--- /dev/null
- #if(GMX_THREADS)
+include_directories(${CMAKE_CURRENT_SOURCE_DIR})
+
+# The nonbonded directory contains subdirectories that are only
+# conditionally built, so we cannot use a GLOB_RECURSE here.
+file(GLOB GMXLIB_SOURCES *.c
+ statistics/*.c nonbonded/*.c nonbonded/nb_kernel_c/*.c
+ nonbonded/nb_kernel_adress_c/*.c)
+
+if(GMX_DOUBLE)
+ set(SSETYPE sse2)
+else()
+ set(SSETYPE sse)
+endif()
+
+if(GMX_IA32_ASM)
+ file(GLOB GMX_SSEKERNEL_C_SRC nonbonded/nb_kernel_ia32_${SSETYPE}/*.c)
+ if(GMX_ASM_USEASM_NASM)
+ file(GLOB GMX_SSEKERNEL_ASM_SRC nonbonded/nb_kernel_ia32_${SSETYPE}/*intel_syntax*.s)
+ else()
+ file(GLOB GMX_SSEKERNEL_ASM_SRC nonbonded/nb_kernel_ia32_${SSETYPE}/*${SSETYPE}.s nonbonded/nb_kernel_ia32_${SSETYPE}/*asm.s)
+ endif()
+endif(GMX_IA32_ASM)
+
+if(GMX_X86_64_ASM)
+ file(GLOB GMX_SSEKERNEL_C_SRC nonbonded/nb_kernel_x86_64_${SSETYPE}/*.c)
+ if(GMX_ASM_USEASM_NASM)
+ file(GLOB GMX_SSEKERNEL_ASM_SRC nonbonded/nb_kernel_x86_64_${SSETYPE}/*intel_syntax*.s)
+ else()
+ file(GLOB GMX_SSEKERNEL_ASM_SRC nonbonded/nb_kernel_x86_64_${SSETYPE}/*${SSETYPE}.s nonbonded/nb_kernel_x86_64_${SSETYPE}/*asm.s)
+ endif()
+endif(GMX_X86_64_ASM)
+
+if(GMX_FORTRAN)
+ if (GMX_DOUBLE)
+ file(GLOB FORTRAN_SOURCES nonbonded/nb_kernel_f77_double/*.[cf])
+ else(GMX_DOUBLE)
+ file(GLOB FORTRAN_SOURCES nonbonded/nb_kernel_f77_single/*.[cf])
+ endif(GMX_DOUBLE)
+endif(GMX_FORTRAN)
+
+if(GMX_POWER6)
+ file(GLOB FORTRAN_SOURCES nonbonded/nb_kernel_power6/*.[cF])
+endif(GMX_POWER6)
+
+if(GMX_BLUEGENE)
+ file(GLOB GMX_BLUEGENE_C_SRC nonbonded/nb_kernel_bluegene/*.c)
+endif(GMX_BLUEGENE)
+
+if(GMX_PPC_ALTIVEC)
+ file(GLOB GMX_PPC_ALTIVEC_SRC nonbonded/nb_kernel_ppc_altivec/*.c)
+endif(GMX_PPC_ALTIVEC)
+
+if(NOT GMX_EXTERNAL_BLAS)
+ file(GLOB BLAS_SOURCES gmx_blas/*.c)
+endif(NOT GMX_EXTERNAL_BLAS)
+
+if(NOT GMX_EXTERNAL_LAPACK)
+ file(GLOB LAPACK_SOURCES gmx_lapack/*.c)
+endif(NOT GMX_EXTERNAL_LAPACK)
+
+# This would be the standard way to include thread_mpi, but we want libgmx
+# to link the functions directly
- #endif(GMX_THREADS)
++#if(GMX_THREAD_MPI)
+# add_subdirectory(thread_mpi)
++#endif(GMX_THREAD_MPI)
+#target_link_libraries(gmx ${GMX_EXTRA_LIBRARIES} ${THREAD_MPI_LIB})
+
+# Files called xxx_test.c are test drivers with a main() function for module xxx.c,
+# so they should not be included in the library
+file(GLOB_RECURSE NOT_GMXLIB_SOURCES *_test.c *\#*)
+list(REMOVE_ITEM GMXLIB_SOURCES ${NOT_GMXLIB_SOURCES})
+
+# An ugly hack to get absolute paths...
+file(GLOB THREAD_MPI_SOURCES ${THREAD_MPI_SRC})
+
+set(GMX_SSEKERNEL_ASM_SRC ${GMX_SSEKERNEL_ASM_SRC} PARENT_SCOPE)
+set(GMXLIB_SOURCES ${GMXLIB_SOURCES} ${BLAS_SOURCES} ${LAPACK_SOURCES}
+ ${GMX_SSEKERNEL_C_SRC} ${FORTRAN_SOURCES}
+ ${GMX_BLUEGENE_C_SRC} ${GMX_PPC_ALTIVEC_SRC} ${THREAD_MPI_SOURCES}
+ PARENT_SCOPE)
--- /dev/null
- #ifdef GMX_THREADS
+/*
+ *
+ * This source code is part of
+ *
+ * G R O M A C S
+ *
+ * GROningen MAchine for Chemical Simulations
+ *
+ * VERSION 3.2.0
+ * Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
+ * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
+ * Copyright (c) 2001-2004, The GROMACS development team,
+ * check out http://www.gromacs.org for more information.
+
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version 2
+ * of the License, or (at your option) any later version.
+ *
+ * If you want to redistribute modifications, please consider that
+ * scientific software is very special. Version control is crucial -
+ * bugs must be traceable. We will be happy to consider code for
+ * inclusion in the official distribution, but derived work must not
+ * be called official GROMACS. Details are found in the README & COPYING
+ * files - if they are missing, get the official version at www.gromacs.org.
+ *
+ * To help us fund GROMACS development, we humbly ask that you cite
+ * the papers on the package - you can find them in the top README file.
+ *
+ * For more info, check our website at http://www.gromacs.org
+ *
+ * And Hey:
+ * GROningen Mixture of Alchemy and Childrens' Stories
+ */
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+#include <string.h>
+#include "gmx_fatal.h"
+#include "main.h"
+#include "smalloc.h"
+#include "network.h"
+#include "copyrite.h"
+#include "statutil.h"
+#include "ctype.h"
+#include "macros.h"
+
+#ifdef GMX_LIB_MPI
+#include <mpi.h>
+#endif
+
- #ifndef GMX_THREADS
++#ifdef GMX_THREAD_MPI
+#include "tmpi.h"
+#endif
+
+
+/* The source code in this file should be thread-safe.
+ Please keep it that way. */
+
+gmx_bool gmx_mpi_initialized(void)
+{
+ int n;
+#ifndef GMX_MPI
+ return 0;
+#else
+ MPI_Initialized(&n);
+
+ return n;
+#endif
+}
+
+int gmx_setup(int *argc,char **argv,int *nnodes)
+{
+#ifndef GMX_MPI
+ gmx_call("gmx_setup");
+ return 0;
+#else
+ char buf[256];
+ int resultlen; /* actual length of node name */
+ int i,flag;
+ int mpi_num_nodes;
+ int mpi_my_rank;
+ char mpi_hostname[MPI_MAX_PROCESSOR_NAME];
+
+ /* Call the MPI routines */
+#ifdef GMX_LIB_MPI
+#ifdef GMX_FAHCORE
+ (void) fah_MPI_Init(argc,&argv);
+#else
+ (void) MPI_Init(argc,&argv);
+#endif
+#endif
+ (void) MPI_Comm_size( MPI_COMM_WORLD, &mpi_num_nodes );
+ (void) MPI_Comm_rank( MPI_COMM_WORLD, &mpi_my_rank );
+ (void) MPI_Get_processor_name( mpi_hostname, &resultlen );
+
+#ifdef GMX_LIB_MPI
+ fprintf(stderr,"NNODES=%d, MYRANK=%d, HOSTNAME=%s\n",
+ mpi_num_nodes,mpi_my_rank,mpi_hostname);
+#endif
+
+ *nnodes=mpi_num_nodes;
+
+ return mpi_my_rank;
+#endif
+}
+
+int gmx_node_num(void)
+{
+#ifndef GMX_MPI
+ return 1;
+#else
+ int i;
+ (void) MPI_Comm_size(MPI_COMM_WORLD, &i);
+ return i;
+#endif
+}
+
+int gmx_node_rank(void)
+{
+#ifndef GMX_MPI
+ return 0;
+#else
+ int i;
+ (void) MPI_Comm_rank(MPI_COMM_WORLD, &i);
+ return i;
+#endif
+}
+
+
+int gmx_hostname_num()
+{
+#ifndef GMX_MPI
+ return 0;
+#else
+ int resultlen,hostnum,i,j;
+ char mpi_hostname[MPI_MAX_PROCESSOR_NAME],hostnum_str[MPI_MAX_PROCESSOR_NAME];
+
+ MPI_Get_processor_name(mpi_hostname,&resultlen);
+ /* This procedure can only differentiate nodes with host names
+ * that end on unique numbers.
+ */
+ i = 0;
+ j = 0;
+ /* Only parse the host name up to the first dot */
+ while(i < resultlen && mpi_hostname[i] != '.') {
+ if (isdigit(mpi_hostname[i])) {
+ hostnum_str[j++] = mpi_hostname[i];
+ }
+ i++;
+ }
+ hostnum_str[j] = '\0';
+ if (j == 0) {
+ hostnum = 0;
+ } else {
+ /* Use only the last 9 decimals, so we don't overflow an int */
+ hostnum = strtol(hostnum_str + max(0,j-9), NULL, 10);
+ }
+
+ if (debug) {
+ fprintf(debug,"In gmx_setup_nodecomm: hostname '%s', hostnum %d\n",
+ mpi_hostname,hostnum);
+ }
+ return hostnum;
+#endif
+}
+
+void gmx_setup_nodecomm(FILE *fplog,t_commrec *cr)
+{
+ gmx_nodecomm_t *nc;
+ int n,rank,hostnum,ng,ni;
+
+ /* Many MPI implementations do not optimize MPI_Allreduce
+ * (and probably also other global communication calls)
+ * for multi-core nodes connected by a network.
+ * We can optimize such communication by using one MPI call
+ * within each node and one between the nodes.
+ * For MVAPICH2 and Intel MPI this reduces the time for
+ * the global_stat communication by 25%
+ * for 2x2-core 3 GHz Woodcrest connected by mixed DDR/SDR Infiniband.
+ * B. Hess, November 2007
+ */
+
+ nc = &cr->nc;
+
+ nc->bUse = FALSE;
- #ifdef GMX_THREADS
++#ifndef GMX_THREAD_MPI
+ if (getenv("GMX_NO_NODECOMM") == NULL) {
+#ifdef GMX_MPI
+ MPI_Comm_size(cr->mpi_comm_mygroup,&n);
+ MPI_Comm_rank(cr->mpi_comm_mygroup,&rank);
+
+ hostnum = gmx_hostname_num();
+
+ if (debug) {
+ fprintf(debug,
+ "In gmx_setup_nodecomm: splitting communicator of size %d\n",
+ n);
+ }
+
+
+ /* The intra-node communicator, split on node number */
+ MPI_Comm_split(cr->mpi_comm_mygroup,hostnum,rank,&nc->comm_intra);
+ MPI_Comm_rank(nc->comm_intra,&nc->rank_intra);
+ if (debug) {
+ fprintf(debug,"In gmx_setup_nodecomm: node rank %d rank_intra %d\n",
+ rank,nc->rank_intra);
+ }
+ /* The inter-node communicator, split on rank_intra.
+ * We actually only need the one for rank=0,
+ * but it is easier to create them all.
+ */
+ MPI_Comm_split(cr->mpi_comm_mygroup,nc->rank_intra,rank,&nc->comm_inter);
+ /* Check if this really created two step communication */
+ MPI_Comm_size(nc->comm_inter,&ng);
+ MPI_Comm_size(nc->comm_intra,&ni);
+ if (debug) {
+ fprintf(debug,"In gmx_setup_nodecomm: groups %d, my group size %d\n",
+ ng,ni);
+ }
+ if ((ng > 1 && ng < n) || (ni > 1 && ni < n)) {
+ nc->bUse = TRUE;
+ if (fplog)
+ fprintf(fplog,"Using two step summing over %d groups of on average %.1f processes\n\n",ng,(real)n/(real)ng);
+ if (nc->rank_intra > 0)
+ MPI_Comm_free(&nc->comm_inter);
+ } else {
+ /* One group or all processes in a separate group, use normal summing */
+ MPI_Comm_free(&nc->comm_inter);
+ MPI_Comm_free(&nc->comm_intra);
+ }
+#endif
+ }
+#endif
+}
+
+void gmx_barrier(const t_commrec *cr)
+{
+#ifndef GMX_MPI
+ gmx_call("gmx_barrier");
+#else
+ MPI_Barrier(cr->mpi_comm_mygroup);
+#endif
+}
+
+void gmx_abort(int noderank,int nnodes,int errorno)
+{
+#ifndef GMX_MPI
+ gmx_call("gmx_abort");
+#else
- #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREADS)
++#ifdef GMX_THREAD_MPI
+ fprintf(stderr,"Halting program %s\n",ShortProgram());
+ thanx(stderr);
+ exit(1);
+#else
+ if (nnodes > 1)
+ {
+ fprintf(stderr,"Halting parallel program %s on CPU %d out of %d\n",
+ ShortProgram(),noderank,nnodes);
+ }
+ else
+ {
+ fprintf(stderr,"Halting program %s\n",ShortProgram());
+ }
+
+ thanx(stderr);
+ MPI_Abort(MPI_COMM_WORLD,errorno);
+ exit(1);
+#endif
+#endif
+}
+
+void gmx_bcast(int nbytes,void *b,const t_commrec *cr)
+{
+#ifndef GMX_MPI
+ gmx_call("gmx_bast");
+#else
+ MPI_Bcast(b,nbytes,MPI_BYTE,MASTERRANK(cr),cr->mpi_comm_mygroup);
+#endif
+}
+
+void gmx_bcast_sim(int nbytes,void *b,const t_commrec *cr)
+{
+#ifndef GMX_MPI
+ gmx_call("gmx_bast");
+#else
+ MPI_Bcast(b,nbytes,MPI_BYTE,MASTERRANK(cr),cr->mpi_comm_mysim);
+#endif
+}
+
+void gmx_sumd(int nr,double r[],const t_commrec *cr)
+{
+#ifndef GMX_MPI
+ gmx_call("gmx_sumd");
+#else
- #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREADS)
++#if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREAD_MPI)
+ if (cr->nc.bUse) {
+ if (cr->nc.rank_intra == 0)
+ {
+ /* Use two step summing. */
+ MPI_Reduce(MPI_IN_PLACE,r,nr,MPI_DOUBLE,MPI_SUM,0,
+ cr->nc.comm_intra);
+ /* Sum the roots of the internal (intra) buffers. */
+ MPI_Allreduce(MPI_IN_PLACE,r,nr,MPI_DOUBLE,MPI_SUM,
+ cr->nc.comm_inter);
+ }
+ else
+ {
+ /* This is here because of the silly MPI specification
+ that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
+ MPI_Reduce(r,NULL,nr,MPI_DOUBLE,MPI_SUM,0,cr->nc.comm_intra);
+ }
+ MPI_Bcast(r,nr,MPI_DOUBLE,0,cr->nc.comm_intra);
+ }
+ else
+ {
+ MPI_Allreduce(MPI_IN_PLACE,r,nr,MPI_DOUBLE,MPI_SUM,
+ cr->mpi_comm_mygroup);
+ }
+#else
+ int i;
+
+ if (nr > cr->mpb->dbuf_alloc) {
+ cr->mpb->dbuf_alloc = nr;
+ srenew(cr->mpb->dbuf,cr->mpb->dbuf_alloc);
+ }
+ if (cr->nc.bUse) {
+ /* Use two step summing */
+ MPI_Allreduce(r,cr->mpb->dbuf,nr,MPI_DOUBLE,MPI_SUM,cr->nc.comm_intra);
+ if (cr->nc.rank_intra == 0) {
+ /* Sum with the buffers reversed */
+ MPI_Allreduce(cr->mpb->dbuf,r,nr,MPI_DOUBLE,MPI_SUM,
+ cr->nc.comm_inter);
+ }
+ MPI_Bcast(r,nr,MPI_DOUBLE,0,cr->nc.comm_intra);
+ } else {
+ MPI_Allreduce(r,cr->mpb->dbuf,nr,MPI_DOUBLE,MPI_SUM,
+ cr->mpi_comm_mygroup);
+ for(i=0; i<nr; i++)
+ r[i] = cr->mpb->dbuf[i];
+ }
+#endif
+#endif
+}
+
+void gmx_sumf(int nr,float r[],const t_commrec *cr)
+{
+#ifndef GMX_MPI
+ gmx_call("gmx_sumf");
+#else
- #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREADS)
++#if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREAD_MPI)
+ if (cr->nc.bUse) {
+ /* Use two step summing. */
+ if (cr->nc.rank_intra == 0)
+ {
+ MPI_Reduce(MPI_IN_PLACE,r,nr,MPI_FLOAT,MPI_SUM,0,
+ cr->nc.comm_intra);
+ /* Sum the roots of the internal (intra) buffers */
+ MPI_Allreduce(MPI_IN_PLACE,r,nr,MPI_FLOAT,MPI_SUM,
+ cr->nc.comm_inter);
+ }
+ else
+ {
+ /* This is here because of the silly MPI specification
+ that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
+ MPI_Reduce(r,NULL,nr,MPI_FLOAT,MPI_SUM,0,cr->nc.comm_intra);
+ }
+ MPI_Bcast(r,nr,MPI_FLOAT,0,cr->nc.comm_intra);
+ }
+ else
+ {
+ MPI_Allreduce(MPI_IN_PLACE,r,nr,MPI_FLOAT,MPI_SUM,cr->mpi_comm_mygroup);
+ }
+#else
+ int i;
+
+ if (nr > cr->mpb->fbuf_alloc) {
+ cr->mpb->fbuf_alloc = nr;
+ srenew(cr->mpb->fbuf,cr->mpb->fbuf_alloc);
+ }
+ if (cr->nc.bUse) {
+ /* Use two step summing */
+ MPI_Allreduce(r,cr->mpb->fbuf,nr,MPI_FLOAT,MPI_SUM,cr->nc.comm_intra);
+ if (cr->nc.rank_intra == 0) {
+ /* Sum with the buffers reversed */
+ MPI_Allreduce(cr->mpb->fbuf,r,nr,MPI_FLOAT,MPI_SUM,
+ cr->nc.comm_inter);
+ }
+ MPI_Bcast(r,nr,MPI_FLOAT,0,cr->nc.comm_intra);
+ } else {
+ MPI_Allreduce(r,cr->mpb->fbuf,nr,MPI_FLOAT,MPI_SUM,
+ cr->mpi_comm_mygroup);
+ for(i=0; i<nr; i++)
+ r[i] = cr->mpb->fbuf[i];
+ }
+#endif
+#endif
+}
+
+void gmx_sumi(int nr,int r[],const t_commrec *cr)
+{
+#ifndef GMX_MPI
+ gmx_call("gmx_sumi");
+#else
- #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREADS)
++#if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREAD_MPI)
+ if (cr->nc.bUse) {
+ /* Use two step summing */
+ if (cr->nc.rank_intra == 0)
+ {
+ MPI_Reduce(MPI_IN_PLACE,r,nr,MPI_INT,MPI_SUM,0,cr->nc.comm_intra);
+ /* Sum with the buffers reversed */
+ MPI_Allreduce(MPI_IN_PLACE,r,nr,MPI_INT,MPI_SUM,cr->nc.comm_inter);
+ }
+ else
+ {
+ /* This is here because of the silly MPI specification
+ that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
+ MPI_Reduce(r,NULL,nr,MPI_INT,MPI_SUM,0,cr->nc.comm_intra);
+ }
+ MPI_Bcast(r,nr,MPI_INT,0,cr->nc.comm_intra);
+ }
+ else
+ {
+ MPI_Allreduce(MPI_IN_PLACE,r,nr,MPI_INT,MPI_SUM,cr->mpi_comm_mygroup);
+ }
+#else
+ int i;
+
+ if (nr > cr->mpb->ibuf_alloc) {
+ cr->mpb->ibuf_alloc = nr;
+ srenew(cr->mpb->ibuf,cr->mpb->ibuf_alloc);
+ }
+ if (cr->nc.bUse) {
+ /* Use two step summing */
+ MPI_Allreduce(r,cr->mpb->ibuf,nr,MPI_INT,MPI_SUM,cr->nc.comm_intra);
+ if (cr->nc.rank_intra == 0) {
+ /* Sum with the buffers reversed */
+ MPI_Allreduce(cr->mpb->ibuf,r,nr,MPI_INT,MPI_SUM,cr->nc.comm_inter);
+ }
+ MPI_Bcast(r,nr,MPI_INT,0,cr->nc.comm_intra);
+ } else {
+ MPI_Allreduce(r,cr->mpb->ibuf,nr,MPI_INT,MPI_SUM,cr->mpi_comm_mygroup);
+ for(i=0; i<nr; i++)
+ r[i] = cr->mpb->ibuf[i];
+ }
+#endif
+#endif
+}
+
+void gmx_sumli(int nr,gmx_large_int_t r[],const t_commrec *cr)
+{
+#ifndef GMX_MPI
+ gmx_call("gmx_sumli");
+#else
- #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREADS)
++#if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREAD_MPI)
+ if (cr->nc.bUse) {
+ /* Use two step summing */
+ if (cr->nc.rank_intra == 0)
+ {
+ MPI_Reduce(MPI_IN_PLACE,r,nr,GMX_MPI_LARGE_INT,MPI_SUM,0,
+ cr->nc.comm_intra);
+ /* Sum with the buffers reversed */
+ MPI_Allreduce(MPI_IN_PLACE,r,nr,GMX_MPI_LARGE_INT,MPI_SUM,
+ cr->nc.comm_inter);
+ }
+ else
+ {
+ /* This is here because of the silly MPI specification
+ that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
+ MPI_Reduce(r,NULL,nr,GMX_MPI_LARGE_INT,MPI_SUM,0,cr->nc.comm_intra);
+ }
+ MPI_Bcast(r,nr,GMX_MPI_LARGE_INT,0,cr->nc.comm_intra);
+ }
+ else
+ {
+ MPI_Allreduce(MPI_IN_PLACE,r,nr,GMX_MPI_LARGE_INT,MPI_SUM,cr->mpi_comm_mygroup);
+ }
+#else
+ int i;
+
+ if (nr > cr->mpb->libuf_alloc) {
+ cr->mpb->libuf_alloc = nr;
+ srenew(cr->mpb->libuf,cr->mpb->libuf_alloc);
+ }
+ if (cr->nc.bUse) {
+ /* Use two step summing */
+ MPI_Allreduce(r,cr->mpb->libuf,nr,GMX_MPI_LARGE_INT,MPI_SUM,
+ cr->nc.comm_intra);
+ if (cr->nc.rank_intra == 0) {
+ /* Sum with the buffers reversed */
+ MPI_Allreduce(cr->mpb->libuf,r,nr,GMX_MPI_LARGE_INT,MPI_SUM,
+ cr->nc.comm_inter);
+ }
+ MPI_Bcast(r,nr,GMX_MPI_LARGE_INT,0,cr->nc.comm_intra);
+ } else {
+ MPI_Allreduce(r,cr->mpb->libuf,nr,GMX_MPI_LARGE_INT,MPI_SUM,
+ cr->mpi_comm_mygroup);
+ for(i=0; i<nr; i++)
+ r[i] = cr->mpb->libuf[i];
+ }
+#endif
+#endif
+}
+
+
+
+#ifdef GMX_MPI
+void gmx_sumd_comm(int nr,double r[],MPI_Comm mpi_comm)
+{
- #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREADS)
++#if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREAD_MPI)
+ MPI_Allreduce(MPI_IN_PLACE,r,nr,MPI_DOUBLE,MPI_SUM,mpi_comm);
+#else
+ /* this function is only used in code that is not performance critical,
+ (during setup, when comm_rec is not the appropriate communication
+ structure), so this isn't as bad as it looks. */
+ double *buf;
+ int i;
+
+ snew(buf, nr);
+ MPI_Allreduce(r,buf,nr,MPI_DOUBLE,MPI_SUM,mpi_comm);
+ for(i=0; i<nr; i++)
+ r[i] = buf[i];
+ sfree(buf);
+#endif
+}
+#endif
+
+#ifdef GMX_MPI
+void gmx_sumf_comm(int nr,float r[],MPI_Comm mpi_comm)
+{
- #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREADS)
++#if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREAD_MPI)
+ MPI_Allreduce(MPI_IN_PLACE,r,nr,MPI_FLOAT,MPI_SUM,mpi_comm);
+#else
+ /* this function is only used in code that is not performance critical,
+ (during setup, when comm_rec is not the appropriate communication
+ structure), so this isn't as bad as it looks. */
+ float *buf;
+ int i;
+
+ snew(buf, nr);
+ MPI_Allreduce(r,buf,nr,MPI_FLOAT,MPI_SUM,mpi_comm);
+ for(i=0; i<nr; i++)
+ r[i] = buf[i];
+ sfree(buf);
+#endif
+}
+#endif
+
+void gmx_sumd_sim(int nr,double r[],const gmx_multisim_t *ms)
+{
+#ifndef GMX_MPI
+ gmx_call("gmx_sumd_sim");
+#else
+ gmx_sumd_comm(nr,r,ms->mpi_comm_masters);
+#endif
+}
+
+void gmx_sumf_sim(int nr,float r[],const gmx_multisim_t *ms)
+{
+#ifndef GMX_MPI
+ gmx_call("gmx_sumf_sim");
+#else
+ gmx_sumf_comm(nr,r,ms->mpi_comm_masters);
+#endif
+}
+
+void gmx_sumi_sim(int nr,int r[], const gmx_multisim_t *ms)
+{
+#ifndef GMX_MPI
+ gmx_call("gmx_sumi_sim");
+#else
- #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREADS)
++#if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREAD_MPI)
+ MPI_Allreduce(MPI_IN_PLACE,r,nr,MPI_INT,MPI_SUM,ms->mpi_comm_masters);
+#else
+ /* this is thread-unsafe, but it will do for now: */
+ int i;
+
+ if (nr > ms->mpb->ibuf_alloc) {
+ ms->mpb->ibuf_alloc = nr;
+ srenew(ms->mpb->ibuf,ms->mpb->ibuf_alloc);
+ }
+ MPI_Allreduce(r,ms->mpb->ibuf,nr,MPI_INT,MPI_SUM,ms->mpi_comm_masters);
+ for(i=0; i<nr; i++)
+ r[i] = ms->mpb->ibuf[i];
+#endif
+#endif
+}
+
+void gmx_sumli_sim(int nr,gmx_large_int_t r[], const gmx_multisim_t *ms)
+{
+#ifndef GMX_MPI
+ gmx_call("gmx_sumli_sim");
+#else
++#if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREAD_MPI)
+ MPI_Allreduce(MPI_IN_PLACE,r,nr,GMX_MPI_LARGE_INT,MPI_SUM,
+ ms->mpi_comm_masters);
+#else
+ /* this is thread-unsafe, but it will do for now: */
+ int i;
+
+ if (nr > ms->mpb->libuf_alloc) {
+ ms->mpb->libuf_alloc = nr;
+ srenew(ms->mpb->libuf,ms->mpb->libuf_alloc);
+ }
+ MPI_Allreduce(r,ms->mpb->libuf,nr,GMX_MPI_LARGE_INT,MPI_SUM,
+ ms->mpi_comm_masters);
+ for(i=0; i<nr; i++)
+ r[i] = ms->mpb->libuf[i];
+#endif
+#endif
+}
+
+
+void gmx_finalize(void)
+{
+#ifndef GMX_MPI
+ gmx_call("gmx_finalize");
+#else
+ int ret;
+
+ /* just as a check; we don't want to finalize twice */
+ int finalized;
+ MPI_Finalized(&finalized);
+ if (finalized)
+ return;
+
+ /* We sync the processes here to try to avoid problems
+ * with buggy MPI implementations that could cause
+ * unfinished processes to terminate.
+ */
+ MPI_Barrier(MPI_COMM_WORLD);
+
+ /*
+ if (DOMAINDECOMP(cr)) {
+ if (cr->npmenodes > 0 || cr->dd->bCartesian)
+ MPI_Comm_free(&cr->mpi_comm_mygroup);
+ if (cr->dd->bCartesian)
+ MPI_Comm_free(&cr->mpi_comm_mysim);
+ }
+ */
+
+ /* Apparently certain mpich implementations cause problems
+ * with MPI_Finalize. In that case comment out MPI_Finalize.
+ */
+ if (debug)
+ fprintf(debug,"Will call MPI_Finalize now\n");
+
+ ret = MPI_Finalize();
+ if (debug)
+ fprintf(debug,"Return code from MPI_Finalize = %d\n",ret);
+#endif
+}
+
--- /dev/null
- #ifdef GMX_THREADS
+/* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
+ *
+ *
+ * This source code is part of
+ *
+ * G R O M A C S
+ *
+ * GROningen MAchine for Chemical Simulations
+ *
+ * VERSION 3.2.0
+ * Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
+ * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
+ * Copyright (c) 2001-2004, The GROMACS development team,
+ * check out http://www.gromacs.org for more information.
+
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version 2
+ * of the License, or (at your option) any later version.
+ *
+ * If you want to redistribute modifications, please consider that
+ * scientific software is very special. Version control is crucial -
+ * bugs must be traceable. We will be happy to consider code for
+ * inclusion in the official distribution, but derived work must not
+ * be called official GROMACS. Details are found in the README & COPYING
+ * files - if they are missing, get the official version at www.gromacs.org.
+ *
+ * To help us fund GROMACS development, we humbly ask that you cite
+ * the papers on the package - you can find them in the top README file.
+ *
+ * For more info, check our website at http://www.gromacs.org
+ *
+ * And Hey:
+ * GROningen Mixture of Alchemy and Childrens' Stories
+ */
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+/* This file is completely threadsafe - please keep it that way! */
++#ifdef GMX_THREAD_MPI
+#include <thread_mpi.h>
+#endif
+
+
+#include <stdio.h>
+#include "smalloc.h"
+#include "typedefs.h"
+#include "names.h"
+#include "txtdump.h"
+#include "string2.h"
+#include "vec.h"
+#include "macros.h"
+
+
+int pr_indent(FILE *fp,int n)
+{
+ int i;
+
+ for (i=0; i<n; i++) (void) fprintf(fp," ");
+ return n;
+}
+
+int available(FILE *fp,void *p,int indent,const char *title)
+{
+ if (!p) {
+ if (indent > 0)
+ pr_indent(fp,indent);
+ (void) fprintf(fp,"%s: not available\n",title);
+ }
+ return (p!=NULL);
+}
+
+int pr_title(FILE *fp,int indent,const char *title)
+{
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"%s:\n",title);
+ return (indent+INDENT);
+}
+
+int pr_title_n(FILE *fp,int indent,const char *title,int n)
+{
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"%s (%d):\n",title,n);
+ return (indent+INDENT);
+}
+
+int pr_title_nxn(FILE *fp,int indent,const char *title,int n1,int n2)
+{
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"%s (%dx%d):\n",title,n1,n2);
+ return (indent+INDENT);
+}
+
+void pr_ivec(FILE *fp,int indent,const char *title,int vec[],int n, gmx_bool bShowNumbers)
+{
+ int i;
+
+ if (available(fp,vec,indent,title))
+ {
+ indent=pr_title_n(fp,indent,title,n);
+ for (i=0; i<n; i++)
+ {
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"%s[%d]=%d\n",title,bShowNumbers?i:-1,vec[i]);
+ }
+ }
+}
+
+void pr_ivec_block(FILE *fp,int indent,const char *title,int vec[],int n, gmx_bool bShowNumbers)
+{
+ int i,j;
+
+ if (available(fp,vec,indent,title))
+ {
+ indent=pr_title_n(fp,indent,title,n);
+ i = 0;
+ while (i < n)
+ {
+ j = i+1;
+ while (j < n && vec[j] == vec[j-1]+1)
+ {
+ j++;
+ }
+ /* Print consecutive groups of 3 or more as blocks */
+ if (j - i < 3)
+ {
+ while(i < j)
+ {
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"%s[%d]=%d\n",
+ title,bShowNumbers?i:-1,vec[i]);
+ i++;
+ }
+ }
+ else
+ {
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"%s[%d,...,%d] = {%d,...,%d}\n",
+ title,
+ bShowNumbers?i:-1,
+ bShowNumbers?j-1:-1,
+ vec[i],vec[j-1]);
+ i = j;
+ }
+ }
+ }
+}
+
+void pr_bvec(FILE *fp,int indent,const char *title,gmx_bool vec[],int n, gmx_bool bShowNumbers)
+{
+ int i;
+
+ if (available(fp,vec,indent,title))
+ {
+ indent=pr_title_n(fp,indent,title,n);
+ for (i=0; i<n; i++)
+ {
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"%s[%d]=%s\n",title,bShowNumbers?i:-1,
+ BOOL(vec[i]));
+ }
+ }
+}
+
+void pr_ivecs(FILE *fp,int indent,const char *title,ivec vec[],int n, gmx_bool bShowNumbers)
+{
+ int i,j;
+
+ if (available(fp,vec,indent,title))
+ {
+ indent=pr_title_nxn(fp,indent,title,n,DIM);
+ for (i=0; i<n; i++)
+ {
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"%s[%d]={",title,bShowNumbers?i:-1);
+ for (j=0; j<DIM; j++)
+ {
+ if (j!=0) (void) fprintf(fp,", ");
+ fprintf(fp,"%d",vec[i][j]);
+ }
+ (void) fprintf(fp,"}\n");
+ }
+ }
+}
+
+void pr_rvec(FILE *fp,int indent,const char *title,real vec[],int n, gmx_bool bShowNumbers)
+{
+ int i;
+
+ if (available(fp,vec,indent,title))
+ {
+ indent=pr_title_n(fp,indent,title,n);
+ for (i=0; i<n; i++)
+ {
+ pr_indent(fp,indent);
+ fprintf(fp,"%s[%d]=%12.5e\n",title,bShowNumbers?i:-1,vec[i]);
+ }
+ }
+}
+
+void pr_dvec(FILE *fp,int indent,const char *title,double vec[],int n, gmx_bool bShowNumbers)
+{
+ int i;
+
+ if (available(fp,vec,indent,title))
+ {
+ indent=pr_title_n(fp,indent,title,n);
+ for (i=0; i<n; i++)
+ {
+ pr_indent(fp,indent);
+ fprintf(fp,"%s[%d]=%12.5e\n",title,bShowNumbers?i:-1,vec[i]);
+ }
+ }
+}
+
+
+/*
+void pr_mat(FILE *fp,int indent,char *title,matrix m)
+{
+ int i,j;
+
+ if (available(fp,m,indent,title)) {
+ indent=pr_title_n(fp,indent,title,n);
+ for(i=0; i<n; i++) {
+ pr_indent(fp,indent);
+ fprintf(fp,"%s[%d]=%12.5e %12.5e %12.5e\n",
+ title,bShowNumbers?i:-1,m[i][XX],m[i][YY],m[i][ZZ]);
+ }
+ }
+}
+*/
+
+void pr_rvecs_len(FILE *fp,int indent,const char *title,rvec vec[],int n)
+{
+ int i,j;
+
+ if (available(fp,vec,indent,title)) {
+ indent=pr_title_nxn(fp,indent,title,n,DIM);
+ for (i=0; i<n; i++) {
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"%s[%5d]={",title,i);
+ for (j=0; j<DIM; j++) {
+ if (j != 0)
+ (void) fprintf(fp,", ");
+ (void) fprintf(fp,"%12.5e",vec[i][j]);
+ }
+ (void) fprintf(fp,"} len=%12.5e\n",norm(vec[i]));
+ }
+ }
+}
+
+void pr_rvecs(FILE *fp,int indent,const char *title,rvec vec[],int n)
+{
+ const char *fshort = "%12.5e";
+ const char *flong = "%15.8e";
+ const char *format;
+ int i,j;
+
+ if (getenv("LONGFORMAT") != NULL)
+ format = flong;
+ else
+ format = fshort;
+
+ if (available(fp,vec,indent,title)) {
+ indent=pr_title_nxn(fp,indent,title,n,DIM);
+ for (i=0; i<n; i++) {
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"%s[%5d]={",title,i);
+ for (j=0; j<DIM; j++) {
+ if (j != 0)
+ (void) fprintf(fp,", ");
+ (void) fprintf(fp,format,vec[i][j]);
+ }
+ (void) fprintf(fp,"}\n");
+ }
+ }
+}
+
+
+void pr_reals(FILE *fp,int indent,const char *title,real *vec,int n)
+{
+ int i;
+
+ if (available(fp,vec,indent,title)) {
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"%s:\t",title);
+ for(i=0; i<n; i++)
+ fprintf(fp," %10g",vec[i]);
+ (void) fprintf(fp,"\n");
+ }
+}
+
+void pr_doubles(FILE *fp,int indent,const char *title,double *vec,int n)
+{
+ int i;
+
+ if (available(fp,vec,indent,title)) {
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"%s:\t",title);
+ for(i=0; i<n; i++)
+ fprintf(fp," %10g",vec[i]);
+ (void) fprintf(fp,"\n");
+ }
+}
+
+static void pr_int(FILE *fp,int indent,const char *title,int i)
+{
+ pr_indent(fp,indent);
+ fprintf(fp,"%-20s = %d\n",title,i);
+}
+
+static void pr_gmx_large_int(FILE *fp,int indent,const char *title,gmx_large_int_t i)
+{
+ char buf[STEPSTRSIZE];
+
+ pr_indent(fp,indent);
+ fprintf(fp,"%-20s = %s\n",title,gmx_step_str(i,buf));
+}
+
+static void pr_real(FILE *fp,int indent,const char *title,real r)
+{
+ pr_indent(fp,indent);
+ fprintf(fp,"%-20s = %g\n",title,r);
+}
+
+static void pr_double(FILE *fp,int indent,const char *title,double d)
+{
+ pr_indent(fp,indent);
+ fprintf(fp,"%-20s = %g\n",title,d);
+}
+
+static void pr_str(FILE *fp,int indent,const char *title,const char *s)
+{
+ pr_indent(fp,indent);
+ fprintf(fp,"%-20s = %s\n",title,s);
+}
+
+void pr_qm_opts(FILE *fp,int indent,const char *title,t_grpopts *opts)
+{
+ int i,m,j;
+
+ fprintf(fp,"%s:\n",title);
+
+ pr_int(fp,indent,"ngQM",opts->ngQM);
+ if (opts->ngQM > 0) {
+ pr_ivec(fp,indent,"QMmethod",opts->QMmethod,opts->ngQM,FALSE);
+ pr_ivec(fp,indent,"QMbasis",opts->QMbasis,opts->ngQM,FALSE);
+ pr_ivec(fp,indent,"QMcharge",opts->QMcharge,opts->ngQM,FALSE);
+ pr_ivec(fp,indent,"QMmult",opts->QMmult,opts->ngQM,FALSE);
+ pr_bvec(fp,indent,"bSH",opts->bSH,opts->ngQM,FALSE);
+ pr_ivec(fp,indent,"CASorbitals",opts->CASorbitals,opts->ngQM,FALSE);
+ pr_ivec(fp,indent,"CASelectrons",opts->CASelectrons,opts->ngQM,FALSE);
+ pr_rvec(fp,indent,"SAon",opts->SAon,opts->ngQM,FALSE);
+ pr_rvec(fp,indent,"SAon",opts->SAon,opts->ngQM,FALSE);
+ pr_ivec(fp,indent,"SAsteps",opts->SAsteps,opts->ngQM,FALSE);
+ pr_bvec(fp,indent,"bOPT",opts->bOPT,opts->ngQM,FALSE);
+ pr_bvec(fp,indent,"bTS",opts->bTS,opts->ngQM,FALSE);
+ }
+}
+
+static void pr_grp_opts(FILE *out,int indent,const char *title,t_grpopts *opts,
+ gmx_bool bMDPformat)
+{
+ int i,m,j;
+
+ if (!bMDPformat)
+ fprintf(out,"%s:\n",title);
+
+ pr_indent(out,indent);
+ fprintf(out,"nrdf%s",bMDPformat ? " = " : ":");
+ for(i=0; (i<opts->ngtc); i++)
+ fprintf(out," %10g",opts->nrdf[i]);
+ fprintf(out,"\n");
+
+ pr_indent(out,indent);
+ fprintf(out,"ref-t%s",bMDPformat ? " = " : ":");
+ for(i=0; (i<opts->ngtc); i++)
+ fprintf(out," %10g",opts->ref_t[i]);
+ fprintf(out,"\n");
+
+ pr_indent(out,indent);
+ fprintf(out,"tau-t%s",bMDPformat ? " = " : ":");
+ for(i=0; (i<opts->ngtc); i++)
+ fprintf(out," %10g",opts->tau_t[i]);
+ fprintf(out,"\n");
+
+ /* Pretty-print the simulated annealing info */
+ fprintf(out,"anneal%s",bMDPformat ? " = " : ":");
+ for(i=0; (i<opts->ngtc); i++)
+ fprintf(out," %10s",EANNEAL(opts->annealing[i]));
+ fprintf(out,"\n");
+
+ fprintf(out,"ann-npoints%s",bMDPformat ? " = " : ":");
+ for(i=0; (i<opts->ngtc); i++)
+ fprintf(out," %10d",opts->anneal_npoints[i]);
+ fprintf(out,"\n");
+
+ for(i=0; (i<opts->ngtc); i++) {
+ if(opts->anneal_npoints[i]>0) {
+ fprintf(out,"ann. times [%d]:\t",i);
+ for(j=0; (j<opts->anneal_npoints[i]); j++)
+ fprintf(out," %10.1f",opts->anneal_time[i][j]);
+ fprintf(out,"\n");
+ fprintf(out,"ann. temps [%d]:\t",i);
+ for(j=0; (j<opts->anneal_npoints[i]); j++)
+ fprintf(out," %10.1f",opts->anneal_temp[i][j]);
+ fprintf(out,"\n");
+ }
+ }
+
+ pr_indent(out,indent);
+ fprintf(out,"acc:\t");
+ for(i=0; (i<opts->ngacc); i++)
+ for(m=0; (m<DIM); m++)
+ fprintf(out," %10g",opts->acc[i][m]);
+ fprintf(out,"\n");
+
+ pr_indent(out,indent);
+ fprintf(out,"nfreeze:");
+ for(i=0; (i<opts->ngfrz); i++)
+ for(m=0; (m<DIM); m++)
+ fprintf(out," %10s",opts->nFreeze[i][m] ? "Y" : "N");
+ fprintf(out,"\n");
+
+
+ for(i=0; (i<opts->ngener); i++) {
+ pr_indent(out,indent);
+ fprintf(out,"energygrp-flags[%3d]:",i);
+ for(m=0; (m<opts->ngener); m++)
+ fprintf(out," %d",opts->egp_flags[opts->ngener*i+m]);
+ fprintf(out,"\n");
+ }
+
+ fflush(out);
+}
+
+static void pr_matrix(FILE *fp,int indent,const char *title,rvec *m,
+ gmx_bool bMDPformat)
+{
+ if (bMDPformat)
+ fprintf(fp,"%-10s = %g %g %g %g %g %g\n",title,
+ m[XX][XX],m[YY][YY],m[ZZ][ZZ],m[XX][YY],m[XX][ZZ],m[YY][ZZ]);
+ else
+ pr_rvecs(fp,indent,title,m,DIM);
+}
+
+static void pr_cosine(FILE *fp,int indent,const char *title,t_cosines *cos,
+ gmx_bool bMDPformat)
+{
+ int j;
+
+ if (bMDPformat) {
+ fprintf(fp,"%s = %d\n",title,cos->n);
+ }
+ else {
+ indent=pr_title(fp,indent,title);
+ (void) pr_indent(fp,indent);
+ fprintf(fp,"n = %d\n",cos->n);
+ if (cos->n > 0) {
+ (void) pr_indent(fp,indent+2);
+ fprintf(fp,"a =");
+ for(j=0; (j<cos->n); j++)
+ fprintf(fp," %e",cos->a[j]);
+ fprintf(fp,"\n");
+ (void) pr_indent(fp,indent+2);
+ fprintf(fp,"phi =");
+ for(j=0; (j<cos->n); j++)
+ fprintf(fp," %e",cos->phi[j]);
+ fprintf(fp,"\n");
+ }
+ }
+}
+
+#define PS(t,s) pr_str(fp,indent,t,s)
+#define PI(t,s) pr_int(fp,indent,t,s)
+#define PSTEP(t,s) pr_gmx_large_int(fp,indent,t,s)
+#define PR(t,s) pr_real(fp,indent,t,s)
+#define PD(t,s) pr_double(fp,indent,t,s)
+
+static void pr_pullgrp(FILE *fp,int indent,int g,t_pullgrp *pg)
+{
+ pr_indent(fp,indent);
+ fprintf(fp,"pull-group %d:\n",g);
+ indent += 2;
+ pr_ivec_block(fp,indent,"atom",pg->ind,pg->nat,TRUE);
+ pr_rvec(fp,indent,"weight",pg->weight,pg->nweight,TRUE);
+ PI("pbcatom",pg->pbcatom);
+ pr_rvec(fp,indent,"vec",pg->vec,DIM,TRUE);
+ pr_rvec(fp,indent,"init",pg->init,DIM,TRUE);
+ PR("rate",pg->rate);
+ PR("k",pg->k);
+ PR("kB",pg->kB);
+}
+
+static void pr_pull(FILE *fp,int indent,t_pull *pull)
+{
+ int g;
+
+ PS("pull-geometry",EPULLGEOM(pull->eGeom));
+ pr_ivec(fp,indent,"pull-dim",pull->dim,DIM,TRUE);
+ PR("pull-r1",pull->cyl_r1);
+ PR("pull-r0",pull->cyl_r0);
+ PR("pull-constr-tol",pull->constr_tol);
+ PI("pull-nstxout",pull->nstxout);
+ PI("pull-nstfout",pull->nstfout);
+ PI("pull-ngrp",pull->ngrp);
+ for(g=0; g<pull->ngrp+1; g++)
+ pr_pullgrp(fp,indent,g,&pull->grp[g]);
+}
+
+static void pr_rotgrp(FILE *fp,int indent,int g,t_rotgrp *rotg)
+{
+ pr_indent(fp,indent);
+ fprintf(fp,"rotation_group %d:\n",g);
+ indent += 2;
+ PS("type",EROTGEOM(rotg->eType));
+ PS("massw",BOOL(rotg->bMassW));
+ pr_ivec_block(fp,indent,"atom",rotg->ind,rotg->nat,TRUE);
+ pr_rvecs(fp,indent,"x_ref",rotg->x_ref,rotg->nat);
+ pr_rvec(fp,indent,"vec",rotg->vec,DIM,TRUE);
+ pr_rvec(fp,indent,"pivot",rotg->pivot,DIM,TRUE);
+ PR("rate",rotg->rate);
+ PR("k",rotg->k);
+ PR("slab_dist",rotg->slab_dist);
+ PR("min_gaussian",rotg->min_gaussian);
+ PR("epsilon",rotg->eps);
+ PS("fit_method",EROTFIT(rotg->eFittype));
+ PI("potfitangle_nstep",rotg->PotAngle_nstep);
+ PR("potfitangle_step",rotg->PotAngle_step);
+}
+
+static void pr_rot(FILE *fp,int indent,t_rot *rot)
+{
+ int g;
+
+ PI("rot_nstrout",rot->nstrout);
+ PI("rot_nstsout",rot->nstsout);
+ PI("rot_ngrp",rot->ngrp);
+ for(g=0; g<rot->ngrp; g++)
+ pr_rotgrp(fp,indent,g,&rot->grp[g]);
+}
+
+void pr_inputrec(FILE *fp,int indent,const char *title,t_inputrec *ir,
+ gmx_bool bMDPformat)
+{
+ const char *infbuf="inf";
+ int i;
+
+ if (available(fp,ir,indent,title)) {
+ if (!bMDPformat)
+ indent=pr_title(fp,indent,title);
+ PS("integrator",EI(ir->eI));
+ PSTEP("nsteps",ir->nsteps);
+ PSTEP("init-step",ir->init_step);
+ PS("ns-type",ENS(ir->ns_type));
+ PI("nstlist",ir->nstlist);
+ PI("ndelta",ir->ndelta);
+ PI("nstcomm",ir->nstcomm);
+ PS("comm-mode",ECOM(ir->comm_mode));
+ PI("nstlog",ir->nstlog);
+ PI("nstxout",ir->nstxout);
+ PI("nstvout",ir->nstvout);
+ PI("nstfout",ir->nstfout);
+ PI("nstcalcenergy",ir->nstcalcenergy);
+ PI("nstenergy",ir->nstenergy);
+ PI("nstxtcout",ir->nstxtcout);
+ PR("init-t",ir->init_t);
+ PR("delta-t",ir->delta_t);
+
+ PR("xtcprec",ir->xtcprec);
+ PI("nkx",ir->nkx);
+ PI("nky",ir->nky);
+ PI("nkz",ir->nkz);
+ PI("pme-order",ir->pme_order);
+ PR("ewald-rtol",ir->ewald_rtol);
+ PR("ewald-geometry",ir->ewald_geometry);
+ PR("epsilon-surface",ir->epsilon_surface);
+ PS("optimize-fft",BOOL(ir->bOptFFT));
+ PS("ePBC",EPBC(ir->ePBC));
+ PS("bPeriodicMols",BOOL(ir->bPeriodicMols));
+ PS("bContinuation",BOOL(ir->bContinuation));
+ PS("bShakeSOR",BOOL(ir->bShakeSOR));
+ PS("etc",ETCOUPLTYPE(ir->etc));
+ PI("nsttcouple",ir->nsttcouple);
+ PS("epc",EPCOUPLTYPE(ir->epc));
+ PS("epctype",EPCOUPLTYPETYPE(ir->epct));
+ PI("nstpcouple",ir->nstpcouple);
+ PR("tau-p",ir->tau_p);
+ pr_matrix(fp,indent,"ref-p",ir->ref_p,bMDPformat);
+ pr_matrix(fp,indent,"compress",ir->compress,bMDPformat);
+ PS("refcoord-scaling",EREFSCALINGTYPE(ir->refcoord_scaling));
+ if (bMDPformat)
+ fprintf(fp,"posres-com = %g %g %g\n",ir->posres_com[XX],
+ ir->posres_com[YY],ir->posres_com[ZZ]);
+ else
+ pr_rvec(fp,indent,"posres-com",ir->posres_com,DIM,TRUE);
+ if (bMDPformat)
+ fprintf(fp,"posres-comB = %g %g %g\n",ir->posres_comB[XX],
+ ir->posres_comB[YY],ir->posres_comB[ZZ]);
+ else
+ pr_rvec(fp,indent,"posres-comB",ir->posres_comB,DIM,TRUE);
+ PI("andersen-seed",ir->andersen_seed);
+ PR("rlist",ir->rlist);
+ PR("rlistlong",ir->rlistlong);
+ PR("rtpi",ir->rtpi);
+ PS("coulombtype",EELTYPE(ir->coulombtype));
+ PR("rcoulomb-switch",ir->rcoulomb_switch);
+ PR("rcoulomb",ir->rcoulomb);
+ PS("vdwtype",EVDWTYPE(ir->vdwtype));
+ PR("rvdw-switch",ir->rvdw_switch);
+ PR("rvdw",ir->rvdw);
+ if (ir->epsilon_r != 0)
+ PR("epsilon-r",ir->epsilon_r);
+ else
+ PS("epsilon-r",infbuf);
+ if (ir->epsilon_rf != 0)
+ PR("epsilon-rf",ir->epsilon_rf);
+ else
+ PS("epsilon-rf",infbuf);
+ PR("tabext",ir->tabext);
+ PS("implicit-solvent",EIMPLICITSOL(ir->implicit_solvent));
+ PS("gb-algorithm",EGBALGORITHM(ir->gb_algorithm));
+ PR("gb-epsilon-solvent",ir->gb_epsilon_solvent);
+ PI("nstgbradii",ir->nstgbradii);
+ PR("rgbradii",ir->rgbradii);
+ PR("gb-saltconc",ir->gb_saltconc);
+ PR("gb-obc-alpha",ir->gb_obc_alpha);
+ PR("gb-obc-beta",ir->gb_obc_beta);
+ PR("gb-obc-gamma",ir->gb_obc_gamma);
+ PR("gb-dielectric-offset",ir->gb_dielectric_offset);
+ PS("sa-algorithm",ESAALGORITHM(ir->gb_algorithm));
+ PR("sa-surface-tension",ir->sa_surface_tension);
+
+ PS("DispCorr",EDISPCORR(ir->eDispCorr));
+ PS("free-energy",EFEPTYPE(ir->efep));
+ PR("init-lambda",ir->init_lambda);
+ PR("delta-lambda",ir->delta_lambda);
+ if (!bMDPformat)
+ {
+ PI("n-foreign-lambda",ir->n_flambda);
+ }
+ if (ir->n_flambda > 0)
+ {
+ pr_indent(fp,indent);
+ fprintf(fp,"foreign-lambda%s",bMDPformat ? " = " : ":");
+ for(i=0; i<ir->n_flambda; i++)
+ {
+ fprintf(fp," %10g",ir->flambda[i]);
+ }
+ fprintf(fp,"\n");
+ }
+ PR("sc-alpha",ir->sc_alpha);
+ PI("sc-power",ir->sc_power);
+ PR("sc-sigma",ir->sc_sigma);
+ PR("sc-sigma-min",ir->sc_sigma_min);
+ PI("nstdhdl", ir->nstdhdl);
+ PS("separate-dhdl-file", SEPDHDLFILETYPE(ir->separate_dhdl_file));
+ PS("dhdl-derivatives", DHDLDERIVATIVESTYPE(ir->dhdl_derivatives));
+ PI("dh-hist-size", ir->dh_hist_size);
+ PD("dh-hist-spacing", ir->dh_hist_spacing);
+
+ PI("nwall",ir->nwall);
+ PS("wall-type",EWALLTYPE(ir->wall_type));
+ PI("wall-atomtype[0]",ir->wall_atomtype[0]);
+ PI("wall-atomtype[1]",ir->wall_atomtype[1]);
+ PR("wall-density[0]",ir->wall_density[0]);
+ PR("wall-density[1]",ir->wall_density[1]);
+ PR("wall-ewald-zfac",ir->wall_ewald_zfac);
+
+ PS("pull",EPULLTYPE(ir->ePull));
+ if (ir->ePull != epullNO)
+ pr_pull(fp,indent,ir->pull);
+
+ PS("rotation",BOOL(ir->bRot));
+ if (ir->bRot)
+ pr_rot(fp,indent,ir->rot);
+
+ PS("disre",EDISRETYPE(ir->eDisre));
+ PS("disre-weighting",EDISREWEIGHTING(ir->eDisreWeighting));
+ PS("disre-mixed",BOOL(ir->bDisreMixed));
+ PR("dr-fc",ir->dr_fc);
+ PR("dr-tau",ir->dr_tau);
+ PR("nstdisreout",ir->nstdisreout);
+ PR("orires-fc",ir->orires_fc);
+ PR("orires-tau",ir->orires_tau);
+ PR("nstorireout",ir->nstorireout);
+
+ PR("dihre-fc",ir->dihre_fc);
+
+ PR("em-stepsize",ir->em_stepsize);
+ PR("em-tol",ir->em_tol);
+ PI("niter",ir->niter);
+ PR("fc-stepsize",ir->fc_stepsize);
+ PI("nstcgsteep",ir->nstcgsteep);
+ PI("nbfgscorr",ir->nbfgscorr);
+
+ PS("ConstAlg",ECONSTRTYPE(ir->eConstrAlg));
+ PR("shake-tol",ir->shake_tol);
+ PI("lincs-order",ir->nProjOrder);
+ PR("lincs-warnangle",ir->LincsWarnAngle);
+ PI("lincs-iter",ir->nLincsIter);
+ PR("bd-fric",ir->bd_fric);
+ PI("ld-seed",ir->ld_seed);
+ PR("cos-accel",ir->cos_accel);
+ pr_matrix(fp,indent,"deform",ir->deform,bMDPformat);
+
+ PS("adress",BOOL(ir->bAdress));
+ if (ir->bAdress){
+ PS("adress_type",EADRESSTYPE(ir->adress->type));
+ PR("adress_const_wf",ir->adress->const_wf);
+ PR("adress_ex_width",ir->adress->ex_width);
+ PR("adress_hy_width",ir->adress->hy_width);
+ PS("adress_interface_correction",EADRESSICTYPE(ir->adress->icor));
+ PS("adress_site",EADRESSSITETYPE(ir->adress->site));
+ PR("adress_ex_force_cap",ir->adress->ex_forcecap);
+ PS("adress_do_hybridpairs", BOOL(ir->adress->do_hybridpairs));
+
+ pr_rvec(fp,indent,"adress_reference_coords",ir->adress->refs,DIM,TRUE);
+ }
+ PI("userint1",ir->userint1);
+ PI("userint2",ir->userint2);
+ PI("userint3",ir->userint3);
+ PI("userint4",ir->userint4);
+ PR("userreal1",ir->userreal1);
+ PR("userreal2",ir->userreal2);
+ PR("userreal3",ir->userreal3);
+ PR("userreal4",ir->userreal4);
+ pr_grp_opts(fp,indent,"grpopts",&(ir->opts),bMDPformat);
+ pr_cosine(fp,indent,"efield-x",&(ir->ex[XX]),bMDPformat);
+ pr_cosine(fp,indent,"efield-xt",&(ir->et[XX]),bMDPformat);
+ pr_cosine(fp,indent,"efield-y",&(ir->ex[YY]),bMDPformat);
+ pr_cosine(fp,indent,"efield-yt",&(ir->et[YY]),bMDPformat);
+ pr_cosine(fp,indent,"efield-z",&(ir->ex[ZZ]),bMDPformat);
+ pr_cosine(fp,indent,"efield-zt",&(ir->et[ZZ]),bMDPformat);
+ PS("bQMMM",BOOL(ir->bQMMM));
+ PI("QMconstraints",ir->QMconstraints);
+ PI("QMMMscheme",ir->QMMMscheme);
+ PR("scalefactor",ir->scalefactor);
+ pr_qm_opts(fp,indent,"qm-opts",&(ir->opts));
+ }
+}
+#undef PS
+#undef PR
+#undef PI
+
+static void pr_harm(FILE *fp,t_iparams *iparams,const char *r,const char *kr)
+{
+ fprintf(fp,"%sA=%12.5e, %sA=%12.5e, %sB=%12.5e, %sB=%12.5e\n",
+ r,iparams->harmonic.rA,kr,iparams->harmonic.krA,
+ r,iparams->harmonic.rB,kr,iparams->harmonic.krB);
+}
+
+void pr_iparams(FILE *fp,t_functype ftype,t_iparams *iparams)
+{
+ int i;
+ real VA[4],VB[4],*rbcA,*rbcB;
+
+ switch (ftype) {
+ case F_ANGLES:
+ case F_G96ANGLES:
+ pr_harm(fp,iparams,"th","ct");
+ break;
+ case F_CROSS_BOND_BONDS:
+ fprintf(fp,"r1e=%15.8e, r2e=%15.8e, krr=%15.8e\n",
+ iparams->cross_bb.r1e,iparams->cross_bb.r2e,
+ iparams->cross_bb.krr);
+ break;
+ case F_CROSS_BOND_ANGLES:
+ fprintf(fp,"r1e=%15.8e, r1e=%15.8e, r3e=%15.8e, krt=%15.8e\n",
+ iparams->cross_ba.r1e,iparams->cross_ba.r2e,
+ iparams->cross_ba.r3e,iparams->cross_ba.krt);
+ break;
+ case F_LINEAR_ANGLES:
+ fprintf(fp,"klinA=%15.8e, aA=%15.8e, klinB=%15.8e, aB=%15.8e\n",
+ iparams->linangle.klinA,iparams->linangle.aA,
+ iparams->linangle.klinB,iparams->linangle.aB);
+ break;
+ case F_UREY_BRADLEY:
+ fprintf(fp,"theta=%15.8e, ktheta=%15.8e, r13=%15.8e, kUB=%15.8e\n",
+ iparams->u_b.theta,iparams->u_b.ktheta,iparams->u_b.r13,iparams->u_b.kUB);
+ break;
+ case F_QUARTIC_ANGLES:
+ fprintf(fp,"theta=%15.8e",iparams->qangle.theta);
+ for(i=0; i<5; i++)
+ fprintf(fp,", c%c=%15.8e",'0'+i,iparams->qangle.c[i]);
+ fprintf(fp,"\n");
+ break;
+ case F_BHAM:
+ fprintf(fp,"a=%15.8e, b=%15.8e, c=%15.8e\n",
+ iparams->bham.a,iparams->bham.b,iparams->bham.c);
+ break;
+ case F_BONDS:
+ case F_G96BONDS:
+ case F_HARMONIC:
+ pr_harm(fp,iparams,"b0","cb");
+ break;
+ case F_IDIHS:
+ pr_harm(fp,iparams,"xi","cx");
+ break;
+ case F_MORSE:
+ fprintf(fp,"b0=%15.8e, cb=%15.8e, beta=%15.8e\n",
+ iparams->morse.b0,iparams->morse.cb,iparams->morse.beta);
+ break;
+ case F_CUBICBONDS:
+ fprintf(fp,"b0=%15.8e, kb=%15.8e, kcub=%15.8e\n",
+ iparams->cubic.b0,iparams->cubic.kb,iparams->cubic.kcub);
+ break;
+ case F_CONNBONDS:
+ fprintf(fp,"\n");
+ break;
+ case F_FENEBONDS:
+ fprintf(fp,"bm=%15.8e, kb=%15.8e\n",iparams->fene.bm,iparams->fene.kb);
+ break;
+ case F_RESTRBONDS:
+ fprintf(fp,"lowA=%15.8e, up1A=%15.8e, up2A=%15.8e, kA=%15.8e, lowB=%15.8e, up1B=%15.8e, up2B=%15.8e, kB=%15.8e,\n",
+ iparams->restraint.lowA,iparams->restraint.up1A,
+ iparams->restraint.up2A,iparams->restraint.kA,
+ iparams->restraint.lowB,iparams->restraint.up1B,
+ iparams->restraint.up2B,iparams->restraint.kB);
+ break;
+ case F_TABBONDS:
+ case F_TABBONDSNC:
+ case F_TABANGLES:
+ case F_TABDIHS:
+ fprintf(fp,"tab=%d, kA=%15.8e, kB=%15.8e\n",
+ iparams->tab.table,iparams->tab.kA,iparams->tab.kB);
+ break;
+ case F_POLARIZATION:
+ fprintf(fp,"alpha=%15.8e\n",iparams->polarize.alpha);
+ break;
+ case F_ANHARM_POL:
+ fprintf(fp,"alpha=%15.8e drcut=%15.8e khyp=%15.8e\n",
+ iparams->anharm_polarize.alpha,
+ iparams->anharm_polarize.drcut,
+ iparams->anharm_polarize.khyp);
+ break;
+ case F_THOLE_POL:
+ fprintf(fp,"a=%15.8e, alpha1=%15.8e, alpha2=%15.8e, rfac=%15.8e\n",
+ iparams->thole.a,iparams->thole.alpha1,iparams->thole.alpha2,
+ iparams->thole.rfac);
+ break;
+ case F_WATER_POL:
+ fprintf(fp,"al_x=%15.8e, al_y=%15.8e, al_z=%15.8e, rOH=%9.6f, rHH=%9.6f, rOD=%9.6f\n",
+ iparams->wpol.al_x,iparams->wpol.al_y,iparams->wpol.al_z,
+ iparams->wpol.rOH,iparams->wpol.rHH,iparams->wpol.rOD);
+ break;
+ case F_LJ:
+ fprintf(fp,"c6=%15.8e, c12=%15.8e\n",iparams->lj.c6,iparams->lj.c12);
+ break;
+ case F_LJ14:
+ fprintf(fp,"c6A=%15.8e, c12A=%15.8e, c6B=%15.8e, c12B=%15.8e\n",
+ iparams->lj14.c6A,iparams->lj14.c12A,
+ iparams->lj14.c6B,iparams->lj14.c12B);
+ break;
+ case F_LJC14_Q:
+ fprintf(fp,"fqq=%15.8e, qi=%15.8e, qj=%15.8e, c6=%15.8e, c12=%15.8e\n",
+ iparams->ljc14.fqq,
+ iparams->ljc14.qi,iparams->ljc14.qj,
+ iparams->ljc14.c6,iparams->ljc14.c12);
+ break;
+ case F_LJC_PAIRS_NB:
+ fprintf(fp,"qi=%15.8e, qj=%15.8e, c6=%15.8e, c12=%15.8e\n",
+ iparams->ljcnb.qi,iparams->ljcnb.qj,
+ iparams->ljcnb.c6,iparams->ljcnb.c12);
+ break;
+ case F_PDIHS:
+ case F_PIDIHS:
+ case F_ANGRES:
+ case F_ANGRESZ:
+ fprintf(fp,"phiA=%15.8e, cpA=%15.8e, phiB=%15.8e, cpB=%15.8e, mult=%d\n",
+ iparams->pdihs.phiA,iparams->pdihs.cpA,
+ iparams->pdihs.phiB,iparams->pdihs.cpB,
+ iparams->pdihs.mult);
+ break;
+ case F_DISRES:
+ fprintf(fp,"label=%4d, type=%1d, low=%15.8e, up1=%15.8e, up2=%15.8e, fac=%15.8e)\n",
+ iparams->disres.label,iparams->disres.type,
+ iparams->disres.low,iparams->disres.up1,
+ iparams->disres.up2,iparams->disres.kfac);
+ break;
+ case F_ORIRES:
+ fprintf(fp,"ex=%4d, label=%d, power=%4d, c=%15.8e, obs=%15.8e, kfac=%15.8e)\n",
+ iparams->orires.ex,iparams->orires.label,iparams->orires.power,
+ iparams->orires.c,iparams->orires.obs,iparams->orires.kfac);
+ break;
+ case F_DIHRES:
+ fprintf(fp,"label=%d, power=%4d phi=%15.8e, dphi=%15.8e, kfac=%15.8e)\n",
+ iparams->dihres.label,iparams->dihres.power,
+ iparams->dihres.phi,iparams->dihres.dphi,iparams->dihres.kfac);
+ break;
+ case F_POSRES:
+ fprintf(fp,"pos0A=(%15.8e,%15.8e,%15.8e), fcA=(%15.8e,%15.8e,%15.8e), pos0B=(%15.8e,%15.8e,%15.8e), fcB=(%15.8e,%15.8e,%15.8e)\n",
+ iparams->posres.pos0A[XX],iparams->posres.pos0A[YY],
+ iparams->posres.pos0A[ZZ],iparams->posres.fcA[XX],
+ iparams->posres.fcA[YY],iparams->posres.fcA[ZZ],
+ iparams->posres.pos0B[XX],iparams->posres.pos0B[YY],
+ iparams->posres.pos0B[ZZ],iparams->posres.fcB[XX],
+ iparams->posres.fcB[YY],iparams->posres.fcB[ZZ]);
+ break;
+ case F_RBDIHS:
+ for (i=0; i<NR_RBDIHS; i++)
+ fprintf(fp,"%srbcA[%d]=%15.8e",i==0?"":", ",i,iparams->rbdihs.rbcA[i]);
+ fprintf(fp,"\n");
+ for (i=0; i<NR_RBDIHS; i++)
+ fprintf(fp,"%srbcB[%d]=%15.8e",i==0?"":", ",i,iparams->rbdihs.rbcB[i]);
+ fprintf(fp,"\n");
+ break;
+ case F_FOURDIHS:
+ /* Use the OPLS -> Ryckaert-Bellemans formula backwards to get the
+ * OPLS potential constants back.
+ */
+ rbcA = iparams->rbdihs.rbcA;
+ rbcB = iparams->rbdihs.rbcB;
+
+ VA[3] = -0.25*rbcA[4];
+ VA[2] = -0.5*rbcA[3];
+ VA[1] = 4.0*VA[3]-rbcA[2];
+ VA[0] = 3.0*VA[2]-2.0*rbcA[1];
+
+ VB[3] = -0.25*rbcB[4];
+ VB[2] = -0.5*rbcB[3];
+ VB[1] = 4.0*VB[3]-rbcB[2];
+ VB[0] = 3.0*VB[2]-2.0*rbcB[1];
+
+ for (i=0; i<NR_FOURDIHS; i++)
+ fprintf(fp,"%sFourA[%d]=%15.8e",i==0?"":", ",i,VA[i]);
+ fprintf(fp,"\n");
+ for (i=0; i<NR_FOURDIHS; i++)
+ fprintf(fp,"%sFourB[%d]=%15.8e",i==0?"":", ",i,VB[i]);
+ fprintf(fp,"\n");
+ break;
+
+ case F_CONSTR:
+ case F_CONSTRNC:
+ fprintf(fp,"dA=%15.8e, dB=%15.8e\n",iparams->constr.dA,iparams->constr.dB);
+ break;
+ case F_SETTLE:
+ fprintf(fp,"doh=%15.8e, dhh=%15.8e\n",iparams->settle.doh,
+ iparams->settle.dhh);
+ break;
+ case F_VSITE2:
+ fprintf(fp,"a=%15.8e\n",iparams->vsite.a);
+ break;
+ case F_VSITE3:
+ case F_VSITE3FD:
+ case F_VSITE3FAD:
+ fprintf(fp,"a=%15.8e, b=%15.8e\n",iparams->vsite.a,iparams->vsite.b);
+ break;
+ case F_VSITE3OUT:
+ case F_VSITE4FD:
+ case F_VSITE4FDN:
+ fprintf(fp,"a=%15.8e, b=%15.8e, c=%15.8e\n",
+ iparams->vsite.a,iparams->vsite.b,iparams->vsite.c);
+ break;
+ case F_VSITEN:
+ fprintf(fp,"n=%2d, a=%15.8e\n",iparams->vsiten.n,iparams->vsiten.a);
+ break;
+ case F_GB12:
+ case F_GB13:
+ case F_GB14:
+ fprintf(fp, "sar=%15.8e, st=%15.8e, pi=%15.8e, gbr=%15.8e, bmlt=%15.8e\n",iparams->gb.sar,iparams->gb.st,iparams->gb.pi,iparams->gb.gbr,iparams->gb.bmlt);
+ break;
+ case F_CMAP:
+ fprintf(fp, "cmapA=%1d, cmapB=%1d\n",iparams->cmap.cmapA, iparams->cmap.cmapB);
+ break;
+ default:
+ gmx_fatal(FARGS,"unknown function type %d (%s) in %s line %d",
+ ftype,interaction_function[ftype].name,__FILE__,__LINE__);
+ }
+}
+
+void pr_ilist(FILE *fp,int indent,const char *title,
+ t_functype *functype,t_ilist *ilist, gmx_bool bShowNumbers)
+{
+ int i,j,k,type,ftype;
+ t_iatom *iatoms;
+
+ if (available(fp,ilist,indent,title) && ilist->nr > 0)
+ {
+ indent=pr_title(fp,indent,title);
+ (void) pr_indent(fp,indent);
+ fprintf(fp,"nr: %d\n",ilist->nr);
+ if (ilist->nr > 0) {
+ (void) pr_indent(fp,indent);
+ fprintf(fp,"iatoms:\n");
+ iatoms=ilist->iatoms;
+ for (i=j=0; i<ilist->nr;) {
+#ifndef DEBUG
+ (void) pr_indent(fp,indent+INDENT);
+ type=*(iatoms++);
+ ftype=functype[type];
+ (void) fprintf(fp,"%d type=%d (%s)",
+ bShowNumbers?j:-1,bShowNumbers?type:-1,
+ interaction_function[ftype].name);
+ j++;
+ for (k=0; k<interaction_function[ftype].nratoms; k++)
+ (void) fprintf(fp," %u",*(iatoms++));
+ (void) fprintf(fp,"\n");
+ i+=1+interaction_function[ftype].nratoms;
+#else
+ fprintf(fp,"%5d%5d\n",i,iatoms[i]);
+ i++;
+#endif
+ }
+ }
+ }
+}
+
+static void pr_cmap(FILE *fp, int indent, const char *title,
+ gmx_cmap_t *cmap_grid, gmx_bool bShowNumbers)
+{
+ int i,j,nelem;
+ real dx,idx;
+
+ dx = 360.0 / cmap_grid->grid_spacing;
+ nelem = cmap_grid->grid_spacing*cmap_grid->grid_spacing;
+
+ if(available(fp,cmap_grid,indent,title))
+ {
+ fprintf(fp,"%s\n",title);
+
+ for(i=0;i<cmap_grid->ngrid;i++)
+ {
+ idx = -180.0;
+ fprintf(fp,"%8s %8s %8s %8s\n","V","dVdx","dVdy","d2dV");
+
+ fprintf(fp,"grid[%3d]={\n",bShowNumbers?i:-1);
+
+ for(j=0;j<nelem;j++)
+ {
+ if( (j%cmap_grid->grid_spacing)==0)
+ {
+ fprintf(fp,"%8.1f\n",idx);
+ idx+=dx;
+ }
+
+ fprintf(fp,"%8.3f ",cmap_grid->cmapdata[i].cmap[j*4]);
+ fprintf(fp,"%8.3f ",cmap_grid->cmapdata[i].cmap[j*4+1]);
+ fprintf(fp,"%8.3f ",cmap_grid->cmapdata[i].cmap[j*4+2]);
+ fprintf(fp,"%8.3f\n",cmap_grid->cmapdata[i].cmap[j*4+3]);
+ }
+ fprintf(fp,"\n");
+ }
+ }
+
+}
+
+void pr_ffparams(FILE *fp,int indent,const char *title,
+ gmx_ffparams_t *ffparams,
+ gmx_bool bShowNumbers)
+{
+ int i,j;
+
+ indent=pr_title(fp,indent,title);
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"atnr=%d\n",ffparams->atnr);
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"ntypes=%d\n",ffparams->ntypes);
+ for (i=0; i<ffparams->ntypes; i++) {
+ (void) pr_indent(fp,indent+INDENT);
+ (void) fprintf(fp,"functype[%d]=%s, ",
+ bShowNumbers?i:-1,
+ interaction_function[ffparams->functype[i]].name);
+ pr_iparams(fp,ffparams->functype[i],&ffparams->iparams[i]);
+ }
+ (void) pr_double(fp,indent,"reppow",ffparams->reppow);
+ (void) pr_real(fp,indent,"fudgeQQ",ffparams->fudgeQQ);
+ pr_cmap(fp,indent,"cmap",&ffparams->cmap_grid,bShowNumbers);
+}
+
+void pr_idef(FILE *fp,int indent,const char *title,t_idef *idef, gmx_bool bShowNumbers)
+{
+ int i,j;
+
+ if (available(fp,idef,indent,title)) {
+ indent=pr_title(fp,indent,title);
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"atnr=%d\n",idef->atnr);
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"ntypes=%d\n",idef->ntypes);
+ for (i=0; i<idef->ntypes; i++) {
+ (void) pr_indent(fp,indent+INDENT);
+ (void) fprintf(fp,"functype[%d]=%s, ",
+ bShowNumbers?i:-1,
+ interaction_function[idef->functype[i]].name);
+ pr_iparams(fp,idef->functype[i],&idef->iparams[i]);
+ }
+ (void) pr_real(fp,indent,"fudgeQQ",idef->fudgeQQ);
+
+ for(j=0; (j<F_NRE); j++)
+ pr_ilist(fp,indent,interaction_function[j].longname,
+ idef->functype,&idef->il[j],bShowNumbers);
+ }
+}
+
+static int pr_block_title(FILE *fp,int indent,const char *title,t_block *block)
+{
+ int i;
+
+ if (available(fp,block,indent,title))
+ {
+ indent=pr_title(fp,indent,title);
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"nr=%d\n",block->nr);
+ }
+ return indent;
+}
+
+static int pr_blocka_title(FILE *fp,int indent,const char *title,t_blocka *block)
+{
+ int i;
+
+ if (available(fp,block,indent,title))
+ {
+ indent=pr_title(fp,indent,title);
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"nr=%d\n",block->nr);
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"nra=%d\n",block->nra);
+ }
+ return indent;
+}
+
+static void low_pr_block(FILE *fp,int indent,const char *title,t_block *block, gmx_bool bShowNumbers)
+{
+ int i;
+
+ if (available(fp,block,indent,title))
+ {
+ indent=pr_block_title(fp,indent,title,block);
+ for (i=0; i<=block->nr; i++)
+ {
+ (void) pr_indent(fp,indent+INDENT);
+ (void) fprintf(fp,"%s->index[%d]=%u\n",
+ title,bShowNumbers?i:-1,block->index[i]);
+ }
+ }
+}
+
+static void low_pr_blocka(FILE *fp,int indent,const char *title,t_blocka *block, gmx_bool bShowNumbers)
+{
+ int i;
+
+ if (available(fp,block,indent,title))
+ {
+ indent=pr_blocka_title(fp,indent,title,block);
+ for (i=0; i<=block->nr; i++)
+ {
+ (void) pr_indent(fp,indent+INDENT);
+ (void) fprintf(fp,"%s->index[%d]=%u\n",
+ title,bShowNumbers?i:-1,block->index[i]);
+ }
+ for (i=0; i<block->nra; i++)
+ {
+ (void) pr_indent(fp,indent+INDENT);
+ (void) fprintf(fp,"%s->a[%d]=%u\n",
+ title,bShowNumbers?i:-1,block->a[i]);
+ }
+ }
+}
+
+void pr_block(FILE *fp,int indent,const char *title,t_block *block,gmx_bool bShowNumbers)
+{
+ int i,j,ok,size,start,end;
+
+ if (available(fp,block,indent,title))
+ {
+ indent=pr_block_title(fp,indent,title,block);
+ start=0;
+ end=start;
+ if ((ok=(block->index[start]==0))==0)
+ (void) fprintf(fp,"block->index[%d] should be 0\n",start);
+ else
+ for (i=0; i<block->nr; i++)
+ {
+ end=block->index[i+1];
+ size=pr_indent(fp,indent);
+ if (end<=start)
+ size+=fprintf(fp,"%s[%d]={}\n",title,i);
+ else
+ size+=fprintf(fp,"%s[%d]={%d..%d}\n",
+ title,bShowNumbers?i:-1,
+ bShowNumbers?start:-1,bShowNumbers?end-1:-1);
+ start=end;
+ }
+ }
+}
+
+void pr_blocka(FILE *fp,int indent,const char *title,t_blocka *block,gmx_bool bShowNumbers)
+{
+ int i,j,ok,size,start,end;
+
+ if (available(fp,block,indent,title))
+ {
+ indent=pr_blocka_title(fp,indent,title,block);
+ start=0;
+ end=start;
+ if ((ok=(block->index[start]==0))==0)
+ (void) fprintf(fp,"block->index[%d] should be 0\n",start);
+ else
+ for (i=0; i<block->nr; i++)
+ {
+ end=block->index[i+1];
+ size=pr_indent(fp,indent);
+ if (end<=start)
+ size+=fprintf(fp,"%s[%d]={",title,i);
+ else
+ size+=fprintf(fp,"%s[%d][%d..%d]={",
+ title,bShowNumbers?i:-1,
+ bShowNumbers?start:-1,bShowNumbers?end-1:-1);
+ for (j=start; j<end; j++)
+ {
+ if (j>start) size+=fprintf(fp,", ");
+ if ((size)>(USE_WIDTH))
+ {
+ (void) fprintf(fp,"\n");
+ size=pr_indent(fp,indent+INDENT);
+ }
+ size+=fprintf(fp,"%u",block->a[j]);
+ }
+ (void) fprintf(fp,"}\n");
+ start=end;
+ }
+ if ((end!=block->nra)||(!ok))
+ {
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"tables inconsistent, dumping complete tables:\n");
+ low_pr_blocka(fp,indent,title,block,bShowNumbers);
+ }
+ }
+}
+
+static void pr_strings(FILE *fp,int indent,const char *title,char ***nm,int n, gmx_bool bShowNumbers)
+{
+ int i;
+
+ if (available(fp,nm,indent,title))
+ {
+ indent=pr_title_n(fp,indent,title,n);
+ for (i=0; i<n; i++)
+ {
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"%s[%d]={name=\"%s\"}\n",
+ title,bShowNumbers?i:-1,*(nm[i]));
+ }
+ }
+}
+
+static void pr_strings2(FILE *fp,int indent,const char *title,
+ char ***nm,char ***nmB,int n, gmx_bool bShowNumbers)
+{
+ int i;
+
+ if (available(fp,nm,indent,title))
+ {
+ indent=pr_title_n(fp,indent,title,n);
+ for (i=0; i<n; i++)
+ {
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"%s[%d]={name=\"%s\",nameB=\"%s\"}\n",
+ title,bShowNumbers?i:-1,*(nm[i]),*(nmB[i]));
+ }
+ }
+}
+
+static void pr_resinfo(FILE *fp,int indent,const char *title,t_resinfo *resinfo,int n, gmx_bool bShowNumbers)
+{
+ int i;
+
+ if (available(fp,resinfo,indent,title))
+ {
+ indent=pr_title_n(fp,indent,title,n);
+ for (i=0; i<n; i++)
+ {
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"%s[%d]={name=\"%s\", nr=%d, ic='%c'}\n",
+ title,bShowNumbers?i:-1,
+ *(resinfo[i].name),resinfo[i].nr,
+ (resinfo[i].ic == '\0') ? ' ' : resinfo[i].ic);
+ }
+ }
+}
+
+static void pr_atom(FILE *fp,int indent,const char *title,t_atom *atom,int n)
+{
+ int i,j;
+
+ if (available(fp,atom,indent,title)) {
+ indent=pr_title_n(fp,indent,title,n);
+ for (i=0; i<n; i++) {
+ (void) pr_indent(fp,indent);
+ fprintf(fp,"%s[%6d]={type=%3d, typeB=%3d, ptype=%8s, m=%12.5e, "
+ "q=%12.5e, mB=%12.5e, qB=%12.5e, resind=%5d, atomnumber=%3d}\n",
+ title,i,atom[i].type,atom[i].typeB,ptype_str[atom[i].ptype],
+ atom[i].m,atom[i].q,atom[i].mB,atom[i].qB,
+ atom[i].resind,atom[i].atomnumber);
+ }
+ }
+}
+
+static void pr_grps(FILE *fp,int indent,const char *title,t_grps grps[],
+ char **grpname[], gmx_bool bShowNumbers)
+{
+ int i,j;
+
+ for(i=0; (i<egcNR); i++)
+ {
+ fprintf(fp,"%s[%-12s] nr=%d, name=[",title,gtypes[i],grps[i].nr);
+ for(j=0; (j<grps[i].nr); j++)
+ {
+ fprintf(fp," %s",*(grpname[grps[i].nm_ind[j]]));
+ }
+ fprintf(fp,"]\n");
+ }
+}
+
+static void pr_groups(FILE *fp,int indent,const char *title,
+ gmx_groups_t *groups,
+ gmx_bool bShowNumbers)
+{
+ int grpnr[egcNR];
+ int nat_max,i,g;
+
+ pr_grps(fp,indent,"grp",groups->grps,groups->grpname,bShowNumbers);
+ pr_strings(fp,indent,"grpname",groups->grpname,groups->ngrpname,bShowNumbers);
+
+ (void) pr_indent(fp,indent);
+ fprintf(fp,"groups ");
+ for(g=0; g<egcNR; g++)
+ {
+ printf(" %5.5s",gtypes[g]);
+ }
+ printf("\n");
+
+ (void) pr_indent(fp,indent);
+ fprintf(fp,"allocated ");
+ nat_max = 0;
+ for(g=0; g<egcNR; g++)
+ {
+ printf(" %5d",groups->ngrpnr[g]);
+ nat_max = max(nat_max,groups->ngrpnr[g]);
+ }
+ printf("\n");
+
+ if (nat_max == 0)
+ {
+ (void) pr_indent(fp,indent);
+ fprintf(fp,"groupnr[%5s] =","*");
+ for(g=0; g<egcNR; g++)
+ {
+ fprintf(fp," %3d ",0);
+ }
+ fprintf(fp,"\n");
+ }
+ else
+ {
+ for(i=0; i<nat_max; i++)
+ {
+ (void) pr_indent(fp,indent);
+ fprintf(fp,"groupnr[%5d] =",i);
+ for(g=0; g<egcNR; g++)
+ {
+ fprintf(fp," %3d ",
+ groups->grpnr[g] ? groups->grpnr[g][i] : 0);
+ }
+ fprintf(fp,"\n");
+ }
+ }
+}
+
+void pr_atoms(FILE *fp,int indent,const char *title,t_atoms *atoms,
+ gmx_bool bShownumbers)
+{
+ if (available(fp,atoms,indent,title))
+ {
+ indent=pr_title(fp,indent,title);
+ pr_atom(fp,indent,"atom",atoms->atom,atoms->nr);
+ pr_strings(fp,indent,"atom",atoms->atomname,atoms->nr,bShownumbers);
+ pr_strings2(fp,indent,"type",atoms->atomtype,atoms->atomtypeB,atoms->nr,bShownumbers);
+ pr_resinfo(fp,indent,"residue",atoms->resinfo,atoms->nres,bShownumbers);
+ }
+}
+
+
+void pr_atomtypes(FILE *fp,int indent,const char *title,t_atomtypes *atomtypes,
+ gmx_bool bShowNumbers)
+{
+ int i;
+ if (available(fp,atomtypes,indent,title))
+ {
+ indent=pr_title(fp,indent,title);
+ for(i=0;i<atomtypes->nr;i++) {
+ pr_indent(fp,indent);
+ fprintf(fp,
+ "atomtype[%3d]={radius=%12.5e, volume=%12.5e, gb_radius=%12.5e, surftens=%12.5e, atomnumber=%4d, S_hct=%12.5e)}\n",
+ bShowNumbers?i:-1,atomtypes->radius[i],atomtypes->vol[i],
+ atomtypes->gb_radius[i],
+ atomtypes->surftens[i],atomtypes->atomnumber[i],atomtypes->S_hct[i]);
+ }
+ }
+}
+
+static void pr_moltype(FILE *fp,int indent,const char *title,
+ gmx_moltype_t *molt,int n,
+ gmx_ffparams_t *ffparams,
+ gmx_bool bShowNumbers)
+{
+ int j;
+
+ indent = pr_title_n(fp,indent,title,n);
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"name=\"%s\"\n",*(molt->name));
+ pr_atoms(fp,indent,"atoms",&(molt->atoms),bShowNumbers);
+ pr_block(fp,indent,"cgs",&molt->cgs, bShowNumbers);
+ pr_blocka(fp,indent,"excls",&molt->excls, bShowNumbers);
+ for(j=0; (j<F_NRE); j++) {
+ pr_ilist(fp,indent,interaction_function[j].longname,
+ ffparams->functype,&molt->ilist[j],bShowNumbers);
+ }
+}
+
+static void pr_molblock(FILE *fp,int indent,const char *title,
+ gmx_molblock_t *molb,int n,
+ gmx_moltype_t *molt,
+ gmx_bool bShowNumbers)
+{
+ indent = pr_title_n(fp,indent,title,n);
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"%-20s = %d \"%s\"\n",
+ "moltype",molb->type,*(molt[molb->type].name));
+ pr_int(fp,indent,"#molecules",molb->nmol);
+ pr_int(fp,indent,"#atoms_mol",molb->natoms_mol);
+ pr_int(fp,indent,"#posres_xA",molb->nposres_xA);
+ if (molb->nposres_xA > 0) {
+ pr_rvecs(fp,indent,"posres_xA",molb->posres_xA,molb->nposres_xA);
+ }
+ pr_int(fp,indent,"#posres_xB",molb->nposres_xB);
+ if (molb->nposres_xB > 0) {
+ pr_rvecs(fp,indent,"posres_xB",molb->posres_xB,molb->nposres_xB);
+ }
+}
+
+void pr_mtop(FILE *fp,int indent,const char *title,gmx_mtop_t *mtop,
+ gmx_bool bShowNumbers)
+{
+ int mt,mb;
+
+ if (available(fp,mtop,indent,title)) {
+ indent=pr_title(fp,indent,title);
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"name=\"%s\"\n",*(mtop->name));
+ pr_int(fp,indent,"#atoms",mtop->natoms);
+ for(mb=0; mb<mtop->nmolblock; mb++) {
+ pr_molblock(fp,indent,"molblock",&mtop->molblock[mb],mb,
+ mtop->moltype,bShowNumbers);
+ }
+ pr_ffparams(fp,indent,"ffparams",&(mtop->ffparams),bShowNumbers);
+ pr_atomtypes(fp,indent,"atomtypes",&(mtop->atomtypes),bShowNumbers);
+ for(mt=0; mt<mtop->nmoltype; mt++) {
+ pr_moltype(fp,indent,"moltype",&mtop->moltype[mt],mt,
+ &mtop->ffparams,bShowNumbers);
+ }
+ pr_groups(fp,indent,"groups",&mtop->groups,bShowNumbers);
+ }
+}
+
+void pr_top(FILE *fp,int indent,const char *title,t_topology *top, gmx_bool bShowNumbers)
+{
+ if (available(fp,top,indent,title)) {
+ indent=pr_title(fp,indent,title);
+ (void) pr_indent(fp,indent);
+ (void) fprintf(fp,"name=\"%s\"\n",*(top->name));
+ pr_atoms(fp,indent,"atoms",&(top->atoms),bShowNumbers);
+ pr_atomtypes(fp,indent,"atomtypes",&(top->atomtypes),bShowNumbers);
+ pr_block(fp,indent,"cgs",&top->cgs, bShowNumbers);
+ pr_block(fp,indent,"mols",&top->mols, bShowNumbers);
+ pr_blocka(fp,indent,"excls",&top->excls, bShowNumbers);
+ pr_idef(fp,indent,"idef",&top->idef,bShowNumbers);
+ }
+}
+
+void pr_header(FILE *fp,int indent,const char *title,t_tpxheader *sh)
+{
+ char buf[22];
+
+ if (available(fp,sh,indent,title))
+ {
+ indent=pr_title(fp,indent,title);
+ pr_indent(fp,indent);
+ fprintf(fp,"bIr = %spresent\n",sh->bIr?"":"not ");
+ pr_indent(fp,indent);
+ fprintf(fp,"bBox = %spresent\n",sh->bBox?"":"not ");
+ pr_indent(fp,indent);
+ fprintf(fp,"bTop = %spresent\n",sh->bTop?"":"not ");
+ pr_indent(fp,indent);
+ fprintf(fp,"bX = %spresent\n",sh->bX?"":"not ");
+ pr_indent(fp,indent);
+ fprintf(fp,"bV = %spresent\n",sh->bV?"":"not ");
+ pr_indent(fp,indent);
+ fprintf(fp,"bF = %spresent\n",sh->bF?"":"not ");
+
+ pr_indent(fp,indent);
+ fprintf(fp,"natoms = %d\n",sh->natoms);
+ pr_indent(fp,indent);
+ fprintf(fp,"lambda = %e\n",sh->lambda);
+ }
+}
+
+void pr_commrec(FILE *fp,int indent,t_commrec *cr)
+{
+ pr_indent(fp,indent);
+ fprintf(fp,"commrec:\n");
+ indent+=2;
+ pr_indent(fp,indent);
+ fprintf(fp,"nodeid = %d\n",cr->nodeid);
+ pr_indent(fp,indent);
+ fprintf(fp,"nnodes = %d\n",cr->nnodes);
+ pr_indent(fp,indent);
+ fprintf(fp,"npmenodes = %d\n",cr->npmenodes);
+ /*
+ pr_indent(fp,indent);
+ fprintf(fp,"threadid = %d\n",cr->threadid);
+ pr_indent(fp,indent);
+ fprintf(fp,"nthreads = %d\n",cr->nthreads);
+ */
+}
--- /dev/null
- #ifdef GMX_THREADS
+/* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
+ *
+ *
+ * This source code is part of
+ *
+ * G R O M A C S
+ *
+ * GROningen MAchine for Chemical Simulations
+ *
+ * VERSION 3.2.0
+ * Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
+ * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
+ * Copyright (c) 2001-2004, The GROMACS development team,
+ * check out http://www.gromacs.org for more information.
+
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version 2
+ * of the License, or (at your option) any later version.
+ *
+ * If you want to redistribute modifications, please consider that
+ * scientific software is very special. Version control is crucial -
+ * bugs must be traceable. We will be happy to consider code for
+ * inclusion in the official distribution, but derived work must not
+ * be called official GROMACS. Details are found in the README & COPYING
+ * files - if they are missing, get the official version at www.gromacs.org.
+ *
+ * To help us fund GROMACS development, we humbly ask that you cite
+ * the papers on the package - you can find them in the top README file.
+ *
+ * For more info, check our website at http://www.gromacs.org
+ *
+ * And Hey:
+ * GROningen Mixture of Alchemy and Childrens' Stories
+ */
+/* This file is completely threadsafe - keep it that way! */
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+#include "smalloc.h"
+#include "symtab.h"
+#include "vec.h"
+#include "pbc.h"
+#include "macros.h"
+#include <string.h>
+
- #ifdef GMX_THREADS
++#ifdef GMX_THREAD_MPI
+#include "thread_mpi.h"
+#endif
+
+/* The source code in this file should be thread-safe.
+ Please keep it that way. */
+
+
+
+static gmx_bool bOverAllocDD=FALSE;
- #ifdef GMX_THREADS
++#ifdef GMX_THREAD_MPI
+static tMPI_Thread_mutex_t over_alloc_mutex=TMPI_THREAD_MUTEX_INITIALIZER;
+#endif
+
+
+void set_over_alloc_dd(gmx_bool set)
+{
- #ifdef GMX_THREADS
++#ifdef GMX_THREAD_MPI
+ tMPI_Thread_mutex_lock(&over_alloc_mutex);
+ /* we just make sure that we don't set this at the same time.
+ We don't worry too much about reading this rarely-set variable */
+#endif
+ bOverAllocDD = set;
++#ifdef GMX_THREAD_MPI
+ tMPI_Thread_mutex_unlock(&over_alloc_mutex);
+#endif
+}
+
+int over_alloc_dd(int n)
+{
+ if (bOverAllocDD)
+ return OVER_ALLOC_FAC*n + 100;
+ else
+ return n;
+}
+
+int gmx_large_int_to_int(gmx_large_int_t step,const char *warn)
+{
+ int i;
+
+ i = (int)step;
+
+ if (warn != NULL && (step < INT_MIN || step > INT_MAX)) {
+ fprintf(stderr,"\nWARNING during %s:\n",warn);
+ fprintf(stderr,"step value ");
+ fprintf(stderr,gmx_large_int_pfmt,step);
+ fprintf(stderr," does not fit in int, converted to %d\n\n",i);
+ }
+
+ return i;
+}
+
+char *gmx_step_str(gmx_large_int_t i,char *buf)
+{
+ sprintf(buf,gmx_large_int_pfmt,i);
+
+ return buf;
+}
+
+void init_block(t_block *block)
+{
+ int i;
+
+ block->nr = 0;
+ block->nalloc_index = 1;
+ snew(block->index,block->nalloc_index);
+ block->index[0] = 0;
+}
+
+void init_blocka(t_blocka *block)
+{
+ int i;
+
+ block->nr = 0;
+ block->nra = 0;
+ block->nalloc_index = 1;
+ snew(block->index,block->nalloc_index);
+ block->index[0] = 0;
+ block->nalloc_a = 0;
+ block->a = NULL;
+}
+
+void init_atom(t_atoms *at)
+{
+ int i;
+
+ at->nr = 0;
+ at->nres = 0;
+ at->atom = NULL;
+ at->resinfo = NULL;
+ at->atomname = NULL;
+ at->atomtype = NULL;
+ at->atomtypeB= NULL;
+ at->pdbinfo = NULL;
+}
+
+void init_atomtypes(t_atomtypes *at)
+{
+ at->nr = 0;
+ at->radius = NULL;
+ at->vol = NULL;
+ at->atomnumber = NULL;
+ at->gb_radius = NULL;
+ at->S_hct = NULL;
+}
+
+void init_groups(gmx_groups_t *groups)
+{
+ int g;
+
+ groups->ngrpname = 0;
+ groups->grpname = NULL;
+ for(g=0; (g<egcNR); g++) {
+ groups->grps[g].nm_ind = NULL;
+ groups->ngrpnr[g] = 0;
+ groups->grpnr[g] = NULL;
+ }
+
+}
+
+void init_mtop(gmx_mtop_t *mtop)
+{
+ mtop->name = NULL;
+ mtop->nmoltype = 0;
+ mtop->moltype = NULL;
+ mtop->nmolblock = 0;
+ mtop->molblock = NULL;
+ mtop->maxres_renum = 0;
+ mtop->maxresnr = -1;
+ init_groups(&mtop->groups);
+ init_block(&mtop->mols);
+ open_symtab(&mtop->symtab);
+}
+
+void init_top (t_topology *top)
+{
+ int i;
+
+ top->name = NULL;
+ init_atom (&(top->atoms));
+ init_atomtypes(&(top->atomtypes));
+ init_block(&top->cgs);
+ init_block(&top->mols);
+ init_blocka(&top->excls);
+ open_symtab(&top->symtab);
+}
+
+void init_inputrec(t_inputrec *ir)
+{
+ memset(ir,0,(size_t)sizeof(*ir));
+}
+
+void stupid_fill_block(t_block *grp,int natom,gmx_bool bOneIndexGroup)
+{
+ int i;
+
+ if (bOneIndexGroup) {
+ grp->nalloc_index = 2;
+ snew(grp->index,grp->nalloc_index);
+ grp->index[0]=0;
+ grp->index[1]=natom;
+ grp->nr=1;
+ }
+ else {
+ grp->nalloc_index = natom+1;
+ snew(grp->index,grp->nalloc_index);
+ snew(grp->index,natom+1);
+ for(i=0; (i<=natom); i++)
+ grp->index[i]=i;
+ grp->nr=natom;
+ }
+}
+
+void stupid_fill_blocka(t_blocka *grp,int natom)
+{
+ int i;
+
+ grp->nalloc_a = natom;
+ snew(grp->a,grp->nalloc_a);
+ for(i=0; (i<natom); i++)
+ grp->a[i]=i;
+ grp->nra=natom;
+
+ grp->nalloc_index = natom + 1;
+ snew(grp->index,grp->nalloc_index);
+ for(i=0; (i<=natom); i++)
+ grp->index[i]=i;
+ grp->nr=natom;
+}
+
+void copy_blocka(const t_blocka *src,t_blocka *dest)
+{
+ int i;
+
+ dest->nr = src->nr;
+ dest->nalloc_index = dest->nr + 1;
+ snew(dest->index,dest->nalloc_index);
+ for(i=0; i<dest->nr+1; i++) {
+ dest->index[i] = src->index[i];
+ }
+ dest->nra = src->nra;
+ dest->nalloc_a = dest->nra + 1;
+ snew(dest->a,dest->nalloc_a);
+ for(i=0; i<dest->nra+1; i++) {
+ dest->a[i] = src->a[i];
+ }
+}
+
+void done_block(t_block *block)
+{
+ block->nr = 0;
+ sfree(block->index);
+ block->nalloc_index = 0;
+}
+
+void done_blocka(t_blocka *block)
+{
+ block->nr = 0;
+ block->nra = 0;
+ sfree(block->index);
+ if (block->a)
+ sfree(block->a);
+ block->nalloc_index = 0;
+ block->nalloc_a = 0;
+}
+
+void done_atom (t_atoms *at)
+{
+ at->nr = 0;
+ at->nres = 0;
+ sfree(at->atom);
+ sfree(at->resinfo);
+ sfree(at->atomname);
+ sfree(at->atomtype);
+ sfree(at->atomtypeB);
+}
+
+void done_atomtypes(t_atomtypes *atype)
+{
+ atype->nr = 0;
+ sfree(atype->radius);
+ sfree(atype->vol);
+ sfree(atype->surftens);
+ sfree(atype->atomnumber);
+ sfree(atype->gb_radius);
+ sfree(atype->S_hct);
+}
+
+void done_moltype(gmx_moltype_t *molt)
+{
+ int f;
+
+ done_atom(&molt->atoms);
+ done_block(&molt->cgs);
+ done_blocka(&molt->excls);
+
+ for(f=0; f<F_NRE; f++) {
+ sfree(molt->ilist[f].iatoms);
+ molt->ilist[f].nalloc = 0;
+ }
+}
+
+void done_molblock(gmx_molblock_t *molb)
+{
+ if (molb->nposres_xA > 0) {
+ molb->nposres_xA = 0;
+ free(molb->posres_xA);
+ }
+ if (molb->nposres_xB > 0) {
+ molb->nposres_xB = 0;
+ free(molb->posres_xB);
+ }
+}
+
+void done_mtop(gmx_mtop_t *mtop,gmx_bool bDoneSymtab)
+{
+ int i;
+
+ if (bDoneSymtab) {
+ done_symtab(&mtop->symtab);
+ }
+
+ sfree(mtop->ffparams.functype);
+ sfree(mtop->ffparams.iparams);
+
+ for(i=0; i<mtop->nmoltype; i++) {
+ done_moltype(&mtop->moltype[i]);
+ }
+ sfree(mtop->moltype);
+ for(i=0; i<mtop->nmolblock; i++) {
+ done_molblock(&mtop->molblock[i]);
+ }
+ sfree(mtop->molblock);
+ done_block(&mtop->mols);
+}
+
+void done_top(t_topology *top)
+{
+ int f;
+
+ sfree(top->idef.functype);
+ sfree(top->idef.iparams);
+ for (f = 0; f < F_NRE; ++f)
+ {
+ sfree(top->idef.il[f].iatoms);
+ top->idef.il[f].iatoms = NULL;
+ top->idef.il[f].nalloc = 0;
+ }
+
+ done_atom (&(top->atoms));
+
+ /* For GB */
+ done_atomtypes(&(top->atomtypes));
+
+ done_symtab(&(top->symtab));
+ done_block(&(top->cgs));
+ done_block(&(top->mols));
+ done_blocka(&(top->excls));
+}
+
+static void done_pullgrp(t_pullgrp *pgrp)
+{
+ sfree(pgrp->ind);
+ sfree(pgrp->ind_loc);
+ sfree(pgrp->weight);
+ sfree(pgrp->weight_loc);
+}
+
+static void done_pull(t_pull *pull)
+{
+ int i;
+
+ for(i=0; i<pull->ngrp+1; i++) {
+ done_pullgrp(pull->grp);
+ done_pullgrp(pull->dyna);
+ }
+}
+
+void done_inputrec(t_inputrec *ir)
+{
+ int m;
+
+ for(m=0; (m<DIM); m++) {
+ if (ir->ex[m].a) sfree(ir->ex[m].a);
+ if (ir->ex[m].phi) sfree(ir->ex[m].phi);
+ if (ir->et[m].a) sfree(ir->et[m].a);
+ if (ir->et[m].phi) sfree(ir->et[m].phi);
+ }
+
+ sfree(ir->opts.nrdf);
+ sfree(ir->opts.ref_t);
+ sfree(ir->opts.annealing);
+ sfree(ir->opts.anneal_npoints);
+ sfree(ir->opts.anneal_time);
+ sfree(ir->opts.anneal_temp);
+ sfree(ir->opts.tau_t);
+ sfree(ir->opts.acc);
+ sfree(ir->opts.nFreeze);
+ sfree(ir->opts.QMmethod);
+ sfree(ir->opts.QMbasis);
+ sfree(ir->opts.QMcharge);
+ sfree(ir->opts.QMmult);
+ sfree(ir->opts.bSH);
+ sfree(ir->opts.CASorbitals);
+ sfree(ir->opts.CASelectrons);
+ sfree(ir->opts.SAon);
+ sfree(ir->opts.SAoff);
+ sfree(ir->opts.SAsteps);
+ sfree(ir->opts.bOPT);
+ sfree(ir->opts.bTS);
+
+ if (ir->pull) {
+ done_pull(ir->pull);
+ sfree(ir->pull);
+ }
+}
+
+static void init_ekinstate(ekinstate_t *eks)
+{
+ eks->ekin_n = 0;
+ eks->ekinh = NULL;
+ eks->ekinf = NULL;
+ eks->ekinh_old = NULL;
+ eks->ekinscalef_nhc = NULL;
+ eks->ekinscaleh_nhc = NULL;
+ eks->vscale_nhc = NULL;
+ eks->dekindl = 0;
+ eks->mvcos = 0;
+}
+
+void init_energyhistory(energyhistory_t * enerhist)
+{
+ enerhist->nener = 0;
+
+ enerhist->ener_ave = NULL;
+ enerhist->ener_sum = NULL;
+ enerhist->ener_sum_sim = NULL;
+ enerhist->dht = NULL;
+
+ enerhist->nsteps = 0;
+ enerhist->nsum = 0;
+ enerhist->nsteps_sim = 0;
+ enerhist->nsum_sim = 0;
+
+ enerhist->dht = NULL;
+}
+
+static void done_delta_h_history(delta_h_history_t *dht)
+{
+ int i;
+
+ for(i=0; i<dht->nndh; i++)
+ {
+ sfree(dht->dh[i]);
+ }
+ sfree(dht->dh);
+ sfree(dht->ndh);
+}
+
+void done_energyhistory(energyhistory_t * enerhist)
+{
+ sfree(enerhist->ener_ave);
+ sfree(enerhist->ener_sum);
+ sfree(enerhist->ener_sum_sim);
+
+ if (enerhist->dht != NULL)
+ {
+ done_delta_h_history(enerhist->dht);
+ sfree(enerhist->dht);
+ }
+}
+
+void init_gtc_state(t_state *state, int ngtc, int nnhpres, int nhchainlength)
+{
+ int i,j;
+
+ state->ngtc = ngtc;
+ state->nnhpres = nnhpres;
+ state->nhchainlength = nhchainlength;
+ if (state->ngtc > 0)
+ {
+ snew(state->nosehoover_xi,state->nhchainlength*state->ngtc);
+ snew(state->nosehoover_vxi,state->nhchainlength*state->ngtc);
+ snew(state->therm_integral,state->ngtc);
+ for(i=0; i<state->ngtc; i++)
+ {
+ for (j=0;j<state->nhchainlength;j++)
+ {
+ state->nosehoover_xi[i*state->nhchainlength + j] = 0.0;
+ state->nosehoover_vxi[i*state->nhchainlength + j] = 0.0;
+ }
+ }
+ for(i=0; i<state->ngtc; i++) {
+ state->therm_integral[i] = 0.0;
+ }
+ }
+ else
+ {
+ state->nosehoover_xi = NULL;
+ state->nosehoover_vxi = NULL;
+ state->therm_integral = NULL;
+ }
+
+ if (state->nnhpres > 0)
+ {
+ snew(state->nhpres_xi,state->nhchainlength*nnhpres);
+ snew(state->nhpres_vxi,state->nhchainlength*nnhpres);
+ for(i=0; i<nnhpres; i++)
+ {
+ for (j=0;j<state->nhchainlength;j++)
+ {
+ state->nhpres_xi[i*nhchainlength + j] = 0.0;
+ state->nhpres_vxi[i*nhchainlength + j] = 0.0;
+ }
+ }
+ }
+ else
+ {
+ state->nhpres_xi = NULL;
+ state->nhpres_vxi = NULL;
+ }
+}
+
+
+void init_state(t_state *state, int natoms, int ngtc, int nnhpres, int nhchainlength)
+{
+ int i;
+
+ state->natoms = natoms;
+ state->nrng = 0;
+ state->flags = 0;
+ state->lambda = 0;
+ state->veta = 0;
+ clear_mat(state->box);
+ clear_mat(state->box_rel);
+ clear_mat(state->boxv);
+ clear_mat(state->pres_prev);
+ clear_mat(state->svir_prev);
+ clear_mat(state->fvir_prev);
+ init_gtc_state(state,ngtc,nnhpres,nhchainlength);
+ state->nalloc = state->natoms;
+ if (state->nalloc > 0) {
+ snew(state->x,state->nalloc);
+ snew(state->v,state->nalloc);
+ } else {
+ state->x = NULL;
+ state->v = NULL;
+ }
+ state->sd_X = NULL;
+ state->cg_p = NULL;
+
+ init_ekinstate(&state->ekinstate);
+
+ init_energyhistory(&state->enerhist);
+
+ state->ddp_count = 0;
+ state->ddp_count_cg_gl = 0;
+ state->cg_gl = NULL;
+ state->cg_gl_nalloc = 0;
+}
+
+void done_state(t_state *state)
+{
+ if (state->nosehoover_xi) sfree(state->nosehoover_xi);
+ if (state->x) sfree(state->x);
+ if (state->v) sfree(state->v);
+ if (state->sd_X) sfree(state->sd_X);
+ if (state->cg_p) sfree(state->cg_p);
+ state->nalloc = 0;
+ if (state->cg_gl) sfree(state->cg_gl);
+ state->cg_gl_nalloc = 0;
+}
+
+static void do_box_rel(t_inputrec *ir,matrix box_rel,matrix b,gmx_bool bInit)
+{
+ int d,d2;
+
+ for(d=YY; d<=ZZ; d++) {
+ for(d2=XX; d2<=(ir->epct==epctSEMIISOTROPIC ? YY : ZZ); d2++) {
+ /* We need to check if this box component is deformed
+ * or if deformation of another component might cause
+ * changes in this component due to box corrections.
+ */
+ if (ir->deform[d][d2] == 0 &&
+ !(d == ZZ && d2 == XX && ir->deform[d][YY] != 0 &&
+ (b[YY][d2] != 0 || ir->deform[YY][d2] != 0))) {
+ if (bInit) {
+ box_rel[d][d2] = b[d][d2]/b[XX][XX];
+ } else {
+ b[d][d2] = b[XX][XX]*box_rel[d][d2];
+ }
+ }
+ }
+ }
+}
+
+void set_box_rel(t_inputrec *ir,t_state *state)
+{
+ /* Make sure the box obeys the restrictions before we fix the ratios */
+ correct_box(NULL,0,state->box,NULL);
+
+ clear_mat(state->box_rel);
+
+ if (PRESERVE_SHAPE(*ir))
+ do_box_rel(ir,state->box_rel,state->box,TRUE);
+}
+
+void preserve_box_shape(t_inputrec *ir,matrix box_rel,matrix b)
+{
+ if (PRESERVE_SHAPE(*ir))
+ do_box_rel(ir,box_rel,b,FALSE);
+}
+
+void add_t_atoms(t_atoms *atoms,int natom_extra,int nres_extra)
+{
+ int i;
+
+ if (natom_extra > 0)
+ {
+ srenew(atoms->atomname,atoms->nr+natom_extra);
+ srenew(atoms->atom,atoms->nr+natom_extra);
+ if (NULL != atoms->pdbinfo)
+ srenew(atoms->pdbinfo,atoms->nr+natom_extra);
+ if (NULL != atoms->atomtype)
+ srenew(atoms->atomtype,atoms->nr+natom_extra);
+ if (NULL != atoms->atomtypeB)
+ srenew(atoms->atomtypeB,atoms->nr+natom_extra);
+ for(i=atoms->nr; (i<atoms->nr+natom_extra); i++) {
+ atoms->atomname[i] = NULL;
+ memset(&atoms->atom[i],0,sizeof(atoms->atom[i]));
+ if (NULL != atoms->pdbinfo)
+ memset(&atoms->pdbinfo[i],0,sizeof(atoms->pdbinfo[i]));
+ if (NULL != atoms->atomtype)
+ atoms->atomtype[i] = NULL;
+ if (NULL != atoms->atomtypeB)
+ atoms->atomtypeB[i] = NULL;
+ }
+ atoms->nr += natom_extra;
+ }
+ if (nres_extra > 0)
+ {
+ srenew(atoms->resinfo,atoms->nres+nres_extra);
+ for(i=atoms->nres; (i<atoms->nres+nres_extra); i++) {
+ memset(&atoms->resinfo[i],0,sizeof(atoms->resinfo[i]));
+ }
+ atoms->nres += nres_extra;
+ }
+}
+
+void init_t_atoms(t_atoms *atoms, int natoms, gmx_bool bPdbinfo)
+{
+ atoms->nr=natoms;
+ atoms->nres=0;
+ snew(atoms->atomname,natoms);
+ atoms->atomtype=NULL;
+ atoms->atomtypeB=NULL;
+ snew(atoms->resinfo,natoms);
+ snew(atoms->atom,natoms);
+ if (bPdbinfo)
+ snew(atoms->pdbinfo,natoms);
+ else
+ atoms->pdbinfo=NULL;
+}
+
+t_atoms *copy_t_atoms(t_atoms *src)
+{
+ t_atoms *dst;
+ int i;
+
+ snew(dst,1);
+ init_t_atoms(dst,src->nr,(NULL != src->pdbinfo));
+ dst->nr = src->nr;
+ if (NULL != src->atomname)
+ snew(dst->atomname,src->nr);
+ if (NULL != src->atomtype)
+ snew(dst->atomtype,src->nr);
+ if (NULL != src->atomtypeB)
+ snew(dst->atomtypeB,src->nr);
+ for(i=0; (i<src->nr); i++) {
+ dst->atom[i] = src->atom[i];
+ if (NULL != src->pdbinfo)
+ dst->pdbinfo[i] = src->pdbinfo[i];
+ if (NULL != src->atomname)
+ dst->atomname[i] = src->atomname[i];
+ if (NULL != src->atomtype)
+ dst->atomtype[i] = src->atomtype[i];
+ if (NULL != src->atomtypeB)
+ dst->atomtypeB[i] = src->atomtypeB[i];
+ }
+ dst->nres = src->nres;
+ for(i=0; (i<src->nres); i++) {
+ dst->resinfo[i] = src->resinfo[i];
+ }
+ return dst;
+}
+
+void t_atoms_set_resinfo(t_atoms *atoms,int atom_ind,t_symtab *symtab,
+ const char *resname,int resnr,unsigned char ic,
+ int chainnum, char chainid)
+{
+ t_resinfo *ri;
+
+ ri = &atoms->resinfo[atoms->atom[atom_ind].resind];
+ ri->name = put_symtab(symtab,resname);
+ ri->rtp = NULL;
+ ri->nr = resnr;
+ ri->ic = ic;
+ ri->chainnum = chainnum;
+ ri->chainid = chainid;
+}
+
+void free_t_atoms(t_atoms *atoms,gmx_bool bFreeNames)
+{
+ int i;
+
+ if (bFreeNames) {
+ for(i=0; i<atoms->nr; i++) {
+ sfree(*atoms->atomname[i]);
+ *atoms->atomname[i]=NULL;
+ }
+ for(i=0; i<atoms->nres; i++) {
+ sfree(*atoms->resinfo[i].name);
+ *atoms->resinfo[i].name=NULL;
+ }
+ }
+ sfree(atoms->atomname);
+ /* Do we need to free atomtype and atomtypeB as well ? */
+ sfree(atoms->resinfo);
+ sfree(atoms->atom);
+ if (atoms->pdbinfo)
+ sfree(atoms->pdbinfo);
+ atoms->nr=0;
+ atoms->nres=0;
+}
+
+real max_cutoff(real cutoff1,real cutoff2)
+{
+ if (cutoff1 == 0 || cutoff2 == 0)
+ {
+ return 0;
+ }
+ else
+ {
+ return max(cutoff1,cutoff2);
+ }
+}
--- /dev/null
- #ifdef GMX_THREADS
+/*
+ *
+ * This source code is part of
+ *
+ * G R O M A C S
+ *
+ * GROningen MAchine for Chemical Simulations
+ *
+ * VERSION 3.2.0
+ * Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
+ * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
+ * Copyright (c) 2001-2004, The GROMACS development team,
+ * check out http://www.gromacs.org for more information.
+
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version 2
+ * of the License, or (at your option) any later version.
+ *
+ * If you want to redistribute modifications, please consider that
+ * scientific software is very special. Version control is crucial -
+ * bugs must be traceable. We will be happy to consider code for
+ * inclusion in the official distribution, but derived work must not
+ * be called official GROMACS. Details are found in the README & COPYING
+ * files - if they are missing, get the official version at www.gromacs.org.
+ *
+ * To help us fund GROMACS development, we humbly ask that you cite
+ * the papers on the package - you can find them in the top README file.
+ *
+ * For more info, check our website at http://www.gromacs.org
+ *
+ * And Hey:
+ * Gromacs Runs On Most of All Computer Systems
+ */
+
+#ifndef _mdrun_h
+#define _mdrun_h
+
+#include <stdio.h>
+#include <time.h>
+#include "typedefs.h"
+#include "network.h"
+#include "tgroup.h"
+#include "filenm.h"
+#include "mshift.h"
+#include "force.h"
+#include "edsam.h"
+#include "mdebin.h"
+#include "vcm.h"
+#include "vsite.h"
+#include "pull.h"
+#include "update.h"
+#include "membed.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MD_POLARISE (1<<2)
+#define MD_IONIZE (1<<3)
+#define MD_RERUN (1<<4)
+#define MD_RERUN_VSITE (1<<5)
+#define MD_FFSCAN (1<<6)
+#define MD_SEPPOT (1<<7)
+#define MD_PARTDEC (1<<9)
+#define MD_DDBONDCHECK (1<<10)
+#define MD_DDBONDCOMM (1<<11)
+#define MD_CONFOUT (1<<12)
+#define MD_REPRODUCIBLE (1<<13)
+#define MD_READ_RNG (1<<14)
+#define MD_APPENDFILES (1<<15)
+#define MD_KEEPANDNUMCPT (1<<16)
+#define MD_READ_EKIN (1<<17)
+#define MD_STARTFROMCPT (1<<18)
+#define MD_RESETCOUNTERSHALFWAY (1<<19)
+
+/* Define a number of flags to better control the information
+ * passed to compute_globals in md.c and global_stat.
+ */
+
+/* We are rerunning the simulation */
+#define CGLO_RERUNMD (1<<1)
+/* we are computing the kinetic energy from average velocities */
+#define CGLO_EKINAVEVEL (1<<2)
+/* we are removing the center of mass momenta */
+#define CGLO_STOPCM (1<<3)
+/* bGStat is defined in do_md */
+#define CGLO_GSTAT (1<<4)
+/* Sum the energy terms in global computation */
+#define CGLO_ENERGY (1<<6)
+/* Sum the kinetic energy terms in global computation */
+#define CGLO_TEMPERATURE (1<<7)
+/* Sum the kinetic energy terms in global computation */
+#define CGLO_PRESSURE (1<<8)
+/* Sum the constraint term in global computation */
+#define CGLO_CONSTRAINT (1<<9)
+/* we are using an integrator that requires iteration over some steps - currently not used*/
+#define CGLO_ITERATE (1<<10)
+/* it is the first time we are iterating (or, only once through is required */
+#define CGLO_FIRSTITERATE (1<<11)
+/* Reading ekin from the trajectory */
+#define CGLO_READEKIN (1<<12)
+/* we need to reset the ekin rescaling factor here */
+#define CGLO_SCALEEKIN (1<<13)
+
+enum {
+ ddnoSEL, ddnoINTERLEAVE, ddnoPP_PME, ddnoCARTESIAN, ddnoNR
+};
+
+typedef struct {
+ double real;
+#ifdef GMX_CRAY_XT3
+ double proc;
+#else
+ clock_t proc;
+#endif
+ double realtime;
+ double proctime;
+ double time_per_step;
+ double last;
+ gmx_large_int_t nsteps_done;
+} gmx_runtime_t;
+
+typedef struct {
+ t_fileio *fp_trn;
+ t_fileio *fp_xtc;
+ int xtc_prec;
+ ener_file_t fp_ene;
+ const char *fn_cpt;
+ gmx_bool bKeepAndNumCPT;
+ int eIntegrator;
+ int simulation_part;
+ FILE *fp_dhdl;
+ FILE *fp_field;
+} gmx_mdoutf_t;
+
+/* Variables for temporary use with the deform option,
+ * used in runner.c and md.c.
+ * (These variables should be stored in the tpx file.)
+ */
+extern gmx_large_int_t deform_init_init_step_tpx;
+extern matrix deform_init_box_tpx;
++#ifdef GMX_THREAD_MPI
+extern tMPI_Thread_mutex_t deform_init_box_mutex;
+
+/* The minimum number of atoms per thread. With fewer atoms than this,
+ * the number of threads will get lowered.
+ */
+#define MIN_ATOMS_PER_THREAD 90
+#endif
+
+
+typedef double gmx_integrator_t(FILE *log,t_commrec *cr,
+ int nfile,const t_filenm fnm[],
+ const output_env_t oenv, gmx_bool bVerbose,
+ gmx_bool bCompact, int nstglobalcomm,
+ gmx_vsite_t *vsite,gmx_constr_t constr,
+ int stepout,
+ t_inputrec *inputrec,
+ gmx_mtop_t *mtop,t_fcdata *fcd,
+ t_state *state,
+ t_mdatoms *mdatoms,
+ t_nrnb *nrnb,gmx_wallcycle_t wcycle,
+ gmx_edsam_t ed,
+ t_forcerec *fr,
+ int repl_ex_nst,int repl_ex_seed,
+ gmx_membed_t *membed,
+ real cpt_period,real max_hours,
+ const char *deviceOptions,
+ unsigned long Flags,
+ gmx_runtime_t *runtime);
+
+typedef struct gmx_global_stat *gmx_global_stat_t;
+
+/* ROUTINES from md.c */
+
+gmx_integrator_t do_md;
+
+gmx_integrator_t do_md_openmm;
+
+
+
+/* ROUTINES from minimize.c */
+
+gmx_integrator_t do_steep;
+/* Do steepest descents EM */
+
+gmx_integrator_t do_cg;
+/* Do conjugate gradient EM */
+
+gmx_integrator_t do_lbfgs;
+/* Do conjugate gradient L-BFGS */
+
+gmx_integrator_t do_nm;
+/* Do normal mode analysis */
+
+/* ROUTINES from tpi.c */
+
+gmx_integrator_t do_tpi;
+/* Do test particle insertion */
+
+
+/* ROUTINES from md_support.c */
+
+/* return the number of steps between global communcations */
+int check_nstglobalcomm(FILE *fplog,t_commrec *cr,
+ int nstglobalcomm,t_inputrec *ir);
+
+/* check whether an 'nst'-style parameter p is a multiple of nst, and
+ set it to be one if not, with a warning. */
+void check_nst_param(FILE *fplog,t_commrec *cr,
+ const char *desc_nst,int nst,
+ const char *desc_p,int *p);
+
+/* check which of the multisim simulations has the shortest number of
+ steps and return that number of nsteps */
+gmx_large_int_t get_multisim_nsteps(const t_commrec *cr,
+ gmx_large_int_t nsteps);
+
+void rerun_parallel_comm(t_commrec *cr,t_trxframe *fr,
+ gmx_bool *bNotLastFrame);
+
+/* get the conserved energy associated with the ensemble type*/
+real compute_conserved_from_auxiliary(t_inputrec *ir, t_state *state,
+ t_extmass *MassQ);
+
+/* reset all cycle and time counters. */
+void reset_all_counters(FILE *fplog,t_commrec *cr,
+ gmx_large_int_t step,
+ gmx_large_int_t *step_rel,t_inputrec *ir,
+ gmx_wallcycle_t wcycle,t_nrnb *nrnb,
+ gmx_runtime_t *runtime);
+
+
+
+/* ROUTINES from sim_util.c */
+void do_pbc_first(FILE *log,matrix box,t_forcerec *fr,
+ t_graph *graph,rvec x[]);
+
+void do_pbc_first_mtop(FILE *fplog,int ePBC,matrix box,
+ gmx_mtop_t *mtop,rvec x[]);
+
+void do_pbc_mtop(FILE *fplog,int ePBC,matrix box,
+ gmx_mtop_t *mtop,rvec x[]);
+
+
+
+/* ROUTINES from stat.c */
+gmx_global_stat_t global_stat_init(t_inputrec *ir);
+
+void global_stat_destroy(gmx_global_stat_t gs);
+
+void global_stat(FILE *log,gmx_global_stat_t gs,
+ t_commrec *cr,gmx_enerdata_t *enerd,
+ tensor fvir,tensor svir,rvec mu_tot,
+ t_inputrec *inputrec,
+ gmx_ekindata_t *ekind,
+ gmx_constr_t constr,t_vcm *vcm,
+ int nsig,real *sig,
+ gmx_mtop_t *top_global, t_state *state_local,
+ gmx_bool bSumEkinhOld, int flags);
+/* Communicate statistics over cr->mpi_comm_mysim */
+
+gmx_mdoutf_t *init_mdoutf(int nfile,const t_filenm fnm[],
+ int mdrun_flags,
+ const t_commrec *cr,const t_inputrec *ir,
+ const output_env_t oenv);
+/* Returns a pointer to a data structure with all output file pointers
+ * and names required by mdrun.
+ */
+
+void done_mdoutf(gmx_mdoutf_t *of);
+/* Close all open output files and free the of pointer */
+
+#define MDOF_X (1<<0)
+#define MDOF_V (1<<1)
+#define MDOF_F (1<<2)
+#define MDOF_XTC (1<<3)
+#define MDOF_CPT (1<<4)
+
+void write_traj(FILE *fplog,t_commrec *cr,
+ gmx_mdoutf_t *of,
+ int mdof_flags,
+ gmx_mtop_t *top_global,
+ gmx_large_int_t step,double t,
+ t_state *state_local,t_state *state_global,
+ rvec *f_local,rvec *f_global,
+ int *n_xtc,rvec **x_xtc);
+/* Routine that writes frames to trn, xtc and/or checkpoint.
+ * What is written is determined by the mdof_flags defined above.
+ * Data is collected to the master node only when necessary.
+ */
+
+int do_per_step(gmx_large_int_t step,gmx_large_int_t nstep);
+/* Return TRUE if io should be done */
+
+int do_any_io(int step, t_inputrec *ir);
+
+/* ROUTINES from sim_util.c */
+
+double gmx_gettime();
+
+void print_time(FILE *out, gmx_runtime_t *runtime,
+ gmx_large_int_t step,t_inputrec *ir, t_commrec *cr);
+
+void runtime_start(gmx_runtime_t *runtime);
+
+void runtime_end(gmx_runtime_t *runtime);
+
+void runtime_upd_proc(gmx_runtime_t *runtime);
+/* The processor time should be updated every once in a while,
+ * since on 32-bit manchines it loops after 72 minutes.
+ */
+
+void print_date_and_time(FILE *log,int pid,const char *title,
+ const gmx_runtime_t *runtime);
+
+void nstop_cm(FILE *log,t_commrec *cr,
+ int start,int nr_atoms,real mass[],rvec x[],rvec v[]);
+
+void finish_run(FILE *log,t_commrec *cr,const char *confout,
+ t_inputrec *inputrec,
+ t_nrnb nrnb[],gmx_wallcycle_t wcycle,
+ gmx_runtime_t *runtime,
+ gmx_bool bWriteStat);
+
+void calc_enervirdiff(FILE *fplog,int eDispCorr,t_forcerec *fr);
+
+void calc_dispcorr(FILE *fplog,t_inputrec *ir,t_forcerec *fr,
+ gmx_large_int_t step, int natoms,
+ matrix box,real lambda,tensor pres,tensor virial,
+ real *prescorr, real *enercorr, real *dvdlcorr);
+
+typedef enum
+{
+ LIST_SCALARS =0001,
+ LIST_INPUTREC =0002,
+ LIST_TOP =0004,
+ LIST_X =0010,
+ LIST_V =0020,
+ LIST_F =0040,
+ LIST_LOAD =0100
+} t_listitem;
+
+void check_nnodes_top(char *fn,t_topology *top);
+/* Reset the tpr file to work with one node if necessary */
+
+
+/* check the version */
+void check_ir_old_tpx_versions(t_commrec *cr,FILE *fplog,
+ t_inputrec *ir,gmx_mtop_t *mtop);
+
+/* Allocate and initialize node-local state entries. */
+void set_state_entries(t_state *state,const t_inputrec *ir,int nnodes);
+
+/* Broadcast the data for a simulation, and allocate node-specific settings
+ such as rng generators. */
+void init_parallel(FILE *log, t_commrec *cr, t_inputrec *inputrec,
+ gmx_mtop_t *mtop);
+
+
+void do_constrain_first(FILE *log,gmx_constr_t constr,
+ t_inputrec *inputrec,t_mdatoms *md,
+ t_state *state,rvec *f,
+ t_graph *graph,t_commrec *cr,t_nrnb *nrnb,
+ t_forcerec *fr, gmx_localtop_t *top, tensor shake_vir);
+
+void dynamic_load_balancing(gmx_bool bVerbose,t_commrec *cr,real capacity[],
+ int dimension,t_mdatoms *md,t_topology *top,
+ rvec x[],rvec v[],matrix box);
+/* Perform load balancing, i.e. split the particles over processors
+ * based on their coordinates in the "dimension" direction.
+ */
+
+int multisim_min(const gmx_multisim_t *ms,int nmin,int n);
+/* Set an appropriate value for n across the whole multi-simulation */
+
+int multisim_nstsimsync(const t_commrec *cr,
+ const t_inputrec *ir,int repl_ex_nst);
+/* Determine the interval for inter-simulation communication */
+
+void init_global_signals(globsig_t *gs,const t_commrec *cr,
+ const t_inputrec *ir,int repl_ex_nst);
+/* Constructor for globsig_t */
+
+void copy_coupling_state(t_state *statea,t_state *stateb,
+ gmx_ekindata_t *ekinda,gmx_ekindata_t *ekindb, t_grpopts* opts);
+/* Copy stuff from state A to state B */
+
+void compute_globals(FILE *fplog, gmx_global_stat_t gstat, t_commrec *cr, t_inputrec *ir,
+ t_forcerec *fr, gmx_ekindata_t *ekind,
+ t_state *state, t_state *state_global, t_mdatoms *mdatoms,
+ t_nrnb *nrnb, t_vcm *vcm, gmx_wallcycle_t wcycle,
+ gmx_enerdata_t *enerd,tensor force_vir, tensor shake_vir, tensor total_vir,
+ tensor pres, rvec mu_tot, gmx_constr_t constr,
+ globsig_t *gs,gmx_bool bInterSimGS,
+ matrix box, gmx_mtop_t *top_global, real *pcurr,
+ int natoms, gmx_bool *bSumEkinhOld, int flags);
+/* Compute global variables during integration */
+
+int mdrunner(int nthreads_requested, 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,
+ int nstepout, int resetstep, int nmultisim, int repl_ex_nst,
+ int repl_ex_seed, real pforce,real cpt_period,real max_hours,
+ const char *deviceOptions, unsigned long Flags);
+/* Driver routine, that calls the different methods */
+
+void md_print_warning(const t_commrec *cr,FILE *fplog,const char *buf);
+/* Print a warning message to stderr on the master node
+ * and to fplog if fplog!=NULL.
+ */
+
+void init_md(FILE *fplog,
+ t_commrec *cr,t_inputrec *ir, const output_env_t oenv,
+ double *t,double *t0,
+ real *lambda,double *lam0,
+ t_nrnb *nrnb,gmx_mtop_t *mtop,
+ gmx_update_t *upd,
+ int nfile,const t_filenm fnm[],
+ gmx_mdoutf_t **outf,t_mdebin **mdebin,
+ tensor force_vir,tensor shake_vir,
+ rvec mu_tot,
+ gmx_bool *bSimAnn,t_vcm **vcm,
+ t_state *state, unsigned long Flags);
+ /* Routine in sim_util.c */
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* _mdrun_h */
--- /dev/null
- #ifdef GMX_THREADS
+/* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
+ *
+ *
+ * This file is part of Gromacs Copyright (c) 1991-2008
+ * David van der Spoel, Erik Lindahl, Berk Hess, University of Groningen.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version 2
+ * of the License, or (at your option) any later version.
+ *
+ * To help us fund GROMACS development, we humbly ask that you cite
+ * the research papers on the package. Check out http://www.gromacs.org
+ *
+ * And Hey:
+ * Gnomes, ROck Monsters And Chili Sauce
+ */
+
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+#include <stdio.h>
+#include <time.h>
+#include <math.h>
+#include <string.h>
+#include <stdlib.h>
+#include "typedefs.h"
+#include "smalloc.h"
+#include "vec.h"
+#include "domdec.h"
+#include "domdec_network.h"
+#include "nrnb.h"
+#include "pbc.h"
+#include "chargegroup.h"
+#include "constr.h"
+#include "mdatoms.h"
+#include "names.h"
+#include "pdbio.h"
+#include "futil.h"
+#include "force.h"
+#include "pme.h"
+#include "pull.h"
+#include "pull_rotation.h"
+#include "gmx_wallcycle.h"
+#include "mdrun.h"
+#include "nsgrid.h"
+#include "shellfc.h"
+#include "mtop_util.h"
+#include "gmxfio.h"
+#include "gmx_ga2la.h"
+#include "gmx_sort.h"
+#include "macros.h"
+
+#ifdef GMX_LIB_MPI
+#include <mpi.h>
+#endif
++#ifdef GMX_THREAD_MPI
+#include "tmpi.h"
+#endif
+
+#define DDRANK(dd,rank) (rank)
+#define DDMASTERRANK(dd) (dd->masterrank)
+
+typedef struct gmx_domdec_master
+{
+ /* The cell boundaries */
+ real **cell_x;
+ /* The global charge group division */
+ int *ncg; /* Number of home charge groups for each node */
+ int *index; /* Index of nnodes+1 into cg */
+ int *cg; /* Global charge group index */
+ int *nat; /* Number of home atoms for each node. */
+ int *ibuf; /* Buffer for communication */
+ rvec *vbuf; /* Buffer for state scattering and gathering */
+} gmx_domdec_master_t;
+
+typedef struct
+{
+ /* The numbers of charge groups to send and receive for each cell
+ * that requires communication, the last entry contains the total
+ * number of atoms that needs to be communicated.
+ */
+ int nsend[DD_MAXIZONE+2];
+ int nrecv[DD_MAXIZONE+2];
+ /* The charge groups to send */
+ int *index;
+ int nalloc;
+ /* The atom range for non-in-place communication */
+ int cell2at0[DD_MAXIZONE];
+ int cell2at1[DD_MAXIZONE];
+} gmx_domdec_ind_t;
+
+typedef struct
+{
+ int np; /* Number of grid pulses in this dimension */
+ int np_dlb; /* For dlb, for use with edlbAUTO */
+ gmx_domdec_ind_t *ind; /* The indices to communicate, size np */
+ int np_nalloc;
+ gmx_bool bInPlace; /* Can we communicate in place? */
+} gmx_domdec_comm_dim_t;
+
+typedef struct
+{
+ gmx_bool *bCellMin; /* Temp. var.: is this cell size at the limit */
+ real *cell_f; /* State var.: cell boundaries, box relative */
+ real *old_cell_f; /* Temp. var.: old cell size */
+ real *cell_f_max0; /* State var.: max lower boundary, incl neighbors */
+ real *cell_f_min1; /* State var.: min upper boundary, incl neighbors */
+ real *bound_min; /* Temp. var.: lower limit for cell boundary */
+ real *bound_max; /* Temp. var.: upper limit for cell boundary */
+ gmx_bool bLimited; /* State var.: is DLB limited in this dim and row */
+ real *buf_ncd; /* Temp. var. */
+} gmx_domdec_root_t;
+
+#define DD_NLOAD_MAX 9
+
+/* Here floats are accurate enough, since these variables
+ * only influence the load balancing, not the actual MD results.
+ */
+typedef struct
+{
+ int nload;
+ float *load;
+ float sum;
+ float max;
+ float sum_m;
+ float cvol_min;
+ float mdf;
+ float pme;
+ int flags;
+} gmx_domdec_load_t;
+
+typedef struct
+{
+ int nsc;
+ int ind_gl;
+ int ind;
+} gmx_cgsort_t;
+
+typedef struct
+{
+ gmx_cgsort_t *sort1,*sort2;
+ int sort_nalloc;
+ gmx_cgsort_t *sort_new;
+ int sort_new_nalloc;
+ int *ibuf;
+ int ibuf_nalloc;
+} gmx_domdec_sort_t;
+
+typedef struct
+{
+ rvec *v;
+ int nalloc;
+} vec_rvec_t;
+
+/* This enum determines the order of the coordinates.
+ * ddnatHOME and ddnatZONE should be first and second,
+ * the others can be ordered as wanted.
+ */
+enum { ddnatHOME, ddnatZONE, ddnatVSITE, ddnatCON, ddnatNR };
+
+enum { edlbAUTO, edlbNO, edlbYES, edlbNR };
+const char *edlb_names[edlbNR] = { "auto", "no", "yes" };
+
+typedef struct
+{
+ int dim; /* The dimension */
+ gmx_bool dim_match;/* Tells if DD and PME dims match */
+ int nslab; /* The number of PME slabs in this dimension */
+ real *slb_dim_f; /* Cell sizes for determining the PME comm. with SLB */
+ int *pp_min; /* The minimum pp node location, size nslab */
+ int *pp_max; /* The maximum pp node location,size nslab */
+ int maxshift; /* The maximum shift for coordinate redistribution in PME */
+} gmx_ddpme_t;
+
+typedef struct
+{
+ real min0; /* The minimum bottom of this zone */
+ real max1; /* The maximum top of this zone */
+ real mch0; /* The maximum bottom communicaton height for this zone */
+ real mch1; /* The maximum top communicaton height for this zone */
+ real p1_0; /* The bottom value of the first cell in this zone */
+ real p1_1; /* The top value of the first cell in this zone */
+} gmx_ddzone_t;
+
+typedef struct gmx_domdec_comm
+{
+ /* All arrays are indexed with 0 to dd->ndim (not Cartesian indexing),
+ * unless stated otherwise.
+ */
+
+ /* The number of decomposition dimensions for PME, 0: no PME */
+ int npmedecompdim;
+ /* The number of nodes doing PME (PP/PME or only PME) */
+ int npmenodes;
+ int npmenodes_x;
+ int npmenodes_y;
+ /* The communication setup including the PME only nodes */
+ gmx_bool bCartesianPP_PME;
+ ivec ntot;
+ int cartpmedim;
+ int *pmenodes; /* size npmenodes */
+ int *ddindex2simnodeid; /* size npmenodes, only with bCartesianPP
+ * but with bCartesianPP_PME */
+ gmx_ddpme_t ddpme[2];
+
+ /* The DD particle-particle nodes only */
+ gmx_bool bCartesianPP;
+ int *ddindex2ddnodeid; /* size npmenode, only with bCartesianPP_PME */
+
+ /* The global charge groups */
+ t_block cgs_gl;
+
+ /* Should we sort the cgs */
+ int nstSortCG;
+ gmx_domdec_sort_t *sort;
+
+ /* Are there bonded and multi-body interactions between charge groups? */
+ gmx_bool bInterCGBondeds;
+ gmx_bool bInterCGMultiBody;
+
+ /* Data for the optional bonded interaction atom communication range */
+ gmx_bool bBondComm;
+ t_blocka *cglink;
+ char *bLocalCG;
+
+ /* The DLB option */
+ int eDLB;
+ /* Are we actually using DLB? */
+ gmx_bool bDynLoadBal;
+
+ /* Cell sizes for static load balancing, first index cartesian */
+ real **slb_frac;
+
+ /* The width of the communicated boundaries */
+ real cutoff_mbody;
+ real cutoff;
+ /* The minimum cell size (including triclinic correction) */
+ rvec cellsize_min;
+ /* For dlb, for use with edlbAUTO */
+ rvec cellsize_min_dlb;
+ /* The lower limit for the DD cell size with DLB */
+ real cellsize_limit;
+ /* Effectively no NB cut-off limit with DLB for systems without PBC? */
+ gmx_bool bVacDLBNoLimit;
+
+ /* tric_dir is only stored here because dd_get_ns_ranges needs it */
+ ivec tric_dir;
+ /* box0 and box_size are required with dim's without pbc and -gcom */
+ rvec box0;
+ rvec box_size;
+
+ /* The cell boundaries */
+ rvec cell_x0;
+ rvec cell_x1;
+
+ /* The old location of the cell boundaries, to check cg displacements */
+ rvec old_cell_x0;
+ rvec old_cell_x1;
+
+ /* The communication setup and charge group boundaries for the zones */
+ gmx_domdec_zones_t zones;
+
+ /* The zone limits for DD dimensions 1 and 2 (not 0), determined from
+ * cell boundaries of neighboring cells for dynamic load balancing.
+ */
+ gmx_ddzone_t zone_d1[2];
+ gmx_ddzone_t zone_d2[2][2];
+
+ /* The coordinate/force communication setup and indices */
+ gmx_domdec_comm_dim_t cd[DIM];
+ /* The maximum number of cells to communicate with in one dimension */
+ int maxpulse;
+
+ /* Which cg distribution is stored on the master node */
+ int master_cg_ddp_count;
+
+ /* The number of cg's received from the direct neighbors */
+ int zone_ncg1[DD_MAXZONE];
+
+ /* The atom counts, the range for each type t is nat[t-1] <= at < nat[t] */
+ int nat[ddnatNR];
+
+ /* Communication buffer for general use */
+ int *buf_int;
+ int nalloc_int;
+
+ /* Communication buffer for general use */
+ vec_rvec_t vbuf;
+
+ /* Communication buffers only used with multiple grid pulses */
+ int *buf_int2;
+ int nalloc_int2;
+ vec_rvec_t vbuf2;
+
+ /* Communication buffers for local redistribution */
+ int **cggl_flag;
+ int cggl_flag_nalloc[DIM*2];
+ rvec **cgcm_state;
+ int cgcm_state_nalloc[DIM*2];
+
+ /* Cell sizes for dynamic load balancing */
+ gmx_domdec_root_t **root;
+ real *cell_f_row;
+ real cell_f0[DIM];
+ real cell_f1[DIM];
+ real cell_f_max0[DIM];
+ real cell_f_min1[DIM];
+
+ /* Stuff for load communication */
+ gmx_bool bRecordLoad;
+ gmx_domdec_load_t *load;
+#ifdef GMX_MPI
+ MPI_Comm *mpi_comm_load;
+#endif
+
+ /* Maximum DLB scaling per load balancing step in percent */
+ int dlb_scale_lim;
+
+ /* Cycle counters */
+ float cycl[ddCyclNr];
+ int cycl_n[ddCyclNr];
+ float cycl_max[ddCyclNr];
+ /* Flop counter (0=no,1=yes,2=with (eFlop-1)*5% noise */
+ int eFlop;
+ double flop;
+ int flop_n;
+ /* Have often have did we have load measurements */
+ int n_load_have;
+ /* Have often have we collected the load measurements */
+ int n_load_collect;
+
+ /* Statistics */
+ double sum_nat[ddnatNR-ddnatZONE];
+ int ndecomp;
+ int nload;
+ double load_step;
+ double load_sum;
+ double load_max;
+ ivec load_lim;
+ double load_mdf;
+ double load_pme;
+
+ /* The last partition step */
+ gmx_large_int_t globalcomm_step;
+
+ /* Debugging */
+ int nstDDDump;
+ int nstDDDumpGrid;
+ int DD_debug;
+} gmx_domdec_comm_t;
+
+/* The size per charge group of the cggl_flag buffer in gmx_domdec_comm_t */
+#define DD_CGIBS 2
+
+/* The flags for the cggl_flag buffer in gmx_domdec_comm_t */
+#define DD_FLAG_NRCG 65535
+#define DD_FLAG_FW(d) (1<<(16+(d)*2))
+#define DD_FLAG_BW(d) (1<<(16+(d)*2+1))
+
+/* Zone permutation required to obtain consecutive charge groups
+ * for neighbor searching.
+ */
+static const int zone_perm[3][4] = { {0,0,0,0},{1,0,0,0},{3,0,1,2} };
+
+/* dd_zo and dd_zp3/dd_zp2 are set up such that i zones with non-zero
+ * components see only j zones with that component 0.
+ */
+
+/* The DD zone order */
+static const ivec dd_zo[DD_MAXZONE] =
+ {{0,0,0},{1,0,0},{1,1,0},{0,1,0},{0,1,1},{0,0,1},{1,0,1},{1,1,1}};
+
+/* The 3D setup */
+#define dd_z3n 8
+#define dd_zp3n 4
+static const ivec dd_zp3[dd_zp3n] = {{0,0,8},{1,3,6},{2,5,6},{3,5,7}};
+
+/* The 2D setup */
+#define dd_z2n 4
+#define dd_zp2n 2
+static const ivec dd_zp2[dd_zp2n] = {{0,0,4},{1,3,4}};
+
+/* The 1D setup */
+#define dd_z1n 2
+#define dd_zp1n 1
+static const ivec dd_zp1[dd_zp1n] = {{0,0,2}};
+
+/* Factors used to avoid problems due to rounding issues */
+#define DD_CELL_MARGIN 1.0001
+#define DD_CELL_MARGIN2 1.00005
+/* Factor to account for pressure scaling during nstlist steps */
+#define DD_PRES_SCALE_MARGIN 1.02
+
+/* Allowed performance loss before we DLB or warn */
+#define DD_PERF_LOSS 0.05
+
+#define DD_CELL_F_SIZE(dd,di) ((dd)->nc[(dd)->dim[(di)]]+1+(di)*2+1+(di))
+
+/* Use separate MPI send and receive commands
+ * when nnodes <= GMX_DD_NNODES_SENDRECV.
+ * This saves memory (and some copying for small nnodes).
+ * For high parallelization scatter and gather calls are used.
+ */
+#define GMX_DD_NNODES_SENDRECV 4
+
+
+/*
+#define dd_index(n,i) ((((i)[ZZ]*(n)[YY] + (i)[YY])*(n)[XX]) + (i)[XX])
+
+static void index2xyz(ivec nc,int ind,ivec xyz)
+{
+ xyz[XX] = ind % nc[XX];
+ xyz[YY] = (ind / nc[XX]) % nc[YY];
+ xyz[ZZ] = ind / (nc[YY]*nc[XX]);
+}
+*/
+
+/* This order is required to minimize the coordinate communication in PME
+ * which uses decomposition in the x direction.
+ */
+#define dd_index(n,i) ((((i)[XX]*(n)[YY] + (i)[YY])*(n)[ZZ]) + (i)[ZZ])
+
+static void ddindex2xyz(ivec nc,int ind,ivec xyz)
+{
+ xyz[XX] = ind / (nc[YY]*nc[ZZ]);
+ xyz[YY] = (ind / nc[ZZ]) % nc[YY];
+ xyz[ZZ] = ind % nc[ZZ];
+}
+
+static int ddcoord2ddnodeid(gmx_domdec_t *dd,ivec c)
+{
+ int ddindex;
+ int ddnodeid=-1;
+
+ ddindex = dd_index(dd->nc,c);
+ if (dd->comm->bCartesianPP_PME)
+ {
+ ddnodeid = dd->comm->ddindex2ddnodeid[ddindex];
+ }
+ else if (dd->comm->bCartesianPP)
+ {
+#ifdef GMX_MPI
+ MPI_Cart_rank(dd->mpi_comm_all,c,&ddnodeid);
+#endif
+ }
+ else
+ {
+ ddnodeid = ddindex;
+ }
+
+ return ddnodeid;
+}
+
+static gmx_bool dynamic_dd_box(gmx_ddbox_t *ddbox,t_inputrec *ir)
+{
+ return (ddbox->nboundeddim < DIM || DYNAMIC_BOX(*ir));
+}
+
+int ddglatnr(gmx_domdec_t *dd,int i)
+{
+ int atnr;
+
+ if (dd == NULL)
+ {
+ atnr = i + 1;
+ }
+ else
+ {
+ if (i >= dd->comm->nat[ddnatNR-1])
+ {
+ gmx_fatal(FARGS,"glatnr called with %d, which is larger than the local number of atoms (%d)",i,dd->comm->nat[ddnatNR-1]);
+ }
+ atnr = dd->gatindex[i] + 1;
+ }
+
+ return atnr;
+}
+
+t_block *dd_charge_groups_global(gmx_domdec_t *dd)
+{
+ return &dd->comm->cgs_gl;
+}
+
+static void vec_rvec_init(vec_rvec_t *v)
+{
+ v->nalloc = 0;
+ v->v = NULL;
+}
+
+static void vec_rvec_check_alloc(vec_rvec_t *v,int n)
+{
+ if (n > v->nalloc)
+ {
+ v->nalloc = over_alloc_dd(n);
+ srenew(v->v,v->nalloc);
+ }
+}
+
+void dd_store_state(gmx_domdec_t *dd,t_state *state)
+{
+ int i;
+
+ if (state->ddp_count != dd->ddp_count)
+ {
+ gmx_incons("The state does not the domain decomposition state");
+ }
+
+ state->ncg_gl = dd->ncg_home;
+ if (state->ncg_gl > state->cg_gl_nalloc)
+ {
+ state->cg_gl_nalloc = over_alloc_dd(state->ncg_gl);
+ srenew(state->cg_gl,state->cg_gl_nalloc);
+ }
+ for(i=0; i<state->ncg_gl; i++)
+ {
+ state->cg_gl[i] = dd->index_gl[i];
+ }
+
+ state->ddp_count_cg_gl = dd->ddp_count;
+}
+
+gmx_domdec_zones_t *domdec_zones(gmx_domdec_t *dd)
+{
+ return &dd->comm->zones;
+}
+
+void dd_get_ns_ranges(gmx_domdec_t *dd,int icg,
+ int *jcg0,int *jcg1,ivec shift0,ivec shift1)
+{
+ gmx_domdec_zones_t *zones;
+ int izone,d,dim;
+
+ zones = &dd->comm->zones;
+
+ izone = 0;
+ while (icg >= zones->izone[izone].cg1)
+ {
+ izone++;
+ }
+
+ if (izone == 0)
+ {
+ *jcg0 = icg;
+ }
+ else if (izone < zones->nizone)
+ {
+ *jcg0 = zones->izone[izone].jcg0;
+ }
+ else
+ {
+ gmx_fatal(FARGS,"DD icg %d out of range: izone (%d) >= nizone (%d)",
+ icg,izone,zones->nizone);
+ }
+
+ *jcg1 = zones->izone[izone].jcg1;
+
+ for(d=0; d<dd->ndim; d++)
+ {
+ dim = dd->dim[d];
+ shift0[dim] = zones->izone[izone].shift0[dim];
+ shift1[dim] = zones->izone[izone].shift1[dim];
+ if (dd->comm->tric_dir[dim] || (dd->bGridJump && d > 0))
+ {
+ /* A conservative approach, this can be optimized */
+ shift0[dim] -= 1;
+ shift1[dim] += 1;
+ }
+ }
+}
+
+int dd_natoms_vsite(gmx_domdec_t *dd)
+{
+ return dd->comm->nat[ddnatVSITE];
+}
+
+void dd_get_constraint_range(gmx_domdec_t *dd,int *at_start,int *at_end)
+{
+ *at_start = dd->comm->nat[ddnatCON-1];
+ *at_end = dd->comm->nat[ddnatCON];
+}
+
+void dd_move_x(gmx_domdec_t *dd,matrix box,rvec x[])
+{
+ int nzone,nat_tot,n,d,p,i,j,at0,at1,zone;
+ int *index,*cgindex;
+ gmx_domdec_comm_t *comm;
+ gmx_domdec_comm_dim_t *cd;
+ gmx_domdec_ind_t *ind;
+ rvec shift={0,0,0},*buf,*rbuf;
+ gmx_bool bPBC,bScrew;
+
+ comm = dd->comm;
+
+ cgindex = dd->cgindex;
+
+ buf = comm->vbuf.v;
+
+ nzone = 1;
+ nat_tot = dd->nat_home;
+ for(d=0; d<dd->ndim; d++)
+ {
+ bPBC = (dd->ci[dd->dim[d]] == 0);
+ bScrew = (bPBC && dd->bScrewPBC && dd->dim[d] == XX);
+ if (bPBC)
+ {
+ copy_rvec(box[dd->dim[d]],shift);
+ }
+ cd = &comm->cd[d];
+ for(p=0; p<cd->np; p++)
+ {
+ ind = &cd->ind[p];
+ index = ind->index;
+ n = 0;
+ if (!bPBC)
+ {
+ for(i=0; i<ind->nsend[nzone]; i++)
+ {
+ at0 = cgindex[index[i]];
+ at1 = cgindex[index[i]+1];
+ for(j=at0; j<at1; j++)
+ {
+ copy_rvec(x[j],buf[n]);
+ n++;
+ }
+ }
+ }
+ else if (!bScrew)
+ {
+ for(i=0; i<ind->nsend[nzone]; i++)
+ {
+ at0 = cgindex[index[i]];
+ at1 = cgindex[index[i]+1];
+ for(j=at0; j<at1; j++)
+ {
+ /* We need to shift the coordinates */
+ rvec_add(x[j],shift,buf[n]);
+ n++;
+ }
+ }
+ }
+ else
+ {
+ for(i=0; i<ind->nsend[nzone]; i++)
+ {
+ at0 = cgindex[index[i]];
+ at1 = cgindex[index[i]+1];
+ for(j=at0; j<at1; j++)
+ {
+ /* Shift x */
+ buf[n][XX] = x[j][XX] + shift[XX];
+ /* Rotate y and z.
+ * This operation requires a special shift force
+ * treatment, which is performed in calc_vir.
+ */
+ buf[n][YY] = box[YY][YY] - x[j][YY];
+ buf[n][ZZ] = box[ZZ][ZZ] - x[j][ZZ];
+ n++;
+ }
+ }
+ }
+
+ if (cd->bInPlace)
+ {
+ rbuf = x + nat_tot;
+ }
+ else
+ {
+ rbuf = comm->vbuf2.v;
+ }
+ /* Send and receive the coordinates */
+ dd_sendrecv_rvec(dd, d, dddirBackward,
+ buf, ind->nsend[nzone+1],
+ rbuf, ind->nrecv[nzone+1]);
+ if (!cd->bInPlace)
+ {
+ j = 0;
+ for(zone=0; zone<nzone; zone++)
+ {
+ for(i=ind->cell2at0[zone]; i<ind->cell2at1[zone]; i++)
+ {
+ copy_rvec(rbuf[j],x[i]);
+ j++;
+ }
+ }
+ }
+ nat_tot += ind->nrecv[nzone+1];
+ }
+ nzone += nzone;
+ }
+}
+
+void dd_move_f(gmx_domdec_t *dd,rvec f[],rvec *fshift)
+{
+ int nzone,nat_tot,n,d,p,i,j,at0,at1,zone;
+ int *index,*cgindex;
+ gmx_domdec_comm_t *comm;
+ gmx_domdec_comm_dim_t *cd;
+ gmx_domdec_ind_t *ind;
+ rvec *buf,*sbuf;
+ ivec vis;
+ int is;
+ gmx_bool bPBC,bScrew;
+
+ comm = dd->comm;
+
+ cgindex = dd->cgindex;
+
+ buf = comm->vbuf.v;
+
+ n = 0;
+ nzone = comm->zones.n/2;
+ nat_tot = dd->nat_tot;
+ for(d=dd->ndim-1; d>=0; d--)
+ {
+ bPBC = (dd->ci[dd->dim[d]] == 0);
+ bScrew = (bPBC && dd->bScrewPBC && dd->dim[d] == XX);
+ if (fshift == NULL && !bScrew)
+ {
+ bPBC = FALSE;
+ }
+ /* Determine which shift vector we need */
+ clear_ivec(vis);
+ vis[dd->dim[d]] = 1;
+ is = IVEC2IS(vis);
+
+ cd = &comm->cd[d];
+ for(p=cd->np-1; p>=0; p--) {
+ ind = &cd->ind[p];
+ nat_tot -= ind->nrecv[nzone+1];
+ if (cd->bInPlace)
+ {
+ sbuf = f + nat_tot;
+ }
+ else
+ {
+ sbuf = comm->vbuf2.v;
+ j = 0;
+ for(zone=0; zone<nzone; zone++)
+ {
+ for(i=ind->cell2at0[zone]; i<ind->cell2at1[zone]; i++)
+ {
+ copy_rvec(f[i],sbuf[j]);
+ j++;
+ }
+ }
+ }
+ /* Communicate the forces */
+ dd_sendrecv_rvec(dd, d, dddirForward,
+ sbuf, ind->nrecv[nzone+1],
+ buf, ind->nsend[nzone+1]);
+ index = ind->index;
+ /* Add the received forces */
+ n = 0;
+ if (!bPBC)
+ {
+ for(i=0; i<ind->nsend[nzone]; i++)
+ {
+ at0 = cgindex[index[i]];
+ at1 = cgindex[index[i]+1];
+ for(j=at0; j<at1; j++)
+ {
+ rvec_inc(f[j],buf[n]);
+ n++;
+ }
+ }
+ }
+ else if (!bScrew)
+ {
+ for(i=0; i<ind->nsend[nzone]; i++)
+ {
+ at0 = cgindex[index[i]];
+ at1 = cgindex[index[i]+1];
+ for(j=at0; j<at1; j++)
+ {
+ rvec_inc(f[j],buf[n]);
+ /* Add this force to the shift force */
+ rvec_inc(fshift[is],buf[n]);
+ n++;
+ }
+ }
+ }
+ else
+ {
+ for(i=0; i<ind->nsend[nzone]; i++)
+ {
+ at0 = cgindex[index[i]];
+ at1 = cgindex[index[i]+1];
+ for(j=at0; j<at1; j++)
+ {
+ /* Rotate the force */
+ f[j][XX] += buf[n][XX];
+ f[j][YY] -= buf[n][YY];
+ f[j][ZZ] -= buf[n][ZZ];
+ if (fshift)
+ {
+ /* Add this force to the shift force */
+ rvec_inc(fshift[is],buf[n]);
+ }
+ n++;
+ }
+ }
+ }
+ }
+ nzone /= 2;
+ }
+}
+
+void dd_atom_spread_real(gmx_domdec_t *dd,real v[])
+{
+ int nzone,nat_tot,n,d,p,i,j,at0,at1,zone;
+ int *index,*cgindex;
+ gmx_domdec_comm_t *comm;
+ gmx_domdec_comm_dim_t *cd;
+ gmx_domdec_ind_t *ind;
+ real *buf,*rbuf;
+
+ comm = dd->comm;
+
+ cgindex = dd->cgindex;
+
+ buf = &comm->vbuf.v[0][0];
+
+ nzone = 1;
+ nat_tot = dd->nat_home;
+ for(d=0; d<dd->ndim; d++)
+ {
+ cd = &comm->cd[d];
+ for(p=0; p<cd->np; p++)
+ {
+ ind = &cd->ind[p];
+ index = ind->index;
+ n = 0;
+ for(i=0; i<ind->nsend[nzone]; i++)
+ {
+ at0 = cgindex[index[i]];
+ at1 = cgindex[index[i]+1];
+ for(j=at0; j<at1; j++)
+ {
+ buf[n] = v[j];
+ n++;
+ }
+ }
+
+ if (cd->bInPlace)
+ {
+ rbuf = v + nat_tot;
+ }
+ else
+ {
+ rbuf = &comm->vbuf2.v[0][0];
+ }
+ /* Send and receive the coordinates */
+ dd_sendrecv_real(dd, d, dddirBackward,
+ buf, ind->nsend[nzone+1],
+ rbuf, ind->nrecv[nzone+1]);
+ if (!cd->bInPlace)
+ {
+ j = 0;
+ for(zone=0; zone<nzone; zone++)
+ {
+ for(i=ind->cell2at0[zone]; i<ind->cell2at1[zone]; i++)
+ {
+ v[i] = rbuf[j];
+ j++;
+ }
+ }
+ }
+ nat_tot += ind->nrecv[nzone+1];
+ }
+ nzone += nzone;
+ }
+}
+
+void dd_atom_sum_real(gmx_domdec_t *dd,real v[])
+{
+ int nzone,nat_tot,n,d,p,i,j,at0,at1,zone;
+ int *index,*cgindex;
+ gmx_domdec_comm_t *comm;
+ gmx_domdec_comm_dim_t *cd;
+ gmx_domdec_ind_t *ind;
+ real *buf,*sbuf;
+
+ comm = dd->comm;
+
+ cgindex = dd->cgindex;
+
+ buf = &comm->vbuf.v[0][0];
+
+ n = 0;
+ nzone = comm->zones.n/2;
+ nat_tot = dd->nat_tot;
+ for(d=dd->ndim-1; d>=0; d--)
+ {
+ cd = &comm->cd[d];
+ for(p=cd->np-1; p>=0; p--) {
+ ind = &cd->ind[p];
+ nat_tot -= ind->nrecv[nzone+1];
+ if (cd->bInPlace)
+ {
+ sbuf = v + nat_tot;
+ }
+ else
+ {
+ sbuf = &comm->vbuf2.v[0][0];
+ j = 0;
+ for(zone=0; zone<nzone; zone++)
+ {
+ for(i=ind->cell2at0[zone]; i<ind->cell2at1[zone]; i++)
+ {
+ sbuf[j] = v[i];
+ j++;
+ }
+ }
+ }
+ /* Communicate the forces */
+ dd_sendrecv_real(dd, d, dddirForward,
+ sbuf, ind->nrecv[nzone+1],
+ buf, ind->nsend[nzone+1]);
+ index = ind->index;
+ /* Add the received forces */
+ n = 0;
+ for(i=0; i<ind->nsend[nzone]; i++)
+ {
+ at0 = cgindex[index[i]];
+ at1 = cgindex[index[i]+1];
+ for(j=at0; j<at1; j++)
+ {
+ v[j] += buf[n];
+ n++;
+ }
+ }
+ }
+ nzone /= 2;
+ }
+}
+
+static void print_ddzone(FILE *fp,int d,int i,int j,gmx_ddzone_t *zone)
+{
+ fprintf(fp,"zone d0 %d d1 %d d2 %d min0 %6.3f max1 %6.3f mch0 %6.3f mch1 %6.3f p1_0 %6.3f p1_1 %6.3f\n",
+ d,i,j,
+ zone->min0,zone->max1,
+ zone->mch0,zone->mch0,
+ zone->p1_0,zone->p1_1);
+}
+
+static void dd_sendrecv_ddzone(const gmx_domdec_t *dd,
+ int ddimind,int direction,
+ gmx_ddzone_t *buf_s,int n_s,
+ gmx_ddzone_t *buf_r,int n_r)
+{
+ rvec vbuf_s[5*2],vbuf_r[5*2];
+ int i;
+
+ for(i=0; i<n_s; i++)
+ {
+ vbuf_s[i*2 ][0] = buf_s[i].min0;
+ vbuf_s[i*2 ][1] = buf_s[i].max1;
+ vbuf_s[i*2 ][2] = buf_s[i].mch0;
+ vbuf_s[i*2+1][0] = buf_s[i].mch1;
+ vbuf_s[i*2+1][1] = buf_s[i].p1_0;
+ vbuf_s[i*2+1][2] = buf_s[i].p1_1;
+ }
+
+ dd_sendrecv_rvec(dd, ddimind, direction,
+ vbuf_s, n_s*2,
+ vbuf_r, n_r*2);
+
+ for(i=0; i<n_r; i++)
+ {
+ buf_r[i].min0 = vbuf_r[i*2 ][0];
+ buf_r[i].max1 = vbuf_r[i*2 ][1];
+ buf_r[i].mch0 = vbuf_r[i*2 ][2];
+ buf_r[i].mch1 = vbuf_r[i*2+1][0];
+ buf_r[i].p1_0 = vbuf_r[i*2+1][1];
+ buf_r[i].p1_1 = vbuf_r[i*2+1][2];
+ }
+}
+
+static void dd_move_cellx(gmx_domdec_t *dd,gmx_ddbox_t *ddbox,
+ rvec cell_ns_x0,rvec cell_ns_x1)
+{
+ int d,d1,dim,dim1,pos,buf_size,i,j,k,p,npulse,npulse_min;
+ gmx_ddzone_t *zp,buf_s[5],buf_r[5],buf_e[5];
+ rvec extr_s[2],extr_r[2];
+ rvec dh;
+ real dist_d,c=0,det;
+ gmx_domdec_comm_t *comm;
+ gmx_bool bPBC,bUse;
+
+ comm = dd->comm;
+
+ for(d=1; d<dd->ndim; d++)
+ {
+ dim = dd->dim[d];
+ zp = (d == 1) ? &comm->zone_d1[0] : &comm->zone_d2[0][0];
+ zp->min0 = cell_ns_x0[dim];
+ zp->max1 = cell_ns_x1[dim];
+ zp->mch0 = cell_ns_x0[dim];
+ zp->mch1 = cell_ns_x1[dim];
+ zp->p1_0 = cell_ns_x0[dim];
+ zp->p1_1 = cell_ns_x1[dim];
+ }
+
+ for(d=dd->ndim-2; d>=0; d--)
+ {
+ dim = dd->dim[d];
+ bPBC = (dim < ddbox->npbcdim);
+
+ /* Use an rvec to store two reals */
+ extr_s[d][0] = comm->cell_f0[d+1];
+ extr_s[d][1] = comm->cell_f1[d+1];
+ extr_s[d][2] = 0;
+
+ pos = 0;
+ /* Store the extremes in the backward sending buffer,
+ * so the get updated separately from the forward communication.
+ */
+ for(d1=d; d1<dd->ndim-1; d1++)
+ {
+ /* We invert the order to be able to use the same loop for buf_e */
+ buf_s[pos].min0 = extr_s[d1][1];
+ buf_s[pos].max1 = extr_s[d1][0];
+ buf_s[pos].mch0 = 0;
+ buf_s[pos].mch1 = 0;
+ /* Store the cell corner of the dimension we communicate along */
+ buf_s[pos].p1_0 = comm->cell_x0[dim];
+ buf_s[pos].p1_1 = 0;
+ pos++;
+ }
+
+ buf_s[pos] = (dd->ndim == 2) ? comm->zone_d1[0] : comm->zone_d2[0][0];
+ pos++;
+
+ if (dd->ndim == 3 && d == 0)
+ {
+ buf_s[pos] = comm->zone_d2[0][1];
+ pos++;
+ buf_s[pos] = comm->zone_d1[0];
+ pos++;
+ }
+
+ /* We only need to communicate the extremes
+ * in the forward direction
+ */
+ npulse = comm->cd[d].np;
+ if (bPBC)
+ {
+ /* Take the minimum to avoid double communication */
+ npulse_min = min(npulse,dd->nc[dim]-1-npulse);
+ }
+ else
+ {
+ /* Without PBC we should really not communicate over
+ * the boundaries, but implementing that complicates
+ * the communication setup and therefore we simply
+ * do all communication, but ignore some data.
+ */
+ npulse_min = npulse;
+ }
+ for(p=0; p<npulse_min; p++)
+ {
+ /* Communicate the extremes forward */
+ bUse = (bPBC || dd->ci[dim] > 0);
+
+ dd_sendrecv_rvec(dd, d, dddirForward,
+ extr_s+d, dd->ndim-d-1,
+ extr_r+d, dd->ndim-d-1);
+
+ if (bUse)
+ {
+ for(d1=d; d1<dd->ndim-1; d1++)
+ {
+ extr_s[d1][0] = max(extr_s[d1][0],extr_r[d1][0]);
+ extr_s[d1][1] = min(extr_s[d1][1],extr_r[d1][1]);
+ }
+ }
+ }
+
+ buf_size = pos;
+ for(p=0; p<npulse; p++)
+ {
+ /* Communicate all the zone information backward */
+ bUse = (bPBC || dd->ci[dim] < dd->nc[dim] - 1);
+
+ dd_sendrecv_ddzone(dd, d, dddirBackward,
+ buf_s, buf_size,
+ buf_r, buf_size);
+
+ clear_rvec(dh);
+ if (p > 0)
+ {
+ for(d1=d+1; d1<dd->ndim; d1++)
+ {
+ /* Determine the decrease of maximum required
+ * communication height along d1 due to the distance along d,
+ * this avoids a lot of useless atom communication.
+ */
+ dist_d = comm->cell_x1[dim] - buf_r[0].p1_0;
+
+ if (ddbox->tric_dir[dim])
+ {
+ /* c is the off-diagonal coupling between the cell planes
+ * along directions d and d1.
+ */
+ c = ddbox->v[dim][dd->dim[d1]][dim];
+ }
+ else
+ {
+ c = 0;
+ }
+ det = (1 + c*c)*comm->cutoff*comm->cutoff - dist_d*dist_d;
+ if (det > 0)
+ {
+ dh[d1] = comm->cutoff - (c*dist_d + sqrt(det))/(1 + c*c);
+ }
+ else
+ {
+ /* A negative value signals out of range */
+ dh[d1] = -1;
+ }
+ }
+ }
+
+ /* Accumulate the extremes over all pulses */
+ for(i=0; i<buf_size; i++)
+ {
+ if (p == 0)
+ {
+ buf_e[i] = buf_r[i];
+ }
+ else
+ {
+ if (bUse)
+ {
+ buf_e[i].min0 = min(buf_e[i].min0,buf_r[i].min0);
+ buf_e[i].max1 = max(buf_e[i].max1,buf_r[i].max1);
+ }
+
+ if (dd->ndim == 3 && d == 0 && i == buf_size - 1)
+ {
+ d1 = 1;
+ }
+ else
+ {
+ d1 = d + 1;
+ }
+ if (bUse && dh[d1] >= 0)
+ {
+ buf_e[i].mch0 = max(buf_e[i].mch0,buf_r[i].mch0-dh[d1]);
+ buf_e[i].mch1 = max(buf_e[i].mch1,buf_r[i].mch1-dh[d1]);
+ }
+ }
+ /* Copy the received buffer to the send buffer,
+ * to pass the data through with the next pulse.
+ */
+ buf_s[i] = buf_r[i];
+ }
+ if (((bPBC || dd->ci[dim]+npulse < dd->nc[dim]) && p == npulse-1) ||
+ (!bPBC && dd->ci[dim]+1+p == dd->nc[dim]-1))
+ {
+ /* Store the extremes */
+ pos = 0;
+
+ for(d1=d; d1<dd->ndim-1; d1++)
+ {
+ extr_s[d1][1] = min(extr_s[d1][1],buf_e[pos].min0);
+ extr_s[d1][0] = max(extr_s[d1][0],buf_e[pos].max1);
+ pos++;
+ }
+
+ if (d == 1 || (d == 0 && dd->ndim == 3))
+ {
+ for(i=d; i<2; i++)
+ {
+ comm->zone_d2[1-d][i] = buf_e[pos];
+ pos++;
+ }
+ }
+ if (d == 0)
+ {
+ comm->zone_d1[1] = buf_e[pos];
+ pos++;
+ }
+ }
+ }
+ }
+
+ if (dd->ndim >= 2)
+ {
+ dim = dd->dim[1];
+ for(i=0; i<2; i++)
+ {
+ if (debug)
+ {
+ print_ddzone(debug,1,i,0,&comm->zone_d1[i]);
+ }
+ cell_ns_x0[dim] = min(cell_ns_x0[dim],comm->zone_d1[i].min0);
+ cell_ns_x1[dim] = max(cell_ns_x1[dim],comm->zone_d1[i].max1);
+ }
+ }
+ if (dd->ndim >= 3)
+ {
+ dim = dd->dim[2];
+ for(i=0; i<2; i++)
+ {
+ for(j=0; j<2; j++)
+ {
+ if (debug)
+ {
+ print_ddzone(debug,2,i,j,&comm->zone_d2[i][j]);
+ }
+ cell_ns_x0[dim] = min(cell_ns_x0[dim],comm->zone_d2[i][j].min0);
+ cell_ns_x1[dim] = max(cell_ns_x1[dim],comm->zone_d2[i][j].max1);
+ }
+ }
+ }
+ for(d=1; d<dd->ndim; d++)
+ {
+ comm->cell_f_max0[d] = extr_s[d-1][0];
+ comm->cell_f_min1[d] = extr_s[d-1][1];
+ if (debug)
+ {
+ fprintf(debug,"Cell fraction d %d, max0 %f, min1 %f\n",
+ d,comm->cell_f_max0[d],comm->cell_f_min1[d]);
+ }
+ }
+}
+
+static void dd_collect_cg(gmx_domdec_t *dd,
+ t_state *state_local)
+{
+ gmx_domdec_master_t *ma=NULL;
+ int buf2[2],*ibuf,i,ncg_home=0,*cg=NULL,nat_home=0;
+ t_block *cgs_gl;
+
+ if (state_local->ddp_count == dd->comm->master_cg_ddp_count)
+ {
+ /* The master has the correct distribution */
+ return;
+ }
+
+ if (state_local->ddp_count == dd->ddp_count)
+ {
+ ncg_home = dd->ncg_home;
+ cg = dd->index_gl;
+ nat_home = dd->nat_home;
+ }
+ else if (state_local->ddp_count_cg_gl == state_local->ddp_count)
+ {
+ cgs_gl = &dd->comm->cgs_gl;
+
+ ncg_home = state_local->ncg_gl;
+ cg = state_local->cg_gl;
+ nat_home = 0;
+ for(i=0; i<ncg_home; i++)
+ {
+ nat_home += cgs_gl->index[cg[i]+1] - cgs_gl->index[cg[i]];
+ }
+ }
+ else
+ {
+ gmx_incons("Attempted to collect a vector for a state for which the charge group distribution is unknown");
+ }
+
+ buf2[0] = dd->ncg_home;
+ buf2[1] = dd->nat_home;
+ if (DDMASTER(dd))
+ {
+ ma = dd->ma;
+ ibuf = ma->ibuf;
+ }
+ else
+ {
+ ibuf = NULL;
+ }
+ /* Collect the charge group and atom counts on the master */
+ dd_gather(dd,2*sizeof(int),buf2,ibuf);
+
+ if (DDMASTER(dd))
+ {
+ ma->index[0] = 0;
+ for(i=0; i<dd->nnodes; i++)
+ {
+ ma->ncg[i] = ma->ibuf[2*i];
+ ma->nat[i] = ma->ibuf[2*i+1];
+ ma->index[i+1] = ma->index[i] + ma->ncg[i];
+
+ }
+ /* Make byte counts and indices */
+ for(i=0; i<dd->nnodes; i++)
+ {
+ ma->ibuf[i] = ma->ncg[i]*sizeof(int);
+ ma->ibuf[dd->nnodes+i] = ma->index[i]*sizeof(int);
+ }
+ if (debug)
+ {
+ fprintf(debug,"Initial charge group distribution: ");
+ for(i=0; i<dd->nnodes; i++)
+ fprintf(debug," %d",ma->ncg[i]);
+ fprintf(debug,"\n");
+ }
+ }
+
+ /* Collect the charge group indices on the master */
+ dd_gatherv(dd,
+ dd->ncg_home*sizeof(int),dd->index_gl,
+ DDMASTER(dd) ? ma->ibuf : NULL,
+ DDMASTER(dd) ? ma->ibuf+dd->nnodes : NULL,
+ DDMASTER(dd) ? ma->cg : NULL);
+
+ dd->comm->master_cg_ddp_count = state_local->ddp_count;
+}
+
+static void dd_collect_vec_sendrecv(gmx_domdec_t *dd,
+ rvec *lv,rvec *v)
+{
+ gmx_domdec_master_t *ma;
+ int n,i,c,a,nalloc=0;
+ rvec *buf=NULL;
+ t_block *cgs_gl;
+
+ ma = dd->ma;
+
+ if (!DDMASTER(dd))
+ {
+#ifdef GMX_MPI
+ MPI_Send(lv,dd->nat_home*sizeof(rvec),MPI_BYTE,DDMASTERRANK(dd),
+ dd->rank,dd->mpi_comm_all);
+#endif
+ } else {
+ /* Copy the master coordinates to the global array */
+ cgs_gl = &dd->comm->cgs_gl;
+
+ n = DDMASTERRANK(dd);
+ a = 0;
+ for(i=ma->index[n]; i<ma->index[n+1]; i++)
+ {
+ for(c=cgs_gl->index[ma->cg[i]]; c<cgs_gl->index[ma->cg[i]+1]; c++)
+ {
+ copy_rvec(lv[a++],v[c]);
+ }
+ }
+
+ for(n=0; n<dd->nnodes; n++)
+ {
+ if (n != dd->rank)
+ {
+ if (ma->nat[n] > nalloc)
+ {
+ nalloc = over_alloc_dd(ma->nat[n]);
+ srenew(buf,nalloc);
+ }
+#ifdef GMX_MPI
+ MPI_Recv(buf,ma->nat[n]*sizeof(rvec),MPI_BYTE,DDRANK(dd,n),
+ n,dd->mpi_comm_all,MPI_STATUS_IGNORE);
+#endif
+ a = 0;
+ for(i=ma->index[n]; i<ma->index[n+1]; i++)
+ {
+ for(c=cgs_gl->index[ma->cg[i]]; c<cgs_gl->index[ma->cg[i]+1]; c++)
+ {
+ copy_rvec(buf[a++],v[c]);
+ }
+ }
+ }
+ }
+ sfree(buf);
+ }
+}
+
+static void get_commbuffer_counts(gmx_domdec_t *dd,
+ int **counts,int **disps)
+{
+ gmx_domdec_master_t *ma;
+ int n;
+
+ ma = dd->ma;
+
+ /* Make the rvec count and displacment arrays */
+ *counts = ma->ibuf;
+ *disps = ma->ibuf + dd->nnodes;
+ for(n=0; n<dd->nnodes; n++)
+ {
+ (*counts)[n] = ma->nat[n]*sizeof(rvec);
+ (*disps)[n] = (n == 0 ? 0 : (*disps)[n-1] + (*counts)[n-1]);
+ }
+}
+
+static void dd_collect_vec_gatherv(gmx_domdec_t *dd,
+ rvec *lv,rvec *v)
+{
+ gmx_domdec_master_t *ma;
+ int *rcounts=NULL,*disps=NULL;
+ int n,i,c,a;
+ rvec *buf=NULL;
+ t_block *cgs_gl;
+
+ ma = dd->ma;
+
+ if (DDMASTER(dd))
+ {
+ get_commbuffer_counts(dd,&rcounts,&disps);
+
+ buf = ma->vbuf;
+ }
+
+ dd_gatherv(dd,dd->nat_home*sizeof(rvec),lv,rcounts,disps,buf);
+
+ if (DDMASTER(dd))
+ {
+ cgs_gl = &dd->comm->cgs_gl;
+
+ a = 0;
+ for(n=0; n<dd->nnodes; n++)
+ {
+ for(i=ma->index[n]; i<ma->index[n+1]; i++)
+ {
+ for(c=cgs_gl->index[ma->cg[i]]; c<cgs_gl->index[ma->cg[i]+1]; c++)
+ {
+ copy_rvec(buf[a++],v[c]);
+ }
+ }
+ }
+ }
+}
+
+void dd_collect_vec(gmx_domdec_t *dd,
+ t_state *state_local,rvec *lv,rvec *v)
+{
+ gmx_domdec_master_t *ma;
+ int n,i,c,a,nalloc=0;
+ rvec *buf=NULL;
+
+ dd_collect_cg(dd,state_local);
+
+ if (dd->nnodes <= GMX_DD_NNODES_SENDRECV)
+ {
+ dd_collect_vec_sendrecv(dd,lv,v);
+ }
+ else
+ {
+ dd_collect_vec_gatherv(dd,lv,v);
+ }
+}
+
+
+void dd_collect_state(gmx_domdec_t *dd,
+ t_state *state_local,t_state *state)
+{
+ int est,i,j,nh;
+
+ nh = state->nhchainlength;
+
+ if (DDMASTER(dd))
+ {
+ state->lambda = state_local->lambda;
+ state->veta = state_local->veta;
+ state->vol0 = state_local->vol0;
+ copy_mat(state_local->box,state->box);
+ copy_mat(state_local->boxv,state->boxv);
+ copy_mat(state_local->svir_prev,state->svir_prev);
+ copy_mat(state_local->fvir_prev,state->fvir_prev);
+ copy_mat(state_local->pres_prev,state->pres_prev);
+
+
+ for(i=0; i<state_local->ngtc; i++)
+ {
+ for(j=0; j<nh; j++) {
+ state->nosehoover_xi[i*nh+j] = state_local->nosehoover_xi[i*nh+j];
+ state->nosehoover_vxi[i*nh+j] = state_local->nosehoover_vxi[i*nh+j];
+ }
+ state->therm_integral[i] = state_local->therm_integral[i];
+ }
+ for(i=0; i<state_local->nnhpres; i++)
+ {
+ for(j=0; j<nh; j++) {
+ state->nhpres_xi[i*nh+j] = state_local->nhpres_xi[i*nh+j];
+ state->nhpres_vxi[i*nh+j] = state_local->nhpres_vxi[i*nh+j];
+ }
+ }
+ }
+ for(est=0; est<estNR; est++)
+ {
+ if (EST_DISTR(est) && (state_local->flags & (1<<est)))
+ {
+ switch (est) {
+ case estX:
+ dd_collect_vec(dd,state_local,state_local->x,state->x);
+ break;
+ case estV:
+ dd_collect_vec(dd,state_local,state_local->v,state->v);
+ break;
+ case estSDX:
+ dd_collect_vec(dd,state_local,state_local->sd_X,state->sd_X);
+ break;
+ case estCGP:
+ dd_collect_vec(dd,state_local,state_local->cg_p,state->cg_p);
+ break;
+ case estLD_RNG:
+ if (state->nrngi == 1)
+ {
+ if (DDMASTER(dd))
+ {
+ for(i=0; i<state_local->nrng; i++)
+ {
+ state->ld_rng[i] = state_local->ld_rng[i];
+ }
+ }
+ }
+ else
+ {
+ dd_gather(dd,state_local->nrng*sizeof(state->ld_rng[0]),
+ state_local->ld_rng,state->ld_rng);
+ }
+ break;
+ case estLD_RNGI:
+ if (state->nrngi == 1)
+ {
+ if (DDMASTER(dd))
+ {
+ state->ld_rngi[0] = state_local->ld_rngi[0];
+ }
+ }
+ else
+ {
+ dd_gather(dd,sizeof(state->ld_rngi[0]),
+ state_local->ld_rngi,state->ld_rngi);
+ }
+ break;
+ case estDISRE_INITF:
+ case estDISRE_RM3TAV:
+ case estORIRE_INITF:
+ case estORIRE_DTAV:
+ break;
+ default:
+ gmx_incons("Unknown state entry encountered in dd_collect_state");
+ }
+ }
+ }
+}
+
+static void dd_realloc_fr_cg(t_forcerec *fr,int nalloc)
+{
+ if (debug)
+ {
+ fprintf(debug,"Reallocating forcerec: currently %d, required %d, allocating %d\n",fr->cg_nalloc,nalloc,over_alloc_dd(nalloc));
+ }
+ fr->cg_nalloc = over_alloc_dd(nalloc);
+ srenew(fr->cg_cm,fr->cg_nalloc);
+ srenew(fr->cginfo,fr->cg_nalloc);
+}
+
+static void dd_realloc_state(t_state *state,rvec **f,int nalloc)
+{
+ int est;
+
+ if (debug)
+ {
+ fprintf(debug,"Reallocating state: currently %d, required %d, allocating %d\n",state->nalloc,nalloc,over_alloc_dd(nalloc));
+ }
+
+ state->nalloc = over_alloc_dd(nalloc);
+
+ for(est=0; est<estNR; est++)
+ {
+ if (EST_DISTR(est) && (state->flags & (1<<est)))
+ {
+ switch(est) {
+ case estX:
+ srenew(state->x,state->nalloc);
+ break;
+ case estV:
+ srenew(state->v,state->nalloc);
+ break;
+ case estSDX:
+ srenew(state->sd_X,state->nalloc);
+ break;
+ case estCGP:
+ srenew(state->cg_p,state->nalloc);
+ break;
+ case estLD_RNG:
+ case estLD_RNGI:
+ case estDISRE_INITF:
+ case estDISRE_RM3TAV:
+ case estORIRE_INITF:
+ case estORIRE_DTAV:
+ /* No reallocation required */
+ break;
+ default:
+ gmx_incons("Unknown state entry encountered in dd_realloc_state");
+ }
+ }
+ }
+
+ if (f != NULL)
+ {
+ srenew(*f,state->nalloc);
+ }
+}
+
+static void dd_distribute_vec_sendrecv(gmx_domdec_t *dd,t_block *cgs,
+ rvec *v,rvec *lv)
+{
+ gmx_domdec_master_t *ma;
+ int n,i,c,a,nalloc=0;
+ rvec *buf=NULL;
+
+ if (DDMASTER(dd))
+ {
+ ma = dd->ma;
+
+ for(n=0; n<dd->nnodes; n++)
+ {
+ if (n != dd->rank)
+ {
+ if (ma->nat[n] > nalloc)
+ {
+ nalloc = over_alloc_dd(ma->nat[n]);
+ srenew(buf,nalloc);
+ }
+ /* Use lv as a temporary buffer */
+ a = 0;
+ for(i=ma->index[n]; i<ma->index[n+1]; i++)
+ {
+ for(c=cgs->index[ma->cg[i]]; c<cgs->index[ma->cg[i]+1]; c++)
+ {
+ copy_rvec(v[c],buf[a++]);
+ }
+ }
+ if (a != ma->nat[n])
+ {
+ gmx_fatal(FARGS,"Internal error a (%d) != nat (%d)",
+ a,ma->nat[n]);
+ }
+
+#ifdef GMX_MPI
+ MPI_Send(buf,ma->nat[n]*sizeof(rvec),MPI_BYTE,
+ DDRANK(dd,n),n,dd->mpi_comm_all);
+#endif
+ }
+ }
+ sfree(buf);
+ n = DDMASTERRANK(dd);
+ a = 0;
+ for(i=ma->index[n]; i<ma->index[n+1]; i++)
+ {
+ for(c=cgs->index[ma->cg[i]]; c<cgs->index[ma->cg[i]+1]; c++)
+ {
+ copy_rvec(v[c],lv[a++]);
+ }
+ }
+ }
+ else
+ {
+#ifdef GMX_MPI
+ MPI_Recv(lv,dd->nat_home*sizeof(rvec),MPI_BYTE,DDMASTERRANK(dd),
+ MPI_ANY_TAG,dd->mpi_comm_all,MPI_STATUS_IGNORE);
+#endif
+ }
+}
+
+static void dd_distribute_vec_scatterv(gmx_domdec_t *dd,t_block *cgs,
+ rvec *v,rvec *lv)
+{
+ gmx_domdec_master_t *ma;
+ int *scounts=NULL,*disps=NULL;
+ int n,i,c,a,nalloc=0;
+ rvec *buf=NULL;
+
+ if (DDMASTER(dd))
+ {
+ ma = dd->ma;
+
+ get_commbuffer_counts(dd,&scounts,&disps);
+
+ buf = ma->vbuf;
+ a = 0;
+ for(n=0; n<dd->nnodes; n++)
+ {
+ for(i=ma->index[n]; i<ma->index[n+1]; i++)
+ {
+ for(c=cgs->index[ma->cg[i]]; c<cgs->index[ma->cg[i]+1]; c++)
+ {
+ copy_rvec(v[c],buf[a++]);
+ }
+ }
+ }
+ }
+
+ dd_scatterv(dd,scounts,disps,buf,dd->nat_home*sizeof(rvec),lv);
+}
+
+static void dd_distribute_vec(gmx_domdec_t *dd,t_block *cgs,rvec *v,rvec *lv)
+{
+ if (dd->nnodes <= GMX_DD_NNODES_SENDRECV)
+ {
+ dd_distribute_vec_sendrecv(dd,cgs,v,lv);
+ }
+ else
+ {
+ dd_distribute_vec_scatterv(dd,cgs,v,lv);
+ }
+}
+
+static void dd_distribute_state(gmx_domdec_t *dd,t_block *cgs,
+ t_state *state,t_state *state_local,
+ rvec **f)
+{
+ int i,j,ngtch,ngtcp,nh;
+
+ nh = state->nhchainlength;
+
+ if (DDMASTER(dd))
+ {
+ state_local->lambda = state->lambda;
+ state_local->veta = state->veta;
+ state_local->vol0 = state->vol0;
+ copy_mat(state->box,state_local->box);
+ copy_mat(state->box_rel,state_local->box_rel);
+ copy_mat(state->boxv,state_local->boxv);
+ copy_mat(state->svir_prev,state_local->svir_prev);
+ copy_mat(state->fvir_prev,state_local->fvir_prev);
+ for(i=0; i<state_local->ngtc; i++)
+ {
+ for(j=0; j<nh; j++) {
+ state_local->nosehoover_xi[i*nh+j] = state->nosehoover_xi[i*nh+j];
+ state_local->nosehoover_vxi[i*nh+j] = state->nosehoover_vxi[i*nh+j];
+ }
+ state_local->therm_integral[i] = state->therm_integral[i];
+ }
+ for(i=0; i<state_local->nnhpres; i++)
+ {
+ for(j=0; j<nh; j++) {
+ state_local->nhpres_xi[i*nh+j] = state->nhpres_xi[i*nh+j];
+ state_local->nhpres_vxi[i*nh+j] = state->nhpres_vxi[i*nh+j];
+ }
+ }
+ }
+ dd_bcast(dd,sizeof(real),&state_local->lambda);
+ dd_bcast(dd,sizeof(real),&state_local->veta);
+ dd_bcast(dd,sizeof(real),&state_local->vol0);
+ dd_bcast(dd,sizeof(state_local->box),state_local->box);
+ dd_bcast(dd,sizeof(state_local->box_rel),state_local->box_rel);
+ dd_bcast(dd,sizeof(state_local->boxv),state_local->boxv);
+ dd_bcast(dd,sizeof(state_local->svir_prev),state_local->svir_prev);
+ dd_bcast(dd,sizeof(state_local->fvir_prev),state_local->fvir_prev);
+ dd_bcast(dd,((state_local->ngtc*nh)*sizeof(double)),state_local->nosehoover_xi);
+ dd_bcast(dd,((state_local->ngtc*nh)*sizeof(double)),state_local->nosehoover_vxi);
+ dd_bcast(dd,state_local->ngtc*sizeof(double),state_local->therm_integral);
+ dd_bcast(dd,((state_local->nnhpres*nh)*sizeof(double)),state_local->nhpres_xi);
+ dd_bcast(dd,((state_local->nnhpres*nh)*sizeof(double)),state_local->nhpres_vxi);
+
+ if (dd->nat_home > state_local->nalloc)
+ {
+ dd_realloc_state(state_local,f,dd->nat_home);
+ }
+ for(i=0; i<estNR; i++)
+ {
+ if (EST_DISTR(i) && (state_local->flags & (1<<i)))
+ {
+ switch (i) {
+ case estX:
+ dd_distribute_vec(dd,cgs,state->x,state_local->x);
+ break;
+ case estV:
+ dd_distribute_vec(dd,cgs,state->v,state_local->v);
+ break;
+ case estSDX:
+ dd_distribute_vec(dd,cgs,state->sd_X,state_local->sd_X);
+ break;
+ case estCGP:
+ dd_distribute_vec(dd,cgs,state->cg_p,state_local->cg_p);
+ break;
+ case estLD_RNG:
+ if (state->nrngi == 1)
+ {
+ dd_bcastc(dd,
+ state_local->nrng*sizeof(state_local->ld_rng[0]),
+ state->ld_rng,state_local->ld_rng);
+ }
+ else
+ {
+ dd_scatter(dd,
+ state_local->nrng*sizeof(state_local->ld_rng[0]),
+ state->ld_rng,state_local->ld_rng);
+ }
+ break;
+ case estLD_RNGI:
+ if (state->nrngi == 1)
+ {
+ dd_bcastc(dd,sizeof(state_local->ld_rngi[0]),
+ state->ld_rngi,state_local->ld_rngi);
+ }
+ else
+ {
+ dd_scatter(dd,sizeof(state_local->ld_rngi[0]),
+ state->ld_rngi,state_local->ld_rngi);
+ }
+ break;
+ case estDISRE_INITF:
+ case estDISRE_RM3TAV:
+ case estORIRE_INITF:
+ case estORIRE_DTAV:
+ /* Not implemented yet */
+ break;
+ default:
+ gmx_incons("Unknown state entry encountered in dd_distribute_state");
+ }
+ }
+ }
+}
+
+static char dim2char(int dim)
+{
+ char c='?';
+
+ switch (dim)
+ {
+ case XX: c = 'X'; break;
+ case YY: c = 'Y'; break;
+ case ZZ: c = 'Z'; break;
+ default: gmx_fatal(FARGS,"Unknown dim %d",dim);
+ }
+
+ return c;
+}
+
+static void write_dd_grid_pdb(const char *fn,gmx_large_int_t step,
+ gmx_domdec_t *dd,matrix box,gmx_ddbox_t *ddbox)
+{
+ rvec grid_s[2],*grid_r=NULL,cx,r;
+ char fname[STRLEN],format[STRLEN],buf[22];
+ FILE *out;
+ int a,i,d,z,y,x;
+ matrix tric;
+ real vol;
+
+ copy_rvec(dd->comm->cell_x0,grid_s[0]);
+ copy_rvec(dd->comm->cell_x1,grid_s[1]);
+
+ if (DDMASTER(dd))
+ {
+ snew(grid_r,2*dd->nnodes);
+ }
+
+ dd_gather(dd,2*sizeof(rvec),grid_s[0],DDMASTER(dd) ? grid_r[0] : NULL);
+
+ if (DDMASTER(dd))
+ {
+ for(d=0; d<DIM; d++)
+ {
+ for(i=0; i<DIM; i++)
+ {
+ if (d == i)
+ {
+ tric[d][i] = 1;
+ }
+ else
+ {
+ if (dd->nc[d] > 1 && d < ddbox->npbcdim)
+ {
+ tric[d][i] = box[i][d]/box[i][i];
+ }
+ else
+ {
+ tric[d][i] = 0;
+ }
+ }
+ }
+ }
+ sprintf(fname,"%s_%s.pdb",fn,gmx_step_str(step,buf));
+ sprintf(format,"%s%s\n",pdbformat,"%6.2f%6.2f");
+ out = gmx_fio_fopen(fname,"w");
+ gmx_write_pdb_box(out,dd->bScrewPBC ? epbcSCREW : epbcXYZ,box);
+ a = 1;
+ for(i=0; i<dd->nnodes; i++)
+ {
+ vol = dd->nnodes/(box[XX][XX]*box[YY][YY]*box[ZZ][ZZ]);
+ for(d=0; d<DIM; d++)
+ {
+ vol *= grid_r[i*2+1][d] - grid_r[i*2][d];
+ }
+ for(z=0; z<2; z++)
+ {
+ for(y=0; y<2; y++)
+ {
+ for(x=0; x<2; x++)
+ {
+ cx[XX] = grid_r[i*2+x][XX];
+ cx[YY] = grid_r[i*2+y][YY];
+ cx[ZZ] = grid_r[i*2+z][ZZ];
+ mvmul(tric,cx,r);
+ fprintf(out,format,"ATOM",a++,"CA","GLY",' ',1+i,
+ 10*r[XX],10*r[YY],10*r[ZZ],1.0,vol);
+ }
+ }
+ }
+ for(d=0; d<DIM; d++)
+ {
+ for(x=0; x<4; x++)
+ {
+ switch(d)
+ {
+ case 0: y = 1 + i*8 + 2*x; break;
+ case 1: y = 1 + i*8 + 2*x - (x % 2); break;
+ case 2: y = 1 + i*8 + x; break;
+ }
+ fprintf(out,"%6s%5d%5d\n","CONECT",y,y+(1<<d));
+ }
+ }
+ }
+ gmx_fio_fclose(out);
+ sfree(grid_r);
+ }
+}
+
+void write_dd_pdb(const char *fn,gmx_large_int_t step,const char *title,
+ gmx_mtop_t *mtop,t_commrec *cr,
+ int natoms,rvec x[],matrix box)
+{
+ char fname[STRLEN],format[STRLEN],format4[STRLEN],buf[22];
+ FILE *out;
+ int i,ii,resnr,c;
+ char *atomname,*resname;
+ real b;
+ gmx_domdec_t *dd;
+
+ dd = cr->dd;
+ if (natoms == -1)
+ {
+ natoms = dd->comm->nat[ddnatVSITE];
+ }
+
+ sprintf(fname,"%s_%s_n%d.pdb",fn,gmx_step_str(step,buf),cr->sim_nodeid);
+
+ sprintf(format,"%s%s\n",pdbformat,"%6.2f%6.2f");
+ sprintf(format4,"%s%s\n",pdbformat4,"%6.2f%6.2f");
+
+ out = gmx_fio_fopen(fname,"w");
+
+ fprintf(out,"TITLE %s\n",title);
+ gmx_write_pdb_box(out,dd->bScrewPBC ? epbcSCREW : epbcXYZ,box);
+ for(i=0; i<natoms; i++)
+ {
+ ii = dd->gatindex[i];
+ gmx_mtop_atominfo_global(mtop,ii,&atomname,&resnr,&resname);
+ if (i < dd->comm->nat[ddnatZONE])
+ {
+ c = 0;
+ while (i >= dd->cgindex[dd->comm->zones.cg_range[c+1]])
+ {
+ c++;
+ }
+ b = c;
+ }
+ else if (i < dd->comm->nat[ddnatVSITE])
+ {
+ b = dd->comm->zones.n;
+ }
+ else
+ {
+ b = dd->comm->zones.n + 1;
+ }
+ fprintf(out,strlen(atomname)<4 ? format : format4,
+ "ATOM",(ii+1)%100000,
+ atomname,resname,' ',resnr%10000,' ',
+ 10*x[i][XX],10*x[i][YY],10*x[i][ZZ],1.0,b);
+ }
+ fprintf(out,"TER\n");
+
+ gmx_fio_fclose(out);
+}
+
+real dd_cutoff_mbody(gmx_domdec_t *dd)
+{
+ gmx_domdec_comm_t *comm;
+ int di;
+ real r;
+
+ comm = dd->comm;
+
+ r = -1;
+ if (comm->bInterCGBondeds)
+ {
+ if (comm->cutoff_mbody > 0)
+ {
+ r = comm->cutoff_mbody;
+ }
+ else
+ {
+ /* cutoff_mbody=0 means we do not have DLB */
+ r = comm->cellsize_min[dd->dim[0]];
+ for(di=1; di<dd->ndim; di++)
+ {
+ r = min(r,comm->cellsize_min[dd->dim[di]]);
+ }
+ if (comm->bBondComm)
+ {
+ r = max(r,comm->cutoff_mbody);
+ }
+ else
+ {
+ r = min(r,comm->cutoff);
+ }
+ }
+ }
+
+ return r;
+}
+
+real dd_cutoff_twobody(gmx_domdec_t *dd)
+{
+ real r_mb;
+
+ r_mb = dd_cutoff_mbody(dd);
+
+ return max(dd->comm->cutoff,r_mb);
+}
+
+
+static void dd_cart_coord2pmecoord(gmx_domdec_t *dd,ivec coord,ivec coord_pme)
+{
+ int nc,ntot;
+
+ nc = dd->nc[dd->comm->cartpmedim];
+ ntot = dd->comm->ntot[dd->comm->cartpmedim];
+ copy_ivec(coord,coord_pme);
+ coord_pme[dd->comm->cartpmedim] =
+ nc + (coord[dd->comm->cartpmedim]*(ntot - nc) + (ntot - nc)/2)/nc;
+}
+
+static int low_ddindex2pmeindex(int ndd,int npme,int ddindex)
+{
+ /* Here we assign a PME node to communicate with this DD node
+ * by assuming that the major index of both is x.
+ * We add cr->npmenodes/2 to obtain an even distribution.
+ */
+ return (ddindex*npme + npme/2)/ndd;
+}
+
+static int ddindex2pmeindex(const gmx_domdec_t *dd,int ddindex)
+{
+ return low_ddindex2pmeindex(dd->nnodes,dd->comm->npmenodes,ddindex);
+}
+
+static int cr_ddindex2pmeindex(const t_commrec *cr,int ddindex)
+{
+ return low_ddindex2pmeindex(cr->dd->nnodes,cr->npmenodes,ddindex);
+}
+
+static int *dd_pmenodes(t_commrec *cr)
+{
+ int *pmenodes;
+ int n,i,p0,p1;
+
+ snew(pmenodes,cr->npmenodes);
+ n = 0;
+ for(i=0; i<cr->dd->nnodes; i++) {
+ p0 = cr_ddindex2pmeindex(cr,i);
+ p1 = cr_ddindex2pmeindex(cr,i+1);
+ if (i+1 == cr->dd->nnodes || p1 > p0) {
+ if (debug)
+ fprintf(debug,"pmenode[%d] = %d\n",n,i+1+n);
+ pmenodes[n] = i + 1 + n;
+ n++;
+ }
+ }
+
+ return pmenodes;
+}
+
+static int gmx_ddcoord2pmeindex(t_commrec *cr,int x,int y,int z)
+{
+ gmx_domdec_t *dd;
+ ivec coords,coords_pme,nc;
+ int slab;
+
+ dd = cr->dd;
+ /*
+ if (dd->comm->bCartesian) {
+ gmx_ddindex2xyz(dd->nc,ddindex,coords);
+ dd_coords2pmecoords(dd,coords,coords_pme);
+ copy_ivec(dd->ntot,nc);
+ nc[dd->cartpmedim] -= dd->nc[dd->cartpmedim];
+ coords_pme[dd->cartpmedim] -= dd->nc[dd->cartpmedim];
+
+ slab = (coords_pme[XX]*nc[YY] + coords_pme[YY])*nc[ZZ] + coords_pme[ZZ];
+ } else {
+ slab = (ddindex*cr->npmenodes + cr->npmenodes/2)/dd->nnodes;
+ }
+ */
+ coords[XX] = x;
+ coords[YY] = y;
+ coords[ZZ] = z;
+ slab = ddindex2pmeindex(dd,dd_index(dd->nc,coords));
+
+ return slab;
+}
+
+static int ddcoord2simnodeid(t_commrec *cr,int x,int y,int z)
+{
+ gmx_domdec_comm_t *comm;
+ ivec coords;
+ int ddindex,nodeid=-1;
+
+ comm = cr->dd->comm;
+
+ coords[XX] = x;
+ coords[YY] = y;
+ coords[ZZ] = z;
+ if (comm->bCartesianPP_PME)
+ {
+#ifdef GMX_MPI
+ MPI_Cart_rank(cr->mpi_comm_mysim,coords,&nodeid);
+#endif
+ }
+ else
+ {
+ ddindex = dd_index(cr->dd->nc,coords);
+ if (comm->bCartesianPP)
+ {
+ nodeid = comm->ddindex2simnodeid[ddindex];
+ }
+ else
+ {
+ if (comm->pmenodes)
+ {
+ nodeid = ddindex + gmx_ddcoord2pmeindex(cr,x,y,z);
+ }
+ else
+ {
+ nodeid = ddindex;
+ }
+ }
+ }
+
+ return nodeid;
+}
+
+static int dd_simnode2pmenode(t_commrec *cr,int sim_nodeid)
+{
+ gmx_domdec_t *dd;
+ gmx_domdec_comm_t *comm;
+ ivec coord,coord_pme;
+ int i;
+ int pmenode=-1;
+
+ dd = cr->dd;
+ comm = dd->comm;
+
+ /* This assumes a uniform x domain decomposition grid cell size */
+ if (comm->bCartesianPP_PME)
+ {
+#ifdef GMX_MPI
+ MPI_Cart_coords(cr->mpi_comm_mysim,sim_nodeid,DIM,coord);
+ if (coord[comm->cartpmedim] < dd->nc[comm->cartpmedim])
+ {
+ /* This is a PP node */
+ dd_cart_coord2pmecoord(dd,coord,coord_pme);
+ MPI_Cart_rank(cr->mpi_comm_mysim,coord_pme,&pmenode);
+ }
+#endif
+ }
+ else if (comm->bCartesianPP)
+ {
+ if (sim_nodeid < dd->nnodes)
+ {
+ pmenode = dd->nnodes + ddindex2pmeindex(dd,sim_nodeid);
+ }
+ }
+ else
+ {
+ /* This assumes DD cells with identical x coordinates
+ * are numbered sequentially.
+ */
+ if (dd->comm->pmenodes == NULL)
+ {
+ if (sim_nodeid < dd->nnodes)
+ {
+ /* The DD index equals the nodeid */
+ pmenode = dd->nnodes + ddindex2pmeindex(dd,sim_nodeid);
+ }
+ }
+ else
+ {
+ i = 0;
+ while (sim_nodeid > dd->comm->pmenodes[i])
+ {
+ i++;
+ }
+ if (sim_nodeid < dd->comm->pmenodes[i])
+ {
+ pmenode = dd->comm->pmenodes[i];
+ }
+ }
+ }
+
+ return pmenode;
+}
+
+gmx_bool gmx_pmeonlynode(t_commrec *cr,int sim_nodeid)
+{
+ gmx_bool bPMEOnlyNode;
+
+ if (DOMAINDECOMP(cr))
+ {
+ bPMEOnlyNode = (dd_simnode2pmenode(cr,sim_nodeid) == -1);
+ }
+ else
+ {
+ bPMEOnlyNode = FALSE;
+ }
+
+ return bPMEOnlyNode;
+}
+
+void get_pme_ddnodes(t_commrec *cr,int pmenodeid,
+ int *nmy_ddnodes,int **my_ddnodes,int *node_peer)
+{
+ gmx_domdec_t *dd;
+ int x,y,z;
+ ivec coord,coord_pme;
+
+ dd = cr->dd;
+
+ snew(*my_ddnodes,(dd->nnodes+cr->npmenodes-1)/cr->npmenodes);
+
+ *nmy_ddnodes = 0;
+ for(x=0; x<dd->nc[XX]; x++)
+ {
+ for(y=0; y<dd->nc[YY]; y++)
+ {
+ for(z=0; z<dd->nc[ZZ]; z++)
+ {
+ if (dd->comm->bCartesianPP_PME)
+ {
+ coord[XX] = x;
+ coord[YY] = y;
+ coord[ZZ] = z;
+ dd_cart_coord2pmecoord(dd,coord,coord_pme);
+ if (dd->ci[XX] == coord_pme[XX] &&
+ dd->ci[YY] == coord_pme[YY] &&
+ dd->ci[ZZ] == coord_pme[ZZ])
+ (*my_ddnodes)[(*nmy_ddnodes)++] = ddcoord2simnodeid(cr,x,y,z);
+ }
+ else
+ {
+ /* The slab corresponds to the nodeid in the PME group */
+ if (gmx_ddcoord2pmeindex(cr,x,y,z) == pmenodeid)
+ {
+ (*my_ddnodes)[(*nmy_ddnodes)++] = ddcoord2simnodeid(cr,x,y,z);
+ }
+ }
+ }
+ }
+ }
+
+ /* The last PP-only node is the peer node */
+ *node_peer = (*my_ddnodes)[*nmy_ddnodes-1];
+
+ if (debug)
+ {
+ fprintf(debug,"Receive coordinates from PP nodes:");
+ for(x=0; x<*nmy_ddnodes; x++)
+ {
+ fprintf(debug," %d",(*my_ddnodes)[x]);
+ }
+ fprintf(debug,"\n");
+ }
+}
+
+static gmx_bool receive_vir_ener(t_commrec *cr)
+{
+ gmx_domdec_comm_t *comm;
+ int pmenode,coords[DIM],rank;
+ gmx_bool bReceive;
+
+ bReceive = TRUE;
+ if (cr->npmenodes < cr->dd->nnodes)
+ {
+ comm = cr->dd->comm;
+ if (comm->bCartesianPP_PME)
+ {
+ pmenode = dd_simnode2pmenode(cr,cr->sim_nodeid);
+#ifdef GMX_MPI
+ MPI_Cart_coords(cr->mpi_comm_mysim,cr->sim_nodeid,DIM,coords);
+ coords[comm->cartpmedim]++;
+ if (coords[comm->cartpmedim] < cr->dd->nc[comm->cartpmedim])
+ {
+ MPI_Cart_rank(cr->mpi_comm_mysim,coords,&rank);
+ if (dd_simnode2pmenode(cr,rank) == pmenode)
+ {
+ /* This is not the last PP node for pmenode */
+ bReceive = FALSE;
+ }
+ }
+#endif
+ }
+ else
+ {
+ pmenode = dd_simnode2pmenode(cr,cr->sim_nodeid);
+ if (cr->sim_nodeid+1 < cr->nnodes &&
+ dd_simnode2pmenode(cr,cr->sim_nodeid+1) == pmenode)
+ {
+ /* This is not the last PP node for pmenode */
+ bReceive = FALSE;
+ }
+ }
+ }
+
+ return bReceive;
+}
+
+static void set_zones_ncg_home(gmx_domdec_t *dd)
+{
+ gmx_domdec_zones_t *zones;
+ int i;
+
+ zones = &dd->comm->zones;
+
+ zones->cg_range[0] = 0;
+ for(i=1; i<zones->n+1; i++)
+ {
+ zones->cg_range[i] = dd->ncg_home;
+ }
+}
+
+static void rebuild_cgindex(gmx_domdec_t *dd,int *gcgs_index,t_state *state)
+{
+ int nat,i,*ind,*dd_cg_gl,*cgindex,cg_gl;
+
+ ind = state->cg_gl;
+ dd_cg_gl = dd->index_gl;
+ cgindex = dd->cgindex;
+ nat = 0;
+ cgindex[0] = nat;
+ for(i=0; i<state->ncg_gl; i++)
+ {
+ cgindex[i] = nat;
+ cg_gl = ind[i];
+ dd_cg_gl[i] = cg_gl;
+ nat += gcgs_index[cg_gl+1] - gcgs_index[cg_gl];
+ }
+ cgindex[i] = nat;
+
+ dd->ncg_home = state->ncg_gl;
+ dd->nat_home = nat;
+
+ set_zones_ncg_home(dd);
+}
+
+static int ddcginfo(const cginfo_mb_t *cginfo_mb,int cg)
+{
+ while (cg >= cginfo_mb->cg_end)
+ {
+ cginfo_mb++;
+ }
+
+ return cginfo_mb->cginfo[(cg - cginfo_mb->cg_start) % cginfo_mb->cg_mod];
+}
+
+static void dd_set_cginfo(int *index_gl,int cg0,int cg1,
+ t_forcerec *fr,char *bLocalCG)
+{
+ cginfo_mb_t *cginfo_mb;
+ int *cginfo;
+ int cg;
+
+ if (fr != NULL)
+ {
+ cginfo_mb = fr->cginfo_mb;
+ cginfo = fr->cginfo;
+
+ for(cg=cg0; cg<cg1; cg++)
+ {
+ cginfo[cg] = ddcginfo(cginfo_mb,index_gl[cg]);
+ }
+ }
+
+ if (bLocalCG != NULL)
+ {
+ for(cg=cg0; cg<cg1; cg++)
+ {
+ bLocalCG[index_gl[cg]] = TRUE;
+ }
+ }
+}
+
+static void make_dd_indices(gmx_domdec_t *dd,int *gcgs_index,int cg_start)
+{
+ int nzone,zone,zone1,cg0,cg,cg_gl,a,a_gl;
+ int *zone2cg,*zone_ncg1,*index_gl,*gatindex;
+ gmx_ga2la_t *ga2la;
+ char *bLocalCG;
+
+ bLocalCG = dd->comm->bLocalCG;
+
+ if (dd->nat_tot > dd->gatindex_nalloc)
+ {
+ dd->gatindex_nalloc = over_alloc_dd(dd->nat_tot);
+ srenew(dd->gatindex,dd->gatindex_nalloc);
+ }
+
+ nzone = dd->comm->zones.n;
+ zone2cg = dd->comm->zones.cg_range;
+ zone_ncg1 = dd->comm->zone_ncg1;
+ index_gl = dd->index_gl;
+ gatindex = dd->gatindex;
+
+ if (zone2cg[1] != dd->ncg_home)
+ {
+ gmx_incons("dd->ncg_zone is not up to date");
+ }
+
+ /* Make the local to global and global to local atom index */
+ a = dd->cgindex[cg_start];
+ for(zone=0; zone<nzone; zone++)
+ {
+ if (zone == 0)
+ {
+ cg0 = cg_start;
+ }
+ else
+ {
+ cg0 = zone2cg[zone];
+ }
+ for(cg=cg0; cg<zone2cg[zone+1]; cg++)
+ {
+ zone1 = zone;
+ if (cg - cg0 >= zone_ncg1[zone])
+ {
+ /* Signal that this cg is from more than one zone away */
+ zone1 += nzone;
+ }
+ cg_gl = index_gl[cg];
+ for(a_gl=gcgs_index[cg_gl]; a_gl<gcgs_index[cg_gl+1]; a_gl++)
+ {
+ gatindex[a] = a_gl;
+ ga2la_set(dd->ga2la,a_gl,a,zone1);
+ a++;
+ }
+ }
+ }
+}
+
+static int check_bLocalCG(gmx_domdec_t *dd,int ncg_sys,const char *bLocalCG,
+ const char *where)
+{
+ int ncg,i,ngl,nerr;
+
+ nerr = 0;
+ if (bLocalCG == NULL)
+ {
+ return nerr;
+ }
+ for(i=0; i<dd->ncg_tot; i++)
+ {
+ if (!bLocalCG[dd->index_gl[i]])
+ {
+ fprintf(stderr,
+ "DD node %d, %s: cg %d, global cg %d is not marked in bLocalCG (ncg_home %d)\n",dd->rank,where,i+1,dd->index_gl[i]+1,dd->ncg_home);
+ nerr++;
+ }
+ }
+ ngl = 0;
+ for(i=0; i<ncg_sys; i++)
+ {
+ if (bLocalCG[i])
+ {
+ ngl++;
+ }
+ }
+ if (ngl != dd->ncg_tot)
+ {
+ fprintf(stderr,"DD node %d, %s: In bLocalCG %d cgs are marked as local, whereas there are %d\n",dd->rank,where,ngl,dd->ncg_tot);
+ nerr++;
+ }
+
+ return nerr;
+}
+
+static void check_index_consistency(gmx_domdec_t *dd,
+ int natoms_sys,int ncg_sys,
+ const char *where)
+{
+ int nerr,ngl,i,a,cell;
+ int *have;
+
+ nerr = 0;
+
+ if (dd->comm->DD_debug > 1)
+ {
+ snew(have,natoms_sys);
+ for(a=0; a<dd->nat_tot; a++)
+ {
+ if (have[dd->gatindex[a]] > 0)
+ {
+ fprintf(stderr,"DD node %d: global atom %d occurs twice: index %d and %d\n",dd->rank,dd->gatindex[a]+1,have[dd->gatindex[a]],a+1);
+ }
+ else
+ {
+ have[dd->gatindex[a]] = a + 1;
+ }
+ }
+ sfree(have);
+ }
+
+ snew(have,dd->nat_tot);
+
+ ngl = 0;
+ for(i=0; i<natoms_sys; i++)
+ {
+ if (ga2la_get(dd->ga2la,i,&a,&cell))
+ {
+ if (a >= dd->nat_tot)
+ {
+ fprintf(stderr,"DD node %d: global atom %d marked as local atom %d, which is larger than nat_tot (%d)\n",dd->rank,i+1,a+1,dd->nat_tot);
+ nerr++;
+ }
+ else
+ {
+ have[a] = 1;
+ if (dd->gatindex[a] != i)
+ {
+ fprintf(stderr,"DD node %d: global atom %d marked as local atom %d, which has global atom index %d\n",dd->rank,i+1,a+1,dd->gatindex[a]+1);
+ nerr++;
+ }
+ }
+ ngl++;
+ }
+ }
+ if (ngl != dd->nat_tot)
+ {
+ fprintf(stderr,
+ "DD node %d, %s: %d global atom indices, %d local atoms\n",
+ dd->rank,where,ngl,dd->nat_tot);
+ }
+ for(a=0; a<dd->nat_tot; a++)
+ {
+ if (have[a] == 0)
+ {
+ fprintf(stderr,
+ "DD node %d, %s: local atom %d, global %d has no global index\n",
+ dd->rank,where,a+1,dd->gatindex[a]+1);
+ }
+ }
+ sfree(have);
+
+ nerr += check_bLocalCG(dd,ncg_sys,dd->comm->bLocalCG,where);
+
+ if (nerr > 0) {
+ gmx_fatal(FARGS,"DD node %d, %s: %d atom/cg index inconsistencies",
+ dd->rank,where,nerr);
+ }
+}
+
+static void clear_dd_indices(gmx_domdec_t *dd,int cg_start,int a_start)
+{
+ int i;
+ char *bLocalCG;
+
+ if (a_start == 0)
+ {
+ /* Clear the whole list without searching */
+ ga2la_clear(dd->ga2la);
+ }
+ else
+ {
+ for(i=a_start; i<dd->nat_tot; i++)
+ {
+ ga2la_del(dd->ga2la,dd->gatindex[i]);
+ }
+ }
+
+ bLocalCG = dd->comm->bLocalCG;
+ if (bLocalCG)
+ {
+ for(i=cg_start; i<dd->ncg_tot; i++)
+ {
+ bLocalCG[dd->index_gl[i]] = FALSE;
+ }
+ }
+
+ dd_clear_local_vsite_indices(dd);
+
+ if (dd->constraints)
+ {
+ dd_clear_local_constraint_indices(dd);
+ }
+}
+
+static real grid_jump_limit(gmx_domdec_comm_t *comm,int dim_ind)
+{
+ real grid_jump_limit;
+
+ /* The distance between the boundaries of cells at distance
+ * x+-1,y+-1 or y+-1,z+-1 is limited by the cut-off restrictions
+ * and by the fact that cells should not be shifted by more than
+ * half their size, such that cg's only shift by one cell
+ * at redecomposition.
+ */
+ grid_jump_limit = comm->cellsize_limit;
+ if (!comm->bVacDLBNoLimit)
+ {
+ grid_jump_limit = max(grid_jump_limit,
+ comm->cutoff/comm->cd[dim_ind].np);
+ }
+
+ return grid_jump_limit;
+}
+
+static void check_grid_jump(gmx_large_int_t step,gmx_domdec_t *dd,gmx_ddbox_t *ddbox)
+{
+ gmx_domdec_comm_t *comm;
+ int d,dim;
+ real limit,bfac;
+
+ comm = dd->comm;
+
+ for(d=1; d<dd->ndim; d++)
+ {
+ dim = dd->dim[d];
+ limit = grid_jump_limit(comm,d);
+ bfac = ddbox->box_size[dim];
+ if (ddbox->tric_dir[dim])
+ {
+ bfac *= ddbox->skew_fac[dim];
+ }
+ if ((comm->cell_f1[d] - comm->cell_f_max0[d])*bfac < limit ||
+ (comm->cell_f0[d] - comm->cell_f_min1[d])*bfac > -limit)
+ {
+ char buf[22];
+ gmx_fatal(FARGS,"Step %s: The domain decomposition grid has shifted too much in the %c-direction around cell %d %d %d\n",
+ gmx_step_str(step,buf),
+ dim2char(dim),dd->ci[XX],dd->ci[YY],dd->ci[ZZ]);
+ }
+ }
+}
+
+static int dd_load_count(gmx_domdec_comm_t *comm)
+{
+ return (comm->eFlop ? comm->flop_n : comm->cycl_n[ddCyclF]);
+}
+
+static float dd_force_load(gmx_domdec_comm_t *comm)
+{
+ float load;
+
+ if (comm->eFlop)
+ {
+ load = comm->flop;
+ if (comm->eFlop > 1)
+ {
+ load *= 1.0 + (comm->eFlop - 1)*(0.1*rand()/RAND_MAX - 0.05);
+ }
+ }
+ else
+ {
+ load = comm->cycl[ddCyclF];
+ if (comm->cycl_n[ddCyclF] > 1)
+ {
+ /* Subtract the maximum of the last n cycle counts
+ * to get rid of possible high counts due to other soures,
+ * for instance system activity, that would otherwise
+ * affect the dynamic load balancing.
+ */
+ load -= comm->cycl_max[ddCyclF];
+ }
+ }
+
+ return load;
+}
+
+static void set_slb_pme_dim_f(gmx_domdec_t *dd,int dim,real **dim_f)
+{
+ gmx_domdec_comm_t *comm;
+ int i;
+
+ comm = dd->comm;
+
+ snew(*dim_f,dd->nc[dim]+1);
+ (*dim_f)[0] = 0;
+ for(i=1; i<dd->nc[dim]; i++)
+ {
+ if (comm->slb_frac[dim])
+ {
+ (*dim_f)[i] = (*dim_f)[i-1] + comm->slb_frac[dim][i-1];
+ }
+ else
+ {
+ (*dim_f)[i] = (real)i/(real)dd->nc[dim];
+ }
+ }
+ (*dim_f)[dd->nc[dim]] = 1;
+}
+
+static void init_ddpme(gmx_domdec_t *dd,gmx_ddpme_t *ddpme,int dimind)
+{
+ int pmeindex,slab,nso,i;
+ ivec xyz;
+
+ if (dimind == 0 && dd->dim[0] == YY && dd->comm->npmenodes_x == 1)
+ {
+ ddpme->dim = YY;
+ }
+ else
+ {
+ ddpme->dim = dimind;
+ }
+ ddpme->dim_match = (ddpme->dim == dd->dim[dimind]);
+
+ ddpme->nslab = (ddpme->dim == 0 ?
+ dd->comm->npmenodes_x :
+ dd->comm->npmenodes_y);
+
+ if (ddpme->nslab <= 1)
+ {
+ return;
+ }
+
+ nso = dd->comm->npmenodes/ddpme->nslab;
+ /* Determine for each PME slab the PP location range for dimension dim */
+ snew(ddpme->pp_min,ddpme->nslab);
+ snew(ddpme->pp_max,ddpme->nslab);
+ for(slab=0; slab<ddpme->nslab; slab++) {
+ ddpme->pp_min[slab] = dd->nc[dd->dim[dimind]] - 1;
+ ddpme->pp_max[slab] = 0;
+ }
+ for(i=0; i<dd->nnodes; i++) {
+ ddindex2xyz(dd->nc,i,xyz);
+ /* For y only use our y/z slab.
+ * This assumes that the PME x grid size matches the DD grid size.
+ */
+ if (dimind == 0 || xyz[XX] == dd->ci[XX]) {
+ pmeindex = ddindex2pmeindex(dd,i);
+ if (dimind == 0) {
+ slab = pmeindex/nso;
+ } else {
+ slab = pmeindex % ddpme->nslab;
+ }
+ ddpme->pp_min[slab] = min(ddpme->pp_min[slab],xyz[dimind]);
+ ddpme->pp_max[slab] = max(ddpme->pp_max[slab],xyz[dimind]);
+ }
+ }
+
+ set_slb_pme_dim_f(dd,ddpme->dim,&ddpme->slb_dim_f);
+}
+
+int dd_pme_maxshift_x(gmx_domdec_t *dd)
+{
+ if (dd->comm->ddpme[0].dim == XX)
+ {
+ return dd->comm->ddpme[0].maxshift;
+ }
+ else
+ {
+ return 0;
+ }
+}
+
+int dd_pme_maxshift_y(gmx_domdec_t *dd)
+{
+ if (dd->comm->ddpme[0].dim == YY)
+ {
+ return dd->comm->ddpme[0].maxshift;
+ }
+ else if (dd->comm->npmedecompdim >= 2 && dd->comm->ddpme[1].dim == YY)
+ {
+ return dd->comm->ddpme[1].maxshift;
+ }
+ else
+ {
+ return 0;
+ }
+}
+
+static void set_pme_maxshift(gmx_domdec_t *dd,gmx_ddpme_t *ddpme,
+ gmx_bool bUniform,gmx_ddbox_t *ddbox,real *cell_f)
+{
+ gmx_domdec_comm_t *comm;
+ int nc,ns,s;
+ int *xmin,*xmax;
+ real range,pme_boundary;
+ int sh;
+
+ comm = dd->comm;
+ nc = dd->nc[ddpme->dim];
+ ns = ddpme->nslab;
+
+ if (!ddpme->dim_match)
+ {
+ /* PP decomposition is not along dim: the worst situation */
+ sh = ns/2;
+ }
+ else if (ns <= 3 || (bUniform && ns == nc))
+ {
+ /* The optimal situation */
+ sh = 1;
+ }
+ else
+ {
+ /* We need to check for all pme nodes which nodes they
+ * could possibly need to communicate with.
+ */
+ xmin = ddpme->pp_min;
+ xmax = ddpme->pp_max;
+ /* Allow for atoms to be maximally 2/3 times the cut-off
+ * out of their DD cell. This is a reasonable balance between
+ * between performance and support for most charge-group/cut-off
+ * combinations.
+ */
+ range = 2.0/3.0*comm->cutoff/ddbox->box_size[ddpme->dim];
+ /* Avoid extra communication when we are exactly at a boundary */
+ range *= 0.999;
+
+ sh = 1;
+ for(s=0; s<ns; s++)
+ {
+ /* PME slab s spreads atoms between box frac. s/ns and (s+1)/ns */
+ pme_boundary = (real)s/ns;
+ while (sh+1 < ns &&
+ ((s-(sh+1) >= 0 &&
+ cell_f[xmax[s-(sh+1) ]+1] + range > pme_boundary) ||
+ (s-(sh+1) < 0 &&
+ cell_f[xmax[s-(sh+1)+ns]+1] - 1 + range > pme_boundary)))
+ {
+ sh++;
+ }
+ pme_boundary = (real)(s+1)/ns;
+ while (sh+1 < ns &&
+ ((s+(sh+1) < ns &&
+ cell_f[xmin[s+(sh+1) ] ] - range < pme_boundary) ||
+ (s+(sh+1) >= ns &&
+ cell_f[xmin[s+(sh+1)-ns] ] + 1 - range < pme_boundary)))
+ {
+ sh++;
+ }
+ }
+ }
+
+ ddpme->maxshift = sh;
+
+ if (debug)
+ {
+ fprintf(debug,"PME slab communication range for dim %d is %d\n",
+ ddpme->dim,ddpme->maxshift);
+ }
+}
+
+static void check_box_size(gmx_domdec_t *dd,gmx_ddbox_t *ddbox)
+{
+ int d,dim;
+
+ for(d=0; d<dd->ndim; d++)
+ {
+ dim = dd->dim[d];
+ if (dim < ddbox->nboundeddim &&
+ ddbox->box_size[dim]*ddbox->skew_fac[dim] <
+ dd->nc[dim]*dd->comm->cellsize_limit*DD_CELL_MARGIN)
+ {
+ gmx_fatal(FARGS,"The %c-size of the box (%f) times the triclinic skew factor (%f) is smaller than the number of DD cells (%d) times the smallest allowed cell size (%f)\n",
+ dim2char(dim),ddbox->box_size[dim],ddbox->skew_fac[dim],
+ dd->nc[dim],dd->comm->cellsize_limit);
+ }
+ }
+}
+
+static void set_dd_cell_sizes_slb(gmx_domdec_t *dd,gmx_ddbox_t *ddbox,
+ gmx_bool bMaster,ivec npulse)
+{
+ gmx_domdec_comm_t *comm;
+ int d,j;
+ rvec cellsize_min;
+ real *cell_x,cell_dx,cellsize;
+
+ comm = dd->comm;
+
+ for(d=0; d<DIM; d++)
+ {
+ cellsize_min[d] = ddbox->box_size[d]*ddbox->skew_fac[d];
+ npulse[d] = 1;
+ if (dd->nc[d] == 1 || comm->slb_frac[d] == NULL)
+ {
+ /* Uniform grid */
+ cell_dx = ddbox->box_size[d]/dd->nc[d];
+ if (bMaster)
+ {
+ for(j=0; j<dd->nc[d]+1; j++)
+ {
+ dd->ma->cell_x[d][j] = ddbox->box0[d] + j*cell_dx;
+ }
+ }
+ else
+ {
+ comm->cell_x0[d] = ddbox->box0[d] + (dd->ci[d] )*cell_dx;
+ comm->cell_x1[d] = ddbox->box0[d] + (dd->ci[d]+1)*cell_dx;
+ }
+ cellsize = cell_dx*ddbox->skew_fac[d];
+ while (cellsize*npulse[d] < comm->cutoff && npulse[d] < dd->nc[d]-1)
+ {
+ npulse[d]++;
+ }
+ cellsize_min[d] = cellsize;
+ }
+ else
+ {
+ /* Statically load balanced grid */
+ /* Also when we are not doing a master distribution we determine
+ * all cell borders in a loop to obtain identical values
+ * to the master distribution case and to determine npulse.
+ */
+ if (bMaster)
+ {
+ cell_x = dd->ma->cell_x[d];
+ }
+ else
+ {
+ snew(cell_x,dd->nc[d]+1);
+ }
+ cell_x[0] = ddbox->box0[d];
+ for(j=0; j<dd->nc[d]; j++)
+ {
+ cell_dx = ddbox->box_size[d]*comm->slb_frac[d][j];
+ cell_x[j+1] = cell_x[j] + cell_dx;
+ cellsize = cell_dx*ddbox->skew_fac[d];
+ while (cellsize*npulse[d] < comm->cutoff &&
+ npulse[d] < dd->nc[d]-1)
+ {
+ npulse[d]++;
+ }
+ cellsize_min[d] = min(cellsize_min[d],cellsize);
+ }
+ if (!bMaster)
+ {
+ comm->cell_x0[d] = cell_x[dd->ci[d]];
+ comm->cell_x1[d] = cell_x[dd->ci[d]+1];
+ sfree(cell_x);
+ }
+ }
+ /* The following limitation is to avoid that a cell would receive
+ * some of its own home charge groups back over the periodic boundary.
+ * Double charge groups cause trouble with the global indices.
+ */
+ if (d < ddbox->npbcdim &&
+ dd->nc[d] > 1 && npulse[d] >= dd->nc[d])
+ {
+ gmx_fatal_collective(FARGS,NULL,dd,
+ "The box size in direction %c (%f) times the triclinic skew factor (%f) is too small for a cut-off of %f with %d domain decomposition cells, use 1 or more than %d %s or increase the box size in this direction",
+ dim2char(d),ddbox->box_size[d],ddbox->skew_fac[d],
+ comm->cutoff,
+ dd->nc[d],dd->nc[d],
+ dd->nnodes > dd->nc[d] ? "cells" : "processors");
+ }
+ }
+
+ if (!comm->bDynLoadBal)
+ {
+ copy_rvec(cellsize_min,comm->cellsize_min);
+ }
+
+ for(d=0; d<comm->npmedecompdim; d++)
+ {
+ set_pme_maxshift(dd,&comm->ddpme[d],
+ comm->slb_frac[dd->dim[d]]==NULL,ddbox,
+ comm->ddpme[d].slb_dim_f);
+ }
+}
+
+
+static void dd_cell_sizes_dlb_root_enforce_limits(gmx_domdec_t *dd,
+ int d,int dim,gmx_domdec_root_t *root,
+ gmx_ddbox_t *ddbox,
+ gmx_bool bUniform,gmx_large_int_t step, real cellsize_limit_f, int range[])
+{
+ gmx_domdec_comm_t *comm;
+ int ncd,i,j,nmin,nmin_old;
+ gmx_bool bLimLo,bLimHi;
+ real *cell_size;
+ real fac,halfway,cellsize_limit_f_i,region_size;
+ gmx_bool bPBC,bLastHi=FALSE;
+ int nrange[]={range[0],range[1]};
+
+ region_size= root->cell_f[range[1]]-root->cell_f[range[0]];
+
+ comm = dd->comm;
+
+ ncd = dd->nc[dim];
+
+ bPBC = (dim < ddbox->npbcdim);
+
+ cell_size = root->buf_ncd;
+
+ if (debug)
+ {
+ fprintf(debug,"enforce_limits: %d %d\n",range[0],range[1]);
+ }
+
+ /* First we need to check if the scaling does not make cells
+ * smaller than the smallest allowed size.
+ * We need to do this iteratively, since if a cell is too small,
+ * it needs to be enlarged, which makes all the other cells smaller,
+ * which could in turn make another cell smaller than allowed.
+ */
+ for(i=range[0]; i<range[1]; i++)
+ {
+ root->bCellMin[i] = FALSE;
+ }
+ nmin = 0;
+ do
+ {
+ nmin_old = nmin;
+ /* We need the total for normalization */
+ fac = 0;
+ for(i=range[0]; i<range[1]; i++)
+ {
+ if (root->bCellMin[i] == FALSE)
+ {
+ fac += cell_size[i];
+ }
+ }
+ fac = ( region_size - nmin*cellsize_limit_f)/fac; /* substracting cells already set to cellsize_limit_f */
+ /* Determine the cell boundaries */
+ for(i=range[0]; i<range[1]; i++)
+ {
+ if (root->bCellMin[i] == FALSE)
+ {
+ cell_size[i] *= fac;
+ if (!bPBC && (i == 0 || i == dd->nc[dim] -1))
+ {
+ cellsize_limit_f_i = 0;
+ }
+ else
+ {
+ cellsize_limit_f_i = cellsize_limit_f;
+ }
+ if (cell_size[i] < cellsize_limit_f_i)
+ {
+ root->bCellMin[i] = TRUE;
+ cell_size[i] = cellsize_limit_f_i;
+ nmin++;
+ }
+ }
+ root->cell_f[i+1] = root->cell_f[i] + cell_size[i];
+ }
+ }
+ while (nmin > nmin_old);
+
+ i=range[1]-1;
+ cell_size[i] = root->cell_f[i+1] - root->cell_f[i];
+ /* For this check we should not use DD_CELL_MARGIN,
+ * but a slightly smaller factor,
+ * since rounding could get use below the limit.
+ */
+ if (bPBC && cell_size[i] < cellsize_limit_f*DD_CELL_MARGIN2/DD_CELL_MARGIN)
+ {
+ char buf[22];
+ gmx_fatal(FARGS,"Step %s: the dynamic load balancing could not balance dimension %c: box size %f, triclinic skew factor %f, #cells %d, minimum cell size %f\n",
+ gmx_step_str(step,buf),
+ dim2char(dim),ddbox->box_size[dim],ddbox->skew_fac[dim],
+ ncd,comm->cellsize_min[dim]);
+ }
+
+ root->bLimited = (nmin > 0) || (range[0]>0) || (range[1]<ncd);
+
+ if (!bUniform)
+ {
+ /* Check if the boundary did not displace more than halfway
+ * each of the cells it bounds, as this could cause problems,
+ * especially when the differences between cell sizes are large.
+ * If changes are applied, they will not make cells smaller
+ * than the cut-off, as we check all the boundaries which
+ * might be affected by a change and if the old state was ok,
+ * the cells will at most be shrunk back to their old size.
+ */
+ for(i=range[0]+1; i<range[1]; i++)
+ {
+ halfway = 0.5*(root->old_cell_f[i] + root->old_cell_f[i-1]);
+ if (root->cell_f[i] < halfway)
+ {
+ root->cell_f[i] = halfway;
+ /* Check if the change also causes shifts of the next boundaries */
+ for(j=i+1; j<range[1]; j++)
+ {
+ if (root->cell_f[j] < root->cell_f[j-1] + cellsize_limit_f)
+ root->cell_f[j] = root->cell_f[j-1] + cellsize_limit_f;
+ }
+ }
+ halfway = 0.5*(root->old_cell_f[i] + root->old_cell_f[i+1]);
+ if (root->cell_f[i] > halfway)
+ {
+ root->cell_f[i] = halfway;
+ /* Check if the change also causes shifts of the next boundaries */
+ for(j=i-1; j>=range[0]+1; j--)
+ {
+ if (root->cell_f[j] > root->cell_f[j+1] - cellsize_limit_f)
+ root->cell_f[j] = root->cell_f[j+1] - cellsize_limit_f;
+ }
+ }
+ }
+ }
+
+ /* nrange is defined as [lower, upper) range for new call to enforce_limits */
+ /* find highest violation of LimLo (a) and the following violation of LimHi (thus the lowest following) (b)
+ * then call enforce_limits for (oldb,a), (a,b). In the next step: (b,nexta). oldb and nexta can be the boundaries.
+ * for a and b nrange is used */
+ if (d > 0)
+ {
+ /* Take care of the staggering of the cell boundaries */
+ if (bUniform)
+ {
+ for(i=range[0]; i<range[1]; i++)
+ {
+ root->cell_f_max0[i] = root->cell_f[i];
+ root->cell_f_min1[i] = root->cell_f[i+1];
+ }
+ }
+ else
+ {
+ for(i=range[0]+1; i<range[1]; i++)
+ {
+ bLimLo = (root->cell_f[i] < root->bound_min[i]);
+ bLimHi = (root->cell_f[i] > root->bound_max[i]);
+ if (bLimLo && bLimHi)
+ {
+ /* Both limits violated, try the best we can */
+ /* For this case we split the original range (range) in two parts and care about the other limitiations in the next iteration. */
+ root->cell_f[i] = 0.5*(root->bound_min[i] + root->bound_max[i]);
+ nrange[0]=range[0];
+ nrange[1]=i;
+ dd_cell_sizes_dlb_root_enforce_limits(dd, d, dim, root, ddbox, bUniform, step, cellsize_limit_f, nrange);
+
+ nrange[0]=i;
+ nrange[1]=range[1];
+ dd_cell_sizes_dlb_root_enforce_limits(dd, d, dim, root, ddbox, bUniform, step, cellsize_limit_f, nrange);
+
+ return;
+ }
+ else if (bLimLo)
+ {
+ /* root->cell_f[i] = root->bound_min[i]; */
+ nrange[1]=i; /* only store violation location. There could be a LimLo violation following with an higher index */
+ bLastHi=FALSE;
+ }
+ else if (bLimHi && !bLastHi)
+ {
+ bLastHi=TRUE;
+ if (nrange[1] < range[1]) /* found a LimLo before */
+ {
+ root->cell_f[nrange[1]] = root->bound_min[nrange[1]];
+ dd_cell_sizes_dlb_root_enforce_limits(dd, d, dim, root, ddbox, bUniform, step, cellsize_limit_f, nrange);
+ nrange[0]=nrange[1];
+ }
+ root->cell_f[i] = root->bound_max[i];
+ nrange[1]=i;
+ dd_cell_sizes_dlb_root_enforce_limits(dd, d, dim, root, ddbox, bUniform, step, cellsize_limit_f, nrange);
+ nrange[0]=i;
+ nrange[1]=range[1];
+ }
+ }
+ if (nrange[1] < range[1]) /* found last a LimLo */
+ {
+ root->cell_f[nrange[1]] = root->bound_min[nrange[1]];
+ dd_cell_sizes_dlb_root_enforce_limits(dd, d, dim, root, ddbox, bUniform, step, cellsize_limit_f, nrange);
+ nrange[0]=nrange[1];
+ nrange[1]=range[1];
+ dd_cell_sizes_dlb_root_enforce_limits(dd, d, dim, root, ddbox, bUniform, step, cellsize_limit_f, nrange);
+ }
+ else if (nrange[0] > range[0]) /* found at least one LimHi */
+ {
+ dd_cell_sizes_dlb_root_enforce_limits(dd, d, dim, root, ddbox, bUniform, step, cellsize_limit_f, nrange);
+ }
+ }
+ }
+}
+
+
+static void set_dd_cell_sizes_dlb_root(gmx_domdec_t *dd,
+ int d,int dim,gmx_domdec_root_t *root,
+ gmx_ddbox_t *ddbox,gmx_bool bDynamicBox,
+ gmx_bool bUniform,gmx_large_int_t step)
+{
+ gmx_domdec_comm_t *comm;
+ int ncd,d1,i,j,pos;
+ real *cell_size;
+ real load_aver,load_i,imbalance,change,change_max,sc;
+ real cellsize_limit_f,dist_min_f,dist_min_f_hard,space;
+ real change_limit;
+ real relax = 0.5;
+ gmx_bool bPBC;
+ int range[] = { 0, 0 };
+
+ comm = dd->comm;
+
+ /* Convert the maximum change from the input percentage to a fraction */
+ change_limit = comm->dlb_scale_lim*0.01;
+
+ ncd = dd->nc[dim];
+
+ bPBC = (dim < ddbox->npbcdim);
+
+ cell_size = root->buf_ncd;
+
+ /* Store the original boundaries */
+ for(i=0; i<ncd+1; i++)
+ {
+ root->old_cell_f[i] = root->cell_f[i];
+ }
+ if (bUniform) {
+ for(i=0; i<ncd; i++)
+ {
+ cell_size[i] = 1.0/ncd;
+ }
+ }
+ else if (dd_load_count(comm))
+ {
+ load_aver = comm->load[d].sum_m/ncd;
+ change_max = 0;
+ for(i=0; i<ncd; i++)
+ {
+ /* Determine the relative imbalance of cell i */
+ load_i = comm->load[d].load[i*comm->load[d].nload+2];
+ imbalance = (load_i - load_aver)/(load_aver>0 ? load_aver : 1);
+ /* Determine the change of the cell size using underrelaxation */
+ change = -relax*imbalance;
+ change_max = max(change_max,max(change,-change));
+ }
+ /* Limit the amount of scaling.
+ * We need to use the same rescaling for all cells in one row,
+ * otherwise the load balancing might not converge.
+ */
+ sc = relax;
+ if (change_max > change_limit)
+ {
+ sc *= change_limit/change_max;
+ }
+ for(i=0; i<ncd; i++)
+ {
+ /* Determine the relative imbalance of cell i */
+ load_i = comm->load[d].load[i*comm->load[d].nload+2];
+ imbalance = (load_i - load_aver)/(load_aver>0 ? load_aver : 1);
+ /* Determine the change of the cell size using underrelaxation */
+ change = -sc*imbalance;
+ cell_size[i] = (root->cell_f[i+1]-root->cell_f[i])*(1 + change);
+ }
+ }
+
+ cellsize_limit_f = comm->cellsize_min[dim]/ddbox->box_size[dim];
+ cellsize_limit_f *= DD_CELL_MARGIN;
+ dist_min_f_hard = grid_jump_limit(comm,d)/ddbox->box_size[dim];
+ dist_min_f = dist_min_f_hard * DD_CELL_MARGIN;
+ if (ddbox->tric_dir[dim])
+ {
+ cellsize_limit_f /= ddbox->skew_fac[dim];
+ dist_min_f /= ddbox->skew_fac[dim];
+ }
+ if (bDynamicBox && d > 0)
+ {
+ dist_min_f *= DD_PRES_SCALE_MARGIN;
+ }
+ if (d > 0 && !bUniform)
+ {
+ /* Make sure that the grid is not shifted too much */
+ for(i=1; i<ncd; i++) {
+ if (root->cell_f_min1[i] - root->cell_f_max0[i-1] < 2 * dist_min_f_hard)
+ {
+ gmx_incons("Inconsistent DD boundary staggering limits!");
+ }
+ root->bound_min[i] = root->cell_f_max0[i-1] + dist_min_f;
+ space = root->cell_f[i] - (root->cell_f_max0[i-1] + dist_min_f);
+ if (space > 0) {
+ root->bound_min[i] += 0.5*space;
+ }
+ root->bound_max[i] = root->cell_f_min1[i] - dist_min_f;
+ space = root->cell_f[i] - (root->cell_f_min1[i] - dist_min_f);
+ if (space < 0) {
+ root->bound_max[i] += 0.5*space;
+ }
+ if (debug)
+ {
+ fprintf(debug,
+ "dim %d boundary %d %.3f < %.3f < %.3f < %.3f < %.3f\n",
+ d,i,
+ root->cell_f_max0[i-1] + dist_min_f,
+ root->bound_min[i],root->cell_f[i],root->bound_max[i],
+ root->cell_f_min1[i] - dist_min_f);
+ }
+ }
+ }
+ range[1]=ncd;
+ root->cell_f[0] = 0;
+ root->cell_f[ncd] = 1;
+ dd_cell_sizes_dlb_root_enforce_limits(dd, d, dim, root, ddbox, bUniform, step, cellsize_limit_f, range);
+
+
+ /* After the checks above, the cells should obey the cut-off
+ * restrictions, but it does not hurt to check.
+ */
+ for(i=0; i<ncd; i++)
+ {
+ if (debug)
+ {
+ fprintf(debug,"Relative bounds dim %d cell %d: %f %f\n",
+ dim,i,root->cell_f[i],root->cell_f[i+1]);
+ }
+
+ if ((bPBC || (i != 0 && i != dd->nc[dim]-1)) &&
+ root->cell_f[i+1] - root->cell_f[i] <
+ cellsize_limit_f/DD_CELL_MARGIN)
+ {
+ char buf[22];
+ fprintf(stderr,
+ "\nWARNING step %s: direction %c, cell %d too small: %f\n",
+ gmx_step_str(step,buf),dim2char(dim),i,
+ (root->cell_f[i+1] - root->cell_f[i])
+ *ddbox->box_size[dim]*ddbox->skew_fac[dim]);
+ }
+ }
+
+ pos = ncd + 1;
+ /* Store the cell boundaries of the lower dimensions at the end */
+ for(d1=0; d1<d; d1++)
+ {
+ root->cell_f[pos++] = comm->cell_f0[d1];
+ root->cell_f[pos++] = comm->cell_f1[d1];
+ }
+
+ if (d < comm->npmedecompdim)
+ {
+ /* The master determines the maximum shift for
+ * the coordinate communication between separate PME nodes.
+ */
+ set_pme_maxshift(dd,&comm->ddpme[d],bUniform,ddbox,root->cell_f);
+ }
+ root->cell_f[pos++] = comm->ddpme[0].maxshift;
+ if (d >= 1)
+ {
+ root->cell_f[pos++] = comm->ddpme[1].maxshift;
+ }
+}
+
+static void relative_to_absolute_cell_bounds(gmx_domdec_t *dd,
+ gmx_ddbox_t *ddbox,int dimind)
+{
+ gmx_domdec_comm_t *comm;
+ int dim;
+
+ comm = dd->comm;
+
+ /* Set the cell dimensions */
+ dim = dd->dim[dimind];
+ comm->cell_x0[dim] = comm->cell_f0[dimind]*ddbox->box_size[dim];
+ comm->cell_x1[dim] = comm->cell_f1[dimind]*ddbox->box_size[dim];
+ if (dim >= ddbox->nboundeddim)
+ {
+ comm->cell_x0[dim] += ddbox->box0[dim];
+ comm->cell_x1[dim] += ddbox->box0[dim];
+ }
+}
+
+static void distribute_dd_cell_sizes_dlb(gmx_domdec_t *dd,
+ int d,int dim,real *cell_f_row,
+ gmx_ddbox_t *ddbox)
+{
+ gmx_domdec_comm_t *comm;
+ int d1,dim1,pos;
+
+ comm = dd->comm;
+
+#ifdef GMX_MPI
+ /* Each node would only need to know two fractions,
+ * but it is probably cheaper to broadcast the whole array.
+ */
+ MPI_Bcast(cell_f_row,DD_CELL_F_SIZE(dd,d)*sizeof(real),MPI_BYTE,
+ 0,comm->mpi_comm_load[d]);
+#endif
+ /* Copy the fractions for this dimension from the buffer */
+ comm->cell_f0[d] = cell_f_row[dd->ci[dim] ];
+ comm->cell_f1[d] = cell_f_row[dd->ci[dim]+1];
+ /* The whole array was communicated, so set the buffer position */
+ pos = dd->nc[dim] + 1;
+ for(d1=0; d1<=d; d1++)
+ {
+ if (d1 < d)
+ {
+ /* Copy the cell fractions of the lower dimensions */
+ comm->cell_f0[d1] = cell_f_row[pos++];
+ comm->cell_f1[d1] = cell_f_row[pos++];
+ }
+ relative_to_absolute_cell_bounds(dd,ddbox,d1);
+ }
+ /* Convert the communicated shift from float to int */
+ comm->ddpme[0].maxshift = (int)(cell_f_row[pos++] + 0.5);
+ if (d >= 1)
+ {
+ comm->ddpme[1].maxshift = (int)(cell_f_row[pos++] + 0.5);
+ }
+}
+
+static void set_dd_cell_sizes_dlb_change(gmx_domdec_t *dd,
+ gmx_ddbox_t *ddbox,gmx_bool bDynamicBox,
+ gmx_bool bUniform,gmx_large_int_t step)
+{
+ gmx_domdec_comm_t *comm;
+ int d,dim,d1;
+ gmx_bool bRowMember,bRowRoot;
+ real *cell_f_row;
+
+ comm = dd->comm;
+
+ for(d=0; d<dd->ndim; d++)
+ {
+ dim = dd->dim[d];
+ bRowMember = TRUE;
+ bRowRoot = TRUE;
+ for(d1=d; d1<dd->ndim; d1++)
+ {
+ if (dd->ci[dd->dim[d1]] > 0)
+ {
+ if (d1 > d)
+ {
+ bRowMember = FALSE;
+ }
+ bRowRoot = FALSE;
+ }
+ }
+ if (bRowMember)
+ {
+ if (bRowRoot)
+ {
+ set_dd_cell_sizes_dlb_root(dd,d,dim,comm->root[d],
+ ddbox,bDynamicBox,bUniform,step);
+ cell_f_row = comm->root[d]->cell_f;
+ }
+ else
+ {
+ cell_f_row = comm->cell_f_row;
+ }
+ distribute_dd_cell_sizes_dlb(dd,d,dim,cell_f_row,ddbox);
+ }
+ }
+}
+
+static void set_dd_cell_sizes_dlb_nochange(gmx_domdec_t *dd,gmx_ddbox_t *ddbox)
+{
+ int d;
+
+ /* This function assumes the box is static and should therefore
+ * not be called when the box has changed since the last
+ * call to dd_partition_system.
+ */
+ for(d=0; d<dd->ndim; d++)
+ {
+ relative_to_absolute_cell_bounds(dd,ddbox,d);
+ }
+}
+
+
+
+static void set_dd_cell_sizes_dlb(gmx_domdec_t *dd,
+ gmx_ddbox_t *ddbox,gmx_bool bDynamicBox,
+ gmx_bool bUniform,gmx_bool bDoDLB,gmx_large_int_t step,
+ gmx_wallcycle_t wcycle)
+{
+ gmx_domdec_comm_t *comm;
+ int dim;
+
+ comm = dd->comm;
+
+ if (bDoDLB)
+ {
+ wallcycle_start(wcycle,ewcDDCOMMBOUND);
+ set_dd_cell_sizes_dlb_change(dd,ddbox,bDynamicBox,bUniform,step);
+ wallcycle_stop(wcycle,ewcDDCOMMBOUND);
+ }
+ else if (bDynamicBox)
+ {
+ set_dd_cell_sizes_dlb_nochange(dd,ddbox);
+ }
+
+ /* Set the dimensions for which no DD is used */
+ for(dim=0; dim<DIM; dim++) {
+ if (dd->nc[dim] == 1) {
+ comm->cell_x0[dim] = 0;
+ comm->cell_x1[dim] = ddbox->box_size[dim];
+ if (dim >= ddbox->nboundeddim)
+ {
+ comm->cell_x0[dim] += ddbox->box0[dim];
+ comm->cell_x1[dim] += ddbox->box0[dim];
+ }
+ }
+ }
+}
+
+static void realloc_comm_ind(gmx_domdec_t *dd,ivec npulse)
+{
+ int d,np,i;
+ gmx_domdec_comm_dim_t *cd;
+
+ for(d=0; d<dd->ndim; d++)
+ {
+ cd = &dd->comm->cd[d];
+ np = npulse[dd->dim[d]];
+ if (np > cd->np_nalloc)
+ {
+ if (debug)
+ {
+ fprintf(debug,"(Re)allocing cd for %c to %d pulses\n",
+ dim2char(dd->dim[d]),np);
+ }
+ if (DDMASTER(dd) && cd->np_nalloc > 0)
+ {
+ fprintf(stderr,"\nIncreasing the number of cell to communicate in dimension %c to %d for the first time\n",dim2char(dd->dim[d]),np);
+ }
+ srenew(cd->ind,np);
+ for(i=cd->np_nalloc; i<np; i++)
+ {
+ cd->ind[i].index = NULL;
+ cd->ind[i].nalloc = 0;
+ }
+ cd->np_nalloc = np;
+ }
+ cd->np = np;
+ }
+}
+
+
+static void set_dd_cell_sizes(gmx_domdec_t *dd,
+ gmx_ddbox_t *ddbox,gmx_bool bDynamicBox,
+ gmx_bool bUniform,gmx_bool bDoDLB,gmx_large_int_t step,
+ gmx_wallcycle_t wcycle)
+{
+ gmx_domdec_comm_t *comm;
+ int d;
+ ivec npulse;
+
+ comm = dd->comm;
+
+ /* Copy the old cell boundaries for the cg displacement check */
+ copy_rvec(comm->cell_x0,comm->old_cell_x0);
+ copy_rvec(comm->cell_x1,comm->old_cell_x1);
+
+ if (comm->bDynLoadBal)
+ {
+ if (DDMASTER(dd))
+ {
+ check_box_size(dd,ddbox);
+ }
+ set_dd_cell_sizes_dlb(dd,ddbox,bDynamicBox,bUniform,bDoDLB,step,wcycle);
+ }
+ else
+ {
+ set_dd_cell_sizes_slb(dd,ddbox,FALSE,npulse);
+ realloc_comm_ind(dd,npulse);
+ }
+
+ if (debug)
+ {
+ for(d=0; d<DIM; d++)
+ {
+ fprintf(debug,"cell_x[%d] %f - %f skew_fac %f\n",
+ d,comm->cell_x0[d],comm->cell_x1[d],ddbox->skew_fac[d]);
+ }
+ }
+}
+
+static void comm_dd_ns_cell_sizes(gmx_domdec_t *dd,
+ gmx_ddbox_t *ddbox,
+ rvec cell_ns_x0,rvec cell_ns_x1,
+ gmx_large_int_t step)
+{
+ gmx_domdec_comm_t *comm;
+ int dim_ind,dim;
+
+ comm = dd->comm;
+
+ for(dim_ind=0; dim_ind<dd->ndim; dim_ind++)
+ {
+ dim = dd->dim[dim_ind];
+
+ /* Without PBC we don't have restrictions on the outer cells */
+ if (!(dim >= ddbox->npbcdim &&
+ (dd->ci[dim] == 0 || dd->ci[dim] == dd->nc[dim] - 1)) &&
+ comm->bDynLoadBal &&
+ (comm->cell_x1[dim] - comm->cell_x0[dim])*ddbox->skew_fac[dim] <
+ comm->cellsize_min[dim])
+ {
+ char buf[22];
+ gmx_fatal(FARGS,"Step %s: The %c-size (%f) times the triclinic skew factor (%f) is smaller than the smallest allowed cell size (%f) for domain decomposition grid cell %d %d %d",
+ gmx_step_str(step,buf),dim2char(dim),
+ comm->cell_x1[dim] - comm->cell_x0[dim],
+ ddbox->skew_fac[dim],
+ dd->comm->cellsize_min[dim],
+ dd->ci[XX],dd->ci[YY],dd->ci[ZZ]);
+ }
+ }
+
+ if ((dd->bGridJump && dd->ndim > 1) || ddbox->nboundeddim < DIM)
+ {
+ /* Communicate the boundaries and update cell_ns_x0/1 */
+ dd_move_cellx(dd,ddbox,cell_ns_x0,cell_ns_x1);
+ if (dd->bGridJump && dd->ndim > 1)
+ {
+ check_grid_jump(step,dd,ddbox);
+ }
+ }
+}
+
+static void make_tric_corr_matrix(int npbcdim,matrix box,matrix tcm)
+{
+ if (YY < npbcdim)
+ {
+ tcm[YY][XX] = -box[YY][XX]/box[YY][YY];
+ }
+ else
+ {
+ tcm[YY][XX] = 0;
+ }
+ if (ZZ < npbcdim)
+ {
+ tcm[ZZ][XX] = -(box[ZZ][YY]*tcm[YY][XX] + box[ZZ][XX])/box[ZZ][ZZ];
+ tcm[ZZ][YY] = -box[ZZ][YY]/box[ZZ][ZZ];
+ }
+ else
+ {
+ tcm[ZZ][XX] = 0;
+ tcm[ZZ][YY] = 0;
+ }
+}
+
+static void check_screw_box(matrix box)
+{
+ /* Mathematical limitation */
+ if (box[YY][XX] != 0 || box[ZZ][XX] != 0)
+ {
+ gmx_fatal(FARGS,"With screw pbc the unit cell can not have non-zero off-diagonal x-components");
+ }
+
+ /* Limitation due to the asymmetry of the eighth shell method */
+ if (box[ZZ][YY] != 0)
+ {
+ gmx_fatal(FARGS,"pbc=screw with non-zero box_zy is not supported");
+ }
+}
+
+static void distribute_cg(FILE *fplog,gmx_large_int_t step,
+ matrix box,ivec tric_dir,t_block *cgs,rvec pos[],
+ gmx_domdec_t *dd)
+{
+ gmx_domdec_master_t *ma;
+ int **tmp_ind=NULL,*tmp_nalloc=NULL;
+ int i,icg,j,k,k0,k1,d,npbcdim;
+ matrix tcm;
+ rvec box_size,cg_cm;
+ ivec ind;
+ real nrcg,inv_ncg,pos_d;
+ atom_id *cgindex;
+ gmx_bool bUnbounded,bScrew;
+
+ ma = dd->ma;
+
+ if (tmp_ind == NULL)
+ {
+ snew(tmp_nalloc,dd->nnodes);
+ snew(tmp_ind,dd->nnodes);
+ for(i=0; i<dd->nnodes; i++)
+ {
+ tmp_nalloc[i] = over_alloc_large(cgs->nr/dd->nnodes+1);
+ snew(tmp_ind[i],tmp_nalloc[i]);
+ }
+ }
+
+ /* Clear the count */
+ for(i=0; i<dd->nnodes; i++)
+ {
+ ma->ncg[i] = 0;
+ ma->nat[i] = 0;
+ }
+
+ make_tric_corr_matrix(dd->npbcdim,box,tcm);
+
+ cgindex = cgs->index;
+
+ /* Compute the center of geometry for all charge groups */
+ for(icg=0; icg<cgs->nr; icg++)
+ {
+ k0 = cgindex[icg];
+ k1 = cgindex[icg+1];
+ nrcg = k1 - k0;
+ if (nrcg == 1)
+ {
+ copy_rvec(pos[k0],cg_cm);
+ }
+ else
+ {
+ inv_ncg = 1.0/nrcg;
+
+ clear_rvec(cg_cm);
+ for(k=k0; (k<k1); k++)
+ {
+ rvec_inc(cg_cm,pos[k]);
+ }
+ for(d=0; (d<DIM); d++)
+ {
+ cg_cm[d] *= inv_ncg;
+ }
+ }
+ /* Put the charge group in the box and determine the cell index */
+ for(d=DIM-1; d>=0; d--) {
+ pos_d = cg_cm[d];
+ if (d < dd->npbcdim)
+ {
+ bScrew = (dd->bScrewPBC && d == XX);
+ if (tric_dir[d] && dd->nc[d] > 1)
+ {
+ /* Use triclinic coordintates for this dimension */
+ for(j=d+1; j<DIM; j++)
+ {
+ pos_d += cg_cm[j]*tcm[j][d];
+ }
+ }
+ while(pos_d >= box[d][d])
+ {
+ pos_d -= box[d][d];
+ rvec_dec(cg_cm,box[d]);
+ if (bScrew)
+ {
+ cg_cm[YY] = box[YY][YY] - cg_cm[YY];
+ cg_cm[ZZ] = box[ZZ][ZZ] - cg_cm[ZZ];
+ }
+ for(k=k0; (k<k1); k++)
+ {
+ rvec_dec(pos[k],box[d]);
+ if (bScrew)
+ {
+ pos[k][YY] = box[YY][YY] - pos[k][YY];
+ pos[k][ZZ] = box[ZZ][ZZ] - pos[k][ZZ];
+ }
+ }
+ }
+ while(pos_d < 0)
+ {
+ pos_d += box[d][d];
+ rvec_inc(cg_cm,box[d]);
+ if (bScrew)
+ {
+ cg_cm[YY] = box[YY][YY] - cg_cm[YY];
+ cg_cm[ZZ] = box[ZZ][ZZ] - cg_cm[ZZ];
+ }
+ for(k=k0; (k<k1); k++)
+ {
+ rvec_inc(pos[k],box[d]);
+ if (bScrew) {
+ pos[k][YY] = box[YY][YY] - pos[k][YY];
+ pos[k][ZZ] = box[ZZ][ZZ] - pos[k][ZZ];
+ }
+ }
+ }
+ }
+ /* This could be done more efficiently */
+ ind[d] = 0;
+ while(ind[d]+1 < dd->nc[d] && pos_d >= ma->cell_x[d][ind[d]+1])
+ {
+ ind[d]++;
+ }
+ }
+ i = dd_index(dd->nc,ind);
+ if (ma->ncg[i] == tmp_nalloc[i])
+ {
+ tmp_nalloc[i] = over_alloc_large(ma->ncg[i]+1);
+ srenew(tmp_ind[i],tmp_nalloc[i]);
+ }
+ tmp_ind[i][ma->ncg[i]] = icg;
+ ma->ncg[i]++;
+ ma->nat[i] += cgindex[icg+1] - cgindex[icg];
+ }
+
+ k1 = 0;
+ for(i=0; i<dd->nnodes; i++)
+ {
+ ma->index[i] = k1;
+ for(k=0; k<ma->ncg[i]; k++)
+ {
+ ma->cg[k1++] = tmp_ind[i][k];
+ }
+ }
+ ma->index[dd->nnodes] = k1;
+
+ for(i=0; i<dd->nnodes; i++)
+ {
+ sfree(tmp_ind[i]);
+ }
+ sfree(tmp_ind);
+ sfree(tmp_nalloc);
+
+ if (fplog)
+ {
+ char buf[22];
+ fprintf(fplog,"Charge group distribution at step %s:",
+ gmx_step_str(step,buf));
+ for(i=0; i<dd->nnodes; i++)
+ {
+ fprintf(fplog," %d",ma->ncg[i]);
+ }
+ fprintf(fplog,"\n");
+ }
+}
+
+static void get_cg_distribution(FILE *fplog,gmx_large_int_t step,gmx_domdec_t *dd,
+ t_block *cgs,matrix box,gmx_ddbox_t *ddbox,
+ rvec pos[])
+{
+ gmx_domdec_master_t *ma=NULL;
+ ivec npulse;
+ int i,cg_gl;
+ int *ibuf,buf2[2] = { 0, 0 };
+
+ if (DDMASTER(dd))
+ {
+ ma = dd->ma;
+
+ if (dd->bScrewPBC)
+ {
+ check_screw_box(box);
+ }
+
+ set_dd_cell_sizes_slb(dd,ddbox,TRUE,npulse);
+
+ distribute_cg(fplog,step,box,ddbox->tric_dir,cgs,pos,dd);
+ for(i=0; i<dd->nnodes; i++)
+ {
+ ma->ibuf[2*i] = ma->ncg[i];
+ ma->ibuf[2*i+1] = ma->nat[i];
+ }
+ ibuf = ma->ibuf;
+ }
+ else
+ {
+ ibuf = NULL;
+ }
+ dd_scatter(dd,2*sizeof(int),ibuf,buf2);
+
+ dd->ncg_home = buf2[0];
+ dd->nat_home = buf2[1];
+ dd->ncg_tot = dd->ncg_home;
+ dd->nat_tot = dd->nat_home;
+ if (dd->ncg_home > dd->cg_nalloc || dd->cg_nalloc == 0)
+ {
+ dd->cg_nalloc = over_alloc_dd(dd->ncg_home);
+ srenew(dd->index_gl,dd->cg_nalloc);
+ srenew(dd->cgindex,dd->cg_nalloc+1);
+ }
+ if (DDMASTER(dd))
+ {
+ for(i=0; i<dd->nnodes; i++)
+ {
+ ma->ibuf[i] = ma->ncg[i]*sizeof(int);
+ ma->ibuf[dd->nnodes+i] = ma->index[i]*sizeof(int);
+ }
+ }
+
+ dd_scatterv(dd,
+ DDMASTER(dd) ? ma->ibuf : NULL,
+ DDMASTER(dd) ? ma->ibuf+dd->nnodes : NULL,
+ DDMASTER(dd) ? ma->cg : NULL,
+ dd->ncg_home*sizeof(int),dd->index_gl);
+
+ /* Determine the home charge group sizes */
+ dd->cgindex[0] = 0;
+ for(i=0; i<dd->ncg_home; i++)
+ {
+ cg_gl = dd->index_gl[i];
+ dd->cgindex[i+1] =
+ dd->cgindex[i] + cgs->index[cg_gl+1] - cgs->index[cg_gl];
+ }
+
+ if (debug)
+ {
+ fprintf(debug,"Home charge groups:\n");
+ for(i=0; i<dd->ncg_home; i++)
+ {
+ fprintf(debug," %d",dd->index_gl[i]);
+ if (i % 10 == 9)
+ fprintf(debug,"\n");
+ }
+ fprintf(debug,"\n");
+ }
+}
+
+static int compact_and_copy_vec_at(int ncg,int *move,
+ int *cgindex,
+ int nvec,int vec,
+ rvec *src,gmx_domdec_comm_t *comm,
+ gmx_bool bCompact)
+{
+ int m,icg,i,i0,i1,nrcg;
+ int home_pos;
+ int pos_vec[DIM*2];
+
+ home_pos = 0;
+
+ for(m=0; m<DIM*2; m++)
+ {
+ pos_vec[m] = 0;
+ }
+
+ i0 = 0;
+ for(icg=0; icg<ncg; icg++)
+ {
+ i1 = cgindex[icg+1];
+ m = move[icg];
+ if (m == -1)
+ {
+ if (bCompact)
+ {
+ /* Compact the home array in place */
+ for(i=i0; i<i1; i++)
+ {
+ copy_rvec(src[i],src[home_pos++]);
+ }
+ }
+ }
+ else
+ {
+ /* Copy to the communication buffer */
+ nrcg = i1 - i0;
+ pos_vec[m] += 1 + vec*nrcg;
+ for(i=i0; i<i1; i++)
+ {
+ copy_rvec(src[i],comm->cgcm_state[m][pos_vec[m]++]);
+ }
+ pos_vec[m] += (nvec - vec - 1)*nrcg;
+ }
+ if (!bCompact)
+ {
+ home_pos += i1 - i0;
+ }
+ i0 = i1;
+ }
+
+ return home_pos;
+}
+
+static int compact_and_copy_vec_cg(int ncg,int *move,
+ int *cgindex,
+ int nvec,rvec *src,gmx_domdec_comm_t *comm,
+ gmx_bool bCompact)
+{
+ int m,icg,i0,i1,nrcg;
+ int home_pos;
+ int pos_vec[DIM*2];
+
+ home_pos = 0;
+
+ for(m=0; m<DIM*2; m++)
+ {
+ pos_vec[m] = 0;
+ }
+
+ i0 = 0;
+ for(icg=0; icg<ncg; icg++)
+ {
+ i1 = cgindex[icg+1];
+ m = move[icg];
+ if (m == -1)
+ {
+ if (bCompact)
+ {
+ /* Compact the home array in place */
+ copy_rvec(src[icg],src[home_pos++]);
+ }
+ }
+ else
+ {
+ nrcg = i1 - i0;
+ /* Copy to the communication buffer */
+ copy_rvec(src[icg],comm->cgcm_state[m][pos_vec[m]]);
+ pos_vec[m] += 1 + nrcg*nvec;
+ }
+ i0 = i1;
+ }
+ if (!bCompact)
+ {
+ home_pos = ncg;
+ }
+
+ return home_pos;
+}
+
+static int compact_ind(int ncg,int *move,
+ int *index_gl,int *cgindex,
+ int *gatindex,
+ gmx_ga2la_t ga2la,char *bLocalCG,
+ int *cginfo)
+{
+ int cg,nat,a0,a1,a,a_gl;
+ int home_pos;
+
+ home_pos = 0;
+ nat = 0;
+ for(cg=0; cg<ncg; cg++)
+ {
+ a0 = cgindex[cg];
+ a1 = cgindex[cg+1];
+ if (move[cg] == -1)
+ {
+ /* Compact the home arrays in place.
+ * Anything that can be done here avoids access to global arrays.
+ */
+ cgindex[home_pos] = nat;
+ for(a=a0; a<a1; a++)
+ {
+ a_gl = gatindex[a];
+ gatindex[nat] = a_gl;
+ /* The cell number stays 0, so we don't need to set it */
+ ga2la_change_la(ga2la,a_gl,nat);
+ nat++;
+ }
+ index_gl[home_pos] = index_gl[cg];
+ cginfo[home_pos] = cginfo[cg];
+ /* The charge group remains local, so bLocalCG does not change */
+ home_pos++;
+ }
+ else
+ {
+ /* Clear the global indices */
+ for(a=a0; a<a1; a++)
+ {
+ ga2la_del(ga2la,gatindex[a]);
+ }
+ if (bLocalCG)
+ {
+ bLocalCG[index_gl[cg]] = FALSE;
+ }
+ }
+ }
+ cgindex[home_pos] = nat;
+
+ return home_pos;
+}
+
+static void clear_and_mark_ind(int ncg,int *move,
+ int *index_gl,int *cgindex,int *gatindex,
+ gmx_ga2la_t ga2la,char *bLocalCG,
+ int *cell_index)
+{
+ int cg,a0,a1,a;
+
+ for(cg=0; cg<ncg; cg++)
+ {
+ if (move[cg] >= 0)
+ {
+ a0 = cgindex[cg];
+ a1 = cgindex[cg+1];
+ /* Clear the global indices */
+ for(a=a0; a<a1; a++)
+ {
+ ga2la_del(ga2la,gatindex[a]);
+ }
+ if (bLocalCG)
+ {
+ bLocalCG[index_gl[cg]] = FALSE;
+ }
+ /* Signal that this cg has moved using the ns cell index.
+ * Here we set it to -1.
+ * fill_grid will change it from -1 to 4*grid->ncells.
+ */
+ cell_index[cg] = -1;
+ }
+ }
+}
+
+static void print_cg_move(FILE *fplog,
+ gmx_domdec_t *dd,
+ gmx_large_int_t step,int cg,int dim,int dir,
+ gmx_bool bHaveLimitdAndCMOld,real limitd,
+ rvec cm_old,rvec cm_new,real pos_d)
+{
+ gmx_domdec_comm_t *comm;
+ char buf[22];
+
+ comm = dd->comm;
+
+ fprintf(fplog,"\nStep %s:\n",gmx_step_str(step,buf));
+ if (bHaveLimitdAndCMOld)
+ {
+ fprintf(fplog,"The charge group starting at atom %d moved than the distance allowed by the domain decomposition (%f) in direction %c\n",
+ ddglatnr(dd,dd->cgindex[cg]),limitd,dim2char(dim));
+ }
+ else
+ {
+ fprintf(fplog,"The charge group starting at atom %d moved than the distance allowed by the domain decomposition in direction %c\n",
+ ddglatnr(dd,dd->cgindex[cg]),dim2char(dim));
+ }
+ fprintf(fplog,"distance out of cell %f\n",
+ dir==1 ? pos_d - comm->cell_x1[dim] : pos_d - comm->cell_x0[dim]);
+ if (bHaveLimitdAndCMOld)
+ {
+ fprintf(fplog,"Old coordinates: %8.3f %8.3f %8.3f\n",
+ cm_old[XX],cm_old[YY],cm_old[ZZ]);
+ }
+ fprintf(fplog,"New coordinates: %8.3f %8.3f %8.3f\n",
+ cm_new[XX],cm_new[YY],cm_new[ZZ]);
+ fprintf(fplog,"Old cell boundaries in direction %c: %8.3f %8.3f\n",
+ dim2char(dim),
+ comm->old_cell_x0[dim],comm->old_cell_x1[dim]);
+ fprintf(fplog,"New cell boundaries in direction %c: %8.3f %8.3f\n",
+ dim2char(dim),
+ comm->cell_x0[dim],comm->cell_x1[dim]);
+}
+
+static void cg_move_error(FILE *fplog,
+ gmx_domdec_t *dd,
+ gmx_large_int_t step,int cg,int dim,int dir,
+ gmx_bool bHaveLimitdAndCMOld,real limitd,
+ rvec cm_old,rvec cm_new,real pos_d)
+{
+ if (fplog)
+ {
+ print_cg_move(fplog, dd,step,cg,dim,dir,
+ bHaveLimitdAndCMOld,limitd,cm_old,cm_new,pos_d);
+ }
+ print_cg_move(stderr,dd,step,cg,dim,dir,
+ bHaveLimitdAndCMOld,limitd,cm_old,cm_new,pos_d);
+ gmx_fatal(FARGS,
+ "A charge group moved too far between two domain decomposition steps\n"
+ "This usually means that your system is not well equilibrated");
+}
+
+static void rotate_state_atom(t_state *state,int a)
+{
+ int est;
+
+ for(est=0; est<estNR; est++)
+ {
+ if (EST_DISTR(est) && (state->flags & (1<<est))) {
+ switch (est) {
+ case estX:
+ /* Rotate the complete state; for a rectangular box only */
+ state->x[a][YY] = state->box[YY][YY] - state->x[a][YY];
+ state->x[a][ZZ] = state->box[ZZ][ZZ] - state->x[a][ZZ];
+ break;
+ case estV:
+ state->v[a][YY] = -state->v[a][YY];
+ state->v[a][ZZ] = -state->v[a][ZZ];
+ break;
+ case estSDX:
+ state->sd_X[a][YY] = -state->sd_X[a][YY];
+ state->sd_X[a][ZZ] = -state->sd_X[a][ZZ];
+ break;
+ case estCGP:
+ state->cg_p[a][YY] = -state->cg_p[a][YY];
+ state->cg_p[a][ZZ] = -state->cg_p[a][ZZ];
+ break;
+ case estDISRE_INITF:
+ case estDISRE_RM3TAV:
+ case estORIRE_INITF:
+ case estORIRE_DTAV:
+ /* These are distances, so not affected by rotation */
+ break;
+ default:
+ gmx_incons("Unknown state entry encountered in rotate_state_atom");
+ }
+ }
+ }
+}
+
+static int dd_redistribute_cg(FILE *fplog,gmx_large_int_t step,
+ gmx_domdec_t *dd,ivec tric_dir,
+ t_state *state,rvec **f,
+ t_forcerec *fr,t_mdatoms *md,
+ gmx_bool bCompact,
+ t_nrnb *nrnb)
+{
+ int *move;
+ int npbcdim;
+ int ncg[DIM*2],nat[DIM*2];
+ int c,i,cg,k,k0,k1,d,dim,dim2,dir,d2,d3,d4,cell_d;
+ int mc,cdd,nrcg,ncg_recv,nat_recv,nvs,nvr,nvec,vec;
+ int sbuf[2],rbuf[2];
+ int home_pos_cg,home_pos_at,ncg_stay_home,buf_pos;
+ int flag;
+ gmx_bool bV=FALSE,bSDX=FALSE,bCGP=FALSE;
+ gmx_bool bScrew;
+ ivec dev;
+ real inv_ncg,pos_d;
+ matrix tcm;
+ rvec *cg_cm,cell_x0,cell_x1,limitd,limit0,limit1,cm_new;
+ atom_id *cgindex;
+ cginfo_mb_t *cginfo_mb;
+ gmx_domdec_comm_t *comm;
+
+ if (dd->bScrewPBC)
+ {
+ check_screw_box(state->box);
+ }
+
+ comm = dd->comm;
+ cg_cm = fr->cg_cm;
+
+ for(i=0; i<estNR; i++)
+ {
+ if (EST_DISTR(i))
+ {
+ switch (i)
+ {
+ case estX: /* Always present */ break;
+ case estV: bV = (state->flags & (1<<i)); break;
+ case estSDX: bSDX = (state->flags & (1<<i)); break;
+ case estCGP: bCGP = (state->flags & (1<<i)); break;
+ case estLD_RNG:
+ case estLD_RNGI:
+ case estDISRE_INITF:
+ case estDISRE_RM3TAV:
+ case estORIRE_INITF:
+ case estORIRE_DTAV:
+ /* No processing required */
+ break;
+ default:
+ gmx_incons("Unknown state entry encountered in dd_redistribute_cg");
+ }
+ }
+ }
+
+ if (dd->ncg_tot > comm->nalloc_int)
+ {
+ comm->nalloc_int = over_alloc_dd(dd->ncg_tot);
+ srenew(comm->buf_int,comm->nalloc_int);
+ }
+ move = comm->buf_int;
+
+ /* Clear the count */
+ for(c=0; c<dd->ndim*2; c++)
+ {
+ ncg[c] = 0;
+ nat[c] = 0;
+ }
+
+ npbcdim = dd->npbcdim;
+
+ for(d=0; (d<DIM); d++)
+ {
+ limitd[d] = dd->comm->cellsize_min[d];
+ if (d >= npbcdim && dd->ci[d] == 0)
+ {
+ cell_x0[d] = -GMX_FLOAT_MAX;
+ }
+ else
+ {
+ cell_x0[d] = comm->cell_x0[d];
+ }
+ if (d >= npbcdim && dd->ci[d] == dd->nc[d] - 1)
+ {
+ cell_x1[d] = GMX_FLOAT_MAX;
+ }
+ else
+ {
+ cell_x1[d] = comm->cell_x1[d];
+ }
+ if (d < npbcdim)
+ {
+ limit0[d] = comm->old_cell_x0[d] - limitd[d];
+ limit1[d] = comm->old_cell_x1[d] + limitd[d];
+ }
+ else
+ {
+ /* We check after communication if a charge group moved
+ * more than one cell. Set the pre-comm check limit to float_max.
+ */
+ limit0[d] = -GMX_FLOAT_MAX;
+ limit1[d] = GMX_FLOAT_MAX;
+ }
+ }
+
+ make_tric_corr_matrix(npbcdim,state->box,tcm);
+
+ cgindex = dd->cgindex;
+
+ /* Compute the center of geometry for all home charge groups
+ * and put them in the box and determine where they should go.
+ */
+ for(cg=0; cg<dd->ncg_home; cg++)
+ {
+ k0 = cgindex[cg];
+ k1 = cgindex[cg+1];
+ nrcg = k1 - k0;
+ if (nrcg == 1)
+ {
+ copy_rvec(state->x[k0],cm_new);
+ }
+ else
+ {
+ inv_ncg = 1.0/nrcg;
+
+ clear_rvec(cm_new);
+ for(k=k0; (k<k1); k++)
+ {
+ rvec_inc(cm_new,state->x[k]);
+ }
+ for(d=0; (d<DIM); d++)
+ {
+ cm_new[d] = inv_ncg*cm_new[d];
+ }
+ }
+
+ clear_ivec(dev);
+ /* Do pbc and check DD cell boundary crossings */
+ for(d=DIM-1; d>=0; d--)
+ {
+ if (dd->nc[d] > 1)
+ {
+ bScrew = (dd->bScrewPBC && d == XX);
+ /* Determine the location of this cg in lattice coordinates */
+ pos_d = cm_new[d];
+ if (tric_dir[d])
+ {
+ for(d2=d+1; d2<DIM; d2++)
+ {
+ pos_d += cm_new[d2]*tcm[d2][d];
+ }
+ }
+ /* Put the charge group in the triclinic unit-cell */
+ if (pos_d >= cell_x1[d])
+ {
+ if (pos_d >= limit1[d])
+ {
+ cg_move_error(fplog,dd,step,cg,d,1,TRUE,limitd[d],
+ cg_cm[cg],cm_new,pos_d);
+ }
+ dev[d] = 1;
+ if (dd->ci[d] == dd->nc[d] - 1)
+ {
+ rvec_dec(cm_new,state->box[d]);
+ if (bScrew)
+ {
+ cm_new[YY] = state->box[YY][YY] - cm_new[YY];
+ cm_new[ZZ] = state->box[ZZ][ZZ] - cm_new[ZZ];
+ }
+ for(k=k0; (k<k1); k++)
+ {
+ rvec_dec(state->x[k],state->box[d]);
+ if (bScrew)
+ {
+ rotate_state_atom(state,k);
+ }
+ }
+ }
+ }
+ else if (pos_d < cell_x0[d])
+ {
+ if (pos_d < limit0[d])
+ {
+ cg_move_error(fplog,dd,step,cg,d,-1,TRUE,limitd[d],
+ cg_cm[cg],cm_new,pos_d);
+ }
+ dev[d] = -1;
+ if (dd->ci[d] == 0)
+ {
+ rvec_inc(cm_new,state->box[d]);
+ if (bScrew)
+ {
+ cm_new[YY] = state->box[YY][YY] - cm_new[YY];
+ cm_new[ZZ] = state->box[ZZ][ZZ] - cm_new[ZZ];
+ }
+ for(k=k0; (k<k1); k++)
+ {
+ rvec_inc(state->x[k],state->box[d]);
+ if (bScrew)
+ {
+ rotate_state_atom(state,k);
+ }
+ }
+ }
+ }
+ }
+ else if (d < npbcdim)
+ {
+ /* Put the charge group in the rectangular unit-cell */
+ while (cm_new[d] >= state->box[d][d])
+ {
+ rvec_dec(cm_new,state->box[d]);
+ for(k=k0; (k<k1); k++)
+ {
+ rvec_dec(state->x[k],state->box[d]);
+ }
+ }
+ while (cm_new[d] < 0)
+ {
+ rvec_inc(cm_new,state->box[d]);
+ for(k=k0; (k<k1); k++)
+ {
+ rvec_inc(state->x[k],state->box[d]);
+ }
+ }
+ }
+ }
+
+ copy_rvec(cm_new,cg_cm[cg]);
+
+ /* Determine where this cg should go */
+ flag = 0;
+ mc = -1;
+ for(d=0; d<dd->ndim; d++)
+ {
+ dim = dd->dim[d];
+ if (dev[dim] == 1)
+ {
+ flag |= DD_FLAG_FW(d);
+ if (mc == -1)
+ {
+ mc = d*2;
+ }
+ }
+ else if (dev[dim] == -1)
+ {
+ flag |= DD_FLAG_BW(d);
+ if (mc == -1) {
+ if (dd->nc[dim] > 2)
+ {
+ mc = d*2 + 1;
+ }
+ else
+ {
+ mc = d*2;
+ }
+ }
+ }
+ }
+ move[cg] = mc;
+ if (mc >= 0)
+ {
+ if (ncg[mc]+1 > comm->cggl_flag_nalloc[mc])
+ {
+ comm->cggl_flag_nalloc[mc] = over_alloc_dd(ncg[mc]+1);
+ srenew(comm->cggl_flag[mc],comm->cggl_flag_nalloc[mc]*DD_CGIBS);
+ }
+ comm->cggl_flag[mc][ncg[mc]*DD_CGIBS ] = dd->index_gl[cg];
+ /* We store the cg size in the lower 16 bits
+ * and the place where the charge group should go
+ * in the next 6 bits. This saves some communication volume.
+ */
+ comm->cggl_flag[mc][ncg[mc]*DD_CGIBS+1] = nrcg | flag;
+ ncg[mc] += 1;
+ nat[mc] += nrcg;
+ }
+ }
+
+ inc_nrnb(nrnb,eNR_CGCM,dd->nat_home);
+ inc_nrnb(nrnb,eNR_RESETX,dd->ncg_home);
+
+ nvec = 1;
+ if (bV)
+ {
+ nvec++;
+ }
+ if (bSDX)
+ {
+ nvec++;
+ }
+ if (bCGP)
+ {
+ nvec++;
+ }
+
+ /* Make sure the communication buffers are large enough */
+ for(mc=0; mc<dd->ndim*2; mc++)
+ {
+ nvr = ncg[mc] + nat[mc]*nvec;
+ if (nvr > comm->cgcm_state_nalloc[mc])
+ {
+ comm->cgcm_state_nalloc[mc] = over_alloc_dd(nvr);
+ srenew(comm->cgcm_state[mc],comm->cgcm_state_nalloc[mc]);
+ }
+ }
+
+ /* Recalculating cg_cm might be cheaper than communicating,
+ * but that could give rise to rounding issues.
+ */
+ home_pos_cg =
+ compact_and_copy_vec_cg(dd->ncg_home,move,cgindex,
+ nvec,cg_cm,comm,bCompact);
+
+ vec = 0;
+ home_pos_at =
+ compact_and_copy_vec_at(dd->ncg_home,move,cgindex,
+ nvec,vec++,state->x,comm,bCompact);
+ if (bV)
+ {
+ compact_and_copy_vec_at(dd->ncg_home,move,cgindex,
+ nvec,vec++,state->v,comm,bCompact);
+ }
+ if (bSDX)
+ {
+ compact_and_copy_vec_at(dd->ncg_home,move,cgindex,
+ nvec,vec++,state->sd_X,comm,bCompact);
+ }
+ if (bCGP)
+ {
+ compact_and_copy_vec_at(dd->ncg_home,move,cgindex,
+ nvec,vec++,state->cg_p,comm,bCompact);
+ }
+
+ if (bCompact)
+ {
+ compact_ind(dd->ncg_home,move,
+ dd->index_gl,dd->cgindex,dd->gatindex,
+ dd->ga2la,comm->bLocalCG,
+ fr->cginfo);
+ }
+ else
+ {
+ clear_and_mark_ind(dd->ncg_home,move,
+ dd->index_gl,dd->cgindex,dd->gatindex,
+ dd->ga2la,comm->bLocalCG,
+ fr->ns.grid->cell_index);
+ }
+
+ cginfo_mb = fr->cginfo_mb;
+
+ ncg_stay_home = home_pos_cg;
+ for(d=0; d<dd->ndim; d++)
+ {
+ dim = dd->dim[d];
+ ncg_recv = 0;
+ nat_recv = 0;
+ nvr = 0;
+ for(dir=0; dir<(dd->nc[dim]==2 ? 1 : 2); dir++)
+ {
+ cdd = d*2 + dir;
+ /* Communicate the cg and atom counts */
+ sbuf[0] = ncg[cdd];
+ sbuf[1] = nat[cdd];
+ if (debug)
+ {
+ fprintf(debug,"Sending ddim %d dir %d: ncg %d nat %d\n",
+ d,dir,sbuf[0],sbuf[1]);
+ }
+ dd_sendrecv_int(dd, d, dir, sbuf, 2, rbuf, 2);
+
+ if ((ncg_recv+rbuf[0])*DD_CGIBS > comm->nalloc_int)
+ {
+ comm->nalloc_int = over_alloc_dd((ncg_recv+rbuf[0])*DD_CGIBS);
+ srenew(comm->buf_int,comm->nalloc_int);
+ }
+
+ /* Communicate the charge group indices, sizes and flags */
+ dd_sendrecv_int(dd, d, dir,
+ comm->cggl_flag[cdd], sbuf[0]*DD_CGIBS,
+ comm->buf_int+ncg_recv*DD_CGIBS, rbuf[0]*DD_CGIBS);
+
+ nvs = ncg[cdd] + nat[cdd]*nvec;
+ i = rbuf[0] + rbuf[1] *nvec;
+ vec_rvec_check_alloc(&comm->vbuf,nvr+i);
+
+ /* Communicate cgcm and state */
+ dd_sendrecv_rvec(dd, d, dir,
+ comm->cgcm_state[cdd], nvs,
+ comm->vbuf.v+nvr, i);
+ ncg_recv += rbuf[0];
+ nat_recv += rbuf[1];
+ nvr += i;
+ }
+
+ /* Process the received charge groups */
+ buf_pos = 0;
+ for(cg=0; cg<ncg_recv; cg++)
+ {
+ flag = comm->buf_int[cg*DD_CGIBS+1];
+
+ if (dim >= npbcdim && dd->nc[dim] > 2)
+ {
+ /* No pbc in this dim and more than one domain boundary.
+ * We to a separate check if a charge did not move too far.
+ */
+ if (((flag & DD_FLAG_FW(d)) &&
+ comm->vbuf.v[buf_pos][d] > cell_x1[dim]) ||
+ ((flag & DD_FLAG_BW(d)) &&
+ comm->vbuf.v[buf_pos][d] < cell_x0[dim]))
+ {
+ cg_move_error(fplog,dd,step,cg,d,
+ (flag & DD_FLAG_FW(d)) ? 1 : 0,
+ FALSE,0,
+ comm->vbuf.v[buf_pos],
+ comm->vbuf.v[buf_pos],
+ comm->vbuf.v[buf_pos][d]);
+ }
+ }
+
+ mc = -1;
+ if (d < dd->ndim-1)
+ {
+ /* Check which direction this cg should go */
+ for(d2=d+1; (d2<dd->ndim && mc==-1); d2++)
+ {
+ if (dd->bGridJump)
+ {
+ /* The cell boundaries for dimension d2 are not equal
+ * for each cell row of the lower dimension(s),
+ * therefore we might need to redetermine where
+ * this cg should go.
+ */
+ dim2 = dd->dim[d2];
+ /* If this cg crosses the box boundary in dimension d2
+ * we can use the communicated flag, so we do not
+ * have to worry about pbc.
+ */
+ if (!((dd->ci[dim2] == dd->nc[dim2]-1 &&
+ (flag & DD_FLAG_FW(d2))) ||
+ (dd->ci[dim2] == 0 &&
+ (flag & DD_FLAG_BW(d2)))))
+ {
+ /* Clear the two flags for this dimension */
+ flag &= ~(DD_FLAG_FW(d2) | DD_FLAG_BW(d2));
+ /* Determine the location of this cg
+ * in lattice coordinates
+ */
+ pos_d = comm->vbuf.v[buf_pos][dim2];
+ if (tric_dir[dim2])
+ {
+ for(d3=dim2+1; d3<DIM; d3++)
+ {
+ pos_d +=
+ comm->vbuf.v[buf_pos][d3]*tcm[d3][dim2];
+ }
+ }
+ /* Check of we are not at the box edge.
+ * pbc is only handled in the first step above,
+ * but this check could move over pbc while
+ * the first step did not due to different rounding.
+ */
+ if (pos_d >= cell_x1[dim2] &&
+ dd->ci[dim2] != dd->nc[dim2]-1)
+ {
+ flag |= DD_FLAG_FW(d2);
+ }
+ else if (pos_d < cell_x0[dim2] &&
+ dd->ci[dim2] != 0)
+ {
+ flag |= DD_FLAG_BW(d2);
+ }
+ comm->buf_int[cg*DD_CGIBS+1] = flag;
+ }
+ }
+ /* Set to which neighboring cell this cg should go */
+ if (flag & DD_FLAG_FW(d2))
+ {
+ mc = d2*2;
+ }
+ else if (flag & DD_FLAG_BW(d2))
+ {
+ if (dd->nc[dd->dim[d2]] > 2)
+ {
+ mc = d2*2+1;
+ }
+ else
+ {
+ mc = d2*2;
+ }
+ }
+ }
+ }
+
+ nrcg = flag & DD_FLAG_NRCG;
+ if (mc == -1)
+ {
+ if (home_pos_cg+1 > dd->cg_nalloc)
+ {
+ dd->cg_nalloc = over_alloc_dd(home_pos_cg+1);
+ srenew(dd->index_gl,dd->cg_nalloc);
+ srenew(dd->cgindex,dd->cg_nalloc+1);
+ }
+ /* Set the global charge group index and size */
+ dd->index_gl[home_pos_cg] = comm->buf_int[cg*DD_CGIBS];
+ dd->cgindex[home_pos_cg+1] = dd->cgindex[home_pos_cg] + nrcg;
+ /* Copy the state from the buffer */
+ if (home_pos_cg >= fr->cg_nalloc)
+ {
+ dd_realloc_fr_cg(fr,home_pos_cg+1);
+ cg_cm = fr->cg_cm;
+ }
+ copy_rvec(comm->vbuf.v[buf_pos++],cg_cm[home_pos_cg]);
+ /* Set the cginfo */
+ fr->cginfo[home_pos_cg] = ddcginfo(cginfo_mb,
+ dd->index_gl[home_pos_cg]);
+ if (comm->bLocalCG)
+ {
+ comm->bLocalCG[dd->index_gl[home_pos_cg]] = TRUE;
+ }
+
+ if (home_pos_at+nrcg > state->nalloc)
+ {
+ dd_realloc_state(state,f,home_pos_at+nrcg);
+ }
+ for(i=0; i<nrcg; i++)
+ {
+ copy_rvec(comm->vbuf.v[buf_pos++],
+ state->x[home_pos_at+i]);
+ }
+ if (bV)
+ {
+ for(i=0; i<nrcg; i++)
+ {
+ copy_rvec(comm->vbuf.v[buf_pos++],
+ state->v[home_pos_at+i]);
+ }
+ }
+ if (bSDX)
+ {
+ for(i=0; i<nrcg; i++)
+ {
+ copy_rvec(comm->vbuf.v[buf_pos++],
+ state->sd_X[home_pos_at+i]);
+ }
+ }
+ if (bCGP)
+ {
+ for(i=0; i<nrcg; i++)
+ {
+ copy_rvec(comm->vbuf.v[buf_pos++],
+ state->cg_p[home_pos_at+i]);
+ }
+ }
+ home_pos_cg += 1;
+ home_pos_at += nrcg;
+ }
+ else
+ {
+ /* Reallocate the buffers if necessary */
+ if (ncg[mc]+1 > comm->cggl_flag_nalloc[mc])
+ {
+ comm->cggl_flag_nalloc[mc] = over_alloc_dd(ncg[mc]+1);
+ srenew(comm->cggl_flag[mc],comm->cggl_flag_nalloc[mc]*DD_CGIBS);
+ }
+ nvr = ncg[mc] + nat[mc]*nvec;
+ if (nvr + 1 + nrcg*nvec > comm->cgcm_state_nalloc[mc])
+ {
+ comm->cgcm_state_nalloc[mc] = over_alloc_dd(nvr + 1 + nrcg*nvec);
+ srenew(comm->cgcm_state[mc],comm->cgcm_state_nalloc[mc]);
+ }
+ /* Copy from the receive to the send buffers */
+ memcpy(comm->cggl_flag[mc] + ncg[mc]*DD_CGIBS,
+ comm->buf_int + cg*DD_CGIBS,
+ DD_CGIBS*sizeof(int));
+ memcpy(comm->cgcm_state[mc][nvr],
+ comm->vbuf.v[buf_pos],
+ (1+nrcg*nvec)*sizeof(rvec));
+ buf_pos += 1 + nrcg*nvec;
+ ncg[mc] += 1;
+ nat[mc] += nrcg;
+ }
+ }
+ }
+
+ /* With sorting (!bCompact) the indices are now only partially up to date
+ * and ncg_home and nat_home are not the real count, since there are
+ * "holes" in the arrays for the charge groups that moved to neighbors.
+ */
+ dd->ncg_home = home_pos_cg;
+ dd->nat_home = home_pos_at;
+
+ if (debug)
+ {
+ fprintf(debug,"Finished repartitioning\n");
+ }
+
+ return ncg_stay_home;
+}
+
+void dd_cycles_add(gmx_domdec_t *dd,float cycles,int ddCycl)
+{
+ dd->comm->cycl[ddCycl] += cycles;
+ dd->comm->cycl_n[ddCycl]++;
+ if (cycles > dd->comm->cycl_max[ddCycl])
+ {
+ dd->comm->cycl_max[ddCycl] = cycles;
+ }
+}
+
+static double force_flop_count(t_nrnb *nrnb)
+{
+ int i;
+ double sum;
+ const char *name;
+
+ sum = 0;
+ for(i=eNR_NBKERNEL010; i<eNR_NBKERNEL_FREE_ENERGY; i++)
+ {
+ /* To get closer to the real timings, we half the count
+ * for the normal loops and again half it for water loops.
+ */
+ name = nrnb_str(i);
+ if (strstr(name,"W3") != NULL || strstr(name,"W4") != NULL)
+ {
+ sum += nrnb->n[i]*0.25*cost_nrnb(i);
+ }
+ else
+ {
+ sum += nrnb->n[i]*0.50*cost_nrnb(i);
+ }
+ }
+ for(i=eNR_NBKERNEL_FREE_ENERGY; i<=eNR_NB14; i++)
+ {
+ name = nrnb_str(i);
+ if (strstr(name,"W3") != NULL || strstr(name,"W4") != NULL)
+ sum += nrnb->n[i]*cost_nrnb(i);
+ }
+ for(i=eNR_BONDS; i<=eNR_WALLS; i++)
+ {
+ sum += nrnb->n[i]*cost_nrnb(i);
+ }
+
+ return sum;
+}
+
+void dd_force_flop_start(gmx_domdec_t *dd,t_nrnb *nrnb)
+{
+ if (dd->comm->eFlop)
+ {
+ dd->comm->flop -= force_flop_count(nrnb);
+ }
+}
+void dd_force_flop_stop(gmx_domdec_t *dd,t_nrnb *nrnb)
+{
+ if (dd->comm->eFlop)
+ {
+ dd->comm->flop += force_flop_count(nrnb);
+ dd->comm->flop_n++;
+ }
+}
+
+static void clear_dd_cycle_counts(gmx_domdec_t *dd)
+{
+ int i;
+
+ for(i=0; i<ddCyclNr; i++)
+ {
+ dd->comm->cycl[i] = 0;
+ dd->comm->cycl_n[i] = 0;
+ dd->comm->cycl_max[i] = 0;
+ }
+ dd->comm->flop = 0;
+ dd->comm->flop_n = 0;
+}
+
+static void get_load_distribution(gmx_domdec_t *dd,gmx_wallcycle_t wcycle)
+{
+ gmx_domdec_comm_t *comm;
+ gmx_domdec_load_t *load;
+ gmx_domdec_root_t *root=NULL;
+ int d,dim,cid,i,pos;
+ float cell_frac=0,sbuf[DD_NLOAD_MAX];
+ gmx_bool bSepPME;
+
+ if (debug)
+ {
+ fprintf(debug,"get_load_distribution start\n");
+ }
+
+ wallcycle_start(wcycle,ewcDDCOMMLOAD);
+
+ comm = dd->comm;
+
+ bSepPME = (dd->pme_nodeid >= 0);
+
+ for(d=dd->ndim-1; d>=0; d--)
+ {
+ dim = dd->dim[d];
+ /* Check if we participate in the communication in this dimension */
+ if (d == dd->ndim-1 ||
+ (dd->ci[dd->dim[d+1]]==0 && dd->ci[dd->dim[dd->ndim-1]]==0))
+ {
+ load = &comm->load[d];
+ if (dd->bGridJump)
+ {
+ cell_frac = comm->cell_f1[d] - comm->cell_f0[d];
+ }
+ pos = 0;
+ if (d == dd->ndim-1)
+ {
+ sbuf[pos++] = dd_force_load(comm);
+ sbuf[pos++] = sbuf[0];
+ if (dd->bGridJump)
+ {
+ sbuf[pos++] = sbuf[0];
+ sbuf[pos++] = cell_frac;
+ if (d > 0)
+ {
+ sbuf[pos++] = comm->cell_f_max0[d];
+ sbuf[pos++] = comm->cell_f_min1[d];
+ }
+ }
+ if (bSepPME)
+ {
+ sbuf[pos++] = comm->cycl[ddCyclPPduringPME];
+ sbuf[pos++] = comm->cycl[ddCyclPME];
+ }
+ }
+ else
+ {
+ sbuf[pos++] = comm->load[d+1].sum;
+ sbuf[pos++] = comm->load[d+1].max;
+ if (dd->bGridJump)
+ {
+ sbuf[pos++] = comm->load[d+1].sum_m;
+ sbuf[pos++] = comm->load[d+1].cvol_min*cell_frac;
+ sbuf[pos++] = comm->load[d+1].flags;
+ if (d > 0)
+ {
+ sbuf[pos++] = comm->cell_f_max0[d];
+ sbuf[pos++] = comm->cell_f_min1[d];
+ }
+ }
+ if (bSepPME)
+ {
+ sbuf[pos++] = comm->load[d+1].mdf;
+ sbuf[pos++] = comm->load[d+1].pme;
+ }
+ }
+ load->nload = pos;
+ /* Communicate a row in DD direction d.
+ * The communicators are setup such that the root always has rank 0.
+ */
+#ifdef GMX_MPI
+ MPI_Gather(sbuf ,load->nload*sizeof(float),MPI_BYTE,
+ load->load,load->nload*sizeof(float),MPI_BYTE,
+ 0,comm->mpi_comm_load[d]);
+#endif
+ if (dd->ci[dim] == dd->master_ci[dim])
+ {
+ /* We are the root, process this row */
+ if (comm->bDynLoadBal)
+ {
+ root = comm->root[d];
+ }
+ load->sum = 0;
+ load->max = 0;
+ load->sum_m = 0;
+ load->cvol_min = 1;
+ load->flags = 0;
+ load->mdf = 0;
+ load->pme = 0;
+ pos = 0;
+ for(i=0; i<dd->nc[dim]; i++)
+ {
+ load->sum += load->load[pos++];
+ load->max = max(load->max,load->load[pos]);
+ pos++;
+ if (dd->bGridJump)
+ {
+ if (root->bLimited)
+ {
+ /* This direction could not be load balanced properly,
+ * therefore we need to use the maximum iso the average load.
+ */
+ load->sum_m = max(load->sum_m,load->load[pos]);
+ }
+ else
+ {
+ load->sum_m += load->load[pos];
+ }
+ pos++;
+ load->cvol_min = min(load->cvol_min,load->load[pos]);
+ pos++;
+ if (d < dd->ndim-1)
+ {
+ load->flags = (int)(load->load[pos++] + 0.5);
+ }
+ if (d > 0)
+ {
+ root->cell_f_max0[i] = load->load[pos++];
+ root->cell_f_min1[i] = load->load[pos++];
+ }
+ }
+ if (bSepPME)
+ {
+ load->mdf = max(load->mdf,load->load[pos]);
+ pos++;
+ load->pme = max(load->pme,load->load[pos]);
+ pos++;
+ }
+ }
+ if (comm->bDynLoadBal && root->bLimited)
+ {
+ load->sum_m *= dd->nc[dim];
+ load->flags |= (1<<d);
+ }
+ }
+ }
+ }
+
+ if (DDMASTER(dd))
+ {
+ comm->nload += dd_load_count(comm);
+ comm->load_step += comm->cycl[ddCyclStep];
+ comm->load_sum += comm->load[0].sum;
+ comm->load_max += comm->load[0].max;
+ if (comm->bDynLoadBal)
+ {
+ for(d=0; d<dd->ndim; d++)
+ {
+ if (comm->load[0].flags & (1<<d))
+ {
+ comm->load_lim[d]++;
+ }
+ }
+ }
+ if (bSepPME)
+ {
+ comm->load_mdf += comm->load[0].mdf;
+ comm->load_pme += comm->load[0].pme;
+ }
+ }
+
+ wallcycle_stop(wcycle,ewcDDCOMMLOAD);
+
+ if (debug)
+ {
+ fprintf(debug,"get_load_distribution finished\n");
+ }
+}
+
+static float dd_force_imb_perf_loss(gmx_domdec_t *dd)
+{
+ /* Return the relative performance loss on the total run time
+ * due to the force calculation load imbalance.
+ */
+ if (dd->comm->nload > 0)
+ {
+ return
+ (dd->comm->load_max*dd->nnodes - dd->comm->load_sum)/
+ (dd->comm->load_step*dd->nnodes);
+ }
+ else
+ {
+ return 0;
+ }
+}
+
+static void print_dd_load_av(FILE *fplog,gmx_domdec_t *dd)
+{
+ char buf[STRLEN];
+ int npp,npme,nnodes,d,limp;
+ float imbal,pme_f_ratio,lossf,lossp=0;
+ gmx_bool bLim;
+ gmx_domdec_comm_t *comm;
+
+ comm = dd->comm;
+ if (DDMASTER(dd) && comm->nload > 0)
+ {
+ npp = dd->nnodes;
+ npme = (dd->pme_nodeid >= 0) ? comm->npmenodes : 0;
+ nnodes = npp + npme;
+ imbal = comm->load_max*npp/comm->load_sum - 1;
+ lossf = dd_force_imb_perf_loss(dd);
+ sprintf(buf," Average load imbalance: %.1f %%\n",imbal*100);
+ fprintf(fplog,"%s",buf);
+ fprintf(stderr,"\n");
+ fprintf(stderr,"%s",buf);
+ sprintf(buf," Part of the total run time spent waiting due to load imbalance: %.1f %%\n",lossf*100);
+ fprintf(fplog,"%s",buf);
+ fprintf(stderr,"%s",buf);
+ bLim = FALSE;
+ if (comm->bDynLoadBal)
+ {
+ sprintf(buf," Steps where the load balancing was limited by -rdd, -rcon and/or -dds:");
+ for(d=0; d<dd->ndim; d++)
+ {
+ limp = (200*comm->load_lim[d]+1)/(2*comm->nload);
+ sprintf(buf+strlen(buf)," %c %d %%",dim2char(dd->dim[d]),limp);
+ if (limp >= 50)
+ {
+ bLim = TRUE;
+ }
+ }
+ sprintf(buf+strlen(buf),"\n");
+ fprintf(fplog,"%s",buf);
+ fprintf(stderr,"%s",buf);
+ }
+ if (npme > 0)
+ {
+ pme_f_ratio = comm->load_pme/comm->load_mdf;
+ lossp = (comm->load_pme -comm->load_mdf)/comm->load_step;
+ if (lossp <= 0)
+ {
+ lossp *= (float)npme/(float)nnodes;
+ }
+ else
+ {
+ lossp *= (float)npp/(float)nnodes;
+ }
+ sprintf(buf," Average PME mesh/force load: %5.3f\n",pme_f_ratio);
+ fprintf(fplog,"%s",buf);
+ fprintf(stderr,"%s",buf);
+ sprintf(buf," Part of the total run time spent waiting due to PP/PME imbalance: %.1f %%\n",fabs(lossp)*100);
+ fprintf(fplog,"%s",buf);
+ fprintf(stderr,"%s",buf);
+ }
+ fprintf(fplog,"\n");
+ fprintf(stderr,"\n");
+
+ if (lossf >= DD_PERF_LOSS)
+ {
+ sprintf(buf,
+ "NOTE: %.1f %% performance was lost due to load imbalance\n"
+ " in the domain decomposition.\n",lossf*100);
+ if (!comm->bDynLoadBal)
+ {
+ sprintf(buf+strlen(buf)," You might want to use dynamic load balancing (option -dlb.)\n");
+ }
+ else if (bLim)
+ {
+ sprintf(buf+strlen(buf)," You might want to decrease the cell size limit (options -rdd, -rcon and/or -dds).\n");
+ }
+ fprintf(fplog,"%s\n",buf);
+ fprintf(stderr,"%s\n",buf);
+ }
+ if (npme > 0 && fabs(lossp) >= DD_PERF_LOSS)
+ {
+ sprintf(buf,
+ "NOTE: %.1f %% performance was lost because the PME nodes\n"
+ " had %s work to do than the PP nodes.\n"
+ " You might want to %s the number of PME nodes\n"
+ " or %s the cut-off and the grid spacing.\n",
+ fabs(lossp*100),
+ (lossp < 0) ? "less" : "more",
+ (lossp < 0) ? "decrease" : "increase",
+ (lossp < 0) ? "decrease" : "increase");
+ fprintf(fplog,"%s\n",buf);
+ fprintf(stderr,"%s\n",buf);
+ }
+ }
+}
+
+static float dd_vol_min(gmx_domdec_t *dd)
+{
+ return dd->comm->load[0].cvol_min*dd->nnodes;
+}
+
+static gmx_bool dd_load_flags(gmx_domdec_t *dd)
+{
+ return dd->comm->load[0].flags;
+}
+
+static float dd_f_imbal(gmx_domdec_t *dd)
+{
+ return dd->comm->load[0].max*dd->nnodes/dd->comm->load[0].sum - 1;
+}
+
+static float dd_pme_f_ratio(gmx_domdec_t *dd)
+{
+ return dd->comm->load[0].pme/dd->comm->load[0].mdf;
+}
+
+static void dd_print_load(FILE *fplog,gmx_domdec_t *dd,gmx_large_int_t step)
+{
+ int flags,d;
+ char buf[22];
+
+ flags = dd_load_flags(dd);
+ if (flags)
+ {
+ fprintf(fplog,
+ "DD load balancing is limited by minimum cell size in dimension");
+ for(d=0; d<dd->ndim; d++)
+ {
+ if (flags & (1<<d))
+ {
+ fprintf(fplog," %c",dim2char(dd->dim[d]));
+ }
+ }
+ fprintf(fplog,"\n");
+ }
+ fprintf(fplog,"DD step %s",gmx_step_str(step,buf));
+ if (dd->comm->bDynLoadBal)
+ {
+ fprintf(fplog," vol min/aver %5.3f%c",
+ dd_vol_min(dd),flags ? '!' : ' ');
+ }
+ fprintf(fplog," load imb.: force %4.1f%%",dd_f_imbal(dd)*100);
+ if (dd->comm->cycl_n[ddCyclPME])
+ {
+ fprintf(fplog," pme mesh/force %5.3f",dd_pme_f_ratio(dd));
+ }
+ fprintf(fplog,"\n\n");
+}
+
+static void dd_print_load_verbose(gmx_domdec_t *dd)
+{
+ if (dd->comm->bDynLoadBal)
+ {
+ fprintf(stderr,"vol %4.2f%c ",
+ dd_vol_min(dd),dd_load_flags(dd) ? '!' : ' ');
+ }
+ fprintf(stderr,"imb F %2d%% ",(int)(dd_f_imbal(dd)*100+0.5));
+ if (dd->comm->cycl_n[ddCyclPME])
+ {
+ fprintf(stderr,"pme/F %4.2f ",dd_pme_f_ratio(dd));
+ }
+}
+
+#ifdef GMX_MPI
+static void make_load_communicator(gmx_domdec_t *dd,MPI_Group g_all,
+ int dim_ind,ivec loc)
+{
+ MPI_Group g_row = MPI_GROUP_EMPTY;
+ MPI_Comm c_row;
+ int dim,i,*rank;
+ ivec loc_c;
+ gmx_domdec_root_t *root;
+ gmx_bool bPartOfGroup = FALSE;
+
+ dim = dd->dim[dim_ind];
+ copy_ivec(loc,loc_c);
+ snew(rank,dd->nc[dim]);
+ for(i=0; i<dd->nc[dim]; i++)
+ {
+ loc_c[dim] = i;
+ rank[i] = dd_index(dd->nc,loc_c);
+ if (rank[i] == dd->rank)
+ {
+ /* This process is part of the group */
+ bPartOfGroup = TRUE;
+ }
+ }
+ if (bPartOfGroup)
+ {
+ MPI_Group_incl(g_all,dd->nc[dim],rank,&g_row);
+ }
+ MPI_Comm_create(dd->mpi_comm_all,g_row,&c_row);
+ if (bPartOfGroup)
+ {
+ dd->comm->mpi_comm_load[dim_ind] = c_row;
+ if (dd->comm->eDLB != edlbNO)
+ {
+ if (dd->ci[dim] == dd->master_ci[dim])
+ {
+ /* This is the root process of this row */
+ snew(dd->comm->root[dim_ind],1);
+ root = dd->comm->root[dim_ind];
+ snew(root->cell_f,DD_CELL_F_SIZE(dd,dim_ind));
+ snew(root->old_cell_f,dd->nc[dim]+1);
+ snew(root->bCellMin,dd->nc[dim]);
+ if (dim_ind > 0)
+ {
+ snew(root->cell_f_max0,dd->nc[dim]);
+ snew(root->cell_f_min1,dd->nc[dim]);
+ snew(root->bound_min,dd->nc[dim]);
+ snew(root->bound_max,dd->nc[dim]);
+ }
+ snew(root->buf_ncd,dd->nc[dim]);
+ }
+ else
+ {
+ /* This is not a root process, we only need to receive cell_f */
+ snew(dd->comm->cell_f_row,DD_CELL_F_SIZE(dd,dim_ind));
+ }
+ }
+ if (dd->ci[dim] == dd->master_ci[dim])
+ {
+ snew(dd->comm->load[dim_ind].load,dd->nc[dim]*DD_NLOAD_MAX);
+ }
+ }
+ sfree(rank);
+}
+#endif
+
+static void make_load_communicators(gmx_domdec_t *dd)
+{
+#ifdef GMX_MPI
+ MPI_Group g_all;
+ int dim0,dim1,i,j;
+ ivec loc;
+
+ if (debug)
+ fprintf(debug,"Making load communicators\n");
+
+ MPI_Comm_group(dd->mpi_comm_all,&g_all);
+
+ snew(dd->comm->load,dd->ndim);
+ snew(dd->comm->mpi_comm_load,dd->ndim);
+
+ clear_ivec(loc);
+ make_load_communicator(dd,g_all,0,loc);
+ if (dd->ndim > 1) {
+ dim0 = dd->dim[0];
+ for(i=0; i<dd->nc[dim0]; i++) {
+ loc[dim0] = i;
+ make_load_communicator(dd,g_all,1,loc);
+ }
+ }
+ if (dd->ndim > 2) {
+ dim0 = dd->dim[0];
+ for(i=0; i<dd->nc[dim0]; i++) {
+ loc[dim0] = i;
+ dim1 = dd->dim[1];
+ for(j=0; j<dd->nc[dim1]; j++) {
+ loc[dim1] = j;
+ make_load_communicator(dd,g_all,2,loc);
+ }
+ }
+ }
+
+ MPI_Group_free(&g_all);
+
+ if (debug)
+ fprintf(debug,"Finished making load communicators\n");
+#endif
+}
+
+void setup_dd_grid(FILE *fplog,gmx_domdec_t *dd)
+{
+ gmx_bool bZYX;
+ int d,dim,i,j,m;
+ ivec tmp,s;
+ int nzone,nzonep;
+ ivec dd_zp[DD_MAXIZONE];
+ gmx_domdec_zones_t *zones;
+ gmx_domdec_ns_ranges_t *izone;
+
+ for(d=0; d<dd->ndim; d++)
+ {
+ dim = dd->dim[d];
+ copy_ivec(dd->ci,tmp);
+ tmp[dim] = (tmp[dim] + 1) % dd->nc[dim];
+ dd->neighbor[d][0] = ddcoord2ddnodeid(dd,tmp);
+ copy_ivec(dd->ci,tmp);
+ tmp[dim] = (tmp[dim] - 1 + dd->nc[dim]) % dd->nc[dim];
+ dd->neighbor[d][1] = ddcoord2ddnodeid(dd,tmp);
+ if (debug)
+ {
+ fprintf(debug,"DD rank %d neighbor ranks in dir %d are + %d - %d\n",
+ dd->rank,dim,
+ dd->neighbor[d][0],
+ dd->neighbor[d][1]);
+ }
+ }
+
+ if (DDMASTER(dd))
+ {
+ fprintf(stderr,"Making %dD domain decomposition %d x %d x %d\n",
+ dd->ndim,dd->nc[XX],dd->nc[YY],dd->nc[ZZ]);
+ }
+ if (fplog)
+ {
+ fprintf(fplog,"\nMaking %dD domain decomposition grid %d x %d x %d, home cell index %d %d %d\n\n",
+ dd->ndim,
+ dd->nc[XX],dd->nc[YY],dd->nc[ZZ],
+ dd->ci[XX],dd->ci[YY],dd->ci[ZZ]);
+ }
+ switch (dd->ndim)
+ {
+ case 3:
+ nzone = dd_z3n;
+ nzonep = dd_zp3n;
+ for(i=0; i<nzonep; i++)
+ {
+ copy_ivec(dd_zp3[i],dd_zp[i]);
+ }
+ break;
+ case 2:
+ nzone = dd_z2n;
+ nzonep = dd_zp2n;
+ for(i=0; i<nzonep; i++)
+ {
+ copy_ivec(dd_zp2[i],dd_zp[i]);
+ }
+ break;
+ case 1:
+ nzone = dd_z1n;
+ nzonep = dd_zp1n;
+ for(i=0; i<nzonep; i++)
+ {
+ copy_ivec(dd_zp1[i],dd_zp[i]);
+ }
+ break;
+ default:
+ gmx_fatal(FARGS,"Can only do 1, 2 or 3D domain decomposition");
+ nzone = 0;
+ nzonep = 0;
+ }
+
+ zones = &dd->comm->zones;
+
+ for(i=0; i<nzone; i++)
+ {
+ m = 0;
+ clear_ivec(zones->shift[i]);
+ for(d=0; d<dd->ndim; d++)
+ {
+ zones->shift[i][dd->dim[d]] = dd_zo[i][m++];
+ }
+ }
+
+ zones->n = nzone;
+ for(i=0; i<nzone; i++)
+ {
+ for(d=0; d<DIM; d++)
+ {
+ s[d] = dd->ci[d] - zones->shift[i][d];
+ if (s[d] < 0)
+ {
+ s[d] += dd->nc[d];
+ }
+ else if (s[d] >= dd->nc[d])
+ {
+ s[d] -= dd->nc[d];
+ }
+ }
+ }
+ zones->nizone = nzonep;
+ for(i=0; i<zones->nizone; i++)
+ {
+ if (dd_zp[i][0] != i)
+ {
+ gmx_fatal(FARGS,"Internal inconsistency in the dd grid setup");
+ }
+ izone = &zones->izone[i];
+ izone->j0 = dd_zp[i][1];
+ izone->j1 = dd_zp[i][2];
+ for(dim=0; dim<DIM; dim++)
+ {
+ if (dd->nc[dim] == 1)
+ {
+ /* All shifts should be allowed */
+ izone->shift0[dim] = -1;
+ izone->shift1[dim] = 1;
+ }
+ else
+ {
+ /*
+ izone->shift0[d] = 0;
+ izone->shift1[d] = 0;
+ for(j=izone->j0; j<izone->j1; j++) {
+ if (dd->shift[j][d] > dd->shift[i][d])
+ izone->shift0[d] = -1;
+ if (dd->shift[j][d] < dd->shift[i][d])
+ izone->shift1[d] = 1;
+ }
+ */
+
+ int shift_diff;
+
+ /* Assume the shift are not more than 1 cell */
+ izone->shift0[dim] = 1;
+ izone->shift1[dim] = -1;
+ for(j=izone->j0; j<izone->j1; j++)
+ {
+ shift_diff = zones->shift[j][dim] - zones->shift[i][dim];
+ if (shift_diff < izone->shift0[dim])
+ {
+ izone->shift0[dim] = shift_diff;
+ }
+ if (shift_diff > izone->shift1[dim])
+ {
+ izone->shift1[dim] = shift_diff;
+ }
+ }
+ }
+ }
+ }
+
+ if (dd->comm->eDLB != edlbNO)
+ {
+ snew(dd->comm->root,dd->ndim);
+ }
+
+ if (dd->comm->bRecordLoad)
+ {
+ make_load_communicators(dd);
+ }
+}
+
+static void make_pp_communicator(FILE *fplog,t_commrec *cr,int reorder)
+{
+ gmx_domdec_t *dd;
+ gmx_domdec_comm_t *comm;
+ int i,rank,*buf;
+ ivec periods;
+#ifdef GMX_MPI
+ MPI_Comm comm_cart;
+#endif
+
+ dd = cr->dd;
+ comm = dd->comm;
+
+#ifdef GMX_MPI
+ if (comm->bCartesianPP)
+ {
+ /* Set up cartesian communication for the particle-particle part */
+ if (fplog)
+ {
+ fprintf(fplog,"Will use a Cartesian communicator: %d x %d x %d\n",
+ dd->nc[XX],dd->nc[YY],dd->nc[ZZ]);
+ }
+
+ for(i=0; i<DIM; i++)
+ {
+ periods[i] = TRUE;
+ }
+ MPI_Cart_create(cr->mpi_comm_mygroup,DIM,dd->nc,periods,reorder,
+ &comm_cart);
+ /* We overwrite the old communicator with the new cartesian one */
+ cr->mpi_comm_mygroup = comm_cart;
+ }
+
+ dd->mpi_comm_all = cr->mpi_comm_mygroup;
+ MPI_Comm_rank(dd->mpi_comm_all,&dd->rank);
+
+ if (comm->bCartesianPP_PME)
+ {
+ /* Since we want to use the original cartesian setup for sim,
+ * and not the one after split, we need to make an index.
+ */
+ snew(comm->ddindex2ddnodeid,dd->nnodes);
+ comm->ddindex2ddnodeid[dd_index(dd->nc,dd->ci)] = dd->rank;
+ gmx_sumi(dd->nnodes,comm->ddindex2ddnodeid,cr);
+ /* Get the rank of the DD master,
+ * above we made sure that the master node is a PP node.
+ */
+ if (MASTER(cr))
+ {
+ rank = dd->rank;
+ }
+ else
+ {
+ rank = 0;
+ }
+ MPI_Allreduce(&rank,&dd->masterrank,1,MPI_INT,MPI_SUM,dd->mpi_comm_all);
+ }
+ else if (comm->bCartesianPP)
+ {
+ if (cr->npmenodes == 0)
+ {
+ /* The PP communicator is also
+ * the communicator for this simulation
+ */
+ cr->mpi_comm_mysim = cr->mpi_comm_mygroup;
+ }
+ cr->nodeid = dd->rank;
+
+ MPI_Cart_coords(dd->mpi_comm_all,dd->rank,DIM,dd->ci);
+
+ /* We need to make an index to go from the coordinates
+ * to the nodeid of this simulation.
+ */
+ snew(comm->ddindex2simnodeid,dd->nnodes);
+ snew(buf,dd->nnodes);
+ if (cr->duty & DUTY_PP)
+ {
+ buf[dd_index(dd->nc,dd->ci)] = cr->sim_nodeid;
+ }
+ /* Communicate the ddindex to simulation nodeid index */
+ MPI_Allreduce(buf,comm->ddindex2simnodeid,dd->nnodes,MPI_INT,MPI_SUM,
+ cr->mpi_comm_mysim);
+ sfree(buf);
+
+ /* Determine the master coordinates and rank.
+ * The DD master should be the same node as the master of this sim.
+ */
+ for(i=0; i<dd->nnodes; i++)
+ {
+ if (comm->ddindex2simnodeid[i] == 0)
+ {
+ ddindex2xyz(dd->nc,i,dd->master_ci);
+ MPI_Cart_rank(dd->mpi_comm_all,dd->master_ci,&dd->masterrank);
+ }
+ }
+ if (debug)
+ {
+ fprintf(debug,"The master rank is %d\n",dd->masterrank);
+ }
+ }
+ else
+ {
+ /* No Cartesian communicators */
+ /* We use the rank in dd->comm->all as DD index */
+ ddindex2xyz(dd->nc,dd->rank,dd->ci);
+ /* The simulation master nodeid is 0, so the DD master rank is also 0 */
+ dd->masterrank = 0;
+ clear_ivec(dd->master_ci);
+ }
+#endif
+
+ if (fplog)
+ {
+ fprintf(fplog,
+ "Domain decomposition nodeid %d, coordinates %d %d %d\n\n",
+ dd->rank,dd->ci[XX],dd->ci[YY],dd->ci[ZZ]);
+ }
+ if (debug)
+ {
+ fprintf(debug,
+ "Domain decomposition nodeid %d, coordinates %d %d %d\n\n",
+ dd->rank,dd->ci[XX],dd->ci[YY],dd->ci[ZZ]);
+ }
+}
+
+static void receive_ddindex2simnodeid(t_commrec *cr)
+{
+ gmx_domdec_t *dd;
+
+ gmx_domdec_comm_t *comm;
+ int *buf;
+
+ dd = cr->dd;
+ comm = dd->comm;
+
+#ifdef GMX_MPI
+ if (!comm->bCartesianPP_PME && comm->bCartesianPP)
+ {
+ snew(comm->ddindex2simnodeid,dd->nnodes);
+ snew(buf,dd->nnodes);
+ if (cr->duty & DUTY_PP)
+ {
+ buf[dd_index(dd->nc,dd->ci)] = cr->sim_nodeid;
+ }
+#ifdef GMX_MPI
+ /* Communicate the ddindex to simulation nodeid index */
+ MPI_Allreduce(buf,comm->ddindex2simnodeid,dd->nnodes,MPI_INT,MPI_SUM,
+ cr->mpi_comm_mysim);
+#endif
+ sfree(buf);
+ }
+#endif
+}
+
+static gmx_domdec_master_t *init_gmx_domdec_master_t(gmx_domdec_t *dd,
+ int ncg,int natoms)
+{
+ gmx_domdec_master_t *ma;
+ int i;
+
+ snew(ma,1);
+
+ snew(ma->ncg,dd->nnodes);
+ snew(ma->index,dd->nnodes+1);
+ snew(ma->cg,ncg);
+ snew(ma->nat,dd->nnodes);
+ snew(ma->ibuf,dd->nnodes*2);
+ snew(ma->cell_x,DIM);
+ for(i=0; i<DIM; i++)
+ {
+ snew(ma->cell_x[i],dd->nc[i]+1);
+ }
+
+ if (dd->nnodes <= GMX_DD_NNODES_SENDRECV)
+ {
+ ma->vbuf = NULL;
+ }
+ else
+ {
+ snew(ma->vbuf,natoms);
+ }
+
+ return ma;
+}
+
+static void split_communicator(FILE *fplog,t_commrec *cr,int dd_node_order,
+ int reorder)
+{
+ gmx_domdec_t *dd;
+ gmx_domdec_comm_t *comm;
+ int i,rank;
+ gmx_bool bDiv[DIM];
+ ivec periods;
+#ifdef GMX_MPI
+ MPI_Comm comm_cart;
+#endif
+
+ dd = cr->dd;
+ comm = dd->comm;
+
+ if (comm->bCartesianPP)
+ {
+ for(i=1; i<DIM; i++)
+ {
+ bDiv[i] = ((cr->npmenodes*dd->nc[i]) % (dd->nnodes) == 0);
+ }
+ if (bDiv[YY] || bDiv[ZZ])
+ {
+ comm->bCartesianPP_PME = TRUE;
+ /* If we have 2D PME decomposition, which is always in x+y,
+ * we stack the PME only nodes in z.
+ * Otherwise we choose the direction that provides the thinnest slab
+ * of PME only nodes as this will have the least effect
+ * on the PP communication.
+ * But for the PME communication the opposite might be better.
+ */
+ if (bDiv[ZZ] && (comm->npmenodes_y > 1 ||
+ !bDiv[YY] ||
+ dd->nc[YY] > dd->nc[ZZ]))
+ {
+ comm->cartpmedim = ZZ;
+ }
+ else
+ {
+ comm->cartpmedim = YY;
+ }
+ comm->ntot[comm->cartpmedim]
+ += (cr->npmenodes*dd->nc[comm->cartpmedim])/dd->nnodes;
+ }
+ else if (fplog)
+ {
+ fprintf(fplog,"#pmenodes (%d) is not a multiple of nx*ny (%d*%d) or nx*nz (%d*%d)\n",cr->npmenodes,dd->nc[XX],dd->nc[YY],dd->nc[XX],dd->nc[ZZ]);
+ fprintf(fplog,
+ "Will not use a Cartesian communicator for PP <-> PME\n\n");
+ }
+ }
+
+#ifdef GMX_MPI
+ if (comm->bCartesianPP_PME)
+ {
+ if (fplog)
+ {
+ fprintf(fplog,"Will use a Cartesian communicator for PP <-> PME: %d x %d x %d\n",comm->ntot[XX],comm->ntot[YY],comm->ntot[ZZ]);
+ }
+
+ for(i=0; i<DIM; i++)
+ {
+ periods[i] = TRUE;
+ }
+ MPI_Cart_create(cr->mpi_comm_mysim,DIM,comm->ntot,periods,reorder,
+ &comm_cart);
+
+ MPI_Comm_rank(comm_cart,&rank);
+ if (MASTERNODE(cr) && rank != 0)
+ {
+ gmx_fatal(FARGS,"MPI rank 0 was renumbered by MPI_Cart_create, we do not allow this");
+ }
+
+ /* With this assigment we loose the link to the original communicator
+ * which will usually be MPI_COMM_WORLD, unless have multisim.
+ */
+ cr->mpi_comm_mysim = comm_cart;
+ cr->sim_nodeid = rank;
+
+ MPI_Cart_coords(cr->mpi_comm_mysim,cr->sim_nodeid,DIM,dd->ci);
+
+ if (fplog)
+ {
+ fprintf(fplog,"Cartesian nodeid %d, coordinates %d %d %d\n\n",
+ cr->sim_nodeid,dd->ci[XX],dd->ci[YY],dd->ci[ZZ]);
+ }
+
+ if (dd->ci[comm->cartpmedim] < dd->nc[comm->cartpmedim])
+ {
+ cr->duty = DUTY_PP;
+ }
+ if (cr->npmenodes == 0 ||
+ dd->ci[comm->cartpmedim] >= dd->nc[comm->cartpmedim])
+ {
+ cr->duty = DUTY_PME;
+ }
+
+ /* Split the sim communicator into PP and PME only nodes */
+ MPI_Comm_split(cr->mpi_comm_mysim,
+ cr->duty,
+ dd_index(comm->ntot,dd->ci),
+ &cr->mpi_comm_mygroup);
+ }
+ else
+ {
+ switch (dd_node_order)
+ {
+ case ddnoPP_PME:
+ if (fplog)
+ {
+ fprintf(fplog,"Order of the nodes: PP first, PME last\n");
+ }
+ break;
+ case ddnoINTERLEAVE:
+ /* Interleave the PP-only and PME-only nodes,
+ * as on clusters with dual-core machines this will double
+ * the communication bandwidth of the PME processes
+ * and thus speed up the PP <-> PME and inter PME communication.
+ */
+ if (fplog)
+ {
+ fprintf(fplog,"Interleaving PP and PME nodes\n");
+ }
+ comm->pmenodes = dd_pmenodes(cr);
+ break;
+ case ddnoCARTESIAN:
+ break;
+ default:
+ gmx_fatal(FARGS,"Unknown dd_node_order=%d",dd_node_order);
+ }
+
+ if (dd_simnode2pmenode(cr,cr->sim_nodeid) == -1)
+ {
+ cr->duty = DUTY_PME;
+ }
+ else
+ {
+ cr->duty = DUTY_PP;
+ }
+
+ /* Split the sim communicator into PP and PME only nodes */
+ MPI_Comm_split(cr->mpi_comm_mysim,
+ cr->duty,
+ cr->nodeid,
+ &cr->mpi_comm_mygroup);
+ MPI_Comm_rank(cr->mpi_comm_mygroup,&cr->nodeid);
+ }
+#endif
+
+ if (fplog)
+ {
+ fprintf(fplog,"This is a %s only node\n\n",
+ (cr->duty & DUTY_PP) ? "particle-particle" : "PME-mesh");
+ }
+}
+
+void make_dd_communicators(FILE *fplog,t_commrec *cr,int dd_node_order)
+{
+ gmx_domdec_t *dd;
+ gmx_domdec_comm_t *comm;
+ int CartReorder;
+
+ dd = cr->dd;
+ comm = dd->comm;
+
+ copy_ivec(dd->nc,comm->ntot);
+
+ comm->bCartesianPP = (dd_node_order == ddnoCARTESIAN);
+ comm->bCartesianPP_PME = FALSE;
+
+ /* Reorder the nodes by default. This might change the MPI ranks.
+ * Real reordering is only supported on very few architectures,
+ * Blue Gene is one of them.
+ */
+ CartReorder = (getenv("GMX_NO_CART_REORDER") == NULL);
+
+ if (cr->npmenodes > 0)
+ {
+ /* Split the communicator into a PP and PME part */
+ split_communicator(fplog,cr,dd_node_order,CartReorder);
+ if (comm->bCartesianPP_PME)
+ {
+ /* We (possibly) reordered the nodes in split_communicator,
+ * so it is no longer required in make_pp_communicator.
+ */
+ CartReorder = FALSE;
+ }
+ }
+ else
+ {
+ /* All nodes do PP and PME */
+#ifdef GMX_MPI
+ /* We do not require separate communicators */
+ cr->mpi_comm_mygroup = cr->mpi_comm_mysim;
+#endif
+ }
+
+ if (cr->duty & DUTY_PP)
+ {
+ /* Copy or make a new PP communicator */
+ make_pp_communicator(fplog,cr,CartReorder);
+ }
+ else
+ {
+ receive_ddindex2simnodeid(cr);
+ }
+
+ if (!(cr->duty & DUTY_PME))
+ {
+ /* Set up the commnuication to our PME node */
+ dd->pme_nodeid = dd_simnode2pmenode(cr,cr->sim_nodeid);
+ dd->pme_receive_vir_ener = receive_vir_ener(cr);
+ if (debug)
+ {
+ fprintf(debug,"My pme_nodeid %d receive ener %d\n",
+ dd->pme_nodeid,dd->pme_receive_vir_ener);
+ }
+ }
+ else
+ {
+ dd->pme_nodeid = -1;
+ }
+
+ if (DDMASTER(dd))
+ {
+ dd->ma = init_gmx_domdec_master_t(dd,
+ comm->cgs_gl.nr,
+ comm->cgs_gl.index[comm->cgs_gl.nr]);
+ }
+}
+
+static real *get_slb_frac(FILE *fplog,const char *dir,int nc,const char *size_string)
+{
+ real *slb_frac,tot;
+ int i,n;
+ double dbl;
+
+ slb_frac = NULL;
+ if (nc > 1 && size_string != NULL)
+ {
+ if (fplog)
+ {
+ fprintf(fplog,"Using static load balancing for the %s direction\n",
+ dir);
+ }
+ snew(slb_frac,nc);
+ tot = 0;
+ for (i=0; i<nc; i++)
+ {
+ dbl = 0;
+ sscanf(size_string,"%lf%n",&dbl,&n);
+ if (dbl == 0)
+ {
+ gmx_fatal(FARGS,"Incorrect or not enough DD cell size entries for direction %s: '%s'",dir,size_string);
+ }
+ slb_frac[i] = dbl;
+ size_string += n;
+ tot += slb_frac[i];
+ }
+ /* Normalize */
+ if (fplog)
+ {
+ fprintf(fplog,"Relative cell sizes:");
+ }
+ for (i=0; i<nc; i++)
+ {
+ slb_frac[i] /= tot;
+ if (fplog)
+ {
+ fprintf(fplog," %5.3f",slb_frac[i]);
+ }
+ }
+ if (fplog)
+ {
+ fprintf(fplog,"\n");
+ }
+ }
+
+ return slb_frac;
+}
+
+static int multi_body_bondeds_count(gmx_mtop_t *mtop)
+{
+ int n,nmol,ftype;
+ gmx_mtop_ilistloop_t iloop;
+ t_ilist *il;
+
+ n = 0;
+ iloop = gmx_mtop_ilistloop_init(mtop);
+ while (gmx_mtop_ilistloop_next(iloop,&il,&nmol))
+ {
+ for(ftype=0; ftype<F_NRE; ftype++)
+ {
+ if ((interaction_function[ftype].flags & IF_BOND) &&
+ NRAL(ftype) > 2)
+ {
+ n += nmol*il[ftype].nr/(1 + NRAL(ftype));
+ }
+ }
+ }
+
+ return n;
+}
+
+static int dd_nst_env(FILE *fplog,const char *env_var,int def)
+{
+ char *val;
+ int nst;
+
+ nst = def;
+ val = getenv(env_var);
+ if (val)
+ {
+ if (sscanf(val,"%d",&nst) <= 0)
+ {
+ nst = 1;
+ }
+ if (fplog)
+ {
+ fprintf(fplog,"Found env.var. %s = %s, using value %d\n",
+ env_var,val,nst);
+ }
+ }
+
+ return nst;
+}
+
+static void dd_warning(t_commrec *cr,FILE *fplog,const char *warn_string)
+{
+ if (MASTER(cr))
+ {
+ fprintf(stderr,"\n%s\n",warn_string);
+ }
+ if (fplog)
+ {
+ fprintf(fplog,"\n%s\n",warn_string);
+ }
+}
+
+static void check_dd_restrictions(t_commrec *cr,gmx_domdec_t *dd,
+ t_inputrec *ir,FILE *fplog)
+{
+ if (ir->ePBC == epbcSCREW &&
+ (dd->nc[XX] == 1 || dd->nc[YY] > 1 || dd->nc[ZZ] > 1))
+ {
+ gmx_fatal(FARGS,"With pbc=%s can only do domain decomposition in the x-direction",epbc_names[ir->ePBC]);
+ }
+
+ if (ir->ns_type == ensSIMPLE)
+ {
+ gmx_fatal(FARGS,"Domain decomposition does not support simple neighbor searching, use grid searching or use particle decomposition");
+ }
+
+ if (ir->nstlist == 0)
+ {
+ gmx_fatal(FARGS,"Domain decomposition does not work with nstlist=0");
+ }
+
+ if (ir->comm_mode == ecmANGULAR && ir->ePBC != epbcNONE)
+ {
+ dd_warning(cr,fplog,"comm-mode angular will give incorrect results when the comm group partially crosses a periodic boundary");
+ }
+}
+
+static real average_cellsize_min(gmx_domdec_t *dd,gmx_ddbox_t *ddbox)
+{
+ int di,d;
+ real r;
+
+ r = ddbox->box_size[XX];
+ for(di=0; di<dd->ndim; di++)
+ {
+ d = dd->dim[di];
+ /* Check using the initial average cell size */
+ r = min(r,ddbox->box_size[d]*ddbox->skew_fac[d]/dd->nc[d]);
+ }
+
+ return r;
+}
+
+static int check_dlb_support(FILE *fplog,t_commrec *cr,
+ const char *dlb_opt,gmx_bool bRecordLoad,
+ unsigned long Flags,t_inputrec *ir)
+{
+ gmx_domdec_t *dd;
+ int eDLB=-1;
+ char buf[STRLEN];
+
+ switch (dlb_opt[0])
+ {
+ case 'a': eDLB = edlbAUTO; break;
+ case 'n': eDLB = edlbNO; break;
+ case 'y': eDLB = edlbYES; break;
+ default: gmx_incons("Unknown dlb_opt");
+ }
+
+ if (Flags & MD_RERUN)
+ {
+ return edlbNO;
+ }
+
+ if (!EI_DYNAMICS(ir->eI))
+ {
+ if (eDLB == edlbYES)
+ {
+ sprintf(buf,"NOTE: dynamic load balancing is only supported with dynamics, not with integrator '%s'\n",EI(ir->eI));
+ dd_warning(cr,fplog,buf);
+ }
+
+ return edlbNO;
+ }
+
+ if (!bRecordLoad)
+ {
+ dd_warning(cr,fplog,"NOTE: Cycle counting is not supported on this architecture, will not use dynamic load balancing\n");
+
+ return edlbNO;
+ }
+
+ if (Flags & MD_REPRODUCIBLE)
+ {
+ switch (eDLB)
+ {
+ case edlbNO:
+ break;
+ case edlbAUTO:
+ dd_warning(cr,fplog,"NOTE: reproducibility requested, will not use dynamic load balancing\n");
+ eDLB = edlbNO;
+ break;
+ case edlbYES:
+ dd_warning(cr,fplog,"WARNING: reproducibility requested with dynamic load balancing, the simulation will NOT be binary reproducible\n");
+ break;
+ default:
+ gmx_fatal(FARGS,"Death horror: undefined case (%d) for load balancing choice",eDLB);
+ break;
+ }
+ }
+
+ return eDLB;
+}
+
+static void set_dd_dim(FILE *fplog,gmx_domdec_t *dd)
+{
+ int dim;
+
+ dd->ndim = 0;
+ if (getenv("GMX_DD_ORDER_ZYX") != NULL)
+ {
+ /* Decomposition order z,y,x */
+ if (fplog)
+ {
+ fprintf(fplog,"Using domain decomposition order z, y, x\n");
+ }
+ for(dim=DIM-1; dim>=0; dim--)
+ {
+ if (dd->nc[dim] > 1)
+ {
+ dd->dim[dd->ndim++] = dim;
+ }
+ }
+ }
+ else
+ {
+ /* Decomposition order x,y,z */
+ for(dim=0; dim<DIM; dim++)
+ {
+ if (dd->nc[dim] > 1)
+ {
+ dd->dim[dd->ndim++] = dim;
+ }
+ }
+ }
+}
+
+static gmx_domdec_comm_t *init_dd_comm()
+{
+ gmx_domdec_comm_t *comm;
+ int i;
+
+ snew(comm,1);
+ snew(comm->cggl_flag,DIM*2);
+ snew(comm->cgcm_state,DIM*2);
+ for(i=0; i<DIM*2; i++)
+ {
+ comm->cggl_flag_nalloc[i] = 0;
+ comm->cgcm_state_nalloc[i] = 0;
+ }
+
+ comm->nalloc_int = 0;
+ comm->buf_int = NULL;
+
+ vec_rvec_init(&comm->vbuf);
+
+ comm->n_load_have = 0;
+ comm->n_load_collect = 0;
+
+ for(i=0; i<ddnatNR-ddnatZONE; i++)
+ {
+ comm->sum_nat[i] = 0;
+ }
+ comm->ndecomp = 0;
+ comm->nload = 0;
+ comm->load_step = 0;
+ comm->load_sum = 0;
+ comm->load_max = 0;
+ clear_ivec(comm->load_lim);
+ comm->load_mdf = 0;
+ comm->load_pme = 0;
+
+ return comm;
+}
+
+gmx_domdec_t *init_domain_decomposition(FILE *fplog,t_commrec *cr,
+ unsigned long Flags,
+ ivec nc,
+ real comm_distance_min,real rconstr,
+ const char *dlb_opt,real dlb_scale,
+ const char *sizex,const char *sizey,const char *sizez,
+ gmx_mtop_t *mtop,t_inputrec *ir,
+ matrix box,rvec *x,
+ gmx_ddbox_t *ddbox,
+ int *npme_x,int *npme_y)
+{
+ gmx_domdec_t *dd;
+ gmx_domdec_comm_t *comm;
+ int recload;
+ int d,i,j;
+ real r_2b,r_mb,r_bonded=-1,r_bonded_limit=-1,limit,acs;
+ gmx_bool bC;
+ char buf[STRLEN];
+
+ if (fplog)
+ {
+ fprintf(fplog,
+ "\nInitializing Domain Decomposition on %d nodes\n",cr->nnodes);
+ }
+
+ snew(dd,1);
+
+ dd->comm = init_dd_comm();
+ comm = dd->comm;
+ snew(comm->cggl_flag,DIM*2);
+ snew(comm->cgcm_state,DIM*2);
+
+ dd->npbcdim = ePBC2npbcdim(ir->ePBC);
+ dd->bScrewPBC = (ir->ePBC == epbcSCREW);
+
+ dd->bSendRecv2 = dd_nst_env(fplog,"GMX_DD_SENDRECV2",0);
+ comm->dlb_scale_lim = dd_nst_env(fplog,"GMX_DLB_MAX",10);
+ comm->eFlop = dd_nst_env(fplog,"GMX_DLB_FLOP",0);
+ recload = dd_nst_env(fplog,"GMX_DD_LOAD",1);
+ comm->nstSortCG = dd_nst_env(fplog,"GMX_DD_SORT",1);
+ comm->nstDDDump = dd_nst_env(fplog,"GMX_DD_DUMP",0);
+ comm->nstDDDumpGrid = dd_nst_env(fplog,"GMX_DD_DUMP_GRID",0);
+ comm->DD_debug = dd_nst_env(fplog,"GMX_DD_DEBUG",0);
+
+ dd->pme_recv_f_alloc = 0;
+ dd->pme_recv_f_buf = NULL;
+
+ if (dd->bSendRecv2 && fplog)
+ {
+ fprintf(fplog,"Will use two sequential MPI_Sendrecv calls instead of two simultaneous non-blocking MPI_Irecv and MPI_Isend pairs for constraint and vsite communication\n");
+ }
+ if (comm->eFlop)
+ {
+ if (fplog)
+ {
+ fprintf(fplog,"Will load balance based on FLOP count\n");
+ }
+ if (comm->eFlop > 1)
+ {
+ srand(1+cr->nodeid);
+ }
+ comm->bRecordLoad = TRUE;
+ }
+ else
+ {
+ comm->bRecordLoad = (wallcycle_have_counter() && recload > 0);
+
+ }
+
+ comm->eDLB = check_dlb_support(fplog,cr,dlb_opt,comm->bRecordLoad,Flags,ir);
+
+ comm->bDynLoadBal = (comm->eDLB == edlbYES);
+ if (fplog)
+ {
+ fprintf(fplog,"Dynamic load balancing: %s\n",edlb_names[comm->eDLB]);
+ }
+ dd->bGridJump = comm->bDynLoadBal;
+
+ if (comm->nstSortCG)
+ {
+ if (fplog)
+ {
+ if (comm->nstSortCG == 1)
+ {
+ fprintf(fplog,"Will sort the charge groups at every domain (re)decomposition\n");
+ }
+ else
+ {
+ fprintf(fplog,"Will sort the charge groups every %d steps\n",
+ comm->nstSortCG);
+ }
+ }
+ snew(comm->sort,1);
+ }
+ else
+ {
+ if (fplog)
+ {
+ fprintf(fplog,"Will not sort the charge groups\n");
+ }
+ }
+
+ comm->bInterCGBondeds = (ncg_mtop(mtop) > mtop->mols.nr);
+ if (comm->bInterCGBondeds)
+ {
+ comm->bInterCGMultiBody = (multi_body_bondeds_count(mtop) > 0);
+ }
+ else
+ {
+ comm->bInterCGMultiBody = FALSE;
+ }
+
+ dd->bInterCGcons = inter_charge_group_constraints(mtop);
+
+ if (ir->rlistlong == 0)
+ {
+ /* Set the cut-off to some very large value,
+ * so we don't need if statements everywhere in the code.
+ * We use sqrt, since the cut-off is squared in some places.
+ */
+ comm->cutoff = GMX_CUTOFF_INF;
+ }
+ else
+ {
+ comm->cutoff = ir->rlistlong;
+ }
+ comm->cutoff_mbody = 0;
+
+ comm->cellsize_limit = 0;
+ comm->bBondComm = FALSE;
+
+ if (comm->bInterCGBondeds)
+ {
+ if (comm_distance_min > 0)
+ {
+ comm->cutoff_mbody = comm_distance_min;
+ if (Flags & MD_DDBONDCOMM)
+ {
+ comm->bBondComm = (comm->cutoff_mbody > comm->cutoff);
+ }
+ else
+ {
+ comm->cutoff = max(comm->cutoff,comm->cutoff_mbody);
+ }
+ r_bonded_limit = comm->cutoff_mbody;
+ }
+ else if (ir->bPeriodicMols)
+ {
+ /* Can not easily determine the required cut-off */
+ dd_warning(cr,fplog,"NOTE: Periodic molecules: can not easily determine the required minimum bonded cut-off, using half the non-bonded cut-off\n");
+ comm->cutoff_mbody = comm->cutoff/2;
+ r_bonded_limit = comm->cutoff_mbody;
+ }
+ else
+ {
+ if (MASTER(cr))
+ {
+ dd_bonded_cg_distance(fplog,dd,mtop,ir,x,box,
+ Flags & MD_DDBONDCHECK,&r_2b,&r_mb);
+ }
+ gmx_bcast(sizeof(r_2b),&r_2b,cr);
+ gmx_bcast(sizeof(r_mb),&r_mb,cr);
+
+ /* We use an initial margin of 10% for the minimum cell size,
+ * except when we are just below the non-bonded cut-off.
+ */
+ if (Flags & MD_DDBONDCOMM)
+ {
+ if (max(r_2b,r_mb) > comm->cutoff)
+ {
+ r_bonded = max(r_2b,r_mb);
+ r_bonded_limit = 1.1*r_bonded;
+ comm->bBondComm = TRUE;
+ }
+ else
+ {
+ r_bonded = r_mb;
+ r_bonded_limit = min(1.1*r_bonded,comm->cutoff);
+ }
+ /* We determine cutoff_mbody later */
+ }
+ else
+ {
+ /* No special bonded communication,
+ * simply increase the DD cut-off.
+ */
+ r_bonded_limit = 1.1*max(r_2b,r_mb);
+ comm->cutoff_mbody = r_bonded_limit;
+ comm->cutoff = max(comm->cutoff,comm->cutoff_mbody);
+ }
+ }
+ comm->cellsize_limit = max(comm->cellsize_limit,r_bonded_limit);
+ if (fplog)
+ {
+ fprintf(fplog,
+ "Minimum cell size due to bonded interactions: %.3f nm\n",
+ comm->cellsize_limit);
+ }
+ }
+
+ if (dd->bInterCGcons && rconstr <= 0)
+ {
+ /* There is a cell size limit due to the constraints (P-LINCS) */
+ rconstr = constr_r_max(fplog,mtop,ir);
+ if (fplog)
+ {
+ fprintf(fplog,
+ "Estimated maximum distance required for P-LINCS: %.3f nm\n",
+ rconstr);
+ if (rconstr > comm->cellsize_limit)
+ {
+ fprintf(fplog,"This distance will limit the DD cell size, you can override this with -rcon\n");
+ }
+ }
+ }
+ else if (rconstr > 0 && fplog)
+ {
+ /* Here we do not check for dd->bInterCGcons,
+ * because one can also set a cell size limit for virtual sites only
+ * and at this point we don't know yet if there are intercg v-sites.
+ */
+ fprintf(fplog,
+ "User supplied maximum distance required for P-LINCS: %.3f nm\n",
+ rconstr);
+ }
+ comm->cellsize_limit = max(comm->cellsize_limit,rconstr);
+
+ comm->cgs_gl = gmx_mtop_global_cgs(mtop);
+
+ if (nc[XX] > 0)
+ {
+ copy_ivec(nc,dd->nc);
+ set_dd_dim(fplog,dd);
+ set_ddbox_cr(cr,&dd->nc,ir,box,&comm->cgs_gl,x,ddbox);
+
+ if (cr->npmenodes == -1)
+ {
+ cr->npmenodes = 0;
+ }
+ acs = average_cellsize_min(dd,ddbox);
+ if (acs < comm->cellsize_limit)
+ {
+ if (fplog)
+ {
+ fprintf(fplog,"ERROR: The initial cell size (%f) is smaller than the cell size limit (%f)\n",acs,comm->cellsize_limit);
+ }
+ gmx_fatal_collective(FARGS,cr,NULL,
+ "The initial cell size (%f) is smaller than the cell size limit (%f), change options -dd, -rdd or -rcon, see the log file for details",
+ acs,comm->cellsize_limit);
+ }
+ }
+ else
+ {
+ set_ddbox_cr(cr,NULL,ir,box,&comm->cgs_gl,x,ddbox);
+
+ /* We need to choose the optimal DD grid and possibly PME nodes */
+ limit = dd_choose_grid(fplog,cr,dd,ir,mtop,box,ddbox,
+ comm->eDLB!=edlbNO,dlb_scale,
+ comm->cellsize_limit,comm->cutoff,
+ comm->bInterCGBondeds,comm->bInterCGMultiBody);
+
+ if (dd->nc[XX] == 0)
+ {
+ bC = (dd->bInterCGcons && rconstr > r_bonded_limit);
+ sprintf(buf,"Change the number of nodes or mdrun option %s%s%s",
+ !bC ? "-rdd" : "-rcon",
+ comm->eDLB!=edlbNO ? " or -dds" : "",
+ bC ? " or your LINCS settings" : "");
+
+ gmx_fatal_collective(FARGS,cr,NULL,
+ "There is no domain decomposition for %d nodes that is compatible with the given box and a minimum cell size of %g nm\n"
+ "%s\n"
+ "Look in the log file for details on the domain decomposition",
+ cr->nnodes-cr->npmenodes,limit,buf);
+ }
+ set_dd_dim(fplog,dd);
+ }
+
+ if (fplog)
+ {
+ fprintf(fplog,
+ "Domain decomposition grid %d x %d x %d, separate PME nodes %d\n",
+ dd->nc[XX],dd->nc[YY],dd->nc[ZZ],cr->npmenodes);
+ }
+
+ dd->nnodes = dd->nc[XX]*dd->nc[YY]*dd->nc[ZZ];
+ if (cr->nnodes - dd->nnodes != cr->npmenodes)
+ {
+ gmx_fatal_collective(FARGS,cr,NULL,
+ "The size of the domain decomposition grid (%d) does not match the number of nodes (%d). The total number of nodes is %d",
+ dd->nnodes,cr->nnodes - cr->npmenodes,cr->nnodes);
+ }
+ if (cr->npmenodes > dd->nnodes)
+ {
+ gmx_fatal_collective(FARGS,cr,NULL,
+ "The number of separate PME nodes (%d) is larger than the number of PP nodes (%d), this is not supported.",cr->npmenodes,dd->nnodes);
+ }
+ if (cr->npmenodes > 0)
+ {
+ comm->npmenodes = cr->npmenodes;
+ }
+ else
+ {
+ comm->npmenodes = dd->nnodes;
+ }
+
+ if (EEL_PME(ir->coulombtype))
+ {
+ /* The following choices should match those
+ * in comm_cost_est in domdec_setup.c.
+ * Note that here the checks have to take into account
+ * that the decomposition might occur in a different order than xyz
+ * (for instance through the env.var. GMX_DD_ORDER_ZYX),
+ * in which case they will not match those in comm_cost_est,
+ * but since that is mainly for testing purposes that's fine.
+ */
+ if (dd->ndim >= 2 && dd->dim[0] == XX && dd->dim[1] == YY &&
+ comm->npmenodes > dd->nc[XX] && comm->npmenodes % dd->nc[XX] == 0 &&
+ getenv("GMX_PMEONEDD") == NULL)
+ {
+ comm->npmedecompdim = 2;
+ comm->npmenodes_x = dd->nc[XX];
+ comm->npmenodes_y = comm->npmenodes/comm->npmenodes_x;
+ }
+ else
+ {
+ /* In case nc is 1 in both x and y we could still choose to
+ * decompose pme in y instead of x, but we use x for simplicity.
+ */
+ comm->npmedecompdim = 1;
+ if (dd->dim[0] == YY)
+ {
+ comm->npmenodes_x = 1;
+ comm->npmenodes_y = comm->npmenodes;
+ }
+ else
+ {
+ comm->npmenodes_x = comm->npmenodes;
+ comm->npmenodes_y = 1;
+ }
+ }
+ if (fplog)
+ {
+ fprintf(fplog,"PME domain decomposition: %d x %d x %d\n",
+ comm->npmenodes_x,comm->npmenodes_y,1);
+ }
+ }
+ else
+ {
+ comm->npmedecompdim = 0;
+ comm->npmenodes_x = 0;
+ comm->npmenodes_y = 0;
+ }
+
+ /* Technically we don't need both of these,
+ * but it simplifies code not having to recalculate it.
+ */
+ *npme_x = comm->npmenodes_x;
+ *npme_y = comm->npmenodes_y;
+
+ snew(comm->slb_frac,DIM);
+ if (comm->eDLB == edlbNO)
+ {
+ comm->slb_frac[XX] = get_slb_frac(fplog,"x",dd->nc[XX],sizex);
+ comm->slb_frac[YY] = get_slb_frac(fplog,"y",dd->nc[YY],sizey);
+ comm->slb_frac[ZZ] = get_slb_frac(fplog,"z",dd->nc[ZZ],sizez);
+ }
+
+ if (comm->bInterCGBondeds && comm->cutoff_mbody == 0)
+ {
+ if (comm->bBondComm || comm->eDLB != edlbNO)
+ {
+ /* Set the bonded communication distance to halfway
+ * the minimum and the maximum,
+ * since the extra communication cost is nearly zero.
+ */
+ acs = average_cellsize_min(dd,ddbox);
+ comm->cutoff_mbody = 0.5*(r_bonded + acs);
+ if (comm->eDLB != edlbNO)
+ {
+ /* Check if this does not limit the scaling */
+ comm->cutoff_mbody = min(comm->cutoff_mbody,dlb_scale*acs);
+ }
+ if (!comm->bBondComm)
+ {
+ /* Without bBondComm do not go beyond the n.b. cut-off */
+ comm->cutoff_mbody = min(comm->cutoff_mbody,comm->cutoff);
+ if (comm->cellsize_limit >= comm->cutoff)
+ {
+ /* We don't loose a lot of efficieny
+ * when increasing it to the n.b. cut-off.
+ * It can even be slightly faster, because we need
+ * less checks for the communication setup.
+ */
+ comm->cutoff_mbody = comm->cutoff;
+ }
+ }
+ /* Check if we did not end up below our original limit */
+ comm->cutoff_mbody = max(comm->cutoff_mbody,r_bonded_limit);
+
+ if (comm->cutoff_mbody > comm->cellsize_limit)
+ {
+ comm->cellsize_limit = comm->cutoff_mbody;
+ }
+ }
+ /* Without DLB and cutoff_mbody<cutoff, cutoff_mbody is dynamic */
+ }
+
+ if (debug)
+ {
+ fprintf(debug,"Bonded atom communication beyond the cut-off: %d\n"
+ "cellsize limit %f\n",
+ comm->bBondComm,comm->cellsize_limit);
+ }
+
+ if (MASTER(cr))
+ {
+ check_dd_restrictions(cr,dd,ir,fplog);
+ }
+
+ comm->globalcomm_step = INT_MIN;
+ dd->ddp_count = 0;
+
+ clear_dd_cycle_counts(dd);
+
+ return dd;
+}
+
+static void set_dlb_limits(gmx_domdec_t *dd)
+
+{
+ int d;
+
+ for(d=0; d<dd->ndim; d++)
+ {
+ dd->comm->cd[d].np = dd->comm->cd[d].np_dlb;
+ dd->comm->cellsize_min[dd->dim[d]] =
+ dd->comm->cellsize_min_dlb[dd->dim[d]];
+ }
+}
+
+
+static void turn_on_dlb(FILE *fplog,t_commrec *cr,gmx_large_int_t step)
+{
+ gmx_domdec_t *dd;
+ gmx_domdec_comm_t *comm;
+ real cellsize_min;
+ int d,nc,i;
+ char buf[STRLEN];
+
+ dd = cr->dd;
+ comm = dd->comm;
+
+ if (fplog)
+ {
+ fprintf(fplog,"At step %s the performance loss due to force load imbalance is %.1f %%\n",gmx_step_str(step,buf),dd_force_imb_perf_loss(dd)*100);
+ }
+
+ cellsize_min = comm->cellsize_min[dd->dim[0]];
+ for(d=1; d<dd->ndim; d++)
+ {
+ cellsize_min = min(cellsize_min,comm->cellsize_min[dd->dim[d]]);
+ }
+
+ if (cellsize_min < comm->cellsize_limit*1.05)
+ {
+ dd_warning(cr,fplog,"NOTE: the minimum cell size is smaller than 1.05 times the cell size limit, will not turn on dynamic load balancing\n");
+
+ /* Change DLB from "auto" to "no". */
+ comm->eDLB = edlbNO;
+
+ return;
+ }
+
+ dd_warning(cr,fplog,"NOTE: Turning on dynamic load balancing\n");
+ comm->bDynLoadBal = TRUE;
+ dd->bGridJump = TRUE;
+
+ set_dlb_limits(dd);
+
+ /* We can set the required cell size info here,
+ * so we do not need to communicate this.
+ * The grid is completely uniform.
+ */
+ for(d=0; d<dd->ndim; d++)
+ {
+ if (comm->root[d])
+ {
+ comm->load[d].sum_m = comm->load[d].sum;
+
+ nc = dd->nc[dd->dim[d]];
+ for(i=0; i<nc; i++)
+ {
+ comm->root[d]->cell_f[i] = i/(real)nc;
+ if (d > 0)
+ {
+ comm->root[d]->cell_f_max0[i] = i /(real)nc;
+ comm->root[d]->cell_f_min1[i] = (i+1)/(real)nc;
+ }
+ }
+ comm->root[d]->cell_f[nc] = 1.0;
+ }
+ }
+}
+
+static char *init_bLocalCG(gmx_mtop_t *mtop)
+{
+ int ncg,cg;
+ char *bLocalCG;
+
+ ncg = ncg_mtop(mtop);
+ snew(bLocalCG,ncg);
+ for(cg=0; cg<ncg; cg++)
+ {
+ bLocalCG[cg] = FALSE;
+ }
+
+ return bLocalCG;
+}
+
+void dd_init_bondeds(FILE *fplog,
+ gmx_domdec_t *dd,gmx_mtop_t *mtop,
+ gmx_vsite_t *vsite,gmx_constr_t constr,
+ t_inputrec *ir,gmx_bool bBCheck,cginfo_mb_t *cginfo_mb)
+{
+ gmx_domdec_comm_t *comm;
+ gmx_bool bBondComm;
+ int d;
+
+ dd_make_reverse_top(fplog,dd,mtop,vsite,constr,ir,bBCheck);
+
+ comm = dd->comm;
+
+ if (comm->bBondComm)
+ {
+ /* Communicate atoms beyond the cut-off for bonded interactions */
+ comm = dd->comm;
+
+ comm->cglink = make_charge_group_links(mtop,dd,cginfo_mb);
+
+ comm->bLocalCG = init_bLocalCG(mtop);
+ }
+ else
+ {
+ /* Only communicate atoms based on cut-off */
+ comm->cglink = NULL;
+ comm->bLocalCG = NULL;
+ }
+}
+
+static void print_dd_settings(FILE *fplog,gmx_domdec_t *dd,
+ t_inputrec *ir,
+ gmx_bool bDynLoadBal,real dlb_scale,
+ gmx_ddbox_t *ddbox)
+{
+ gmx_domdec_comm_t *comm;
+ int d;
+ ivec np;
+ real limit,shrink;
+ char buf[64];
+
+ if (fplog == NULL)
+ {
+ return;
+ }
+
+ comm = dd->comm;
+
+ if (bDynLoadBal)
+ {
+ fprintf(fplog,"The maximum number of communication pulses is:");
+ for(d=0; d<dd->ndim; d++)
+ {
+ fprintf(fplog," %c %d",dim2char(dd->dim[d]),comm->cd[d].np_dlb);
+ }
+ fprintf(fplog,"\n");
+ fprintf(fplog,"The minimum size for domain decomposition cells is %.3f nm\n",comm->cellsize_limit);
+ fprintf(fplog,"The requested allowed shrink of DD cells (option -dds) is: %.2f\n",dlb_scale);
+ fprintf(fplog,"The allowed shrink of domain decomposition cells is:");
+ for(d=0; d<DIM; d++)
+ {
+ if (dd->nc[d] > 1)
+ {
+ if (d >= ddbox->npbcdim && dd->nc[d] == 2)
+ {
+ shrink = 0;
+ }
+ else
+ {
+ shrink =
+ comm->cellsize_min_dlb[d]/
+ (ddbox->box_size[d]*ddbox->skew_fac[d]/dd->nc[d]);
+ }
+ fprintf(fplog," %c %.2f",dim2char(d),shrink);
+ }
+ }
+ fprintf(fplog,"\n");
+ }
+ else
+ {
+ set_dd_cell_sizes_slb(dd,ddbox,FALSE,np);
+ fprintf(fplog,"The initial number of communication pulses is:");
+ for(d=0; d<dd->ndim; d++)
+ {
+ fprintf(fplog," %c %d",dim2char(dd->dim[d]),np[dd->dim[d]]);
+ }
+ fprintf(fplog,"\n");
+ fprintf(fplog,"The initial domain decomposition cell size is:");
+ for(d=0; d<DIM; d++) {
+ if (dd->nc[d] > 1)
+ {
+ fprintf(fplog," %c %.2f nm",
+ dim2char(d),dd->comm->cellsize_min[d]);
+ }
+ }
+ fprintf(fplog,"\n\n");
+ }
+
+ if (comm->bInterCGBondeds || dd->vsite_comm || dd->constraint_comm)
+ {
+ fprintf(fplog,"The maximum allowed distance for charge groups involved in interactions is:\n");
+ fprintf(fplog,"%40s %-7s %6.3f nm\n",
+ "non-bonded interactions","",comm->cutoff);
+
+ if (bDynLoadBal)
+ {
+ limit = dd->comm->cellsize_limit;
+ }
+ else
+ {
+ if (dynamic_dd_box(ddbox,ir))
+ {
+ fprintf(fplog,"(the following are initial values, they could change due to box deformation)\n");
+ }
+ limit = dd->comm->cellsize_min[XX];
+ for(d=1; d<DIM; d++)
+ {
+ limit = min(limit,dd->comm->cellsize_min[d]);
+ }
+ }
+
+ if (comm->bInterCGBondeds)
+ {
+ fprintf(fplog,"%40s %-7s %6.3f nm\n",
+ "two-body bonded interactions","(-rdd)",
+ max(comm->cutoff,comm->cutoff_mbody));
+ fprintf(fplog,"%40s %-7s %6.3f nm\n",
+ "multi-body bonded interactions","(-rdd)",
+ (comm->bBondComm || dd->bGridJump) ? comm->cutoff_mbody : min(comm->cutoff,limit));
+ }
+ if (dd->vsite_comm)
+ {
+ fprintf(fplog,"%40s %-7s %6.3f nm\n",
+ "virtual site constructions","(-rcon)",limit);
+ }
+ if (dd->constraint_comm)
+ {
+ sprintf(buf,"atoms separated by up to %d constraints",
+ 1+ir->nProjOrder);
+ fprintf(fplog,"%40s %-7s %6.3f nm\n",
+ buf,"(-rcon)",limit);
+ }
+ fprintf(fplog,"\n");
+ }
+
+ fflush(fplog);
+}
+
+void set_dd_parameters(FILE *fplog,gmx_domdec_t *dd,real dlb_scale,
+ t_inputrec *ir,t_forcerec *fr,
+ gmx_ddbox_t *ddbox)
+{
+ gmx_domdec_comm_t *comm;
+ int d,dim,npulse,npulse_d_max,npulse_d;
+ gmx_bool bNoCutOff;
+ int natoms_tot;
+ real vol_frac;
+
+ comm = dd->comm;
+
+ bNoCutOff = (ir->rvdw == 0 || ir->rcoulomb == 0);
+
+ if (EEL_PME(ir->coulombtype))
+ {
+ init_ddpme(dd,&comm->ddpme[0],0);
+ if (comm->npmedecompdim >= 2)
+ {
+ init_ddpme(dd,&comm->ddpme[1],1);
+ }
+ }
+ else
+ {
+ comm->npmenodes = 0;
+ if (dd->pme_nodeid >= 0)
+ {
+ gmx_fatal_collective(FARGS,NULL,dd,
+ "Can not have separate PME nodes without PME electrostatics");
+ }
+ }
+
+ /* If each molecule is a single charge group
+ * or we use domain decomposition for each periodic dimension,
+ * we do not need to take pbc into account for the bonded interactions.
+ */
+ if (fr->ePBC == epbcNONE || !comm->bInterCGBondeds ||
+ (dd->nc[XX]>1 && dd->nc[YY]>1 && (dd->nc[ZZ]>1 || fr->ePBC==epbcXY)))
+ {
+ fr->bMolPBC = FALSE;
+ }
+ else
+ {
+ fr->bMolPBC = TRUE;
+ }
+
+ if (debug)
+ {
+ fprintf(debug,"The DD cut-off is %f\n",comm->cutoff);
+ }
+ if (comm->eDLB != edlbNO)
+ {
+ /* Determine the maximum number of comm. pulses in one dimension */
+
+ comm->cellsize_limit = max(comm->cellsize_limit,comm->cutoff_mbody);
+
+ /* Determine the maximum required number of grid pulses */
+ if (comm->cellsize_limit >= comm->cutoff)
+ {
+ /* Only a single pulse is required */
+ npulse = 1;
+ }
+ else if (!bNoCutOff && comm->cellsize_limit > 0)
+ {
+ /* We round down slightly here to avoid overhead due to the latency
+ * of extra communication calls when the cut-off
+ * would be only slightly longer than the cell size.
+ * Later cellsize_limit is redetermined,
+ * so we can not miss interactions due to this rounding.
+ */
+ npulse = (int)(0.96 + comm->cutoff/comm->cellsize_limit);
+ }
+ else
+ {
+ /* There is no cell size limit */
+ npulse = max(dd->nc[XX]-1,max(dd->nc[YY]-1,dd->nc[ZZ]-1));
+ }
+
+ if (!bNoCutOff && npulse > 1)
+ {
+ /* See if we can do with less pulses, based on dlb_scale */
+ npulse_d_max = 0;
+ for(d=0; d<dd->ndim; d++)
+ {
+ dim = dd->dim[d];
+ npulse_d = (int)(1 + dd->nc[dim]*comm->cutoff
+ /(ddbox->box_size[dim]*ddbox->skew_fac[dim]*dlb_scale));
+ npulse_d_max = max(npulse_d_max,npulse_d);
+ }
+ npulse = min(npulse,npulse_d_max);
+ }
+
+ /* This env var can override npulse */
+ d = dd_nst_env(fplog,"GMX_DD_NPULSE",0);
+ if (d > 0)
+ {
+ npulse = d;
+ }
+
+ comm->maxpulse = 1;
+ comm->bVacDLBNoLimit = (ir->ePBC == epbcNONE);
+ for(d=0; d<dd->ndim; d++)
+ {
+ comm->cd[d].np_dlb = min(npulse,dd->nc[dd->dim[d]]-1);
+ comm->cd[d].np_nalloc = comm->cd[d].np_dlb;
+ snew(comm->cd[d].ind,comm->cd[d].np_nalloc);
+ comm->maxpulse = max(comm->maxpulse,comm->cd[d].np_dlb);
+ if (comm->cd[d].np_dlb < dd->nc[dd->dim[d]]-1)
+ {
+ comm->bVacDLBNoLimit = FALSE;
+ }
+ }
+
+ /* cellsize_limit is set for LINCS in init_domain_decomposition */
+ if (!comm->bVacDLBNoLimit)
+ {
+ comm->cellsize_limit = max(comm->cellsize_limit,
+ comm->cutoff/comm->maxpulse);
+ }
+ comm->cellsize_limit = max(comm->cellsize_limit,comm->cutoff_mbody);
+ /* Set the minimum cell size for each DD dimension */
+ for(d=0; d<dd->ndim; d++)
+ {
+ if (comm->bVacDLBNoLimit ||
+ comm->cd[d].np_dlb*comm->cellsize_limit >= comm->cutoff)
+ {
+ comm->cellsize_min_dlb[dd->dim[d]] = comm->cellsize_limit;
+ }
+ else
+ {
+ comm->cellsize_min_dlb[dd->dim[d]] =
+ comm->cutoff/comm->cd[d].np_dlb;
+ }
+ }
+ if (comm->cutoff_mbody <= 0)
+ {
+ comm->cutoff_mbody = min(comm->cutoff,comm->cellsize_limit);
+ }
+ if (comm->bDynLoadBal)
+ {
+ set_dlb_limits(dd);
+ }
+ }
+
+ print_dd_settings(fplog,dd,ir,comm->bDynLoadBal,dlb_scale,ddbox);
+ if (comm->eDLB == edlbAUTO)
+ {
+ if (fplog)
+ {
+ fprintf(fplog,"When dynamic load balancing gets turned on, these settings will change to:\n");
+ }
+ print_dd_settings(fplog,dd,ir,TRUE,dlb_scale,ddbox);
+ }
+
+ if (ir->ePBC == epbcNONE)
+ {
+ vol_frac = 1 - 1/(double)dd->nnodes;
+ }
+ else
+ {
+ vol_frac =
+ (1 + comm_box_frac(dd->nc,comm->cutoff,ddbox))/(double)dd->nnodes;
+ }
+ if (debug)
+ {
+ fprintf(debug,"Volume fraction for all DD zones: %f\n",vol_frac);
+ }
+ natoms_tot = comm->cgs_gl.index[comm->cgs_gl.nr];
+
+ dd->ga2la = ga2la_init(natoms_tot,vol_frac*natoms_tot);
+}
+
+static void merge_cg_buffers(int ncell,
+ gmx_domdec_comm_dim_t *cd, int pulse,
+ int *ncg_cell,
+ int *index_gl, int *recv_i,
+ rvec *cg_cm, rvec *recv_vr,
+ int *cgindex,
+ cginfo_mb_t *cginfo_mb,int *cginfo)
+{
+ gmx_domdec_ind_t *ind,*ind_p;
+ int p,cell,c,cg,cg0,cg1,cg_gl,nat;
+ int shift,shift_at;
+
+ ind = &cd->ind[pulse];
+
+ /* First correct the already stored data */
+ shift = ind->nrecv[ncell];
+ for(cell=ncell-1; cell>=0; cell--)
+ {
+ shift -= ind->nrecv[cell];
+ if (shift > 0)
+ {
+ /* Move the cg's present from previous grid pulses */
+ cg0 = ncg_cell[ncell+cell];
+ cg1 = ncg_cell[ncell+cell+1];
+ cgindex[cg1+shift] = cgindex[cg1];
+ for(cg=cg1-1; cg>=cg0; cg--)
+ {
+ index_gl[cg+shift] = index_gl[cg];
+ copy_rvec(cg_cm[cg],cg_cm[cg+shift]);
+ cgindex[cg+shift] = cgindex[cg];
+ cginfo[cg+shift] = cginfo[cg];
+ }
+ /* Correct the already stored send indices for the shift */
+ for(p=1; p<=pulse; p++)
+ {
+ ind_p = &cd->ind[p];
+ cg0 = 0;
+ for(c=0; c<cell; c++)
+ {
+ cg0 += ind_p->nsend[c];
+ }
+ cg1 = cg0 + ind_p->nsend[cell];
+ for(cg=cg0; cg<cg1; cg++)
+ {
+ ind_p->index[cg] += shift;
+ }
+ }
+ }
+ }
+
+ /* Merge in the communicated buffers */
+ shift = 0;
+ shift_at = 0;
+ cg0 = 0;
+ for(cell=0; cell<ncell; cell++)
+ {
+ cg1 = ncg_cell[ncell+cell+1] + shift;
+ if (shift_at > 0)
+ {
+ /* Correct the old cg indices */
+ for(cg=ncg_cell[ncell+cell]; cg<cg1; cg++)
+ {
+ cgindex[cg+1] += shift_at;
+ }
+ }
+ for(cg=0; cg<ind->nrecv[cell]; cg++)
+ {
+ /* Copy this charge group from the buffer */
+ index_gl[cg1] = recv_i[cg0];
+ copy_rvec(recv_vr[cg0],cg_cm[cg1]);
+ /* Add it to the cgindex */
+ cg_gl = index_gl[cg1];
+ cginfo[cg1] = ddcginfo(cginfo_mb,cg_gl);
+ nat = GET_CGINFO_NATOMS(cginfo[cg1]);
+ cgindex[cg1+1] = cgindex[cg1] + nat;
+ cg0++;
+ cg1++;
+ shift_at += nat;
+ }
+ shift += ind->nrecv[cell];
+ ncg_cell[ncell+cell+1] = cg1;
+ }
+}
+
+static void make_cell2at_index(gmx_domdec_comm_dim_t *cd,
+ int nzone,int cg0,const int *cgindex)
+{
+ int cg,zone,p;
+
+ /* Store the atom block boundaries for easy copying of communication buffers
+ */
+ cg = cg0;
+ for(zone=0; zone<nzone; zone++)
+ {
+ for(p=0; p<cd->np; p++) {
+ cd->ind[p].cell2at0[zone] = cgindex[cg];
+ cg += cd->ind[p].nrecv[zone];
+ cd->ind[p].cell2at1[zone] = cgindex[cg];
+ }
+ }
+}
+
+static gmx_bool missing_link(t_blocka *link,int cg_gl,char *bLocalCG)
+{
+ int i;
+ gmx_bool bMiss;
+
+ bMiss = FALSE;
+ for(i=link->index[cg_gl]; i<link->index[cg_gl+1]; i++)
+ {
+ if (!bLocalCG[link->a[i]])
+ {
+ bMiss = TRUE;
+ }
+ }
+
+ return bMiss;
+}
+
+static void setup_dd_communication(gmx_domdec_t *dd,
+ matrix box,gmx_ddbox_t *ddbox,t_forcerec *fr)
+{
+ int dim_ind,dim,dim0,dim1=-1,dim2=-1,dimd,p,nat_tot;
+ int nzone,nzone_send,zone,zonei,cg0,cg1;
+ int c,i,j,cg,cg_gl,nrcg;
+ int *zone_cg_range,pos_cg,*index_gl,*cgindex,*recv_i;
+ gmx_domdec_comm_t *comm;
+ gmx_domdec_zones_t *zones;
+ gmx_domdec_comm_dim_t *cd;
+ gmx_domdec_ind_t *ind;
+ cginfo_mb_t *cginfo_mb;
+ gmx_bool bBondComm,bDist2B,bDistMB,bDistMB_pulse,bDistBonded,bScrew;
+ real r_mb,r_comm2,r_scomm2,r_bcomm2,r,r_0,r_1,r2,rb2,r2inc,inv_ncg,tric_sh;
+ rvec rb,rn;
+ real corner[DIM][4],corner_round_0=0,corner_round_1[4];
+ real bcorner[DIM],bcorner_round_1=0;
+ ivec tric_dist;
+ rvec *cg_cm,*normal,*v_d,*v_0=NULL,*v_1=NULL,*recv_vr;
+ real skew_fac2_d,skew_fac_01;
+ rvec sf2_round;
+ int nsend,nat;
+
+ if (debug)
+ {
+ fprintf(debug,"Setting up DD communication\n");
+ }
+
+ comm = dd->comm;
+ cg_cm = fr->cg_cm;
+
+ for(dim_ind=0; dim_ind<dd->ndim; dim_ind++)
+ {
+ dim = dd->dim[dim_ind];
+
+ /* Check if we need to use triclinic distances */
+ tric_dist[dim_ind] = 0;
+ for(i=0; i<=dim_ind; i++)
+ {
+ if (ddbox->tric_dir[dd->dim[i]])
+ {
+ tric_dist[dim_ind] = 1;
+ }
+ }
+ }
+
+ bBondComm = comm->bBondComm;
+
+ /* Do we need to determine extra distances for multi-body bondeds? */
+ bDistMB = (comm->bInterCGMultiBody && dd->bGridJump && dd->ndim > 1);
+
+ /* Do we need to determine extra distances for only two-body bondeds? */
+ bDist2B = (bBondComm && !bDistMB);
+
+ r_comm2 = sqr(comm->cutoff);
+ r_bcomm2 = sqr(comm->cutoff_mbody);
+
+ if (debug)
+ {
+ fprintf(debug,"bBondComm %d, r_bc %f\n",bBondComm,sqrt(r_bcomm2));
+ }
+
+ zones = &comm->zones;
+
+ dim0 = dd->dim[0];
+ /* The first dimension is equal for all cells */
+ corner[0][0] = comm->cell_x0[dim0];
+ if (bDistMB)
+ {
+ bcorner[0] = corner[0][0];
+ }
+ if (dd->ndim >= 2)
+ {
+ dim1 = dd->dim[1];
+ /* This cell row is only seen from the first row */
+ corner[1][0] = comm->cell_x0[dim1];
+ /* All rows can see this row */
+ corner[1][1] = comm->cell_x0[dim1];
+ if (dd->bGridJump)
+ {
+ corner[1][1] = max(comm->cell_x0[dim1],comm->zone_d1[1].mch0);
+ if (bDistMB)
+ {
+ /* For the multi-body distance we need the maximum */
+ bcorner[1] = max(comm->cell_x0[dim1],comm->zone_d1[1].p1_0);
+ }
+ }
+ /* Set the upper-right corner for rounding */
+ corner_round_0 = comm->cell_x1[dim0];
+
+ if (dd->ndim >= 3)
+ {
+ dim2 = dd->dim[2];
+ for(j=0; j<4; j++)
+ {
+ corner[2][j] = comm->cell_x0[dim2];
+ }
+ if (dd->bGridJump)
+ {
+ /* Use the maximum of the i-cells that see a j-cell */
+ for(i=0; i<zones->nizone; i++)
+ {
+ for(j=zones->izone[i].j0; j<zones->izone[i].j1; j++)
+ {
+ if (j >= 4)
+ {
+ corner[2][j-4] =
+ max(corner[2][j-4],
+ comm->zone_d2[zones->shift[i][dim0]][zones->shift[i][dim1]].mch0);
+ }
+ }
+ }
+ if (bDistMB)
+ {
+ /* For the multi-body distance we need the maximum */
+ bcorner[2] = comm->cell_x0[dim2];
+ for(i=0; i<2; i++)
+ {
+ for(j=0; j<2; j++)
+ {
+ bcorner[2] = max(bcorner[2],
+ comm->zone_d2[i][j].p1_0);
+ }
+ }
+ }
+ }
+
+ /* Set the upper-right corner for rounding */
+ /* Cell (0,0,0) and cell (1,0,0) can see cell 4 (0,1,1)
+ * Only cell (0,0,0) can see cell 7 (1,1,1)
+ */
+ corner_round_1[0] = comm->cell_x1[dim1];
+ corner_round_1[3] = comm->cell_x1[dim1];
+ if (dd->bGridJump)
+ {
+ corner_round_1[0] = max(comm->cell_x1[dim1],
+ comm->zone_d1[1].mch1);
+ if (bDistMB)
+ {
+ /* For the multi-body distance we need the maximum */
+ bcorner_round_1 = max(comm->cell_x1[dim1],
+ comm->zone_d1[1].p1_1);
+ }
+ }
+ }
+ }
+
+ /* Triclinic stuff */
+ normal = ddbox->normal;
+ skew_fac_01 = 0;
+ if (dd->ndim >= 2)
+ {
+ v_0 = ddbox->v[dim0];
+ if (ddbox->tric_dir[dim0] && ddbox->tric_dir[dim1])
+ {
+ /* Determine the coupling coefficient for the distances
+ * to the cell planes along dim0 and dim1 through dim2.
+ * This is required for correct rounding.
+ */
+ skew_fac_01 =
+ ddbox->v[dim0][dim1+1][dim0]*ddbox->v[dim1][dim1+1][dim1];
+ if (debug)
+ {
+ fprintf(debug,"\nskew_fac_01 %f\n",skew_fac_01);
+ }
+ }
+ }
+ if (dd->ndim >= 3)
+ {
+ v_1 = ddbox->v[dim1];
+ }
+
+ zone_cg_range = zones->cg_range;
+ index_gl = dd->index_gl;
+ cgindex = dd->cgindex;
+ cginfo_mb = fr->cginfo_mb;
+
+ zone_cg_range[0] = 0;
+ zone_cg_range[1] = dd->ncg_home;
+ comm->zone_ncg1[0] = dd->ncg_home;
+ pos_cg = dd->ncg_home;
+
+ nat_tot = dd->nat_home;
+ nzone = 1;
+ for(dim_ind=0; dim_ind<dd->ndim; dim_ind++)
+ {
+ dim = dd->dim[dim_ind];
+ cd = &comm->cd[dim_ind];
+
+ if (dim >= ddbox->npbcdim && dd->ci[dim] == 0)
+ {
+ /* No pbc in this dimension, the first node should not comm. */
+ nzone_send = 0;
+ }
+ else
+ {
+ nzone_send = nzone;
+ }
+
+ bScrew = (dd->bScrewPBC && dim == XX);
+
+ v_d = ddbox->v[dim];
+ skew_fac2_d = sqr(ddbox->skew_fac[dim]);
+
+ cd->bInPlace = TRUE;
+ for(p=0; p<cd->np; p++)
+ {
+ /* Only atoms communicated in the first pulse are used
+ * for multi-body bonded interactions or for bBondComm.
+ */
+ bDistBonded = ((bDistMB || bDist2B) && p == 0);
+ bDistMB_pulse = (bDistMB && bDistBonded);
+
+ ind = &cd->ind[p];
+ nsend = 0;
+ nat = 0;
+ for(zone=0; zone<nzone_send; zone++)
+ {
+ if (tric_dist[dim_ind] && dim_ind > 0)
+ {
+ /* Determine slightly more optimized skew_fac's
+ * for rounding.
+ * This reduces the number of communicated atoms
+ * by about 10% for 3D DD of rhombic dodecahedra.
+ */
+ for(dimd=0; dimd<dim; dimd++)
+ {
+ sf2_round[dimd] = 1;
+ if (ddbox->tric_dir[dimd])
+ {
+ for(i=dd->dim[dimd]+1; i<DIM; i++)
+ {
+ /* If we are shifted in dimension i
+ * and the cell plane is tilted forward
+ * in dimension i, skip this coupling.
+ */
+ if (!(zones->shift[nzone+zone][i] &&
+ ddbox->v[dimd][i][dimd] >= 0))
+ {
+ sf2_round[dimd] +=
+ sqr(ddbox->v[dimd][i][dimd]);
+ }
+ }
+ sf2_round[dimd] = 1/sf2_round[dimd];
+ }
+ }
+ }
+
+ zonei = zone_perm[dim_ind][zone];
+ if (p == 0)
+ {
+ /* Here we permutate the zones to obtain a convenient order
+ * for neighbor searching
+ */
+ cg0 = zone_cg_range[zonei];
+ cg1 = zone_cg_range[zonei+1];
+ }
+ else
+ {
+ /* Look only at the cg's received in the previous grid pulse
+ */
+ cg1 = zone_cg_range[nzone+zone+1];
+ cg0 = cg1 - cd->ind[p-1].nrecv[zone];
+ }
+ ind->nsend[zone] = 0;
+ for(cg=cg0; cg<cg1; cg++)
+ {
+ r2 = 0;
+ rb2 = 0;
+ if (tric_dist[dim_ind] == 0)
+ {
+ /* Rectangular direction, easy */
+ r = cg_cm[cg][dim] - corner[dim_ind][zone];
+ if (r > 0)
+ {
+ r2 += r*r;
+ }
+ if (bDistMB_pulse)
+ {
+ r = cg_cm[cg][dim] - bcorner[dim_ind];
+ if (r > 0)
+ {
+ rb2 += r*r;
+ }
+ }
+ /* Rounding gives at most a 16% reduction
+ * in communicated atoms
+ */
+ if (dim_ind >= 1 && (zonei == 1 || zonei == 2))
+ {
+ r = cg_cm[cg][dim0] - corner_round_0;
+ /* This is the first dimension, so always r >= 0 */
+ r2 += r*r;
+ if (bDistMB_pulse)
+ {
+ rb2 += r*r;
+ }
+ }
+ if (dim_ind == 2 && (zonei == 2 || zonei == 3))
+ {
+ r = cg_cm[cg][dim1] - corner_round_1[zone];
+ if (r > 0)
+ {
+ r2 += r*r;
+ }
+ if (bDistMB_pulse)
+ {
+ r = cg_cm[cg][dim1] - bcorner_round_1;
+ if (r > 0)
+ {
+ rb2 += r*r;
+ }
+ }
+ }
+ }
+ else
+ {
+ /* Triclinic direction, more complicated */
+ clear_rvec(rn);
+ clear_rvec(rb);
+ /* Rounding, conservative as the skew_fac multiplication
+ * will slightly underestimate the distance.
+ */
+ if (dim_ind >= 1 && (zonei == 1 || zonei == 2))
+ {
+ rn[dim0] = cg_cm[cg][dim0] - corner_round_0;
+ for(i=dim0+1; i<DIM; i++)
+ {
+ rn[dim0] -= cg_cm[cg][i]*v_0[i][dim0];
+ }
+ r2 = rn[dim0]*rn[dim0]*sf2_round[dim0];
+ if (bDistMB_pulse)
+ {
+ rb[dim0] = rn[dim0];
+ rb2 = r2;
+ }
+ /* Take care that the cell planes along dim0 might not
+ * be orthogonal to those along dim1 and dim2.
+ */
+ for(i=1; i<=dim_ind; i++)
+ {
+ dimd = dd->dim[i];
+ if (normal[dim0][dimd] > 0)
+ {
+ rn[dimd] -= rn[dim0]*normal[dim0][dimd];
+ if (bDistMB_pulse)
+ {
+ rb[dimd] -= rb[dim0]*normal[dim0][dimd];
+ }
+ }
+ }
+ }
+ if (dim_ind == 2 && (zonei == 2 || zonei == 3))
+ {
+ rn[dim1] += cg_cm[cg][dim1] - corner_round_1[zone];
+ tric_sh = 0;
+ for(i=dim1+1; i<DIM; i++)
+ {
+ tric_sh -= cg_cm[cg][i]*v_1[i][dim1];
+ }
+ rn[dim1] += tric_sh;
+ if (rn[dim1] > 0)
+ {
+ r2 += rn[dim1]*rn[dim1]*sf2_round[dim1];
+ /* Take care of coupling of the distances
+ * to the planes along dim0 and dim1 through dim2.
+ */
+ r2 -= rn[dim0]*rn[dim1]*skew_fac_01;
+ /* Take care that the cell planes along dim1
+ * might not be orthogonal to that along dim2.
+ */
+ if (normal[dim1][dim2] > 0)
+ {
+ rn[dim2] -= rn[dim1]*normal[dim1][dim2];
+ }
+ }
+ if (bDistMB_pulse)
+ {
+ rb[dim1] +=
+ cg_cm[cg][dim1] - bcorner_round_1 + tric_sh;
+ if (rb[dim1] > 0)
+ {
+ rb2 += rb[dim1]*rb[dim1]*sf2_round[dim1];
+ /* Take care of coupling of the distances
+ * to the planes along dim0 and dim1 through dim2.
+ */
+ rb2 -= rb[dim0]*rb[dim1]*skew_fac_01;
+ /* Take care that the cell planes along dim1
+ * might not be orthogonal to that along dim2.
+ */
+ if (normal[dim1][dim2] > 0)
+ {
+ rb[dim2] -= rb[dim1]*normal[dim1][dim2];
+ }
+ }
+ }
+ }
+ /* The distance along the communication direction */
+ rn[dim] += cg_cm[cg][dim] - corner[dim_ind][zone];
+ tric_sh = 0;
+ for(i=dim+1; i<DIM; i++)
+ {
+ tric_sh -= cg_cm[cg][i]*v_d[i][dim];
+ }
+ rn[dim] += tric_sh;
+ if (rn[dim] > 0)
+ {
+ r2 += rn[dim]*rn[dim]*skew_fac2_d;
+ /* Take care of coupling of the distances
+ * to the planes along dim0 and dim1 through dim2.
+ */
+ if (dim_ind == 1 && zonei == 1)
+ {
+ r2 -= rn[dim0]*rn[dim]*skew_fac_01;
+ }
+ }
+ if (bDistMB_pulse)
+ {
+ clear_rvec(rb);
+ rb[dim] += cg_cm[cg][dim] - bcorner[dim_ind] + tric_sh;
+ if (rb[dim] > 0)
+ {
+ rb2 += rb[dim]*rb[dim]*skew_fac2_d;
+ /* Take care of coupling of the distances
+ * to the planes along dim0 and dim1 through dim2.
+ */
+ if (dim_ind == 1 && zonei == 1)
+ {
+ rb2 -= rb[dim0]*rb[dim]*skew_fac_01;
+ }
+ }
+ }
+ }
+
+ if (r2 < r_comm2 ||
+ (bDistBonded &&
+ ((bDistMB && rb2 < r_bcomm2) ||
+ (bDist2B && r2 < r_bcomm2)) &&
+ (!bBondComm ||
+ (GET_CGINFO_BOND_INTER(fr->cginfo[cg]) &&
+ missing_link(comm->cglink,index_gl[cg],
+ comm->bLocalCG)))))
+ {
+ /* Make an index to the local charge groups */
+ if (nsend+1 > ind->nalloc)
+ {
+ ind->nalloc = over_alloc_large(nsend+1);
+ srenew(ind->index,ind->nalloc);
+ }
+ if (nsend+1 > comm->nalloc_int)
+ {
+ comm->nalloc_int = over_alloc_large(nsend+1);
+ srenew(comm->buf_int,comm->nalloc_int);
+ }
+ ind->index[nsend] = cg;
+ comm->buf_int[nsend] = index_gl[cg];
+ ind->nsend[zone]++;
+ vec_rvec_check_alloc(&comm->vbuf,nsend+1);
+
+ if (dd->ci[dim] == 0)
+ {
+ /* Correct cg_cm for pbc */
+ rvec_add(cg_cm[cg],box[dim],comm->vbuf.v[nsend]);
+ if (bScrew)
+ {
+ comm->vbuf.v[nsend][YY] =
+ box[YY][YY]-comm->vbuf.v[nsend][YY];
+ comm->vbuf.v[nsend][ZZ] =
+ box[ZZ][ZZ]-comm->vbuf.v[nsend][ZZ];
+ }
+ }
+ else
+ {
+ copy_rvec(cg_cm[cg],comm->vbuf.v[nsend]);
+ }
+ nsend++;
+ nat += cgindex[cg+1] - cgindex[cg];
+ }
+ }
+ }
+ /* Clear the counts in case we do not have pbc */
+ for(zone=nzone_send; zone<nzone; zone++)
+ {
+ ind->nsend[zone] = 0;
+ }
+ ind->nsend[nzone] = nsend;
+ ind->nsend[nzone+1] = nat;
+ /* Communicate the number of cg's and atoms to receive */
+ dd_sendrecv_int(dd, dim_ind, dddirBackward,
+ ind->nsend, nzone+2,
+ ind->nrecv, nzone+2);
+
+ /* The rvec buffer is also required for atom buffers of size nsend
+ * in dd_move_x and dd_move_f.
+ */
+ vec_rvec_check_alloc(&comm->vbuf,ind->nsend[nzone+1]);
+
+ if (p > 0)
+ {
+ /* We can receive in place if only the last zone is not empty */
+ for(zone=0; zone<nzone-1; zone++)
+ {
+ if (ind->nrecv[zone] > 0)
+ {
+ cd->bInPlace = FALSE;
+ }
+ }
+ if (!cd->bInPlace)
+ {
+ /* The int buffer is only required here for the cg indices */
+ if (ind->nrecv[nzone] > comm->nalloc_int2)
+ {
+ comm->nalloc_int2 = over_alloc_dd(ind->nrecv[nzone]);
+ srenew(comm->buf_int2,comm->nalloc_int2);
+ }
+ /* The rvec buffer is also required for atom buffers
+ * of size nrecv in dd_move_x and dd_move_f.
+ */
+ i = max(cd->ind[0].nrecv[nzone+1],ind->nrecv[nzone+1]);
+ vec_rvec_check_alloc(&comm->vbuf2,i);
+ }
+ }
+
+ /* Make space for the global cg indices */
+ if (pos_cg + ind->nrecv[nzone] > dd->cg_nalloc
+ || dd->cg_nalloc == 0)
+ {
+ dd->cg_nalloc = over_alloc_dd(pos_cg + ind->nrecv[nzone]);
+ srenew(index_gl,dd->cg_nalloc);
+ srenew(cgindex,dd->cg_nalloc+1);
+ }
+ /* Communicate the global cg indices */
+ if (cd->bInPlace)
+ {
+ recv_i = index_gl + pos_cg;
+ }
+ else
+ {
+ recv_i = comm->buf_int2;
+ }
+ dd_sendrecv_int(dd, dim_ind, dddirBackward,
+ comm->buf_int, nsend,
+ recv_i, ind->nrecv[nzone]);
+
+ /* Make space for cg_cm */
+ if (pos_cg + ind->nrecv[nzone] > fr->cg_nalloc)
+ {
+ dd_realloc_fr_cg(fr,pos_cg + ind->nrecv[nzone]);
+ cg_cm = fr->cg_cm;
+ }
+ /* Communicate cg_cm */
+ if (cd->bInPlace)
+ {
+ recv_vr = cg_cm + pos_cg;
+ }
+ else
+ {
+ recv_vr = comm->vbuf2.v;
+ }
+ dd_sendrecv_rvec(dd, dim_ind, dddirBackward,
+ comm->vbuf.v, nsend,
+ recv_vr, ind->nrecv[nzone]);
+
+ /* Make the charge group index */
+ if (cd->bInPlace)
+ {
+ zone = (p == 0 ? 0 : nzone - 1);
+ while (zone < nzone)
+ {
+ for(cg=0; cg<ind->nrecv[zone]; cg++)
+ {
+ cg_gl = index_gl[pos_cg];
+ fr->cginfo[pos_cg] = ddcginfo(cginfo_mb,cg_gl);
+ nrcg = GET_CGINFO_NATOMS(fr->cginfo[pos_cg]);
+ cgindex[pos_cg+1] = cgindex[pos_cg] + nrcg;
+ if (bBondComm)
+ {
+ /* Update the charge group presence,
+ * so we can use it in the next pass of the loop.
+ */
+ comm->bLocalCG[cg_gl] = TRUE;
+ }
+ pos_cg++;
+ }
+ if (p == 0)
+ {
+ comm->zone_ncg1[nzone+zone] = ind->nrecv[zone];
+ }
+ zone++;
+ zone_cg_range[nzone+zone] = pos_cg;
+ }
+ }
+ else
+ {
+ /* This part of the code is never executed with bBondComm. */
+ merge_cg_buffers(nzone,cd,p,zone_cg_range,
+ index_gl,recv_i,cg_cm,recv_vr,
+ cgindex,fr->cginfo_mb,fr->cginfo);
+ pos_cg += ind->nrecv[nzone];
+ }
+ nat_tot += ind->nrecv[nzone+1];
+ }
+ if (!cd->bInPlace)
+ {
+ /* Store the atom block for easy copying of communication buffers */
+ make_cell2at_index(cd,nzone,zone_cg_range[nzone],cgindex);
+ }
+ nzone += nzone;
+ }
+ dd->index_gl = index_gl;
+ dd->cgindex = cgindex;
+
+ dd->ncg_tot = zone_cg_range[zones->n];
+ dd->nat_tot = nat_tot;
+ comm->nat[ddnatHOME] = dd->nat_home;
+ for(i=ddnatZONE; i<ddnatNR; i++)
+ {
+ comm->nat[i] = dd->nat_tot;
+ }
+
+ if (!bBondComm)
+ {
+ /* We don't need to update cginfo, since that was alrady done above.
+ * So we pass NULL for the forcerec.
+ */
+ dd_set_cginfo(dd->index_gl,dd->ncg_home,dd->ncg_tot,
+ NULL,comm->bLocalCG);
+ }
+
+ if (debug)
+ {
+ fprintf(debug,"Finished setting up DD communication, zones:");
+ for(c=0; c<zones->n; c++)
+ {
+ fprintf(debug," %d",zones->cg_range[c+1]-zones->cg_range[c]);
+ }
+ fprintf(debug,"\n");
+ }
+}
+
+static void set_cg_boundaries(gmx_domdec_zones_t *zones)
+{
+ int c;
+
+ for(c=0; c<zones->nizone; c++)
+ {
+ zones->izone[c].cg1 = zones->cg_range[c+1];
+ zones->izone[c].jcg0 = zones->cg_range[zones->izone[c].j0];
+ zones->izone[c].jcg1 = zones->cg_range[zones->izone[c].j1];
+ }
+}
+
+static int comp_cgsort(const void *a,const void *b)
+{
+ int comp;
+
+ gmx_cgsort_t *cga,*cgb;
+ cga = (gmx_cgsort_t *)a;
+ cgb = (gmx_cgsort_t *)b;
+
+ comp = cga->nsc - cgb->nsc;
+ if (comp == 0)
+ {
+ comp = cga->ind_gl - cgb->ind_gl;
+ }
+
+ return comp;
+}
+
+static void order_int_cg(int n,gmx_cgsort_t *sort,
+ int *a,int *buf)
+{
+ int i;
+
+ /* Order the data */
+ for(i=0; i<n; i++)
+ {
+ buf[i] = a[sort[i].ind];
+ }
+
+ /* Copy back to the original array */
+ for(i=0; i<n; i++)
+ {
+ a[i] = buf[i];
+ }
+}
+
+static void order_vec_cg(int n,gmx_cgsort_t *sort,
+ rvec *v,rvec *buf)
+{
+ int i;
+
+ /* Order the data */
+ for(i=0; i<n; i++)
+ {
+ copy_rvec(v[sort[i].ind],buf[i]);
+ }
+
+ /* Copy back to the original array */
+ for(i=0; i<n; i++)
+ {
+ copy_rvec(buf[i],v[i]);
+ }
+}
+
+static void order_vec_atom(int ncg,int *cgindex,gmx_cgsort_t *sort,
+ rvec *v,rvec *buf)
+{
+ int a,atot,cg,cg0,cg1,i;
+
+ /* Order the data */
+ a = 0;
+ for(cg=0; cg<ncg; cg++)
+ {
+ cg0 = cgindex[sort[cg].ind];
+ cg1 = cgindex[sort[cg].ind+1];
+ for(i=cg0; i<cg1; i++)
+ {
+ copy_rvec(v[i],buf[a]);
+ a++;
+ }
+ }
+ atot = a;
+
+ /* Copy back to the original array */
+ for(a=0; a<atot; a++)
+ {
+ copy_rvec(buf[a],v[a]);
+ }
+}
+
+static void ordered_sort(int nsort2,gmx_cgsort_t *sort2,
+ int nsort_new,gmx_cgsort_t *sort_new,
+ gmx_cgsort_t *sort1)
+{
+ int i1,i2,i_new;
+
+ /* The new indices are not very ordered, so we qsort them */
+ qsort_threadsafe(sort_new,nsort_new,sizeof(sort_new[0]),comp_cgsort);
+
+ /* sort2 is already ordered, so now we can merge the two arrays */
+ i1 = 0;
+ i2 = 0;
+ i_new = 0;
+ while(i2 < nsort2 || i_new < nsort_new)
+ {
+ if (i2 == nsort2)
+ {
+ sort1[i1++] = sort_new[i_new++];
+ }
+ else if (i_new == nsort_new)
+ {
+ sort1[i1++] = sort2[i2++];
+ }
+ else if (sort2[i2].nsc < sort_new[i_new].nsc ||
+ (sort2[i2].nsc == sort_new[i_new].nsc &&
+ sort2[i2].ind_gl < sort_new[i_new].ind_gl))
+ {
+ sort1[i1++] = sort2[i2++];
+ }
+ else
+ {
+ sort1[i1++] = sort_new[i_new++];
+ }
+ }
+}
+
+static void dd_sort_state(gmx_domdec_t *dd,int ePBC,
+ rvec *cgcm,t_forcerec *fr,t_state *state,
+ int ncg_home_old)
+{
+ gmx_domdec_sort_t *sort;
+ gmx_cgsort_t *cgsort,*sort_i;
+ int ncg_new,nsort2,nsort_new,i,cell_index,*ibuf,cgsize;
+ rvec *vbuf;
+
+ sort = dd->comm->sort;
+
+ if (dd->ncg_home > sort->sort_nalloc)
+ {
+ sort->sort_nalloc = over_alloc_dd(dd->ncg_home);
+ srenew(sort->sort1,sort->sort_nalloc);
+ srenew(sort->sort2,sort->sort_nalloc);
+ }
+
+ if (ncg_home_old >= 0)
+ {
+ /* The charge groups that remained in the same ns grid cell
+ * are completely ordered. So we can sort efficiently by sorting
+ * the charge groups that did move into the stationary list.
+ */
+ ncg_new = 0;
+ nsort2 = 0;
+ nsort_new = 0;
+ for(i=0; i<dd->ncg_home; i++)
+ {
+ /* Check if this cg did not move to another node */
+ cell_index = fr->ns.grid->cell_index[i];
+ if (cell_index != 4*fr->ns.grid->ncells)
+ {
+ if (i >= ncg_home_old || cell_index != sort->sort1[i].nsc)
+ {
+ /* This cg is new on this node or moved ns grid cell */
+ if (nsort_new >= sort->sort_new_nalloc)
+ {
+ sort->sort_new_nalloc = over_alloc_dd(nsort_new+1);
+ srenew(sort->sort_new,sort->sort_new_nalloc);
+ }
+ sort_i = &(sort->sort_new[nsort_new++]);
+ }
+ else
+ {
+ /* This cg did not move */
+ sort_i = &(sort->sort2[nsort2++]);
+ }
+ /* Sort on the ns grid cell indices
+ * and the global topology index
+ */
+ sort_i->nsc = cell_index;
+ sort_i->ind_gl = dd->index_gl[i];
+ sort_i->ind = i;
+ ncg_new++;
+ }
+ }
+ if (debug)
+ {
+ fprintf(debug,"ordered sort cgs: stationary %d moved %d\n",
+ nsort2,nsort_new);
+ }
+ /* Sort efficiently */
+ ordered_sort(nsort2,sort->sort2,nsort_new,sort->sort_new,sort->sort1);
+ }
+ else
+ {
+ cgsort = sort->sort1;
+ ncg_new = 0;
+ for(i=0; i<dd->ncg_home; i++)
+ {
+ /* Sort on the ns grid cell indices
+ * and the global topology index
+ */
+ cgsort[i].nsc = fr->ns.grid->cell_index[i];
+ cgsort[i].ind_gl = dd->index_gl[i];
+ cgsort[i].ind = i;
+ if (cgsort[i].nsc != 4*fr->ns.grid->ncells)
+ {
+ ncg_new++;
+ }
+ }
+ if (debug)
+ {
+ fprintf(debug,"qsort cgs: %d new home %d\n",dd->ncg_home,ncg_new);
+ }
+ /* Determine the order of the charge groups using qsort */
+ qsort_threadsafe(cgsort,dd->ncg_home,sizeof(cgsort[0]),comp_cgsort);
+ }
+ cgsort = sort->sort1;
+
+ /* We alloc with the old size, since cgindex is still old */
+ vec_rvec_check_alloc(&dd->comm->vbuf,dd->cgindex[dd->ncg_home]);
+ vbuf = dd->comm->vbuf.v;
+
+ /* Remove the charge groups which are no longer at home here */
+ dd->ncg_home = ncg_new;
+
+ /* Reorder the state */
+ for(i=0; i<estNR; i++)
+ {
+ if (EST_DISTR(i) && (state->flags & (1<<i)))
+ {
+ switch (i)
+ {
+ case estX:
+ order_vec_atom(dd->ncg_home,dd->cgindex,cgsort,state->x,vbuf);
+ break;
+ case estV:
+ order_vec_atom(dd->ncg_home,dd->cgindex,cgsort,state->v,vbuf);
+ break;
+ case estSDX:
+ order_vec_atom(dd->ncg_home,dd->cgindex,cgsort,state->sd_X,vbuf);
+ break;
+ case estCGP:
+ order_vec_atom(dd->ncg_home,dd->cgindex,cgsort,state->cg_p,vbuf);
+ break;
+ case estLD_RNG:
+ case estLD_RNGI:
+ case estDISRE_INITF:
+ case estDISRE_RM3TAV:
+ case estORIRE_INITF:
+ case estORIRE_DTAV:
+ /* No ordering required */
+ break;
+ default:
+ gmx_incons("Unknown state entry encountered in dd_sort_state");
+ break;
+ }
+ }
+ }
+ /* Reorder cgcm */
+ order_vec_cg(dd->ncg_home,cgsort,cgcm,vbuf);
+
+ if (dd->ncg_home+1 > sort->ibuf_nalloc)
+ {
+ sort->ibuf_nalloc = over_alloc_dd(dd->ncg_home+1);
+ srenew(sort->ibuf,sort->ibuf_nalloc);
+ }
+ ibuf = sort->ibuf;
+ /* Reorder the global cg index */
+ order_int_cg(dd->ncg_home,cgsort,dd->index_gl,ibuf);
+ /* Reorder the cginfo */
+ order_int_cg(dd->ncg_home,cgsort,fr->cginfo,ibuf);
+ /* Rebuild the local cg index */
+ ibuf[0] = 0;
+ for(i=0; i<dd->ncg_home; i++)
+ {
+ cgsize = dd->cgindex[cgsort[i].ind+1] - dd->cgindex[cgsort[i].ind];
+ ibuf[i+1] = ibuf[i] + cgsize;
+ }
+ for(i=0; i<dd->ncg_home+1; i++)
+ {
+ dd->cgindex[i] = ibuf[i];
+ }
+ /* Set the home atom number */
+ dd->nat_home = dd->cgindex[dd->ncg_home];
+
+ /* Copy the sorted ns cell indices back to the ns grid struct */
+ for(i=0; i<dd->ncg_home; i++)
+ {
+ fr->ns.grid->cell_index[i] = cgsort[i].nsc;
+ }
+ fr->ns.grid->nr = dd->ncg_home;
+}
+
+static void add_dd_statistics(gmx_domdec_t *dd)
+{
+ gmx_domdec_comm_t *comm;
+ int ddnat;
+
+ comm = dd->comm;
+
+ for(ddnat=ddnatZONE; ddnat<ddnatNR; ddnat++)
+ {
+ comm->sum_nat[ddnat-ddnatZONE] +=
+ comm->nat[ddnat] - comm->nat[ddnat-1];
+ }
+ comm->ndecomp++;
+}
+
+void reset_dd_statistics_counters(gmx_domdec_t *dd)
+{
+ gmx_domdec_comm_t *comm;
+ int ddnat;
+
+ comm = dd->comm;
+
+ /* Reset all the statistics and counters for total run counting */
+ for(ddnat=ddnatZONE; ddnat<ddnatNR; ddnat++)
+ {
+ comm->sum_nat[ddnat-ddnatZONE] = 0;
+ }
+ comm->ndecomp = 0;
+ comm->nload = 0;
+ comm->load_step = 0;
+ comm->load_sum = 0;
+ comm->load_max = 0;
+ clear_ivec(comm->load_lim);
+ comm->load_mdf = 0;
+ comm->load_pme = 0;
+}
+
+void print_dd_statistics(t_commrec *cr,t_inputrec *ir,FILE *fplog)
+{
+ gmx_domdec_comm_t *comm;
+ int ddnat;
+ double av;
+
+ comm = cr->dd->comm;
+
+ gmx_sumd(ddnatNR-ddnatZONE,comm->sum_nat,cr);
+
+ if (fplog == NULL)
+ {
+ return;
+ }
+
+ fprintf(fplog,"\n D O M A I N D E C O M P O S I T I O N S T A T I S T I C S\n\n");
+
+ for(ddnat=ddnatZONE; ddnat<ddnatNR; ddnat++)
+ {
+ av = comm->sum_nat[ddnat-ddnatZONE]/comm->ndecomp;
+ switch(ddnat)
+ {
+ case ddnatZONE:
+ fprintf(fplog,
+ " av. #atoms communicated per step for force: %d x %.1f\n",
+ 2,av);
+ break;
+ case ddnatVSITE:
+ if (cr->dd->vsite_comm)
+ {
+ fprintf(fplog,
+ " av. #atoms communicated per step for vsites: %d x %.1f\n",
+ (EEL_PME(ir->coulombtype) || ir->coulombtype==eelEWALD) ? 3 : 2,
+ av);
+ }
+ break;
+ case ddnatCON:
+ if (cr->dd->constraint_comm)
+ {
+ fprintf(fplog,
+ " av. #atoms communicated per step for LINCS: %d x %.1f\n",
+ 1 + ir->nLincsIter,av);
+ }
+ break;
+ default:
+ gmx_incons(" Unknown type for DD statistics");
+ }
+ }
+ fprintf(fplog,"\n");
+
+ if (comm->bRecordLoad && EI_DYNAMICS(ir->eI))
+ {
+ print_dd_load_av(fplog,cr->dd);
+ }
+}
+
+void dd_partition_system(FILE *fplog,
+ gmx_large_int_t step,
+ t_commrec *cr,
+ gmx_bool bMasterState,
+ int nstglobalcomm,
+ t_state *state_global,
+ gmx_mtop_t *top_global,
+ t_inputrec *ir,
+ t_state *state_local,
+ rvec **f,
+ t_mdatoms *mdatoms,
+ gmx_localtop_t *top_local,
+ t_forcerec *fr,
+ gmx_vsite_t *vsite,
+ gmx_shellfc_t shellfc,
+ gmx_constr_t constr,
+ t_nrnb *nrnb,
+ gmx_wallcycle_t wcycle,
+ gmx_bool bVerbose)
+{
+ gmx_domdec_t *dd;
+ gmx_domdec_comm_t *comm;
+ gmx_ddbox_t ddbox={0};
+ t_block *cgs_gl;
+ gmx_large_int_t step_pcoupl;
+ rvec cell_ns_x0,cell_ns_x1;
+ int i,j,n,cg0=0,ncg_home_old=-1,nat_f_novirsum;
+ gmx_bool bBoxChanged,bNStGlobalComm,bDoDLB,bCheckDLB,bTurnOnDLB,bLogLoad;
+ gmx_bool bRedist,bSortCG,bResortAll;
+ ivec ncells_old,np;
+ real grid_density;
+ char sbuf[22];
+
+ dd = cr->dd;
+ comm = dd->comm;
+
+ bBoxChanged = (bMasterState || DEFORM(*ir));
+ if (ir->epc != epcNO)
+ {
+ /* With nstpcouple > 1 pressure coupling happens.
+ * one step after calculating the pressure.
+ * Box scaling happens at the end of the MD step,
+ * after the DD partitioning.
+ * We therefore have to do DLB in the first partitioning
+ * after an MD step where P-coupling occured.
+ * We need to determine the last step in which p-coupling occurred.
+ * MRS -- need to validate this for vv?
+ */
+ n = ir->nstpcouple;
+ if (n == 1)
+ {
+ step_pcoupl = step - 1;
+ }
+ else
+ {
+ step_pcoupl = ((step - 1)/n)*n + 1;
+ }
+ if (step_pcoupl >= comm->globalcomm_step)
+ {
+ bBoxChanged = TRUE;
+ }
+ }
+
+ bNStGlobalComm = (step >= comm->globalcomm_step + nstglobalcomm);
+
+ if (!comm->bDynLoadBal)
+ {
+ bDoDLB = FALSE;
+ }
+ else
+ {
+ /* Should we do dynamic load balacing this step?
+ * Since it requires (possibly expensive) global communication,
+ * we might want to do DLB less frequently.
+ */
+ if (bBoxChanged || ir->epc != epcNO)
+ {
+ bDoDLB = bBoxChanged;
+ }
+ else
+ {
+ bDoDLB = bNStGlobalComm;
+ }
+ }
+
+ /* Check if we have recorded loads on the nodes */
+ if (comm->bRecordLoad && dd_load_count(comm))
+ {
+ if (comm->eDLB == edlbAUTO && !comm->bDynLoadBal)
+ {
+ /* Check if we should use DLB at the second partitioning
+ * and every 100 partitionings,
+ * so the extra communication cost is negligible.
+ */
+ n = max(100,nstglobalcomm);
+ bCheckDLB = (comm->n_load_collect == 0 ||
+ comm->n_load_have % n == n-1);
+ }
+ else
+ {
+ bCheckDLB = FALSE;
+ }
+
+ /* Print load every nstlog, first and last step to the log file */
+ bLogLoad = ((ir->nstlog > 0 && step % ir->nstlog == 0) ||
+ comm->n_load_collect == 0 ||
+ (ir->nsteps >= 0 &&
+ (step + ir->nstlist > ir->init_step + ir->nsteps)));
+
+ /* Avoid extra communication due to verbose screen output
+ * when nstglobalcomm is set.
+ */
+ if (bDoDLB || bLogLoad || bCheckDLB ||
+ (bVerbose && (ir->nstlist == 0 || nstglobalcomm <= ir->nstlist)))
+ {
+ get_load_distribution(dd,wcycle);
+ if (DDMASTER(dd))
+ {
+ if (bLogLoad)
+ {
+ dd_print_load(fplog,dd,step-1);
+ }
+ if (bVerbose)
+ {
+ dd_print_load_verbose(dd);
+ }
+ }
+ comm->n_load_collect++;
+
+ if (bCheckDLB) {
+ /* Since the timings are node dependent, the master decides */
+ if (DDMASTER(dd))
+ {
+ bTurnOnDLB =
+ (dd_force_imb_perf_loss(dd) >= DD_PERF_LOSS);
+ if (debug)
+ {
+ fprintf(debug,"step %s, imb loss %f\n",
+ gmx_step_str(step,sbuf),
+ dd_force_imb_perf_loss(dd));
+ }
+ }
+ dd_bcast(dd,sizeof(bTurnOnDLB),&bTurnOnDLB);
+ if (bTurnOnDLB)
+ {
+ turn_on_dlb(fplog,cr,step);
+ bDoDLB = TRUE;
+ }
+ }
+ }
+ comm->n_load_have++;
+ }
+
+ cgs_gl = &comm->cgs_gl;
+
+ bRedist = FALSE;
+ if (bMasterState)
+ {
+ /* Clear the old state */
+ clear_dd_indices(dd,0,0);
+
+ set_ddbox(dd,bMasterState,cr,ir,state_global->box,
+ TRUE,cgs_gl,state_global->x,&ddbox);
+
+ get_cg_distribution(fplog,step,dd,cgs_gl,
+ state_global->box,&ddbox,state_global->x);
+
+ dd_distribute_state(dd,cgs_gl,
+ state_global,state_local,f);
+
+ dd_make_local_cgs(dd,&top_local->cgs);
+
+ if (dd->ncg_home > fr->cg_nalloc)
+ {
+ dd_realloc_fr_cg(fr,dd->ncg_home);
+ }
+ calc_cgcm(fplog,0,dd->ncg_home,
+ &top_local->cgs,state_local->x,fr->cg_cm);
+
+ inc_nrnb(nrnb,eNR_CGCM,dd->nat_home);
+
+ dd_set_cginfo(dd->index_gl,0,dd->ncg_home,fr,comm->bLocalCG);
+
+ cg0 = 0;
+ }
+ else if (state_local->ddp_count != dd->ddp_count)
+ {
+ if (state_local->ddp_count > dd->ddp_count)
+ {
+ gmx_fatal(FARGS,"Internal inconsistency state_local->ddp_count (%d) > dd->ddp_count (%d)",state_local->ddp_count,dd->ddp_count);
+ }
+
+ if (state_local->ddp_count_cg_gl != state_local->ddp_count)
+ {
+ gmx_fatal(FARGS,"Internal inconsistency state_local->ddp_count_cg_gl (%d) != state_local->ddp_count (%d)",state_local->ddp_count_cg_gl,state_local->ddp_count);
+ }
+
+ /* Clear the old state */
+ clear_dd_indices(dd,0,0);
+
+ /* Build the new indices */
+ rebuild_cgindex(dd,cgs_gl->index,state_local);
+ make_dd_indices(dd,cgs_gl->index,0);
+
+ /* Redetermine the cg COMs */
+ calc_cgcm(fplog,0,dd->ncg_home,
+ &top_local->cgs,state_local->x,fr->cg_cm);
+
+ inc_nrnb(nrnb,eNR_CGCM,dd->nat_home);
+
+ dd_set_cginfo(dd->index_gl,0,dd->ncg_home,fr,comm->bLocalCG);
+
+ set_ddbox(dd,bMasterState,cr,ir,state_local->box,
+ TRUE,&top_local->cgs,state_local->x,&ddbox);
+
+ bRedist = comm->bDynLoadBal;
+ }
+ else
+ {
+ /* We have the full state, only redistribute the cgs */
+
+ /* Clear the non-home indices */
+ clear_dd_indices(dd,dd->ncg_home,dd->nat_home);
+
+ /* Avoid global communication for dim's without pbc and -gcom */
+ if (!bNStGlobalComm)
+ {
+ copy_rvec(comm->box0 ,ddbox.box0 );
+ copy_rvec(comm->box_size,ddbox.box_size);
+ }
+ set_ddbox(dd,bMasterState,cr,ir,state_local->box,
+ bNStGlobalComm,&top_local->cgs,state_local->x,&ddbox);
+
+ bBoxChanged = TRUE;
+ bRedist = TRUE;
+ }
+ /* For dim's without pbc and -gcom */
+ copy_rvec(ddbox.box0 ,comm->box0 );
+ copy_rvec(ddbox.box_size,comm->box_size);
+
+ set_dd_cell_sizes(dd,&ddbox,dynamic_dd_box(&ddbox,ir),bMasterState,bDoDLB,
+ step,wcycle);
+
+ if (comm->nstDDDumpGrid > 0 && step % comm->nstDDDumpGrid == 0)
+ {
+ write_dd_grid_pdb("dd_grid",step,dd,state_local->box,&ddbox);
+ }
+
+ /* Check if we should sort the charge groups */
+ if (comm->nstSortCG > 0)
+ {
+ bSortCG = (bMasterState ||
+ (bRedist && (step % comm->nstSortCG == 0)));
+ }
+ else
+ {
+ bSortCG = FALSE;
+ }
+
+ ncg_home_old = dd->ncg_home;
+
+ if (bRedist)
+ {
+ cg0 = dd_redistribute_cg(fplog,step,dd,ddbox.tric_dir,
+ state_local,f,fr,mdatoms,
+ !bSortCG,nrnb);
+ }
+
+ get_nsgrid_boundaries(fr->ns.grid,dd,
+ state_local->box,&ddbox,&comm->cell_x0,&comm->cell_x1,
+ dd->ncg_home,fr->cg_cm,
+ cell_ns_x0,cell_ns_x1,&grid_density);
+
+ if (bBoxChanged)
+ {
+ comm_dd_ns_cell_sizes(dd,&ddbox,cell_ns_x0,cell_ns_x1,step);
+ }
+
+ copy_ivec(fr->ns.grid->n,ncells_old);
+ grid_first(fplog,fr->ns.grid,dd,&ddbox,fr->ePBC,
+ state_local->box,cell_ns_x0,cell_ns_x1,
+ fr->rlistlong,grid_density);
+ /* We need to store tric_dir for dd_get_ns_ranges called from ns.c */
+ copy_ivec(ddbox.tric_dir,comm->tric_dir);
+
+ if (bSortCG)
+ {
+ /* Sort the state on charge group position.
+ * This enables exact restarts from this step.
+ * It also improves performance by about 15% with larger numbers
+ * of atoms per node.
+ */
+
+ /* Fill the ns grid with the home cell,
+ * so we can sort with the indices.
+ */
+ set_zones_ncg_home(dd);
+ fill_grid(fplog,&comm->zones,fr->ns.grid,dd->ncg_home,
+ 0,dd->ncg_home,fr->cg_cm);
+
+ /* Check if we can user the old order and ns grid cell indices
+ * of the charge groups to sort the charge groups efficiently.
+ */
+ bResortAll = (bMasterState ||
+ fr->ns.grid->n[XX] != ncells_old[XX] ||
+ fr->ns.grid->n[YY] != ncells_old[YY] ||
+ fr->ns.grid->n[ZZ] != ncells_old[ZZ]);
+
+ if (debug)
+ {
+ fprintf(debug,"Step %s, sorting the %d home charge groups\n",
+ gmx_step_str(step,sbuf),dd->ncg_home);
+ }
+ dd_sort_state(dd,ir->ePBC,fr->cg_cm,fr,state_local,
+ bResortAll ? -1 : ncg_home_old);
+ /* Rebuild all the indices */
+ cg0 = 0;
+ ga2la_clear(dd->ga2la);
+ }
+
+ /* Setup up the communication and communicate the coordinates */
+ setup_dd_communication(dd,state_local->box,&ddbox,fr);
+
+ /* Set the indices */
+ make_dd_indices(dd,cgs_gl->index,cg0);
+
+ /* Set the charge group boundaries for neighbor searching */
+ set_cg_boundaries(&comm->zones);
+
+ /*
+ write_dd_pdb("dd_home",step,"dump",top_global,cr,
+ -1,state_local->x,state_local->box);
+ */
+
+ /* Extract a local topology from the global topology */
+ for(i=0; i<dd->ndim; i++)
+ {
+ np[dd->dim[i]] = comm->cd[i].np;
+ }
+ dd_make_local_top(fplog,dd,&comm->zones,dd->npbcdim,state_local->box,
+ comm->cellsize_min,np,
+ fr,vsite,top_global,top_local);
+
+ /* Set up the special atom communication */
+ n = comm->nat[ddnatZONE];
+ for(i=ddnatZONE+1; i<ddnatNR; i++)
+ {
+ switch(i)
+ {
+ case ddnatVSITE:
+ if (vsite && vsite->n_intercg_vsite)
+ {
+ n = dd_make_local_vsites(dd,n,top_local->idef.il);
+ }
+ break;
+ case ddnatCON:
+ if (dd->bInterCGcons)
+ {
+ /* Only for inter-cg constraints we need special code */
+ n = dd_make_local_constraints(dd,n,top_global,
+ constr,ir->nProjOrder,
+ &top_local->idef.il[F_CONSTR]);
+ }
+ break;
+ default:
+ gmx_incons("Unknown special atom type setup");
+ }
+ comm->nat[i] = n;
+ }
+
+ /* Make space for the extra coordinates for virtual site
+ * or constraint communication.
+ */
+ state_local->natoms = comm->nat[ddnatNR-1];
+ if (state_local->natoms > state_local->nalloc)
+ {
+ dd_realloc_state(state_local,f,state_local->natoms);
+ }
+
+ if (fr->bF_NoVirSum)
+ {
+ if (vsite && vsite->n_intercg_vsite)
+ {
+ nat_f_novirsum = comm->nat[ddnatVSITE];
+ }
+ else
+ {
+ if (EEL_FULL(ir->coulombtype) && dd->n_intercg_excl > 0)
+ {
+ nat_f_novirsum = dd->nat_tot;
+ }
+ else
+ {
+ nat_f_novirsum = dd->nat_home;
+ }
+ }
+ }
+ else
+ {
+ nat_f_novirsum = 0;
+ }
+
+ /* Set the number of atoms required for the force calculation.
+ * Forces need to be constrained when using a twin-range setup
+ * or with energy minimization. For simple simulations we could
+ * avoid some allocation, zeroing and copying, but this is
+ * probably not worth the complications ande checking.
+ */
+ forcerec_set_ranges(fr,dd->ncg_home,dd->ncg_tot,
+ dd->nat_tot,comm->nat[ddnatCON],nat_f_novirsum);
+
+ /* We make the all mdatoms up to nat_tot_con.
+ * We could save some work by only setting invmass
+ * between nat_tot and nat_tot_con.
+ */
+ /* This call also sets the new number of home particles to dd->nat_home */
+ atoms2md(top_global,ir,
+ comm->nat[ddnatCON],dd->gatindex,0,dd->nat_home,mdatoms);
+
+ /* Now we have the charges we can sort the FE interactions */
+ dd_sort_local_top(dd,mdatoms,top_local);
+
+ if (shellfc)
+ {
+ /* Make the local shell stuff, currently no communication is done */
+ make_local_shells(cr,mdatoms,shellfc);
+ }
+
+ if (ir->implicit_solvent)
+ {
+ make_local_gb(cr,fr->born,ir->gb_algorithm);
+ }
+
+ if (!(cr->duty & DUTY_PME))
+ {
+ /* Send the charges to our PME only node */
+ gmx_pme_send_q(cr,mdatoms->nChargePerturbed,
+ mdatoms->chargeA,mdatoms->chargeB,
+ dd_pme_maxshift_x(dd),dd_pme_maxshift_y(dd));
+ }
+
+ if (constr)
+ {
+ set_constraints(constr,top_local,ir,mdatoms,cr);
+ }
+
+ if (ir->ePull != epullNO)
+ {
+ /* Update the local pull groups */
+ dd_make_local_pull_groups(dd,ir->pull,mdatoms);
+ }
+
+ if (ir->bRot)
+ {
+ /* Update the local rotation groups */
+ dd_make_local_rotation_groups(dd,ir->rot);
+ }
+
+
+ add_dd_statistics(dd);
+
+ /* Make sure we only count the cycles for this DD partitioning */
+ clear_dd_cycle_counts(dd);
+
+ /* Because the order of the atoms might have changed since
+ * the last vsite construction, we need to communicate the constructing
+ * atom coordinates again (for spreading the forces this MD step).
+ */
+ dd_move_x_vsites(dd,state_local->box,state_local->x);
+
+ if (comm->nstDDDump > 0 && step % comm->nstDDDump == 0)
+ {
+ dd_move_x(dd,state_local->box,state_local->x);
+ write_dd_pdb("dd_dump",step,"dump",top_global,cr,
+ -1,state_local->x,state_local->box);
+ }
+
+ if (bNStGlobalComm)
+ {
+ /* Store the global communication step */
+ comm->globalcomm_step = step;
+ }
+
+ /* Increase the DD partitioning counter */
+ dd->ddp_count++;
+ /* The state currently matches this DD partitioning count, store it */
+ state_local->ddp_count = dd->ddp_count;
+ if (bMasterState)
+ {
+ /* The DD master node knows the complete cg distribution,
+ * store the count so we can possibly skip the cg info communication.
+ */
+ comm->master_cg_ddp_count = (bSortCG ? 0 : dd->ddp_count);
+ }
+
+ if (comm->DD_debug > 0)
+ {
+ /* Set the env var GMX_DD_DEBUG if you suspect corrupted indices */
+ check_index_consistency(dd,top_global->natoms,ncg_mtop(top_global),
+ "after partitioning");
+ }
+}
--- /dev/null
- #ifdef GMX_THREADS
+/* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
+ *
+ *
+ * This source code is part of
+ *
+ * G R O M A C S
+ *
+ * GROningen MAchine for Chemical Simulations
+ *
+ * VERSION 3.2.0
+ * Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
+ * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
+ * Copyright (c) 2001-2004, The GROMACS development team,
+ * check out http://www.gromacs.org for more information.
+
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version 2
+ * of the License, or (at your option) any later version.
+ *
+ * If you want to redistribute modifications, please consider that
+ * scientific software is very special. Version control is crucial -
+ * bugs must be traceable. We will be happy to consider code for
+ * inclusion in the official distribution, but derived work must not
+ * be called official GROMACS. Details are found in the README & COPYING
+ * files - if they are missing, get the official version at www.gromacs.org.
+ *
+ * To help us fund GROMACS development, we humbly ask that you cite
+ * the papers on the package - you can find them in the top README file.
+ *
+ * For more info, check our website at http://www.gromacs.org
+ *
+ * And Hey:
+ * GROwing Monsters And Cloning Shrimps
+ */
+/* IMPORTANT FOR DEVELOPERS:
+ *
+ * Triclinic pme stuff isn't entirely trivial, and we've experienced
+ * some bugs during development (many of them due to me). To avoid
+ * this in the future, please check the following things if you make
+ * changes in this file:
+ *
+ * 1. You should obtain identical (at least to the PME precision)
+ * energies, forces, and virial for
+ * a rectangular box and a triclinic one where the z (or y) axis is
+ * tilted a whole box side. For instance you could use these boxes:
+ *
+ * rectangular triclinic
+ * 2 0 0 2 0 0
+ * 0 2 0 0 2 0
+ * 0 0 6 2 2 6
+ *
+ * 2. You should check the energy conservation in a triclinic box.
+ *
+ * It might seem an overkill, but better safe than sorry.
+ * /Erik 001109
+ */
+
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+#ifdef GMX_LIB_MPI
+#include <mpi.h>
+#endif
++#ifdef GMX_THREAD_MPI
+#include "tmpi.h"
+#endif
+
+#ifdef GMX_OPENMP
+#include <omp.h>
+#endif
+
+#include <stdio.h>
+#include <string.h>
+#include <math.h>
+#include "typedefs.h"
+#include "txtdump.h"
+#include "vec.h"
+#include "gmxcomplex.h"
+#include "smalloc.h"
+#include "futil.h"
+#include "coulomb.h"
+#include "gmx_fatal.h"
+#include "pme.h"
+#include "network.h"
+#include "physics.h"
+#include "nrnb.h"
+#include "copyrite.h"
+#include "gmx_wallcycle.h"
+#include "gmx_parallel_3dfft.h"
+#include "pdbio.h"
+#include "gmx_cyclecounter.h"
+#include "macros.h"
+
+#if ( !defined(GMX_DOUBLE) && ( defined(GMX_IA32_SSE) || defined(GMX_X86_64_SSE) || defined(GMX_X86_64_SSE2) ) )
+#include "gmx_sse2_single.h"
+
+#define PME_SSE
+/* Some old AMD processors could have problems with unaligned loads+stores */
+#ifndef GMX_FAHCORE
+#define PME_SSE_UNALIGNED
+#endif
+#endif
+
+#define DFT_TOL 1e-7
+/* #define PRT_FORCE */
+/* conditions for on the fly time-measurement */
+/* #define TAKETIME (step > 1 && timesteps < 10) */
+#define TAKETIME FALSE
+
+/* #define PME_TIME_THREADS */
+
+#ifdef GMX_DOUBLE
+#define mpi_type MPI_DOUBLE
+#else
+#define mpi_type MPI_FLOAT
+#endif
+
+/* GMX_CACHE_SEP should be a multiple of 16 to preserve alignment */
+#define GMX_CACHE_SEP 64
+
+/* We only define a maximum to be able to use local arrays without allocation.
+ * An order larger than 12 should never be needed, even for test cases.
+ * If needed it can be changed here.
+ */
+#define PME_ORDER_MAX 12
+
+/* Internal datastructures */
+typedef struct {
+ int send_index0;
+ int send_nindex;
+ int recv_index0;
+ int recv_nindex;
+} pme_grid_comm_t;
+
+typedef struct {
+#ifdef GMX_MPI
+ MPI_Comm mpi_comm;
+#endif
+ int nnodes,nodeid;
+ int *s2g0;
+ int *s2g1;
+ int noverlap_nodes;
+ int *send_id,*recv_id;
+ pme_grid_comm_t *comm_data;
+ real *sendbuf;
+ real *recvbuf;
+} pme_overlap_t;
+
+typedef struct {
+ int *n; /* Cumulative counts of the number of particles per thread */
+ int nalloc; /* Allocation size of i */
+ int *i; /* Particle indices ordered on thread index (n) */
+} thread_plist_t;
+
+typedef struct {
+ int n;
+ int *ind;
+ splinevec theta;
+ splinevec dtheta;
+} splinedata_t;
+
+typedef struct {
+ int dimind; /* The index of the dimension, 0=x, 1=y */
+ int nslab;
+ int nodeid;
+#ifdef GMX_MPI
+ MPI_Comm mpi_comm;
+#endif
+
+ int *node_dest; /* The nodes to send x and q to with DD */
+ int *node_src; /* The nodes to receive x and q from with DD */
+ int *buf_index; /* Index for commnode into the buffers */
+
+ int maxshift;
+
+ int npd;
+ int pd_nalloc;
+ int *pd;
+ int *count; /* The number of atoms to send to each node */
+ int **count_thread;
+ int *rcount; /* The number of atoms to receive */
+
+ int n;
+ int nalloc;
+ rvec *x;
+ real *q;
+ rvec *f;
+ gmx_bool bSpread; /* These coordinates are used for spreading */
+ int pme_order;
+ ivec *idx;
+ rvec *fractx; /* Fractional coordinate relative to the
+ * lower cell boundary
+ */
+ int nthread;
+ int *thread_idx; /* Which thread should spread which charge */
+ thread_plist_t *thread_plist;
+ splinedata_t *spline;
+} pme_atomcomm_t;
+
+#define FLBS 3
+#define FLBSZ 4
+
+typedef struct {
+ ivec ci; /* The spatial location of this grid */
+ ivec n; /* The size of *grid, including order-1 */
+ ivec offset; /* The grid offset from the full node grid */
+ int order; /* PME spreading order */
+ real *grid; /* The grid local thread, size n */
+} pmegrid_t;
+
+typedef struct {
+ pmegrid_t grid; /* The full node grid (non thread-local) */
+ int nthread; /* The number of threads operating on this grid */
+ ivec nc; /* The local spatial decomposition over the threads */
+ pmegrid_t *grid_th; /* Array of grids for each thread */
+ int **g2t; /* The grid to thread index */
+ ivec nthread_comm; /* The number of threads to communicate with */
+} pmegrids_t;
+
+
+typedef struct {
+#ifdef PME_SSE
+ /* Masks for SSE aligned spreading and gathering */
+ __m128 mask_SSE0[6],mask_SSE1[6];
+#else
+ int dummy; /* C89 requires that struct has at least one member */
+#endif
+} pme_spline_work_t;
+
+typedef struct {
+ /* work data for solve_pme */
+ int nalloc;
+ real * mhx;
+ real * mhy;
+ real * mhz;
+ real * m2;
+ real * denom;
+ real * tmp1_alloc;
+ real * tmp1;
+ real * eterm;
+ real * m2inv;
+
+ real energy;
+ matrix vir;
+} pme_work_t;
+
+typedef struct gmx_pme {
+ int ndecompdim; /* The number of decomposition dimensions */
+ int nodeid; /* Our nodeid in mpi->mpi_comm */
+ int nodeid_major;
+ int nodeid_minor;
+ int nnodes; /* The number of nodes doing PME */
+ int nnodes_major;
+ int nnodes_minor;
+
+ MPI_Comm mpi_comm;
+ MPI_Comm mpi_comm_d[2]; /* Indexed on dimension, 0=x, 1=y */
+#ifdef GMX_MPI
+ MPI_Datatype rvec_mpi; /* the pme vector's MPI type */
+#endif
+
+ int nthread; /* The number of threads doing PME */
+
+ gmx_bool bPPnode; /* Node also does particle-particle forces */
+ gmx_bool bFEP; /* Compute Free energy contribution */
+ int nkx,nky,nkz; /* Grid dimensions */
+ int pme_order;
+ real epsilon_r;
+
+ pmegrids_t pmegridA; /* Grids on which we do spreading/interpolation, includes overlap */
+ pmegrids_t pmegridB;
+ /* The PME charge spreading grid sizes/strides, includes pme_order-1 */
+ int pmegrid_nx,pmegrid_ny,pmegrid_nz;
+ /* pmegrid_nz might be larger than strictly necessary to ensure
+ * memory alignment, pmegrid_nz_base gives the real base size.
+ */
+ int pmegrid_nz_base;
+ /* The local PME grid starting indices */
+ int pmegrid_start_ix,pmegrid_start_iy,pmegrid_start_iz;
+
+ /* Work data for spreading and gathering */
+ pme_spline_work_t spline_work;
+
+ real *fftgridA; /* Grids for FFT. With 1D FFT decomposition this can be a pointer */
+ real *fftgridB; /* inside the interpolation grid, but separate for 2D PME decomp. */
+ int fftgrid_nx,fftgrid_ny,fftgrid_nz;
+
+ t_complex *cfftgridA; /* Grids for complex FFT data */
+ t_complex *cfftgridB;
+ int cfftgrid_nx,cfftgrid_ny,cfftgrid_nz;
+
+ gmx_parallel_3dfft_t pfft_setupA;
+ gmx_parallel_3dfft_t pfft_setupB;
+
+ int *nnx,*nny,*nnz;
+ real *fshx,*fshy,*fshz;
+
+ pme_atomcomm_t atc[2]; /* Indexed on decomposition index */
+ matrix recipbox;
+ splinevec bsp_mod;
+
+ pme_overlap_t overlap[2]; /* Indexed on dimension, 0=x, 1=y */
+
+ pme_atomcomm_t atc_energy; /* Only for gmx_pme_calc_energy */
+
+ rvec *bufv; /* Communication buffer */
+ real *bufr; /* Communication buffer */
+ int buf_nalloc; /* The communication buffer size */
+
+ /* thread local work data for solve_pme */
+ pme_work_t *work;
+
+ /* Work data for PME_redist */
+ gmx_bool redist_init;
+ int * scounts;
+ int * rcounts;
+ int * sdispls;
+ int * rdispls;
+ int * sidx;
+ int * idxa;
+ real * redist_buf;
+ int redist_buf_nalloc;
+
+ /* Work data for sum_qgrid */
+ real * sum_qgrid_tmp;
+ real * sum_qgrid_dd_tmp;
+} t_gmx_pme;
+
+
+static void calc_interpolation_idx(gmx_pme_t pme,pme_atomcomm_t *atc,
+ int start,int end,int thread)
+{
+ int i;
+ int *idxptr,tix,tiy,tiz;
+ real *xptr,*fptr,tx,ty,tz;
+ real rxx,ryx,ryy,rzx,rzy,rzz;
+ int nx,ny,nz;
+ int start_ix,start_iy,start_iz;
+ int *g2tx,*g2ty,*g2tz;
+ gmx_bool bThreads;
+ int *thread_idx=NULL;
+ thread_plist_t *tpl=NULL;
+ int *tpl_n=NULL;
+ int thread_i;
+
+ nx = pme->nkx;
+ ny = pme->nky;
+ nz = pme->nkz;
+
+ start_ix = pme->pmegrid_start_ix;
+ start_iy = pme->pmegrid_start_iy;
+ start_iz = pme->pmegrid_start_iz;
+
+ rxx = pme->recipbox[XX][XX];
+ ryx = pme->recipbox[YY][XX];
+ ryy = pme->recipbox[YY][YY];
+ rzx = pme->recipbox[ZZ][XX];
+ rzy = pme->recipbox[ZZ][YY];
+ rzz = pme->recipbox[ZZ][ZZ];
+
+ g2tx = pme->pmegridA.g2t[XX];
+ g2ty = pme->pmegridA.g2t[YY];
+ g2tz = pme->pmegridA.g2t[ZZ];
+
+ bThreads = (atc->nthread > 1);
+ if (bThreads)
+ {
+ thread_idx = atc->thread_idx;
+
+ tpl = &atc->thread_plist[thread];
+ tpl_n = tpl->n;
+ for(i=0; i<atc->nthread; i++)
+ {
+ tpl_n[i] = 0;
+ }
+ }
+
+ for(i=start; i<end; i++) {
+ xptr = atc->x[i];
+ idxptr = atc->idx[i];
+ fptr = atc->fractx[i];
+
+ /* Fractional coordinates along box vectors, add 2.0 to make 100% sure we are positive for triclinic boxes */
+ tx = nx * ( xptr[XX] * rxx + xptr[YY] * ryx + xptr[ZZ] * rzx + 2.0 );
+ ty = ny * ( xptr[YY] * ryy + xptr[ZZ] * rzy + 2.0 );
+ tz = nz * ( xptr[ZZ] * rzz + 2.0 );
+
+ tix = (int)(tx);
+ tiy = (int)(ty);
+ tiz = (int)(tz);
+
+ /* Because decomposition only occurs in x and y,
+ * we never have a fraction correction in z.
+ */
+ fptr[XX] = tx - tix + pme->fshx[tix];
+ fptr[YY] = ty - tiy + pme->fshy[tiy];
+ fptr[ZZ] = tz - tiz;
+
+ idxptr[XX] = pme->nnx[tix];
+ idxptr[YY] = pme->nny[tiy];
+ idxptr[ZZ] = pme->nnz[tiz];
+
+#ifdef DEBUG
+ range_check(idxptr[XX],0,pme->pmegrid_nx);
+ range_check(idxptr[YY],0,pme->pmegrid_ny);
+ range_check(idxptr[ZZ],0,pme->pmegrid_nz);
+#endif
+
+ if (bThreads)
+ {
+ thread_i = g2tx[idxptr[XX]] + g2ty[idxptr[YY]] + g2tz[idxptr[ZZ]];
+ thread_idx[i] = thread_i;
+ tpl_n[thread_i]++;
+ }
+ }
+
+ if (bThreads)
+ {
+ /* Make a list of particle indices sorted on thread */
+
+ /* Get the cumulative count */
+ for(i=1; i<atc->nthread; i++)
+ {
+ tpl_n[i] += tpl_n[i-1];
+ }
+ /* The current implementation distributes particles equally
+ * over the threads, so we could actually allocate for that
+ * in pme_realloc_atomcomm_things.
+ */
+ if (tpl_n[atc->nthread-1] > tpl->nalloc)
+ {
+ tpl->nalloc = over_alloc_large(tpl_n[atc->nthread-1]);
+ srenew(tpl->i,tpl->nalloc);
+ }
+ /* Set tpl_n to the cumulative start */
+ for(i=atc->nthread-1; i>=1; i--)
+ {
+ tpl_n[i] = tpl_n[i-1];
+ }
+ tpl_n[0] = 0;
+
+ /* Fill our thread local array with indices sorted on thread */
+ for(i=start; i<end; i++)
+ {
+ tpl->i[tpl_n[atc->thread_idx[i]]++] = i;
+ }
+ /* Now tpl_n contains the cummulative count again */
+ }
+}
+
+static void make_thread_local_ind(pme_atomcomm_t *atc,
+ int thread,splinedata_t *spline)
+{
+ int n,t,i,start,end;
+ thread_plist_t *tpl;
+
+ /* Combine the indices made by each thread into one index */
+
+ n = 0;
+ start = 0;
+ for(t=0; t<atc->nthread; t++)
+ {
+ tpl = &atc->thread_plist[t];
+ /* Copy our part (start - end) from the list of thread t */
+ if (thread > 0)
+ {
+ start = tpl->n[thread-1];
+ }
+ end = tpl->n[thread];
+ for(i=start; i<end; i++)
+ {
+ spline->ind[n++] = tpl->i[i];
+ }
+ }
+
+ spline->n = n;
+}
+
+
+static void pme_calc_pidx(int start, int end,
+ matrix recipbox, rvec x[],
+ pme_atomcomm_t *atc, int *count)
+{
+ int nslab,i;
+ int si;
+ real *xptr,s;
+ real rxx,ryx,rzx,ryy,rzy;
+ int *pd;
+
+ /* Calculate PME task index (pidx) for each grid index.
+ * Here we always assign equally sized slabs to each node
+ * for load balancing reasons (the PME grid spacing is not used).
+ */
+
+ nslab = atc->nslab;
+ pd = atc->pd;
+
+ /* Reset the count */
+ for(i=0; i<nslab; i++)
+ {
+ count[i] = 0;
+ }
+
+ if (atc->dimind == 0)
+ {
+ rxx = recipbox[XX][XX];
+ ryx = recipbox[YY][XX];
+ rzx = recipbox[ZZ][XX];
+ /* Calculate the node index in x-dimension */
+ for(i=start; i<end; i++)
+ {
+ xptr = x[i];
+ /* Fractional coordinates along box vectors */
+ s = nslab*(xptr[XX]*rxx + xptr[YY]*ryx + xptr[ZZ]*rzx);
+ si = (int)(s + 2*nslab) % nslab;
+ pd[i] = si;
+ count[si]++;
+ }
+ }
+ else
+ {
+ ryy = recipbox[YY][YY];
+ rzy = recipbox[ZZ][YY];
+ /* Calculate the node index in y-dimension */
+ for(i=start; i<end; i++)
+ {
+ xptr = x[i];
+ /* Fractional coordinates along box vectors */
+ s = nslab*(xptr[YY]*ryy + xptr[ZZ]*rzy);
+ si = (int)(s + 2*nslab) % nslab;
+ pd[i] = si;
+ count[si]++;
+ }
+ }
+}
+
+static void pme_calc_pidx_wrapper(int natoms, matrix recipbox, rvec x[],
+ pme_atomcomm_t *atc)
+{
+ int nthread,thread,slab;
+
+ nthread = atc->nthread;
+
+#pragma omp parallel for num_threads(nthread) schedule(static)
+ for(thread=0; thread<nthread; thread++)
+ {
+ pme_calc_pidx(natoms* thread /nthread,
+ natoms*(thread+1)/nthread,
+ recipbox,x,atc,atc->count_thread[thread]);
+ }
+ /* Non-parallel reduction, since nslab is small */
+
+ for(thread=1; thread<nthread; thread++)
+ {
+ for(slab=0; slab<atc->nslab; slab++)
+ {
+ atc->count_thread[0][slab] += atc->count_thread[thread][slab];
+ }
+ }
+}
+
+static void pme_realloc_splinedata(splinedata_t *spline, pme_atomcomm_t *atc)
+{
+ int i,d;
+
+ srenew(spline->ind,atc->nalloc);
+ /* Initialize the index to identity so it works without threads */
+ for(i=0; i<atc->nalloc; i++)
+ {
+ spline->ind[i] = i;
+ }
+
+ for(d=0;d<DIM;d++)
+ {
+ srenew(spline->theta[d] ,atc->pme_order*atc->nalloc);
+ srenew(spline->dtheta[d],atc->pme_order*atc->nalloc);
+ }
+}
+
+static void pme_realloc_atomcomm_things(pme_atomcomm_t *atc)
+{
+ int nalloc_old,i,j,nalloc_tpl;
+
+ /* We have to avoid a NULL pointer for atc->x to avoid
+ * possible fatal errors in MPI routines.
+ */
+ if (atc->n > atc->nalloc || atc->nalloc == 0)
+ {
+ nalloc_old = atc->nalloc;
+ atc->nalloc = over_alloc_dd(max(atc->n,1));
+
+ if (atc->nslab > 1) {
+ srenew(atc->x,atc->nalloc);
+ srenew(atc->q,atc->nalloc);
+ srenew(atc->f,atc->nalloc);
+ for(i=nalloc_old; i<atc->nalloc; i++)
+ {
+ clear_rvec(atc->f[i]);
+ }
+ }
+ if (atc->bSpread) {
+ srenew(atc->fractx,atc->nalloc);
+ srenew(atc->idx ,atc->nalloc);
+
+ if (atc->nthread > 1)
+ {
+ srenew(atc->thread_idx,atc->nalloc);
+ }
+
+ for(i=0; i<atc->nthread; i++)
+ {
+ pme_realloc_splinedata(&atc->spline[i],atc);
+ }
+ }
+ }
+}
+
+static void pmeredist_pd(gmx_pme_t pme, gmx_bool forw,
+ int n, gmx_bool bXF, rvec *x_f, real *charge,
+ pme_atomcomm_t *atc)
+/* Redistribute particle data for PME calculation */
+/* domain decomposition by x coordinate */
+{
+ int *idxa;
+ int i, ii;
+
+ if(FALSE == pme->redist_init) {
+ snew(pme->scounts,atc->nslab);
+ snew(pme->rcounts,atc->nslab);
+ snew(pme->sdispls,atc->nslab);
+ snew(pme->rdispls,atc->nslab);
+ snew(pme->sidx,atc->nslab);
+ pme->redist_init = TRUE;
+ }
+ if (n > pme->redist_buf_nalloc) {
+ pme->redist_buf_nalloc = over_alloc_dd(n);
+ srenew(pme->redist_buf,pme->redist_buf_nalloc*DIM);
+ }
+
+ pme->idxa = atc->pd;
+
+#ifdef GMX_MPI
+ if (forw && bXF) {
+ /* forward, redistribution from pp to pme */
+
+ /* Calculate send counts and exchange them with other nodes */
+ for(i=0; (i<atc->nslab); i++) pme->scounts[i]=0;
+ for(i=0; (i<n); i++) pme->scounts[pme->idxa[i]]++;
+ MPI_Alltoall( pme->scounts, 1, MPI_INT, pme->rcounts, 1, MPI_INT, atc->mpi_comm);
+
+ /* Calculate send and receive displacements and index into send
+ buffer */
+ pme->sdispls[0]=0;
+ pme->rdispls[0]=0;
+ pme->sidx[0]=0;
+ for(i=1; i<atc->nslab; i++) {
+ pme->sdispls[i]=pme->sdispls[i-1]+pme->scounts[i-1];
+ pme->rdispls[i]=pme->rdispls[i-1]+pme->rcounts[i-1];
+ pme->sidx[i]=pme->sdispls[i];
+ }
+ /* Total # of particles to be received */
+ atc->n = pme->rdispls[atc->nslab-1] + pme->rcounts[atc->nslab-1];
+
+ pme_realloc_atomcomm_things(atc);
+
+ /* Copy particle coordinates into send buffer and exchange*/
+ for(i=0; (i<n); i++) {
+ ii=DIM*pme->sidx[pme->idxa[i]];
+ pme->sidx[pme->idxa[i]]++;
+ pme->redist_buf[ii+XX]=x_f[i][XX];
+ pme->redist_buf[ii+YY]=x_f[i][YY];
+ pme->redist_buf[ii+ZZ]=x_f[i][ZZ];
+ }
+ MPI_Alltoallv(pme->redist_buf, pme->scounts, pme->sdispls,
+ pme->rvec_mpi, atc->x, pme->rcounts, pme->rdispls,
+ pme->rvec_mpi, atc->mpi_comm);
+ }
+ if (forw) {
+ /* Copy charge into send buffer and exchange*/
+ for(i=0; i<atc->nslab; i++) pme->sidx[i]=pme->sdispls[i];
+ for(i=0; (i<n); i++) {
+ ii=pme->sidx[pme->idxa[i]];
+ pme->sidx[pme->idxa[i]]++;
+ pme->redist_buf[ii]=charge[i];
+ }
+ MPI_Alltoallv(pme->redist_buf, pme->scounts, pme->sdispls, mpi_type,
+ atc->q, pme->rcounts, pme->rdispls, mpi_type,
+ atc->mpi_comm);
+ }
+ else { /* backward, redistribution from pme to pp */
+ MPI_Alltoallv(atc->f, pme->rcounts, pme->rdispls, pme->rvec_mpi,
+ pme->redist_buf, pme->scounts, pme->sdispls,
+ pme->rvec_mpi, atc->mpi_comm);
+
+ /* Copy data from receive buffer */
+ for(i=0; i<atc->nslab; i++)
+ pme->sidx[i] = pme->sdispls[i];
+ for(i=0; (i<n); i++) {
+ ii = DIM*pme->sidx[pme->idxa[i]];
+ x_f[i][XX] += pme->redist_buf[ii+XX];
+ x_f[i][YY] += pme->redist_buf[ii+YY];
+ x_f[i][ZZ] += pme->redist_buf[ii+ZZ];
+ pme->sidx[pme->idxa[i]]++;
+ }
+ }
+#endif
+}
+
+static void pme_dd_sendrecv(pme_atomcomm_t *atc,
+ gmx_bool bBackward,int shift,
+ void *buf_s,int nbyte_s,
+ void *buf_r,int nbyte_r)
+{
+#ifdef GMX_MPI
+ int dest,src;
+ MPI_Status stat;
+
+ if (bBackward == FALSE) {
+ dest = atc->node_dest[shift];
+ src = atc->node_src[shift];
+ } else {
+ dest = atc->node_src[shift];
+ src = atc->node_dest[shift];
+ }
+
+ if (nbyte_s > 0 && nbyte_r > 0) {
+ MPI_Sendrecv(buf_s,nbyte_s,MPI_BYTE,
+ dest,shift,
+ buf_r,nbyte_r,MPI_BYTE,
+ src,shift,
+ atc->mpi_comm,&stat);
+ } else if (nbyte_s > 0) {
+ MPI_Send(buf_s,nbyte_s,MPI_BYTE,
+ dest,shift,
+ atc->mpi_comm);
+ } else if (nbyte_r > 0) {
+ MPI_Recv(buf_r,nbyte_r,MPI_BYTE,
+ src,shift,
+ atc->mpi_comm,&stat);
+ }
+#endif
+}
+
+static void dd_pmeredist_x_q(gmx_pme_t pme,
+ int n, gmx_bool bX, rvec *x, real *charge,
+ pme_atomcomm_t *atc)
+{
+ int *commnode,*buf_index;
+ int nnodes_comm,i,nsend,local_pos,buf_pos,node,scount,rcount;
+
+ commnode = atc->node_dest;
+ buf_index = atc->buf_index;
+
+ nnodes_comm = min(2*atc->maxshift,atc->nslab-1);
+
+ nsend = 0;
+ for(i=0; i<nnodes_comm; i++) {
+ buf_index[commnode[i]] = nsend;
+ nsend += atc->count[commnode[i]];
+ }
+ if (bX) {
+ if (atc->count[atc->nodeid] + nsend != n)
+ gmx_fatal(FARGS,"%d particles communicated to PME node %d are more than 2/3 times the cut-off out of the domain decomposition cell of their charge group in dimension %c.\n"
+ "This usually means that your system is not well equilibrated.",
+ n - (atc->count[atc->nodeid] + nsend),
+ pme->nodeid,'x'+atc->dimind);
+
+ if (nsend > pme->buf_nalloc) {
+ pme->buf_nalloc = over_alloc_dd(nsend);
+ srenew(pme->bufv,pme->buf_nalloc);
+ srenew(pme->bufr,pme->buf_nalloc);
+ }
+
+ atc->n = atc->count[atc->nodeid];
+ for(i=0; i<nnodes_comm; i++) {
+ scount = atc->count[commnode[i]];
+ /* Communicate the count */
+ if (debug)
+ fprintf(debug,"dimind %d PME node %d send to node %d: %d\n",
+ atc->dimind,atc->nodeid,commnode[i],scount);
+ pme_dd_sendrecv(atc,FALSE,i,
+ &scount,sizeof(int),
+ &atc->rcount[i],sizeof(int));
+ atc->n += atc->rcount[i];
+ }
+
+ pme_realloc_atomcomm_things(atc);
+ }
+
+ local_pos = 0;
+ for(i=0; i<n; i++) {
+ node = atc->pd[i];
+ if (node == atc->nodeid) {
+ /* Copy direct to the receive buffer */
+ if (bX) {
+ copy_rvec(x[i],atc->x[local_pos]);
+ }
+ atc->q[local_pos] = charge[i];
+ local_pos++;
+ } else {
+ /* Copy to the send buffer */
+ if (bX) {
+ copy_rvec(x[i],pme->bufv[buf_index[node]]);
+ }
+ pme->bufr[buf_index[node]] = charge[i];
+ buf_index[node]++;
+ }
+ }
+
+ buf_pos = 0;
+ for(i=0; i<nnodes_comm; i++) {
+ scount = atc->count[commnode[i]];
+ rcount = atc->rcount[i];
+ if (scount > 0 || rcount > 0) {
+ if (bX) {
+ /* Communicate the coordinates */
+ pme_dd_sendrecv(atc,FALSE,i,
+ pme->bufv[buf_pos],scount*sizeof(rvec),
+ atc->x[local_pos],rcount*sizeof(rvec));
+ }
+ /* Communicate the charges */
+ pme_dd_sendrecv(atc,FALSE,i,
+ pme->bufr+buf_pos,scount*sizeof(real),
+ atc->q+local_pos,rcount*sizeof(real));
+ buf_pos += scount;
+ local_pos += atc->rcount[i];
+ }
+ }
+}
+
+static void dd_pmeredist_f(gmx_pme_t pme, pme_atomcomm_t *atc,
+ int n, rvec *f,
+ gmx_bool bAddF)
+{
+ int *commnode,*buf_index;
+ int nnodes_comm,local_pos,buf_pos,i,scount,rcount,node;
+
+ commnode = atc->node_dest;
+ buf_index = atc->buf_index;
+
+ nnodes_comm = min(2*atc->maxshift,atc->nslab-1);
+
+ local_pos = atc->count[atc->nodeid];
+ buf_pos = 0;
+ for(i=0; i<nnodes_comm; i++) {
+ scount = atc->rcount[i];
+ rcount = atc->count[commnode[i]];
+ if (scount > 0 || rcount > 0) {
+ /* Communicate the forces */
+ pme_dd_sendrecv(atc,TRUE,i,
+ atc->f[local_pos],scount*sizeof(rvec),
+ pme->bufv[buf_pos],rcount*sizeof(rvec));
+ local_pos += scount;
+ }
+ buf_index[commnode[i]] = buf_pos;
+ buf_pos += rcount;
+ }
+
+ local_pos = 0;
+ if (bAddF)
+ {
+ for(i=0; i<n; i++)
+ {
+ node = atc->pd[i];
+ if (node == atc->nodeid)
+ {
+ /* Add from the local force array */
+ rvec_inc(f[i],atc->f[local_pos]);
+ local_pos++;
+ }
+ else
+ {
+ /* Add from the receive buffer */
+ rvec_inc(f[i],pme->bufv[buf_index[node]]);
+ buf_index[node]++;
+ }
+ }
+ }
+ else
+ {
+ for(i=0; i<n; i++)
+ {
+ node = atc->pd[i];
+ if (node == atc->nodeid)
+ {
+ /* Copy from the local force array */
+ copy_rvec(atc->f[local_pos],f[i]);
+ local_pos++;
+ }
+ else
+ {
+ /* Copy from the receive buffer */
+ copy_rvec(pme->bufv[buf_index[node]],f[i]);
+ buf_index[node]++;
+ }
+ }
+ }
+}
+
+#ifdef GMX_MPI
+static void
+gmx_sum_qgrid_dd(gmx_pme_t pme, real *grid, int direction)
+{
+ pme_overlap_t *overlap;
+ int send_index0,send_nindex;
+ int recv_index0,recv_nindex;
+ MPI_Status stat;
+ int i,j,k,ix,iy,iz,icnt;
+ int ipulse,send_id,recv_id,datasize;
+ real *p;
+ real *sendptr,*recvptr;
+
+ /* Start with minor-rank communication. This is a bit of a pain since it is not contiguous */
+ overlap = &pme->overlap[1];
+
+ for(ipulse=0;ipulse<overlap->noverlap_nodes;ipulse++)
+ {
+ /* Since we have already (un)wrapped the overlap in the z-dimension,
+ * we only have to communicate 0 to nkz (not pmegrid_nz).
+ */
+ if (direction==GMX_SUM_QGRID_FORWARD)
+ {
+ send_id = overlap->send_id[ipulse];
+ recv_id = overlap->recv_id[ipulse];
+ send_index0 = overlap->comm_data[ipulse].send_index0;
+ send_nindex = overlap->comm_data[ipulse].send_nindex;
+ recv_index0 = overlap->comm_data[ipulse].recv_index0;
+ recv_nindex = overlap->comm_data[ipulse].recv_nindex;
+ }
+ else
+ {
+ send_id = overlap->recv_id[ipulse];
+ recv_id = overlap->send_id[ipulse];
+ send_index0 = overlap->comm_data[ipulse].recv_index0;
+ send_nindex = overlap->comm_data[ipulse].recv_nindex;
+ recv_index0 = overlap->comm_data[ipulse].send_index0;
+ recv_nindex = overlap->comm_data[ipulse].send_nindex;
+ }
+
+ /* Copy data to contiguous send buffer */
+ if (debug)
+ {
+ fprintf(debug,"PME send node %d %d -> %d grid start %d Communicating %d to %d\n",
+ pme->nodeid,overlap->nodeid,send_id,
+ pme->pmegrid_start_iy,
+ send_index0-pme->pmegrid_start_iy,
+ send_index0-pme->pmegrid_start_iy+send_nindex);
+ }
+ icnt = 0;
+ for(i=0;i<pme->pmegrid_nx;i++)
+ {
+ ix = i;
+ for(j=0;j<send_nindex;j++)
+ {
+ iy = j + send_index0 - pme->pmegrid_start_iy;
+ for(k=0;k<pme->nkz;k++)
+ {
+ iz = k;
+ overlap->sendbuf[icnt++] = grid[ix*(pme->pmegrid_ny*pme->pmegrid_nz)+iy*(pme->pmegrid_nz)+iz];
+ }
+ }
+ }
+
+ datasize = pme->pmegrid_nx * pme->nkz;
+
+ MPI_Sendrecv(overlap->sendbuf,send_nindex*datasize,GMX_MPI_REAL,
+ send_id,ipulse,
+ overlap->recvbuf,recv_nindex*datasize,GMX_MPI_REAL,
+ recv_id,ipulse,
+ overlap->mpi_comm,&stat);
+
+ /* Get data from contiguous recv buffer */
+ if (debug)
+ {
+ fprintf(debug,"PME recv node %d %d <- %d grid start %d Communicating %d to %d\n",
+ pme->nodeid,overlap->nodeid,recv_id,
+ pme->pmegrid_start_iy,
+ recv_index0-pme->pmegrid_start_iy,
+ recv_index0-pme->pmegrid_start_iy+recv_nindex);
+ }
+ icnt = 0;
+ for(i=0;i<pme->pmegrid_nx;i++)
+ {
+ ix = i;
+ for(j=0;j<recv_nindex;j++)
+ {
+ iy = j + recv_index0 - pme->pmegrid_start_iy;
+ for(k=0;k<pme->nkz;k++)
+ {
+ iz = k;
+ if(direction==GMX_SUM_QGRID_FORWARD)
+ {
+ grid[ix*(pme->pmegrid_ny*pme->pmegrid_nz)+iy*(pme->pmegrid_nz)+iz] += overlap->recvbuf[icnt++];
+ }
+ else
+ {
+ grid[ix*(pme->pmegrid_ny*pme->pmegrid_nz)+iy*(pme->pmegrid_nz)+iz] = overlap->recvbuf[icnt++];
+ }
+ }
+ }
+ }
+ }
+
+ /* Major dimension is easier, no copying required,
+ * but we might have to sum to separate array.
+ * Since we don't copy, we have to communicate up to pmegrid_nz,
+ * not nkz as for the minor direction.
+ */
+ overlap = &pme->overlap[0];
+
+ for(ipulse=0;ipulse<overlap->noverlap_nodes;ipulse++)
+ {
+ if(direction==GMX_SUM_QGRID_FORWARD)
+ {
+ send_id = overlap->send_id[ipulse];
+ recv_id = overlap->recv_id[ipulse];
+ send_index0 = overlap->comm_data[ipulse].send_index0;
+ send_nindex = overlap->comm_data[ipulse].send_nindex;
+ recv_index0 = overlap->comm_data[ipulse].recv_index0;
+ recv_nindex = overlap->comm_data[ipulse].recv_nindex;
+ recvptr = overlap->recvbuf;
+ }
+ else
+ {
+ send_id = overlap->recv_id[ipulse];
+ recv_id = overlap->send_id[ipulse];
+ send_index0 = overlap->comm_data[ipulse].recv_index0;
+ send_nindex = overlap->comm_data[ipulse].recv_nindex;
+ recv_index0 = overlap->comm_data[ipulse].send_index0;
+ recv_nindex = overlap->comm_data[ipulse].send_nindex;
+ recvptr = grid + (recv_index0-pme->pmegrid_start_ix)*(pme->pmegrid_ny*pme->pmegrid_nz);
+ }
+
+ sendptr = grid + (send_index0-pme->pmegrid_start_ix)*(pme->pmegrid_ny*pme->pmegrid_nz);
+ datasize = pme->pmegrid_ny * pme->pmegrid_nz;
+
+ if (debug)
+ {
+ fprintf(debug,"PME send node %d %d -> %d grid start %d Communicating %d to %d\n",
+ pme->nodeid,overlap->nodeid,send_id,
+ pme->pmegrid_start_ix,
+ send_index0-pme->pmegrid_start_ix,
+ send_index0-pme->pmegrid_start_ix+send_nindex);
+ fprintf(debug,"PME recv node %d %d <- %d grid start %d Communicating %d to %d\n",
+ pme->nodeid,overlap->nodeid,recv_id,
+ pme->pmegrid_start_ix,
+ recv_index0-pme->pmegrid_start_ix,
+ recv_index0-pme->pmegrid_start_ix+recv_nindex);
+ }
+
+ MPI_Sendrecv(sendptr,send_nindex*datasize,GMX_MPI_REAL,
+ send_id,ipulse,
+ recvptr,recv_nindex*datasize,GMX_MPI_REAL,
+ recv_id,ipulse,
+ overlap->mpi_comm,&stat);
+
+ /* ADD data from contiguous recv buffer */
+ if(direction==GMX_SUM_QGRID_FORWARD)
+ {
+ p = grid + (recv_index0-pme->pmegrid_start_ix)*(pme->pmegrid_ny*pme->pmegrid_nz);
+ for(i=0;i<recv_nindex*datasize;i++)
+ {
+ p[i] += overlap->recvbuf[i];
+ }
+ }
+ }
+}
+#endif
+
+
+static int
+copy_pmegrid_to_fftgrid(gmx_pme_t pme, real *pmegrid, real *fftgrid)
+{
+ ivec local_fft_ndata,local_fft_offset,local_fft_size;
+ ivec local_pme_size;
+ int i,ix,iy,iz;
+ int pmeidx,fftidx;
+
+ /* Dimensions should be identical for A/B grid, so we just use A here */
+ gmx_parallel_3dfft_real_limits(pme->pfft_setupA,
+ local_fft_ndata,
+ local_fft_offset,
+ local_fft_size);
+
+ local_pme_size[0] = pme->pmegrid_nx;
+ local_pme_size[1] = pme->pmegrid_ny;
+ local_pme_size[2] = pme->pmegrid_nz;
+
+ /* The fftgrid is always 'justified' to the lower-left corner of the PME grid,
+ the offset is identical, and the PME grid always has more data (due to overlap)
+ */
+ {
+#ifdef DEBUG_PME
+ FILE *fp,*fp2;
+ char fn[STRLEN],format[STRLEN];
+ real val;
+ sprintf(fn,"pmegrid%d.pdb",pme->nodeid);
+ fp = ffopen(fn,"w");
+ sprintf(fn,"pmegrid%d.txt",pme->nodeid);
+ fp2 = ffopen(fn,"w");
+ sprintf(format,"%s%s\n",pdbformat,"%6.2f%6.2f");
+#endif
+
+ for(ix=0;ix<local_fft_ndata[XX];ix++)
+ {
+ for(iy=0;iy<local_fft_ndata[YY];iy++)
+ {
+ for(iz=0;iz<local_fft_ndata[ZZ];iz++)
+ {
+ pmeidx = ix*(local_pme_size[YY]*local_pme_size[ZZ])+iy*(local_pme_size[ZZ])+iz;
+ fftidx = ix*(local_fft_size[YY]*local_fft_size[ZZ])+iy*(local_fft_size[ZZ])+iz;
+ fftgrid[fftidx] = pmegrid[pmeidx];
+#ifdef DEBUG_PME
+ val = 100*pmegrid[pmeidx];
+ if (pmegrid[pmeidx] != 0)
+ fprintf(fp,format,"ATOM",pmeidx,"CA","GLY",' ',pmeidx,' ',
+ 5.0*ix,5.0*iy,5.0*iz,1.0,val);
+ if (pmegrid[pmeidx] != 0)
+ fprintf(fp2,"%-12s %5d %5d %5d %12.5e\n",
+ "qgrid",
+ pme->pmegrid_start_ix + ix,
+ pme->pmegrid_start_iy + iy,
+ pme->pmegrid_start_iz + iz,
+ pmegrid[pmeidx]);
+#endif
+ }
+ }
+ }
+#ifdef DEBUG_PME
+ fclose(fp);
+ fclose(fp2);
+#endif
+ }
+ return 0;
+}
+
+
+static gmx_cycles_t omp_cyc_start()
+{
+ return gmx_cycles_read();
+}
+
+static gmx_cycles_t omp_cyc_end(gmx_cycles_t c)
+{
+ return gmx_cycles_read() - c;
+}
+
+
+static int
+copy_fftgrid_to_pmegrid(gmx_pme_t pme, const real *fftgrid, real *pmegrid,
+ int nthread,int thread)
+{
+ ivec local_fft_ndata,local_fft_offset,local_fft_size;
+ ivec local_pme_size;
+ int ixy0,ixy1,ixy,ix,iy,iz;
+ int pmeidx,fftidx;
+#ifdef PME_TIME_THREADS
+ gmx_cycles_t c1;
+ static double cs1=0;
+ static int cnt=0;
+#endif
+
+#ifdef PME_TIME_THREADS
+ c1 = omp_cyc_start();
+#endif
+ /* Dimensions should be identical for A/B grid, so we just use A here */
+ gmx_parallel_3dfft_real_limits(pme->pfft_setupA,
+ local_fft_ndata,
+ local_fft_offset,
+ local_fft_size);
+
+ local_pme_size[0] = pme->pmegrid_nx;
+ local_pme_size[1] = pme->pmegrid_ny;
+ local_pme_size[2] = pme->pmegrid_nz;
+
+ /* The fftgrid is always 'justified' to the lower-left corner of the PME grid,
+ the offset is identical, and the PME grid always has more data (due to overlap)
+ */
+ ixy0 = ((thread )*local_fft_ndata[XX]*local_fft_ndata[YY])/nthread;
+ ixy1 = ((thread+1)*local_fft_ndata[XX]*local_fft_ndata[YY])/nthread;
+
+ for(ixy=ixy0;ixy<ixy1;ixy++)
+ {
+ ix = ixy/local_fft_ndata[YY];
+ iy = ixy - ix*local_fft_ndata[YY];
+
+ pmeidx = (ix*local_pme_size[YY] + iy)*local_pme_size[ZZ];
+ fftidx = (ix*local_fft_size[YY] + iy)*local_fft_size[ZZ];
+ for(iz=0;iz<local_fft_ndata[ZZ];iz++)
+ {
+ pmegrid[pmeidx+iz] = fftgrid[fftidx+iz];
+ }
+ }
+
+#ifdef PME_TIME_THREADS
+ c1 = omp_cyc_end(c1);
+ cs1 += (double)c1;
+ cnt++;
+ if (cnt % 20 == 0)
+ {
+ printf("copy %.2f\n",cs1*1e-9);
+ }
+#endif
+
+ return 0;
+}
+
+
+static void
+wrap_periodic_pmegrid(gmx_pme_t pme, real *pmegrid)
+{
+ int nx,ny,nz,pnx,pny,pnz,ny_x,overlap,ix,iy,iz;
+
+ nx = pme->nkx;
+ ny = pme->nky;
+ nz = pme->nkz;
+
+ pnx = pme->pmegrid_nx;
+ pny = pme->pmegrid_ny;
+ pnz = pme->pmegrid_nz;
+
+ overlap = pme->pme_order - 1;
+
+ /* Add periodic overlap in z */
+ for(ix=0; ix<pme->pmegrid_nx; ix++)
+ {
+ for(iy=0; iy<pme->pmegrid_ny; iy++)
+ {
+ for(iz=0; iz<overlap; iz++)
+ {
+ pmegrid[(ix*pny+iy)*pnz+iz] +=
+ pmegrid[(ix*pny+iy)*pnz+nz+iz];
+ }
+ }
+ }
+
+ if (pme->nnodes_minor == 1)
+ {
+ for(ix=0; ix<pme->pmegrid_nx; ix++)
+ {
+ for(iy=0; iy<overlap; iy++)
+ {
+ for(iz=0; iz<nz; iz++)
+ {
+ pmegrid[(ix*pny+iy)*pnz+iz] +=
+ pmegrid[(ix*pny+ny+iy)*pnz+iz];
+ }
+ }
+ }
+ }
+
+ if (pme->nnodes_major == 1)
+ {
+ ny_x = (pme->nnodes_minor == 1 ? ny : pme->pmegrid_ny);
+
+ for(ix=0; ix<overlap; ix++)
+ {
+ for(iy=0; iy<ny_x; iy++)
+ {
+ for(iz=0; iz<nz; iz++)
+ {
+ pmegrid[(ix*pny+iy)*pnz+iz] +=
+ pmegrid[((nx+ix)*pny+iy)*pnz+iz];
+ }
+ }
+ }
+ }
+}
+
+
+static void
+unwrap_periodic_pmegrid(gmx_pme_t pme, real *pmegrid)
+{
+ int nx,ny,nz,pnx,pny,pnz,ny_x,overlap,ix;
+
+ nx = pme->nkx;
+ ny = pme->nky;
+ nz = pme->nkz;
+
+ pnx = pme->pmegrid_nx;
+ pny = pme->pmegrid_ny;
+ pnz = pme->pmegrid_nz;
+
+ overlap = pme->pme_order - 1;
+
+ if (pme->nnodes_major == 1)
+ {
+ ny_x = (pme->nnodes_minor == 1 ? ny : pme->pmegrid_ny);
+
+ for(ix=0; ix<overlap; ix++)
+ {
+ int iy,iz;
+
+ for(iy=0; iy<ny_x; iy++)
+ {
+ for(iz=0; iz<nz; iz++)
+ {
+ pmegrid[((nx+ix)*pny+iy)*pnz+iz] =
+ pmegrid[(ix*pny+iy)*pnz+iz];
+ }
+ }
+ }
+ }
+
+ if (pme->nnodes_minor == 1)
+ {
+#pragma omp parallel for num_threads(pme->nthread) schedule(static)
+ for(ix=0; ix<pme->pmegrid_nx; ix++)
+ {
+ int iy,iz;
+
+ for(iy=0; iy<overlap; iy++)
+ {
+ for(iz=0; iz<nz; iz++)
+ {
+ pmegrid[(ix*pny+ny+iy)*pnz+iz] =
+ pmegrid[(ix*pny+iy)*pnz+iz];
+ }
+ }
+ }
+ }
+
+ /* Copy periodic overlap in z */
+#pragma omp parallel for num_threads(pme->nthread) schedule(static)
+ for(ix=0; ix<pme->pmegrid_nx; ix++)
+ {
+ int iy,iz;
+
+ for(iy=0; iy<pme->pmegrid_ny; iy++)
+ {
+ for(iz=0; iz<overlap; iz++)
+ {
+ pmegrid[(ix*pny+iy)*pnz+nz+iz] =
+ pmegrid[(ix*pny+iy)*pnz+iz];
+ }
+ }
+ }
+}
+
+static void clear_grid(int nx,int ny,int nz,real *grid,
+ ivec fs,int *flag,
+ int fx,int fy,int fz,
+ int order)
+{
+ int nc,ncz;
+ int fsx,fsy,fsz,gx,gy,gz,g0x,g0y,x,y,z;
+ int flind;
+
+ nc = 2 + (order - 2)/FLBS;
+ ncz = 2 + (order - 2)/FLBSZ;
+
+ for(fsx=fx; fsx<fx+nc; fsx++)
+ {
+ for(fsy=fy; fsy<fy+nc; fsy++)
+ {
+ for(fsz=fz; fsz<fz+ncz; fsz++)
+ {
+ flind = (fsx*fs[YY] + fsy)*fs[ZZ] + fsz;
+ if (flag[flind] == 0)
+ {
+ gx = fsx*FLBS;
+ gy = fsy*FLBS;
+ gz = fsz*FLBSZ;
+ g0x = (gx*ny + gy)*nz + gz;
+ for(x=0; x<FLBS; x++)
+ {
+ g0y = g0x;
+ for(y=0; y<FLBS; y++)
+ {
+ for(z=0; z<FLBSZ; z++)
+ {
+ grid[g0y+z] = 0;
+ }
+ g0y += nz;
+ }
+ g0x += ny*nz;
+ }
+
+ flag[flind] = 1;
+ }
+ }
+ }
+ }
+}
+
+/* This has to be a macro to enable full compiler optimization with xlC (and probably others too) */
+#define DO_BSPLINE(order) \
+for(ithx=0; (ithx<order); ithx++) \
+{ \
+ index_x = (i0+ithx)*pny*pnz; \
+ valx = qn*thx[ithx]; \
+ \
+ for(ithy=0; (ithy<order); ithy++) \
+ { \
+ valxy = valx*thy[ithy]; \
+ index_xy = index_x+(j0+ithy)*pnz; \
+ \
+ for(ithz=0; (ithz<order); ithz++) \
+ { \
+ index_xyz = index_xy+(k0+ithz); \
+ grid[index_xyz] += valxy*thz[ithz]; \
+ } \
+ } \
+}
+
+
+static void spread_q_bsplines_thread(pmegrid_t *pmegrid,
+ pme_atomcomm_t *atc, splinedata_t *spline,
+ pme_spline_work_t *work)
+{
+
+ /* spread charges from home atoms to local grid */
+ real *grid;
+ pme_overlap_t *ol;
+ int b,i,nn,n,ithx,ithy,ithz,i0,j0,k0;
+ int * idxptr;
+ int order,norder,index_x,index_xy,index_xyz;
+ real valx,valxy,qn;
+ real *thx,*thy,*thz;
+ int localsize, bndsize;
+ int pnx,pny,pnz,ndatatot;
+ int offx,offy,offz;
+
+ pnx = pmegrid->n[XX];
+ pny = pmegrid->n[YY];
+ pnz = pmegrid->n[ZZ];
+
+ offx = pmegrid->offset[XX];
+ offy = pmegrid->offset[YY];
+ offz = pmegrid->offset[ZZ];
+
+ ndatatot = pnx*pny*pnz;
+ grid = pmegrid->grid;
+ for(i=0;i<ndatatot;i++)
+ {
+ grid[i] = 0;
+ }
+
+ order = pmegrid->order;
+
+ for(nn=0; nn<spline->n; nn++)
+ {
+ n = spline->ind[nn];
+ qn = atc->q[n];
+
+ if (qn != 0)
+ {
+ idxptr = atc->idx[n];
+ norder = nn*order;
+
+ i0 = idxptr[XX] - offx;
+ j0 = idxptr[YY] - offy;
+ k0 = idxptr[ZZ] - offz;
+
+ thx = spline->theta[XX] + norder;
+ thy = spline->theta[YY] + norder;
+ thz = spline->theta[ZZ] + norder;
+
+ switch (order) {
+ case 4:
+#ifdef PME_SSE
+#ifdef PME_SSE_UNALIGNED
+#define PME_SPREAD_SSE_ORDER4
+#else
+#define PME_SPREAD_SSE_ALIGNED
+#define PME_ORDER 4
+#endif
+#include "pme_sse_single.h"
+#else
+ DO_BSPLINE(4);
+#endif
+ break;
+ case 5:
+#ifdef PME_SSE
+#define PME_SPREAD_SSE_ALIGNED
+#define PME_ORDER 5
+#include "pme_sse_single.h"
+#else
+ DO_BSPLINE(5);
+#endif
+ break;
+ default:
+ DO_BSPLINE(order);
+ break;
+ }
+ }
+ }
+}
+
+
+static void alloc_real_aligned(int n,real **ptr_raw,real **ptr)
+{
+ snew(*ptr_raw,n+8);
+
+ *ptr = (real *) (((size_t) *ptr_raw + 16) & (~((size_t) 15)));
+
+}
+static void set_grid_alignment(int *pmegrid_nz,int pme_order)
+{
+#ifdef PME_SSE
+ if (pme_order == 5
+#ifndef PME_SSE_UNALIGNED
+ || pme_order == 4
+#endif
+ )
+ {
+ /* Round nz up to a multiple of 4 to ensure alignment */
+ *pmegrid_nz = ((*pmegrid_nz + 3) & ~3);
+ }
+#endif
+}
+
+static void set_gridsize_alignment(int *gridsize,int pme_order)
+{
+#ifdef PME_SSE
+#ifndef PME_SSE_UNALIGNED
+ if (pme_order == 4)
+ {
+ /* Add extra elements to ensured aligned operations do not go
+ * beyond the allocated grid size.
+ * Note that for pme_order=5, the pme grid z-size alignment
+ * ensures that we will not go beyond the grid size.
+ */
+ *gridsize += 4;
+ }
+#endif
+#endif
+}
+
+static void pmegrid_init(pmegrid_t *grid,
+ int cx, int cy, int cz,
+ int x0, int y0, int z0,
+ int x1, int y1, int z1,
+ gmx_bool set_alignment,
+ int pme_order,
+ real *ptr)
+{
+ int nz,gridsize;
+
+ grid->ci[XX] = cx;
+ grid->ci[YY] = cy;
+ grid->ci[ZZ] = cz;
+ grid->offset[XX] = x0;
+ grid->offset[YY] = y0;
+ grid->offset[ZZ] = z0;
+ grid->n[XX] = x1 - x0 + pme_order - 1;
+ grid->n[YY] = y1 - y0 + pme_order - 1;
+ grid->n[ZZ] = z1 - z0 + pme_order - 1;
+
+ nz = grid->n[ZZ];
+ set_grid_alignment(&nz,pme_order);
+ if (set_alignment)
+ {
+ grid->n[ZZ] = nz;
+ }
+ else if (nz != grid->n[ZZ])
+ {
+ gmx_incons("pmegrid_init call with an unaligned z size");
+ }
+
+ grid->order = pme_order;
+ if (ptr == NULL)
+ {
+ gridsize = grid->n[XX]*grid->n[YY]*grid->n[ZZ];
+ set_gridsize_alignment(&gridsize,pme_order);
+ snew_aligned(grid->grid,gridsize,16);
+ }
+ else
+ {
+ grid->grid = ptr;
+ }
+}
+
+static int div_round_up(int enumerator,int denominator)
+{
+ return (enumerator + denominator - 1)/denominator;
+}
+
+static void make_subgrid_division(const ivec n,int ovl,int nthread,
+ ivec nsub)
+{
+ int gsize_opt,gsize;
+ int nsx,nsy,nsz;
+ char *env;
+
+ gsize_opt = -1;
+ for(nsx=1; nsx<=nthread; nsx++)
+ {
+ if (nthread % nsx == 0)
+ {
+ for(nsy=1; nsy<=nthread; nsy++)
+ {
+ if (nsx*nsy <= nthread && nthread % (nsx*nsy) == 0)
+ {
+ nsz = nthread/(nsx*nsy);
+
+ /* Determine the number of grid points per thread */
+ gsize =
+ (div_round_up(n[XX],nsx) + ovl)*
+ (div_round_up(n[YY],nsy) + ovl)*
+ (div_round_up(n[ZZ],nsz) + ovl);
+
+ /* Minimize the number of grids points per thread
+ * and, secondarily, the number of cuts in minor dimensions.
+ */
+ if (gsize_opt == -1 ||
+ gsize < gsize_opt ||
+ (gsize == gsize_opt &&
+ (nsz < nsub[ZZ] || (nsz == nsub[ZZ] && nsy < nsub[YY]))))
+ {
+ nsub[XX] = nsx;
+ nsub[YY] = nsy;
+ nsub[ZZ] = nsz;
+ gsize_opt = gsize;
+ }
+ }
+ }
+ }
+ }
+
+ env = getenv("GMX_PME_THREAD_DIVISION");
+ if (env != NULL)
+ {
+ sscanf(env,"%d %d %d",&nsub[XX],&nsub[YY],&nsub[ZZ]);
+ }
+
+ if (nsub[XX]*nsub[YY]*nsub[ZZ] != nthread)
+ {
+ gmx_fatal(FARGS,"PME grid thread division (%d x %d x %d) does not match the total number of threads (%d)",nsub[XX],nsub[YY],nsub[ZZ],nthread);
+ }
+}
+
+static void pmegrids_init(pmegrids_t *grids,
+ int nx,int ny,int nz,int nz_base,
+ int pme_order,
+ int nthread,
+ int overlap_x,
+ int overlap_y)
+{
+ ivec n,n_base,g0,g1;
+ int t,x,y,z,d,i,tfac;
+ int max_comm_lines;
+
+ n[XX] = nx - (pme_order - 1);
+ n[YY] = ny - (pme_order - 1);
+ n[ZZ] = nz - (pme_order - 1);
+
+ copy_ivec(n,n_base);
+ n_base[ZZ] = nz_base;
+
+ pmegrid_init(&grids->grid,0,0,0,0,0,0,n[XX],n[YY],n[ZZ],FALSE,pme_order,
+ NULL);
+
+ grids->nthread = nthread;
+
+ make_subgrid_division(n_base,pme_order-1,grids->nthread,grids->nc);
+
+ if (grids->nthread > 1)
+ {
+ ivec nst;
+ int gridsize;
+ real *grid_all;
+
+ for(d=0; d<DIM; d++)
+ {
+ nst[d] = div_round_up(n[d],grids->nc[d]) + pme_order - 1;
+ }
+ set_grid_alignment(&nst[ZZ],pme_order);
+
+ if (debug)
+ {
+ fprintf(debug,"pmegrid thread local division: %d x %d x %d\n",
+ grids->nc[XX],grids->nc[YY],grids->nc[ZZ]);
+ fprintf(debug,"pmegrid %d %d %d max thread pmegrid %d %d %d\n",
+ nx,ny,nz,
+ nst[XX],nst[YY],nst[ZZ]);
+ }
+
+ snew(grids->grid_th,grids->nthread);
+ t = 0;
+ gridsize = nst[XX]*nst[YY]*nst[ZZ];
+ set_gridsize_alignment(&gridsize,pme_order);
+ snew_aligned(grid_all,
+ grids->nthread*gridsize+(grids->nthread+1)*GMX_CACHE_SEP,
+ 16);
+
+ for(x=0; x<grids->nc[XX]; x++)
+ {
+ for(y=0; y<grids->nc[YY]; y++)
+ {
+ for(z=0; z<grids->nc[ZZ]; z++)
+ {
+ pmegrid_init(&grids->grid_th[t],
+ x,y,z,
+ (n[XX]*(x ))/grids->nc[XX],
+ (n[YY]*(y ))/grids->nc[YY],
+ (n[ZZ]*(z ))/grids->nc[ZZ],
+ (n[XX]*(x+1))/grids->nc[XX],
+ (n[YY]*(y+1))/grids->nc[YY],
+ (n[ZZ]*(z+1))/grids->nc[ZZ],
+ TRUE,
+ pme_order,
+ grid_all+GMX_CACHE_SEP+t*(gridsize+GMX_CACHE_SEP));
+ t++;
+ }
+ }
+ }
+ }
+
+ snew(grids->g2t,DIM);
+ tfac = 1;
+ for(d=DIM-1; d>=0; d--)
+ {
+ snew(grids->g2t[d],n[d]);
+ t = 0;
+ for(i=0; i<n[d]; i++)
+ {
+ /* The second check should match the parameters
+ * of the pmegrid_init call above.
+ */
+ while (t + 1 < grids->nc[d] && i >= (n[d]*(t+1))/grids->nc[d])
+ {
+ t++;
+ }
+ grids->g2t[d][i] = t*tfac;
+ }
+
+ tfac *= grids->nc[d];
+
+ switch (d)
+ {
+ case XX: max_comm_lines = overlap_x; break;
+ case YY: max_comm_lines = overlap_y; break;
+ case ZZ: max_comm_lines = pme_order - 1; break;
+ }
+ grids->nthread_comm[d] = 0;
+ while ((n[d]*grids->nthread_comm[d])/grids->nc[d] < max_comm_lines)
+ {
+ grids->nthread_comm[d]++;
+ }
+ if (debug != NULL)
+ {
+ fprintf(debug,"pmegrid thread grid communication range in %c: %d\n",
+ 'x'+d,grids->nthread_comm[d]);
+ }
+ /* It should be possible to make grids->nthread_comm[d]==grids->nc[d]
+ * work, but this is not a problematic restriction.
+ */
+ if (grids->nc[d] > 1 && grids->nthread_comm[d] > grids->nc[d])
+ {
+ gmx_fatal(FARGS,"Too many threads for PME (%d) compared to the number of grid lines, reduce the number of threads doing PME",grids->nthread);
+ }
+ }
+}
+
+
+static void pmegrids_destroy(pmegrids_t *grids)
+{
+ int t;
+
+ if (grids->grid.grid != NULL)
+ {
+ sfree(grids->grid.grid);
+
+ if (grids->nthread > 0)
+ {
+ for(t=0; t<grids->nthread; t++)
+ {
+ sfree(grids->grid_th[t].grid);
+ }
+ sfree(grids->grid_th);
+ }
+ }
+}
+
+
+static void realloc_work(pme_work_t *work,int nkx)
+{
+ if (nkx > work->nalloc)
+ {
+ work->nalloc = nkx;
+ srenew(work->mhx ,work->nalloc);
+ srenew(work->mhy ,work->nalloc);
+ srenew(work->mhz ,work->nalloc);
+ srenew(work->m2 ,work->nalloc);
+ srenew(work->denom,work->nalloc);
+ /* Allocate an aligned pointer for SSE operations, including 3 extra
+ * elements at the end since SSE operates on 4 elements at a time.
+ */
+ sfree_aligned(work->denom);
+ sfree_aligned(work->tmp1);
+ sfree_aligned(work->eterm);
+ snew_aligned(work->denom,work->nalloc+3,16);
+ snew_aligned(work->tmp1 ,work->nalloc+3,16);
+ snew_aligned(work->eterm,work->nalloc+3,16);
+ srenew(work->m2inv,work->nalloc);
+ }
+}
+
+
+static void free_work(pme_work_t *work)
+{
+ sfree(work->mhx);
+ sfree(work->mhy);
+ sfree(work->mhz);
+ sfree(work->m2);
+ sfree_aligned(work->denom);
+ sfree_aligned(work->tmp1);
+ sfree_aligned(work->eterm);
+ sfree(work->m2inv);
+}
+
+
+#ifdef PME_SSE
+ /* Calculate exponentials through SSE in float precision */
+inline static void calc_exponentials(int start, int end, real f, real *d_aligned, real *r_aligned, real *e_aligned)
+{
+ {
+ const __m128 two = _mm_set_ps(2.0f,2.0f,2.0f,2.0f);
+ __m128 f_sse;
+ __m128 lu;
+ __m128 tmp_d1,d_inv,tmp_r,tmp_e;
+ int kx;
+ f_sse = _mm_load1_ps(&f);
+ for(kx=0; kx<end; kx+=4)
+ {
+ tmp_d1 = _mm_load_ps(d_aligned+kx);
+ lu = _mm_rcp_ps(tmp_d1);
+ d_inv = _mm_mul_ps(lu,_mm_sub_ps(two,_mm_mul_ps(lu,tmp_d1)));
+ tmp_r = _mm_load_ps(r_aligned+kx);
+ tmp_r = gmx_mm_exp_ps(tmp_r);
+ tmp_e = _mm_mul_ps(f_sse,d_inv);
+ tmp_e = _mm_mul_ps(tmp_e,tmp_r);
+ _mm_store_ps(e_aligned+kx,tmp_e);
+ }
+ }
+}
+#else
+inline static void calc_exponentials(int start, int end, real f, real *d, real *r, real *e)
+{
+ int kx;
+ for(kx=start; kx<end; kx++)
+ {
+ d[kx] = 1.0/d[kx];
+ }
+ for(kx=start; kx<end; kx++)
+ {
+ r[kx] = exp(r[kx]);
+ }
+ for(kx=start; kx<end; kx++)
+ {
+ e[kx] = f*r[kx]*d[kx];
+ }
+}
+#endif
+
+
+static int solve_pme_yzx(gmx_pme_t pme,t_complex *grid,
+ real ewaldcoeff,real vol,
+ gmx_bool bEnerVir,
+ int nthread,int thread)
+{
+ /* do recip sum over local cells in grid */
+ /* y major, z middle, x minor or continuous */
+ t_complex *p0;
+ int kx,ky,kz,maxkx,maxky,maxkz;
+ int nx,ny,nz,iyz0,iyz1,iyz,iy,iz,kxstart,kxend;
+ real mx,my,mz;
+ real factor=M_PI*M_PI/(ewaldcoeff*ewaldcoeff);
+ real ets2,struct2,vfactor,ets2vf;
+ real d1,d2,energy=0;
+ real by,bz;
+ real virxx=0,virxy=0,virxz=0,viryy=0,viryz=0,virzz=0;
+ real rxx,ryx,ryy,rzx,rzy,rzz;
+ pme_work_t *work;
+ real *mhx,*mhy,*mhz,*m2,*denom,*tmp1,*eterm,*m2inv;
+ real mhxk,mhyk,mhzk,m2k;
+ real corner_fac;
+ ivec complex_order;
+ ivec local_ndata,local_offset,local_size;
+ real elfac;
+
+ elfac = ONE_4PI_EPS0/pme->epsilon_r;
+
+ nx = pme->nkx;
+ ny = pme->nky;
+ nz = pme->nkz;
+
+ /* Dimensions should be identical for A/B grid, so we just use A here */
+ gmx_parallel_3dfft_complex_limits(pme->pfft_setupA,
+ complex_order,
+ local_ndata,
+ local_offset,
+ local_size);
+
+ rxx = pme->recipbox[XX][XX];
+ ryx = pme->recipbox[YY][XX];
+ ryy = pme->recipbox[YY][YY];
+ rzx = pme->recipbox[ZZ][XX];
+ rzy = pme->recipbox[ZZ][YY];
+ rzz = pme->recipbox[ZZ][ZZ];
+
+ maxkx = (nx+1)/2;
+ maxky = (ny+1)/2;
+ maxkz = nz/2+1;
+
+ work = &pme->work[thread];
+ mhx = work->mhx;
+ mhy = work->mhy;
+ mhz = work->mhz;
+ m2 = work->m2;
+ denom = work->denom;
+ tmp1 = work->tmp1;
+ eterm = work->eterm;
+ m2inv = work->m2inv;
+
+ iyz0 = local_ndata[YY]*local_ndata[ZZ]* thread /nthread;
+ iyz1 = local_ndata[YY]*local_ndata[ZZ]*(thread+1)/nthread;
+
+ for(iyz=iyz0; iyz<iyz1; iyz++)
+ {
+ iy = iyz/local_ndata[ZZ];
+ iz = iyz - iy*local_ndata[ZZ];
+
+ ky = iy + local_offset[YY];
+
+ if (ky < maxky)
+ {
+ my = ky;
+ }
+ else
+ {
+ my = (ky - ny);
+ }
+
+ by = M_PI*vol*pme->bsp_mod[YY][ky];
+
+ kz = iz + local_offset[ZZ];
+
+ mz = kz;
+
+ bz = pme->bsp_mod[ZZ][kz];
+
+ /* 0.5 correction for corner points */
+ corner_fac = 1;
+ if (kz == 0 || kz == (nz+1)/2)
+ {
+ corner_fac = 0.5;
+ }
+
+ p0 = grid + iy*local_size[ZZ]*local_size[XX] + iz*local_size[XX];
+
+ /* We should skip the k-space point (0,0,0) */
+ if (local_offset[XX] > 0 || ky > 0 || kz > 0)
+ {
+ kxstart = local_offset[XX];
+ }
+ else
+ {
+ kxstart = local_offset[XX] + 1;
+ p0++;
+ }
+ kxend = local_offset[XX] + local_ndata[XX];
+
+ if (bEnerVir)
+ {
+ /* More expensive inner loop, especially because of the storage
+ * of the mh elements in array's.
+ * Because x is the minor grid index, all mh elements
+ * depend on kx for triclinic unit cells.
+ */
+
+ /* Two explicit loops to avoid a conditional inside the loop */
+ for(kx=kxstart; kx<maxkx; kx++)
+ {
+ mx = kx;
+
+ mhxk = mx * rxx;
+ mhyk = mx * ryx + my * ryy;
+ mhzk = mx * rzx + my * rzy + mz * rzz;
+ m2k = mhxk*mhxk + mhyk*mhyk + mhzk*mhzk;
+ mhx[kx] = mhxk;
+ mhy[kx] = mhyk;
+ mhz[kx] = mhzk;
+ m2[kx] = m2k;
+ denom[kx] = m2k*bz*by*pme->bsp_mod[XX][kx];
+ tmp1[kx] = -factor*m2k;
+ }
+
+ for(kx=maxkx; kx<kxend; kx++)
+ {
+ mx = (kx - nx);
+
+ mhxk = mx * rxx;
+ mhyk = mx * ryx + my * ryy;
+ mhzk = mx * rzx + my * rzy + mz * rzz;
+ m2k = mhxk*mhxk + mhyk*mhyk + mhzk*mhzk;
+ mhx[kx] = mhxk;
+ mhy[kx] = mhyk;
+ mhz[kx] = mhzk;
+ m2[kx] = m2k;
+ denom[kx] = m2k*bz*by*pme->bsp_mod[XX][kx];
+ tmp1[kx] = -factor*m2k;
+ }
+
+ for(kx=kxstart; kx<kxend; kx++)
+ {
+ m2inv[kx] = 1.0/m2[kx];
+ }
+
+ calc_exponentials(kxstart,kxend,elfac,denom,tmp1,eterm);
+
+ for(kx=kxstart; kx<kxend; kx++,p0++)
+ {
+ d1 = p0->re;
+ d2 = p0->im;
+
+ p0->re = d1*eterm[kx];
+ p0->im = d2*eterm[kx];
+
+ struct2 = 2.0*(d1*d1+d2*d2);
+
+ tmp1[kx] = eterm[kx]*struct2;
+ }
+
+ for(kx=kxstart; kx<kxend; kx++)
+ {
+ ets2 = corner_fac*tmp1[kx];
+ vfactor = (factor*m2[kx] + 1.0)*2.0*m2inv[kx];
+ energy += ets2;
+
+ ets2vf = ets2*vfactor;
+ virxx += ets2vf*mhx[kx]*mhx[kx] - ets2;
+ virxy += ets2vf*mhx[kx]*mhy[kx];
+ virxz += ets2vf*mhx[kx]*mhz[kx];
+ viryy += ets2vf*mhy[kx]*mhy[kx] - ets2;
+ viryz += ets2vf*mhy[kx]*mhz[kx];
+ virzz += ets2vf*mhz[kx]*mhz[kx] - ets2;
+ }
+ }
+ else
+ {
+ /* We don't need to calculate the energy and the virial.
+ * In this case the triclinic overhead is small.
+ */
+
+ /* Two explicit loops to avoid a conditional inside the loop */
+
+ for(kx=kxstart; kx<maxkx; kx++)
+ {
+ mx = kx;
+
+ mhxk = mx * rxx;
+ mhyk = mx * ryx + my * ryy;
+ mhzk = mx * rzx + my * rzy + mz * rzz;
+ m2k = mhxk*mhxk + mhyk*mhyk + mhzk*mhzk;
+ denom[kx] = m2k*bz*by*pme->bsp_mod[XX][kx];
+ tmp1[kx] = -factor*m2k;
+ }
+
+ for(kx=maxkx; kx<kxend; kx++)
+ {
+ mx = (kx - nx);
+
+ mhxk = mx * rxx;
+ mhyk = mx * ryx + my * ryy;
+ mhzk = mx * rzx + my * rzy + mz * rzz;
+ m2k = mhxk*mhxk + mhyk*mhyk + mhzk*mhzk;
+ denom[kx] = m2k*bz*by*pme->bsp_mod[XX][kx];
+ tmp1[kx] = -factor*m2k;
+ }
+
+ calc_exponentials(kxstart,kxend,elfac,denom,tmp1,eterm);
+
+ for(kx=kxstart; kx<kxend; kx++,p0++)
+ {
+ d1 = p0->re;
+ d2 = p0->im;
+
+ p0->re = d1*eterm[kx];
+ p0->im = d2*eterm[kx];
+ }
+ }
+ }
+
+ if (bEnerVir)
+ {
+ /* Update virial with local values.
+ * The virial is symmetric by definition.
+ * this virial seems ok for isotropic scaling, but I'm
+ * experiencing problems on semiisotropic membranes.
+ * IS THAT COMMENT STILL VALID??? (DvdS, 2001/02/07).
+ */
+ work->vir[XX][XX] = 0.25*virxx;
+ work->vir[YY][YY] = 0.25*viryy;
+ work->vir[ZZ][ZZ] = 0.25*virzz;
+ work->vir[XX][YY] = work->vir[YY][XX] = 0.25*virxy;
+ work->vir[XX][ZZ] = work->vir[ZZ][XX] = 0.25*virxz;
+ work->vir[YY][ZZ] = work->vir[ZZ][YY] = 0.25*viryz;
+
+ /* This energy should be corrected for a charged system */
+ work->energy = 0.5*energy;
+ }
+
+ /* Return the loop count */
+ return local_ndata[YY]*local_ndata[XX];
+}
+
+static void get_pme_ener_vir(const gmx_pme_t pme,int nthread,
+ real *mesh_energy,matrix vir)
+{
+ /* This function sums output over threads
+ * and should therefore only be called after thread synchronization.
+ */
+ int thread;
+
+ *mesh_energy = pme->work[0].energy;
+ copy_mat(pme->work[0].vir,vir);
+
+ for(thread=1; thread<nthread; thread++)
+ {
+ *mesh_energy += pme->work[thread].energy;
+ m_add(vir,pme->work[thread].vir,vir);
+ }
+}
+
+static int solve_pme_yzx_wrapper(gmx_pme_t pme,t_complex *grid,
+ real ewaldcoeff,real vol,
+ gmx_bool bEnerVir,real *mesh_energy,matrix vir)
+{
+ int nthread,thread;
+ int nelements=0;
+
+ nthread = pme->nthread;
+
+#pragma omp parallel for num_threads(nthread) schedule(static)
+ for(thread=0; thread<nthread; thread++)
+ {
+ int n;
+
+ n = solve_pme_yzx(pme,grid,ewaldcoeff,vol,bEnerVir,nthread,thread);
+ if (thread == 0)
+ {
+ nelements = n;
+ }
+ }
+
+ if (bEnerVir)
+ {
+ get_pme_ener_vir(pme,nthread,mesh_energy,vir);
+ }
+
+ return nelements;
+}
+
+
+#define DO_FSPLINE(order) \
+for(ithx=0; (ithx<order); ithx++) \
+{ \
+ index_x = (i0+ithx)*pny*pnz; \
+ tx = thx[ithx]; \
+ dx = dthx[ithx]; \
+ \
+ for(ithy=0; (ithy<order); ithy++) \
+ { \
+ index_xy = index_x+(j0+ithy)*pnz; \
+ ty = thy[ithy]; \
+ dy = dthy[ithy]; \
+ fxy1 = fz1 = 0; \
+ \
+ for(ithz=0; (ithz<order); ithz++) \
+ { \
+ gval = grid[index_xy+(k0+ithz)]; \
+ fxy1 += thz[ithz]*gval; \
+ fz1 += dthz[ithz]*gval; \
+ } \
+ fx += dx*ty*fxy1; \
+ fy += tx*dy*fxy1; \
+ fz += tx*ty*fz1; \
+ } \
+}
+
+
+void gather_f_bsplines(gmx_pme_t pme,real *grid,
+ gmx_bool bClearF,pme_atomcomm_t *atc,
+ splinedata_t *spline,
+ real scale)
+{
+ /* sum forces for local particles */
+ int nn,n,ithx,ithy,ithz,i0,j0,k0;
+ int index_x,index_xy;
+ int nx,ny,nz,pnx,pny,pnz;
+ int * idxptr;
+ real tx,ty,dx,dy,qn;
+ real fx,fy,fz,gval;
+ real fxy1,fz1;
+ real *thx,*thy,*thz,*dthx,*dthy,*dthz;
+ int norder;
+ real rxx,ryx,ryy,rzx,rzy,rzz;
+ int order;
+
+ pme_spline_work_t *work;
+
+ work = &pme->spline_work;
+
+ order = pme->pme_order;
+ thx = spline->theta[XX];
+ thy = spline->theta[YY];
+ thz = spline->theta[ZZ];
+ dthx = spline->dtheta[XX];
+ dthy = spline->dtheta[YY];
+ dthz = spline->dtheta[ZZ];
+ nx = pme->nkx;
+ ny = pme->nky;
+ nz = pme->nkz;
+ pnx = pme->pmegrid_nx;
+ pny = pme->pmegrid_ny;
+ pnz = pme->pmegrid_nz;
+
+ rxx = pme->recipbox[XX][XX];
+ ryx = pme->recipbox[YY][XX];
+ ryy = pme->recipbox[YY][YY];
+ rzx = pme->recipbox[ZZ][XX];
+ rzy = pme->recipbox[ZZ][YY];
+ rzz = pme->recipbox[ZZ][ZZ];
+
+ for(nn=0; nn<spline->n; nn++)
+ {
+ n = spline->ind[nn];
+ qn = atc->q[n];
+
+ if (bClearF)
+ {
+ atc->f[n][XX] = 0;
+ atc->f[n][YY] = 0;
+ atc->f[n][ZZ] = 0;
+ }
+ if (qn != 0)
+ {
+ fx = 0;
+ fy = 0;
+ fz = 0;
+ idxptr = atc->idx[n];
+ norder = nn*order;
+
+ i0 = idxptr[XX];
+ j0 = idxptr[YY];
+ k0 = idxptr[ZZ];
+
+ /* Pointer arithmetic alert, next six statements */
+ thx = spline->theta[XX] + norder;
+ thy = spline->theta[YY] + norder;
+ thz = spline->theta[ZZ] + norder;
+ dthx = spline->dtheta[XX] + norder;
+ dthy = spline->dtheta[YY] + norder;
+ dthz = spline->dtheta[ZZ] + norder;
+
+ switch (order) {
+ case 4:
+#ifdef PME_SSE
+#ifdef PME_SSE_UNALIGNED
+#define PME_GATHER_F_SSE_ORDER4
+#else
+#define PME_GATHER_F_SSE_ALIGNED
+#define PME_ORDER 4
+#endif
+#include "pme_sse_single.h"
+#else
+ DO_FSPLINE(4);
+#endif
+ break;
+ case 5:
+#ifdef PME_SSE
+#define PME_GATHER_F_SSE_ALIGNED
+#define PME_ORDER 5
+#include "pme_sse_single.h"
+#else
+ DO_FSPLINE(5);
+#endif
+ break;
+ default:
+ DO_FSPLINE(order);
+ break;
+ }
+
+ atc->f[n][XX] += -qn*( fx*nx*rxx );
+ atc->f[n][YY] += -qn*( fx*nx*ryx + fy*ny*ryy );
+ atc->f[n][ZZ] += -qn*( fx*nx*rzx + fy*ny*rzy + fz*nz*rzz );
+ }
+ }
+ /* Since the energy and not forces are interpolated
+ * the net force might not be exactly zero.
+ * This can be solved by also interpolating F, but
+ * that comes at a cost.
+ * A better hack is to remove the net force every
+ * step, but that must be done at a higher level
+ * since this routine doesn't see all atoms if running
+ * in parallel. Don't know how important it is? EL 990726
+ */
+}
+
+
+static real gather_energy_bsplines(gmx_pme_t pme,real *grid,
+ pme_atomcomm_t *atc)
+{
+ splinedata_t *spline;
+ int n,ithx,ithy,ithz,i0,j0,k0;
+ int index_x,index_xy;
+ int * idxptr;
+ real energy,pot,tx,ty,qn,gval;
+ real *thx,*thy,*thz;
+ int norder;
+ int order;
+
+ spline = &atc->spline[0];
+
+ order = pme->pme_order;
+
+ energy = 0;
+ for(n=0; (n<atc->n); n++) {
+ qn = atc->q[n];
+
+ if (qn != 0) {
+ idxptr = atc->idx[n];
+ norder = n*order;
+
+ i0 = idxptr[XX];
+ j0 = idxptr[YY];
+ k0 = idxptr[ZZ];
+
+ /* Pointer arithmetic alert, next three statements */
+ thx = spline->theta[XX] + norder;
+ thy = spline->theta[YY] + norder;
+ thz = spline->theta[ZZ] + norder;
+
+ pot = 0;
+ for(ithx=0; (ithx<order); ithx++)
+ {
+ index_x = (i0+ithx)*pme->pmegrid_ny*pme->pmegrid_nz;
+ tx = thx[ithx];
+
+ for(ithy=0; (ithy<order); ithy++)
+ {
+ index_xy = index_x+(j0+ithy)*pme->pmegrid_nz;
+ ty = thy[ithy];
+
+ for(ithz=0; (ithz<order); ithz++)
+ {
+ gval = grid[index_xy+(k0+ithz)];
+ pot += tx*ty*thz[ithz]*gval;
+ }
+
+ }
+ }
+
+ energy += pot*qn;
+ }
+ }
+
+ return energy;
+}
+
+/* Macro to force loop unrolling by fixing order.
+ * This gives a significant performance gain.
+ */
+#define CALC_SPLINE(order) \
+{ \
+ int j,k,l; \
+ real dr,div; \
+ real data[PME_ORDER_MAX]; \
+ real ddata[PME_ORDER_MAX]; \
+ \
+ for(j=0; (j<DIM); j++) \
+ { \
+ dr = xptr[j]; \
+ \
+ /* dr is relative offset from lower cell limit */ \
+ data[order-1] = 0; \
+ data[1] = dr; \
+ data[0] = 1 - dr; \
+ \
+ for(k=3; (k<order); k++) \
+ { \
+ div = 1.0/(k - 1.0); \
+ data[k-1] = div*dr*data[k-2]; \
+ for(l=1; (l<(k-1)); l++) \
+ { \
+ data[k-l-1] = div*((dr+l)*data[k-l-2]+(k-l-dr)* \
+ data[k-l-1]); \
+ } \
+ data[0] = div*(1-dr)*data[0]; \
+ } \
+ /* differentiate */ \
+ ddata[0] = -data[0]; \
+ for(k=1; (k<order); k++) \
+ { \
+ ddata[k] = data[k-1] - data[k]; \
+ } \
+ \
+ div = 1.0/(order - 1); \
+ data[order-1] = div*dr*data[order-2]; \
+ for(l=1; (l<(order-1)); l++) \
+ { \
+ data[order-l-1] = div*((dr+l)*data[order-l-2]+ \
+ (order-l-dr)*data[order-l-1]); \
+ } \
+ data[0] = div*(1 - dr)*data[0]; \
+ \
+ for(k=0; k<order; k++) \
+ { \
+ theta[j][i*order+k] = data[k]; \
+ dtheta[j][i*order+k] = ddata[k]; \
+ } \
+ } \
+}
+
+void make_bsplines(splinevec theta,splinevec dtheta,int order,
+ rvec fractx[],int nr,int ind[],real charge[],
+ gmx_bool bFreeEnergy)
+{
+ /* construct splines for local atoms */
+ int i,ii;
+ real *xptr;
+
+ for(i=0; i<nr; i++)
+ {
+ /* With free energy we do not use the charge check.
+ * In most cases this will be more efficient than calling make_bsplines
+ * twice, since usually more than half the particles have charges.
+ */
+ ii = ind[i];
+ if (bFreeEnergy || charge[ii] != 0.0) {
+ xptr = fractx[ii];
+ switch(order) {
+ case 4: CALC_SPLINE(4); break;
+ case 5: CALC_SPLINE(5); break;
+ default: CALC_SPLINE(order); break;
+ }
+ }
+ }
+}
+
+
+void make_dft_mod(real *mod,real *data,int ndata)
+{
+ int i,j;
+ real sc,ss,arg;
+
+ for(i=0;i<ndata;i++) {
+ sc=ss=0;
+ for(j=0;j<ndata;j++) {
+ arg=(2.0*M_PI*i*j)/ndata;
+ sc+=data[j]*cos(arg);
+ ss+=data[j]*sin(arg);
+ }
+ mod[i]=sc*sc+ss*ss;
+ }
+ for(i=0;i<ndata;i++)
+ if(mod[i]<1e-7)
+ mod[i]=(mod[i-1]+mod[i+1])*0.5;
+}
+
+
+
+void make_bspline_moduli(splinevec bsp_mod,int nx,int ny,int nz,int order)
+{
+ int nmax=max(nx,max(ny,nz));
+ real *data,*ddata,*bsp_data;
+ int i,k,l;
+ real div;
+
+ snew(data,order);
+ snew(ddata,order);
+ snew(bsp_data,nmax);
+
+ data[order-1]=0;
+ data[1]=0;
+ data[0]=1;
+
+ for(k=3;k<order;k++) {
+ div=1.0/(k-1.0);
+ data[k-1]=0;
+ for(l=1;l<(k-1);l++)
+ data[k-l-1]=div*(l*data[k-l-2]+(k-l)*data[k-l-1]);
+ data[0]=div*data[0];
+ }
+ /* differentiate */
+ ddata[0]=-data[0];
+ for(k=1;k<order;k++)
+ ddata[k]=data[k-1]-data[k];
+ div=1.0/(order-1);
+ data[order-1]=0;
+ for(l=1;l<(order-1);l++)
+ data[order-l-1]=div*(l*data[order-l-2]+(order-l)*data[order-l-1]);
+ data[0]=div*data[0];
+
+ for(i=0;i<nmax;i++)
+ bsp_data[i]=0;
+ for(i=1;i<=order;i++)
+ bsp_data[i]=data[i-1];
+
+ make_dft_mod(bsp_mod[XX],bsp_data,nx);
+ make_dft_mod(bsp_mod[YY],bsp_data,ny);
+ make_dft_mod(bsp_mod[ZZ],bsp_data,nz);
+
+ sfree(data);
+ sfree(ddata);
+ sfree(bsp_data);
+}
+
+static void setup_coordinate_communication(pme_atomcomm_t *atc)
+{
+ int nslab,n,i;
+ int fw,bw;
+
+ nslab = atc->nslab;
+
+ n = 0;
+ for(i=1; i<=nslab/2; i++) {
+ fw = (atc->nodeid + i) % nslab;
+ bw = (atc->nodeid - i + nslab) % nslab;
+ if (n < nslab - 1) {
+ atc->node_dest[n] = fw;
+ atc->node_src[n] = bw;
+ n++;
+ }
+ if (n < nslab - 1) {
+ atc->node_dest[n] = bw;
+ atc->node_src[n] = fw;
+ n++;
+ }
+ }
+}
+
+int gmx_pme_destroy(FILE *log,gmx_pme_t *pmedata)
+{
+ int thread;
+
+ if(NULL != log)
+ {
+ fprintf(log,"Destroying PME data structures.\n");
+ }
+
+ sfree((*pmedata)->nnx);
+ sfree((*pmedata)->nny);
+ sfree((*pmedata)->nnz);
+
+ pmegrids_destroy(&(*pmedata)->pmegridA);
+
+ sfree((*pmedata)->fftgridA);
+ sfree((*pmedata)->cfftgridA);
+ gmx_parallel_3dfft_destroy((*pmedata)->pfft_setupA);
+
+ if ((*pmedata)->pmegridB.grid.grid != NULL)
+ {
+ pmegrids_destroy(&(*pmedata)->pmegridB);
+ sfree((*pmedata)->fftgridB);
+ sfree((*pmedata)->cfftgridB);
+ gmx_parallel_3dfft_destroy((*pmedata)->pfft_setupB);
+ }
+ for(thread=0; thread<(*pmedata)->nthread; thread++)
+ {
+ free_work(&(*pmedata)->work[thread]);
+ }
+ sfree((*pmedata)->work);
+
+ sfree(*pmedata);
+ *pmedata = NULL;
+
+ return 0;
+}
+
+static int mult_up(int n,int f)
+{
+ return ((n + f - 1)/f)*f;
+}
+
+
+static double pme_load_imbalance(gmx_pme_t pme)
+{
+ int nma,nmi;
+ double n1,n2,n3;
+
+ nma = pme->nnodes_major;
+ nmi = pme->nnodes_minor;
+
+ n1 = mult_up(pme->nkx,nma)*mult_up(pme->nky,nmi)*pme->nkz;
+ n2 = mult_up(pme->nkx,nma)*mult_up(pme->nkz,nmi)*pme->nky;
+ n3 = mult_up(pme->nky,nma)*mult_up(pme->nkz,nmi)*pme->nkx;
+
+ /* pme_solve is roughly double the cost of an fft */
+
+ return (n1 + n2 + 3*n3)/(double)(6*pme->nkx*pme->nky*pme->nkz);
+}
+
+static void init_atomcomm(gmx_pme_t pme,pme_atomcomm_t *atc, t_commrec *cr,
+ int dimind,gmx_bool bSpread)
+{
+ int nk,k,s,thread;
+
+ atc->dimind = dimind;
+ atc->nslab = 1;
+ atc->nodeid = 0;
+ atc->pd_nalloc = 0;
+#ifdef GMX_MPI
+ if (pme->nnodes > 1)
+ {
+ atc->mpi_comm = pme->mpi_comm_d[dimind];
+ MPI_Comm_size(atc->mpi_comm,&atc->nslab);
+ MPI_Comm_rank(atc->mpi_comm,&atc->nodeid);
+ }
+ if (debug)
+ {
+ fprintf(debug,"For PME atom communication in dimind %d: nslab %d rank %d\n",atc->dimind,atc->nslab,atc->nodeid);
+ }
+#endif
+
+ atc->bSpread = bSpread;
+ atc->pme_order = pme->pme_order;
+
+ if (atc->nslab > 1)
+ {
+ /* These three allocations are not required for particle decomp. */
+ snew(atc->node_dest,atc->nslab);
+ snew(atc->node_src,atc->nslab);
+ setup_coordinate_communication(atc);
+
+ snew(atc->count_thread,pme->nthread);
+ for(thread=0; thread<pme->nthread; thread++)
+ {
+ snew(atc->count_thread[thread],atc->nslab);
+ }
+ atc->count = atc->count_thread[0];
+ snew(atc->rcount,atc->nslab);
+ snew(atc->buf_index,atc->nslab);
+ }
+
+ atc->nthread = pme->nthread;
+ if (atc->nthread > 1)
+ {
+ snew(atc->thread_plist,atc->nthread);
+ }
+ snew(atc->spline,atc->nthread);
+ for(thread=0; thread<atc->nthread; thread++)
+ {
+ if (atc->nthread > 1)
+ {
+ snew(atc->thread_plist[thread].n,atc->nthread+2*GMX_CACHE_SEP);
+ atc->thread_plist[thread].n += GMX_CACHE_SEP;
+ }
+ }
+}
+
+static void
+init_overlap_comm(pme_overlap_t * ol,
+ int norder,
+#ifdef GMX_MPI
+ MPI_Comm comm,
+#endif
+ int nnodes,
+ int nodeid,
+ int ndata,
+ int commplainsize)
+{
+ int lbnd,rbnd,maxlr,b,i;
+ int exten;
+ int nn,nk;
+ pme_grid_comm_t *pgc;
+ gmx_bool bCont;
+ int fft_start,fft_end,send_index1,recv_index1;
+
+#ifdef GMX_MPI
+ ol->mpi_comm = comm;
+#endif
+
+ ol->nnodes = nnodes;
+ ol->nodeid = nodeid;
+
+ /* Linear translation of the PME grid wo'nt affect reciprocal space
+ * calculations, so to optimize we only interpolate "upwards",
+ * which also means we only have to consider overlap in one direction.
+ * I.e., particles on this node might also be spread to grid indices
+ * that belong to higher nodes (modulo nnodes)
+ */
+
+ snew(ol->s2g0,ol->nnodes+1);
+ snew(ol->s2g1,ol->nnodes);
+ if (debug) { fprintf(debug,"PME slab boundaries:"); }
+ for(i=0; i<nnodes; i++)
+ {
+ /* s2g0 the local interpolation grid start.
+ * s2g1 the local interpolation grid end.
+ * Because grid overlap communication only goes forward,
+ * the grid the slabs for fft's should be rounded down.
+ */
+ ol->s2g0[i] = ( i *ndata + 0 )/nnodes;
+ ol->s2g1[i] = ((i+1)*ndata + nnodes-1)/nnodes + norder - 1;
+
+ if (debug)
+ {
+ fprintf(debug," %3d %3d",ol->s2g0[i],ol->s2g1[i]);
+ }
+ }
+ ol->s2g0[nnodes] = ndata;
+ if (debug) { fprintf(debug,"\n"); }
+
+ /* Determine with how many nodes we need to communicate the grid overlap */
+ b = 0;
+ do
+ {
+ b++;
+ bCont = FALSE;
+ for(i=0; i<nnodes; i++)
+ {
+ if ((i+b < nnodes && ol->s2g1[i] > ol->s2g0[i+b]) ||
+ (i+b >= nnodes && ol->s2g1[i] > ol->s2g0[i+b-nnodes] + ndata))
+ {
+ bCont = TRUE;
+ }
+ }
+ }
+ while (bCont && b < nnodes);
+ ol->noverlap_nodes = b - 1;
+
+ snew(ol->send_id,ol->noverlap_nodes);
+ snew(ol->recv_id,ol->noverlap_nodes);
+ for(b=0; b<ol->noverlap_nodes; b++)
+ {
+ ol->send_id[b] = (ol->nodeid + (b + 1)) % ol->nnodes;
+ ol->recv_id[b] = (ol->nodeid - (b + 1) + ol->nnodes) % ol->nnodes;
+ }
+ snew(ol->comm_data, ol->noverlap_nodes);
+
+ for(b=0; b<ol->noverlap_nodes; b++)
+ {
+ pgc = &ol->comm_data[b];
+ /* Send */
+ fft_start = ol->s2g0[ol->send_id[b]];
+ fft_end = ol->s2g0[ol->send_id[b]+1];
+ if (ol->send_id[b] < nodeid)
+ {
+ fft_start += ndata;
+ fft_end += ndata;
+ }
+ send_index1 = ol->s2g1[nodeid];
+ send_index1 = min(send_index1,fft_end);
+ pgc->send_index0 = fft_start;
+ pgc->send_nindex = max(0,send_index1 - pgc->send_index0);
+
+ /* We always start receiving to the first index of our slab */
+ fft_start = ol->s2g0[ol->nodeid];
+ fft_end = ol->s2g0[ol->nodeid+1];
+ recv_index1 = ol->s2g1[ol->recv_id[b]];
+ if (ol->recv_id[b] > nodeid)
+ {
+ recv_index1 -= ndata;
+ }
+ recv_index1 = min(recv_index1,fft_end);
+ pgc->recv_index0 = fft_start;
+ pgc->recv_nindex = max(0,recv_index1 - pgc->recv_index0);
+ }
+
+ /* For non-divisible grid we need pme_order iso pme_order-1 */
+ snew(ol->sendbuf,norder*commplainsize);
+ snew(ol->recvbuf,norder*commplainsize);
+}
+
+static void
+make_gridindex5_to_localindex(int n,int local_start,int local_range,
+ int **global_to_local,
+ real **fraction_shift)
+{
+ int i;
+ int * gtl;
+ real * fsh;
+
+ snew(gtl,5*n);
+ snew(fsh,5*n);
+ for(i=0; (i<5*n); i++)
+ {
+ /* Determine the global to local grid index */
+ gtl[i] = (i - local_start + n) % n;
+ /* For coordinates that fall within the local grid the fraction
+ * is correct, we don't need to shift it.
+ */
+ fsh[i] = 0;
+ if (local_range < n)
+ {
+ /* Due to rounding issues i could be 1 beyond the lower or
+ * upper boundary of the local grid. Correct the index for this.
+ * If we shift the index, we need to shift the fraction by
+ * the same amount in the other direction to not affect
+ * the weights.
+ * Note that due to this shifting the weights at the end of
+ * the spline might change, but that will only involve values
+ * between zero and values close to the precision of a real,
+ * which is anyhow the accuracy of the whole mesh calculation.
+ */
+ /* With local_range=0 we should not change i=local_start */
+ if (i % n != local_start)
+ {
+ if (gtl[i] == n-1)
+ {
+ gtl[i] = 0;
+ fsh[i] = -1;
+ }
+ else if (gtl[i] == local_range)
+ {
+ gtl[i] = local_range - 1;
+ fsh[i] = 1;
+ }
+ }
+ }
+ }
+
+ *global_to_local = gtl;
+ *fraction_shift = fsh;
+}
+
+static void sse_mask_init(pme_spline_work_t *work,int order)
+{
+#ifdef PME_SSE
+ float tmp[8];
+ __m128 zero_SSE;
+ int of,i;
+
+ zero_SSE = _mm_setzero_ps();
+
+ for(of=0; of<8-(order-1); of++)
+ {
+ for(i=0; i<8; i++)
+ {
+ tmp[i] = (i >= of && i < of+order ? 1 : 0);
+ }
+ work->mask_SSE0[of] = _mm_loadu_ps(tmp);
+ work->mask_SSE1[of] = _mm_loadu_ps(tmp+4);
+ work->mask_SSE0[of] = _mm_cmpgt_ps(work->mask_SSE0[of],zero_SSE);
+ work->mask_SSE1[of] = _mm_cmpgt_ps(work->mask_SSE1[of],zero_SSE);
+ }
+#endif
+}
+
+static void
+gmx_pme_check_grid_restrictions(FILE *fplog,char dim,int nnodes,int *nk)
+{
+ int nk_new;
+
+ if (*nk % nnodes != 0)
+ {
+ nk_new = nnodes*(*nk/nnodes + 1);
+
+ if (2*nk_new >= 3*(*nk))
+ {
+ gmx_fatal(FARGS,"The PME grid size in dim %c (%d) is not divisble by the number of nodes doing PME in dim %c (%d). The grid size would have to be increased by more than 50%% to make the grid divisible. Change the total number of nodes or the number of domain decomposition cells in x or the PME grid %c dimension (and the cut-off).",
+ dim,*nk,dim,nnodes,dim);
+ }
+
+ if (fplog != NULL)
+ {
+ fprintf(fplog,"\nNOTE: The PME grid size in dim %c (%d) is not divisble by the number of nodes doing PME in dim %c (%d). Increasing the PME grid size in dim %c to %d. This will increase the accuracy and will not decrease the performance significantly on this number of nodes. For optimal performance change the total number of nodes or the number of domain decomposition cells in x or the PME grid %c dimension (and the cut-off).\n\n",
+ dim,*nk,dim,nnodes,dim,nk_new,dim);
+ }
+
+ *nk = nk_new;
+ }
+}
+
+int gmx_pme_init(gmx_pme_t * pmedata,
+ t_commrec * cr,
+ int nnodes_major,
+ int nnodes_minor,
+ t_inputrec * ir,
+ int homenr,
+ gmx_bool bFreeEnergy,
+ gmx_bool bReproducible,
+ int nthread)
+{
+ gmx_pme_t pme=NULL;
+
+ pme_atomcomm_t *atc;
+ ivec ndata;
+
+ if (debug)
+ fprintf(debug,"Creating PME data structures.\n");
+ snew(pme,1);
+
+ pme->redist_init = FALSE;
+ pme->sum_qgrid_tmp = NULL;
+ pme->sum_qgrid_dd_tmp = NULL;
+ pme->buf_nalloc = 0;
+ pme->redist_buf_nalloc = 0;
+
+ pme->nnodes = 1;
+ pme->bPPnode = TRUE;
+
+ pme->nnodes_major = nnodes_major;
+ pme->nnodes_minor = nnodes_minor;
+
+#ifdef GMX_MPI
+ if (PAR(cr))
+ {
+ pme->mpi_comm = cr->mpi_comm_mygroup;
+
+ MPI_Comm_rank(pme->mpi_comm,&pme->nodeid);
+ MPI_Comm_size(pme->mpi_comm,&pme->nnodes);
+ }
+#endif
+
+ if (pme->nnodes == 1)
+ {
+ pme->ndecompdim = 0;
+ pme->nodeid_major = 0;
+ pme->nodeid_minor = 0;
+#ifdef GMX_MPI
+ pme->mpi_comm_d[0] = pme->mpi_comm_d[1] = MPI_COMM_NULL;
+#endif
+ }
+ else
+ {
+ if (nnodes_minor == 1)
+ {
+#ifdef GMX_MPI
+ pme->mpi_comm_d[0] = pme->mpi_comm;
+ pme->mpi_comm_d[1] = MPI_COMM_NULL;
+#endif
+ pme->ndecompdim = 1;
+ pme->nodeid_major = pme->nodeid;
+ pme->nodeid_minor = 0;
+
+ }
+ else if (nnodes_major == 1)
+ {
+#ifdef GMX_MPI
+ pme->mpi_comm_d[0] = MPI_COMM_NULL;
+ pme->mpi_comm_d[1] = pme->mpi_comm;
+#endif
+ pme->ndecompdim = 1;
+ pme->nodeid_major = 0;
+ pme->nodeid_minor = pme->nodeid;
+ }
+ else
+ {
+ if (pme->nnodes % nnodes_major != 0)
+ {
+ gmx_incons("For 2D PME decomposition, #PME nodes must be divisible by the number of nodes in the major dimension");
+ }
+ pme->ndecompdim = 2;
+
+#ifdef GMX_MPI
+ MPI_Comm_split(pme->mpi_comm,pme->nodeid % nnodes_minor,
+ pme->nodeid,&pme->mpi_comm_d[0]); /* My communicator along major dimension */
+ MPI_Comm_split(pme->mpi_comm,pme->nodeid/nnodes_minor,
+ pme->nodeid,&pme->mpi_comm_d[1]); /* My communicator along minor dimension */
+
+ MPI_Comm_rank(pme->mpi_comm_d[0],&pme->nodeid_major);
+ MPI_Comm_size(pme->mpi_comm_d[0],&pme->nnodes_major);
+ MPI_Comm_rank(pme->mpi_comm_d[1],&pme->nodeid_minor);
+ MPI_Comm_size(pme->mpi_comm_d[1],&pme->nnodes_minor);
+#endif
+ }
+ pme->bPPnode = (cr->duty & DUTY_PP);
+ }
+
+ pme->nthread = nthread;
+
+ if (ir->ePBC == epbcSCREW)
+ {
+ gmx_fatal(FARGS,"pme does not (yet) work with pbc = screw");
+ }
+
+ pme->bFEP = ((ir->efep != efepNO) && bFreeEnergy);
+ pme->nkx = ir->nkx;
+ pme->nky = ir->nky;
+ pme->nkz = ir->nkz;
+ pme->pme_order = ir->pme_order;
+ pme->epsilon_r = ir->epsilon_r;
+
+ if (pme->pme_order > PME_ORDER_MAX)
+ {
+ gmx_fatal(FARGS,"pme_order (%d) is larger than the maximum allowed value (%d). Modify and recompile the code if you really need such a high order.",
+ pme->pme_order,PME_ORDER_MAX);
+ }
+
+ /* Currently pme.c supports only the fft5d FFT code.
+ * Therefore the grid always needs to be divisible by nnodes.
+ * When the old 1D code is also supported again, change this check.
+ *
+ * This check should be done before calling gmx_pme_init
+ * and fplog should be passed iso stderr.
+ *
+ if (pme->ndecompdim >= 2)
+ */
+ if (pme->ndecompdim >= 1)
+ {
+ /*
+ gmx_pme_check_grid_restrictions(pme->nodeid==0 ? stderr : NULL,
+ 'x',nnodes_major,&pme->nkx);
+ gmx_pme_check_grid_restrictions(pme->nodeid==0 ? stderr : NULL,
+ 'y',nnodes_minor,&pme->nky);
+ */
+ }
+
+ if (pme->nkx <= pme->pme_order*(pme->nnodes_major > 1 ? 2 : 1) ||
+ pme->nky <= pme->pme_order*(pme->nnodes_minor > 1 ? 2 : 1) ||
+ pme->nkz <= pme->pme_order)
+ {
+ gmx_fatal(FARGS,"The pme grid dimensions need to be larger than pme_order (%d) and in parallel larger than 2*pme_ordern for x and/or y",pme->pme_order);
+ }
+
+ if (pme->nnodes > 1) {
+ double imbal;
+
+#ifdef GMX_MPI
+ MPI_Type_contiguous(DIM, mpi_type, &(pme->rvec_mpi));
+ MPI_Type_commit(&(pme->rvec_mpi));
+#endif
+
+ /* Note that the charge spreading and force gathering, which usually
+ * takes about the same amount of time as FFT+solve_pme,
+ * is always fully load balanced
+ * (unless the charge distribution is inhomogeneous).
+ */
+
+ imbal = pme_load_imbalance(pme);
+ if (imbal >= 1.2 && pme->nodeid_major == 0 && pme->nodeid_minor == 0)
+ {
+ fprintf(stderr,
+ "\n"
+ "NOTE: The load imbalance in PME FFT and solve is %d%%.\n"
+ " For optimal PME load balancing\n"
+ " PME grid_x (%d) and grid_y (%d) should be divisible by #PME_nodes_x (%d)\n"
+ " and PME grid_y (%d) and grid_z (%d) should be divisible by #PME_nodes_y (%d)\n"
+ "\n",
+ (int)((imbal-1)*100 + 0.5),
+ pme->nkx,pme->nky,pme->nnodes_major,
+ pme->nky,pme->nkz,pme->nnodes_minor);
+ }
+ }
+
+ /* For non-divisible grid we need pme_order iso pme_order-1 */
+ /* In sum_qgrid_dd x overlap is copied in place: take padding into account.
+ * y is always copied through a buffer: we don't need padding in z,
+ * but we do need the overlap in x because of the communication order.
+ */
+ init_overlap_comm(&pme->overlap[0],pme->pme_order,
+#ifdef GMX_MPI
+ pme->mpi_comm_d[0],
+#endif
+ pme->nnodes_major,pme->nodeid_major,
+ pme->nkx,
+ (div_round_up(pme->nky,pme->nnodes_minor)+pme->pme_order)*(pme->nkz+pme->pme_order-1));
+
+ init_overlap_comm(&pme->overlap[1],pme->pme_order,
+#ifdef GMX_MPI
+ pme->mpi_comm_d[1],
+#endif
+ pme->nnodes_minor,pme->nodeid_minor,
+ pme->nky,
+ (div_round_up(pme->nkx,pme->nnodes_major)+pme->pme_order)*pme->nkz);
+
+ /* Check for a limitation of the (current) sum_fftgrid_dd code */
+ if (pme->nthread > 1 &&
+ (pme->overlap[0].noverlap_nodes > 1 ||
+ pme->overlap[1].noverlap_nodes > 1))
+ {
+ gmx_fatal(FARGS,"With threads the number of grid lines per node along x and or y should be pme_order (%d) or more or exactly pme_order-1",pme->pme_order);
+ }
+
+ snew(pme->bsp_mod[XX],pme->nkx);
+ snew(pme->bsp_mod[YY],pme->nky);
+ snew(pme->bsp_mod[ZZ],pme->nkz);
+
+ /* The required size of the interpolation grid, including overlap.
+ * The allocated size (pmegrid_n?) might be slightly larger.
+ */
+ pme->pmegrid_nx = pme->overlap[0].s2g1[pme->nodeid_major] -
+ pme->overlap[0].s2g0[pme->nodeid_major];
+ pme->pmegrid_ny = pme->overlap[1].s2g1[pme->nodeid_minor] -
+ pme->overlap[1].s2g0[pme->nodeid_minor];
+ pme->pmegrid_nz_base = pme->nkz;
+ pme->pmegrid_nz = pme->pmegrid_nz_base + pme->pme_order - 1;
+ set_grid_alignment(&pme->pmegrid_nz,pme->pme_order);
+
+ pme->pmegrid_start_ix = pme->overlap[0].s2g0[pme->nodeid_major];
+ pme->pmegrid_start_iy = pme->overlap[1].s2g0[pme->nodeid_minor];
+ pme->pmegrid_start_iz = 0;
+
+ make_gridindex5_to_localindex(pme->nkx,
+ pme->pmegrid_start_ix,
+ pme->pmegrid_nx - (pme->pme_order-1),
+ &pme->nnx,&pme->fshx);
+ make_gridindex5_to_localindex(pme->nky,
+ pme->pmegrid_start_iy,
+ pme->pmegrid_ny - (pme->pme_order-1),
+ &pme->nny,&pme->fshy);
+ make_gridindex5_to_localindex(pme->nkz,
+ pme->pmegrid_start_iz,
+ pme->pmegrid_nz_base,
+ &pme->nnz,&pme->fshz);
+
+ pmegrids_init(&pme->pmegridA,
+ pme->pmegrid_nx,pme->pmegrid_ny,pme->pmegrid_nz,
+ pme->pmegrid_nz_base,
+ pme->pme_order,
+ pme->nthread,
+ pme->overlap[0].s2g1[pme->nodeid_major]-pme->overlap[0].s2g0[pme->nodeid_major+1],
+ pme->overlap[1].s2g1[pme->nodeid_minor]-pme->overlap[1].s2g0[pme->nodeid_minor+1]);
+
+ sse_mask_init(&pme->spline_work,pme->pme_order);
+
+ ndata[0] = pme->nkx;
+ ndata[1] = pme->nky;
+ ndata[2] = pme->nkz;
+
+ /* This routine will allocate the grid data to fit the FFTs */
+ gmx_parallel_3dfft_init(&pme->pfft_setupA,ndata,
+ &pme->fftgridA,&pme->cfftgridA,
+ pme->mpi_comm_d,
+ pme->overlap[0].s2g0,pme->overlap[1].s2g0,
+ bReproducible,pme->nthread);
+
+ if (bFreeEnergy)
+ {
+ pmegrids_init(&pme->pmegridB,
+ pme->pmegrid_nx,pme->pmegrid_ny,pme->pmegrid_nz,
+ pme->pmegrid_nz_base,
+ pme->pme_order,
+ pme->nthread,
+ pme->nkx % pme->nnodes_major != 0,
+ pme->nky % pme->nnodes_minor != 0);
+
+ gmx_parallel_3dfft_init(&pme->pfft_setupB,ndata,
+ &pme->fftgridB,&pme->cfftgridB,
+ pme->mpi_comm_d,
+ pme->overlap[0].s2g0,pme->overlap[1].s2g0,
+ bReproducible,pme->nthread);
+ }
+ else
+ {
+ pme->pmegridB.grid.grid = NULL;
+ pme->fftgridB = NULL;
+ pme->cfftgridB = NULL;
+ }
+
+ make_bspline_moduli(pme->bsp_mod,pme->nkx,pme->nky,pme->nkz,pme->pme_order);
+
+ /* Use atc[0] for spreading */
+ init_atomcomm(pme,&pme->atc[0],cr,nnodes_major > 1 ? 0 : 1,TRUE);
+ if (pme->ndecompdim >= 2)
+ {
+ init_atomcomm(pme,&pme->atc[1],cr,1,FALSE);
+ }
+
+ if (pme->nnodes == 1) {
+ pme->atc[0].n = homenr;
+ pme_realloc_atomcomm_things(&pme->atc[0]);
+ }
+
+ {
+ int thread;
+
+ /* Use fft5d, order after FFT is y major, z, x minor */
+
+ snew(pme->work,pme->nthread);
+ for(thread=0; thread<pme->nthread; thread++)
+ {
+ realloc_work(&pme->work[thread],pme->nkx);
+ }
+ }
+
+ *pmedata = pme;
+
+ return 0;
+}
+
+
+static void copy_local_grid(gmx_pme_t pme,
+ pmegrids_t *pmegrids,int thread,real *fftgrid)
+{
+ ivec local_fft_ndata,local_fft_offset,local_fft_size;
+ int fft_my,fft_mz;
+ int nsx,nsy,nsz;
+ ivec nf;
+ int offx,offy,offz,x,y,z,i0,i0t;
+ int d;
+ pmegrid_t *pmegrid;
+ real *grid_th;
+
+ gmx_parallel_3dfft_real_limits(pme->pfft_setupA,
+ local_fft_ndata,
+ local_fft_offset,
+ local_fft_size);
+ fft_my = local_fft_size[YY];
+ fft_mz = local_fft_size[ZZ];
+
+ pmegrid = &pmegrids->grid_th[thread];
+
+ nsx = pmegrid->n[XX];
+ nsy = pmegrid->n[YY];
+ nsz = pmegrid->n[ZZ];
+
+ for(d=0; d<DIM; d++)
+ {
+ nf[d] = min(pmegrid->n[d] - (pmegrid->order - 1),
+ local_fft_ndata[d] - pmegrid->offset[d]);
+ }
+
+ offx = pmegrid->offset[XX];
+ offy = pmegrid->offset[YY];
+ offz = pmegrid->offset[ZZ];
+
+ /* Directly copy the non-overlapping parts of the local grids.
+ * This also initializes the full grid.
+ */
+ grid_th = pmegrid->grid;
+ for(x=0; x<nf[XX]; x++)
+ {
+ for(y=0; y<nf[YY]; y++)
+ {
+ i0 = ((offx + x)*fft_my + (offy + y))*fft_mz + offz;
+ i0t = (x*nsy + y)*nsz;
+ for(z=0; z<nf[ZZ]; z++)
+ {
+ fftgrid[i0+z] = grid_th[i0t+z];
+ }
+ }
+ }
+}
+
+static void print_sendbuf(gmx_pme_t pme,real *sendbuf)
+{
+ ivec local_fft_ndata,local_fft_offset,local_fft_size;
+ pme_overlap_t *overlap;
+ int datasize,nind;
+ int i,x,y,z,n;
+
+ gmx_parallel_3dfft_real_limits(pme->pfft_setupA,
+ local_fft_ndata,
+ local_fft_offset,
+ local_fft_size);
+ /* Major dimension */
+ overlap = &pme->overlap[0];
+
+ nind = overlap->comm_data[0].send_nindex;
+
+ for(y=0; y<local_fft_ndata[YY]; y++) {
+ printf(" %2d",y);
+ }
+ printf("\n");
+
+ i = 0;
+ for(x=0; x<nind; x++) {
+ for(y=0; y<local_fft_ndata[YY]; y++) {
+ n = 0;
+ for(z=0; z<local_fft_ndata[ZZ]; z++) {
+ if (sendbuf[i] != 0) n++;
+ i++;
+ }
+ printf(" %2d",n);
+ }
+ printf("\n");
+ }
+}
+
+static void
+reduce_threadgrid_overlap(gmx_pme_t pme,
+ const pmegrids_t *pmegrids,int thread,
+ real *fftgrid,real *commbuf_x,real *commbuf_y)
+{
+ ivec local_fft_ndata,local_fft_offset,local_fft_size;
+ int fft_nx,fft_ny,fft_nz;
+ int fft_my,fft_mz;
+ int buf_my=-1;
+ int nsx,nsy,nsz;
+ ivec ne;
+ int offx,offy,offz,x,y,z,i0,i0t;
+ int sx,sy,sz,fx,fy,fz,tx1,ty1,tz1,ox,oy,oz;
+ gmx_bool bClearBufX,bClearBufY,bClearBufXY,bClearBuf;
+ gmx_bool bCommX,bCommY;
+ int d;
+ int thread_f;
+ const pmegrid_t *pmegrid,*pmegrid_g,*pmegrid_f;
+ const real *grid_th;
+ real *commbuf=NULL;
+
+ gmx_parallel_3dfft_real_limits(pme->pfft_setupA,
+ local_fft_ndata,
+ local_fft_offset,
+ local_fft_size);
+ fft_nx = local_fft_ndata[XX];
+ fft_ny = local_fft_ndata[YY];
+ fft_nz = local_fft_ndata[ZZ];
+
+ fft_my = local_fft_size[YY];
+ fft_mz = local_fft_size[ZZ];
+
+ /* This routine is called when all thread have finished spreading.
+ * Here each thread sums grid contributions calculated by other threads
+ * to the thread local grid volume.
+ * To minimize the number of grid copying operations,
+ * this routines sums immediately from the pmegrid to the fftgrid.
+ */
+
+ /* Determine which part of the full node grid we should operate on,
+ * this is our thread local part of the full grid.
+ */
+ pmegrid = &pmegrids->grid_th[thread];
+
+ for(d=0; d<DIM; d++)
+ {
+ ne[d] = min(pmegrid->offset[d]+pmegrid->n[d]-(pmegrid->order-1),
+ local_fft_ndata[d]);
+ }
+
+ offx = pmegrid->offset[XX];
+ offy = pmegrid->offset[YY];
+ offz = pmegrid->offset[ZZ];
+
+
+ bClearBufX = TRUE;
+ bClearBufY = TRUE;
+ bClearBufXY = TRUE;
+
+ /* Now loop over all the thread data blocks that contribute
+ * to the grid region we (our thread) are operating on.
+ */
+ /* Note that ffy_nx/y is equal to the number of grid points
+ * between the first point of our node grid and the one of the next node.
+ */
+ for(sx=0; sx>=-pmegrids->nthread_comm[XX]; sx--)
+ {
+ fx = pmegrid->ci[XX] + sx;
+ ox = 0;
+ bCommX = FALSE;
+ if (fx < 0) {
+ fx += pmegrids->nc[XX];
+ ox -= fft_nx;
+ bCommX = (pme->nnodes_major > 1);
+ }
+ pmegrid_g = &pmegrids->grid_th[fx*pmegrids->nc[YY]*pmegrids->nc[ZZ]];
+ ox += pmegrid_g->offset[XX];
+ if (!bCommX)
+ {
+ tx1 = min(ox + pmegrid_g->n[XX],ne[XX]);
+ }
+ else
+ {
+ tx1 = min(ox + pmegrid_g->n[XX],pme->pme_order);
+ }
+
+ for(sy=0; sy>=-pmegrids->nthread_comm[YY]; sy--)
+ {
+ fy = pmegrid->ci[YY] + sy;
+ oy = 0;
+ bCommY = FALSE;
+ if (fy < 0) {
+ fy += pmegrids->nc[YY];
+ oy -= fft_ny;
+ bCommY = (pme->nnodes_minor > 1);
+ }
+ pmegrid_g = &pmegrids->grid_th[fy*pmegrids->nc[ZZ]];
+ oy += pmegrid_g->offset[YY];
+ if (!bCommY)
+ {
+ ty1 = min(oy + pmegrid_g->n[YY],ne[YY]);
+ }
+ else
+ {
+ ty1 = min(oy + pmegrid_g->n[YY],pme->pme_order);
+ }
+
+ for(sz=0; sz>=-pmegrids->nthread_comm[ZZ]; sz--)
+ {
+ fz = pmegrid->ci[ZZ] + sz;
+ oz = 0;
+ if (fz < 0)
+ {
+ fz += pmegrids->nc[ZZ];
+ oz -= fft_nz;
+ }
+ pmegrid_g = &pmegrids->grid_th[fz];
+ oz += pmegrid_g->offset[ZZ];
+ tz1 = min(oz + pmegrid_g->n[ZZ],ne[ZZ]);
+
+ if (sx == 0 && sy == 0 && sz == 0)
+ {
+ /* We have already added our local contribution
+ * before calling this routine, so skip it here.
+ */
+ continue;
+ }
+
+ thread_f = (fx*pmegrids->nc[YY] + fy)*pmegrids->nc[ZZ] + fz;
+
+ pmegrid_f = &pmegrids->grid_th[thread_f];
+
+ grid_th = pmegrid_f->grid;
+
+ nsx = pmegrid_f->n[XX];
+ nsy = pmegrid_f->n[YY];
+ nsz = pmegrid_f->n[ZZ];
+
+#ifdef DEBUG_PME_REDUCE
+ printf("n%d t%d add %d %2d %2d %2d %2d %2d %2d %2d-%2d %2d-%2d, %2d-%2d %2d-%2d, %2d-%2d %2d-%2d\n",
+ pme->nodeid,thread,thread_f,
+ pme->pmegrid_start_ix,
+ pme->pmegrid_start_iy,
+ pme->pmegrid_start_iz,
+ sx,sy,sz,
+ offx-ox,tx1-ox,offx,tx1,
+ offy-oy,ty1-oy,offy,ty1,
+ offz-oz,tz1-oz,offz,tz1);
+#endif
+
+ if (!(bCommX || bCommY))
+ {
+ /* Copy from the thread local grid to the node grid */
+ for(x=offx; x<tx1; x++)
+ {
+ for(y=offy; y<ty1; y++)
+ {
+ i0 = (x*fft_my + y)*fft_mz;
+ i0t = ((x - ox)*nsy + (y - oy))*nsz - oz;
+ for(z=offz; z<tz1; z++)
+ {
+ fftgrid[i0+z] += grid_th[i0t+z];
+ }
+ }
+ }
+ }
+ else
+ {
+ /* The order of this conditional decides
+ * where the corner volume gets stored with x+y decomp.
+ */
+ if (bCommY)
+ {
+ commbuf = commbuf_y;
+ buf_my = ty1 - offy;
+ if (bCommX)
+ {
+ /* We index commbuf modulo the local grid size */
+ commbuf += buf_my*fft_nx*fft_nz;
+
+ bClearBuf = bClearBufXY;
+ bClearBufXY = FALSE;
+ }
+ else
+ {
+ bClearBuf = bClearBufY;
+ bClearBufY = FALSE;
+ }
+ }
+ else
+ {
+ commbuf = commbuf_x;
+ buf_my = fft_ny;
+ bClearBuf = bClearBufX;
+ bClearBufX = FALSE;
+ }
+
+ /* Copy to the communication buffer */
+ for(x=offx; x<tx1; x++)
+ {
+ for(y=offy; y<ty1; y++)
+ {
+ i0 = (x*buf_my + y)*fft_nz;
+ i0t = ((x - ox)*nsy + (y - oy))*nsz - oz;
+
+ if (bClearBuf)
+ {
+ /* First access of commbuf, initialize it */
+ for(z=offz; z<tz1; z++)
+ {
+ commbuf[i0+z] = grid_th[i0t+z];
+ }
+ }
+ else
+ {
+ for(z=offz; z<tz1; z++)
+ {
+ commbuf[i0+z] += grid_th[i0t+z];
+ }
+ }
+ }
+ }
+ }
+ }
+ }
+ }
+}
+
+
+static void sum_fftgrid_dd(gmx_pme_t pme,real *fftgrid)
+{
+ ivec local_fft_ndata,local_fft_offset,local_fft_size;
+ pme_overlap_t *overlap;
+ int send_nindex;
+ int recv_index0,recv_nindex;
+#ifdef GMX_MPI
+ MPI_Status stat;
+#endif
+ int ipulse,send_id,recv_id,datasize,gridsize,size_yx;
+ real *sendptr,*recvptr;
+ int x,y,z,indg,indb;
+
+ /* Note that this routine is only used for forward communication.
+ * Since the force gathering, unlike the charge spreading,
+ * can be trivially parallelized over the particles,
+ * the backwards process is much simpler and can use the "old"
+ * communication setup.
+ */
+
+ gmx_parallel_3dfft_real_limits(pme->pfft_setupA,
+ local_fft_ndata,
+ local_fft_offset,
+ local_fft_size);
+
+ /* Currently supports only a single communication pulse */
+
+/* for(ipulse=0;ipulse<overlap->noverlap_nodes;ipulse++) */
+ if (pme->nnodes_minor > 1)
+ {
+ /* Major dimension */
+ overlap = &pme->overlap[1];
+
+ if (pme->nnodes_major > 1)
+ {
+ size_yx = pme->overlap[0].comm_data[0].send_nindex;
+ }
+ else
+ {
+ size_yx = 0;
+ }
+ datasize = (local_fft_ndata[XX]+size_yx)*local_fft_ndata[ZZ];
+
+ ipulse = 0;
+
+ send_id = overlap->send_id[ipulse];
+ recv_id = overlap->recv_id[ipulse];
+ send_nindex = overlap->comm_data[ipulse].send_nindex;
+ /* recv_index0 = overlap->comm_data[ipulse].recv_index0; */
+ recv_index0 = 0;
+ recv_nindex = overlap->comm_data[ipulse].recv_nindex;
+
+ sendptr = overlap->sendbuf;
+ recvptr = overlap->recvbuf;
+
+ /*
+ printf("node %d comm %2d x %2d x %2d\n",pme->nodeid,
+ local_fft_ndata[XX]+size_yx,send_nindex,local_fft_ndata[ZZ]);
+ printf("node %d send %f, %f\n",pme->nodeid,
+ sendptr[0],sendptr[send_nindex*datasize-1]);
+ */
+
+#ifdef GMX_MPI
+ MPI_Sendrecv(sendptr,send_nindex*datasize,GMX_MPI_REAL,
+ send_id,ipulse,
+ recvptr,recv_nindex*datasize,GMX_MPI_REAL,
+ recv_id,ipulse,
+ overlap->mpi_comm,&stat);
+#endif
+
+ for(x=0; x<local_fft_ndata[XX]; x++)
+ {
+ for(y=0; y<recv_nindex; y++)
+ {
+ indg = (x*local_fft_size[YY] + y)*local_fft_size[ZZ];
+ indb = (x*recv_nindex + y)*local_fft_ndata[ZZ];
+ for(z=0; z<local_fft_ndata[ZZ]; z++)
+ {
+ fftgrid[indg+z] += recvptr[indb+z];
+ }
+ }
+ }
+ if (pme->nnodes_major > 1)
+ {
+ sendptr = pme->overlap[0].sendbuf;
+ for(x=0; x<size_yx; x++)
+ {
+ for(y=0; y<recv_nindex; y++)
+ {
+ indg = (x*local_fft_ndata[YY] + y)*local_fft_ndata[ZZ];
+ indb = ((local_fft_ndata[XX] + x)*recv_nindex +y)*local_fft_ndata[ZZ];
+ for(z=0; z<local_fft_ndata[ZZ]; z++)
+ {
+ sendptr[indg+z] += recvptr[indb+z];
+ }
+ }
+ }
+ }
+ }
+
+ /* for(ipulse=0;ipulse<overlap->noverlap_nodes;ipulse++) */
+ if (pme->nnodes_major > 1)
+ {
+ /* Major dimension */
+ overlap = &pme->overlap[0];
+
+ datasize = local_fft_ndata[YY]*local_fft_ndata[ZZ];
+ gridsize = local_fft_size[YY] *local_fft_size[ZZ];
+
+ ipulse = 0;
+
+ send_id = overlap->send_id[ipulse];
+ recv_id = overlap->recv_id[ipulse];
+ send_nindex = overlap->comm_data[ipulse].send_nindex;
+ /* recv_index0 = overlap->comm_data[ipulse].recv_index0; */
+ recv_index0 = 0;
+ recv_nindex = overlap->comm_data[ipulse].recv_nindex;
+
+ sendptr = overlap->sendbuf;
+ recvptr = overlap->recvbuf;
+
+ if (debug != NULL)
+ {
+ fprintf(debug,"PME fftgrid comm %2d x %2d x %2d\n",
+ send_nindex,local_fft_ndata[YY],local_fft_ndata[ZZ]);
+ }
+
+#ifdef GMX_MPI
+ MPI_Sendrecv(sendptr,send_nindex*datasize,GMX_MPI_REAL,
+ send_id,ipulse,
+ recvptr,recv_nindex*datasize,GMX_MPI_REAL,
+ recv_id,ipulse,
+ overlap->mpi_comm,&stat);
+#endif
+
+ for(x=0; x<recv_nindex; x++)
+ {
+ for(y=0; y<local_fft_ndata[YY]; y++)
+ {
+ indg = (x*local_fft_size[YY] + y)*local_fft_size[ZZ];
+ indb = (x*local_fft_ndata[YY] + y)*local_fft_ndata[ZZ];
+ for(z=0; z<local_fft_ndata[ZZ]; z++)
+ {
+ fftgrid[indg+z] += recvptr[indb+z];
+ }
+ }
+ }
+ }
+}
+
+
+static void spread_on_grid(gmx_pme_t pme,
+ pme_atomcomm_t *atc,pmegrids_t *grids,
+ gmx_bool bCalcSplines,gmx_bool bSpread,
+ real *fftgrid)
+{
+ int nthread,thread;
+#ifdef PME_TIME_THREADS
+ gmx_cycles_t c1,c2,c3,ct1a,ct1b,ct1c;
+ static double cs1=0,cs2=0,cs3=0;
+ static double cs1a[6]={0,0,0,0,0,0};
+ static int cnt=0;
+#endif
+
+ nthread = pme->nthread;
+
+#ifdef PME_TIME_THREADS
+ c1 = omp_cyc_start();
+#endif
+ if (bCalcSplines)
+ {
+#pragma omp parallel for num_threads(nthread) schedule(static)
+ for(thread=0; thread<nthread; thread++)
+ {
+ int start,end;
+
+ start = atc->n* thread /nthread;
+ end = atc->n*(thread+1)/nthread;
+
+ /* Compute fftgrid index for all atoms,
+ * with help of some extra variables.
+ */
+ calc_interpolation_idx(pme,atc,start,end,thread);
+ }
+ }
+#ifdef PME_TIME_THREADS
+ c1 = omp_cyc_end(c1);
+ cs1 += (double)c1;
+#endif
+
+#ifdef PME_TIME_THREADS
+ c2 = omp_cyc_start();
+#endif
+#pragma omp parallel for num_threads(nthread) schedule(static)
+ for(thread=0; thread<nthread; thread++)
+ {
+ splinedata_t *spline;
+ pmegrid_t *grid;
+
+ /* make local bsplines */
+ if (grids->nthread == 1)
+ {
+ spline = &atc->spline[0];
+
+ spline->n = atc->n;
+
+ grid = &grids->grid;
+ }
+ else
+ {
+ spline = &atc->spline[thread];
+
+ make_thread_local_ind(atc,thread,spline);
+
+ grid = &grids->grid_th[thread];
+ }
+
+ if (bCalcSplines)
+ {
+ make_bsplines(spline->theta,spline->dtheta,pme->pme_order,
+ atc->fractx,spline->n,spline->ind,atc->q,pme->bFEP);
+ }
+
+ if (bSpread)
+ {
+ /* put local atoms on grid. */
+#ifdef PME_TIME_SPREAD
+ ct1a = omp_cyc_start();
+#endif
+ spread_q_bsplines_thread(grid,atc,spline,&pme->spline_work);
+
+ if (grids->nthread > 1)
+ {
+ copy_local_grid(pme,grids,thread,fftgrid);
+ }
+#ifdef PME_TIME_SPREAD
+ ct1a = omp_cyc_end(ct1a);
+ cs1a[thread] += (double)ct1a;
+#endif
+ }
+ }
+#ifdef PME_TIME_THREADS
+ c2 = omp_cyc_end(c2);
+ cs2 += (double)c2;
+#endif
+
+ if (grids->nthread > 1)
+ {
+#ifdef PME_TIME_THREADS
+ c3 = omp_cyc_start();
+#endif
+#pragma omp parallel for num_threads(grids->nthread) schedule(static)
+ for(thread=0; thread<grids->nthread; thread++)
+ {
+ reduce_threadgrid_overlap(pme,grids,thread,
+ fftgrid,
+ pme->overlap[0].sendbuf,
+ pme->overlap[1].sendbuf);
+#ifdef PRINT_PME_SENDBUF
+ print_sendbuf(pme,pme->overlap[0].sendbuf);
+#endif
+ }
+#ifdef PME_TIME_THREADS
+ c3 = omp_cyc_end(c3);
+ cs3 += (double)c3;
+#endif
+
+ if (pme->nnodes > 1)
+ {
+ /* Communicate the overlapping part of the fftgrid */
+ sum_fftgrid_dd(pme,fftgrid);
+ }
+ }
+
+#ifdef PME_TIME_THREADS
+ cnt++;
+ if (cnt % 20 == 0)
+ {
+ printf("idx %.2f spread %.2f red %.2f",
+ cs1*1e-9,cs2*1e-9,cs3*1e-9);
+#ifdef PME_TIME_SPREAD
+ for(thread=0; thread<nthread; thread++)
+ printf(" %.2f",cs1a[thread]*1e-9);
+#endif
+ printf("\n");
+ }
+#endif
+}
+
+
+static void dump_grid(FILE *fp,
+ int sx,int sy,int sz,int nx,int ny,int nz,
+ int my,int mz,const real *g)
+{
+ int x,y,z;
+
+ for(x=0; x<nx; x++)
+ {
+ for(y=0; y<ny; y++)
+ {
+ for(z=0; z<nz; z++)
+ {
+ fprintf(fp,"%2d %2d %2d %6.3f\n",
+ sx+x,sy+y,sz+z,g[(x*my + y)*mz + z]);
+ }
+ }
+ }
+}
+
+static void dump_local_fftgrid(gmx_pme_t pme,const real *fftgrid)
+{
+ ivec local_fft_ndata,local_fft_offset,local_fft_size;
+
+ gmx_parallel_3dfft_real_limits(pme->pfft_setupA,
+ local_fft_ndata,
+ local_fft_offset,
+ local_fft_size);
+
+ dump_grid(stderr,
+ pme->pmegrid_start_ix,
+ pme->pmegrid_start_iy,
+ pme->pmegrid_start_iz,
+ pme->pmegrid_nx-pme->pme_order+1,
+ pme->pmegrid_ny-pme->pme_order+1,
+ pme->pmegrid_nz-pme->pme_order+1,
+ local_fft_size[YY],
+ local_fft_size[ZZ],
+ fftgrid);
+}
+
+
+void gmx_pme_calc_energy(gmx_pme_t pme,int n,rvec *x,real *q,real *V)
+{
+ pme_atomcomm_t *atc;
+ pmegrids_t *grid;
+
+ if (pme->nnodes > 1)
+ {
+ gmx_incons("gmx_pme_calc_energy called in parallel");
+ }
+ if (pme->bFEP > 1)
+ {
+ gmx_incons("gmx_pme_calc_energy with free energy");
+ }
+
+ atc = &pme->atc_energy;
+ atc->nslab = 1;
+ atc->bSpread = TRUE;
+ atc->pme_order = pme->pme_order;
+ atc->n = n;
+ pme_realloc_atomcomm_things(atc);
+ atc->x = x;
+ atc->q = q;
+
+ /* We only use the A-charges grid */
+ grid = &pme->pmegridA;
+
+ spread_on_grid(pme,atc,NULL,TRUE,FALSE,pme->fftgridA);
+
+ *V = gather_energy_bsplines(pme,grid->grid.grid,atc);
+}
+
+
+static void reset_pmeonly_counters(t_commrec *cr,gmx_wallcycle_t wcycle,
+ t_nrnb *nrnb,t_inputrec *ir, gmx_large_int_t step_rel)
+{
+ /* Reset all the counters related to performance over the run */
+ wallcycle_stop(wcycle,ewcRUN);
+ wallcycle_reset_all(wcycle);
+ init_nrnb(nrnb);
+ ir->init_step += step_rel;
+ ir->nsteps -= step_rel;
+ wallcycle_start(wcycle,ewcRUN);
+}
+
+
+int gmx_pmeonly(gmx_pme_t pme,
+ t_commrec *cr, t_nrnb *nrnb,
+ gmx_wallcycle_t wcycle,
+ real ewaldcoeff, gmx_bool bGatherOnly,
+ t_inputrec *ir)
+{
+ gmx_pme_pp_t pme_pp;
+ int natoms;
+ matrix box;
+ rvec *x_pp=NULL,*f_pp=NULL;
+ real *chargeA=NULL,*chargeB=NULL;
+ real lambda=0;
+ int maxshift_x=0,maxshift_y=0;
+ real energy,dvdlambda;
+ matrix vir;
+ float cycles;
+ int count;
+ gmx_bool bEnerVir;
+ gmx_large_int_t step,step_rel;
+
+
+ pme_pp = gmx_pme_pp_init(cr);
+
+ init_nrnb(nrnb);
+
+ count = 0;
+ do /****** this is a quasi-loop over time steps! */
+ {
+ /* Domain decomposition */
+ natoms = gmx_pme_recv_q_x(pme_pp,
+ &chargeA,&chargeB,box,&x_pp,&f_pp,
+ &maxshift_x,&maxshift_y,
+ &pme->bFEP,&lambda,
+ &bEnerVir,
+ &step);
+
+ if (natoms == -1) {
+ /* We should stop: break out of the loop */
+ break;
+ }
+
+ step_rel = step - ir->init_step;
+
+ if (count == 0)
+ wallcycle_start(wcycle,ewcRUN);
+
+ wallcycle_start(wcycle,ewcPMEMESH);
+
+ dvdlambda = 0;
+ clear_mat(vir);
+ gmx_pme_do(pme,0,natoms,x_pp,f_pp,chargeA,chargeB,box,
+ cr,maxshift_x,maxshift_y,nrnb,wcycle,vir,ewaldcoeff,
+ &energy,lambda,&dvdlambda,
+ GMX_PME_DO_ALL_F | (bEnerVir ? GMX_PME_CALC_ENER_VIR : 0));
+
+ cycles = wallcycle_stop(wcycle,ewcPMEMESH);
+
+ gmx_pme_send_force_vir_ener(pme_pp,
+ f_pp,vir,energy,dvdlambda,
+ cycles);
+
+ count++;
+
+ if (step_rel == wcycle_get_reset_counters(wcycle))
+ {
+ /* Reset all the counters related to performance over the run */
+ reset_pmeonly_counters(cr,wcycle,nrnb,ir,step_rel);
+ wcycle_set_reset_counters(wcycle, 0);
+ }
+
+ } /***** end of quasi-loop, we stop with the break above */
+ while (TRUE);
+
+ return 0;
+}
+
+int gmx_pme_do(gmx_pme_t pme,
+ int start, int homenr,
+ rvec x[], rvec f[],
+ real *chargeA, real *chargeB,
+ matrix box, t_commrec *cr,
+ int maxshift_x, int maxshift_y,
+ t_nrnb *nrnb, gmx_wallcycle_t wcycle,
+ matrix vir, real ewaldcoeff,
+ real *energy, real lambda,
+ real *dvdlambda, int flags)
+{
+ int q,d,i,j,ntot,npme;
+ int nx,ny,nz;
+ int n_d,local_ny;
+ pme_atomcomm_t *atc=NULL;
+ pmegrids_t *pmegrid=NULL;
+ real *grid=NULL;
+ real *ptr;
+ rvec *x_d,*f_d;
+ real *charge=NULL,*q_d;
+ real energy_AB[2];
+ matrix vir_AB[2];
+ gmx_bool bClearF;
+ gmx_parallel_3dfft_t pfft_setup;
+ real * fftgrid;
+ t_complex * cfftgrid;
+ int thread;
+
+ if (pme->nnodes > 1) {
+ atc = &pme->atc[0];
+ atc->npd = homenr;
+ if (atc->npd > atc->pd_nalloc) {
+ atc->pd_nalloc = over_alloc_dd(atc->npd);
+ srenew(atc->pd,atc->pd_nalloc);
+ }
+ atc->maxshift = (atc->dimind==0 ? maxshift_x : maxshift_y);
+ }
+ else
+ {
+ /* This could be necessary for TPI */
+ pme->atc[0].n = homenr;
+ }
+
+ for(q=0; q<(pme->bFEP ? 2 : 1); q++) {
+ if (q == 0) {
+ pmegrid = &pme->pmegridA;
+ fftgrid = pme->fftgridA;
+ cfftgrid = pme->cfftgridA;
+ pfft_setup = pme->pfft_setupA;
+ charge = chargeA+start;
+ } else {
+ pmegrid = &pme->pmegridB;
+ fftgrid = pme->fftgridB;
+ cfftgrid = pme->cfftgridB;
+ pfft_setup = pme->pfft_setupB;
+ charge = chargeB+start;
+ }
+ grid = pmegrid->grid.grid;
+ /* Unpack structure */
+ if (debug) {
+ fprintf(debug,"PME: nnodes = %d, nodeid = %d\n",
+ cr->nnodes,cr->nodeid);
+ fprintf(debug,"Grid = %p\n",(void*)grid);
+ if (grid == NULL)
+ gmx_fatal(FARGS,"No grid!");
+ }
+ where();
+
+ m_inv_ur0(box,pme->recipbox);
+
+ if (pme->nnodes == 1) {
+ atc = &pme->atc[0];
+ if (DOMAINDECOMP(cr)) {
+ atc->n = homenr;
+ pme_realloc_atomcomm_things(atc);
+ }
+ atc->x = x;
+ atc->q = charge;
+ atc->f = f;
+ } else {
+ wallcycle_start(wcycle,ewcPME_REDISTXF);
+ for(d=pme->ndecompdim-1; d>=0; d--)
+ {
+ if (d == pme->ndecompdim-1)
+ {
+ n_d = homenr;
+ x_d = x + start;
+ q_d = charge;
+ }
+ else
+ {
+ n_d = pme->atc[d+1].n;
+ x_d = atc->x;
+ q_d = atc->q;
+ }
+ atc = &pme->atc[d];
+ atc->npd = n_d;
+ if (atc->npd > atc->pd_nalloc) {
+ atc->pd_nalloc = over_alloc_dd(atc->npd);
+ srenew(atc->pd,atc->pd_nalloc);
+ }
+ atc->maxshift = (atc->dimind==0 ? maxshift_x : maxshift_y);
+ pme_calc_pidx_wrapper(n_d,pme->recipbox,x_d,atc);
+ where();
+
+ /* Redistribute x (only once) and qA or qB */
+ if (DOMAINDECOMP(cr)) {
+ dd_pmeredist_x_q(pme, n_d, q==0, x_d, q_d, atc);
+ } else {
+ pmeredist_pd(pme, TRUE, n_d, q==0, x_d, q_d, atc);
+ }
+ }
+ where();
+
+ wallcycle_stop(wcycle,ewcPME_REDISTXF);
+ }
+
+ if (debug)
+ fprintf(debug,"Node= %6d, pme local particles=%6d\n",
+ cr->nodeid,atc->n);
+
+ if (flags & GMX_PME_SPREAD_Q)
+ {
+ wallcycle_start(wcycle,ewcPME_SPREADGATHER);
+
+ /* Spread the charges on a grid */
+ spread_on_grid(pme,&pme->atc[0],pmegrid,q==0,TRUE,fftgrid);
+
+ if (q == 0)
+ {
+ inc_nrnb(nrnb,eNR_WEIGHTS,DIM*atc->n);
+ }
+ inc_nrnb(nrnb,eNR_SPREADQBSP,
+ pme->pme_order*pme->pme_order*pme->pme_order*atc->n);
+
+ if (pme->nthread == 1)
+ {
+ wrap_periodic_pmegrid(pme,grid);
+
+ /* sum contributions to local grid from other nodes */
+#ifdef GMX_MPI
+ if (pme->nnodes > 1)
+ {
+ gmx_sum_qgrid_dd(pme,grid,GMX_SUM_QGRID_FORWARD);
+ where();
+ }
+#endif
+
+ copy_pmegrid_to_fftgrid(pme,grid,fftgrid);
+ }
+
+ wallcycle_stop(wcycle,ewcPME_SPREADGATHER);
+
+ /*
+ dump_local_fftgrid(pme,fftgrid);
+ exit(0);
+ */
+ }
+
+ /* Here we start a large thread parallel region */
+#pragma omp parallel for num_threads(pme->nthread) schedule(static)
+ for(thread=0; thread<pme->nthread; thread++)
+ {
+ if (flags & GMX_PME_SOLVE)
+ {
+ int loop_count;
+
+ /* do 3d-fft */
+ if (thread == 0)
+ {
+ wallcycle_start(wcycle,ewcPME_FFT);
+ }
+ gmx_parallel_3dfft_execute(pfft_setup,GMX_FFT_REAL_TO_COMPLEX,
+ fftgrid,cfftgrid,thread,wcycle);
+ if (thread == 0)
+ {
+ wallcycle_stop(wcycle,ewcPME_FFT);
+ }
+ where();
+
+ /* solve in k-space for our local cells */
+ if (thread == 0)
+ {
+ wallcycle_start(wcycle,ewcPME_SOLVE);
+ }
+ loop_count =
+ solve_pme_yzx(pme,cfftgrid,ewaldcoeff,
+ box[XX][XX]*box[YY][YY]*box[ZZ][ZZ],
+ flags & GMX_PME_CALC_ENER_VIR,
+ pme->nthread,thread);
+ if (thread == 0)
+ {
+ wallcycle_stop(wcycle,ewcPME_SOLVE);
+ where();
+ inc_nrnb(nrnb,eNR_SOLVEPME,loop_count);
+ }
+ }
+
+ if (flags & GMX_PME_CALC_F)
+ {
+ /* do 3d-invfft */
+ if (thread == 0)
+ {
+ where();
+ wallcycle_start(wcycle,ewcPME_FFT);
+ }
+ gmx_parallel_3dfft_execute(pfft_setup,GMX_FFT_COMPLEX_TO_REAL,
+ cfftgrid,fftgrid,thread,wcycle);
+ if (thread == 0)
+ {
+ wallcycle_stop(wcycle,ewcPME_FFT);
+
+ where();
+
+ if (pme->nodeid == 0)
+ {
+ ntot = pme->nkx*pme->nky*pme->nkz;
+ npme = ntot*log((real)ntot)/log(2.0);
+ inc_nrnb(nrnb,eNR_FFT,2*npme);
+ }
+
+ wallcycle_start(wcycle,ewcPME_SPREADGATHER);
+ }
+
+ copy_fftgrid_to_pmegrid(pme,fftgrid,grid,pme->nthread,thread);
+ }
+ }
+ /* End of thread parallel section.
+ * With MPI we have to synchronize here before gmx_sum_qgrid_dd.
+ */
+
+ if (flags & GMX_PME_CALC_F)
+ {
+ /* distribute local grid to all nodes */
+#ifdef GMX_MPI
+ if (pme->nnodes > 1) {
+ gmx_sum_qgrid_dd(pme,grid,GMX_SUM_QGRID_BACKWARD);
+ }
+#endif
+ where();
+
+ unwrap_periodic_pmegrid(pme,grid);
+
+ /* interpolate forces for our local atoms */
+
+ where();
+
+ /* If we are running without parallelization,
+ * atc->f is the actual force array, not a buffer,
+ * therefore we should not clear it.
+ */
+ bClearF = (q == 0 && PAR(cr));
+#pragma omp parallel for num_threads(pme->nthread) schedule(static)
+ for(thread=0; thread<pme->nthread; thread++)
+ {
+ gather_f_bsplines(pme,grid,bClearF,atc,
+ &atc->spline[thread],
+ pme->bFEP ? (q==0 ? 1.0-lambda : lambda) : 1.0);
+ }
+
+ where();
+
+ inc_nrnb(nrnb,eNR_GATHERFBSP,
+ pme->pme_order*pme->pme_order*pme->pme_order*pme->atc[0].n);
+ wallcycle_stop(wcycle,ewcPME_SPREADGATHER);
+ }
+
+ if (flags & GMX_PME_CALC_ENER_VIR)
+ {
+ /* This should only be called on the master thread
+ * and after the threads have synchronized.
+ */
+ get_pme_ener_vir(pme,pme->nthread,&energy_AB[q],vir_AB[q]);
+ }
+ } /* of q-loop */
+
+ if ((flags & GMX_PME_CALC_F) && pme->nnodes > 1) {
+ wallcycle_start(wcycle,ewcPME_REDISTXF);
+ for(d=0; d<pme->ndecompdim; d++)
+ {
+ atc = &pme->atc[d];
+ if (d == pme->ndecompdim - 1)
+ {
+ n_d = homenr;
+ f_d = f + start;
+ }
+ else
+ {
+ n_d = pme->atc[d+1].n;
+ f_d = pme->atc[d+1].f;
+ }
+ if (DOMAINDECOMP(cr)) {
+ dd_pmeredist_f(pme,atc,n_d,f_d,
+ d==pme->ndecompdim-1 && pme->bPPnode);
+ } else {
+ pmeredist_pd(pme, FALSE, n_d, TRUE, f_d, NULL, atc);
+ }
+ }
+
+ wallcycle_stop(wcycle,ewcPME_REDISTXF);
+ }
+ where();
+
+ if (!pme->bFEP) {
+ *energy = energy_AB[0];
+ m_add(vir,vir_AB[0],vir);
+ } else {
+ *energy = (1.0-lambda)*energy_AB[0] + lambda*energy_AB[1];
+ *dvdlambda += energy_AB[1] - energy_AB[0];
+ for(i=0; i<DIM; i++)
+ for(j=0; j<DIM; j++)
+ vir[i][j] += (1.0-lambda)*vir_AB[0][i][j] + lambda*vir_AB[1][i][j];
+ }
+
+ if (debug)
+ {
+ fprintf(debug,"PME mesh energy: %g\n",*energy);
+ }
+
+ return 0;
+}
--- /dev/null
- #ifdef GMX_THREADS
+/*
+ *
+ * This source code is part of
+ *
+ * G R O M A C S
+ *
+ * GROningen MAchine for Chemical Simulations
+ *
+ * VERSION 3.2.0
+ * Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
+ * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
+ * Copyright (c) 2001-2004, The GROMACS development team,
+ * check out http://www.gromacs.org for more information.
+
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version 2
+ * of the License, or (at your option) any later version.
+ *
+ * If you want to redistribute modifications, please consider that
+ * scientific software is very special. Version control is crucial -
+ * bugs must be traceable. We will be happy to consider code for
+ * inclusion in the official distribution, but derived work must not
+ * be called official GROMACS. Details are found in the README & COPYING
+ * files - if they are missing, get the official version at www.gromacs.org.
+ *
+ * To help us fund GROMACS development, we humbly ask that you cite
+ * the papers on the package - you can find them in the top README file.
+ *
+ * For more info, check our website at http://www.gromacs.org
+ *
+ * And Hey:
+ * GROwing Monsters And Cloning Shrimps
+ */
+
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+
+#include <stdio.h>
+#include <string.h>
+#include <math.h>
+#include "typedefs.h"
+#include "smalloc.h"
+#include "gmx_fatal.h"
+#include "vec.h"
+#include "pme.h"
+#include "network.h"
+#include "domdec.h"
+#include "sighandler.h"
+
+#ifdef GMX_LIB_MPI
+#include <mpi.h>
+#endif
++#ifdef GMX_THREAD_MPI
+#include "tmpi.h"
+#endif
+
+#define PP_PME_CHARGE (1<<0)
+#define PP_PME_CHARGEB (1<<1)
+#define PP_PME_COORD (1<<2)
+#define PP_PME_FEP (1<<3)
+#define PP_PME_ENER_VIR (1<<4)
+#define PP_PME_FINISH (1<<5)
+
+#define PME_PP_SIGSTOP (1<<0)
+#define PME_PP_SIGSTOPNSS (1<<1)
+
+typedef struct gmx_pme_pp {
+#ifdef GMX_MPI
+ MPI_Comm mpi_comm_mysim;
+#endif
+ int nnode; /* The number of PP node to communicate with */
+ int *node; /* The PP node ranks */
+ int node_peer; /* The peer PP node rank */
+ int *nat; /* The number of atom for each PP node */
+ int flags_charge; /* The flags sent along with the last charges */
+ real *chargeA;
+ real *chargeB;
+ rvec *x;
+ rvec *f;
+ int nalloc;
+#ifdef GMX_MPI
+ MPI_Request *req;
+ MPI_Status *stat;
+#endif
+} t_gmx_pme_pp;
+
+typedef struct gmx_pme_comm_n_box {
+ int natoms;
+ matrix box;
+ int maxshift_x;
+ int maxshift_y;
+ real lambda;
+ int flags;
+ gmx_large_int_t step;
+} gmx_pme_comm_n_box_t;
+
+typedef struct {
+ matrix vir;
+ real energy;
+ real dvdlambda;
+ float cycles;
+ gmx_stop_cond_t stop_cond;
+} gmx_pme_comm_vir_ene_t;
+
+
+
+
+gmx_pme_pp_t gmx_pme_pp_init(t_commrec *cr)
+{
+ struct gmx_pme_pp *pme_pp;
+ int rank;
+
+ snew(pme_pp,1);
+
+#ifdef GMX_MPI
+ pme_pp->mpi_comm_mysim = cr->mpi_comm_mysim;
+ MPI_Comm_rank(cr->mpi_comm_mygroup,&rank);
+ get_pme_ddnodes(cr,rank,&pme_pp->nnode,&pme_pp->node,&pme_pp->node_peer);
+ snew(pme_pp->nat,pme_pp->nnode);
+ snew(pme_pp->req,2*pme_pp->nnode);
+ snew(pme_pp->stat,2*pme_pp->nnode);
+ pme_pp->nalloc = 0;
+ pme_pp->flags_charge = 0;
+#endif
+
+ return pme_pp;
+}
+
+/* This should be faster with a real non-blocking MPI implementation */
+/* #define GMX_PME_DELAYED_WAIT */
+
+static void gmx_pme_send_q_x_wait(gmx_domdec_t *dd)
+{
+#ifdef GMX_MPI
+ if (dd->nreq_pme) {
+ MPI_Waitall(dd->nreq_pme,dd->req_pme,MPI_STATUSES_IGNORE);
+ dd->nreq_pme = 0;
+ }
+#endif
+}
+
+static void gmx_pme_send_q_x(t_commrec *cr, int flags,
+ real *chargeA, real *chargeB,
+ matrix box, rvec *x,
+ real lambda,
+ int maxshift_x, int maxshift_y,
+ gmx_large_int_t step)
+{
+ gmx_domdec_t *dd;
+ gmx_pme_comm_n_box_t *cnb;
+ int n;
+
+ dd = cr->dd;
+ n = dd->nat_home;
+
+ if (debug)
+ fprintf(debug,"PP node %d sending to PME node %d: %d%s%s\n",
+ cr->sim_nodeid,dd->pme_nodeid,n,
+ flags & PP_PME_CHARGE ? " charges" : "",
+ flags & PP_PME_COORD ? " coordinates" : "");
+
+#ifdef GMX_PME_DELAYED_WAIT
+ /* When can not use cnb until pending communication has finished */
+ gmx_pme_send_x_q_wait(dd);
+#endif
+
+ if (dd->pme_receive_vir_ener) {
+ /* Peer PP node: communicate all data */
+ if (dd->cnb == NULL)
+ snew(dd->cnb,1);
+ cnb = dd->cnb;
+
+ cnb->flags = flags;
+ cnb->natoms = n;
+ cnb->maxshift_x = maxshift_x;
+ cnb->maxshift_y = maxshift_y;
+ cnb->lambda = lambda;
+ cnb->step = step;
+ if (flags & PP_PME_COORD)
+ copy_mat(box,cnb->box);
+#ifdef GMX_MPI
+ MPI_Isend(cnb,sizeof(*cnb),MPI_BYTE,
+ dd->pme_nodeid,0,cr->mpi_comm_mysim,
+ &dd->req_pme[dd->nreq_pme++]);
+#endif
+ } else if (flags & PP_PME_CHARGE) {
+#ifdef GMX_MPI
+ /* Communicate only the number of atoms */
+ MPI_Isend(&n,sizeof(n),MPI_BYTE,
+ dd->pme_nodeid,0,cr->mpi_comm_mysim,
+ &dd->req_pme[dd->nreq_pme++]);
+#endif
+ }
+
+#ifdef GMX_MPI
+ if (n > 0) {
+ if (flags & PP_PME_CHARGE) {
+ MPI_Isend(chargeA,n*sizeof(real),MPI_BYTE,
+ dd->pme_nodeid,1,cr->mpi_comm_mysim,
+ &dd->req_pme[dd->nreq_pme++]);
+ }
+ if (flags & PP_PME_CHARGEB) {
+ MPI_Isend(chargeB,n*sizeof(real),MPI_BYTE,
+ dd->pme_nodeid,2,cr->mpi_comm_mysim,
+ &dd->req_pme[dd->nreq_pme++]);
+ }
+ if (flags & PP_PME_COORD) {
+ MPI_Isend(x[0],n*sizeof(rvec),MPI_BYTE,
+ dd->pme_nodeid,3,cr->mpi_comm_mysim,
+ &dd->req_pme[dd->nreq_pme++]);
+ }
+ }
+
+#ifndef GMX_PME_DELAYED_WAIT
+ /* Wait for the data to arrive */
+ /* We can skip this wait as we are sure x and q will not be modified
+ * before the next call to gmx_pme_send_x_q or gmx_pme_receive_f.
+ */
+ gmx_pme_send_q_x_wait(dd);
+#endif
+#endif
+}
+
+void gmx_pme_send_q(t_commrec *cr,
+ gmx_bool bFreeEnergy, real *chargeA, real *chargeB,
+ int maxshift_x, int maxshift_y)
+{
+ int flags;
+
+ flags = PP_PME_CHARGE;
+ if (bFreeEnergy)
+ flags |= PP_PME_CHARGEB;
+
+ gmx_pme_send_q_x(cr,flags,
+ chargeA,chargeB,NULL,NULL,0,maxshift_x,maxshift_y,-1);
+}
+
+void gmx_pme_send_x(t_commrec *cr, matrix box, rvec *x,
+ gmx_bool bFreeEnergy, real lambda,
+ gmx_bool bEnerVir,
+ gmx_large_int_t step)
+{
+ int flags;
+
+ flags = PP_PME_COORD;
+ if (bFreeEnergy)
+ flags |= PP_PME_FEP;
+ if (bEnerVir)
+ flags |= PP_PME_ENER_VIR;
+
+ gmx_pme_send_q_x(cr,flags,NULL,NULL,box,x,lambda,0,0,step);
+}
+
+void gmx_pme_finish(t_commrec *cr)
+{
+ int flags;
+
+ flags = PP_PME_FINISH;
+
+ gmx_pme_send_q_x(cr,flags,NULL,NULL,NULL,NULL,0,0,0,-1);
+}
+
+int gmx_pme_recv_q_x(struct gmx_pme_pp *pme_pp,
+ real **chargeA, real **chargeB,
+ matrix box, rvec **x,rvec **f,
+ int *maxshift_x, int *maxshift_y,
+ gmx_bool *bFreeEnergy,real *lambda,
+ gmx_bool *bEnerVir,
+ gmx_large_int_t *step)
+{
+ gmx_pme_comm_n_box_t cnb;
+ int nat=0,q,messages,sender;
+ real *charge_pp;
+
+ messages = 0;
+
+ /* avoid compiler warning about unused variable without MPI support */
+ cnb.flags = 0;
+#ifdef GMX_MPI
+ do {
+ /* Receive the send count, box and time step from the peer PP node */
+ MPI_Recv(&cnb,sizeof(cnb),MPI_BYTE,
+ pme_pp->node_peer,0,
+ pme_pp->mpi_comm_mysim,MPI_STATUS_IGNORE);
+
+ if (debug)
+ fprintf(debug,"PME only node receiving:%s%s%s\n",
+ (cnb.flags & PP_PME_CHARGE) ? " charges" : "",
+ (cnb.flags & PP_PME_COORD ) ? " coordinates" : "",
+ (cnb.flags & PP_PME_FINISH) ? " finish" : "");
+
+ if (cnb.flags & PP_PME_CHARGE) {
+ /* Receive the send counts from the other PP nodes */
+ for(sender=0; sender<pme_pp->nnode; sender++) {
+ if (pme_pp->node[sender] == pme_pp->node_peer) {
+ pme_pp->nat[sender] = cnb.natoms;
+ } else {
+ MPI_Irecv(&(pme_pp->nat[sender]),sizeof(pme_pp->nat[0]),
+ MPI_BYTE,
+ pme_pp->node[sender],0,
+ pme_pp->mpi_comm_mysim,&pme_pp->req[messages++]);
+ }
+ }
+ MPI_Waitall(messages, pme_pp->req, pme_pp->stat);
+ messages = 0;
+
+ nat = 0;
+ for(sender=0; sender<pme_pp->nnode; sender++)
+ nat += pme_pp->nat[sender];
+
+ if (nat > pme_pp->nalloc) {
+ pme_pp->nalloc = over_alloc_dd(nat);
+ srenew(pme_pp->chargeA,pme_pp->nalloc);
+ if (cnb.flags & PP_PME_CHARGEB)
+ srenew(pme_pp->chargeB,pme_pp->nalloc);
+ srenew(pme_pp->x,pme_pp->nalloc);
+ srenew(pme_pp->f,pme_pp->nalloc);
+ }
+
+ /* maxshift is sent when the charges are sent */
+ *maxshift_x = cnb.maxshift_x;
+ *maxshift_y = cnb.maxshift_y;
+
+ /* Receive the charges in place */
+ for(q=0; q<((cnb.flags & PP_PME_CHARGEB) ? 2 : 1); q++) {
+ if (q == 0)
+ charge_pp = pme_pp->chargeA;
+ else
+ charge_pp = pme_pp->chargeB;
+ nat = 0;
+ for(sender=0; sender<pme_pp->nnode; sender++) {
+ if (pme_pp->nat[sender] > 0) {
+ MPI_Irecv(charge_pp+nat,
+ pme_pp->nat[sender]*sizeof(real),
+ MPI_BYTE,
+ pme_pp->node[sender],1+q,
+ pme_pp->mpi_comm_mysim,
+ &pme_pp->req[messages++]);
+ nat += pme_pp->nat[sender];
+ if (debug)
+ fprintf(debug,"Received from PP node %d: %d "
+ "charges\n",
+ pme_pp->node[sender],pme_pp->nat[sender]);
+ }
+ }
+ }
+
+ pme_pp->flags_charge = cnb.flags;
+ }
+
+ if (cnb.flags & PP_PME_COORD) {
+ if (!(pme_pp->flags_charge & PP_PME_CHARGE))
+ gmx_incons("PME-only node received coordinates before charges"
+ );
+
+ /* The box, FE flag and lambda are sent along with the coordinates
+ * */
+ copy_mat(cnb.box,box);
+ *bFreeEnergy = (cnb.flags & PP_PME_FEP);
+ *lambda = cnb.lambda;
+ *bEnerVir = (cnb.flags & PP_PME_ENER_VIR);
+
+ if (*bFreeEnergy && !(pme_pp->flags_charge & PP_PME_CHARGEB))
+ gmx_incons("PME-only node received free energy request, but "
+ "did not receive B-state charges");
+
+ /* Receive the coordinates in place */
+ nat = 0;
+ for(sender=0; sender<pme_pp->nnode; sender++) {
+ if (pme_pp->nat[sender] > 0) {
+ MPI_Irecv(pme_pp->x[nat],pme_pp->nat[sender]*sizeof(rvec),
+ MPI_BYTE,
+ pme_pp->node[sender],3,
+ pme_pp->mpi_comm_mysim,&pme_pp->req[messages++]);
+ nat += pme_pp->nat[sender];
+ if (debug)
+ fprintf(debug,"Received from PP node %d: %d "
+ "coordinates\n",
+ pme_pp->node[sender],pme_pp->nat[sender]);
+ }
+ }
+ }
+
+ /* Wait for the coordinates and/or charges to arrive */
+ MPI_Waitall(messages, pme_pp->req, pme_pp->stat);
+ messages = 0;
+ } while (!(cnb.flags & (PP_PME_COORD | PP_PME_FINISH)));
+
+ *step = cnb.step;
+#endif
+
+ *chargeA = pme_pp->chargeA;
+ *chargeB = pme_pp->chargeB;
+ *x = pme_pp->x;
+ *f = pme_pp->f;
+
+
+ return ((cnb.flags & PP_PME_FINISH) ? -1 : nat);
+}
+
+static void receive_virial_energy(t_commrec *cr,
+ matrix vir,real *energy,real *dvdlambda,
+ float *pme_cycles)
+{
+ gmx_pme_comm_vir_ene_t cve;
+
+ if (cr->dd->pme_receive_vir_ener) {
+ if (debug)
+ fprintf(debug,
+ "PP node %d receiving from PME node %d: virial and energy\n",
+ cr->sim_nodeid,cr->dd->pme_nodeid);
+#ifdef GMX_MPI
+ MPI_Recv(&cve,sizeof(cve),MPI_BYTE,cr->dd->pme_nodeid,1,cr->mpi_comm_mysim,
+ MPI_STATUS_IGNORE);
+#else
+ memset(&cve,0,sizeof(cve));
+#endif
+
+ m_add(vir,cve.vir,vir);
+ *energy = cve.energy;
+ *dvdlambda += cve.dvdlambda;
+ *pme_cycles = cve.cycles;
+
+ if ( cve.stop_cond != gmx_stop_cond_none )
+ {
+ gmx_set_stop_condition(cve.stop_cond);
+ }
+ } else {
+ *energy = 0;
+ *pme_cycles = 0;
+ }
+}
+
+void gmx_pme_receive_f(t_commrec *cr,
+ rvec f[], matrix vir,
+ real *energy, real *dvdlambda,
+ float *pme_cycles)
+{
+ int natoms,i;
+
+#ifdef GMX_PME_DELAYED_WAIT
+ /* Wait for the x request to finish */
+ gmx_pme_send_q_x_wait(cr->dd);
+#endif
+
+ natoms = cr->dd->nat_home;
+
+ if (natoms > cr->dd->pme_recv_f_alloc)
+ {
+ cr->dd->pme_recv_f_alloc = over_alloc_dd(natoms);
+ srenew(cr->dd->pme_recv_f_buf, cr->dd->pme_recv_f_alloc);
+ }
+
+#ifdef GMX_MPI
+ MPI_Recv(cr->dd->pme_recv_f_buf[0],
+ natoms*sizeof(rvec),MPI_BYTE,
+ cr->dd->pme_nodeid,0,cr->mpi_comm_mysim,
+ MPI_STATUS_IGNORE);
+#endif
+
+ for(i=0; i<natoms; i++)
+ rvec_inc(f[i],cr->dd->pme_recv_f_buf[i]);
+
+
+ receive_virial_energy(cr,vir,energy,dvdlambda,pme_cycles);
+}
+
+void gmx_pme_send_force_vir_ener(struct gmx_pme_pp *pme_pp,
+ rvec *f, matrix vir,
+ real energy, real dvdlambda,
+ float cycles)
+{
+ gmx_pme_comm_vir_ene_t cve;
+ int messages,ind_start,ind_end,receiver;
+
+ cve.cycles = cycles;
+
+ /* Now the evaluated forces have to be transferred to the PP nodes */
+ messages = 0;
+ ind_end = 0;
+ for (receiver=0; receiver<pme_pp->nnode; receiver++) {
+ ind_start = ind_end;
+ ind_end = ind_start + pme_pp->nat[receiver];
+#ifdef GMX_MPI
+ if (MPI_Isend(f[ind_start],(ind_end-ind_start)*sizeof(rvec),MPI_BYTE,
+ pme_pp->node[receiver],0,
+ pme_pp->mpi_comm_mysim,&pme_pp->req[messages++]) != 0)
+ gmx_comm("MPI_Isend failed in do_pmeonly");
+#endif
+ }
+
+ /* send virial and energy to our last PP node */
+ copy_mat(vir,cve.vir);
+ cve.energy = energy;
+ cve.dvdlambda = dvdlambda;
+ /* check for the signals to send back to a PP node */
+ cve.stop_cond = gmx_get_stop_condition();
+
+ cve.cycles = cycles;
+
+ if (debug)
+ fprintf(debug,"PME node sending to PP node %d: virial and energy\n",
+ pme_pp->node_peer);
+#ifdef GMX_MPI
+ MPI_Isend(&cve,sizeof(cve),MPI_BYTE,
+ pme_pp->node_peer,1,
+ pme_pp->mpi_comm_mysim,&pme_pp->req[messages++]);
+
+ /* Wait for the forces to arrive */
+ MPI_Waitall(messages, pme_pp->req, pme_pp->stat);
+#endif
+}
--- /dev/null
- #ifdef GMX_THREADS
+/* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
+ *
+ *
+ * This source code is part of
+ *
+ * G R O M A C S
+ *
+ * GROningen MAchine for Chemical Simulations
+ *
+ * VERSION 3.2.0
+ * Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
+ * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
+ * Copyright (c) 2001-2004, The GROMACS development team,
+ * check out http://www.gromacs.org for more information.
+
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version 2
+ * of the License, or (at your option) any later version.
+ *
+ * If you want to redistribute modifications, please consider that
+ * scientific software is very special. Version control is crucial -
+ * bugs must be traceable. We will be happy to consider code for
+ * inclusion in the official distribution, but derived work must not
+ * be called official GROMACS. Details are found in the README & COPYING
+ * files - if they are missing, get the official version at www.gromacs.org.
+ *
+ * To help us fund GROMACS development, we humbly ask that you cite
+ * the papers on the package - you can find them in the top README file.
+ *
+ * For more info, check our website at http://www.gromacs.org
+ *
+ * And Hey:
+ * GROwing Monsters And Cloning Shrimps
+ */
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+#ifdef GMX_CRAY_XT3
+#include<catamount/dclock.h>
+#endif
+
+
+#include <stdio.h>
+#include <time.h>
+#ifdef HAVE_SYS_TIME_H
+#include <sys/time.h>
+#endif
+#include <math.h>
+#include "typedefs.h"
+#include "string2.h"
+#include "gmxfio.h"
+#include "smalloc.h"
+#include "names.h"
+#include "confio.h"
+#include "mvdata.h"
+#include "txtdump.h"
+#include "pbc.h"
+#include "chargegroup.h"
+#include "vec.h"
+#include "time.h"
+#include "nrnb.h"
+#include "mshift.h"
+#include "mdrun.h"
+#include "update.h"
+#include "physics.h"
+#include "main.h"
+#include "mdatoms.h"
+#include "force.h"
+#include "bondf.h"
+#include "pme.h"
+#include "pppm.h"
+#include "disre.h"
+#include "orires.h"
+#include "network.h"
+#include "calcmu.h"
+#include "constr.h"
+#include "xvgr.h"
+#include "trnio.h"
+#include "xtcio.h"
+#include "copyrite.h"
+#include "pull_rotation.h"
+#include "domdec.h"
+#include "partdec.h"
+#include "gmx_wallcycle.h"
+#include "genborn.h"
+
+#ifdef GMX_LIB_MPI
+#include <mpi.h>
+#endif
- #ifndef GMX_THREADS
++#ifdef GMX_THREAD_MPI
+#include "tmpi.h"
+#endif
+
+#include "adress.h"
+#include "qmmm.h"
+
+#if 0
+typedef struct gmx_timeprint {
+
+} t_gmx_timeprint;
+#endif
+
+/* Portable version of ctime_r implemented in src/gmxlib/string2.c, but we do not want it declared in public installed headers */
+char *
+gmx_ctime_r(const time_t *clock,char *buf, int n);
+
+
+double
+gmx_gettime()
+{
+#ifdef HAVE_GETTIMEOFDAY
+ struct timeval t;
+ double seconds;
+
+ gettimeofday(&t,NULL);
+
+ seconds = (double) t.tv_sec + 1e-6*(double)t.tv_usec;
+
+ return seconds;
+#else
+ double seconds;
+
+ seconds = time(NULL);
+
+ return seconds;
+#endif
+}
+
+
+#define difftime(end,start) ((double)(end)-(double)(start))
+
+void print_time(FILE *out,gmx_runtime_t *runtime,gmx_large_int_t step,
+ t_inputrec *ir, t_commrec *cr)
+{
+ time_t finish;
+ char timebuf[STRLEN];
+ double dt;
+ char buf[48];
+
- #ifndef GMX_THREADS
++#ifndef GMX_THREAD_MPI
+ if (!PAR(cr))
+#endif
+ {
+ fprintf(out,"\r");
+ }
+ fprintf(out,"step %s",gmx_step_str(step,buf));
+ if ((step >= ir->nstlist))
+ {
+ if ((ir->nstlist == 0) || ((step % ir->nstlist) == 0))
+ {
+ /* We have done a full cycle let's update time_per_step */
+ runtime->last = gmx_gettime();
+ dt = difftime(runtime->last,runtime->real);
+ runtime->time_per_step = dt/(step - ir->init_step + 1);
+ }
+ dt = (ir->nsteps + ir->init_step - step)*runtime->time_per_step;
+
+ if (ir->nsteps >= 0)
+ {
+ if (dt >= 300)
+ {
+ finish = (time_t) (runtime->last + dt);
+ gmx_ctime_r(&finish,timebuf,STRLEN);
+ sprintf(buf,"%s",timebuf);
+ buf[strlen(buf)-1]='\0';
+ fprintf(out,", will finish %s",buf);
+ }
+ else
+ fprintf(out,", remaining runtime: %5d s ",(int)dt);
+ }
+ else
+ {
+ fprintf(out," performance: %.1f ns/day ",
+ ir->delta_t/1000*24*60*60/runtime->time_per_step);
+ }
+ }
++#ifndef GMX_THREAD_MPI
+ if (PAR(cr))
+ {
+ fprintf(out,"\n");
+ }
+#endif
+
+ fflush(out);
+}
+
+#ifdef NO_CLOCK
+#define clock() -1
+#endif
+
+static double set_proctime(gmx_runtime_t *runtime)
+{
+ double diff;
+#ifdef GMX_CRAY_XT3
+ double prev;
+
+ prev = runtime->proc;
+ runtime->proc = dclock();
+
+ diff = runtime->proc - prev;
+#else
+ clock_t prev;
+
+ prev = runtime->proc;
+ runtime->proc = clock();
+
+ diff = (double)(runtime->proc - prev)/(double)CLOCKS_PER_SEC;
+#endif
+ if (diff < 0)
+ {
+ /* The counter has probably looped, ignore this data */
+ diff = 0;
+ }
+
+ return diff;
+}
+
+void runtime_start(gmx_runtime_t *runtime)
+{
+ runtime->real = gmx_gettime();
+ runtime->proc = 0;
+ set_proctime(runtime);
+ runtime->realtime = 0;
+ runtime->proctime = 0;
+ runtime->last = 0;
+ runtime->time_per_step = 0;
+}
+
+void runtime_end(gmx_runtime_t *runtime)
+{
+ double now;
+
+ now = gmx_gettime();
+
+ runtime->proctime += set_proctime(runtime);
+ runtime->realtime = now - runtime->real;
+ runtime->real = now;
+}
+
+void runtime_upd_proc(gmx_runtime_t *runtime)
+{
+ runtime->proctime += set_proctime(runtime);
+}
+
+void print_date_and_time(FILE *fplog,int nodeid,const char *title,
+ const gmx_runtime_t *runtime)
+{
+ int i;
+ char timebuf[STRLEN];
+ char time_string[STRLEN];
+ time_t tmptime;
+
+ if (fplog)
+ {
+ if (runtime != NULL)
+ {
+ tmptime = (time_t) runtime->real;
+ gmx_ctime_r(&tmptime,timebuf,STRLEN);
+ }
+ else
+ {
+ tmptime = (time_t) gmx_gettime();
+ gmx_ctime_r(&tmptime,timebuf,STRLEN);
+ }
+ for(i=0; timebuf[i]>=' '; i++)
+ {
+ time_string[i]=timebuf[i];
+ }
+ time_string[i]='\0';
+
+ fprintf(fplog,"%s on node %d %s\n",title,nodeid,time_string);
+ }
+}
+
+static void sum_forces(int start,int end,rvec f[],rvec flr[])
+{
+ int i;
+
+ if (gmx_debug_at) {
+ pr_rvecs(debug,0,"fsr",f+start,end-start);
+ pr_rvecs(debug,0,"flr",flr+start,end-start);
+ }
+ for(i=start; (i<end); i++)
+ rvec_inc(f[i],flr[i]);
+}
+
+/*
+ * calc_f_el calculates forces due to an electric field.
+ *
+ * force is kJ mol^-1 nm^-1 = e * kJ mol^-1 nm^-1 / e
+ *
+ * Et[] contains the parameters for the time dependent
+ * part of the field (not yet used).
+ * Ex[] contains the parameters for
+ * the spatial dependent part of the field. You can have cool periodic
+ * fields in principle, but only a constant field is supported
+ * now.
+ * The function should return the energy due to the electric field
+ * (if any) but for now returns 0.
+ *
+ * WARNING:
+ * There can be problems with the virial.
+ * Since the field is not self-consistent this is unavoidable.
+ * For neutral molecules the virial is correct within this approximation.
+ * For neutral systems with many charged molecules the error is small.
+ * But for systems with a net charge or a few charged molecules
+ * the error can be significant when the field is high.
+ * Solution: implement a self-consitent electric field into PME.
+ */
+static void calc_f_el(FILE *fp,int start,int homenr,
+ real charge[],rvec x[],rvec f[],
+ t_cosines Ex[],t_cosines Et[],double t)
+{
+ rvec Ext;
+ real t0;
+ int i,m;
+
+ for(m=0; (m<DIM); m++)
+ {
+ if (Et[m].n > 0)
+ {
+ if (Et[m].n == 3)
+ {
+ t0 = Et[m].a[1];
+ Ext[m] = cos(Et[m].a[0]*(t-t0))*exp(-sqr(t-t0)/(2.0*sqr(Et[m].a[2])));
+ }
+ else
+ {
+ Ext[m] = cos(Et[m].a[0]*t);
+ }
+ }
+ else
+ {
+ Ext[m] = 1.0;
+ }
+ if (Ex[m].n > 0)
+ {
+ /* Convert the field strength from V/nm to MD-units */
+ Ext[m] *= Ex[m].a[0]*FIELDFAC;
+ for(i=start; (i<start+homenr); i++)
+ f[i][m] += charge[i]*Ext[m];
+ }
+ else
+ {
+ Ext[m] = 0;
+ }
+ }
+ if (fp != NULL)
+ {
+ fprintf(fp,"%10g %10g %10g %10g #FIELD\n",t,
+ Ext[XX]/FIELDFAC,Ext[YY]/FIELDFAC,Ext[ZZ]/FIELDFAC);
+ }
+}
+
+static void calc_virial(FILE *fplog,int start,int homenr,rvec x[],rvec f[],
+ tensor vir_part,t_graph *graph,matrix box,
+ t_nrnb *nrnb,const t_forcerec *fr,int ePBC)
+{
+ int i,j;
+ tensor virtest;
+
+ /* The short-range virial from surrounding boxes */
+ clear_mat(vir_part);
+ calc_vir(fplog,SHIFTS,fr->shift_vec,fr->fshift,vir_part,ePBC==epbcSCREW,box);
+ inc_nrnb(nrnb,eNR_VIRIAL,SHIFTS);
+
+ /* Calculate partial virial, for local atoms only, based on short range.
+ * Total virial is computed in global_stat, called from do_md
+ */
+ f_calc_vir(fplog,start,start+homenr,x,f,vir_part,graph,box);
+ inc_nrnb(nrnb,eNR_VIRIAL,homenr);
+
+ /* Add position restraint contribution */
+ for(i=0; i<DIM; i++) {
+ vir_part[i][i] += fr->vir_diag_posres[i];
+ }
+
+ /* Add wall contribution */
+ for(i=0; i<DIM; i++) {
+ vir_part[i][ZZ] += fr->vir_wall_z[i];
+ }
+
+ if (debug)
+ pr_rvecs(debug,0,"vir_part",vir_part,DIM);
+}
+
+static void print_large_forces(FILE *fp,t_mdatoms *md,t_commrec *cr,
+ gmx_large_int_t step,real pforce,rvec *x,rvec *f)
+{
+ int i;
+ real pf2,fn2;
+ char buf[STEPSTRSIZE];
+
+ pf2 = sqr(pforce);
+ for(i=md->start; i<md->start+md->homenr; i++) {
+ fn2 = norm2(f[i]);
+ /* We also catch NAN, if the compiler does not optimize this away. */
+ if (fn2 >= pf2 || fn2 != fn2) {
+ fprintf(fp,"step %s atom %6d x %8.3f %8.3f %8.3f force %12.5e\n",
+ gmx_step_str(step,buf),
+ ddglatnr(cr->dd,i),x[i][XX],x[i][YY],x[i][ZZ],sqrt(fn2));
+ }
+ }
+}
+
+void do_force(FILE *fplog,t_commrec *cr,
+ t_inputrec *inputrec,
+ gmx_large_int_t step,t_nrnb *nrnb,gmx_wallcycle_t wcycle,
+ gmx_localtop_t *top,
+ gmx_mtop_t *mtop,
+ gmx_groups_t *groups,
+ matrix box,rvec x[],history_t *hist,
+ rvec f[],
+ tensor vir_force,
+ t_mdatoms *mdatoms,
+ gmx_enerdata_t *enerd,t_fcdata *fcd,
+ real lambda,t_graph *graph,
+ t_forcerec *fr,gmx_vsite_t *vsite,rvec mu_tot,
+ double t,FILE *field,gmx_edsam_t ed,
+ gmx_bool bBornRadii,
+ int flags)
+{
+ int cg0,cg1,i,j;
+ int start,homenr;
+ double mu[2*DIM];
+ gmx_bool bSepDVDL,bStateChanged,bNS,bFillGrid,bCalcCGCM,bBS;
+ gmx_bool bDoLongRange,bDoForces,bSepLRF;
+ gmx_bool bDoAdressWF;
+ matrix boxs;
+ real e,v,dvdl;
+ t_pbc pbc;
+ float cycles_ppdpme,cycles_pme,cycles_seppme,cycles_force;
+
+ start = mdatoms->start;
+ homenr = mdatoms->homenr;
+
+ bSepDVDL = (fr->bSepDVDL && do_per_step(step,inputrec->nstlog));
+
+ clear_mat(vir_force);
+
+ if (PARTDECOMP(cr))
+ {
+ pd_cg_range(cr,&cg0,&cg1);
+ }
+ else
+ {
+ cg0 = 0;
+ if (DOMAINDECOMP(cr))
+ {
+ cg1 = cr->dd->ncg_tot;
+ }
+ else
+ {
+ cg1 = top->cgs.nr;
+ }
+ if (fr->n_tpi > 0)
+ {
+ cg1--;
+ }
+ }
+
+ bStateChanged = (flags & GMX_FORCE_STATECHANGED);
+ bNS = (flags & GMX_FORCE_NS) && (fr->bAllvsAll==FALSE);
+ bFillGrid = (bNS && bStateChanged);
+ bCalcCGCM = (bFillGrid && !DOMAINDECOMP(cr));
+ bDoLongRange = (fr->bTwinRange && bNS && (flags & GMX_FORCE_DOLR));
+ bDoForces = (flags & GMX_FORCE_FORCES);
+ bSepLRF = (bDoLongRange && bDoForces && (flags & GMX_FORCE_SEPLRF));
+ /* should probably move this to the forcerec since it doesn't change */
+ bDoAdressWF = ((fr->adress_type!=eAdressOff));
+
+ if (bStateChanged)
+ {
+ update_forcerec(fplog,fr,box);
+
+ /* Calculate total (local) dipole moment in a temporary common array.
+ * This makes it possible to sum them over nodes faster.
+ */
+ calc_mu(start,homenr,
+ x,mdatoms->chargeA,mdatoms->chargeB,mdatoms->nChargePerturbed,
+ mu,mu+DIM);
+ }
+
+ if (fr->ePBC != epbcNONE) {
+ /* Compute shift vectors every step,
+ * because of pressure coupling or box deformation!
+ */
+ if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
+ calc_shifts(box,fr->shift_vec);
+
+ if (bCalcCGCM) {
+ put_charge_groups_in_box(fplog,cg0,cg1,fr->ePBC,box,
+ &(top->cgs),x,fr->cg_cm);
+ inc_nrnb(nrnb,eNR_CGCM,homenr);
+ inc_nrnb(nrnb,eNR_RESETX,cg1-cg0);
+ }
+ else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph) {
+ unshift_self(graph,box,x);
+ }
+ }
+ else if (bCalcCGCM) {
+ calc_cgcm(fplog,cg0,cg1,&(top->cgs),x,fr->cg_cm);
+ inc_nrnb(nrnb,eNR_CGCM,homenr);
+ }
+
+ if (bCalcCGCM) {
+ if (PAR(cr)) {
+ move_cgcm(fplog,cr,fr->cg_cm);
+ }
+ if (gmx_debug_at)
+ pr_rvecs(debug,0,"cgcm",fr->cg_cm,top->cgs.nr);
+ }
+
+#ifdef GMX_MPI
+ if (!(cr->duty & DUTY_PME)) {
+ /* Send particle coordinates to the pme nodes.
+ * Since this is only implemented for domain decomposition
+ * and domain decomposition does not use the graph,
+ * we do not need to worry about shifting.
+ */
+
+ wallcycle_start(wcycle,ewcPP_PMESENDX);
+
+ bBS = (inputrec->nwall == 2);
+ if (bBS) {
+ copy_mat(box,boxs);
+ svmul(inputrec->wall_ewald_zfac,boxs[ZZ],boxs[ZZ]);
+ }
+
+ gmx_pme_send_x(cr,bBS ? boxs : box,x,
+ mdatoms->nChargePerturbed,lambda,
+ ( flags & GMX_FORCE_VIRIAL),step);
+
+ wallcycle_stop(wcycle,ewcPP_PMESENDX);
+ }
+#endif /* GMX_MPI */
+
+ /* Communicate coordinates and sum dipole if necessary */
+ if (PAR(cr))
+ {
+ wallcycle_start(wcycle,ewcMOVEX);
+ if (DOMAINDECOMP(cr))
+ {
+ dd_move_x(cr->dd,box,x);
+ }
+ else
+ {
+ move_x(fplog,cr,GMX_LEFT,GMX_RIGHT,x,nrnb);
+ }
+ /* When we don't need the total dipole we sum it in global_stat */
+ if (bStateChanged && NEED_MUTOT(*inputrec))
+ {
+ gmx_sumd(2*DIM,mu,cr);
+ }
+ wallcycle_stop(wcycle,ewcMOVEX);
+ }
+ if (bStateChanged)
+ {
+
+ /* update adress weight beforehand */
+ if(bDoAdressWF)
+ {
+ /* need pbc for adress weight calculation with pbc_dx */
+ set_pbc(&pbc,inputrec->ePBC,box);
+ if(fr->adress_site == eAdressSITEcog)
+ {
+ update_adress_weights_cog(top->idef.iparams,top->idef.il,x,fr,mdatoms,
+ inputrec->ePBC==epbcNONE ? NULL : &pbc);
+ }
+ else if (fr->adress_site == eAdressSITEcom)
+ {
+ update_adress_weights_com(fplog,cg0,cg1,&(top->cgs),x,fr,mdatoms,
+ inputrec->ePBC==epbcNONE ? NULL : &pbc);
+ }
+ else if (fr->adress_site == eAdressSITEatomatom){
+ update_adress_weights_atom_per_atom(cg0,cg1,&(top->cgs),x,fr,mdatoms,
+ inputrec->ePBC==epbcNONE ? NULL : &pbc);
+ }
+ else
+ {
+ update_adress_weights_atom(cg0,cg1,&(top->cgs),x,fr,mdatoms,
+ inputrec->ePBC==epbcNONE ? NULL : &pbc);
+ }
+ }
+
+ for(i=0; i<2; i++)
+ {
+ for(j=0;j<DIM;j++)
+ {
+ fr->mu_tot[i][j] = mu[i*DIM + j];
+ }
+ }
+ }
+ if (fr->efep == efepNO)
+ {
+ copy_rvec(fr->mu_tot[0],mu_tot);
+ }
+ else
+ {
+ for(j=0; j<DIM; j++)
+ {
+ mu_tot[j] =
+ (1.0 - lambda)*fr->mu_tot[0][j] + lambda*fr->mu_tot[1][j];
+ }
+ }
+
+ /* Reset energies */
+ reset_enerdata(&(inputrec->opts),fr,bNS,enerd,MASTER(cr));
+ clear_rvecs(SHIFTS,fr->fshift);
+
+ if (bNS)
+ {
+ wallcycle_start(wcycle,ewcNS);
+
+ if (graph && bStateChanged)
+ {
+ /* Calculate intramolecular shift vectors to make molecules whole */
+ mk_mshift(fplog,graph,fr->ePBC,box,x);
+ }
+
+ /* Reset long range forces if necessary */
+ if (fr->bTwinRange)
+ {
+ /* Reset the (long-range) forces if necessary */
+ clear_rvecs(fr->natoms_force_constr,bSepLRF ? fr->f_twin : f);
+ }
+
+ /* Do the actual neighbour searching and if twin range electrostatics
+ * also do the calculation of long range forces and energies.
+ */
+ dvdl = 0;
+ ns(fplog,fr,x,box,
+ groups,&(inputrec->opts),top,mdatoms,
+ cr,nrnb,lambda,&dvdl,&enerd->grpp,bFillGrid,
+ bDoLongRange,bDoForces,bSepLRF ? fr->f_twin : f);
+ if (bSepDVDL)
+ {
+ fprintf(fplog,sepdvdlformat,"LR non-bonded",0.0,dvdl);
+ }
+ enerd->dvdl_lin += dvdl;
+
+ wallcycle_stop(wcycle,ewcNS);
+ }
+
+ if (inputrec->implicit_solvent && bNS)
+ {
+ make_gb_nblist(cr,inputrec->gb_algorithm,inputrec->rlist,
+ x,box,fr,&top->idef,graph,fr->born);
+ }
+
+ if (DOMAINDECOMP(cr))
+ {
+ if (!(cr->duty & DUTY_PME))
+ {
+ wallcycle_start(wcycle,ewcPPDURINGPME);
+ dd_force_flop_start(cr->dd,nrnb);
+ }
+ }
+
+ if (inputrec->bRot)
+ {
+ /* Enforced rotation has its own cycle counter that starts after the collective
+ * coordinates have been communicated. It is added to ddCyclF to allow
+ * for proper load-balancing */
+ wallcycle_start(wcycle,ewcROT);
+ do_rotation(cr,inputrec,box,x,t,step,wcycle,bNS);
+ wallcycle_stop(wcycle,ewcROT);
+ }
+
+ /* Start the force cycle counter.
+ * This counter is stopped in do_forcelow_level.
+ * No parallel communication should occur while this counter is running,
+ * since that will interfere with the dynamic load balancing.
+ */
+ wallcycle_start(wcycle,ewcFORCE);
+
+ if (bDoForces)
+ {
+ /* Reset forces for which the virial is calculated separately:
+ * PME/Ewald forces if necessary */
+ if (fr->bF_NoVirSum)
+ {
+ if (flags & GMX_FORCE_VIRIAL)
+ {
+ fr->f_novirsum = fr->f_novirsum_alloc;
+ if (fr->bDomDec)
+ {
+ clear_rvecs(fr->f_novirsum_n,fr->f_novirsum);
+ }
+ else
+ {
+ clear_rvecs(homenr,fr->f_novirsum+start);
+ }
+ }
+ else
+ {
+ /* We are not calculating the pressure so we do not need
+ * a separate array for forces that do not contribute
+ * to the pressure.
+ */
+ fr->f_novirsum = f;
+ }
+ }
+
+ if (bSepLRF)
+ {
+ /* Add the long range forces to the short range forces */
+ for(i=0; i<fr->natoms_force_constr; i++)
+ {
+ copy_rvec(fr->f_twin[i],f[i]);
+ }
+ }
+ else if (!(fr->bTwinRange && bNS))
+ {
+ /* Clear the short-range forces */
+ clear_rvecs(fr->natoms_force_constr,f);
+ }
+
+ clear_rvec(fr->vir_diag_posres);
+ }
+ if (inputrec->ePull == epullCONSTRAINT)
+ {
+ clear_pull_forces(inputrec->pull);
+ }
+
+ /* update QMMMrec, if necessary */
+ if(fr->bQMMM)
+ {
+ update_QMMMrec(cr,fr,x,mdatoms,box,top);
+ }
+
+ if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_POSRES].nr > 0)
+ {
+ /* Position restraints always require full pbc. Check if we already did it for Adress */
+ if(!(bStateChanged && bDoAdressWF))
+ {
+ set_pbc(&pbc,inputrec->ePBC,box);
+ }
+ v = posres(top->idef.il[F_POSRES].nr,top->idef.il[F_POSRES].iatoms,
+ top->idef.iparams_posres,
+ (const rvec*)x,fr->f_novirsum,fr->vir_diag_posres,
+ inputrec->ePBC==epbcNONE ? NULL : &pbc,lambda,&dvdl,
+ fr->rc_scaling,fr->ePBC,fr->posres_com,fr->posres_comB);
+ if (bSepDVDL)
+ {
+ fprintf(fplog,sepdvdlformat,
+ interaction_function[F_POSRES].longname,v,dvdl);
+ }
+ enerd->term[F_POSRES] += v;
+ /* This linear lambda dependence assumption is only correct
+ * when only k depends on lambda,
+ * not when the reference position depends on lambda.
+ * grompp checks for this.
+ */
+ enerd->dvdl_lin += dvdl;
+ inc_nrnb(nrnb,eNR_POSRES,top->idef.il[F_POSRES].nr/2);
+ }
+
+ /* Compute the bonded and non-bonded energies and optionally forces */
+ do_force_lowlevel(fplog,step,fr,inputrec,&(top->idef),
+ cr,nrnb,wcycle,mdatoms,&(inputrec->opts),
+ x,hist,f,enerd,fcd,mtop,top,fr->born,
+ &(top->atomtypes),bBornRadii,box,
+ lambda,graph,&(top->excls),fr->mu_tot,
+ flags,&cycles_pme);
+
+ cycles_force = wallcycle_stop(wcycle,ewcFORCE);
+
+ if (ed)
+ {
+ do_flood(fplog,cr,x,f,ed,box,step);
+ }
+
+ if (DOMAINDECOMP(cr))
+ {
+ dd_force_flop_stop(cr->dd,nrnb);
+ if (wcycle)
+ {
+ dd_cycles_add(cr->dd,cycles_force-cycles_pme,ddCyclF);
+ }
+ }
+
+ if (bDoForces)
+ {
+ if (IR_ELEC_FIELD(*inputrec))
+ {
+ /* Compute forces due to electric field */
+ calc_f_el(MASTER(cr) ? field : NULL,
+ start,homenr,mdatoms->chargeA,x,fr->f_novirsum,
+ inputrec->ex,inputrec->et,t);
+ }
+
+ if (bDoAdressWF && fr->adress_icor == eAdressICThermoForce)
+ {
+ /* Compute thermodynamic force in hybrid AdResS region */
+ adress_thermo_force(start,homenr,&(top->cgs),x,fr->f_novirsum,fr,mdatoms,
+ inputrec->ePBC==epbcNONE ? NULL : &pbc);
+ }
+
+ /* Communicate the forces */
+ if (PAR(cr))
+ {
+ wallcycle_start(wcycle,ewcMOVEF);
+ if (DOMAINDECOMP(cr))
+ {
+ dd_move_f(cr->dd,f,fr->fshift);
+ /* Do we need to communicate the separate force array
+ * for terms that do not contribute to the single sum virial?
+ * Position restraints and electric fields do not introduce
+ * inter-cg forces, only full electrostatics methods do.
+ * When we do not calculate the virial, fr->f_novirsum = f,
+ * so we have already communicated these forces.
+ */
+ if (EEL_FULL(fr->eeltype) && cr->dd->n_intercg_excl &&
+ (flags & GMX_FORCE_VIRIAL))
+ {
+ dd_move_f(cr->dd,fr->f_novirsum,NULL);
+ }
+ if (bSepLRF)
+ {
+ /* We should not update the shift forces here,
+ * since f_twin is already included in f.
+ */
+ dd_move_f(cr->dd,fr->f_twin,NULL);
+ }
+ }
+ else
+ {
+ pd_move_f(cr,f,nrnb);
+ if (bSepLRF)
+ {
+ pd_move_f(cr,fr->f_twin,nrnb);
+ }
+ }
+ wallcycle_stop(wcycle,ewcMOVEF);
+ }
+
+ /* If we have NoVirSum forces, but we do not calculate the virial,
+ * we sum fr->f_novirum=f later.
+ */
+ if (vsite && !(fr->bF_NoVirSum && !(flags & GMX_FORCE_VIRIAL)))
+ {
+ wallcycle_start(wcycle,ewcVSITESPREAD);
+ spread_vsite_f(fplog,vsite,x,f,fr->fshift,nrnb,
+ &top->idef,fr->ePBC,fr->bMolPBC,graph,box,cr);
+ wallcycle_stop(wcycle,ewcVSITESPREAD);
+
+ if (bSepLRF)
+ {
+ wallcycle_start(wcycle,ewcVSITESPREAD);
+ spread_vsite_f(fplog,vsite,x,fr->f_twin,NULL,
+ nrnb,
+ &top->idef,fr->ePBC,fr->bMolPBC,graph,box,cr);
+ wallcycle_stop(wcycle,ewcVSITESPREAD);
+ }
+ }
+
+ if (flags & GMX_FORCE_VIRIAL)
+ {
+ /* Calculation of the virial must be done after vsites! */
+ calc_virial(fplog,mdatoms->start,mdatoms->homenr,x,f,
+ vir_force,graph,box,nrnb,fr,inputrec->ePBC);
+ }
+ }
+
+ enerd->term[F_COM_PULL] = 0;
+ if (inputrec->ePull == epullUMBRELLA || inputrec->ePull == epullCONST_F)
+ {
+ /* Calculate the center of mass forces, this requires communication,
+ * which is why pull_potential is called close to other communication.
+ * The virial contribution is calculated directly,
+ * which is why we call pull_potential after calc_virial.
+ */
+ set_pbc(&pbc,inputrec->ePBC,box);
+ dvdl = 0;
+ enerd->term[F_COM_PULL] +=
+ pull_potential(inputrec->ePull,inputrec->pull,mdatoms,&pbc,
+ cr,t,lambda,x,f,vir_force,&dvdl);
+ if (bSepDVDL)
+ {
+ fprintf(fplog,sepdvdlformat,"Com pull",enerd->term[F_COM_PULL],dvdl);
+ }
+ enerd->dvdl_lin += dvdl;
+ }
+
+ /* Add the forces from enforced rotation potentials (if any) */
+ if (inputrec->bRot)
+ {
+ wallcycle_start(wcycle,ewcROTadd);
+ enerd->term[F_COM_PULL] += add_rot_forces(inputrec->rot, f, cr,step,t);
+ wallcycle_stop(wcycle,ewcROTadd);
+ }
+
+ if (PAR(cr) && !(cr->duty & DUTY_PME))
+ {
+ cycles_ppdpme = wallcycle_stop(wcycle,ewcPPDURINGPME);
+ dd_cycles_add(cr->dd,cycles_ppdpme,ddCyclPPduringPME);
+
+ /* In case of node-splitting, the PP nodes receive the long-range
+ * forces, virial and energy from the PME nodes here.
+ */
+ wallcycle_start(wcycle,ewcPP_PMEWAITRECVF);
+ dvdl = 0;
+ gmx_pme_receive_f(cr,fr->f_novirsum,fr->vir_el_recip,&e,&dvdl,
+ &cycles_seppme);
+ if (bSepDVDL)
+ {
+ fprintf(fplog,sepdvdlformat,"PME mesh",e,dvdl);
+ }
+ enerd->term[F_COUL_RECIP] += e;
+ enerd->dvdl_lin += dvdl;
+ if (wcycle)
+ {
+ dd_cycles_add(cr->dd,cycles_seppme,ddCyclPME);
+ }
+ wallcycle_stop(wcycle,ewcPP_PMEWAITRECVF);
+ }
+
+ if (bDoForces && fr->bF_NoVirSum)
+ {
+ if (vsite)
+ {
+ /* Spread the mesh force on virtual sites to the other particles...
+ * This is parallellized. MPI communication is performed
+ * if the constructing atoms aren't local.
+ */
+ wallcycle_start(wcycle,ewcVSITESPREAD);
+ spread_vsite_f(fplog,vsite,x,fr->f_novirsum,NULL,nrnb,
+ &top->idef,fr->ePBC,fr->bMolPBC,graph,box,cr);
+ wallcycle_stop(wcycle,ewcVSITESPREAD);
+ }
+ if (flags & GMX_FORCE_VIRIAL)
+ {
+ /* Now add the forces, this is local */
+ if (fr->bDomDec)
+ {
+ sum_forces(0,fr->f_novirsum_n,f,fr->f_novirsum);
+ }
+ else
+ {
+ sum_forces(start,start+homenr,f,fr->f_novirsum);
+ }
+ if (EEL_FULL(fr->eeltype))
+ {
+ /* Add the mesh contribution to the virial */
+ m_add(vir_force,fr->vir_el_recip,vir_force);
+ }
+ if (debug)
+ {
+ pr_rvecs(debug,0,"vir_force",vir_force,DIM);
+ }
+ }
+ }
+
+ /* Sum the potential energy terms from group contributions */
+ sum_epot(&(inputrec->opts),enerd);
+
+ if (fr->print_force >= 0 && bDoForces)
+ {
+ print_large_forces(stderr,mdatoms,cr,step,fr->print_force,x,f);
+ }
+}
+
+void do_constrain_first(FILE *fplog,gmx_constr_t constr,
+ t_inputrec *ir,t_mdatoms *md,
+ t_state *state,rvec *f,
+ t_graph *graph,t_commrec *cr,t_nrnb *nrnb,
+ t_forcerec *fr, gmx_localtop_t *top, tensor shake_vir)
+{
+ int i,m,start,end;
+ gmx_large_int_t step;
+ double mass,tmass,vcm[4];
+ real dt=ir->delta_t;
+ real dvdlambda;
+ rvec *savex;
+
+ snew(savex,state->natoms);
+
+ start = md->start;
+ end = md->homenr + start;
+
+ if (debug)
+ fprintf(debug,"vcm: start=%d, homenr=%d, end=%d\n",
+ start,md->homenr,end);
+ /* Do a first constrain to reset particles... */
+ step = ir->init_step;
+ if (fplog)
+ {
+ char buf[STEPSTRSIZE];
+ fprintf(fplog,"\nConstraining the starting coordinates (step %s)\n",
+ gmx_step_str(step,buf));
+ }
+ dvdlambda = 0;
+
+ /* constrain the current position */
+ constrain(NULL,TRUE,FALSE,constr,&(top->idef),
+ ir,NULL,cr,step,0,md,
+ state->x,state->x,NULL,
+ state->box,state->lambda,&dvdlambda,
+ NULL,NULL,nrnb,econqCoord,ir->epc==epcMTTK,state->veta,state->veta);
+ if (EI_VV(ir->eI))
+ {
+ /* constrain the inital velocity, and save it */
+ /* also may be useful if we need the ekin from the halfstep for velocity verlet */
+ /* might not yet treat veta correctly */
+ constrain(NULL,TRUE,FALSE,constr,&(top->idef),
+ ir,NULL,cr,step,0,md,
+ state->x,state->v,state->v,
+ state->box,state->lambda,&dvdlambda,
+ NULL,NULL,nrnb,econqVeloc,ir->epc==epcMTTK,state->veta,state->veta);
+ }
+ /* constrain the inital velocities at t-dt/2 */
+ if (EI_STATE_VELOCITY(ir->eI) && ir->eI!=eiVV)
+ {
+ for(i=start; (i<end); i++)
+ {
+ for(m=0; (m<DIM); m++)
+ {
+ /* Reverse the velocity */
+ state->v[i][m] = -state->v[i][m];
+ /* Store the position at t-dt in buf */
+ savex[i][m] = state->x[i][m] + dt*state->v[i][m];
+ }
+ }
+ /* Shake the positions at t=-dt with the positions at t=0
+ * as reference coordinates.
+ */
+ if (fplog)
+ {
+ char buf[STEPSTRSIZE];
+ fprintf(fplog,"\nConstraining the coordinates at t0-dt (step %s)\n",
+ gmx_step_str(step,buf));
+ }
+ dvdlambda = 0;
+ constrain(NULL,TRUE,FALSE,constr,&(top->idef),
+ ir,NULL,cr,step,-1,md,
+ state->x,savex,NULL,
+ state->box,state->lambda,&dvdlambda,
+ state->v,NULL,nrnb,econqCoord,ir->epc==epcMTTK,state->veta,state->veta);
+
+ for(i=start; i<end; i++) {
+ for(m=0; m<DIM; m++) {
+ /* Re-reverse the velocities */
+ state->v[i][m] = -state->v[i][m];
+ }
+ }
+ }
+
+ for(m=0; (m<4); m++)
+ vcm[m] = 0;
+ for(i=start; i<end; i++) {
+ mass = md->massT[i];
+ for(m=0; m<DIM; m++) {
+ vcm[m] += state->v[i][m]*mass;
+ }
+ vcm[3] += mass;
+ }
+
+ if (ir->nstcomm != 0 || debug) {
+ /* Compute the global sum of vcm */
+ if (debug)
+ fprintf(debug,"vcm: %8.3f %8.3f %8.3f,"
+ " total mass = %12.5e\n",vcm[XX],vcm[YY],vcm[ZZ],vcm[3]);
+ if (PAR(cr))
+ gmx_sumd(4,vcm,cr);
+ tmass = vcm[3];
+ for(m=0; (m<DIM); m++)
+ vcm[m] /= tmass;
+ if (debug)
+ fprintf(debug,"vcm: %8.3f %8.3f %8.3f,"
+ " total mass = %12.5e\n",vcm[XX],vcm[YY],vcm[ZZ],tmass);
+ if (ir->nstcomm != 0) {
+ /* Now we have the velocity of center of mass, let's remove it */
+ for(i=start; (i<end); i++) {
+ for(m=0; (m<DIM); m++)
+ state->v[i][m] -= vcm[m];
+ }
+
+ }
+ }
+ sfree(savex);
+}
+
+void calc_enervirdiff(FILE *fplog,int eDispCorr,t_forcerec *fr)
+{
+ double eners[2],virs[2],enersum,virsum,y0,f,g,h;
+ double r0,r1,r,rc3,rc9,ea,eb,ec,pa,pb,pc,pd;
+ double invscale,invscale2,invscale3;
+ int ri0,ri1,ri,i,offstart,offset;
+ real scale,*vdwtab;
+
+ fr->enershiftsix = 0;
+ fr->enershifttwelve = 0;
+ fr->enerdiffsix = 0;
+ fr->enerdifftwelve = 0;
+ fr->virdiffsix = 0;
+ fr->virdifftwelve = 0;
+
+ if (eDispCorr != edispcNO) {
+ for(i=0; i<2; i++) {
+ eners[i] = 0;
+ virs[i] = 0;
+ }
+ if ((fr->vdwtype == evdwSWITCH) || (fr->vdwtype == evdwSHIFT)) {
+ if (fr->rvdw_switch == 0)
+ gmx_fatal(FARGS,
+ "With dispersion correction rvdw-switch can not be zero "
+ "for vdw-type = %s",evdw_names[fr->vdwtype]);
+
+ scale = fr->nblists[0].tab.scale;
+ vdwtab = fr->nblists[0].vdwtab;
+
+ /* Round the cut-offs to exact table values for precision */
+ ri0 = floor(fr->rvdw_switch*scale);
+ ri1 = ceil(fr->rvdw*scale);
+ r0 = ri0/scale;
+ r1 = ri1/scale;
+ rc3 = r0*r0*r0;
+ rc9 = rc3*rc3*rc3;
+
+ if (fr->vdwtype == evdwSHIFT) {
+ /* Determine the constant energy shift below rvdw_switch */
+ fr->enershiftsix = (real)(-1.0/(rc3*rc3)) - vdwtab[8*ri0];
+ fr->enershifttwelve = (real)( 1.0/(rc9*rc3)) - vdwtab[8*ri0 + 4];
+ }
+ /* Add the constant part from 0 to rvdw_switch.
+ * This integration from 0 to rvdw_switch overcounts the number
+ * of interactions by 1, as it also counts the self interaction.
+ * We will correct for this later.
+ */
+ eners[0] += 4.0*M_PI*fr->enershiftsix*rc3/3.0;
+ eners[1] += 4.0*M_PI*fr->enershifttwelve*rc3/3.0;
+
+ invscale = 1.0/(scale);
+ invscale2 = invscale*invscale;
+ invscale3 = invscale*invscale2;
+
+ /* following summation derived from cubic spline definition,
+ Numerical Recipies in C, second edition, p. 113-116. Exact
+ for the cubic spline. We first calculate the negative of
+ the energy from rvdw to rvdw_switch, assuming that g(r)=1,
+ and then add the more standard, abrupt cutoff correction to
+ that result, yielding the long-range correction for a
+ switched function. We perform both the pressure and energy
+ loops at the same time for simplicity, as the computational
+ cost is low. */
+
+ for (i=0;i<2;i++) {
+ enersum = 0.0; virsum = 0.0;
+ if (i==0)
+ offstart = 0;
+ else
+ offstart = 4;
+ for (ri=ri0; ri<ri1; ri++) {
+ r = ri*invscale;
+ ea = invscale3;
+ eb = 2.0*invscale2*r;
+ ec = invscale*r*r;
+
+ pa = invscale3;
+ pb = 3.0*invscale2*r;
+ pc = 3.0*invscale*r*r;
+ pd = r*r*r;
+
+ /* this "8" is from the packing in the vdwtab array - perhaps
+ should be #define'ed? */
+ offset = 8*ri + offstart;
+ y0 = vdwtab[offset];
+ f = vdwtab[offset+1];
+ g = vdwtab[offset+2];
+ h = vdwtab[offset+3];
+
+ enersum += y0*(ea/3 + eb/2 + ec) + f*(ea/4 + eb/3 + ec/2)+
+ g*(ea/5 + eb/4 + ec/3) + h*(ea/6 + eb/5 + ec/4);
+ virsum += f*(pa/4 + pb/3 + pc/2 + pd) +
+ 2*g*(pa/5 + pb/4 + pc/3 + pd/2) + 3*h*(pa/6 + pb/5 + pc/4 + pd/3);
+
+ }
+ enersum *= 4.0*M_PI;
+ virsum *= 4.0*M_PI;
+ eners[i] -= enersum;
+ virs[i] -= virsum;
+ }
+
+ /* now add the correction for rvdw_switch to infinity */
+ eners[0] += -4.0*M_PI/(3.0*rc3);
+ eners[1] += 4.0*M_PI/(9.0*rc9);
+ virs[0] += 8.0*M_PI/rc3;
+ virs[1] += -16.0*M_PI/(3.0*rc9);
+ }
+ else if ((fr->vdwtype == evdwCUT) || (fr->vdwtype == evdwUSER)) {
+ if (fr->vdwtype == evdwUSER && fplog)
+ fprintf(fplog,
+ "WARNING: using dispersion correction with user tables\n");
+ rc3 = fr->rvdw*fr->rvdw*fr->rvdw;
+ rc9 = rc3*rc3*rc3;
+ eners[0] += -4.0*M_PI/(3.0*rc3);
+ eners[1] += 4.0*M_PI/(9.0*rc9);
+ virs[0] += 8.0*M_PI/rc3;
+ virs[1] += -16.0*M_PI/(3.0*rc9);
+ } else {
+ gmx_fatal(FARGS,
+ "Dispersion correction is not implemented for vdw-type = %s",
+ evdw_names[fr->vdwtype]);
+ }
+ fr->enerdiffsix = eners[0];
+ fr->enerdifftwelve = eners[1];
+ /* The 0.5 is due to the Gromacs definition of the virial */
+ fr->virdiffsix = 0.5*virs[0];
+ fr->virdifftwelve = 0.5*virs[1];
+ }
+}
+
+void calc_dispcorr(FILE *fplog,t_inputrec *ir,t_forcerec *fr,
+ gmx_large_int_t step,int natoms,
+ matrix box,real lambda,tensor pres,tensor virial,
+ real *prescorr, real *enercorr, real *dvdlcorr)
+{
+ gmx_bool bCorrAll,bCorrPres;
+ real dvdlambda,invvol,dens,ninter,avcsix,avctwelve,enerdiff,svir=0,spres=0;
+ int m;
+
+ *prescorr = 0;
+ *enercorr = 0;
+ *dvdlcorr = 0;
+
+ clear_mat(virial);
+ clear_mat(pres);
+
+ if (ir->eDispCorr != edispcNO) {
+ bCorrAll = (ir->eDispCorr == edispcAllEner ||
+ ir->eDispCorr == edispcAllEnerPres);
+ bCorrPres = (ir->eDispCorr == edispcEnerPres ||
+ ir->eDispCorr == edispcAllEnerPres);
+
+ invvol = 1/det(box);
+ if (fr->n_tpi)
+ {
+ /* Only correct for the interactions with the inserted molecule */
+ dens = (natoms - fr->n_tpi)*invvol;
+ ninter = fr->n_tpi;
+ }
+ else
+ {
+ dens = natoms*invvol;
+ ninter = 0.5*natoms;
+ }
+
+ if (ir->efep == efepNO)
+ {
+ avcsix = fr->avcsix[0];
+ avctwelve = fr->avctwelve[0];
+ }
+ else
+ {
+ avcsix = (1 - lambda)*fr->avcsix[0] + lambda*fr->avcsix[1];
+ avctwelve = (1 - lambda)*fr->avctwelve[0] + lambda*fr->avctwelve[1];
+ }
+
+ enerdiff = ninter*(dens*fr->enerdiffsix - fr->enershiftsix);
+ *enercorr += avcsix*enerdiff;
+ dvdlambda = 0.0;
+ if (ir->efep != efepNO)
+ {
+ dvdlambda += (fr->avcsix[1] - fr->avcsix[0])*enerdiff;
+ }
+ if (bCorrAll)
+ {
+ enerdiff = ninter*(dens*fr->enerdifftwelve - fr->enershifttwelve);
+ *enercorr += avctwelve*enerdiff;
+ if (fr->efep != efepNO)
+ {
+ dvdlambda += (fr->avctwelve[1] - fr->avctwelve[0])*enerdiff;
+ }
+ }
+
+ if (bCorrPres)
+ {
+ svir = ninter*dens*avcsix*fr->virdiffsix/3.0;
+ if (ir->eDispCorr == edispcAllEnerPres)
+ {
+ svir += ninter*dens*avctwelve*fr->virdifftwelve/3.0;
+ }
+ /* The factor 2 is because of the Gromacs virial definition */
+ spres = -2.0*invvol*svir*PRESFAC;
+
+ for(m=0; m<DIM; m++) {
+ virial[m][m] += svir;
+ pres[m][m] += spres;
+ }
+ *prescorr += spres;
+ }
+
+ /* Can't currently control when it prints, for now, just print when degugging */
+ if (debug)
+ {
+ if (bCorrAll) {
+ fprintf(debug,"Long Range LJ corr.: <C6> %10.4e, <C12> %10.4e\n",
+ avcsix,avctwelve);
+ }
+ if (bCorrPres)
+ {
+ fprintf(debug,
+ "Long Range LJ corr.: Epot %10g, Pres: %10g, Vir: %10g\n",
+ *enercorr,spres,svir);
+ }
+ else
+ {
+ fprintf(debug,"Long Range LJ corr.: Epot %10g\n",*enercorr);
+ }
+ }
+
+ if (fr->bSepDVDL && do_per_step(step,ir->nstlog))
+ {
+ fprintf(fplog,sepdvdlformat,"Dispersion correction",
+ *enercorr,dvdlambda);
+ }
+ if (fr->efep != efepNO)
+ {
+ *dvdlcorr += dvdlambda;
+ }
+ }
+}
+
+void do_pbc_first(FILE *fplog,matrix box,t_forcerec *fr,
+ t_graph *graph,rvec x[])
+{
+ if (fplog)
+ fprintf(fplog,"Removing pbc first time\n");
+ calc_shifts(box,fr->shift_vec);
+ if (graph) {
+ mk_mshift(fplog,graph,fr->ePBC,box,x);
+ if (gmx_debug_at)
+ p_graph(debug,"do_pbc_first 1",graph);
+ shift_self(graph,box,x);
+ /* By doing an extra mk_mshift the molecules that are broken
+ * because they were e.g. imported from another software
+ * will be made whole again. Such are the healing powers
+ * of GROMACS.
+ */
+ mk_mshift(fplog,graph,fr->ePBC,box,x);
+ if (gmx_debug_at)
+ p_graph(debug,"do_pbc_first 2",graph);
+ }
+ if (fplog)
+ fprintf(fplog,"Done rmpbc\n");
+}
+
+static void low_do_pbc_mtop(FILE *fplog,int ePBC,matrix box,
+ gmx_mtop_t *mtop,rvec x[],
+ gmx_bool bFirst)
+{
+ t_graph *graph;
+ int mb,as,mol;
+ gmx_molblock_t *molb;
+
+ if (bFirst && fplog)
+ fprintf(fplog,"Removing pbc first time\n");
+
+ snew(graph,1);
+ as = 0;
+ for(mb=0; mb<mtop->nmolblock; mb++) {
+ molb = &mtop->molblock[mb];
+ if (molb->natoms_mol == 1 ||
+ (!bFirst && mtop->moltype[molb->type].cgs.nr == 1)) {
+ /* Just one atom or charge group in the molecule, no PBC required */
+ as += molb->nmol*molb->natoms_mol;
+ } else {
+ /* Pass NULL iso fplog to avoid graph prints for each molecule type */
+ mk_graph_ilist(NULL,mtop->moltype[molb->type].ilist,
+ 0,molb->natoms_mol,FALSE,FALSE,graph);
+
+ for(mol=0; mol<molb->nmol; mol++) {
+ mk_mshift(fplog,graph,ePBC,box,x+as);
+
+ shift_self(graph,box,x+as);
+ /* The molecule is whole now.
+ * We don't need the second mk_mshift call as in do_pbc_first,
+ * since we no longer need this graph.
+ */
+
+ as += molb->natoms_mol;
+ }
+ done_graph(graph);
+ }
+ }
+ sfree(graph);
+}
+
+void do_pbc_first_mtop(FILE *fplog,int ePBC,matrix box,
+ gmx_mtop_t *mtop,rvec x[])
+{
+ low_do_pbc_mtop(fplog,ePBC,box,mtop,x,TRUE);
+}
+
+void do_pbc_mtop(FILE *fplog,int ePBC,matrix box,
+ gmx_mtop_t *mtop,rvec x[])
+{
+ low_do_pbc_mtop(fplog,ePBC,box,mtop,x,FALSE);
+}
+
+void finish_run(FILE *fplog,t_commrec *cr,const char *confout,
+ t_inputrec *inputrec,
+ t_nrnb nrnb[],gmx_wallcycle_t wcycle,
+ gmx_runtime_t *runtime,
+ gmx_bool bWriteStat)
+{
+ int i,j;
+ t_nrnb *nrnb_tot=NULL;
+ real delta_t;
+ double nbfs,mflop;
+ double cycles[ewcNR];
+
+ wallcycle_sum(cr,wcycle,cycles);
+
+ if (cr->nnodes > 1) {
+ if (SIMMASTER(cr))
+ snew(nrnb_tot,1);
+#ifdef GMX_MPI
+ MPI_Reduce(nrnb->n,nrnb_tot->n,eNRNB,MPI_DOUBLE,MPI_SUM,
+ MASTERRANK(cr),cr->mpi_comm_mysim);
+#endif
+ } else {
+ nrnb_tot = nrnb;
+ }
+
+ if (SIMMASTER(cr)) {
+ print_flop(fplog,nrnb_tot,&nbfs,&mflop);
+ if (cr->nnodes > 1) {
+ sfree(nrnb_tot);
+ }
+ }
+
+ if ((cr->duty & DUTY_PP) && DOMAINDECOMP(cr)) {
+ print_dd_statistics(cr,inputrec,fplog);
+ }
+
+#ifdef GMX_MPI
+ if (PARTDECOMP(cr))
+ {
+ if (MASTER(cr))
+ {
+ t_nrnb *nrnb_all;
+ int s;
+ MPI_Status stat;
+
+ snew(nrnb_all,cr->nnodes);
+ nrnb_all[0] = *nrnb;
+ for(s=1; s<cr->nnodes; s++)
+ {
+ MPI_Recv(nrnb_all[s].n,eNRNB,MPI_DOUBLE,s,0,
+ cr->mpi_comm_mysim,&stat);
+ }
+ pr_load(fplog,cr,nrnb_all);
+ sfree(nrnb_all);
+ }
+ else
+ {
+ MPI_Send(nrnb->n,eNRNB,MPI_DOUBLE,MASTERRANK(cr),0,
+ cr->mpi_comm_mysim);
+ }
+ }
+#endif
+
+ if (SIMMASTER(cr)) {
+ wallcycle_print(fplog,cr->nnodes,cr->npmenodes,runtime->realtime,
+ wcycle,cycles);
+
+ if (EI_DYNAMICS(inputrec->eI)) {
+ delta_t = inputrec->delta_t;
+ } else {
+ delta_t = 0;
+ }
+
+ if (fplog) {
+ print_perf(fplog,runtime->proctime,runtime->realtime,
+ cr->nnodes-cr->npmenodes,
+ runtime->nsteps_done,delta_t,nbfs,mflop);
+ }
+ if (bWriteStat) {
+ print_perf(stderr,runtime->proctime,runtime->realtime,
+ cr->nnodes-cr->npmenodes,
+ runtime->nsteps_done,delta_t,nbfs,mflop);
+ }
+
+ /*
+ runtime=inputrec->nsteps*inputrec->delta_t;
+ if (bWriteStat) {
+ if (cr->nnodes == 1)
+ fprintf(stderr,"\n\n");
+ print_perf(stderr,nodetime,realtime,runtime,&ntot,
+ cr->nnodes-cr->npmenodes,FALSE);
+ }
+ wallcycle_print(fplog,cr->nnodes,cr->npmenodes,realtime,wcycle,cycles);
+ print_perf(fplog,nodetime,realtime,runtime,&ntot,cr->nnodes-cr->npmenodes,
+ TRUE);
+ if (PARTDECOMP(cr))
+ pr_load(fplog,cr,nrnb_all);
+ if (cr->nnodes > 1)
+ sfree(nrnb_all);
+ */
+ }
+}
+
+void init_md(FILE *fplog,
+ t_commrec *cr,t_inputrec *ir,const output_env_t oenv,
+ double *t,double *t0,
+ real *lambda,double *lam0,
+ t_nrnb *nrnb,gmx_mtop_t *mtop,
+ gmx_update_t *upd,
+ int nfile,const t_filenm fnm[],
+ gmx_mdoutf_t **outf,t_mdebin **mdebin,
+ tensor force_vir,tensor shake_vir,rvec mu_tot,
+ gmx_bool *bSimAnn,t_vcm **vcm, t_state *state, unsigned long Flags)
+{
+ int i,j,n;
+ real tmpt,mod;
+
+ /* Initial values */
+ *t = *t0 = ir->init_t;
+ if (ir->efep != efepNO)
+ {
+ *lam0 = ir->init_lambda;
+ *lambda = *lam0 + ir->init_step*ir->delta_lambda;
+ }
+ else
+ {
+ *lambda = *lam0 = 0.0;
+ }
+
+ *bSimAnn=FALSE;
+ for(i=0;i<ir->opts.ngtc;i++)
+ {
+ /* set bSimAnn if any group is being annealed */
+ if(ir->opts.annealing[i]!=eannNO)
+ {
+ *bSimAnn = TRUE;
+ }
+ }
+ if (*bSimAnn)
+ {
+ update_annealing_target_temp(&(ir->opts),ir->init_t);
+ }
+
+ if (upd)
+ {
+ *upd = init_update(fplog,ir);
+ }
+
+ if (vcm != NULL)
+ {
+ *vcm = init_vcm(fplog,&mtop->groups,ir);
+ }
+
+ if (EI_DYNAMICS(ir->eI) && !(Flags & MD_APPENDFILES))
+ {
+ if (ir->etc == etcBERENDSEN)
+ {
+ please_cite(fplog,"Berendsen84a");
+ }
+ if (ir->etc == etcVRESCALE)
+ {
+ please_cite(fplog,"Bussi2007a");
+ }
+ }
+
+ init_nrnb(nrnb);
+
+ if (nfile != -1)
+ {
+ *outf = init_mdoutf(nfile,fnm,Flags,cr,ir,oenv);
+
+ *mdebin = init_mdebin((Flags & MD_APPENDFILES) ? NULL : (*outf)->fp_ene,
+ mtop,ir, (*outf)->fp_dhdl);
+ }
+
+ if (ir->bAdress)
+ {
+ please_cite(fplog,"Fritsch12");
+ please_cite(fplog,"Junghans10");
+ }
+ /* Initiate variables */
+ clear_mat(force_vir);
+ clear_mat(shake_vir);
+ clear_rvec(mu_tot);
+
+ debug_gmx();
+}
+
--- /dev/null
- #ifdef GMX_THREADS
+/* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
+ *
+ *
+ * This source code is part of
+ *
+ * G R O M A C S
+ *
+ * GROningen MAchine for Chemical Simulations
+ *
+ * VERSION 3.2.0
+ * Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
+ * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
+ * Copyright (c) 2001-2004, The GROMACS development team,
+ * check out http://www.gromacs.org for more information.
+
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version 2
+ * of the License, or (at your option) any later version.
+ *
+ * If you want to redistribute modifications, please consider that
+ * scientific software is very special. Version control is crucial -
+ * bugs must be traceable. We will be happy to consider code for
+ * inclusion in the official distribution, but derived work must not
+ * be called official GROMACS. Details are found in the README & COPYING
+ * files - if they are missing, get the official version at www.gromacs.org.
+ *
+ * To help us fund GROMACS development, we humbly ask that you cite
+ * the papers on the package - you can find them in the top README file.
+ *
+ * For more info, check our website at http://www.gromacs.org
+ *
+ * And Hey:
+ * Gallium Rubidium Oxygen Manganese Argon Carbon Silicon
+ */
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+#if ((defined WIN32 || defined _WIN32 || defined WIN64 || defined _WIN64) && !defined __CYGWIN__ && !defined __CYGWIN32__)
+/* _isnan() */
+#include <float.h>
+#endif
+
+#include "typedefs.h"
+#include "smalloc.h"
+#include "sysstuff.h"
+#include "vec.h"
+#include "statutil.h"
+#include "vcm.h"
+#include "mdebin.h"
+#include "nrnb.h"
+#include "calcmu.h"
+#include "index.h"
+#include "vsite.h"
+#include "update.h"
+#include "ns.h"
+#include "trnio.h"
+#include "xtcio.h"
+#include "mdrun.h"
+#include "confio.h"
+#include "network.h"
+#include "pull.h"
+#include "xvgr.h"
+#include "physics.h"
+#include "names.h"
+#include "xmdrun.h"
+#include "ionize.h"
+#include "disre.h"
+#include "orires.h"
+#include "dihre.h"
+#include "pppm.h"
+#include "pme.h"
+#include "mdatoms.h"
+#include "repl_ex.h"
+#include "qmmm.h"
+#include "domdec.h"
+#include "partdec.h"
+#include "topsort.h"
+#include "coulomb.h"
+#include "constr.h"
+#include "shellfc.h"
+#include "compute_io.h"
+#include "mvdata.h"
+#include "checkpoint.h"
+#include "mtop_util.h"
+#include "sighandler.h"
+#include "membed.h"
+#include "string2.h"
+
+#ifdef GMX_LIB_MPI
+#include <mpi.h>
+#endif
- #ifdef GMX_THREADS
++#ifdef GMX_THREAD_MPI
+#include "tmpi.h"
+#endif
+
+#ifdef GMX_FAHCORE
+#include "corewrap.h"
+#endif
+
+
+double do_md(FILE *fplog,t_commrec *cr,int nfile,const t_filenm fnm[],
+ const output_env_t oenv, gmx_bool bVerbose,gmx_bool bCompact,
+ int nstglobalcomm,
+ gmx_vsite_t *vsite,gmx_constr_t constr,
+ int stepout,t_inputrec *ir,
+ gmx_mtop_t *top_global,
+ t_fcdata *fcd,
+ t_state *state_global,
+ t_mdatoms *mdatoms,
+ t_nrnb *nrnb,gmx_wallcycle_t wcycle,
+ gmx_edsam_t ed,t_forcerec *fr,
+ int repl_ex_nst,int repl_ex_seed,gmx_membed_t *membed,
+ real cpt_period,real max_hours,
+ const char *deviceOptions,
+ unsigned long Flags,
+ gmx_runtime_t *runtime)
+{
+ gmx_mdoutf_t *outf;
+ gmx_large_int_t step,step_rel;
+ double run_time;
+ double t,t0,lam0;
+ gmx_bool bGStatEveryStep,bGStat,bNstEner,bCalcEnerPres;
+ gmx_bool bNS,bNStList,bSimAnn,bStopCM,bRerunMD,bNotLastFrame=FALSE,
+ bFirstStep,bStateFromTPX,bInitStep,bLastStep,
+ bBornRadii,bStartingFromCpt;
+ gmx_bool bDoDHDL=FALSE;
+ gmx_bool do_ene,do_log,do_verbose,bRerunWarnNoV=TRUE,
+ bForceUpdate=FALSE,bCPT;
+ int mdof_flags;
+ gmx_bool bMasterState;
+ int force_flags,cglo_flags;
+ tensor force_vir,shake_vir,total_vir,tmp_vir,pres;
+ int i,m;
+ t_trxstatus *status;
+ rvec mu_tot;
+ t_vcm *vcm;
+ t_state *bufstate=NULL;
+ matrix *scale_tot,pcoupl_mu,M,ebox;
+ gmx_nlheur_t nlh;
+ t_trxframe rerun_fr;
+ gmx_repl_ex_t repl_ex=NULL;
+ int nchkpt=1;
+
+ gmx_localtop_t *top;
+ t_mdebin *mdebin=NULL;
+ t_state *state=NULL;
+ rvec *f_global=NULL;
+ int n_xtc=-1;
+ rvec *x_xtc=NULL;
+ gmx_enerdata_t *enerd;
+ rvec *f=NULL;
+ gmx_global_stat_t gstat;
+ gmx_update_t upd=NULL;
+ t_graph *graph=NULL;
+ globsig_t gs;
+
+ gmx_bool bFFscan;
+ gmx_groups_t *groups;
+ gmx_ekindata_t *ekind, *ekind_save;
+ gmx_shellfc_t shellfc;
+ int count,nconverged=0;
+ real timestep=0;
+ double tcount=0;
+ gmx_bool bIonize=FALSE;
+ gmx_bool bTCR=FALSE,bConverged=TRUE,bOK,bSumEkinhOld,bExchanged;
+ gmx_bool bAppend;
+ gmx_bool bResetCountersHalfMaxH=FALSE;
+ gmx_bool bVV,bIterations,bFirstIterate,bTemp,bPres,bTrotter;
+ real temp0,mu_aver=0,dvdl;
+ int a0,a1,gnx=0,ii;
+ atom_id *grpindex=NULL;
+ char *grpname;
+ t_coupl_rec *tcr=NULL;
+ rvec *xcopy=NULL,*vcopy=NULL,*cbuf=NULL;
+ matrix boxcopy={{0}},lastbox;
+ tensor tmpvir;
+ real fom,oldfom,veta_save,pcurr,scalevir,tracevir;
+ real vetanew = 0;
+ double cycles;
+ real saved_conserved_quantity = 0;
+ real last_ekin = 0;
+ int iter_i;
+ t_extmass MassQ;
+ int **trotter_seq;
+ char sbuf[STEPSTRSIZE],sbuf2[STEPSTRSIZE];
+ int handled_stop_condition=gmx_stop_cond_none; /* compare to get_stop_condition*/
+ gmx_iterate_t iterate;
+ gmx_large_int_t multisim_nsteps=-1; /* number of steps to do before first multisim
+ simulation stops. If equal to zero, don't
+ communicate any more between multisims.*/
+#ifdef GMX_FAHCORE
+ /* Temporary addition for FAHCORE checkpointing */
+ int chkpt_ret;
+#endif
+
+ /* Check for special mdrun options */
+ bRerunMD = (Flags & MD_RERUN);
+ bIonize = (Flags & MD_IONIZE);
+ bFFscan = (Flags & MD_FFSCAN);
+ bAppend = (Flags & MD_APPENDFILES);
+ if (Flags & MD_RESETCOUNTERSHALFWAY)
+ {
+ if (ir->nsteps > 0)
+ {
+ /* Signal to reset the counters half the simulation steps. */
+ wcycle_set_reset_counters(wcycle,ir->nsteps/2);
+ }
+ /* Signal to reset the counters halfway the simulation time. */
+ bResetCountersHalfMaxH = (max_hours > 0);
+ }
+
+ /* md-vv uses averaged full step velocities for T-control
+ md-vv-avek uses averaged half step velocities for T-control (but full step ekin for P control)
+ md uses averaged half step kinetic energies to determine temperature unless defined otherwise by GMX_EKIN_AVE_VEL; */
+ bVV = EI_VV(ir->eI);
+ if (bVV) /* to store the initial velocities while computing virial */
+ {
+ snew(cbuf,top_global->natoms);
+ }
+ /* all the iteratative cases - only if there are constraints */
+ bIterations = ((IR_NPT_TROTTER(ir)) && (constr) && (!bRerunMD));
+ bTrotter = (bVV && (IR_NPT_TROTTER(ir) || (IR_NVT_TROTTER(ir))));
+
+ if (bRerunMD)
+ {
+ /* Since we don't know if the frames read are related in any way,
+ * rebuild the neighborlist at every step.
+ */
+ ir->nstlist = 1;
+ ir->nstcalcenergy = 1;
+ nstglobalcomm = 1;
+ }
+
+ check_ir_old_tpx_versions(cr,fplog,ir,top_global);
+
+ nstglobalcomm = check_nstglobalcomm(fplog,cr,nstglobalcomm,ir);
+ bGStatEveryStep = (nstglobalcomm == 1);
+
+ if (!bGStatEveryStep && ir->nstlist == -1 && fplog != NULL)
+ {
+ fprintf(fplog,
+ "To reduce the energy communication with nstlist = -1\n"
+ "the neighbor list validity should not be checked at every step,\n"
+ "this means that exact integration is not guaranteed.\n"
+ "The neighbor list validity is checked after:\n"
+ " <n.list life time> - 2*std.dev.(n.list life time) steps.\n"
+ "In most cases this will result in exact integration.\n"
+ "This reduces the energy communication by a factor of 2 to 3.\n"
+ "If you want less energy communication, set nstlist > 3.\n\n");
+ }
+
+ if (bRerunMD || bFFscan)
+ {
+ ir->nstxtcout = 0;
+ }
+ groups = &top_global->groups;
+
+ /* Initial values */
+ init_md(fplog,cr,ir,oenv,&t,&t0,&state_global->lambda,&lam0,
+ nrnb,top_global,&upd,
+ nfile,fnm,&outf,&mdebin,
+ force_vir,shake_vir,mu_tot,&bSimAnn,&vcm,state_global,Flags);
+
+ clear_mat(total_vir);
+ clear_mat(pres);
+ /* Energy terms and groups */
+ snew(enerd,1);
+ init_enerdata(top_global->groups.grps[egcENER].nr,ir->n_flambda,enerd);
+ if (DOMAINDECOMP(cr))
+ {
+ f = NULL;
+ }
+ else
+ {
+ snew(f,top_global->natoms);
+ }
+
+ /* Kinetic energy data */
+ snew(ekind,1);
+ init_ekindata(fplog,top_global,&(ir->opts),ekind);
+ /* needed for iteration of constraints */
+ snew(ekind_save,1);
+ init_ekindata(fplog,top_global,&(ir->opts),ekind_save);
+ /* Copy the cos acceleration to the groups struct */
+ ekind->cosacc.cos_accel = ir->cos_accel;
+
+ gstat = global_stat_init(ir);
+ debug_gmx();
+
+ /* Check for polarizable models and flexible constraints */
+ shellfc = init_shell_flexcon(fplog,
+ top_global,n_flexible_constraints(constr),
+ (ir->bContinuation ||
+ (DOMAINDECOMP(cr) && !MASTER(cr))) ?
+ NULL : state_global->x);
+
+ if (DEFORM(*ir))
+ {
- #ifdef GMX_THREADS
++#ifdef GMX_THREAD_MPI
+ tMPI_Thread_mutex_lock(&deform_init_box_mutex);
+#endif
+ set_deform_reference_box(upd,
+ deform_init_init_step_tpx,
+ deform_init_box_tpx);
- #ifdef GMX_THREADS
++#ifdef GMX_THREAD_MPI
+ tMPI_Thread_mutex_unlock(&deform_init_box_mutex);
+#endif
+ }
+
+ {
+ double io = compute_io(ir,top_global->natoms,groups,mdebin->ebin->nener,1);
+ if ((io > 2000) && MASTER(cr))
+ fprintf(stderr,
+ "\nWARNING: This run will generate roughly %.0f Mb of data\n\n",
+ io);
+ }
+
+ if (DOMAINDECOMP(cr)) {
+ top = dd_init_local_top(top_global);
+
+ snew(state,1);
+ dd_init_local_state(cr->dd,state_global,state);
+
+ if (DDMASTER(cr->dd) && ir->nstfout) {
+ snew(f_global,state_global->natoms);
+ }
+ } else {
+ if (PAR(cr)) {
+ /* Initialize the particle decomposition and split the topology */
+ top = split_system(fplog,top_global,ir,cr);
+
+ pd_cg_range(cr,&fr->cg0,&fr->hcg);
+ pd_at_range(cr,&a0,&a1);
+ } else {
+ top = gmx_mtop_generate_local_top(top_global,ir);
+
+ a0 = 0;
+ a1 = top_global->natoms;
+ }
+
+ state = partdec_init_local_state(cr,state_global);
+ f_global = f;
+
+ atoms2md(top_global,ir,0,NULL,a0,a1-a0,mdatoms);
+
+ if (vsite) {
+ set_vsite_top(vsite,top,mdatoms,cr);
+ }
+
+ if (ir->ePBC != epbcNONE && !ir->bPeriodicMols) {
+ graph = mk_graph(fplog,&(top->idef),0,top_global->natoms,FALSE,FALSE);
+ }
+
+ if (shellfc) {
+ make_local_shells(cr,mdatoms,shellfc);
+ }
+
+ if (ir->pull && PAR(cr)) {
+ dd_make_local_pull_groups(NULL,ir->pull,mdatoms);
+ }
+ }
+
+ if (DOMAINDECOMP(cr))
+ {
+ /* Distribute the charge groups over the nodes from the master node */
+ dd_partition_system(fplog,ir->init_step,cr,TRUE,1,
+ state_global,top_global,ir,
+ state,&f,mdatoms,top,fr,
+ vsite,shellfc,constr,
+ nrnb,wcycle,FALSE);
+ }
+
+ update_mdatoms(mdatoms,state->lambda);
+
+ if (MASTER(cr))
+ {
+ if (opt2bSet("-cpi",nfile,fnm))
+ {
+ /* Update mdebin with energy history if appending to output files */
+ if ( Flags & MD_APPENDFILES )
+ {
+ restore_energyhistory_from_state(mdebin,&state_global->enerhist);
+ }
+ else
+ {
+ /* We might have read an energy history from checkpoint,
+ * free the allocated memory and reset the counts.
+ */
+ done_energyhistory(&state_global->enerhist);
+ init_energyhistory(&state_global->enerhist);
+ }
+ }
+ /* Set the initial energy history in state by updating once */
+ update_energyhistory(&state_global->enerhist,mdebin);
+ }
+
+ if ((state->flags & (1<<estLD_RNG)) && (Flags & MD_READ_RNG)) {
+ /* Set the random state if we read a checkpoint file */
+ set_stochd_state(upd,state);
+ }
+
+ /* Initialize constraints */
+ if (constr) {
+ if (!DOMAINDECOMP(cr))
+ set_constraints(constr,top,ir,mdatoms,cr);
+ }
+
+ /* Check whether we have to GCT stuff */
+ bTCR = ftp2bSet(efGCT,nfile,fnm);
+ if (bTCR) {
+ if (MASTER(cr)) {
+ fprintf(stderr,"Will do General Coupling Theory!\n");
+ }
+ gnx = top_global->mols.nr;
+ snew(grpindex,gnx);
+ for(i=0; (i<gnx); i++) {
+ grpindex[i] = i;
+ }
+ }
+
+ if (repl_ex_nst > 0)
+ {
+ /* We need to be sure replica exchange can only occur
+ * when the energies are current */
+ check_nst_param(fplog,cr,"nstcalcenergy",ir->nstcalcenergy,
+ "repl_ex_nst",&repl_ex_nst);
+ /* This check needs to happen before inter-simulation
+ * signals are initialized, too */
+ }
+ if (repl_ex_nst > 0 && MASTER(cr))
+ repl_ex = init_replica_exchange(fplog,cr->ms,state_global,ir,
+ repl_ex_nst,repl_ex_seed);
+
+ if (!ir->bContinuation && !bRerunMD)
+ {
+ if (mdatoms->cFREEZE && (state->flags & (1<<estV)))
+ {
+ /* Set the velocities of frozen particles to zero */
+ for(i=mdatoms->start; i<mdatoms->start+mdatoms->homenr; i++)
+ {
+ for(m=0; m<DIM; m++)
+ {
+ if (ir->opts.nFreeze[mdatoms->cFREEZE[i]][m])
+ {
+ state->v[i][m] = 0;
+ }
+ }
+ }
+ }
+
+ if (constr)
+ {
+ /* Constrain the initial coordinates and velocities */
+ do_constrain_first(fplog,constr,ir,mdatoms,state,f,
+ graph,cr,nrnb,fr,top,shake_vir);
+ }
+ if (vsite)
+ {
+ /* Construct the virtual sites for the initial configuration */
+ construct_vsites(fplog,vsite,state->x,nrnb,ir->delta_t,NULL,
+ top->idef.iparams,top->idef.il,
+ fr->ePBC,fr->bMolPBC,graph,cr,state->box);
+ }
+ }
+
+ debug_gmx();
+
+ /* I'm assuming we need global communication the first time! MRS */
+ cglo_flags = (CGLO_TEMPERATURE | CGLO_GSTAT
+ | (bVV ? CGLO_PRESSURE:0)
+ | (bVV ? CGLO_CONSTRAINT:0)
+ | (bRerunMD ? CGLO_RERUNMD:0)
+ | ((Flags & MD_READ_EKIN) ? CGLO_READEKIN:0));
+
+ bSumEkinhOld = FALSE;
+ compute_globals(fplog,gstat,cr,ir,fr,ekind,state,state_global,mdatoms,nrnb,vcm,
+ NULL,enerd,force_vir,shake_vir,total_vir,pres,mu_tot,
+ constr,NULL,FALSE,state->box,
+ top_global,&pcurr,top_global->natoms,&bSumEkinhOld,cglo_flags);
+ if (ir->eI == eiVVAK) {
+ /* a second call to get the half step temperature initialized as well */
+ /* we do the same call as above, but turn the pressure off -- internally to
+ compute_globals, this is recognized as a velocity verlet half-step
+ kinetic energy calculation. This minimized excess variables, but
+ perhaps loses some logic?*/
+
+ compute_globals(fplog,gstat,cr,ir,fr,ekind,state,state_global,mdatoms,nrnb,vcm,
+ NULL,enerd,force_vir,shake_vir,total_vir,pres,mu_tot,
+ constr,NULL,FALSE,state->box,
+ top_global,&pcurr,top_global->natoms,&bSumEkinhOld,
+ cglo_flags &~ CGLO_PRESSURE);
+ }
+
+ /* Calculate the initial half step temperature, and save the ekinh_old */
+ if (!(Flags & MD_STARTFROMCPT))
+ {
+ for(i=0; (i<ir->opts.ngtc); i++)
+ {
+ copy_mat(ekind->tcstat[i].ekinh,ekind->tcstat[i].ekinh_old);
+ }
+ }
+ if (ir->eI != eiVV)
+ {
+ enerd->term[F_TEMP] *= 2; /* result of averages being done over previous and current step,
+ and there is no previous step */
+ }
+ temp0 = enerd->term[F_TEMP];
+
+ /* if using an iterative algorithm, we need to create a working directory for the state. */
+ if (bIterations)
+ {
+ bufstate = init_bufstate(state);
+ }
+ if (bFFscan)
+ {
+ snew(xcopy,state->natoms);
+ snew(vcopy,state->natoms);
+ copy_rvecn(state->x,xcopy,0,state->natoms);
+ copy_rvecn(state->v,vcopy,0,state->natoms);
+ copy_mat(state->box,boxcopy);
+ }
+
+ /* need to make an initiation call to get the Trotter variables set, as well as other constants for non-trotter
+ temperature control */
+ trotter_seq = init_npt_vars(ir,state,&MassQ,bTrotter);
+
+ if (MASTER(cr))
+ {
+ if (constr && !ir->bContinuation && ir->eConstrAlg == econtLINCS)
+ {
+ fprintf(fplog,
+ "RMS relative constraint deviation after constraining: %.2e\n",
+ constr_rmsd(constr,FALSE));
+ }
+ fprintf(fplog,"Initial temperature: %g K\n",enerd->term[F_TEMP]);
+ if (bRerunMD)
+ {
+ fprintf(stderr,"starting md rerun '%s', reading coordinates from"
+ " input trajectory '%s'\n\n",
+ *(top_global->name),opt2fn("-rerun",nfile,fnm));
+ if (bVerbose)
+ {
+ fprintf(stderr,"Calculated time to finish depends on nsteps from "
+ "run input file,\nwhich may not correspond to the time "
+ "needed to process input trajectory.\n\n");
+ }
+ }
+ else
+ {
+ char tbuf[20];
+ fprintf(stderr,"starting mdrun '%s'\n",
+ *(top_global->name));
+ if (ir->nsteps >= 0)
+ {
+ sprintf(tbuf,"%8.1f",(ir->init_step+ir->nsteps)*ir->delta_t);
+ }
+ else
+ {
+ sprintf(tbuf,"%s","infinite");
+ }
+ if (ir->init_step > 0)
+ {
+ fprintf(stderr,"%s steps, %s ps (continuing from step %s, %8.1f ps).\n",
+ gmx_step_str(ir->init_step+ir->nsteps,sbuf),tbuf,
+ gmx_step_str(ir->init_step,sbuf2),
+ ir->init_step*ir->delta_t);
+ }
+ else
+ {
+ fprintf(stderr,"%s steps, %s ps.\n",
+ gmx_step_str(ir->nsteps,sbuf),tbuf);
+ }
+ }
+ fprintf(fplog,"\n");
+ }
+
+ /* Set and write start time */
+ runtime_start(runtime);
+ print_date_and_time(fplog,cr->nodeid,"Started mdrun",runtime);
+ wallcycle_start(wcycle,ewcRUN);
+ if (fplog)
+ fprintf(fplog,"\n");
+
+ /* safest point to do file checkpointing is here. More general point would be immediately before integrator call */
+#ifdef GMX_FAHCORE
+ chkpt_ret=fcCheckPointParallel( cr->nodeid,
+ NULL,0);
+ if ( chkpt_ret == 0 )
+ gmx_fatal( 3,__FILE__,__LINE__, "Checkpoint error on step %d\n", 0 );
+#endif
+
+ debug_gmx();
+ /***********************************************************
+ *
+ * Loop over MD steps
+ *
+ ************************************************************/
+
+ /* if rerunMD then read coordinates and velocities from input trajectory */
+ if (bRerunMD)
+ {
+ if (getenv("GMX_FORCE_UPDATE"))
+ {
+ bForceUpdate = TRUE;
+ }
+
+ rerun_fr.natoms = 0;
+ if (MASTER(cr))
+ {
+ bNotLastFrame = read_first_frame(oenv,&status,
+ opt2fn("-rerun",nfile,fnm),
+ &rerun_fr,TRX_NEED_X | TRX_READ_V);
+ if (rerun_fr.natoms != top_global->natoms)
+ {
+ gmx_fatal(FARGS,
+ "Number of atoms in trajectory (%d) does not match the "
+ "run input file (%d)\n",
+ rerun_fr.natoms,top_global->natoms);
+ }
+ if (ir->ePBC != epbcNONE)
+ {
+ if (!rerun_fr.bBox)
+ {
+ gmx_fatal(FARGS,"Rerun trajectory frame step %d time %f does not contain a box, while pbc is used",rerun_fr.step,rerun_fr.time);
+ }
+ if (max_cutoff2(ir->ePBC,rerun_fr.box) < sqr(fr->rlistlong))
+ {
+ gmx_fatal(FARGS,"Rerun trajectory frame step %d time %f has too small box dimensions",rerun_fr.step,rerun_fr.time);
+ }
+ }
+ }
+
+ if (PAR(cr))
+ {
+ rerun_parallel_comm(cr,&rerun_fr,&bNotLastFrame);
+ }
+
+ if (ir->ePBC != epbcNONE)
+ {
+ /* Set the shift vectors.
+ * Necessary here when have a static box different from the tpr box.
+ */
+ calc_shifts(rerun_fr.box,fr->shift_vec);
+ }
+ }
+
+ /* loop over MD steps or if rerunMD to end of input trajectory */
+ bFirstStep = TRUE;
+ /* Skip the first Nose-Hoover integration when we get the state from tpx */
+ bStateFromTPX = !opt2bSet("-cpi",nfile,fnm);
+ bInitStep = bFirstStep && (bStateFromTPX || bVV);
+ bStartingFromCpt = (Flags & MD_STARTFROMCPT) && bInitStep;
+ bLastStep = FALSE;
+ bSumEkinhOld = FALSE;
+ bExchanged = FALSE;
+
+ init_global_signals(&gs,cr,ir,repl_ex_nst);
+
+ step = ir->init_step;
+ step_rel = 0;
+
+ if (ir->nstlist == -1)
+ {
+ init_nlistheuristics(&nlh,bGStatEveryStep,step);
+ }
+
+ if (MULTISIM(cr) && (repl_ex_nst <=0 ))
+ {
+ /* check how many steps are left in other sims */
+ multisim_nsteps=get_multisim_nsteps(cr, ir->nsteps);
+ }
+
+
+ /* and stop now if we should */
+ bLastStep = (bRerunMD || (ir->nsteps >= 0 && step_rel > ir->nsteps) ||
+ ((multisim_nsteps >= 0) && (step_rel >= multisim_nsteps )));
+ while (!bLastStep || (bRerunMD && bNotLastFrame)) {
+
+ wallcycle_start(wcycle,ewcSTEP);
+
+ if (bRerunMD) {
+ if (rerun_fr.bStep) {
+ step = rerun_fr.step;
+ step_rel = step - ir->init_step;
+ }
+ if (rerun_fr.bTime) {
+ t = rerun_fr.time;
+ }
+ else
+ {
+ t = step;
+ }
+ }
+ else
+ {
+ bLastStep = (step_rel == ir->nsteps);
+ t = t0 + step*ir->delta_t;
+ }
+
+ if (ir->efep != efepNO)
+ {
+ if (bRerunMD && rerun_fr.bLambda && (ir->delta_lambda!=0))
+ {
+ state_global->lambda = rerun_fr.lambda;
+ }
+ else
+ {
+ state_global->lambda = lam0 + step*ir->delta_lambda;
+ }
+ state->lambda = state_global->lambda;
+ bDoDHDL = do_per_step(step,ir->nstdhdl);
+ }
+
+ if (bSimAnn)
+ {
+ update_annealing_target_temp(&(ir->opts),t);
+ }
+
+ if (bRerunMD)
+ {
+ if (!(DOMAINDECOMP(cr) && !MASTER(cr)))
+ {
+ for(i=0; i<state_global->natoms; i++)
+ {
+ copy_rvec(rerun_fr.x[i],state_global->x[i]);
+ }
+ if (rerun_fr.bV)
+ {
+ for(i=0; i<state_global->natoms; i++)
+ {
+ copy_rvec(rerun_fr.v[i],state_global->v[i]);
+ }
+ }
+ else
+ {
+ for(i=0; i<state_global->natoms; i++)
+ {
+ clear_rvec(state_global->v[i]);
+ }
+ if (bRerunWarnNoV)
+ {
+ fprintf(stderr,"\nWARNING: Some frames do not contain velocities.\n"
+ " Ekin, temperature and pressure are incorrect,\n"
+ " the virial will be incorrect when constraints are present.\n"
+ "\n");
+ bRerunWarnNoV = FALSE;
+ }
+ }
+ }
+ copy_mat(rerun_fr.box,state_global->box);
+ copy_mat(state_global->box,state->box);
+
+ if (vsite && (Flags & MD_RERUN_VSITE))
+ {
+ if (DOMAINDECOMP(cr))
+ {
+ gmx_fatal(FARGS,"Vsite recalculation with -rerun is not implemented for domain decomposition, use particle decomposition");
+ }
+ if (graph)
+ {
+ /* Following is necessary because the graph may get out of sync
+ * with the coordinates if we only have every N'th coordinate set
+ */
+ mk_mshift(fplog,graph,fr->ePBC,state->box,state->x);
+ shift_self(graph,state->box,state->x);
+ }
+ construct_vsites(fplog,vsite,state->x,nrnb,ir->delta_t,state->v,
+ top->idef.iparams,top->idef.il,
+ fr->ePBC,fr->bMolPBC,graph,cr,state->box);
+ if (graph)
+ {
+ unshift_self(graph,state->box,state->x);
+ }
+ }
+ }
+
+ /* Stop Center of Mass motion */
+ bStopCM = (ir->comm_mode != ecmNO && do_per_step(step,ir->nstcomm));
+
+ /* Copy back starting coordinates in case we're doing a forcefield scan */
+ if (bFFscan)
+ {
+ for(ii=0; (ii<state->natoms); ii++)
+ {
+ copy_rvec(xcopy[ii],state->x[ii]);
+ copy_rvec(vcopy[ii],state->v[ii]);
+ }
+ copy_mat(boxcopy,state->box);
+ }
+
+ if (bRerunMD)
+ {
+ /* for rerun MD always do Neighbour Searching */
+ bNS = (bFirstStep || ir->nstlist != 0);
+ bNStList = bNS;
+ }
+ else
+ {
+ /* Determine whether or not to do Neighbour Searching and LR */
+ bNStList = (ir->nstlist > 0 && step % ir->nstlist == 0);
+
+ bNS = (bFirstStep || bExchanged || bNStList ||
+ (ir->nstlist == -1 && nlh.nabnsb > 0));
+
+ if (bNS && ir->nstlist == -1)
+ {
+ set_nlistheuristics(&nlh,bFirstStep || bExchanged,step);
+ }
+ }
+
+ /* check whether we should stop because another simulation has
+ stopped. */
+ if (MULTISIM(cr))
+ {
+ if ( (multisim_nsteps >= 0) && (step_rel >= multisim_nsteps) &&
+ (multisim_nsteps != ir->nsteps) )
+ {
+ if (bNS)
+ {
+ if (MASTER(cr))
+ {
+ fprintf(stderr,
+ "Stopping simulation %d because another one has finished\n",
+ cr->ms->sim);
+ }
+ bLastStep=TRUE;
+ gs.sig[eglsCHKPT] = 1;
+ }
+ }
+ }
+
+ /* < 0 means stop at next step, > 0 means stop at next NS step */
+ if ( (gs.set[eglsSTOPCOND] < 0 ) ||
+ ( (gs.set[eglsSTOPCOND] > 0 ) && ( bNS || ir->nstlist==0)) )
+ {
+ bLastStep = TRUE;
+ }
+
+ /* Determine whether or not to update the Born radii if doing GB */
+ bBornRadii=bFirstStep;
+ if (ir->implicit_solvent && (step % ir->nstgbradii==0))
+ {
+ bBornRadii=TRUE;
+ }
+
+ do_log = do_per_step(step,ir->nstlog) || bFirstStep || bLastStep;
+ do_verbose = bVerbose &&
+ (step % stepout == 0 || bFirstStep || bLastStep);
+
+ if (bNS && !(bFirstStep && ir->bContinuation && !bRerunMD))
+ {
+ if (bRerunMD)
+ {
+ bMasterState = TRUE;
+ }
+ else
+ {
+ bMasterState = FALSE;
+ /* Correct the new box if it is too skewed */
+ if (DYNAMIC_BOX(*ir))
+ {
+ if (correct_box(fplog,step,state->box,graph))
+ {
+ bMasterState = TRUE;
+ }
+ }
+ if (DOMAINDECOMP(cr) && bMasterState)
+ {
+ dd_collect_state(cr->dd,state,state_global);
+ }
+ }
+
+ if (DOMAINDECOMP(cr))
+ {
+ /* Repartition the domain decomposition */
+ wallcycle_start(wcycle,ewcDOMDEC);
+ dd_partition_system(fplog,step,cr,
+ bMasterState,nstglobalcomm,
+ state_global,top_global,ir,
+ state,&f,mdatoms,top,fr,
+ vsite,shellfc,constr,
+ nrnb,wcycle,do_verbose);
+ wallcycle_stop(wcycle,ewcDOMDEC);
+ /* If using an iterative integrator, reallocate space to match the decomposition */
+ }
+ }
+
+ if (MASTER(cr) && do_log && !bFFscan)
+ {
+ print_ebin_header(fplog,step,t,state->lambda);
+ }
+
+ if (ir->efep != efepNO)
+ {
+ update_mdatoms(mdatoms,state->lambda);
+ }
+
+ if (bRerunMD && rerun_fr.bV)
+ {
+
+ /* We need the kinetic energy at minus the half step for determining
+ * the full step kinetic energy and possibly for T-coupling.*/
+ /* This may not be quite working correctly yet . . . . */
+ compute_globals(fplog,gstat,cr,ir,fr,ekind,state,state_global,mdatoms,nrnb,vcm,
+ wcycle,enerd,NULL,NULL,NULL,NULL,mu_tot,
+ constr,NULL,FALSE,state->box,
+ top_global,&pcurr,top_global->natoms,&bSumEkinhOld,
+ CGLO_RERUNMD | CGLO_GSTAT | CGLO_TEMPERATURE);
+ }
+ clear_mat(force_vir);
+
+ /* Ionize the atoms if necessary */
+ if (bIonize)
+ {
+ ionize(fplog,oenv,mdatoms,top_global,t,ir,state->x,state->v,
+ mdatoms->start,mdatoms->start+mdatoms->homenr,state->box,cr);
+ }
+
+ /* Update force field in ffscan program */
+ if (bFFscan)
+ {
+ if (update_forcefield(fplog,
+ nfile,fnm,fr,
+ mdatoms->nr,state->x,state->box)) {
+ if (gmx_parallel_env_initialized())
+ {
+ gmx_finalize();
+ }
+ exit(0);
+ }
+ }
+
+ /* We write a checkpoint at this MD step when:
+ * either at an NS step when we signalled through gs,
+ * or at the last step (but not when we do not want confout),
+ * but never at the first step or with rerun.
+ */
+ bCPT = (((gs.set[eglsCHKPT] && (bNS || ir->nstlist == 0)) ||
+ (bLastStep && (Flags & MD_CONFOUT))) &&
+ step > ir->init_step && !bRerunMD);
+ if (bCPT)
+ {
+ gs.set[eglsCHKPT] = 0;
+ }
+
+ /* Determine the energy and pressure:
+ * at nstcalcenergy steps and at energy output steps (set below).
+ */
+ bNstEner = do_per_step(step,ir->nstcalcenergy);
+ bCalcEnerPres =
+ (bNstEner ||
+ (ir->epc != epcNO && do_per_step(step,ir->nstpcouple)));
+
+ /* Do we need global communication ? */
+ bGStat = (bCalcEnerPres || bStopCM ||
+ do_per_step(step,nstglobalcomm) ||
+ (ir->nstlist == -1 && !bRerunMD && step >= nlh.step_nscheck));
+
+ do_ene = (do_per_step(step,ir->nstenergy) || bLastStep);
+
+ if (do_ene || do_log)
+ {
+ bCalcEnerPres = TRUE;
+ bGStat = TRUE;
+ }
+
+ /* these CGLO_ options remain the same throughout the iteration */
+ cglo_flags = ((bRerunMD ? CGLO_RERUNMD : 0) |
+ (bStopCM ? CGLO_STOPCM : 0) |
+ (bGStat ? CGLO_GSTAT : 0)
+ );
+
+ force_flags = (GMX_FORCE_STATECHANGED |
+ ((DYNAMIC_BOX(*ir) || bRerunMD) ? GMX_FORCE_DYNAMICBOX : 0) |
+ GMX_FORCE_ALLFORCES |
+ (bNStList ? GMX_FORCE_DOLR : 0) |
+ GMX_FORCE_SEPLRF |
+ (bCalcEnerPres ? GMX_FORCE_VIRIAL : 0) |
+ (bDoDHDL ? GMX_FORCE_DHDL : 0)
+ );
+
+ if (shellfc)
+ {
+ /* Now is the time to relax the shells */
+ count=relax_shell_flexcon(fplog,cr,bVerbose,bFFscan ? step+1 : step,
+ ir,bNS,force_flags,
+ bStopCM,top,top_global,
+ constr,enerd,fcd,
+ state,f,force_vir,mdatoms,
+ nrnb,wcycle,graph,groups,
+ shellfc,fr,bBornRadii,t,mu_tot,
+ state->natoms,&bConverged,vsite,
+ outf->fp_field);
+ tcount+=count;
+
+ if (bConverged)
+ {
+ nconverged++;
+ }
+ }
+ else
+ {
+ /* The coordinates (x) are shifted (to get whole molecules)
+ * in do_force.
+ * This is parallellized as well, and does communication too.
+ * Check comments in sim_util.c
+ */
+
+ do_force(fplog,cr,ir,step,nrnb,wcycle,top,top_global,groups,
+ state->box,state->x,&state->hist,
+ f,force_vir,mdatoms,enerd,fcd,
+ state->lambda,graph,
+ fr,vsite,mu_tot,t,outf->fp_field,ed,bBornRadii,
+ (bNS ? GMX_FORCE_NS : 0) | force_flags);
+ }
+
+ if (bTCR)
+ {
+ mu_aver = calc_mu_aver(cr,state->x,mdatoms->chargeA,
+ mu_tot,&top_global->mols,mdatoms,gnx,grpindex);
+ }
+
+ if (bTCR && bFirstStep)
+ {
+ tcr=init_coupling(fplog,nfile,fnm,cr,fr,mdatoms,&(top->idef));
+ fprintf(fplog,"Done init_coupling\n");
+ fflush(fplog);
+ }
+
+ if (bVV && !bStartingFromCpt && !bRerunMD)
+ /* ############### START FIRST UPDATE HALF-STEP FOR VV METHODS############### */
+ {
+ if (ir->eI==eiVV && bInitStep)
+ {
+ /* if using velocity verlet with full time step Ekin,
+ * take the first half step only to compute the
+ * virial for the first step. From there,
+ * revert back to the initial coordinates
+ * so that the input is actually the initial step.
+ */
+ copy_rvecn(state->v,cbuf,0,state->natoms); /* should make this better for parallelizing? */
+ } else {
+ /* this is for NHC in the Ekin(t+dt/2) version of vv */
+ trotter_update(ir,step,ekind,enerd,state,total_vir,mdatoms,&MassQ,trotter_seq,ettTSEQ1);
+ }
+
+ update_coords(fplog,step,ir,mdatoms,state,
+ f,fr->bTwinRange && bNStList,fr->f_twin,fcd,
+ ekind,M,wcycle,upd,bInitStep,etrtVELOCITY1,
+ cr,nrnb,constr,&top->idef);
+
+ if (bIterations)
+ {
+ gmx_iterate_init(&iterate,bIterations && !bInitStep);
+ }
+ /* for iterations, we save these vectors, as we will be self-consistently iterating
+ the calculations */
+
+ /*#### UPDATE EXTENDED VARIABLES IN TROTTER FORMULATION */
+
+ /* save the state */
+ if (bIterations && iterate.bIterate) {
+ copy_coupling_state(state,bufstate,ekind,ekind_save,&(ir->opts));
+ }
+
+ bFirstIterate = TRUE;
+ while (bFirstIterate || (bIterations && iterate.bIterate))
+ {
+ if (bIterations && iterate.bIterate)
+ {
+ copy_coupling_state(bufstate,state,ekind_save,ekind,&(ir->opts));
+ if (bFirstIterate && bTrotter)
+ {
+ /* The first time through, we need a decent first estimate
+ of veta(t+dt) to compute the constraints. Do
+ this by computing the box volume part of the
+ trotter integration at this time. Nothing else
+ should be changed by this routine here. If
+ !(first time), we start with the previous value
+ of veta. */
+
+ veta_save = state->veta;
+ trotter_update(ir,step,ekind,enerd,state,total_vir,mdatoms,&MassQ,trotter_seq,ettTSEQ0);
+ vetanew = state->veta;
+ state->veta = veta_save;
+ }
+ }
+
+ bOK = TRUE;
+ if ( !bRerunMD || rerun_fr.bV || bForceUpdate) { /* Why is rerun_fr.bV here? Unclear. */
+ dvdl = 0;
+
+ update_constraints(fplog,step,&dvdl,ir,ekind,mdatoms,state,graph,f,
+ &top->idef,shake_vir,NULL,
+ cr,nrnb,wcycle,upd,constr,
+ bInitStep,TRUE,bCalcEnerPres,vetanew);
+
+ if (!bOK && !bFFscan)
+ {
+ gmx_fatal(FARGS,"Constraint error: Shake, Lincs or Settle could not solve the constrains");
+ }
+
+ }
+ else if (graph)
+ { /* Need to unshift here if a do_force has been
+ called in the previous step */
+ unshift_self(graph,state->box,state->x);
+ }
+
+
+ /* if VV, compute the pressure and constraints */
+ /* For VV2, we strictly only need this if using pressure
+ * control, but we really would like to have accurate pressures
+ * printed out.
+ * Think about ways around this in the future?
+ * For now, keep this choice in comments.
+ */
+ /*bPres = (ir->eI==eiVV || IR_NPT_TROTTER(ir)); */
+ /*bTemp = ((ir->eI==eiVV &&(!bInitStep)) || (ir->eI==eiVVAK && IR_NPT_TROTTER(ir)));*/
+ bPres = TRUE;
+ bTemp = ((ir->eI==eiVV &&(!bInitStep)) || (ir->eI==eiVVAK));
+ compute_globals(fplog,gstat,cr,ir,fr,ekind,state,state_global,mdatoms,nrnb,vcm,
+ wcycle,enerd,force_vir,shake_vir,total_vir,pres,mu_tot,
+ constr,NULL,FALSE,state->box,
+ top_global,&pcurr,top_global->natoms,&bSumEkinhOld,
+ cglo_flags
+ | CGLO_ENERGY
+ | (bTemp ? CGLO_TEMPERATURE:0)
+ | (bPres ? CGLO_PRESSURE : 0)
+ | (bPres ? CGLO_CONSTRAINT : 0)
+ | ((bIterations && iterate.bIterate) ? CGLO_ITERATE : 0)
+ | (bFirstIterate ? CGLO_FIRSTITERATE : 0)
+ | CGLO_SCALEEKIN
+ );
+ /* explanation of above:
+ a) We compute Ekin at the full time step
+ if 1) we are using the AveVel Ekin, and it's not the
+ initial step, or 2) if we are using AveEkin, but need the full
+ time step kinetic energy for the pressure (always true now, since we want accurate statistics).
+ b) If we are using EkinAveEkin for the kinetic energy for the temperture control, we still feed in
+ EkinAveVel because it's needed for the pressure */
+
+ /* temperature scaling and pressure scaling to produce the extended variables at t+dt */
+ if (!bInitStep)
+ {
+ if (bTrotter)
+ {
+ trotter_update(ir,step,ekind,enerd,state,total_vir,mdatoms,&MassQ,trotter_seq,ettTSEQ2);
+ }
+ else
+ {
+ update_tcouple(fplog,step,ir,state,ekind,wcycle,upd,&MassQ,mdatoms);
+ }
+ }
+
+ if (bIterations &&
+ done_iterating(cr,fplog,step,&iterate,bFirstIterate,
+ state->veta,&vetanew))
+ {
+ break;
+ }
+ bFirstIterate = FALSE;
+ }
+
+ if (bTrotter && !bInitStep) {
+ copy_mat(shake_vir,state->svir_prev);
+ copy_mat(force_vir,state->fvir_prev);
+ if (IR_NVT_TROTTER(ir) && ir->eI==eiVV) {
+ /* update temperature and kinetic energy now that step is over - this is the v(t+dt) point */
+ enerd->term[F_TEMP] = sum_ekin(&(ir->opts),ekind,NULL,(ir->eI==eiVV),FALSE,FALSE);
+ enerd->term[F_EKIN] = trace(ekind->ekin);
+ }
+ }
+ /* if it's the initial step, we performed this first step just to get the constraint virial */
+ if (bInitStep && ir->eI==eiVV) {
+ copy_rvecn(cbuf,state->v,0,state->natoms);
+ }
+
+ if (fr->bSepDVDL && fplog && do_log)
+ {
+ fprintf(fplog,sepdvdlformat,"Constraint",0.0,dvdl);
+ }
+ enerd->term[F_DHDL_CON] += dvdl;
+ }
+
+ /* MRS -- now done iterating -- compute the conserved quantity */
+ if (bVV) {
+ saved_conserved_quantity = compute_conserved_from_auxiliary(ir,state,&MassQ);
+ if (ir->eI==eiVV)
+ {
+ last_ekin = enerd->term[F_EKIN]; /* does this get preserved through checkpointing? */
+ }
+ if ((ir->eDispCorr != edispcEnerPres) && (ir->eDispCorr != edispcAllEnerPres))
+ {
+ saved_conserved_quantity -= enerd->term[F_DISPCORR];
+ }
+ }
+
+ /* ######## END FIRST UPDATE STEP ############## */
+ /* ######## If doing VV, we now have v(dt) ###### */
+
+ /* ################## START TRAJECTORY OUTPUT ################# */
+
+ /* Now we have the energies and forces corresponding to the
+ * coordinates at time t. We must output all of this before
+ * the update.
+ * for RerunMD t is read from input trajectory
+ */
+ mdof_flags = 0;
+ if (do_per_step(step,ir->nstxout)) { mdof_flags |= MDOF_X; }
+ if (do_per_step(step,ir->nstvout)) { mdof_flags |= MDOF_V; }
+ if (do_per_step(step,ir->nstfout)) { mdof_flags |= MDOF_F; }
+ if (do_per_step(step,ir->nstxtcout)) { mdof_flags |= MDOF_XTC; }
+ if (bCPT) { mdof_flags |= MDOF_CPT; };
+
+#if defined(GMX_FAHCORE) || defined(GMX_WRITELASTSTEP)
+ if (bLastStep)
+ {
+ /* Enforce writing positions and velocities at end of run */
+ mdof_flags |= (MDOF_X | MDOF_V);
+ }
+#endif
+#ifdef GMX_FAHCORE
+ if (MASTER(cr))
+ fcReportProgress( ir->nsteps, step );
+
+ /* sync bCPT and fc record-keeping */
+ if (bCPT && MASTER(cr))
+ fcRequestCheckPoint();
+#endif
+
+ if (mdof_flags != 0)
+ {
+ wallcycle_start(wcycle,ewcTRAJ);
+ if (bCPT)
+ {
+ if (state->flags & (1<<estLD_RNG))
+ {
+ get_stochd_state(upd,state);
+ }
+ if (MASTER(cr))
+ {
+ if (bSumEkinhOld)
+ {
+ state_global->ekinstate.bUpToDate = FALSE;
+ }
+ else
+ {
+ update_ekinstate(&state_global->ekinstate,ekind);
+ state_global->ekinstate.bUpToDate = TRUE;
+ }
+ update_energyhistory(&state_global->enerhist,mdebin);
+ }
+ }
+ write_traj(fplog,cr,outf,mdof_flags,top_global,
+ step,t,state,state_global,f,f_global,&n_xtc,&x_xtc);
+ if (bCPT)
+ {
+ nchkpt++;
+ bCPT = FALSE;
+ }
+ debug_gmx();
+ if (bLastStep && step_rel == ir->nsteps &&
+ (Flags & MD_CONFOUT) && MASTER(cr) &&
+ !bRerunMD && !bFFscan)
+ {
+ /* x and v have been collected in write_traj,
+ * because a checkpoint file will always be written
+ * at the last step.
+ */
+ fprintf(stderr,"\nWriting final coordinates.\n");
+ if (ir->ePBC != epbcNONE && !ir->bPeriodicMols &&
+ DOMAINDECOMP(cr))
+ {
+ /* Make molecules whole only for confout writing */
+ do_pbc_mtop(fplog,ir->ePBC,state->box,top_global,state_global->x);
+ }
+ write_sto_conf_mtop(ftp2fn(efSTO,nfile,fnm),
+ *top_global->name,top_global,
+ state_global->x,state_global->v,
+ ir->ePBC,state->box);
+ debug_gmx();
+ }
+ wallcycle_stop(wcycle,ewcTRAJ);
+ }
+
+ /* kludge -- virial is lost with restart for NPT control. Must restart */
+ if (bStartingFromCpt && bVV)
+ {
+ copy_mat(state->svir_prev,shake_vir);
+ copy_mat(state->fvir_prev,force_vir);
+ }
+ /* ################## END TRAJECTORY OUTPUT ################ */
+
+ /* Determine the wallclock run time up till now */
+ run_time = gmx_gettime() - (double)runtime->real;
+
+ /* Check whether everything is still allright */
+ if (((int)gmx_get_stop_condition() > handled_stop_condition)
++#ifdef GMX_THREAD_MPI
+ && MASTER(cr)
+#endif
+ )
+ {
+ /* this is just make gs.sig compatible with the hack
+ of sending signals around by MPI_Reduce with together with
+ other floats */
+ if ( gmx_get_stop_condition() == gmx_stop_cond_next_ns )
+ gs.sig[eglsSTOPCOND]=1;
+ if ( gmx_get_stop_condition() == gmx_stop_cond_next )
+ gs.sig[eglsSTOPCOND]=-1;
+ /* < 0 means stop at next step, > 0 means stop at next NS step */
+ if (fplog)
+ {
+ fprintf(fplog,
+ "\n\nReceived the %s signal, stopping at the next %sstep\n\n",
+ gmx_get_signal_name(),
+ gs.sig[eglsSTOPCOND]==1 ? "NS " : "");
+ fflush(fplog);
+ }
+ fprintf(stderr,
+ "\n\nReceived the %s signal, stopping at the next %sstep\n\n",
+ gmx_get_signal_name(),
+ gs.sig[eglsSTOPCOND]==1 ? "NS " : "");
+ fflush(stderr);
+ handled_stop_condition=(int)gmx_get_stop_condition();
+ }
+ else if (MASTER(cr) && (bNS || ir->nstlist <= 0) &&
+ (max_hours > 0 && run_time > max_hours*60.0*60.0*0.99) &&
+ gs.sig[eglsSTOPCOND] == 0 && gs.set[eglsSTOPCOND] == 0)
+ {
+ /* Signal to terminate the run */
+ gs.sig[eglsSTOPCOND] = 1;
+ if (fplog)
+ {
+ fprintf(fplog,"\nStep %s: Run time exceeded %.3f hours, will terminate the run\n",gmx_step_str(step,sbuf),max_hours*0.99);
+ }
+ fprintf(stderr, "\nStep %s: Run time exceeded %.3f hours, will terminate the run\n",gmx_step_str(step,sbuf),max_hours*0.99);
+ }
+
+ if (bResetCountersHalfMaxH && MASTER(cr) &&
+ run_time > max_hours*60.0*60.0*0.495)
+ {
+ gs.sig[eglsRESETCOUNTERS] = 1;
+ }
+
+ if (ir->nstlist == -1 && !bRerunMD)
+ {
+ /* When bGStatEveryStep=FALSE, global_stat is only called
+ * when we check the atom displacements, not at NS steps.
+ * This means that also the bonded interaction count check is not
+ * performed immediately after NS. Therefore a few MD steps could
+ * be performed with missing interactions.
+ * But wrong energies are never written to file,
+ * since energies are only written after global_stat
+ * has been called.
+ */
+ if (step >= nlh.step_nscheck)
+ {
+ nlh.nabnsb = natoms_beyond_ns_buffer(ir,fr,&top->cgs,
+ nlh.scale_tot,state->x);
+ }
+ else
+ {
+ /* This is not necessarily true,
+ * but step_nscheck is determined quite conservatively.
+ */
+ nlh.nabnsb = 0;
+ }
+ }
+
+ /* In parallel we only have to check for checkpointing in steps
+ * where we do global communication,
+ * otherwise the other nodes don't know.
+ */
+ if (MASTER(cr) && ((bGStat || !PAR(cr)) &&
+ cpt_period >= 0 &&
+ (cpt_period == 0 ||
+ run_time >= nchkpt*cpt_period*60.0)) &&
+ gs.set[eglsCHKPT] == 0)
+ {
+ gs.sig[eglsCHKPT] = 1;
+ }
+
+ if (bIterations)
+ {
+ gmx_iterate_init(&iterate,bIterations);
+ }
+
+ /* for iterations, we save these vectors, as we will be redoing the calculations */
+ if (bIterations && iterate.bIterate)
+ {
+ copy_coupling_state(state,bufstate,ekind,ekind_save,&(ir->opts));
+ }
+ bFirstIterate = TRUE;
+ while (bFirstIterate || (bIterations && iterate.bIterate))
+ {
+ /* We now restore these vectors to redo the calculation with improved extended variables */
+ if (bIterations)
+ {
+ copy_coupling_state(bufstate,state,ekind_save,ekind,&(ir->opts));
+ }
+
+ /* We make the decision to break or not -after- the calculation of Ekin and Pressure,
+ so scroll down for that logic */
+
+ /* ######### START SECOND UPDATE STEP ################# */
+ /* Box is changed in update() when we do pressure coupling,
+ * but we should still use the old box for energy corrections and when
+ * writing it to the energy file, so it matches the trajectory files for
+ * the same timestep above. Make a copy in a separate array.
+ */
+ copy_mat(state->box,lastbox);
+
+ bOK = TRUE;
+ if (!(bRerunMD && !rerun_fr.bV && !bForceUpdate))
+ {
+ wallcycle_start(wcycle,ewcUPDATE);
+ dvdl = 0;
+ /* UPDATE PRESSURE VARIABLES IN TROTTER FORMULATION WITH CONSTRAINTS */
+ if (bTrotter)
+ {
+ if (bIterations && iterate.bIterate)
+ {
+ if (bFirstIterate)
+ {
+ scalevir = 1;
+ }
+ else
+ {
+ /* we use a new value of scalevir to converge the iterations faster */
+ scalevir = tracevir/trace(shake_vir);
+ }
+ msmul(shake_vir,scalevir,shake_vir);
+ m_add(force_vir,shake_vir,total_vir);
+ clear_mat(shake_vir);
+ }
+ trotter_update(ir,step,ekind,enerd,state,total_vir,mdatoms,&MassQ,trotter_seq,ettTSEQ3);
+ /* We can only do Berendsen coupling after we have summed
+ * the kinetic energy or virial. Since the happens
+ * in global_state after update, we should only do it at
+ * step % nstlist = 1 with bGStatEveryStep=FALSE.
+ */
+ }
+ else
+ {
+ update_tcouple(fplog,step,ir,state,ekind,wcycle,upd,&MassQ,mdatoms);
+ update_pcouple(fplog,step,ir,state,pcoupl_mu,M,wcycle,
+ upd,bInitStep);
+ }
+
+ if (bVV)
+ {
+ /* velocity half-step update */
+ update_coords(fplog,step,ir,mdatoms,state,f,
+ fr->bTwinRange && bNStList,fr->f_twin,fcd,
+ ekind,M,wcycle,upd,FALSE,etrtVELOCITY2,
+ cr,nrnb,constr,&top->idef);
+ }
+
+ /* Above, initialize just copies ekinh into ekin,
+ * it doesn't copy position (for VV),
+ * and entire integrator for MD.
+ */
+
+ if (ir->eI==eiVVAK)
+ {
+ copy_rvecn(state->x,cbuf,0,state->natoms);
+ }
+
+ update_coords(fplog,step,ir,mdatoms,state,f,fr->bTwinRange && bNStList,fr->f_twin,fcd,
+ ekind,M,wcycle,upd,bInitStep,etrtPOSITION,cr,nrnb,constr,&top->idef);
+ wallcycle_stop(wcycle,ewcUPDATE);
+
+ update_constraints(fplog,step,&dvdl,ir,ekind,mdatoms,state,graph,f,
+ &top->idef,shake_vir,force_vir,
+ cr,nrnb,wcycle,upd,constr,
+ bInitStep,FALSE,bCalcEnerPres,state->veta);
+
+ if (ir->eI==eiVVAK)
+ {
+ /* erase F_EKIN and F_TEMP here? */
+ /* just compute the kinetic energy at the half step to perform a trotter step */
+ compute_globals(fplog,gstat,cr,ir,fr,ekind,state,state_global,mdatoms,nrnb,vcm,
+ wcycle,enerd,force_vir,shake_vir,total_vir,pres,mu_tot,
+ constr,NULL,FALSE,lastbox,
+ top_global,&pcurr,top_global->natoms,&bSumEkinhOld,
+ cglo_flags | CGLO_TEMPERATURE
+ );
+ wallcycle_start(wcycle,ewcUPDATE);
+ trotter_update(ir,step,ekind,enerd,state,total_vir,mdatoms,&MassQ,trotter_seq,ettTSEQ4);
+ /* now we know the scaling, we can compute the positions again again */
+ copy_rvecn(cbuf,state->x,0,state->natoms);
+
+ update_coords(fplog,step,ir,mdatoms,state,f,fr->bTwinRange && bNStList,fr->f_twin,fcd,
+ ekind,M,wcycle,upd,bInitStep,etrtPOSITION,cr,nrnb,constr,&top->idef);
+ wallcycle_stop(wcycle,ewcUPDATE);
+
+ /* do we need an extra constraint here? just need to copy out of state->v to upd->xp? */
+ /* are the small terms in the shake_vir here due
+ * to numerical errors, or are they important
+ * physically? I'm thinking they are just errors, but not completely sure.
+ * For now, will call without actually constraining, constr=NULL*/
+ update_constraints(fplog,step,&dvdl,ir,ekind,mdatoms,state,graph,f,
+ &top->idef,tmp_vir,force_vir,
+ cr,nrnb,wcycle,upd,NULL,
+ bInitStep,FALSE,bCalcEnerPres,
+ state->veta);
+ }
+ if (!bOK && !bFFscan)
+ {
+ gmx_fatal(FARGS,"Constraint error: Shake, Lincs or Settle could not solve the constrains");
+ }
+
+ if (fr->bSepDVDL && fplog && do_log)
+ {
+ fprintf(fplog,sepdvdlformat,"Constraint",0.0,dvdl);
+ }
+ enerd->term[F_DHDL_CON] += dvdl;
+ }
+ else if (graph)
+ {
+ /* Need to unshift here */
+ unshift_self(graph,state->box,state->x);
+ }
+
+ if (vsite != NULL)
+ {
+ wallcycle_start(wcycle,ewcVSITECONSTR);
+ if (graph != NULL)
+ {
+ shift_self(graph,state->box,state->x);
+ }
+ construct_vsites(fplog,vsite,state->x,nrnb,ir->delta_t,state->v,
+ top->idef.iparams,top->idef.il,
+ fr->ePBC,fr->bMolPBC,graph,cr,state->box);
+
+ if (graph != NULL)
+ {
+ unshift_self(graph,state->box,state->x);
+ }
+ wallcycle_stop(wcycle,ewcVSITECONSTR);
+ }
+
+ /* ############## IF NOT VV, Calculate globals HERE, also iterate constraints ############ */
+ if (ir->nstlist == -1 && bFirstIterate)
+ {
+ gs.sig[eglsNABNSB] = nlh.nabnsb;
+ }
+ compute_globals(fplog,gstat,cr,ir,fr,ekind,state,state_global,mdatoms,nrnb,vcm,
+ wcycle,enerd,force_vir,shake_vir,total_vir,pres,mu_tot,
+ constr,
+ bFirstIterate ? &gs : NULL,
+ (step_rel % gs.nstms == 0) &&
+ (multisim_nsteps<0 || (step_rel<multisim_nsteps)),
+ lastbox,
+ top_global,&pcurr,top_global->natoms,&bSumEkinhOld,
+ cglo_flags
+ | (!EI_VV(ir->eI) ? CGLO_ENERGY : 0)
+ | (!EI_VV(ir->eI) ? CGLO_TEMPERATURE : 0)
+ | (!EI_VV(ir->eI) || bRerunMD ? CGLO_PRESSURE : 0)
+ | (bIterations && iterate.bIterate ? CGLO_ITERATE : 0)
+ | (bFirstIterate ? CGLO_FIRSTITERATE : 0)
+ | CGLO_CONSTRAINT
+ );
+ if (ir->nstlist == -1 && bFirstIterate)
+ {
+ nlh.nabnsb = gs.set[eglsNABNSB];
+ gs.set[eglsNABNSB] = 0;
+ }
+ /* bIterate is set to keep it from eliminating the old ekin kinetic energy terms */
+ /* ############# END CALC EKIN AND PRESSURE ################# */
+
+ /* Note: this is OK, but there are some numerical precision issues with using the convergence of
+ the virial that should probably be addressed eventually. state->veta has better properies,
+ but what we actually need entering the new cycle is the new shake_vir value. Ideally, we could
+ generate the new shake_vir, but test the veta value for convergence. This will take some thought. */
+
+ if (bIterations &&
+ done_iterating(cr,fplog,step,&iterate,bFirstIterate,
+ trace(shake_vir),&tracevir))
+ {
+ break;
+ }
+ bFirstIterate = FALSE;
+ }
+
+ update_box(fplog,step,ir,mdatoms,state,graph,f,
+ ir->nstlist==-1 ? &nlh.scale_tot : NULL,pcoupl_mu,nrnb,wcycle,upd,bInitStep,FALSE);
+
+ /* ################# END UPDATE STEP 2 ################# */
+ /* #### We now have r(t+dt) and v(t+dt/2) ############# */
+
+ /* The coordinates (x) were unshifted in update */
+ if (bFFscan && (shellfc==NULL || bConverged))
+ {
+ if (print_forcefield(fplog,enerd->term,mdatoms->homenr,
+ f,NULL,xcopy,
+ &(top_global->mols),mdatoms->massT,pres))
+ {
+ if (gmx_parallel_env_initialized())
+ {
+ gmx_finalize();
+ }
+ fprintf(stderr,"\n");
+ exit(0);
+ }
+ }
+ if (!bGStat)
+ {
+ /* We will not sum ekinh_old,
+ * so signal that we still have to do it.
+ */
+ bSumEkinhOld = TRUE;
+ }
+
+ if (bTCR)
+ {
+ /* Only do GCT when the relaxation of shells (minimization) has converged,
+ * otherwise we might be coupling to bogus energies.
+ * In parallel we must always do this, because the other sims might
+ * update the FF.
+ */
+
+ /* Since this is called with the new coordinates state->x, I assume
+ * we want the new box state->box too. / EL 20040121
+ */
+ do_coupling(fplog,oenv,nfile,fnm,tcr,t,step,enerd->term,fr,
+ ir,MASTER(cr),
+ mdatoms,&(top->idef),mu_aver,
+ top_global->mols.nr,cr,
+ state->box,total_vir,pres,
+ mu_tot,state->x,f,bConverged);
+ debug_gmx();
+ }
+
+ /* ######### BEGIN PREPARING EDR OUTPUT ########### */
+
+ /* sum up the foreign energy and dhdl terms */
+ sum_dhdl(enerd,state->lambda,ir);
+
+ /* use the directly determined last velocity, not actually the averaged half steps */
+ if (bTrotter && ir->eI==eiVV)
+ {
+ enerd->term[F_EKIN] = last_ekin;
+ }
+ enerd->term[F_ETOT] = enerd->term[F_EPOT] + enerd->term[F_EKIN];
+
+ if (bVV)
+ {
+ enerd->term[F_ECONSERVED] = enerd->term[F_ETOT] + saved_conserved_quantity;
+ }
+ else
+ {
+ enerd->term[F_ECONSERVED] = enerd->term[F_ETOT] + compute_conserved_from_auxiliary(ir,state,&MassQ);
+ }
+ /* Check for excessively large energies */
+ if (bIonize)
+ {
+#ifdef GMX_DOUBLE
+ real etot_max = 1e200;
+#else
+ real etot_max = 1e30;
+#endif
+ if (fabs(enerd->term[F_ETOT]) > etot_max)
+ {
+ fprintf(stderr,"Energy too large (%g), giving up\n",
+ enerd->term[F_ETOT]);
+ }
+ }
+ /* ######### END PREPARING EDR OUTPUT ########### */
+
+ /* Time for performance */
+ if (((step % stepout) == 0) || bLastStep)
+ {
+ runtime_upd_proc(runtime);
+ }
+
+ /* Output stuff */
+ if (MASTER(cr))
+ {
+ gmx_bool do_dr,do_or;
+
+ if (!(bStartingFromCpt && (EI_VV(ir->eI))))
+ {
+ if (bNstEner)
+ {
+ upd_mdebin(mdebin,bDoDHDL, TRUE,
+ t,mdatoms->tmass,enerd,state,lastbox,
+ shake_vir,force_vir,total_vir,pres,
+ ekind,mu_tot,constr);
+ }
+ else
+ {
+ upd_mdebin_step(mdebin);
+ }
+
+ do_dr = do_per_step(step,ir->nstdisreout);
+ do_or = do_per_step(step,ir->nstorireout);
+
+ print_ebin(outf->fp_ene,do_ene,do_dr,do_or,do_log?fplog:NULL,
+ step,t,
+ eprNORMAL,bCompact,mdebin,fcd,groups,&(ir->opts));
+ }
+ if (ir->ePull != epullNO)
+ {
+ pull_print_output(ir->pull,step,t);
+ }
+
+ if (do_per_step(step,ir->nstlog))
+ {
+ if(fflush(fplog) != 0)
+ {
+ gmx_fatal(FARGS,"Cannot flush logfile - maybe you are out of quota?");
+ }
+ }
+ }
+
+
+ /* Remaining runtime */
+ if (MULTIMASTER(cr) && (do_verbose || gmx_got_usr_signal() ))
+ {
+ if (shellfc)
+ {
+ fprintf(stderr,"\n");
+ }
+ print_time(stderr,runtime,step,ir,cr);
+ }
+
+ /* Replica exchange */
+ bExchanged = FALSE;
+ if ((repl_ex_nst > 0) && (step > 0) && !bLastStep &&
+ do_per_step(step,repl_ex_nst))
+ {
+ bExchanged = replica_exchange(fplog,cr,repl_ex,
+ state_global,enerd->term,
+ state,step,t);
+
+ if (bExchanged && DOMAINDECOMP(cr))
+ {
+ dd_partition_system(fplog,step,cr,TRUE,1,
+ state_global,top_global,ir,
+ state,&f,mdatoms,top,fr,
+ vsite,shellfc,constr,
+ nrnb,wcycle,FALSE);
+ }
+ }
+
+ bFirstStep = FALSE;
+ bInitStep = FALSE;
+ bStartingFromCpt = FALSE;
+
+ /* ####### SET VARIABLES FOR NEXT ITERATION IF THEY STILL NEED IT ###### */
+ /* With all integrators, except VV, we need to retain the pressure
+ * at the current step for coupling at the next step.
+ */
+ if ((state->flags & (1<<estPRES_PREV)) &&
+ (bGStatEveryStep ||
+ (ir->nstpcouple > 0 && step % ir->nstpcouple == 0)))
+ {
+ /* Store the pressure in t_state for pressure coupling
+ * at the next MD step.
+ */
+ copy_mat(pres,state->pres_prev);
+ }
+
+ /* ####### END SET VARIABLES FOR NEXT ITERATION ###### */
+
+ if ( (membed!=NULL) && (!bLastStep) )
+ rescale_membed(step_rel,membed,state_global->x);
+
+ if (bRerunMD)
+ {
+ if (MASTER(cr))
+ {
+ /* read next frame from input trajectory */
+ bNotLastFrame = read_next_frame(oenv,status,&rerun_fr);
+ }
+
+ if (PAR(cr))
+ {
+ rerun_parallel_comm(cr,&rerun_fr,&bNotLastFrame);
+ }
+ }
+
+ if (!bRerunMD || !rerun_fr.bStep)
+ {
+ /* increase the MD step number */
+ step++;
+ step_rel++;
+ }
+
+ cycles = wallcycle_stop(wcycle,ewcSTEP);
+ if (DOMAINDECOMP(cr) && wcycle)
+ {
+ dd_cycles_add(cr->dd,cycles,ddCyclStep);
+ }
+
+ if (step_rel == wcycle_get_reset_counters(wcycle) ||
+ gs.set[eglsRESETCOUNTERS] != 0)
+ {
+ /* Reset all the counters related to performance over the run */
+ reset_all_counters(fplog,cr,step,&step_rel,ir,wcycle,nrnb,runtime);
+ wcycle_set_reset_counters(wcycle,-1);
+ /* Correct max_hours for the elapsed time */
+ max_hours -= run_time/(60.0*60.0);
+ bResetCountersHalfMaxH = FALSE;
+ gs.set[eglsRESETCOUNTERS] = 0;
+ }
+
+ }
+ /* End of main MD loop */
+ debug_gmx();
+
+ /* Stop the time */
+ runtime_end(runtime);
+
+ if (bRerunMD && MASTER(cr))
+ {
+ close_trj(status);
+ }
+
+ if (!(cr->duty & DUTY_PME))
+ {
+ /* Tell the PME only node to finish */
+ gmx_pme_finish(cr);
+ }
+
+ if (MASTER(cr))
+ {
+ if (ir->nstcalcenergy > 0 && !bRerunMD)
+ {
+ print_ebin(outf->fp_ene,FALSE,FALSE,FALSE,fplog,step,t,
+ eprAVER,FALSE,mdebin,fcd,groups,&(ir->opts));
+ }
+ }
+
+ done_mdoutf(outf);
+
+ debug_gmx();
+
+ if (ir->nstlist == -1 && nlh.nns > 0 && fplog)
+ {
+ fprintf(fplog,"Average neighborlist lifetime: %.1f steps, std.dev.: %.1f steps\n",nlh.s1/nlh.nns,sqrt(nlh.s2/nlh.nns - sqr(nlh.s1/nlh.nns)));
+ fprintf(fplog,"Average number of atoms that crossed the half buffer length: %.1f\n\n",nlh.ab/nlh.nns);
+ }
+
+ if (shellfc && fplog)
+ {
+ fprintf(fplog,"Fraction of iterations that converged: %.2f %%\n",
+ (nconverged*100.0)/step_rel);
+ fprintf(fplog,"Average number of force evaluations per MD step: %.2f\n\n",
+ tcount/step_rel);
+ }
+
+ if (repl_ex_nst > 0 && MASTER(cr))
+ {
+ print_replica_exchange_statistics(fplog,repl_ex);
+ }
+
+ runtime->nsteps_done = step_rel;
+
+ return 0;
+}
--- /dev/null
- #ifdef GMX_THREADS
+/* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
+ *
+ *
+ * This source code is part of
+ *
+ * G R O M A C S
+ *
+ * GROningen MAchine for Chemical Simulations
+ *
+ * Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
+ * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
+ * Copyright (c) 2001-2010, The GROMACS development team,
+ * check out http://www.gromacs.org for more information.
+
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version 2
+ * of the License, or (at your option) any later version.
+ *
+ * If you want to redistribute modifications, please consider that
+ * scientific software is very special. Version control is crucial -
+ * bugs must be traceable. We will be happy to consider code for
+ * inclusion in the official distribution, but derived work must not
+ * be called official GROMACS. Details are found in the README & COPYING
+ * files - if they are missing, get the official version at www.gromacs.org.
+ *
+ * To help us fund GROMACS development, we humbly ask that you cite
+ * the papers on the package - you can find them in the top README file.
+ *
+ * For more info, check our website at http://www.gromacs.org
+ *
+ * And Hey:
+ * Gallium Rubidium Oxygen Manganese Argon Carbon Silicon
+ */
+
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+#include <signal.h>
+#include <stdlib.h>
+
+#if ((defined WIN32 || defined _WIN32 || defined WIN64 || defined _WIN64) && !defined __CYGWIN__ && !defined __CYGWIN32__)
+/* _isnan() */
+#include <float.h>
+#endif
+
+#include "typedefs.h"
+#include "smalloc.h"
+#include "sysstuff.h"
+#include "vec.h"
+#include "statutil.h"
+#include "vcm.h"
+#include "mdebin.h"
+#include "nrnb.h"
+#include "calcmu.h"
+#include "index.h"
+#include "vsite.h"
+#include "update.h"
+#include "ns.h"
+#include "trnio.h"
+#include "xtcio.h"
+#include "mdrun.h"
+#include "confio.h"
+#include "network.h"
+#include "pull.h"
+#include "xvgr.h"
+#include "physics.h"
+#include "names.h"
+#include "xmdrun.h"
+#include "ionize.h"
+#include "disre.h"
+#include "orires.h"
+#include "dihre.h"
+#include "pppm.h"
+#include "pme.h"
+#include "mdatoms.h"
+#include "qmmm.h"
+#include "domdec.h"
+#include "partdec.h"
+#include "topsort.h"
+#include "coulomb.h"
+#include "constr.h"
+#include "compute_io.h"
+#include "mvdata.h"
+#include "checkpoint.h"
+#include "mtop_util.h"
+#include "sighandler.h"
+#include "genborn.h"
+#include "string2.h"
+#include "copyrite.h"
+#include "membed.h"
+
- #ifdef GMX_THREADS
++#ifdef GMX_THREAD_MPI
+#include "tmpi.h"
+#endif
+
+/* include even when OpenMM not used to force compilation of do_md_openmm */
+#include "openmm_wrapper.h"
+
+double do_md_openmm(FILE *fplog,t_commrec *cr,int nfile,const t_filenm fnm[],
+ const output_env_t oenv, gmx_bool bVerbose,gmx_bool bCompact,
+ int nstglobalcomm,
+ gmx_vsite_t *vsite,gmx_constr_t constr,
+ int stepout,t_inputrec *ir,
+ gmx_mtop_t *top_global,
+ t_fcdata *fcd,
+ t_state *state_global,
+ t_mdatoms *mdatoms,
+ t_nrnb *nrnb,gmx_wallcycle_t wcycle,
+ gmx_edsam_t ed,t_forcerec *fr,
+ int repl_ex_nst,int repl_ex_seed,
+ gmx_membed_t *membed,
+ real cpt_period,real max_hours,
+ const char *deviceOptions,
+ unsigned long Flags,
+ gmx_runtime_t *runtime)
+{
+ gmx_mdoutf_t *outf;
+ gmx_large_int_t step,step_rel;
+ double run_time;
+ double t,t0,lam0;
+ gmx_bool bSimAnn,
+ bFirstStep,bStateFromTPX,bLastStep,bStartingFromCpt;
+ gmx_bool bInitStep=TRUE;
+ gmx_bool do_ene,do_log, do_verbose,
+ bX,bV,bF,bCPT;
+ tensor force_vir,shake_vir,total_vir,pres;
+ int i,m;
+ int mdof_flags;
+ rvec mu_tot;
+ t_vcm *vcm;
+ int nchkpt=1;
+ gmx_localtop_t *top;
+ t_mdebin *mdebin=NULL;
+ t_state *state=NULL;
+ rvec *f_global=NULL;
+ int n_xtc=-1;
+ rvec *x_xtc=NULL;
+ gmx_enerdata_t *enerd;
+ rvec *f=NULL;
+ gmx_global_stat_t gstat;
+ gmx_update_t upd=NULL;
+ t_graph *graph=NULL;
+ globsig_t gs;
+
+ gmx_groups_t *groups;
+ gmx_ekindata_t *ekind, *ekind_save;
+ gmx_bool bAppend;
+ int a0,a1;
+ matrix lastbox;
+ real reset_counters=0,reset_counters_now=0;
+ char sbuf[STEPSTRSIZE],sbuf2[STEPSTRSIZE];
+ int handled_stop_condition=gmx_stop_cond_none;
+
+ const char *ommOptions = NULL;
+ void *openmmData;
+
+#ifdef GMX_DOUBLE
+ /* Checks in cmake should prevent the compilation in double precision
+ * with OpenMM, but just to be sure we check here.
+ */
+ gmx_fatal(FARGS,"Compilation was performed in double precision, but OpenMM only supports single precision. If you want to use to OpenMM, compile in single precision.");
+#endif
+
+ bAppend = (Flags & MD_APPENDFILES);
+ check_ir_old_tpx_versions(cr,fplog,ir,top_global);
+
+ groups = &top_global->groups;
+
+ /* Initial values */
+ init_md(fplog,cr,ir,oenv,&t,&t0,&state_global->lambda,&lam0,
+ nrnb,top_global,&upd,
+ nfile,fnm,&outf,&mdebin,
+ force_vir,shake_vir,mu_tot,&bSimAnn,&vcm,state_global,Flags);
+
+ clear_mat(total_vir);
+ clear_mat(pres);
+ /* Energy terms and groups */
+ snew(enerd,1);
+ init_enerdata(top_global->groups.grps[egcENER].nr,ir->n_flambda,enerd);
+ snew(f,top_global->natoms);
+
+ /* Kinetic energy data */
+ snew(ekind,1);
+ init_ekindata(fplog,top_global,&(ir->opts),ekind);
+ /* needed for iteration of constraints */
+ snew(ekind_save,1);
+ init_ekindata(fplog,top_global,&(ir->opts),ekind_save);
+ /* Copy the cos acceleration to the groups struct */
+ ekind->cosacc.cos_accel = ir->cos_accel;
+
+ gstat = global_stat_init(ir);
+ debug_gmx();
+
+ {
+ double io = compute_io(ir,top_global->natoms,groups,mdebin->ebin->nener,1);
+ if ((io > 2000) && MASTER(cr))
+ fprintf(stderr,
+ "\nWARNING: This run will generate roughly %.0f Mb of data\n\n",
+ io);
+ }
+
+ top = gmx_mtop_generate_local_top(top_global,ir);
+
+ a0 = 0;
+ a1 = top_global->natoms;
+
+ state = partdec_init_local_state(cr,state_global);
+ f_global = f;
+
+ atoms2md(top_global,ir,0,NULL,a0,a1-a0,mdatoms);
+
+ if (vsite)
+ {
+ set_vsite_top(vsite,top,mdatoms,cr);
+ }
+
+ if (ir->ePBC != epbcNONE && !ir->bPeriodicMols)
+ {
+ graph = mk_graph(fplog,&(top->idef),0,top_global->natoms,FALSE,FALSE);
+ }
+
+ update_mdatoms(mdatoms,state->lambda);
+
+ if (deviceOptions[0]=='\0')
+ {
+ /* empty options, which should default to OpenMM in this build */
+ ommOptions=deviceOptions;
+ }
+ else
+ {
+ if (gmx_strncasecmp(deviceOptions,"OpenMM",6)!=0)
+ {
+ gmx_fatal(FARGS, "This Gromacs version currently only works with OpenMM. Use -device \"OpenMM:<options>\"");
+ }
+ else
+ {
+ ommOptions=strchr(deviceOptions,':');
+ if (NULL!=ommOptions)
+ {
+ /* Increase the pointer to skip the colon */
+ ommOptions++;
+ }
+ }
+ }
+
+ openmmData = openmm_init(fplog, ommOptions, ir, top_global, top, mdatoms, fr, state);
+ please_cite(fplog,"Friedrichs2009");
+
+ if (MASTER(cr))
+ {
+ /* Update mdebin with energy history if appending to output files */
+ if ( Flags & MD_APPENDFILES )
+ {
+ restore_energyhistory_from_state(mdebin,&state_global->enerhist);
+ }
+ /* Set the initial energy history in state to zero by updating once */
+ update_energyhistory(&state_global->enerhist,mdebin);
+ }
+
+ if (constr)
+ {
+ set_constraints(constr,top,ir,mdatoms,cr);
+ }
+
+ if (!ir->bContinuation)
+ {
+ if (mdatoms->cFREEZE && (state->flags & (1<<estV)))
+ {
+ /* Set the velocities of frozen particles to zero */
+ for (i=mdatoms->start; i<mdatoms->start+mdatoms->homenr; i++)
+ {
+ for (m=0; m<DIM; m++)
+ {
+ if (ir->opts.nFreeze[mdatoms->cFREEZE[i]][m])
+ {
+ state->v[i][m] = 0;
+ }
+ }
+ }
+ }
+
+ if (constr)
+ {
+ /* Constrain the initial coordinates and velocities */
+ do_constrain_first(fplog,constr,ir,mdatoms,state,f,
+ graph,cr,nrnb,fr,top,shake_vir);
+ }
+ if (vsite)
+ {
+ /* Construct the virtual sites for the initial configuration */
+ construct_vsites(fplog,vsite,state->x,nrnb,ir->delta_t,NULL,
+ top->idef.iparams,top->idef.il,
+ fr->ePBC,fr->bMolPBC,graph,cr,state->box);
+ }
+ }
+
+ debug_gmx();
+
+ if (MASTER(cr))
+ {
+ char tbuf[20];
+ fprintf(fplog,"Initial temperature: %g K\n",enerd->term[F_TEMP]);
+ fprintf(stderr,"starting mdrun '%s'\n",
+ *(top_global->name));
+ if (ir->nsteps >= 0)
+ {
+ sprintf(tbuf,"%8.1f",(ir->init_step+ir->nsteps)*ir->delta_t);
+ }
+ else
+ {
+ sprintf(tbuf,"%s","infinite");
+ }
+ if (ir->init_step > 0)
+ {
+ fprintf(stderr,"%s steps, %s ps (continuing from step %s, %8.1f ps).\n",
+ gmx_step_str(ir->init_step+ir->nsteps,sbuf),tbuf,
+ gmx_step_str(ir->init_step,sbuf2),
+ ir->init_step*ir->delta_t);
+ }
+ else
+ {
+ fprintf(stderr,"%s steps, %s ps.\n",
+ gmx_step_str(ir->nsteps,sbuf),tbuf);
+ }
+ }
+
+ fprintf(fplog,"\n");
+
+ /* Set and write start time */
+ runtime_start(runtime);
+ print_date_and_time(fplog,cr->nodeid,"Started mdrun",runtime);
+ wallcycle_start(wcycle,ewcRUN);
+ if (fplog)
+ fprintf(fplog,"\n");
+
+ /* safest point to do file checkpointing is here. More general point would be immediately before integrator call */
+
+ debug_gmx();
+ /***********************************************************
+ *
+ * Loop over MD steps
+ *
+ ************************************************************/
+
+ /* loop over MD steps or if rerunMD to end of input trajectory */
+ bFirstStep = TRUE;
+ /* Skip the first Nose-Hoover integration when we get the state from tpx */
+ bStateFromTPX = !opt2bSet("-cpi",nfile,fnm);
+ bInitStep = bFirstStep && bStateFromTPX;
+ bStartingFromCpt = (Flags & MD_STARTFROMCPT) && bInitStep;
+ bLastStep = FALSE;
+
+ init_global_signals(&gs,cr,ir,repl_ex_nst);
+
+ step = ir->init_step;
+ step_rel = 0;
+
+ while (!bLastStep)
+ {
+ wallcycle_start(wcycle,ewcSTEP);
+
+ bLastStep = (step_rel == ir->nsteps);
+ t = t0 + step*ir->delta_t;
+
+ if (gs.set[eglsSTOPCOND] != 0)
+ {
+ bLastStep = TRUE;
+ }
+
+ do_log = do_per_step(step,ir->nstlog) || bFirstStep || bLastStep;
+ do_verbose = bVerbose &&
+ (step % stepout == 0 || bFirstStep || bLastStep);
+
+ if (MASTER(cr) && do_log)
+ {
+ print_ebin_header(fplog,step,t,state->lambda);
+ }
+
+ clear_mat(force_vir);
+
+ /* We write a checkpoint at this MD step when:
+ * either when we signalled through gs (in OpenMM NS works different),
+ * or at the last step (but not when we do not want confout),
+ * but never at the first step.
+ */
+ bCPT = ((gs.set[eglsCHKPT] ||
+ (bLastStep && (Flags & MD_CONFOUT))) &&
+ step > ir->init_step );
+ if (bCPT)
+ {
+ gs.set[eglsCHKPT] = 0;
+ }
+
+ /* Now we have the energies and forces corresponding to the
+ * coordinates at time t. We must output all of this before
+ * the update.
+ * for RerunMD t is read from input trajectory
+ */
+ mdof_flags = 0;
+ if (do_per_step(step,ir->nstxout))
+ {
+ mdof_flags |= MDOF_X;
+ }
+ if (do_per_step(step,ir->nstvout))
+ {
+ mdof_flags |= MDOF_V;
+ }
+ if (do_per_step(step,ir->nstfout))
+ {
+ mdof_flags |= MDOF_F;
+ }
+ if (do_per_step(step,ir->nstxtcout))
+ {
+ mdof_flags |= MDOF_XTC;
+ }
+ if (bCPT)
+ {
+ mdof_flags |= MDOF_CPT;
+ };
+ do_ene = (do_per_step(step,ir->nstenergy) || bLastStep);
+
+ if (mdof_flags != 0 || do_ene || do_log)
+ {
+ wallcycle_start(wcycle,ewcTRAJ);
+ bF = (mdof_flags & MDOF_F);
+ bX = (mdof_flags & (MDOF_X | MDOF_XTC | MDOF_CPT));
+ bV = (mdof_flags & (MDOF_V | MDOF_CPT));
+
+ openmm_copy_state(openmmData, state, &t, f, enerd, bX, bV, bF, do_ene);
+
+ upd_mdebin(mdebin,FALSE,TRUE,
+ t,mdatoms->tmass,enerd,state,lastbox,
+ shake_vir,force_vir,total_vir,pres,
+ ekind,mu_tot,constr);
+ print_ebin(outf->fp_ene,do_ene,FALSE,FALSE,do_log?fplog:NULL,
+ step,t,
+ eprNORMAL,bCompact,mdebin,fcd,groups,&(ir->opts));
+ write_traj(fplog,cr,outf,mdof_flags,top_global,
+ step,t,state,state_global,f,f_global,&n_xtc,&x_xtc);
+ if (bCPT)
+ {
+ nchkpt++;
+ bCPT = FALSE;
+ }
+ debug_gmx();
+ if (bLastStep && step_rel == ir->nsteps &&
+ (Flags & MD_CONFOUT) && MASTER(cr))
+ {
+ /* x and v have been collected in write_traj,
+ * because a checkpoint file will always be written
+ * at the last step.
+ */
+ fprintf(stderr,"\nWriting final coordinates.\n");
+ if (ir->ePBC != epbcNONE && !ir->bPeriodicMols)
+ {
+ /* Make molecules whole only for confout writing */
+ do_pbc_mtop(fplog,ir->ePBC,state->box,top_global,state_global->x);
+ }
+ write_sto_conf_mtop(ftp2fn(efSTO,nfile,fnm),
+ *top_global->name,top_global,
+ state_global->x,state_global->v,
+ ir->ePBC,state->box);
+ debug_gmx();
+ }
+ wallcycle_stop(wcycle,ewcTRAJ);
+ }
+
+ /* Determine the wallclock run time up till now */
+ run_time = gmx_gettime() - (double)runtime->real;
+
+ /* Check whether everything is still allright */
+ if (((int)gmx_get_stop_condition() > handled_stop_condition)
++#ifdef GMX_THREAD_MPI
+ && MASTER(cr)
+#endif
+ )
+ {
+ /* this is just make gs.sig compatible with the hack
+ of sending signals around by MPI_Reduce with together with
+ other floats */
+ /* NOTE: this only works for serial code. For code that allows
+ MPI nodes to propagate their condition, see kernel/md.c*/
+ if ( gmx_get_stop_condition() == gmx_stop_cond_next_ns )
+ gs.set[eglsSTOPCOND]=1;
+ if ( gmx_get_stop_condition() == gmx_stop_cond_next )
+ gs.set[eglsSTOPCOND]=1;
+ /* < 0 means stop at next step, > 0 means stop at next NS step */
+ if (fplog)
+ {
+ fprintf(fplog,
+ "\n\nReceived the %s signal, stopping at the next %sstep\n\n",
+ gmx_get_signal_name(),
+ gs.sig[eglsSTOPCOND]==1 ? "NS " : "");
+ fflush(fplog);
+ }
+ fprintf(stderr,
+ "\n\nReceived the %s signal, stopping at the next %sstep\n\n",
+ gmx_get_signal_name(),
+ gs.sig[eglsSTOPCOND]==1 ? "NS " : "");
+ fflush(stderr);
+ handled_stop_condition=(int)gmx_get_stop_condition();
+ }
+ else if (MASTER(cr) &&
+ (max_hours > 0 && run_time > max_hours*60.0*60.0*0.99) &&
+ gs.set[eglsSTOPCOND] == 0)
+ {
+ /* Signal to terminate the run */
+ gs.set[eglsSTOPCOND] = 1;
+ if (fplog)
+ {
+ fprintf(fplog,"\nStep %s: Run time exceeded %.3f hours, will terminate the run\n",gmx_step_str(step,sbuf),max_hours*0.99);
+ }
+ fprintf(stderr, "\nStep %s: Run time exceeded %.3f hours, will terminate the run\n",gmx_step_str(step,sbuf),max_hours*0.99);
+ }
+
+ /* checkpoints */
+ if (MASTER(cr) && (cpt_period >= 0 &&
+ (cpt_period == 0 ||
+ run_time >= nchkpt*cpt_period*60.0)) &&
+ gs.set[eglsCHKPT] == 0)
+ {
+ gs.set[eglsCHKPT] = 1;
+ }
+
+ /* Time for performance */
+ if (((step % stepout) == 0) || bLastStep)
+ {
+ runtime_upd_proc(runtime);
+ }
+
+ if (do_per_step(step,ir->nstlog))
+ {
+ if (fflush(fplog) != 0)
+ {
+ gmx_fatal(FARGS,"Cannot flush logfile - maybe you are out of quota?");
+ }
+ }
+
+ /* Remaining runtime */
+ if (MULTIMASTER(cr) && (do_verbose || gmx_got_usr_signal() ))
+ {
+ print_time(stderr,runtime,step,ir,cr);
+ }
+
+ bFirstStep = FALSE;
+ bInitStep = FALSE;
+ bStartingFromCpt = FALSE;
+ step++;
+ step_rel++;
+
+ openmm_take_one_step(openmmData);
+ }
+ /* End of main MD loop */
+ debug_gmx();
+
+ /* Stop the time */
+ runtime_end(runtime);
+
+ if (MASTER(cr))
+ {
+ if (ir->nstcalcenergy > 0)
+ {
+ print_ebin(outf->fp_ene,FALSE,FALSE,FALSE,fplog,step,t,
+ eprAVER,FALSE,mdebin,fcd,groups,&(ir->opts));
+ }
+ }
+
+ openmm_cleanup(fplog, openmmData);
+
+ done_mdoutf(outf);
+
+ debug_gmx();
+
+ runtime->nsteps_done = step_rel;
+
+ return 0;
+}
--- /dev/null
- #ifdef GMX_THREADS
+/* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
+ *
+ *
+ * This source code is part of
+ *
+ * G R O M A C S
+ *
+ * GROningen MAchine for Chemical Simulations
+ *
+ * VERSION 3.2.0
+ * Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
+ * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
+ * Copyright (c) 2001-2004, The GROMACS development team,
+ * check out http://www.gromacs.org for more information.
+
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version 2
+ * of the License, or (at your option) any later version.
+ *
+ * If you want to redistribute modifications, please consider that
+ * scientific software is very special. Version control is crucial -
+ * bugs must be traceable. We will be happy to consider code for
+ * inclusion in the official distribution, but derived work must not
+ * be called official GROMACS. Details are found in the README & COPYING
+ * files - if they are missing, get the official version at www.gromacs.org.
+ *
+ * To help us fund GROMACS development, we humbly ask that you cite
+ * the papers on the package - you can find them in the top README file.
+ *
+ * For more info, check our website at http://www.gromacs.org
+ *
+ * And Hey:
+ * Gallium Rubidium Oxygen Manganese Argon Carbon Silicon
+ */
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+#include "typedefs.h"
+#include "macros.h"
+#include "copyrite.h"
+#include "main.h"
+#include "statutil.h"
+#include "smalloc.h"
+#include "futil.h"
+#include "smalloc.h"
+#include "edsam.h"
+#include "mdrun.h"
+#include "xmdrun.h"
+#include "checkpoint.h"
- #ifdef GMX_THREADS
++#ifdef GMX_THREAD_MPI
+#include "thread_mpi.h"
+#endif
+
+/* afm stuf */
+#include "pull.h"
+
+int main(int argc,char *argv[])
+{
+ const char *desc[] = {
+ #ifdef GMX_OPENMM
+ "This is an experimental release of GROMACS for accelerated",
+ "Molecular Dynamics simulations on GPU processors. Support is provided",
+ "by the OpenMM library (https://simtk.org/home/openmm).[PAR]",
+ "*Warning*[BR]",
+ "This release is targeted at developers and advanced users and",
+ "care should be taken before production use. The following should be",
+ "noted before using the program:[PAR]",
+ " * The current release runs only on modern nVidia GPU hardware with CUDA support.",
+ "Make sure that the necessary CUDA drivers and libraries for your operating system",
+ "are already installed. The CUDA SDK also should be installed in order to compile",
+ "the program from source (http://www.nvidia.com/object/cuda_home.html).[PAR]",
+ " * Multiple GPU cards are not supported.[PAR]",
+ " * Only a small subset of the GROMACS features and options are supported on the GPUs.",
+ "See below for a detailed list.[PAR]",
+ " * Consumer level GPU cards are known to often have problems with faulty memory.",
+ "It is recommended that a full memory check of the cards is done at least once",
+ "(for example, using the memtest=full option).",
+ "A partial memory check (for example, memtest=15) before and",
+ "after the simulation run would help spot",
+ "problems resulting from processor overheating.[PAR]",
+ " * The maximum size of the simulated systems depends on the available",
+ "GPU memory,for example, a GTX280 with 1GB memory has been tested with systems",
+ "of up to about 100,000 atoms.[PAR]",
+ " * In order to take a full advantage of the GPU platform features, many algorithms",
+ "have been implemented in a very different way than they are on the CPUs.",
+ "Therefore numercal correspondence between properties of the state of",
+ "simulated systems should not be expected. Moreover, the values will likely vary",
+ "when simulations are done on different GPU hardware.[PAR]",
+ " * Frequent retrieval of system state information such as",
+ "trajectory coordinates and energies can greatly influence the performance",
+ "of the program due to slow CPU<->GPU memory transfer speed.[PAR]",
+ " * MD algorithms are complex, and although the Gromacs code is highly tuned for them,",
+ "they often do not translate very well onto the streaming architetures.",
+ "Realistic expectations about the achievable speed-up from test with GTX280:",
+ "For small protein systems in implicit solvent using all-vs-all kernels the acceleration",
+ "can be as high as 20 times, but in most other setups involving cutoffs and PME the",
+ "acceleration is usually only ~4 times relative to a 3GHz CPU.[PAR]",
+ "Supported features:[PAR]",
+ " * Integrators: md/md-vv/md-vv-avek, sd/sd1 and bd.\n",
+ " * Long-range interactions (option coulombtype): Reaction-Field, Ewald, PME, and cut-off (for Implicit Solvent only)\n",
+ " * Temperature control: Supported only with the md/md-vv/md-vv-avek, sd/sd1 and bd integrators.\n",
+ " * Pressure control: Supported.\n",
+ " * Implicit solvent: Supported.\n",
+ "A detailed description can be found on the GROMACS website:\n",
+ "http://www.gromacs.org/gpu[PAR]",
+/* From the original mdrun documentaion */
+ "The [TT]mdrun[tt] program reads the run input file ([TT]-s[tt])",
+ "and distributes the topology over nodes if needed.",
+ "[TT]mdrun[tt] produces at least four output files.",
+ "A single log file ([TT]-g[tt]) is written, unless the option",
+ "[TT]-seppot[tt] is used, in which case each node writes a log file.",
+ "The trajectory file ([TT]-o[tt]), contains coordinates, velocities and",
+ "optionally forces.",
+ "The structure file ([TT]-c[tt]) contains the coordinates and",
+ "velocities of the last step.",
+ "The energy file ([TT]-e[tt]) contains energies, the temperature,",
+ "pressure, etc, a lot of these things are also printed in the log file.",
+ "Optionally coordinates can be written to a compressed trajectory file",
+ "([TT]-x[tt]).[PAR]",
+/* openmm specific information */
+ "Usage with OpenMM:[BR]",
+ "[TT]mdrun -device \"OpenMM:platform=Cuda,memtest=15,deviceid=0,force-device=no\"[tt][PAR]",
+ "Options:[PAR]",
+ " [TT]platform[tt] = Cuda\t\t:\tThe only available value. OpenCL support will be available in future.\n",
+ " [TT]memtest[tt] = 15\t\t:\tRun a partial, random GPU memory test for the given amount of seconds. A full test",
+ "(recommended!) can be run with \"memtest=full\". Memory testing can be disabled with \"memtest=off\".\n",
+ " [TT]deviceid[tt] = 0\t\t:\tSpecify the target device when multiple cards are present.",
+ "Only one card can be used at any given time though.\n",
+ " [TT]force-device[tt] = no\t\t:\tIf set to \"yes\" [TT]mdrun[tt] will be forced to execute on",
+ "hardware that is not officially supported. GPU acceleration can also be achieved on older",
+ "but Cuda capable cards, although the simulation might be too slow, and the memory limits too strict.",
+#else
+ "The [TT]mdrun[tt] program is the main computational chemistry engine",
+ "within GROMACS. Obviously, it performs Molecular Dynamics simulations,",
+ "but it can also perform Stochastic Dynamics, Energy Minimization,",
+ "test particle insertion or (re)calculation of energies.",
+ "Normal mode analysis is another option. In this case [TT]mdrun[tt]",
+ "builds a Hessian matrix from single conformation.",
+ "For usual Normal Modes-like calculations, make sure that",
+ "the structure provided is properly energy-minimized.",
+ "The generated matrix can be diagonalized by [TT]g_nmeig[tt].[PAR]",
+ "The [TT]mdrun[tt] program reads the run input file ([TT]-s[tt])",
+ "and distributes the topology over nodes if needed.",
+ "[TT]mdrun[tt] produces at least four output files.",
+ "A single log file ([TT]-g[tt]) is written, unless the option",
+ "[TT]-seppot[tt] is used, in which case each node writes a log file.",
+ "The trajectory file ([TT]-o[tt]), contains coordinates, velocities and",
+ "optionally forces.",
+ "The structure file ([TT]-c[tt]) contains the coordinates and",
+ "velocities of the last step.",
+ "The energy file ([TT]-e[tt]) contains energies, the temperature,",
+ "pressure, etc, a lot of these things are also printed in the log file.",
+ "Optionally coordinates can be written to a compressed trajectory file",
+ "([TT]-x[tt]).[PAR]",
+ "The option [TT]-dhdl[tt] is only used when free energy calculation is",
+ "turned on.[PAR]",
+ "When [TT]mdrun[tt] is started using MPI with more than 1 node, parallelization",
+ "is used. By default domain decomposition is used, unless the [TT]-pd[tt]",
+ "option is set, which selects particle decomposition.[PAR]",
+ "With domain decomposition, the spatial decomposition can be set",
+ "with option [TT]-dd[tt]. By default [TT]mdrun[tt] selects a good decomposition.",
+ "The user only needs to change this when the system is very inhomogeneous.",
+ "Dynamic load balancing is set with the option [TT]-dlb[tt],",
+ "which can give a significant performance improvement,",
+ "especially for inhomogeneous systems. The only disadvantage of",
+ "dynamic load balancing is that runs are no longer binary reproducible,",
+ "but in most cases this is not important.",
+ "By default the dynamic load balancing is automatically turned on",
+ "when the measured performance loss due to load imbalance is 5% or more.",
+ "At low parallelization these are the only important options",
+ "for domain decomposition.",
+ "At high parallelization the options in the next two sections",
+ "could be important for increasing the performace.",
+ "[PAR]",
+ "When PME is used with domain decomposition, separate nodes can",
+ "be assigned to do only the PME mesh calculation;",
+ "this is computationally more efficient starting at about 12 nodes.",
+ "The number of PME nodes is set with option [TT]-npme[tt],",
+ "this can not be more than half of the nodes.",
+ "By default [TT]mdrun[tt] makes a guess for the number of PME",
+ "nodes when the number of nodes is larger than 11 or performance wise",
+ "not compatible with the PME grid x dimension.",
+ "But the user should optimize npme. Performance statistics on this issue",
+ "are written at the end of the log file.",
+ "For good load balancing at high parallelization, the PME grid x and y",
+ "dimensions should be divisible by the number of PME nodes",
+ "(the simulation will run correctly also when this is not the case).",
+ "[PAR]",
+ "This section lists all options that affect the domain decomposition.",
+ "[PAR]",
+ "Option [TT]-rdd[tt] can be used to set the required maximum distance",
+ "for inter charge-group bonded interactions.",
+ "Communication for two-body bonded interactions below the non-bonded",
+ "cut-off distance always comes for free with the non-bonded communication.",
+ "Atoms beyond the non-bonded cut-off are only communicated when they have",
+ "missing bonded interactions; this means that the extra cost is minor",
+ "and nearly indepedent of the value of [TT]-rdd[tt].",
+ "With dynamic load balancing option [TT]-rdd[tt] also sets",
+ "the lower limit for the domain decomposition cell sizes.",
+ "By default [TT]-rdd[tt] is determined by [TT]mdrun[tt] based on",
+ "the initial coordinates. The chosen value will be a balance",
+ "between interaction range and communication cost.",
+ "[PAR]",
+ "When inter charge-group bonded interactions are beyond",
+ "the bonded cut-off distance, [TT]mdrun[tt] terminates with an error message.",
+ "For pair interactions and tabulated bonds",
+ "that do not generate exclusions, this check can be turned off",
+ "with the option [TT]-noddcheck[tt].",
+ "[PAR]",
+ "When constraints are present, option [TT]-rcon[tt] influences",
+ "the cell size limit as well.",
+ "Atoms connected by NC constraints, where NC is the LINCS order plus 1,",
+ "should not be beyond the smallest cell size. A error message is",
+ "generated when this happens and the user should change the decomposition",
+ "or decrease the LINCS order and increase the number of LINCS iterations.",
+ "By default [TT]mdrun[tt] estimates the minimum cell size required for P-LINCS",
+ "in a conservative fashion. For high parallelization it can be useful",
+ "to set the distance required for P-LINCS with the option [TT]-rcon[tt].",
+ "[PAR]",
+ "The [TT]-dds[tt] option sets the minimum allowed x, y and/or z scaling",
+ "of the cells with dynamic load balancing. [TT]mdrun[tt] will ensure that",
+ "the cells can scale down by at least this factor. This option is used",
+ "for the automated spatial decomposition (when not using [TT]-dd[tt])",
+ "as well as for determining the number of grid pulses, which in turn",
+ "sets the minimum allowed cell size. Under certain circumstances",
+ "the value of [TT]-dds[tt] might need to be adjusted to account for",
+ "high or low spatial inhomogeneity of the system.",
+ "[PAR]",
+ "The option [TT]-gcom[tt] can be used to only do global communication",
+ "every n steps.",
+ "This can improve performance for highly parallel simulations",
+ "where this global communication step becomes the bottleneck.",
+ "For a global thermostat and/or barostat the temperature",
+ "and/or pressure will also only be updated every [TT]-gcom[tt] steps.",
+ "By default it is set to the minimum of nstcalcenergy and nstlist.[PAR]",
+ "With [TT]-rerun[tt] an input trajectory can be given for which ",
+ "forces and energies will be (re)calculated. Neighbor searching will be",
+ "performed for every frame, unless [TT]nstlist[tt] is zero",
+ "(see the [TT].mdp[tt] file).[PAR]",
+ "ED (essential dynamics) sampling is switched on by using the [TT]-ei[tt]",
+ "flag followed by an [TT].edi[tt] file.",
+ "The [TT].edi[tt] file can be produced using options in the essdyn",
+ "menu of the WHAT IF program. [TT]mdrun[tt] produces a [TT].edo[tt] file that",
+ "contains projections of positions, velocities and forces onto selected",
+ "eigenvectors.[PAR]",
+ "When user-defined potential functions have been selected in the",
+ "[TT].mdp[tt] file the [TT]-table[tt] option is used to pass [TT]mdrun[tt]",
+ "a formatted table with potential functions. The file is read from",
+ "either the current directory or from the [TT]GMXLIB[tt] directory.",
+ "A number of pre-formatted tables are presented in the [TT]GMXLIB[tt] dir,",
+ "for 6-8, 6-9, 6-10, 6-11, 6-12 Lennard-Jones potentials with",
+ "normal Coulomb.",
+ "When pair interactions are present, a separate table for pair interaction",
+ "functions is read using the [TT]-tablep[tt] option.[PAR]",
+ "When tabulated bonded functions are present in the topology,",
+ "interaction functions are read using the [TT]-tableb[tt] option.",
+ "For each different tabulated interaction type the table file name is",
+ "modified in a different way: before the file extension an underscore is",
+ "appended, then a 'b' for bonds, an 'a' for angles or a 'd' for dihedrals",
+ "and finally the table number of the interaction type.[PAR]",
+ "The options [TT]-px[tt] and [TT]-pf[tt] are used for writing pull COM",
+ "coordinates and forces when pulling is selected",
+ "in the [TT].mdp[tt] file.[PAR]",
+ "With [TT]-multi[tt] or [TT]-multidir[tt], multiple systems can be ",
+ "simulated in parallel.",
+ "As many input files/directories are required as the number of systems. ",
+ "The [TT]-multidir[tt] option takes a list of directories (one for each ",
+ "system) and runs in each of them, using the input/output file names, ",
+ "such as specified by e.g. the [TT]-s[tt] option, relative to these ",
+ "directories.",
+ "With [TT]-multi[tt], the system number is appended to the run input ",
+ "and each output filename, for instance [TT]topol.tpr[tt] becomes",
+ "[TT]topol0.tpr[tt], [TT]topol1.tpr[tt] etc.",
+ "The number of nodes per system is the total number of nodes",
+ "divided by the number of systems.",
+ "One use of this option is for NMR refinement: when distance",
+ "or orientation restraints are present these can be ensemble averaged",
+ "over all the systems.[PAR]",
+ "With [TT]-replex[tt] replica exchange is attempted every given number",
+ "of steps. The number of replicas is set with the [TT]-multi[tt] or ",
+ "[TT]-multidir[tt] option, described above.",
+ "All run input files should use a different coupling temperature,",
+ "the order of the files is not important. The random seed is set with",
+ "[TT]-reseed[tt]. The velocities are scaled and neighbor searching",
+ "is performed after every exchange.[PAR]",
+ "Finally some experimental algorithms can be tested when the",
+ "appropriate options have been given. Currently under",
+ "investigation are: polarizability and X-ray bombardments.",
+ "[PAR]",
+ "The option [TT]-membed[dd] does what used to be g_membed, i.e. embed",
+ "a protein into a membrane. The data file should contain the options",
+ "that where passed to g_membed before. The [TT]-mn[tt] and [TT]-mp[tt]",
+ "both apply to this as well.",
+ "[PAR]",
+ "The option [TT]-pforce[tt] is useful when you suspect a simulation",
+ "crashes due to too large forces. With this option coordinates and",
+ "forces of atoms with a force larger than a certain value will",
+ "be printed to stderr.",
+ "[PAR]",
+ "Checkpoints containing the complete state of the system are written",
+ "at regular intervals (option [TT]-cpt[tt]) to the file [TT]-cpo[tt],",
+ "unless option [TT]-cpt[tt] is set to -1.",
+ "The previous checkpoint is backed up to [TT]state_prev.cpt[tt] to",
+ "make sure that a recent state of the system is always available,",
+ "even when the simulation is terminated while writing a checkpoint.",
+ "With [TT]-cpnum[tt] all checkpoint files are kept and appended",
+ "with the step number.",
+ "A simulation can be continued by reading the full state from file",
+ "with option [TT]-cpi[tt]. This option is intelligent in the way that",
+ "if no checkpoint file is found, Gromacs just assumes a normal run and",
+ "starts from the first step of the [TT].tpr[tt] file. By default the output",
+ "will be appending to the existing output files. The checkpoint file",
+ "contains checksums of all output files, such that you will never",
+ "loose data when some output files are modified, corrupt or removed.",
+ "There are three scenarios with [TT]-cpi[tt]:[PAR]",
+ "[TT]*[tt] no files with matching names are present: new output files are written[PAR]",
+ "[TT]*[tt] all files are present with names and checksums matching those stored",
+ "in the checkpoint file: files are appended[PAR]",
+ "[TT]*[tt] otherwise no files are modified and a fatal error is generated[PAR]",
+ "With [TT]-noappend[tt] new output files are opened and the simulation",
+ "part number is added to all output file names.",
+ "Note that in all cases the checkpoint file itself is not renamed",
+ "and will be overwritten, unless its name does not match",
+ "the [TT]-cpo[tt] option.",
+ "[PAR]",
+ "With checkpointing the output is appended to previously written",
+ "output files, unless [TT]-noappend[tt] is used or none of the previous",
+ "output files are present (except for the checkpoint file).",
+ "The integrity of the files to be appended is verified using checksums",
+ "which are stored in the checkpoint file. This ensures that output can",
+ "not be mixed up or corrupted due to file appending. When only some",
+ "of the previous output files are present, a fatal error is generated",
+ "and no old output files are modified and no new output files are opened.",
+ "The result with appending will be the same as from a single run.",
+ "The contents will be binary identical, unless you use a different number",
+ "of nodes or dynamic load balancing or the FFT library uses optimizations",
+ "through timing.",
+ "[PAR]",
+ "With option [TT]-maxh[tt] a simulation is terminated and a checkpoint",
+ "file is written at the first neighbor search step where the run time",
+ "exceeds [TT]-maxh[tt]*0.99 hours.",
+ "[PAR]",
+ "When [TT]mdrun[tt] receives a TERM signal, it will set nsteps to the current",
+ "step plus one. When [TT]mdrun[tt] receives an INT signal (e.g. when ctrl+C is",
+ "pressed), it will stop after the next neighbor search step ",
+ "(with nstlist=0 at the next step).",
+ "In both cases all the usual output will be written to file.",
+ "When running with MPI, a signal to one of the [TT]mdrun[tt] processes",
+ "is sufficient, this signal should not be sent to mpirun or",
+ "the [TT]mdrun[tt] process that is the parent of the others.",
+ "[PAR]",
+ "When [TT]mdrun[tt] is started with MPI, it does not run niced by default."
+#endif
+ };
+ t_commrec *cr;
+ t_filenm fnm[] = {
+ { efTPX, NULL, NULL, ffREAD },
+ { efTRN, "-o", NULL, ffWRITE },
+ { efXTC, "-x", NULL, ffOPTWR },
+ { efCPT, "-cpi", NULL, ffOPTRD },
+ { efCPT, "-cpo", NULL, ffOPTWR },
+ { efSTO, "-c", "confout", ffWRITE },
+ { efEDR, "-e", "ener", ffWRITE },
+ { efLOG, "-g", "md", ffWRITE },
+ { efXVG, "-dhdl", "dhdl", ffOPTWR },
+ { efXVG, "-field", "field", ffOPTWR },
+ { efXVG, "-table", "table", ffOPTRD },
+ { efXVG, "-tabletf", "tabletf", ffOPTRD },
+ { efXVG, "-tablep", "tablep", ffOPTRD },
+ { efXVG, "-tableb", "table", ffOPTRD },
+ { efTRX, "-rerun", "rerun", ffOPTRD },
+ { efXVG, "-tpi", "tpi", ffOPTWR },
+ { efXVG, "-tpid", "tpidist", ffOPTWR },
+ { efEDI, "-ei", "sam", ffOPTRD },
+ { efEDO, "-eo", "sam", ffOPTWR },
+ { efGCT, "-j", "wham", ffOPTRD },
+ { efGCT, "-jo", "bam", ffOPTWR },
+ { efXVG, "-ffout", "gct", ffOPTWR },
+ { efXVG, "-devout", "deviatie", ffOPTWR },
+ { efXVG, "-runav", "runaver", ffOPTWR },
+ { efXVG, "-px", "pullx", ffOPTWR },
+ { efXVG, "-pf", "pullf", ffOPTWR },
+ { efXVG, "-ro", "rotation", ffOPTWR },
+ { efLOG, "-ra", "rotangles",ffOPTWR },
+ { efLOG, "-rs", "rotslabs", ffOPTWR },
+ { efLOG, "-rt", "rottorque",ffOPTWR },
+ { efMTX, "-mtx", "nm", ffOPTWR },
+ { efNDX, "-dn", "dipole", ffOPTWR },
+ { efDAT, "-membed", "membed", ffOPTRD },
+ { efTOP, "-mp", "membed", ffOPTRD },
+ { efNDX, "-mn", "membed", ffOPTRD },
+ { efRND, "-multidir",NULL, ffOPTRDMULT}
+ };
+#define NFILE asize(fnm)
+
+ /* Command line options ! */
+ gmx_bool bCart = FALSE;
+ gmx_bool bPPPME = FALSE;
+ gmx_bool bPartDec = FALSE;
+ gmx_bool bDDBondCheck = TRUE;
+ gmx_bool bDDBondComm = TRUE;
+ gmx_bool bVerbose = FALSE;
+ gmx_bool bCompact = TRUE;
+ gmx_bool bSepPot = FALSE;
+ gmx_bool bRerunVSite = FALSE;
+ gmx_bool bIonize = FALSE;
+ gmx_bool bConfout = TRUE;
+ gmx_bool bReproducible = FALSE;
+
+ int npme=-1;
+ int nmultisim=0;
+ int nstglobalcomm=-1;
+ int repl_ex_nst=0;
+ int repl_ex_seed=-1;
+ int nstepout=100;
+ int nthreads=0; /* set to determine # of threads automatically */
+ int resetstep=-1;
+
+ rvec realddxyz={0,0,0};
+ const char *ddno_opt[ddnoNR+1] =
+ { NULL, "interleave", "pp_pme", "cartesian", NULL };
+ const char *dddlb_opt[] =
+ { NULL, "auto", "no", "yes", NULL };
+ real rdd=0.0,rconstr=0.0,dlb_scale=0.8,pforce=-1;
+ char *ddcsx=NULL,*ddcsy=NULL,*ddcsz=NULL;
+ real cpt_period=15.0,max_hours=-1;
+ gmx_bool bAppendFiles=TRUE;
+ gmx_bool bKeepAndNumCPT=FALSE;
+ gmx_bool bResetCountersHalfWay=FALSE;
+ output_env_t oenv=NULL;
+ const char *deviceOptions = "";
+
+ t_pargs pa[] = {
+
+ { "-pd", FALSE, etBOOL,{&bPartDec},
+ "Use particle decompostion" },
+ { "-dd", FALSE, etRVEC,{&realddxyz},
+ "Domain decomposition grid, 0 is optimize" },
- #ifndef GMX_THREADS
++#ifdef GMX_THREAD_MPI
+ { "-nt", FALSE, etINT, {&nthreads},
+ "Number of threads to start (0 is guess)" },
+#endif
+ { "-npme", FALSE, etINT, {&npme},
+ "Number of separate nodes to be used for PME, -1 is guess" },
+ { "-ddorder", FALSE, etENUM, {ddno_opt},
+ "DD node order" },
+ { "-ddcheck", FALSE, etBOOL, {&bDDBondCheck},
+ "Check for all bonded interactions with DD" },
+ { "-ddbondcomm", FALSE, etBOOL, {&bDDBondComm},
+ "HIDDENUse special bonded atom communication when [TT]-rdd[tt] > cut-off" },
+ { "-rdd", FALSE, etREAL, {&rdd},
+ "The maximum distance for bonded interactions with DD (nm), 0 is determine from initial coordinates" },
+ { "-rcon", FALSE, etREAL, {&rconstr},
+ "Maximum distance for P-LINCS (nm), 0 is estimate" },
+ { "-dlb", FALSE, etENUM, {dddlb_opt},
+ "Dynamic load balancing (with DD)" },
+ { "-dds", FALSE, etREAL, {&dlb_scale},
+ "Minimum allowed dlb scaling of the DD cell size" },
+ { "-ddcsx", FALSE, etSTR, {&ddcsx},
+ "HIDDENThe DD cell sizes in x" },
+ { "-ddcsy", FALSE, etSTR, {&ddcsy},
+ "HIDDENThe DD cell sizes in y" },
+ { "-ddcsz", FALSE, etSTR, {&ddcsz},
+ "HIDDENThe DD cell sizes in z" },
+ { "-gcom", FALSE, etINT,{&nstglobalcomm},
+ "Global communication frequency" },
+ { "-v", FALSE, etBOOL,{&bVerbose},
+ "Be loud and noisy" },
+ { "-compact", FALSE, etBOOL,{&bCompact},
+ "Write a compact log file" },
+ { "-seppot", FALSE, etBOOL, {&bSepPot},
+ "Write separate V and dVdl terms for each interaction type and node to the log file(s)" },
+ { "-pforce", FALSE, etREAL, {&pforce},
+ "Print all forces larger than this (kJ/mol nm)" },
+ { "-reprod", FALSE, etBOOL,{&bReproducible},
+ "Try to avoid optimizations that affect binary reproducibility" },
+ { "-cpt", FALSE, etREAL, {&cpt_period},
+ "Checkpoint interval (minutes)" },
+ { "-cpnum", FALSE, etBOOL, {&bKeepAndNumCPT},
+ "Keep and number checkpoint files" },
+ { "-append", FALSE, etBOOL, {&bAppendFiles},
+ "Append to previous output files when continuing from checkpoint instead of adding the simulation part number to all file names" },
+ { "-maxh", FALSE, etREAL, {&max_hours},
+ "Terminate after 0.99 times this time (hours)" },
+ { "-multi", FALSE, etINT,{&nmultisim},
+ "Do multiple simulations in parallel" },
+ { "-replex", FALSE, etINT, {&repl_ex_nst},
+ "Attempt replica exchange every # steps" },
+ { "-reseed", FALSE, etINT, {&repl_ex_seed},
+ "Seed for replica exchange, -1 is generate a seed" },
+ { "-rerunvsite", FALSE, etBOOL, {&bRerunVSite},
+ "HIDDENRecalculate virtual site coordinates with [TT]-rerun[tt]" },
+ { "-ionize", FALSE, etBOOL,{&bIonize},
+ "Do a simulation including the effect of an X-Ray bombardment on your system" },
+ { "-confout", FALSE, etBOOL, {&bConfout},
+ "HIDDENWrite the last configuration with [TT]-c[tt] and force checkpointing at the last step" },
+ { "-stepout", FALSE, etINT, {&nstepout},
+ "HIDDENFrequency of writing the remaining runtime" },
+ { "-resetstep", FALSE, etINT, {&resetstep},
+ "HIDDENReset cycle counters after these many time steps" },
+ { "-resethway", FALSE, etBOOL, {&bResetCountersHalfWay},
+ "HIDDENReset the cycle counters after half the number of steps or halfway [TT]-maxh[tt]" }
+#ifdef GMX_OPENMM
+ ,
+ { "-device", FALSE, etSTR, {&deviceOptions},
+ "Device option string" }
+#endif
+ };
+ gmx_edsam_t ed;
+ unsigned long Flags, PCA_Flags;
+ ivec ddxyz;
+ int dd_node_order;
+ gmx_bool bAddPart;
+ FILE *fplog,*fptest;
+ int sim_part,sim_part_fn;
+ const char *part_suffix=".part";
+ char suffix[STRLEN];
+ int rc;
+ char **multidir=NULL;
+
+
+ cr = init_par(&argc,&argv);
+
+ if (MASTER(cr))
+ CopyRight(stderr, argv[0]);
+
+ PCA_Flags = (PCA_KEEP_ARGS | PCA_NOEXIT_ON_ARGS | PCA_CAN_SET_DEFFNM
+ | (MASTER(cr) ? 0 : PCA_QUIET));
+
+
+ /* Comment this in to do fexist calls only on master
+ * works not with rerun or tables at the moment
+ * also comment out the version of init_forcerec in md.c
+ * with NULL instead of opt2fn
+ */
+ /*
+ if (!MASTER(cr))
+ {
+ PCA_Flags |= PCA_NOT_READ_NODE;
+ }
+ */
+
+ parse_common_args(&argc,argv,PCA_Flags, NFILE,fnm,asize(pa),pa,
+ asize(desc),desc,0,NULL, &oenv);
+
+
+
+ /* we set these early because they might be used in init_multisystem()
+ Note that there is the potential for npme>nnodes until the number of
+ threads is set later on, if there's thread parallelization. That shouldn't
+ lead to problems. */
+ dd_node_order = nenum(ddno_opt);
+ cr->npmenodes = npme;
+
- #ifndef GMX_THREADS
++#ifndef GMX_THREAD_MPI
+ nthreads=1;
+#endif
+
+ /* now check the -multi and -multidir option */
+ if (opt2bSet("-multidir", NFILE, fnm))
+ {
+ int i;
+ if (nmultisim > 0)
+ {
+ gmx_fatal(FARGS, "mdrun -multi and -multidir options are mutually exclusive.");
+ }
+ nmultisim = opt2fns(&multidir, "-multidir", NFILE, fnm);
+ }
+
+
+ if (repl_ex_nst != 0 && nmultisim < 2)
+ gmx_fatal(FARGS,"Need at least two replicas for replica exchange (option -multi)");
+
+ if (nmultisim > 1) {
++#ifndef GMX_THREAD_MPI
+ gmx_bool bParFn = (multidir == NULL);
+ init_multisystem(cr, nmultisim, multidir, NFILE, fnm, bParFn);
+#else
+ gmx_fatal(FARGS,"mdrun -multi is not supported with the thread library.Please compile GROMACS with MPI support");
+#endif
+ }
+
+ bAddPart = !bAppendFiles;
+
+ /* Check if there is ANY checkpoint file available */
+ sim_part = 1;
+ sim_part_fn = sim_part;
+ if (opt2bSet("-cpi",NFILE,fnm))
+ {
+ if (bSepPot && bAppendFiles)
+ {
+ gmx_fatal(FARGS,"Output file appending is not supported with -seppot");
+ }
+
+ bAppendFiles =
+ read_checkpoint_simulation_part(opt2fn_master("-cpi", NFILE,
+ fnm,cr),
+ &sim_part_fn,NULL,cr,
+ bAppendFiles,NFILE,fnm,
+ part_suffix,&bAddPart);
+ if (sim_part_fn==0 && MASTER(cr))
+ {
+ fprintf(stdout,"No previous checkpoint file present, assuming this is a new run.\n");
+ }
+ else
+ {
+ sim_part = sim_part_fn + 1;
+ }
+
+ if (MULTISIM(cr) && MASTER(cr))
+ {
+ check_multi_int(stdout,cr->ms,sim_part,"simulation part");
+ }
+ }
+ else
+ {
+ bAppendFiles = FALSE;
+ }
+
+ if (!bAppendFiles)
+ {
+ sim_part_fn = sim_part;
+ }
+
+ if (bAddPart)
+ {
+ /* Rename all output files (except checkpoint files) */
+ /* create new part name first (zero-filled) */
+ sprintf(suffix,"%s%04d",part_suffix,sim_part_fn);
+
+ add_suffix_to_output_names(fnm,NFILE,suffix);
+ if (MASTER(cr))
+ {
+ fprintf(stdout,"Checkpoint file is from part %d, new output files will be suffixed '%s'.\n",sim_part-1,suffix);
+ }
+ }
+
+ Flags = opt2bSet("-rerun",NFILE,fnm) ? MD_RERUN : 0;
+ Flags = Flags | (bSepPot ? MD_SEPPOT : 0);
+ Flags = Flags | (bIonize ? MD_IONIZE : 0);
+ Flags = Flags | (bPartDec ? MD_PARTDEC : 0);
+ Flags = Flags | (bDDBondCheck ? MD_DDBONDCHECK : 0);
+ Flags = Flags | (bDDBondComm ? MD_DDBONDCOMM : 0);
+ Flags = Flags | (bConfout ? MD_CONFOUT : 0);
+ Flags = Flags | (bRerunVSite ? MD_RERUN_VSITE : 0);
+ Flags = Flags | (bReproducible ? MD_REPRODUCIBLE : 0);
+ Flags = Flags | (bAppendFiles ? MD_APPENDFILES : 0);
+ Flags = Flags | (bKeepAndNumCPT ? MD_KEEPANDNUMCPT : 0);
+ Flags = Flags | (sim_part>1 ? MD_STARTFROMCPT : 0);
+ Flags = Flags | (bResetCountersHalfWay ? MD_RESETCOUNTERSHALFWAY : 0);
+
+
+ /* We postpone opening the log file if we are appending, so we can
+ first truncate the old log file and append to the correct position
+ there instead. */
+ if ((MASTER(cr) || bSepPot) && !bAppendFiles)
+ {
+ gmx_log_open(ftp2fn(efLOG,NFILE,fnm),cr,!bSepPot,Flags,&fplog);
+ CopyRight(fplog,argv[0]);
+ please_cite(fplog,"Hess2008b");
+ please_cite(fplog,"Spoel2005a");
+ please_cite(fplog,"Lindahl2001a");
+ please_cite(fplog,"Berendsen95a");
+ }
+ else if (!MASTER(cr) && bSepPot)
+ {
+ gmx_log_open(ftp2fn(efLOG,NFILE,fnm),cr,!bSepPot,Flags,&fplog);
+ }
+ else
+ {
+ fplog = NULL;
+ }
+
+ ddxyz[XX] = (int)(realddxyz[XX] + 0.5);
+ ddxyz[YY] = (int)(realddxyz[YY] + 0.5);
+ ddxyz[ZZ] = (int)(realddxyz[ZZ] + 0.5);
+
+ rc = mdrunner(nthreads, fplog,cr,NFILE,fnm,oenv,bVerbose,bCompact,
+ nstglobalcomm, ddxyz,dd_node_order,rdd,rconstr,
+ dddlb_opt[0],dlb_scale,ddcsx,ddcsy,ddcsz,
+ nstepout,resetstep,nmultisim,repl_ex_nst,repl_ex_seed,
+ pforce, cpt_period,max_hours,deviceOptions,Flags);
+
+ if (gmx_parallel_env_initialized())
+ gmx_finalize();
+
+ if (MULTIMASTER(cr)) {
+ thanx(stderr);
+ }
+
+ /* Log file has to be closed in mdrunner if we are appending to it
+ (fplog not set here) */
+ if (MASTER(cr) && !bAppendFiles)
+ {
+ gmx_log_close(fplog);
+ }
+
+ return rc;
+}
+
--- /dev/null
- #ifdef GMX_THREADS
+/* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
+ *
+ *
+ * This source code is part of
+ *
+ * G R O M A C S
+ *
+ * GROningen MAchine for Chemical Simulations
+ *
+ * VERSION 3.2.0
+ * Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
+ * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
+ * Copyright (c) 2001-2004, The GROMACS development team,
+ * check out http://www.gromacs.org for more information.
+
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version 2
+ * of the License, or (at your option) any later version.
+ *
+ * If you want to redistribute modifications, please consider that
+ * scientific software is very special. Version control is crucial -
+ * bugs must be traceable. We will be happy to consider code for
+ * inclusion in the official distribution, but derived work must not
+ * be called official GROMACS. Details are found in the README & COPYING
+ * files - if they are missing, get the official version at www.gromacs.org.
+ *
+ * To help us fund GROMACS development, we humbly ask that you cite
+ * the papers on the package - you can find them in the top README file.
+ *
+ * For more info, check our website at http://www.gromacs.org
+ *
+ * And Hey:
+ * Gallium Rubidium Oxygen Manganese Argon Carbon Silicon
+ */
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+#ifdef __linux
+#define _GNU_SOURCE
+#include <sched.h>
+#include <sys/syscall.h>
+#endif
+#include <signal.h>
+#include <stdlib.h>
+
+#if ((defined WIN32 || defined _WIN32 || defined WIN64 || defined _WIN64) && !defined __CYGWIN__ && !defined __CYGWIN32__)
+/* _isnan() */
+#include <float.h>
+#endif
+
+#include "typedefs.h"
+#include "smalloc.h"
+#include "sysstuff.h"
+#include "statutil.h"
+#include "mdrun.h"
+#include "network.h"
+#include "pull.h"
+#include "pull_rotation.h"
+#include "names.h"
+#include "disre.h"
+#include "orires.h"
+#include "dihre.h"
+#include "pppm.h"
+#include "pme.h"
+#include "mdatoms.h"
+#include "repl_ex.h"
+#include "qmmm.h"
+#include "domdec.h"
+#include "partdec.h"
+#include "coulomb.h"
+#include "constr.h"
+#include "mvdata.h"
+#include "checkpoint.h"
+#include "mtop_util.h"
+#include "sighandler.h"
+#include "tpxio.h"
+#include "txtdump.h"
+#include "pull_rotation.h"
+#include "membed.h"
+#include "macros.h"
+
+#ifdef GMX_LIB_MPI
+#include <mpi.h>
+#endif
- #ifdef GMX_THREADS
++#ifdef GMX_THREAD_MPI
+#include "tmpi.h"
+#endif
+
+#ifdef GMX_FAHCORE
+#include "corewrap.h"
+#endif
+
+#ifdef GMX_OPENMM
+#include "md_openmm.h"
+#endif
+
+#ifdef GMX_OPENMP
+#include <omp.h>
+#endif
+
+
+typedef struct {
+ gmx_integrator_t *func;
+} gmx_intp_t;
+
+/* The array should match the eI array in include/types/enums.h */
+#ifdef GMX_OPENMM /* FIXME do_md_openmm needs fixing */
+const gmx_intp_t integrator[eiNR] = { {do_md_openmm}, {do_md_openmm}, {do_md_openmm}, {do_md_openmm}, {do_md_openmm}, {do_md_openmm}, {do_md_openmm}, {do_md_openmm}, {do_md_openmm}, {do_md_openmm}, {do_md_openmm},{do_md_openmm}};
+#else
+const gmx_intp_t integrator[eiNR] = { {do_md}, {do_steep}, {do_cg}, {do_md}, {do_md}, {do_nm}, {do_lbfgs}, {do_tpi}, {do_tpi}, {do_md}, {do_md},{do_md}};
+#endif
+
+gmx_large_int_t deform_init_init_step_tpx;
+matrix deform_init_box_tpx;
- #ifdef GMX_THREADS
++#ifdef GMX_THREAD_MPI
+tMPI_Thread_mutex_t deform_init_box_mutex=TMPI_THREAD_MUTEX_INITIALIZER;
+#endif
+
+
- #ifdef GMX_THREADS
++#ifdef GMX_THREAD_MPI
+struct mdrunner_arglist
+{
+ FILE *fplog;
+ t_commrec *cr;
+ int nfile;
+ const t_filenm *fnm;
+ 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;
+ int nstepout;
+ int resetstep;
+ int nmultisim;
+ int repl_ex_nst;
+ int repl_ex_seed;
+ real pforce;
+ real cpt_period;
+ real max_hours;
+ const char *deviceOptions;
+ unsigned long Flags;
+ int ret; /* return value */
+};
+
+
+/* The function used for spawning threads. Extracts the mdrunner()
+ arguments from its one argument and calls mdrunner(), after making
+ a commrec. */
+static void mdrunner_start_fn(void *arg)
+{
+ struct mdrunner_arglist *mda=(struct mdrunner_arglist*)arg;
+ struct mdrunner_arglist mc=*mda; /* copy the arg list to make sure
+ that it's thread-local. This doesn't
+ copy pointed-to items, of course,
+ but those are all const. */
+ t_commrec *cr; /* we need a local version of this */
+ FILE *fplog=NULL;
+ t_filenm *fnm;
+
+ fnm = dup_tfn(mc.nfile, mc.fnm);
+
+ cr = init_par_threads(mc.cr);
+
+ if (MASTER(cr))
+ {
+ fplog=mc.fplog;
+ }
+
+ mda->ret=mdrunner(cr->nnodes, fplog, cr, mc.nfile, fnm, mc.oenv,
+ mc.bVerbose, mc.bCompact, mc.nstglobalcomm,
+ mc.ddxyz, mc.dd_node_order, mc.rdd,
+ mc.rconstr, mc.dddlb_opt, mc.dlb_scale,
+ mc.ddcsx, mc.ddcsy, mc.ddcsz, mc.nstepout, mc.resetstep,
+ mc.nmultisim, mc.repl_ex_nst, mc.repl_ex_seed, mc.pforce,
+ mc.cpt_period, mc.max_hours, mc.deviceOptions, mc.Flags);
+}
+
+/* called by mdrunner() to start a specific number of threads (including
+ the main thread) for thread-parallel runs. This in turn calls mdrunner()
+ for each thread.
+ All options besides nthreads are the same as for mdrunner(). */
+static t_commrec *mdrunner_start_threads(int nthreads,
+ 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,
+ int nstepout,int resetstep,int nmultisim,int repl_ex_nst,
+ int repl_ex_seed, real pforce,real cpt_period, real max_hours,
+ const char *deviceOptions, unsigned long Flags)
+{
+ int ret;
+ struct mdrunner_arglist *mda;
+ t_commrec *crn; /* the new commrec */
+ t_filenm *fnmn;
+
+ /* first check whether we even need to start tMPI */
+ if (nthreads<2)
+ return cr;
+
+ /* a few small, one-time, almost unavoidable memory leaks: */
+ snew(mda,1);
+ fnmn=dup_tfn(nfile, fnm);
+
+ /* fill the data structure to pass as void pointer to thread start fn */
+ mda->fplog=fplog;
+ mda->cr=cr;
+ mda->nfile=nfile;
+ mda->fnm=fnmn;
+ mda->oenv=oenv;
+ mda->bVerbose=bVerbose;
+ mda->bCompact=bCompact;
+ mda->nstglobalcomm=nstglobalcomm;
+ mda->ddxyz[XX]=ddxyz[XX];
+ mda->ddxyz[YY]=ddxyz[YY];
+ mda->ddxyz[ZZ]=ddxyz[ZZ];
+ mda->dd_node_order=dd_node_order;
+ mda->rdd=rdd;
+ mda->rconstr=rconstr;
+ mda->dddlb_opt=dddlb_opt;
+ mda->dlb_scale=dlb_scale;
+ mda->ddcsx=ddcsx;
+ mda->ddcsy=ddcsy;
+ mda->ddcsz=ddcsz;
+ mda->nstepout=nstepout;
+ mda->resetstep=resetstep;
+ mda->nmultisim=nmultisim;
+ mda->repl_ex_nst=repl_ex_nst;
+ mda->repl_ex_seed=repl_ex_seed;
+ mda->pforce=pforce;
+ mda->cpt_period=cpt_period;
+ mda->max_hours=max_hours;
+ mda->deviceOptions=deviceOptions;
+ mda->Flags=Flags;
+
+ fprintf(stderr, "Starting %d threads\n",nthreads);
+ fflush(stderr);
+ /* now spawn new threads that start mdrunner_start_fn(), while
+ the main thread returns */
+ ret=tMPI_Init_fn(TRUE, nthreads, mdrunner_start_fn, (void*)(mda) );
+ if (ret!=TMPI_SUCCESS)
+ return NULL;
+
+ /* make a new comm_rec to reflect the new situation */
+ crn=init_par_threads(cr);
+ return crn;
+}
+
+
+/* Get the number of threads to use for thread-MPI based on how many
+ * were requested, which algorithms we're using,
+ * and how many particles there are.
+ */
+static int get_nthreads_mpi(int nthreads_requested, t_inputrec *inputrec,
+ gmx_mtop_t *mtop)
+{
+ int nthreads,nthreads_new;
+ int min_atoms_per_thread;
+ char *env;
+
+ nthreads = nthreads_requested;
+
+ /* determine # of hardware threads. */
+ if (nthreads_requested < 1)
+ {
+ if ((env = getenv("GMX_MAX_THREADS")) != NULL)
+ {
+ nthreads = 0;
+ sscanf(env,"%d",&nthreads);
+ if (nthreads < 1)
+ {
+ gmx_fatal(FARGS,"GMX_MAX_THREADS (%d) should be larger than 0",
+ nthreads);
+ }
+ }
+ else
+ {
+ nthreads = tMPI_Thread_get_hw_number();
+ }
+ }
+
+ if (inputrec->eI == eiNM || EI_TPI(inputrec->eI))
+ {
+ /* Steps are divided over the nodes iso splitting the atoms */
+ min_atoms_per_thread = 0;
+ }
+ else
+ {
+ min_atoms_per_thread = MIN_ATOMS_PER_THREAD;
+ }
+
+ /* Check if an algorithm does not support parallel simulation. */
+ if (nthreads != 1 &&
+ ( inputrec->eI == eiLBFGS ||
+ inputrec->coulombtype == eelEWALD ) )
+ {
+ fprintf(stderr,"\nThe integration or electrostatics algorithm doesn't support parallel runs. Not starting any threads.\n");
+ nthreads = 1;
+ }
+ else if (nthreads_requested < 1 &&
+ mtop->natoms/nthreads < min_atoms_per_thread)
+ {
+ /* the thread number was chosen automatically, but there are too many
+ threads (too few atoms per thread) */
+ nthreads_new = max(1,mtop->natoms/min_atoms_per_thread);
+
+ if (nthreads_new > 8 || (nthreads == 8 && nthreads_new > 4))
+ {
+ /* Use only multiples of 4 above 8 threads
+ * or with an 8-core processor
+ * (to avoid 6 threads on 8 core processors with 4 real cores).
+ */
+ nthreads_new = (nthreads_new/4)*4;
+ }
+ else if (nthreads_new > 4)
+ {
+ /* Avoid 5 or 7 threads */
+ nthreads_new = (nthreads_new/2)*2;
+ }
+
+ nthreads = nthreads_new;
+
+ fprintf(stderr,"\n");
+ fprintf(stderr,"NOTE: Parallelization is limited by the small number of atoms,\n");
+ fprintf(stderr," only starting %d threads.\n",nthreads);
+ fprintf(stderr," You can use the -nt option to optimize the number of threads.\n\n");
+ }
+ return nthreads;
+}
+#endif
+
+
+int mdrunner(int nthreads_requested, 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,
+ int nstepout,int resetstep,int nmultisim,int repl_ex_nst,
+ int repl_ex_seed, real pforce,real cpt_period,real max_hours,
+ const char *deviceOptions, unsigned long Flags)
+{
+ double nodetime=0,realtime;
+ t_inputrec *inputrec;
+ t_state *state=NULL;
+ matrix box;
+ gmx_ddbox_t ddbox={0};
+ int npme_major,npme_minor;
+ real tmpr1,tmpr2;
+ t_nrnb *nrnb;
+ gmx_mtop_t *mtop=NULL;
+ t_mdatoms *mdatoms=NULL;
+ t_forcerec *fr=NULL;
+ t_fcdata *fcd=NULL;
+ real ewaldcoeff=0;
+ gmx_pme_t *pmedata=NULL;
+ gmx_vsite_t *vsite=NULL;
+ gmx_constr_t constr;
+ int i,m,nChargePerturbed=-1,status,nalloc;
+ char *gro;
+ gmx_wallcycle_t wcycle;
+ gmx_bool bReadRNG,bReadEkin;
+ int list;
+ gmx_runtime_t runtime;
+ int rc;
+ gmx_large_int_t reset_counters;
+ gmx_edsam_t ed=NULL;
+ t_commrec *cr_old=cr;
+ int nthreads_mpi=1;
+ int nthreads_pme=1;
+ gmx_membed_t *membed=NULL;
+
+ /* 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 (bVerbose && SIMMASTER(cr))
+ {
+ fprintf(stderr,"Getting Loaded...\n");
+ }
+
+ if (Flags & MD_APPENDFILES)
+ {
+ fplog = NULL;
+ }
+
+ snew(state,1);
+ if (MASTER(cr))
+ {
+ /* Read (nearly) all data required for the simulation */
+ read_tpx_state(ftp2fn(efTPX,nfile,fnm),inputrec,state,NULL,mtop);
+
+ /* NOW the threads will be started: */
- #ifdef GMX_THREADS
++#ifdef GMX_THREAD_MPI
+ nthreads_mpi = get_nthreads_mpi(nthreads_requested, inputrec, mtop);
+
+ if (nthreads_mpi > 1)
+ {
+ /* now start the threads. */
+ cr=mdrunner_start_threads(nthreads_mpi, fplog, cr_old, nfile, fnm,
+ oenv, bVerbose, bCompact, nstglobalcomm,
+ ddxyz, dd_node_order, rdd, rconstr,
+ dddlb_opt, dlb_scale, ddcsx, ddcsy, ddcsz,
+ nstepout, resetstep, nmultisim,
+ repl_ex_nst, repl_ex_seed, pforce,
+ cpt_period, max_hours, deviceOptions,
+ 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))
+ {
+ fprintf(stderr,"Entering membed code");
+ snew(membed,1);
+ init_membed(fplog,membed,nfile,fnm,mtop,inputrec,state,cr,&cpt_period);
+ }
+
+ if (PAR(cr))
+ {
+ /* now broadcast everything to the non-master nodes/threads: */
+ init_parallel(fplog, cr, inputrec, mtop);
+ }
+ if (fplog != NULL)
+ {
+ pr_inputrec(fplog,0,"Input Parameters",inputrec,FALSE);
+ }
+
+ /* now make sure the state is initialized and propagated */
+ set_state_entries(state,inputrec,cr->nnodes);
+
+ /* 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 mdrun was compiled without threads or MPI enabled"
+#else
- #ifdef GMX_THREADS
++#ifdef GMX_THREAD_MPI
+ "but the number of threads (option -nt) is 1"
+#else
+ "but mdrun was not started through mpirun/mpiexec or only one process was requested through mpirun/mpiexec"
+#endif
+#endif
+ );
+ }
+
+ if ((Flags & MD_RERUN) &&
+ (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,mtop,TRUE,cr,fplog))
+ {
+ /* All-vs-all loops do not work with domain decomposition */
+ Flags |= MD_PARTDEC;
+ }
+
+ if (!EEL_PME(inputrec->coulombtype) || (Flags & MD_PARTDEC))
+ {
+ cr->npmenodes = 0;
+ }
+
+#ifdef GMX_FAHCORE
+ 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,Flags & MD_PARTDEC,fcd,state);
+
+ if (gmx_mtop_ftype_count(mtop,F_ORIRES) > 0)
+ {
+ if (PAR(cr) && !(Flags & MD_PARTDEC))
+ {
+ gmx_fatal(FARGS,"Orientation restraints do not work (yet) with domain decomposition, use particle decomposition (mdrun option -pd)");
+ }
+ /* Orientation restraints */
+ if (MASTER(cr))
+ {
+ init_orires(fplog,mtop,state->x,inputrec,cr->ms,&(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.
+ */
- #ifdef GMX_THREADS
++#ifdef GMX_THREAD_MPI
+ tMPI_Thread_mutex_lock(&deform_init_box_mutex);
+#endif
+ deform_init_init_step_tpx = inputrec->init_step;
+ copy_mat(box,deform_init_box_tpx);
- #ifdef GMX_THREADS
++#ifdef GMX_THREAD_MPI
+ tMPI_Thread_mutex_unlock(&deform_init_box_mutex);
+#endif
+ }
+
+ 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,Flags & MD_PARTDEC,ddxyz,
+ inputrec,state,&bReadRNG,&bReadEkin,
+ (Flags & MD_APPENDFILES));
+
+ if (bReadRNG)
+ {
+ Flags |= MD_READ_RNG;
+ }
+ if (bReadEkin)
+ {
+ Flags |= MD_READ_EKIN;
+ }
+ }
+ }
+
+ if (((MASTER(cr) || (Flags & MD_SEPPOT)) && (Flags & MD_APPENDFILES))
- #ifdef GMX_THREADS
++#ifdef GMX_THREAD_MPI
+ /* With thread MPI only the master node/thread exists in mdrun.c,
+ * therefore non-master nodes need to open the "seppot" log file here.
+ */
+ || (!MASTER(cr) && (Flags & MD_SEPPOT))
+#endif
+ )
+ {
+ gmx_log_open(ftp2fn(efLOG,nfile,fnm),cr,!(Flags & MD_SEPPOT),
+ Flags,&fplog);
+ }
+
+ 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(nfile,fnm,Flags,cr);
+ }
+
+ if (bVerbose && SIMMASTER(cr))
+ {
+ fprintf(stderr,"Loaded with Money\n\n");
+ }
+
+ if (PAR(cr) && !((Flags & MD_PARTDEC) ||
+ 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 (!EI_TPI(inputrec->eI))
+ {
+ npme_major = cr->nnodes;
+ }
+
+ 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);
+ }
+
+ /* get number of OpenMP/PME threads
+ * env variable should be read only on one node to make sure it is identical everywhere */
+#ifdef GMX_OPENMP
+ if (EEL_PME(inputrec->coulombtype))
+ {
+ if (MASTER(cr))
+ {
+ char *ptr;
+ if ((ptr=getenv("GMX_PME_NTHREADS")) != NULL)
+ {
+ sscanf(ptr,"%d",&nthreads_pme);
+ }
+ if (fplog != NULL && nthreads_pme > 1)
+ {
+ fprintf(fplog,"Using %d threads for PME\n",nthreads_pme);
+ }
+ }
+ if (PAR(cr))
+ {
+ gmx_bcast_sim(sizeof(nthreads_pme),&nthreads_pme,cr);
+ }
+ }
+#endif
+
+ wcycle = wallcycle_init(fplog,resetstep,cr,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)
+ {
+ /* For domain decomposition we allocate dynamically
+ * in dd_partition_system.
+ */
+ if (DOMAINDECOMP(cr))
+ {
+ bcast_state_setup(cr,state);
+ }
+ else
+ {
+ if (PAR(cr))
+ {
+ bcast_state(cr,state,TRUE);
+ }
+ }
+
+ /* Dihedral Restraints */
+ if (gmx_mtop_ftype_count(mtop,F_DIHRES) > 0)
+ {
+ init_dihres(fplog,mtop,inputrec,fcd);
+ }
+
+ /* Initiate forcerecord */
+ fr = mk_forcerec();
+ init_forcerec(fplog,oenv,fr,fcd,inputrec,mtop,cr,box,FALSE,
+ opt2fn("-table",nfile,fnm),
+ opt2fn("-tabletf",nfile,fnm),
+ opt2fn("-tablep",nfile,fnm),
+ opt2fn("-tableb",nfile,fnm),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);
+ */
+ fr->bSepDVDL = ((Flags & MD_SEPPOT) == MD_SEPPOT);
+
+ /* Initialize QM-MM */
+ if(fr->bQMMM)
+ {
+ init_QMMMrec(cr,box,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);
+
+ 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(fplog,vsite,mtop,state->x);
+ }
+ }
+
+ /* Initiate PPPM if necessary */
+ if (fr->eeltype == eelPPPM)
+ {
+ if (mdatoms->nChargePerturbed)
+ {
+ gmx_fatal(FARGS,"Free energy with %s is not implemented",
+ eel_names[fr->eeltype]);
+ }
+ status = gmx_pppm_init(fplog,cr,oenv,FALSE,TRUE,box,
+ getenv("GMXGHAT"),inputrec, (Flags & MD_REPRODUCIBLE));
+ if (status != 0)
+ {
+ gmx_fatal(FARGS,"Error %d initializing PPPM",status);
+ }
+ }
+
+ if (EEL_PME(fr->eeltype))
+ {
+ ewaldcoeff = fr->ewaldcoeff;
+ pmedata = &fr->pmedata;
+ }
+ else
+ {
+ pmedata = NULL;
+ }
+ }
+ else
+ {
+ /* This is a PME only node */
+
+ /* We don't need the state */
+ done_state(state);
+
+ ewaldcoeff = calc_ewaldcoeff(inputrec->rcoulomb, inputrec->ewald_rtol);
+ snew(pmedata,1);
+ }
+
+ /* Initiate PME if necessary,
+ * either on all nodes or on dedicated PME nodes only. */
+ if (EEL_PME(inputrec->coulombtype))
+ {
+ if (mdatoms)
+ {
+ nChargePerturbed = mdatoms->nChargePerturbed;
+ }
+ if (cr->npmenodes > 0)
+ {
+ /* The PME only nodes need to know nChargePerturbed */
+ gmx_bcast_sim(sizeof(nChargePerturbed),&nChargePerturbed,cr);
+ }
+
+
+ /* 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 at
+ if only PME is using threads.
+ */
+
+#ifdef GMX_OPENMP
+#ifdef __linux
+#ifdef GMX_LIB_MPI
+ {
+ int core;
+ MPI_Comm comm_intra; /* intra communicator (but different to nc.comm_intra includes PME nodes) */
+ MPI_Comm_split(MPI_COMM_WORLD,gmx_hostname_num(),gmx_node_rank(),&comm_intra);
+ int local_omp_nthreads = (cr->duty & DUTY_PME) ? nthreads_pme : 1; /* threads on this node */
+ MPI_Scan(&local_omp_nthreads,&core, 1, MPI_INT, MPI_SUM, comm_intra);
+ core-=local_omp_nthreads; /* make exclusive scan */
+#pragma omp parallel firstprivate(core) num_threads(local_omp_nthreads)
+ {
+ cpu_set_t mask;
+ CPU_ZERO(&mask);
+ core+=omp_get_thread_num();
+ CPU_SET(core,&mask);
+ sched_setaffinity((pid_t) syscall (SYS_gettid),sizeof(cpu_set_t),&mask);
+ }
+ }
+#endif /*GMX_MPI*/
+#endif /*__linux*/
+#endif /*GMX_OPENMP*/
+
+ if (cr->duty & DUTY_PME)
+ {
+ status = gmx_pme_init(pmedata,cr,npme_major,npme_minor,inputrec,
+ mtop ? mtop->natoms : 0,nChargePerturbed,
+ (Flags & MD_REPRODUCIBLE),nthreads_pme);
+ if (status != 0)
+ {
+ gmx_fatal(FARGS,"Error %d initializing PME",status);
+ }
+ }
+ }
+
+
+ if (integrator[inputrec->eI].func == do_md
+#ifdef GMX_OPENMM
+ ||
+ integrator[inputrec->eI].func == do_md_openmm
+#endif
+ )
+ {
+ /* 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)
+ {
+ if (inputrec->ePull != epullNO)
+ {
+ /* Initialize pull code */
+ init_pull(fplog,inputrec,nfile,fnm,mtop,cr,oenv,
+ EI_DYNAMICS(inputrec->eI) && MASTER(cr),Flags);
+ }
+
+ if (inputrec->bRot)
+ {
+ /* Initialize enforced rotation code */
+ init_rot(fplog,inputrec,nfile,fnm,cr,state->x,state->box,mtop,oenv,
+ bVerbose,Flags);
+ }
+
+ constr = init_constraints(fplog,mtop,inputrec,ed,state,cr);
+
+ if (DOMAINDECOMP(cr))
+ {
+ dd_init_bondeds(fplog,cr->dd,mtop,vsite,constr,inputrec,
+ Flags & MD_DDBONDCHECK,fr->cginfo_mb);
+
+ set_dd_parameters(fplog,cr->dd,dlb_scale,inputrec,fr,&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_seed,
+ membed,
+ cpt_period,max_hours,
+ deviceOptions,
+ Flags,
+ &runtime);
+
+ if (inputrec->ePull != epullNO)
+ {
+ finish_pull(fplog,inputrec->pull);
+ }
+
+ if (inputrec->bRot)
+ {
+ finish_rot(fplog,inputrec->rot);
+ }
+
+ }
+ else
+ {
+ /* do PME only */
+ gmx_pmeonly(*pmedata,cr,nrnb,wcycle,ewaldcoeff,FALSE,inputrec);
+ }
+
+ if (EI_DYNAMICS(inputrec->eI) || EI_TPI(inputrec->eI))
+ {
+ /* Some timing stats */
+ if (SIMMASTER(cr))
+ {
+ if (runtime.proc == 0)
+ {
+ runtime.proc = runtime.real;
+ }
+ }
+ else
+ {
+ runtime.real = 0;
+ }
+ }
+
+ wallcycle_stop(wcycle,ewcRUN);
+
+ /* Finish up, write some stuff
+ * if rerunMD, don't write last frame again
+ */
+ finish_run(fplog,cr,ftp2fn(efSTO,nfile,fnm),
+ inputrec,nrnb,wcycle,&runtime,
+ EI_DYNAMICS(inputrec->eI) && !MULTISIM(cr));
+
+ if (opt2bSet("-membed",nfile,fnm))
+ {
+ sfree(membed);
+ }
+
+ /* Does what it says */
+ print_date_and_time(fplog,cr->nodeid,"Finished mdrun",&runtime);
+
+ /* 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();
+
++#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;
+}
}
- if (repl_ex_nst != 0 && nmultisim < 2)
- gmx_fatal(FARGS,"Need at least two replicas for replica exchange (option -multi)");
+ parse_common_args(&argc,argv,0, NFILE,fnm,asize(pa),pa,
+ asize(desc),desc,0,NULL, &oenv);
if (nmultisim > 1) {
- #ifndef GMX_THREADS
+ #ifndef GMX_THREAD_MPI
gmx_bool bParFn = (multidir == NULL);
init_multisystem(cr,nmultisim,multidir,NFILE,fnm,TRUE);
#else
biod.pnpi.spb.ru / alexxy / gromacs.git / commitdiff