node) can be different from the
502 * number of GPU IDs selected.
504 static int gmx_count_gpu_dev_unique(const gmx_gpu_info_t *gpu_info,
505 const gmx_gpu_opt_t *gpu_opt)
507 int i, uniq_count, ngpu;
513 ngpu = gpu_info->ncuda_dev;
516 snew(uniq_ids, ngpu);
518 /* Each element in uniq_ids will be set to 0 or 1. The n-th element set
519 * to 1 indicates that the respective GPU was selected to be used. */
520 for (i = 0; i < gpu_opt->ncuda_dev_use; i++)
522 uniq_ids[get_gpu_device_id(gpu_info, gpu_opt, i)] = 1;
524 /* Count the devices used. */
525 for (i = 0; i < ngpu; i++)
527 uniq_count += uniq_ids[i];
536 /* Return the number of hardware threads supported by the current CPU.
537 * We assume that this is equal with the number of "processors"
538 * reported to be online by the OS at the time of the call. The
539 * definition of "processor" is according to an old POSIX standard.
541 * Note that the number of hardware threads is generally greater than
542 * the number of cores (e.g. x86 hyper-threading, Power). Managing the
543 * mapping of software threads to hardware threads is managed
545 static int get_nthreads_hw_avail(FILE gmx_unused *fplog, const t_commrec gmx_unused *cr)
549 #if ((defined(WIN32) || defined( _WIN32 ) || defined(WIN64) || defined( _WIN64 )) && !(defined (__CYGWIN__) || defined (__CYGWIN32__)))
552 GetSystemInfo( &sysinfo );
553 ret = sysinfo.dwNumberOfProcessors;
554 #elif defined HAVE_SYSCONF
555 /* We are probably on Unix.
556 * Now check if we have the argument to use before executing the call
558 #if defined(_SC_NPROCESSORS_ONLN)
559 ret = sysconf(_SC_NPROCESSORS_ONLN);
560 #elif defined(_SC_NPROC_ONLN)
561 ret = sysconf(_SC_NPROC_ONLN);
562 #elif defined(_SC_NPROCESSORS_CONF)
563 ret = sysconf(_SC_NPROCESSORS_CONF);
564 #elif defined(_SC_NPROC_CONF)
565 ret = sysconf(_SC_NPROC_CONF);
567 #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!"
568 #endif /* End of check for sysconf argument values */
571 /* Neither windows nor Unix. No fscking idea how many hardware threads we have! */
577 fprintf(debug, "Detected %d hardware threads to use.\n", ret);
581 if (ret != gmx_omp_get_num_procs())
583 md_print_warn(cr, fplog,
584 "Number of hardware threads detected (%d) does not match the number reported by OpenMP (%d).\n"
585 "Consider setting the launch configuration manually!",
586 ret, gmx_omp_get_num_procs());
593 static void gmx_detect_gpus(FILE *fplog, const t_commrec *cr)
597 MPI_Comm physicalnode_comm;
601 /* Under certain circumstances MPI ranks on the same physical node
602 * can not simultaneously access the same GPU(s). Therefore we run
603 * the detection only on one MPI rank per node and broadcast the info.
604 * Note that with thread-MPI only a single thread runs this code.
606 * TODO: We should also do CPU hardware detection only once on each
607 * physical node and broadcast it, instead of do it on every MPI rank.
610 /* A split of MPI_COMM_WORLD over physical nodes is only required here,
611 * so we create and destroy it locally.
613 MPI_Comm_rank(MPI_COMM_WORLD, &rank_world);
614 MPI_Comm_split(MPI_COMM_WORLD, gmx_physicalnode_id_hash(),
615 rank_world, &physicalnode_comm);
616 MPI_Comm_rank(physicalnode_comm, &rank_local);
618 /* Here there should be only one process, check this */
619 assert(cr->nnodes == 1 && cr->sim_nodeid == 0);
626 char detection_error[STRLEN] = "", sbuf[STRLEN];
628 if (detect_cuda_gpus(&hwinfo_g->gpu_info, detection_error) != 0)
630 if (detection_error[0] != '\0')
632 sprintf(sbuf, ":\n %s\n", detection_error);
638 md_print_warn(cr, fplog,
639 "NOTE: Error occurred during GPU detection%s"
640 " Can not use GPU acceleration, will fall back to CPU kernels.\n",
646 /* Broadcast the GPU info to the other ranks within this node */
647 MPI_Bcast(&hwinfo_g->gpu_info.ncuda_dev, 1, MPI_INT, 0, physicalnode_comm);
649 if (hwinfo_g->gpu_info.ncuda_dev > 0)
653 cuda_dev_size = hwinfo_g->gpu_info.ncuda_dev*sizeof_cuda_dev_info();
657 hwinfo_g->gpu_info.cuda_dev =
658 (cuda_dev_info_ptr_t)malloc(cuda_dev_size);
660 MPI_Bcast(hwinfo_g->gpu_info.cuda_dev, cuda_dev_size, MPI_BYTE,
661 0, physicalnode_comm);
662 MPI_Bcast(&hwinfo_g->gpu_info.ncuda_dev_compatible, 1, MPI_INT,
663 0, physicalnode_comm);
666 MPI_Comm_free(&physicalnode_comm);
670 gmx_hw_info_t *gmx_detect_hardware(FILE *fplog, const t_commrec *cr,
671 gmx_bool bDetectGPUs)
675 /* make sure no one else is doing the same thing */
676 ret = tMPI_Thread_mutex_lock(&hw_info_lock);
679 gmx_fatal(FARGS, "Error locking hwinfo mutex: %s", strerror(errno));
682 /* only initialize the hwinfo structure if it is not already initalized */
687 /* detect CPUID info; no fuss, we don't detect system-wide
688 * -- sloppy, but that's it for now */
689 if (gmx_cpuid_init(&hwinfo_g->cpuid_info) != 0)
691 gmx_fatal_collective(FARGS, cr, NULL, "CPUID detection failed!");
694 /* detect number of hardware threads */
695 hwinfo_g->nthreads_hw_avail = get_nthreads_hw_avail(fplog, cr);
698 hwinfo_g->gpu_info.ncuda_dev = 0;
699 hwinfo_g->gpu_info.cuda_dev = NULL;
700 hwinfo_g->gpu_info.ncuda_dev_compatible = 0;
702 /* Run the detection if the binary was compiled with GPU support
703 * and we requested detection.
705 hwinfo_g->gpu_info.bDetectGPUs =
706 (bGPUBinary && bDetectGPUs &&
707 getenv("GMX_DISABLE_GPU_DETECTION") == NULL);
708 if (hwinfo_g->gpu_info.bDetectGPUs)
710 gmx_detect_gpus(fplog, cr);
713 /* increase the reference counter */
716 ret = tMPI_Thread_mutex_unlock(&hw_info_lock);
719 gmx_fatal(FARGS, "Error unlocking hwinfo mutex: %s", strerror(errno));
725 void gmx_parse_gpu_ids(gmx_gpu_opt_t *gpu_opt)
729 if (gpu_opt->gpu_id != NULL && !bGPUBinary)
731 gmx_fatal(FARGS, "GPU ID string set, but %s was compiled without GPU support!", ShortProgram());
734 env = getenv("GMX_GPU_ID");
735 if (env != NULL && gpu_opt->gpu_id != NULL)
737 gmx_fatal(FARGS, "GMX_GPU_ID and -gpu_id can not be used at the same time");
741 env = gpu_opt->gpu_id;
744 /* parse GPU IDs if the user passed any */
747 /* Parse a "plain" GPU ID string which contains a sequence of
748 * digits corresponding to GPU IDs; the order will indicate
749 * the process/tMPI thread - GPU assignment. */
750 parse_digits_from_plain_string(env,
751 &gpu_opt->ncuda_dev_use,
752 &gpu_opt->cuda_dev_use);
754 if (gpu_opt->ncuda_dev_use == 0)
756 gmx_fatal(FARGS, "Empty GPU ID string encountered.\n%s\n",
760 gpu_opt->bUserSet = TRUE;
764 void gmx_select_gpu_ids(FILE *fplog, const t_commrec *cr,
765 const gmx_gpu_info_t *gpu_info,
766 gmx_bool bForceUseGPU,
767 gmx_gpu_opt_t *gpu_opt)
770 char sbuf[STRLEN], stmp[STRLEN];
772 /* Bail if binary is not compiled with GPU acceleration, but this is either
773 * explicitly (-nb gpu) or implicitly (gpu ID passed) requested. */
774 if (bForceUseGPU && !bGPUBinary)
776 gmx_fatal(FARGS, "GPU acceleration requested, but %s was compiled without GPU support!", ShortProgram());
779 if (gpu_opt->bUserSet)
781 /* Check the GPU IDs passed by the user.
782 * (GPU IDs have been parsed by gmx_parse_gpu_ids before)
787 snew(checkres, gpu_opt->ncuda_dev_use);
789 res = check_selected_cuda_gpus(checkres, gpu_info, gpu_opt);
793 print_gpu_detection_stats(fplog, gpu_info, cr);
795 sprintf(sbuf, "Some of the requested GPUs do not exist, behave strangely, or are not compatible:\n");
796 for (i = 0; i < gpu_opt->ncuda_dev_use; i++)
798 if (checkres[i] != egpuCompatible)
800 sprintf(stmp, " GPU #%d: %s\n",
801 gpu_opt->cuda_dev_use[i],
802 gpu_detect_res_str[checkres[i]]);
806 gmx_fatal(FARGS, "%s", sbuf);
813 pick_compatible_gpus(&hwinfo_g->gpu_info, gpu_opt);
815 if (gpu_opt->ncuda_dev_use > cr->nrank_pp_intranode)
817 /* We picked more GPUs than we can use: limit the number.
818 * We print detailed messages about this later in
819 * gmx_check_hw_runconf_consistency.
821 limit_num_gpus_used(gpu_opt, cr->nrank_pp_intranode);
824 gpu_opt->bUserSet = FALSE;
827 /* If the user asked for a GPU, check whether we have a GPU */
828 if (bForceUseGPU && gpu_info->ncuda_dev_compatible == 0)
830 gmx_fatal(FARGS, "GPU acceleration requested, but no compatible GPUs were detected.");
834 static void limit_num_gpus_used(gmx_gpu_opt_t *gpu_opt, int count)
840 ndev_use = gpu_opt->ncuda_dev_use;
842 if (count > ndev_use)
844 /* won't increase the # of GPUs */
851 sprintf(sbuf, "Limiting the number of GPUs to <1 doesn't make sense (detected %d, %d requested)!",
856 /* TODO: improve this implementation: either sort GPUs or remove the weakest here */
857 gpu_opt->ncuda_dev_use = count;
860 void gmx_hardware_info_free(gmx_hw_info_t *hwinfo)
864 ret = tMPI_Thread_mutex_lock(&hw_info_lock);
867 gmx_fatal(FARGS, "Error locking hwinfo mutex: %s", strerror(errno));
870 /* decrease the reference counter */
874 if (hwinfo != hwinfo_g)
876 gmx_incons("hwinfo < hwinfo_g");
881 gmx_incons("n_hwinfo < 0");
886 gmx_cpuid_done(hwinfo_g->cpuid_info);
887 free_gpu_info(&hwinfo_g->gpu_info);
891 ret = tMPI_Thread_mutex_unlock(&hw_info_lock);
894 gmx_fatal(FARGS, "Error unlocking hwinfo mutex: %s", strerror(errno));