* To help us fund GROMACS development, we humbly ask that you cite
* the research papers on the package. Check out http://www.gromacs.org.
*/
-#ifdef HAVE_CONFIG_H
-#include <config.h>
-#endif
+#include "gmxpre.h"
+
+#include "config.h"
+#include <assert.h>
#include <math.h>
#include <string.h>
-#include <assert.h>
-#include "sysstuff.h"
-#include "typedefs.h"
-#include "types/commrec.h"
-#include "vec.h"
+
+#include "gromacs/legacyheaders/copyrite.h"
+#include "gromacs/legacyheaders/coulomb.h"
+#include "gromacs/legacyheaders/domdec.h"
+#include "gromacs/legacyheaders/force.h"
+#include "gromacs/legacyheaders/gmx_detect_hardware.h"
+#include "gromacs/legacyheaders/gmx_omp_nthreads.h"
+#include "gromacs/legacyheaders/gpu_utils.h"
+#include "gromacs/legacyheaders/inputrec.h"
+#include "gromacs/legacyheaders/macros.h"
+#include "gromacs/legacyheaders/md_logging.h"
+#include "gromacs/legacyheaders/md_support.h"
+#include "gromacs/legacyheaders/names.h"
+#include "gromacs/legacyheaders/network.h"
+#include "gromacs/legacyheaders/nonbonded.h"
+#include "gromacs/legacyheaders/ns.h"
+#include "gromacs/legacyheaders/pmalloc_cuda.h"
+#include "gromacs/legacyheaders/qmmm.h"
+#include "gromacs/legacyheaders/tables.h"
+#include "gromacs/legacyheaders/txtdump.h"
+#include "gromacs/legacyheaders/typedefs.h"
+#include "gromacs/legacyheaders/types/commrec.h"
+#include "gromacs/legacyheaders/types/nbnxn_cuda_types_ext.h"
+#include "gromacs/math/units.h"
#include "gromacs/math/utilities.h"
-#include "macros.h"
+#include "gromacs/math/vec.h"
+#include "gromacs/mdlib/nb_verlet.h"
+#include "gromacs/mdlib/nbnxn_atomdata.h"
+#include "gromacs/mdlib/nbnxn_consts.h"
+#include "gromacs/mdlib/nbnxn_search.h"
+#include "gromacs/mdlib/nbnxn_simd.h"
+#include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_data_mgmt.h"
+#include "gromacs/pbcutil/ishift.h"
+#include "gromacs/pbcutil/pbc.h"
+#include "gromacs/topology/mtop_util.h"
+#include "gromacs/utility/fatalerror.h"
#include "gromacs/utility/smalloc.h"
-#include "macros.h"
-#include "gmx_fatal.h"
-#include "physics.h"
-#include "force.h"
-#include "tables.h"
-#include "nonbonded.h"
-#include "invblock.h"
-#include "names.h"
-#include "network.h"
-#include "pbc.h"
-#include "ns.h"
-#include "mshift.h"
-#include "txtdump.h"
-#include "coulomb.h"
-#include "md_support.h"
-#include "md_logging.h"
-#include "domdec.h"
-#include "qmmm.h"
-#include "copyrite.h"
-#include "mtop_util.h"
-#include "nbnxn_simd.h"
-#include "nbnxn_search.h"
-#include "nbnxn_atomdata.h"
-#include "nbnxn_consts.h"
-#include "gmx_omp_nthreads.h"
-#include "gmx_detect_hardware.h"
-#include "inputrec.h"
-
-#include "types/nbnxn_cuda_types_ext.h"
-#include "gpu_utils.h"
-#include "nbnxn_cuda_data_mgmt.h"
-#include "pmalloc_cuda.h"
t_forcerec *mk_forcerec(void)
{
sigmaj = pow(c12j / c6j, 1.0/6.0);
epsi = c6i * c6i / c12i;
epsj = c6j * c6j / c12j;
- c6 = epsi * epsj * pow(0.5*(sigmai+sigmaj), 6);
- c12 = epsi * epsj * pow(0.5*(sigmai+sigmaj), 12);
+ c6 = sqrt(epsi * epsj) * pow(0.5*(sigmai+sigmaj), 6);
+ c12 = sqrt(epsi * epsj) * pow(0.5*(sigmai+sigmaj), 12);
}
C6(nbfp, atnr, i, j) = c6*6.0;
C12(nbfp, atnr, i, j) = c12*12.0;
bestsol = esolNO;
}
-#ifdef DISABLE_WATER_NLIST
- bestsol = esolNO;
-#endif
-
fr->nWatMol = 0;
for (mb = 0; mb < mtop->nmolblock; mb++)
{
{
/* At this point the init should never fail as we made sure that
* we have all the GPUs we need. If it still does, we'll bail. */
- gmx_fatal(FARGS, "On node %d failed to initialize GPU #%d: %s",
+ gmx_fatal(FARGS, "On rank %d failed to initialize GPU #%d: %s",
cr->nodeid,
get_gpu_device_id(&hwinfo->gpu_info, gpu_opt,
cr->rank_pp_intranode),
if (bUsesSimpleTables)
{
- /* With a spacing of 0.0005 we are at the force summation accuracy
- * for the SSE kernels for "normal" atomistic simulations.
+ /* Get the Ewald table spacing based on Coulomb and/or LJ
+ * Ewald coefficients and rtol.
*/
- ic->tabq_scale = ewald_spline3_table_scale(ic->ewaldcoeff_q,
- ic->rcoulomb);
+ ic->tabq_scale = ewald_spline3_table_scale(ic);
maxr = (rtab > ic->rcoulomb) ? rtab : ic->rcoulomb;
ic->tabq_size = (int)(maxr*ic->tabq_scale) + 2;
DOMAINDECOMP(cr) ? &cr->dd->nc : NULL,
DOMAINDECOMP(cr) ? domdec_zones(cr->dd) : NULL,
bFEP_NonBonded,
- gmx_omp_nthreads_get(emntNonbonded));
+ gmx_omp_nthreads_get(emntPairsearch));
for (i = 0; i < nbv->ngrp; i++)
{
fr->excl_load[t] = i;
}
}
+
+/* Frees GPU memory and destroys the CUDA context.
+ *
+ * Note that this function needs to be called even if GPUs are not used
+ * in this run because the PME ranks have no knowledge of whether GPUs
+ * are used or not, but all ranks need to enter the barrier below.
+ */
+void free_gpu_resources(const t_forcerec *fr,
+ const t_commrec *cr)
+{
+ gmx_bool bIsPPrankUsingGPU;
+ char gpu_err_str[STRLEN];
+
+ bIsPPrankUsingGPU = (cr->duty & DUTY_PP) && fr && fr->nbv && fr->nbv->bUseGPU;
+
+ if (bIsPPrankUsingGPU)
+ {
+ /* free nbnxn data in GPU memory */
+ nbnxn_cuda_free(fr->nbv->cu_nbv);
+
+ /* With tMPI we need to wait for all ranks to finish deallocation before
+ * destroying the context in free_gpu() as some ranks may be sharing
+ * GPU and context.
+ * Note: as only PP ranks need to free GPU resources, so it is safe to
+ * not call the barrier on PME ranks.
+ */
+#ifdef GMX_THREAD_MPI
+ if (PAR(cr))
+ {
+ gmx_barrier(cr);
+ }
+#endif /* GMX_THREAD_MPI */
+
+ /* uninitialize GPU (by destroying the context) */
+ if (!free_gpu(gpu_err_str))
+ {
+ gmx_warning("On rank %d failed to free GPU #%d: %s",
+ cr->nodeid, get_current_gpu_device_id(), gpu_err_str);
+ }
+ }
+}
biod.pnpi.spb.ru / alexxy / gromacs.git / blobdiff