void gmx_check_hw_runconf_consistency(FILE *fplog,
const gmx_hw_info_t *hwinfo,
const t_commrec *cr,
const gmx_hw_opt_t *hw_opt,
gmx_bool bUseGPU)
{
int npppn;
char th_or_proc[STRLEN], th_or_proc_plural[STRLEN], pernode[STRLEN];
gmx_bool btMPI, bMPI, bNthreadsAuto, bEmulateGPU;
GMX_RELEASE_ASSERT(hwinfo, "hwinfo must be a non-NULL pointer");
GMX_RELEASE_ASSERT(cr, "cr must be a non-NULL pointer");
/* Below we only do consistency checks for PP and GPUs,
* this is irrelevant for PME only nodes, so in that case we return
* here.
*/
if (!(cr->duty & DUTY_PP))
{
return;
}
#if GMX_THREAD_MPI
bMPI = FALSE;
btMPI = TRUE;
bNthreadsAuto = (hw_opt->nthreads_tmpi < 1);
#elif GMX_LIB_MPI
bMPI = TRUE;
btMPI = FALSE;
bNthreadsAuto = FALSE;
#else
bMPI = FALSE;
btMPI = FALSE;
bNthreadsAuto = FALSE;
#endif
/* GPU emulation detection is done later, but we need here as well
* -- uncool, but there's no elegant workaround */
bEmulateGPU = (getenv("GMX_EMULATE_GPU") != NULL);
if (hwinfo->gpu_info.n_dev_compatible > 0)
{
std::string gpuUseageReport;
try
{
gpuUseageReport = makeGpuUsageReport(&hwinfo->gpu_info,
&hw_opt->gpu_opt,
cr->nrank_pp_intranode,
bMPI && cr->nnodes > 1);
}
GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
/* NOTE: this print is only for and on one physical node */
md_print_info(cr, fplog, "%s\n", gpuUseageReport.c_str());
}
void gmx_check_hw_runconf_consistency(FILE *fplog,
const gmx_hw_info_t *hwinfo,
const t_commrec *cr,
const gmx_hw_opt_t *hw_opt,
gmx_bool bUseGPU)
{
int npppn;
char th_or_proc[STRLEN], th_or_proc_plural[STRLEN], pernode[STRLEN];
gmx_bool btMPI, bMPI, bMaxMpiThreadsSet, bNthreadsAuto, bEmulateGPU;
assert(hwinfo);
assert(cr);
/* Below we only do consistency checks for PP and GPUs,
* this is irrelevant for PME only nodes, so in that case we return
* here.
*/
if (!(cr->duty & DUTY_PP))
{
return;
}
#if defined(GMX_THREAD_MPI)
bMPI = FALSE;
btMPI = TRUE;
bNthreadsAuto = (hw_opt->nthreads_tmpi < 1);
#elif defined(GMX_LIB_MPI)
bMPI = TRUE;
btMPI = FALSE;
bNthreadsAuto = FALSE;
#else
bMPI = FALSE;
btMPI = FALSE;
bNthreadsAuto = FALSE;
#endif
/* GPU emulation detection is done later, but we need here as well
* -- uncool, but there's no elegant workaround */
bEmulateGPU = (getenv("GMX_EMULATE_GPU") != NULL);
bMaxMpiThreadsSet = (getenv("GMX_MAX_MPI_THREADS") != NULL);
/* check the SIMD level mdrun is compiled with against hardware
capabilities */
/* TODO: Here we assume homogeneous hardware which is not necessarily
the case! Might not hurt to add an extra check over MPI. */
gmx_cpuid_simd_check(hwinfo->cpuid_info, fplog, SIMMASTER(cr));
check_use_of_rdtscp_on_this_cpu(fplog, cr, hwinfo);
/* NOTE: this print is only for and on one physical node */
print_gpu_detection_stats(fplog, &hwinfo->gpu_info, cr);
if (hwinfo->gpu_info.ncuda_dev_compatible > 0)
{
std::string gpuUseageReport;
try
{
gpuUseageReport = makeGpuUsageReport(&hwinfo->gpu_info,
&hw_opt->gpu_opt,
cr->nrank_pp_intranode);
}
GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
/* NOTE: this print is only for and on one physical node */
md_print_info(cr, fplog, gpuUseageReport.c_str());
}
/* Set CPU affinity. Can be important for performance.
On some systems (e.g. Cray) CPU Affinity is set by default.
But default assigning doesn't work (well) with only some ranks
having threads. This causes very low performance.
External tools have cumbersome syntax for setting affinity
in the case that only some ranks have threads.
Thus it is important that GROMACS sets the affinity internally
if only PME is using threads.
*/
void
gmx_set_thread_affinity(FILE *fplog,
const t_commrec *cr,
gmx_hw_opt_t *hw_opt,
int nthreads_pme,
const gmx_hw_info_t *hwinfo,
const t_inputrec *inputrec)
{
int nth_affinity_set, thread_id_node, thread_id,
nthread_local, nthread_node, nthread_hw_max, nphyscore;
int offset;
const int *locality_order;
int rc;
if (hw_opt->thread_affinity == threadaffOFF)
{
/* Nothing to do */
return;
}
/* If the tMPI thread affinity setting is not supported encourage the user
* to report it as it's either a bug or an exotic platform which we might
* want to support. */
if (tMPI_Thread_setaffinity_support() != TMPI_SETAFFINITY_SUPPORT_YES)
{
/* we know Mac OS doesn't support setting thread affinity, so there's
no point in warning the user in that case. In any other case
the user might be able to do something about it. */
#ifndef __APPLE__
md_print_warn(NULL, fplog,
"Can not set thread affinities on the current platform. On NUMA systems this\n"
"can cause performance degradation. If you think your platform should support\n"
"setting affinities, contact the GROMACS developers.");
#endif /* __APPLE__ */
return;
}
/* threads on this MPI process or TMPI thread */
if (cr->duty & DUTY_PP)
{
nthread_local = gmx_omp_nthreads_get(emntNonbonded);
}
else
{
nthread_local = gmx_omp_nthreads_get(emntPME);
}
/* map the current process to cores */
thread_id_node = 0;
nthread_node = nthread_local;
#ifdef GMX_MPI
if (PAR(cr) || MULTISIM(cr))
{
/* We need to determine a scan of the thread counts in this
* compute node.
*/
MPI_Comm comm_intra;
MPI_Comm_split(MPI_COMM_WORLD, gmx_hostname_num(), cr->rank_intranode,
&comm_intra);
MPI_Scan(&nthread_local, &thread_id_node, 1, MPI_INT, MPI_SUM, comm_intra);
/* MPI_Scan is inclusive, but here we need exclusive */
thread_id_node -= nthread_local;
/* Get the total number of threads on this physical node */
MPI_Allreduce(&nthread_local, &nthread_node, 1, MPI_INT, MPI_SUM, comm_intra);
MPI_Comm_free(&comm_intra);
}
#endif
if (hw_opt->thread_affinity == threadaffAUTO &&
nthread_node != hwinfo->nthreads_hw_avail)
{
if (nthread_node > 1 && nthread_node < hwinfo->nthreads_hw_avail)
{
md_print_warn(cr, fplog,
"NOTE: The number of threads is not equal to the number of (logical) cores\n"
" and the -pin option is set to auto: will not pin thread to cores.\n"
" This can lead to significant performance degradation.\n"
" Consider using -pin on (and -pinoffset in case you run multiple jobs).\n");
}
return;
}
offset = 0;
if (hw_opt->core_pinning_offset != 0)
{
offset = hw_opt->core_pinning_offset;
md_print_info(cr, fplog, "Applying core pinning offset %d\n", offset);
}
rc = get_thread_affinity_layout(fplog, cr, hwinfo,
nthread_node,
offset, &hw_opt->core_pinning_stride,
&locality_order);
if (rc != 0)
{
/* Incompatible layout, don't pin, warning was already issued */
return;
}
/* Set the per-thread affinity. In order to be able to check the success
* of affinity settings, we will set nth_affinity_set to 1 on threads
* where the affinity setting succeded and to 0 where it failed.
* Reducing these 0/1 values over the threads will give the total number
* of threads on which we succeeded.
*/
nth_affinity_set = 0;
#pragma omp parallel firstprivate(thread_id_node) num_threads(nthread_local) \
reduction(+:nth_affinity_set)
{
int index, core;
gmx_bool setaffinity_ret;
thread_id = gmx_omp_get_thread_num();
thread_id_node += thread_id;
index = offset + thread_id_node*hw_opt->core_pinning_stride;
if (locality_order != NULL)
{
core = locality_order[index];
}
else
{
core = index;
}
setaffinity_ret = tMPI_Thread_setaffinity_single(tMPI_Thread_self(), core);
/* store the per-thread success-values of the setaffinity */
nth_affinity_set = (setaffinity_ret == 0);
if (debug)
{
fprintf(debug, "On rank %2d, thread %2d, core %2d the affinity setting returned %d\n",
cr->nodeid, gmx_omp_get_thread_num(), core, setaffinity_ret);
}
}
if (nth_affinity_set > nthread_local)
{
char msg[STRLEN];
sprintf(msg, "Looks like we have set affinity for more threads than "
"we have (%d > %d)!\n", nth_affinity_set, nthread_local);
gmx_incons(msg);
}
else
{
/* check & warn if some threads failed to set their affinities */
if (nth_affinity_set != nthread_local)
{
char sbuf1[STRLEN], sbuf2[STRLEN];
/* sbuf1 contains rank info, while sbuf2 OpenMP thread info */
sbuf1[0] = sbuf2[0] = '\0';
/* Only add rank info if we have more than one rank. */
if (cr->nnodes > 1)
{
#ifdef GMX_MPI
#ifdef GMX_THREAD_MPI
sprintf(sbuf1, "In tMPI thread #%d: ", cr->nodeid);
#else /* GMX_LIB_MPI */
sprintf(sbuf1, "In MPI process #%d: ", cr->nodeid);
#endif
#endif /* GMX_MPI */
}
if (nthread_local > 1)
{
sprintf(sbuf2, "for %d/%d thread%s ",
nthread_local - nth_affinity_set, nthread_local,
nthread_local > 1 ? "s" : "");
}
md_print_warn(NULL, fplog,
"WARNING: %sAffinity setting %sfailed.\n"
" This can cause performance degradation! If you think your setting are\n"
" correct, contact the GROMACS developers.",
sbuf1, sbuf2);
}
}
return;
}
/* Set CPU affinity. Can be important for performance.
On some systems (e.g. Cray) CPU Affinity is set by default.
But default assigning doesn't work (well) with only some ranks
having threads. This causes very low performance.
External tools have cumbersome syntax for setting affinity
in the case that only some ranks have threads.
Thus it is important that GROMACS sets the affinity internally
if only PME is using threads.
*/
void
gmx_set_thread_affinity(FILE *fplog,
const t_commrec *cr,
const gmx_hw_opt_t *hw_opt,
const gmx_hw_info_t *hwinfo)
{
int nth_affinity_set, thread0_id_node,
nthread_local, nthread_node;
int offset;
int * localityOrder = nullptr;
int rc;
if (hw_opt->thread_affinity == threadaffOFF)
{
/* Nothing to do */
return;
}
/* If the tMPI thread affinity setting is not supported encourage the user
* to report it as it's either a bug or an exotic platform which we might
* want to support. */
if (tMPI_Thread_setaffinity_support() != TMPI_SETAFFINITY_SUPPORT_YES)
{
/* we know Mac OS & BlueGene do not support setting thread affinity, so there's
no point in warning the user in that case. In any other case
the user might be able to do something about it. */
#if !defined(__APPLE__) && !defined(__bg__)
md_print_warn(cr, fplog,
"NOTE: Cannot set thread affinities on the current platform.\n");
#endif /* __APPLE__ */
return;
}
/* threads on this MPI process or TMPI thread */
if (cr->duty & DUTY_PP)
{
nthread_local = gmx_omp_nthreads_get(emntNonbonded);
}
else
{
nthread_local = gmx_omp_nthreads_get(emntPME);
}
/* map the current process to cores */
thread0_id_node = 0;
nthread_node = nthread_local;
#if GMX_MPI
if (PAR(cr) || MULTISIM(cr))
{
/* We need to determine a scan of the thread counts in this
* compute node.
*/
MPI_Comm comm_intra;
MPI_Comm_split(MPI_COMM_WORLD,
gmx_physicalnode_id_hash(), cr->rank_intranode,
&comm_intra);
MPI_Scan(&nthread_local, &thread0_id_node, 1, MPI_INT, MPI_SUM, comm_intra);
/* MPI_Scan is inclusive, but here we need exclusive */
thread0_id_node -= nthread_local;
/* Get the total number of threads on this physical node */
MPI_Allreduce(&nthread_local, &nthread_node, 1, MPI_INT, MPI_SUM, comm_intra);
MPI_Comm_free(&comm_intra);
}
#endif
if (hw_opt->thread_affinity == threadaffAUTO &&
nthread_node != hwinfo->nthreads_hw_avail)
{
if (nthread_node > 1 && nthread_node < hwinfo->nthreads_hw_avail)
{
md_print_warn(cr, fplog,
"NOTE: The number of threads is not equal to the number of (logical) cores\n"
" and the -pin option is set to auto: will not pin thread to cores.\n"
" This can lead to significant performance degradation.\n"
" Consider using -pin on (and -pinoffset in case you run multiple jobs).\n");
}
return;
}
offset = 0;
if (hw_opt->core_pinning_offset != 0)
{
offset = hw_opt->core_pinning_offset;
md_print_info(cr, fplog, "Applying core pinning offset %d\n", offset);
}
int core_pinning_stride = hw_opt->core_pinning_stride;
rc = get_thread_affinity_layout(fplog, cr, hwinfo,
nthread_node,
offset, &core_pinning_stride,
&localityOrder);
gmx::scoped_guard_sfree localityOrderGuard(localityOrder);
if (rc != 0)
{
/* Incompatible layout, don't pin, warning was already issued */
return;
}
/* Set the per-thread affinity. In order to be able to check the success
* of affinity settings, we will set nth_affinity_set to 1 on threads
* where the affinity setting succeded and to 0 where it failed.
* Reducing these 0/1 values over the threads will give the total number
* of threads on which we succeeded.
*/
// To avoid warnings from the static analyzer we initialize nth_affinity_set
// to zero outside the OpenMP block, and then add to it inside the block.
// The value will still always be 0 or 1 from each thread.
nth_affinity_set = 0;
#pragma omp parallel num_threads(nthread_local) reduction(+:nth_affinity_set)
{
try
{
int thread_id, thread_id_node;
int index, core;
gmx_bool setaffinity_ret;
thread_id = gmx_omp_get_thread_num();
thread_id_node = thread0_id_node + thread_id;
index = offset + thread_id_node*core_pinning_stride;
if (localityOrder != nullptr)
{
core = localityOrder[index];
}
else
{
core = index;
}
setaffinity_ret = tMPI_Thread_setaffinity_single(tMPI_Thread_self(), core);
/* store the per-thread success-values of the setaffinity */
nth_affinity_set += (setaffinity_ret == 0);
if (debug)
{
fprintf(debug, "On rank %2d, thread %2d, index %2d, core %2d the affinity setting returned %d\n",
cr->nodeid, gmx_omp_get_thread_num(), index, core, setaffinity_ret);
}
}
GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
}
if (nth_affinity_set > nthread_local)
{
char msg[STRLEN];
sprintf(msg, "Looks like we have set affinity for more threads than "
"we have (%d > %d)!\n", nth_affinity_set, nthread_local);
gmx_incons(msg);
}
else
{
/* check & warn if some threads failed to set their affinities */
const bool allAffinitiesSet = (nth_affinity_set == nthread_local);
if (!allAffinitiesSet)
{
char sbuf1[STRLEN], sbuf2[STRLEN];
/* sbuf1 contains rank info, while sbuf2 OpenMP thread info */
sbuf1[0] = sbuf2[0] = '\0';
/* Only add rank info if we have more than one rank. */
if (cr->nnodes > 1)
{
#if GMX_MPI
#if GMX_THREAD_MPI
sprintf(sbuf1, "In tMPI thread #%d: ", cr->nodeid);
#else /* GMX_LIB_MPI */
sprintf(sbuf1, "In MPI process #%d: ", cr->nodeid);
#endif
#endif /* GMX_MPI */
}
if (nthread_local > 1)
{
sprintf(sbuf2, "for %d/%d thread%s ",
nthread_local - nth_affinity_set, nthread_local,
nthread_local > 1 ? "s" : "");
}
fprintf(stderr, "NOTE: %sAffinity setting %sfailed.\n", sbuf1, sbuf2);
}
if (invalidWithinSimulation(cr, !allAffinitiesSet))
{
md_print_warn(cr, fplog,
"NOTE: Thread affinity setting failed. This can cause performance degradation.\n"
" If you think your settings are correct, ask on the gmx-users list.\n");
}
}
}
void gmx_omp_nthreads_init(FILE *fplog, t_commrec *cr,
int nthreads_hw_avail,
int omp_nthreads_req,
int omp_nthreads_pme_req,
gmx_bool gmx_unused bThisNodePMEOnly,
gmx_bool bFullOmpSupport)
{
int nth, nth_pmeonly, gmx_maxth, nppn;
char *env;
gmx_bool bSepPME, bOMP;
#ifdef GMX_OPENMP
bOMP = TRUE;
#else
bOMP = FALSE;
#endif /* GMX_OPENMP */
/* number of MPI processes/threads per physical node */
nppn = cr->nrank_intranode;
bSepPME = ( (cr->duty & DUTY_PP) && !(cr->duty & DUTY_PME)) ||
(!(cr->duty & DUTY_PP) && (cr->duty & DUTY_PME));
#ifdef GMX_THREAD_MPI
/* modth is shared among tMPI threads, so for thread safety do the
* detection is done on the master only. It is not thread-safe with
* multiple simulations, but that's anyway not supported by tMPI. */
if (SIMMASTER(cr))
#endif
{
/* just return if the initialization has already been done */
if (modth.initialized)
{
return;
}
/* With full OpenMP support (verlet scheme) set the number of threads
* per process / default:
* - 1 if not compiled with OpenMP or
* - OMP_NUM_THREADS if the env. var is set, or
* - omp_nthreads_req = #of threads requested by the user on the mdrun
* command line, otherwise
* - take the max number of available threads and distribute them
* on the processes/tMPI threads.
* ~ The GMX_*_NUM_THREADS env var overrides the number of threads of
* the respective module and it has to be used in conjunction with
* OMP_NUM_THREADS.
*
* With the group scheme OpenMP multithreading is only supported in PME,
* for all other modules nthreads is set to 1.
* The number of PME threads is equal to:
* - 1 if not compiled with OpenMP or
* - GMX_PME_NUM_THREADS if defined, otherwise
* - OMP_NUM_THREADS if defined, otherwise
* - 1
*/
nth = 1;
if ((env = getenv("OMP_NUM_THREADS")) != NULL)
{
if (!bOMP && (strncmp(env, "1", 1) != 0))
{
gmx_warning("OMP_NUM_THREADS is set, but %s was compiled without OpenMP support!",
ShortProgram());
}
else
{
nth = gmx_omp_get_max_threads();
}
}
else if (omp_nthreads_req > 0)
{
nth = omp_nthreads_req;
}
else if (bFullOmpSupport && bOMP)
{
/* max available threads per node */
nth = nthreads_hw_avail;
/* divide the threads among the MPI processes/tMPI threads */
if (nth >= nppn)
{
nth /= nppn;
}
else
{
nth = 1;
}
}
/* now we have the global values, set them:
* - 1 if not compiled with OpenMP and for the group scheme
* - nth for the verlet scheme when compiled with OpenMP
*/
if (bFullOmpSupport && bOMP)
{
modth.gnth = nth;
}
else
{
modth.gnth = 1;
}
if (bSepPME)
{
if (omp_nthreads_pme_req > 0)
{
modth.gnth_pme = omp_nthreads_pme_req;
}
else
{
modth.gnth_pme = nth;
}
}
else
{
modth.gnth_pme = 0;
}
/* now set the per-module values */
modth.nth[emntDefault] = modth.gnth;
pick_module_nthreads(fplog, emntDomdec, SIMMASTER(cr), bFullOmpSupport, bSepPME);
pick_module_nthreads(fplog, emntPairsearch, SIMMASTER(cr), bFullOmpSupport, bSepPME);
pick_module_nthreads(fplog, emntNonbonded, SIMMASTER(cr), bFullOmpSupport, bSepPME);
pick_module_nthreads(fplog, emntBonded, SIMMASTER(cr), bFullOmpSupport, bSepPME);
pick_module_nthreads(fplog, emntPME, SIMMASTER(cr), bFullOmpSupport, bSepPME);
pick_module_nthreads(fplog, emntUpdate, SIMMASTER(cr), bFullOmpSupport, bSepPME);
pick_module_nthreads(fplog, emntVSITE, SIMMASTER(cr), bFullOmpSupport, bSepPME);
pick_module_nthreads(fplog, emntLINCS, SIMMASTER(cr), bFullOmpSupport, bSepPME);
pick_module_nthreads(fplog, emntSETTLE, SIMMASTER(cr), bFullOmpSupport, bSepPME);
/* set the number of threads globally */
if (bOMP)
{
#ifndef GMX_THREAD_MPI
if (bThisNodePMEOnly)
{
gmx_omp_set_num_threads(modth.gnth_pme);
}
else
#endif /* GMX_THREAD_MPI */
{
if (bFullOmpSupport)
{
gmx_omp_set_num_threads(nth);
}
else
{
gmx_omp_set_num_threads(1);
}
}
}
modth.initialized = TRUE;
}
#ifdef GMX_THREAD_MPI
/* Non-master threads have to wait for the detection to be done. */
if (PAR(cr))
{
MPI_Barrier(cr->mpi_comm_mysim);
}
#endif
/* inform the user about the settings */
if (bOMP)
{
#ifdef GMX_THREAD_MPI
const char *mpi_str = "per tMPI thread";
#else
const char *mpi_str = "per MPI process";
#endif
/* for group scheme we print PME threads info only */
if (bFullOmpSupport)
{
md_print_info(cr, fplog, "Using %d OpenMP thread%s %s\n",
modth.gnth, modth.gnth > 1 ? "s" : "",
cr->nnodes > 1 ? mpi_str : "");
}
if (bSepPME && modth.gnth_pme != modth.gnth)
{
md_print_info(cr, fplog, "Using %d OpenMP thread%s %s for PME\n",
modth.gnth_pme, modth.gnth_pme > 1 ? "s" : "",
cr->nnodes > 1 ? mpi_str : "");
}
}
/* detect and warn about oversubscription
* TODO: enable this for separate PME nodes as well! */
if (!bSepPME && cr->rank_pp_intranode == 0)
{
char sbuf[STRLEN], sbuf1[STRLEN], sbuf2[STRLEN];
if (modth.gnth*nppn > nthreads_hw_avail)
{
sprintf(sbuf, "threads");
sbuf1[0] = '\0';
sprintf(sbuf2, "O");
#ifdef GMX_MPI
if (modth.gnth == 1)
{
#ifdef GMX_THREAD_MPI
sprintf(sbuf, "thread-MPI threads");
#else
sprintf(sbuf, "MPI processes");
sprintf(sbuf1, " per node");
sprintf(sbuf2, "On node %d: o", cr->sim_nodeid);
#endif
}
#endif
md_print_warn(cr, fplog,
"WARNING: %sversubscribing the available %d logical CPU cores%s with %d %s.\n"
" This will cause considerable performance loss!",
sbuf2, nthreads_hw_avail, sbuf1, nppn*modth.gnth, sbuf);
}
}
}
int mdrunner(gmx_hw_opt_t *hw_opt,
FILE *fplog, t_commrec *cr, int nfile,
const t_filenm fnm[], const output_env_t oenv, gmx_bool bVerbose,
gmx_bool bCompact, int nstglobalcomm,
ivec ddxyz, int dd_node_order, real rdd, real rconstr,
const char *dddlb_opt, real dlb_scale,
const char *ddcsx, const char *ddcsy, const char *ddcsz,
const char *nbpu_opt, int nstlist_cmdline,
gmx_int64_t nsteps_cmdline, int nstepout, int resetstep,
int gmx_unused nmultisim, int repl_ex_nst, int repl_ex_nex,
int repl_ex_seed, real pforce, real cpt_period, real max_hours,
int imdport, unsigned long Flags)
{
gmx_bool bForceUseGPU, bTryUseGPU, bRerunMD;
t_inputrec *inputrec;
t_state *state = NULL;
matrix box;
gmx_ddbox_t ddbox = {0};
int npme_major, npme_minor;
t_nrnb *nrnb;
gmx_mtop_t *mtop = NULL;
t_mdatoms *mdatoms = NULL;
t_forcerec *fr = NULL;
t_fcdata *fcd = NULL;
real ewaldcoeff_q = 0;
real ewaldcoeff_lj = 0;
struct gmx_pme_t **pmedata = NULL;
gmx_vsite_t *vsite = NULL;
gmx_constr_t constr;
int nChargePerturbed = -1, nTypePerturbed = 0, status;
gmx_wallcycle_t wcycle;
gmx_bool bReadEkin;
gmx_walltime_accounting_t walltime_accounting = NULL;
int rc;
gmx_int64_t reset_counters;
gmx_edsam_t ed = NULL;
int nthreads_pme = 1;
int nthreads_pp = 1;
gmx_membed_t membed = NULL;
gmx_hw_info_t *hwinfo = NULL;
/* The master rank decides early on bUseGPU and broadcasts this later */
gmx_bool bUseGPU = FALSE;
/* CAUTION: threads may be started later on in this function, so
cr doesn't reflect the final parallel state right now */
snew(inputrec, 1);
snew(mtop, 1);
if (Flags & MD_APPENDFILES)
{
fplog = NULL;
}
bRerunMD = (Flags & MD_RERUN);
bForceUseGPU = (strncmp(nbpu_opt, "gpu", 3) == 0);
bTryUseGPU = (strncmp(nbpu_opt, "auto", 4) == 0) || bForceUseGPU;
/* Detect hardware, gather information. This is an operation that is
* global for this process (MPI rank). */
hwinfo = gmx_detect_hardware(fplog, cr, bTryUseGPU);
gmx_print_detected_hardware(fplog, cr, hwinfo);
if (fplog != NULL)
{
/* Print references after all software/hardware printing */
please_cite(fplog, "Abraham2015");
please_cite(fplog, "Pall2015");
please_cite(fplog, "Pronk2013");
please_cite(fplog, "Hess2008b");
please_cite(fplog, "Spoel2005a");
please_cite(fplog, "Lindahl2001a");
please_cite(fplog, "Berendsen95a");
}
snew(state, 1);
if (SIMMASTER(cr))
{
/* Read (nearly) all data required for the simulation */
read_tpx_state(ftp2fn(efTPR, nfile, fnm), inputrec, state, NULL, mtop);
if (inputrec->cutoff_scheme == ecutsVERLET)
{
/* Here the master rank decides if all ranks will use GPUs */
bUseGPU = (hwinfo->gpu_info.n_dev_compatible > 0 ||
getenv("GMX_EMULATE_GPU") != NULL);
/* TODO add GPU kernels for this and replace this check by:
* (bUseGPU && (ir->vdwtype == evdwPME &&
* ir->ljpme_combination_rule == eljpmeLB))
* update the message text and the content of nbnxn_acceleration_supported.
*/
if (bUseGPU &&
!nbnxn_gpu_acceleration_supported(fplog, cr, inputrec, bRerunMD))
{
/* Fallback message printed by nbnxn_acceleration_supported */
if (bForceUseGPU)
{
gmx_fatal(FARGS, "GPU acceleration requested, but not supported with the given input settings");
}
bUseGPU = FALSE;
}
prepare_verlet_scheme(fplog, cr,
inputrec, nstlist_cmdline, mtop, state->box,
bUseGPU);
}
else
{
if (nstlist_cmdline > 0)
{
gmx_fatal(FARGS, "Can not set nstlist with the group cut-off scheme");
}
if (hwinfo->gpu_info.n_dev_compatible > 0)
{
md_print_warn(cr, fplog,
"NOTE: GPU(s) found, but the current simulation can not use GPUs\n"
" To use a GPU, set the mdp option: cutoff-scheme = Verlet\n");
}
if (bForceUseGPU)
{
gmx_fatal(FARGS, "GPU requested, but can't be used without cutoff-scheme=Verlet");
}
#ifdef GMX_TARGET_BGQ
md_print_warn(cr, fplog,
"NOTE: There is no SIMD implementation of the group scheme kernels on\n"
" BlueGene/Q. You will observe better performance from using the\n"
" Verlet cut-off scheme.\n");
#endif
}
if (inputrec->eI == eiSD2)
{
md_print_warn(cr, fplog, "The stochastic dynamics integrator %s is deprecated, since\n"
"it is slower than integrator %s and is slightly less accurate\n"
"with constraints. Use the %s integrator.",
ei_names[inputrec->eI], ei_names[eiSD1], ei_names[eiSD1]);
}
}
/* Check and update the hardware options for internal consistency */
check_and_update_hw_opt_1(hw_opt, cr);
/* Early check for externally set process affinity. */
gmx_check_thread_affinity_set(fplog, cr,
hw_opt, hwinfo->nthreads_hw_avail, FALSE);
#ifdef GMX_THREAD_MPI
if (SIMMASTER(cr))
{
if (cr->npmenodes > 0 && hw_opt->nthreads_tmpi <= 0)
{
gmx_fatal(FARGS, "You need to explicitly specify the number of MPI threads (-ntmpi) when using separate PME ranks");
}
/* Since the master knows the cut-off scheme, update hw_opt for this.
* This is done later for normal MPI and also once more with tMPI
* for all tMPI ranks.
*/
check_and_update_hw_opt_2(hw_opt, inputrec->cutoff_scheme);
/* NOW the threads will be started: */
hw_opt->nthreads_tmpi = get_nthreads_mpi(hwinfo,
hw_opt,
inputrec, mtop,
cr, fplog, bUseGPU);
if (hw_opt->nthreads_tmpi > 1)
{
t_commrec *cr_old = cr;
/* now start the threads. */
cr = mdrunner_start_threads(hw_opt, fplog, cr_old, nfile, fnm,
oenv, bVerbose, bCompact, nstglobalcomm,
ddxyz, dd_node_order, rdd, rconstr,
dddlb_opt, dlb_scale, ddcsx, ddcsy, ddcsz,
nbpu_opt, nstlist_cmdline,
nsteps_cmdline, nstepout, resetstep, nmultisim,
repl_ex_nst, repl_ex_nex, repl_ex_seed, pforce,
cpt_period, max_hours,
Flags);
/* the main thread continues here with a new cr. We don't deallocate
the old cr because other threads may still be reading it. */
if (cr == NULL)
{
gmx_comm("Failed to spawn threads");
}
}
}
#endif
/* END OF CAUTION: cr is now reliable */
/* g_membed initialisation *
* Because we change the mtop, init_membed is called before the init_parallel *
* (in case we ever want to make it run in parallel) */
if (opt2bSet("-membed", nfile, fnm))
{
if (MASTER(cr))
{
fprintf(stderr, "Initializing membed");
}
membed = init_membed(fplog, nfile, fnm, mtop, inputrec, state, cr, &cpt_period);
}
if (PAR(cr))
{
/* now broadcast everything to the non-master nodes/threads: */
init_parallel(cr, inputrec, mtop);
/* The master rank decided on the use of GPUs,
* broadcast this information to all ranks.
*/
gmx_bcast_sim(sizeof(bUseGPU), &bUseGPU, cr);
}
if (fplog != NULL)
{
pr_inputrec(fplog, 0, "Input Parameters", inputrec, FALSE);
fprintf(fplog, "\n");
}
/* now make sure the state is initialized and propagated */
set_state_entries(state, inputrec);
/* A parallel command line option consistency check that we can
only do after any threads have started. */
if (!PAR(cr) &&
(ddxyz[XX] > 1 || ddxyz[YY] > 1 || ddxyz[ZZ] > 1 || cr->npmenodes > 0))
{
gmx_fatal(FARGS,
"The -dd or -npme option request a parallel simulation, "
#ifndef GMX_MPI
"but %s was compiled without threads or MPI enabled"
#else
#ifdef GMX_THREAD_MPI
"but the number of threads (option -nt) is 1"
#else
"but %s was not started through mpirun/mpiexec or only one rank was requested through mpirun/mpiexec"
#endif
#endif
, output_env_get_program_display_name(oenv)
);
}
if (bRerunMD &&
(EI_ENERGY_MINIMIZATION(inputrec->eI) || eiNM == inputrec->eI))
{
gmx_fatal(FARGS, "The .mdp file specified an energy mininization or normal mode algorithm, and these are not compatible with mdrun -rerun");
}
if (can_use_allvsall(inputrec, TRUE, cr, fplog) && DOMAINDECOMP(cr))
{
gmx_fatal(FARGS, "All-vs-all loops do not work with domain decomposition, use a single MPI rank");
}
if (!(EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype)))
{
if (cr->npmenodes > 0)
{
gmx_fatal_collective(FARGS, cr, NULL,
"PME-only ranks are requested, but the system does not use PME for electrostatics or LJ");
}
cr->npmenodes = 0;
}
if (bUseGPU && cr->npmenodes < 0)
{
/* With GPUs we don't automatically use PME-only ranks. PME ranks can
* improve performance with many threads per GPU, since our OpenMP
* scaling is bad, but it's difficult to automate the setup.
*/
cr->npmenodes = 0;
}
#ifdef GMX_FAHCORE
if (MASTER(cr))
{
fcRegisterSteps(inputrec->nsteps, inputrec->init_step);
}
#endif
/* NMR restraints must be initialized before load_checkpoint,
* since with time averaging the history is added to t_state.
* For proper consistency check we therefore need to extend
* t_state here.
* So the PME-only nodes (if present) will also initialize
* the distance restraints.
*/
snew(fcd, 1);
/* This needs to be called before read_checkpoint to extend the state */
init_disres(fplog, mtop, inputrec, cr, fcd, state, repl_ex_nst > 0);
init_orires(fplog, mtop, state->x, inputrec, cr, &(fcd->orires),
state);
if (DEFORM(*inputrec))
{
/* Store the deform reference box before reading the checkpoint */
if (SIMMASTER(cr))
{
copy_mat(state->box, box);
}
if (PAR(cr))
{
gmx_bcast(sizeof(box), box, cr);
}
/* Because we do not have the update struct available yet
* in which the reference values should be stored,
* we store them temporarily in static variables.
* This should be thread safe, since they are only written once
* and with identical values.
*/
tMPI_Thread_mutex_lock(&deform_init_box_mutex);
deform_init_init_step_tpx = inputrec->init_step;
copy_mat(box, deform_init_box_tpx);
tMPI_Thread_mutex_unlock(&deform_init_box_mutex);
}
if (opt2bSet("-cpi", nfile, fnm))
{
/* Check if checkpoint file exists before doing continuation.
* This way we can use identical input options for the first and subsequent runs...
*/
if (gmx_fexist_master(opt2fn_master("-cpi", nfile, fnm, cr), cr) )
{
load_checkpoint(opt2fn_master("-cpi", nfile, fnm, cr), &fplog,
cr, ddxyz,
inputrec, state, &bReadEkin,
(Flags & MD_APPENDFILES),
(Flags & MD_APPENDFILESSET));
if (bReadEkin)
{
Flags |= MD_READ_EKIN;
}
}
}
if (MASTER(cr) && (Flags & MD_APPENDFILES))
{
gmx_log_open(ftp2fn(efLOG, nfile, fnm), cr,
Flags, &fplog);
}
/* override nsteps with value from cmdline */
override_nsteps_cmdline(fplog, nsteps_cmdline, inputrec, cr);
if (SIMMASTER(cr))
{
copy_mat(state->box, box);
}
if (PAR(cr))
{
gmx_bcast(sizeof(box), box, cr);
}
/* Essential dynamics */
if (opt2bSet("-ei", nfile, fnm))
{
/* Open input and output files, allocate space for ED data structure */
ed = ed_open(mtop->natoms, &state->edsamstate, nfile, fnm, Flags, oenv, cr);
}
if (PAR(cr) && !(EI_TPI(inputrec->eI) ||
inputrec->eI == eiNM))
{
cr->dd = init_domain_decomposition(fplog, cr, Flags, ddxyz, rdd, rconstr,
dddlb_opt, dlb_scale,
ddcsx, ddcsy, ddcsz,
mtop, inputrec,
box, state->x,
&ddbox, &npme_major, &npme_minor);
make_dd_communicators(fplog, cr, dd_node_order);
/* Set overallocation to avoid frequent reallocation of arrays */
set_over_alloc_dd(TRUE);
}
else
{
/* PME, if used, is done on all nodes with 1D decomposition */
cr->npmenodes = 0;
cr->duty = (DUTY_PP | DUTY_PME);
npme_major = 1;
npme_minor = 1;
if (inputrec->ePBC == epbcSCREW)
{
gmx_fatal(FARGS,
"pbc=%s is only implemented with domain decomposition",
epbc_names[inputrec->ePBC]);
}
}
if (PAR(cr))
{
/* After possible communicator splitting in make_dd_communicators.
* we can set up the intra/inter node communication.
*/
gmx_setup_nodecomm(fplog, cr);
}
/* Initialize per-physical-node MPI process/thread ID and counters. */
gmx_init_intranode_counters(cr);
#ifdef GMX_MPI
if (MULTISIM(cr))
{
md_print_info(cr, fplog,
"This is simulation %d out of %d running as a composite GROMACS\n"
"multi-simulation job. Setup for this simulation:\n\n",
cr->ms->sim, cr->ms->nsim);
}
md_print_info(cr, fplog, "Using %d MPI %s\n",
cr->nnodes,
#ifdef GMX_THREAD_MPI
cr->nnodes == 1 ? "thread" : "threads"
#else
cr->nnodes == 1 ? "process" : "processes"
#endif
);
fflush(stderr);
#endif
/* Check and update hw_opt for the cut-off scheme */
check_and_update_hw_opt_2(hw_opt, inputrec->cutoff_scheme);
/* Check and update hw_opt for the number of MPI ranks */
check_and_update_hw_opt_3(hw_opt);
gmx_omp_nthreads_init(fplog, cr,
hwinfo->nthreads_hw_avail,
hw_opt->nthreads_omp,
hw_opt->nthreads_omp_pme,
(cr->duty & DUTY_PP) == 0,
inputrec->cutoff_scheme == ecutsVERLET);
#ifndef NDEBUG
if (integrator[inputrec->eI].func != do_tpi &&
inputrec->cutoff_scheme == ecutsVERLET)
{
gmx_feenableexcept();
}
#endif
if (bUseGPU)
{
/* Select GPU id's to use */
gmx_select_gpu_ids(fplog, cr, &hwinfo->gpu_info, bForceUseGPU,
&hw_opt->gpu_opt);
}
else
{
/* Ignore (potentially) manually selected GPUs */
hw_opt->gpu_opt.n_dev_use = 0;
}
/* check consistency across ranks of things like SIMD
* support and number of GPUs selected */
gmx_check_hw_runconf_consistency(fplog, hwinfo, cr, hw_opt, bUseGPU);
/* Now that we know the setup is consistent, check for efficiency */
check_resource_division_efficiency(hwinfo, hw_opt, Flags & MD_NTOMPSET,
cr, fplog);
if (DOMAINDECOMP(cr))
{
/* When we share GPUs over ranks, we need to know this for the DLB */
dd_setup_dlb_resource_sharing(cr, hwinfo, hw_opt);
}
/* getting number of PP/PME threads
PME: env variable should be read only on one node to make sure it is
identical everywhere;
*/
/* TODO nthreads_pp is only used for pinning threads.
* This is a temporary solution until we have a hw topology library.
*/
nthreads_pp = gmx_omp_nthreads_get(emntNonbonded);
nthreads_pme = gmx_omp_nthreads_get(emntPME);
wcycle = wallcycle_init(fplog, resetstep, cr, nthreads_pp, nthreads_pme);
if (PAR(cr))
{
/* Master synchronizes its value of reset_counters with all nodes
* including PME only nodes */
reset_counters = wcycle_get_reset_counters(wcycle);
gmx_bcast_sim(sizeof(reset_counters), &reset_counters, cr);
wcycle_set_reset_counters(wcycle, reset_counters);
}
snew(nrnb, 1);
if (cr->duty & DUTY_PP)
{
bcast_state(cr, state);
/* Initiate forcerecord */
fr = mk_forcerec();
fr->hwinfo = hwinfo;
fr->gpu_opt = &hw_opt->gpu_opt;
init_forcerec(fplog, oenv, fr, fcd, inputrec, mtop, cr, box,
opt2fn("-table", nfile, fnm),
opt2fn("-tabletf", nfile, fnm),
opt2fn("-tablep", nfile, fnm),
opt2fn("-tableb", nfile, fnm),
nbpu_opt,
FALSE,
pforce);
/* version for PCA_NOT_READ_NODE (see md.c) */
/*init_forcerec(fplog,fr,fcd,inputrec,mtop,cr,box,FALSE,
"nofile","nofile","nofile","nofile",FALSE,pforce);
*/
/* Initialize QM-MM */
if (fr->bQMMM)
{
init_QMMMrec(cr, mtop, inputrec, fr);
}
/* Initialize the mdatoms structure.
* mdatoms is not filled with atom data,
* as this can not be done now with domain decomposition.
*/
mdatoms = init_mdatoms(fplog, mtop, inputrec->efep != efepNO);
/* Initialize the virtual site communication */
vsite = init_vsite(mtop, cr, FALSE);
calc_shifts(box, fr->shift_vec);
/* With periodic molecules the charge groups should be whole at start up
* and the virtual sites should not be far from their proper positions.
*/
if (!inputrec->bContinuation && MASTER(cr) &&
!(inputrec->ePBC != epbcNONE && inputrec->bPeriodicMols))
{
/* Make molecules whole at start of run */
if (fr->ePBC != epbcNONE)
{
do_pbc_first_mtop(fplog, inputrec->ePBC, box, mtop, state->x);
}
if (vsite)
{
/* Correct initial vsite positions are required
* for the initial distribution in the domain decomposition
* and for the initial shell prediction.
*/
construct_vsites_mtop(vsite, mtop, state->x);
}
}
if (EEL_PME(fr->eeltype) || EVDW_PME(fr->vdwtype))
{
ewaldcoeff_q = fr->ewaldcoeff_q;
ewaldcoeff_lj = fr->ewaldcoeff_lj;
pmedata = &fr->pmedata;
}
else
{
pmedata = NULL;
}
}
else
{
/* This is a PME only node */
/* We don't need the state */
done_state(state);
ewaldcoeff_q = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
snew(pmedata, 1);
}
if (hw_opt->thread_affinity != threadaffOFF)
{
/* Before setting affinity, check whether the affinity has changed
* - which indicates that probably the OpenMP library has changed it
* since we first checked).
*/
gmx_check_thread_affinity_set(fplog, cr,
hw_opt, hwinfo->nthreads_hw_avail, TRUE);
/* Set the CPU affinity */
gmx_set_thread_affinity(fplog, cr, hw_opt, hwinfo);
}
/* Initiate PME if necessary,
* either on all nodes or on dedicated PME nodes only. */
if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
{
if (mdatoms)
{
nChargePerturbed = mdatoms->nChargePerturbed;
if (EVDW_PME(inputrec->vdwtype))
{
nTypePerturbed = mdatoms->nTypePerturbed;
}
}
if (cr->npmenodes > 0)
{
/* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
gmx_bcast_sim(sizeof(nChargePerturbed), &nChargePerturbed, cr);
gmx_bcast_sim(sizeof(nTypePerturbed), &nTypePerturbed, cr);
}
if (cr->duty & DUTY_PME)
{
status = gmx_pme_init(pmedata, cr, npme_major, npme_minor, inputrec,
mtop ? mtop->natoms : 0, nChargePerturbed, nTypePerturbed,
(Flags & MD_REPRODUCIBLE), nthreads_pme);
if (status != 0)
{
gmx_fatal(FARGS, "Error %d initializing PME", status);
}
}
}
if (integrator[inputrec->eI].func == do_md)
{
/* Turn on signal handling on all nodes */
/*
* (A user signal from the PME nodes (if any)
* is communicated to the PP nodes.
*/
signal_handler_install();
}
if (cr->duty & DUTY_PP)
{
/* Assumes uniform use of the number of OpenMP threads */
walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntDefault));
if (inputrec->bPull)
{
/* Initialize pull code */
inputrec->pull_work =
init_pull(fplog, inputrec->pull, inputrec, nfile, fnm,
mtop, cr, oenv, inputrec->fepvals->init_lambda,
EI_DYNAMICS(inputrec->eI) && MASTER(cr), Flags);
}
if (inputrec->bRot)
{
/* Initialize enforced rotation code */
init_rot(fplog, inputrec, nfile, fnm, cr, state->x, box, mtop, oenv,
bVerbose, Flags);
}
if (inputrec->eSwapCoords != eswapNO)
{
/* Initialize ion swapping code */
init_swapcoords(fplog, bVerbose, inputrec, opt2fn_master("-swap", nfile, fnm, cr),
mtop, state->x, state->box, &state->swapstate, cr, oenv, Flags);
}
constr = init_constraints(fplog, mtop, inputrec, ed, state, cr);
if (DOMAINDECOMP(cr))
{
GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
dd_init_bondeds(fplog, cr->dd, mtop, vsite, inputrec,
Flags & MD_DDBONDCHECK, fr->cginfo_mb);
set_dd_parameters(fplog, cr->dd, dlb_scale, inputrec, &ddbox);
setup_dd_grid(fplog, cr->dd);
}
/* Now do whatever the user wants us to do (how flexible...) */
integrator[inputrec->eI].func(fplog, cr, nfile, fnm,
oenv, bVerbose, bCompact,
nstglobalcomm,
vsite, constr,
nstepout, inputrec, mtop,
fcd, state,
mdatoms, nrnb, wcycle, ed, fr,
repl_ex_nst, repl_ex_nex, repl_ex_seed,
membed,
cpt_period, max_hours,
imdport,
Flags,
walltime_accounting);
if (inputrec->bPull)
{
finish_pull(inputrec->pull_work);
}
if (inputrec->bRot)
{
finish_rot(inputrec->rot);
}
}
else
{
GMX_RELEASE_ASSERT(pmedata, "pmedata was NULL while cr->duty was not DUTY_PP");
/* do PME only */
walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntPME));
gmx_pmeonly(*pmedata, cr, nrnb, wcycle, walltime_accounting, ewaldcoeff_q, ewaldcoeff_lj, inputrec);
}
wallcycle_stop(wcycle, ewcRUN);
/* Finish up, write some stuff
* if rerunMD, don't write last frame again
*/
finish_run(fplog, cr,
inputrec, nrnb, wcycle, walltime_accounting,
fr ? fr->nbv : NULL,
EI_DYNAMICS(inputrec->eI) && !MULTISIM(cr));
/* Free GPU memory and context */
free_gpu_resources(fr, cr, &hwinfo->gpu_info, fr ? fr->gpu_opt : NULL);
if (opt2bSet("-membed", nfile, fnm))
{
sfree(membed);
}
gmx_hardware_info_free(hwinfo);
/* Does what it says */
print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
walltime_accounting_destroy(walltime_accounting);
/* PLUMED */
if(plumedswitch){
plumed_finalize(plumedmain);
}
/* END PLUMED */
/* Close logfile already here if we were appending to it */
if (MASTER(cr) && (Flags & MD_APPENDFILES))
{
gmx_log_close(fplog);
}
rc = (int)gmx_get_stop_condition();
done_ed(&ed);
#ifdef GMX_THREAD_MPI
/* we need to join all threads. The sub-threads join when they
exit this function, but the master thread needs to be told to
wait for that. */
if (PAR(cr) && MASTER(cr))
{
tMPI_Finalize();
}
#endif
return rc;
}
