diff --git a/patches/gromacs-2016-beta1.config b/patches/gromacs-2016-beta1.config
new file mode 100644
index 0000000000000000000000000000000000000000..03f65f9e10a72823dcc4648c695166cffa631af6
--- /dev/null
+++ b/patches/gromacs-2016-beta1.config
@@ -0,0 +1,24 @@
+
+
+function plumed_preliminary_test(){
+# check if the README contains the word GROMACS and if gromacs has been already configured
+ grep -q GROMACS README 1>/dev/null 2>/dev/null
+}
+
+function plumed_patch_info(){
+cat << EOF
+
+PLUMED can be incorporated into gromacs using the standard patching procedure.
+Patching must be done in the gromacs root directory _before_ the cmake command is invoked.
+
+To enable PLUMED in a gromacs simulation one should use
+mdrun with an extra -plumed flag. The flag can be used to
+specify the name of the PLUMED input file, e.g.:
+
+gmx mdrun -plumed plumed.dat
+
+For more information on gromacs you should visit http://www.gromacs.org
+
+EOF
+}
+
diff --git a/patches/gromacs-2016-beta1.diff/src/gromacs/CMakeLists.txt b/patches/gromacs-2016-beta1.diff/src/gromacs/CMakeLists.txt
new file mode 100644
index 0000000000000000000000000000000000000000..2f8a59296e79cbc4c9f1b47034389ac8c2730816
--- /dev/null
+++ b/patches/gromacs-2016-beta1.diff/src/gromacs/CMakeLists.txt
@@ -0,0 +1,261 @@
+#
+# This file is part of the GROMACS molecular simulation package.
+#
+# Copyright (c) 2010,2011,2012,2013,2014,2015,2016, by the GROMACS development team, led by
+# Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
+# and including many others, as listed in the AUTHORS file in the
+# top-level source directory and at http://www.gromacs.org.
+#
+# GROMACS is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Lesser General Public License
+# as published by the Free Software Foundation; either version 2.1
+# of the License, or (at your option) any later version.
+#
+# GROMACS is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Lesser General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public
+# License along with GROMACS; if not, see
+# http://www.gnu.org/licenses, or write to the Free Software Foundation,
+# Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+#
+# If you want to redistribute modifications to GROMACS, 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 http://www.gromacs.org.
+#
+# To help us fund GROMACS development, we humbly ask that you cite
+# the research papers on the package. Check out http://www.gromacs.org.
+
+include(${CMAKE_SOURCE_DIR}/Plumed.cmake)
+
+set(LIBGROMACS_SOURCES)
+
+set_property(GLOBAL PROPERTY GMX_LIBGROMACS_SOURCES)
+set_property(GLOBAL PROPERTY GMX_INSTALLED_HEADERS)
+
+function (_gmx_add_files_to_property PROPERTY)
+ foreach (_file ${ARGN})
+ if (IS_ABSOLUTE "${_file}")
+ set_property(GLOBAL APPEND PROPERTY ${PROPERTY} ${_file})
+ else()
+ set_property(GLOBAL APPEND PROPERTY ${PROPERTY}
+ ${CMAKE_CURRENT_LIST_DIR}/${_file})
+ endif()
+ endforeach()
+endfunction ()
+
+function (gmx_add_libgromacs_sources)
+ _gmx_add_files_to_property(GMX_LIBGROMACS_SOURCES ${ARGN})
+endfunction ()
+
+function (gmx_install_headers)
+ if (NOT GMX_BUILD_MDRUN_ONLY)
+ file(RELATIVE_PATH _dest ${PROJECT_SOURCE_DIR}/src ${CMAKE_CURRENT_LIST_DIR})
+ install(FILES ${ARGN}
+ DESTINATION "${INCL_INSTALL_DIR}/${_dest}"
+ COMPONENT development)
+ endif()
+ _gmx_add_files_to_property(GMX_INSTALLED_HEADERS ${ARGN})
+endfunction ()
+
+function (gmx_write_installed_header_list)
+ get_property(_list GLOBAL PROPERTY GMX_INSTALLED_HEADERS)
+ string(REPLACE ";" "\n" _list "${_list}")
+ # TODO: Make this only update the file timestamp if the contents actually change.
+ file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/installed-headers.txt "${_list}")
+endfunction()
+
+if(GMX_USE_TNG)
+ option(GMX_EXTERNAL_TNG "Use external TNG instead of compiling the version shipped with GROMACS."
+ OFF)
+ # Detect TNG if GMX_EXTERNAL_TNG is explicitly ON
+ if(GMX_EXTERNAL_TNG)
+ find_package(TNG_IO 1.6.0)
+ if(NOT TNG_IO_FOUND)
+ message(FATAL_ERROR
+ "TNG >= 1.6.0 not found. "
+ "You can set GMX_EXTERNAL_TNG=OFF to compile TNG.")
+ endif()
+ include_directories(SYSTEM ${TNG_IO_INCLUDE_DIRS})
+ endif()
+ if(NOT GMX_EXTERNAL_TNG)
+ include(${CMAKE_SOURCE_DIR}/src/external/tng_io/BuildTNG.cmake)
+ tng_get_source_list(TNG_SOURCES TNG_IO_DEFINITIONS)
+ list(APPEND LIBGROMACS_SOURCES ${TNG_SOURCES})
+ tng_set_source_properties(WITH_ZLIB ${HAVE_ZLIB})
+
+ if (HAVE_ZLIB)
+ list(APPEND GMX_EXTRA_LIBRARIES ${ZLIB_LIBRARIES})
+ include_directories(SYSTEM ${ZLIB_INCLUDE_DIRS})
+ endif()
+ endif()
+else()
+ # We still need to get tng/tng_io_fwd.h from somewhere!
+ include_directories(BEFORE ${CMAKE_SOURCE_DIR}/src/external/tng_io/include)
+endif()
+
+add_subdirectory(gmxlib)
+add_subdirectory(mdlib)
+add_subdirectory(listed-forces)
+add_subdirectory(commandline)
+add_subdirectory(domdec)
+add_subdirectory(ewald)
+add_subdirectory(fft)
+add_subdirectory(gpu_utils)
+add_subdirectory(hardware)
+add_subdirectory(linearalgebra)
+add_subdirectory(math)
+add_subdirectory(mdrunutility)
+add_subdirectory(mdtypes)
+add_subdirectory(onlinehelp)
+add_subdirectory(options)
+add_subdirectory(pbcutil)
+add_subdirectory(random)
+add_subdirectory(tables)
+add_subdirectory(timing)
+add_subdirectory(topology)
+add_subdirectory(trajectory)
+add_subdirectory(utility)
+add_subdirectory(fileio)
+add_subdirectory(swap)
+add_subdirectory(essentialdynamics)
+add_subdirectory(pulling)
+add_subdirectory(simd)
+add_subdirectory(imd)
+if (NOT GMX_BUILD_MDRUN_ONLY)
+ add_subdirectory(gmxana)
+ add_subdirectory(gmxpreprocess)
+ add_subdirectory(correlationfunctions)
+ add_subdirectory(statistics)
+ add_subdirectory(analysisdata)
+ add_subdirectory(selection)
+ add_subdirectory(trajectoryanalysis)
+ add_subdirectory(tools)
+endif()
+
+get_property(PROPERTY_SOURCES GLOBAL PROPERTY GMX_LIBGROMACS_SOURCES)
+list(APPEND LIBGROMACS_SOURCES ${GMXLIB_SOURCES} ${MDLIB_SOURCES} ${PROPERTY_SOURCES})
+
+# This would be the standard way to include thread_mpi, but
+# we want libgromacs to link the functions directly
+#if(GMX_THREAD_MPI)
+# add_subdirectory(thread_mpi)
+#endif()
+#target_link_libraries(gmx ${GMX_EXTRA_LIBRARIES} ${THREAD_MPI_LIB})
+
+tmpi_get_source_list(THREAD_MPI_SOURCES ${CMAKE_SOURCE_DIR}/src/external/thread_mpi/src)
+list(APPEND LIBGROMACS_SOURCES ${THREAD_MPI_SOURCES})
+
+configure_file(version.h.cmakein version.h)
+gmx_install_headers(
+ analysisdata.h
+ commandline.h
+ options.h
+ random.h
+ selection.h
+ trajectoryanalysis.h
+ utility.h
+ ${CMAKE_CURRENT_BINARY_DIR}/version.h
+ )
+
+# This code is here instead of utility/CMakeLists.txt, because CMake
+# custom commands and source file properties can only be set in the directory
+# that contains the target that uses them.
+# TODO: Generate a header instead that can be included from baseversion.c.
+# That probably simplifies things somewhat.
+set(GENERATED_VERSION_FILE utility/baseversion-gen.c)
+gmx_configure_version_file(
+ utility/baseversion-gen.c.cmakein ${GENERATED_VERSION_FILE}
+ REMOTE_HASH SOURCE_FILE)
+list(APPEND LIBGROMACS_SOURCES ${GENERATED_VERSION_FILE})
+
+if (GMX_USE_CUDA)
+ cuda_add_library(libgromacs ${LIBGROMACS_SOURCES})
+else()
+ add_library(libgromacs ${LIBGROMACS_SOURCES})
+endif()
+
+# Recent versions of gcc and clang give warnings on scanner.cpp, which
+# is a generated source file. These are awkward to suppress inline, so
+# we do it in the compilation command (after testing that the compiler
+# supports the suppressions).
+include(CheckCXXCompilerFlag)
+check_cxx_compiler_flag(-Wno-unused-parameter HAS_NO_UNUSED_PARAMETER)
+if (HAS_NO_UNUSED_PARAMETER)
+ set(_scanner_cpp_compiler_flags "${_scanner_cpp_compiler_flags} -Wno-unused-parameter")
+endif()
+check_cxx_compiler_flag(-Wno-deprecated-register HAS_NO_DEPRECATED_REGISTER)
+if (HAS_NO_DEPRECATED_REGISTER)
+ set(_scanner_cpp_compiler_flags "${_scanner_cpp_compiler_flags} -Wno-deprecated-register")
+else()
+ check_cxx_compiler_flag(-Wno-deprecated HAS_NO_DEPRECATED)
+ if (HAS_NO_DEPRECATED)
+ set(_scanner_cpp_compiler_flags "${_scanner_cpp_compiler_flags} -Wno-deprecated")
+ endif()
+endif()
+set_source_files_properties(selection/scanner.cpp PROPERTIES COMPILE_FLAGS "${_scanner_cpp_compiler_flags}")
+
+target_link_libraries(libgromacs ${PLUMED_LOAD})
+
+target_link_libraries(libgromacs
+ ${EXTRAE_LIBRARIES}
+ ${GMX_EXTRA_LIBRARIES}
+ ${TNG_IO_LIBRARIES}
+ ${FFT_LIBRARIES} ${LINEAR_ALGEBRA_LIBRARIES}
+ ${XML_LIBRARIES}
+ ${THREAD_LIB} ${GMX_SHARED_LINKER_FLAGS} ${OPENCL_LIBRARIES}
+ ${GMX_STDLIB_LIBRARIES})
+set_target_properties(libgromacs PROPERTIES
+ OUTPUT_NAME "gromacs${GMX_LIBS_SUFFIX}"
+ SOVERSION ${LIBRARY_SOVERSION_MAJOR}
+ VERSION ${LIBRARY_VERSION}
+ COMPILE_FLAGS "${OpenMP_C_FLAGS}")
+
+gmx_write_installed_header_list()
+
+# Only install the library in mdrun-only mode if it is actually necessary
+# for the binary
+if (NOT GMX_BUILD_MDRUN_ONLY OR BUILD_SHARED_LIBS)
+ install(TARGETS libgromacs
+ EXPORT libgromacs
+ LIBRARY DESTINATION ${LIB_INSTALL_DIR}
+ RUNTIME DESTINATION ${BIN_INSTALL_DIR}
+ ARCHIVE DESTINATION ${LIB_INSTALL_DIR}
+ COMPONENT libraries)
+endif()
+
+if (NOT GMX_BUILD_MDRUN_ONLY)
+ include(InstallLibInfo.cmake)
+endif()
+
+# Technically, the user could want to do this for an OpenCL build
+# using the CUDA runtime, but currently there's no reason to want to
+# do that.
+if (INSTALL_CUDART_LIB) #can be set manual by user
+ if (GMX_USE_CUDA)
+ foreach(CUDA_LIB ${CUDA_LIBRARIES})
+ string(REGEX MATCH "cudart" IS_CUDART ${CUDA_LIB})
+ if(IS_CUDART) #libcuda should not be installed
+ #install also name-links (linker uses those)
+ file(GLOB CUDA_LIBS ${CUDA_LIB}*)
+ install(FILES ${CUDA_LIBS} DESTINATION
+ ${LIB_INSTALL_DIR} COMPONENT libraries)
+ endif()
+ endforeach()
+ else()
+ message(WARNING "INSTALL_CUDART_LIB only makes sense when configuring for CUDA support")
+ endif()
+endif()
+
+if(GMX_USE_OPENCL)
+ set(OPENCL_KERNELS ${MDLIB_OPENCL_KERNELS})
+
+ install(FILES ${OPENCL_KERNELS} DESTINATION
+ ${OCL_INSTALL_DIR} COMPONENT libraries)
+endif()
diff --git a/patches/gromacs-2016-beta1.diff/src/gromacs/CMakeLists.txt.preplumed b/patches/gromacs-2016-beta1.diff/src/gromacs/CMakeLists.txt.preplumed
new file mode 100644
index 0000000000000000000000000000000000000000..edc051fe8328dcb0f2ff5604048c6e1a07bc66b1
--- /dev/null
+++ b/patches/gromacs-2016-beta1.diff/src/gromacs/CMakeLists.txt.preplumed
@@ -0,0 +1,257 @@
+#
+# This file is part of the GROMACS molecular simulation package.
+#
+# Copyright (c) 2010,2011,2012,2013,2014,2015,2016, by the GROMACS development team, led by
+# Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
+# and including many others, as listed in the AUTHORS file in the
+# top-level source directory and at http://www.gromacs.org.
+#
+# GROMACS is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Lesser General Public License
+# as published by the Free Software Foundation; either version 2.1
+# of the License, or (at your option) any later version.
+#
+# GROMACS is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+# Lesser General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public
+# License along with GROMACS; if not, see
+# http://www.gnu.org/licenses, or write to the Free Software Foundation,
+# Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+#
+# If you want to redistribute modifications to GROMACS, 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 http://www.gromacs.org.
+#
+# To help us fund GROMACS development, we humbly ask that you cite
+# the research papers on the package. Check out http://www.gromacs.org.
+
+set(LIBGROMACS_SOURCES)
+
+set_property(GLOBAL PROPERTY GMX_LIBGROMACS_SOURCES)
+set_property(GLOBAL PROPERTY GMX_INSTALLED_HEADERS)
+
+function (_gmx_add_files_to_property PROPERTY)
+ foreach (_file ${ARGN})
+ if (IS_ABSOLUTE "${_file}")
+ set_property(GLOBAL APPEND PROPERTY ${PROPERTY} ${_file})
+ else()
+ set_property(GLOBAL APPEND PROPERTY ${PROPERTY}
+ ${CMAKE_CURRENT_LIST_DIR}/${_file})
+ endif()
+ endforeach()
+endfunction ()
+
+function (gmx_add_libgromacs_sources)
+ _gmx_add_files_to_property(GMX_LIBGROMACS_SOURCES ${ARGN})
+endfunction ()
+
+function (gmx_install_headers)
+ if (NOT GMX_BUILD_MDRUN_ONLY)
+ file(RELATIVE_PATH _dest ${PROJECT_SOURCE_DIR}/src ${CMAKE_CURRENT_LIST_DIR})
+ install(FILES ${ARGN}
+ DESTINATION "${INCL_INSTALL_DIR}/${_dest}"
+ COMPONENT development)
+ endif()
+ _gmx_add_files_to_property(GMX_INSTALLED_HEADERS ${ARGN})
+endfunction ()
+
+function (gmx_write_installed_header_list)
+ get_property(_list GLOBAL PROPERTY GMX_INSTALLED_HEADERS)
+ string(REPLACE ";" "\n" _list "${_list}")
+ # TODO: Make this only update the file timestamp if the contents actually change.
+ file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/installed-headers.txt "${_list}")
+endfunction()
+
+if(GMX_USE_TNG)
+ option(GMX_EXTERNAL_TNG "Use external TNG instead of compiling the version shipped with GROMACS."
+ OFF)
+ # Detect TNG if GMX_EXTERNAL_TNG is explicitly ON
+ if(GMX_EXTERNAL_TNG)
+ find_package(TNG_IO 1.6.0)
+ if(NOT TNG_IO_FOUND)
+ message(FATAL_ERROR
+ "TNG >= 1.6.0 not found. "
+ "You can set GMX_EXTERNAL_TNG=OFF to compile TNG.")
+ endif()
+ include_directories(SYSTEM ${TNG_IO_INCLUDE_DIRS})
+ endif()
+ if(NOT GMX_EXTERNAL_TNG)
+ include(${CMAKE_SOURCE_DIR}/src/external/tng_io/BuildTNG.cmake)
+ tng_get_source_list(TNG_SOURCES TNG_IO_DEFINITIONS)
+ list(APPEND LIBGROMACS_SOURCES ${TNG_SOURCES})
+ tng_set_source_properties(WITH_ZLIB ${HAVE_ZLIB})
+
+ if (HAVE_ZLIB)
+ list(APPEND GMX_EXTRA_LIBRARIES ${ZLIB_LIBRARIES})
+ include_directories(SYSTEM ${ZLIB_INCLUDE_DIRS})
+ endif()
+ endif()
+else()
+ # We still need to get tng/tng_io_fwd.h from somewhere!
+ include_directories(BEFORE ${CMAKE_SOURCE_DIR}/src/external/tng_io/include)
+endif()
+
+add_subdirectory(gmxlib)
+add_subdirectory(mdlib)
+add_subdirectory(listed-forces)
+add_subdirectory(commandline)
+add_subdirectory(domdec)
+add_subdirectory(ewald)
+add_subdirectory(fft)
+add_subdirectory(gpu_utils)
+add_subdirectory(hardware)
+add_subdirectory(linearalgebra)
+add_subdirectory(math)
+add_subdirectory(mdrunutility)
+add_subdirectory(mdtypes)
+add_subdirectory(onlinehelp)
+add_subdirectory(options)
+add_subdirectory(pbcutil)
+add_subdirectory(random)
+add_subdirectory(tables)
+add_subdirectory(timing)
+add_subdirectory(topology)
+add_subdirectory(trajectory)
+add_subdirectory(utility)
+add_subdirectory(fileio)
+add_subdirectory(swap)
+add_subdirectory(essentialdynamics)
+add_subdirectory(pulling)
+add_subdirectory(simd)
+add_subdirectory(imd)
+if (NOT GMX_BUILD_MDRUN_ONLY)
+ add_subdirectory(gmxana)
+ add_subdirectory(gmxpreprocess)
+ add_subdirectory(correlationfunctions)
+ add_subdirectory(statistics)
+ add_subdirectory(analysisdata)
+ add_subdirectory(selection)
+ add_subdirectory(trajectoryanalysis)
+ add_subdirectory(tools)
+endif()
+
+get_property(PROPERTY_SOURCES GLOBAL PROPERTY GMX_LIBGROMACS_SOURCES)
+list(APPEND LIBGROMACS_SOURCES ${GMXLIB_SOURCES} ${MDLIB_SOURCES} ${PROPERTY_SOURCES})
+
+# This would be the standard way to include thread_mpi, but
+# we want libgromacs to link the functions directly
+#if(GMX_THREAD_MPI)
+# add_subdirectory(thread_mpi)
+#endif()
+#target_link_libraries(gmx ${GMX_EXTRA_LIBRARIES} ${THREAD_MPI_LIB})
+
+tmpi_get_source_list(THREAD_MPI_SOURCES ${CMAKE_SOURCE_DIR}/src/external/thread_mpi/src)
+list(APPEND LIBGROMACS_SOURCES ${THREAD_MPI_SOURCES})
+
+configure_file(version.h.cmakein version.h)
+gmx_install_headers(
+ analysisdata.h
+ commandline.h
+ options.h
+ random.h
+ selection.h
+ trajectoryanalysis.h
+ utility.h
+ ${CMAKE_CURRENT_BINARY_DIR}/version.h
+ )
+
+# This code is here instead of utility/CMakeLists.txt, because CMake
+# custom commands and source file properties can only be set in the directory
+# that contains the target that uses them.
+# TODO: Generate a header instead that can be included from baseversion.c.
+# That probably simplifies things somewhat.
+set(GENERATED_VERSION_FILE utility/baseversion-gen.c)
+gmx_configure_version_file(
+ utility/baseversion-gen.c.cmakein ${GENERATED_VERSION_FILE}
+ REMOTE_HASH SOURCE_FILE)
+list(APPEND LIBGROMACS_SOURCES ${GENERATED_VERSION_FILE})
+
+if (GMX_USE_CUDA)
+ cuda_add_library(libgromacs ${LIBGROMACS_SOURCES})
+else()
+ add_library(libgromacs ${LIBGROMACS_SOURCES})
+endif()
+
+# Recent versions of gcc and clang give warnings on scanner.cpp, which
+# is a generated source file. These are awkward to suppress inline, so
+# we do it in the compilation command (after testing that the compiler
+# supports the suppressions).
+include(CheckCXXCompilerFlag)
+check_cxx_compiler_flag(-Wno-unused-parameter HAS_NO_UNUSED_PARAMETER)
+if (HAS_NO_UNUSED_PARAMETER)
+ set(_scanner_cpp_compiler_flags "${_scanner_cpp_compiler_flags} -Wno-unused-parameter")
+endif()
+check_cxx_compiler_flag(-Wno-deprecated-register HAS_NO_DEPRECATED_REGISTER)
+if (HAS_NO_DEPRECATED_REGISTER)
+ set(_scanner_cpp_compiler_flags "${_scanner_cpp_compiler_flags} -Wno-deprecated-register")
+else()
+ check_cxx_compiler_flag(-Wno-deprecated HAS_NO_DEPRECATED)
+ if (HAS_NO_DEPRECATED)
+ set(_scanner_cpp_compiler_flags "${_scanner_cpp_compiler_flags} -Wno-deprecated")
+ endif()
+endif()
+set_source_files_properties(selection/scanner.cpp PROPERTIES COMPILE_FLAGS "${_scanner_cpp_compiler_flags}")
+
+target_link_libraries(libgromacs
+ ${EXTRAE_LIBRARIES}
+ ${GMX_EXTRA_LIBRARIES}
+ ${TNG_IO_LIBRARIES}
+ ${FFT_LIBRARIES} ${LINEAR_ALGEBRA_LIBRARIES}
+ ${XML_LIBRARIES}
+ ${THREAD_LIB} ${GMX_SHARED_LINKER_FLAGS} ${OPENCL_LIBRARIES}
+ ${GMX_STDLIB_LIBRARIES})
+set_target_properties(libgromacs PROPERTIES
+ OUTPUT_NAME "gromacs${GMX_LIBS_SUFFIX}"
+ SOVERSION ${LIBRARY_SOVERSION_MAJOR}
+ VERSION ${LIBRARY_VERSION}
+ COMPILE_FLAGS "${OpenMP_C_FLAGS}")
+
+gmx_write_installed_header_list()
+
+# Only install the library in mdrun-only mode if it is actually necessary
+# for the binary
+if (NOT GMX_BUILD_MDRUN_ONLY OR BUILD_SHARED_LIBS)
+ install(TARGETS libgromacs
+ EXPORT libgromacs
+ LIBRARY DESTINATION ${LIB_INSTALL_DIR}
+ RUNTIME DESTINATION ${BIN_INSTALL_DIR}
+ ARCHIVE DESTINATION ${LIB_INSTALL_DIR}
+ COMPONENT libraries)
+endif()
+
+if (NOT GMX_BUILD_MDRUN_ONLY)
+ include(InstallLibInfo.cmake)
+endif()
+
+# Technically, the user could want to do this for an OpenCL build
+# using the CUDA runtime, but currently there's no reason to want to
+# do that.
+if (INSTALL_CUDART_LIB) #can be set manual by user
+ if (GMX_USE_CUDA)
+ foreach(CUDA_LIB ${CUDA_LIBRARIES})
+ string(REGEX MATCH "cudart" IS_CUDART ${CUDA_LIB})
+ if(IS_CUDART) #libcuda should not be installed
+ #install also name-links (linker uses those)
+ file(GLOB CUDA_LIBS ${CUDA_LIB}*)
+ install(FILES ${CUDA_LIBS} DESTINATION
+ ${LIB_INSTALL_DIR} COMPONENT libraries)
+ endif()
+ endforeach()
+ else()
+ message(WARNING "INSTALL_CUDART_LIB only makes sense when configuring for CUDA support")
+ endif()
+endif()
+
+if(GMX_USE_OPENCL)
+ set(OPENCL_KERNELS ${MDLIB_OPENCL_KERNELS})
+
+ install(FILES ${OPENCL_KERNELS} DESTINATION
+ ${OCL_INSTALL_DIR} COMPONENT libraries)
+endif()
diff --git a/patches/gromacs-2016-beta1.diff/src/gromacs/mdlib/force.cpp b/patches/gromacs-2016-beta1.diff/src/gromacs/mdlib/force.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..3bee45971a7e763451ebb77a43f40c0c2f2cb3eb
--- /dev/null
+++ b/patches/gromacs-2016-beta1.diff/src/gromacs/mdlib/force.cpp
@@ -0,0 +1,913 @@
+/*
+ * This file is part of the GROMACS molecular simulation package.
+ *
+ * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
+ * Copyright (c) 2001-2004, The GROMACS development team.
+ * Copyright (c) 2013,2014,2015,2016, by the GROMACS development team, led by
+ * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
+ * and including many others, as listed in the AUTHORS file in the
+ * top-level source directory and at http://www.gromacs.org.
+ *
+ * GROMACS is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public License
+ * as published by the Free Software Foundation; either version 2.1
+ * of the License, or (at your option) any later version.
+ *
+ * GROMACS is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with GROMACS; if not, see
+ * http://www.gnu.org/licenses, or write to the Free Software Foundation,
+ * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ * If you want to redistribute modifications to GROMACS, 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 http://www.gromacs.org.
+ *
+ * To help us fund GROMACS development, we humbly ask that you cite
+ * the research papers on the package. Check out http://www.gromacs.org.
+ */
+#include "gmxpre.h"
+
+#include "force.h"
+
+#include "config.h"
+
+#include <assert.h>
+#include <math.h>
+#include <string.h>
+
+#include "gromacs/domdec/domdec.h"
+#include "gromacs/domdec/domdec_struct.h"
+#include "gromacs/ewald/ewald.h"
+#include "gromacs/ewald/long-range-correction.h"
+#include "gromacs/ewald/pme.h"
+#include "gromacs/gmxlib/network.h"
+#include "gromacs/gmxlib/nrnb.h"
+#include "gromacs/gmxlib/nonbonded/nonbonded.h"
+#include "gromacs/listed-forces/listed-forces.h"
+#include "gromacs/math/vec.h"
+#include "gromacs/math/vecdump.h"
+#include "gromacs/mdlib/forcerec-threading.h"
+#include "gromacs/mdlib/genborn.h"
+#include "gromacs/mdlib/mdrun.h"
+#include "gromacs/mdlib/ns.h"
+#include "gromacs/mdlib/qmmm.h"
+#include "gromacs/mdtypes/commrec.h"
+#include "gromacs/mdtypes/inputrec.h"
+#include "gromacs/mdtypes/md_enums.h"
+#include "gromacs/pbcutil/ishift.h"
+#include "gromacs/pbcutil/mshift.h"
+#include "gromacs/pbcutil/pbc.h"
+#include "gromacs/timing/wallcycle.h"
+#include "gromacs/utility/cstringutil.h"
+#include "gromacs/utility/exceptions.h"
+#include "gromacs/utility/fatalerror.h"
+#include "gromacs/utility/smalloc.h"
+/* PLUMED */
+#include "../../../Plumed.h"
+int plumedswitch=0;
+plumed plumedmain;
+void(*plumedcmd)(plumed,const char*,const void*)=NULL;
+/* END PLUMED */
+
+
+void ns(FILE *fp,
+ t_forcerec *fr,
+ matrix box,
+ gmx_groups_t *groups,
+ gmx_localtop_t *top,
+ t_mdatoms *md,
+ t_commrec *cr,
+ t_nrnb *nrnb,
+ gmx_bool bFillGrid)
+{
+ int nsearch;
+
+
+ if (!fr->ns->nblist_initialized)
+ {
+ init_neighbor_list(fp, fr, md->homenr);
+ }
+
+ nsearch = search_neighbours(fp, fr, box, top, groups, cr, nrnb, md,
+ bFillGrid);
+ if (debug)
+ {
+ fprintf(debug, "nsearch = %d\n", nsearch);
+ }
+
+ /* Check whether we have to do dynamic load balancing */
+ /*if ((nsb->nstDlb > 0) && (mod(step,nsb->nstDlb) == 0))
+ count_nb(cr,nsb,&(top->blocks[ebCGS]),nns,fr->nlr,
+ &(top->idef),opts->ngener);
+ */
+ if (fr->ns->dump_nl > 0)
+ {
+ dump_nblist(fp, cr, fr, fr->ns->dump_nl);
+ }
+}
+
+static void reduce_thread_energies(tensor vir_q, tensor vir_lj,
+ real *Vcorr_q, real *Vcorr_lj,
+ real *dvdl_q, real *dvdl_lj,
+ int nthreads,
+ ewald_corr_thread_t *ewc_t)
+{
+ int t;
+
+ for (t = 1; t < nthreads; t++)
+ {
+ *Vcorr_q += ewc_t[t].Vcorr_q;
+ *Vcorr_lj += ewc_t[t].Vcorr_lj;
+ *dvdl_q += ewc_t[t].dvdl[efptCOUL];
+ *dvdl_lj += ewc_t[t].dvdl[efptVDW];
+ m_add(vir_q, ewc_t[t].vir_q, vir_q);
+ m_add(vir_lj, ewc_t[t].vir_lj, vir_lj);
+ }
+}
+
+void do_force_lowlevel(t_forcerec *fr, t_inputrec *ir,
+ t_idef *idef, t_commrec *cr,
+ t_nrnb *nrnb, gmx_wallcycle_t wcycle,
+ t_mdatoms *md,
+ rvec x[], history_t *hist,
+ rvec f[],
+ gmx_enerdata_t *enerd,
+ t_fcdata *fcd,
+ gmx_localtop_t *top,
+ gmx_genborn_t *born,
+ gmx_bool bBornRadii,
+ matrix box,
+ t_lambda *fepvals,
+ real *lambda,
+ t_graph *graph,
+ t_blocka *excl,
+ rvec mu_tot[],
+ int flags,
+ float *cycles_pme)
+{
+ int i, j;
+ int donb_flags;
+ gmx_bool bSB;
+ int pme_flags;
+ matrix boxs;
+ rvec box_size;
+ t_pbc pbc;
+ real dvdl_dum[efptNR], dvdl_nb[efptNR];
+
+#if GMX_MPI
+ double t0 = 0.0, t1, t2, t3; /* time measurement for coarse load balancing */
+#endif
+
+ set_pbc(&pbc, fr->ePBC, box);
+
+ /* reset free energy components */
+ for (i = 0; i < efptNR; i++)
+ {
+ dvdl_nb[i] = 0;
+ dvdl_dum[i] = 0;
+ }
+
+ /* Reset box */
+ for (i = 0; (i < DIM); i++)
+ {
+ box_size[i] = box[i][i];
+ }
+
+ /* do QMMM first if requested */
+ if (fr->bQMMM)
+ {
+ enerd->term[F_EQM] = calculate_QMMM(cr, x, f, fr);
+ }
+
+ /* Call the short range functions all in one go. */
+
+#if GMX_MPI
+ /*#define TAKETIME ((cr->npmenodes) && (fr->timesteps < 12))*/
+#define TAKETIME FALSE
+ if (TAKETIME)
+ {
+ MPI_Barrier(cr->mpi_comm_mygroup);
+ t0 = MPI_Wtime();
+ }
+#endif
+
+ if (ir->nwall)
+ {
+ /* foreign lambda component for walls */
+ real dvdl_walls = do_walls(ir, fr, box, md, x, f, lambda[efptVDW],
+ enerd->grpp.ener[egLJSR], nrnb);
+ enerd->dvdl_lin[efptVDW] += dvdl_walls;
+ }
+
+ /* If doing GB, reset dvda and calculate the Born radii */
+ if (ir->implicit_solvent)
+ {
+ wallcycle_sub_start(wcycle, ewcsNONBONDED);
+
+ for (i = 0; i < born->nr; i++)
+ {
+ fr->dvda[i] = 0;
+ }
+
+ if (bBornRadii)
+ {
+ calc_gb_rad(cr, fr, ir, top, x, fr->gblist, born, md, nrnb);
+ }
+
+ wallcycle_sub_stop(wcycle, ewcsNONBONDED);
+ }
+
+ where();
+ /* We only do non-bonded calculation with group scheme here, the verlet
+ * calls are done from do_force_cutsVERLET(). */
+ if (fr->cutoff_scheme == ecutsGROUP && (flags & GMX_FORCE_NONBONDED))
+ {
+ donb_flags = 0;
+ /* Add short-range interactions */
+ donb_flags |= GMX_NONBONDED_DO_SR;
+
+ /* Currently all group scheme kernels always calculate (shift-)forces */
+ if (flags & GMX_FORCE_FORCES)
+ {
+ donb_flags |= GMX_NONBONDED_DO_FORCE;
+ }
+ if (flags & GMX_FORCE_VIRIAL)
+ {
+ donb_flags |= GMX_NONBONDED_DO_SHIFTFORCE;
+ }
+ if (flags & GMX_FORCE_ENERGY)
+ {
+ donb_flags |= GMX_NONBONDED_DO_POTENTIAL;
+ }
+
+ wallcycle_sub_start(wcycle, ewcsNONBONDED);
+ do_nonbonded(fr, x, f, md, excl,
+ &enerd->grpp, nrnb,
+ lambda, dvdl_nb, -1, -1, donb_flags);
+
+ /* If we do foreign lambda and we have soft-core interactions
+ * we have to recalculate the (non-linear) energies contributions.
+ */
+ if (fepvals->n_lambda > 0 && (flags & GMX_FORCE_DHDL) && fepvals->sc_alpha != 0)
+ {
+ for (i = 0; i < enerd->n_lambda; i++)
+ {
+ real lam_i[efptNR];
+
+ for (j = 0; j < efptNR; j++)
+ {
+ lam_i[j] = (i == 0 ? lambda[j] : fepvals->all_lambda[j][i-1]);
+ }
+ reset_foreign_enerdata(enerd);
+ do_nonbonded(fr, x, f, md, excl,
+ &(enerd->foreign_grpp), nrnb,
+ lam_i, dvdl_dum, -1, -1,
+ (donb_flags & ~GMX_NONBONDED_DO_FORCE) | GMX_NONBONDED_DO_FOREIGNLAMBDA);
+ sum_epot(&(enerd->foreign_grpp), enerd->foreign_term);
+ enerd->enerpart_lambda[i] += enerd->foreign_term[F_EPOT];
+ }
+ }
+ wallcycle_sub_stop(wcycle, ewcsNONBONDED);
+ where();
+ }
+
+ /* If we are doing GB, calculate bonded forces and apply corrections
+ * to the solvation forces */
+ /* MRS: Eventually, many need to include free energy contribution here! */
+ if (ir->implicit_solvent)
+ {
+ wallcycle_sub_start(wcycle, ewcsLISTED);
+ calc_gb_forces(cr, md, born, top, x, f, fr, idef,
+ ir->gb_algorithm, ir->sa_algorithm, nrnb, &pbc, graph, enerd);
+ wallcycle_sub_stop(wcycle, ewcsLISTED);
+ }
+
+#if GMX_MPI
+ if (TAKETIME)
+ {
+ t1 = MPI_Wtime();
+ fr->t_fnbf += t1-t0;
+ }
+#endif
+
+ if (fepvals->sc_alpha != 0)
+ {
+ enerd->dvdl_nonlin[efptVDW] += dvdl_nb[efptVDW];
+ }
+ else
+ {
+ enerd->dvdl_lin[efptVDW] += dvdl_nb[efptVDW];
+ }
+
+ if (fepvals->sc_alpha != 0)
+
+ /* even though coulomb part is linear, we already added it, beacuse we
+ need to go through the vdw calculation anyway */
+ {
+ enerd->dvdl_nonlin[efptCOUL] += dvdl_nb[efptCOUL];
+ }
+ else
+ {
+ enerd->dvdl_lin[efptCOUL] += dvdl_nb[efptCOUL];
+ }
+
+ if (debug)
+ {
+ pr_rvecs(debug, 0, "fshift after SR", fr->fshift, SHIFTS);
+ }
+
+ /* Shift the coordinates. Must be done before listed forces and PPPM,
+ * but is also necessary for SHAKE and update, therefore it can NOT
+ * go when no listed forces have to be evaluated.
+ *
+ * The shifting and PBC code is deliberately not timed, since with
+ * the Verlet scheme it only takes non-zero time with triclinic
+ * boxes, and even then the time is around a factor of 100 less
+ * than the next smallest counter.
+ */
+
+
+ /* Here sometimes we would not need to shift with NBFonly,
+ * but we do so anyhow for consistency of the returned coordinates.
+ */
+ if (graph)
+ {
+ shift_self(graph, box, x);
+ if (TRICLINIC(box))
+ {
+ inc_nrnb(nrnb, eNR_SHIFTX, 2*graph->nnodes);
+ }
+ else
+ {
+ inc_nrnb(nrnb, eNR_SHIFTX, graph->nnodes);
+ }
+ }
+ /* Check whether we need to do listed interactions or correct for exclusions */
+ if (fr->bMolPBC &&
+ ((flags & GMX_FORCE_LISTED)
+ || EEL_RF(fr->eeltype) || EEL_FULL(fr->eeltype) || EVDW_PME(fr->vdwtype)))
+ {
+ /* TODO There are no electrostatics methods that require this
+ transformation, when using the Verlet scheme, so update the
+ above conditional. */
+ /* Since all atoms are in the rectangular or triclinic unit-cell,
+ * only single box vector shifts (2 in x) are required.
+ */
+ set_pbc_dd(&pbc, fr->ePBC, DOMAINDECOMP(cr) ? cr->dd->nc : nullptr,
+ TRUE, box);
+ }
+
+ do_force_listed(wcycle, box, ir->fepvals, cr->ms,
+ idef, (const rvec *) x, hist, f, fr,
+ &pbc, graph, enerd, nrnb, lambda, md, fcd,
+ DOMAINDECOMP(cr) ? cr->dd->gatindex : NULL,
+ flags);
+
+ where();
+
+ *cycles_pme = 0;
+ clear_mat(fr->vir_el_recip);
+ clear_mat(fr->vir_lj_recip);
+
+ /* Do long-range electrostatics and/or LJ-PME, including related short-range
+ * corrections.
+ */
+ if (EEL_FULL(fr->eeltype) || EVDW_PME(fr->vdwtype))
+ {
+ int status = 0;
+ real Vlr_q = 0, Vlr_lj = 0, Vcorr_q = 0, Vcorr_lj = 0;
+ real dvdl_long_range_q = 0, dvdl_long_range_lj = 0;
+
+ bSB = (ir->nwall == 2);
+ if (bSB)
+ {
+ copy_mat(box, boxs);
+ svmul(ir->wall_ewald_zfac, boxs[ZZ], boxs[ZZ]);
+ box_size[ZZ] *= ir->wall_ewald_zfac;
+ }
+
+ if (EEL_PME_EWALD(fr->eeltype) || EVDW_PME(fr->vdwtype))
+ {
+ real dvdl_long_range_correction_q = 0;
+ real dvdl_long_range_correction_lj = 0;
+ /* With the Verlet scheme exclusion forces are calculated
+ * in the non-bonded kernel.
+ */
+ /* The TPI molecule does not have exclusions with the rest
+ * of the system and no intra-molecular PME grid
+ * contributions will be calculated in
+ * gmx_pme_calc_energy.
+ */
+ if ((ir->cutoff_scheme == ecutsGROUP && fr->n_tpi == 0) ||
+ ir->ewald_geometry != eewg3D ||
+ ir->epsilon_surface != 0)
+ {
+ int nthreads, t;
+
+ wallcycle_sub_start(wcycle, ewcsEWALD_CORRECTION);
+
+ if (fr->n_tpi > 0)
+ {
+ gmx_fatal(FARGS, "TPI with PME currently only works in a 3D geometry with tin-foil boundary conditions");
+ }
+
+ nthreads = fr->nthread_ewc;
+#pragma omp parallel for num_threads(nthreads) schedule(static)
+ for (t = 0; t < nthreads; t++)
+ {
+ try
+ {
+ tensor *vir_q, *vir_lj;
+ real *Vcorrt_q, *Vcorrt_lj, *dvdlt_q, *dvdlt_lj;
+ if (t == 0)
+ {
+ vir_q = &fr->vir_el_recip;
+ vir_lj = &fr->vir_lj_recip;
+ Vcorrt_q = &Vcorr_q;
+ Vcorrt_lj = &Vcorr_lj;
+ dvdlt_q = &dvdl_long_range_correction_q;
+ dvdlt_lj = &dvdl_long_range_correction_lj;
+ }
+ else
+ {
+ vir_q = &fr->ewc_t[t].vir_q;
+ vir_lj = &fr->ewc_t[t].vir_lj;
+ Vcorrt_q = &fr->ewc_t[t].Vcorr_q;
+ Vcorrt_lj = &fr->ewc_t[t].Vcorr_lj;
+ dvdlt_q = &fr->ewc_t[t].dvdl[efptCOUL];
+ dvdlt_lj = &fr->ewc_t[t].dvdl[efptVDW];
+ clear_mat(*vir_q);
+ clear_mat(*vir_lj);
+ }
+ *dvdlt_q = 0;
+ *dvdlt_lj = 0;
+
+ /* Threading is only supported with the Verlet cut-off
+ * scheme and then only single particle forces (no
+ * exclusion forces) are calculated, so we can store
+ * the forces in the normal, single fr->f_novirsum array.
+ */
+ ewald_LRcorrection(fr->excl_load[t], fr->excl_load[t+1],
+ cr, t, fr,
+ md->chargeA, md->chargeB,
+ md->sqrt_c6A, md->sqrt_c6B,
+ md->sigmaA, md->sigmaB,
+ md->sigma3A, md->sigma3B,
+ md->nChargePerturbed || md->nTypePerturbed,
+ ir->cutoff_scheme != ecutsVERLET,
+ excl, x, bSB ? boxs : box, mu_tot,
+ ir->ewald_geometry,
+ ir->epsilon_surface,
+ fr->f_novirsum, *vir_q, *vir_lj,
+ Vcorrt_q, Vcorrt_lj,
+ lambda[efptCOUL], lambda[efptVDW],
+ dvdlt_q, dvdlt_lj);
+ }
+ GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+ }
+ if (nthreads > 1)
+ {
+ reduce_thread_energies(fr->vir_el_recip, fr->vir_lj_recip,
+ &Vcorr_q, &Vcorr_lj,
+ &dvdl_long_range_correction_q,
+ &dvdl_long_range_correction_lj,
+ nthreads, fr->ewc_t);
+ }
+ wallcycle_sub_stop(wcycle, ewcsEWALD_CORRECTION);
+ }
+
+ if (EEL_PME_EWALD(fr->eeltype) && fr->n_tpi == 0)
+ {
+ /* This is not in a subcounter because it takes a
+ negligible and constant-sized amount of time */
+ Vcorr_q += ewald_charge_correction(cr, fr, lambda[efptCOUL], box,
+ &dvdl_long_range_correction_q,
+ fr->vir_el_recip);
+ }
+
+ enerd->dvdl_lin[efptCOUL] += dvdl_long_range_correction_q;
+ enerd->dvdl_lin[efptVDW] += dvdl_long_range_correction_lj;
+
+ if ((EEL_PME(fr->eeltype) || EVDW_PME(fr->vdwtype)) && (cr->duty & DUTY_PME))
+ {
+ /* Do reciprocal PME for Coulomb and/or LJ. */
+ assert(fr->n_tpi >= 0);
+ if (fr->n_tpi == 0 || (flags & GMX_FORCE_STATECHANGED))
+ {
+ pme_flags = GMX_PME_SPREAD | GMX_PME_SOLVE;
+ if (EEL_PME(fr->eeltype))
+ {
+ pme_flags |= GMX_PME_DO_COULOMB;
+ }
+ if (EVDW_PME(fr->vdwtype))
+ {
+ pme_flags |= GMX_PME_DO_LJ;
+ }
+ if (flags & GMX_FORCE_FORCES)
+ {
+ pme_flags |= GMX_PME_CALC_F;
+ }
+ if (flags & GMX_FORCE_VIRIAL)
+ {
+ pme_flags |= GMX_PME_CALC_ENER_VIR;
+ }
+ if (fr->n_tpi > 0)
+ {
+ /* We don't calculate f, but we do want the potential */
+ pme_flags |= GMX_PME_CALC_POT;
+ }
+ wallcycle_start(wcycle, ewcPMEMESH);
+ status = gmx_pme_do(fr->pmedata,
+ 0, md->homenr - fr->n_tpi,
+ x, fr->f_novirsum,
+ md->chargeA, md->chargeB,
+ md->sqrt_c6A, md->sqrt_c6B,
+ md->sigmaA, md->sigmaB,
+ bSB ? boxs : box, cr,
+ DOMAINDECOMP(cr) ? dd_pme_maxshift_x(cr->dd) : 0,
+ DOMAINDECOMP(cr) ? dd_pme_maxshift_y(cr->dd) : 0,
+ nrnb, wcycle,
+ fr->vir_el_recip, fr->ewaldcoeff_q,
+ fr->vir_lj_recip, fr->ewaldcoeff_lj,
+ &Vlr_q, &Vlr_lj,
+ lambda[efptCOUL], lambda[efptVDW],
+ &dvdl_long_range_q, &dvdl_long_range_lj, pme_flags);
+ *cycles_pme = wallcycle_stop(wcycle, ewcPMEMESH);
+ if (status != 0)
+ {
+ gmx_fatal(FARGS, "Error %d in reciprocal PME routine", status);
+ }
+ /* We should try to do as little computation after
+ * this as possible, because parallel PME synchronizes
+ * the nodes, so we want all load imbalance of the
+ * rest of the force calculation to be before the PME
+ * call. DD load balancing is done on the whole time
+ * of the force call (without PME).
+ */
+ }
+ if (fr->n_tpi > 0)
+ {
+ if (EVDW_PME(ir->vdwtype))
+ {
+
+ gmx_fatal(FARGS, "Test particle insertion not implemented with LJ-PME");
+ }
+ /* Determine the PME grid energy of the test molecule
+ * with the PME grid potential of the other charges.
+ */
+ gmx_pme_calc_energy(fr->pmedata, fr->n_tpi,
+ x + md->homenr - fr->n_tpi,
+ md->chargeA + md->homenr - fr->n_tpi,
+ &Vlr_q);
+ }
+ }
+ }
+
+ if (!EEL_PME(fr->eeltype) && EEL_PME_EWALD(fr->eeltype))
+ {
+ Vlr_q = do_ewald(ir, x, fr->f_novirsum,
+ md->chargeA, md->chargeB,
+ box_size, cr, md->homenr,
+ fr->vir_el_recip, fr->ewaldcoeff_q,
+ lambda[efptCOUL], &dvdl_long_range_q, fr->ewald_table);
+ }
+
+ /* Note that with separate PME nodes we get the real energies later */
+ enerd->dvdl_lin[efptCOUL] += dvdl_long_range_q;
+ enerd->dvdl_lin[efptVDW] += dvdl_long_range_lj;
+ enerd->term[F_COUL_RECIP] = Vlr_q + Vcorr_q;
+ enerd->term[F_LJ_RECIP] = Vlr_lj + Vcorr_lj;
+ if (debug)
+ {
+ fprintf(debug, "Vlr_q = %g, Vcorr_q = %g, Vlr_corr_q = %g\n",
+ Vlr_q, Vcorr_q, enerd->term[F_COUL_RECIP]);
+ pr_rvecs(debug, 0, "vir_el_recip after corr", fr->vir_el_recip, DIM);
+ pr_rvecs(debug, 0, "fshift after LR Corrections", fr->fshift, SHIFTS);
+ fprintf(debug, "Vlr_lj: %g, Vcorr_lj = %g, Vlr_corr_lj = %g\n",
+ Vlr_lj, Vcorr_lj, enerd->term[F_LJ_RECIP]);
+ pr_rvecs(debug, 0, "vir_lj_recip after corr", fr->vir_lj_recip, DIM);
+ }
+ }
+ else
+ {
+ /* Is there a reaction-field exclusion correction needed?
+ * With the Verlet scheme, exclusion forces are calculated
+ * in the non-bonded kernel.
+ */
+ if (ir->cutoff_scheme != ecutsVERLET && EEL_RF(fr->eeltype))
+ {
+ real dvdl_rf_excl = 0;
+ enerd->term[F_RF_EXCL] =
+ RF_excl_correction(fr, graph, md, excl, x, f,
+ fr->fshift, &pbc, lambda[efptCOUL], &dvdl_rf_excl);
+
+ enerd->dvdl_lin[efptCOUL] += dvdl_rf_excl;
+ }
+ }
+ where();
+
+ if (debug)
+ {
+ print_nrnb(debug, nrnb);
+ }
+
+#if GMX_MPI
+ if (TAKETIME)
+ {
+ t2 = MPI_Wtime();
+ MPI_Barrier(cr->mpi_comm_mygroup);
+ t3 = MPI_Wtime();
+ fr->t_wait += t3-t2;
+ if (fr->timesteps == 11)
+ {
+ char buf[22];
+ fprintf(stderr, "* PP load balancing info: rank %d, step %s, rel wait time=%3.0f%% , load string value: %7.2f\n",
+ cr->nodeid, gmx_step_str(fr->timesteps, buf),
+ 100*fr->t_wait/(fr->t_wait+fr->t_fnbf),
+ (fr->t_fnbf+fr->t_wait)/fr->t_fnbf);
+ }
+ fr->timesteps++;
+ }
+#endif
+
+ if (debug)
+ {
+ pr_rvecs(debug, 0, "fshift after bondeds", fr->fshift, SHIFTS);
+ }
+
+ /* PLUMED */
+ if(plumedswitch){
+ int plumedNeedsEnergy;
+ (*plumedcmd)(plumedmain,"isEnergyNeeded",&plumedNeedsEnergy);
+ if(!plumedNeedsEnergy) (*plumedcmd)(plumedmain,"performCalc",NULL);
+ }
+ /* END PLUMED */
+}
+
+void init_enerdata(int ngener, int n_lambda, gmx_enerdata_t *enerd)
+{
+ int i, n2;
+
+ for (i = 0; i < F_NRE; i++)
+ {
+ enerd->term[i] = 0;
+ enerd->foreign_term[i] = 0;
+ }
+
+
+ for (i = 0; i < efptNR; i++)
+ {
+ enerd->dvdl_lin[i] = 0;
+ enerd->dvdl_nonlin[i] = 0;
+ }
+
+ n2 = ngener*ngener;
+ if (debug)
+ {
+ fprintf(debug, "Creating %d sized group matrix for energies\n", n2);
+ }
+ enerd->grpp.nener = n2;
+ enerd->foreign_grpp.nener = n2;
+ for (i = 0; (i < egNR); i++)
+ {
+ snew(enerd->grpp.ener[i], n2);
+ snew(enerd->foreign_grpp.ener[i], n2);
+ }
+
+ if (n_lambda)
+ {
+ enerd->n_lambda = 1 + n_lambda;
+ snew(enerd->enerpart_lambda, enerd->n_lambda);
+ }
+ else
+ {
+ enerd->n_lambda = 0;
+ }
+}
+
+void destroy_enerdata(gmx_enerdata_t *enerd)
+{
+ int i;
+
+ for (i = 0; (i < egNR); i++)
+ {
+ sfree(enerd->grpp.ener[i]);
+ }
+
+ for (i = 0; (i < egNR); i++)
+ {
+ sfree(enerd->foreign_grpp.ener[i]);
+ }
+
+ if (enerd->n_lambda)
+ {
+ sfree(enerd->enerpart_lambda);
+ }
+}
+
+static real sum_v(int n, real v[])
+{
+ real t;
+ int i;
+
+ t = 0.0;
+ for (i = 0; (i < n); i++)
+ {
+ t = t + v[i];
+ }
+
+ return t;
+}
+
+void sum_epot(gmx_grppairener_t *grpp, real *epot)
+{
+ int i;
+
+ /* Accumulate energies */
+ epot[F_COUL_SR] = sum_v(grpp->nener, grpp->ener[egCOULSR]);
+ epot[F_LJ] = sum_v(grpp->nener, grpp->ener[egLJSR]);
+ epot[F_LJ14] = sum_v(grpp->nener, grpp->ener[egLJ14]);
+ epot[F_COUL14] = sum_v(grpp->nener, grpp->ener[egCOUL14]);
+ /* We have already added 1-2,1-3, and 1-4 terms to F_GBPOL */
+ epot[F_GBPOL] += sum_v(grpp->nener, grpp->ener[egGB]);
+
+/* lattice part of LR doesnt belong to any group
+ * and has been added earlier
+ */
+ epot[F_BHAM] = sum_v(grpp->nener, grpp->ener[egBHAMSR]);
+
+ epot[F_EPOT] = 0;
+ for (i = 0; (i < F_EPOT); i++)
+ {
+ if (i != F_DISRESVIOL && i != F_ORIRESDEV)
+ {
+ epot[F_EPOT] += epot[i];
+ }
+ }
+}
+
+void sum_dhdl(gmx_enerdata_t *enerd, real *lambda, t_lambda *fepvals)
+{
+ int i, j, index;
+ double dlam;
+
+ enerd->dvdl_lin[efptVDW] += enerd->term[F_DVDL_VDW]; /* include dispersion correction */
+ enerd->term[F_DVDL] = 0.0;
+ for (i = 0; i < efptNR; i++)
+ {
+ if (fepvals->separate_dvdl[i])
+ {
+ /* could this be done more readably/compactly? */
+ switch (i)
+ {
+ case (efptMASS):
+ index = F_DKDL;
+ break;
+ case (efptCOUL):
+ index = F_DVDL_COUL;
+ break;
+ case (efptVDW):
+ index = F_DVDL_VDW;
+ break;
+ case (efptBONDED):
+ index = F_DVDL_BONDED;
+ break;
+ case (efptRESTRAINT):
+ index = F_DVDL_RESTRAINT;
+ break;
+ default:
+ index = F_DVDL;
+ break;
+ }
+ enerd->term[index] = enerd->dvdl_lin[i] + enerd->dvdl_nonlin[i];
+ if (debug)
+ {
+ fprintf(debug, "dvdl-%s[%2d]: %f: non-linear %f + linear %f\n",
+ efpt_names[i], i, enerd->term[index], enerd->dvdl_nonlin[i], enerd->dvdl_lin[i]);
+ }
+ }
+ else
+ {
+ enerd->term[F_DVDL] += enerd->dvdl_lin[i] + enerd->dvdl_nonlin[i];
+ if (debug)
+ {
+ fprintf(debug, "dvd-%sl[%2d]: %f: non-linear %f + linear %f\n",
+ efpt_names[0], i, enerd->term[F_DVDL], enerd->dvdl_nonlin[i], enerd->dvdl_lin[i]);
+ }
+ }
+ }
+
+ /* Notes on the foreign lambda free energy difference evaluation:
+ * Adding the potential and ekin terms that depend linearly on lambda
+ * as delta lam * dvdl to the energy differences is exact.
+ * For the constraints this is not exact, but we have no other option
+ * without literally changing the lengths and reevaluating the energies at each step.
+ * (try to remedy this post 4.6 - MRS)
+ */
+ if (fepvals->separate_dvdl[efptBONDED])
+ {
+ enerd->term[F_DVDL_BONDED] += enerd->term[F_DVDL_CONSTR];
+ }
+ else
+ {
+ enerd->term[F_DVDL] += enerd->term[F_DVDL_CONSTR];
+ }
+ enerd->term[F_DVDL_CONSTR] = 0;
+
+ for (i = 0; i < fepvals->n_lambda; i++)
+ {
+ /* note we are iterating over fepvals here!
+ For the current lam, dlam = 0 automatically,
+ so we don't need to add anything to the
+ enerd->enerpart_lambda[0] */
+
+ /* we don't need to worry about dvdl_lin contributions to dE at
+ current lambda, because the contributions to the current
+ lambda are automatically zeroed */
+
+ for (j = 0; j < efptNR; j++)
+ {
+ /* Note that this loop is over all dhdl components, not just the separated ones */
+ dlam = (fepvals->all_lambda[j][i]-lambda[j]);
+ enerd->enerpart_lambda[i+1] += dlam*enerd->dvdl_lin[j];
+ if (debug)
+ {
+ fprintf(debug, "enerdiff lam %g: (%15s), non-linear %f linear %f*%f\n",
+ fepvals->all_lambda[j][i], efpt_names[j],
+ (enerd->enerpart_lambda[i+1] - enerd->enerpart_lambda[0]),
+ dlam, enerd->dvdl_lin[j]);
+ }
+ }
+ }
+}
+
+
+void reset_foreign_enerdata(gmx_enerdata_t *enerd)
+{
+ int i, j;
+
+ /* First reset all foreign energy components. Foreign energies always called on
+ neighbor search steps */
+ for (i = 0; (i < egNR); i++)
+ {
+ for (j = 0; (j < enerd->grpp.nener); j++)
+ {
+ enerd->foreign_grpp.ener[i][j] = 0.0;
+ }
+ }
+
+ /* potential energy components */
+ for (i = 0; (i <= F_EPOT); i++)
+ {
+ enerd->foreign_term[i] = 0.0;
+ }
+}
+
+void reset_enerdata(gmx_enerdata_t *enerd)
+{
+ int i, j;
+
+ /* First reset all energy components. */
+ for (i = 0; (i < egNR); i++)
+ {
+ for (j = 0; (j < enerd->grpp.nener); j++)
+ {
+ enerd->grpp.ener[i][j] = 0.0;
+ }
+ }
+ for (i = 0; i < efptNR; i++)
+ {
+ enerd->dvdl_lin[i] = 0.0;
+ enerd->dvdl_nonlin[i] = 0.0;
+ }
+
+ /* Normal potential energy components */
+ for (i = 0; (i <= F_EPOT); i++)
+ {
+ enerd->term[i] = 0.0;
+ }
+ enerd->term[F_DVDL] = 0.0;
+ enerd->term[F_DVDL_COUL] = 0.0;
+ enerd->term[F_DVDL_VDW] = 0.0;
+ enerd->term[F_DVDL_BONDED] = 0.0;
+ enerd->term[F_DVDL_RESTRAINT] = 0.0;
+ enerd->term[F_DKDL] = 0.0;
+ if (enerd->n_lambda > 0)
+ {
+ for (i = 0; i < enerd->n_lambda; i++)
+ {
+ enerd->enerpart_lambda[i] = 0.0;
+ }
+ }
+ /* reset foreign energy data - separate function since we also call it elsewhere */
+ reset_foreign_enerdata(enerd);
+}
diff --git a/patches/gromacs-2016-beta1.diff/src/gromacs/mdlib/force.cpp.preplumed b/patches/gromacs-2016-beta1.diff/src/gromacs/mdlib/force.cpp.preplumed
new file mode 100644
index 0000000000000000000000000000000000000000..107e00a8e7b3393ec0eaf42a07f584d29c6ea4eb
--- /dev/null
+++ b/patches/gromacs-2016-beta1.diff/src/gromacs/mdlib/force.cpp.preplumed
@@ -0,0 +1,899 @@
+/*
+ * This file is part of the GROMACS molecular simulation package.
+ *
+ * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
+ * Copyright (c) 2001-2004, The GROMACS development team.
+ * Copyright (c) 2013,2014,2015,2016, by the GROMACS development team, led by
+ * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
+ * and including many others, as listed in the AUTHORS file in the
+ * top-level source directory and at http://www.gromacs.org.
+ *
+ * GROMACS is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public License
+ * as published by the Free Software Foundation; either version 2.1
+ * of the License, or (at your option) any later version.
+ *
+ * GROMACS is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with GROMACS; if not, see
+ * http://www.gnu.org/licenses, or write to the Free Software Foundation,
+ * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ * If you want to redistribute modifications to GROMACS, 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 http://www.gromacs.org.
+ *
+ * To help us fund GROMACS development, we humbly ask that you cite
+ * the research papers on the package. Check out http://www.gromacs.org.
+ */
+#include "gmxpre.h"
+
+#include "force.h"
+
+#include "config.h"
+
+#include <assert.h>
+#include <math.h>
+#include <string.h>
+
+#include "gromacs/domdec/domdec.h"
+#include "gromacs/domdec/domdec_struct.h"
+#include "gromacs/ewald/ewald.h"
+#include "gromacs/ewald/long-range-correction.h"
+#include "gromacs/ewald/pme.h"
+#include "gromacs/gmxlib/network.h"
+#include "gromacs/gmxlib/nrnb.h"
+#include "gromacs/gmxlib/nonbonded/nonbonded.h"
+#include "gromacs/listed-forces/listed-forces.h"
+#include "gromacs/math/vec.h"
+#include "gromacs/math/vecdump.h"
+#include "gromacs/mdlib/forcerec-threading.h"
+#include "gromacs/mdlib/genborn.h"
+#include "gromacs/mdlib/mdrun.h"
+#include "gromacs/mdlib/ns.h"
+#include "gromacs/mdlib/qmmm.h"
+#include "gromacs/mdtypes/commrec.h"
+#include "gromacs/mdtypes/inputrec.h"
+#include "gromacs/mdtypes/md_enums.h"
+#include "gromacs/pbcutil/ishift.h"
+#include "gromacs/pbcutil/mshift.h"
+#include "gromacs/pbcutil/pbc.h"
+#include "gromacs/timing/wallcycle.h"
+#include "gromacs/utility/cstringutil.h"
+#include "gromacs/utility/exceptions.h"
+#include "gromacs/utility/fatalerror.h"
+#include "gromacs/utility/smalloc.h"
+
+void ns(FILE *fp,
+ t_forcerec *fr,
+ matrix box,
+ gmx_groups_t *groups,
+ gmx_localtop_t *top,
+ t_mdatoms *md,
+ t_commrec *cr,
+ t_nrnb *nrnb,
+ gmx_bool bFillGrid)
+{
+ int nsearch;
+
+
+ if (!fr->ns->nblist_initialized)
+ {
+ init_neighbor_list(fp, fr, md->homenr);
+ }
+
+ nsearch = search_neighbours(fp, fr, box, top, groups, cr, nrnb, md,
+ bFillGrid);
+ if (debug)
+ {
+ fprintf(debug, "nsearch = %d\n", nsearch);
+ }
+
+ /* Check whether we have to do dynamic load balancing */
+ /*if ((nsb->nstDlb > 0) && (mod(step,nsb->nstDlb) == 0))
+ count_nb(cr,nsb,&(top->blocks[ebCGS]),nns,fr->nlr,
+ &(top->idef),opts->ngener);
+ */
+ if (fr->ns->dump_nl > 0)
+ {
+ dump_nblist(fp, cr, fr, fr->ns->dump_nl);
+ }
+}
+
+static void reduce_thread_energies(tensor vir_q, tensor vir_lj,
+ real *Vcorr_q, real *Vcorr_lj,
+ real *dvdl_q, real *dvdl_lj,
+ int nthreads,
+ ewald_corr_thread_t *ewc_t)
+{
+ int t;
+
+ for (t = 1; t < nthreads; t++)
+ {
+ *Vcorr_q += ewc_t[t].Vcorr_q;
+ *Vcorr_lj += ewc_t[t].Vcorr_lj;
+ *dvdl_q += ewc_t[t].dvdl[efptCOUL];
+ *dvdl_lj += ewc_t[t].dvdl[efptVDW];
+ m_add(vir_q, ewc_t[t].vir_q, vir_q);
+ m_add(vir_lj, ewc_t[t].vir_lj, vir_lj);
+ }
+}
+
+void do_force_lowlevel(t_forcerec *fr, t_inputrec *ir,
+ t_idef *idef, t_commrec *cr,
+ t_nrnb *nrnb, gmx_wallcycle_t wcycle,
+ t_mdatoms *md,
+ rvec x[], history_t *hist,
+ rvec f[],
+ gmx_enerdata_t *enerd,
+ t_fcdata *fcd,
+ gmx_localtop_t *top,
+ gmx_genborn_t *born,
+ gmx_bool bBornRadii,
+ matrix box,
+ t_lambda *fepvals,
+ real *lambda,
+ t_graph *graph,
+ t_blocka *excl,
+ rvec mu_tot[],
+ int flags,
+ float *cycles_pme)
+{
+ int i, j;
+ int donb_flags;
+ gmx_bool bSB;
+ int pme_flags;
+ matrix boxs;
+ rvec box_size;
+ t_pbc pbc;
+ real dvdl_dum[efptNR], dvdl_nb[efptNR];
+
+#if GMX_MPI
+ double t0 = 0.0, t1, t2, t3; /* time measurement for coarse load balancing */
+#endif
+
+ set_pbc(&pbc, fr->ePBC, box);
+
+ /* reset free energy components */
+ for (i = 0; i < efptNR; i++)
+ {
+ dvdl_nb[i] = 0;
+ dvdl_dum[i] = 0;
+ }
+
+ /* Reset box */
+ for (i = 0; (i < DIM); i++)
+ {
+ box_size[i] = box[i][i];
+ }
+
+ /* do QMMM first if requested */
+ if (fr->bQMMM)
+ {
+ enerd->term[F_EQM] = calculate_QMMM(cr, x, f, fr);
+ }
+
+ /* Call the short range functions all in one go. */
+
+#if GMX_MPI
+ /*#define TAKETIME ((cr->npmenodes) && (fr->timesteps < 12))*/
+#define TAKETIME FALSE
+ if (TAKETIME)
+ {
+ MPI_Barrier(cr->mpi_comm_mygroup);
+ t0 = MPI_Wtime();
+ }
+#endif
+
+ if (ir->nwall)
+ {
+ /* foreign lambda component for walls */
+ real dvdl_walls = do_walls(ir, fr, box, md, x, f, lambda[efptVDW],
+ enerd->grpp.ener[egLJSR], nrnb);
+ enerd->dvdl_lin[efptVDW] += dvdl_walls;
+ }
+
+ /* If doing GB, reset dvda and calculate the Born radii */
+ if (ir->implicit_solvent)
+ {
+ wallcycle_sub_start(wcycle, ewcsNONBONDED);
+
+ for (i = 0; i < born->nr; i++)
+ {
+ fr->dvda[i] = 0;
+ }
+
+ if (bBornRadii)
+ {
+ calc_gb_rad(cr, fr, ir, top, x, fr->gblist, born, md, nrnb);
+ }
+
+ wallcycle_sub_stop(wcycle, ewcsNONBONDED);
+ }
+
+ where();
+ /* We only do non-bonded calculation with group scheme here, the verlet
+ * calls are done from do_force_cutsVERLET(). */
+ if (fr->cutoff_scheme == ecutsGROUP && (flags & GMX_FORCE_NONBONDED))
+ {
+ donb_flags = 0;
+ /* Add short-range interactions */
+ donb_flags |= GMX_NONBONDED_DO_SR;
+
+ /* Currently all group scheme kernels always calculate (shift-)forces */
+ if (flags & GMX_FORCE_FORCES)
+ {
+ donb_flags |= GMX_NONBONDED_DO_FORCE;
+ }
+ if (flags & GMX_FORCE_VIRIAL)
+ {
+ donb_flags |= GMX_NONBONDED_DO_SHIFTFORCE;
+ }
+ if (flags & GMX_FORCE_ENERGY)
+ {
+ donb_flags |= GMX_NONBONDED_DO_POTENTIAL;
+ }
+
+ wallcycle_sub_start(wcycle, ewcsNONBONDED);
+ do_nonbonded(fr, x, f, md, excl,
+ &enerd->grpp, nrnb,
+ lambda, dvdl_nb, -1, -1, donb_flags);
+
+ /* If we do foreign lambda and we have soft-core interactions
+ * we have to recalculate the (non-linear) energies contributions.
+ */
+ if (fepvals->n_lambda > 0 && (flags & GMX_FORCE_DHDL) && fepvals->sc_alpha != 0)
+ {
+ for (i = 0; i < enerd->n_lambda; i++)
+ {
+ real lam_i[efptNR];
+
+ for (j = 0; j < efptNR; j++)
+ {
+ lam_i[j] = (i == 0 ? lambda[j] : fepvals->all_lambda[j][i-1]);
+ }
+ reset_foreign_enerdata(enerd);
+ do_nonbonded(fr, x, f, md, excl,
+ &(enerd->foreign_grpp), nrnb,
+ lam_i, dvdl_dum, -1, -1,
+ (donb_flags & ~GMX_NONBONDED_DO_FORCE) | GMX_NONBONDED_DO_FOREIGNLAMBDA);
+ sum_epot(&(enerd->foreign_grpp), enerd->foreign_term);
+ enerd->enerpart_lambda[i] += enerd->foreign_term[F_EPOT];
+ }
+ }
+ wallcycle_sub_stop(wcycle, ewcsNONBONDED);
+ where();
+ }
+
+ /* If we are doing GB, calculate bonded forces and apply corrections
+ * to the solvation forces */
+ /* MRS: Eventually, many need to include free energy contribution here! */
+ if (ir->implicit_solvent)
+ {
+ wallcycle_sub_start(wcycle, ewcsLISTED);
+ calc_gb_forces(cr, md, born, top, x, f, fr, idef,
+ ir->gb_algorithm, ir->sa_algorithm, nrnb, &pbc, graph, enerd);
+ wallcycle_sub_stop(wcycle, ewcsLISTED);
+ }
+
+#if GMX_MPI
+ if (TAKETIME)
+ {
+ t1 = MPI_Wtime();
+ fr->t_fnbf += t1-t0;
+ }
+#endif
+
+ if (fepvals->sc_alpha != 0)
+ {
+ enerd->dvdl_nonlin[efptVDW] += dvdl_nb[efptVDW];
+ }
+ else
+ {
+ enerd->dvdl_lin[efptVDW] += dvdl_nb[efptVDW];
+ }
+
+ if (fepvals->sc_alpha != 0)
+
+ /* even though coulomb part is linear, we already added it, beacuse we
+ need to go through the vdw calculation anyway */
+ {
+ enerd->dvdl_nonlin[efptCOUL] += dvdl_nb[efptCOUL];
+ }
+ else
+ {
+ enerd->dvdl_lin[efptCOUL] += dvdl_nb[efptCOUL];
+ }
+
+ if (debug)
+ {
+ pr_rvecs(debug, 0, "fshift after SR", fr->fshift, SHIFTS);
+ }
+
+ /* Shift the coordinates. Must be done before listed forces and PPPM,
+ * but is also necessary for SHAKE and update, therefore it can NOT
+ * go when no listed forces have to be evaluated.
+ *
+ * The shifting and PBC code is deliberately not timed, since with
+ * the Verlet scheme it only takes non-zero time with triclinic
+ * boxes, and even then the time is around a factor of 100 less
+ * than the next smallest counter.
+ */
+
+
+ /* Here sometimes we would not need to shift with NBFonly,
+ * but we do so anyhow for consistency of the returned coordinates.
+ */
+ if (graph)
+ {
+ shift_self(graph, box, x);
+ if (TRICLINIC(box))
+ {
+ inc_nrnb(nrnb, eNR_SHIFTX, 2*graph->nnodes);
+ }
+ else
+ {
+ inc_nrnb(nrnb, eNR_SHIFTX, graph->nnodes);
+ }
+ }
+ /* Check whether we need to do listed interactions or correct for exclusions */
+ if (fr->bMolPBC &&
+ ((flags & GMX_FORCE_LISTED)
+ || EEL_RF(fr->eeltype) || EEL_FULL(fr->eeltype) || EVDW_PME(fr->vdwtype)))
+ {
+ /* TODO There are no electrostatics methods that require this
+ transformation, when using the Verlet scheme, so update the
+ above conditional. */
+ /* Since all atoms are in the rectangular or triclinic unit-cell,
+ * only single box vector shifts (2 in x) are required.
+ */
+ set_pbc_dd(&pbc, fr->ePBC, DOMAINDECOMP(cr) ? cr->dd->nc : nullptr,
+ TRUE, box);
+ }
+
+ do_force_listed(wcycle, box, ir->fepvals, cr->ms,
+ idef, (const rvec *) x, hist, f, fr,
+ &pbc, graph, enerd, nrnb, lambda, md, fcd,
+ DOMAINDECOMP(cr) ? cr->dd->gatindex : NULL,
+ flags);
+
+ where();
+
+ *cycles_pme = 0;
+ clear_mat(fr->vir_el_recip);
+ clear_mat(fr->vir_lj_recip);
+
+ /* Do long-range electrostatics and/or LJ-PME, including related short-range
+ * corrections.
+ */
+ if (EEL_FULL(fr->eeltype) || EVDW_PME(fr->vdwtype))
+ {
+ int status = 0;
+ real Vlr_q = 0, Vlr_lj = 0, Vcorr_q = 0, Vcorr_lj = 0;
+ real dvdl_long_range_q = 0, dvdl_long_range_lj = 0;
+
+ bSB = (ir->nwall == 2);
+ if (bSB)
+ {
+ copy_mat(box, boxs);
+ svmul(ir->wall_ewald_zfac, boxs[ZZ], boxs[ZZ]);
+ box_size[ZZ] *= ir->wall_ewald_zfac;
+ }
+
+ if (EEL_PME_EWALD(fr->eeltype) || EVDW_PME(fr->vdwtype))
+ {
+ real dvdl_long_range_correction_q = 0;
+ real dvdl_long_range_correction_lj = 0;
+ /* With the Verlet scheme exclusion forces are calculated
+ * in the non-bonded kernel.
+ */
+ /* The TPI molecule does not have exclusions with the rest
+ * of the system and no intra-molecular PME grid
+ * contributions will be calculated in
+ * gmx_pme_calc_energy.
+ */
+ if ((ir->cutoff_scheme == ecutsGROUP && fr->n_tpi == 0) ||
+ ir->ewald_geometry != eewg3D ||
+ ir->epsilon_surface != 0)
+ {
+ int nthreads, t;
+
+ wallcycle_sub_start(wcycle, ewcsEWALD_CORRECTION);
+
+ if (fr->n_tpi > 0)
+ {
+ gmx_fatal(FARGS, "TPI with PME currently only works in a 3D geometry with tin-foil boundary conditions");
+ }
+
+ nthreads = fr->nthread_ewc;
+#pragma omp parallel for num_threads(nthreads) schedule(static)
+ for (t = 0; t < nthreads; t++)
+ {
+ try
+ {
+ tensor *vir_q, *vir_lj;
+ real *Vcorrt_q, *Vcorrt_lj, *dvdlt_q, *dvdlt_lj;
+ if (t == 0)
+ {
+ vir_q = &fr->vir_el_recip;
+ vir_lj = &fr->vir_lj_recip;
+ Vcorrt_q = &Vcorr_q;
+ Vcorrt_lj = &Vcorr_lj;
+ dvdlt_q = &dvdl_long_range_correction_q;
+ dvdlt_lj = &dvdl_long_range_correction_lj;
+ }
+ else
+ {
+ vir_q = &fr->ewc_t[t].vir_q;
+ vir_lj = &fr->ewc_t[t].vir_lj;
+ Vcorrt_q = &fr->ewc_t[t].Vcorr_q;
+ Vcorrt_lj = &fr->ewc_t[t].Vcorr_lj;
+ dvdlt_q = &fr->ewc_t[t].dvdl[efptCOUL];
+ dvdlt_lj = &fr->ewc_t[t].dvdl[efptVDW];
+ clear_mat(*vir_q);
+ clear_mat(*vir_lj);
+ }
+ *dvdlt_q = 0;
+ *dvdlt_lj = 0;
+
+ /* Threading is only supported with the Verlet cut-off
+ * scheme and then only single particle forces (no
+ * exclusion forces) are calculated, so we can store
+ * the forces in the normal, single fr->f_novirsum array.
+ */
+ ewald_LRcorrection(fr->excl_load[t], fr->excl_load[t+1],
+ cr, t, fr,
+ md->chargeA, md->chargeB,
+ md->sqrt_c6A, md->sqrt_c6B,
+ md->sigmaA, md->sigmaB,
+ md->sigma3A, md->sigma3B,
+ md->nChargePerturbed || md->nTypePerturbed,
+ ir->cutoff_scheme != ecutsVERLET,
+ excl, x, bSB ? boxs : box, mu_tot,
+ ir->ewald_geometry,
+ ir->epsilon_surface,
+ fr->f_novirsum, *vir_q, *vir_lj,
+ Vcorrt_q, Vcorrt_lj,
+ lambda[efptCOUL], lambda[efptVDW],
+ dvdlt_q, dvdlt_lj);
+ }
+ GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+ }
+ if (nthreads > 1)
+ {
+ reduce_thread_energies(fr->vir_el_recip, fr->vir_lj_recip,
+ &Vcorr_q, &Vcorr_lj,
+ &dvdl_long_range_correction_q,
+ &dvdl_long_range_correction_lj,
+ nthreads, fr->ewc_t);
+ }
+ wallcycle_sub_stop(wcycle, ewcsEWALD_CORRECTION);
+ }
+
+ if (EEL_PME_EWALD(fr->eeltype) && fr->n_tpi == 0)
+ {
+ /* This is not in a subcounter because it takes a
+ negligible and constant-sized amount of time */
+ Vcorr_q += ewald_charge_correction(cr, fr, lambda[efptCOUL], box,
+ &dvdl_long_range_correction_q,
+ fr->vir_el_recip);
+ }
+
+ enerd->dvdl_lin[efptCOUL] += dvdl_long_range_correction_q;
+ enerd->dvdl_lin[efptVDW] += dvdl_long_range_correction_lj;
+
+ if ((EEL_PME(fr->eeltype) || EVDW_PME(fr->vdwtype)) && (cr->duty & DUTY_PME))
+ {
+ /* Do reciprocal PME for Coulomb and/or LJ. */
+ assert(fr->n_tpi >= 0);
+ if (fr->n_tpi == 0 || (flags & GMX_FORCE_STATECHANGED))
+ {
+ pme_flags = GMX_PME_SPREAD | GMX_PME_SOLVE;
+ if (EEL_PME(fr->eeltype))
+ {
+ pme_flags |= GMX_PME_DO_COULOMB;
+ }
+ if (EVDW_PME(fr->vdwtype))
+ {
+ pme_flags |= GMX_PME_DO_LJ;
+ }
+ if (flags & GMX_FORCE_FORCES)
+ {
+ pme_flags |= GMX_PME_CALC_F;
+ }
+ if (flags & GMX_FORCE_VIRIAL)
+ {
+ pme_flags |= GMX_PME_CALC_ENER_VIR;
+ }
+ if (fr->n_tpi > 0)
+ {
+ /* We don't calculate f, but we do want the potential */
+ pme_flags |= GMX_PME_CALC_POT;
+ }
+ wallcycle_start(wcycle, ewcPMEMESH);
+ status = gmx_pme_do(fr->pmedata,
+ 0, md->homenr - fr->n_tpi,
+ x, fr->f_novirsum,
+ md->chargeA, md->chargeB,
+ md->sqrt_c6A, md->sqrt_c6B,
+ md->sigmaA, md->sigmaB,
+ bSB ? boxs : box, cr,
+ DOMAINDECOMP(cr) ? dd_pme_maxshift_x(cr->dd) : 0,
+ DOMAINDECOMP(cr) ? dd_pme_maxshift_y(cr->dd) : 0,
+ nrnb, wcycle,
+ fr->vir_el_recip, fr->ewaldcoeff_q,
+ fr->vir_lj_recip, fr->ewaldcoeff_lj,
+ &Vlr_q, &Vlr_lj,
+ lambda[efptCOUL], lambda[efptVDW],
+ &dvdl_long_range_q, &dvdl_long_range_lj, pme_flags);
+ *cycles_pme = wallcycle_stop(wcycle, ewcPMEMESH);
+ if (status != 0)
+ {
+ gmx_fatal(FARGS, "Error %d in reciprocal PME routine", status);
+ }
+ /* We should try to do as little computation after
+ * this as possible, because parallel PME synchronizes
+ * the nodes, so we want all load imbalance of the
+ * rest of the force calculation to be before the PME
+ * call. DD load balancing is done on the whole time
+ * of the force call (without PME).
+ */
+ }
+ if (fr->n_tpi > 0)
+ {
+ if (EVDW_PME(ir->vdwtype))
+ {
+
+ gmx_fatal(FARGS, "Test particle insertion not implemented with LJ-PME");
+ }
+ /* Determine the PME grid energy of the test molecule
+ * with the PME grid potential of the other charges.
+ */
+ gmx_pme_calc_energy(fr->pmedata, fr->n_tpi,
+ x + md->homenr - fr->n_tpi,
+ md->chargeA + md->homenr - fr->n_tpi,
+ &Vlr_q);
+ }
+ }
+ }
+
+ if (!EEL_PME(fr->eeltype) && EEL_PME_EWALD(fr->eeltype))
+ {
+ Vlr_q = do_ewald(ir, x, fr->f_novirsum,
+ md->chargeA, md->chargeB,
+ box_size, cr, md->homenr,
+ fr->vir_el_recip, fr->ewaldcoeff_q,
+ lambda[efptCOUL], &dvdl_long_range_q, fr->ewald_table);
+ }
+
+ /* Note that with separate PME nodes we get the real energies later */
+ enerd->dvdl_lin[efptCOUL] += dvdl_long_range_q;
+ enerd->dvdl_lin[efptVDW] += dvdl_long_range_lj;
+ enerd->term[F_COUL_RECIP] = Vlr_q + Vcorr_q;
+ enerd->term[F_LJ_RECIP] = Vlr_lj + Vcorr_lj;
+ if (debug)
+ {
+ fprintf(debug, "Vlr_q = %g, Vcorr_q = %g, Vlr_corr_q = %g\n",
+ Vlr_q, Vcorr_q, enerd->term[F_COUL_RECIP]);
+ pr_rvecs(debug, 0, "vir_el_recip after corr", fr->vir_el_recip, DIM);
+ pr_rvecs(debug, 0, "fshift after LR Corrections", fr->fshift, SHIFTS);
+ fprintf(debug, "Vlr_lj: %g, Vcorr_lj = %g, Vlr_corr_lj = %g\n",
+ Vlr_lj, Vcorr_lj, enerd->term[F_LJ_RECIP]);
+ pr_rvecs(debug, 0, "vir_lj_recip after corr", fr->vir_lj_recip, DIM);
+ }
+ }
+ else
+ {
+ /* Is there a reaction-field exclusion correction needed?
+ * With the Verlet scheme, exclusion forces are calculated
+ * in the non-bonded kernel.
+ */
+ if (ir->cutoff_scheme != ecutsVERLET && EEL_RF(fr->eeltype))
+ {
+ real dvdl_rf_excl = 0;
+ enerd->term[F_RF_EXCL] =
+ RF_excl_correction(fr, graph, md, excl, x, f,
+ fr->fshift, &pbc, lambda[efptCOUL], &dvdl_rf_excl);
+
+ enerd->dvdl_lin[efptCOUL] += dvdl_rf_excl;
+ }
+ }
+ where();
+
+ if (debug)
+ {
+ print_nrnb(debug, nrnb);
+ }
+
+#if GMX_MPI
+ if (TAKETIME)
+ {
+ t2 = MPI_Wtime();
+ MPI_Barrier(cr->mpi_comm_mygroup);
+ t3 = MPI_Wtime();
+ fr->t_wait += t3-t2;
+ if (fr->timesteps == 11)
+ {
+ char buf[22];
+ fprintf(stderr, "* PP load balancing info: rank %d, step %s, rel wait time=%3.0f%% , load string value: %7.2f\n",
+ cr->nodeid, gmx_step_str(fr->timesteps, buf),
+ 100*fr->t_wait/(fr->t_wait+fr->t_fnbf),
+ (fr->t_fnbf+fr->t_wait)/fr->t_fnbf);
+ }
+ fr->timesteps++;
+ }
+#endif
+
+ if (debug)
+ {
+ pr_rvecs(debug, 0, "fshift after bondeds", fr->fshift, SHIFTS);
+ }
+
+}
+
+void init_enerdata(int ngener, int n_lambda, gmx_enerdata_t *enerd)
+{
+ int i, n2;
+
+ for (i = 0; i < F_NRE; i++)
+ {
+ enerd->term[i] = 0;
+ enerd->foreign_term[i] = 0;
+ }
+
+
+ for (i = 0; i < efptNR; i++)
+ {
+ enerd->dvdl_lin[i] = 0;
+ enerd->dvdl_nonlin[i] = 0;
+ }
+
+ n2 = ngener*ngener;
+ if (debug)
+ {
+ fprintf(debug, "Creating %d sized group matrix for energies\n", n2);
+ }
+ enerd->grpp.nener = n2;
+ enerd->foreign_grpp.nener = n2;
+ for (i = 0; (i < egNR); i++)
+ {
+ snew(enerd->grpp.ener[i], n2);
+ snew(enerd->foreign_grpp.ener[i], n2);
+ }
+
+ if (n_lambda)
+ {
+ enerd->n_lambda = 1 + n_lambda;
+ snew(enerd->enerpart_lambda, enerd->n_lambda);
+ }
+ else
+ {
+ enerd->n_lambda = 0;
+ }
+}
+
+void destroy_enerdata(gmx_enerdata_t *enerd)
+{
+ int i;
+
+ for (i = 0; (i < egNR); i++)
+ {
+ sfree(enerd->grpp.ener[i]);
+ }
+
+ for (i = 0; (i < egNR); i++)
+ {
+ sfree(enerd->foreign_grpp.ener[i]);
+ }
+
+ if (enerd->n_lambda)
+ {
+ sfree(enerd->enerpart_lambda);
+ }
+}
+
+static real sum_v(int n, real v[])
+{
+ real t;
+ int i;
+
+ t = 0.0;
+ for (i = 0; (i < n); i++)
+ {
+ t = t + v[i];
+ }
+
+ return t;
+}
+
+void sum_epot(gmx_grppairener_t *grpp, real *epot)
+{
+ int i;
+
+ /* Accumulate energies */
+ epot[F_COUL_SR] = sum_v(grpp->nener, grpp->ener[egCOULSR]);
+ epot[F_LJ] = sum_v(grpp->nener, grpp->ener[egLJSR]);
+ epot[F_LJ14] = sum_v(grpp->nener, grpp->ener[egLJ14]);
+ epot[F_COUL14] = sum_v(grpp->nener, grpp->ener[egCOUL14]);
+ /* We have already added 1-2,1-3, and 1-4 terms to F_GBPOL */
+ epot[F_GBPOL] += sum_v(grpp->nener, grpp->ener[egGB]);
+
+/* lattice part of LR doesnt belong to any group
+ * and has been added earlier
+ */
+ epot[F_BHAM] = sum_v(grpp->nener, grpp->ener[egBHAMSR]);
+
+ epot[F_EPOT] = 0;
+ for (i = 0; (i < F_EPOT); i++)
+ {
+ if (i != F_DISRESVIOL && i != F_ORIRESDEV)
+ {
+ epot[F_EPOT] += epot[i];
+ }
+ }
+}
+
+void sum_dhdl(gmx_enerdata_t *enerd, real *lambda, t_lambda *fepvals)
+{
+ int i, j, index;
+ double dlam;
+
+ enerd->dvdl_lin[efptVDW] += enerd->term[F_DVDL_VDW]; /* include dispersion correction */
+ enerd->term[F_DVDL] = 0.0;
+ for (i = 0; i < efptNR; i++)
+ {
+ if (fepvals->separate_dvdl[i])
+ {
+ /* could this be done more readably/compactly? */
+ switch (i)
+ {
+ case (efptMASS):
+ index = F_DKDL;
+ break;
+ case (efptCOUL):
+ index = F_DVDL_COUL;
+ break;
+ case (efptVDW):
+ index = F_DVDL_VDW;
+ break;
+ case (efptBONDED):
+ index = F_DVDL_BONDED;
+ break;
+ case (efptRESTRAINT):
+ index = F_DVDL_RESTRAINT;
+ break;
+ default:
+ index = F_DVDL;
+ break;
+ }
+ enerd->term[index] = enerd->dvdl_lin[i] + enerd->dvdl_nonlin[i];
+ if (debug)
+ {
+ fprintf(debug, "dvdl-%s[%2d]: %f: non-linear %f + linear %f\n",
+ efpt_names[i], i, enerd->term[index], enerd->dvdl_nonlin[i], enerd->dvdl_lin[i]);
+ }
+ }
+ else
+ {
+ enerd->term[F_DVDL] += enerd->dvdl_lin[i] + enerd->dvdl_nonlin[i];
+ if (debug)
+ {
+ fprintf(debug, "dvd-%sl[%2d]: %f: non-linear %f + linear %f\n",
+ efpt_names[0], i, enerd->term[F_DVDL], enerd->dvdl_nonlin[i], enerd->dvdl_lin[i]);
+ }
+ }
+ }
+
+ /* Notes on the foreign lambda free energy difference evaluation:
+ * Adding the potential and ekin terms that depend linearly on lambda
+ * as delta lam * dvdl to the energy differences is exact.
+ * For the constraints this is not exact, but we have no other option
+ * without literally changing the lengths and reevaluating the energies at each step.
+ * (try to remedy this post 4.6 - MRS)
+ */
+ if (fepvals->separate_dvdl[efptBONDED])
+ {
+ enerd->term[F_DVDL_BONDED] += enerd->term[F_DVDL_CONSTR];
+ }
+ else
+ {
+ enerd->term[F_DVDL] += enerd->term[F_DVDL_CONSTR];
+ }
+ enerd->term[F_DVDL_CONSTR] = 0;
+
+ for (i = 0; i < fepvals->n_lambda; i++)
+ {
+ /* note we are iterating over fepvals here!
+ For the current lam, dlam = 0 automatically,
+ so we don't need to add anything to the
+ enerd->enerpart_lambda[0] */
+
+ /* we don't need to worry about dvdl_lin contributions to dE at
+ current lambda, because the contributions to the current
+ lambda are automatically zeroed */
+
+ for (j = 0; j < efptNR; j++)
+ {
+ /* Note that this loop is over all dhdl components, not just the separated ones */
+ dlam = (fepvals->all_lambda[j][i]-lambda[j]);
+ enerd->enerpart_lambda[i+1] += dlam*enerd->dvdl_lin[j];
+ if (debug)
+ {
+ fprintf(debug, "enerdiff lam %g: (%15s), non-linear %f linear %f*%f\n",
+ fepvals->all_lambda[j][i], efpt_names[j],
+ (enerd->enerpart_lambda[i+1] - enerd->enerpart_lambda[0]),
+ dlam, enerd->dvdl_lin[j]);
+ }
+ }
+ }
+}
+
+
+void reset_foreign_enerdata(gmx_enerdata_t *enerd)
+{
+ int i, j;
+
+ /* First reset all foreign energy components. Foreign energies always called on
+ neighbor search steps */
+ for (i = 0; (i < egNR); i++)
+ {
+ for (j = 0; (j < enerd->grpp.nener); j++)
+ {
+ enerd->foreign_grpp.ener[i][j] = 0.0;
+ }
+ }
+
+ /* potential energy components */
+ for (i = 0; (i <= F_EPOT); i++)
+ {
+ enerd->foreign_term[i] = 0.0;
+ }
+}
+
+void reset_enerdata(gmx_enerdata_t *enerd)
+{
+ int i, j;
+
+ /* First reset all energy components. */
+ for (i = 0; (i < egNR); i++)
+ {
+ for (j = 0; (j < enerd->grpp.nener); j++)
+ {
+ enerd->grpp.ener[i][j] = 0.0;
+ }
+ }
+ for (i = 0; i < efptNR; i++)
+ {
+ enerd->dvdl_lin[i] = 0.0;
+ enerd->dvdl_nonlin[i] = 0.0;
+ }
+
+ /* Normal potential energy components */
+ for (i = 0; (i <= F_EPOT); i++)
+ {
+ enerd->term[i] = 0.0;
+ }
+ enerd->term[F_DVDL] = 0.0;
+ enerd->term[F_DVDL_COUL] = 0.0;
+ enerd->term[F_DVDL_VDW] = 0.0;
+ enerd->term[F_DVDL_BONDED] = 0.0;
+ enerd->term[F_DVDL_RESTRAINT] = 0.0;
+ enerd->term[F_DKDL] = 0.0;
+ if (enerd->n_lambda > 0)
+ {
+ for (i = 0; i < enerd->n_lambda; i++)
+ {
+ enerd->enerpart_lambda[i] = 0.0;
+ }
+ }
+ /* reset foreign energy data - separate function since we also call it elsewhere */
+ reset_foreign_enerdata(enerd);
+}
diff --git a/patches/gromacs-2016-beta1.diff/src/gromacs/mdlib/minimize.cpp b/patches/gromacs-2016-beta1.diff/src/gromacs/mdlib/minimize.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..3ce7832812fc7b72d446ab0c9985e23832d66fda
--- /dev/null
+++ b/patches/gromacs-2016-beta1.diff/src/gromacs/mdlib/minimize.cpp
@@ -0,0 +1,3128 @@
+/*
+ * This file is part of the GROMACS molecular simulation package.
+ *
+ * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
+ * Copyright (c) 2001-2004, The GROMACS development team.
+ * Copyright (c) 2013,2014,2015,2016, by the GROMACS development team, led by
+ * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
+ * and including many others, as listed in the AUTHORS file in the
+ * top-level source directory and at http://www.gromacs.org.
+ *
+ * GROMACS is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public License
+ * as published by the Free Software Foundation; either version 2.1
+ * of the License, or (at your option) any later version.
+ *
+ * GROMACS is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with GROMACS; if not, see
+ * http://www.gnu.org/licenses, or write to the Free Software Foundation,
+ * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ * If you want to redistribute modifications to GROMACS, 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 http://www.gromacs.org.
+ *
+ * To help us fund GROMACS development, we humbly ask that you cite
+ * the research papers on the package. Check out http://www.gromacs.org.
+ */
+/*! \internal \file
+ *
+ * \brief This file defines integrators for energy minimization
+ *
+ * \author Berk Hess <hess@kth.se>
+ * \author Erik Lindahl <erik@kth.se>
+ * \ingroup module_mdlib
+ */
+#include "gmxpre.h"
+
+#include "minimize.h"
+
+#include "config.h"
+
+#include <cmath>
+#include <cstring>
+#include <ctime>
+
+#include <algorithm>
+#include <vector>
+
+#include "gromacs/commandline/filenm.h"
+#include "gromacs/domdec/domdec.h"
+#include "gromacs/domdec/domdec_struct.h"
+#include "gromacs/ewald/pme.h"
+#include "gromacs/fileio/confio.h"
+#include "gromacs/fileio/mtxio.h"
+#include "gromacs/gmxlib/md_logging.h"
+#include "gromacs/gmxlib/network.h"
+#include "gromacs/gmxlib/nrnb.h"
+#include "gromacs/imd/imd.h"
+#include "gromacs/linearalgebra/sparsematrix.h"
+#include "gromacs/listed-forces/manage-threading.h"
+#include "gromacs/math/functions.h"
+#include "gromacs/math/vec.h"
+#include "gromacs/mdlib/constr.h"
+#include "gromacs/mdlib/force.h"
+#include "gromacs/mdlib/forcerec.h"
+#include "gromacs/mdlib/gmx_omp_nthreads.h"
+#include "gromacs/mdlib/md_support.h"
+#include "gromacs/mdlib/mdatoms.h"
+#include "gromacs/mdlib/mdebin.h"
+#include "gromacs/mdlib/mdrun.h"
+#include "gromacs/mdlib/ns.h"
+#include "gromacs/mdlib/shellfc.h"
+#include "gromacs/mdlib/sim_util.h"
+#include "gromacs/mdlib/tgroup.h"
+#include "gromacs/mdlib/trajectory_writing.h"
+/* PLUMED */
+#include "../../../Plumed.h"
+extern int plumedswitch;
+extern plumed plumedmain;
+extern void(*plumedcmd)(plumed,const char*,const void*);
+/* END PLUMED */
+
+#include "gromacs/mdlib/update.h"
+#include "gromacs/mdlib/vsite.h"
+#include "gromacs/mdtypes/commrec.h"
+#include "gromacs/mdtypes/inputrec.h"
+#include "gromacs/mdtypes/md_enums.h"
+#include "gromacs/pbcutil/mshift.h"
+#include "gromacs/pbcutil/pbc.h"
+#include "gromacs/timing/wallcycle.h"
+#include "gromacs/timing/walltime_accounting.h"
+#include "gromacs/topology/mtop_util.h"
+#include "gromacs/utility/cstringutil.h"
+#include "gromacs/utility/exceptions.h"
+#include "gromacs/utility/fatalerror.h"
+#include "gromacs/utility/smalloc.h"
+
+//! Utility structure for manipulating states during EM
+typedef struct {
+ //! Copy of the global state
+ t_state s;
+ //! Force array
+ rvec *f;
+ //! Potential energy
+ real epot;
+ //! Norm of the force
+ real fnorm;
+ //! Maximum force
+ real fmax;
+ //! Direction
+ int a_fmax;
+} em_state_t;
+
+//! Initiate em_state_t structure and return pointer to it
+static em_state_t *init_em_state()
+{
+ em_state_t *ems;
+
+ snew(ems, 1);
+
+ /* does this need to be here? Should the array be declared differently (staticaly)in the state definition? */
+ snew(ems->s.lambda, efptNR);
+
+ return ems;
+}
+
+//! Print the EM starting conditions
+static void print_em_start(FILE *fplog,
+ t_commrec *cr,
+ gmx_walltime_accounting_t walltime_accounting,
+ gmx_wallcycle_t wcycle,
+ const char *name)
+{
+ walltime_accounting_start(walltime_accounting);
+ wallcycle_start(wcycle, ewcRUN);
+ print_start(fplog, cr, walltime_accounting, name);
+}
+
+//! Stop counting time for EM
+static void em_time_end(gmx_walltime_accounting_t walltime_accounting,
+ gmx_wallcycle_t wcycle)
+{
+ wallcycle_stop(wcycle, ewcRUN);
+
+ walltime_accounting_end(walltime_accounting);
+}
+
+//! Printing a log file and console header
+static void sp_header(FILE *out, const char *minimizer, real ftol, int nsteps)
+{
+ fprintf(out, "\n");
+ fprintf(out, "%s:\n", minimizer);
+ fprintf(out, " Tolerance (Fmax) = %12.5e\n", ftol);
+ fprintf(out, " Number of steps = %12d\n", nsteps);
+}
+
+//! Print warning message
+static void warn_step(FILE *fp, real ftol, gmx_bool bLastStep, gmx_bool bConstrain)
+{
+ char buffer[2048];
+ if (bLastStep)
+ {
+ sprintf(buffer,
+ "\nEnergy minimization reached the maximum number "
+ "of steps before the forces reached the requested "
+ "precision Fmax < %g.\n", ftol);
+ }
+ else
+ {
+ sprintf(buffer,
+ "\nEnergy minimization has stopped, but the forces have "
+ "not converged to the requested precision Fmax < %g (which "
+ "may not be possible for your system). It stopped "
+ "because the algorithm tried to make a new step whose size "
+ "was too small, or there was no change in the energy since "
+ "last step. Either way, we regard the minimization as "
+ "converged to within the available machine precision, "
+ "given your starting configuration and EM parameters.\n%s%s",
+ ftol,
+ sizeof(real) < sizeof(double) ?
+ "\nDouble precision normally gives you higher accuracy, but "
+ "this is often not needed for preparing to run molecular "
+ "dynamics.\n" :
+ "",
+ bConstrain ?
+ "You might need to increase your constraint accuracy, or turn\n"
+ "off constraints altogether (set constraints = none in mdp file)\n" :
+ "");
+ }
+ fputs(wrap_lines(buffer, 78, 0, FALSE), fp);
+}
+
+//! Print message about convergence of the EM
+static void print_converged(FILE *fp, const char *alg, real ftol,
+ gmx_int64_t count, gmx_bool bDone, gmx_int64_t nsteps,
+ real epot, real fmax, int nfmax, real fnorm)
+{
+ char buf[STEPSTRSIZE];
+
+ if (bDone)
+ {
+ fprintf(fp, "\n%s converged to Fmax < %g in %s steps\n",
+ alg, ftol, gmx_step_str(count, buf));
+ }
+ else if (count < nsteps)
+ {
+ fprintf(fp, "\n%s converged to machine precision in %s steps,\n"
+ "but did not reach the requested Fmax < %g.\n",
+ alg, gmx_step_str(count, buf), ftol);
+ }
+ else
+ {
+ fprintf(fp, "\n%s did not converge to Fmax < %g in %s steps.\n",
+ alg, ftol, gmx_step_str(count, buf));
+ }
+
+#if GMX_DOUBLE
+ fprintf(fp, "Potential Energy = %21.14e\n", epot);
+ fprintf(fp, "Maximum force = %21.14e on atom %d\n", fmax, nfmax+1);
+ fprintf(fp, "Norm of force = %21.14e\n", fnorm);
+#else
+ fprintf(fp, "Potential Energy = %14.7e\n", epot);
+ fprintf(fp, "Maximum force = %14.7e on atom %d\n", fmax, nfmax+1);
+ fprintf(fp, "Norm of force = %14.7e\n", fnorm);
+#endif
+}
+
+//! Compute the norm and max of the force array in parallel
+static void get_f_norm_max(t_commrec *cr,
+ t_grpopts *opts, t_mdatoms *mdatoms, rvec *f,
+ real *fnorm, real *fmax, int *a_fmax)
+{
+ double fnorm2, *sum;
+ real fmax2, fam;
+ int la_max, a_max, start, end, i, m, gf;
+
+ /* This routine finds the largest force and returns it.
+ * On parallel machines the global max is taken.
+ */
+ fnorm2 = 0;
+ fmax2 = 0;
+ la_max = -1;
+ start = 0;
+ end = mdatoms->homenr;
+ if (mdatoms->cFREEZE)
+ {
+ for (i = start; i < end; i++)
+ {
+ gf = mdatoms->cFREEZE[i];
+ fam = 0;
+ for (m = 0; m < DIM; m++)
+ {
+ if (!opts->nFreeze[gf][m])
+ {
+ fam += gmx::square(f[i][m]);
+ }
+ }
+ fnorm2 += fam;
+ if (fam > fmax2)
+ {
+ fmax2 = fam;
+ la_max = i;
+ }
+ }
+ }
+ else
+ {
+ for (i = start; i < end; i++)
+ {
+ fam = norm2(f[i]);
+ fnorm2 += fam;
+ if (fam > fmax2)
+ {
+ fmax2 = fam;
+ la_max = i;
+ }
+ }
+ }
+
+ if (la_max >= 0 && DOMAINDECOMP(cr))
+ {
+ a_max = cr->dd->gatindex[la_max];
+ }
+ else
+ {
+ a_max = la_max;
+ }
+ if (PAR(cr))
+ {
+ snew(sum, 2*cr->nnodes+1);
+ sum[2*cr->nodeid] = fmax2;
+ sum[2*cr->nodeid+1] = a_max;
+ sum[2*cr->nnodes] = fnorm2;
+ gmx_sumd(2*cr->nnodes+1, sum, cr);
+ fnorm2 = sum[2*cr->nnodes];
+ /* Determine the global maximum */
+ for (i = 0; i < cr->nnodes; i++)
+ {
+ if (sum[2*i] > fmax2)
+ {
+ fmax2 = sum[2*i];
+ a_max = (int)(sum[2*i+1] + 0.5);
+ }
+ }
+ sfree(sum);
+ }
+
+ if (fnorm)
+ {
+ *fnorm = sqrt(fnorm2);
+ }
+ if (fmax)
+ {
+ *fmax = sqrt(fmax2);
+ }
+ if (a_fmax)
+ {
+ *a_fmax = a_max;
+ }
+}
+
+//! Compute the norm of the force
+static void get_state_f_norm_max(t_commrec *cr,
+ t_grpopts *opts, t_mdatoms *mdatoms,
+ em_state_t *ems)
+{
+ get_f_norm_max(cr, opts, mdatoms, ems->f, &ems->fnorm, &ems->fmax, &ems->a_fmax);
+}
+
+//! Initialize the energy minimization
+void init_em(FILE *fplog, const char *title,
+ t_commrec *cr, t_inputrec *ir,
+ t_state *state_global, gmx_mtop_t *top_global,
+ em_state_t *ems, gmx_localtop_t **top,
+ rvec **f,
+ t_nrnb *nrnb, rvec mu_tot,
+ t_forcerec *fr, gmx_enerdata_t **enerd,
+ t_graph **graph, t_mdatoms *mdatoms, gmx_global_stat_t *gstat,
+ gmx_vsite_t *vsite, gmx_constr_t constr,
+ int nfile, const t_filenm fnm[],
+ gmx_mdoutf_t *outf, t_mdebin **mdebin,
+ int imdport, unsigned long gmx_unused Flags,
+ gmx_wallcycle_t wcycle)
+{
+ int i;
+ real dvdl_constr;
+
+ if (fplog)
+ {
+ fprintf(fplog, "Initiating %s\n", title);
+ }
+
+ state_global->ngtc = 0;
+
+ /* Initialize lambda variables */
+ initialize_lambdas(fplog, ir, &(state_global->fep_state), state_global->lambda, NULL);
+
+ init_nrnb(nrnb);
+
+ /* Interactive molecular dynamics */
+ init_IMD(ir, cr, top_global, fplog, 1, state_global->x,
+ nfile, fnm, NULL, imdport, Flags);
+
+ if (DOMAINDECOMP(cr))
+ {
+ *top = dd_init_local_top(top_global);
+
+ dd_init_local_state(cr->dd, state_global, &ems->s);
+
+ *f = NULL;
+
+ /* 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,
+ &ems->s, &ems->f, mdatoms, *top,
+ fr, vsite, constr,
+ nrnb, NULL, FALSE);
+ dd_store_state(cr->dd, &ems->s);
+
+ *graph = NULL;
+ }
+ else
+ {
+ snew(*f, top_global->natoms);
+
+ /* Just copy the state */
+ ems->s = *state_global;
+ /* We need to allocate one element extra, since we might use
+ * (unaligned) 4-wide SIMD loads to access rvec entries.
+ */
+ snew(ems->s.x, ems->s.nalloc + 1);
+ snew(ems->f, ems->s.nalloc+1);
+ snew(ems->s.v, ems->s.nalloc+1);
+ for (i = 0; i < state_global->natoms; i++)
+ {
+ copy_rvec(state_global->x[i], ems->s.x[i]);
+ }
+ copy_mat(state_global->box, ems->s.box);
+
+ *top = gmx_mtop_generate_local_top(top_global, ir->efep != efepNO);
+
+ forcerec_set_excl_load(fr, *top);
+
+ setup_bonded_threading(fr, &(*top)->idef);
+
+ if (ir->ePBC != epbcNONE && !fr->bMolPBC)
+ {
+ *graph = mk_graph(fplog, &((*top)->idef), 0, top_global->natoms, FALSE, FALSE);
+ }
+ else
+ {
+ *graph = NULL;
+ }
+
+ atoms2md(top_global, ir, 0, NULL, top_global->natoms, mdatoms);
+ update_mdatoms(mdatoms, state_global->lambda[efptFEP]);
+
+ if (vsite)
+ {
+ set_vsite_top(vsite, *top, mdatoms, cr);
+ }
+ }
+
+ if (constr)
+ {
+ if (ir->eConstrAlg == econtSHAKE &&
+ gmx_mtop_ftype_count(top_global, F_CONSTR) > 0)
+ {
+ gmx_fatal(FARGS, "Can not do energy minimization with %s, use %s\n",
+ econstr_names[econtSHAKE], econstr_names[econtLINCS]);
+ }
+
+ if (!DOMAINDECOMP(cr))
+ {
+ set_constraints(constr, *top, ir, mdatoms, cr);
+ }
+
+ if (!ir->bContinuation)
+ {
+ /* Constrain the starting coordinates */
+ dvdl_constr = 0;
+ constrain(PAR(cr) ? NULL : fplog, TRUE, TRUE, constr, &(*top)->idef,
+ ir, cr, -1, 0, 1.0, mdatoms,
+ ems->s.x, ems->s.x, NULL, fr->bMolPBC, ems->s.box,
+ ems->s.lambda[efptFEP], &dvdl_constr,
+ NULL, NULL, nrnb, econqCoord);
+ }
+ }
+
+ if (PAR(cr))
+ {
+ *gstat = global_stat_init(ir);
+ }
+ else
+ {
+ *gstat = NULL;
+ }
+
+ *outf = init_mdoutf(fplog, nfile, fnm, 0, cr, ir, top_global, NULL, wcycle);
+
+ snew(*enerd, 1);
+ init_enerdata(top_global->groups.grps[egcENER].nr, ir->fepvals->n_lambda,
+ *enerd);
+
+ if (mdebin != NULL)
+ {
+ /* Init bin for energy stuff */
+ *mdebin = init_mdebin(mdoutf_get_fp_ene(*outf), top_global, ir, NULL);
+ }
+
+ clear_rvec(mu_tot);
+ calc_shifts(ems->s.box, fr->shift_vec);
+
+ /* PLUMED */
+ if(plumedswitch){
+ if(cr->ms && cr->ms->nsim>1) {
+ if(MASTER(cr)) (*plumedcmd) (plumedmain,"GREX setMPIIntercomm",&cr->ms->mpi_comm_masters);
+ if(PAR(cr)){
+ if(DOMAINDECOMP(cr)) {
+ (*plumedcmd) (plumedmain,"GREX setMPIIntracomm",&cr->dd->mpi_comm_all);
+ }else{
+ (*plumedcmd) (plumedmain,"GREX setMPIIntracomm",&cr->mpi_comm_mysim);
+ }
+ }
+ (*plumedcmd) (plumedmain,"GREX init",NULL);
+ }
+ if(PAR(cr)){
+ if(DOMAINDECOMP(cr)) {
+ (*plumedcmd) (plumedmain,"setMPIComm",&cr->dd->mpi_comm_all);
+ }else{
+ (*plumedcmd) (plumedmain,"setMPIComm",&cr->mpi_comm_mysim);
+ }
+ }
+ (*plumedcmd) (plumedmain,"setNatoms",&top_global->natoms);
+ (*plumedcmd) (plumedmain,"setMDEngine","gromacs");
+ (*plumedcmd) (plumedmain,"setLog",fplog);
+ real real_delta_t;
+ real_delta_t=ir->delta_t;
+ (*plumedcmd) (plumedmain,"setTimestep",&real_delta_t);
+ (*plumedcmd) (plumedmain,"init",NULL);
+
+ if(PAR(cr)){
+ if(DOMAINDECOMP(cr)) {
+ (*plumedcmd) (plumedmain,"setAtomsNlocal",&cr->dd->nat_home);
+ (*plumedcmd) (plumedmain,"setAtomsGatindex",cr->dd->gatindex);
+ }
+ }
+ }
+ /* END PLUMED */
+}
+
+//! Finalize the minimization
+static void finish_em(t_commrec *cr, gmx_mdoutf_t outf,
+ gmx_walltime_accounting_t walltime_accounting,
+ gmx_wallcycle_t wcycle)
+{
+ if (!(cr->duty & DUTY_PME))
+ {
+ /* Tell the PME only node to finish */
+ gmx_pme_send_finish(cr);
+ }
+
+ done_mdoutf(outf);
+
+ em_time_end(walltime_accounting, wcycle);
+}
+
+//! Swap two different EM states during minimization
+static void swap_em_state(em_state_t *ems1, em_state_t *ems2)
+{
+ em_state_t tmp;
+
+ tmp = *ems1;
+ *ems1 = *ems2;
+ *ems2 = tmp;
+}
+
+//! Copy coordinate from an EM state to a "normal" state structure
+static void copy_em_coords(em_state_t *ems, t_state *state)
+{
+ int i;
+
+ for (i = 0; (i < state->natoms); i++)
+ {
+ copy_rvec(ems->s.x[i], state->x[i]);
+ }
+}
+
+//! Save the EM trajectory
+static void write_em_traj(FILE *fplog, t_commrec *cr,
+ gmx_mdoutf_t outf,
+ gmx_bool bX, gmx_bool bF, const char *confout,
+ gmx_mtop_t *top_global,
+ t_inputrec *ir, gmx_int64_t step,
+ em_state_t *state,
+ t_state *state_global)
+{
+ int mdof_flags;
+ gmx_bool bIMDout = FALSE;
+
+
+ /* Shall we do IMD output? */
+ if (ir->bIMD)
+ {
+ bIMDout = do_per_step(step, IMD_get_step(ir->imd->setup));
+ }
+
+ if ((bX || bF || bIMDout || confout != NULL) && !DOMAINDECOMP(cr))
+ {
+ copy_em_coords(state, state_global);
+ }
+
+ mdof_flags = 0;
+ if (bX)
+ {
+ mdof_flags |= MDOF_X;
+ }
+ if (bF)
+ {
+ mdof_flags |= MDOF_F;
+ }
+
+ /* If we want IMD output, set appropriate MDOF flag */
+ if (ir->bIMD)
+ {
+ mdof_flags |= MDOF_IMD;
+ }
+
+ mdoutf_write_to_trajectory_files(fplog, cr, outf, mdof_flags,
+ top_global, step, (double)step,
+ &state->s, state_global, state->f);
+
+ if (confout != NULL && MASTER(cr))
+ {
+ if (ir->ePBC != epbcNONE && !ir->bPeriodicMols && DOMAINDECOMP(cr))
+ {
+ /* Make molecules whole only for confout writing */
+ do_pbc_mtop(fplog, ir->ePBC, state_global->box, top_global,
+ state_global->x);
+ }
+
+ write_sto_conf_mtop(confout,
+ *top_global->name, top_global,
+ state_global->x, NULL, ir->ePBC, state_global->box);
+ }
+}
+
+//! Do one minimization step
+static void do_em_step(t_commrec *cr, t_inputrec *ir, t_mdatoms *md,
+ gmx_bool bMolPBC,
+ em_state_t *ems1, real a, rvec *f, em_state_t *ems2,
+ gmx_constr_t constr, gmx_localtop_t *top,
+ t_nrnb *nrnb, gmx_wallcycle_t wcycle,
+ gmx_int64_t count)
+
+{
+ t_state *s1, *s2;
+ int i;
+ int start, end;
+ rvec *x1, *x2;
+ real dvdl_constr;
+ int nthreads gmx_unused;
+
+ s1 = &ems1->s;
+ s2 = &ems2->s;
+
+ if (DOMAINDECOMP(cr) && s1->ddp_count != cr->dd->ddp_count)
+ {
+ gmx_incons("state mismatch in do_em_step");
+ }
+
+ s2->flags = s1->flags;
+
+ if (s2->nalloc != s1->nalloc)
+ {
+ s2->nalloc = s1->nalloc;
+ /* We need to allocate one element extra, since we might use
+ * (unaligned) 4-wide SIMD loads to access rvec entries.
+ */
+ srenew(s2->x, s1->nalloc + 1);
+ srenew(ems2->f, s1->nalloc);
+ if (s2->flags & (1<<estCGP))
+ {
+ srenew(s2->cg_p, s1->nalloc + 1);
+ }
+ }
+
+ s2->natoms = s1->natoms;
+ copy_mat(s1->box, s2->box);
+ /* Copy free energy state */
+ for (i = 0; i < efptNR; i++)
+ {
+ s2->lambda[i] = s1->lambda[i];
+ }
+ copy_mat(s1->box, s2->box);
+
+ start = 0;
+ end = md->homenr;
+
+ x1 = s1->x;
+ x2 = s2->x;
+
+ // cppcheck-suppress unreadVariable
+ nthreads = gmx_omp_nthreads_get(emntUpdate);
+#pragma omp parallel num_threads(nthreads)
+ {
+ int gf, i, m;
+
+ gf = 0;
+#pragma omp for schedule(static) nowait
+ for (i = start; i < end; i++)
+ {
+ try
+ {
+ if (md->cFREEZE)
+ {
+ gf = md->cFREEZE[i];
+ }
+ for (m = 0; m < DIM; m++)
+ {
+ if (ir->opts.nFreeze[gf][m])
+ {
+ x2[i][m] = x1[i][m];
+ }
+ else
+ {
+ x2[i][m] = x1[i][m] + a*f[i][m];
+ }
+ }
+ }
+ GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+ }
+
+ if (s2->flags & (1<<estCGP))
+ {
+ /* Copy the CG p vector */
+ x1 = s1->cg_p;
+ x2 = s2->cg_p;
+#pragma omp for schedule(static) nowait
+ for (i = start; i < end; i++)
+ {
+ // Trivial OpenMP block that does not throw
+ copy_rvec(x1[i], x2[i]);
+ }
+ }
+
+ if (DOMAINDECOMP(cr))
+ {
+ s2->ddp_count = s1->ddp_count;
+ if (s2->cg_gl_nalloc < s1->cg_gl_nalloc)
+ {
+#pragma omp barrier
+ s2->cg_gl_nalloc = s1->cg_gl_nalloc;
+ try
+ {
+ /* We need to allocate one element extra, since we might use
+ * (unaligned) 4-wide SIMD loads to access rvec entries.
+ */
+ srenew(s2->cg_gl, s2->cg_gl_nalloc + 1);
+ }
+ GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+#pragma omp barrier
+ }
+ s2->ncg_gl = s1->ncg_gl;
+#pragma omp for schedule(static) nowait
+ for (i = 0; i < s2->ncg_gl; i++)
+ {
+ s2->cg_gl[i] = s1->cg_gl[i];
+ }
+ s2->ddp_count_cg_gl = s1->ddp_count_cg_gl;
+ }
+ }
+
+ if (constr)
+ {
+ wallcycle_start(wcycle, ewcCONSTR);
+ dvdl_constr = 0;
+ constrain(NULL, TRUE, TRUE, constr, &top->idef,
+ ir, cr, count, 0, 1.0, md,
+ s1->x, s2->x, NULL, bMolPBC, s2->box,
+ s2->lambda[efptBONDED], &dvdl_constr,
+ NULL, NULL, nrnb, econqCoord);
+ wallcycle_stop(wcycle, ewcCONSTR);
+ }
+}
+
+//! Prepare EM for using domain decomposition parallellization
+static void em_dd_partition_system(FILE *fplog, int step, t_commrec *cr,
+ gmx_mtop_t *top_global, t_inputrec *ir,
+ em_state_t *ems, gmx_localtop_t *top,
+ t_mdatoms *mdatoms, t_forcerec *fr,
+ gmx_vsite_t *vsite, gmx_constr_t constr,
+ t_nrnb *nrnb, gmx_wallcycle_t wcycle)
+{
+ /* Repartition the domain decomposition */
+ dd_partition_system(fplog, step, cr, FALSE, 1,
+ NULL, top_global, ir,
+ &ems->s, &ems->f,
+ mdatoms, top, fr, vsite, constr,
+ nrnb, wcycle, FALSE);
+ dd_store_state(cr->dd, &ems->s);
+}
+
+//! De one energy evaluation
+static void evaluate_energy(FILE *fplog, t_commrec *cr,
+ gmx_mtop_t *top_global,
+ em_state_t *ems, gmx_localtop_t *top,
+ t_inputrec *inputrec,
+ t_nrnb *nrnb, gmx_wallcycle_t wcycle,
+ gmx_global_stat_t gstat,
+ gmx_vsite_t *vsite, gmx_constr_t constr,
+ t_fcdata *fcd,
+ t_graph *graph, t_mdatoms *mdatoms,
+ t_forcerec *fr, rvec mu_tot,
+ gmx_enerdata_t *enerd, tensor vir, tensor pres,
+ gmx_int64_t count, gmx_bool bFirst)
+{
+ real t;
+ gmx_bool bNS;
+ tensor force_vir, shake_vir, ekin;
+ real dvdl_constr, prescorr, enercorr, dvdlcorr;
+ real terminate = 0;
+
+ /* Set the time to the initial time, the time does not change during EM */
+ t = inputrec->init_t;
+
+ if (bFirst ||
+ (DOMAINDECOMP(cr) && ems->s.ddp_count < cr->dd->ddp_count))
+ {
+ /* This is the first state or an old state used before the last ns */
+ bNS = TRUE;
+ }
+ else
+ {
+ bNS = FALSE;
+ if (inputrec->nstlist > 0)
+ {
+ bNS = TRUE;
+ }
+ }
+
+ if (vsite)
+ {
+ construct_vsites(vsite, ems->s.x, 1, NULL,
+ top->idef.iparams, top->idef.il,
+ fr->ePBC, fr->bMolPBC, cr, ems->s.box);
+ }
+
+ if (DOMAINDECOMP(cr) && bNS)
+ {
+ /* Repartition the domain decomposition */
+ em_dd_partition_system(fplog, count, cr, top_global, inputrec,
+ ems, top, mdatoms, fr, vsite, constr,
+ nrnb, wcycle);
+ }
+
+ /* Calc force & energy on new trial position */
+ /* do_force always puts the charge groups in the box and shifts again
+ * We do not unshift, so molecules are always whole in congrad.c
+ */
+ /* PLUMED */
+ int plumedNeedsEnergy=0;
+ matrix plumed_vir;
+ if(plumedswitch){
+ long int lstep=count; (*plumedcmd)(plumedmain,"setStepLong",&lstep);
+ (*plumedcmd) (plumedmain,"setPositions",&ems->s.x[0][0]);
+ (*plumedcmd) (plumedmain,"setMasses",&mdatoms->massT[0]);
+ (*plumedcmd) (plumedmain,"setCharges",&mdatoms->chargeA[0]);
+ (*plumedcmd) (plumedmain,"setBox",&ems->s.box[0][0]);
+ (*plumedcmd) (plumedmain,"prepareCalc",NULL);
+ (*plumedcmd) (plumedmain,"setForces",&ems->f[0][0]);
+ (*plumedcmd) (plumedmain,"isEnergyNeeded",&plumedNeedsEnergy);
+ clear_mat(plumed_vir);
+ (*plumedcmd) (plumedmain,"setVirial",&plumed_vir[0][0]);
+ }
+ /* END PLUMED */
+
+ do_force(fplog, cr, inputrec,
+ count, nrnb, wcycle, top, &top_global->groups,
+ ems->s.box, ems->s.x, &ems->s.hist,
+ ems->f, force_vir, mdatoms, enerd, fcd,
+ ems->s.lambda, graph, fr, vsite, mu_tot, t, NULL, NULL, TRUE,
+ GMX_FORCE_STATECHANGED | GMX_FORCE_ALLFORCES |
+ GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY |
+ (bNS ? GMX_FORCE_NS : 0));
+ /* PLUMED */
+ if(plumedswitch){
+ if(plumedNeedsEnergy) {
+ msmul(force_vir,2.0,plumed_vir);
+ (*plumedcmd) (plumedmain,"setEnergy",&enerd->term[F_EPOT]);
+ (*plumedcmd) (plumedmain,"performCalc",NULL);
+ msmul(plumed_vir,0.5,force_vir);
+ } else {
+ msmul(plumed_vir,0.5,plumed_vir);
+ m_add(force_vir,plumed_vir,force_vir);
+ }
+ }
+ /* END PLUMED */
+
+ /* Clear the unused shake virial and pressure */
+ clear_mat(shake_vir);
+ clear_mat(pres);
+
+ /* Communicate stuff when parallel */
+ if (PAR(cr) && inputrec->eI != eiNM)
+ {
+ wallcycle_start(wcycle, ewcMoveE);
+
+ global_stat(gstat, cr, enerd, force_vir, shake_vir, mu_tot,
+ inputrec, NULL, NULL, NULL, 1, &terminate,
+ NULL, FALSE,
+ CGLO_ENERGY |
+ CGLO_PRESSURE |
+ CGLO_CONSTRAINT);
+
+ wallcycle_stop(wcycle, ewcMoveE);
+ }
+
+ /* Calculate long range corrections to pressure and energy */
+ calc_dispcorr(inputrec, fr, ems->s.box, ems->s.lambda[efptVDW],
+ pres, force_vir, &prescorr, &enercorr, &dvdlcorr);
+ enerd->term[F_DISPCORR] = enercorr;
+ enerd->term[F_EPOT] += enercorr;
+ enerd->term[F_PRES] += prescorr;
+ enerd->term[F_DVDL] += dvdlcorr;
+
+ ems->epot = enerd->term[F_EPOT];
+
+ if (constr)
+ {
+ /* Project out the constraint components of the force */
+ wallcycle_start(wcycle, ewcCONSTR);
+ dvdl_constr = 0;
+ constrain(NULL, FALSE, FALSE, constr, &top->idef,
+ inputrec, cr, count, 0, 1.0, mdatoms,
+ ems->s.x, ems->f, ems->f, fr->bMolPBC, ems->s.box,
+ ems->s.lambda[efptBONDED], &dvdl_constr,
+ NULL, &shake_vir, nrnb, econqForceDispl);
+ enerd->term[F_DVDL_CONSTR] += dvdl_constr;
+ m_add(force_vir, shake_vir, vir);
+ wallcycle_stop(wcycle, ewcCONSTR);
+ }
+ else
+ {
+ copy_mat(force_vir, vir);
+ }
+
+ clear_mat(ekin);
+ enerd->term[F_PRES] =
+ calc_pres(fr->ePBC, inputrec->nwall, ems->s.box, ekin, vir, pres);
+
+ sum_dhdl(enerd, ems->s.lambda, inputrec->fepvals);
+
+ if (EI_ENERGY_MINIMIZATION(inputrec->eI))
+ {
+ get_state_f_norm_max(cr, &(inputrec->opts), mdatoms, ems);
+ }
+}
+
+//! Parallel utility summing energies and forces
+static double reorder_partsum(t_commrec *cr, t_grpopts *opts, t_mdatoms *mdatoms,
+ gmx_mtop_t *top_global,
+ em_state_t *s_min, em_state_t *s_b)
+{
+ rvec *fm, *fb, *fmg;
+ t_block *cgs_gl;
+ int ncg, *cg_gl, *index, c, cg, i, a0, a1, a, gf, m;
+ double partsum;
+ unsigned char *grpnrFREEZE;
+
+ if (debug)
+ {
+ fprintf(debug, "Doing reorder_partsum\n");
+ }
+
+ fm = s_min->f;
+ fb = s_b->f;
+
+ cgs_gl = dd_charge_groups_global(cr->dd);
+ index = cgs_gl->index;
+
+ /* Collect fm in a global vector fmg.
+ * This conflicts with the spirit of domain decomposition,
+ * but to fully optimize this a much more complicated algorithm is required.
+ */
+ snew(fmg, top_global->natoms);
+
+ ncg = s_min->s.ncg_gl;
+ cg_gl = s_min->s.cg_gl;
+ i = 0;
+ for (c = 0; c < ncg; c++)
+ {
+ cg = cg_gl[c];
+ a0 = index[cg];
+ a1 = index[cg+1];
+ for (a = a0; a < a1; a++)
+ {
+ copy_rvec(fm[i], fmg[a]);
+ i++;
+ }
+ }
+ gmx_sum(top_global->natoms*3, fmg[0], cr);
+
+ /* Now we will determine the part of the sum for the cgs in state s_b */
+ ncg = s_b->s.ncg_gl;
+ cg_gl = s_b->s.cg_gl;
+ partsum = 0;
+ i = 0;
+ gf = 0;
+ grpnrFREEZE = top_global->groups.grpnr[egcFREEZE];
+ for (c = 0; c < ncg; c++)
+ {
+ cg = cg_gl[c];
+ a0 = index[cg];
+ a1 = index[cg+1];
+ for (a = a0; a < a1; a++)
+ {
+ if (mdatoms->cFREEZE && grpnrFREEZE)
+ {
+ gf = grpnrFREEZE[i];
+ }
+ for (m = 0; m < DIM; m++)
+ {
+ if (!opts->nFreeze[gf][m])
+ {
+ partsum += (fb[i][m] - fmg[a][m])*fb[i][m];
+ }
+ }
+ i++;
+ }
+ }
+
+ sfree(fmg);
+
+ return partsum;
+}
+
+//! Print some stuff, like beta, whatever that means.
+static real pr_beta(t_commrec *cr, t_grpopts *opts, t_mdatoms *mdatoms,
+ gmx_mtop_t *top_global,
+ em_state_t *s_min, em_state_t *s_b)
+{
+ rvec *fm, *fb;
+ double sum;
+ int gf, i, m;
+
+ /* This is just the classical Polak-Ribiere calculation of beta;
+ * it looks a bit complicated since we take freeze groups into account,
+ * and might have to sum it in parallel runs.
+ */
+
+ if (!DOMAINDECOMP(cr) ||
+ (s_min->s.ddp_count == cr->dd->ddp_count &&
+ s_b->s.ddp_count == cr->dd->ddp_count))
+ {
+ fm = s_min->f;
+ fb = s_b->f;
+ sum = 0;
+ gf = 0;
+ /* This part of code can be incorrect with DD,
+ * since the atom ordering in s_b and s_min might differ.
+ */
+ for (i = 0; i < mdatoms->homenr; i++)
+ {
+ if (mdatoms->cFREEZE)
+ {
+ gf = mdatoms->cFREEZE[i];
+ }
+ for (m = 0; m < DIM; m++)
+ {
+ if (!opts->nFreeze[gf][m])
+ {
+ sum += (fb[i][m] - fm[i][m])*fb[i][m];
+ }
+ }
+ }
+ }
+ else
+ {
+ /* We need to reorder cgs while summing */
+ sum = reorder_partsum(cr, opts, mdatoms, top_global, s_min, s_b);
+ }
+ if (PAR(cr))
+ {
+ gmx_sumd(1, &sum, cr);
+ }
+
+ return sum/gmx::square(s_min->fnorm);
+}
+
+namespace gmx
+{
+
+/*! \brief Do conjugate gradients minimization
+ \copydoc integrator_t (FILE *fplog, t_commrec *cr,
+ int nfile, const t_filenm fnm[],
+ const gmx_output_env_t *oenv, gmx_bool bVerbose,
+ int nstglobalcomm,
+ gmx_vsite_t *vsite, gmx_constr_t constr,
+ int stepout,
+ t_inputrec *inputrec,
+ 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_nex, int repl_ex_seed,
+ real cpt_period, real max_hours,
+ int imdport,
+ unsigned long Flags,
+ gmx_walltime_accounting_t walltime_accounting)
+ */
+double do_cg(FILE *fplog, t_commrec *cr,
+ int nfile, const t_filenm fnm[],
+ const gmx_output_env_t gmx_unused *oenv, gmx_bool bVerbose,
+ int gmx_unused nstglobalcomm,
+ gmx_vsite_t *vsite, gmx_constr_t constr,
+ int gmx_unused stepout,
+ t_inputrec *inputrec,
+ 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 gmx_unused ed,
+ t_forcerec *fr,
+ int gmx_unused repl_ex_nst, int gmx_unused repl_ex_nex, int gmx_unused repl_ex_seed,
+ real gmx_unused cpt_period, real gmx_unused max_hours,
+ int imdport,
+ unsigned long gmx_unused Flags,
+ gmx_walltime_accounting_t walltime_accounting)
+{
+ const char *CG = "Polak-Ribiere Conjugate Gradients";
+
+ em_state_t *s_min, *s_a, *s_b, *s_c;
+ gmx_localtop_t *top;
+ gmx_enerdata_t *enerd;
+ rvec *f;
+ gmx_global_stat_t gstat;
+ t_graph *graph;
+ rvec *p, *sf;
+ double gpa, gpb, gpc, tmp, minstep;
+ real fnormn;
+ real stepsize;
+ real a, b, c, beta = 0.0;
+ real epot_repl = 0;
+ real pnorm;
+ t_mdebin *mdebin;
+ gmx_bool converged, foundlower;
+ rvec mu_tot;
+ gmx_bool do_log = FALSE, do_ene = FALSE, do_x, do_f;
+ tensor vir, pres;
+ int number_steps, neval = 0, nstcg = inputrec->nstcgsteep;
+ gmx_mdoutf_t outf;
+ int i, m, gf, step, nminstep;
+
+ step = 0;
+
+ s_min = init_em_state();
+ s_a = init_em_state();
+ s_b = init_em_state();
+ s_c = init_em_state();
+
+ /* Init em and store the local state in s_min */
+ init_em(fplog, CG, cr, inputrec,
+ state_global, top_global, s_min, &top, &f,
+ nrnb, mu_tot, fr, &enerd, &graph, mdatoms, &gstat, vsite, constr,
+ nfile, fnm, &outf, &mdebin, imdport, Flags, wcycle);
+
+ /* Print to log file */
+ print_em_start(fplog, cr, walltime_accounting, wcycle, CG);
+
+ /* Max number of steps */
+ number_steps = inputrec->nsteps;
+
+ if (MASTER(cr))
+ {
+ sp_header(stderr, CG, inputrec->em_tol, number_steps);
+ }
+ if (fplog)
+ {
+ sp_header(fplog, CG, inputrec->em_tol, number_steps);
+ }
+
+ /* Call the force routine and some auxiliary (neighboursearching etc.) */
+ /* do_force always puts the charge groups in the box and shifts again
+ * We do not unshift, so molecules are always whole in congrad.c
+ */
+ evaluate_energy(fplog, cr,
+ top_global, s_min, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph, mdatoms, fr,
+ mu_tot, enerd, vir, pres, -1, TRUE);
+ where();
+
+ if (MASTER(cr))
+ {
+ /* Copy stuff to the energy bin for easy printing etc. */
+ upd_mdebin(mdebin, FALSE, FALSE, (double)step,
+ mdatoms->tmass, enerd, &s_min->s, inputrec->fepvals, inputrec->expandedvals, s_min->s.box,
+ NULL, NULL, vir, pres, NULL, mu_tot, constr);
+
+ print_ebin_header(fplog, step, step);
+ print_ebin(mdoutf_get_fp_ene(outf), TRUE, FALSE, FALSE, fplog, step, step, eprNORMAL,
+ mdebin, fcd, &(top_global->groups), &(inputrec->opts));
+ }
+ where();
+
+ /* Estimate/guess the initial stepsize */
+ stepsize = inputrec->em_stepsize/s_min->fnorm;
+
+ if (MASTER(cr))
+ {
+ double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
+ fprintf(stderr, " F-max = %12.5e on atom %d\n",
+ s_min->fmax, s_min->a_fmax+1);
+ fprintf(stderr, " F-Norm = %12.5e\n",
+ s_min->fnorm/sqrtNumAtoms);
+ fprintf(stderr, "\n");
+ /* and copy to the log file too... */
+ fprintf(fplog, " F-max = %12.5e on atom %d\n",
+ s_min->fmax, s_min->a_fmax+1);
+ fprintf(fplog, " F-Norm = %12.5e\n",
+ s_min->fnorm/sqrtNumAtoms);
+ fprintf(fplog, "\n");
+ }
+ /* Start the loop over CG steps.
+ * Each successful step is counted, and we continue until
+ * we either converge or reach the max number of steps.
+ */
+ converged = FALSE;
+ for (step = 0; (number_steps < 0 || step <= number_steps) && !converged; step++)
+ {
+
+ /* start taking steps in a new direction
+ * First time we enter the routine, beta=0, and the direction is
+ * simply the negative gradient.
+ */
+
+ /* Calculate the new direction in p, and the gradient in this direction, gpa */
+ p = s_min->s.cg_p;
+ sf = s_min->f;
+ gpa = 0;
+ gf = 0;
+ for (i = 0; i < mdatoms->homenr; i++)
+ {
+ if (mdatoms->cFREEZE)
+ {
+ gf = mdatoms->cFREEZE[i];
+ }
+ for (m = 0; m < DIM; m++)
+ {
+ if (!inputrec->opts.nFreeze[gf][m])
+ {
+ p[i][m] = sf[i][m] + beta*p[i][m];
+ gpa -= p[i][m]*sf[i][m];
+ /* f is negative gradient, thus the sign */
+ }
+ else
+ {
+ p[i][m] = 0;
+ }
+ }
+ }
+
+ /* Sum the gradient along the line across CPUs */
+ if (PAR(cr))
+ {
+ gmx_sumd(1, &gpa, cr);
+ }
+
+ /* Calculate the norm of the search vector */
+ get_f_norm_max(cr, &(inputrec->opts), mdatoms, p, &pnorm, NULL, NULL);
+
+ /* Just in case stepsize reaches zero due to numerical precision... */
+ if (stepsize <= 0)
+ {
+ stepsize = inputrec->em_stepsize/pnorm;
+ }
+
+ /*
+ * Double check the value of the derivative in the search direction.
+ * If it is positive it must be due to the old information in the
+ * CG formula, so just remove that and start over with beta=0.
+ * This corresponds to a steepest descent step.
+ */
+ if (gpa > 0)
+ {
+ beta = 0;
+ step--; /* Don't count this step since we are restarting */
+ continue; /* Go back to the beginning of the big for-loop */
+ }
+
+ /* Calculate minimum allowed stepsize, before the average (norm)
+ * relative change in coordinate is smaller than precision
+ */
+ minstep = 0;
+ for (i = 0; i < mdatoms->homenr; i++)
+ {
+ for (m = 0; m < DIM; m++)
+ {
+ tmp = fabs(s_min->s.x[i][m]);
+ if (tmp < 1.0)
+ {
+ tmp = 1.0;
+ }
+ tmp = p[i][m]/tmp;
+ minstep += tmp*tmp;
+ }
+ }
+ /* Add up from all CPUs */
+ if (PAR(cr))
+ {
+ gmx_sumd(1, &minstep, cr);
+ }
+
+ minstep = GMX_REAL_EPS/sqrt(minstep/(3*state_global->natoms));
+
+ if (stepsize < minstep)
+ {
+ converged = TRUE;
+ break;
+ }
+
+ /* Write coordinates if necessary */
+ do_x = do_per_step(step, inputrec->nstxout);
+ do_f = do_per_step(step, inputrec->nstfout);
+
+ write_em_traj(fplog, cr, outf, do_x, do_f, NULL,
+ top_global, inputrec, step,
+ s_min, state_global);
+
+ /* Take a step downhill.
+ * In theory, we should minimize the function along this direction.
+ * That is quite possible, but it turns out to take 5-10 function evaluations
+ * for each line. However, we dont really need to find the exact minimum -
+ * it is much better to start a new CG step in a modified direction as soon
+ * as we are close to it. This will save a lot of energy evaluations.
+ *
+ * In practice, we just try to take a single step.
+ * If it worked (i.e. lowered the energy), we increase the stepsize but
+ * the continue straight to the next CG step without trying to find any minimum.
+ * If it didn't work (higher energy), there must be a minimum somewhere between
+ * the old position and the new one.
+ *
+ * Due to the finite numerical accuracy, it turns out that it is a good idea
+ * to even accept a SMALL increase in energy, if the derivative is still downhill.
+ * This leads to lower final energies in the tests I've done. / Erik
+ */
+ s_a->epot = s_min->epot;
+ a = 0.0;
+ c = a + stepsize; /* reference position along line is zero */
+
+ if (DOMAINDECOMP(cr) && s_min->s.ddp_count < cr->dd->ddp_count)
+ {
+ em_dd_partition_system(fplog, step, cr, top_global, inputrec,
+ s_min, top, mdatoms, fr, vsite, constr,
+ nrnb, wcycle);
+ }
+
+ /* Take a trial step (new coords in s_c) */
+ do_em_step(cr, inputrec, mdatoms, fr->bMolPBC, s_min, c, s_min->s.cg_p, s_c,
+ constr, top, nrnb, wcycle, -1);
+
+ neval++;
+ /* Calculate energy for the trial step */
+ evaluate_energy(fplog, cr,
+ top_global, s_c, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph, mdatoms, fr,
+ mu_tot, enerd, vir, pres, -1, FALSE);
+
+ /* Calc derivative along line */
+ p = s_c->s.cg_p;
+ sf = s_c->f;
+ gpc = 0;
+ for (i = 0; i < mdatoms->homenr; i++)
+ {
+ for (m = 0; m < DIM; m++)
+ {
+ gpc -= p[i][m]*sf[i][m]; /* f is negative gradient, thus the sign */
+ }
+ }
+ /* Sum the gradient along the line across CPUs */
+ if (PAR(cr))
+ {
+ gmx_sumd(1, &gpc, cr);
+ }
+
+ /* This is the max amount of increase in energy we tolerate */
+ tmp = sqrt(GMX_REAL_EPS)*fabs(s_a->epot);
+
+ /* Accept the step if the energy is lower, or if it is not significantly higher
+ * and the line derivative is still negative.
+ */
+ if (s_c->epot < s_a->epot || (gpc < 0 && s_c->epot < (s_a->epot + tmp)))
+ {
+ foundlower = TRUE;
+ /* Great, we found a better energy. Increase step for next iteration
+ * if we are still going down, decrease it otherwise
+ */
+ if (gpc < 0)
+ {
+ stepsize *= 1.618034; /* The golden section */
+ }
+ else
+ {
+ stepsize *= 0.618034; /* 1/golden section */
+ }
+ }
+ else
+ {
+ /* New energy is the same or higher. We will have to do some work
+ * to find a smaller value in the interval. Take smaller step next time!
+ */
+ foundlower = FALSE;
+ stepsize *= 0.618034;
+ }
+
+
+
+
+ /* OK, if we didn't find a lower value we will have to locate one now - there must
+ * be one in the interval [a=0,c].
+ * The same thing is valid here, though: Don't spend dozens of iterations to find
+ * the line minimum. We try to interpolate based on the derivative at the endpoints,
+ * and only continue until we find a lower value. In most cases this means 1-2 iterations.
+ *
+ * I also have a safeguard for potentially really pathological functions so we never
+ * take more than 20 steps before we give up ...
+ *
+ * If we already found a lower value we just skip this step and continue to the update.
+ */
+ if (!foundlower)
+ {
+ nminstep = 0;
+
+ do
+ {
+ /* Select a new trial point.
+ * If the derivatives at points a & c have different sign we interpolate to zero,
+ * otherwise just do a bisection.
+ */
+ if (gpa < 0 && gpc > 0)
+ {
+ b = a + gpa*(a-c)/(gpc-gpa);
+ }
+ else
+ {
+ b = 0.5*(a+c);
+ }
+
+ /* safeguard if interpolation close to machine accuracy causes errors:
+ * never go outside the interval
+ */
+ if (b <= a || b >= c)
+ {
+ b = 0.5*(a+c);
+ }
+
+ if (DOMAINDECOMP(cr) && s_min->s.ddp_count != cr->dd->ddp_count)
+ {
+ /* Reload the old state */
+ em_dd_partition_system(fplog, -1, cr, top_global, inputrec,
+ s_min, top, mdatoms, fr, vsite, constr,
+ nrnb, wcycle);
+ }
+
+ /* Take a trial step to this new point - new coords in s_b */
+ do_em_step(cr, inputrec, mdatoms, fr->bMolPBC, s_min, b, s_min->s.cg_p, s_b,
+ constr, top, nrnb, wcycle, -1);
+
+ neval++;
+ /* Calculate energy for the trial step */
+ evaluate_energy(fplog, cr,
+ top_global, s_b, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph, mdatoms, fr,
+ mu_tot, enerd, vir, pres, -1, FALSE);
+
+ /* p does not change within a step, but since the domain decomposition
+ * might change, we have to use cg_p of s_b here.
+ */
+ p = s_b->s.cg_p;
+ sf = s_b->f;
+ gpb = 0;
+ for (i = 0; i < mdatoms->homenr; i++)
+ {
+ for (m = 0; m < DIM; m++)
+ {
+ gpb -= p[i][m]*sf[i][m]; /* f is negative gradient, thus the sign */
+ }
+ }
+ /* Sum the gradient along the line across CPUs */
+ if (PAR(cr))
+ {
+ gmx_sumd(1, &gpb, cr);
+ }
+
+ if (debug)
+ {
+ fprintf(debug, "CGE: EpotA %f EpotB %f EpotC %f gpb %f\n",
+ s_a->epot, s_b->epot, s_c->epot, gpb);
+ }
+
+ epot_repl = s_b->epot;
+
+ /* Keep one of the intervals based on the value of the derivative at the new point */
+ if (gpb > 0)
+ {
+ /* Replace c endpoint with b */
+ swap_em_state(s_b, s_c);
+ c = b;
+ gpc = gpb;
+ }
+ else
+ {
+ /* Replace a endpoint with b */
+ swap_em_state(s_b, s_a);
+ a = b;
+ gpa = gpb;
+ }
+
+ /*
+ * Stop search as soon as we find a value smaller than the endpoints.
+ * Never run more than 20 steps, no matter what.
+ */
+ nminstep++;
+ }
+ while ((epot_repl > s_a->epot || epot_repl > s_c->epot) &&
+ (nminstep < 20));
+
+ if (fabs(epot_repl - s_min->epot) < fabs(s_min->epot)*GMX_REAL_EPS ||
+ nminstep >= 20)
+ {
+ /* OK. We couldn't find a significantly lower energy.
+ * If beta==0 this was steepest descent, and then we give up.
+ * If not, set beta=0 and restart with steepest descent before quitting.
+ */
+ if (beta == 0.0)
+ {
+ /* Converged */
+ converged = TRUE;
+ break;
+ }
+ else
+ {
+ /* Reset memory before giving up */
+ beta = 0.0;
+ continue;
+ }
+ }
+
+ /* Select min energy state of A & C, put the best in B.
+ */
+ if (s_c->epot < s_a->epot)
+ {
+ if (debug)
+ {
+ fprintf(debug, "CGE: C (%f) is lower than A (%f), moving C to B\n",
+ s_c->epot, s_a->epot);
+ }
+ swap_em_state(s_b, s_c);
+ gpb = gpc;
+ }
+ else
+ {
+ if (debug)
+ {
+ fprintf(debug, "CGE: A (%f) is lower than C (%f), moving A to B\n",
+ s_a->epot, s_c->epot);
+ }
+ swap_em_state(s_b, s_a);
+ gpb = gpa;
+ }
+
+ }
+ else
+ {
+ if (debug)
+ {
+ fprintf(debug, "CGE: Found a lower energy %f, moving C to B\n",
+ s_c->epot);
+ }
+ swap_em_state(s_b, s_c);
+ gpb = gpc;
+ }
+
+ /* new search direction */
+ /* beta = 0 means forget all memory and restart with steepest descents. */
+ if (nstcg && ((step % nstcg) == 0))
+ {
+ beta = 0.0;
+ }
+ else
+ {
+ /* s_min->fnorm cannot be zero, because then we would have converged
+ * and broken out.
+ */
+
+ /* Polak-Ribiere update.
+ * Change to fnorm2/fnorm2_old for Fletcher-Reeves
+ */
+ beta = pr_beta(cr, &inputrec->opts, mdatoms, top_global, s_min, s_b);
+ }
+ /* Limit beta to prevent oscillations */
+ if (fabs(beta) > 5.0)
+ {
+ beta = 0.0;
+ }
+
+
+ /* update positions */
+ swap_em_state(s_min, s_b);
+ gpa = gpb;
+
+ /* Print it if necessary */
+ if (MASTER(cr))
+ {
+ if (bVerbose)
+ {
+ double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
+ fprintf(stderr, "\rStep %d, Epot=%12.6e, Fnorm=%9.3e, Fmax=%9.3e (atom %d)\n",
+ step, s_min->epot, s_min->fnorm/sqrtNumAtoms,
+ s_min->fmax, s_min->a_fmax+1);
+ fflush(stderr);
+ }
+ /* Store the new (lower) energies */
+ upd_mdebin(mdebin, FALSE, FALSE, (double)step,
+ mdatoms->tmass, enerd, &s_min->s, inputrec->fepvals, inputrec->expandedvals, s_min->s.box,
+ NULL, NULL, vir, pres, NULL, mu_tot, constr);
+
+ do_log = do_per_step(step, inputrec->nstlog);
+ do_ene = do_per_step(step, inputrec->nstenergy);
+
+ /* Prepare IMD energy record, if bIMD is TRUE. */
+ IMD_fill_energy_record(inputrec->bIMD, inputrec->imd, enerd, step, TRUE);
+
+ if (do_log)
+ {
+ print_ebin_header(fplog, step, step);
+ }
+ print_ebin(mdoutf_get_fp_ene(outf), do_ene, FALSE, FALSE,
+ do_log ? fplog : NULL, step, step, eprNORMAL,
+ mdebin, fcd, &(top_global->groups), &(inputrec->opts));
+ }
+
+ /* Send energies and positions to the IMD client if bIMD is TRUE. */
+ if (do_IMD(inputrec->bIMD, step, cr, TRUE, state_global->box, state_global->x, inputrec, 0, wcycle) && MASTER(cr))
+ {
+ IMD_send_positions(inputrec->imd);
+ }
+
+ /* Stop when the maximum force lies below tolerance.
+ * If we have reached machine precision, converged is already set to true.
+ */
+ converged = converged || (s_min->fmax < inputrec->em_tol);
+
+ } /* End of the loop */
+
+ /* IMD cleanup, if bIMD is TRUE. */
+ IMD_finalize(inputrec->bIMD, inputrec->imd);
+
+ if (converged)
+ {
+ step--; /* we never took that last step in this case */
+
+ }
+ if (s_min->fmax > inputrec->em_tol)
+ {
+ if (MASTER(cr))
+ {
+ warn_step(stderr, inputrec->em_tol, step-1 == number_steps, FALSE);
+ warn_step(fplog, inputrec->em_tol, step-1 == number_steps, FALSE);
+ }
+ converged = FALSE;
+ }
+
+ if (MASTER(cr))
+ {
+ /* If we printed energy and/or logfile last step (which was the last step)
+ * we don't have to do it again, but otherwise print the final values.
+ */
+ if (!do_log)
+ {
+ /* Write final value to log since we didn't do anything the last step */
+ print_ebin_header(fplog, step, step);
+ }
+ if (!do_ene || !do_log)
+ {
+ /* Write final energy file entries */
+ print_ebin(mdoutf_get_fp_ene(outf), !do_ene, FALSE, FALSE,
+ !do_log ? fplog : NULL, step, step, eprNORMAL,
+ mdebin, fcd, &(top_global->groups), &(inputrec->opts));
+ }
+ }
+
+ /* Print some stuff... */
+ if (MASTER(cr))
+ {
+ fprintf(stderr, "\nwriting lowest energy coordinates.\n");
+ }
+
+ /* IMPORTANT!
+ * For accurate normal mode calculation it is imperative that we
+ * store the last conformation into the full precision binary trajectory.
+ *
+ * However, we should only do it if we did NOT already write this step
+ * above (which we did if do_x or do_f was true).
+ */
+ do_x = !do_per_step(step, inputrec->nstxout);
+ do_f = (inputrec->nstfout > 0 && !do_per_step(step, inputrec->nstfout));
+
+ write_em_traj(fplog, cr, outf, do_x, do_f, ftp2fn(efSTO, nfile, fnm),
+ top_global, inputrec, step,
+ s_min, state_global);
+
+
+ if (MASTER(cr))
+ {
+ double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
+ fnormn = s_min->fnorm/sqrtNumAtoms;
+ print_converged(stderr, CG, inputrec->em_tol, step, converged, number_steps,
+ s_min->epot, s_min->fmax, s_min->a_fmax, fnormn);
+ print_converged(fplog, CG, inputrec->em_tol, step, converged, number_steps,
+ s_min->epot, s_min->fmax, s_min->a_fmax, fnormn);
+
+ fprintf(fplog, "\nPerformed %d energy evaluations in total.\n", neval);
+ }
+
+ finish_em(cr, outf, walltime_accounting, wcycle);
+
+ /* To print the actual number of steps we needed somewhere */
+ walltime_accounting_set_nsteps_done(walltime_accounting, step);
+
+ return 0;
+} /* That's all folks */
+
+
+/*! \brief Do L-BFGS conjugate gradients minimization
+ \copydoc integrator_t (FILE *fplog, t_commrec *cr,
+ int nfile, const t_filenm fnm[],
+ const gmx_output_env_t *oenv, gmx_bool bVerbose,
+ int nstglobalcomm,
+ gmx_vsite_t *vsite, gmx_constr_t constr,
+ int stepout,
+ t_inputrec *inputrec,
+ 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_nex, int repl_ex_seed,
+ real cpt_period, real max_hours,
+ int imdport,
+ unsigned long Flags,
+ gmx_walltime_accounting_t walltime_accounting)
+ */
+double do_lbfgs(FILE *fplog, t_commrec *cr,
+ int nfile, const t_filenm fnm[],
+ const gmx_output_env_t gmx_unused *oenv, gmx_bool bVerbose,
+ int gmx_unused nstglobalcomm,
+ gmx_vsite_t *vsite, gmx_constr_t constr,
+ int gmx_unused stepout,
+ t_inputrec *inputrec,
+ 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 gmx_unused ed,
+ t_forcerec *fr,
+ int gmx_unused repl_ex_nst, int gmx_unused repl_ex_nex, int gmx_unused repl_ex_seed,
+ real gmx_unused cpt_period, real gmx_unused max_hours,
+ int imdport,
+ unsigned long gmx_unused Flags,
+ gmx_walltime_accounting_t walltime_accounting)
+{
+ static const char *LBFGS = "Low-Memory BFGS Minimizer";
+ em_state_t ems;
+ gmx_localtop_t *top;
+ gmx_enerdata_t *enerd;
+ rvec *f;
+ gmx_global_stat_t gstat;
+ t_graph *graph;
+ int ncorr, nmaxcorr, point, cp, neval, nminstep;
+ double stepsize, step_taken, gpa, gpb, gpc, tmp, minstep;
+ real *rho, *alpha, *ff, *xx, *p, *s, *lastx, *lastf, **dx, **dg;
+ real *xa, *xb, *xc, *fa, *fb, *fc, *xtmp, *ftmp;
+ real a, b, c, maxdelta, delta;
+ real diag, Epot0, Epot, EpotA, EpotB, EpotC;
+ real dgdx, dgdg, sq, yr, beta;
+ t_mdebin *mdebin;
+ gmx_bool converged;
+ rvec mu_tot;
+ real fnorm, fmax;
+ gmx_bool do_log, do_ene, do_x, do_f, foundlower, *frozen;
+ tensor vir, pres;
+ int start, end, number_steps;
+ gmx_mdoutf_t outf;
+ int i, k, m, n, nfmax, gf, step;
+ int mdof_flags;
+
+ if (PAR(cr))
+ {
+ gmx_fatal(FARGS, "Cannot do parallel L-BFGS Minimization - yet.\n");
+ }
+
+ if (NULL != constr)
+ {
+ gmx_fatal(FARGS, "The combination of constraints and L-BFGS minimization is not implemented. Either do not use constraints, or use another minimizer (e.g. steepest descent).");
+ }
+
+ n = 3*state_global->natoms;
+ nmaxcorr = inputrec->nbfgscorr;
+
+ /* Allocate memory */
+ /* Use pointers to real so we dont have to loop over both atoms and
+ * dimensions all the time...
+ * x/f are allocated as rvec *, so make new x0/f0 pointers-to-real
+ * that point to the same memory.
+ */
+ snew(xa, n);
+ snew(xb, n);
+ snew(xc, n);
+ snew(fa, n);
+ snew(fb, n);
+ snew(fc, n);
+ snew(frozen, n);
+
+ snew(p, n);
+ snew(lastx, n);
+ snew(lastf, n);
+ snew(rho, nmaxcorr);
+ snew(alpha, nmaxcorr);
+
+ snew(dx, nmaxcorr);
+ for (i = 0; i < nmaxcorr; i++)
+ {
+ snew(dx[i], n);
+ }
+
+ snew(dg, nmaxcorr);
+ for (i = 0; i < nmaxcorr; i++)
+ {
+ snew(dg[i], n);
+ }
+
+ step = 0;
+ neval = 0;
+
+ /* Init em */
+ init_em(fplog, LBFGS, cr, inputrec,
+ state_global, top_global, &ems, &top, &f,
+ nrnb, mu_tot, fr, &enerd, &graph, mdatoms, &gstat, vsite, constr,
+ nfile, fnm, &outf, &mdebin, imdport, Flags, wcycle);
+ /* Do_lbfgs is not completely updated like do_steep and do_cg,
+ * so we free some memory again.
+ */
+ sfree(ems.s.x);
+ sfree(ems.f);
+
+ xx = (real *)state_global->x;
+ ff = (real *)f;
+
+ start = 0;
+ end = mdatoms->homenr;
+
+ /* Print to log file */
+ print_em_start(fplog, cr, walltime_accounting, wcycle, LBFGS);
+
+ do_log = do_ene = do_x = do_f = TRUE;
+
+ /* Max number of steps */
+ number_steps = inputrec->nsteps;
+
+ /* Create a 3*natoms index to tell whether each degree of freedom is frozen */
+ gf = 0;
+ for (i = start; i < end; i++)
+ {
+ if (mdatoms->cFREEZE)
+ {
+ gf = mdatoms->cFREEZE[i];
+ }
+ for (m = 0; m < DIM; m++)
+ {
+ frozen[3*i+m] = inputrec->opts.nFreeze[gf][m];
+ }
+ }
+ if (MASTER(cr))
+ {
+ sp_header(stderr, LBFGS, inputrec->em_tol, number_steps);
+ }
+ if (fplog)
+ {
+ sp_header(fplog, LBFGS, inputrec->em_tol, number_steps);
+ }
+
+ if (vsite)
+ {
+ construct_vsites(vsite, state_global->x, 1, NULL,
+ top->idef.iparams, top->idef.il,
+ fr->ePBC, fr->bMolPBC, cr, state_global->box);
+ }
+
+ /* Call the force routine and some auxiliary (neighboursearching etc.) */
+ /* do_force always puts the charge groups in the box and shifts again
+ * We do not unshift, so molecules are always whole
+ */
+ neval++;
+ ems.s.x = state_global->x;
+ ems.f = f;
+ evaluate_energy(fplog, cr,
+ top_global, &ems, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph, mdatoms, fr,
+ mu_tot, enerd, vir, pres, -1, TRUE);
+ where();
+
+ if (MASTER(cr))
+ {
+ /* Copy stuff to the energy bin for easy printing etc. */
+ upd_mdebin(mdebin, FALSE, FALSE, (double)step,
+ mdatoms->tmass, enerd, state_global, inputrec->fepvals, inputrec->expandedvals, state_global->box,
+ NULL, NULL, vir, pres, NULL, mu_tot, constr);
+
+ print_ebin_header(fplog, step, step);
+ print_ebin(mdoutf_get_fp_ene(outf), TRUE, FALSE, FALSE, fplog, step, step, eprNORMAL,
+ mdebin, fcd, &(top_global->groups), &(inputrec->opts));
+ }
+ where();
+
+ /* This is the starting energy */
+ Epot = enerd->term[F_EPOT];
+
+ fnorm = ems.fnorm;
+ fmax = ems.fmax;
+ nfmax = ems.a_fmax;
+
+ /* Set the initial step.
+ * since it will be multiplied by the non-normalized search direction
+ * vector (force vector the first time), we scale it by the
+ * norm of the force.
+ */
+
+ if (MASTER(cr))
+ {
+ double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
+ fprintf(stderr, "Using %d BFGS correction steps.\n\n", nmaxcorr);
+ fprintf(stderr, " F-max = %12.5e on atom %d\n", fmax, nfmax+1);
+ fprintf(stderr, " F-Norm = %12.5e\n", fnorm/sqrtNumAtoms);
+ fprintf(stderr, "\n");
+ /* and copy to the log file too... */
+ fprintf(fplog, "Using %d BFGS correction steps.\n\n", nmaxcorr);
+ fprintf(fplog, " F-max = %12.5e on atom %d\n", fmax, nfmax+1);
+ fprintf(fplog, " F-Norm = %12.5e\n", fnorm/sqrtNumAtoms);
+ fprintf(fplog, "\n");
+ }
+
+ // Point is an index to the memory of search directions, where 0 is the first one.
+ point = 0;
+
+ // Set initial search direction to the force (-gradient), or 0 for frozen particles.
+ for (i = 0; i < n; i++)
+ {
+ if (!frozen[i])
+ {
+ dx[point][i] = ff[i]; /* Initial search direction */
+ }
+ else
+ {
+ dx[point][i] = 0;
+ }
+ }
+
+ // Stepsize will be modified during the search, and actually it is not critical
+ // (the main efficiency in the algorithm comes from changing directions), but
+ // we still need an initial value, so estimate it as the inverse of the norm
+ // so we take small steps where the potential fluctuates a lot.
+ stepsize = 1.0/fnorm;
+
+ /* Start the loop over BFGS steps.
+ * Each successful step is counted, and we continue until
+ * we either converge or reach the max number of steps.
+ */
+
+ ncorr = 0;
+
+ /* Set the gradient from the force */
+ converged = FALSE;
+ for (step = 0; (number_steps < 0 || step <= number_steps) && !converged; step++)
+ {
+
+ /* Write coordinates if necessary */
+ do_x = do_per_step(step, inputrec->nstxout);
+ do_f = do_per_step(step, inputrec->nstfout);
+
+ mdof_flags = 0;
+ if (do_x)
+ {
+ mdof_flags |= MDOF_X;
+ }
+
+ if (do_f)
+ {
+ mdof_flags |= MDOF_F;
+ }
+
+ if (inputrec->bIMD)
+ {
+ mdof_flags |= MDOF_IMD;
+ }
+
+ mdoutf_write_to_trajectory_files(fplog, cr, outf, mdof_flags,
+ top_global, step, (real)step, state_global, state_global, f);
+
+ /* Do the linesearching in the direction dx[point][0..(n-1)] */
+
+ /* make s a pointer to current search direction - point=0 first time we get here */
+ s = dx[point];
+
+ // calculate line gradient in position A
+ for (gpa = 0, i = 0; i < n; i++)
+ {
+ gpa -= s[i]*ff[i];
+ }
+
+ /* Calculate minimum allowed stepsize along the line, before the average (norm)
+ * relative change in coordinate is smaller than precision
+ */
+ for (minstep = 0, i = 0; i < n; i++)
+ {
+ tmp = fabs(xx[i]);
+ if (tmp < 1.0)
+ {
+ tmp = 1.0;
+ }
+ tmp = s[i]/tmp;
+ minstep += tmp*tmp;
+ }
+ minstep = GMX_REAL_EPS/sqrt(minstep/n);
+
+ if (stepsize < minstep)
+ {
+ converged = TRUE;
+ break;
+ }
+
+ // Before taking any steps along the line, store the old position
+ for (i = 0; i < n; i++)
+ {
+ lastx[i] = xx[i];
+ lastf[i] = ff[i];
+ }
+ Epot0 = Epot;
+
+ for (i = 0; i < n; i++)
+ {
+ xa[i] = xx[i];
+ }
+
+ /* Take a step downhill.
+ * In theory, we should find the actual minimum of the function in this
+ * direction, somewhere along the line.
+ * That is quite possible, but it turns out to take 5-10 function evaluations
+ * for each line. However, we dont really need to find the exact minimum -
+ * it is much better to start a new BFGS step in a modified direction as soon
+ * as we are close to it. This will save a lot of energy evaluations.
+ *
+ * In practice, we just try to take a single step.
+ * If it worked (i.e. lowered the energy), we increase the stepsize but
+ * continue straight to the next BFGS step without trying to find any minimum,
+ * i.e. we change the search direction too. If the line was smooth, it is
+ * likely we are in a smooth region, and then it makes sense to take longer
+ * steps in the modified search direction too.
+ *
+ * If it didn't work (higher energy), there must be a minimum somewhere between
+ * the old position and the new one. Then we need to start by finding a lower
+ * value before we change search direction. Since the energy was apparently
+ * quite rough, we need to decrease the step size.
+ *
+ * Due to the finite numerical accuracy, it turns out that it is a good idea
+ * to accept a SMALL increase in energy, if the derivative is still downhill.
+ * This leads to lower final energies in the tests I've done. / Erik
+ */
+
+ // State "A" is the first position along the line.
+ // reference position along line is initially zero
+ EpotA = Epot0;
+ a = 0.0;
+
+ // Check stepsize first. We do not allow displacements
+ // larger than emstep.
+ //
+ do
+ {
+ // Pick a new position C by adding stepsize to A.
+ c = a + stepsize;
+
+ // Calculate what the largest change in any individual coordinate
+ // would be (translation along line * gradient along line)
+ maxdelta = 0;
+ for (i = 0; i < n; i++)
+ {
+ delta = c*s[i];
+ if (delta > maxdelta)
+ {
+ maxdelta = delta;
+ }
+ }
+ // If any displacement is larger than the stepsize limit, reduce the step
+ if (maxdelta > inputrec->em_stepsize)
+ {
+ stepsize *= 0.1;
+ }
+ }
+ while (maxdelta > inputrec->em_stepsize);
+
+ // Take a trial step and move the coordinate array xc[] to position C
+ for (i = 0; i < n; i++)
+ {
+ xc[i] = lastx[i] + c*s[i];
+ }
+
+ neval++;
+ // Calculate energy for the trial step in position C
+ ems.s.x = (rvec *)xc;
+ ems.f = (rvec *)fc;
+ evaluate_energy(fplog, cr,
+ top_global, &ems, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph, mdatoms, fr,
+ mu_tot, enerd, vir, pres, step, FALSE);
+ EpotC = ems.epot;
+
+ // Calc line gradient in position C
+ for (gpc = 0, i = 0; i < n; i++)
+ {
+ gpc -= s[i]*fc[i]; /* f is negative gradient, thus the sign */
+ }
+ /* Sum the gradient along the line across CPUs */
+ if (PAR(cr))
+ {
+ gmx_sumd(1, &gpc, cr);
+ }
+
+ // This is the max amount of increase in energy we tolerate.
+ // By allowing VERY small changes (close to numerical precision) we
+ // frequently find even better (lower) final energies.
+ tmp = sqrt(GMX_REAL_EPS)*fabs(EpotA);
+
+ // Accept the step if the energy is lower in the new position C (compared to A),
+ // or if it is not significantly higher and the line derivative is still negative.
+ if (EpotC < EpotA || (gpc < 0 && EpotC < (EpotA+tmp)))
+ {
+ // Great, we found a better energy. We no longer try to alter the
+ // stepsize, but simply accept this new better position. The we select a new
+ // search direction instead, which will be much more efficient than continuing
+ // to take smaller steps along a line. Set fnorm based on the new C position,
+ // which will be used to update the stepsize to 1/fnorm further down.
+ foundlower = TRUE;
+ fnorm = ems.fnorm;
+ }
+ else
+ {
+ // If we got here, the energy is NOT lower in point C, i.e. it will be the same
+ // or higher than in point A. In this case it is pointless to move to point C,
+ // so we will have to do more iterations along the same line to find a smaller
+ // value in the interval [A=0.0,C].
+ // Here, A is still 0.0, but that will change when we do a search in the interval
+ // [0.0,C] below. That search we will do by interpolation or bisection rather
+ // than with the stepsize, so no need to modify it. For the next search direction
+ // it will be reset to 1/fnorm anyway.
+ foundlower = FALSE;
+ }
+
+ if (!foundlower)
+ {
+ // OK, if we didn't find a lower value we will have to locate one now - there must
+ // be one in the interval [a,c].
+ // The same thing is valid here, though: Don't spend dozens of iterations to find
+ // the line minimum. We try to interpolate based on the derivative at the endpoints,
+ // and only continue until we find a lower value. In most cases this means 1-2 iterations.
+ // I also have a safeguard for potentially really pathological functions so we never
+ // take more than 20 steps before we give up.
+ // If we already found a lower value we just skip this step and continue to the update.
+ nminstep = 0;
+ do
+ {
+ // Select a new trial point B in the interval [A,C].
+ // If the derivatives at points a & c have different sign we interpolate to zero,
+ // otherwise just do a bisection since there might be multiple minima/maxima
+ // inside the interval.
+ if (gpa < 0 && gpc > 0)
+ {
+ b = a + gpa*(a-c)/(gpc-gpa);
+ }
+ else
+ {
+ b = 0.5*(a+c);
+ }
+
+ /* safeguard if interpolation close to machine accuracy causes errors:
+ * never go outside the interval
+ */
+ if (b <= a || b >= c)
+ {
+ b = 0.5*(a+c);
+ }
+
+ // Take a trial step to point B
+ for (i = 0; i < n; i++)
+ {
+ xb[i] = lastx[i] + b*s[i];
+ }
+
+ neval++;
+ // Calculate energy for the trial step in point B
+ ems.s.x = (rvec *)xb;
+ ems.f = (rvec *)fb;
+ evaluate_energy(fplog, cr,
+ top_global, &ems, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph, mdatoms, fr,
+ mu_tot, enerd, vir, pres, step, FALSE);
+ EpotB = ems.epot;
+ fnorm = ems.fnorm;
+
+ // Calculate gradient in point B
+ for (gpb = 0, i = 0; i < n; i++)
+ {
+ gpb -= s[i]*fb[i]; /* f is negative gradient, thus the sign */
+
+ }
+ /* Sum the gradient along the line across CPUs */
+ if (PAR(cr))
+ {
+ gmx_sumd(1, &gpb, cr);
+ }
+
+ // Keep one of the intervals [A,B] or [B,C] based on the value of the derivative
+ // at the new point B, and rename the endpoints of this new interval A and C.
+ if (gpb > 0)
+ {
+ /* Replace c endpoint with b */
+ EpotC = EpotB;
+ c = b;
+ gpc = gpb;
+ /* swap coord pointers b/c */
+ xtmp = xb;
+ ftmp = fb;
+ xb = xc;
+ fb = fc;
+ xc = xtmp;
+ fc = ftmp;
+ }
+ else
+ {
+ /* Replace a endpoint with b */
+ EpotA = EpotB;
+ a = b;
+ gpa = gpb;
+ /* swap coord pointers a/b */
+ xtmp = xb;
+ ftmp = fb;
+ xb = xa;
+ fb = fa;
+ xa = xtmp;
+ fa = ftmp;
+ }
+
+ /*
+ * Stop search as soon as we find a value smaller than the endpoints,
+ * or if the tolerance is below machine precision.
+ * Never run more than 20 steps, no matter what.
+ */
+ nminstep++;
+ }
+ while ((EpotB > EpotA || EpotB > EpotC) && (nminstep < 20));
+
+ if (fabs(EpotB-Epot0) < GMX_REAL_EPS || nminstep >= 20)
+ {
+ /* OK. We couldn't find a significantly lower energy.
+ * If ncorr==0 this was steepest descent, and then we give up.
+ * If not, reset memory to restart as steepest descent before quitting.
+ */
+ if (ncorr == 0)
+ {
+ /* Converged */
+ converged = TRUE;
+ break;
+ }
+ else
+ {
+ /* Reset memory */
+ ncorr = 0;
+ /* Search in gradient direction */
+ for (i = 0; i < n; i++)
+ {
+ dx[point][i] = ff[i];
+ }
+ /* Reset stepsize */
+ stepsize = 1.0/fnorm;
+ continue;
+ }
+ }
+
+ /* Select min energy state of A & C, put the best in xx/ff/Epot
+ */
+ if (EpotC < EpotA)
+ {
+ Epot = EpotC;
+ /* Use state C */
+ for (i = 0; i < n; i++)
+ {
+ xx[i] = xc[i];
+ ff[i] = fc[i];
+ }
+ step_taken = c;
+ }
+ else
+ {
+ Epot = EpotA;
+ /* Use state A */
+ for (i = 0; i < n; i++)
+ {
+ xx[i] = xa[i];
+ ff[i] = fa[i];
+ }
+ step_taken = a;
+ }
+
+ }
+ else
+ {
+ /* found lower */
+ Epot = EpotC;
+ /* Use state C */
+ for (i = 0; i < n; i++)
+ {
+ xx[i] = xc[i];
+ ff[i] = fc[i];
+ }
+ step_taken = c;
+ }
+
+ /* Update the memory information, and calculate a new
+ * approximation of the inverse hessian
+ */
+
+ /* Have new data in Epot, xx, ff */
+ if (ncorr < nmaxcorr)
+ {
+ ncorr++;
+ }
+
+ for (i = 0; i < n; i++)
+ {
+ dg[point][i] = lastf[i]-ff[i];
+ dx[point][i] *= step_taken;
+ }
+
+ dgdg = 0;
+ dgdx = 0;
+ for (i = 0; i < n; i++)
+ {
+ dgdg += dg[point][i]*dg[point][i];
+ dgdx += dg[point][i]*dx[point][i];
+ }
+
+ diag = dgdx/dgdg;
+
+ rho[point] = 1.0/dgdx;
+ point++;
+
+ if (point >= nmaxcorr)
+ {
+ point = 0;
+ }
+
+ /* Update */
+ for (i = 0; i < n; i++)
+ {
+ p[i] = ff[i];
+ }
+
+ cp = point;
+
+ /* Recursive update. First go back over the memory points */
+ for (k = 0; k < ncorr; k++)
+ {
+ cp--;
+ if (cp < 0)
+ {
+ cp = ncorr-1;
+ }
+
+ sq = 0;
+ for (i = 0; i < n; i++)
+ {
+ sq += dx[cp][i]*p[i];
+ }
+
+ alpha[cp] = rho[cp]*sq;
+
+ for (i = 0; i < n; i++)
+ {
+ p[i] -= alpha[cp]*dg[cp][i];
+ }
+ }
+
+ for (i = 0; i < n; i++)
+ {
+ p[i] *= diag;
+ }
+
+ /* And then go forward again */
+ for (k = 0; k < ncorr; k++)
+ {
+ yr = 0;
+ for (i = 0; i < n; i++)
+ {
+ yr += p[i]*dg[cp][i];
+ }
+
+ beta = rho[cp]*yr;
+ beta = alpha[cp]-beta;
+
+ for (i = 0; i < n; i++)
+ {
+ p[i] += beta*dx[cp][i];
+ }
+
+ cp++;
+ if (cp >= ncorr)
+ {
+ cp = 0;
+ }
+ }
+
+ for (i = 0; i < n; i++)
+ {
+ if (!frozen[i])
+ {
+ dx[point][i] = p[i];
+ }
+ else
+ {
+ dx[point][i] = 0;
+ }
+ }
+
+ /* Test whether the convergence criterion is met */
+ get_f_norm_max(cr, &(inputrec->opts), mdatoms, f, &fnorm, &fmax, &nfmax);
+
+ /* Print it if necessary */
+ if (MASTER(cr))
+ {
+ if (bVerbose)
+ {
+ double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
+ fprintf(stderr, "\rStep %d, Epot=%12.6e, Fnorm=%9.3e, Fmax=%9.3e (atom %d)\n",
+ step, Epot, fnorm/sqrtNumAtoms, fmax, nfmax+1);
+ fflush(stderr);
+ }
+ /* Store the new (lower) energies */
+ upd_mdebin(mdebin, FALSE, FALSE, (double)step,
+ mdatoms->tmass, enerd, state_global, inputrec->fepvals, inputrec->expandedvals, state_global->box,
+ NULL, NULL, vir, pres, NULL, mu_tot, constr);
+ do_log = do_per_step(step, inputrec->nstlog);
+ do_ene = do_per_step(step, inputrec->nstenergy);
+ if (do_log)
+ {
+ print_ebin_header(fplog, step, step);
+ }
+ print_ebin(mdoutf_get_fp_ene(outf), do_ene, FALSE, FALSE,
+ do_log ? fplog : NULL, step, step, eprNORMAL,
+ mdebin, fcd, &(top_global->groups), &(inputrec->opts));
+ }
+
+ /* Send x and E to IMD client, if bIMD is TRUE. */
+ if (do_IMD(inputrec->bIMD, step, cr, TRUE, state_global->box, state_global->x, inputrec, 0, wcycle) && MASTER(cr))
+ {
+ IMD_send_positions(inputrec->imd);
+ }
+
+ // Reset stepsize in we are doing more iterations
+ stepsize = 1.0/fnorm;
+
+ /* Stop when the maximum force lies below tolerance.
+ * If we have reached machine precision, converged is already set to true.
+ */
+ converged = converged || (fmax < inputrec->em_tol);
+
+ } /* End of the loop */
+
+ /* IMD cleanup, if bIMD is TRUE. */
+ IMD_finalize(inputrec->bIMD, inputrec->imd);
+
+ if (converged)
+ {
+ step--; /* we never took that last step in this case */
+
+ }
+ if (fmax > inputrec->em_tol)
+ {
+ if (MASTER(cr))
+ {
+ warn_step(stderr, inputrec->em_tol, step-1 == number_steps, FALSE);
+ warn_step(fplog, inputrec->em_tol, step-1 == number_steps, FALSE);
+ }
+ converged = FALSE;
+ }
+
+ /* If we printed energy and/or logfile last step (which was the last step)
+ * we don't have to do it again, but otherwise print the final values.
+ */
+ if (!do_log) /* Write final value to log since we didn't do anythin last step */
+ {
+ print_ebin_header(fplog, step, step);
+ }
+ if (!do_ene || !do_log) /* Write final energy file entries */
+ {
+ print_ebin(mdoutf_get_fp_ene(outf), !do_ene, FALSE, FALSE,
+ !do_log ? fplog : NULL, step, step, eprNORMAL,
+ mdebin, fcd, &(top_global->groups), &(inputrec->opts));
+ }
+
+ /* Print some stuff... */
+ if (MASTER(cr))
+ {
+ fprintf(stderr, "\nwriting lowest energy coordinates.\n");
+ }
+
+ /* IMPORTANT!
+ * For accurate normal mode calculation it is imperative that we
+ * store the last conformation into the full precision binary trajectory.
+ *
+ * However, we should only do it if we did NOT already write this step
+ * above (which we did if do_x or do_f was true).
+ */
+ do_x = !do_per_step(step, inputrec->nstxout);
+ do_f = !do_per_step(step, inputrec->nstfout);
+ write_em_traj(fplog, cr, outf, do_x, do_f, ftp2fn(efSTO, nfile, fnm),
+ top_global, inputrec, step,
+ &ems, state_global);
+
+ if (MASTER(cr))
+ {
+ double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
+ print_converged(stderr, LBFGS, inputrec->em_tol, step, converged,
+ number_steps, Epot, fmax, nfmax, fnorm/sqrtNumAtoms);
+ print_converged(fplog, LBFGS, inputrec->em_tol, step, converged,
+ number_steps, Epot, fmax, nfmax, fnorm/sqrtNumAtoms);
+
+ fprintf(fplog, "\nPerformed %d energy evaluations in total.\n", neval);
+ }
+
+ finish_em(cr, outf, walltime_accounting, wcycle);
+
+ /* To print the actual number of steps we needed somewhere */
+ walltime_accounting_set_nsteps_done(walltime_accounting, step);
+
+ return 0;
+} /* That's all folks */
+
+/*! \brief Do steepest descents minimization
+ \copydoc integrator_t (FILE *fplog, t_commrec *cr,
+ int nfile, const t_filenm fnm[],
+ const gmx_output_env_t *oenv, gmx_bool bVerbose,
+ int nstglobalcomm,
+ gmx_vsite_t *vsite, gmx_constr_t constr,
+ int stepout,
+ t_inputrec *inputrec,
+ 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_nex, int repl_ex_seed,
+ real cpt_period, real max_hours,
+ int imdport,
+ unsigned long Flags,
+ gmx_walltime_accounting_t walltime_accounting)
+ */
+double do_steep(FILE *fplog, t_commrec *cr,
+ int nfile, const t_filenm fnm[],
+ const gmx_output_env_t gmx_unused *oenv, gmx_bool bVerbose,
+ int gmx_unused nstglobalcomm,
+ gmx_vsite_t *vsite, gmx_constr_t constr,
+ int gmx_unused stepout,
+ t_inputrec *inputrec,
+ 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 gmx_unused ed,
+ t_forcerec *fr,
+ int gmx_unused repl_ex_nst, int gmx_unused repl_ex_nex, int gmx_unused repl_ex_seed,
+ real gmx_unused cpt_period, real gmx_unused max_hours,
+ int imdport,
+ unsigned long gmx_unused Flags,
+ gmx_walltime_accounting_t walltime_accounting)
+{
+ const char *SD = "Steepest Descents";
+ em_state_t *s_min, *s_try;
+ gmx_localtop_t *top;
+ gmx_enerdata_t *enerd;
+ rvec *f;
+ gmx_global_stat_t gstat;
+ t_graph *graph;
+ real stepsize;
+ real ustep, fnormn;
+ gmx_mdoutf_t outf;
+ t_mdebin *mdebin;
+ gmx_bool bDone, bAbort, do_x, do_f;
+ tensor vir, pres;
+ rvec mu_tot;
+ int nsteps;
+ int count = 0;
+ int steps_accepted = 0;
+
+ s_min = init_em_state();
+ s_try = init_em_state();
+
+ /* Init em and store the local state in s_try */
+ init_em(fplog, SD, cr, inputrec,
+ state_global, top_global, s_try, &top, &f,
+ nrnb, mu_tot, fr, &enerd, &graph, mdatoms, &gstat, vsite, constr,
+ nfile, fnm, &outf, &mdebin, imdport, Flags, wcycle);
+
+ /* Print to log file */
+ print_em_start(fplog, cr, walltime_accounting, wcycle, SD);
+
+ /* Set variables for stepsize (in nm). This is the largest
+ * step that we are going to make in any direction.
+ */
+ ustep = inputrec->em_stepsize;
+ stepsize = 0;
+
+ /* Max number of steps */
+ nsteps = inputrec->nsteps;
+
+ if (MASTER(cr))
+ {
+ /* Print to the screen */
+ sp_header(stderr, SD, inputrec->em_tol, nsteps);
+ }
+ if (fplog)
+ {
+ sp_header(fplog, SD, inputrec->em_tol, nsteps);
+ }
+
+ /**** HERE STARTS THE LOOP ****
+ * count is the counter for the number of steps
+ * bDone will be TRUE when the minimization has converged
+ * bAbort will be TRUE when nsteps steps have been performed or when
+ * the stepsize becomes smaller than is reasonable for machine precision
+ */
+ count = 0;
+ bDone = FALSE;
+ bAbort = FALSE;
+ while (!bDone && !bAbort)
+ {
+ bAbort = (nsteps >= 0) && (count == nsteps);
+
+ /* set new coordinates, except for first step */
+ if (count > 0)
+ {
+ do_em_step(cr, inputrec, mdatoms, fr->bMolPBC,
+ s_min, stepsize, s_min->f, s_try,
+ constr, top, nrnb, wcycle, count);
+ }
+
+ evaluate_energy(fplog, cr,
+ top_global, s_try, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph, mdatoms, fr,
+ mu_tot, enerd, vir, pres, count, count == 0);
+
+ if (MASTER(cr))
+ {
+ print_ebin_header(fplog, count, count);
+ }
+
+ if (count == 0)
+ {
+ s_min->epot = s_try->epot;
+ }
+
+ /* Print it if necessary */
+ if (MASTER(cr))
+ {
+ if (bVerbose)
+ {
+ fprintf(stderr, "Step=%5d, Dmax= %6.1e nm, Epot= %12.5e Fmax= %11.5e, atom= %d%c",
+ count, ustep, s_try->epot, s_try->fmax, s_try->a_fmax+1,
+ ( (count == 0) || (s_try->epot < s_min->epot) ) ? '\n' : '\r');
+ fflush(stderr);
+ }
+
+ if ( (count == 0) || (s_try->epot < s_min->epot) )
+ {
+ /* Store the new (lower) energies */
+ upd_mdebin(mdebin, FALSE, FALSE, (double)count,
+ mdatoms->tmass, enerd, &s_try->s, inputrec->fepvals, inputrec->expandedvals,
+ s_try->s.box, NULL, NULL, vir, pres, NULL, mu_tot, constr);
+
+ /* Prepare IMD energy record, if bIMD is TRUE. */
+ IMD_fill_energy_record(inputrec->bIMD, inputrec->imd, enerd, count, TRUE);
+
+ print_ebin(mdoutf_get_fp_ene(outf), TRUE,
+ do_per_step(steps_accepted, inputrec->nstdisreout),
+ do_per_step(steps_accepted, inputrec->nstorireout),
+ fplog, count, count, eprNORMAL,
+ mdebin, fcd, &(top_global->groups), &(inputrec->opts));
+ fflush(fplog);
+ }
+ }
+
+ /* Now if the new energy is smaller than the previous...
+ * or if this is the first step!
+ * or if we did random steps!
+ */
+
+ if ( (count == 0) || (s_try->epot < s_min->epot) )
+ {
+ steps_accepted++;
+
+ /* Test whether the convergence criterion is met... */
+ bDone = (s_try->fmax < inputrec->em_tol);
+
+ /* Copy the arrays for force, positions and energy */
+ /* The 'Min' array always holds the coords and forces of the minimal
+ sampled energy */
+ swap_em_state(s_min, s_try);
+ if (count > 0)
+ {
+ ustep *= 1.2;
+ }
+
+ /* Write to trn, if necessary */
+ do_x = do_per_step(steps_accepted, inputrec->nstxout);
+ do_f = do_per_step(steps_accepted, inputrec->nstfout);
+ write_em_traj(fplog, cr, outf, do_x, do_f, NULL,
+ top_global, inputrec, count,
+ s_min, state_global);
+ }
+ else
+ {
+ /* If energy is not smaller make the step smaller... */
+ ustep *= 0.5;
+
+ if (DOMAINDECOMP(cr) && s_min->s.ddp_count != cr->dd->ddp_count)
+ {
+ /* Reload the old state */
+ em_dd_partition_system(fplog, count, cr, top_global, inputrec,
+ s_min, top, mdatoms, fr, vsite, constr,
+ nrnb, wcycle);
+ }
+ }
+
+ /* Determine new step */
+ stepsize = ustep/s_min->fmax;
+
+ /* Check if stepsize is too small, with 1 nm as a characteristic length */
+#if GMX_DOUBLE
+ if (count == nsteps || ustep < 1e-12)
+#else
+ if (count == nsteps || ustep < 1e-6)
+#endif
+ {
+ if (MASTER(cr))
+ {
+ warn_step(stderr, inputrec->em_tol, count == nsteps, constr != NULL);
+ warn_step(fplog, inputrec->em_tol, count == nsteps, constr != NULL);
+ }
+ bAbort = TRUE;
+ }
+
+ /* Send IMD energies and positions, if bIMD is TRUE. */
+ if (do_IMD(inputrec->bIMD, count, cr, TRUE, state_global->box, state_global->x, inputrec, 0, wcycle) && MASTER(cr))
+ {
+ IMD_send_positions(inputrec->imd);
+ }
+
+ count++;
+ } /* End of the loop */
+
+ /* IMD cleanup, if bIMD is TRUE. */
+ IMD_finalize(inputrec->bIMD, inputrec->imd);
+
+ /* Print some data... */
+ if (MASTER(cr))
+ {
+ fprintf(stderr, "\nwriting lowest energy coordinates.\n");
+ }
+ write_em_traj(fplog, cr, outf, TRUE, inputrec->nstfout, ftp2fn(efSTO, nfile, fnm),
+ top_global, inputrec, count,
+ s_min, state_global);
+
+ if (MASTER(cr))
+ {
+ double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
+ fnormn = s_min->fnorm/sqrtNumAtoms;
+
+ print_converged(stderr, SD, inputrec->em_tol, count, bDone, nsteps,
+ s_min->epot, s_min->fmax, s_min->a_fmax, fnormn);
+ print_converged(fplog, SD, inputrec->em_tol, count, bDone, nsteps,
+ s_min->epot, s_min->fmax, s_min->a_fmax, fnormn);
+ }
+
+ finish_em(cr, outf, walltime_accounting, wcycle);
+
+ /* To print the actual number of steps we needed somewhere */
+ inputrec->nsteps = count;
+
+ walltime_accounting_set_nsteps_done(walltime_accounting, count);
+
+ return 0;
+} /* That's all folks */
+
+/*! \brief Do normal modes analysis
+ \copydoc integrator_t (FILE *fplog, t_commrec *cr,
+ int nfile, const t_filenm fnm[],
+ const gmx_output_env_t *oenv, gmx_bool bVerbose,
+ int nstglobalcomm,
+ gmx_vsite_t *vsite, gmx_constr_t constr,
+ int stepout,
+ t_inputrec *inputrec,
+ 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_nex, int repl_ex_seed,
+ real cpt_period, real max_hours,
+ int imdport,
+ unsigned long Flags,
+ gmx_walltime_accounting_t walltime_accounting)
+ */
+double do_nm(FILE *fplog, t_commrec *cr,
+ int nfile, const t_filenm fnm[],
+ const gmx_output_env_t gmx_unused *oenv, gmx_bool bVerbose,
+ int gmx_unused nstglobalcomm,
+ gmx_vsite_t *vsite, gmx_constr_t constr,
+ int gmx_unused stepout,
+ t_inputrec *inputrec,
+ 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 gmx_unused ed,
+ t_forcerec *fr,
+ int gmx_unused repl_ex_nst, int gmx_unused repl_ex_nex, int gmx_unused repl_ex_seed,
+ real gmx_unused cpt_period, real gmx_unused max_hours,
+ int imdport,
+ unsigned long gmx_unused Flags,
+ gmx_walltime_accounting_t walltime_accounting)
+{
+ const char *NM = "Normal Mode Analysis";
+ gmx_mdoutf_t outf;
+ int nnodes, node;
+ gmx_localtop_t *top;
+ gmx_enerdata_t *enerd;
+ rvec *f;
+ gmx_global_stat_t gstat;
+ t_graph *graph;
+ tensor vir, pres;
+ rvec mu_tot;
+ rvec *fneg, *dfdx;
+ gmx_bool bSparse; /* use sparse matrix storage format */
+ size_t sz;
+ gmx_sparsematrix_t * sparse_matrix = NULL;
+ real * full_matrix = NULL;
+ em_state_t * state_work;
+
+ /* added with respect to mdrun */
+ int row, col;
+ real der_range = 10.0*sqrt(GMX_REAL_EPS);
+ real x_min;
+ bool bIsMaster = MASTER(cr);
+
+ if (constr != NULL)
+ {
+ gmx_fatal(FARGS, "Constraints present with Normal Mode Analysis, this combination is not supported");
+ }
+
+ state_work = init_em_state();
+
+ /* Init em and store the local state in state_minimum */
+ init_em(fplog, NM, cr, inputrec,
+ state_global, top_global, state_work, &top,
+ &f,
+ nrnb, mu_tot, fr, &enerd, &graph, mdatoms, &gstat, vsite, constr,
+ nfile, fnm, &outf, NULL, imdport, Flags, wcycle);
+
+ gmx_shellfc_t *shellfc = init_shell_flexcon(stdout,
+ top_global,
+ n_flexible_constraints(constr),
+ inputrec->nstcalcenergy,
+ DOMAINDECOMP(cr));
+
+ if (shellfc)
+ {
+ make_local_shells(cr, mdatoms, shellfc);
+ }
+ std::vector<size_t> atom_index = get_atom_index(top_global);
+ snew(fneg, atom_index.size());
+ snew(dfdx, atom_index.size());
+
+#if !GMX_DOUBLE
+ if (bIsMaster)
+ {
+ fprintf(stderr,
+ "NOTE: This version of GROMACS has been compiled in single precision,\n"
+ " which MIGHT not be accurate enough for normal mode analysis.\n"
+ " GROMACS now uses sparse matrix storage, so the memory requirements\n"
+ " are fairly modest even if you recompile in double precision.\n\n");
+ }
+#endif
+
+ /* Check if we can/should use sparse storage format.
+ *
+ * Sparse format is only useful when the Hessian itself is sparse, which it
+ * will be when we use a cutoff.
+ * For small systems (n<1000) it is easier to always use full matrix format, though.
+ */
+ if (EEL_FULL(fr->eeltype) || fr->rlist == 0.0)
+ {
+ md_print_info(cr, fplog, "Non-cutoff electrostatics used, forcing full Hessian format.\n");
+ bSparse = FALSE;
+ }
+ else if (atom_index.size() < 1000)
+ {
+ md_print_info(cr, fplog, "Small system size (N=%d), using full Hessian format.\n", atom_index.size());
+ bSparse = FALSE;
+ }
+ else
+ {
+ md_print_info(cr, fplog, "Using compressed symmetric sparse Hessian format.\n");
+ bSparse = TRUE;
+ }
+
+ /* Number of dimensions, based on real atoms, that is not vsites or shell */
+ sz = DIM*atom_index.size();
+
+ fprintf(stderr, "Allocating Hessian memory...\n\n");
+
+ if (bSparse)
+ {
+ sparse_matrix = gmx_sparsematrix_init(sz);
+ sparse_matrix->compressed_symmetric = TRUE;
+ }
+ else
+ {
+ snew(full_matrix, sz*sz);
+ }
+
+ init_nrnb(nrnb);
+
+ where();
+
+ /* Write start time and temperature */
+ print_em_start(fplog, cr, walltime_accounting, wcycle, NM);
+
+ /* fudge nr of steps to nr of atoms */
+ inputrec->nsteps = atom_index.size()*2;
+
+ if (bIsMaster)
+ {
+ fprintf(stderr, "starting normal mode calculation '%s'\n%d steps.\n\n",
+ *(top_global->name), (int)inputrec->nsteps);
+ }
+
+ nnodes = cr->nnodes;
+
+ /* Make evaluate_energy do a single node force calculation */
+ cr->nnodes = 1;
+ evaluate_energy(fplog, cr,
+ top_global, state_work, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph, mdatoms, fr,
+ mu_tot, enerd, vir, pres, -1, TRUE);
+ cr->nnodes = nnodes;
+
+ /* if forces are not small, warn user */
+ get_state_f_norm_max(cr, &(inputrec->opts), mdatoms, state_work);
+
+ md_print_info(cr, fplog, "Maximum force:%12.5e\n", state_work->fmax);
+ if (state_work->fmax > 1.0e-3)
+ {
+ md_print_info(cr, fplog,
+ "The force is probably not small enough to "
+ "ensure that you are at a minimum.\n"
+ "Be aware that negative eigenvalues may occur\n"
+ "when the resulting matrix is diagonalized.\n\n");
+ }
+
+ /***********************************************************
+ *
+ * Loop over all pairs in matrix
+ *
+ * do_force called twice. Once with positive and
+ * once with negative displacement
+ *
+ ************************************************************/
+
+ /* Steps are divided one by one over the nodes */
+ bool bNS = true;
+ for (unsigned int aid = cr->nodeid; aid < atom_index.size(); aid += nnodes)
+ {
+ size_t atom = atom_index[aid];
+ for (size_t d = 0; d < DIM; d++)
+ {
+ gmx_bool bBornRadii = FALSE;
+ gmx_int64_t step = 0;
+ int force_flags = GMX_FORCE_STATECHANGED | GMX_FORCE_ALLFORCES;
+ double t = 0;
+
+ x_min = state_work->s.x[atom][d];
+
+ for (unsigned int dx = 0; (dx < 2); dx++)
+ {
+ if (dx == 0)
+ {
+ state_work->s.x[atom][d] = x_min - der_range;
+ }
+ else
+ {
+ state_work->s.x[atom][d] = x_min + der_range;
+ }
+
+ /* Make evaluate_energy do a single node force calculation */
+ cr->nnodes = 1;
+ if (shellfc)
+ {
+ /* Now is the time to relax the shells */
+ (void) relax_shell_flexcon(fplog, cr, bVerbose, step,
+ inputrec, bNS, force_flags,
+ top,
+ constr, enerd, fcd,
+ &state_work->s, state_work->f, vir, mdatoms,
+ nrnb, wcycle, graph, &top_global->groups,
+ shellfc, fr, bBornRadii, t, mu_tot,
+ vsite, NULL);
+ bNS = false;
+ step++;
+ }
+ else
+ {
+ evaluate_energy(fplog, cr,
+ top_global, state_work, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph, mdatoms, fr,
+ mu_tot, enerd, vir, pres, atom*2+dx, FALSE);
+ }
+
+ cr->nnodes = nnodes;
+
+ if (dx == 0)
+ {
+ for (size_t i = 0; i < atom_index.size(); i++)
+ {
+ copy_rvec(state_work->f[atom_index[i]], fneg[i]);
+ }
+ }
+ }
+
+ /* x is restored to original */
+ state_work->s.x[atom][d] = x_min;
+
+ for (size_t j = 0; j < atom_index.size(); j++)
+ {
+ for (size_t k = 0; (k < DIM); k++)
+ {
+ dfdx[j][k] =
+ -(state_work->f[atom_index[j]][k] - fneg[j][k])/(2*der_range);
+ }
+ }
+
+ if (!bIsMaster)
+ {
+#if GMX_MPI
+#define mpi_type GMX_MPI_REAL
+ MPI_Send(dfdx[0], atom_index.size()*DIM, mpi_type, MASTER(cr),
+ cr->nodeid, cr->mpi_comm_mygroup);
+#endif
+ }
+ else
+ {
+ for (node = 0; (node < nnodes && atom+node < atom_index.size()); node++)
+ {
+ if (node > 0)
+ {
+#if GMX_MPI
+ MPI_Status stat;
+ MPI_Recv(dfdx[0], atom_index.size()*DIM, mpi_type, node, node,
+ cr->mpi_comm_mygroup, &stat);
+#undef mpi_type
+#endif
+ }
+
+ row = (atom + node)*DIM + d;
+
+ for (size_t j = 0; j < atom_index.size(); j++)
+ {
+ for (size_t k = 0; k < DIM; k++)
+ {
+ col = j*DIM + k;
+
+ if (bSparse)
+ {
+ if (col >= row && dfdx[j][k] != 0.0)
+ {
+ gmx_sparsematrix_increment_value(sparse_matrix,
+ row, col, dfdx[j][k]);
+ }
+ }
+ else
+ {
+ full_matrix[row*sz+col] = dfdx[j][k];
+ }
+ }
+ }
+ }
+ }
+
+ if (bVerbose && fplog)
+ {
+ fflush(fplog);
+ }
+ }
+ /* write progress */
+ if (bIsMaster && bVerbose)
+ {
+ fprintf(stderr, "\rFinished step %d out of %d",
+ static_cast<int>(std::min(atom+nnodes, atom_index.size())),
+ static_cast<int>(atom_index.size()));
+ fflush(stderr);
+ }
+ }
+
+ if (bIsMaster)
+ {
+ fprintf(stderr, "\n\nWriting Hessian...\n");
+ gmx_mtxio_write(ftp2fn(efMTX, nfile, fnm), sz, sz, full_matrix, sparse_matrix);
+ }
+
+ finish_em(cr, outf, walltime_accounting, wcycle);
+
+ walltime_accounting_set_nsteps_done(walltime_accounting, atom_index.size()*2);
+
+ return 0;
+}
+
+} // namespace gmx
diff --git a/patches/gromacs-2016-beta1.diff/src/gromacs/mdlib/minimize.cpp.preplumed b/patches/gromacs-2016-beta1.diff/src/gromacs/mdlib/minimize.cpp.preplumed
new file mode 100644
index 0000000000000000000000000000000000000000..5a205694090705e66fb3b3a07da9748f185c8bd7
--- /dev/null
+++ b/patches/gromacs-2016-beta1.diff/src/gromacs/mdlib/minimize.cpp.preplumed
@@ -0,0 +1,3054 @@
+/*
+ * This file is part of the GROMACS molecular simulation package.
+ *
+ * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
+ * Copyright (c) 2001-2004, The GROMACS development team.
+ * Copyright (c) 2013,2014,2015,2016, by the GROMACS development team, led by
+ * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
+ * and including many others, as listed in the AUTHORS file in the
+ * top-level source directory and at http://www.gromacs.org.
+ *
+ * GROMACS is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public License
+ * as published by the Free Software Foundation; either version 2.1
+ * of the License, or (at your option) any later version.
+ *
+ * GROMACS is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with GROMACS; if not, see
+ * http://www.gnu.org/licenses, or write to the Free Software Foundation,
+ * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ * If you want to redistribute modifications to GROMACS, 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 http://www.gromacs.org.
+ *
+ * To help us fund GROMACS development, we humbly ask that you cite
+ * the research papers on the package. Check out http://www.gromacs.org.
+ */
+/*! \internal \file
+ *
+ * \brief This file defines integrators for energy minimization
+ *
+ * \author Berk Hess <hess@kth.se>
+ * \author Erik Lindahl <erik@kth.se>
+ * \ingroup module_mdlib
+ */
+#include "gmxpre.h"
+
+#include "minimize.h"
+
+#include "config.h"
+
+#include <cmath>
+#include <cstring>
+#include <ctime>
+
+#include <algorithm>
+#include <vector>
+
+#include "gromacs/commandline/filenm.h"
+#include "gromacs/domdec/domdec.h"
+#include "gromacs/domdec/domdec_struct.h"
+#include "gromacs/ewald/pme.h"
+#include "gromacs/fileio/confio.h"
+#include "gromacs/fileio/mtxio.h"
+#include "gromacs/gmxlib/md_logging.h"
+#include "gromacs/gmxlib/network.h"
+#include "gromacs/gmxlib/nrnb.h"
+#include "gromacs/imd/imd.h"
+#include "gromacs/linearalgebra/sparsematrix.h"
+#include "gromacs/listed-forces/manage-threading.h"
+#include "gromacs/math/functions.h"
+#include "gromacs/math/vec.h"
+#include "gromacs/mdlib/constr.h"
+#include "gromacs/mdlib/force.h"
+#include "gromacs/mdlib/forcerec.h"
+#include "gromacs/mdlib/gmx_omp_nthreads.h"
+#include "gromacs/mdlib/md_support.h"
+#include "gromacs/mdlib/mdatoms.h"
+#include "gromacs/mdlib/mdebin.h"
+#include "gromacs/mdlib/mdrun.h"
+#include "gromacs/mdlib/ns.h"
+#include "gromacs/mdlib/shellfc.h"
+#include "gromacs/mdlib/sim_util.h"
+#include "gromacs/mdlib/tgroup.h"
+#include "gromacs/mdlib/trajectory_writing.h"
+#include "gromacs/mdlib/update.h"
+#include "gromacs/mdlib/vsite.h"
+#include "gromacs/mdtypes/commrec.h"
+#include "gromacs/mdtypes/inputrec.h"
+#include "gromacs/mdtypes/md_enums.h"
+#include "gromacs/pbcutil/mshift.h"
+#include "gromacs/pbcutil/pbc.h"
+#include "gromacs/timing/wallcycle.h"
+#include "gromacs/timing/walltime_accounting.h"
+#include "gromacs/topology/mtop_util.h"
+#include "gromacs/utility/cstringutil.h"
+#include "gromacs/utility/exceptions.h"
+#include "gromacs/utility/fatalerror.h"
+#include "gromacs/utility/smalloc.h"
+
+//! Utility structure for manipulating states during EM
+typedef struct {
+ //! Copy of the global state
+ t_state s;
+ //! Force array
+ rvec *f;
+ //! Potential energy
+ real epot;
+ //! Norm of the force
+ real fnorm;
+ //! Maximum force
+ real fmax;
+ //! Direction
+ int a_fmax;
+} em_state_t;
+
+//! Initiate em_state_t structure and return pointer to it
+static em_state_t *init_em_state()
+{
+ em_state_t *ems;
+
+ snew(ems, 1);
+
+ /* does this need to be here? Should the array be declared differently (staticaly)in the state definition? */
+ snew(ems->s.lambda, efptNR);
+
+ return ems;
+}
+
+//! Print the EM starting conditions
+static void print_em_start(FILE *fplog,
+ t_commrec *cr,
+ gmx_walltime_accounting_t walltime_accounting,
+ gmx_wallcycle_t wcycle,
+ const char *name)
+{
+ walltime_accounting_start(walltime_accounting);
+ wallcycle_start(wcycle, ewcRUN);
+ print_start(fplog, cr, walltime_accounting, name);
+}
+
+//! Stop counting time for EM
+static void em_time_end(gmx_walltime_accounting_t walltime_accounting,
+ gmx_wallcycle_t wcycle)
+{
+ wallcycle_stop(wcycle, ewcRUN);
+
+ walltime_accounting_end(walltime_accounting);
+}
+
+//! Printing a log file and console header
+static void sp_header(FILE *out, const char *minimizer, real ftol, int nsteps)
+{
+ fprintf(out, "\n");
+ fprintf(out, "%s:\n", minimizer);
+ fprintf(out, " Tolerance (Fmax) = %12.5e\n", ftol);
+ fprintf(out, " Number of steps = %12d\n", nsteps);
+}
+
+//! Print warning message
+static void warn_step(FILE *fp, real ftol, gmx_bool bLastStep, gmx_bool bConstrain)
+{
+ char buffer[2048];
+ if (bLastStep)
+ {
+ sprintf(buffer,
+ "\nEnergy minimization reached the maximum number "
+ "of steps before the forces reached the requested "
+ "precision Fmax < %g.\n", ftol);
+ }
+ else
+ {
+ sprintf(buffer,
+ "\nEnergy minimization has stopped, but the forces have "
+ "not converged to the requested precision Fmax < %g (which "
+ "may not be possible for your system). It stopped "
+ "because the algorithm tried to make a new step whose size "
+ "was too small, or there was no change in the energy since "
+ "last step. Either way, we regard the minimization as "
+ "converged to within the available machine precision, "
+ "given your starting configuration and EM parameters.\n%s%s",
+ ftol,
+ sizeof(real) < sizeof(double) ?
+ "\nDouble precision normally gives you higher accuracy, but "
+ "this is often not needed for preparing to run molecular "
+ "dynamics.\n" :
+ "",
+ bConstrain ?
+ "You might need to increase your constraint accuracy, or turn\n"
+ "off constraints altogether (set constraints = none in mdp file)\n" :
+ "");
+ }
+ fputs(wrap_lines(buffer, 78, 0, FALSE), fp);
+}
+
+//! Print message about convergence of the EM
+static void print_converged(FILE *fp, const char *alg, real ftol,
+ gmx_int64_t count, gmx_bool bDone, gmx_int64_t nsteps,
+ real epot, real fmax, int nfmax, real fnorm)
+{
+ char buf[STEPSTRSIZE];
+
+ if (bDone)
+ {
+ fprintf(fp, "\n%s converged to Fmax < %g in %s steps\n",
+ alg, ftol, gmx_step_str(count, buf));
+ }
+ else if (count < nsteps)
+ {
+ fprintf(fp, "\n%s converged to machine precision in %s steps,\n"
+ "but did not reach the requested Fmax < %g.\n",
+ alg, gmx_step_str(count, buf), ftol);
+ }
+ else
+ {
+ fprintf(fp, "\n%s did not converge to Fmax < %g in %s steps.\n",
+ alg, ftol, gmx_step_str(count, buf));
+ }
+
+#if GMX_DOUBLE
+ fprintf(fp, "Potential Energy = %21.14e\n", epot);
+ fprintf(fp, "Maximum force = %21.14e on atom %d\n", fmax, nfmax+1);
+ fprintf(fp, "Norm of force = %21.14e\n", fnorm);
+#else
+ fprintf(fp, "Potential Energy = %14.7e\n", epot);
+ fprintf(fp, "Maximum force = %14.7e on atom %d\n", fmax, nfmax+1);
+ fprintf(fp, "Norm of force = %14.7e\n", fnorm);
+#endif
+}
+
+//! Compute the norm and max of the force array in parallel
+static void get_f_norm_max(t_commrec *cr,
+ t_grpopts *opts, t_mdatoms *mdatoms, rvec *f,
+ real *fnorm, real *fmax, int *a_fmax)
+{
+ double fnorm2, *sum;
+ real fmax2, fam;
+ int la_max, a_max, start, end, i, m, gf;
+
+ /* This routine finds the largest force and returns it.
+ * On parallel machines the global max is taken.
+ */
+ fnorm2 = 0;
+ fmax2 = 0;
+ la_max = -1;
+ start = 0;
+ end = mdatoms->homenr;
+ if (mdatoms->cFREEZE)
+ {
+ for (i = start; i < end; i++)
+ {
+ gf = mdatoms->cFREEZE[i];
+ fam = 0;
+ for (m = 0; m < DIM; m++)
+ {
+ if (!opts->nFreeze[gf][m])
+ {
+ fam += gmx::square(f[i][m]);
+ }
+ }
+ fnorm2 += fam;
+ if (fam > fmax2)
+ {
+ fmax2 = fam;
+ la_max = i;
+ }
+ }
+ }
+ else
+ {
+ for (i = start; i < end; i++)
+ {
+ fam = norm2(f[i]);
+ fnorm2 += fam;
+ if (fam > fmax2)
+ {
+ fmax2 = fam;
+ la_max = i;
+ }
+ }
+ }
+
+ if (la_max >= 0 && DOMAINDECOMP(cr))
+ {
+ a_max = cr->dd->gatindex[la_max];
+ }
+ else
+ {
+ a_max = la_max;
+ }
+ if (PAR(cr))
+ {
+ snew(sum, 2*cr->nnodes+1);
+ sum[2*cr->nodeid] = fmax2;
+ sum[2*cr->nodeid+1] = a_max;
+ sum[2*cr->nnodes] = fnorm2;
+ gmx_sumd(2*cr->nnodes+1, sum, cr);
+ fnorm2 = sum[2*cr->nnodes];
+ /* Determine the global maximum */
+ for (i = 0; i < cr->nnodes; i++)
+ {
+ if (sum[2*i] > fmax2)
+ {
+ fmax2 = sum[2*i];
+ a_max = (int)(sum[2*i+1] + 0.5);
+ }
+ }
+ sfree(sum);
+ }
+
+ if (fnorm)
+ {
+ *fnorm = sqrt(fnorm2);
+ }
+ if (fmax)
+ {
+ *fmax = sqrt(fmax2);
+ }
+ if (a_fmax)
+ {
+ *a_fmax = a_max;
+ }
+}
+
+//! Compute the norm of the force
+static void get_state_f_norm_max(t_commrec *cr,
+ t_grpopts *opts, t_mdatoms *mdatoms,
+ em_state_t *ems)
+{
+ get_f_norm_max(cr, opts, mdatoms, ems->f, &ems->fnorm, &ems->fmax, &ems->a_fmax);
+}
+
+//! Initialize the energy minimization
+void init_em(FILE *fplog, const char *title,
+ t_commrec *cr, t_inputrec *ir,
+ t_state *state_global, gmx_mtop_t *top_global,
+ em_state_t *ems, gmx_localtop_t **top,
+ rvec **f,
+ t_nrnb *nrnb, rvec mu_tot,
+ t_forcerec *fr, gmx_enerdata_t **enerd,
+ t_graph **graph, t_mdatoms *mdatoms, gmx_global_stat_t *gstat,
+ gmx_vsite_t *vsite, gmx_constr_t constr,
+ int nfile, const t_filenm fnm[],
+ gmx_mdoutf_t *outf, t_mdebin **mdebin,
+ int imdport, unsigned long gmx_unused Flags,
+ gmx_wallcycle_t wcycle)
+{
+ int i;
+ real dvdl_constr;
+
+ if (fplog)
+ {
+ fprintf(fplog, "Initiating %s\n", title);
+ }
+
+ state_global->ngtc = 0;
+
+ /* Initialize lambda variables */
+ initialize_lambdas(fplog, ir, &(state_global->fep_state), state_global->lambda, NULL);
+
+ init_nrnb(nrnb);
+
+ /* Interactive molecular dynamics */
+ init_IMD(ir, cr, top_global, fplog, 1, state_global->x,
+ nfile, fnm, NULL, imdport, Flags);
+
+ if (DOMAINDECOMP(cr))
+ {
+ *top = dd_init_local_top(top_global);
+
+ dd_init_local_state(cr->dd, state_global, &ems->s);
+
+ *f = NULL;
+
+ /* 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,
+ &ems->s, &ems->f, mdatoms, *top,
+ fr, vsite, constr,
+ nrnb, NULL, FALSE);
+ dd_store_state(cr->dd, &ems->s);
+
+ *graph = NULL;
+ }
+ else
+ {
+ snew(*f, top_global->natoms);
+
+ /* Just copy the state */
+ ems->s = *state_global;
+ /* We need to allocate one element extra, since we might use
+ * (unaligned) 4-wide SIMD loads to access rvec entries.
+ */
+ snew(ems->s.x, ems->s.nalloc + 1);
+ snew(ems->f, ems->s.nalloc+1);
+ snew(ems->s.v, ems->s.nalloc+1);
+ for (i = 0; i < state_global->natoms; i++)
+ {
+ copy_rvec(state_global->x[i], ems->s.x[i]);
+ }
+ copy_mat(state_global->box, ems->s.box);
+
+ *top = gmx_mtop_generate_local_top(top_global, ir->efep != efepNO);
+
+ forcerec_set_excl_load(fr, *top);
+
+ setup_bonded_threading(fr, &(*top)->idef);
+
+ if (ir->ePBC != epbcNONE && !fr->bMolPBC)
+ {
+ *graph = mk_graph(fplog, &((*top)->idef), 0, top_global->natoms, FALSE, FALSE);
+ }
+ else
+ {
+ *graph = NULL;
+ }
+
+ atoms2md(top_global, ir, 0, NULL, top_global->natoms, mdatoms);
+ update_mdatoms(mdatoms, state_global->lambda[efptFEP]);
+
+ if (vsite)
+ {
+ set_vsite_top(vsite, *top, mdatoms, cr);
+ }
+ }
+
+ if (constr)
+ {
+ if (ir->eConstrAlg == econtSHAKE &&
+ gmx_mtop_ftype_count(top_global, F_CONSTR) > 0)
+ {
+ gmx_fatal(FARGS, "Can not do energy minimization with %s, use %s\n",
+ econstr_names[econtSHAKE], econstr_names[econtLINCS]);
+ }
+
+ if (!DOMAINDECOMP(cr))
+ {
+ set_constraints(constr, *top, ir, mdatoms, cr);
+ }
+
+ if (!ir->bContinuation)
+ {
+ /* Constrain the starting coordinates */
+ dvdl_constr = 0;
+ constrain(PAR(cr) ? NULL : fplog, TRUE, TRUE, constr, &(*top)->idef,
+ ir, cr, -1, 0, 1.0, mdatoms,
+ ems->s.x, ems->s.x, NULL, fr->bMolPBC, ems->s.box,
+ ems->s.lambda[efptFEP], &dvdl_constr,
+ NULL, NULL, nrnb, econqCoord);
+ }
+ }
+
+ if (PAR(cr))
+ {
+ *gstat = global_stat_init(ir);
+ }
+ else
+ {
+ *gstat = NULL;
+ }
+
+ *outf = init_mdoutf(fplog, nfile, fnm, 0, cr, ir, top_global, NULL, wcycle);
+
+ snew(*enerd, 1);
+ init_enerdata(top_global->groups.grps[egcENER].nr, ir->fepvals->n_lambda,
+ *enerd);
+
+ if (mdebin != NULL)
+ {
+ /* Init bin for energy stuff */
+ *mdebin = init_mdebin(mdoutf_get_fp_ene(*outf), top_global, ir, NULL);
+ }
+
+ clear_rvec(mu_tot);
+ calc_shifts(ems->s.box, fr->shift_vec);
+}
+
+//! Finalize the minimization
+static void finish_em(t_commrec *cr, gmx_mdoutf_t outf,
+ gmx_walltime_accounting_t walltime_accounting,
+ gmx_wallcycle_t wcycle)
+{
+ if (!(cr->duty & DUTY_PME))
+ {
+ /* Tell the PME only node to finish */
+ gmx_pme_send_finish(cr);
+ }
+
+ done_mdoutf(outf);
+
+ em_time_end(walltime_accounting, wcycle);
+}
+
+//! Swap two different EM states during minimization
+static void swap_em_state(em_state_t *ems1, em_state_t *ems2)
+{
+ em_state_t tmp;
+
+ tmp = *ems1;
+ *ems1 = *ems2;
+ *ems2 = tmp;
+}
+
+//! Copy coordinate from an EM state to a "normal" state structure
+static void copy_em_coords(em_state_t *ems, t_state *state)
+{
+ int i;
+
+ for (i = 0; (i < state->natoms); i++)
+ {
+ copy_rvec(ems->s.x[i], state->x[i]);
+ }
+}
+
+//! Save the EM trajectory
+static void write_em_traj(FILE *fplog, t_commrec *cr,
+ gmx_mdoutf_t outf,
+ gmx_bool bX, gmx_bool bF, const char *confout,
+ gmx_mtop_t *top_global,
+ t_inputrec *ir, gmx_int64_t step,
+ em_state_t *state,
+ t_state *state_global)
+{
+ int mdof_flags;
+ gmx_bool bIMDout = FALSE;
+
+
+ /* Shall we do IMD output? */
+ if (ir->bIMD)
+ {
+ bIMDout = do_per_step(step, IMD_get_step(ir->imd->setup));
+ }
+
+ if ((bX || bF || bIMDout || confout != NULL) && !DOMAINDECOMP(cr))
+ {
+ copy_em_coords(state, state_global);
+ }
+
+ mdof_flags = 0;
+ if (bX)
+ {
+ mdof_flags |= MDOF_X;
+ }
+ if (bF)
+ {
+ mdof_flags |= MDOF_F;
+ }
+
+ /* If we want IMD output, set appropriate MDOF flag */
+ if (ir->bIMD)
+ {
+ mdof_flags |= MDOF_IMD;
+ }
+
+ mdoutf_write_to_trajectory_files(fplog, cr, outf, mdof_flags,
+ top_global, step, (double)step,
+ &state->s, state_global, state->f);
+
+ if (confout != NULL && MASTER(cr))
+ {
+ if (ir->ePBC != epbcNONE && !ir->bPeriodicMols && DOMAINDECOMP(cr))
+ {
+ /* Make molecules whole only for confout writing */
+ do_pbc_mtop(fplog, ir->ePBC, state_global->box, top_global,
+ state_global->x);
+ }
+
+ write_sto_conf_mtop(confout,
+ *top_global->name, top_global,
+ state_global->x, NULL, ir->ePBC, state_global->box);
+ }
+}
+
+//! Do one minimization step
+static void do_em_step(t_commrec *cr, t_inputrec *ir, t_mdatoms *md,
+ gmx_bool bMolPBC,
+ em_state_t *ems1, real a, rvec *f, em_state_t *ems2,
+ gmx_constr_t constr, gmx_localtop_t *top,
+ t_nrnb *nrnb, gmx_wallcycle_t wcycle,
+ gmx_int64_t count)
+
+{
+ t_state *s1, *s2;
+ int i;
+ int start, end;
+ rvec *x1, *x2;
+ real dvdl_constr;
+ int nthreads gmx_unused;
+
+ s1 = &ems1->s;
+ s2 = &ems2->s;
+
+ if (DOMAINDECOMP(cr) && s1->ddp_count != cr->dd->ddp_count)
+ {
+ gmx_incons("state mismatch in do_em_step");
+ }
+
+ s2->flags = s1->flags;
+
+ if (s2->nalloc != s1->nalloc)
+ {
+ s2->nalloc = s1->nalloc;
+ /* We need to allocate one element extra, since we might use
+ * (unaligned) 4-wide SIMD loads to access rvec entries.
+ */
+ srenew(s2->x, s1->nalloc + 1);
+ srenew(ems2->f, s1->nalloc);
+ if (s2->flags & (1<<estCGP))
+ {
+ srenew(s2->cg_p, s1->nalloc + 1);
+ }
+ }
+
+ s2->natoms = s1->natoms;
+ copy_mat(s1->box, s2->box);
+ /* Copy free energy state */
+ for (i = 0; i < efptNR; i++)
+ {
+ s2->lambda[i] = s1->lambda[i];
+ }
+ copy_mat(s1->box, s2->box);
+
+ start = 0;
+ end = md->homenr;
+
+ x1 = s1->x;
+ x2 = s2->x;
+
+ // cppcheck-suppress unreadVariable
+ nthreads = gmx_omp_nthreads_get(emntUpdate);
+#pragma omp parallel num_threads(nthreads)
+ {
+ int gf, i, m;
+
+ gf = 0;
+#pragma omp for schedule(static) nowait
+ for (i = start; i < end; i++)
+ {
+ try
+ {
+ if (md->cFREEZE)
+ {
+ gf = md->cFREEZE[i];
+ }
+ for (m = 0; m < DIM; m++)
+ {
+ if (ir->opts.nFreeze[gf][m])
+ {
+ x2[i][m] = x1[i][m];
+ }
+ else
+ {
+ x2[i][m] = x1[i][m] + a*f[i][m];
+ }
+ }
+ }
+ GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+ }
+
+ if (s2->flags & (1<<estCGP))
+ {
+ /* Copy the CG p vector */
+ x1 = s1->cg_p;
+ x2 = s2->cg_p;
+#pragma omp for schedule(static) nowait
+ for (i = start; i < end; i++)
+ {
+ // Trivial OpenMP block that does not throw
+ copy_rvec(x1[i], x2[i]);
+ }
+ }
+
+ if (DOMAINDECOMP(cr))
+ {
+ s2->ddp_count = s1->ddp_count;
+ if (s2->cg_gl_nalloc < s1->cg_gl_nalloc)
+ {
+#pragma omp barrier
+ s2->cg_gl_nalloc = s1->cg_gl_nalloc;
+ try
+ {
+ /* We need to allocate one element extra, since we might use
+ * (unaligned) 4-wide SIMD loads to access rvec entries.
+ */
+ srenew(s2->cg_gl, s2->cg_gl_nalloc + 1);
+ }
+ GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+#pragma omp barrier
+ }
+ s2->ncg_gl = s1->ncg_gl;
+#pragma omp for schedule(static) nowait
+ for (i = 0; i < s2->ncg_gl; i++)
+ {
+ s2->cg_gl[i] = s1->cg_gl[i];
+ }
+ s2->ddp_count_cg_gl = s1->ddp_count_cg_gl;
+ }
+ }
+
+ if (constr)
+ {
+ wallcycle_start(wcycle, ewcCONSTR);
+ dvdl_constr = 0;
+ constrain(NULL, TRUE, TRUE, constr, &top->idef,
+ ir, cr, count, 0, 1.0, md,
+ s1->x, s2->x, NULL, bMolPBC, s2->box,
+ s2->lambda[efptBONDED], &dvdl_constr,
+ NULL, NULL, nrnb, econqCoord);
+ wallcycle_stop(wcycle, ewcCONSTR);
+ }
+}
+
+//! Prepare EM for using domain decomposition parallellization
+static void em_dd_partition_system(FILE *fplog, int step, t_commrec *cr,
+ gmx_mtop_t *top_global, t_inputrec *ir,
+ em_state_t *ems, gmx_localtop_t *top,
+ t_mdatoms *mdatoms, t_forcerec *fr,
+ gmx_vsite_t *vsite, gmx_constr_t constr,
+ t_nrnb *nrnb, gmx_wallcycle_t wcycle)
+{
+ /* Repartition the domain decomposition */
+ dd_partition_system(fplog, step, cr, FALSE, 1,
+ NULL, top_global, ir,
+ &ems->s, &ems->f,
+ mdatoms, top, fr, vsite, constr,
+ nrnb, wcycle, FALSE);
+ dd_store_state(cr->dd, &ems->s);
+}
+
+//! De one energy evaluation
+static void evaluate_energy(FILE *fplog, t_commrec *cr,
+ gmx_mtop_t *top_global,
+ em_state_t *ems, gmx_localtop_t *top,
+ t_inputrec *inputrec,
+ t_nrnb *nrnb, gmx_wallcycle_t wcycle,
+ gmx_global_stat_t gstat,
+ gmx_vsite_t *vsite, gmx_constr_t constr,
+ t_fcdata *fcd,
+ t_graph *graph, t_mdatoms *mdatoms,
+ t_forcerec *fr, rvec mu_tot,
+ gmx_enerdata_t *enerd, tensor vir, tensor pres,
+ gmx_int64_t count, gmx_bool bFirst)
+{
+ real t;
+ gmx_bool bNS;
+ tensor force_vir, shake_vir, ekin;
+ real dvdl_constr, prescorr, enercorr, dvdlcorr;
+ real terminate = 0;
+
+ /* Set the time to the initial time, the time does not change during EM */
+ t = inputrec->init_t;
+
+ if (bFirst ||
+ (DOMAINDECOMP(cr) && ems->s.ddp_count < cr->dd->ddp_count))
+ {
+ /* This is the first state or an old state used before the last ns */
+ bNS = TRUE;
+ }
+ else
+ {
+ bNS = FALSE;
+ if (inputrec->nstlist > 0)
+ {
+ bNS = TRUE;
+ }
+ }
+
+ if (vsite)
+ {
+ construct_vsites(vsite, ems->s.x, 1, NULL,
+ top->idef.iparams, top->idef.il,
+ fr->ePBC, fr->bMolPBC, cr, ems->s.box);
+ }
+
+ if (DOMAINDECOMP(cr) && bNS)
+ {
+ /* Repartition the domain decomposition */
+ em_dd_partition_system(fplog, count, cr, top_global, inputrec,
+ ems, top, mdatoms, fr, vsite, constr,
+ nrnb, wcycle);
+ }
+
+ /* Calc force & energy on new trial position */
+ /* do_force always puts the charge groups in the box and shifts again
+ * We do not unshift, so molecules are always whole in congrad.c
+ */
+ do_force(fplog, cr, inputrec,
+ count, nrnb, wcycle, top, &top_global->groups,
+ ems->s.box, ems->s.x, &ems->s.hist,
+ ems->f, force_vir, mdatoms, enerd, fcd,
+ ems->s.lambda, graph, fr, vsite, mu_tot, t, NULL, NULL, TRUE,
+ GMX_FORCE_STATECHANGED | GMX_FORCE_ALLFORCES |
+ GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY |
+ (bNS ? GMX_FORCE_NS : 0));
+
+ /* Clear the unused shake virial and pressure */
+ clear_mat(shake_vir);
+ clear_mat(pres);
+
+ /* Communicate stuff when parallel */
+ if (PAR(cr) && inputrec->eI != eiNM)
+ {
+ wallcycle_start(wcycle, ewcMoveE);
+
+ global_stat(gstat, cr, enerd, force_vir, shake_vir, mu_tot,
+ inputrec, NULL, NULL, NULL, 1, &terminate,
+ NULL, FALSE,
+ CGLO_ENERGY |
+ CGLO_PRESSURE |
+ CGLO_CONSTRAINT);
+
+ wallcycle_stop(wcycle, ewcMoveE);
+ }
+
+ /* Calculate long range corrections to pressure and energy */
+ calc_dispcorr(inputrec, fr, ems->s.box, ems->s.lambda[efptVDW],
+ pres, force_vir, &prescorr, &enercorr, &dvdlcorr);
+ enerd->term[F_DISPCORR] = enercorr;
+ enerd->term[F_EPOT] += enercorr;
+ enerd->term[F_PRES] += prescorr;
+ enerd->term[F_DVDL] += dvdlcorr;
+
+ ems->epot = enerd->term[F_EPOT];
+
+ if (constr)
+ {
+ /* Project out the constraint components of the force */
+ wallcycle_start(wcycle, ewcCONSTR);
+ dvdl_constr = 0;
+ constrain(NULL, FALSE, FALSE, constr, &top->idef,
+ inputrec, cr, count, 0, 1.0, mdatoms,
+ ems->s.x, ems->f, ems->f, fr->bMolPBC, ems->s.box,
+ ems->s.lambda[efptBONDED], &dvdl_constr,
+ NULL, &shake_vir, nrnb, econqForceDispl);
+ enerd->term[F_DVDL_CONSTR] += dvdl_constr;
+ m_add(force_vir, shake_vir, vir);
+ wallcycle_stop(wcycle, ewcCONSTR);
+ }
+ else
+ {
+ copy_mat(force_vir, vir);
+ }
+
+ clear_mat(ekin);
+ enerd->term[F_PRES] =
+ calc_pres(fr->ePBC, inputrec->nwall, ems->s.box, ekin, vir, pres);
+
+ sum_dhdl(enerd, ems->s.lambda, inputrec->fepvals);
+
+ if (EI_ENERGY_MINIMIZATION(inputrec->eI))
+ {
+ get_state_f_norm_max(cr, &(inputrec->opts), mdatoms, ems);
+ }
+}
+
+//! Parallel utility summing energies and forces
+static double reorder_partsum(t_commrec *cr, t_grpopts *opts, t_mdatoms *mdatoms,
+ gmx_mtop_t *top_global,
+ em_state_t *s_min, em_state_t *s_b)
+{
+ rvec *fm, *fb, *fmg;
+ t_block *cgs_gl;
+ int ncg, *cg_gl, *index, c, cg, i, a0, a1, a, gf, m;
+ double partsum;
+ unsigned char *grpnrFREEZE;
+
+ if (debug)
+ {
+ fprintf(debug, "Doing reorder_partsum\n");
+ }
+
+ fm = s_min->f;
+ fb = s_b->f;
+
+ cgs_gl = dd_charge_groups_global(cr->dd);
+ index = cgs_gl->index;
+
+ /* Collect fm in a global vector fmg.
+ * This conflicts with the spirit of domain decomposition,
+ * but to fully optimize this a much more complicated algorithm is required.
+ */
+ snew(fmg, top_global->natoms);
+
+ ncg = s_min->s.ncg_gl;
+ cg_gl = s_min->s.cg_gl;
+ i = 0;
+ for (c = 0; c < ncg; c++)
+ {
+ cg = cg_gl[c];
+ a0 = index[cg];
+ a1 = index[cg+1];
+ for (a = a0; a < a1; a++)
+ {
+ copy_rvec(fm[i], fmg[a]);
+ i++;
+ }
+ }
+ gmx_sum(top_global->natoms*3, fmg[0], cr);
+
+ /* Now we will determine the part of the sum for the cgs in state s_b */
+ ncg = s_b->s.ncg_gl;
+ cg_gl = s_b->s.cg_gl;
+ partsum = 0;
+ i = 0;
+ gf = 0;
+ grpnrFREEZE = top_global->groups.grpnr[egcFREEZE];
+ for (c = 0; c < ncg; c++)
+ {
+ cg = cg_gl[c];
+ a0 = index[cg];
+ a1 = index[cg+1];
+ for (a = a0; a < a1; a++)
+ {
+ if (mdatoms->cFREEZE && grpnrFREEZE)
+ {
+ gf = grpnrFREEZE[i];
+ }
+ for (m = 0; m < DIM; m++)
+ {
+ if (!opts->nFreeze[gf][m])
+ {
+ partsum += (fb[i][m] - fmg[a][m])*fb[i][m];
+ }
+ }
+ i++;
+ }
+ }
+
+ sfree(fmg);
+
+ return partsum;
+}
+
+//! Print some stuff, like beta, whatever that means.
+static real pr_beta(t_commrec *cr, t_grpopts *opts, t_mdatoms *mdatoms,
+ gmx_mtop_t *top_global,
+ em_state_t *s_min, em_state_t *s_b)
+{
+ rvec *fm, *fb;
+ double sum;
+ int gf, i, m;
+
+ /* This is just the classical Polak-Ribiere calculation of beta;
+ * it looks a bit complicated since we take freeze groups into account,
+ * and might have to sum it in parallel runs.
+ */
+
+ if (!DOMAINDECOMP(cr) ||
+ (s_min->s.ddp_count == cr->dd->ddp_count &&
+ s_b->s.ddp_count == cr->dd->ddp_count))
+ {
+ fm = s_min->f;
+ fb = s_b->f;
+ sum = 0;
+ gf = 0;
+ /* This part of code can be incorrect with DD,
+ * since the atom ordering in s_b and s_min might differ.
+ */
+ for (i = 0; i < mdatoms->homenr; i++)
+ {
+ if (mdatoms->cFREEZE)
+ {
+ gf = mdatoms->cFREEZE[i];
+ }
+ for (m = 0; m < DIM; m++)
+ {
+ if (!opts->nFreeze[gf][m])
+ {
+ sum += (fb[i][m] - fm[i][m])*fb[i][m];
+ }
+ }
+ }
+ }
+ else
+ {
+ /* We need to reorder cgs while summing */
+ sum = reorder_partsum(cr, opts, mdatoms, top_global, s_min, s_b);
+ }
+ if (PAR(cr))
+ {
+ gmx_sumd(1, &sum, cr);
+ }
+
+ return sum/gmx::square(s_min->fnorm);
+}
+
+namespace gmx
+{
+
+/*! \brief Do conjugate gradients minimization
+ \copydoc integrator_t (FILE *fplog, t_commrec *cr,
+ int nfile, const t_filenm fnm[],
+ const gmx_output_env_t *oenv, gmx_bool bVerbose,
+ int nstglobalcomm,
+ gmx_vsite_t *vsite, gmx_constr_t constr,
+ int stepout,
+ t_inputrec *inputrec,
+ 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_nex, int repl_ex_seed,
+ real cpt_period, real max_hours,
+ int imdport,
+ unsigned long Flags,
+ gmx_walltime_accounting_t walltime_accounting)
+ */
+double do_cg(FILE *fplog, t_commrec *cr,
+ int nfile, const t_filenm fnm[],
+ const gmx_output_env_t gmx_unused *oenv, gmx_bool bVerbose,
+ int gmx_unused nstglobalcomm,
+ gmx_vsite_t *vsite, gmx_constr_t constr,
+ int gmx_unused stepout,
+ t_inputrec *inputrec,
+ 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 gmx_unused ed,
+ t_forcerec *fr,
+ int gmx_unused repl_ex_nst, int gmx_unused repl_ex_nex, int gmx_unused repl_ex_seed,
+ real gmx_unused cpt_period, real gmx_unused max_hours,
+ int imdport,
+ unsigned long gmx_unused Flags,
+ gmx_walltime_accounting_t walltime_accounting)
+{
+ const char *CG = "Polak-Ribiere Conjugate Gradients";
+
+ em_state_t *s_min, *s_a, *s_b, *s_c;
+ gmx_localtop_t *top;
+ gmx_enerdata_t *enerd;
+ rvec *f;
+ gmx_global_stat_t gstat;
+ t_graph *graph;
+ rvec *p, *sf;
+ double gpa, gpb, gpc, tmp, minstep;
+ real fnormn;
+ real stepsize;
+ real a, b, c, beta = 0.0;
+ real epot_repl = 0;
+ real pnorm;
+ t_mdebin *mdebin;
+ gmx_bool converged, foundlower;
+ rvec mu_tot;
+ gmx_bool do_log = FALSE, do_ene = FALSE, do_x, do_f;
+ tensor vir, pres;
+ int number_steps, neval = 0, nstcg = inputrec->nstcgsteep;
+ gmx_mdoutf_t outf;
+ int i, m, gf, step, nminstep;
+
+ step = 0;
+
+ s_min = init_em_state();
+ s_a = init_em_state();
+ s_b = init_em_state();
+ s_c = init_em_state();
+
+ /* Init em and store the local state in s_min */
+ init_em(fplog, CG, cr, inputrec,
+ state_global, top_global, s_min, &top, &f,
+ nrnb, mu_tot, fr, &enerd, &graph, mdatoms, &gstat, vsite, constr,
+ nfile, fnm, &outf, &mdebin, imdport, Flags, wcycle);
+
+ /* Print to log file */
+ print_em_start(fplog, cr, walltime_accounting, wcycle, CG);
+
+ /* Max number of steps */
+ number_steps = inputrec->nsteps;
+
+ if (MASTER(cr))
+ {
+ sp_header(stderr, CG, inputrec->em_tol, number_steps);
+ }
+ if (fplog)
+ {
+ sp_header(fplog, CG, inputrec->em_tol, number_steps);
+ }
+
+ /* Call the force routine and some auxiliary (neighboursearching etc.) */
+ /* do_force always puts the charge groups in the box and shifts again
+ * We do not unshift, so molecules are always whole in congrad.c
+ */
+ evaluate_energy(fplog, cr,
+ top_global, s_min, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph, mdatoms, fr,
+ mu_tot, enerd, vir, pres, -1, TRUE);
+ where();
+
+ if (MASTER(cr))
+ {
+ /* Copy stuff to the energy bin for easy printing etc. */
+ upd_mdebin(mdebin, FALSE, FALSE, (double)step,
+ mdatoms->tmass, enerd, &s_min->s, inputrec->fepvals, inputrec->expandedvals, s_min->s.box,
+ NULL, NULL, vir, pres, NULL, mu_tot, constr);
+
+ print_ebin_header(fplog, step, step);
+ print_ebin(mdoutf_get_fp_ene(outf), TRUE, FALSE, FALSE, fplog, step, step, eprNORMAL,
+ mdebin, fcd, &(top_global->groups), &(inputrec->opts));
+ }
+ where();
+
+ /* Estimate/guess the initial stepsize */
+ stepsize = inputrec->em_stepsize/s_min->fnorm;
+
+ if (MASTER(cr))
+ {
+ double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
+ fprintf(stderr, " F-max = %12.5e on atom %d\n",
+ s_min->fmax, s_min->a_fmax+1);
+ fprintf(stderr, " F-Norm = %12.5e\n",
+ s_min->fnorm/sqrtNumAtoms);
+ fprintf(stderr, "\n");
+ /* and copy to the log file too... */
+ fprintf(fplog, " F-max = %12.5e on atom %d\n",
+ s_min->fmax, s_min->a_fmax+1);
+ fprintf(fplog, " F-Norm = %12.5e\n",
+ s_min->fnorm/sqrtNumAtoms);
+ fprintf(fplog, "\n");
+ }
+ /* Start the loop over CG steps.
+ * Each successful step is counted, and we continue until
+ * we either converge or reach the max number of steps.
+ */
+ converged = FALSE;
+ for (step = 0; (number_steps < 0 || step <= number_steps) && !converged; step++)
+ {
+
+ /* start taking steps in a new direction
+ * First time we enter the routine, beta=0, and the direction is
+ * simply the negative gradient.
+ */
+
+ /* Calculate the new direction in p, and the gradient in this direction, gpa */
+ p = s_min->s.cg_p;
+ sf = s_min->f;
+ gpa = 0;
+ gf = 0;
+ for (i = 0; i < mdatoms->homenr; i++)
+ {
+ if (mdatoms->cFREEZE)
+ {
+ gf = mdatoms->cFREEZE[i];
+ }
+ for (m = 0; m < DIM; m++)
+ {
+ if (!inputrec->opts.nFreeze[gf][m])
+ {
+ p[i][m] = sf[i][m] + beta*p[i][m];
+ gpa -= p[i][m]*sf[i][m];
+ /* f is negative gradient, thus the sign */
+ }
+ else
+ {
+ p[i][m] = 0;
+ }
+ }
+ }
+
+ /* Sum the gradient along the line across CPUs */
+ if (PAR(cr))
+ {
+ gmx_sumd(1, &gpa, cr);
+ }
+
+ /* Calculate the norm of the search vector */
+ get_f_norm_max(cr, &(inputrec->opts), mdatoms, p, &pnorm, NULL, NULL);
+
+ /* Just in case stepsize reaches zero due to numerical precision... */
+ if (stepsize <= 0)
+ {
+ stepsize = inputrec->em_stepsize/pnorm;
+ }
+
+ /*
+ * Double check the value of the derivative in the search direction.
+ * If it is positive it must be due to the old information in the
+ * CG formula, so just remove that and start over with beta=0.
+ * This corresponds to a steepest descent step.
+ */
+ if (gpa > 0)
+ {
+ beta = 0;
+ step--; /* Don't count this step since we are restarting */
+ continue; /* Go back to the beginning of the big for-loop */
+ }
+
+ /* Calculate minimum allowed stepsize, before the average (norm)
+ * relative change in coordinate is smaller than precision
+ */
+ minstep = 0;
+ for (i = 0; i < mdatoms->homenr; i++)
+ {
+ for (m = 0; m < DIM; m++)
+ {
+ tmp = fabs(s_min->s.x[i][m]);
+ if (tmp < 1.0)
+ {
+ tmp = 1.0;
+ }
+ tmp = p[i][m]/tmp;
+ minstep += tmp*tmp;
+ }
+ }
+ /* Add up from all CPUs */
+ if (PAR(cr))
+ {
+ gmx_sumd(1, &minstep, cr);
+ }
+
+ minstep = GMX_REAL_EPS/sqrt(minstep/(3*state_global->natoms));
+
+ if (stepsize < minstep)
+ {
+ converged = TRUE;
+ break;
+ }
+
+ /* Write coordinates if necessary */
+ do_x = do_per_step(step, inputrec->nstxout);
+ do_f = do_per_step(step, inputrec->nstfout);
+
+ write_em_traj(fplog, cr, outf, do_x, do_f, NULL,
+ top_global, inputrec, step,
+ s_min, state_global);
+
+ /* Take a step downhill.
+ * In theory, we should minimize the function along this direction.
+ * That is quite possible, but it turns out to take 5-10 function evaluations
+ * for each line. However, we dont really need to find the exact minimum -
+ * it is much better to start a new CG step in a modified direction as soon
+ * as we are close to it. This will save a lot of energy evaluations.
+ *
+ * In practice, we just try to take a single step.
+ * If it worked (i.e. lowered the energy), we increase the stepsize but
+ * the continue straight to the next CG step without trying to find any minimum.
+ * If it didn't work (higher energy), there must be a minimum somewhere between
+ * the old position and the new one.
+ *
+ * Due to the finite numerical accuracy, it turns out that it is a good idea
+ * to even accept a SMALL increase in energy, if the derivative is still downhill.
+ * This leads to lower final energies in the tests I've done. / Erik
+ */
+ s_a->epot = s_min->epot;
+ a = 0.0;
+ c = a + stepsize; /* reference position along line is zero */
+
+ if (DOMAINDECOMP(cr) && s_min->s.ddp_count < cr->dd->ddp_count)
+ {
+ em_dd_partition_system(fplog, step, cr, top_global, inputrec,
+ s_min, top, mdatoms, fr, vsite, constr,
+ nrnb, wcycle);
+ }
+
+ /* Take a trial step (new coords in s_c) */
+ do_em_step(cr, inputrec, mdatoms, fr->bMolPBC, s_min, c, s_min->s.cg_p, s_c,
+ constr, top, nrnb, wcycle, -1);
+
+ neval++;
+ /* Calculate energy for the trial step */
+ evaluate_energy(fplog, cr,
+ top_global, s_c, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph, mdatoms, fr,
+ mu_tot, enerd, vir, pres, -1, FALSE);
+
+ /* Calc derivative along line */
+ p = s_c->s.cg_p;
+ sf = s_c->f;
+ gpc = 0;
+ for (i = 0; i < mdatoms->homenr; i++)
+ {
+ for (m = 0; m < DIM; m++)
+ {
+ gpc -= p[i][m]*sf[i][m]; /* f is negative gradient, thus the sign */
+ }
+ }
+ /* Sum the gradient along the line across CPUs */
+ if (PAR(cr))
+ {
+ gmx_sumd(1, &gpc, cr);
+ }
+
+ /* This is the max amount of increase in energy we tolerate */
+ tmp = sqrt(GMX_REAL_EPS)*fabs(s_a->epot);
+
+ /* Accept the step if the energy is lower, or if it is not significantly higher
+ * and the line derivative is still negative.
+ */
+ if (s_c->epot < s_a->epot || (gpc < 0 && s_c->epot < (s_a->epot + tmp)))
+ {
+ foundlower = TRUE;
+ /* Great, we found a better energy. Increase step for next iteration
+ * if we are still going down, decrease it otherwise
+ */
+ if (gpc < 0)
+ {
+ stepsize *= 1.618034; /* The golden section */
+ }
+ else
+ {
+ stepsize *= 0.618034; /* 1/golden section */
+ }
+ }
+ else
+ {
+ /* New energy is the same or higher. We will have to do some work
+ * to find a smaller value in the interval. Take smaller step next time!
+ */
+ foundlower = FALSE;
+ stepsize *= 0.618034;
+ }
+
+
+
+
+ /* OK, if we didn't find a lower value we will have to locate one now - there must
+ * be one in the interval [a=0,c].
+ * The same thing is valid here, though: Don't spend dozens of iterations to find
+ * the line minimum. We try to interpolate based on the derivative at the endpoints,
+ * and only continue until we find a lower value. In most cases this means 1-2 iterations.
+ *
+ * I also have a safeguard for potentially really pathological functions so we never
+ * take more than 20 steps before we give up ...
+ *
+ * If we already found a lower value we just skip this step and continue to the update.
+ */
+ if (!foundlower)
+ {
+ nminstep = 0;
+
+ do
+ {
+ /* Select a new trial point.
+ * If the derivatives at points a & c have different sign we interpolate to zero,
+ * otherwise just do a bisection.
+ */
+ if (gpa < 0 && gpc > 0)
+ {
+ b = a + gpa*(a-c)/(gpc-gpa);
+ }
+ else
+ {
+ b = 0.5*(a+c);
+ }
+
+ /* safeguard if interpolation close to machine accuracy causes errors:
+ * never go outside the interval
+ */
+ if (b <= a || b >= c)
+ {
+ b = 0.5*(a+c);
+ }
+
+ if (DOMAINDECOMP(cr) && s_min->s.ddp_count != cr->dd->ddp_count)
+ {
+ /* Reload the old state */
+ em_dd_partition_system(fplog, -1, cr, top_global, inputrec,
+ s_min, top, mdatoms, fr, vsite, constr,
+ nrnb, wcycle);
+ }
+
+ /* Take a trial step to this new point - new coords in s_b */
+ do_em_step(cr, inputrec, mdatoms, fr->bMolPBC, s_min, b, s_min->s.cg_p, s_b,
+ constr, top, nrnb, wcycle, -1);
+
+ neval++;
+ /* Calculate energy for the trial step */
+ evaluate_energy(fplog, cr,
+ top_global, s_b, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph, mdatoms, fr,
+ mu_tot, enerd, vir, pres, -1, FALSE);
+
+ /* p does not change within a step, but since the domain decomposition
+ * might change, we have to use cg_p of s_b here.
+ */
+ p = s_b->s.cg_p;
+ sf = s_b->f;
+ gpb = 0;
+ for (i = 0; i < mdatoms->homenr; i++)
+ {
+ for (m = 0; m < DIM; m++)
+ {
+ gpb -= p[i][m]*sf[i][m]; /* f is negative gradient, thus the sign */
+ }
+ }
+ /* Sum the gradient along the line across CPUs */
+ if (PAR(cr))
+ {
+ gmx_sumd(1, &gpb, cr);
+ }
+
+ if (debug)
+ {
+ fprintf(debug, "CGE: EpotA %f EpotB %f EpotC %f gpb %f\n",
+ s_a->epot, s_b->epot, s_c->epot, gpb);
+ }
+
+ epot_repl = s_b->epot;
+
+ /* Keep one of the intervals based on the value of the derivative at the new point */
+ if (gpb > 0)
+ {
+ /* Replace c endpoint with b */
+ swap_em_state(s_b, s_c);
+ c = b;
+ gpc = gpb;
+ }
+ else
+ {
+ /* Replace a endpoint with b */
+ swap_em_state(s_b, s_a);
+ a = b;
+ gpa = gpb;
+ }
+
+ /*
+ * Stop search as soon as we find a value smaller than the endpoints.
+ * Never run more than 20 steps, no matter what.
+ */
+ nminstep++;
+ }
+ while ((epot_repl > s_a->epot || epot_repl > s_c->epot) &&
+ (nminstep < 20));
+
+ if (fabs(epot_repl - s_min->epot) < fabs(s_min->epot)*GMX_REAL_EPS ||
+ nminstep >= 20)
+ {
+ /* OK. We couldn't find a significantly lower energy.
+ * If beta==0 this was steepest descent, and then we give up.
+ * If not, set beta=0 and restart with steepest descent before quitting.
+ */
+ if (beta == 0.0)
+ {
+ /* Converged */
+ converged = TRUE;
+ break;
+ }
+ else
+ {
+ /* Reset memory before giving up */
+ beta = 0.0;
+ continue;
+ }
+ }
+
+ /* Select min energy state of A & C, put the best in B.
+ */
+ if (s_c->epot < s_a->epot)
+ {
+ if (debug)
+ {
+ fprintf(debug, "CGE: C (%f) is lower than A (%f), moving C to B\n",
+ s_c->epot, s_a->epot);
+ }
+ swap_em_state(s_b, s_c);
+ gpb = gpc;
+ }
+ else
+ {
+ if (debug)
+ {
+ fprintf(debug, "CGE: A (%f) is lower than C (%f), moving A to B\n",
+ s_a->epot, s_c->epot);
+ }
+ swap_em_state(s_b, s_a);
+ gpb = gpa;
+ }
+
+ }
+ else
+ {
+ if (debug)
+ {
+ fprintf(debug, "CGE: Found a lower energy %f, moving C to B\n",
+ s_c->epot);
+ }
+ swap_em_state(s_b, s_c);
+ gpb = gpc;
+ }
+
+ /* new search direction */
+ /* beta = 0 means forget all memory and restart with steepest descents. */
+ if (nstcg && ((step % nstcg) == 0))
+ {
+ beta = 0.0;
+ }
+ else
+ {
+ /* s_min->fnorm cannot be zero, because then we would have converged
+ * and broken out.
+ */
+
+ /* Polak-Ribiere update.
+ * Change to fnorm2/fnorm2_old for Fletcher-Reeves
+ */
+ beta = pr_beta(cr, &inputrec->opts, mdatoms, top_global, s_min, s_b);
+ }
+ /* Limit beta to prevent oscillations */
+ if (fabs(beta) > 5.0)
+ {
+ beta = 0.0;
+ }
+
+
+ /* update positions */
+ swap_em_state(s_min, s_b);
+ gpa = gpb;
+
+ /* Print it if necessary */
+ if (MASTER(cr))
+ {
+ if (bVerbose)
+ {
+ double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
+ fprintf(stderr, "\rStep %d, Epot=%12.6e, Fnorm=%9.3e, Fmax=%9.3e (atom %d)\n",
+ step, s_min->epot, s_min->fnorm/sqrtNumAtoms,
+ s_min->fmax, s_min->a_fmax+1);
+ fflush(stderr);
+ }
+ /* Store the new (lower) energies */
+ upd_mdebin(mdebin, FALSE, FALSE, (double)step,
+ mdatoms->tmass, enerd, &s_min->s, inputrec->fepvals, inputrec->expandedvals, s_min->s.box,
+ NULL, NULL, vir, pres, NULL, mu_tot, constr);
+
+ do_log = do_per_step(step, inputrec->nstlog);
+ do_ene = do_per_step(step, inputrec->nstenergy);
+
+ /* Prepare IMD energy record, if bIMD is TRUE. */
+ IMD_fill_energy_record(inputrec->bIMD, inputrec->imd, enerd, step, TRUE);
+
+ if (do_log)
+ {
+ print_ebin_header(fplog, step, step);
+ }
+ print_ebin(mdoutf_get_fp_ene(outf), do_ene, FALSE, FALSE,
+ do_log ? fplog : NULL, step, step, eprNORMAL,
+ mdebin, fcd, &(top_global->groups), &(inputrec->opts));
+ }
+
+ /* Send energies and positions to the IMD client if bIMD is TRUE. */
+ if (do_IMD(inputrec->bIMD, step, cr, TRUE, state_global->box, state_global->x, inputrec, 0, wcycle) && MASTER(cr))
+ {
+ IMD_send_positions(inputrec->imd);
+ }
+
+ /* Stop when the maximum force lies below tolerance.
+ * If we have reached machine precision, converged is already set to true.
+ */
+ converged = converged || (s_min->fmax < inputrec->em_tol);
+
+ } /* End of the loop */
+
+ /* IMD cleanup, if bIMD is TRUE. */
+ IMD_finalize(inputrec->bIMD, inputrec->imd);
+
+ if (converged)
+ {
+ step--; /* we never took that last step in this case */
+
+ }
+ if (s_min->fmax > inputrec->em_tol)
+ {
+ if (MASTER(cr))
+ {
+ warn_step(stderr, inputrec->em_tol, step-1 == number_steps, FALSE);
+ warn_step(fplog, inputrec->em_tol, step-1 == number_steps, FALSE);
+ }
+ converged = FALSE;
+ }
+
+ if (MASTER(cr))
+ {
+ /* If we printed energy and/or logfile last step (which was the last step)
+ * we don't have to do it again, but otherwise print the final values.
+ */
+ if (!do_log)
+ {
+ /* Write final value to log since we didn't do anything the last step */
+ print_ebin_header(fplog, step, step);
+ }
+ if (!do_ene || !do_log)
+ {
+ /* Write final energy file entries */
+ print_ebin(mdoutf_get_fp_ene(outf), !do_ene, FALSE, FALSE,
+ !do_log ? fplog : NULL, step, step, eprNORMAL,
+ mdebin, fcd, &(top_global->groups), &(inputrec->opts));
+ }
+ }
+
+ /* Print some stuff... */
+ if (MASTER(cr))
+ {
+ fprintf(stderr, "\nwriting lowest energy coordinates.\n");
+ }
+
+ /* IMPORTANT!
+ * For accurate normal mode calculation it is imperative that we
+ * store the last conformation into the full precision binary trajectory.
+ *
+ * However, we should only do it if we did NOT already write this step
+ * above (which we did if do_x or do_f was true).
+ */
+ do_x = !do_per_step(step, inputrec->nstxout);
+ do_f = (inputrec->nstfout > 0 && !do_per_step(step, inputrec->nstfout));
+
+ write_em_traj(fplog, cr, outf, do_x, do_f, ftp2fn(efSTO, nfile, fnm),
+ top_global, inputrec, step,
+ s_min, state_global);
+
+
+ if (MASTER(cr))
+ {
+ double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
+ fnormn = s_min->fnorm/sqrtNumAtoms;
+ print_converged(stderr, CG, inputrec->em_tol, step, converged, number_steps,
+ s_min->epot, s_min->fmax, s_min->a_fmax, fnormn);
+ print_converged(fplog, CG, inputrec->em_tol, step, converged, number_steps,
+ s_min->epot, s_min->fmax, s_min->a_fmax, fnormn);
+
+ fprintf(fplog, "\nPerformed %d energy evaluations in total.\n", neval);
+ }
+
+ finish_em(cr, outf, walltime_accounting, wcycle);
+
+ /* To print the actual number of steps we needed somewhere */
+ walltime_accounting_set_nsteps_done(walltime_accounting, step);
+
+ return 0;
+} /* That's all folks */
+
+
+/*! \brief Do L-BFGS conjugate gradients minimization
+ \copydoc integrator_t (FILE *fplog, t_commrec *cr,
+ int nfile, const t_filenm fnm[],
+ const gmx_output_env_t *oenv, gmx_bool bVerbose,
+ int nstglobalcomm,
+ gmx_vsite_t *vsite, gmx_constr_t constr,
+ int stepout,
+ t_inputrec *inputrec,
+ 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_nex, int repl_ex_seed,
+ real cpt_period, real max_hours,
+ int imdport,
+ unsigned long Flags,
+ gmx_walltime_accounting_t walltime_accounting)
+ */
+double do_lbfgs(FILE *fplog, t_commrec *cr,
+ int nfile, const t_filenm fnm[],
+ const gmx_output_env_t gmx_unused *oenv, gmx_bool bVerbose,
+ int gmx_unused nstglobalcomm,
+ gmx_vsite_t *vsite, gmx_constr_t constr,
+ int gmx_unused stepout,
+ t_inputrec *inputrec,
+ 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 gmx_unused ed,
+ t_forcerec *fr,
+ int gmx_unused repl_ex_nst, int gmx_unused repl_ex_nex, int gmx_unused repl_ex_seed,
+ real gmx_unused cpt_period, real gmx_unused max_hours,
+ int imdport,
+ unsigned long gmx_unused Flags,
+ gmx_walltime_accounting_t walltime_accounting)
+{
+ static const char *LBFGS = "Low-Memory BFGS Minimizer";
+ em_state_t ems;
+ gmx_localtop_t *top;
+ gmx_enerdata_t *enerd;
+ rvec *f;
+ gmx_global_stat_t gstat;
+ t_graph *graph;
+ int ncorr, nmaxcorr, point, cp, neval, nminstep;
+ double stepsize, step_taken, gpa, gpb, gpc, tmp, minstep;
+ real *rho, *alpha, *ff, *xx, *p, *s, *lastx, *lastf, **dx, **dg;
+ real *xa, *xb, *xc, *fa, *fb, *fc, *xtmp, *ftmp;
+ real a, b, c, maxdelta, delta;
+ real diag, Epot0, Epot, EpotA, EpotB, EpotC;
+ real dgdx, dgdg, sq, yr, beta;
+ t_mdebin *mdebin;
+ gmx_bool converged;
+ rvec mu_tot;
+ real fnorm, fmax;
+ gmx_bool do_log, do_ene, do_x, do_f, foundlower, *frozen;
+ tensor vir, pres;
+ int start, end, number_steps;
+ gmx_mdoutf_t outf;
+ int i, k, m, n, nfmax, gf, step;
+ int mdof_flags;
+
+ if (PAR(cr))
+ {
+ gmx_fatal(FARGS, "Cannot do parallel L-BFGS Minimization - yet.\n");
+ }
+
+ if (NULL != constr)
+ {
+ gmx_fatal(FARGS, "The combination of constraints and L-BFGS minimization is not implemented. Either do not use constraints, or use another minimizer (e.g. steepest descent).");
+ }
+
+ n = 3*state_global->natoms;
+ nmaxcorr = inputrec->nbfgscorr;
+
+ /* Allocate memory */
+ /* Use pointers to real so we dont have to loop over both atoms and
+ * dimensions all the time...
+ * x/f are allocated as rvec *, so make new x0/f0 pointers-to-real
+ * that point to the same memory.
+ */
+ snew(xa, n);
+ snew(xb, n);
+ snew(xc, n);
+ snew(fa, n);
+ snew(fb, n);
+ snew(fc, n);
+ snew(frozen, n);
+
+ snew(p, n);
+ snew(lastx, n);
+ snew(lastf, n);
+ snew(rho, nmaxcorr);
+ snew(alpha, nmaxcorr);
+
+ snew(dx, nmaxcorr);
+ for (i = 0; i < nmaxcorr; i++)
+ {
+ snew(dx[i], n);
+ }
+
+ snew(dg, nmaxcorr);
+ for (i = 0; i < nmaxcorr; i++)
+ {
+ snew(dg[i], n);
+ }
+
+ step = 0;
+ neval = 0;
+
+ /* Init em */
+ init_em(fplog, LBFGS, cr, inputrec,
+ state_global, top_global, &ems, &top, &f,
+ nrnb, mu_tot, fr, &enerd, &graph, mdatoms, &gstat, vsite, constr,
+ nfile, fnm, &outf, &mdebin, imdport, Flags, wcycle);
+ /* Do_lbfgs is not completely updated like do_steep and do_cg,
+ * so we free some memory again.
+ */
+ sfree(ems.s.x);
+ sfree(ems.f);
+
+ xx = (real *)state_global->x;
+ ff = (real *)f;
+
+ start = 0;
+ end = mdatoms->homenr;
+
+ /* Print to log file */
+ print_em_start(fplog, cr, walltime_accounting, wcycle, LBFGS);
+
+ do_log = do_ene = do_x = do_f = TRUE;
+
+ /* Max number of steps */
+ number_steps = inputrec->nsteps;
+
+ /* Create a 3*natoms index to tell whether each degree of freedom is frozen */
+ gf = 0;
+ for (i = start; i < end; i++)
+ {
+ if (mdatoms->cFREEZE)
+ {
+ gf = mdatoms->cFREEZE[i];
+ }
+ for (m = 0; m < DIM; m++)
+ {
+ frozen[3*i+m] = inputrec->opts.nFreeze[gf][m];
+ }
+ }
+ if (MASTER(cr))
+ {
+ sp_header(stderr, LBFGS, inputrec->em_tol, number_steps);
+ }
+ if (fplog)
+ {
+ sp_header(fplog, LBFGS, inputrec->em_tol, number_steps);
+ }
+
+ if (vsite)
+ {
+ construct_vsites(vsite, state_global->x, 1, NULL,
+ top->idef.iparams, top->idef.il,
+ fr->ePBC, fr->bMolPBC, cr, state_global->box);
+ }
+
+ /* Call the force routine and some auxiliary (neighboursearching etc.) */
+ /* do_force always puts the charge groups in the box and shifts again
+ * We do not unshift, so molecules are always whole
+ */
+ neval++;
+ ems.s.x = state_global->x;
+ ems.f = f;
+ evaluate_energy(fplog, cr,
+ top_global, &ems, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph, mdatoms, fr,
+ mu_tot, enerd, vir, pres, -1, TRUE);
+ where();
+
+ if (MASTER(cr))
+ {
+ /* Copy stuff to the energy bin for easy printing etc. */
+ upd_mdebin(mdebin, FALSE, FALSE, (double)step,
+ mdatoms->tmass, enerd, state_global, inputrec->fepvals, inputrec->expandedvals, state_global->box,
+ NULL, NULL, vir, pres, NULL, mu_tot, constr);
+
+ print_ebin_header(fplog, step, step);
+ print_ebin(mdoutf_get_fp_ene(outf), TRUE, FALSE, FALSE, fplog, step, step, eprNORMAL,
+ mdebin, fcd, &(top_global->groups), &(inputrec->opts));
+ }
+ where();
+
+ /* This is the starting energy */
+ Epot = enerd->term[F_EPOT];
+
+ fnorm = ems.fnorm;
+ fmax = ems.fmax;
+ nfmax = ems.a_fmax;
+
+ /* Set the initial step.
+ * since it will be multiplied by the non-normalized search direction
+ * vector (force vector the first time), we scale it by the
+ * norm of the force.
+ */
+
+ if (MASTER(cr))
+ {
+ double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
+ fprintf(stderr, "Using %d BFGS correction steps.\n\n", nmaxcorr);
+ fprintf(stderr, " F-max = %12.5e on atom %d\n", fmax, nfmax+1);
+ fprintf(stderr, " F-Norm = %12.5e\n", fnorm/sqrtNumAtoms);
+ fprintf(stderr, "\n");
+ /* and copy to the log file too... */
+ fprintf(fplog, "Using %d BFGS correction steps.\n\n", nmaxcorr);
+ fprintf(fplog, " F-max = %12.5e on atom %d\n", fmax, nfmax+1);
+ fprintf(fplog, " F-Norm = %12.5e\n", fnorm/sqrtNumAtoms);
+ fprintf(fplog, "\n");
+ }
+
+ // Point is an index to the memory of search directions, where 0 is the first one.
+ point = 0;
+
+ // Set initial search direction to the force (-gradient), or 0 for frozen particles.
+ for (i = 0; i < n; i++)
+ {
+ if (!frozen[i])
+ {
+ dx[point][i] = ff[i]; /* Initial search direction */
+ }
+ else
+ {
+ dx[point][i] = 0;
+ }
+ }
+
+ // Stepsize will be modified during the search, and actually it is not critical
+ // (the main efficiency in the algorithm comes from changing directions), but
+ // we still need an initial value, so estimate it as the inverse of the norm
+ // so we take small steps where the potential fluctuates a lot.
+ stepsize = 1.0/fnorm;
+
+ /* Start the loop over BFGS steps.
+ * Each successful step is counted, and we continue until
+ * we either converge or reach the max number of steps.
+ */
+
+ ncorr = 0;
+
+ /* Set the gradient from the force */
+ converged = FALSE;
+ for (step = 0; (number_steps < 0 || step <= number_steps) && !converged; step++)
+ {
+
+ /* Write coordinates if necessary */
+ do_x = do_per_step(step, inputrec->nstxout);
+ do_f = do_per_step(step, inputrec->nstfout);
+
+ mdof_flags = 0;
+ if (do_x)
+ {
+ mdof_flags |= MDOF_X;
+ }
+
+ if (do_f)
+ {
+ mdof_flags |= MDOF_F;
+ }
+
+ if (inputrec->bIMD)
+ {
+ mdof_flags |= MDOF_IMD;
+ }
+
+ mdoutf_write_to_trajectory_files(fplog, cr, outf, mdof_flags,
+ top_global, step, (real)step, state_global, state_global, f);
+
+ /* Do the linesearching in the direction dx[point][0..(n-1)] */
+
+ /* make s a pointer to current search direction - point=0 first time we get here */
+ s = dx[point];
+
+ // calculate line gradient in position A
+ for (gpa = 0, i = 0; i < n; i++)
+ {
+ gpa -= s[i]*ff[i];
+ }
+
+ /* Calculate minimum allowed stepsize along the line, before the average (norm)
+ * relative change in coordinate is smaller than precision
+ */
+ for (minstep = 0, i = 0; i < n; i++)
+ {
+ tmp = fabs(xx[i]);
+ if (tmp < 1.0)
+ {
+ tmp = 1.0;
+ }
+ tmp = s[i]/tmp;
+ minstep += tmp*tmp;
+ }
+ minstep = GMX_REAL_EPS/sqrt(minstep/n);
+
+ if (stepsize < minstep)
+ {
+ converged = TRUE;
+ break;
+ }
+
+ // Before taking any steps along the line, store the old position
+ for (i = 0; i < n; i++)
+ {
+ lastx[i] = xx[i];
+ lastf[i] = ff[i];
+ }
+ Epot0 = Epot;
+
+ for (i = 0; i < n; i++)
+ {
+ xa[i] = xx[i];
+ }
+
+ /* Take a step downhill.
+ * In theory, we should find the actual minimum of the function in this
+ * direction, somewhere along the line.
+ * That is quite possible, but it turns out to take 5-10 function evaluations
+ * for each line. However, we dont really need to find the exact minimum -
+ * it is much better to start a new BFGS step in a modified direction as soon
+ * as we are close to it. This will save a lot of energy evaluations.
+ *
+ * In practice, we just try to take a single step.
+ * If it worked (i.e. lowered the energy), we increase the stepsize but
+ * continue straight to the next BFGS step without trying to find any minimum,
+ * i.e. we change the search direction too. If the line was smooth, it is
+ * likely we are in a smooth region, and then it makes sense to take longer
+ * steps in the modified search direction too.
+ *
+ * If it didn't work (higher energy), there must be a minimum somewhere between
+ * the old position and the new one. Then we need to start by finding a lower
+ * value before we change search direction. Since the energy was apparently
+ * quite rough, we need to decrease the step size.
+ *
+ * Due to the finite numerical accuracy, it turns out that it is a good idea
+ * to accept a SMALL increase in energy, if the derivative is still downhill.
+ * This leads to lower final energies in the tests I've done. / Erik
+ */
+
+ // State "A" is the first position along the line.
+ // reference position along line is initially zero
+ EpotA = Epot0;
+ a = 0.0;
+
+ // Check stepsize first. We do not allow displacements
+ // larger than emstep.
+ //
+ do
+ {
+ // Pick a new position C by adding stepsize to A.
+ c = a + stepsize;
+
+ // Calculate what the largest change in any individual coordinate
+ // would be (translation along line * gradient along line)
+ maxdelta = 0;
+ for (i = 0; i < n; i++)
+ {
+ delta = c*s[i];
+ if (delta > maxdelta)
+ {
+ maxdelta = delta;
+ }
+ }
+ // If any displacement is larger than the stepsize limit, reduce the step
+ if (maxdelta > inputrec->em_stepsize)
+ {
+ stepsize *= 0.1;
+ }
+ }
+ while (maxdelta > inputrec->em_stepsize);
+
+ // Take a trial step and move the coordinate array xc[] to position C
+ for (i = 0; i < n; i++)
+ {
+ xc[i] = lastx[i] + c*s[i];
+ }
+
+ neval++;
+ // Calculate energy for the trial step in position C
+ ems.s.x = (rvec *)xc;
+ ems.f = (rvec *)fc;
+ evaluate_energy(fplog, cr,
+ top_global, &ems, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph, mdatoms, fr,
+ mu_tot, enerd, vir, pres, step, FALSE);
+ EpotC = ems.epot;
+
+ // Calc line gradient in position C
+ for (gpc = 0, i = 0; i < n; i++)
+ {
+ gpc -= s[i]*fc[i]; /* f is negative gradient, thus the sign */
+ }
+ /* Sum the gradient along the line across CPUs */
+ if (PAR(cr))
+ {
+ gmx_sumd(1, &gpc, cr);
+ }
+
+ // This is the max amount of increase in energy we tolerate.
+ // By allowing VERY small changes (close to numerical precision) we
+ // frequently find even better (lower) final energies.
+ tmp = sqrt(GMX_REAL_EPS)*fabs(EpotA);
+
+ // Accept the step if the energy is lower in the new position C (compared to A),
+ // or if it is not significantly higher and the line derivative is still negative.
+ if (EpotC < EpotA || (gpc < 0 && EpotC < (EpotA+tmp)))
+ {
+ // Great, we found a better energy. We no longer try to alter the
+ // stepsize, but simply accept this new better position. The we select a new
+ // search direction instead, which will be much more efficient than continuing
+ // to take smaller steps along a line. Set fnorm based on the new C position,
+ // which will be used to update the stepsize to 1/fnorm further down.
+ foundlower = TRUE;
+ fnorm = ems.fnorm;
+ }
+ else
+ {
+ // If we got here, the energy is NOT lower in point C, i.e. it will be the same
+ // or higher than in point A. In this case it is pointless to move to point C,
+ // so we will have to do more iterations along the same line to find a smaller
+ // value in the interval [A=0.0,C].
+ // Here, A is still 0.0, but that will change when we do a search in the interval
+ // [0.0,C] below. That search we will do by interpolation or bisection rather
+ // than with the stepsize, so no need to modify it. For the next search direction
+ // it will be reset to 1/fnorm anyway.
+ foundlower = FALSE;
+ }
+
+ if (!foundlower)
+ {
+ // OK, if we didn't find a lower value we will have to locate one now - there must
+ // be one in the interval [a,c].
+ // The same thing is valid here, though: Don't spend dozens of iterations to find
+ // the line minimum. We try to interpolate based on the derivative at the endpoints,
+ // and only continue until we find a lower value. In most cases this means 1-2 iterations.
+ // I also have a safeguard for potentially really pathological functions so we never
+ // take more than 20 steps before we give up.
+ // If we already found a lower value we just skip this step and continue to the update.
+ nminstep = 0;
+ do
+ {
+ // Select a new trial point B in the interval [A,C].
+ // If the derivatives at points a & c have different sign we interpolate to zero,
+ // otherwise just do a bisection since there might be multiple minima/maxima
+ // inside the interval.
+ if (gpa < 0 && gpc > 0)
+ {
+ b = a + gpa*(a-c)/(gpc-gpa);
+ }
+ else
+ {
+ b = 0.5*(a+c);
+ }
+
+ /* safeguard if interpolation close to machine accuracy causes errors:
+ * never go outside the interval
+ */
+ if (b <= a || b >= c)
+ {
+ b = 0.5*(a+c);
+ }
+
+ // Take a trial step to point B
+ for (i = 0; i < n; i++)
+ {
+ xb[i] = lastx[i] + b*s[i];
+ }
+
+ neval++;
+ // Calculate energy for the trial step in point B
+ ems.s.x = (rvec *)xb;
+ ems.f = (rvec *)fb;
+ evaluate_energy(fplog, cr,
+ top_global, &ems, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph, mdatoms, fr,
+ mu_tot, enerd, vir, pres, step, FALSE);
+ EpotB = ems.epot;
+ fnorm = ems.fnorm;
+
+ // Calculate gradient in point B
+ for (gpb = 0, i = 0; i < n; i++)
+ {
+ gpb -= s[i]*fb[i]; /* f is negative gradient, thus the sign */
+
+ }
+ /* Sum the gradient along the line across CPUs */
+ if (PAR(cr))
+ {
+ gmx_sumd(1, &gpb, cr);
+ }
+
+ // Keep one of the intervals [A,B] or [B,C] based on the value of the derivative
+ // at the new point B, and rename the endpoints of this new interval A and C.
+ if (gpb > 0)
+ {
+ /* Replace c endpoint with b */
+ EpotC = EpotB;
+ c = b;
+ gpc = gpb;
+ /* swap coord pointers b/c */
+ xtmp = xb;
+ ftmp = fb;
+ xb = xc;
+ fb = fc;
+ xc = xtmp;
+ fc = ftmp;
+ }
+ else
+ {
+ /* Replace a endpoint with b */
+ EpotA = EpotB;
+ a = b;
+ gpa = gpb;
+ /* swap coord pointers a/b */
+ xtmp = xb;
+ ftmp = fb;
+ xb = xa;
+ fb = fa;
+ xa = xtmp;
+ fa = ftmp;
+ }
+
+ /*
+ * Stop search as soon as we find a value smaller than the endpoints,
+ * or if the tolerance is below machine precision.
+ * Never run more than 20 steps, no matter what.
+ */
+ nminstep++;
+ }
+ while ((EpotB > EpotA || EpotB > EpotC) && (nminstep < 20));
+
+ if (fabs(EpotB-Epot0) < GMX_REAL_EPS || nminstep >= 20)
+ {
+ /* OK. We couldn't find a significantly lower energy.
+ * If ncorr==0 this was steepest descent, and then we give up.
+ * If not, reset memory to restart as steepest descent before quitting.
+ */
+ if (ncorr == 0)
+ {
+ /* Converged */
+ converged = TRUE;
+ break;
+ }
+ else
+ {
+ /* Reset memory */
+ ncorr = 0;
+ /* Search in gradient direction */
+ for (i = 0; i < n; i++)
+ {
+ dx[point][i] = ff[i];
+ }
+ /* Reset stepsize */
+ stepsize = 1.0/fnorm;
+ continue;
+ }
+ }
+
+ /* Select min energy state of A & C, put the best in xx/ff/Epot
+ */
+ if (EpotC < EpotA)
+ {
+ Epot = EpotC;
+ /* Use state C */
+ for (i = 0; i < n; i++)
+ {
+ xx[i] = xc[i];
+ ff[i] = fc[i];
+ }
+ step_taken = c;
+ }
+ else
+ {
+ Epot = EpotA;
+ /* Use state A */
+ for (i = 0; i < n; i++)
+ {
+ xx[i] = xa[i];
+ ff[i] = fa[i];
+ }
+ step_taken = a;
+ }
+
+ }
+ else
+ {
+ /* found lower */
+ Epot = EpotC;
+ /* Use state C */
+ for (i = 0; i < n; i++)
+ {
+ xx[i] = xc[i];
+ ff[i] = fc[i];
+ }
+ step_taken = c;
+ }
+
+ /* Update the memory information, and calculate a new
+ * approximation of the inverse hessian
+ */
+
+ /* Have new data in Epot, xx, ff */
+ if (ncorr < nmaxcorr)
+ {
+ ncorr++;
+ }
+
+ for (i = 0; i < n; i++)
+ {
+ dg[point][i] = lastf[i]-ff[i];
+ dx[point][i] *= step_taken;
+ }
+
+ dgdg = 0;
+ dgdx = 0;
+ for (i = 0; i < n; i++)
+ {
+ dgdg += dg[point][i]*dg[point][i];
+ dgdx += dg[point][i]*dx[point][i];
+ }
+
+ diag = dgdx/dgdg;
+
+ rho[point] = 1.0/dgdx;
+ point++;
+
+ if (point >= nmaxcorr)
+ {
+ point = 0;
+ }
+
+ /* Update */
+ for (i = 0; i < n; i++)
+ {
+ p[i] = ff[i];
+ }
+
+ cp = point;
+
+ /* Recursive update. First go back over the memory points */
+ for (k = 0; k < ncorr; k++)
+ {
+ cp--;
+ if (cp < 0)
+ {
+ cp = ncorr-1;
+ }
+
+ sq = 0;
+ for (i = 0; i < n; i++)
+ {
+ sq += dx[cp][i]*p[i];
+ }
+
+ alpha[cp] = rho[cp]*sq;
+
+ for (i = 0; i < n; i++)
+ {
+ p[i] -= alpha[cp]*dg[cp][i];
+ }
+ }
+
+ for (i = 0; i < n; i++)
+ {
+ p[i] *= diag;
+ }
+
+ /* And then go forward again */
+ for (k = 0; k < ncorr; k++)
+ {
+ yr = 0;
+ for (i = 0; i < n; i++)
+ {
+ yr += p[i]*dg[cp][i];
+ }
+
+ beta = rho[cp]*yr;
+ beta = alpha[cp]-beta;
+
+ for (i = 0; i < n; i++)
+ {
+ p[i] += beta*dx[cp][i];
+ }
+
+ cp++;
+ if (cp >= ncorr)
+ {
+ cp = 0;
+ }
+ }
+
+ for (i = 0; i < n; i++)
+ {
+ if (!frozen[i])
+ {
+ dx[point][i] = p[i];
+ }
+ else
+ {
+ dx[point][i] = 0;
+ }
+ }
+
+ /* Test whether the convergence criterion is met */
+ get_f_norm_max(cr, &(inputrec->opts), mdatoms, f, &fnorm, &fmax, &nfmax);
+
+ /* Print it if necessary */
+ if (MASTER(cr))
+ {
+ if (bVerbose)
+ {
+ double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
+ fprintf(stderr, "\rStep %d, Epot=%12.6e, Fnorm=%9.3e, Fmax=%9.3e (atom %d)\n",
+ step, Epot, fnorm/sqrtNumAtoms, fmax, nfmax+1);
+ fflush(stderr);
+ }
+ /* Store the new (lower) energies */
+ upd_mdebin(mdebin, FALSE, FALSE, (double)step,
+ mdatoms->tmass, enerd, state_global, inputrec->fepvals, inputrec->expandedvals, state_global->box,
+ NULL, NULL, vir, pres, NULL, mu_tot, constr);
+ do_log = do_per_step(step, inputrec->nstlog);
+ do_ene = do_per_step(step, inputrec->nstenergy);
+ if (do_log)
+ {
+ print_ebin_header(fplog, step, step);
+ }
+ print_ebin(mdoutf_get_fp_ene(outf), do_ene, FALSE, FALSE,
+ do_log ? fplog : NULL, step, step, eprNORMAL,
+ mdebin, fcd, &(top_global->groups), &(inputrec->opts));
+ }
+
+ /* Send x and E to IMD client, if bIMD is TRUE. */
+ if (do_IMD(inputrec->bIMD, step, cr, TRUE, state_global->box, state_global->x, inputrec, 0, wcycle) && MASTER(cr))
+ {
+ IMD_send_positions(inputrec->imd);
+ }
+
+ // Reset stepsize in we are doing more iterations
+ stepsize = 1.0/fnorm;
+
+ /* Stop when the maximum force lies below tolerance.
+ * If we have reached machine precision, converged is already set to true.
+ */
+ converged = converged || (fmax < inputrec->em_tol);
+
+ } /* End of the loop */
+
+ /* IMD cleanup, if bIMD is TRUE. */
+ IMD_finalize(inputrec->bIMD, inputrec->imd);
+
+ if (converged)
+ {
+ step--; /* we never took that last step in this case */
+
+ }
+ if (fmax > inputrec->em_tol)
+ {
+ if (MASTER(cr))
+ {
+ warn_step(stderr, inputrec->em_tol, step-1 == number_steps, FALSE);
+ warn_step(fplog, inputrec->em_tol, step-1 == number_steps, FALSE);
+ }
+ converged = FALSE;
+ }
+
+ /* If we printed energy and/or logfile last step (which was the last step)
+ * we don't have to do it again, but otherwise print the final values.
+ */
+ if (!do_log) /* Write final value to log since we didn't do anythin last step */
+ {
+ print_ebin_header(fplog, step, step);
+ }
+ if (!do_ene || !do_log) /* Write final energy file entries */
+ {
+ print_ebin(mdoutf_get_fp_ene(outf), !do_ene, FALSE, FALSE,
+ !do_log ? fplog : NULL, step, step, eprNORMAL,
+ mdebin, fcd, &(top_global->groups), &(inputrec->opts));
+ }
+
+ /* Print some stuff... */
+ if (MASTER(cr))
+ {
+ fprintf(stderr, "\nwriting lowest energy coordinates.\n");
+ }
+
+ /* IMPORTANT!
+ * For accurate normal mode calculation it is imperative that we
+ * store the last conformation into the full precision binary trajectory.
+ *
+ * However, we should only do it if we did NOT already write this step
+ * above (which we did if do_x or do_f was true).
+ */
+ do_x = !do_per_step(step, inputrec->nstxout);
+ do_f = !do_per_step(step, inputrec->nstfout);
+ write_em_traj(fplog, cr, outf, do_x, do_f, ftp2fn(efSTO, nfile, fnm),
+ top_global, inputrec, step,
+ &ems, state_global);
+
+ if (MASTER(cr))
+ {
+ double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
+ print_converged(stderr, LBFGS, inputrec->em_tol, step, converged,
+ number_steps, Epot, fmax, nfmax, fnorm/sqrtNumAtoms);
+ print_converged(fplog, LBFGS, inputrec->em_tol, step, converged,
+ number_steps, Epot, fmax, nfmax, fnorm/sqrtNumAtoms);
+
+ fprintf(fplog, "\nPerformed %d energy evaluations in total.\n", neval);
+ }
+
+ finish_em(cr, outf, walltime_accounting, wcycle);
+
+ /* To print the actual number of steps we needed somewhere */
+ walltime_accounting_set_nsteps_done(walltime_accounting, step);
+
+ return 0;
+} /* That's all folks */
+
+/*! \brief Do steepest descents minimization
+ \copydoc integrator_t (FILE *fplog, t_commrec *cr,
+ int nfile, const t_filenm fnm[],
+ const gmx_output_env_t *oenv, gmx_bool bVerbose,
+ int nstglobalcomm,
+ gmx_vsite_t *vsite, gmx_constr_t constr,
+ int stepout,
+ t_inputrec *inputrec,
+ 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_nex, int repl_ex_seed,
+ real cpt_period, real max_hours,
+ int imdport,
+ unsigned long Flags,
+ gmx_walltime_accounting_t walltime_accounting)
+ */
+double do_steep(FILE *fplog, t_commrec *cr,
+ int nfile, const t_filenm fnm[],
+ const gmx_output_env_t gmx_unused *oenv, gmx_bool bVerbose,
+ int gmx_unused nstglobalcomm,
+ gmx_vsite_t *vsite, gmx_constr_t constr,
+ int gmx_unused stepout,
+ t_inputrec *inputrec,
+ 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 gmx_unused ed,
+ t_forcerec *fr,
+ int gmx_unused repl_ex_nst, int gmx_unused repl_ex_nex, int gmx_unused repl_ex_seed,
+ real gmx_unused cpt_period, real gmx_unused max_hours,
+ int imdport,
+ unsigned long gmx_unused Flags,
+ gmx_walltime_accounting_t walltime_accounting)
+{
+ const char *SD = "Steepest Descents";
+ em_state_t *s_min, *s_try;
+ gmx_localtop_t *top;
+ gmx_enerdata_t *enerd;
+ rvec *f;
+ gmx_global_stat_t gstat;
+ t_graph *graph;
+ real stepsize;
+ real ustep, fnormn;
+ gmx_mdoutf_t outf;
+ t_mdebin *mdebin;
+ gmx_bool bDone, bAbort, do_x, do_f;
+ tensor vir, pres;
+ rvec mu_tot;
+ int nsteps;
+ int count = 0;
+ int steps_accepted = 0;
+
+ s_min = init_em_state();
+ s_try = init_em_state();
+
+ /* Init em and store the local state in s_try */
+ init_em(fplog, SD, cr, inputrec,
+ state_global, top_global, s_try, &top, &f,
+ nrnb, mu_tot, fr, &enerd, &graph, mdatoms, &gstat, vsite, constr,
+ nfile, fnm, &outf, &mdebin, imdport, Flags, wcycle);
+
+ /* Print to log file */
+ print_em_start(fplog, cr, walltime_accounting, wcycle, SD);
+
+ /* Set variables for stepsize (in nm). This is the largest
+ * step that we are going to make in any direction.
+ */
+ ustep = inputrec->em_stepsize;
+ stepsize = 0;
+
+ /* Max number of steps */
+ nsteps = inputrec->nsteps;
+
+ if (MASTER(cr))
+ {
+ /* Print to the screen */
+ sp_header(stderr, SD, inputrec->em_tol, nsteps);
+ }
+ if (fplog)
+ {
+ sp_header(fplog, SD, inputrec->em_tol, nsteps);
+ }
+
+ /**** HERE STARTS THE LOOP ****
+ * count is the counter for the number of steps
+ * bDone will be TRUE when the minimization has converged
+ * bAbort will be TRUE when nsteps steps have been performed or when
+ * the stepsize becomes smaller than is reasonable for machine precision
+ */
+ count = 0;
+ bDone = FALSE;
+ bAbort = FALSE;
+ while (!bDone && !bAbort)
+ {
+ bAbort = (nsteps >= 0) && (count == nsteps);
+
+ /* set new coordinates, except for first step */
+ if (count > 0)
+ {
+ do_em_step(cr, inputrec, mdatoms, fr->bMolPBC,
+ s_min, stepsize, s_min->f, s_try,
+ constr, top, nrnb, wcycle, count);
+ }
+
+ evaluate_energy(fplog, cr,
+ top_global, s_try, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph, mdatoms, fr,
+ mu_tot, enerd, vir, pres, count, count == 0);
+
+ if (MASTER(cr))
+ {
+ print_ebin_header(fplog, count, count);
+ }
+
+ if (count == 0)
+ {
+ s_min->epot = s_try->epot;
+ }
+
+ /* Print it if necessary */
+ if (MASTER(cr))
+ {
+ if (bVerbose)
+ {
+ fprintf(stderr, "Step=%5d, Dmax= %6.1e nm, Epot= %12.5e Fmax= %11.5e, atom= %d%c",
+ count, ustep, s_try->epot, s_try->fmax, s_try->a_fmax+1,
+ ( (count == 0) || (s_try->epot < s_min->epot) ) ? '\n' : '\r');
+ fflush(stderr);
+ }
+
+ if ( (count == 0) || (s_try->epot < s_min->epot) )
+ {
+ /* Store the new (lower) energies */
+ upd_mdebin(mdebin, FALSE, FALSE, (double)count,
+ mdatoms->tmass, enerd, &s_try->s, inputrec->fepvals, inputrec->expandedvals,
+ s_try->s.box, NULL, NULL, vir, pres, NULL, mu_tot, constr);
+
+ /* Prepare IMD energy record, if bIMD is TRUE. */
+ IMD_fill_energy_record(inputrec->bIMD, inputrec->imd, enerd, count, TRUE);
+
+ print_ebin(mdoutf_get_fp_ene(outf), TRUE,
+ do_per_step(steps_accepted, inputrec->nstdisreout),
+ do_per_step(steps_accepted, inputrec->nstorireout),
+ fplog, count, count, eprNORMAL,
+ mdebin, fcd, &(top_global->groups), &(inputrec->opts));
+ fflush(fplog);
+ }
+ }
+
+ /* Now if the new energy is smaller than the previous...
+ * or if this is the first step!
+ * or if we did random steps!
+ */
+
+ if ( (count == 0) || (s_try->epot < s_min->epot) )
+ {
+ steps_accepted++;
+
+ /* Test whether the convergence criterion is met... */
+ bDone = (s_try->fmax < inputrec->em_tol);
+
+ /* Copy the arrays for force, positions and energy */
+ /* The 'Min' array always holds the coords and forces of the minimal
+ sampled energy */
+ swap_em_state(s_min, s_try);
+ if (count > 0)
+ {
+ ustep *= 1.2;
+ }
+
+ /* Write to trn, if necessary */
+ do_x = do_per_step(steps_accepted, inputrec->nstxout);
+ do_f = do_per_step(steps_accepted, inputrec->nstfout);
+ write_em_traj(fplog, cr, outf, do_x, do_f, NULL,
+ top_global, inputrec, count,
+ s_min, state_global);
+ }
+ else
+ {
+ /* If energy is not smaller make the step smaller... */
+ ustep *= 0.5;
+
+ if (DOMAINDECOMP(cr) && s_min->s.ddp_count != cr->dd->ddp_count)
+ {
+ /* Reload the old state */
+ em_dd_partition_system(fplog, count, cr, top_global, inputrec,
+ s_min, top, mdatoms, fr, vsite, constr,
+ nrnb, wcycle);
+ }
+ }
+
+ /* Determine new step */
+ stepsize = ustep/s_min->fmax;
+
+ /* Check if stepsize is too small, with 1 nm as a characteristic length */
+#if GMX_DOUBLE
+ if (count == nsteps || ustep < 1e-12)
+#else
+ if (count == nsteps || ustep < 1e-6)
+#endif
+ {
+ if (MASTER(cr))
+ {
+ warn_step(stderr, inputrec->em_tol, count == nsteps, constr != NULL);
+ warn_step(fplog, inputrec->em_tol, count == nsteps, constr != NULL);
+ }
+ bAbort = TRUE;
+ }
+
+ /* Send IMD energies and positions, if bIMD is TRUE. */
+ if (do_IMD(inputrec->bIMD, count, cr, TRUE, state_global->box, state_global->x, inputrec, 0, wcycle) && MASTER(cr))
+ {
+ IMD_send_positions(inputrec->imd);
+ }
+
+ count++;
+ } /* End of the loop */
+
+ /* IMD cleanup, if bIMD is TRUE. */
+ IMD_finalize(inputrec->bIMD, inputrec->imd);
+
+ /* Print some data... */
+ if (MASTER(cr))
+ {
+ fprintf(stderr, "\nwriting lowest energy coordinates.\n");
+ }
+ write_em_traj(fplog, cr, outf, TRUE, inputrec->nstfout, ftp2fn(efSTO, nfile, fnm),
+ top_global, inputrec, count,
+ s_min, state_global);
+
+ if (MASTER(cr))
+ {
+ double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
+ fnormn = s_min->fnorm/sqrtNumAtoms;
+
+ print_converged(stderr, SD, inputrec->em_tol, count, bDone, nsteps,
+ s_min->epot, s_min->fmax, s_min->a_fmax, fnormn);
+ print_converged(fplog, SD, inputrec->em_tol, count, bDone, nsteps,
+ s_min->epot, s_min->fmax, s_min->a_fmax, fnormn);
+ }
+
+ finish_em(cr, outf, walltime_accounting, wcycle);
+
+ /* To print the actual number of steps we needed somewhere */
+ inputrec->nsteps = count;
+
+ walltime_accounting_set_nsteps_done(walltime_accounting, count);
+
+ return 0;
+} /* That's all folks */
+
+/*! \brief Do normal modes analysis
+ \copydoc integrator_t (FILE *fplog, t_commrec *cr,
+ int nfile, const t_filenm fnm[],
+ const gmx_output_env_t *oenv, gmx_bool bVerbose,
+ int nstglobalcomm,
+ gmx_vsite_t *vsite, gmx_constr_t constr,
+ int stepout,
+ t_inputrec *inputrec,
+ 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_nex, int repl_ex_seed,
+ real cpt_period, real max_hours,
+ int imdport,
+ unsigned long Flags,
+ gmx_walltime_accounting_t walltime_accounting)
+ */
+double do_nm(FILE *fplog, t_commrec *cr,
+ int nfile, const t_filenm fnm[],
+ const gmx_output_env_t gmx_unused *oenv, gmx_bool bVerbose,
+ int gmx_unused nstglobalcomm,
+ gmx_vsite_t *vsite, gmx_constr_t constr,
+ int gmx_unused stepout,
+ t_inputrec *inputrec,
+ 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 gmx_unused ed,
+ t_forcerec *fr,
+ int gmx_unused repl_ex_nst, int gmx_unused repl_ex_nex, int gmx_unused repl_ex_seed,
+ real gmx_unused cpt_period, real gmx_unused max_hours,
+ int imdport,
+ unsigned long gmx_unused Flags,
+ gmx_walltime_accounting_t walltime_accounting)
+{
+ const char *NM = "Normal Mode Analysis";
+ gmx_mdoutf_t outf;
+ int nnodes, node;
+ gmx_localtop_t *top;
+ gmx_enerdata_t *enerd;
+ rvec *f;
+ gmx_global_stat_t gstat;
+ t_graph *graph;
+ tensor vir, pres;
+ rvec mu_tot;
+ rvec *fneg, *dfdx;
+ gmx_bool bSparse; /* use sparse matrix storage format */
+ size_t sz;
+ gmx_sparsematrix_t * sparse_matrix = NULL;
+ real * full_matrix = NULL;
+ em_state_t * state_work;
+
+ /* added with respect to mdrun */
+ int row, col;
+ real der_range = 10.0*sqrt(GMX_REAL_EPS);
+ real x_min;
+ bool bIsMaster = MASTER(cr);
+
+ if (constr != NULL)
+ {
+ gmx_fatal(FARGS, "Constraints present with Normal Mode Analysis, this combination is not supported");
+ }
+
+ state_work = init_em_state();
+
+ /* Init em and store the local state in state_minimum */
+ init_em(fplog, NM, cr, inputrec,
+ state_global, top_global, state_work, &top,
+ &f,
+ nrnb, mu_tot, fr, &enerd, &graph, mdatoms, &gstat, vsite, constr,
+ nfile, fnm, &outf, NULL, imdport, Flags, wcycle);
+
+ gmx_shellfc_t *shellfc = init_shell_flexcon(stdout,
+ top_global,
+ n_flexible_constraints(constr),
+ inputrec->nstcalcenergy,
+ DOMAINDECOMP(cr));
+
+ if (shellfc)
+ {
+ make_local_shells(cr, mdatoms, shellfc);
+ }
+ std::vector<size_t> atom_index = get_atom_index(top_global);
+ snew(fneg, atom_index.size());
+ snew(dfdx, atom_index.size());
+
+#if !GMX_DOUBLE
+ if (bIsMaster)
+ {
+ fprintf(stderr,
+ "NOTE: This version of GROMACS has been compiled in single precision,\n"
+ " which MIGHT not be accurate enough for normal mode analysis.\n"
+ " GROMACS now uses sparse matrix storage, so the memory requirements\n"
+ " are fairly modest even if you recompile in double precision.\n\n");
+ }
+#endif
+
+ /* Check if we can/should use sparse storage format.
+ *
+ * Sparse format is only useful when the Hessian itself is sparse, which it
+ * will be when we use a cutoff.
+ * For small systems (n<1000) it is easier to always use full matrix format, though.
+ */
+ if (EEL_FULL(fr->eeltype) || fr->rlist == 0.0)
+ {
+ md_print_info(cr, fplog, "Non-cutoff electrostatics used, forcing full Hessian format.\n");
+ bSparse = FALSE;
+ }
+ else if (atom_index.size() < 1000)
+ {
+ md_print_info(cr, fplog, "Small system size (N=%d), using full Hessian format.\n", atom_index.size());
+ bSparse = FALSE;
+ }
+ else
+ {
+ md_print_info(cr, fplog, "Using compressed symmetric sparse Hessian format.\n");
+ bSparse = TRUE;
+ }
+
+ /* Number of dimensions, based on real atoms, that is not vsites or shell */
+ sz = DIM*atom_index.size();
+
+ fprintf(stderr, "Allocating Hessian memory...\n\n");
+
+ if (bSparse)
+ {
+ sparse_matrix = gmx_sparsematrix_init(sz);
+ sparse_matrix->compressed_symmetric = TRUE;
+ }
+ else
+ {
+ snew(full_matrix, sz*sz);
+ }
+
+ init_nrnb(nrnb);
+
+ where();
+
+ /* Write start time and temperature */
+ print_em_start(fplog, cr, walltime_accounting, wcycle, NM);
+
+ /* fudge nr of steps to nr of atoms */
+ inputrec->nsteps = atom_index.size()*2;
+
+ if (bIsMaster)
+ {
+ fprintf(stderr, "starting normal mode calculation '%s'\n%d steps.\n\n",
+ *(top_global->name), (int)inputrec->nsteps);
+ }
+
+ nnodes = cr->nnodes;
+
+ /* Make evaluate_energy do a single node force calculation */
+ cr->nnodes = 1;
+ evaluate_energy(fplog, cr,
+ top_global, state_work, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph, mdatoms, fr,
+ mu_tot, enerd, vir, pres, -1, TRUE);
+ cr->nnodes = nnodes;
+
+ /* if forces are not small, warn user */
+ get_state_f_norm_max(cr, &(inputrec->opts), mdatoms, state_work);
+
+ md_print_info(cr, fplog, "Maximum force:%12.5e\n", state_work->fmax);
+ if (state_work->fmax > 1.0e-3)
+ {
+ md_print_info(cr, fplog,
+ "The force is probably not small enough to "
+ "ensure that you are at a minimum.\n"
+ "Be aware that negative eigenvalues may occur\n"
+ "when the resulting matrix is diagonalized.\n\n");
+ }
+
+ /***********************************************************
+ *
+ * Loop over all pairs in matrix
+ *
+ * do_force called twice. Once with positive and
+ * once with negative displacement
+ *
+ ************************************************************/
+
+ /* Steps are divided one by one over the nodes */
+ bool bNS = true;
+ for (unsigned int aid = cr->nodeid; aid < atom_index.size(); aid += nnodes)
+ {
+ size_t atom = atom_index[aid];
+ for (size_t d = 0; d < DIM; d++)
+ {
+ gmx_bool bBornRadii = FALSE;
+ gmx_int64_t step = 0;
+ int force_flags = GMX_FORCE_STATECHANGED | GMX_FORCE_ALLFORCES;
+ double t = 0;
+
+ x_min = state_work->s.x[atom][d];
+
+ for (unsigned int dx = 0; (dx < 2); dx++)
+ {
+ if (dx == 0)
+ {
+ state_work->s.x[atom][d] = x_min - der_range;
+ }
+ else
+ {
+ state_work->s.x[atom][d] = x_min + der_range;
+ }
+
+ /* Make evaluate_energy do a single node force calculation */
+ cr->nnodes = 1;
+ if (shellfc)
+ {
+ /* Now is the time to relax the shells */
+ (void) relax_shell_flexcon(fplog, cr, bVerbose, step,
+ inputrec, bNS, force_flags,
+ top,
+ constr, enerd, fcd,
+ &state_work->s, state_work->f, vir, mdatoms,
+ nrnb, wcycle, graph, &top_global->groups,
+ shellfc, fr, bBornRadii, t, mu_tot,
+ vsite, NULL);
+ bNS = false;
+ step++;
+ }
+ else
+ {
+ evaluate_energy(fplog, cr,
+ top_global, state_work, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph, mdatoms, fr,
+ mu_tot, enerd, vir, pres, atom*2+dx, FALSE);
+ }
+
+ cr->nnodes = nnodes;
+
+ if (dx == 0)
+ {
+ for (size_t i = 0; i < atom_index.size(); i++)
+ {
+ copy_rvec(state_work->f[atom_index[i]], fneg[i]);
+ }
+ }
+ }
+
+ /* x is restored to original */
+ state_work->s.x[atom][d] = x_min;
+
+ for (size_t j = 0; j < atom_index.size(); j++)
+ {
+ for (size_t k = 0; (k < DIM); k++)
+ {
+ dfdx[j][k] =
+ -(state_work->f[atom_index[j]][k] - fneg[j][k])/(2*der_range);
+ }
+ }
+
+ if (!bIsMaster)
+ {
+#if GMX_MPI
+#define mpi_type GMX_MPI_REAL
+ MPI_Send(dfdx[0], atom_index.size()*DIM, mpi_type, MASTER(cr),
+ cr->nodeid, cr->mpi_comm_mygroup);
+#endif
+ }
+ else
+ {
+ for (node = 0; (node < nnodes && atom+node < atom_index.size()); node++)
+ {
+ if (node > 0)
+ {
+#if GMX_MPI
+ MPI_Status stat;
+ MPI_Recv(dfdx[0], atom_index.size()*DIM, mpi_type, node, node,
+ cr->mpi_comm_mygroup, &stat);
+#undef mpi_type
+#endif
+ }
+
+ row = (atom + node)*DIM + d;
+
+ for (size_t j = 0; j < atom_index.size(); j++)
+ {
+ for (size_t k = 0; k < DIM; k++)
+ {
+ col = j*DIM + k;
+
+ if (bSparse)
+ {
+ if (col >= row && dfdx[j][k] != 0.0)
+ {
+ gmx_sparsematrix_increment_value(sparse_matrix,
+ row, col, dfdx[j][k]);
+ }
+ }
+ else
+ {
+ full_matrix[row*sz+col] = dfdx[j][k];
+ }
+ }
+ }
+ }
+ }
+
+ if (bVerbose && fplog)
+ {
+ fflush(fplog);
+ }
+ }
+ /* write progress */
+ if (bIsMaster && bVerbose)
+ {
+ fprintf(stderr, "\rFinished step %d out of %d",
+ static_cast<int>(std::min(atom+nnodes, atom_index.size())),
+ static_cast<int>(atom_index.size()));
+ fflush(stderr);
+ }
+ }
+
+ if (bIsMaster)
+ {
+ fprintf(stderr, "\n\nWriting Hessian...\n");
+ gmx_mtxio_write(ftp2fn(efMTX, nfile, fnm), sz, sz, full_matrix, sparse_matrix);
+ }
+
+ finish_em(cr, outf, walltime_accounting, wcycle);
+
+ walltime_accounting_set_nsteps_done(walltime_accounting, atom_index.size()*2);
+
+ return 0;
+}
+
+} // namespace gmx
diff --git a/patches/gromacs-2016-beta1.diff/src/programs/mdrun/md.cpp b/patches/gromacs-2016-beta1.diff/src/programs/mdrun/md.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..2bd872565f10b34eaa07f80101ae3691a30be7ca
--- /dev/null
+++ b/patches/gromacs-2016-beta1.diff/src/programs/mdrun/md.cpp
@@ -0,0 +1,1940 @@
+/*
+ * This file is part of the GROMACS molecular simulation package.
+ *
+ * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
+ * Copyright (c) 2001-2004, The GROMACS development team.
+ * Copyright (c) 2011,2012,2013,2014,2015,2016, by the GROMACS development team, led by
+ * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
+ * and including many others, as listed in the AUTHORS file in the
+ * top-level source directory and at http://www.gromacs.org.
+ *
+ * GROMACS is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public License
+ * as published by the Free Software Foundation; either version 2.1
+ * of the License, or (at your option) any later version.
+ *
+ * GROMACS is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with GROMACS; if not, see
+ * http://www.gnu.org/licenses, or write to the Free Software Foundation,
+ * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ * If you want to redistribute modifications to GROMACS, 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 http://www.gromacs.org.
+ *
+ * To help us fund GROMACS development, we humbly ask that you cite
+ * the research papers on the package. Check out http://www.gromacs.org.
+ */
+#include "gmxpre.h"
+
+#include "md.h"
+
+#include "config.h"
+
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#include <algorithm>
+
+#include "thread_mpi/threads.h"
+
+#include "gromacs/commandline/filenm.h"
+#include "gromacs/domdec/domdec.h"
+#include "gromacs/domdec/domdec_network.h"
+#include "gromacs/domdec/domdec_struct.h"
+#include "gromacs/ewald/pme.h"
+#include "gromacs/ewald/pme-load-balancing.h"
+#include "gromacs/fileio/trxio.h"
+#include "gromacs/gmxlib/md_logging.h"
+#include "gromacs/gmxlib/network.h"
+#include "gromacs/gmxlib/nrnb.h"
+#include "gromacs/gpu_utils/gpu_utils.h"
+#include "gromacs/imd/imd.h"
+#include "gromacs/listed-forces/manage-threading.h"
+#include "gromacs/math/functions.h"
+#include "gromacs/math/units.h"
+#include "gromacs/math/utilities.h"
+#include "gromacs/math/vec.h"
+#include "gromacs/math/vectypes.h"
+#include "gromacs/mdlib/compute_io.h"
+#include "gromacs/mdlib/constr.h"
+#include "gromacs/mdlib/ebin.h"
+#include "gromacs/mdlib/force.h"
+#include "gromacs/mdlib/force_flags.h"
+#include "gromacs/mdlib/forcerec.h"
+#include "gromacs/mdlib/md_support.h"
+#include "gromacs/mdlib/mdatoms.h"
+#include "gromacs/mdlib/mdebin.h"
+#include "gromacs/mdlib/mdoutf.h"
+#include "gromacs/mdlib/mdrun.h"
+#include "gromacs/mdlib/mdrun_signalling.h"
+#include "gromacs/mdlib/nb_verlet.h"
+#include "gromacs/mdlib/nbnxn_gpu_data_mgmt.h"
+#include "gromacs/mdlib/ns.h"
+#include "gromacs/mdlib/shellfc.h"
+#include "gromacs/mdlib/sighandler.h"
+#include "gromacs/mdlib/sim_util.h"
+#include "gromacs/mdlib/tgroup.h"
+#include "gromacs/mdlib/trajectory_writing.h"
+#include "gromacs/mdlib/update.h"
+#include "gromacs/mdlib/vcm.h"
+#include "gromacs/mdlib/vsite.h"
+#include "gromacs/mdtypes/commrec.h"
+#include "gromacs/mdtypes/df_history.h"
+#include "gromacs/mdtypes/energyhistory.h"
+#include "gromacs/mdtypes/fcdata.h"
+#include "gromacs/mdtypes/forcerec.h"
+#include "gromacs/mdtypes/group.h"
+#include "gromacs/mdtypes/inputrec.h"
+#include "gromacs/mdtypes/interaction_const.h"
+#include "gromacs/mdtypes/md_enums.h"
+#include "gromacs/mdtypes/mdatom.h"
+#include "gromacs/mdtypes/state.h"
+#include "gromacs/pbcutil/mshift.h"
+#include "gromacs/pbcutil/pbc.h"
+#include "gromacs/pulling/pull.h"
+#include "gromacs/swap/swapcoords.h"
+#include "gromacs/timing/wallcycle.h"
+#include "gromacs/timing/walltime_accounting.h"
+#include "gromacs/topology/atoms.h"
+#include "gromacs/topology/idef.h"
+#include "gromacs/topology/mtop_util.h"
+#include "gromacs/topology/topology.h"
+#include "gromacs/trajectory/trajectoryframe.h"
+#include "gromacs/utility/basedefinitions.h"
+#include "gromacs/utility/cstringutil.h"
+#include "gromacs/utility/fatalerror.h"
+#include "gromacs/utility/real.h"
+#include "gromacs/utility/smalloc.h"
+
+#include "deform.h"
+#include "membed.h"
+#include "repl_ex.h"
+
+/* PLUMED */
+#include "../../../Plumed.h"
+extern int plumedswitch;
+extern plumed plumedmain;
+/* END PLUMED */
+
+#ifdef GMX_FAHCORE
+#include "corewrap.h"
+#endif
+
+/*! \brief Check whether bonded interactions are missing, if appropriate
+ *
+ * \param[in] fplog Log file pointer
+ * \param[in] cr Communication object
+ * \param[in] totalNumberOfBondedInteractions Result of the global reduction over the number of bonds treated in each domain
+ * \param[in] top_global Global topology for the error message
+ * \param[in] top_local Local topology for the error message
+ * \param[in] state Global state for the error message
+ * \param[inout] shouldCheckNumberOfBondedInteractions Whether we should do the check.
+ *
+ * \return Nothing, except that shouldCheckNumberOfBondedInteractions
+ * is always set to false after exit.
+ */
+static void checkNumberOfBondedInteractions(FILE *fplog, t_commrec *cr, int totalNumberOfBondedInteractions,
+ gmx_mtop_t *top_global, gmx_localtop_t *top_local, t_state *state,
+ bool *shouldCheckNumberOfBondedInteractions)
+{
+ if (*shouldCheckNumberOfBondedInteractions)
+ {
+ if (totalNumberOfBondedInteractions != cr->dd->nbonded_global)
+ {
+ dd_print_missing_interactions(fplog, cr, totalNumberOfBondedInteractions, top_global, top_local, state); // Does not return
+ }
+ *shouldCheckNumberOfBondedInteractions = false;
+ }
+}
+
+static void reset_all_counters(FILE *fplog, t_commrec *cr,
+ gmx_int64_t step,
+ gmx_int64_t *step_rel, t_inputrec *ir,
+ gmx_wallcycle_t wcycle, t_nrnb *nrnb,
+ gmx_walltime_accounting_t walltime_accounting,
+ struct nonbonded_verlet_t *nbv)
+{
+ char sbuf[STEPSTRSIZE];
+
+ /* Reset all the counters related to performance over the run */
+ md_print_warn(cr, fplog, "step %s: resetting all time and cycle counters\n",
+ gmx_step_str(step, sbuf));
+
+ if (use_GPU(nbv))
+ {
+ nbnxn_gpu_reset_timings(nbv);
+ resetGpuProfiler();
+ }
+
+ wallcycle_stop(wcycle, ewcRUN);
+ wallcycle_reset_all(wcycle);
+ if (DOMAINDECOMP(cr))
+ {
+ reset_dd_statistics_counters(cr->dd);
+ }
+ init_nrnb(nrnb);
+ ir->init_step += *step_rel;
+ ir->nsteps -= *step_rel;
+ *step_rel = 0;
+ wallcycle_start(wcycle, ewcRUN);
+ walltime_accounting_start(walltime_accounting);
+ print_date_and_time(fplog, cr->nodeid, "Restarted time", gmx_gettime());
+}
+
+/*! \libinternal
+ \copydoc integrator_t (FILE *fplog, t_commrec *cr,
+ int nfile, const t_filenm fnm[],
+ const gmx_output_env_t *oenv, gmx_bool bVerbose,
+ int nstglobalcomm,
+ gmx_vsite_t *vsite, gmx_constr_t constr,
+ int stepout,
+ t_inputrec *inputrec,
+ 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_nex, int repl_ex_seed,
+ real cpt_period, real max_hours,
+ int imdport,
+ unsigned long Flags,
+ gmx_walltime_accounting_t walltime_accounting)
+ */
+double gmx::do_md(FILE *fplog, t_commrec *cr, int nfile, const t_filenm fnm[],
+ const gmx_output_env_t *oenv, gmx_bool bVerbose,
+ 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_nex, int repl_ex_seed,
+ real cpt_period, real max_hours,
+ int imdport,
+ unsigned long Flags,
+ gmx_walltime_accounting_t walltime_accounting)
+{
+ gmx_mdoutf_t outf = NULL;
+ gmx_int64_t step, step_rel;
+ double elapsed_time;
+ double t, t0, lam0[efptNR];
+ gmx_bool bGStatEveryStep, bGStat, bCalcVir, bCalcEnerStep, bCalcEner;
+ gmx_bool bNS, bNStList, bSimAnn, bStopCM, bRerunMD,
+ bFirstStep, startingFromCheckpoint, bInitStep, bLastStep = FALSE,
+ bBornRadii;
+ gmx_bool bDoDHDL = FALSE, bDoFEP = FALSE, bDoExpanded = FALSE;
+ gmx_bool do_ene, do_log, do_verbose, bRerunWarnNoV = TRUE,
+ bForceUpdate = FALSE, bCPT;
+ 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;
+ matrix pcoupl_mu, M;
+ 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;
+ gmx_enerdata_t *enerd;
+ rvec *f = NULL;
+ gmx_global_stat_t gstat;
+ gmx_update_t *upd = NULL;
+ t_graph *graph = NULL;
+ gmx_signalling_t gs;
+ gmx_groups_t *groups;
+ gmx_ekindata_t *ekind;
+ gmx_shellfc_t *shellfc;
+ gmx_bool bSumEkinhOld, bDoReplEx, bExchanged, bNeedRepartition;
+ gmx_bool bResetCountersHalfMaxH = FALSE;
+ gmx_bool bTemp, bPres, bTrotter;
+ real dvdl_constr;
+ rvec *cbuf = NULL;
+ int cbuf_nalloc = 0;
+ matrix lastbox;
+ int lamnew = 0;
+ /* for FEP */
+ int nstfep = 0;
+ double cycles;
+ real saved_conserved_quantity = 0;
+ real last_ekin = 0;
+ 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_int64_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.*/
+ /* PME load balancing data for GPU kernels */
+ pme_load_balancing_t *pme_loadbal = NULL;
+ gmx_bool bPMETune = FALSE;
+ gmx_bool bPMETunePrinting = FALSE;
+
+ /* Interactive MD */
+ gmx_bool bIMDstep = FALSE;
+ gmx_membed_t *membed = NULL;
+
+ /* PLUMED */
+ int plumedNeedsEnergy=0;
+ int plumedWantsToStop=0;
+ matrix plumed_vir;
+ /* END PLUMED */
+
+#ifdef GMX_FAHCORE
+ /* Temporary addition for FAHCORE checkpointing */
+ int chkpt_ret;
+#endif
+ /* Domain decomposition could incorrectly miss a bonded
+ interaction, but checking for that requires a global
+ communication stage, which does not otherwise happen in DD
+ code. So we do that alongside the first global energy reduction
+ after a new DD is made. These variables handle whether the
+ check happens, and the result it returns. */
+ bool shouldCheckNumberOfBondedInteractions = false;
+ int totalNumberOfBondedInteractions = -1;
+
+ /* Check for special mdrun options */
+ bRerunMD = (Flags & MD_RERUN);
+ 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; */
+ bTrotter = (EI_VV(ir->eI) && (inputrecNptTrotter(ir) || inputrecNphTrotter(ir) || inputrecNvtTrotter(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;
+ }
+
+ nstglobalcomm = check_nstglobalcomm(fplog, cr, nstglobalcomm, ir);
+ bGStatEveryStep = (nstglobalcomm == 1);
+
+ if (bRerunMD)
+ {
+ ir->nstxout_compressed = 0;
+ }
+ groups = &top_global->groups;
+
+ if (opt2bSet("-membed", nfile, fnm))
+ {
+ if (MASTER(cr))
+ {
+ fprintf(stderr, "Initializing membed");
+ }
+ /* Note that membed cannot work in parallel because mtop is
+ * changed here. Fix this if we ever want to make it run with
+ * multiple ranks. */
+ membed = init_membed(fplog, nfile, fnm, top_global, ir, state_global, cr, &cpt_period);
+ }
+
+ if (ir->eSwapCoords != eswapNO)
+ {
+ /* Initialize ion swapping code */
+ init_swapcoords(fplog, bVerbose, ir, opt2fn_master("-swap", nfile, fnm, cr),
+ top_global, state_global->x, state_global->box, &state_global->swapstate, cr, oenv, Flags);
+ }
+
+ /* Initial values */
+ init_md(fplog, cr, ir, oenv, &t, &t0, state_global->lambda,
+ &(state_global->fep_state), lam0,
+ nrnb, top_global, &upd,
+ nfile, fnm, &outf, &mdebin,
+ force_vir, shake_vir, mu_tot, &bSimAnn, &vcm, Flags, wcycle);
+
+ clear_mat(total_vir);
+ clear_mat(pres);
+ /* Energy terms and groups */
+ snew(enerd, 1);
+ init_enerdata(top_global->groups.grps[egcENER].nr, ir->fepvals->n_lambda,
+ 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);
+ /* Copy the cos acceleration to the groups struct */
+ ekind->cosacc.cos_accel = ir->cos_accel;
+
+ gstat = global_stat_init(ir);
+
+ /* Check for polarizable models and flexible constraints */
+ shellfc = init_shell_flexcon(fplog,
+ top_global, n_flexible_constraints(constr),
+ ir->nstcalcenergy, DOMAINDECOMP(cr));
+
+ if (shellfc && ir->nstcalcenergy != 1)
+ {
+ gmx_fatal(FARGS, "You have nstcalcenergy set to a value (%d) that is different from 1.\nThis is not supported in combinations with shell particles.\nPlease make a new tpr file.", ir->nstcalcenergy);
+ }
+ if (shellfc && DOMAINDECOMP(cr))
+ {
+ gmx_fatal(FARGS, "Shell particles are not implemented with domain decomposition, use a single rank");
+ }
+
+ if (inputrecDeform(ir))
+ {
+ tMPI_Thread_mutex_lock(&deform_init_box_mutex);
+ set_deform_reference_box(upd,
+ deform_init_init_step_tpx,
+ deform_init_box_tpx);
+ tMPI_Thread_mutex_unlock(&deform_init_box_mutex);
+ }
+
+ {
+ 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);
+ }
+ else
+ {
+ top = gmx_mtop_generate_local_top(top_global, ir->efep != efepNO);
+
+ forcerec_set_excl_load(fr, top);
+
+ state = serial_init_local_state(state_global);
+
+ atoms2md(top_global, ir, 0, NULL, top_global->natoms, mdatoms);
+
+ if (vsite)
+ {
+ set_vsite_top(vsite, top, mdatoms, cr);
+ }
+
+ if (ir->ePBC != epbcNONE && !fr->bMolPBC)
+ {
+ graph = mk_graph(fplog, &(top->idef), 0, top_global->natoms, FALSE, FALSE);
+ }
+
+ if (shellfc)
+ {
+ make_local_shells(cr, mdatoms, shellfc);
+ }
+
+ setup_bonded_threading(fr, &top->idef);
+
+ update_realloc(upd, state->nalloc);
+ }
+
+ /* Set up interactive MD (IMD) */
+ init_IMD(ir, cr, top_global, fplog, ir->nstcalcenergy, state_global->x,
+ nfile, fnm, oenv, imdport, Flags);
+
+ 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, constr,
+ nrnb, NULL, FALSE);
+ shouldCheckNumberOfBondedInteractions = true;
+ update_realloc(upd, state->nalloc);
+ }
+
+ update_mdatoms(mdatoms, state->lambda[efptMASS]);
+
+ startingFromCheckpoint = Flags & MD_STARTFROMCPT;
+
+ if (ir->bExpanded)
+ {
+ init_expanded_ensemble(startingFromCheckpoint, ir, &state->dfhist);
+ }
+
+ if (MASTER(cr))
+ {
+ if (startingFromCheckpoint)
+ {
+ /* 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);
+ }
+
+ /* Initialize constraints */
+ if (constr && !DOMAINDECOMP(cr))
+ {
+ set_constraints(constr, top, ir, mdatoms, cr);
+ }
+
+ if (repl_ex_nst > 0 && MASTER(cr))
+ {
+ repl_ex = init_replica_exchange(fplog, cr->ms, state_global, ir,
+ repl_ex_nst, repl_ex_nex, repl_ex_seed);
+ }
+
+ /* PME tuning is only supported with PME for Coulomb. Is is not supported
+ * with only LJ PME, or for reruns.
+ */
+ bPMETune = ((Flags & MD_TUNEPME) && EEL_PME(fr->eeltype) && !bRerunMD &&
+ !(Flags & MD_REPRODUCIBLE));
+ if (bPMETune)
+ {
+ pme_loadbal_init(&pme_loadbal, cr, fplog, ir, state->box,
+ fr->ic, fr->pmedata, use_GPU(fr->nbv),
+ &bPMETunePrinting);
+ }
+
+ if (!ir->bContinuation && !bRerunMD)
+ {
+ if (mdatoms->cFREEZE && (state->flags & (1<<estV)))
+ {
+ /* Set the velocities of frozen particles to zero */
+ for (i = 0; i < 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,
+ cr, nrnb, fr, top);
+ }
+ if (vsite)
+ {
+ /* Construct the virtual sites for the initial configuration */
+ construct_vsites(vsite, state->x, ir->delta_t, NULL,
+ top->idef.iparams, top->idef.il,
+ fr->ePBC, fr->bMolPBC, cr, state->box);
+ }
+ }
+
+ if (ir->efep != efepNO)
+ {
+ /* Set free energy calculation frequency as the greatest common
+ * denominator of nstdhdl and repl_ex_nst. */
+ nstfep = ir->fepvals->nstdhdl;
+ if (ir->bExpanded)
+ {
+ nstfep = gmx_greatest_common_divisor(ir->expandedvals->nstexpanded, nstfep);
+ }
+ if (repl_ex_nst > 0)
+ {
+ nstfep = gmx_greatest_common_divisor(repl_ex_nst, nstfep);
+ }
+ }
+
+ /* Be REALLY careful about what flags you set here. You CANNOT assume
+ * this is the first step, since we might be restarting from a checkpoint,
+ * and in that case we should not do any modifications to the state.
+ */
+ bStopCM = (ir->comm_mode != ecmNO && !ir->bContinuation);
+
+ if (Flags & MD_READ_EKIN)
+ {
+ restore_ekinstate_from_state(cr, ekind, &state_global->ekinstate);
+ }
+
+ cglo_flags = (CGLO_TEMPERATURE | CGLO_GSTAT
+ | (bStopCM ? CGLO_STOPCM : 0)
+ | (EI_VV(ir->eI) ? CGLO_PRESSURE : 0)
+ | (EI_VV(ir->eI) ? CGLO_CONSTRAINT : 0)
+ | ((Flags & MD_READ_EKIN) ? CGLO_READEKIN : 0));
+
+ bSumEkinhOld = FALSE;
+ compute_globals(fplog, gstat, cr, ir, fr, ekind, state, mdatoms, nrnb, vcm,
+ NULL, enerd, force_vir, shake_vir, total_vir, pres, mu_tot,
+ constr, NULL, FALSE, state->box,
+ &totalNumberOfBondedInteractions, &bSumEkinhOld, cglo_flags
+ | (shouldCheckNumberOfBondedInteractions ? CGLO_CHECK_NUMBER_OF_BONDED_INTERACTIONS : 0));
+ checkNumberOfBondedInteractions(fplog, cr, totalNumberOfBondedInteractions,
+ top_global, top, state,
+ &shouldCheckNumberOfBondedInteractions);
+ 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, mdatoms, nrnb, vcm,
+ NULL, enerd, force_vir, shake_vir, total_vir, pres, mu_tot,
+ constr, NULL, FALSE, state->box,
+ NULL, &bSumEkinhOld,
+ cglo_flags &~(CGLO_STOPCM | 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 */
+ }
+
+ /* 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));
+ }
+ if (EI_STATE_VELOCITY(ir->eI))
+ {
+ 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");
+ }
+
+ /* PLUMED */
+ if(plumedswitch){
+ /* detect plumed API version */
+ int pversion=0;
+ plumed_cmd(plumedmain,"getApiVersion",&pversion);
+ /* setting kbT is only implemented with api>1) */
+ real kbT=ir->opts.ref_t[0]*BOLTZ;
+ if(pversion>1) plumed_cmd(plumedmain,"setKbT",&kbT);
+ if(pversion>2){
+ int res=1;
+ if( (Flags & MD_STARTFROMCPT) ) plumed_cmd(plumedmain,"setRestart",&res);
+ }
+
+ if(cr->ms && cr->ms->nsim>1) {
+ if(MASTER(cr)) plumed_cmd(plumedmain,"GREX setMPIIntercomm",&cr->ms->mpi_comm_masters);
+ if(PAR(cr)){
+ if(DOMAINDECOMP(cr)) {
+ plumed_cmd(plumedmain,"GREX setMPIIntracomm",&cr->dd->mpi_comm_all);
+ }else{
+ plumed_cmd(plumedmain,"GREX setMPIIntracomm",&cr->mpi_comm_mysim);
+ }
+ }
+ plumed_cmd(plumedmain,"GREX init",NULL);
+ }
+ if(PAR(cr)){
+ if(DOMAINDECOMP(cr)) {
+ plumed_cmd(plumedmain,"setMPIComm",&cr->dd->mpi_comm_all);
+ }
+ }
+ plumed_cmd(plumedmain,"setNatoms",&top_global->natoms);
+ plumed_cmd(plumedmain,"setMDEngine","gromacs");
+ plumed_cmd(plumedmain,"setLog",fplog);
+ real real_delta_t;
+ real_delta_t=ir->delta_t;
+ plumed_cmd(plumedmain,"setTimestep",&real_delta_t);
+ plumed_cmd(plumedmain,"init",NULL);
+
+ if(PAR(cr)){
+ if(DOMAINDECOMP(cr)) {
+ plumed_cmd(plumedmain,"setAtomsNlocal",&cr->dd->nat_home);
+ plumed_cmd(plumedmain,"setAtomsGatindex",cr->dd->gatindex);
+ }
+ }
+ }
+ /* END PLUMED */
+
+ walltime_accounting_start(walltime_accounting);
+ wallcycle_start(wcycle, ewcRUN);
+ print_start(fplog, cr, walltime_accounting, "mdrun");
+
+ /* 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
+
+ /***********************************************************
+ *
+ * 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))
+ {
+ bLastStep = !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) < gmx::square(fr->rlist))
+ {
+ 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, &bLastStep);
+ }
+
+ 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 */
+ bInitStep = !startingFromCheckpoint || EI_VV(ir->eI);
+ bSumEkinhOld = FALSE;
+ bExchanged = FALSE;
+ bNeedRepartition = FALSE;
+
+ init_global_signals(&gs, cr, ir, repl_ex_nst);
+
+ step = ir->init_step;
+ step_rel = 0;
+
+ 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 = (bLastStep || (ir->nsteps >= 0 && step_rel > ir->nsteps) ||
+ ((multisim_nsteps >= 0) && (step_rel >= multisim_nsteps )));
+ while (!bLastStep)
+ {
+
+ /* Determine if this is a neighbor search step */
+ bNStList = (ir->nstlist > 0 && step % ir->nstlist == 0);
+
+ if (bPMETune && bNStList)
+ {
+ /* PME grid + cut-off optimization with GPUs or PME nodes */
+ pme_loadbal_do(pme_loadbal, cr,
+ (bVerbose && MASTER(cr)) ? stderr : NULL,
+ fplog,
+ ir, fr, state,
+ wcycle,
+ step, step_rel,
+ &bPMETunePrinting);
+ }
+
+ 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 || ir->bSimTemp)
+ {
+ /* find and set the current lambdas. If rerunning, we either read in a state, or a lambda value,
+ requiring different logic. */
+
+ set_current_lambdas(step, ir->fepvals, bRerunMD, &rerun_fr, state_global, state, lam0);
+ bDoDHDL = do_per_step(step, ir->fepvals->nstdhdl);
+ bDoFEP = ((ir->efep != efepNO) && do_per_step(step, nstfep));
+ bDoExpanded = (do_per_step(step, ir->expandedvals->nstexpanded)
+ && (ir->bExpanded) && (step > 0) && (!startingFromCheckpoint));
+ }
+
+ bDoReplEx = ((repl_ex_nst > 0) && (step > 0) && !bLastStep &&
+ do_per_step(step, repl_ex_nst));
+
+ if (bSimAnn)
+ {
+ update_annealing_target_temp(ir, t, upd);
+ }
+
+ 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 with domain decomposition, use a single rank");
+ }
+ 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(vsite, state->x, ir->delta_t, state->v,
+ top->idef.iparams, top->idef.il,
+ fr->ePBC, fr->bMolPBC, 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));
+
+ 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 */
+ bNS = (bFirstStep || bNStList || bExchanged || bNeedRepartition);
+ }
+
+ /* 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 after this step, > 0 means stop at next NS step */
+ if ( (gs.set[eglsSTOPCOND] < 0) ||
+ ( (gs.set[eglsSTOPCOND] > 0) && (bNStList || 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 triggers energy and virial calculation. Because this leads
+ * to different code paths, forces can be different. Thus for exact
+ * continuation we should avoid extra log output.
+ * Note that the || bLastStep can result in non-exact continuation
+ * beyond the last step. But we don't consider that to be an issue.
+ */
+ do_log = do_per_step(step, ir->nstlog) || (bFirstStep && !startingFromCheckpoint) || bLastStep || bRerunMD;
+ do_verbose = bVerbose &&
+ (step % stepout == 0 || bFirstStep || bLastStep || bRerunMD);
+
+ if (bNS && !(bFirstStep && ir->bContinuation && !bRerunMD))
+ {
+ if (bRerunMD)
+ {
+ bMasterState = TRUE;
+ }
+ else
+ {
+ bMasterState = FALSE;
+ /* Correct the new box if it is too skewed */
+ if (inputrecDynamicBox(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 */
+ dd_partition_system(fplog, step, cr,
+ bMasterState, nstglobalcomm,
+ state_global, top_global, ir,
+ state, &f, mdatoms, top, fr,
+ vsite, constr,
+ nrnb, wcycle,
+ do_verbose && !bPMETunePrinting);
+ shouldCheckNumberOfBondedInteractions = true;
+ update_realloc(upd, state->nalloc);
+
+ /* PLUMED */
+ if(plumedswitch){
+ plumed_cmd(plumedmain,"setAtomsNlocal",&cr->dd->nat_home);
+ plumed_cmd(plumedmain,"setAtomsGatindex",cr->dd->gatindex);
+ }
+ /* END PLUMED */
+ }
+ }
+
+ if (MASTER(cr) && do_log)
+ {
+ print_ebin_header(fplog, step, t); /* can we improve the information printed here? */
+ }
+
+ if (ir->efep != efepNO)
+ {
+ update_mdatoms(mdatoms, state->lambda[efptMASS]);
+ }
+
+ if ((bRerunMD && rerun_fr.bV) || bExchanged)
+ {
+
+ /* 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, mdatoms, nrnb, vcm,
+ wcycle, enerd, NULL, NULL, NULL, NULL, mu_tot,
+ constr, NULL, FALSE, state->box,
+ &totalNumberOfBondedInteractions, &bSumEkinhOld,
+ CGLO_GSTAT | CGLO_TEMPERATURE | CGLO_CHECK_NUMBER_OF_BONDED_INTERACTIONS);
+ checkNumberOfBondedInteractions(fplog, cr, totalNumberOfBondedInteractions,
+ top_global, top, state,
+ &shouldCheckNumberOfBondedInteractions);
+ }
+ clear_mat(force_vir);
+
+ /* 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).
+ */
+ if (EI_VV(ir->eI) && (!bInitStep))
+ {
+ /* for vv, the first half of the integration actually corresponds
+ to the previous step. bCalcEner is only required to be evaluated on the 'next' step,
+ but the virial needs to be calculated on both the current step and the 'next' step. Future
+ reorganization may be able to get rid of one of the bCalcVir=TRUE steps. */
+
+ /* TODO: This is probably not what we want, we will write to energy file one step after nstcalcenergy steps. */
+ bCalcEnerStep = do_per_step(step - 1, ir->nstcalcenergy);
+ bCalcVir = bCalcEnerStep ||
+ (ir->epc != epcNO && (do_per_step(step, ir->nstpcouple) || do_per_step(step-1, ir->nstpcouple)));
+ }
+ else
+ {
+ bCalcEnerStep = do_per_step(step, ir->nstcalcenergy);
+ bCalcVir = bCalcEnerStep ||
+ (ir->epc != epcNO && do_per_step(step, ir->nstpcouple));
+ }
+ bCalcEner = bCalcEnerStep;
+
+ do_ene = (do_per_step(step, ir->nstenergy) || bLastStep || bRerunMD);
+
+ if (do_ene || do_log || bDoReplEx)
+ {
+ bCalcVir = TRUE;
+ bCalcEner = TRUE;
+ }
+
+ /* Do we need global communication ? */
+ bGStat = (bCalcVir || bCalcEner || bStopCM ||
+ do_per_step(step, nstglobalcomm) ||
+ (EI_VV(ir->eI) && inputrecNvtTrotter(ir) && do_per_step(step-1, nstglobalcomm)));
+
+ force_flags = (GMX_FORCE_STATECHANGED |
+ ((inputrecDynamicBox(ir) || bRerunMD) ? GMX_FORCE_DYNAMICBOX : 0) |
+ GMX_FORCE_ALLFORCES |
+ (bCalcVir ? GMX_FORCE_VIRIAL : 0) |
+ (bCalcEner ? GMX_FORCE_ENERGY : 0) |
+ (bDoFEP ? GMX_FORCE_DHDL : 0)
+ );
+
+ if (shellfc)
+ {
+ /* Now is the time to relax the shells */
+ relax_shell_flexcon(fplog, cr, bVerbose, step,
+ ir, bNS, force_flags, top,
+ constr, enerd, fcd,
+ state, f, force_vir, mdatoms,
+ nrnb, wcycle, graph, groups,
+ shellfc, fr, bBornRadii, t, mu_tot,
+ vsite, mdoutf_get_fp_field(outf));
+ }
+ 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
+ */
+
+ /* PLUMED */
+ plumedNeedsEnergy=0;
+ if(plumedswitch){
+ long int lstep=step; plumed_cmd(plumedmain,"setStepLong",&lstep);
+ plumed_cmd(plumedmain,"setPositions",&state->x[0][0]);
+ plumed_cmd(plumedmain,"setMasses",&mdatoms->massT[0]);
+ plumed_cmd(plumedmain,"setCharges",&mdatoms->chargeA[0]);
+ plumed_cmd(plumedmain,"setBox",&state->box[0][0]);
+ plumed_cmd(plumedmain,"prepareCalc",NULL);
+ plumed_cmd(plumedmain,"setStopFlag",&plumedWantsToStop);
+ plumed_cmd(plumedmain,"setForces",&f[0][0]);
+ plumed_cmd(plumedmain,"isEnergyNeeded",&plumedNeedsEnergy);
+ clear_mat(plumed_vir);
+ plumed_cmd(plumedmain,"setVirial",&plumed_vir[0][0]);
+ }
+ /* END PLUMED */
+ do_force(fplog, cr, ir, step, nrnb, wcycle, top, groups,
+ state->box, state->x, &state->hist,
+ f, force_vir, mdatoms, enerd, fcd,
+ state->lambda, graph,
+ fr, vsite, mu_tot, t, mdoutf_get_fp_field(outf), ed, bBornRadii,
+ (bNS ? GMX_FORCE_NS : 0) | force_flags);
+ /* PLUMED */
+ if(plumedswitch){
+ if(plumedNeedsEnergy){
+ msmul(force_vir,2.0,plumed_vir);
+ plumed_cmd(plumedmain,"setEnergy",&enerd->term[F_EPOT]);
+ plumed_cmd(plumedmain,"performCalc",NULL);
+ msmul(plumed_vir,0.5,force_vir);
+ } else {
+ msmul(plumed_vir,0.5,plumed_vir);
+ m_add(force_vir,plumed_vir,force_vir);
+ }
+ if ((repl_ex_nst > 0) && (step > 0) && !bLastStep &&
+ do_per_step(step,repl_ex_nst)) plumed_cmd(plumedmain,"GREX savePositions",NULL);
+ if(plumedWantsToStop) ir->nsteps=step_rel+1;
+ }
+ /* END PLUMED */
+ }
+
+ if (EI_VV(ir->eI) && !startingFromCheckpoint && !bRerunMD)
+ /* ############### START FIRST UPDATE HALF-STEP FOR VV METHODS############### */
+ {
+ rvec *vbuf = NULL;
+
+ wallcycle_start(wcycle, ewcUPDATE);
+ 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.
+ */
+ snew(vbuf, state->natoms);
+ copy_rvecn(state->v, vbuf, 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, fcd,
+ ekind, M, upd, etrtVELOCITY1,
+ cr, constr);
+
+ if (!bRerunMD || rerun_fr.bV || bForceUpdate) /* Why is rerun_fr.bV here? Unclear. */
+ {
+ wallcycle_stop(wcycle, ewcUPDATE);
+ update_constraints(fplog, step, NULL, ir, mdatoms,
+ state, fr->bMolPBC, graph, f,
+ &top->idef, shake_vir,
+ cr, nrnb, wcycle, upd, constr,
+ TRUE, bCalcVir);
+ wallcycle_start(wcycle, ewcUPDATE);
+ }
+ 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 || inputrecNptTrotter(ir)); */
+ /*bTemp = ((ir->eI==eiVV &&(!bInitStep)) || (ir->eI==eiVVAK && inputrecNptTrotter(ir)));*/
+ bPres = TRUE;
+ bTemp = ((ir->eI == eiVV && (!bInitStep)) || (ir->eI == eiVVAK));
+ if (bCalcEner && ir->eI == eiVVAK)
+ {
+ bSumEkinhOld = TRUE;
+ }
+ /* for vv, the first half of the integration actually corresponds to the previous step.
+ So we need information from the last step in the first half of the integration */
+ if (bGStat || do_per_step(step-1, nstglobalcomm))
+ {
+ wallcycle_stop(wcycle, ewcUPDATE);
+ compute_globals(fplog, gstat, cr, ir, fr, ekind, state, mdatoms, nrnb, vcm,
+ wcycle, enerd, force_vir, shake_vir, total_vir, pres, mu_tot,
+ constr, NULL, FALSE, state->box,
+ &totalNumberOfBondedInteractions, &bSumEkinhOld,
+ (bGStat ? CGLO_GSTAT : 0)
+ | CGLO_ENERGY
+ | (bTemp ? CGLO_TEMPERATURE : 0)
+ | (bPres ? CGLO_PRESSURE : 0)
+ | (bPres ? CGLO_CONSTRAINT : 0)
+ | (bStopCM ? CGLO_STOPCM : 0)
+ | (shouldCheckNumberOfBondedInteractions ? CGLO_CHECK_NUMBER_OF_BONDED_INTERACTIONS : 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 temperature control, we still feed in
+ EkinAveVel because it's needed for the pressure */
+ checkNumberOfBondedInteractions(fplog, cr, totalNumberOfBondedInteractions,
+ top_global, top, state,
+ &shouldCheckNumberOfBondedInteractions);
+ wallcycle_start(wcycle, ewcUPDATE);
+ }
+ /* temperature scaling and pressure scaling to produce the extended variables at t+dt */
+ if (!bInitStep)
+ {
+ if (bTrotter)
+ {
+ m_add(force_vir, shake_vir, total_vir); /* we need the un-dispersion corrected total vir here */
+ trotter_update(ir, step, ekind, enerd, state, total_vir, mdatoms, &MassQ, trotter_seq, ettTSEQ2);
+
+ copy_mat(shake_vir, state->svir_prev);
+ copy_mat(force_vir, state->fvir_prev);
+ if (inputrecNvtTrotter(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);
+ enerd->term[F_EKIN] = trace(ekind->ekin);
+ }
+ }
+ else if (bExchanged)
+ {
+ wallcycle_stop(wcycle, ewcUPDATE);
+ /* 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, mdatoms, nrnb, vcm,
+ wcycle, enerd, NULL, NULL, NULL, NULL, mu_tot,
+ constr, NULL, FALSE, state->box,
+ NULL, &bSumEkinhOld,
+ CGLO_GSTAT | CGLO_TEMPERATURE);
+ wallcycle_start(wcycle, ewcUPDATE);
+ }
+ }
+ /* if it's the initial step, we performed this first step just to get the constraint virial */
+ if (ir->eI == eiVV && bInitStep)
+ {
+ copy_rvecn(vbuf, state->v, 0, state->natoms);
+ sfree(vbuf);
+ }
+ wallcycle_stop(wcycle, ewcUPDATE);
+ }
+
+ /* compute the conserved quantity */
+ if (EI_VV(ir->eI))
+ {
+ saved_conserved_quantity = compute_conserved_from_auxiliary(ir, state, &MassQ);
+ if (ir->eI == eiVV)
+ {
+ last_ekin = enerd->term[F_EKIN];
+ }
+ if ((ir->eDispCorr != edispcEnerPres) && (ir->eDispCorr != edispcAllEnerPres))
+ {
+ saved_conserved_quantity -= enerd->term[F_DISPCORR];
+ }
+ /* sum up the foreign energy and dhdl terms for vv. currently done every step so that dhdl is correct in the .edr */
+ if (ir->efep != efepNO && !bRerunMD)
+ {
+ sum_dhdl(enerd, state->lambda, ir->fepvals);
+ }
+ }
+
+ /* ######## END FIRST UPDATE STEP ############## */
+ /* ######## If doing VV, we now have v(dt) ###### */
+ if (bDoExpanded)
+ {
+ /* perform extended ensemble sampling in lambda - we don't
+ actually move to the new state before outputting
+ statistics, but if performing simulated tempering, we
+ do update the velocities and the tau_t. */
+
+ lamnew = ExpandedEnsembleDynamics(fplog, ir, enerd, state, &MassQ, state->fep_state, &state->dfhist, step, state->v, mdatoms);
+ /* history is maintained in state->dfhist, but state_global is what is sent to trajectory and log output */
+ copy_df_history(&state_global->dfhist, &state->dfhist);
+ }
+
+ /* Now we have the energies and forces corresponding to the
+ * coordinates at time t. We must output all of this before
+ * the update.
+ */
+ do_md_trajectory_writing(fplog, cr, nfile, fnm, step, step_rel, t,
+ ir, state, state_global, top_global, fr,
+ outf, mdebin, ekind, f,
+ &nchkpt,
+ bCPT, bRerunMD, bLastStep, (Flags & MD_CONFOUT),
+ bSumEkinhOld);
+ /* Check if IMD step and do IMD communication, if bIMD is TRUE. */
+ bIMDstep = do_IMD(ir->bIMD, step, cr, bNS, state->box, state->x, ir, t, wcycle);
+
+ /* kludge -- virial is lost with restart for MTTK NPT control. Must reload (saved earlier). */
+ if (startingFromCheckpoint && bTrotter)
+ {
+ copy_mat(state->svir_prev, shake_vir);
+ copy_mat(state->fvir_prev, force_vir);
+ }
+
+ elapsed_time = walltime_accounting_get_current_elapsed_time(walltime_accounting);
+
+ /* Check whether everything is still allright */
+ if (((int)gmx_get_stop_condition() > handled_stop_condition)
+#if GMX_THREAD_MPI
+ && MASTER(cr)
+#endif
+ )
+ {
+ int nsteps_stop = -1;
+
+ /* 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;
+ nsteps_stop = std::max(ir->nstlist, 2*nstglobalcomm);
+ }
+ if (gmx_get_stop_condition() == gmx_stop_cond_next)
+ {
+ gs.sig[eglsSTOPCOND] = -1;
+ nsteps_stop = nstglobalcomm + 1;
+ }
+ if (fplog)
+ {
+ fprintf(fplog,
+ "\n\nReceived the %s signal, stopping within %d steps\n\n",
+ gmx_get_signal_name(), nsteps_stop);
+ fflush(fplog);
+ }
+ fprintf(stderr,
+ "\n\nReceived the %s signal, stopping within %d steps\n\n",
+ gmx_get_signal_name(), nsteps_stop);
+ fflush(stderr);
+ handled_stop_condition = (int)gmx_get_stop_condition();
+ }
+ else if (MASTER(cr) && (bNS || ir->nstlist <= 0) &&
+ (max_hours > 0 && elapsed_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) &&
+ elapsed_time > max_hours*60.0*60.0*0.495)
+ {
+ /* Set flag that will communicate the signal to all ranks in the simulation */
+ gs.sig[eglsRESETCOUNTERS] = 1;
+ }
+
+ /* 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 ||
+ elapsed_time >= nchkpt*cpt_period*60.0)) &&
+ gs.set[eglsCHKPT] == 0)
+ {
+ gs.sig[eglsCHKPT] = 1;
+ }
+
+ /* ######### START SECOND UPDATE STEP ################# */
+
+ /* at the start of step, randomize or scale the velocities ((if vv. Restriction of Andersen controlled
+ in preprocessing */
+
+ if (ETC_ANDERSEN(ir->etc)) /* keep this outside of update_tcouple because of the extra info required to pass */
+ {
+ gmx_bool bIfRandomize;
+ bIfRandomize = update_randomize_velocities(ir, step, cr, mdatoms, state, upd, constr);
+ /* if we have constraints, we have to remove the kinetic energy parallel to the bonds */
+ if (constr && bIfRandomize)
+ {
+ update_constraints(fplog, step, NULL, ir, mdatoms,
+ state, fr->bMolPBC, graph, f,
+ &top->idef, tmp_vir,
+ cr, nrnb, wcycle, upd, constr,
+ TRUE, bCalcVir);
+ }
+ }
+ /* 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);
+
+ dvdl_constr = 0;
+
+ if (!bRerunMD || rerun_fr.bV || bForceUpdate)
+ {
+ wallcycle_start(wcycle, ewcUPDATE);
+ /* UPDATE PRESSURE VARIABLES IN TROTTER FORMULATION WITH CONSTRAINTS */
+ if (bTrotter)
+ {
+ 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(step, ir, state, ekind, &MassQ, mdatoms);
+ update_pcouple(fplog, step, ir, state, pcoupl_mu, M, bInitStep);
+ }
+
+ if (EI_VV(ir->eI))
+ {
+ /* velocity half-step update */
+ update_coords(fplog, step, ir, mdatoms, state, f, fcd,
+ ekind, M, upd, etrtVELOCITY2,
+ cr, constr);
+ }
+
+ /* Above, initialize just copies ekinh into ekin,
+ * it doesn't copy position (for VV),
+ * and entire integrator for MD.
+ */
+
+ if (ir->eI == eiVVAK)
+ {
+ /* We probably only need md->homenr, not state->natoms */
+ if (state->natoms > cbuf_nalloc)
+ {
+ cbuf_nalloc = state->natoms;
+ srenew(cbuf, cbuf_nalloc);
+ }
+ copy_rvecn(state->x, cbuf, 0, state->natoms);
+ }
+
+ update_coords(fplog, step, ir, mdatoms, state, f, fcd,
+ ekind, M, upd, etrtPOSITION, cr, constr);
+ wallcycle_stop(wcycle, ewcUPDATE);
+
+ update_constraints(fplog, step, &dvdl_constr, ir, mdatoms, state,
+ fr->bMolPBC, graph, f,
+ &top->idef, shake_vir,
+ cr, nrnb, wcycle, upd, constr,
+ FALSE, bCalcVir);
+
+ 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, mdatoms, nrnb, vcm,
+ wcycle, enerd, force_vir, shake_vir, total_vir, pres, mu_tot,
+ constr, NULL, FALSE, lastbox,
+ NULL, &bSumEkinhOld,
+ (bGStat ? CGLO_GSTAT : 0) | 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, fcd,
+ ekind, M, upd, etrtPOSITION, cr, constr);
+ 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, NULL, ir, mdatoms,
+ state, fr->bMolPBC, graph, f,
+ &top->idef, tmp_vir,
+ cr, nrnb, wcycle, upd, NULL,
+ FALSE, bCalcVir);
+ }
+ if (EI_VV(ir->eI))
+ {
+ /* this factor or 2 correction is necessary
+ because half of the constraint force is removed
+ in the vv step, so we have to double it. See
+ the Redmine issue #1255. It is not yet clear
+ if the factor of 2 is exact, or just a very
+ good approximation, and this will be
+ investigated. The next step is to see if this
+ can be done adding a dhdl contribution from the
+ rattle step, but this is somewhat more
+ complicated with the current code. Will be
+ investigated, hopefully for 4.6.3. However,
+ this current solution is much better than
+ having it completely wrong.
+ */
+ enerd->term[F_DVDL_CONSTR] += 2*dvdl_constr;
+ }
+ else
+ {
+ enerd->term[F_DVDL_CONSTR] += dvdl_constr;
+ }
+ }
+ 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(vsite, state->x, ir->delta_t, state->v,
+ top->idef.iparams, top->idef.il,
+ fr->ePBC, fr->bMolPBC, cr, state->box);
+
+ if (graph != NULL)
+ {
+ unshift_self(graph, state->box, state->x);
+ }
+ wallcycle_stop(wcycle, ewcVSITECONSTR);
+ }
+
+ /* ############## IF NOT VV, Calculate globals HERE ############ */
+ /* With Leap-Frog we can skip compute_globals at
+ * non-communication steps, but we need to calculate
+ * the kinetic energy one step before communication.
+ */
+ if (bGStat || (!EI_VV(ir->eI) && do_per_step(step+1, nstglobalcomm)))
+ {
+ compute_globals(fplog, gstat, cr, ir, fr, ekind, state, mdatoms, nrnb, vcm,
+ wcycle, enerd, force_vir, shake_vir, total_vir, pres, mu_tot,
+ constr, &gs,
+ (step_rel % gs.nstms == 0) &&
+ (multisim_nsteps < 0 || (step_rel < multisim_nsteps)),
+ lastbox,
+ &totalNumberOfBondedInteractions, &bSumEkinhOld,
+ (bGStat ? CGLO_GSTAT : 0)
+ | (!EI_VV(ir->eI) || bRerunMD ? CGLO_ENERGY : 0)
+ | (!EI_VV(ir->eI) && bStopCM ? CGLO_STOPCM : 0)
+ | (!EI_VV(ir->eI) ? CGLO_TEMPERATURE : 0)
+ | (!EI_VV(ir->eI) || bRerunMD ? CGLO_PRESSURE : 0)
+ | CGLO_CONSTRAINT
+ | (shouldCheckNumberOfBondedInteractions ? CGLO_CHECK_NUMBER_OF_BONDED_INTERACTIONS : 0)
+ );
+ checkNumberOfBondedInteractions(fplog, cr, totalNumberOfBondedInteractions,
+ top_global, top, state,
+ &shouldCheckNumberOfBondedInteractions);
+ }
+
+ /* ############# 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 (ir->efep != efepNO && (!EI_VV(ir->eI) || bRerunMD))
+ {
+ /* Sum up the foreign energy and dhdl terms for md and sd.
+ Currently done every step so that dhdl is correct in the .edr */
+ sum_dhdl(enerd, state->lambda, ir->fepvals);
+ }
+ update_box(fplog, step, ir, mdatoms, state, f,
+ pcoupl_mu, nrnb, upd);
+
+ /* ################# END UPDATE STEP 2 ################# */
+ /* #### We now have r(t+dt) and v(t+dt/2) ############# */
+
+ /* The coordinates (x) were unshifted in update */
+ if (!bGStat)
+ {
+ /* We will not sum ekinh_old,
+ * so signal that we still have to do it.
+ */
+ bSumEkinhOld = TRUE;
+ }
+
+ /* ######### BEGIN PREPARING EDR OUTPUT ########### */
+
+ /* 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 (EI_VV(ir->eI))
+ {
+ 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);
+ }
+ /* ######### END PREPARING EDR OUTPUT ########### */
+
+ /* Output stuff */
+ if (MASTER(cr))
+ {
+ if (fplog && do_log && bDoExpanded)
+ {
+ /* only needed if doing expanded ensemble */
+ PrintFreeEnergyInfoToFile(fplog, ir->fepvals, ir->expandedvals, ir->bSimTemp ? ir->simtempvals : NULL,
+ &state_global->dfhist, state->fep_state, ir->nstlog, step);
+ }
+ if (bCalcEner)
+ {
+ upd_mdebin(mdebin, bDoDHDL, bCalcEnerStep,
+ t, mdatoms->tmass, enerd, state,
+ ir->fepvals, ir->expandedvals, lastbox,
+ shake_vir, force_vir, total_vir, pres,
+ ekind, mu_tot, constr);
+ }
+ else
+ {
+ upd_mdebin_step(mdebin);
+ }
+
+ gmx_bool do_dr = do_per_step(step, ir->nstdisreout);
+ gmx_bool do_or = do_per_step(step, ir->nstorireout);
+
+ print_ebin(mdoutf_get_fp_ene(outf), do_ene, do_dr, do_or, do_log ? fplog : NULL,
+ step, t,
+ eprNORMAL, mdebin, fcd, groups, &(ir->opts));
+
+ if (ir->bPull)
+ {
+ pull_print_output(ir->pull_work, step, t);
+ }
+
+ if (do_per_step(step, ir->nstlog))
+ {
+ if (fflush(fplog) != 0)
+ {
+ gmx_fatal(FARGS, "Cannot flush logfile - maybe you are out of disk space?");
+ }
+ }
+ }
+ if (bDoExpanded)
+ {
+ /* Have to do this part _after_ outputting the logfile and the edr file */
+ /* Gets written into the state at the beginning of next loop*/
+ state->fep_state = lamnew;
+ }
+ /* Print the remaining wall clock time for the run */
+ if (MULTIMASTER(cr) &&
+ (do_verbose || gmx_got_usr_signal()) &&
+ !bPMETunePrinting)
+ {
+ if (shellfc)
+ {
+ fprintf(stderr, "\n");
+ }
+ print_time(stderr, walltime_accounting, step, ir, cr);
+ }
+
+ /* Ion/water position swapping.
+ * Not done in last step since trajectory writing happens before this call
+ * in the MD loop and exchanges would be lost anyway. */
+ bNeedRepartition = FALSE;
+ if ((ir->eSwapCoords != eswapNO) && (step > 0) && !bLastStep &&
+ do_per_step(step, ir->swap->nstswap))
+ {
+ bNeedRepartition = do_swapcoords(cr, step, t, ir, wcycle,
+ bRerunMD ? rerun_fr.x : state->x,
+ bRerunMD ? rerun_fr.box : state->box,
+ top_global, MASTER(cr) && bVerbose, bRerunMD);
+
+ if (bNeedRepartition && DOMAINDECOMP(cr))
+ {
+ dd_collect_state(cr->dd, state, state_global);
+ }
+ }
+
+ /* Replica exchange */
+ bExchanged = FALSE;
+ if (bDoReplEx)
+ {
+ bExchanged = replica_exchange(fplog, cr, repl_ex,
+ state_global, enerd,
+ state, step, t);
+ }
+
+ if ( (bExchanged || bNeedRepartition) && DOMAINDECOMP(cr) )
+ {
+ dd_partition_system(fplog, step, cr, TRUE, 1,
+ state_global, top_global, ir,
+ state, &f, mdatoms, top, fr,
+ vsite, constr,
+ nrnb, wcycle, FALSE);
+ shouldCheckNumberOfBondedInteractions = true;
+ update_realloc(upd, state->nalloc);
+ }
+
+ bFirstStep = FALSE;
+ bInitStep = FALSE;
+ startingFromCheckpoint = 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 */
+ bLastStep = !read_next_frame(oenv, status, &rerun_fr);
+ }
+
+ if (PAR(cr))
+ {
+ rerun_parallel_comm(cr, &rerun_fr, &bLastStep);
+ }
+ }
+
+ cycles = wallcycle_stop(wcycle, ewcSTEP);
+ if (DOMAINDECOMP(cr) && wcycle)
+ {
+ dd_cycles_add(cr->dd, cycles, ddCyclStep);
+ }
+
+ if (!bRerunMD || !rerun_fr.bStep)
+ {
+ /* increase the MD step number */
+ step++;
+ step_rel++;
+ }
+
+ /* TODO make a counter-reset module */
+ /* If it is time to reset counters, set a flag that remains
+ true until counters actually get reset */
+ if (step_rel == wcycle_get_reset_counters(wcycle) ||
+ gs.set[eglsRESETCOUNTERS] != 0)
+ {
+ if (pme_loadbal_is_active(pme_loadbal))
+ {
+ /* Do not permit counter reset while PME load
+ * balancing is active. The only purpose for resetting
+ * counters is to measure reliable performance data,
+ * and that can't be done before balancing
+ * completes.
+ *
+ * TODO consider fixing this by delaying the reset
+ * until after load balancing completes,
+ * e.g. https://gerrit.gromacs.org/#/c/4964/2 */
+ gmx_fatal(FARGS, "PME tuning was still active when attempting to "
+ "reset mdrun counters at step %" GMX_PRId64 ". Try "
+ "resetting counters later in the run, e.g. with gmx "
+ "mdrun -resetstep.", step);
+ }
+ reset_all_counters(fplog, cr, step, &step_rel, ir, wcycle, nrnb, walltime_accounting,
+ use_GPU(fr->nbv) ? fr->nbv : NULL);
+ wcycle_set_reset_counters(wcycle, -1);
+ if (!(cr->duty & DUTY_PME))
+ {
+ /* Tell our PME node to reset its counters */
+ gmx_pme_send_resetcounters(cr, step);
+ }
+ /* Correct max_hours for the elapsed time */
+ max_hours -= elapsed_time/(60.0*60.0);
+ /* If mdrun -maxh -resethway was active, it can only trigger once */
+ bResetCountersHalfMaxH = FALSE; /* TODO move this to where gs.sig[eglsRESETCOUNTERS] is set */
+ /* Reset can only happen once, so clear the triggering flag. */
+ gs.set[eglsRESETCOUNTERS] = 0;
+ }
+
+ /* If bIMD is TRUE, the master updates the IMD energy record and sends positions to VMD client */
+ IMD_prep_energies_send_positions(ir->bIMD && MASTER(cr), bIMDstep, ir->imd, enerd, step, bCalcEner, wcycle);
+
+ }
+ /* End of main MD loop */
+
+ /* Closing TNG files can include compressing data. Therefore it is good to do that
+ * before stopping the time measurements. */
+ mdoutf_tng_close(outf);
+
+ /* Stop measuring walltime */
+ walltime_accounting_end(walltime_accounting);
+
+ if (bRerunMD && MASTER(cr))
+ {
+ close_trj(status);
+ }
+
+ if (!(cr->duty & DUTY_PME))
+ {
+ /* Tell the PME only node to finish */
+ gmx_pme_send_finish(cr);
+ }
+
+ if (MASTER(cr))
+ {
+ if (ir->nstcalcenergy > 0 && !bRerunMD)
+ {
+ print_ebin(mdoutf_get_fp_ene(outf), FALSE, FALSE, FALSE, fplog, step, t,
+ eprAVER, mdebin, fcd, groups, &(ir->opts));
+ }
+ }
+
+ done_mdoutf(outf);
+
+ if (bPMETune)
+ {
+ pme_loadbal_done(pme_loadbal, cr, fplog, use_GPU(fr->nbv));
+ }
+
+ done_shellfc(fplog, shellfc, step_rel);
+
+ if (repl_ex_nst > 0 && MASTER(cr))
+ {
+ print_replica_exchange_statistics(fplog, repl_ex);
+ }
+
+ // Clean up swapcoords
+ if (ir->eSwapCoords != eswapNO)
+ {
+ finish_swapcoords(ir->swap);
+ }
+
+ if (membed != nullptr)
+ {
+ free_membed(membed);
+ }
+
+ /* IMD cleanup, if bIMD is TRUE. */
+ IMD_finalize(ir->bIMD, ir->imd);
+
+ walltime_accounting_set_nsteps_done(walltime_accounting, step_rel);
+
+ return 0;
+}
diff --git a/patches/gromacs-2016-beta1.diff/src/programs/mdrun/md.cpp.preplumed b/patches/gromacs-2016-beta1.diff/src/programs/mdrun/md.cpp.preplumed
new file mode 100644
index 0000000000000000000000000000000000000000..92fa911eb42e08119a3cb688cadea70fab76148f
--- /dev/null
+++ b/patches/gromacs-2016-beta1.diff/src/programs/mdrun/md.cpp.preplumed
@@ -0,0 +1,1841 @@
+/*
+ * This file is part of the GROMACS molecular simulation package.
+ *
+ * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
+ * Copyright (c) 2001-2004, The GROMACS development team.
+ * Copyright (c) 2011,2012,2013,2014,2015,2016, by the GROMACS development team, led by
+ * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
+ * and including many others, as listed in the AUTHORS file in the
+ * top-level source directory and at http://www.gromacs.org.
+ *
+ * GROMACS is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public License
+ * as published by the Free Software Foundation; either version 2.1
+ * of the License, or (at your option) any later version.
+ *
+ * GROMACS is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with GROMACS; if not, see
+ * http://www.gnu.org/licenses, or write to the Free Software Foundation,
+ * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ * If you want to redistribute modifications to GROMACS, 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 http://www.gromacs.org.
+ *
+ * To help us fund GROMACS development, we humbly ask that you cite
+ * the research papers on the package. Check out http://www.gromacs.org.
+ */
+#include "gmxpre.h"
+
+#include "md.h"
+
+#include "config.h"
+
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#include <algorithm>
+
+#include "thread_mpi/threads.h"
+
+#include "gromacs/commandline/filenm.h"
+#include "gromacs/domdec/domdec.h"
+#include "gromacs/domdec/domdec_network.h"
+#include "gromacs/domdec/domdec_struct.h"
+#include "gromacs/ewald/pme.h"
+#include "gromacs/ewald/pme-load-balancing.h"
+#include "gromacs/fileio/trxio.h"
+#include "gromacs/gmxlib/md_logging.h"
+#include "gromacs/gmxlib/network.h"
+#include "gromacs/gmxlib/nrnb.h"
+#include "gromacs/gpu_utils/gpu_utils.h"
+#include "gromacs/imd/imd.h"
+#include "gromacs/listed-forces/manage-threading.h"
+#include "gromacs/math/functions.h"
+#include "gromacs/math/utilities.h"
+#include "gromacs/math/vec.h"
+#include "gromacs/math/vectypes.h"
+#include "gromacs/mdlib/compute_io.h"
+#include "gromacs/mdlib/constr.h"
+#include "gromacs/mdlib/ebin.h"
+#include "gromacs/mdlib/force.h"
+#include "gromacs/mdlib/force_flags.h"
+#include "gromacs/mdlib/forcerec.h"
+#include "gromacs/mdlib/md_support.h"
+#include "gromacs/mdlib/mdatoms.h"
+#include "gromacs/mdlib/mdebin.h"
+#include "gromacs/mdlib/mdoutf.h"
+#include "gromacs/mdlib/mdrun.h"
+#include "gromacs/mdlib/mdrun_signalling.h"
+#include "gromacs/mdlib/nb_verlet.h"
+#include "gromacs/mdlib/nbnxn_gpu_data_mgmt.h"
+#include "gromacs/mdlib/ns.h"
+#include "gromacs/mdlib/shellfc.h"
+#include "gromacs/mdlib/sighandler.h"
+#include "gromacs/mdlib/sim_util.h"
+#include "gromacs/mdlib/tgroup.h"
+#include "gromacs/mdlib/trajectory_writing.h"
+#include "gromacs/mdlib/update.h"
+#include "gromacs/mdlib/vcm.h"
+#include "gromacs/mdlib/vsite.h"
+#include "gromacs/mdtypes/commrec.h"
+#include "gromacs/mdtypes/df_history.h"
+#include "gromacs/mdtypes/energyhistory.h"
+#include "gromacs/mdtypes/fcdata.h"
+#include "gromacs/mdtypes/forcerec.h"
+#include "gromacs/mdtypes/group.h"
+#include "gromacs/mdtypes/inputrec.h"
+#include "gromacs/mdtypes/interaction_const.h"
+#include "gromacs/mdtypes/md_enums.h"
+#include "gromacs/mdtypes/mdatom.h"
+#include "gromacs/mdtypes/state.h"
+#include "gromacs/pbcutil/mshift.h"
+#include "gromacs/pbcutil/pbc.h"
+#include "gromacs/pulling/pull.h"
+#include "gromacs/swap/swapcoords.h"
+#include "gromacs/timing/wallcycle.h"
+#include "gromacs/timing/walltime_accounting.h"
+#include "gromacs/topology/atoms.h"
+#include "gromacs/topology/idef.h"
+#include "gromacs/topology/mtop_util.h"
+#include "gromacs/topology/topology.h"
+#include "gromacs/trajectory/trajectoryframe.h"
+#include "gromacs/utility/basedefinitions.h"
+#include "gromacs/utility/cstringutil.h"
+#include "gromacs/utility/fatalerror.h"
+#include "gromacs/utility/real.h"
+#include "gromacs/utility/smalloc.h"
+
+#include "deform.h"
+#include "membed.h"
+#include "repl_ex.h"
+
+#ifdef GMX_FAHCORE
+#include "corewrap.h"
+#endif
+
+/*! \brief Check whether bonded interactions are missing, if appropriate
+ *
+ * \param[in] fplog Log file pointer
+ * \param[in] cr Communication object
+ * \param[in] totalNumberOfBondedInteractions Result of the global reduction over the number of bonds treated in each domain
+ * \param[in] top_global Global topology for the error message
+ * \param[in] top_local Local topology for the error message
+ * \param[in] state Global state for the error message
+ * \param[inout] shouldCheckNumberOfBondedInteractions Whether we should do the check.
+ *
+ * \return Nothing, except that shouldCheckNumberOfBondedInteractions
+ * is always set to false after exit.
+ */
+static void checkNumberOfBondedInteractions(FILE *fplog, t_commrec *cr, int totalNumberOfBondedInteractions,
+ gmx_mtop_t *top_global, gmx_localtop_t *top_local, t_state *state,
+ bool *shouldCheckNumberOfBondedInteractions)
+{
+ if (*shouldCheckNumberOfBondedInteractions)
+ {
+ if (totalNumberOfBondedInteractions != cr->dd->nbonded_global)
+ {
+ dd_print_missing_interactions(fplog, cr, totalNumberOfBondedInteractions, top_global, top_local, state); // Does not return
+ }
+ *shouldCheckNumberOfBondedInteractions = false;
+ }
+}
+
+static void reset_all_counters(FILE *fplog, t_commrec *cr,
+ gmx_int64_t step,
+ gmx_int64_t *step_rel, t_inputrec *ir,
+ gmx_wallcycle_t wcycle, t_nrnb *nrnb,
+ gmx_walltime_accounting_t walltime_accounting,
+ struct nonbonded_verlet_t *nbv)
+{
+ char sbuf[STEPSTRSIZE];
+
+ /* Reset all the counters related to performance over the run */
+ md_print_warn(cr, fplog, "step %s: resetting all time and cycle counters\n",
+ gmx_step_str(step, sbuf));
+
+ if (use_GPU(nbv))
+ {
+ nbnxn_gpu_reset_timings(nbv);
+ resetGpuProfiler();
+ }
+
+ wallcycle_stop(wcycle, ewcRUN);
+ wallcycle_reset_all(wcycle);
+ if (DOMAINDECOMP(cr))
+ {
+ reset_dd_statistics_counters(cr->dd);
+ }
+ init_nrnb(nrnb);
+ ir->init_step += *step_rel;
+ ir->nsteps -= *step_rel;
+ *step_rel = 0;
+ wallcycle_start(wcycle, ewcRUN);
+ walltime_accounting_start(walltime_accounting);
+ print_date_and_time(fplog, cr->nodeid, "Restarted time", gmx_gettime());
+}
+
+/*! \libinternal
+ \copydoc integrator_t (FILE *fplog, t_commrec *cr,
+ int nfile, const t_filenm fnm[],
+ const gmx_output_env_t *oenv, gmx_bool bVerbose,
+ int nstglobalcomm,
+ gmx_vsite_t *vsite, gmx_constr_t constr,
+ int stepout,
+ t_inputrec *inputrec,
+ 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_nex, int repl_ex_seed,
+ real cpt_period, real max_hours,
+ int imdport,
+ unsigned long Flags,
+ gmx_walltime_accounting_t walltime_accounting)
+ */
+double gmx::do_md(FILE *fplog, t_commrec *cr, int nfile, const t_filenm fnm[],
+ const gmx_output_env_t *oenv, gmx_bool bVerbose,
+ 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_nex, int repl_ex_seed,
+ real cpt_period, real max_hours,
+ int imdport,
+ unsigned long Flags,
+ gmx_walltime_accounting_t walltime_accounting)
+{
+ gmx_mdoutf_t outf = NULL;
+ gmx_int64_t step, step_rel;
+ double elapsed_time;
+ double t, t0, lam0[efptNR];
+ gmx_bool bGStatEveryStep, bGStat, bCalcVir, bCalcEnerStep, bCalcEner;
+ gmx_bool bNS, bNStList, bSimAnn, bStopCM, bRerunMD,
+ bFirstStep, startingFromCheckpoint, bInitStep, bLastStep = FALSE,
+ bBornRadii;
+ gmx_bool bDoDHDL = FALSE, bDoFEP = FALSE, bDoExpanded = FALSE;
+ gmx_bool do_ene, do_log, do_verbose, bRerunWarnNoV = TRUE,
+ bForceUpdate = FALSE, bCPT;
+ 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;
+ matrix pcoupl_mu, M;
+ 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;
+ gmx_enerdata_t *enerd;
+ rvec *f = NULL;
+ gmx_global_stat_t gstat;
+ gmx_update_t *upd = NULL;
+ t_graph *graph = NULL;
+ gmx_signalling_t gs;
+ gmx_groups_t *groups;
+ gmx_ekindata_t *ekind;
+ gmx_shellfc_t *shellfc;
+ gmx_bool bSumEkinhOld, bDoReplEx, bExchanged, bNeedRepartition;
+ gmx_bool bResetCountersHalfMaxH = FALSE;
+ gmx_bool bTemp, bPres, bTrotter;
+ real dvdl_constr;
+ rvec *cbuf = NULL;
+ int cbuf_nalloc = 0;
+ matrix lastbox;
+ int lamnew = 0;
+ /* for FEP */
+ int nstfep = 0;
+ double cycles;
+ real saved_conserved_quantity = 0;
+ real last_ekin = 0;
+ 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_int64_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.*/
+ /* PME load balancing data for GPU kernels */
+ pme_load_balancing_t *pme_loadbal = NULL;
+ gmx_bool bPMETune = FALSE;
+ gmx_bool bPMETunePrinting = FALSE;
+
+ /* Interactive MD */
+ gmx_bool bIMDstep = FALSE;
+ gmx_membed_t *membed = NULL;
+
+#ifdef GMX_FAHCORE
+ /* Temporary addition for FAHCORE checkpointing */
+ int chkpt_ret;
+#endif
+ /* Domain decomposition could incorrectly miss a bonded
+ interaction, but checking for that requires a global
+ communication stage, which does not otherwise happen in DD
+ code. So we do that alongside the first global energy reduction
+ after a new DD is made. These variables handle whether the
+ check happens, and the result it returns. */
+ bool shouldCheckNumberOfBondedInteractions = false;
+ int totalNumberOfBondedInteractions = -1;
+
+ /* Check for special mdrun options */
+ bRerunMD = (Flags & MD_RERUN);
+ 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; */
+ bTrotter = (EI_VV(ir->eI) && (inputrecNptTrotter(ir) || inputrecNphTrotter(ir) || inputrecNvtTrotter(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;
+ }
+
+ nstglobalcomm = check_nstglobalcomm(fplog, cr, nstglobalcomm, ir);
+ bGStatEveryStep = (nstglobalcomm == 1);
+
+ if (bRerunMD)
+ {
+ ir->nstxout_compressed = 0;
+ }
+ groups = &top_global->groups;
+
+ if (opt2bSet("-membed", nfile, fnm))
+ {
+ if (MASTER(cr))
+ {
+ fprintf(stderr, "Initializing membed");
+ }
+ /* Note that membed cannot work in parallel because mtop is
+ * changed here. Fix this if we ever want to make it run with
+ * multiple ranks. */
+ membed = init_membed(fplog, nfile, fnm, top_global, ir, state_global, cr, &cpt_period);
+ }
+
+ if (ir->eSwapCoords != eswapNO)
+ {
+ /* Initialize ion swapping code */
+ init_swapcoords(fplog, bVerbose, ir, opt2fn_master("-swap", nfile, fnm, cr),
+ top_global, state_global->x, state_global->box, &state_global->swapstate, cr, oenv, Flags);
+ }
+
+ /* Initial values */
+ init_md(fplog, cr, ir, oenv, &t, &t0, state_global->lambda,
+ &(state_global->fep_state), lam0,
+ nrnb, top_global, &upd,
+ nfile, fnm, &outf, &mdebin,
+ force_vir, shake_vir, mu_tot, &bSimAnn, &vcm, Flags, wcycle);
+
+ clear_mat(total_vir);
+ clear_mat(pres);
+ /* Energy terms and groups */
+ snew(enerd, 1);
+ init_enerdata(top_global->groups.grps[egcENER].nr, ir->fepvals->n_lambda,
+ 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);
+ /* Copy the cos acceleration to the groups struct */
+ ekind->cosacc.cos_accel = ir->cos_accel;
+
+ gstat = global_stat_init(ir);
+
+ /* Check for polarizable models and flexible constraints */
+ shellfc = init_shell_flexcon(fplog,
+ top_global, n_flexible_constraints(constr),
+ ir->nstcalcenergy, DOMAINDECOMP(cr));
+
+ if (shellfc && ir->nstcalcenergy != 1)
+ {
+ gmx_fatal(FARGS, "You have nstcalcenergy set to a value (%d) that is different from 1.\nThis is not supported in combinations with shell particles.\nPlease make a new tpr file.", ir->nstcalcenergy);
+ }
+ if (shellfc && DOMAINDECOMP(cr))
+ {
+ gmx_fatal(FARGS, "Shell particles are not implemented with domain decomposition, use a single rank");
+ }
+
+ if (inputrecDeform(ir))
+ {
+ tMPI_Thread_mutex_lock(&deform_init_box_mutex);
+ set_deform_reference_box(upd,
+ deform_init_init_step_tpx,
+ deform_init_box_tpx);
+ tMPI_Thread_mutex_unlock(&deform_init_box_mutex);
+ }
+
+ {
+ 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);
+ }
+ else
+ {
+ top = gmx_mtop_generate_local_top(top_global, ir->efep != efepNO);
+
+ forcerec_set_excl_load(fr, top);
+
+ state = serial_init_local_state(state_global);
+
+ atoms2md(top_global, ir, 0, NULL, top_global->natoms, mdatoms);
+
+ if (vsite)
+ {
+ set_vsite_top(vsite, top, mdatoms, cr);
+ }
+
+ if (ir->ePBC != epbcNONE && !fr->bMolPBC)
+ {
+ graph = mk_graph(fplog, &(top->idef), 0, top_global->natoms, FALSE, FALSE);
+ }
+
+ if (shellfc)
+ {
+ make_local_shells(cr, mdatoms, shellfc);
+ }
+
+ setup_bonded_threading(fr, &top->idef);
+
+ update_realloc(upd, state->nalloc);
+ }
+
+ /* Set up interactive MD (IMD) */
+ init_IMD(ir, cr, top_global, fplog, ir->nstcalcenergy, state_global->x,
+ nfile, fnm, oenv, imdport, Flags);
+
+ 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, constr,
+ nrnb, NULL, FALSE);
+ shouldCheckNumberOfBondedInteractions = true;
+ update_realloc(upd, state->nalloc);
+ }
+
+ update_mdatoms(mdatoms, state->lambda[efptMASS]);
+
+ startingFromCheckpoint = Flags & MD_STARTFROMCPT;
+
+ if (ir->bExpanded)
+ {
+ init_expanded_ensemble(startingFromCheckpoint, ir, &state->dfhist);
+ }
+
+ if (MASTER(cr))
+ {
+ if (startingFromCheckpoint)
+ {
+ /* 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);
+ }
+
+ /* Initialize constraints */
+ if (constr && !DOMAINDECOMP(cr))
+ {
+ set_constraints(constr, top, ir, mdatoms, cr);
+ }
+
+ if (repl_ex_nst > 0 && MASTER(cr))
+ {
+ repl_ex = init_replica_exchange(fplog, cr->ms, state_global, ir,
+ repl_ex_nst, repl_ex_nex, repl_ex_seed);
+ }
+
+ /* PME tuning is only supported with PME for Coulomb. Is is not supported
+ * with only LJ PME, or for reruns.
+ */
+ bPMETune = ((Flags & MD_TUNEPME) && EEL_PME(fr->eeltype) && !bRerunMD &&
+ !(Flags & MD_REPRODUCIBLE));
+ if (bPMETune)
+ {
+ pme_loadbal_init(&pme_loadbal, cr, fplog, ir, state->box,
+ fr->ic, fr->pmedata, use_GPU(fr->nbv),
+ &bPMETunePrinting);
+ }
+
+ if (!ir->bContinuation && !bRerunMD)
+ {
+ if (mdatoms->cFREEZE && (state->flags & (1<<estV)))
+ {
+ /* Set the velocities of frozen particles to zero */
+ for (i = 0; i < 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,
+ cr, nrnb, fr, top);
+ }
+ if (vsite)
+ {
+ /* Construct the virtual sites for the initial configuration */
+ construct_vsites(vsite, state->x, ir->delta_t, NULL,
+ top->idef.iparams, top->idef.il,
+ fr->ePBC, fr->bMolPBC, cr, state->box);
+ }
+ }
+
+ if (ir->efep != efepNO)
+ {
+ /* Set free energy calculation frequency as the greatest common
+ * denominator of nstdhdl and repl_ex_nst. */
+ nstfep = ir->fepvals->nstdhdl;
+ if (ir->bExpanded)
+ {
+ nstfep = gmx_greatest_common_divisor(ir->expandedvals->nstexpanded, nstfep);
+ }
+ if (repl_ex_nst > 0)
+ {
+ nstfep = gmx_greatest_common_divisor(repl_ex_nst, nstfep);
+ }
+ }
+
+ /* Be REALLY careful about what flags you set here. You CANNOT assume
+ * this is the first step, since we might be restarting from a checkpoint,
+ * and in that case we should not do any modifications to the state.
+ */
+ bStopCM = (ir->comm_mode != ecmNO && !ir->bContinuation);
+
+ if (Flags & MD_READ_EKIN)
+ {
+ restore_ekinstate_from_state(cr, ekind, &state_global->ekinstate);
+ }
+
+ cglo_flags = (CGLO_TEMPERATURE | CGLO_GSTAT
+ | (bStopCM ? CGLO_STOPCM : 0)
+ | (EI_VV(ir->eI) ? CGLO_PRESSURE : 0)
+ | (EI_VV(ir->eI) ? CGLO_CONSTRAINT : 0)
+ | ((Flags & MD_READ_EKIN) ? CGLO_READEKIN : 0));
+
+ bSumEkinhOld = FALSE;
+ compute_globals(fplog, gstat, cr, ir, fr, ekind, state, mdatoms, nrnb, vcm,
+ NULL, enerd, force_vir, shake_vir, total_vir, pres, mu_tot,
+ constr, NULL, FALSE, state->box,
+ &totalNumberOfBondedInteractions, &bSumEkinhOld, cglo_flags
+ | (shouldCheckNumberOfBondedInteractions ? CGLO_CHECK_NUMBER_OF_BONDED_INTERACTIONS : 0));
+ checkNumberOfBondedInteractions(fplog, cr, totalNumberOfBondedInteractions,
+ top_global, top, state,
+ &shouldCheckNumberOfBondedInteractions);
+ 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, mdatoms, nrnb, vcm,
+ NULL, enerd, force_vir, shake_vir, total_vir, pres, mu_tot,
+ constr, NULL, FALSE, state->box,
+ NULL, &bSumEkinhOld,
+ cglo_flags &~(CGLO_STOPCM | 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 */
+ }
+
+ /* 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));
+ }
+ if (EI_STATE_VELOCITY(ir->eI))
+ {
+ 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");
+ }
+
+ walltime_accounting_start(walltime_accounting);
+ wallcycle_start(wcycle, ewcRUN);
+ print_start(fplog, cr, walltime_accounting, "mdrun");
+
+ /* 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
+
+ /***********************************************************
+ *
+ * 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))
+ {
+ bLastStep = !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) < gmx::square(fr->rlist))
+ {
+ 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, &bLastStep);
+ }
+
+ 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 */
+ bInitStep = !startingFromCheckpoint || EI_VV(ir->eI);
+ bSumEkinhOld = FALSE;
+ bExchanged = FALSE;
+ bNeedRepartition = FALSE;
+
+ init_global_signals(&gs, cr, ir, repl_ex_nst);
+
+ step = ir->init_step;
+ step_rel = 0;
+
+ 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 = (bLastStep || (ir->nsteps >= 0 && step_rel > ir->nsteps) ||
+ ((multisim_nsteps >= 0) && (step_rel >= multisim_nsteps )));
+ while (!bLastStep)
+ {
+
+ /* Determine if this is a neighbor search step */
+ bNStList = (ir->nstlist > 0 && step % ir->nstlist == 0);
+
+ if (bPMETune && bNStList)
+ {
+ /* PME grid + cut-off optimization with GPUs or PME nodes */
+ pme_loadbal_do(pme_loadbal, cr,
+ (bVerbose && MASTER(cr)) ? stderr : NULL,
+ fplog,
+ ir, fr, state,
+ wcycle,
+ step, step_rel,
+ &bPMETunePrinting);
+ }
+
+ 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 || ir->bSimTemp)
+ {
+ /* find and set the current lambdas. If rerunning, we either read in a state, or a lambda value,
+ requiring different logic. */
+
+ set_current_lambdas(step, ir->fepvals, bRerunMD, &rerun_fr, state_global, state, lam0);
+ bDoDHDL = do_per_step(step, ir->fepvals->nstdhdl);
+ bDoFEP = ((ir->efep != efepNO) && do_per_step(step, nstfep));
+ bDoExpanded = (do_per_step(step, ir->expandedvals->nstexpanded)
+ && (ir->bExpanded) && (step > 0) && (!startingFromCheckpoint));
+ }
+
+ bDoReplEx = ((repl_ex_nst > 0) && (step > 0) && !bLastStep &&
+ do_per_step(step, repl_ex_nst));
+
+ if (bSimAnn)
+ {
+ update_annealing_target_temp(ir, t, upd);
+ }
+
+ 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 with domain decomposition, use a single rank");
+ }
+ 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(vsite, state->x, ir->delta_t, state->v,
+ top->idef.iparams, top->idef.il,
+ fr->ePBC, fr->bMolPBC, 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));
+
+ 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 */
+ bNS = (bFirstStep || bNStList || bExchanged || bNeedRepartition);
+ }
+
+ /* 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 after this step, > 0 means stop at next NS step */
+ if ( (gs.set[eglsSTOPCOND] < 0) ||
+ ( (gs.set[eglsSTOPCOND] > 0) && (bNStList || 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 triggers energy and virial calculation. Because this leads
+ * to different code paths, forces can be different. Thus for exact
+ * continuation we should avoid extra log output.
+ * Note that the || bLastStep can result in non-exact continuation
+ * beyond the last step. But we don't consider that to be an issue.
+ */
+ do_log = do_per_step(step, ir->nstlog) || (bFirstStep && !startingFromCheckpoint) || bLastStep || bRerunMD;
+ do_verbose = bVerbose &&
+ (step % stepout == 0 || bFirstStep || bLastStep || bRerunMD);
+
+ if (bNS && !(bFirstStep && ir->bContinuation && !bRerunMD))
+ {
+ if (bRerunMD)
+ {
+ bMasterState = TRUE;
+ }
+ else
+ {
+ bMasterState = FALSE;
+ /* Correct the new box if it is too skewed */
+ if (inputrecDynamicBox(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 */
+ dd_partition_system(fplog, step, cr,
+ bMasterState, nstglobalcomm,
+ state_global, top_global, ir,
+ state, &f, mdatoms, top, fr,
+ vsite, constr,
+ nrnb, wcycle,
+ do_verbose && !bPMETunePrinting);
+ shouldCheckNumberOfBondedInteractions = true;
+ update_realloc(upd, state->nalloc);
+ }
+ }
+
+ if (MASTER(cr) && do_log)
+ {
+ print_ebin_header(fplog, step, t); /* can we improve the information printed here? */
+ }
+
+ if (ir->efep != efepNO)
+ {
+ update_mdatoms(mdatoms, state->lambda[efptMASS]);
+ }
+
+ if ((bRerunMD && rerun_fr.bV) || bExchanged)
+ {
+
+ /* 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, mdatoms, nrnb, vcm,
+ wcycle, enerd, NULL, NULL, NULL, NULL, mu_tot,
+ constr, NULL, FALSE, state->box,
+ &totalNumberOfBondedInteractions, &bSumEkinhOld,
+ CGLO_GSTAT | CGLO_TEMPERATURE | CGLO_CHECK_NUMBER_OF_BONDED_INTERACTIONS);
+ checkNumberOfBondedInteractions(fplog, cr, totalNumberOfBondedInteractions,
+ top_global, top, state,
+ &shouldCheckNumberOfBondedInteractions);
+ }
+ clear_mat(force_vir);
+
+ /* 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).
+ */
+ if (EI_VV(ir->eI) && (!bInitStep))
+ {
+ /* for vv, the first half of the integration actually corresponds
+ to the previous step. bCalcEner is only required to be evaluated on the 'next' step,
+ but the virial needs to be calculated on both the current step and the 'next' step. Future
+ reorganization may be able to get rid of one of the bCalcVir=TRUE steps. */
+
+ /* TODO: This is probably not what we want, we will write to energy file one step after nstcalcenergy steps. */
+ bCalcEnerStep = do_per_step(step - 1, ir->nstcalcenergy);
+ bCalcVir = bCalcEnerStep ||
+ (ir->epc != epcNO && (do_per_step(step, ir->nstpcouple) || do_per_step(step-1, ir->nstpcouple)));
+ }
+ else
+ {
+ bCalcEnerStep = do_per_step(step, ir->nstcalcenergy);
+ bCalcVir = bCalcEnerStep ||
+ (ir->epc != epcNO && do_per_step(step, ir->nstpcouple));
+ }
+ bCalcEner = bCalcEnerStep;
+
+ do_ene = (do_per_step(step, ir->nstenergy) || bLastStep || bRerunMD);
+
+ if (do_ene || do_log || bDoReplEx)
+ {
+ bCalcVir = TRUE;
+ bCalcEner = TRUE;
+ }
+
+ /* Do we need global communication ? */
+ bGStat = (bCalcVir || bCalcEner || bStopCM ||
+ do_per_step(step, nstglobalcomm) ||
+ (EI_VV(ir->eI) && inputrecNvtTrotter(ir) && do_per_step(step-1, nstglobalcomm)));
+
+ force_flags = (GMX_FORCE_STATECHANGED |
+ ((inputrecDynamicBox(ir) || bRerunMD) ? GMX_FORCE_DYNAMICBOX : 0) |
+ GMX_FORCE_ALLFORCES |
+ (bCalcVir ? GMX_FORCE_VIRIAL : 0) |
+ (bCalcEner ? GMX_FORCE_ENERGY : 0) |
+ (bDoFEP ? GMX_FORCE_DHDL : 0)
+ );
+
+ if (shellfc)
+ {
+ /* Now is the time to relax the shells */
+ relax_shell_flexcon(fplog, cr, bVerbose, step,
+ ir, bNS, force_flags, top,
+ constr, enerd, fcd,
+ state, f, force_vir, mdatoms,
+ nrnb, wcycle, graph, groups,
+ shellfc, fr, bBornRadii, t, mu_tot,
+ vsite, mdoutf_get_fp_field(outf));
+ }
+ 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, groups,
+ state->box, state->x, &state->hist,
+ f, force_vir, mdatoms, enerd, fcd,
+ state->lambda, graph,
+ fr, vsite, mu_tot, t, mdoutf_get_fp_field(outf), ed, bBornRadii,
+ (bNS ? GMX_FORCE_NS : 0) | force_flags);
+ }
+
+ if (EI_VV(ir->eI) && !startingFromCheckpoint && !bRerunMD)
+ /* ############### START FIRST UPDATE HALF-STEP FOR VV METHODS############### */
+ {
+ rvec *vbuf = NULL;
+
+ wallcycle_start(wcycle, ewcUPDATE);
+ 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.
+ */
+ snew(vbuf, state->natoms);
+ copy_rvecn(state->v, vbuf, 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, fcd,
+ ekind, M, upd, etrtVELOCITY1,
+ cr, constr);
+
+ if (!bRerunMD || rerun_fr.bV || bForceUpdate) /* Why is rerun_fr.bV here? Unclear. */
+ {
+ wallcycle_stop(wcycle, ewcUPDATE);
+ update_constraints(fplog, step, NULL, ir, mdatoms,
+ state, fr->bMolPBC, graph, f,
+ &top->idef, shake_vir,
+ cr, nrnb, wcycle, upd, constr,
+ TRUE, bCalcVir);
+ wallcycle_start(wcycle, ewcUPDATE);
+ }
+ 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 || inputrecNptTrotter(ir)); */
+ /*bTemp = ((ir->eI==eiVV &&(!bInitStep)) || (ir->eI==eiVVAK && inputrecNptTrotter(ir)));*/
+ bPres = TRUE;
+ bTemp = ((ir->eI == eiVV && (!bInitStep)) || (ir->eI == eiVVAK));
+ if (bCalcEner && ir->eI == eiVVAK)
+ {
+ bSumEkinhOld = TRUE;
+ }
+ /* for vv, the first half of the integration actually corresponds to the previous step.
+ So we need information from the last step in the first half of the integration */
+ if (bGStat || do_per_step(step-1, nstglobalcomm))
+ {
+ wallcycle_stop(wcycle, ewcUPDATE);
+ compute_globals(fplog, gstat, cr, ir, fr, ekind, state, mdatoms, nrnb, vcm,
+ wcycle, enerd, force_vir, shake_vir, total_vir, pres, mu_tot,
+ constr, NULL, FALSE, state->box,
+ &totalNumberOfBondedInteractions, &bSumEkinhOld,
+ (bGStat ? CGLO_GSTAT : 0)
+ | CGLO_ENERGY
+ | (bTemp ? CGLO_TEMPERATURE : 0)
+ | (bPres ? CGLO_PRESSURE : 0)
+ | (bPres ? CGLO_CONSTRAINT : 0)
+ | (bStopCM ? CGLO_STOPCM : 0)
+ | (shouldCheckNumberOfBondedInteractions ? CGLO_CHECK_NUMBER_OF_BONDED_INTERACTIONS : 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 temperature control, we still feed in
+ EkinAveVel because it's needed for the pressure */
+ checkNumberOfBondedInteractions(fplog, cr, totalNumberOfBondedInteractions,
+ top_global, top, state,
+ &shouldCheckNumberOfBondedInteractions);
+ wallcycle_start(wcycle, ewcUPDATE);
+ }
+ /* temperature scaling and pressure scaling to produce the extended variables at t+dt */
+ if (!bInitStep)
+ {
+ if (bTrotter)
+ {
+ m_add(force_vir, shake_vir, total_vir); /* we need the un-dispersion corrected total vir here */
+ trotter_update(ir, step, ekind, enerd, state, total_vir, mdatoms, &MassQ, trotter_seq, ettTSEQ2);
+
+ copy_mat(shake_vir, state->svir_prev);
+ copy_mat(force_vir, state->fvir_prev);
+ if (inputrecNvtTrotter(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);
+ enerd->term[F_EKIN] = trace(ekind->ekin);
+ }
+ }
+ else if (bExchanged)
+ {
+ wallcycle_stop(wcycle, ewcUPDATE);
+ /* 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, mdatoms, nrnb, vcm,
+ wcycle, enerd, NULL, NULL, NULL, NULL, mu_tot,
+ constr, NULL, FALSE, state->box,
+ NULL, &bSumEkinhOld,
+ CGLO_GSTAT | CGLO_TEMPERATURE);
+ wallcycle_start(wcycle, ewcUPDATE);
+ }
+ }
+ /* if it's the initial step, we performed this first step just to get the constraint virial */
+ if (ir->eI == eiVV && bInitStep)
+ {
+ copy_rvecn(vbuf, state->v, 0, state->natoms);
+ sfree(vbuf);
+ }
+ wallcycle_stop(wcycle, ewcUPDATE);
+ }
+
+ /* compute the conserved quantity */
+ if (EI_VV(ir->eI))
+ {
+ saved_conserved_quantity = compute_conserved_from_auxiliary(ir, state, &MassQ);
+ if (ir->eI == eiVV)
+ {
+ last_ekin = enerd->term[F_EKIN];
+ }
+ if ((ir->eDispCorr != edispcEnerPres) && (ir->eDispCorr != edispcAllEnerPres))
+ {
+ saved_conserved_quantity -= enerd->term[F_DISPCORR];
+ }
+ /* sum up the foreign energy and dhdl terms for vv. currently done every step so that dhdl is correct in the .edr */
+ if (ir->efep != efepNO && !bRerunMD)
+ {
+ sum_dhdl(enerd, state->lambda, ir->fepvals);
+ }
+ }
+
+ /* ######## END FIRST UPDATE STEP ############## */
+ /* ######## If doing VV, we now have v(dt) ###### */
+ if (bDoExpanded)
+ {
+ /* perform extended ensemble sampling in lambda - we don't
+ actually move to the new state before outputting
+ statistics, but if performing simulated tempering, we
+ do update the velocities and the tau_t. */
+
+ lamnew = ExpandedEnsembleDynamics(fplog, ir, enerd, state, &MassQ, state->fep_state, &state->dfhist, step, state->v, mdatoms);
+ /* history is maintained in state->dfhist, but state_global is what is sent to trajectory and log output */
+ copy_df_history(&state_global->dfhist, &state->dfhist);
+ }
+
+ /* Now we have the energies and forces corresponding to the
+ * coordinates at time t. We must output all of this before
+ * the update.
+ */
+ do_md_trajectory_writing(fplog, cr, nfile, fnm, step, step_rel, t,
+ ir, state, state_global, top_global, fr,
+ outf, mdebin, ekind, f,
+ &nchkpt,
+ bCPT, bRerunMD, bLastStep, (Flags & MD_CONFOUT),
+ bSumEkinhOld);
+ /* Check if IMD step and do IMD communication, if bIMD is TRUE. */
+ bIMDstep = do_IMD(ir->bIMD, step, cr, bNS, state->box, state->x, ir, t, wcycle);
+
+ /* kludge -- virial is lost with restart for MTTK NPT control. Must reload (saved earlier). */
+ if (startingFromCheckpoint && bTrotter)
+ {
+ copy_mat(state->svir_prev, shake_vir);
+ copy_mat(state->fvir_prev, force_vir);
+ }
+
+ elapsed_time = walltime_accounting_get_current_elapsed_time(walltime_accounting);
+
+ /* Check whether everything is still allright */
+ if (((int)gmx_get_stop_condition() > handled_stop_condition)
+#if GMX_THREAD_MPI
+ && MASTER(cr)
+#endif
+ )
+ {
+ int nsteps_stop = -1;
+
+ /* 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;
+ nsteps_stop = std::max(ir->nstlist, 2*nstglobalcomm);
+ }
+ if (gmx_get_stop_condition() == gmx_stop_cond_next)
+ {
+ gs.sig[eglsSTOPCOND] = -1;
+ nsteps_stop = nstglobalcomm + 1;
+ }
+ if (fplog)
+ {
+ fprintf(fplog,
+ "\n\nReceived the %s signal, stopping within %d steps\n\n",
+ gmx_get_signal_name(), nsteps_stop);
+ fflush(fplog);
+ }
+ fprintf(stderr,
+ "\n\nReceived the %s signal, stopping within %d steps\n\n",
+ gmx_get_signal_name(), nsteps_stop);
+ fflush(stderr);
+ handled_stop_condition = (int)gmx_get_stop_condition();
+ }
+ else if (MASTER(cr) && (bNS || ir->nstlist <= 0) &&
+ (max_hours > 0 && elapsed_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) &&
+ elapsed_time > max_hours*60.0*60.0*0.495)
+ {
+ /* Set flag that will communicate the signal to all ranks in the simulation */
+ gs.sig[eglsRESETCOUNTERS] = 1;
+ }
+
+ /* 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 ||
+ elapsed_time >= nchkpt*cpt_period*60.0)) &&
+ gs.set[eglsCHKPT] == 0)
+ {
+ gs.sig[eglsCHKPT] = 1;
+ }
+
+ /* ######### START SECOND UPDATE STEP ################# */
+
+ /* at the start of step, randomize or scale the velocities ((if vv. Restriction of Andersen controlled
+ in preprocessing */
+
+ if (ETC_ANDERSEN(ir->etc)) /* keep this outside of update_tcouple because of the extra info required to pass */
+ {
+ gmx_bool bIfRandomize;
+ bIfRandomize = update_randomize_velocities(ir, step, cr, mdatoms, state, upd, constr);
+ /* if we have constraints, we have to remove the kinetic energy parallel to the bonds */
+ if (constr && bIfRandomize)
+ {
+ update_constraints(fplog, step, NULL, ir, mdatoms,
+ state, fr->bMolPBC, graph, f,
+ &top->idef, tmp_vir,
+ cr, nrnb, wcycle, upd, constr,
+ TRUE, bCalcVir);
+ }
+ }
+ /* 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);
+
+ dvdl_constr = 0;
+
+ if (!bRerunMD || rerun_fr.bV || bForceUpdate)
+ {
+ wallcycle_start(wcycle, ewcUPDATE);
+ /* UPDATE PRESSURE VARIABLES IN TROTTER FORMULATION WITH CONSTRAINTS */
+ if (bTrotter)
+ {
+ 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(step, ir, state, ekind, &MassQ, mdatoms);
+ update_pcouple(fplog, step, ir, state, pcoupl_mu, M, bInitStep);
+ }
+
+ if (EI_VV(ir->eI))
+ {
+ /* velocity half-step update */
+ update_coords(fplog, step, ir, mdatoms, state, f, fcd,
+ ekind, M, upd, etrtVELOCITY2,
+ cr, constr);
+ }
+
+ /* Above, initialize just copies ekinh into ekin,
+ * it doesn't copy position (for VV),
+ * and entire integrator for MD.
+ */
+
+ if (ir->eI == eiVVAK)
+ {
+ /* We probably only need md->homenr, not state->natoms */
+ if (state->natoms > cbuf_nalloc)
+ {
+ cbuf_nalloc = state->natoms;
+ srenew(cbuf, cbuf_nalloc);
+ }
+ copy_rvecn(state->x, cbuf, 0, state->natoms);
+ }
+
+ update_coords(fplog, step, ir, mdatoms, state, f, fcd,
+ ekind, M, upd, etrtPOSITION, cr, constr);
+ wallcycle_stop(wcycle, ewcUPDATE);
+
+ update_constraints(fplog, step, &dvdl_constr, ir, mdatoms, state,
+ fr->bMolPBC, graph, f,
+ &top->idef, shake_vir,
+ cr, nrnb, wcycle, upd, constr,
+ FALSE, bCalcVir);
+
+ 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, mdatoms, nrnb, vcm,
+ wcycle, enerd, force_vir, shake_vir, total_vir, pres, mu_tot,
+ constr, NULL, FALSE, lastbox,
+ NULL, &bSumEkinhOld,
+ (bGStat ? CGLO_GSTAT : 0) | 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, fcd,
+ ekind, M, upd, etrtPOSITION, cr, constr);
+ 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, NULL, ir, mdatoms,
+ state, fr->bMolPBC, graph, f,
+ &top->idef, tmp_vir,
+ cr, nrnb, wcycle, upd, NULL,
+ FALSE, bCalcVir);
+ }
+ if (EI_VV(ir->eI))
+ {
+ /* this factor or 2 correction is necessary
+ because half of the constraint force is removed
+ in the vv step, so we have to double it. See
+ the Redmine issue #1255. It is not yet clear
+ if the factor of 2 is exact, or just a very
+ good approximation, and this will be
+ investigated. The next step is to see if this
+ can be done adding a dhdl contribution from the
+ rattle step, but this is somewhat more
+ complicated with the current code. Will be
+ investigated, hopefully for 4.6.3. However,
+ this current solution is much better than
+ having it completely wrong.
+ */
+ enerd->term[F_DVDL_CONSTR] += 2*dvdl_constr;
+ }
+ else
+ {
+ enerd->term[F_DVDL_CONSTR] += dvdl_constr;
+ }
+ }
+ 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(vsite, state->x, ir->delta_t, state->v,
+ top->idef.iparams, top->idef.il,
+ fr->ePBC, fr->bMolPBC, cr, state->box);
+
+ if (graph != NULL)
+ {
+ unshift_self(graph, state->box, state->x);
+ }
+ wallcycle_stop(wcycle, ewcVSITECONSTR);
+ }
+
+ /* ############## IF NOT VV, Calculate globals HERE ############ */
+ /* With Leap-Frog we can skip compute_globals at
+ * non-communication steps, but we need to calculate
+ * the kinetic energy one step before communication.
+ */
+ if (bGStat || (!EI_VV(ir->eI) && do_per_step(step+1, nstglobalcomm)))
+ {
+ compute_globals(fplog, gstat, cr, ir, fr, ekind, state, mdatoms, nrnb, vcm,
+ wcycle, enerd, force_vir, shake_vir, total_vir, pres, mu_tot,
+ constr, &gs,
+ (step_rel % gs.nstms == 0) &&
+ (multisim_nsteps < 0 || (step_rel < multisim_nsteps)),
+ lastbox,
+ &totalNumberOfBondedInteractions, &bSumEkinhOld,
+ (bGStat ? CGLO_GSTAT : 0)
+ | (!EI_VV(ir->eI) || bRerunMD ? CGLO_ENERGY : 0)
+ | (!EI_VV(ir->eI) && bStopCM ? CGLO_STOPCM : 0)
+ | (!EI_VV(ir->eI) ? CGLO_TEMPERATURE : 0)
+ | (!EI_VV(ir->eI) || bRerunMD ? CGLO_PRESSURE : 0)
+ | CGLO_CONSTRAINT
+ | (shouldCheckNumberOfBondedInteractions ? CGLO_CHECK_NUMBER_OF_BONDED_INTERACTIONS : 0)
+ );
+ checkNumberOfBondedInteractions(fplog, cr, totalNumberOfBondedInteractions,
+ top_global, top, state,
+ &shouldCheckNumberOfBondedInteractions);
+ }
+
+ /* ############# 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 (ir->efep != efepNO && (!EI_VV(ir->eI) || bRerunMD))
+ {
+ /* Sum up the foreign energy and dhdl terms for md and sd.
+ Currently done every step so that dhdl is correct in the .edr */
+ sum_dhdl(enerd, state->lambda, ir->fepvals);
+ }
+ update_box(fplog, step, ir, mdatoms, state, f,
+ pcoupl_mu, nrnb, upd);
+
+ /* ################# END UPDATE STEP 2 ################# */
+ /* #### We now have r(t+dt) and v(t+dt/2) ############# */
+
+ /* The coordinates (x) were unshifted in update */
+ if (!bGStat)
+ {
+ /* We will not sum ekinh_old,
+ * so signal that we still have to do it.
+ */
+ bSumEkinhOld = TRUE;
+ }
+
+ /* ######### BEGIN PREPARING EDR OUTPUT ########### */
+
+ /* 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 (EI_VV(ir->eI))
+ {
+ 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);
+ }
+ /* ######### END PREPARING EDR OUTPUT ########### */
+
+ /* Output stuff */
+ if (MASTER(cr))
+ {
+ if (fplog && do_log && bDoExpanded)
+ {
+ /* only needed if doing expanded ensemble */
+ PrintFreeEnergyInfoToFile(fplog, ir->fepvals, ir->expandedvals, ir->bSimTemp ? ir->simtempvals : NULL,
+ &state_global->dfhist, state->fep_state, ir->nstlog, step);
+ }
+ if (bCalcEner)
+ {
+ upd_mdebin(mdebin, bDoDHDL, bCalcEnerStep,
+ t, mdatoms->tmass, enerd, state,
+ ir->fepvals, ir->expandedvals, lastbox,
+ shake_vir, force_vir, total_vir, pres,
+ ekind, mu_tot, constr);
+ }
+ else
+ {
+ upd_mdebin_step(mdebin);
+ }
+
+ gmx_bool do_dr = do_per_step(step, ir->nstdisreout);
+ gmx_bool do_or = do_per_step(step, ir->nstorireout);
+
+ print_ebin(mdoutf_get_fp_ene(outf), do_ene, do_dr, do_or, do_log ? fplog : NULL,
+ step, t,
+ eprNORMAL, mdebin, fcd, groups, &(ir->opts));
+
+ if (ir->bPull)
+ {
+ pull_print_output(ir->pull_work, step, t);
+ }
+
+ if (do_per_step(step, ir->nstlog))
+ {
+ if (fflush(fplog) != 0)
+ {
+ gmx_fatal(FARGS, "Cannot flush logfile - maybe you are out of disk space?");
+ }
+ }
+ }
+ if (bDoExpanded)
+ {
+ /* Have to do this part _after_ outputting the logfile and the edr file */
+ /* Gets written into the state at the beginning of next loop*/
+ state->fep_state = lamnew;
+ }
+ /* Print the remaining wall clock time for the run */
+ if (MULTIMASTER(cr) &&
+ (do_verbose || gmx_got_usr_signal()) &&
+ !bPMETunePrinting)
+ {
+ if (shellfc)
+ {
+ fprintf(stderr, "\n");
+ }
+ print_time(stderr, walltime_accounting, step, ir, cr);
+ }
+
+ /* Ion/water position swapping.
+ * Not done in last step since trajectory writing happens before this call
+ * in the MD loop and exchanges would be lost anyway. */
+ bNeedRepartition = FALSE;
+ if ((ir->eSwapCoords != eswapNO) && (step > 0) && !bLastStep &&
+ do_per_step(step, ir->swap->nstswap))
+ {
+ bNeedRepartition = do_swapcoords(cr, step, t, ir, wcycle,
+ bRerunMD ? rerun_fr.x : state->x,
+ bRerunMD ? rerun_fr.box : state->box,
+ top_global, MASTER(cr) && bVerbose, bRerunMD);
+
+ if (bNeedRepartition && DOMAINDECOMP(cr))
+ {
+ dd_collect_state(cr->dd, state, state_global);
+ }
+ }
+
+ /* Replica exchange */
+ bExchanged = FALSE;
+ if (bDoReplEx)
+ {
+ bExchanged = replica_exchange(fplog, cr, repl_ex,
+ state_global, enerd,
+ state, step, t);
+ }
+
+ if ( (bExchanged || bNeedRepartition) && DOMAINDECOMP(cr) )
+ {
+ dd_partition_system(fplog, step, cr, TRUE, 1,
+ state_global, top_global, ir,
+ state, &f, mdatoms, top, fr,
+ vsite, constr,
+ nrnb, wcycle, FALSE);
+ shouldCheckNumberOfBondedInteractions = true;
+ update_realloc(upd, state->nalloc);
+ }
+
+ bFirstStep = FALSE;
+ bInitStep = FALSE;
+ startingFromCheckpoint = 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 */
+ bLastStep = !read_next_frame(oenv, status, &rerun_fr);
+ }
+
+ if (PAR(cr))
+ {
+ rerun_parallel_comm(cr, &rerun_fr, &bLastStep);
+ }
+ }
+
+ cycles = wallcycle_stop(wcycle, ewcSTEP);
+ if (DOMAINDECOMP(cr) && wcycle)
+ {
+ dd_cycles_add(cr->dd, cycles, ddCyclStep);
+ }
+
+ if (!bRerunMD || !rerun_fr.bStep)
+ {
+ /* increase the MD step number */
+ step++;
+ step_rel++;
+ }
+
+ /* TODO make a counter-reset module */
+ /* If it is time to reset counters, set a flag that remains
+ true until counters actually get reset */
+ if (step_rel == wcycle_get_reset_counters(wcycle) ||
+ gs.set[eglsRESETCOUNTERS] != 0)
+ {
+ if (pme_loadbal_is_active(pme_loadbal))
+ {
+ /* Do not permit counter reset while PME load
+ * balancing is active. The only purpose for resetting
+ * counters is to measure reliable performance data,
+ * and that can't be done before balancing
+ * completes.
+ *
+ * TODO consider fixing this by delaying the reset
+ * until after load balancing completes,
+ * e.g. https://gerrit.gromacs.org/#/c/4964/2 */
+ gmx_fatal(FARGS, "PME tuning was still active when attempting to "
+ "reset mdrun counters at step %" GMX_PRId64 ". Try "
+ "resetting counters later in the run, e.g. with gmx "
+ "mdrun -resetstep.", step);
+ }
+ reset_all_counters(fplog, cr, step, &step_rel, ir, wcycle, nrnb, walltime_accounting,
+ use_GPU(fr->nbv) ? fr->nbv : NULL);
+ wcycle_set_reset_counters(wcycle, -1);
+ if (!(cr->duty & DUTY_PME))
+ {
+ /* Tell our PME node to reset its counters */
+ gmx_pme_send_resetcounters(cr, step);
+ }
+ /* Correct max_hours for the elapsed time */
+ max_hours -= elapsed_time/(60.0*60.0);
+ /* If mdrun -maxh -resethway was active, it can only trigger once */
+ bResetCountersHalfMaxH = FALSE; /* TODO move this to where gs.sig[eglsRESETCOUNTERS] is set */
+ /* Reset can only happen once, so clear the triggering flag. */
+ gs.set[eglsRESETCOUNTERS] = 0;
+ }
+
+ /* If bIMD is TRUE, the master updates the IMD energy record and sends positions to VMD client */
+ IMD_prep_energies_send_positions(ir->bIMD && MASTER(cr), bIMDstep, ir->imd, enerd, step, bCalcEner, wcycle);
+
+ }
+ /* End of main MD loop */
+
+ /* Closing TNG files can include compressing data. Therefore it is good to do that
+ * before stopping the time measurements. */
+ mdoutf_tng_close(outf);
+
+ /* Stop measuring walltime */
+ walltime_accounting_end(walltime_accounting);
+
+ if (bRerunMD && MASTER(cr))
+ {
+ close_trj(status);
+ }
+
+ if (!(cr->duty & DUTY_PME))
+ {
+ /* Tell the PME only node to finish */
+ gmx_pme_send_finish(cr);
+ }
+
+ if (MASTER(cr))
+ {
+ if (ir->nstcalcenergy > 0 && !bRerunMD)
+ {
+ print_ebin(mdoutf_get_fp_ene(outf), FALSE, FALSE, FALSE, fplog, step, t,
+ eprAVER, mdebin, fcd, groups, &(ir->opts));
+ }
+ }
+
+ done_mdoutf(outf);
+
+ if (bPMETune)
+ {
+ pme_loadbal_done(pme_loadbal, cr, fplog, use_GPU(fr->nbv));
+ }
+
+ done_shellfc(fplog, shellfc, step_rel);
+
+ if (repl_ex_nst > 0 && MASTER(cr))
+ {
+ print_replica_exchange_statistics(fplog, repl_ex);
+ }
+
+ // Clean up swapcoords
+ if (ir->eSwapCoords != eswapNO)
+ {
+ finish_swapcoords(ir->swap);
+ }
+
+ if (membed != nullptr)
+ {
+ free_membed(membed);
+ }
+
+ /* IMD cleanup, if bIMD is TRUE. */
+ IMD_finalize(ir->bIMD, ir->imd);
+
+ walltime_accounting_set_nsteps_done(walltime_accounting, step_rel);
+
+ return 0;
+}
diff --git a/patches/gromacs-2016-beta1.diff/src/programs/mdrun/mdrun.cpp b/patches/gromacs-2016-beta1.diff/src/programs/mdrun/mdrun.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..55f3529faaba83f8ee5e66744004b578c9bce6d5
--- /dev/null
+++ b/patches/gromacs-2016-beta1.diff/src/programs/mdrun/mdrun.cpp
@@ -0,0 +1,577 @@
+/*
+ * This file is part of the GROMACS molecular simulation package.
+ *
+ * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
+ * Copyright (c) 2001-2004, The GROMACS development team.
+ * Copyright (c) 2011,2012,2013,2014,2015,2016, by the GROMACS development team, led by
+ * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
+ * and including many others, as listed in the AUTHORS file in the
+ * top-level source directory and at http://www.gromacs.org.
+ *
+ * GROMACS is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public License
+ * as published by the Free Software Foundation; either version 2.1
+ * of the License, or (at your option) any later version.
+ *
+ * GROMACS is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with GROMACS; if not, see
+ * http://www.gnu.org/licenses, or write to the Free Software Foundation,
+ * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ * If you want to redistribute modifications to GROMACS, 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 http://www.gromacs.org.
+ *
+ * To help us fund GROMACS development, we humbly ask that you cite
+ * the research papers on the package. Check out http://www.gromacs.org.
+ */
+/*! \defgroup module_mdrun Implementation of mdrun
+ * \ingroup group_mdrun
+ *
+ * \brief This module contains code that implements mdrun.
+ */
+/*! \internal \file
+ *
+ * \brief This file implements mdrun
+ *
+ * \author Berk Hess <hess@kth.se>
+ * \author David van der Spoel <david.vanderspoel@icm.uu.se>
+ * \author Erik Lindahl <erik@kth.se>
+ * \author Mark Abraham <mark.j.abraham@gmail.com>
+ *
+ * \ingroup module_mdrun
+ */
+#include "gmxpre.h"
+
+#include "config.h"
+
+#include <stdio.h>
+#include <string.h>
+
+#include "gromacs/commandline/filenm.h"
+#include "gromacs/commandline/pargs.h"
+#include "gromacs/fileio/readinp.h"
+#include "gromacs/gmxlib/network.h"
+#include "gromacs/mdlib/main.h"
+#include "gromacs/mdlib/mdrun.h"
+#include "gromacs/mdrunutility/handlerestart.h"
+#include "gromacs/mdtypes/commrec.h"
+#include "gromacs/utility/arraysize.h"
+#include "gromacs/utility/fatalerror.h"
+
+#include "mdrun_main.h"
+#include "runner.h"
+
+/* PLUMED */
+#include "../../../Plumed.h"
+extern int plumedswitch;
+extern plumed plumedmain;
+extern void(*plumedcmd)(plumed,const char*,const void*);
+/* END PLUMED */
+
+/*! \brief Return whether either of the command-line parameters that
+ * will trigger a multi-simulation is set */
+static bool is_multisim_option_set(int argc, const char *const argv[])
+{
+ for (int i = 0; i < argc; ++i)
+ {
+ if (strcmp(argv[i], "-multi") == 0 || strcmp(argv[i], "-multidir") == 0)
+ {
+ return true;
+ }
+ }
+ return false;
+}
+
+//! Implements C-style main function for mdrun
+int gmx_mdrun(int argc, char *argv[])
+{
+ const char *desc[] = {
+ "[THISMODULE] 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 [gmx-nmeig].[PAR]",
+ "The [TT]mdrun[tt] program reads the run input file ([TT]-s[tt])",
+ "and distributes the topology over ranks if needed.",
+ "[TT]mdrun[tt] produces at least four output files.",
+ "A single log file ([TT]-g[tt]) is written.",
+ "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]",
+ "Running mdrun efficiently in parallel is a complex topic topic,",
+ "many aspects of which are covered in the online User Guide. You",
+ "should look there for practical advice on using many of the options",
+ "available in mdrun.[PAR]",
+ "ED (essential dynamics) sampling and/or additional flooding potentials",
+ "are switched on by using the [TT]-ei[tt] flag followed by an [REF].edi[ref]",
+ "file. The [REF].edi[ref] file can be produced with the [TT]make_edi[tt] tool",
+ "or by using options in the essdyn menu of the WHAT IF program.",
+ "[TT]mdrun[tt] produces a [REF].xvg[ref] output file that",
+ "contains projections of positions, velocities and forces onto selected",
+ "eigenvectors.[PAR]",
+ "When user-defined potential functions have been selected in the",
+ "[REF].mdp[ref] 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 [REF].mdp[ref] file.[PAR]",
+ "Finally some experimental algorithms can be tested when the",
+ "appropriate options have been given. Currently under",
+ "investigation are: polarizability.",
+ "[PAR]",
+ "The option [TT]-membed[tt] does what used to be g_membed, i.e. embed",
+ "a protein into a membrane. This module requires a number of settings",
+ "that are provided in a data file that is the argument of this option.",
+ "For more details in membrane embedding, see the documentation in the",
+ "user guide. The options [TT]-mn[tt] and [TT]-mp[tt] are used to provide",
+ "the index and topology files used for the embedding.",
+ "[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 [REF].tpr[ref] 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 ranks 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. This option is particularly useful in",
+ "combination with setting [TT]nsteps[tt] to -1 either in the mdp or using the",
+ "similarly named command line option. This results in an infinite run,",
+ "terminated only when the time limit set by [TT]-maxh[tt] is reached (if any)"
+ "or upon receiving a signal."
+ "[PAR]",
+ "When [TT]mdrun[tt] receives a TERM signal, it will stop as soon as",
+ "checkpoint file can be written, i.e. after the next global communication step.",
+ "When [TT]mdrun[tt] receives an INT signal (e.g. when ctrl+C is",
+ "pressed), it will stop at the next neighbor search step or at the",
+ "second global communication step, whichever happens later.",
+ "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] ranks",
+ "is sufficient, this signal should not be sent to mpirun or",
+ "the [TT]mdrun[tt] process that is the parent of the others.",
+ "[PAR]",
+ "Interactive molecular dynamics (IMD) can be activated by using at least one",
+ "of the three IMD switches: The [TT]-imdterm[tt] switch allows one to terminate",
+ "the simulation from the molecular viewer (e.g. VMD). With [TT]-imdwait[tt],",
+ "[TT]mdrun[tt] pauses whenever no IMD client is connected. Pulling from the",
+ "IMD remote can be turned on by [TT]-imdpull[tt].",
+ "The port [TT]mdrun[tt] listens to can be altered by [TT]-imdport[tt].The",
+ "file pointed to by [TT]-if[tt] contains atom indices and forces if IMD",
+ "pulling is used."
+ "[PAR]",
+ "When [TT]mdrun[tt] is started with MPI, it does not run niced by default."
+ };
+ t_commrec *cr;
+ t_filenm fnm[] = {
+ { efTPR, NULL, NULL, ffREAD },
+ { efTRN, "-o", NULL, ffWRITE },
+ { efCOMPRESSED, "-x", NULL, ffOPTWR },
+ { efCPT, "-cpi", NULL, ffOPTRD | ffALLOW_MISSING },
+ { 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, "-tablep", "tablep", ffOPTRD },
+ { efXVG, "-tableb", "table", ffOPTRD },
+ { efTRX, "-rerun", "rerun", ffOPTRD },
+ { efXVG, "-tpi", "tpi", ffOPTWR },
+ { efXVG, "-tpid", "tpidist", ffOPTWR },
+ { efEDI, "-ei", "sam", ffOPTRD },
+ { efXVG, "-eo", "edsam", 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 },
+ { efRND, "-multidir", NULL, ffOPTRDMULT},
+ { efDAT, "-plumed", "plumed", ffOPTRD }, /* PLUMED */
+ { efDAT, "-membed", "membed", ffOPTRD },
+ { efTOP, "-mp", "membed", ffOPTRD },
+ { efNDX, "-mn", "membed", ffOPTRD },
+ { efXVG, "-if", "imdforces", ffOPTWR },
+ { efXVG, "-swap", "swapions", ffOPTWR }
+ };
+ const int NFILE = asize(fnm);
+
+ /* Command line options ! */
+ gmx_bool bDDBondCheck = TRUE;
+ gmx_bool bDDBondComm = TRUE;
+ gmx_bool bTunePME = TRUE;
+ gmx_bool bVerbose = FALSE;
+ gmx_bool bRerunVSite = FALSE;
+ gmx_bool bConfout = TRUE;
+ gmx_bool bReproducible = FALSE;
+ gmx_bool bIMDwait = FALSE;
+ gmx_bool bIMDterm = FALSE;
+ gmx_bool bIMDpull = FALSE;
+
+ int npme = -1;
+ int nstlist = 0;
+ int nmultisim = 0;
+ int nstglobalcomm = -1;
+ int repl_ex_nst = 0;
+ int repl_ex_seed = -1;
+ int repl_ex_nex = 0;
+ int nstepout = 100;
+ int resetstep = -1;
+ gmx_int64_t nsteps = -2; /* the value -2 means that the mdp option will be used */
+ int imdport = 8888; /* can be almost anything, 8888 is easy to remember */
+
+ rvec realddxyz = {0, 0, 0};
+ const char *ddrank_opt[ddrankorderNR+1] =
+ { NULL, "interleave", "pp_pme", "cartesian", NULL };
+ const char *dddlb_opt[] =
+ { NULL, "auto", "no", "yes", NULL };
+ const char *thread_aff_opt[threadaffNR+1] =
+ { NULL, "auto", "on", "off", NULL };
+ const char *nbpu_opt[] =
+ { NULL, "auto", "cpu", "gpu", "gpu_cpu", 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 bTryToAppendFiles = TRUE;
+ gmx_bool bKeepAndNumCPT = FALSE;
+ gmx_bool bResetCountersHalfWay = FALSE;
+ gmx_output_env_t *oenv = NULL;
+
+ /* Non transparent initialization of a complex gmx_hw_opt_t struct.
+ * But unfortunately we are not allowed to call a function here,
+ * since declarations follow below.
+ */
+ gmx_hw_opt_t hw_opt = {
+ 0, 0, 0, 0, threadaffSEL, 0, 0,
+ { NULL, FALSE, 0, NULL }
+ };
+
+ t_pargs pa[] = {
+
+ { "-dd", FALSE, etRVEC, {&realddxyz},
+ "Domain decomposition grid, 0 is optimize" },
+ { "-ddorder", FALSE, etENUM, {ddrank_opt},
+ "DD rank order" },
+ { "-npme", FALSE, etINT, {&npme},
+ "Number of separate ranks to be used for PME, -1 is guess" },
+ { "-nt", FALSE, etINT, {&hw_opt.nthreads_tot},
+ "Total number of threads to start (0 is guess)" },
+ { "-ntmpi", FALSE, etINT, {&hw_opt.nthreads_tmpi},
+ "Number of thread-MPI threads to start (0 is guess)" },
+ { "-ntomp", FALSE, etINT, {&hw_opt.nthreads_omp},
+ "Number of OpenMP threads per MPI rank to start (0 is guess)" },
+ { "-ntomp_pme", FALSE, etINT, {&hw_opt.nthreads_omp_pme},
+ "Number of OpenMP threads per MPI rank to start (0 is -ntomp)" },
+ { "-pin", FALSE, etENUM, {thread_aff_opt},
+ "Whether mdrun should try to set thread affinities" },
+ { "-pinoffset", FALSE, etINT, {&hw_opt.core_pinning_offset},
+ "The lowest logical core number to which mdrun should pin the first thread" },
+ { "-pinstride", FALSE, etINT, {&hw_opt.core_pinning_stride},
+ "Pinning distance in logical cores for threads, use 0 to minimize the number of threads per physical core" },
+ { "-gpu_id", FALSE, etSTR, {&hw_opt.gpu_opt.gpu_id},
+ "List of GPU device id-s to use, specifies the per-node PP rank to GPU mapping" },
+ { "-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},
+ "Fraction in (0,1) by whose reciprocal the initial DD cell size will be increased in order to "
+ "provide a margin in which dynamic load balancing can act while preserving the minimum cell size." },
+ { "-ddcsx", FALSE, etSTR, {&ddcsx},
+ "HIDDENA string containing a vector of the relative sizes in the x "
+ "direction of the corresponding DD cells. Only effective with static "
+ "load balancing." },
+ { "-ddcsy", FALSE, etSTR, {&ddcsy},
+ "HIDDENA string containing a vector of the relative sizes in the y "
+ "direction of the corresponding DD cells. Only effective with static "
+ "load balancing." },
+ { "-ddcsz", FALSE, etSTR, {&ddcsz},
+ "HIDDENA string containing a vector of the relative sizes in the z "
+ "direction of the corresponding DD cells. Only effective with static "
+ "load balancing." },
+ { "-gcom", FALSE, etINT, {&nstglobalcomm},
+ "Global communication frequency" },
+ { "-nb", FALSE, etENUM, {&nbpu_opt},
+ "Calculate non-bonded interactions on" },
+ { "-nstlist", FALSE, etINT, {&nstlist},
+ "Set nstlist when using a Verlet buffer tolerance (0 is guess)" },
+ { "-tunepme", FALSE, etBOOL, {&bTunePME},
+ "Optimize PME load between PP/PME ranks or GPU/CPU" },
+ { "-v", FALSE, etBOOL, {&bVerbose},
+ "Be loud and noisy" },
+ { "-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, {&bTryToAppendFiles},
+ "Append to previous output files when continuing from checkpoint instead of adding the simulation part number to all file names" },
+ { "-nsteps", FALSE, etINT64, {&nsteps},
+ "Run this number of steps, overrides .mdp file option (-1 means infinite, -2 means use mdp option, smaller is invalid)" },
+ { "-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 periodically with this period (steps)" },
+ { "-nex", FALSE, etINT, {&repl_ex_nex},
+ "Number of random exchanges to carry out each exchange interval (N^3 is one suggestion). -nex zero or not specified gives neighbor replica exchange." },
+ { "-reseed", FALSE, etINT, {&repl_ex_seed},
+ "Seed for replica exchange, -1 is generate a seed" },
+ { "-imdport", FALSE, etINT, {&imdport},
+ "HIDDENIMD listening port" },
+ { "-imdwait", FALSE, etBOOL, {&bIMDwait},
+ "HIDDENPause the simulation while no IMD client is connected" },
+ { "-imdterm", FALSE, etBOOL, {&bIMDterm},
+ "HIDDENAllow termination of the simulation from IMD client" },
+ { "-imdpull", FALSE, etBOOL, {&bIMDpull},
+ "HIDDENAllow pulling in the simulation from IMD client" },
+ { "-rerunvsite", FALSE, etBOOL, {&bRerunVSite},
+ "HIDDENRecalculate virtual site coordinates with [TT]-rerun[tt]" },
+ { "-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 wall clock time for the run" },
+ { "-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]" }
+ };
+ unsigned long Flags;
+ ivec ddxyz;
+ int dd_rank_order;
+ gmx_bool bDoAppendFiles, bStartFromCpt;
+ FILE *fplog;
+ int rc;
+ char **multidir = NULL;
+
+ cr = init_commrec();
+
+ unsigned long PCA_Flags = PCA_CAN_SET_DEFFNM;
+ // With -multi or -multidir, the file names are going to get processed
+ // further (or the working directory changed), so we can't check for their
+ // existence during parsing. It isn't useful to do any completion based on
+ // file system contents, either.
+ if (is_multisim_option_set(argc, argv))
+ {
+ PCA_Flags |= PCA_DISABLE_INPUT_FILE_CHECKING;
+ }
+
+ /* 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;
+ }
+ */
+
+ if (!parse_common_args(&argc, argv, PCA_Flags, NFILE, fnm, asize(pa), pa,
+ asize(desc), desc, 0, NULL, &oenv))
+ {
+ return 0;
+ }
+
+
+ dd_rank_order = nenum(ddrank_opt);
+
+ hw_opt.thread_affinity = nenum(thread_aff_opt);
+
+ /* now check the -multi and -multidir option */
+ if (opt2bSet("-multidir", NFILE, fnm))
+ {
+ 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 (repl_ex_nex < 0)
+ {
+ gmx_fatal(FARGS, "Replica exchange number of exchanges needs to be positive");
+ }
+
+ if (nmultisim >= 1)
+ {
+#if !GMX_THREAD_MPI
+ gmx_bool bParFn = (multidir == NULL);
+ init_multisystem(cr, nmultisim, multidir, NFILE, fnm, bParFn);
+#else
+ gmx_fatal(FARGS, "mdrun -multi or -multidir are not supported with the thread-MPI library. "
+ "Please compile GROMACS with a proper external MPI library.");
+#endif
+ }
+
+ handleRestart(cr, bTryToAppendFiles, NFILE, fnm,
+ &bDoAppendFiles, &bStartFromCpt);
+
+ Flags = opt2bSet("-rerun", NFILE, fnm) ? MD_RERUN : 0;
+ Flags = Flags | (bDDBondCheck ? MD_DDBONDCHECK : 0);
+ Flags = Flags | (bDDBondComm ? MD_DDBONDCOMM : 0);
+ Flags = Flags | (bTunePME ? MD_TUNEPME : 0);
+ Flags = Flags | (bConfout ? MD_CONFOUT : 0);
+ Flags = Flags | (bRerunVSite ? MD_RERUN_VSITE : 0);
+ Flags = Flags | (bReproducible ? MD_REPRODUCIBLE : 0);
+ Flags = Flags | (bDoAppendFiles ? MD_APPENDFILES : 0);
+ Flags = Flags | (opt2parg_bSet("-append", asize(pa), pa) ? MD_APPENDFILESSET : 0);
+ Flags = Flags | (bKeepAndNumCPT ? MD_KEEPANDNUMCPT : 0);
+ Flags = Flags | (bStartFromCpt ? MD_STARTFROMCPT : 0);
+ Flags = Flags | (bResetCountersHalfWay ? MD_RESETCOUNTERSHALFWAY : 0);
+ Flags = Flags | (opt2parg_bSet("-ntomp", asize(pa), pa) ? MD_NTOMPSET : 0);
+ Flags = Flags | (bIMDwait ? MD_IMDWAIT : 0);
+ Flags = Flags | (bIMDterm ? MD_IMDTERM : 0);
+ Flags = Flags | (bIMDpull ? MD_IMDPULL : 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) && !bDoAppendFiles)
+ {
+ gmx_log_open(ftp2fn(efLOG, NFILE, fnm), cr,
+ Flags & MD_APPENDFILES, &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);
+
+ /* PLUMED */
+ plumedswitch=0;
+ if (opt2bSet("-plumed",NFILE,fnm)) plumedswitch=1;
+ if(plumedswitch){
+ plumedcmd=plumed_cmd;
+ int real_precision=sizeof(real);
+ real energyUnits=1.0;
+ real lengthUnits=1.0;
+ real timeUnits=1.0;
+
+ if(!plumed_installed()){
+ gmx_fatal(FARGS,"Plumed is not available. Check your PLUMED_KERNEL variable.");
+ }
+ plumedmain=plumed_create();
+ plumed_cmd(plumedmain,"setRealPrecision",&real_precision);
+ // this is not necessary for gromacs units:
+ plumed_cmd(plumedmain,"setMDEnergyUnits",&energyUnits);
+ plumed_cmd(plumedmain,"setMDLengthUnits",&lengthUnits);
+ plumed_cmd(plumedmain,"setMDTimeUnits",&timeUnits);
+ //
+ plumed_cmd(plumedmain,"setPlumedDat",ftp2fn(efDAT,NFILE,fnm));
+ plumedswitch=1;
+ }
+ /* END PLUMED */
+
+ rc = gmx::mdrunner(&hw_opt, fplog, cr, NFILE, fnm, oenv, bVerbose,
+ nstglobalcomm, ddxyz, dd_rank_order, npme, rdd, rconstr,
+ dddlb_opt[0], dlb_scale, ddcsx, ddcsy, ddcsz,
+ nbpu_opt[0], nstlist,
+ nsteps, nstepout, resetstep,
+ nmultisim, repl_ex_nst, repl_ex_nex, repl_ex_seed,
+ pforce, cpt_period, max_hours, imdport, Flags);
+
+ /* Log file has to be closed in mdrunner if we are appending to it
+ (fplog not set here) */
+ if (MASTER(cr) && !bDoAppendFiles)
+ {
+ gmx_log_close(fplog);
+ }
+
+ return rc;
+}
diff --git a/patches/gromacs-2016-beta1.diff/src/programs/mdrun/mdrun.cpp.preplumed b/patches/gromacs-2016-beta1.diff/src/programs/mdrun/mdrun.cpp.preplumed
new file mode 100644
index 0000000000000000000000000000000000000000..9458dfbf91803983e33b4ae66f8661057fb2d4db
--- /dev/null
+++ b/patches/gromacs-2016-beta1.diff/src/programs/mdrun/mdrun.cpp.preplumed
@@ -0,0 +1,544 @@
+/*
+ * This file is part of the GROMACS molecular simulation package.
+ *
+ * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
+ * Copyright (c) 2001-2004, The GROMACS development team.
+ * Copyright (c) 2011,2012,2013,2014,2015,2016, by the GROMACS development team, led by
+ * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
+ * and including many others, as listed in the AUTHORS file in the
+ * top-level source directory and at http://www.gromacs.org.
+ *
+ * GROMACS is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public License
+ * as published by the Free Software Foundation; either version 2.1
+ * of the License, or (at your option) any later version.
+ *
+ * GROMACS is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with GROMACS; if not, see
+ * http://www.gnu.org/licenses, or write to the Free Software Foundation,
+ * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ * If you want to redistribute modifications to GROMACS, 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 http://www.gromacs.org.
+ *
+ * To help us fund GROMACS development, we humbly ask that you cite
+ * the research papers on the package. Check out http://www.gromacs.org.
+ */
+/*! \defgroup module_mdrun Implementation of mdrun
+ * \ingroup group_mdrun
+ *
+ * \brief This module contains code that implements mdrun.
+ */
+/*! \internal \file
+ *
+ * \brief This file implements mdrun
+ *
+ * \author Berk Hess <hess@kth.se>
+ * \author David van der Spoel <david.vanderspoel@icm.uu.se>
+ * \author Erik Lindahl <erik@kth.se>
+ * \author Mark Abraham <mark.j.abraham@gmail.com>
+ *
+ * \ingroup module_mdrun
+ */
+#include "gmxpre.h"
+
+#include "config.h"
+
+#include <stdio.h>
+#include <string.h>
+
+#include "gromacs/commandline/filenm.h"
+#include "gromacs/commandline/pargs.h"
+#include "gromacs/fileio/readinp.h"
+#include "gromacs/gmxlib/network.h"
+#include "gromacs/mdlib/main.h"
+#include "gromacs/mdlib/mdrun.h"
+#include "gromacs/mdrunutility/handlerestart.h"
+#include "gromacs/mdtypes/commrec.h"
+#include "gromacs/utility/arraysize.h"
+#include "gromacs/utility/fatalerror.h"
+
+#include "mdrun_main.h"
+#include "runner.h"
+
+/*! \brief Return whether either of the command-line parameters that
+ * will trigger a multi-simulation is set */
+static bool is_multisim_option_set(int argc, const char *const argv[])
+{
+ for (int i = 0; i < argc; ++i)
+ {
+ if (strcmp(argv[i], "-multi") == 0 || strcmp(argv[i], "-multidir") == 0)
+ {
+ return true;
+ }
+ }
+ return false;
+}
+
+//! Implements C-style main function for mdrun
+int gmx_mdrun(int argc, char *argv[])
+{
+ const char *desc[] = {
+ "[THISMODULE] 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 [gmx-nmeig].[PAR]",
+ "The [TT]mdrun[tt] program reads the run input file ([TT]-s[tt])",
+ "and distributes the topology over ranks if needed.",
+ "[TT]mdrun[tt] produces at least four output files.",
+ "A single log file ([TT]-g[tt]) is written.",
+ "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]",
+ "Running mdrun efficiently in parallel is a complex topic topic,",
+ "many aspects of which are covered in the online User Guide. You",
+ "should look there for practical advice on using many of the options",
+ "available in mdrun.[PAR]",
+ "ED (essential dynamics) sampling and/or additional flooding potentials",
+ "are switched on by using the [TT]-ei[tt] flag followed by an [REF].edi[ref]",
+ "file. The [REF].edi[ref] file can be produced with the [TT]make_edi[tt] tool",
+ "or by using options in the essdyn menu of the WHAT IF program.",
+ "[TT]mdrun[tt] produces a [REF].xvg[ref] output file that",
+ "contains projections of positions, velocities and forces onto selected",
+ "eigenvectors.[PAR]",
+ "When user-defined potential functions have been selected in the",
+ "[REF].mdp[ref] 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 [REF].mdp[ref] file.[PAR]",
+ "Finally some experimental algorithms can be tested when the",
+ "appropriate options have been given. Currently under",
+ "investigation are: polarizability.",
+ "[PAR]",
+ "The option [TT]-membed[tt] does what used to be g_membed, i.e. embed",
+ "a protein into a membrane. This module requires a number of settings",
+ "that are provided in a data file that is the argument of this option.",
+ "For more details in membrane embedding, see the documentation in the",
+ "user guide. The options [TT]-mn[tt] and [TT]-mp[tt] are used to provide",
+ "the index and topology files used for the embedding.",
+ "[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 [REF].tpr[ref] 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 ranks 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. This option is particularly useful in",
+ "combination with setting [TT]nsteps[tt] to -1 either in the mdp or using the",
+ "similarly named command line option. This results in an infinite run,",
+ "terminated only when the time limit set by [TT]-maxh[tt] is reached (if any)"
+ "or upon receiving a signal."
+ "[PAR]",
+ "When [TT]mdrun[tt] receives a TERM signal, it will stop as soon as",
+ "checkpoint file can be written, i.e. after the next global communication step.",
+ "When [TT]mdrun[tt] receives an INT signal (e.g. when ctrl+C is",
+ "pressed), it will stop at the next neighbor search step or at the",
+ "second global communication step, whichever happens later.",
+ "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] ranks",
+ "is sufficient, this signal should not be sent to mpirun or",
+ "the [TT]mdrun[tt] process that is the parent of the others.",
+ "[PAR]",
+ "Interactive molecular dynamics (IMD) can be activated by using at least one",
+ "of the three IMD switches: The [TT]-imdterm[tt] switch allows one to terminate",
+ "the simulation from the molecular viewer (e.g. VMD). With [TT]-imdwait[tt],",
+ "[TT]mdrun[tt] pauses whenever no IMD client is connected. Pulling from the",
+ "IMD remote can be turned on by [TT]-imdpull[tt].",
+ "The port [TT]mdrun[tt] listens to can be altered by [TT]-imdport[tt].The",
+ "file pointed to by [TT]-if[tt] contains atom indices and forces if IMD",
+ "pulling is used."
+ "[PAR]",
+ "When [TT]mdrun[tt] is started with MPI, it does not run niced by default."
+ };
+ t_commrec *cr;
+ t_filenm fnm[] = {
+ { efTPR, NULL, NULL, ffREAD },
+ { efTRN, "-o", NULL, ffWRITE },
+ { efCOMPRESSED, "-x", NULL, ffOPTWR },
+ { efCPT, "-cpi", NULL, ffOPTRD | ffALLOW_MISSING },
+ { 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, "-tablep", "tablep", ffOPTRD },
+ { efXVG, "-tableb", "table", ffOPTRD },
+ { efTRX, "-rerun", "rerun", ffOPTRD },
+ { efXVG, "-tpi", "tpi", ffOPTWR },
+ { efXVG, "-tpid", "tpidist", ffOPTWR },
+ { efEDI, "-ei", "sam", ffOPTRD },
+ { efXVG, "-eo", "edsam", 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 },
+ { efRND, "-multidir", NULL, ffOPTRDMULT},
+ { efDAT, "-membed", "membed", ffOPTRD },
+ { efTOP, "-mp", "membed", ffOPTRD },
+ { efNDX, "-mn", "membed", ffOPTRD },
+ { efXVG, "-if", "imdforces", ffOPTWR },
+ { efXVG, "-swap", "swapions", ffOPTWR }
+ };
+ const int NFILE = asize(fnm);
+
+ /* Command line options ! */
+ gmx_bool bDDBondCheck = TRUE;
+ gmx_bool bDDBondComm = TRUE;
+ gmx_bool bTunePME = TRUE;
+ gmx_bool bVerbose = FALSE;
+ gmx_bool bRerunVSite = FALSE;
+ gmx_bool bConfout = TRUE;
+ gmx_bool bReproducible = FALSE;
+ gmx_bool bIMDwait = FALSE;
+ gmx_bool bIMDterm = FALSE;
+ gmx_bool bIMDpull = FALSE;
+
+ int npme = -1;
+ int nstlist = 0;
+ int nmultisim = 0;
+ int nstglobalcomm = -1;
+ int repl_ex_nst = 0;
+ int repl_ex_seed = -1;
+ int repl_ex_nex = 0;
+ int nstepout = 100;
+ int resetstep = -1;
+ gmx_int64_t nsteps = -2; /* the value -2 means that the mdp option will be used */
+ int imdport = 8888; /* can be almost anything, 8888 is easy to remember */
+
+ rvec realddxyz = {0, 0, 0};
+ const char *ddrank_opt[ddrankorderNR+1] =
+ { NULL, "interleave", "pp_pme", "cartesian", NULL };
+ const char *dddlb_opt[] =
+ { NULL, "auto", "no", "yes", NULL };
+ const char *thread_aff_opt[threadaffNR+1] =
+ { NULL, "auto", "on", "off", NULL };
+ const char *nbpu_opt[] =
+ { NULL, "auto", "cpu", "gpu", "gpu_cpu", 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 bTryToAppendFiles = TRUE;
+ gmx_bool bKeepAndNumCPT = FALSE;
+ gmx_bool bResetCountersHalfWay = FALSE;
+ gmx_output_env_t *oenv = NULL;
+
+ /* Non transparent initialization of a complex gmx_hw_opt_t struct.
+ * But unfortunately we are not allowed to call a function here,
+ * since declarations follow below.
+ */
+ gmx_hw_opt_t hw_opt = {
+ 0, 0, 0, 0, threadaffSEL, 0, 0,
+ { NULL, FALSE, 0, NULL }
+ };
+
+ t_pargs pa[] = {
+
+ { "-dd", FALSE, etRVEC, {&realddxyz},
+ "Domain decomposition grid, 0 is optimize" },
+ { "-ddorder", FALSE, etENUM, {ddrank_opt},
+ "DD rank order" },
+ { "-npme", FALSE, etINT, {&npme},
+ "Number of separate ranks to be used for PME, -1 is guess" },
+ { "-nt", FALSE, etINT, {&hw_opt.nthreads_tot},
+ "Total number of threads to start (0 is guess)" },
+ { "-ntmpi", FALSE, etINT, {&hw_opt.nthreads_tmpi},
+ "Number of thread-MPI threads to start (0 is guess)" },
+ { "-ntomp", FALSE, etINT, {&hw_opt.nthreads_omp},
+ "Number of OpenMP threads per MPI rank to start (0 is guess)" },
+ { "-ntomp_pme", FALSE, etINT, {&hw_opt.nthreads_omp_pme},
+ "Number of OpenMP threads per MPI rank to start (0 is -ntomp)" },
+ { "-pin", FALSE, etENUM, {thread_aff_opt},
+ "Whether mdrun should try to set thread affinities" },
+ { "-pinoffset", FALSE, etINT, {&hw_opt.core_pinning_offset},
+ "The lowest logical core number to which mdrun should pin the first thread" },
+ { "-pinstride", FALSE, etINT, {&hw_opt.core_pinning_stride},
+ "Pinning distance in logical cores for threads, use 0 to minimize the number of threads per physical core" },
+ { "-gpu_id", FALSE, etSTR, {&hw_opt.gpu_opt.gpu_id},
+ "List of GPU device id-s to use, specifies the per-node PP rank to GPU mapping" },
+ { "-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},
+ "Fraction in (0,1) by whose reciprocal the initial DD cell size will be increased in order to "
+ "provide a margin in which dynamic load balancing can act while preserving the minimum cell size." },
+ { "-ddcsx", FALSE, etSTR, {&ddcsx},
+ "HIDDENA string containing a vector of the relative sizes in the x "
+ "direction of the corresponding DD cells. Only effective with static "
+ "load balancing." },
+ { "-ddcsy", FALSE, etSTR, {&ddcsy},
+ "HIDDENA string containing a vector of the relative sizes in the y "
+ "direction of the corresponding DD cells. Only effective with static "
+ "load balancing." },
+ { "-ddcsz", FALSE, etSTR, {&ddcsz},
+ "HIDDENA string containing a vector of the relative sizes in the z "
+ "direction of the corresponding DD cells. Only effective with static "
+ "load balancing." },
+ { "-gcom", FALSE, etINT, {&nstglobalcomm},
+ "Global communication frequency" },
+ { "-nb", FALSE, etENUM, {&nbpu_opt},
+ "Calculate non-bonded interactions on" },
+ { "-nstlist", FALSE, etINT, {&nstlist},
+ "Set nstlist when using a Verlet buffer tolerance (0 is guess)" },
+ { "-tunepme", FALSE, etBOOL, {&bTunePME},
+ "Optimize PME load between PP/PME ranks or GPU/CPU" },
+ { "-v", FALSE, etBOOL, {&bVerbose},
+ "Be loud and noisy" },
+ { "-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, {&bTryToAppendFiles},
+ "Append to previous output files when continuing from checkpoint instead of adding the simulation part number to all file names" },
+ { "-nsteps", FALSE, etINT64, {&nsteps},
+ "Run this number of steps, overrides .mdp file option (-1 means infinite, -2 means use mdp option, smaller is invalid)" },
+ { "-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 periodically with this period (steps)" },
+ { "-nex", FALSE, etINT, {&repl_ex_nex},
+ "Number of random exchanges to carry out each exchange interval (N^3 is one suggestion). -nex zero or not specified gives neighbor replica exchange." },
+ { "-reseed", FALSE, etINT, {&repl_ex_seed},
+ "Seed for replica exchange, -1 is generate a seed" },
+ { "-imdport", FALSE, etINT, {&imdport},
+ "HIDDENIMD listening port" },
+ { "-imdwait", FALSE, etBOOL, {&bIMDwait},
+ "HIDDENPause the simulation while no IMD client is connected" },
+ { "-imdterm", FALSE, etBOOL, {&bIMDterm},
+ "HIDDENAllow termination of the simulation from IMD client" },
+ { "-imdpull", FALSE, etBOOL, {&bIMDpull},
+ "HIDDENAllow pulling in the simulation from IMD client" },
+ { "-rerunvsite", FALSE, etBOOL, {&bRerunVSite},
+ "HIDDENRecalculate virtual site coordinates with [TT]-rerun[tt]" },
+ { "-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 wall clock time for the run" },
+ { "-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]" }
+ };
+ unsigned long Flags;
+ ivec ddxyz;
+ int dd_rank_order;
+ gmx_bool bDoAppendFiles, bStartFromCpt;
+ FILE *fplog;
+ int rc;
+ char **multidir = NULL;
+
+ cr = init_commrec();
+
+ unsigned long PCA_Flags = PCA_CAN_SET_DEFFNM;
+ // With -multi or -multidir, the file names are going to get processed
+ // further (or the working directory changed), so we can't check for their
+ // existence during parsing. It isn't useful to do any completion based on
+ // file system contents, either.
+ if (is_multisim_option_set(argc, argv))
+ {
+ PCA_Flags |= PCA_DISABLE_INPUT_FILE_CHECKING;
+ }
+
+ /* 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;
+ }
+ */
+
+ if (!parse_common_args(&argc, argv, PCA_Flags, NFILE, fnm, asize(pa), pa,
+ asize(desc), desc, 0, NULL, &oenv))
+ {
+ return 0;
+ }
+
+
+ dd_rank_order = nenum(ddrank_opt);
+
+ hw_opt.thread_affinity = nenum(thread_aff_opt);
+
+ /* now check the -multi and -multidir option */
+ if (opt2bSet("-multidir", NFILE, fnm))
+ {
+ 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 (repl_ex_nex < 0)
+ {
+ gmx_fatal(FARGS, "Replica exchange number of exchanges needs to be positive");
+ }
+
+ if (nmultisim >= 1)
+ {
+#if !GMX_THREAD_MPI
+ gmx_bool bParFn = (multidir == NULL);
+ init_multisystem(cr, nmultisim, multidir, NFILE, fnm, bParFn);
+#else
+ gmx_fatal(FARGS, "mdrun -multi or -multidir are not supported with the thread-MPI library. "
+ "Please compile GROMACS with a proper external MPI library.");
+#endif
+ }
+
+ handleRestart(cr, bTryToAppendFiles, NFILE, fnm,
+ &bDoAppendFiles, &bStartFromCpt);
+
+ Flags = opt2bSet("-rerun", NFILE, fnm) ? MD_RERUN : 0;
+ Flags = Flags | (bDDBondCheck ? MD_DDBONDCHECK : 0);
+ Flags = Flags | (bDDBondComm ? MD_DDBONDCOMM : 0);
+ Flags = Flags | (bTunePME ? MD_TUNEPME : 0);
+ Flags = Flags | (bConfout ? MD_CONFOUT : 0);
+ Flags = Flags | (bRerunVSite ? MD_RERUN_VSITE : 0);
+ Flags = Flags | (bReproducible ? MD_REPRODUCIBLE : 0);
+ Flags = Flags | (bDoAppendFiles ? MD_APPENDFILES : 0);
+ Flags = Flags | (opt2parg_bSet("-append", asize(pa), pa) ? MD_APPENDFILESSET : 0);
+ Flags = Flags | (bKeepAndNumCPT ? MD_KEEPANDNUMCPT : 0);
+ Flags = Flags | (bStartFromCpt ? MD_STARTFROMCPT : 0);
+ Flags = Flags | (bResetCountersHalfWay ? MD_RESETCOUNTERSHALFWAY : 0);
+ Flags = Flags | (opt2parg_bSet("-ntomp", asize(pa), pa) ? MD_NTOMPSET : 0);
+ Flags = Flags | (bIMDwait ? MD_IMDWAIT : 0);
+ Flags = Flags | (bIMDterm ? MD_IMDTERM : 0);
+ Flags = Flags | (bIMDpull ? MD_IMDPULL : 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) && !bDoAppendFiles)
+ {
+ gmx_log_open(ftp2fn(efLOG, NFILE, fnm), cr,
+ Flags & MD_APPENDFILES, &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 = gmx::mdrunner(&hw_opt, fplog, cr, NFILE, fnm, oenv, bVerbose,
+ nstglobalcomm, ddxyz, dd_rank_order, npme, rdd, rconstr,
+ dddlb_opt[0], dlb_scale, ddcsx, ddcsy, ddcsz,
+ nbpu_opt[0], nstlist,
+ nsteps, nstepout, resetstep,
+ nmultisim, repl_ex_nst, repl_ex_nex, repl_ex_seed,
+ pforce, cpt_period, max_hours, imdport, Flags);
+
+ /* Log file has to be closed in mdrunner if we are appending to it
+ (fplog not set here) */
+ if (MASTER(cr) && !bDoAppendFiles)
+ {
+ gmx_log_close(fplog);
+ }
+
+ return rc;
+}
diff --git a/patches/gromacs-2016-beta1.diff/src/programs/mdrun/repl_ex.cpp b/patches/gromacs-2016-beta1.diff/src/programs/mdrun/repl_ex.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..2b870b96d1ec09aa25e998809568da441d706738
--- /dev/null
+++ b/patches/gromacs-2016-beta1.diff/src/programs/mdrun/repl_ex.cpp
@@ -0,0 +1,1479 @@
+/*
+ * This file is part of the GROMACS molecular simulation package.
+ *
+ * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
+ * Copyright (c) 2001-2004, The GROMACS development team.
+ * Copyright (c) 2011,2012,2013,2014,2015,2016, by the GROMACS development team, led by
+ * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
+ * and including many others, as listed in the AUTHORS file in the
+ * top-level source directory and at http://www.gromacs.org.
+ *
+ * GROMACS is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public License
+ * as published by the Free Software Foundation; either version 2.1
+ * of the License, or (at your option) any later version.
+ *
+ * GROMACS is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with GROMACS; if not, see
+ * http://www.gnu.org/licenses, or write to the Free Software Foundation,
+ * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ * If you want to redistribute modifications to GROMACS, 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 http://www.gromacs.org.
+ *
+ * To help us fund GROMACS development, we humbly ask that you cite
+ * the research papers on the package. Check out http://www.gromacs.org.
+ */
+
+#include "gmxpre.h"
+
+#include "repl_ex.h"
+
+#include "config.h"
+
+#include <math.h>
+
+#include <random>
+
+#include "gromacs/domdec/domdec.h"
+#include "gromacs/gmxlib/network.h"
+#include "gromacs/math/units.h"
+#include "gromacs/math/vec.h"
+#include "gromacs/mdlib/main.h"
+#include "gromacs/mdtypes/commrec.h"
+#include "gromacs/mdtypes/inputrec.h"
+#include "gromacs/mdtypes/md_enums.h"
+#include "gromacs/random/threefry.h"
+#include "gromacs/random/uniformintdistribution.h"
+#include "gromacs/random/uniformrealdistribution.h"
+#include "gromacs/utility/fatalerror.h"
+#include "gromacs/utility/pleasecite.h"
+#include "gromacs/utility/smalloc.h"
+
+
+/* PLUMED */
+#include "../../../Plumed.h"
+extern int plumedswitch;
+extern plumed plumedmain;
+/* END PLUMED */
+
+#define PROBABILITYCUTOFF 100
+/* we don't bother evaluating if events are more rare than exp(-100) = 3.7x10^-44 */
+
+//! Rank in the multisimulaiton
+#define MSRANK(ms, nodeid) (nodeid)
+
+enum {
+ ereTEMP, ereLAMBDA, ereENDSINGLE, ereTL, ereNR
+};
+const char *erename[ereNR] = { "temperature", "lambda", "end_single_marker", "temperature and lambda"};
+/* end_single_marker merely notes the end of single variable replica exchange. All types higher than
+ it are multiple replica exchange methods */
+/* Eventually, should add 'pressure', 'temperature and pressure', 'lambda_and_pressure', 'temperature_lambda_pressure'?;
+ Let's wait until we feel better about the pressure control methods giving exact ensembles. Right now, we assume constant pressure */
+
+typedef struct gmx_repl_ex
+{
+ int repl; /* replica ID */
+ int nrepl; /* total number of replica */
+ real temp; /* temperature */
+ int type; /* replica exchange type from ere enum */
+ real **q; /* quantity, e.g. temperature or lambda; first index is ere, second index is replica ID */
+ gmx_bool bNPT; /* use constant pressure and temperature */
+ real *pres; /* replica pressures */
+ int *ind; /* replica indices */
+ int *allswaps; /* used for keeping track of all the replica swaps */
+ int nst; /* replica exchange interval (number of steps) */
+ int nex; /* number of exchanges per interval */
+ int seed; /* random seed */
+ int nattempt[2]; /* number of even and odd replica change attempts */
+ real *prob_sum; /* sum of probabilities */
+ int **nmoves; /* number of moves between replicas i and j */
+ int *nexchange; /* i-th element of the array is the number of exchanges between replica i-1 and i */
+
+ /* these are helper arrays for replica exchange; allocated here so they
+ don't have to be allocated each time */
+ int *destinations;
+ int **cyclic;
+ int **order;
+ int *tmpswap;
+ gmx_bool *incycle;
+ gmx_bool *bEx;
+
+ /* helper arrays to hold the quantities that are exchanged */
+ real *prob;
+ real *Epot;
+ real *beta;
+ real *Vol;
+ real **de;
+
+} t_gmx_repl_ex;
+
+static gmx_bool repl_quantity(const gmx_multisim_t *ms,
+ struct gmx_repl_ex *re, int ere, real q)
+{
+ real *qall;
+ gmx_bool bDiff;
+ int s;
+
+ snew(qall, ms->nsim);
+ qall[re->repl] = q;
+ gmx_sum_sim(ms->nsim, qall, ms);
+
+ /* PLUMED */
+ //bDiff = FALSE;
+ //for (s = 1; s < ms->nsim; s++)
+ //{
+ // if (qall[s] != qall[0])
+ // {
+ bDiff = TRUE;
+ // }
+ //}
+ /* END PLUMED */
+
+ if (bDiff)
+ {
+ /* Set the replica exchange type and quantities */
+ re->type = ere;
+
+ snew(re->q[ere], re->nrepl);
+ for (s = 0; s < ms->nsim; s++)
+ {
+ re->q[ere][s] = qall[s];
+ }
+ }
+ sfree(qall);
+ return bDiff;
+}
+
+gmx_repl_ex_t init_replica_exchange(FILE *fplog,
+ const gmx_multisim_t *ms,
+ const t_state *state,
+ const t_inputrec *ir,
+ int nst, int nex, int init_seed)
+{
+ real pres;
+ int i, j, k;
+ struct gmx_repl_ex *re;
+ gmx_bool bTemp;
+ gmx_bool bLambda = FALSE;
+
+ fprintf(fplog, "\nInitializing Replica Exchange\n");
+
+ if (ms == NULL || ms->nsim == 1)
+ {
+ gmx_fatal(FARGS, "Nothing to exchange with only one replica, maybe you forgot to set the -multi option of mdrun?");
+ }
+ if (!EI_DYNAMICS(ir->eI))
+ {
+ gmx_fatal(FARGS, "Replica exchange is only supported by dynamical simulations");
+ /* Note that PAR(cr) is defined by cr->nnodes > 1, which is
+ * distinct from MULTISIM(cr). A multi-simulation only runs
+ * with real MPI parallelism, but this does not imply PAR(cr)
+ * is true!
+ *
+ * Since we are using a dynamical integrator, the only
+ * decomposition is DD, so PAR(cr) and DOMAINDECOMP(cr) are
+ * synonymous. The only way for cr->nnodes > 1 to be true is
+ * if we are using DD. */
+ }
+
+ snew(re, 1);
+
+ re->repl = ms->sim;
+ re->nrepl = ms->nsim;
+ snew(re->q, ereENDSINGLE);
+
+ fprintf(fplog, "Repl There are %d replicas:\n", re->nrepl);
+
+ check_multi_int(fplog, ms, state->natoms, "the number of atoms", FALSE);
+ check_multi_int(fplog, ms, ir->eI, "the integrator", FALSE);
+ check_multi_int64(fplog, ms, ir->init_step+ir->nsteps, "init_step+nsteps", FALSE);
+ check_multi_int64(fplog, ms, (ir->init_step+nst-1)/nst,
+ "first exchange step: init_step/-replex", FALSE);
+ check_multi_int(fplog, ms, ir->etc, "the temperature coupling", FALSE);
+ check_multi_int(fplog, ms, ir->opts.ngtc,
+ "the number of temperature coupling groups", FALSE);
+ check_multi_int(fplog, ms, ir->epc, "the pressure coupling", FALSE);
+ check_multi_int(fplog, ms, ir->efep, "free energy", FALSE);
+ check_multi_int(fplog, ms, ir->fepvals->n_lambda, "number of lambda states", FALSE);
+
+ re->temp = ir->opts.ref_t[0];
+ for (i = 1; (i < ir->opts.ngtc); i++)
+ {
+ if (ir->opts.ref_t[i] != re->temp)
+ {
+ fprintf(fplog, "\nWARNING: The temperatures of the different temperature coupling groups are not identical\n\n");
+ fprintf(stderr, "\nWARNING: The temperatures of the different temperature coupling groups are not identical\n\n");
+ }
+ }
+
+ re->type = -1;
+ bTemp = repl_quantity(ms, re, ereTEMP, re->temp);
+ if (ir->efep != efepNO)
+ {
+ bLambda = repl_quantity(ms, re, ereLAMBDA, (real)ir->fepvals->init_fep_state);
+ }
+ if (re->type == -1) /* nothing was assigned */
+ {
+ gmx_fatal(FARGS, "The properties of the %d systems are all the same, there is nothing to exchange", re->nrepl);
+ }
+ if (bLambda && bTemp)
+ {
+ re->type = ereTL;
+ }
+
+ if (bTemp)
+ {
+ please_cite(fplog, "Sugita1999a");
+ if (ir->epc != epcNO)
+ {
+ re->bNPT = TRUE;
+ fprintf(fplog, "Repl Using Constant Pressure REMD.\n");
+ please_cite(fplog, "Okabe2001a");
+ }
+ if (ir->etc == etcBERENDSEN)
+ {
+ gmx_fatal(FARGS, "REMD with the %s thermostat does not produce correct potential energy distributions, consider using the %s thermostat instead",
+ ETCOUPLTYPE(ir->etc), ETCOUPLTYPE(etcVRESCALE));
+ }
+ }
+ if (bLambda)
+ {
+ if (ir->fepvals->delta_lambda != 0) /* check this? */
+ {
+ gmx_fatal(FARGS, "delta_lambda is not zero");
+ }
+ }
+ if (re->bNPT)
+ {
+ snew(re->pres, re->nrepl);
+ if (ir->epct == epctSURFACETENSION)
+ {
+ pres = ir->ref_p[ZZ][ZZ];
+ }
+ else
+ {
+ pres = 0;
+ j = 0;
+ for (i = 0; i < DIM; i++)
+ {
+ if (ir->compress[i][i] != 0)
+ {
+ pres += ir->ref_p[i][i];
+ j++;
+ }
+ }
+ pres /= j;
+ }
+ re->pres[re->repl] = pres;
+ gmx_sum_sim(re->nrepl, re->pres, ms);
+ }
+
+ /* Make an index for increasing replica order */
+ /* only makes sense if one or the other is varying, not both!
+ if both are varying, we trust the order the person gave. */
+ snew(re->ind, re->nrepl);
+ for (i = 0; i < re->nrepl; i++)
+ {
+ re->ind[i] = i;
+ }
+
+ /* PLUMED */
+ // plumed2: check if we want alternative patterns (i.e. for bias-exchange metaD)
+ // in those cases replicas can share the same temperature.
+ /*
+ if (re->type < ereENDSINGLE)
+ {
+
+ for (i = 0; i < re->nrepl; i++)
+ {
+ for (j = i+1; j < re->nrepl; j++)
+ {
+ if (re->q[re->type][re->ind[j]] < re->q[re->type][re->ind[i]])
+ {*/
+ /* Unordered replicas are supposed to work, but there
+ * is still an issues somewhere.
+ * Note that at this point still re->ind[i]=i.
+ */
+ /*
+ gmx_fatal(FARGS, "Replicas with indices %d < %d have %ss %g > %g, please order your replicas on increasing %s",
+ i, j,
+ erename[re->type],
+ re->q[re->type][i], re->q[re->type][j],
+ erename[re->type]);
+
+ k = re->ind[i];
+ re->ind[i] = re->ind[j];
+ re->ind[j] = k;
+ }
+ else if (re->q[re->type][re->ind[j]] == re->q[re->type][re->ind[i]])
+ {
+ gmx_fatal(FARGS, "Two replicas have identical %ss", erename[re->type]);
+ }
+ }
+ }
+ }
+ */
+ /* END PLUMED */
+
+ /* keep track of all the swaps, starting with the initial placement. */
+ snew(re->allswaps, re->nrepl);
+ for (i = 0; i < re->nrepl; i++)
+ {
+ re->allswaps[i] = re->ind[i];
+ }
+
+ switch (re->type)
+ {
+ case ereTEMP:
+ fprintf(fplog, "\nReplica exchange in temperature\n");
+ for (i = 0; i < re->nrepl; i++)
+ {
+ fprintf(fplog, " %5.1f", re->q[re->type][re->ind[i]]);
+ }
+ fprintf(fplog, "\n");
+ break;
+ case ereLAMBDA:
+ fprintf(fplog, "\nReplica exchange in lambda\n");
+ for (i = 0; i < re->nrepl; i++)
+ {
+ fprintf(fplog, " %3d", (int)re->q[re->type][re->ind[i]]);
+ }
+ fprintf(fplog, "\n");
+ break;
+ case ereTL:
+ fprintf(fplog, "\nReplica exchange in temperature and lambda state\n");
+ for (i = 0; i < re->nrepl; i++)
+ {
+ fprintf(fplog, " %5.1f", re->q[ereTEMP][re->ind[i]]);
+ }
+ fprintf(fplog, "\n");
+ for (i = 0; i < re->nrepl; i++)
+ {
+ fprintf(fplog, " %5d", (int)re->q[ereLAMBDA][re->ind[i]]);
+ }
+ fprintf(fplog, "\n");
+ break;
+ default:
+ gmx_incons("Unknown replica exchange quantity");
+ }
+ if (re->bNPT)
+ {
+ fprintf(fplog, "\nRepl p");
+ for (i = 0; i < re->nrepl; i++)
+ {
+ fprintf(fplog, " %5.2f", re->pres[re->ind[i]]);
+ }
+
+ for (i = 0; i < re->nrepl; i++)
+ {
+ if ((i > 0) && (re->pres[re->ind[i]] < re->pres[re->ind[i-1]]))
+ {
+ fprintf(fplog, "\nWARNING: The reference pressures decrease with increasing temperatures\n\n");
+ fprintf(stderr, "\nWARNING: The reference pressures decrease with increasing temperatures\n\n");
+ }
+ }
+ }
+ re->nst = nst;
+ if (init_seed == -1)
+ {
+ if (MASTERSIM(ms))
+ {
+ re->seed = static_cast<int>(gmx::makeRandomSeed());
+ }
+ else
+ {
+ re->seed = 0;
+ }
+ gmx_sumi_sim(1, &(re->seed), ms);
+ }
+ else
+ {
+ re->seed = init_seed;
+ }
+ fprintf(fplog, "\nReplica exchange interval: %d\n", re->nst);
+ fprintf(fplog, "\nReplica random seed: %d\n", re->seed);
+
+ re->nattempt[0] = 0;
+ re->nattempt[1] = 0;
+
+ snew(re->prob_sum, re->nrepl);
+ snew(re->nexchange, re->nrepl);
+ snew(re->nmoves, re->nrepl);
+ for (i = 0; i < re->nrepl; i++)
+ {
+ snew(re->nmoves[i], re->nrepl);
+ }
+ fprintf(fplog, "Replica exchange information below: ex and x = exchange, pr = probability\n");
+
+ /* generate space for the helper functions so we don't have to snew each time */
+
+ snew(re->destinations, re->nrepl);
+ snew(re->incycle, re->nrepl);
+ snew(re->tmpswap, re->nrepl);
+ snew(re->cyclic, re->nrepl);
+ snew(re->order, re->nrepl);
+ for (i = 0; i < re->nrepl; i++)
+ {
+ snew(re->cyclic[i], re->nrepl+1);
+ snew(re->order[i], re->nrepl);
+ }
+ /* allocate space for the functions storing the data for the replicas */
+ /* not all of these arrays needed in all cases, but they don't take
+ up much space, since the max size is nrepl**2 */
+ snew(re->prob, re->nrepl);
+ snew(re->bEx, re->nrepl);
+ snew(re->beta, re->nrepl);
+ snew(re->Vol, re->nrepl);
+ snew(re->Epot, re->nrepl);
+ snew(re->de, re->nrepl);
+ for (i = 0; i < re->nrepl; i++)
+ {
+ snew(re->de[i], re->nrepl);
+ }
+ re->nex = nex;
+ return re;
+}
+
+static void exchange_reals(const gmx_multisim_t gmx_unused *ms, int gmx_unused b, real *v, int n)
+{
+ real *buf;
+ int i;
+
+ if (v)
+ {
+ snew(buf, n);
+#if GMX_MPI
+ /*
+ MPI_Sendrecv(v, n*sizeof(real),MPI_BYTE,MSRANK(ms,b),0,
+ buf,n*sizeof(real),MPI_BYTE,MSRANK(ms,b),0,
+ ms->mpi_comm_masters,MPI_STATUS_IGNORE);
+ */
+ {
+ MPI_Request mpi_req;
+
+ MPI_Isend(v, n*sizeof(real), MPI_BYTE, MSRANK(ms, b), 0,
+ ms->mpi_comm_masters, &mpi_req);
+ MPI_Recv(buf, n*sizeof(real), MPI_BYTE, MSRANK(ms, b), 0,
+ ms->mpi_comm_masters, MPI_STATUS_IGNORE);
+ MPI_Wait(&mpi_req, MPI_STATUS_IGNORE);
+ }
+#endif
+ for (i = 0; i < n; i++)
+ {
+ v[i] = buf[i];
+ }
+ sfree(buf);
+ }
+}
+
+
+static void exchange_doubles(const gmx_multisim_t gmx_unused *ms, int gmx_unused b, double *v, int n)
+{
+ double *buf;
+ int i;
+
+ if (v)
+ {
+ snew(buf, n);
+#if GMX_MPI
+ /*
+ MPI_Sendrecv(v, n*sizeof(double),MPI_BYTE,MSRANK(ms,b),0,
+ buf,n*sizeof(double),MPI_BYTE,MSRANK(ms,b),0,
+ ms->mpi_comm_masters,MPI_STATUS_IGNORE);
+ */
+ {
+ MPI_Request mpi_req;
+
+ MPI_Isend(v, n*sizeof(double), MPI_BYTE, MSRANK(ms, b), 0,
+ ms->mpi_comm_masters, &mpi_req);
+ MPI_Recv(buf, n*sizeof(double), MPI_BYTE, MSRANK(ms, b), 0,
+ ms->mpi_comm_masters, MPI_STATUS_IGNORE);
+ MPI_Wait(&mpi_req, MPI_STATUS_IGNORE);
+ }
+#endif
+ for (i = 0; i < n; i++)
+ {
+ v[i] = buf[i];
+ }
+ sfree(buf);
+ }
+}
+
+static void exchange_rvecs(const gmx_multisim_t gmx_unused *ms, int gmx_unused b, rvec *v, int n)
+{
+ rvec *buf;
+ int i;
+
+ if (v)
+ {
+ snew(buf, n);
+#if GMX_MPI
+ /*
+ MPI_Sendrecv(v[0], n*sizeof(rvec),MPI_BYTE,MSRANK(ms,b),0,
+ buf[0],n*sizeof(rvec),MPI_BYTE,MSRANK(ms,b),0,
+ ms->mpi_comm_masters,MPI_STATUS_IGNORE);
+ */
+ {
+ MPI_Request mpi_req;
+
+ MPI_Isend(v[0], n*sizeof(rvec), MPI_BYTE, MSRANK(ms, b), 0,
+ ms->mpi_comm_masters, &mpi_req);
+ MPI_Recv(buf[0], n*sizeof(rvec), MPI_BYTE, MSRANK(ms, b), 0,
+ ms->mpi_comm_masters, MPI_STATUS_IGNORE);
+ MPI_Wait(&mpi_req, MPI_STATUS_IGNORE);
+ }
+#endif
+ for (i = 0; i < n; i++)
+ {
+ copy_rvec(buf[i], v[i]);
+ }
+ sfree(buf);
+ }
+}
+
+static void exchange_state(const gmx_multisim_t *ms, int b, t_state *state)
+{
+ /* When t_state changes, this code should be updated. */
+ int ngtc, nnhpres;
+ ngtc = state->ngtc * state->nhchainlength;
+ nnhpres = state->nnhpres* state->nhchainlength;
+ exchange_rvecs(ms, b, state->box, DIM);
+ exchange_rvecs(ms, b, state->box_rel, DIM);
+ exchange_rvecs(ms, b, state->boxv, DIM);
+ exchange_reals(ms, b, &(state->veta), 1);
+ exchange_reals(ms, b, &(state->vol0), 1);
+ exchange_rvecs(ms, b, state->svir_prev, DIM);
+ exchange_rvecs(ms, b, state->fvir_prev, DIM);
+ exchange_rvecs(ms, b, state->pres_prev, DIM);
+ exchange_doubles(ms, b, state->nosehoover_xi, ngtc);
+ exchange_doubles(ms, b, state->nosehoover_vxi, ngtc);
+ exchange_doubles(ms, b, state->nhpres_xi, nnhpres);
+ exchange_doubles(ms, b, state->nhpres_vxi, nnhpres);
+ exchange_doubles(ms, b, state->therm_integral, state->ngtc);
+ exchange_rvecs(ms, b, state->x, state->natoms);
+ exchange_rvecs(ms, b, state->v, state->natoms);
+}
+
+static void copy_rvecs(rvec *s, rvec *d, int n)
+{
+ int i;
+
+ if (d != NULL)
+ {
+ for (i = 0; i < n; i++)
+ {
+ copy_rvec(s[i], d[i]);
+ }
+ }
+}
+
+static void copy_doubles(const double *s, double *d, int n)
+{
+ int i;
+
+ if (d != NULL)
+ {
+ for (i = 0; i < n; i++)
+ {
+ d[i] = s[i];
+ }
+ }
+}
+
+static void copy_reals(const real *s, real *d, int n)
+{
+ int i;
+
+ if (d != NULL)
+ {
+ for (i = 0; i < n; i++)
+ {
+ d[i] = s[i];
+ }
+ }
+}
+
+static void copy_ints(const int *s, int *d, int n)
+{
+ int i;
+
+ if (d != NULL)
+ {
+ for (i = 0; i < n; i++)
+ {
+ d[i] = s[i];
+ }
+ }
+}
+
+#define scopy_rvecs(v, n) copy_rvecs(state->v, state_local->v, n);
+#define scopy_doubles(v, n) copy_doubles(state->v, state_local->v, n);
+#define scopy_reals(v, n) copy_reals(state->v, state_local->v, n);
+#define scopy_ints(v, n) copy_ints(state->v, state_local->v, n);
+
+static void copy_state_nonatomdata(t_state *state, t_state *state_local)
+{
+ /* When t_state changes, this code should be updated. */
+ int ngtc, nnhpres;
+ ngtc = state->ngtc * state->nhchainlength;
+ nnhpres = state->nnhpres* state->nhchainlength;
+ scopy_rvecs(box, DIM);
+ scopy_rvecs(box_rel, DIM);
+ scopy_rvecs(boxv, DIM);
+ state_local->veta = state->veta;
+ state_local->vol0 = state->vol0;
+ scopy_rvecs(svir_prev, DIM);
+ scopy_rvecs(fvir_prev, DIM);
+ scopy_rvecs(pres_prev, DIM);
+ scopy_doubles(nosehoover_xi, ngtc);
+ scopy_doubles(nosehoover_vxi, ngtc);
+ scopy_doubles(nhpres_xi, nnhpres);
+ scopy_doubles(nhpres_vxi, nnhpres);
+ scopy_doubles(therm_integral, state->ngtc);
+ scopy_rvecs(x, state->natoms);
+ scopy_rvecs(v, state->natoms);
+ copy_ints(&(state->fep_state), &(state_local->fep_state), 1);
+ scopy_reals(lambda, efptNR);
+}
+
+static void scale_velocities(t_state *state, real fac)
+{
+ int i;
+
+ if (state->v)
+ {
+ for (i = 0; i < state->natoms; i++)
+ {
+ svmul(fac, state->v[i], state->v[i]);
+ }
+ }
+}
+
+static void print_transition_matrix(FILE *fplog, int n, int **nmoves, int *nattempt)
+{
+ int i, j, ntot;
+ float Tprint;
+
+ ntot = nattempt[0] + nattempt[1];
+ fprintf(fplog, "\n");
+ fprintf(fplog, "Repl");
+ for (i = 0; i < n; i++)
+ {
+ fprintf(fplog, " "); /* put the title closer to the center */
+ }
+ fprintf(fplog, "Empirical Transition Matrix\n");
+
+ fprintf(fplog, "Repl");
+ for (i = 0; i < n; i++)
+ {
+ fprintf(fplog, "%8d", (i+1));
+ }
+ fprintf(fplog, "\n");
+
+ for (i = 0; i < n; i++)
+ {
+ fprintf(fplog, "Repl");
+ for (j = 0; j < n; j++)
+ {
+ Tprint = 0.0;
+ if (nmoves[i][j] > 0)
+ {
+ Tprint = nmoves[i][j]/(2.0*ntot);
+ }
+ fprintf(fplog, "%8.4f", Tprint);
+ }
+ fprintf(fplog, "%3d\n", i);
+ }
+}
+
+static void print_ind(FILE *fplog, const char *leg, int n, int *ind, gmx_bool *bEx)
+{
+ int i;
+
+ fprintf(fplog, "Repl %2s %2d", leg, ind[0]);
+ for (i = 1; i < n; i++)
+ {
+ fprintf(fplog, " %c %2d", (bEx != 0 && bEx[i]) ? 'x' : ' ', ind[i]);
+ }
+ fprintf(fplog, "\n");
+}
+
+static void print_allswitchind(FILE *fplog, int n, int *pind, int *allswaps, int *tmpswap)
+{
+ int i;
+
+ for (i = 0; i < n; i++)
+ {
+ tmpswap[i] = allswaps[i];
+ }
+ for (i = 0; i < n; i++)
+ {
+ allswaps[i] = tmpswap[pind[i]];
+ }
+
+ fprintf(fplog, "\nAccepted Exchanges: ");
+ for (i = 0; i < n; i++)
+ {
+ fprintf(fplog, "%d ", pind[i]);
+ }
+ fprintf(fplog, "\n");
+
+ /* the "Order After Exchange" is the state label corresponding to the configuration that
+ started in state listed in order, i.e.
+
+ 3 0 1 2
+
+ means that the:
+ configuration starting in simulation 3 is now in simulation 0,
+ configuration starting in simulation 0 is now in simulation 1,
+ configuration starting in simulation 1 is now in simulation 2,
+ configuration starting in simulation 2 is now in simulation 3
+ */
+ fprintf(fplog, "Order After Exchange: ");
+ for (i = 0; i < n; i++)
+ {
+ fprintf(fplog, "%d ", allswaps[i]);
+ }
+ fprintf(fplog, "\n\n");
+}
+
+static void print_prob(FILE *fplog, const char *leg, int n, real *prob)
+{
+ int i;
+ char buf[8];
+
+ fprintf(fplog, "Repl %2s ", leg);
+ for (i = 1; i < n; i++)
+ {
+ if (prob[i] >= 0)
+ {
+ sprintf(buf, "%4.2f", prob[i]);
+ fprintf(fplog, " %3s", buf[0] == '1' ? "1.0" : buf+1);
+ }
+ else
+ {
+ fprintf(fplog, " ");
+ }
+ }
+ fprintf(fplog, "\n");
+}
+
+static void print_count(FILE *fplog, const char *leg, int n, int *count)
+{
+ int i;
+
+ fprintf(fplog, "Repl %2s ", leg);
+ for (i = 1; i < n; i++)
+ {
+ fprintf(fplog, " %4d", count[i]);
+ }
+ fprintf(fplog, "\n");
+}
+
+static real calc_delta(FILE *fplog, gmx_bool bPrint, struct gmx_repl_ex *re, int a, int b, int ap, int bp)
+{
+
+ real ediff, dpV, delta = 0;
+ real *Epot = re->Epot;
+ real *Vol = re->Vol;
+ real **de = re->de;
+ real *beta = re->beta;
+
+ /* Two cases; we are permuted and not. In all cases, setting ap = a and bp = b will reduce
+ to the non permuted case */
+
+ switch (re->type)
+ {
+ case ereTEMP:
+ /*
+ * Okabe et. al. Chem. Phys. Lett. 335 (2001) 435-439
+ */
+ ediff = Epot[b] - Epot[a];
+ delta = -(beta[bp] - beta[ap])*ediff;
+ break;
+ case ereLAMBDA:
+ /* two cases: when we are permuted, and not. */
+ /* non-permuted:
+ ediff = E_new - E_old
+ = [H_b(x_a) + H_a(x_b)] - [H_b(x_b) + H_a(x_a)]
+ = [H_b(x_a) - H_a(x_a)] + [H_a(x_b) - H_b(x_b)]
+ = de[b][a] + de[a][b] */
+
+ /* permuted:
+ ediff = E_new - E_old
+ = [H_bp(x_a) + H_ap(x_b)] - [H_bp(x_b) + H_ap(x_a)]
+ = [H_bp(x_a) - H_ap(x_a)] + [H_ap(x_b) - H_bp(x_b)]
+ = [H_bp(x_a) - H_a(x_a) + H_a(x_a) - H_ap(x_a)] + [H_ap(x_b) - H_b(x_b) + H_b(x_b) - H_bp(x_b)]
+ = [H_bp(x_a) - H_a(x_a)] - [H_ap(x_a) - H_a(x_a)] + [H_ap(x_b) - H_b(x_b)] - H_bp(x_b) - H_b(x_b)]
+ = (de[bp][a] - de[ap][a]) + (de[ap][b] - de[bp][b]) */
+ /* but, in the current code implementation, we flip configurations, not indices . . .
+ So let's examine that.
+ = [H_b(x_ap) - H_a(x_a)] - [H_a(x_ap) - H_a(x_a)] + [H_a(x_bp) - H_b(x_b)] - H_b(x_bp) - H_b(x_b)]
+ = [H_b(x_ap) - H_a(x_ap)] + [H_a(x_bp) - H_b(x_pb)]
+ = (de[b][ap] - de[a][ap]) + (de[a][bp] - de[b][bp]
+ So, if we exchange b<=> bp and a<=> ap, we return to the same result.
+ So the simple solution is to flip the
+ position of perturbed and original indices in the tests.
+ */
+
+ ediff = (de[bp][a] - de[ap][a]) + (de[ap][b] - de[bp][b]);
+ delta = ediff*beta[a]; /* assume all same temperature in this case */
+ break;
+ case ereTL:
+ /* not permuted: */
+ /* delta = reduced E_new - reduced E_old
+ = [beta_b H_b(x_a) + beta_a H_a(x_b)] - [beta_b H_b(x_b) + beta_a H_a(x_a)]
+ = [beta_b H_b(x_a) - beta_a H_a(x_a)] + [beta_a H_a(x_b) - beta_b H_b(x_b)]
+ = [beta_b dH_b(x_a) + beta_b H_a(x_a) - beta_a H_a(x_a)] +
+ [beta_a dH_a(x_b) + beta_a H_b(x_b) - beta_b H_b(x_b)]
+ = [beta_b dH_b(x_a) + [beta_a dH_a(x_b) +
+ beta_b (H_a(x_a) - H_b(x_b)]) - beta_a (H_a(x_a) - H_b(x_b))
+ = beta_b dH_b(x_a) + beta_a dH_a(x_b) - (beta_b - beta_a)(H_b(x_b) - H_a(x_a) */
+ /* delta = beta[b]*de[b][a] + beta[a]*de[a][b] - (beta[b] - beta[a])*(Epot[b] - Epot[a]; */
+ /* permuted (big breath!) */
+ /* delta = reduced E_new - reduced E_old
+ = [beta_bp H_bp(x_a) + beta_ap H_ap(x_b)] - [beta_bp H_bp(x_b) + beta_ap H_ap(x_a)]
+ = [beta_bp H_bp(x_a) - beta_ap H_ap(x_a)] + [beta_ap H_ap(x_b) - beta_bp H_bp(x_b)]
+ = [beta_bp H_bp(x_a) - beta_ap H_ap(x_a)] + [beta_ap H_ap(x_b) - beta_bp H_bp(x_b)]
+ - beta_pb H_a(x_a) + beta_ap H_a(x_a) + beta_pb H_a(x_a) - beta_ap H_a(x_a)
+ - beta_ap H_b(x_b) + beta_bp H_b(x_b) + beta_ap H_b(x_b) - beta_bp H_b(x_b)
+ = [(beta_bp H_bp(x_a) - beta_bp H_a(x_a)) - (beta_ap H_ap(x_a) - beta_ap H_a(x_a))] +
+ [(beta_ap H_ap(x_b) - beta_ap H_b(x_b)) - (beta_bp H_bp(x_b) - beta_bp H_b(x_b))]
+ + beta_pb H_a(x_a) - beta_ap H_a(x_a) + beta_ap H_b(x_b) - beta_bp H_b(x_b)
+ = [beta_bp (H_bp(x_a) - H_a(x_a)) - beta_ap (H_ap(x_a) - H_a(x_a))] +
+ [beta_ap (H_ap(x_b) - H_b(x_b)) - beta_bp (H_bp(x_b) - H_b(x_b))]
+ + beta_pb (H_a(x_a) - H_b(x_b)) - beta_ap (H_a(x_a) - H_b(x_b))
+ = ([beta_bp de[bp][a] - beta_ap de[ap][a]) + beta_ap de[ap][b] - beta_bp de[bp][b])
+ + (beta_pb-beta_ap)(H_a(x_a) - H_b(x_b)) */
+ delta = beta[bp]*(de[bp][a] - de[bp][b]) + beta[ap]*(de[ap][b] - de[ap][a]) - (beta[bp]-beta[ap])*(Epot[b]-Epot[a]);
+ break;
+ default:
+ gmx_incons("Unknown replica exchange quantity");
+ }
+ if (bPrint)
+ {
+ fprintf(fplog, "Repl %d <-> %d dE_term = %10.3e (kT)\n", a, b, delta);
+ }
+ if (re->bNPT)
+ {
+ /* revist the calculation for 5.0. Might be some improvements. */
+ dpV = (beta[ap]*re->pres[ap]-beta[bp]*re->pres[bp])*(Vol[b]-Vol[a])/PRESFAC;
+ if (bPrint)
+ {
+ fprintf(fplog, " dpV = %10.3e d = %10.3e\n", dpV, delta + dpV);
+ }
+ delta += dpV;
+ }
+ return delta;
+}
+
+static void
+test_for_replica_exchange(FILE *fplog,
+ const gmx_multisim_t *ms,
+ struct gmx_repl_ex *re,
+ gmx_enerdata_t *enerd,
+ real vol,
+ gmx_int64_t step,
+ real time)
+{
+ int m, i, j, a, b, ap, bp, i0, i1, tmp;
+ real delta = 0;
+ gmx_bool bPrint, bMultiEx;
+ gmx_bool *bEx = re->bEx;
+ real *prob = re->prob;
+ int *pind = re->destinations; /* permuted index */
+ gmx_bool bEpot = FALSE;
+ gmx_bool bDLambda = FALSE;
+ gmx_bool bVol = FALSE;
+ gmx::ThreeFry2x64<64> rng(re->seed, gmx::RandomDomain::ReplicaExchange);
+ gmx::UniformRealDistribution<real> uniformRealDist;
+ gmx::UniformIntDistribution<int> uniformNreplDist(0, re->nrepl-1);
+
+ bMultiEx = (re->nex > 1); /* multiple exchanges at each state */
+ fprintf(fplog, "Replica exchange at step %" GMX_PRId64 " time %.5f\n", step, time);
+
+ if (re->bNPT)
+ {
+ for (i = 0; i < re->nrepl; i++)
+ {
+ re->Vol[i] = 0;
+ }
+ bVol = TRUE;
+ re->Vol[re->repl] = vol;
+ }
+ if ((re->type == ereTEMP || re->type == ereTL))
+ {
+ for (i = 0; i < re->nrepl; i++)
+ {
+ re->Epot[i] = 0;
+ }
+ bEpot = TRUE;
+ re->Epot[re->repl] = enerd->term[F_EPOT];
+ /* temperatures of different states*/
+ for (i = 0; i < re->nrepl; i++)
+ {
+ re->beta[i] = 1.0/(re->q[ereTEMP][i]*BOLTZ);
+ }
+ }
+ else
+ {
+ for (i = 0; i < re->nrepl; i++)
+ {
+ re->beta[i] = 1.0/(re->temp*BOLTZ); /* we have a single temperature */
+ }
+ }
+ if (re->type == ereLAMBDA || re->type == ereTL)
+ {
+ bDLambda = TRUE;
+ /* lambda differences. */
+ /* de[i][j] is the energy of the jth simulation in the ith Hamiltonian
+ minus the energy of the jth simulation in the jth Hamiltonian */
+ for (i = 0; i < re->nrepl; i++)
+ {
+ for (j = 0; j < re->nrepl; j++)
+ {
+ re->de[i][j] = 0;
+ }
+ }
+ for (i = 0; i < re->nrepl; i++)
+ {
+ re->de[i][re->repl] = (enerd->enerpart_lambda[(int)re->q[ereLAMBDA][i]+1]-enerd->enerpart_lambda[0]);
+ }
+ }
+
+ /* now actually do the communication */
+ if (bVol)
+ {
+ gmx_sum_sim(re->nrepl, re->Vol, ms);
+ }
+ if (bEpot)
+ {
+ gmx_sum_sim(re->nrepl, re->Epot, ms);
+ }
+ if (bDLambda)
+ {
+ for (i = 0; i < re->nrepl; i++)
+ {
+ gmx_sum_sim(re->nrepl, re->de[i], ms);
+ }
+ }
+
+ /* make a duplicate set of indices for shuffling */
+ for (i = 0; i < re->nrepl; i++)
+ {
+ pind[i] = re->ind[i];
+ }
+
+ /* PLUMED */
+ int plumed_test_exchange_pattern=0;
+ /* END PLUMED */
+
+ if (bMultiEx)
+ {
+ /* multiple random switch exchange */
+ int nself = 0;
+
+ rng.restart( step, 0 );
+
+ for (i = 0; i < re->nex + nself; i++)
+ {
+ /* randomly select a pair */
+ /* in theory, could reduce this by identifying only which switches had a nonneglibible
+ probability of occurring (log p > -100) and only operate on those switches */
+ /* find out which state it is from, and what label that state currently has. Likely
+ more work that useful. */
+ i0 = uniformNreplDist(rng);
+ i1 = uniformNreplDist(rng);
+ if (i0 == i1)
+ {
+ nself++;
+ continue; /* self-exchange, back up and do it again */
+ }
+
+ a = re->ind[i0]; /* what are the indices of these states? */
+ b = re->ind[i1];
+ ap = pind[i0];
+ bp = pind[i1];
+
+ bPrint = FALSE; /* too noisy */
+ /* calculate the energy difference */
+ /* if the code changes to flip the STATES, rather than the configurations,
+ use the commented version of the code */
+ /* delta = calc_delta(fplog,bPrint,re,a,b,ap,bp); */
+ delta = calc_delta(fplog, bPrint, re, ap, bp, a, b);
+
+ /* we actually only use the first space in the prob and bEx array,
+ since there are actually many switches between pairs. */
+
+ if (delta <= 0)
+ {
+ /* accepted */
+ prob[0] = 1;
+ bEx[0] = TRUE;
+ }
+ else
+ {
+ if (delta > PROBABILITYCUTOFF)
+ {
+ prob[0] = 0;
+ }
+ else
+ {
+ prob[0] = exp(-delta);
+ }
+ /* roll a number to determine if accepted */
+ bEx[0] = uniformRealDist(rng) < prob[0];
+ }
+ re->prob_sum[0] += prob[0];
+
+ if (bEx[0])
+ {
+ /* swap the states */
+ tmp = pind[i0];
+ pind[i0] = pind[i1];
+ pind[i1] = tmp;
+ }
+ }
+ re->nattempt[0]++; /* keep track of total permutation trials here */
+ print_allswitchind(fplog, re->nrepl, pind, re->allswaps, re->tmpswap);
+ }
+ else
+ {
+ /* standard nearest neighbor replica exchange */
+
+ m = (step / re->nst) % 2;
+ /* PLUMED */
+ if(plumedswitch){
+ int partner=re->repl;
+ plumed_cmd(plumedmain,"getExchangesFlag",&plumed_test_exchange_pattern);
+ if(plumed_test_exchange_pattern>0){
+ int *list;
+ snew(list,re->nrepl);
+ plumed_cmd(plumedmain,"setNumberOfReplicas",&(re->nrepl));
+ plumed_cmd(plumedmain,"getExchangesList",list);
+ for(i=0; i<re->nrepl; i++) re->ind[i]=list[i];
+ sfree(list);
+ }
+
+ for(i=1; i<re->nrepl; i++) {
+ if (i % 2 != m) continue;
+ a = re->ind[i-1];
+ b = re->ind[i];
+ if(re->repl==a) partner=b;
+ if(re->repl==b) partner=a;
+ }
+ plumed_cmd(plumedmain,"GREX setPartner",&partner);
+ plumed_cmd(plumedmain,"GREX calculate",NULL);
+ plumed_cmd(plumedmain,"GREX shareAllDeltaBias",NULL);
+ }
+ /* END PLUMED */
+ for (i = 1; i < re->nrepl; i++)
+ {
+ a = re->ind[i-1];
+ b = re->ind[i];
+
+ bPrint = (re->repl == a || re->repl == b);
+ if (i % 2 == m)
+ {
+ delta = calc_delta(fplog, bPrint, re, a, b, a, b);
+ /* PLUMED */
+ if(plumedswitch){
+ real adb,bdb,dplumed;
+ char buf[300];
+ sprintf(buf,"GREX getDeltaBias %d",a); plumed_cmd(plumedmain,buf,&adb);
+ sprintf(buf,"GREX getDeltaBias %d",b); plumed_cmd(plumedmain,buf,&bdb);
+ dplumed=adb*re->beta[a]+bdb*re->beta[b];
+ delta+=dplumed;
+ if (bPrint)
+ fprintf(fplog,"dplumed = %10.3e dE_Term = %10.3e (kT)\n",dplumed,delta);
+ }
+ /* END PLUMED */
+ if (delta <= 0)
+ {
+ /* accepted */
+ prob[i] = 1;
+ bEx[i] = TRUE;
+ }
+ else
+ {
+ if (delta > PROBABILITYCUTOFF)
+ {
+ prob[i] = 0;
+ }
+ else
+ {
+ prob[i] = exp(-delta);
+ }
+ /* roll a number to determine if accepted */
+ bEx[i] = uniformRealDist(rng) < prob[i];
+ }
+ re->prob_sum[i] += prob[i];
+
+ if (bEx[i])
+ {
+ /* PLUMED */
+ if(!plumed_test_exchange_pattern) {
+ /* standard neighbour swapping */
+ /* swap these two */
+ tmp = pind[i-1];
+ pind[i-1] = pind[i];
+ pind[i] = tmp;
+ re->nexchange[i]++; /* statistics for back compatibility */
+ } else {
+ /* alternative swapping patterns */
+ tmp = pind[a];
+ pind[a] = pind[b];
+ pind[b] = tmp;
+ re->nexchange[i]++; /* statistics for back compatibility */
+ }
+ /* END PLUMED */
+ }
+ }
+ else
+ {
+ prob[i] = -1;
+ bEx[i] = FALSE;
+ }
+ }
+ /* print some statistics */
+ print_ind(fplog, "ex", re->nrepl, re->ind, bEx);
+ print_prob(fplog, "pr", re->nrepl, prob);
+ fprintf(fplog, "\n");
+ re->nattempt[m]++;
+ }
+
+ /* PLUMED */
+ if(plumed_test_exchange_pattern>0) {
+ for (i = 0; i < re->nrepl; i++)
+ {
+ re->ind[i] = i;
+ }
+ }
+ /* END PLUMED */
+
+ /* record which moves were made and accepted */
+ for (i = 0; i < re->nrepl; i++)
+ {
+ re->nmoves[re->ind[i]][pind[i]] += 1;
+ re->nmoves[pind[i]][re->ind[i]] += 1;
+ }
+ fflush(fplog); /* make sure we can see what the last exchange was */
+}
+
+static void
+cyclic_decomposition(const int *destinations,
+ int **cyclic,
+ gmx_bool *incycle,
+ const int nrepl,
+ int *nswap)
+{
+
+ int i, j, c, p;
+ int maxlen = 1;
+ for (i = 0; i < nrepl; i++)
+ {
+ incycle[i] = FALSE;
+ }
+ for (i = 0; i < nrepl; i++) /* one cycle for each replica */
+ {
+ if (incycle[i])
+ {
+ cyclic[i][0] = -1;
+ continue;
+ }
+ cyclic[i][0] = i;
+ incycle[i] = TRUE;
+ c = 1;
+ p = i;
+ for (j = 0; j < nrepl; j++) /* potentially all cycles are part, but we will break first */
+ {
+ p = destinations[p]; /* start permuting */
+ if (p == i)
+ {
+ cyclic[i][c] = -1;
+ if (c > maxlen)
+ {
+ maxlen = c;
+ }
+ break; /* we've reached the original element, the cycle is complete, and we marked the end. */
+ }
+ else
+ {
+ cyclic[i][c] = p; /* each permutation gives a new member of the cycle */
+ incycle[p] = TRUE;
+ c++;
+ }
+ }
+ }
+ *nswap = maxlen - 1;
+
+ if (debug)
+ {
+ for (i = 0; i < nrepl; i++)
+ {
+ fprintf(debug, "Cycle %d:", i);
+ for (j = 0; j < nrepl; j++)
+ {
+ if (cyclic[i][j] < 0)
+ {
+ break;
+ }
+ fprintf(debug, "%2d", cyclic[i][j]);
+ }
+ fprintf(debug, "\n");
+ }
+ fflush(debug);
+ }
+}
+
+static void
+compute_exchange_order(int **cyclic,
+ int **order,
+ const int nrepl,
+ const int maxswap)
+{
+ int i, j;
+
+ for (j = 0; j < maxswap; j++)
+ {
+ for (i = 0; i < nrepl; i++)
+ {
+ if (cyclic[i][j+1] >= 0)
+ {
+ order[cyclic[i][j+1]][j] = cyclic[i][j];
+ order[cyclic[i][j]][j] = cyclic[i][j+1];
+ }
+ }
+ for (i = 0; i < nrepl; i++)
+ {
+ if (order[i][j] < 0)
+ {
+ order[i][j] = i; /* if it's not exchanging, it should stay this round*/
+ }
+ }
+ }
+
+ if (debug)
+ {
+ fprintf(debug, "Replica Exchange Order\n");
+ for (i = 0; i < nrepl; i++)
+ {
+ fprintf(debug, "Replica %d:", i);
+ for (j = 0; j < maxswap; j++)
+ {
+ if (order[i][j] < 0)
+ {
+ break;
+ }
+ fprintf(debug, "%2d", order[i][j]);
+ }
+ fprintf(debug, "\n");
+ }
+ fflush(debug);
+ }
+}
+
+static void
+prepare_to_do_exchange(struct gmx_repl_ex *re,
+ const int replica_id,
+ int *maxswap,
+ gmx_bool *bThisReplicaExchanged)
+{
+ int i, j;
+ /* Hold the cyclic decomposition of the (multiple) replica
+ * exchange. */
+ gmx_bool bAnyReplicaExchanged = FALSE;
+ *bThisReplicaExchanged = FALSE;
+
+ for (i = 0; i < re->nrepl; i++)
+ {
+ if (re->destinations[i] != re->ind[i])
+ {
+ /* only mark as exchanged if the index has been shuffled */
+ bAnyReplicaExchanged = TRUE;
+ break;
+ }
+ }
+ if (bAnyReplicaExchanged)
+ {
+ /* reinitialize the placeholder arrays */
+ for (i = 0; i < re->nrepl; i++)
+ {
+ for (j = 0; j < re->nrepl; j++)
+ {
+ re->cyclic[i][j] = -1;
+ re->order[i][j] = -1;
+ }
+ }
+
+ /* Identify the cyclic decomposition of the permutation (very
+ * fast if neighbor replica exchange). */
+ cyclic_decomposition(re->destinations, re->cyclic, re->incycle, re->nrepl, maxswap);
+
+ /* Now translate the decomposition into a replica exchange
+ * order at each step. */
+ compute_exchange_order(re->cyclic, re->order, re->nrepl, *maxswap);
+
+ /* Did this replica do any exchange at any point? */
+ for (j = 0; j < *maxswap; j++)
+ {
+ if (replica_id != re->order[replica_id][j])
+ {
+ *bThisReplicaExchanged = TRUE;
+ break;
+ }
+ }
+ }
+}
+
+gmx_bool replica_exchange(FILE *fplog, const t_commrec *cr, struct gmx_repl_ex *re,
+ t_state *state, gmx_enerdata_t *enerd,
+ t_state *state_local, gmx_int64_t step, real time)
+{
+ int j;
+ int replica_id = 0;
+ int exchange_partner;
+ int maxswap = 0;
+ /* Number of rounds of exchanges needed to deal with any multiple
+ * exchanges. */
+ /* Where each replica ends up after the exchange attempt(s). */
+ /* The order in which multiple exchanges will occur. */
+ gmx_bool bThisReplicaExchanged = FALSE;
+
+ /* PLUMED */
+ if(plumedswitch)plumed_cmd(plumedmain,"GREX prepare",NULL);
+ /* END PLUMED */
+
+ if (MASTER(cr))
+ {
+ replica_id = re->repl;
+ test_for_replica_exchange(fplog, cr->ms, re, enerd, det(state_local->box), step, time);
+ prepare_to_do_exchange(re, replica_id, &maxswap, &bThisReplicaExchanged);
+ }
+ /* Do intra-simulation broadcast so all processors belonging to
+ * each simulation know whether they need to participate in
+ * collecting the state. Otherwise, they might as well get on with
+ * the next thing to do. */
+ if (DOMAINDECOMP(cr))
+ {
+#if GMX_MPI
+ MPI_Bcast(&bThisReplicaExchanged, sizeof(gmx_bool), MPI_BYTE, MASTERRANK(cr),
+ cr->mpi_comm_mygroup);
+#endif
+ }
+
+ if (bThisReplicaExchanged)
+ {
+ /* Exchange the states */
+ /* Collect the global state on the master node */
+ if (DOMAINDECOMP(cr))
+ {
+ dd_collect_state(cr->dd, state_local, state);
+ }
+ else
+ {
+ copy_state_nonatomdata(state_local, state);
+ }
+
+ if (MASTER(cr))
+ {
+ /* There will be only one swap cycle with standard replica
+ * exchange, but there may be multiple swap cycles if we
+ * allow multiple swaps. */
+
+ for (j = 0; j < maxswap; j++)
+ {
+ exchange_partner = re->order[replica_id][j];
+
+ if (exchange_partner != replica_id)
+ {
+ /* Exchange the global states between the master nodes */
+ if (debug)
+ {
+ fprintf(debug, "Exchanging %d with %d\n", replica_id, exchange_partner);
+ }
+ exchange_state(cr->ms, exchange_partner, state);
+ }
+ }
+ /* For temperature-type replica exchange, we need to scale
+ * the velocities. */
+ if (re->type == ereTEMP || re->type == ereTL)
+ {
+ scale_velocities(state, sqrt(re->q[ereTEMP][replica_id]/re->q[ereTEMP][re->destinations[replica_id]]));
+ }
+
+ }
+
+ /* With domain decomposition the global state is distributed later */
+ if (!DOMAINDECOMP(cr))
+ {
+ /* Copy the global state to the local state data structure */
+ copy_state_nonatomdata(state, state_local);
+ }
+ }
+
+ return bThisReplicaExchanged;
+}
+
+void print_replica_exchange_statistics(FILE *fplog, struct gmx_repl_ex *re)
+{
+ int i;
+
+ fprintf(fplog, "\nReplica exchange statistics\n");
+
+ if (re->nex == 0)
+ {
+ fprintf(fplog, "Repl %d attempts, %d odd, %d even\n",
+ re->nattempt[0]+re->nattempt[1], re->nattempt[1], re->nattempt[0]);
+
+ fprintf(fplog, "Repl average probabilities:\n");
+ for (i = 1; i < re->nrepl; i++)
+ {
+ if (re->nattempt[i%2] == 0)
+ {
+ re->prob[i] = 0;
+ }
+ else
+ {
+ re->prob[i] = re->prob_sum[i]/re->nattempt[i%2];
+ }
+ }
+ print_ind(fplog, "", re->nrepl, re->ind, NULL);
+ print_prob(fplog, "", re->nrepl, re->prob);
+
+ fprintf(fplog, "Repl number of exchanges:\n");
+ print_ind(fplog, "", re->nrepl, re->ind, NULL);
+ print_count(fplog, "", re->nrepl, re->nexchange);
+
+ fprintf(fplog, "Repl average number of exchanges:\n");
+ for (i = 1; i < re->nrepl; i++)
+ {
+ if (re->nattempt[i%2] == 0)
+ {
+ re->prob[i] = 0;
+ }
+ else
+ {
+ re->prob[i] = ((real)re->nexchange[i])/re->nattempt[i%2];
+ }
+ }
+ print_ind(fplog, "", re->nrepl, re->ind, NULL);
+ print_prob(fplog, "", re->nrepl, re->prob);
+
+ fprintf(fplog, "\n");
+ }
+ /* print the transition matrix */
+ print_transition_matrix(fplog, re->nrepl, re->nmoves, re->nattempt);
+}
diff --git a/patches/gromacs-2016-beta1.diff/src/programs/mdrun/repl_ex.cpp.preplumed b/patches/gromacs-2016-beta1.diff/src/programs/mdrun/repl_ex.cpp.preplumed
new file mode 100644
index 0000000000000000000000000000000000000000..ab12fc4f1789ddbc42674ed56b6268069fdc369a
--- /dev/null
+++ b/patches/gromacs-2016-beta1.diff/src/programs/mdrun/repl_ex.cpp.preplumed
@@ -0,0 +1,1399 @@
+/*
+ * This file is part of the GROMACS molecular simulation package.
+ *
+ * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
+ * Copyright (c) 2001-2004, The GROMACS development team.
+ * Copyright (c) 2011,2012,2013,2014,2015,2016, by the GROMACS development team, led by
+ * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
+ * and including many others, as listed in the AUTHORS file in the
+ * top-level source directory and at http://www.gromacs.org.
+ *
+ * GROMACS is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public License
+ * as published by the Free Software Foundation; either version 2.1
+ * of the License, or (at your option) any later version.
+ *
+ * GROMACS is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with GROMACS; if not, see
+ * http://www.gnu.org/licenses, or write to the Free Software Foundation,
+ * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ * If you want to redistribute modifications to GROMACS, 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 http://www.gromacs.org.
+ *
+ * To help us fund GROMACS development, we humbly ask that you cite
+ * the research papers on the package. Check out http://www.gromacs.org.
+ */
+
+#include "gmxpre.h"
+
+#include "repl_ex.h"
+
+#include "config.h"
+
+#include <math.h>
+
+#include <random>
+
+#include "gromacs/domdec/domdec.h"
+#include "gromacs/gmxlib/network.h"
+#include "gromacs/math/units.h"
+#include "gromacs/math/vec.h"
+#include "gromacs/mdlib/main.h"
+#include "gromacs/mdtypes/commrec.h"
+#include "gromacs/mdtypes/inputrec.h"
+#include "gromacs/mdtypes/md_enums.h"
+#include "gromacs/random/threefry.h"
+#include "gromacs/random/uniformintdistribution.h"
+#include "gromacs/random/uniformrealdistribution.h"
+#include "gromacs/utility/fatalerror.h"
+#include "gromacs/utility/pleasecite.h"
+#include "gromacs/utility/smalloc.h"
+
+
+#define PROBABILITYCUTOFF 100
+/* we don't bother evaluating if events are more rare than exp(-100) = 3.7x10^-44 */
+
+//! Rank in the multisimulaiton
+#define MSRANK(ms, nodeid) (nodeid)
+
+enum {
+ ereTEMP, ereLAMBDA, ereENDSINGLE, ereTL, ereNR
+};
+const char *erename[ereNR] = { "temperature", "lambda", "end_single_marker", "temperature and lambda"};
+/* end_single_marker merely notes the end of single variable replica exchange. All types higher than
+ it are multiple replica exchange methods */
+/* Eventually, should add 'pressure', 'temperature and pressure', 'lambda_and_pressure', 'temperature_lambda_pressure'?;
+ Let's wait until we feel better about the pressure control methods giving exact ensembles. Right now, we assume constant pressure */
+
+typedef struct gmx_repl_ex
+{
+ int repl; /* replica ID */
+ int nrepl; /* total number of replica */
+ real temp; /* temperature */
+ int type; /* replica exchange type from ere enum */
+ real **q; /* quantity, e.g. temperature or lambda; first index is ere, second index is replica ID */
+ gmx_bool bNPT; /* use constant pressure and temperature */
+ real *pres; /* replica pressures */
+ int *ind; /* replica indices */
+ int *allswaps; /* used for keeping track of all the replica swaps */
+ int nst; /* replica exchange interval (number of steps) */
+ int nex; /* number of exchanges per interval */
+ int seed; /* random seed */
+ int nattempt[2]; /* number of even and odd replica change attempts */
+ real *prob_sum; /* sum of probabilities */
+ int **nmoves; /* number of moves between replicas i and j */
+ int *nexchange; /* i-th element of the array is the number of exchanges between replica i-1 and i */
+
+ /* these are helper arrays for replica exchange; allocated here so they
+ don't have to be allocated each time */
+ int *destinations;
+ int **cyclic;
+ int **order;
+ int *tmpswap;
+ gmx_bool *incycle;
+ gmx_bool *bEx;
+
+ /* helper arrays to hold the quantities that are exchanged */
+ real *prob;
+ real *Epot;
+ real *beta;
+ real *Vol;
+ real **de;
+
+} t_gmx_repl_ex;
+
+static gmx_bool repl_quantity(const gmx_multisim_t *ms,
+ struct gmx_repl_ex *re, int ere, real q)
+{
+ real *qall;
+ gmx_bool bDiff;
+ int s;
+
+ snew(qall, ms->nsim);
+ qall[re->repl] = q;
+ gmx_sum_sim(ms->nsim, qall, ms);
+
+ bDiff = FALSE;
+ for (s = 1; s < ms->nsim; s++)
+ {
+ if (qall[s] != qall[0])
+ {
+ bDiff = TRUE;
+ }
+ }
+
+ if (bDiff)
+ {
+ /* Set the replica exchange type and quantities */
+ re->type = ere;
+
+ snew(re->q[ere], re->nrepl);
+ for (s = 0; s < ms->nsim; s++)
+ {
+ re->q[ere][s] = qall[s];
+ }
+ }
+ sfree(qall);
+ return bDiff;
+}
+
+gmx_repl_ex_t init_replica_exchange(FILE *fplog,
+ const gmx_multisim_t *ms,
+ const t_state *state,
+ const t_inputrec *ir,
+ int nst, int nex, int init_seed)
+{
+ real pres;
+ int i, j, k;
+ struct gmx_repl_ex *re;
+ gmx_bool bTemp;
+ gmx_bool bLambda = FALSE;
+
+ fprintf(fplog, "\nInitializing Replica Exchange\n");
+
+ if (ms == NULL || ms->nsim == 1)
+ {
+ gmx_fatal(FARGS, "Nothing to exchange with only one replica, maybe you forgot to set the -multi option of mdrun?");
+ }
+ if (!EI_DYNAMICS(ir->eI))
+ {
+ gmx_fatal(FARGS, "Replica exchange is only supported by dynamical simulations");
+ /* Note that PAR(cr) is defined by cr->nnodes > 1, which is
+ * distinct from MULTISIM(cr). A multi-simulation only runs
+ * with real MPI parallelism, but this does not imply PAR(cr)
+ * is true!
+ *
+ * Since we are using a dynamical integrator, the only
+ * decomposition is DD, so PAR(cr) and DOMAINDECOMP(cr) are
+ * synonymous. The only way for cr->nnodes > 1 to be true is
+ * if we are using DD. */
+ }
+
+ snew(re, 1);
+
+ re->repl = ms->sim;
+ re->nrepl = ms->nsim;
+ snew(re->q, ereENDSINGLE);
+
+ fprintf(fplog, "Repl There are %d replicas:\n", re->nrepl);
+
+ check_multi_int(fplog, ms, state->natoms, "the number of atoms", FALSE);
+ check_multi_int(fplog, ms, ir->eI, "the integrator", FALSE);
+ check_multi_int64(fplog, ms, ir->init_step+ir->nsteps, "init_step+nsteps", FALSE);
+ check_multi_int64(fplog, ms, (ir->init_step+nst-1)/nst,
+ "first exchange step: init_step/-replex", FALSE);
+ check_multi_int(fplog, ms, ir->etc, "the temperature coupling", FALSE);
+ check_multi_int(fplog, ms, ir->opts.ngtc,
+ "the number of temperature coupling groups", FALSE);
+ check_multi_int(fplog, ms, ir->epc, "the pressure coupling", FALSE);
+ check_multi_int(fplog, ms, ir->efep, "free energy", FALSE);
+ check_multi_int(fplog, ms, ir->fepvals->n_lambda, "number of lambda states", FALSE);
+
+ re->temp = ir->opts.ref_t[0];
+ for (i = 1; (i < ir->opts.ngtc); i++)
+ {
+ if (ir->opts.ref_t[i] != re->temp)
+ {
+ fprintf(fplog, "\nWARNING: The temperatures of the different temperature coupling groups are not identical\n\n");
+ fprintf(stderr, "\nWARNING: The temperatures of the different temperature coupling groups are not identical\n\n");
+ }
+ }
+
+ re->type = -1;
+ bTemp = repl_quantity(ms, re, ereTEMP, re->temp);
+ if (ir->efep != efepNO)
+ {
+ bLambda = repl_quantity(ms, re, ereLAMBDA, (real)ir->fepvals->init_fep_state);
+ }
+ if (re->type == -1) /* nothing was assigned */
+ {
+ gmx_fatal(FARGS, "The properties of the %d systems are all the same, there is nothing to exchange", re->nrepl);
+ }
+ if (bLambda && bTemp)
+ {
+ re->type = ereTL;
+ }
+
+ if (bTemp)
+ {
+ please_cite(fplog, "Sugita1999a");
+ if (ir->epc != epcNO)
+ {
+ re->bNPT = TRUE;
+ fprintf(fplog, "Repl Using Constant Pressure REMD.\n");
+ please_cite(fplog, "Okabe2001a");
+ }
+ if (ir->etc == etcBERENDSEN)
+ {
+ gmx_fatal(FARGS, "REMD with the %s thermostat does not produce correct potential energy distributions, consider using the %s thermostat instead",
+ ETCOUPLTYPE(ir->etc), ETCOUPLTYPE(etcVRESCALE));
+ }
+ }
+ if (bLambda)
+ {
+ if (ir->fepvals->delta_lambda != 0) /* check this? */
+ {
+ gmx_fatal(FARGS, "delta_lambda is not zero");
+ }
+ }
+ if (re->bNPT)
+ {
+ snew(re->pres, re->nrepl);
+ if (ir->epct == epctSURFACETENSION)
+ {
+ pres = ir->ref_p[ZZ][ZZ];
+ }
+ else
+ {
+ pres = 0;
+ j = 0;
+ for (i = 0; i < DIM; i++)
+ {
+ if (ir->compress[i][i] != 0)
+ {
+ pres += ir->ref_p[i][i];
+ j++;
+ }
+ }
+ pres /= j;
+ }
+ re->pres[re->repl] = pres;
+ gmx_sum_sim(re->nrepl, re->pres, ms);
+ }
+
+ /* Make an index for increasing replica order */
+ /* only makes sense if one or the other is varying, not both!
+ if both are varying, we trust the order the person gave. */
+ snew(re->ind, re->nrepl);
+ for (i = 0; i < re->nrepl; i++)
+ {
+ re->ind[i] = i;
+ }
+
+ if (re->type < ereENDSINGLE)
+ {
+
+ for (i = 0; i < re->nrepl; i++)
+ {
+ for (j = i+1; j < re->nrepl; j++)
+ {
+ if (re->q[re->type][re->ind[j]] < re->q[re->type][re->ind[i]])
+ {
+ /* Unordered replicas are supposed to work, but there
+ * is still an issues somewhere.
+ * Note that at this point still re->ind[i]=i.
+ */
+ gmx_fatal(FARGS, "Replicas with indices %d < %d have %ss %g > %g, please order your replicas on increasing %s",
+ i, j,
+ erename[re->type],
+ re->q[re->type][i], re->q[re->type][j],
+ erename[re->type]);
+
+ k = re->ind[i];
+ re->ind[i] = re->ind[j];
+ re->ind[j] = k;
+ }
+ else if (re->q[re->type][re->ind[j]] == re->q[re->type][re->ind[i]])
+ {
+ gmx_fatal(FARGS, "Two replicas have identical %ss", erename[re->type]);
+ }
+ }
+ }
+ }
+
+ /* keep track of all the swaps, starting with the initial placement. */
+ snew(re->allswaps, re->nrepl);
+ for (i = 0; i < re->nrepl; i++)
+ {
+ re->allswaps[i] = re->ind[i];
+ }
+
+ switch (re->type)
+ {
+ case ereTEMP:
+ fprintf(fplog, "\nReplica exchange in temperature\n");
+ for (i = 0; i < re->nrepl; i++)
+ {
+ fprintf(fplog, " %5.1f", re->q[re->type][re->ind[i]]);
+ }
+ fprintf(fplog, "\n");
+ break;
+ case ereLAMBDA:
+ fprintf(fplog, "\nReplica exchange in lambda\n");
+ for (i = 0; i < re->nrepl; i++)
+ {
+ fprintf(fplog, " %3d", (int)re->q[re->type][re->ind[i]]);
+ }
+ fprintf(fplog, "\n");
+ break;
+ case ereTL:
+ fprintf(fplog, "\nReplica exchange in temperature and lambda state\n");
+ for (i = 0; i < re->nrepl; i++)
+ {
+ fprintf(fplog, " %5.1f", re->q[ereTEMP][re->ind[i]]);
+ }
+ fprintf(fplog, "\n");
+ for (i = 0; i < re->nrepl; i++)
+ {
+ fprintf(fplog, " %5d", (int)re->q[ereLAMBDA][re->ind[i]]);
+ }
+ fprintf(fplog, "\n");
+ break;
+ default:
+ gmx_incons("Unknown replica exchange quantity");
+ }
+ if (re->bNPT)
+ {
+ fprintf(fplog, "\nRepl p");
+ for (i = 0; i < re->nrepl; i++)
+ {
+ fprintf(fplog, " %5.2f", re->pres[re->ind[i]]);
+ }
+
+ for (i = 0; i < re->nrepl; i++)
+ {
+ if ((i > 0) && (re->pres[re->ind[i]] < re->pres[re->ind[i-1]]))
+ {
+ fprintf(fplog, "\nWARNING: The reference pressures decrease with increasing temperatures\n\n");
+ fprintf(stderr, "\nWARNING: The reference pressures decrease with increasing temperatures\n\n");
+ }
+ }
+ }
+ re->nst = nst;
+ if (init_seed == -1)
+ {
+ if (MASTERSIM(ms))
+ {
+ re->seed = static_cast<int>(gmx::makeRandomSeed());
+ }
+ else
+ {
+ re->seed = 0;
+ }
+ gmx_sumi_sim(1, &(re->seed), ms);
+ }
+ else
+ {
+ re->seed = init_seed;
+ }
+ fprintf(fplog, "\nReplica exchange interval: %d\n", re->nst);
+ fprintf(fplog, "\nReplica random seed: %d\n", re->seed);
+
+ re->nattempt[0] = 0;
+ re->nattempt[1] = 0;
+
+ snew(re->prob_sum, re->nrepl);
+ snew(re->nexchange, re->nrepl);
+ snew(re->nmoves, re->nrepl);
+ for (i = 0; i < re->nrepl; i++)
+ {
+ snew(re->nmoves[i], re->nrepl);
+ }
+ fprintf(fplog, "Replica exchange information below: ex and x = exchange, pr = probability\n");
+
+ /* generate space for the helper functions so we don't have to snew each time */
+
+ snew(re->destinations, re->nrepl);
+ snew(re->incycle, re->nrepl);
+ snew(re->tmpswap, re->nrepl);
+ snew(re->cyclic, re->nrepl);
+ snew(re->order, re->nrepl);
+ for (i = 0; i < re->nrepl; i++)
+ {
+ snew(re->cyclic[i], re->nrepl+1);
+ snew(re->order[i], re->nrepl);
+ }
+ /* allocate space for the functions storing the data for the replicas */
+ /* not all of these arrays needed in all cases, but they don't take
+ up much space, since the max size is nrepl**2 */
+ snew(re->prob, re->nrepl);
+ snew(re->bEx, re->nrepl);
+ snew(re->beta, re->nrepl);
+ snew(re->Vol, re->nrepl);
+ snew(re->Epot, re->nrepl);
+ snew(re->de, re->nrepl);
+ for (i = 0; i < re->nrepl; i++)
+ {
+ snew(re->de[i], re->nrepl);
+ }
+ re->nex = nex;
+ return re;
+}
+
+static void exchange_reals(const gmx_multisim_t gmx_unused *ms, int gmx_unused b, real *v, int n)
+{
+ real *buf;
+ int i;
+
+ if (v)
+ {
+ snew(buf, n);
+#if GMX_MPI
+ /*
+ MPI_Sendrecv(v, n*sizeof(real),MPI_BYTE,MSRANK(ms,b),0,
+ buf,n*sizeof(real),MPI_BYTE,MSRANK(ms,b),0,
+ ms->mpi_comm_masters,MPI_STATUS_IGNORE);
+ */
+ {
+ MPI_Request mpi_req;
+
+ MPI_Isend(v, n*sizeof(real), MPI_BYTE, MSRANK(ms, b), 0,
+ ms->mpi_comm_masters, &mpi_req);
+ MPI_Recv(buf, n*sizeof(real), MPI_BYTE, MSRANK(ms, b), 0,
+ ms->mpi_comm_masters, MPI_STATUS_IGNORE);
+ MPI_Wait(&mpi_req, MPI_STATUS_IGNORE);
+ }
+#endif
+ for (i = 0; i < n; i++)
+ {
+ v[i] = buf[i];
+ }
+ sfree(buf);
+ }
+}
+
+
+static void exchange_doubles(const gmx_multisim_t gmx_unused *ms, int gmx_unused b, double *v, int n)
+{
+ double *buf;
+ int i;
+
+ if (v)
+ {
+ snew(buf, n);
+#if GMX_MPI
+ /*
+ MPI_Sendrecv(v, n*sizeof(double),MPI_BYTE,MSRANK(ms,b),0,
+ buf,n*sizeof(double),MPI_BYTE,MSRANK(ms,b),0,
+ ms->mpi_comm_masters,MPI_STATUS_IGNORE);
+ */
+ {
+ MPI_Request mpi_req;
+
+ MPI_Isend(v, n*sizeof(double), MPI_BYTE, MSRANK(ms, b), 0,
+ ms->mpi_comm_masters, &mpi_req);
+ MPI_Recv(buf, n*sizeof(double), MPI_BYTE, MSRANK(ms, b), 0,
+ ms->mpi_comm_masters, MPI_STATUS_IGNORE);
+ MPI_Wait(&mpi_req, MPI_STATUS_IGNORE);
+ }
+#endif
+ for (i = 0; i < n; i++)
+ {
+ v[i] = buf[i];
+ }
+ sfree(buf);
+ }
+}
+
+static void exchange_rvecs(const gmx_multisim_t gmx_unused *ms, int gmx_unused b, rvec *v, int n)
+{
+ rvec *buf;
+ int i;
+
+ if (v)
+ {
+ snew(buf, n);
+#if GMX_MPI
+ /*
+ MPI_Sendrecv(v[0], n*sizeof(rvec),MPI_BYTE,MSRANK(ms,b),0,
+ buf[0],n*sizeof(rvec),MPI_BYTE,MSRANK(ms,b),0,
+ ms->mpi_comm_masters,MPI_STATUS_IGNORE);
+ */
+ {
+ MPI_Request mpi_req;
+
+ MPI_Isend(v[0], n*sizeof(rvec), MPI_BYTE, MSRANK(ms, b), 0,
+ ms->mpi_comm_masters, &mpi_req);
+ MPI_Recv(buf[0], n*sizeof(rvec), MPI_BYTE, MSRANK(ms, b), 0,
+ ms->mpi_comm_masters, MPI_STATUS_IGNORE);
+ MPI_Wait(&mpi_req, MPI_STATUS_IGNORE);
+ }
+#endif
+ for (i = 0; i < n; i++)
+ {
+ copy_rvec(buf[i], v[i]);
+ }
+ sfree(buf);
+ }
+}
+
+static void exchange_state(const gmx_multisim_t *ms, int b, t_state *state)
+{
+ /* When t_state changes, this code should be updated. */
+ int ngtc, nnhpres;
+ ngtc = state->ngtc * state->nhchainlength;
+ nnhpres = state->nnhpres* state->nhchainlength;
+ exchange_rvecs(ms, b, state->box, DIM);
+ exchange_rvecs(ms, b, state->box_rel, DIM);
+ exchange_rvecs(ms, b, state->boxv, DIM);
+ exchange_reals(ms, b, &(state->veta), 1);
+ exchange_reals(ms, b, &(state->vol0), 1);
+ exchange_rvecs(ms, b, state->svir_prev, DIM);
+ exchange_rvecs(ms, b, state->fvir_prev, DIM);
+ exchange_rvecs(ms, b, state->pres_prev, DIM);
+ exchange_doubles(ms, b, state->nosehoover_xi, ngtc);
+ exchange_doubles(ms, b, state->nosehoover_vxi, ngtc);
+ exchange_doubles(ms, b, state->nhpres_xi, nnhpres);
+ exchange_doubles(ms, b, state->nhpres_vxi, nnhpres);
+ exchange_doubles(ms, b, state->therm_integral, state->ngtc);
+ exchange_rvecs(ms, b, state->x, state->natoms);
+ exchange_rvecs(ms, b, state->v, state->natoms);
+}
+
+static void copy_rvecs(rvec *s, rvec *d, int n)
+{
+ int i;
+
+ if (d != NULL)
+ {
+ for (i = 0; i < n; i++)
+ {
+ copy_rvec(s[i], d[i]);
+ }
+ }
+}
+
+static void copy_doubles(const double *s, double *d, int n)
+{
+ int i;
+
+ if (d != NULL)
+ {
+ for (i = 0; i < n; i++)
+ {
+ d[i] = s[i];
+ }
+ }
+}
+
+static void copy_reals(const real *s, real *d, int n)
+{
+ int i;
+
+ if (d != NULL)
+ {
+ for (i = 0; i < n; i++)
+ {
+ d[i] = s[i];
+ }
+ }
+}
+
+static void copy_ints(const int *s, int *d, int n)
+{
+ int i;
+
+ if (d != NULL)
+ {
+ for (i = 0; i < n; i++)
+ {
+ d[i] = s[i];
+ }
+ }
+}
+
+#define scopy_rvecs(v, n) copy_rvecs(state->v, state_local->v, n);
+#define scopy_doubles(v, n) copy_doubles(state->v, state_local->v, n);
+#define scopy_reals(v, n) copy_reals(state->v, state_local->v, n);
+#define scopy_ints(v, n) copy_ints(state->v, state_local->v, n);
+
+static void copy_state_nonatomdata(t_state *state, t_state *state_local)
+{
+ /* When t_state changes, this code should be updated. */
+ int ngtc, nnhpres;
+ ngtc = state->ngtc * state->nhchainlength;
+ nnhpres = state->nnhpres* state->nhchainlength;
+ scopy_rvecs(box, DIM);
+ scopy_rvecs(box_rel, DIM);
+ scopy_rvecs(boxv, DIM);
+ state_local->veta = state->veta;
+ state_local->vol0 = state->vol0;
+ scopy_rvecs(svir_prev, DIM);
+ scopy_rvecs(fvir_prev, DIM);
+ scopy_rvecs(pres_prev, DIM);
+ scopy_doubles(nosehoover_xi, ngtc);
+ scopy_doubles(nosehoover_vxi, ngtc);
+ scopy_doubles(nhpres_xi, nnhpres);
+ scopy_doubles(nhpres_vxi, nnhpres);
+ scopy_doubles(therm_integral, state->ngtc);
+ scopy_rvecs(x, state->natoms);
+ scopy_rvecs(v, state->natoms);
+ copy_ints(&(state->fep_state), &(state_local->fep_state), 1);
+ scopy_reals(lambda, efptNR);
+}
+
+static void scale_velocities(t_state *state, real fac)
+{
+ int i;
+
+ if (state->v)
+ {
+ for (i = 0; i < state->natoms; i++)
+ {
+ svmul(fac, state->v[i], state->v[i]);
+ }
+ }
+}
+
+static void print_transition_matrix(FILE *fplog, int n, int **nmoves, int *nattempt)
+{
+ int i, j, ntot;
+ float Tprint;
+
+ ntot = nattempt[0] + nattempt[1];
+ fprintf(fplog, "\n");
+ fprintf(fplog, "Repl");
+ for (i = 0; i < n; i++)
+ {
+ fprintf(fplog, " "); /* put the title closer to the center */
+ }
+ fprintf(fplog, "Empirical Transition Matrix\n");
+
+ fprintf(fplog, "Repl");
+ for (i = 0; i < n; i++)
+ {
+ fprintf(fplog, "%8d", (i+1));
+ }
+ fprintf(fplog, "\n");
+
+ for (i = 0; i < n; i++)
+ {
+ fprintf(fplog, "Repl");
+ for (j = 0; j < n; j++)
+ {
+ Tprint = 0.0;
+ if (nmoves[i][j] > 0)
+ {
+ Tprint = nmoves[i][j]/(2.0*ntot);
+ }
+ fprintf(fplog, "%8.4f", Tprint);
+ }
+ fprintf(fplog, "%3d\n", i);
+ }
+}
+
+static void print_ind(FILE *fplog, const char *leg, int n, int *ind, gmx_bool *bEx)
+{
+ int i;
+
+ fprintf(fplog, "Repl %2s %2d", leg, ind[0]);
+ for (i = 1; i < n; i++)
+ {
+ fprintf(fplog, " %c %2d", (bEx != 0 && bEx[i]) ? 'x' : ' ', ind[i]);
+ }
+ fprintf(fplog, "\n");
+}
+
+static void print_allswitchind(FILE *fplog, int n, int *pind, int *allswaps, int *tmpswap)
+{
+ int i;
+
+ for (i = 0; i < n; i++)
+ {
+ tmpswap[i] = allswaps[i];
+ }
+ for (i = 0; i < n; i++)
+ {
+ allswaps[i] = tmpswap[pind[i]];
+ }
+
+ fprintf(fplog, "\nAccepted Exchanges: ");
+ for (i = 0; i < n; i++)
+ {
+ fprintf(fplog, "%d ", pind[i]);
+ }
+ fprintf(fplog, "\n");
+
+ /* the "Order After Exchange" is the state label corresponding to the configuration that
+ started in state listed in order, i.e.
+
+ 3 0 1 2
+
+ means that the:
+ configuration starting in simulation 3 is now in simulation 0,
+ configuration starting in simulation 0 is now in simulation 1,
+ configuration starting in simulation 1 is now in simulation 2,
+ configuration starting in simulation 2 is now in simulation 3
+ */
+ fprintf(fplog, "Order After Exchange: ");
+ for (i = 0; i < n; i++)
+ {
+ fprintf(fplog, "%d ", allswaps[i]);
+ }
+ fprintf(fplog, "\n\n");
+}
+
+static void print_prob(FILE *fplog, const char *leg, int n, real *prob)
+{
+ int i;
+ char buf[8];
+
+ fprintf(fplog, "Repl %2s ", leg);
+ for (i = 1; i < n; i++)
+ {
+ if (prob[i] >= 0)
+ {
+ sprintf(buf, "%4.2f", prob[i]);
+ fprintf(fplog, " %3s", buf[0] == '1' ? "1.0" : buf+1);
+ }
+ else
+ {
+ fprintf(fplog, " ");
+ }
+ }
+ fprintf(fplog, "\n");
+}
+
+static void print_count(FILE *fplog, const char *leg, int n, int *count)
+{
+ int i;
+
+ fprintf(fplog, "Repl %2s ", leg);
+ for (i = 1; i < n; i++)
+ {
+ fprintf(fplog, " %4d", count[i]);
+ }
+ fprintf(fplog, "\n");
+}
+
+static real calc_delta(FILE *fplog, gmx_bool bPrint, struct gmx_repl_ex *re, int a, int b, int ap, int bp)
+{
+
+ real ediff, dpV, delta = 0;
+ real *Epot = re->Epot;
+ real *Vol = re->Vol;
+ real **de = re->de;
+ real *beta = re->beta;
+
+ /* Two cases; we are permuted and not. In all cases, setting ap = a and bp = b will reduce
+ to the non permuted case */
+
+ switch (re->type)
+ {
+ case ereTEMP:
+ /*
+ * Okabe et. al. Chem. Phys. Lett. 335 (2001) 435-439
+ */
+ ediff = Epot[b] - Epot[a];
+ delta = -(beta[bp] - beta[ap])*ediff;
+ break;
+ case ereLAMBDA:
+ /* two cases: when we are permuted, and not. */
+ /* non-permuted:
+ ediff = E_new - E_old
+ = [H_b(x_a) + H_a(x_b)] - [H_b(x_b) + H_a(x_a)]
+ = [H_b(x_a) - H_a(x_a)] + [H_a(x_b) - H_b(x_b)]
+ = de[b][a] + de[a][b] */
+
+ /* permuted:
+ ediff = E_new - E_old
+ = [H_bp(x_a) + H_ap(x_b)] - [H_bp(x_b) + H_ap(x_a)]
+ = [H_bp(x_a) - H_ap(x_a)] + [H_ap(x_b) - H_bp(x_b)]
+ = [H_bp(x_a) - H_a(x_a) + H_a(x_a) - H_ap(x_a)] + [H_ap(x_b) - H_b(x_b) + H_b(x_b) - H_bp(x_b)]
+ = [H_bp(x_a) - H_a(x_a)] - [H_ap(x_a) - H_a(x_a)] + [H_ap(x_b) - H_b(x_b)] - H_bp(x_b) - H_b(x_b)]
+ = (de[bp][a] - de[ap][a]) + (de[ap][b] - de[bp][b]) */
+ /* but, in the current code implementation, we flip configurations, not indices . . .
+ So let's examine that.
+ = [H_b(x_ap) - H_a(x_a)] - [H_a(x_ap) - H_a(x_a)] + [H_a(x_bp) - H_b(x_b)] - H_b(x_bp) - H_b(x_b)]
+ = [H_b(x_ap) - H_a(x_ap)] + [H_a(x_bp) - H_b(x_pb)]
+ = (de[b][ap] - de[a][ap]) + (de[a][bp] - de[b][bp]
+ So, if we exchange b<=> bp and a<=> ap, we return to the same result.
+ So the simple solution is to flip the
+ position of perturbed and original indices in the tests.
+ */
+
+ ediff = (de[bp][a] - de[ap][a]) + (de[ap][b] - de[bp][b]);
+ delta = ediff*beta[a]; /* assume all same temperature in this case */
+ break;
+ case ereTL:
+ /* not permuted: */
+ /* delta = reduced E_new - reduced E_old
+ = [beta_b H_b(x_a) + beta_a H_a(x_b)] - [beta_b H_b(x_b) + beta_a H_a(x_a)]
+ = [beta_b H_b(x_a) - beta_a H_a(x_a)] + [beta_a H_a(x_b) - beta_b H_b(x_b)]
+ = [beta_b dH_b(x_a) + beta_b H_a(x_a) - beta_a H_a(x_a)] +
+ [beta_a dH_a(x_b) + beta_a H_b(x_b) - beta_b H_b(x_b)]
+ = [beta_b dH_b(x_a) + [beta_a dH_a(x_b) +
+ beta_b (H_a(x_a) - H_b(x_b)]) - beta_a (H_a(x_a) - H_b(x_b))
+ = beta_b dH_b(x_a) + beta_a dH_a(x_b) - (beta_b - beta_a)(H_b(x_b) - H_a(x_a) */
+ /* delta = beta[b]*de[b][a] + beta[a]*de[a][b] - (beta[b] - beta[a])*(Epot[b] - Epot[a]; */
+ /* permuted (big breath!) */
+ /* delta = reduced E_new - reduced E_old
+ = [beta_bp H_bp(x_a) + beta_ap H_ap(x_b)] - [beta_bp H_bp(x_b) + beta_ap H_ap(x_a)]
+ = [beta_bp H_bp(x_a) - beta_ap H_ap(x_a)] + [beta_ap H_ap(x_b) - beta_bp H_bp(x_b)]
+ = [beta_bp H_bp(x_a) - beta_ap H_ap(x_a)] + [beta_ap H_ap(x_b) - beta_bp H_bp(x_b)]
+ - beta_pb H_a(x_a) + beta_ap H_a(x_a) + beta_pb H_a(x_a) - beta_ap H_a(x_a)
+ - beta_ap H_b(x_b) + beta_bp H_b(x_b) + beta_ap H_b(x_b) - beta_bp H_b(x_b)
+ = [(beta_bp H_bp(x_a) - beta_bp H_a(x_a)) - (beta_ap H_ap(x_a) - beta_ap H_a(x_a))] +
+ [(beta_ap H_ap(x_b) - beta_ap H_b(x_b)) - (beta_bp H_bp(x_b) - beta_bp H_b(x_b))]
+ + beta_pb H_a(x_a) - beta_ap H_a(x_a) + beta_ap H_b(x_b) - beta_bp H_b(x_b)
+ = [beta_bp (H_bp(x_a) - H_a(x_a)) - beta_ap (H_ap(x_a) - H_a(x_a))] +
+ [beta_ap (H_ap(x_b) - H_b(x_b)) - beta_bp (H_bp(x_b) - H_b(x_b))]
+ + beta_pb (H_a(x_a) - H_b(x_b)) - beta_ap (H_a(x_a) - H_b(x_b))
+ = ([beta_bp de[bp][a] - beta_ap de[ap][a]) + beta_ap de[ap][b] - beta_bp de[bp][b])
+ + (beta_pb-beta_ap)(H_a(x_a) - H_b(x_b)) */
+ delta = beta[bp]*(de[bp][a] - de[bp][b]) + beta[ap]*(de[ap][b] - de[ap][a]) - (beta[bp]-beta[ap])*(Epot[b]-Epot[a]);
+ break;
+ default:
+ gmx_incons("Unknown replica exchange quantity");
+ }
+ if (bPrint)
+ {
+ fprintf(fplog, "Repl %d <-> %d dE_term = %10.3e (kT)\n", a, b, delta);
+ }
+ if (re->bNPT)
+ {
+ /* revist the calculation for 5.0. Might be some improvements. */
+ dpV = (beta[ap]*re->pres[ap]-beta[bp]*re->pres[bp])*(Vol[b]-Vol[a])/PRESFAC;
+ if (bPrint)
+ {
+ fprintf(fplog, " dpV = %10.3e d = %10.3e\n", dpV, delta + dpV);
+ }
+ delta += dpV;
+ }
+ return delta;
+}
+
+static void
+test_for_replica_exchange(FILE *fplog,
+ const gmx_multisim_t *ms,
+ struct gmx_repl_ex *re,
+ gmx_enerdata_t *enerd,
+ real vol,
+ gmx_int64_t step,
+ real time)
+{
+ int m, i, j, a, b, ap, bp, i0, i1, tmp;
+ real delta = 0;
+ gmx_bool bPrint, bMultiEx;
+ gmx_bool *bEx = re->bEx;
+ real *prob = re->prob;
+ int *pind = re->destinations; /* permuted index */
+ gmx_bool bEpot = FALSE;
+ gmx_bool bDLambda = FALSE;
+ gmx_bool bVol = FALSE;
+ gmx::ThreeFry2x64<64> rng(re->seed, gmx::RandomDomain::ReplicaExchange);
+ gmx::UniformRealDistribution<real> uniformRealDist;
+ gmx::UniformIntDistribution<int> uniformNreplDist(0, re->nrepl-1);
+
+ bMultiEx = (re->nex > 1); /* multiple exchanges at each state */
+ fprintf(fplog, "Replica exchange at step %" GMX_PRId64 " time %.5f\n", step, time);
+
+ if (re->bNPT)
+ {
+ for (i = 0; i < re->nrepl; i++)
+ {
+ re->Vol[i] = 0;
+ }
+ bVol = TRUE;
+ re->Vol[re->repl] = vol;
+ }
+ if ((re->type == ereTEMP || re->type == ereTL))
+ {
+ for (i = 0; i < re->nrepl; i++)
+ {
+ re->Epot[i] = 0;
+ }
+ bEpot = TRUE;
+ re->Epot[re->repl] = enerd->term[F_EPOT];
+ /* temperatures of different states*/
+ for (i = 0; i < re->nrepl; i++)
+ {
+ re->beta[i] = 1.0/(re->q[ereTEMP][i]*BOLTZ);
+ }
+ }
+ else
+ {
+ for (i = 0; i < re->nrepl; i++)
+ {
+ re->beta[i] = 1.0/(re->temp*BOLTZ); /* we have a single temperature */
+ }
+ }
+ if (re->type == ereLAMBDA || re->type == ereTL)
+ {
+ bDLambda = TRUE;
+ /* lambda differences. */
+ /* de[i][j] is the energy of the jth simulation in the ith Hamiltonian
+ minus the energy of the jth simulation in the jth Hamiltonian */
+ for (i = 0; i < re->nrepl; i++)
+ {
+ for (j = 0; j < re->nrepl; j++)
+ {
+ re->de[i][j] = 0;
+ }
+ }
+ for (i = 0; i < re->nrepl; i++)
+ {
+ re->de[i][re->repl] = (enerd->enerpart_lambda[(int)re->q[ereLAMBDA][i]+1]-enerd->enerpart_lambda[0]);
+ }
+ }
+
+ /* now actually do the communication */
+ if (bVol)
+ {
+ gmx_sum_sim(re->nrepl, re->Vol, ms);
+ }
+ if (bEpot)
+ {
+ gmx_sum_sim(re->nrepl, re->Epot, ms);
+ }
+ if (bDLambda)
+ {
+ for (i = 0; i < re->nrepl; i++)
+ {
+ gmx_sum_sim(re->nrepl, re->de[i], ms);
+ }
+ }
+
+ /* make a duplicate set of indices for shuffling */
+ for (i = 0; i < re->nrepl; i++)
+ {
+ pind[i] = re->ind[i];
+ }
+
+ if (bMultiEx)
+ {
+ /* multiple random switch exchange */
+ int nself = 0;
+
+ rng.restart( step, 0 );
+
+ for (i = 0; i < re->nex + nself; i++)
+ {
+ /* randomly select a pair */
+ /* in theory, could reduce this by identifying only which switches had a nonneglibible
+ probability of occurring (log p > -100) and only operate on those switches */
+ /* find out which state it is from, and what label that state currently has. Likely
+ more work that useful. */
+ i0 = uniformNreplDist(rng);
+ i1 = uniformNreplDist(rng);
+ if (i0 == i1)
+ {
+ nself++;
+ continue; /* self-exchange, back up and do it again */
+ }
+
+ a = re->ind[i0]; /* what are the indices of these states? */
+ b = re->ind[i1];
+ ap = pind[i0];
+ bp = pind[i1];
+
+ bPrint = FALSE; /* too noisy */
+ /* calculate the energy difference */
+ /* if the code changes to flip the STATES, rather than the configurations,
+ use the commented version of the code */
+ /* delta = calc_delta(fplog,bPrint,re,a,b,ap,bp); */
+ delta = calc_delta(fplog, bPrint, re, ap, bp, a, b);
+
+ /* we actually only use the first space in the prob and bEx array,
+ since there are actually many switches between pairs. */
+
+ if (delta <= 0)
+ {
+ /* accepted */
+ prob[0] = 1;
+ bEx[0] = TRUE;
+ }
+ else
+ {
+ if (delta > PROBABILITYCUTOFF)
+ {
+ prob[0] = 0;
+ }
+ else
+ {
+ prob[0] = exp(-delta);
+ }
+ /* roll a number to determine if accepted */
+ bEx[0] = uniformRealDist(rng) < prob[0];
+ }
+ re->prob_sum[0] += prob[0];
+
+ if (bEx[0])
+ {
+ /* swap the states */
+ tmp = pind[i0];
+ pind[i0] = pind[i1];
+ pind[i1] = tmp;
+ }
+ }
+ re->nattempt[0]++; /* keep track of total permutation trials here */
+ print_allswitchind(fplog, re->nrepl, pind, re->allswaps, re->tmpswap);
+ }
+ else
+ {
+ /* standard nearest neighbor replica exchange */
+
+ m = (step / re->nst) % 2;
+ for (i = 1; i < re->nrepl; i++)
+ {
+ a = re->ind[i-1];
+ b = re->ind[i];
+
+ bPrint = (re->repl == a || re->repl == b);
+ if (i % 2 == m)
+ {
+ delta = calc_delta(fplog, bPrint, re, a, b, a, b);
+ if (delta <= 0)
+ {
+ /* accepted */
+ prob[i] = 1;
+ bEx[i] = TRUE;
+ }
+ else
+ {
+ if (delta > PROBABILITYCUTOFF)
+ {
+ prob[i] = 0;
+ }
+ else
+ {
+ prob[i] = exp(-delta);
+ }
+ /* roll a number to determine if accepted */
+ bEx[i] = uniformRealDist(rng) < prob[i];
+ }
+ re->prob_sum[i] += prob[i];
+
+ if (bEx[i])
+ {
+ /* swap these two */
+ tmp = pind[i-1];
+ pind[i-1] = pind[i];
+ pind[i] = tmp;
+ re->nexchange[i]++; /* statistics for back compatibility */
+ }
+ }
+ else
+ {
+ prob[i] = -1;
+ bEx[i] = FALSE;
+ }
+ }
+ /* print some statistics */
+ print_ind(fplog, "ex", re->nrepl, re->ind, bEx);
+ print_prob(fplog, "pr", re->nrepl, prob);
+ fprintf(fplog, "\n");
+ re->nattempt[m]++;
+ }
+
+ /* record which moves were made and accepted */
+ for (i = 0; i < re->nrepl; i++)
+ {
+ re->nmoves[re->ind[i]][pind[i]] += 1;
+ re->nmoves[pind[i]][re->ind[i]] += 1;
+ }
+ fflush(fplog); /* make sure we can see what the last exchange was */
+}
+
+static void
+cyclic_decomposition(const int *destinations,
+ int **cyclic,
+ gmx_bool *incycle,
+ const int nrepl,
+ int *nswap)
+{
+
+ int i, j, c, p;
+ int maxlen = 1;
+ for (i = 0; i < nrepl; i++)
+ {
+ incycle[i] = FALSE;
+ }
+ for (i = 0; i < nrepl; i++) /* one cycle for each replica */
+ {
+ if (incycle[i])
+ {
+ cyclic[i][0] = -1;
+ continue;
+ }
+ cyclic[i][0] = i;
+ incycle[i] = TRUE;
+ c = 1;
+ p = i;
+ for (j = 0; j < nrepl; j++) /* potentially all cycles are part, but we will break first */
+ {
+ p = destinations[p]; /* start permuting */
+ if (p == i)
+ {
+ cyclic[i][c] = -1;
+ if (c > maxlen)
+ {
+ maxlen = c;
+ }
+ break; /* we've reached the original element, the cycle is complete, and we marked the end. */
+ }
+ else
+ {
+ cyclic[i][c] = p; /* each permutation gives a new member of the cycle */
+ incycle[p] = TRUE;
+ c++;
+ }
+ }
+ }
+ *nswap = maxlen - 1;
+
+ if (debug)
+ {
+ for (i = 0; i < nrepl; i++)
+ {
+ fprintf(debug, "Cycle %d:", i);
+ for (j = 0; j < nrepl; j++)
+ {
+ if (cyclic[i][j] < 0)
+ {
+ break;
+ }
+ fprintf(debug, "%2d", cyclic[i][j]);
+ }
+ fprintf(debug, "\n");
+ }
+ fflush(debug);
+ }
+}
+
+static void
+compute_exchange_order(int **cyclic,
+ int **order,
+ const int nrepl,
+ const int maxswap)
+{
+ int i, j;
+
+ for (j = 0; j < maxswap; j++)
+ {
+ for (i = 0; i < nrepl; i++)
+ {
+ if (cyclic[i][j+1] >= 0)
+ {
+ order[cyclic[i][j+1]][j] = cyclic[i][j];
+ order[cyclic[i][j]][j] = cyclic[i][j+1];
+ }
+ }
+ for (i = 0; i < nrepl; i++)
+ {
+ if (order[i][j] < 0)
+ {
+ order[i][j] = i; /* if it's not exchanging, it should stay this round*/
+ }
+ }
+ }
+
+ if (debug)
+ {
+ fprintf(debug, "Replica Exchange Order\n");
+ for (i = 0; i < nrepl; i++)
+ {
+ fprintf(debug, "Replica %d:", i);
+ for (j = 0; j < maxswap; j++)
+ {
+ if (order[i][j] < 0)
+ {
+ break;
+ }
+ fprintf(debug, "%2d", order[i][j]);
+ }
+ fprintf(debug, "\n");
+ }
+ fflush(debug);
+ }
+}
+
+static void
+prepare_to_do_exchange(struct gmx_repl_ex *re,
+ const int replica_id,
+ int *maxswap,
+ gmx_bool *bThisReplicaExchanged)
+{
+ int i, j;
+ /* Hold the cyclic decomposition of the (multiple) replica
+ * exchange. */
+ gmx_bool bAnyReplicaExchanged = FALSE;
+ *bThisReplicaExchanged = FALSE;
+
+ for (i = 0; i < re->nrepl; i++)
+ {
+ if (re->destinations[i] != re->ind[i])
+ {
+ /* only mark as exchanged if the index has been shuffled */
+ bAnyReplicaExchanged = TRUE;
+ break;
+ }
+ }
+ if (bAnyReplicaExchanged)
+ {
+ /* reinitialize the placeholder arrays */
+ for (i = 0; i < re->nrepl; i++)
+ {
+ for (j = 0; j < re->nrepl; j++)
+ {
+ re->cyclic[i][j] = -1;
+ re->order[i][j] = -1;
+ }
+ }
+
+ /* Identify the cyclic decomposition of the permutation (very
+ * fast if neighbor replica exchange). */
+ cyclic_decomposition(re->destinations, re->cyclic, re->incycle, re->nrepl, maxswap);
+
+ /* Now translate the decomposition into a replica exchange
+ * order at each step. */
+ compute_exchange_order(re->cyclic, re->order, re->nrepl, *maxswap);
+
+ /* Did this replica do any exchange at any point? */
+ for (j = 0; j < *maxswap; j++)
+ {
+ if (replica_id != re->order[replica_id][j])
+ {
+ *bThisReplicaExchanged = TRUE;
+ break;
+ }
+ }
+ }
+}
+
+gmx_bool replica_exchange(FILE *fplog, const t_commrec *cr, struct gmx_repl_ex *re,
+ t_state *state, gmx_enerdata_t *enerd,
+ t_state *state_local, gmx_int64_t step, real time)
+{
+ int j;
+ int replica_id = 0;
+ int exchange_partner;
+ int maxswap = 0;
+ /* Number of rounds of exchanges needed to deal with any multiple
+ * exchanges. */
+ /* Where each replica ends up after the exchange attempt(s). */
+ /* The order in which multiple exchanges will occur. */
+ gmx_bool bThisReplicaExchanged = FALSE;
+
+ if (MASTER(cr))
+ {
+ replica_id = re->repl;
+ test_for_replica_exchange(fplog, cr->ms, re, enerd, det(state_local->box), step, time);
+ prepare_to_do_exchange(re, replica_id, &maxswap, &bThisReplicaExchanged);
+ }
+ /* Do intra-simulation broadcast so all processors belonging to
+ * each simulation know whether they need to participate in
+ * collecting the state. Otherwise, they might as well get on with
+ * the next thing to do. */
+ if (DOMAINDECOMP(cr))
+ {
+#if GMX_MPI
+ MPI_Bcast(&bThisReplicaExchanged, sizeof(gmx_bool), MPI_BYTE, MASTERRANK(cr),
+ cr->mpi_comm_mygroup);
+#endif
+ }
+
+ if (bThisReplicaExchanged)
+ {
+ /* Exchange the states */
+ /* Collect the global state on the master node */
+ if (DOMAINDECOMP(cr))
+ {
+ dd_collect_state(cr->dd, state_local, state);
+ }
+ else
+ {
+ copy_state_nonatomdata(state_local, state);
+ }
+
+ if (MASTER(cr))
+ {
+ /* There will be only one swap cycle with standard replica
+ * exchange, but there may be multiple swap cycles if we
+ * allow multiple swaps. */
+
+ for (j = 0; j < maxswap; j++)
+ {
+ exchange_partner = re->order[replica_id][j];
+
+ if (exchange_partner != replica_id)
+ {
+ /* Exchange the global states between the master nodes */
+ if (debug)
+ {
+ fprintf(debug, "Exchanging %d with %d\n", replica_id, exchange_partner);
+ }
+ exchange_state(cr->ms, exchange_partner, state);
+ }
+ }
+ /* For temperature-type replica exchange, we need to scale
+ * the velocities. */
+ if (re->type == ereTEMP || re->type == ereTL)
+ {
+ scale_velocities(state, sqrt(re->q[ereTEMP][replica_id]/re->q[ereTEMP][re->destinations[replica_id]]));
+ }
+
+ }
+
+ /* With domain decomposition the global state is distributed later */
+ if (!DOMAINDECOMP(cr))
+ {
+ /* Copy the global state to the local state data structure */
+ copy_state_nonatomdata(state, state_local);
+ }
+ }
+
+ return bThisReplicaExchanged;
+}
+
+void print_replica_exchange_statistics(FILE *fplog, struct gmx_repl_ex *re)
+{
+ int i;
+
+ fprintf(fplog, "\nReplica exchange statistics\n");
+
+ if (re->nex == 0)
+ {
+ fprintf(fplog, "Repl %d attempts, %d odd, %d even\n",
+ re->nattempt[0]+re->nattempt[1], re->nattempt[1], re->nattempt[0]);
+
+ fprintf(fplog, "Repl average probabilities:\n");
+ for (i = 1; i < re->nrepl; i++)
+ {
+ if (re->nattempt[i%2] == 0)
+ {
+ re->prob[i] = 0;
+ }
+ else
+ {
+ re->prob[i] = re->prob_sum[i]/re->nattempt[i%2];
+ }
+ }
+ print_ind(fplog, "", re->nrepl, re->ind, NULL);
+ print_prob(fplog, "", re->nrepl, re->prob);
+
+ fprintf(fplog, "Repl number of exchanges:\n");
+ print_ind(fplog, "", re->nrepl, re->ind, NULL);
+ print_count(fplog, "", re->nrepl, re->nexchange);
+
+ fprintf(fplog, "Repl average number of exchanges:\n");
+ for (i = 1; i < re->nrepl; i++)
+ {
+ if (re->nattempt[i%2] == 0)
+ {
+ re->prob[i] = 0;
+ }
+ else
+ {
+ re->prob[i] = ((real)re->nexchange[i])/re->nattempt[i%2];
+ }
+ }
+ print_ind(fplog, "", re->nrepl, re->ind, NULL);
+ print_prob(fplog, "", re->nrepl, re->prob);
+
+ fprintf(fplog, "\n");
+ }
+ /* print the transition matrix */
+ print_transition_matrix(fplog, re->nrepl, re->nmoves, re->nattempt);
+}
diff --git a/patches/gromacs-2016-beta1.diff/src/programs/mdrun/runner.cpp b/patches/gromacs-2016-beta1.diff/src/programs/mdrun/runner.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..d642a42e20b80cc6b6432ed38f566d72b16012ca
--- /dev/null
+++ b/patches/gromacs-2016-beta1.diff/src/programs/mdrun/runner.cpp
@@ -0,0 +1,1398 @@
+/*
+ * This file is part of the GROMACS molecular simulation package.
+ *
+ * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
+ * Copyright (c) 2001-2004, The GROMACS development team.
+ * Copyright (c) 2011,2012,2013,2014,2015,2016, by the GROMACS development team, led by
+ * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
+ * and including many others, as listed in the AUTHORS file in the
+ * top-level source directory and at http://www.gromacs.org.
+ *
+ * GROMACS is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public License
+ * as published by the Free Software Foundation; either version 2.1
+ * of the License, or (at your option) any later version.
+ *
+ * GROMACS is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with GROMACS; if not, see
+ * http://www.gnu.org/licenses, or write to the Free Software Foundation,
+ * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ * If you want to redistribute modifications to GROMACS, 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 http://www.gromacs.org.
+ *
+ * To help us fund GROMACS development, we humbly ask that you cite
+ * the research papers on the package. Check out http://www.gromacs.org.
+ */
+/*! \internal \file
+ *
+ * \brief Implements the MD runner routine calling all integrators.
+ *
+ * \author David van der Spoel <david.vanderspoel@icm.uu.se>
+ * \ingroup module_mdlib
+ */
+#include "gmxpre.h"
+
+#include "runner.h"
+
+#include "config.h"
+
+#include <assert.h>
+#include <signal.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include <algorithm>
+
+#include "gromacs/commandline/filenm.h"
+#include "gromacs/domdec/domdec.h"
+#include "gromacs/domdec/domdec_struct.h"
+#include "gromacs/essentialdynamics/edsam.h"
+#include "gromacs/ewald/pme.h"
+#include "gromacs/fileio/checkpoint.h"
+#include "gromacs/fileio/oenv.h"
+#include "gromacs/fileio/tpxio.h"
+#include "gromacs/gmxlib/md_logging.h"
+#include "gromacs/gmxlib/network.h"
+#include "gromacs/gpu_utils/gpu_utils.h"
+#include "gromacs/hardware/cpuinfo.h"
+#include "gromacs/hardware/detecthardware.h"
+#include "gromacs/listed-forces/disre.h"
+#include "gromacs/listed-forces/orires.h"
+#include "gromacs/math/calculate-ewald-splitting-coefficient.h"
+#include "gromacs/math/functions.h"
+#include "gromacs/math/utilities.h"
+#include "gromacs/math/vec.h"
+#include "gromacs/mdlib/calc_verletbuf.h"
+#include "gromacs/mdlib/constr.h"
+#include "gromacs/mdlib/force.h"
+#include "gromacs/mdlib/forcerec.h"
+#include "gromacs/mdlib/gmx_omp_nthreads.h"
+#include "gromacs/mdlib/integrator.h"
+#include "gromacs/mdlib/main.h"
+#include "gromacs/mdlib/md_support.h"
+#include "gromacs/mdlib/mdatoms.h"
+#include "gromacs/mdlib/mdrun.h"
+#include "gromacs/mdlib/minimize.h"
+#include "gromacs/mdlib/nbnxn_search.h"
+#include "gromacs/mdlib/qmmm.h"
+#include "gromacs/mdlib/sighandler.h"
+#include "gromacs/mdlib/sim_util.h"
+#include "gromacs/mdlib/tpi.h"
+#include "gromacs/mdrunutility/threadaffinity.h"
+#include "gromacs/mdtypes/commrec.h"
+#include "gromacs/mdtypes/inputrec.h"
+#include "gromacs/mdtypes/md_enums.h"
+#include "gromacs/mdtypes/state.h"
+#include "gromacs/pbcutil/pbc.h"
+#include "gromacs/pulling/pull.h"
+#include "gromacs/pulling/pull_rotation.h"
+#include "gromacs/timing/wallcycle.h"
+#include "gromacs/topology/mtop_util.h"
+#include "gromacs/trajectory/trajectoryframe.h"
+#include "gromacs/utility/cstringutil.h"
+#include "gromacs/utility/exceptions.h"
+#include "gromacs/utility/fatalerror.h"
+#include "gromacs/utility/gmxassert.h"
+#include "gromacs/utility/gmxmpi.h"
+#include "gromacs/utility/pleasecite.h"
+#include "gromacs/utility/smalloc.h"
+
+#include "deform.h"
+#include "md.h"
+#include "repl_ex.h"
+#include "resource-division.h"
+
+/* PLUMED */
+#include "../../../Plumed.h"
+extern int plumedswitch;
+extern plumed plumedmain;
+/* END PLUMED */
+
+#ifdef GMX_FAHCORE
+#include "corewrap.h"
+#endif
+
+//! First step used in pressure scaling
+gmx_int64_t deform_init_init_step_tpx;
+//! Initial box for pressure scaling
+matrix deform_init_box_tpx;
+//! MPI variable for use in pressure scaling
+tMPI_Thread_mutex_t deform_init_box_mutex = TMPI_THREAD_MUTEX_INITIALIZER;
+
+#if GMX_THREAD_MPI
+/* The minimum number of atoms per tMPI thread. With fewer atoms than this,
+ * the number of threads will get lowered.
+ */
+#define MIN_ATOMS_PER_MPI_THREAD 90
+#define MIN_ATOMS_PER_GPU 900
+
+struct mdrunner_arglist
+{
+ gmx_hw_opt_t hw_opt;
+ FILE *fplog;
+ t_commrec *cr;
+ int nfile;
+ const t_filenm *fnm;
+ const gmx_output_env_t *oenv;
+ gmx_bool bVerbose;
+ int nstglobalcomm;
+ ivec ddxyz;
+ int dd_rank_order;
+ int npme;
+ real rdd;
+ real rconstr;
+ const char *dddlb_opt;
+ real dlb_scale;
+ const char *ddcsx;
+ const char *ddcsy;
+ const char *ddcsz;
+ const char *nbpu_opt;
+ int nstlist_cmdline;
+ gmx_int64_t nsteps_cmdline;
+ int nstepout;
+ int resetstep;
+ int nmultisim;
+ int repl_ex_nst;
+ int repl_ex_nex;
+ int repl_ex_seed;
+ real pforce;
+ real cpt_period;
+ real max_hours;
+ int imdport;
+ unsigned long Flags;
+};
+
+
+/* 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)
+{
+ try
+ {
+ 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 = reinitialize_commrec_for_this_thread(mc.cr);
+
+ if (MASTER(cr))
+ {
+ fplog = mc.fplog;
+ }
+
+ gmx::mdrunner(&mc.hw_opt, fplog, cr, mc.nfile, fnm, mc.oenv,
+ mc.bVerbose, mc.nstglobalcomm,
+ mc.ddxyz, mc.dd_rank_order, mc.npme, mc.rdd,
+ mc.rconstr, mc.dddlb_opt, mc.dlb_scale,
+ mc.ddcsx, mc.ddcsy, mc.ddcsz,
+ mc.nbpu_opt, mc.nstlist_cmdline,
+ mc.nsteps_cmdline, mc.nstepout, mc.resetstep,
+ mc.nmultisim, mc.repl_ex_nst, mc.repl_ex_nex, mc.repl_ex_seed, mc.pforce,
+ mc.cpt_period, mc.max_hours, mc.imdport, mc.Flags);
+ }
+ GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+}
+
+
+/* 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(gmx_hw_opt_t *hw_opt,
+ FILE *fplog, t_commrec *cr, int nfile,
+ const t_filenm fnm[], const gmx_output_env_t *oenv, gmx_bool bVerbose,
+ int nstglobalcomm,
+ ivec ddxyz, int dd_rank_order, int npme,
+ real rdd, real rconstr,
+ const char *dddlb_opt, real dlb_scale,
+ const char *ddcsx, const char *ddcsy, const char *ddcsz,
+ const char *nbpu_opt, int nstlist_cmdline,
+ gmx_int64_t nsteps_cmdline,
+ int nstepout, int resetstep,
+ int nmultisim, int repl_ex_nst, int repl_ex_nex, int repl_ex_seed,
+ real pforce, real cpt_period, real max_hours,
+ 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 (hw_opt->nthreads_tmpi < 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 */
+ /* hw_opt contains pointers, which should all be NULL at this stage */
+ mda->hw_opt = *hw_opt;
+ mda->fplog = fplog;
+ mda->cr = cr;
+ mda->nfile = nfile;
+ mda->fnm = fnmn;
+ mda->oenv = oenv;
+ mda->bVerbose = bVerbose;
+ mda->nstglobalcomm = nstglobalcomm;
+ mda->ddxyz[XX] = ddxyz[XX];
+ mda->ddxyz[YY] = ddxyz[YY];
+ mda->ddxyz[ZZ] = ddxyz[ZZ];
+ mda->dd_rank_order = dd_rank_order;
+ mda->npme = npme;
+ 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->nbpu_opt = nbpu_opt;
+ mda->nstlist_cmdline = nstlist_cmdline;
+ mda->nsteps_cmdline = nsteps_cmdline;
+ mda->nstepout = nstepout;
+ mda->resetstep = resetstep;
+ mda->nmultisim = nmultisim;
+ mda->repl_ex_nst = repl_ex_nst;
+ mda->repl_ex_nex = repl_ex_nex;
+ mda->repl_ex_seed = repl_ex_seed;
+ mda->pforce = pforce;
+ mda->cpt_period = cpt_period;
+ mda->max_hours = max_hours;
+ mda->Flags = Flags;
+
+ /* now spawn new threads that start mdrunner_start_fn(), while
+ the main thread returns, we set thread affinity later */
+ ret = tMPI_Init_fn(TRUE, hw_opt->nthreads_tmpi, TMPI_AFFINITY_NONE,
+ mdrunner_start_fn, (void*)(mda) );
+ if (ret != TMPI_SUCCESS)
+ {
+ return NULL;
+ }
+
+ crn = reinitialize_commrec_for_this_thread(cr);
+ return crn;
+}
+
+#endif /* GMX_THREAD_MPI */
+
+
+/*! \brief Cost of non-bonded kernels
+ *
+ * We determine the extra cost of the non-bonded kernels compared to
+ * a reference nstlist value of 10 (which is the default in grompp).
+ */
+static const int nbnxnReferenceNstlist = 10;
+//! The values to try when switching
+const int nstlist_try[] = { 20, 25, 40 };
+//! Number of elements in the neighborsearch list trials.
+#define NNSTL sizeof(nstlist_try)/sizeof(nstlist_try[0])
+/* Increase nstlist until the non-bonded cost increases more than listfac_ok,
+ * but never more than listfac_max.
+ * A standard (protein+)water system at 300K with PME ewald_rtol=1e-5
+ * needs 1.28 at rcoulomb=0.9 and 1.24 at rcoulomb=1.0 to get to nstlist=40.
+ * Note that both CPU and GPU factors are conservative. Performance should
+ * not go down due to this tuning, except with a relatively slow GPU.
+ * On the other hand, at medium/high parallelization or with fast GPUs
+ * nstlist will not be increased enough to reach optimal performance.
+ */
+/* CPU: pair-search is about a factor 1.5 slower than the non-bonded kernel */
+//! Max OK performance ratio beween force calc and neighbor searching
+static const float nbnxn_cpu_listfac_ok = 1.05;
+//! Too high performance ratio beween force calc and neighbor searching
+static const float nbnxn_cpu_listfac_max = 1.09;
+/* CPU: pair-search is about a factor 2-3 slower than the non-bonded kernel */
+//! Max OK performance ratio beween force calc and neighbor searching
+static const float nbnxn_knl_listfac_ok = 1.22;
+//! Too high performance ratio beween force calc and neighbor searching
+static const float nbnxn_knl_listfac_max = 1.3;
+/* GPU: pair-search is a factor 1.5-3 slower than the non-bonded kernel */
+//! Max OK performance ratio beween force calc and neighbor searching
+static const float nbnxn_gpu_listfac_ok = 1.20;
+//! Too high performance ratio beween force calc and neighbor searching
+static const float nbnxn_gpu_listfac_max = 1.30;
+
+/*! \brief Try to increase nstlist when using the Verlet cut-off scheme */
+static void increase_nstlist(FILE *fp, t_commrec *cr,
+ t_inputrec *ir, int nstlist_cmdline,
+ const gmx_mtop_t *mtop, matrix box,
+ gmx_bool bGPU, const gmx::CpuInfo &cpuinfo)
+{
+ float listfac_ok, listfac_max;
+ int nstlist_orig, nstlist_prev;
+ verletbuf_list_setup_t ls;
+ real rlistWithReferenceNstlist, rlist_inc, rlist_ok, rlist_max;
+ real rlist_new, rlist_prev;
+ size_t nstlist_ind = 0;
+ t_state state_tmp;
+ gmx_bool bBox, bDD, bCont;
+ const char *nstl_gpu = "\nFor optimal performance with a GPU nstlist (now %d) should be larger.\nThe optimum depends on your CPU and GPU resources.\nYou might want to try several nstlist values.\n";
+ const char *nve_err = "Can not increase nstlist because an NVE ensemble is used";
+ const char *vbd_err = "Can not increase nstlist because verlet-buffer-tolerance is not set or used";
+ const char *box_err = "Can not increase nstlist because the box is too small";
+ const char *dd_err = "Can not increase nstlist because of domain decomposition limitations";
+ char buf[STRLEN];
+
+ if (nstlist_cmdline <= 0)
+ {
+ if (ir->nstlist == 1)
+ {
+ /* The user probably set nstlist=1 for a reason,
+ * don't mess with the settings.
+ */
+ return;
+ }
+
+ if (fp != NULL && bGPU && ir->nstlist < nstlist_try[0])
+ {
+ fprintf(fp, nstl_gpu, ir->nstlist);
+ }
+ nstlist_ind = 0;
+ while (nstlist_ind < NNSTL && ir->nstlist >= nstlist_try[nstlist_ind])
+ {
+ nstlist_ind++;
+ }
+ if (nstlist_ind == NNSTL)
+ {
+ /* There are no larger nstlist value to try */
+ return;
+ }
+ }
+
+ if (EI_MD(ir->eI) && ir->etc == etcNO)
+ {
+ if (MASTER(cr))
+ {
+ fprintf(stderr, "%s\n", nve_err);
+ }
+ if (fp != NULL)
+ {
+ fprintf(fp, "%s\n", nve_err);
+ }
+
+ return;
+ }
+
+ if (ir->verletbuf_tol == 0 && bGPU)
+ {
+ gmx_fatal(FARGS, "You are using an old tpr file with a GPU, please generate a new tpr file with an up to date version of grompp");
+ }
+
+ if (ir->verletbuf_tol < 0)
+ {
+ if (MASTER(cr))
+ {
+ fprintf(stderr, "%s\n", vbd_err);
+ }
+ if (fp != NULL)
+ {
+ fprintf(fp, "%s\n", vbd_err);
+ }
+
+ return;
+ }
+
+ if (bGPU)
+ {
+ listfac_ok = nbnxn_gpu_listfac_ok;
+ listfac_max = nbnxn_gpu_listfac_max;
+ }
+ else if (cpuinfo.feature(gmx::CpuInfo::Feature::X86_Avx512ER))
+ {
+ listfac_ok = nbnxn_knl_listfac_ok;
+ listfac_max = nbnxn_knl_listfac_max;
+ }
+ else
+ {
+ listfac_ok = nbnxn_cpu_listfac_ok;
+ listfac_max = nbnxn_cpu_listfac_max;
+ }
+
+ nstlist_orig = ir->nstlist;
+ if (nstlist_cmdline > 0)
+ {
+ if (fp)
+ {
+ sprintf(buf, "Getting nstlist=%d from command line option",
+ nstlist_cmdline);
+ }
+ ir->nstlist = nstlist_cmdline;
+ }
+
+ verletbuf_get_list_setup(TRUE, bGPU, &ls);
+
+ /* Allow rlist to make the list a given factor larger than the list
+ * would be with the reference value for nstlist (10).
+ */
+ nstlist_prev = ir->nstlist;
+ ir->nstlist = nbnxnReferenceNstlist;
+ calc_verlet_buffer_size(mtop, det(box), ir, -1, &ls, NULL,
+ &rlistWithReferenceNstlist);
+ ir->nstlist = nstlist_prev;
+
+ /* Determine the pair list size increase due to zero interactions */
+ rlist_inc = nbnxn_get_rlist_effective_inc(ls.cluster_size_j,
+ mtop->natoms/det(box));
+ rlist_ok = (rlistWithReferenceNstlist + rlist_inc)*std::cbrt(listfac_ok) - rlist_inc;
+ rlist_max = (rlistWithReferenceNstlist + rlist_inc)*std::cbrt(listfac_max) - rlist_inc;
+ if (debug)
+ {
+ fprintf(debug, "nstlist tuning: rlist_inc %.3f rlist_ok %.3f rlist_max %.3f\n",
+ rlist_inc, rlist_ok, rlist_max);
+ }
+
+ nstlist_prev = nstlist_orig;
+ rlist_prev = ir->rlist;
+ do
+ {
+ if (nstlist_cmdline <= 0)
+ {
+ ir->nstlist = nstlist_try[nstlist_ind];
+ }
+
+ /* Set the pair-list buffer size in ir */
+ calc_verlet_buffer_size(mtop, det(box), ir, -1, &ls, NULL, &rlist_new);
+
+ /* Does rlist fit in the box? */
+ bBox = (gmx::square(rlist_new) < max_cutoff2(ir->ePBC, box));
+ bDD = TRUE;
+ if (bBox && DOMAINDECOMP(cr))
+ {
+ /* Check if rlist fits in the domain decomposition */
+ if (inputrec2nboundeddim(ir) < DIM)
+ {
+ gmx_incons("Changing nstlist with domain decomposition and unbounded dimensions is not implemented yet");
+ }
+ copy_mat(box, state_tmp.box);
+ bDD = change_dd_cutoff(cr, &state_tmp, ir, rlist_new);
+ }
+
+ if (debug)
+ {
+ fprintf(debug, "nstlist %d rlist %.3f bBox %d bDD %d\n",
+ ir->nstlist, rlist_new, bBox, bDD);
+ }
+
+ bCont = FALSE;
+
+ if (nstlist_cmdline <= 0)
+ {
+ if (bBox && bDD && rlist_new <= rlist_max)
+ {
+ /* Increase nstlist */
+ nstlist_prev = ir->nstlist;
+ rlist_prev = rlist_new;
+ bCont = (nstlist_ind+1 < NNSTL && rlist_new < rlist_ok);
+ }
+ else
+ {
+ /* Stick with the previous nstlist */
+ ir->nstlist = nstlist_prev;
+ rlist_new = rlist_prev;
+ bBox = TRUE;
+ bDD = TRUE;
+ }
+ }
+
+ nstlist_ind++;
+ }
+ while (bCont);
+
+ if (!bBox || !bDD)
+ {
+ gmx_warning(!bBox ? box_err : dd_err);
+ if (fp != NULL)
+ {
+ fprintf(fp, "\n%s\n", bBox ? box_err : dd_err);
+ }
+ ir->nstlist = nstlist_orig;
+ }
+ else if (ir->nstlist != nstlist_orig || rlist_new != ir->rlist)
+ {
+ sprintf(buf, "Changing nstlist from %d to %d, rlist from %g to %g",
+ nstlist_orig, ir->nstlist,
+ ir->rlist, rlist_new);
+ if (MASTER(cr))
+ {
+ fprintf(stderr, "%s\n\n", buf);
+ }
+ if (fp != NULL)
+ {
+ fprintf(fp, "%s\n\n", buf);
+ }
+ ir->rlist = rlist_new;
+ }
+}
+
+/*! \brief Initialize variables for Verlet scheme simulation */
+static void prepare_verlet_scheme(FILE *fplog,
+ t_commrec *cr,
+ t_inputrec *ir,
+ int nstlist_cmdline,
+ const gmx_mtop_t *mtop,
+ matrix box,
+ gmx_bool bUseGPU,
+ const gmx::CpuInfo &cpuinfo)
+{
+ /* For NVE simulations, we will retain the initial list buffer */
+ if (EI_DYNAMICS(ir->eI) &&
+ ir->verletbuf_tol > 0 &&
+ !(EI_MD(ir->eI) && ir->etc == etcNO))
+ {
+ /* Update the Verlet buffer size for the current run setup */
+ verletbuf_list_setup_t ls;
+ real rlist_new;
+
+ /* Here we assume SIMD-enabled kernels are being used. But as currently
+ * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
+ * and 4x2 gives a larger buffer than 4x4, this is ok.
+ */
+ verletbuf_get_list_setup(TRUE, bUseGPU, &ls);
+
+ calc_verlet_buffer_size(mtop, det(box), ir, -1, &ls, NULL, &rlist_new);
+
+ if (rlist_new != ir->rlist)
+ {
+ if (fplog != NULL)
+ {
+ fprintf(fplog, "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
+ ir->rlist, rlist_new,
+ ls.cluster_size_i, ls.cluster_size_j);
+ }
+ ir->rlist = rlist_new;
+ }
+ }
+
+ if (nstlist_cmdline > 0 && (!EI_DYNAMICS(ir->eI) || ir->verletbuf_tol <= 0))
+ {
+ gmx_fatal(FARGS, "Can not set nstlist without %s",
+ !EI_DYNAMICS(ir->eI) ? "dynamics" : "verlet-buffer-tolerance");
+ }
+
+ if (EI_DYNAMICS(ir->eI))
+ {
+ /* Set or try nstlist values */
+ increase_nstlist(fplog, cr, ir, nstlist_cmdline, mtop, box, bUseGPU, cpuinfo);
+ }
+}
+
+/*! \brief Override the nslist value in inputrec
+ *
+ * with value passed on the command line (if any)
+ */
+static void override_nsteps_cmdline(FILE *fplog,
+ gmx_int64_t nsteps_cmdline,
+ t_inputrec *ir,
+ const t_commrec *cr)
+{
+ assert(ir);
+ assert(cr);
+
+ /* override with anything else than the default -2 */
+ if (nsteps_cmdline > -2)
+ {
+ char sbuf_steps[STEPSTRSIZE];
+ char sbuf_msg[STRLEN];
+
+ ir->nsteps = nsteps_cmdline;
+ if (EI_DYNAMICS(ir->eI) && nsteps_cmdline != -1)
+ {
+ sprintf(sbuf_msg, "Overriding nsteps with value passed on the command line: %s steps, %.3g ps",
+ gmx_step_str(nsteps_cmdline, sbuf_steps),
+ fabs(nsteps_cmdline*ir->delta_t));
+ }
+ else
+ {
+ sprintf(sbuf_msg, "Overriding nsteps with value passed on the command line: %s steps",
+ gmx_step_str(nsteps_cmdline, sbuf_steps));
+ }
+
+ md_print_warn(cr, fplog, "%s\n", sbuf_msg);
+ }
+ else if (nsteps_cmdline < -2)
+ {
+ gmx_fatal(FARGS, "Invalid nsteps value passed on the command line: %d",
+ nsteps_cmdline);
+ }
+ /* Do nothing if nsteps_cmdline == -2 */
+}
+
+namespace gmx
+{
+
+//! \brief Return the correct integrator function.
+static integrator_t *my_integrator(unsigned int ei)
+{
+ switch (ei)
+ {
+ case eiMD:
+ case eiBD:
+ case eiSD1:
+ case eiVV:
+ case eiVVAK:
+ if (!EI_DYNAMICS(ei))
+ {
+ GMX_THROW(APIError("do_md integrator would be called for a non-dynamical integrator"));
+ }
+ return do_md;
+ case eiSteep:
+ return do_steep;
+ case eiCG:
+ return do_cg;
+ case eiNM:
+ return do_nm;
+ case eiLBFGS:
+ return do_lbfgs;
+ case eiTPI:
+ case eiTPIC:
+ if (!EI_TPI(ei))
+ {
+ GMX_THROW(APIError("do_tpi integrator would be called for a non-TPI integrator"));
+ }
+ return do_tpi;
+ case eiSD2_REMOVED:
+ GMX_THROW(NotImplementedError("SD2 integrator has been removed"));
+ default:
+ GMX_THROW(APIError("Non existing integrator selected"));
+ }
+}
+
+int mdrunner(gmx_hw_opt_t *hw_opt,
+ FILE *fplog, t_commrec *cr, int nfile,
+ const t_filenm fnm[], const gmx_output_env_t *oenv, gmx_bool bVerbose,
+ int nstglobalcomm,
+ ivec ddxyz, int dd_rank_order, int npme, real rdd, real rconstr,
+ const char *dddlb_opt, real dlb_scale,
+ const char *ddcsx, const char *ddcsy, const char *ddcsz,
+ const char *nbpu_opt, int nstlist_cmdline,
+ gmx_int64_t nsteps_cmdline, int nstepout, int resetstep,
+ int gmx_unused nmultisim, int repl_ex_nst, int repl_ex_nex,
+ int repl_ex_seed, real pforce, real cpt_period, real max_hours,
+ int imdport, unsigned long Flags)
+{
+ gmx_bool bForceUseGPU, bTryUseGPU, bRerunMD;
+ t_inputrec *inputrec;
+ t_state *state = NULL;
+ matrix box;
+ gmx_ddbox_t ddbox = {0};
+ int npme_major, npme_minor;
+ t_nrnb *nrnb;
+ gmx_mtop_t *mtop = NULL;
+ t_mdatoms *mdatoms = NULL;
+ t_forcerec *fr = NULL;
+ t_fcdata *fcd = NULL;
+ real ewaldcoeff_q = 0;
+ real ewaldcoeff_lj = 0;
+ struct gmx_pme_t **pmedata = NULL;
+ gmx_vsite_t *vsite = NULL;
+ gmx_constr_t constr;
+ int nChargePerturbed = -1, nTypePerturbed = 0, status;
+ gmx_wallcycle_t wcycle;
+ gmx_walltime_accounting_t walltime_accounting = NULL;
+ int rc;
+ gmx_int64_t reset_counters;
+ gmx_edsam_t ed = NULL;
+ int nthreads_pme = 1;
+ gmx_hw_info_t *hwinfo = NULL;
+ /* The master rank decides early on bUseGPU and broadcasts this later */
+ gmx_bool bUseGPU = FALSE;
+
+ /* CAUTION: threads may be started later on in this function, so
+ cr doesn't reflect the final parallel state right now */
+ snew(inputrec, 1);
+ snew(mtop, 1);
+
+ if (Flags & MD_APPENDFILES)
+ {
+ fplog = NULL;
+ }
+
+ bRerunMD = (Flags & MD_RERUN);
+ bForceUseGPU = (strncmp(nbpu_opt, "gpu", 3) == 0);
+ bTryUseGPU = (strncmp(nbpu_opt, "auto", 4) == 0) || bForceUseGPU;
+
+ /* Detect hardware, gather information. This is an operation that is
+ * global for this process (MPI rank). */
+ hwinfo = gmx_detect_hardware(fplog, cr, bTryUseGPU);
+
+ gmx_print_detected_hardware(fplog, cr, hwinfo);
+
+ if (fplog != NULL)
+ {
+ /* Print references after all software/hardware printing */
+ please_cite(fplog, "Abraham2015");
+ please_cite(fplog, "Pall2015");
+ please_cite(fplog, "Pronk2013");
+ please_cite(fplog, "Hess2008b");
+ please_cite(fplog, "Spoel2005a");
+ please_cite(fplog, "Lindahl2001a");
+ please_cite(fplog, "Berendsen95a");
+ }
+
+ snew(state, 1);
+ if (SIMMASTER(cr))
+ {
+ /* Read (nearly) all data required for the simulation */
+ read_tpx_state(ftp2fn(efTPR, nfile, fnm), inputrec, state, mtop);
+
+ if (inputrec->cutoff_scheme == ecutsVERLET)
+ {
+ /* Here the master rank decides if all ranks will use GPUs */
+ bUseGPU = (hwinfo->gpu_info.n_dev_compatible > 0 ||
+ getenv("GMX_EMULATE_GPU") != NULL);
+
+ /* TODO add GPU kernels for this and replace this check by:
+ * (bUseGPU && (ir->vdwtype == evdwPME &&
+ * ir->ljpme_combination_rule == eljpmeLB))
+ * update the message text and the content of nbnxn_acceleration_supported.
+ */
+ if (bUseGPU &&
+ !nbnxn_gpu_acceleration_supported(fplog, cr, inputrec, bRerunMD))
+ {
+ /* Fallback message printed by nbnxn_acceleration_supported */
+ if (bForceUseGPU)
+ {
+ gmx_fatal(FARGS, "GPU acceleration requested, but not supported with the given input settings");
+ }
+ bUseGPU = FALSE;
+ }
+
+ prepare_verlet_scheme(fplog, cr,
+ inputrec, nstlist_cmdline, mtop, state->box,
+ bUseGPU, *hwinfo->cpuInfo);
+ }
+ else
+ {
+ if (nstlist_cmdline > 0)
+ {
+ gmx_fatal(FARGS, "Can not set nstlist with the group cut-off scheme");
+ }
+
+ if (hwinfo->gpu_info.n_dev_compatible > 0)
+ {
+ md_print_warn(cr, fplog,
+ "NOTE: GPU(s) found, but the current simulation can not use GPUs\n"
+ " To use a GPU, set the mdp option: cutoff-scheme = Verlet\n");
+ }
+
+ if (bForceUseGPU)
+ {
+ gmx_fatal(FARGS, "GPU requested, but can't be used without cutoff-scheme=Verlet");
+ }
+
+#if GMX_TARGET_BGQ
+ md_print_warn(cr, fplog,
+ "NOTE: There is no SIMD implementation of the group scheme kernels on\n"
+ " BlueGene/Q. You will observe better performance from using the\n"
+ " Verlet cut-off scheme.\n");
+#endif
+ }
+ }
+
+ /* Check and update the hardware options for internal consistency */
+ check_and_update_hw_opt_1(hw_opt, cr, npme);
+
+ /* Early check for externally set process affinity. */
+ gmx_check_thread_affinity_set(fplog, cr,
+ hw_opt, hwinfo->nthreads_hw_avail, FALSE);
+
+#if GMX_THREAD_MPI
+ if (SIMMASTER(cr))
+ {
+ if (npme > 0 && hw_opt->nthreads_tmpi <= 0)
+ {
+ gmx_fatal(FARGS, "You need to explicitly specify the number of MPI threads (-ntmpi) when using separate PME ranks");
+ }
+
+ /* Since the master knows the cut-off scheme, update hw_opt for this.
+ * This is done later for normal MPI and also once more with tMPI
+ * for all tMPI ranks.
+ */
+ check_and_update_hw_opt_2(hw_opt, inputrec->cutoff_scheme);
+
+ /* NOW the threads will be started: */
+ hw_opt->nthreads_tmpi = get_nthreads_mpi(hwinfo,
+ hw_opt,
+ inputrec, mtop,
+ cr, fplog, bUseGPU);
+
+ if (hw_opt->nthreads_tmpi > 1)
+ {
+ t_commrec *cr_old = cr;
+ /* now start the threads. */
+ cr = mdrunner_start_threads(hw_opt, fplog, cr_old, nfile, fnm,
+ oenv, bVerbose, nstglobalcomm,
+ ddxyz, dd_rank_order, npme, rdd, rconstr,
+ dddlb_opt, dlb_scale, ddcsx, ddcsy, ddcsz,
+ nbpu_opt, nstlist_cmdline,
+ nsteps_cmdline, nstepout, resetstep, nmultisim,
+ repl_ex_nst, repl_ex_nex, repl_ex_seed, pforce,
+ cpt_period, max_hours,
+ Flags);
+ /* the main thread continues here with a new cr. We don't deallocate
+ the old cr because other threads may still be reading it. */
+ if (cr == NULL)
+ {
+ gmx_comm("Failed to spawn threads");
+ }
+ }
+ }
+#endif
+ /* END OF CAUTION: cr is now reliable */
+
+ if (PAR(cr))
+ {
+ /* now broadcast everything to the non-master nodes/threads: */
+ init_parallel(cr, inputrec, mtop);
+
+ /* The master rank decided on the use of GPUs,
+ * broadcast this information to all ranks.
+ */
+ gmx_bcast_sim(sizeof(bUseGPU), &bUseGPU, cr);
+ }
+
+ if (fplog != NULL)
+ {
+ pr_inputrec(fplog, 0, "Input Parameters", inputrec, FALSE);
+ fprintf(fplog, "\n");
+ }
+
+ /* now make sure the state is initialized and propagated */
+ set_state_entries(state, inputrec);
+
+ /* A parallel command line option consistency check that we can
+ only do after any threads have started. */
+ if (!PAR(cr) &&
+ (ddxyz[XX] > 1 || ddxyz[YY] > 1 || ddxyz[ZZ] > 1 || npme > 0))
+ {
+ gmx_fatal(FARGS,
+ "The -dd or -npme option request a parallel simulation, "
+#if !GMX_MPI
+ "but %s was compiled without threads or MPI enabled"
+#else
+#if GMX_THREAD_MPI
+ "but the number of MPI-threads (option -ntmpi) is not set or is 1"
+#else
+ "but %s was not started through mpirun/mpiexec or only one rank was requested through mpirun/mpiexec"
+#endif
+#endif
+ , output_env_get_program_display_name(oenv)
+ );
+ }
+
+ if (bRerunMD &&
+ (EI_ENERGY_MINIMIZATION(inputrec->eI) || eiNM == inputrec->eI))
+ {
+ gmx_fatal(FARGS, "The .mdp file specified an energy mininization or normal mode algorithm, and these are not compatible with mdrun -rerun");
+ }
+
+ if (can_use_allvsall(inputrec, TRUE, cr, fplog) && DOMAINDECOMP(cr))
+ {
+ gmx_fatal(FARGS, "All-vs-all loops do not work with domain decomposition, use a single MPI rank");
+ }
+
+ if (!(EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype)))
+ {
+ if (npme > 0)
+ {
+ gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
+ "PME-only ranks are requested, but the system does not use PME for electrostatics or LJ");
+ }
+
+ npme = 0;
+ }
+
+ if (bUseGPU && npme < 0)
+ {
+ /* With GPUs we don't automatically use PME-only ranks. PME ranks can
+ * improve performance with many threads per GPU, since our OpenMP
+ * scaling is bad, but it's difficult to automate the setup.
+ */
+ npme = 0;
+ }
+
+#ifdef GMX_FAHCORE
+ if (MASTER(cr))
+ {
+ fcRegisterSteps(inputrec->nsteps, inputrec->init_step);
+ }
+#endif
+
+ /* NMR restraints must be initialized before load_checkpoint,
+ * since with time averaging the history is added to t_state.
+ * For proper consistency check we therefore need to extend
+ * t_state here.
+ * So the PME-only nodes (if present) will also initialize
+ * the distance restraints.
+ */
+ snew(fcd, 1);
+
+ /* This needs to be called before read_checkpoint to extend the state */
+ init_disres(fplog, mtop, inputrec, cr, fcd, state, repl_ex_nst > 0);
+
+ init_orires(fplog, mtop, state->x, inputrec, cr, &(fcd->orires),
+ state);
+
+ if (inputrecDeform(inputrec))
+ {
+ /* Store the deform reference box before reading the checkpoint */
+ if (SIMMASTER(cr))
+ {
+ copy_mat(state->box, box);
+ }
+ if (PAR(cr))
+ {
+ gmx_bcast(sizeof(box), box, cr);
+ }
+ /* Because we do not have the update struct available yet
+ * in which the reference values should be stored,
+ * we store them temporarily in static variables.
+ * This should be thread safe, since they are only written once
+ * and with identical values.
+ */
+ tMPI_Thread_mutex_lock(&deform_init_box_mutex);
+ deform_init_init_step_tpx = inputrec->init_step;
+ copy_mat(box, deform_init_box_tpx);
+ tMPI_Thread_mutex_unlock(&deform_init_box_mutex);
+ }
+
+ if (Flags & MD_STARTFROMCPT)
+ {
+ /* Check if checkpoint file exists before doing continuation.
+ * This way we can use identical input options for the first and subsequent runs...
+ */
+ gmx_bool bReadEkin;
+
+ load_checkpoint(opt2fn_master("-cpi", nfile, fnm, cr), &fplog,
+ cr, ddxyz, &npme,
+ inputrec, state, &bReadEkin,
+ (Flags & MD_APPENDFILES),
+ (Flags & MD_APPENDFILESSET));
+
+ if (bReadEkin)
+ {
+ Flags |= MD_READ_EKIN;
+ }
+ }
+
+ if (MASTER(cr) && (Flags & MD_APPENDFILES))
+ {
+ gmx_log_open(ftp2fn(efLOG, nfile, fnm), cr,
+ Flags, &fplog);
+ }
+
+ /* override nsteps with value from cmdline */
+ override_nsteps_cmdline(fplog, nsteps_cmdline, inputrec, cr);
+
+ if (SIMMASTER(cr))
+ {
+ copy_mat(state->box, box);
+ }
+
+ if (PAR(cr))
+ {
+ gmx_bcast(sizeof(box), box, cr);
+ }
+
+ // TODO This should move to do_md(), because it only makes sense
+ // with dynamical integrators, but there is no test coverage and
+ // it interacts with constraints, somehow.
+ /* Essential dynamics */
+ if (opt2bSet("-ei", nfile, fnm))
+ {
+ /* Open input and output files, allocate space for ED data structure */
+ ed = ed_open(mtop->natoms, &state->edsamstate, nfile, fnm, Flags, oenv, cr);
+ }
+
+ if (PAR(cr) && !(EI_TPI(inputrec->eI) ||
+ inputrec->eI == eiNM))
+ {
+ cr->dd = init_domain_decomposition(fplog, cr, Flags, ddxyz, npme,
+ dd_rank_order,
+ rdd, rconstr,
+ dddlb_opt, dlb_scale,
+ ddcsx, ddcsy, ddcsz,
+ mtop, inputrec,
+ box, state->x,
+ &ddbox, &npme_major, &npme_minor);
+ }
+ else
+ {
+ /* PME, if used, is done on all nodes with 1D decomposition */
+ cr->npmenodes = 0;
+ cr->duty = (DUTY_PP | DUTY_PME);
+ npme_major = 1;
+ npme_minor = 1;
+
+ if (inputrec->ePBC == epbcSCREW)
+ {
+ gmx_fatal(FARGS,
+ "pbc=%s is only implemented with domain decomposition",
+ epbc_names[inputrec->ePBC]);
+ }
+ }
+
+ if (PAR(cr))
+ {
+ /* After possible communicator splitting in make_dd_communicators.
+ * we can set up the intra/inter node communication.
+ */
+ gmx_setup_nodecomm(fplog, cr);
+ }
+
+ /* Initialize per-physical-node MPI process/thread ID and counters. */
+ gmx_init_intranode_counters(cr);
+#if GMX_MPI
+ if (MULTISIM(cr))
+ {
+ md_print_info(cr, fplog,
+ "This is simulation %d out of %d running as a composite GROMACS\n"
+ "multi-simulation job. Setup for this simulation:\n\n",
+ cr->ms->sim, cr->ms->nsim);
+ }
+ md_print_info(cr, fplog, "Using %d MPI %s\n",
+ cr->nnodes,
+#if GMX_THREAD_MPI
+ cr->nnodes == 1 ? "thread" : "threads"
+#else
+ cr->nnodes == 1 ? "process" : "processes"
+#endif
+ );
+ fflush(stderr);
+#endif
+
+ /* Check and update hw_opt for the cut-off scheme */
+ check_and_update_hw_opt_2(hw_opt, inputrec->cutoff_scheme);
+
+ /* Check and update hw_opt for the number of MPI ranks */
+ check_and_update_hw_opt_3(hw_opt);
+
+ gmx_omp_nthreads_init(fplog, cr,
+ hwinfo->nthreads_hw_avail,
+ hw_opt->nthreads_omp,
+ hw_opt->nthreads_omp_pme,
+ (cr->duty & DUTY_PP) == 0,
+ inputrec->cutoff_scheme == ecutsVERLET);
+
+#ifndef NDEBUG
+ if (EI_TPI(inputrec->eI) &&
+ inputrec->cutoff_scheme == ecutsVERLET)
+ {
+ gmx_feenableexcept();
+ }
+#endif
+
+ if (bUseGPU)
+ {
+ /* Select GPU id's to use */
+ gmx_select_gpu_ids(fplog, cr, &hwinfo->gpu_info, bForceUseGPU,
+ &hw_opt->gpu_opt);
+ }
+ else
+ {
+ /* Ignore (potentially) manually selected GPUs */
+ hw_opt->gpu_opt.n_dev_use = 0;
+ }
+
+ /* check consistency across ranks of things like SIMD
+ * support and number of GPUs selected */
+ gmx_check_hw_runconf_consistency(fplog, hwinfo, cr, hw_opt, bUseGPU);
+
+ /* Now that we know the setup is consistent, check for efficiency */
+ check_resource_division_efficiency(hwinfo, hw_opt, Flags & MD_NTOMPSET,
+ cr, fplog);
+
+ if (DOMAINDECOMP(cr))
+ {
+ /* When we share GPUs over ranks, we need to know this for the DLB */
+ dd_setup_dlb_resource_sharing(cr, hwinfo, hw_opt);
+ }
+
+ /* getting number of PP/PME threads
+ PME: env variable should be read only on one node to make sure it is
+ identical everywhere;
+ */
+ nthreads_pme = gmx_omp_nthreads_get(emntPME);
+
+ wcycle = wallcycle_init(fplog, resetstep, cr);
+
+ if (PAR(cr))
+ {
+ /* Master synchronizes its value of reset_counters with all nodes
+ * including PME only nodes */
+ reset_counters = wcycle_get_reset_counters(wcycle);
+ gmx_bcast_sim(sizeof(reset_counters), &reset_counters, cr);
+ wcycle_set_reset_counters(wcycle, reset_counters);
+ }
+
+ snew(nrnb, 1);
+ if (cr->duty & DUTY_PP)
+ {
+ bcast_state(cr, state);
+
+ /* Initiate forcerecord */
+ fr = mk_forcerec();
+ fr->hwinfo = hwinfo;
+ fr->gpu_opt = &hw_opt->gpu_opt;
+ init_forcerec(fplog, fr, fcd, inputrec, mtop, cr, box,
+ opt2fn("-table", nfile, fnm),
+ opt2fn("-tablep", nfile, fnm),
+ opt2fn("-tableb", nfile, fnm),
+ nbpu_opt,
+ FALSE,
+ pforce);
+
+ /* Initialize QM-MM */
+ if (fr->bQMMM)
+ {
+ init_QMMMrec(cr, mtop, inputrec, fr);
+ }
+
+ /* Initialize the mdatoms structure.
+ * mdatoms is not filled with atom data,
+ * as this can not be done now with domain decomposition.
+ */
+ mdatoms = init_mdatoms(fplog, mtop, inputrec->efep != efepNO);
+
+ /* Initialize the virtual site communication */
+ vsite = init_vsite(mtop, cr, FALSE);
+
+ calc_shifts(box, fr->shift_vec);
+
+ /* With periodic molecules the charge groups should be whole at start up
+ * and the virtual sites should not be far from their proper positions.
+ */
+ if (!inputrec->bContinuation && MASTER(cr) &&
+ !(inputrec->ePBC != epbcNONE && inputrec->bPeriodicMols))
+ {
+ /* Make molecules whole at start of run */
+ if (fr->ePBC != epbcNONE)
+ {
+ do_pbc_first_mtop(fplog, inputrec->ePBC, box, mtop, state->x);
+ }
+ if (vsite)
+ {
+ /* Correct initial vsite positions are required
+ * for the initial distribution in the domain decomposition
+ * and for the initial shell prediction.
+ */
+ construct_vsites_mtop(vsite, mtop, state->x);
+ }
+ }
+
+ if (EEL_PME(fr->eeltype) || EVDW_PME(fr->vdwtype))
+ {
+ ewaldcoeff_q = fr->ewaldcoeff_q;
+ ewaldcoeff_lj = fr->ewaldcoeff_lj;
+ pmedata = &fr->pmedata;
+ }
+ else
+ {
+ pmedata = NULL;
+ }
+ }
+ else
+ {
+ /* This is a PME only node */
+
+ /* We don't need the state */
+ done_state(state);
+
+ ewaldcoeff_q = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
+ ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
+ snew(pmedata, 1);
+ }
+
+ if (hw_opt->thread_affinity != threadaffOFF)
+ {
+ /* Before setting affinity, check whether the affinity has changed
+ * - which indicates that probably the OpenMP library has changed it
+ * since we first checked).
+ */
+ gmx_check_thread_affinity_set(fplog, cr,
+ hw_opt, hwinfo->nthreads_hw_avail, TRUE);
+
+ /* Set the CPU affinity */
+ gmx_set_thread_affinity(fplog, cr, hw_opt, hwinfo);
+ }
+
+ /* Initiate PME if necessary,
+ * either on all nodes or on dedicated PME nodes only. */
+ if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
+ {
+ if (mdatoms)
+ {
+ nChargePerturbed = mdatoms->nChargePerturbed;
+ if (EVDW_PME(inputrec->vdwtype))
+ {
+ nTypePerturbed = mdatoms->nTypePerturbed;
+ }
+ }
+ if (cr->npmenodes > 0)
+ {
+ /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
+ gmx_bcast_sim(sizeof(nChargePerturbed), &nChargePerturbed, cr);
+ gmx_bcast_sim(sizeof(nTypePerturbed), &nTypePerturbed, cr);
+ }
+
+ if (cr->duty & DUTY_PME)
+ {
+ status = gmx_pme_init(pmedata, cr, npme_major, npme_minor, inputrec,
+ mtop ? mtop->natoms : 0, nChargePerturbed, nTypePerturbed,
+ (Flags & MD_REPRODUCIBLE), nthreads_pme);
+ if (status != 0)
+ {
+ gmx_fatal(FARGS, "Error %d initializing PME", status);
+ }
+ }
+ }
+
+
+ if (EI_DYNAMICS(inputrec->eI))
+ {
+ /* Turn on signal handling on all nodes */
+ /*
+ * (A user signal from the PME nodes (if any)
+ * is communicated to the PP nodes.
+ */
+ signal_handler_install();
+ }
+
+ if (cr->duty & DUTY_PP)
+ {
+ /* Assumes uniform use of the number of OpenMP threads */
+ walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntDefault));
+
+ if (inputrec->bPull)
+ {
+ /* Initialize pull code */
+ inputrec->pull_work =
+ init_pull(fplog, inputrec->pull, inputrec, nfile, fnm,
+ mtop, cr, oenv, inputrec->fepvals->init_lambda,
+ EI_DYNAMICS(inputrec->eI) && MASTER(cr), Flags);
+ }
+
+ if (inputrec->bRot)
+ {
+ /* Initialize enforced rotation code */
+ init_rot(fplog, inputrec, nfile, fnm, cr, state->x, state->box, mtop, oenv,
+ bVerbose, Flags);
+ }
+
+ constr = init_constraints(fplog, mtop, inputrec, ed, state, cr);
+
+ if (DOMAINDECOMP(cr))
+ {
+ GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
+ /* This call is not included in init_domain_decomposition mainly
+ * because fr->cginfo_mb is set later.
+ */
+ dd_init_bondeds(fplog, cr->dd, mtop, vsite, inputrec,
+ Flags & MD_DDBONDCHECK, fr->cginfo_mb);
+ }
+
+ /* Now do whatever the user wants us to do (how flexible...) */
+ my_integrator(inputrec->eI) (fplog, cr, nfile, fnm,
+ oenv, bVerbose,
+ nstglobalcomm,
+ vsite, constr,
+ nstepout, inputrec, mtop,
+ fcd, state,
+ mdatoms, nrnb, wcycle, ed, fr,
+ repl_ex_nst, repl_ex_nex, repl_ex_seed,
+ cpt_period, max_hours,
+ imdport,
+ Flags,
+ walltime_accounting);
+
+ if (inputrec->bRot)
+ {
+ finish_rot(inputrec->rot);
+ }
+
+ if (inputrec->bPull)
+ {
+ finish_pull(inputrec->pull_work);
+ }
+
+ }
+ else
+ {
+ GMX_RELEASE_ASSERT(pmedata, "pmedata was NULL while cr->duty was not DUTY_PP");
+ /* do PME only */
+ walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntPME));
+ gmx_pmeonly(*pmedata, cr, nrnb, wcycle, walltime_accounting, ewaldcoeff_q, ewaldcoeff_lj, inputrec);
+ }
+
+ wallcycle_stop(wcycle, ewcRUN);
+
+ /* Finish up, write some stuff
+ * if rerunMD, don't write last frame again
+ */
+ finish_run(fplog, cr,
+ inputrec, nrnb, wcycle, walltime_accounting,
+ fr ? fr->nbv : NULL,
+ EI_DYNAMICS(inputrec->eI) && !MULTISIM(cr));
+
+
+ /* Free GPU memory and context */
+ free_gpu_resources(fr, cr, &hwinfo->gpu_info, fr ? fr->gpu_opt : NULL);
+
+ gmx_hardware_info_free(hwinfo);
+
+ /* Does what it says */
+ print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
+ walltime_accounting_destroy(walltime_accounting);
+
+ /* PLUMED */
+ if(plumedswitch){
+ plumed_finalize(plumedmain);
+ }
+ /* END PLUMED */
+
+ /* Close logfile already here if we were appending to it */
+ if (MASTER(cr) && (Flags & MD_APPENDFILES))
+ {
+ gmx_log_close(fplog);
+ }
+
+ rc = (int)gmx_get_stop_condition();
+
+ done_ed(&ed);
+
+#if 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;
+}
+
+} // namespace gmx
diff --git a/patches/gromacs-2016-beta1.diff/src/programs/mdrun/runner.cpp.preplumed b/patches/gromacs-2016-beta1.diff/src/programs/mdrun/runner.cpp.preplumed
new file mode 100644
index 0000000000000000000000000000000000000000..3294776ff121cf4b9ee740de07be997e6636fe2d
--- /dev/null
+++ b/patches/gromacs-2016-beta1.diff/src/programs/mdrun/runner.cpp.preplumed
@@ -0,0 +1,1386 @@
+/*
+ * This file is part of the GROMACS molecular simulation package.
+ *
+ * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
+ * Copyright (c) 2001-2004, The GROMACS development team.
+ * Copyright (c) 2011,2012,2013,2014,2015,2016, by the GROMACS development team, led by
+ * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
+ * and including many others, as listed in the AUTHORS file in the
+ * top-level source directory and at http://www.gromacs.org.
+ *
+ * GROMACS is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public License
+ * as published by the Free Software Foundation; either version 2.1
+ * of the License, or (at your option) any later version.
+ *
+ * GROMACS is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with GROMACS; if not, see
+ * http://www.gnu.org/licenses, or write to the Free Software Foundation,
+ * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ * If you want to redistribute modifications to GROMACS, 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 http://www.gromacs.org.
+ *
+ * To help us fund GROMACS development, we humbly ask that you cite
+ * the research papers on the package. Check out http://www.gromacs.org.
+ */
+/*! \internal \file
+ *
+ * \brief Implements the MD runner routine calling all integrators.
+ *
+ * \author David van der Spoel <david.vanderspoel@icm.uu.se>
+ * \ingroup module_mdlib
+ */
+#include "gmxpre.h"
+
+#include "runner.h"
+
+#include "config.h"
+
+#include <assert.h>
+#include <signal.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include <algorithm>
+
+#include "gromacs/commandline/filenm.h"
+#include "gromacs/domdec/domdec.h"
+#include "gromacs/domdec/domdec_struct.h"
+#include "gromacs/essentialdynamics/edsam.h"
+#include "gromacs/ewald/pme.h"
+#include "gromacs/fileio/checkpoint.h"
+#include "gromacs/fileio/oenv.h"
+#include "gromacs/fileio/tpxio.h"
+#include "gromacs/gmxlib/md_logging.h"
+#include "gromacs/gmxlib/network.h"
+#include "gromacs/gpu_utils/gpu_utils.h"
+#include "gromacs/hardware/cpuinfo.h"
+#include "gromacs/hardware/detecthardware.h"
+#include "gromacs/listed-forces/disre.h"
+#include "gromacs/listed-forces/orires.h"
+#include "gromacs/math/calculate-ewald-splitting-coefficient.h"
+#include "gromacs/math/functions.h"
+#include "gromacs/math/utilities.h"
+#include "gromacs/math/vec.h"
+#include "gromacs/mdlib/calc_verletbuf.h"
+#include "gromacs/mdlib/constr.h"
+#include "gromacs/mdlib/force.h"
+#include "gromacs/mdlib/forcerec.h"
+#include "gromacs/mdlib/gmx_omp_nthreads.h"
+#include "gromacs/mdlib/integrator.h"
+#include "gromacs/mdlib/main.h"
+#include "gromacs/mdlib/md_support.h"
+#include "gromacs/mdlib/mdatoms.h"
+#include "gromacs/mdlib/mdrun.h"
+#include "gromacs/mdlib/minimize.h"
+#include "gromacs/mdlib/nbnxn_search.h"
+#include "gromacs/mdlib/qmmm.h"
+#include "gromacs/mdlib/sighandler.h"
+#include "gromacs/mdlib/sim_util.h"
+#include "gromacs/mdlib/tpi.h"
+#include "gromacs/mdrunutility/threadaffinity.h"
+#include "gromacs/mdtypes/commrec.h"
+#include "gromacs/mdtypes/inputrec.h"
+#include "gromacs/mdtypes/md_enums.h"
+#include "gromacs/mdtypes/state.h"
+#include "gromacs/pbcutil/pbc.h"
+#include "gromacs/pulling/pull.h"
+#include "gromacs/pulling/pull_rotation.h"
+#include "gromacs/timing/wallcycle.h"
+#include "gromacs/topology/mtop_util.h"
+#include "gromacs/trajectory/trajectoryframe.h"
+#include "gromacs/utility/cstringutil.h"
+#include "gromacs/utility/exceptions.h"
+#include "gromacs/utility/fatalerror.h"
+#include "gromacs/utility/gmxassert.h"
+#include "gromacs/utility/gmxmpi.h"
+#include "gromacs/utility/pleasecite.h"
+#include "gromacs/utility/smalloc.h"
+
+#include "deform.h"
+#include "md.h"
+#include "repl_ex.h"
+#include "resource-division.h"
+
+#ifdef GMX_FAHCORE
+#include "corewrap.h"
+#endif
+
+//! First step used in pressure scaling
+gmx_int64_t deform_init_init_step_tpx;
+//! Initial box for pressure scaling
+matrix deform_init_box_tpx;
+//! MPI variable for use in pressure scaling
+tMPI_Thread_mutex_t deform_init_box_mutex = TMPI_THREAD_MUTEX_INITIALIZER;
+
+#if GMX_THREAD_MPI
+/* The minimum number of atoms per tMPI thread. With fewer atoms than this,
+ * the number of threads will get lowered.
+ */
+#define MIN_ATOMS_PER_MPI_THREAD 90
+#define MIN_ATOMS_PER_GPU 900
+
+struct mdrunner_arglist
+{
+ gmx_hw_opt_t hw_opt;
+ FILE *fplog;
+ t_commrec *cr;
+ int nfile;
+ const t_filenm *fnm;
+ const gmx_output_env_t *oenv;
+ gmx_bool bVerbose;
+ int nstglobalcomm;
+ ivec ddxyz;
+ int dd_rank_order;
+ int npme;
+ real rdd;
+ real rconstr;
+ const char *dddlb_opt;
+ real dlb_scale;
+ const char *ddcsx;
+ const char *ddcsy;
+ const char *ddcsz;
+ const char *nbpu_opt;
+ int nstlist_cmdline;
+ gmx_int64_t nsteps_cmdline;
+ int nstepout;
+ int resetstep;
+ int nmultisim;
+ int repl_ex_nst;
+ int repl_ex_nex;
+ int repl_ex_seed;
+ real pforce;
+ real cpt_period;
+ real max_hours;
+ int imdport;
+ unsigned long Flags;
+};
+
+
+/* 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)
+{
+ try
+ {
+ 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 = reinitialize_commrec_for_this_thread(mc.cr);
+
+ if (MASTER(cr))
+ {
+ fplog = mc.fplog;
+ }
+
+ gmx::mdrunner(&mc.hw_opt, fplog, cr, mc.nfile, fnm, mc.oenv,
+ mc.bVerbose, mc.nstglobalcomm,
+ mc.ddxyz, mc.dd_rank_order, mc.npme, mc.rdd,
+ mc.rconstr, mc.dddlb_opt, mc.dlb_scale,
+ mc.ddcsx, mc.ddcsy, mc.ddcsz,
+ mc.nbpu_opt, mc.nstlist_cmdline,
+ mc.nsteps_cmdline, mc.nstepout, mc.resetstep,
+ mc.nmultisim, mc.repl_ex_nst, mc.repl_ex_nex, mc.repl_ex_seed, mc.pforce,
+ mc.cpt_period, mc.max_hours, mc.imdport, mc.Flags);
+ }
+ GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+}
+
+
+/* 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(gmx_hw_opt_t *hw_opt,
+ FILE *fplog, t_commrec *cr, int nfile,
+ const t_filenm fnm[], const gmx_output_env_t *oenv, gmx_bool bVerbose,
+ int nstglobalcomm,
+ ivec ddxyz, int dd_rank_order, int npme,
+ real rdd, real rconstr,
+ const char *dddlb_opt, real dlb_scale,
+ const char *ddcsx, const char *ddcsy, const char *ddcsz,
+ const char *nbpu_opt, int nstlist_cmdline,
+ gmx_int64_t nsteps_cmdline,
+ int nstepout, int resetstep,
+ int nmultisim, int repl_ex_nst, int repl_ex_nex, int repl_ex_seed,
+ real pforce, real cpt_period, real max_hours,
+ 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 (hw_opt->nthreads_tmpi < 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 */
+ /* hw_opt contains pointers, which should all be NULL at this stage */
+ mda->hw_opt = *hw_opt;
+ mda->fplog = fplog;
+ mda->cr = cr;
+ mda->nfile = nfile;
+ mda->fnm = fnmn;
+ mda->oenv = oenv;
+ mda->bVerbose = bVerbose;
+ mda->nstglobalcomm = nstglobalcomm;
+ mda->ddxyz[XX] = ddxyz[XX];
+ mda->ddxyz[YY] = ddxyz[YY];
+ mda->ddxyz[ZZ] = ddxyz[ZZ];
+ mda->dd_rank_order = dd_rank_order;
+ mda->npme = npme;
+ 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->nbpu_opt = nbpu_opt;
+ mda->nstlist_cmdline = nstlist_cmdline;
+ mda->nsteps_cmdline = nsteps_cmdline;
+ mda->nstepout = nstepout;
+ mda->resetstep = resetstep;
+ mda->nmultisim = nmultisim;
+ mda->repl_ex_nst = repl_ex_nst;
+ mda->repl_ex_nex = repl_ex_nex;
+ mda->repl_ex_seed = repl_ex_seed;
+ mda->pforce = pforce;
+ mda->cpt_period = cpt_period;
+ mda->max_hours = max_hours;
+ mda->Flags = Flags;
+
+ /* now spawn new threads that start mdrunner_start_fn(), while
+ the main thread returns, we set thread affinity later */
+ ret = tMPI_Init_fn(TRUE, hw_opt->nthreads_tmpi, TMPI_AFFINITY_NONE,
+ mdrunner_start_fn, (void*)(mda) );
+ if (ret != TMPI_SUCCESS)
+ {
+ return NULL;
+ }
+
+ crn = reinitialize_commrec_for_this_thread(cr);
+ return crn;
+}
+
+#endif /* GMX_THREAD_MPI */
+
+
+/*! \brief Cost of non-bonded kernels
+ *
+ * We determine the extra cost of the non-bonded kernels compared to
+ * a reference nstlist value of 10 (which is the default in grompp).
+ */
+static const int nbnxnReferenceNstlist = 10;
+//! The values to try when switching
+const int nstlist_try[] = { 20, 25, 40 };
+//! Number of elements in the neighborsearch list trials.
+#define NNSTL sizeof(nstlist_try)/sizeof(nstlist_try[0])
+/* Increase nstlist until the non-bonded cost increases more than listfac_ok,
+ * but never more than listfac_max.
+ * A standard (protein+)water system at 300K with PME ewald_rtol=1e-5
+ * needs 1.28 at rcoulomb=0.9 and 1.24 at rcoulomb=1.0 to get to nstlist=40.
+ * Note that both CPU and GPU factors are conservative. Performance should
+ * not go down due to this tuning, except with a relatively slow GPU.
+ * On the other hand, at medium/high parallelization or with fast GPUs
+ * nstlist will not be increased enough to reach optimal performance.
+ */
+/* CPU: pair-search is about a factor 1.5 slower than the non-bonded kernel */
+//! Max OK performance ratio beween force calc and neighbor searching
+static const float nbnxn_cpu_listfac_ok = 1.05;
+//! Too high performance ratio beween force calc and neighbor searching
+static const float nbnxn_cpu_listfac_max = 1.09;
+/* CPU: pair-search is about a factor 2-3 slower than the non-bonded kernel */
+//! Max OK performance ratio beween force calc and neighbor searching
+static const float nbnxn_knl_listfac_ok = 1.22;
+//! Too high performance ratio beween force calc and neighbor searching
+static const float nbnxn_knl_listfac_max = 1.3;
+/* GPU: pair-search is a factor 1.5-3 slower than the non-bonded kernel */
+//! Max OK performance ratio beween force calc and neighbor searching
+static const float nbnxn_gpu_listfac_ok = 1.20;
+//! Too high performance ratio beween force calc and neighbor searching
+static const float nbnxn_gpu_listfac_max = 1.30;
+
+/*! \brief Try to increase nstlist when using the Verlet cut-off scheme */
+static void increase_nstlist(FILE *fp, t_commrec *cr,
+ t_inputrec *ir, int nstlist_cmdline,
+ const gmx_mtop_t *mtop, matrix box,
+ gmx_bool bGPU, const gmx::CpuInfo &cpuinfo)
+{
+ float listfac_ok, listfac_max;
+ int nstlist_orig, nstlist_prev;
+ verletbuf_list_setup_t ls;
+ real rlistWithReferenceNstlist, rlist_inc, rlist_ok, rlist_max;
+ real rlist_new, rlist_prev;
+ size_t nstlist_ind = 0;
+ t_state state_tmp;
+ gmx_bool bBox, bDD, bCont;
+ const char *nstl_gpu = "\nFor optimal performance with a GPU nstlist (now %d) should be larger.\nThe optimum depends on your CPU and GPU resources.\nYou might want to try several nstlist values.\n";
+ const char *nve_err = "Can not increase nstlist because an NVE ensemble is used";
+ const char *vbd_err = "Can not increase nstlist because verlet-buffer-tolerance is not set or used";
+ const char *box_err = "Can not increase nstlist because the box is too small";
+ const char *dd_err = "Can not increase nstlist because of domain decomposition limitations";
+ char buf[STRLEN];
+
+ if (nstlist_cmdline <= 0)
+ {
+ if (ir->nstlist == 1)
+ {
+ /* The user probably set nstlist=1 for a reason,
+ * don't mess with the settings.
+ */
+ return;
+ }
+
+ if (fp != NULL && bGPU && ir->nstlist < nstlist_try[0])
+ {
+ fprintf(fp, nstl_gpu, ir->nstlist);
+ }
+ nstlist_ind = 0;
+ while (nstlist_ind < NNSTL && ir->nstlist >= nstlist_try[nstlist_ind])
+ {
+ nstlist_ind++;
+ }
+ if (nstlist_ind == NNSTL)
+ {
+ /* There are no larger nstlist value to try */
+ return;
+ }
+ }
+
+ if (EI_MD(ir->eI) && ir->etc == etcNO)
+ {
+ if (MASTER(cr))
+ {
+ fprintf(stderr, "%s\n", nve_err);
+ }
+ if (fp != NULL)
+ {
+ fprintf(fp, "%s\n", nve_err);
+ }
+
+ return;
+ }
+
+ if (ir->verletbuf_tol == 0 && bGPU)
+ {
+ gmx_fatal(FARGS, "You are using an old tpr file with a GPU, please generate a new tpr file with an up to date version of grompp");
+ }
+
+ if (ir->verletbuf_tol < 0)
+ {
+ if (MASTER(cr))
+ {
+ fprintf(stderr, "%s\n", vbd_err);
+ }
+ if (fp != NULL)
+ {
+ fprintf(fp, "%s\n", vbd_err);
+ }
+
+ return;
+ }
+
+ if (bGPU)
+ {
+ listfac_ok = nbnxn_gpu_listfac_ok;
+ listfac_max = nbnxn_gpu_listfac_max;
+ }
+ else if (cpuinfo.feature(gmx::CpuInfo::Feature::X86_Avx512ER))
+ {
+ listfac_ok = nbnxn_knl_listfac_ok;
+ listfac_max = nbnxn_knl_listfac_max;
+ }
+ else
+ {
+ listfac_ok = nbnxn_cpu_listfac_ok;
+ listfac_max = nbnxn_cpu_listfac_max;
+ }
+
+ nstlist_orig = ir->nstlist;
+ if (nstlist_cmdline > 0)
+ {
+ if (fp)
+ {
+ sprintf(buf, "Getting nstlist=%d from command line option",
+ nstlist_cmdline);
+ }
+ ir->nstlist = nstlist_cmdline;
+ }
+
+ verletbuf_get_list_setup(TRUE, bGPU, &ls);
+
+ /* Allow rlist to make the list a given factor larger than the list
+ * would be with the reference value for nstlist (10).
+ */
+ nstlist_prev = ir->nstlist;
+ ir->nstlist = nbnxnReferenceNstlist;
+ calc_verlet_buffer_size(mtop, det(box), ir, -1, &ls, NULL,
+ &rlistWithReferenceNstlist);
+ ir->nstlist = nstlist_prev;
+
+ /* Determine the pair list size increase due to zero interactions */
+ rlist_inc = nbnxn_get_rlist_effective_inc(ls.cluster_size_j,
+ mtop->natoms/det(box));
+ rlist_ok = (rlistWithReferenceNstlist + rlist_inc)*std::cbrt(listfac_ok) - rlist_inc;
+ rlist_max = (rlistWithReferenceNstlist + rlist_inc)*std::cbrt(listfac_max) - rlist_inc;
+ if (debug)
+ {
+ fprintf(debug, "nstlist tuning: rlist_inc %.3f rlist_ok %.3f rlist_max %.3f\n",
+ rlist_inc, rlist_ok, rlist_max);
+ }
+
+ nstlist_prev = nstlist_orig;
+ rlist_prev = ir->rlist;
+ do
+ {
+ if (nstlist_cmdline <= 0)
+ {
+ ir->nstlist = nstlist_try[nstlist_ind];
+ }
+
+ /* Set the pair-list buffer size in ir */
+ calc_verlet_buffer_size(mtop, det(box), ir, -1, &ls, NULL, &rlist_new);
+
+ /* Does rlist fit in the box? */
+ bBox = (gmx::square(rlist_new) < max_cutoff2(ir->ePBC, box));
+ bDD = TRUE;
+ if (bBox && DOMAINDECOMP(cr))
+ {
+ /* Check if rlist fits in the domain decomposition */
+ if (inputrec2nboundeddim(ir) < DIM)
+ {
+ gmx_incons("Changing nstlist with domain decomposition and unbounded dimensions is not implemented yet");
+ }
+ copy_mat(box, state_tmp.box);
+ bDD = change_dd_cutoff(cr, &state_tmp, ir, rlist_new);
+ }
+
+ if (debug)
+ {
+ fprintf(debug, "nstlist %d rlist %.3f bBox %d bDD %d\n",
+ ir->nstlist, rlist_new, bBox, bDD);
+ }
+
+ bCont = FALSE;
+
+ if (nstlist_cmdline <= 0)
+ {
+ if (bBox && bDD && rlist_new <= rlist_max)
+ {
+ /* Increase nstlist */
+ nstlist_prev = ir->nstlist;
+ rlist_prev = rlist_new;
+ bCont = (nstlist_ind+1 < NNSTL && rlist_new < rlist_ok);
+ }
+ else
+ {
+ /* Stick with the previous nstlist */
+ ir->nstlist = nstlist_prev;
+ rlist_new = rlist_prev;
+ bBox = TRUE;
+ bDD = TRUE;
+ }
+ }
+
+ nstlist_ind++;
+ }
+ while (bCont);
+
+ if (!bBox || !bDD)
+ {
+ gmx_warning(!bBox ? box_err : dd_err);
+ if (fp != NULL)
+ {
+ fprintf(fp, "\n%s\n", bBox ? box_err : dd_err);
+ }
+ ir->nstlist = nstlist_orig;
+ }
+ else if (ir->nstlist != nstlist_orig || rlist_new != ir->rlist)
+ {
+ sprintf(buf, "Changing nstlist from %d to %d, rlist from %g to %g",
+ nstlist_orig, ir->nstlist,
+ ir->rlist, rlist_new);
+ if (MASTER(cr))
+ {
+ fprintf(stderr, "%s\n\n", buf);
+ }
+ if (fp != NULL)
+ {
+ fprintf(fp, "%s\n\n", buf);
+ }
+ ir->rlist = rlist_new;
+ }
+}
+
+/*! \brief Initialize variables for Verlet scheme simulation */
+static void prepare_verlet_scheme(FILE *fplog,
+ t_commrec *cr,
+ t_inputrec *ir,
+ int nstlist_cmdline,
+ const gmx_mtop_t *mtop,
+ matrix box,
+ gmx_bool bUseGPU,
+ const gmx::CpuInfo &cpuinfo)
+{
+ /* For NVE simulations, we will retain the initial list buffer */
+ if (EI_DYNAMICS(ir->eI) &&
+ ir->verletbuf_tol > 0 &&
+ !(EI_MD(ir->eI) && ir->etc == etcNO))
+ {
+ /* Update the Verlet buffer size for the current run setup */
+ verletbuf_list_setup_t ls;
+ real rlist_new;
+
+ /* Here we assume SIMD-enabled kernels are being used. But as currently
+ * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
+ * and 4x2 gives a larger buffer than 4x4, this is ok.
+ */
+ verletbuf_get_list_setup(TRUE, bUseGPU, &ls);
+
+ calc_verlet_buffer_size(mtop, det(box), ir, -1, &ls, NULL, &rlist_new);
+
+ if (rlist_new != ir->rlist)
+ {
+ if (fplog != NULL)
+ {
+ fprintf(fplog, "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
+ ir->rlist, rlist_new,
+ ls.cluster_size_i, ls.cluster_size_j);
+ }
+ ir->rlist = rlist_new;
+ }
+ }
+
+ if (nstlist_cmdline > 0 && (!EI_DYNAMICS(ir->eI) || ir->verletbuf_tol <= 0))
+ {
+ gmx_fatal(FARGS, "Can not set nstlist without %s",
+ !EI_DYNAMICS(ir->eI) ? "dynamics" : "verlet-buffer-tolerance");
+ }
+
+ if (EI_DYNAMICS(ir->eI))
+ {
+ /* Set or try nstlist values */
+ increase_nstlist(fplog, cr, ir, nstlist_cmdline, mtop, box, bUseGPU, cpuinfo);
+ }
+}
+
+/*! \brief Override the nslist value in inputrec
+ *
+ * with value passed on the command line (if any)
+ */
+static void override_nsteps_cmdline(FILE *fplog,
+ gmx_int64_t nsteps_cmdline,
+ t_inputrec *ir,
+ const t_commrec *cr)
+{
+ assert(ir);
+ assert(cr);
+
+ /* override with anything else than the default -2 */
+ if (nsteps_cmdline > -2)
+ {
+ char sbuf_steps[STEPSTRSIZE];
+ char sbuf_msg[STRLEN];
+
+ ir->nsteps = nsteps_cmdline;
+ if (EI_DYNAMICS(ir->eI) && nsteps_cmdline != -1)
+ {
+ sprintf(sbuf_msg, "Overriding nsteps with value passed on the command line: %s steps, %.3g ps",
+ gmx_step_str(nsteps_cmdline, sbuf_steps),
+ fabs(nsteps_cmdline*ir->delta_t));
+ }
+ else
+ {
+ sprintf(sbuf_msg, "Overriding nsteps with value passed on the command line: %s steps",
+ gmx_step_str(nsteps_cmdline, sbuf_steps));
+ }
+
+ md_print_warn(cr, fplog, "%s\n", sbuf_msg);
+ }
+ else if (nsteps_cmdline < -2)
+ {
+ gmx_fatal(FARGS, "Invalid nsteps value passed on the command line: %d",
+ nsteps_cmdline);
+ }
+ /* Do nothing if nsteps_cmdline == -2 */
+}
+
+namespace gmx
+{
+
+//! \brief Return the correct integrator function.
+static integrator_t *my_integrator(unsigned int ei)
+{
+ switch (ei)
+ {
+ case eiMD:
+ case eiBD:
+ case eiSD1:
+ case eiVV:
+ case eiVVAK:
+ if (!EI_DYNAMICS(ei))
+ {
+ GMX_THROW(APIError("do_md integrator would be called for a non-dynamical integrator"));
+ }
+ return do_md;
+ case eiSteep:
+ return do_steep;
+ case eiCG:
+ return do_cg;
+ case eiNM:
+ return do_nm;
+ case eiLBFGS:
+ return do_lbfgs;
+ case eiTPI:
+ case eiTPIC:
+ if (!EI_TPI(ei))
+ {
+ GMX_THROW(APIError("do_tpi integrator would be called for a non-TPI integrator"));
+ }
+ return do_tpi;
+ case eiSD2_REMOVED:
+ GMX_THROW(NotImplementedError("SD2 integrator has been removed"));
+ default:
+ GMX_THROW(APIError("Non existing integrator selected"));
+ }
+}
+
+int mdrunner(gmx_hw_opt_t *hw_opt,
+ FILE *fplog, t_commrec *cr, int nfile,
+ const t_filenm fnm[], const gmx_output_env_t *oenv, gmx_bool bVerbose,
+ int nstglobalcomm,
+ ivec ddxyz, int dd_rank_order, int npme, real rdd, real rconstr,
+ const char *dddlb_opt, real dlb_scale,
+ const char *ddcsx, const char *ddcsy, const char *ddcsz,
+ const char *nbpu_opt, int nstlist_cmdline,
+ gmx_int64_t nsteps_cmdline, int nstepout, int resetstep,
+ int gmx_unused nmultisim, int repl_ex_nst, int repl_ex_nex,
+ int repl_ex_seed, real pforce, real cpt_period, real max_hours,
+ int imdport, unsigned long Flags)
+{
+ gmx_bool bForceUseGPU, bTryUseGPU, bRerunMD;
+ t_inputrec *inputrec;
+ t_state *state = NULL;
+ matrix box;
+ gmx_ddbox_t ddbox = {0};
+ int npme_major, npme_minor;
+ t_nrnb *nrnb;
+ gmx_mtop_t *mtop = NULL;
+ t_mdatoms *mdatoms = NULL;
+ t_forcerec *fr = NULL;
+ t_fcdata *fcd = NULL;
+ real ewaldcoeff_q = 0;
+ real ewaldcoeff_lj = 0;
+ struct gmx_pme_t **pmedata = NULL;
+ gmx_vsite_t *vsite = NULL;
+ gmx_constr_t constr;
+ int nChargePerturbed = -1, nTypePerturbed = 0, status;
+ gmx_wallcycle_t wcycle;
+ gmx_walltime_accounting_t walltime_accounting = NULL;
+ int rc;
+ gmx_int64_t reset_counters;
+ gmx_edsam_t ed = NULL;
+ int nthreads_pme = 1;
+ gmx_hw_info_t *hwinfo = NULL;
+ /* The master rank decides early on bUseGPU and broadcasts this later */
+ gmx_bool bUseGPU = FALSE;
+
+ /* CAUTION: threads may be started later on in this function, so
+ cr doesn't reflect the final parallel state right now */
+ snew(inputrec, 1);
+ snew(mtop, 1);
+
+ if (Flags & MD_APPENDFILES)
+ {
+ fplog = NULL;
+ }
+
+ bRerunMD = (Flags & MD_RERUN);
+ bForceUseGPU = (strncmp(nbpu_opt, "gpu", 3) == 0);
+ bTryUseGPU = (strncmp(nbpu_opt, "auto", 4) == 0) || bForceUseGPU;
+
+ /* Detect hardware, gather information. This is an operation that is
+ * global for this process (MPI rank). */
+ hwinfo = gmx_detect_hardware(fplog, cr, bTryUseGPU);
+
+ gmx_print_detected_hardware(fplog, cr, hwinfo);
+
+ if (fplog != NULL)
+ {
+ /* Print references after all software/hardware printing */
+ please_cite(fplog, "Abraham2015");
+ please_cite(fplog, "Pall2015");
+ please_cite(fplog, "Pronk2013");
+ please_cite(fplog, "Hess2008b");
+ please_cite(fplog, "Spoel2005a");
+ please_cite(fplog, "Lindahl2001a");
+ please_cite(fplog, "Berendsen95a");
+ }
+
+ snew(state, 1);
+ if (SIMMASTER(cr))
+ {
+ /* Read (nearly) all data required for the simulation */
+ read_tpx_state(ftp2fn(efTPR, nfile, fnm), inputrec, state, mtop);
+
+ if (inputrec->cutoff_scheme == ecutsVERLET)
+ {
+ /* Here the master rank decides if all ranks will use GPUs */
+ bUseGPU = (hwinfo->gpu_info.n_dev_compatible > 0 ||
+ getenv("GMX_EMULATE_GPU") != NULL);
+
+ /* TODO add GPU kernels for this and replace this check by:
+ * (bUseGPU && (ir->vdwtype == evdwPME &&
+ * ir->ljpme_combination_rule == eljpmeLB))
+ * update the message text and the content of nbnxn_acceleration_supported.
+ */
+ if (bUseGPU &&
+ !nbnxn_gpu_acceleration_supported(fplog, cr, inputrec, bRerunMD))
+ {
+ /* Fallback message printed by nbnxn_acceleration_supported */
+ if (bForceUseGPU)
+ {
+ gmx_fatal(FARGS, "GPU acceleration requested, but not supported with the given input settings");
+ }
+ bUseGPU = FALSE;
+ }
+
+ prepare_verlet_scheme(fplog, cr,
+ inputrec, nstlist_cmdline, mtop, state->box,
+ bUseGPU, *hwinfo->cpuInfo);
+ }
+ else
+ {
+ if (nstlist_cmdline > 0)
+ {
+ gmx_fatal(FARGS, "Can not set nstlist with the group cut-off scheme");
+ }
+
+ if (hwinfo->gpu_info.n_dev_compatible > 0)
+ {
+ md_print_warn(cr, fplog,
+ "NOTE: GPU(s) found, but the current simulation can not use GPUs\n"
+ " To use a GPU, set the mdp option: cutoff-scheme = Verlet\n");
+ }
+
+ if (bForceUseGPU)
+ {
+ gmx_fatal(FARGS, "GPU requested, but can't be used without cutoff-scheme=Verlet");
+ }
+
+#if GMX_TARGET_BGQ
+ md_print_warn(cr, fplog,
+ "NOTE: There is no SIMD implementation of the group scheme kernels on\n"
+ " BlueGene/Q. You will observe better performance from using the\n"
+ " Verlet cut-off scheme.\n");
+#endif
+ }
+ }
+
+ /* Check and update the hardware options for internal consistency */
+ check_and_update_hw_opt_1(hw_opt, cr, npme);
+
+ /* Early check for externally set process affinity. */
+ gmx_check_thread_affinity_set(fplog, cr,
+ hw_opt, hwinfo->nthreads_hw_avail, FALSE);
+
+#if GMX_THREAD_MPI
+ if (SIMMASTER(cr))
+ {
+ if (npme > 0 && hw_opt->nthreads_tmpi <= 0)
+ {
+ gmx_fatal(FARGS, "You need to explicitly specify the number of MPI threads (-ntmpi) when using separate PME ranks");
+ }
+
+ /* Since the master knows the cut-off scheme, update hw_opt for this.
+ * This is done later for normal MPI and also once more with tMPI
+ * for all tMPI ranks.
+ */
+ check_and_update_hw_opt_2(hw_opt, inputrec->cutoff_scheme);
+
+ /* NOW the threads will be started: */
+ hw_opt->nthreads_tmpi = get_nthreads_mpi(hwinfo,
+ hw_opt,
+ inputrec, mtop,
+ cr, fplog, bUseGPU);
+
+ if (hw_opt->nthreads_tmpi > 1)
+ {
+ t_commrec *cr_old = cr;
+ /* now start the threads. */
+ cr = mdrunner_start_threads(hw_opt, fplog, cr_old, nfile, fnm,
+ oenv, bVerbose, nstglobalcomm,
+ ddxyz, dd_rank_order, npme, rdd, rconstr,
+ dddlb_opt, dlb_scale, ddcsx, ddcsy, ddcsz,
+ nbpu_opt, nstlist_cmdline,
+ nsteps_cmdline, nstepout, resetstep, nmultisim,
+ repl_ex_nst, repl_ex_nex, repl_ex_seed, pforce,
+ cpt_period, max_hours,
+ Flags);
+ /* the main thread continues here with a new cr. We don't deallocate
+ the old cr because other threads may still be reading it. */
+ if (cr == NULL)
+ {
+ gmx_comm("Failed to spawn threads");
+ }
+ }
+ }
+#endif
+ /* END OF CAUTION: cr is now reliable */
+
+ if (PAR(cr))
+ {
+ /* now broadcast everything to the non-master nodes/threads: */
+ init_parallel(cr, inputrec, mtop);
+
+ /* The master rank decided on the use of GPUs,
+ * broadcast this information to all ranks.
+ */
+ gmx_bcast_sim(sizeof(bUseGPU), &bUseGPU, cr);
+ }
+
+ if (fplog != NULL)
+ {
+ pr_inputrec(fplog, 0, "Input Parameters", inputrec, FALSE);
+ fprintf(fplog, "\n");
+ }
+
+ /* now make sure the state is initialized and propagated */
+ set_state_entries(state, inputrec);
+
+ /* A parallel command line option consistency check that we can
+ only do after any threads have started. */
+ if (!PAR(cr) &&
+ (ddxyz[XX] > 1 || ddxyz[YY] > 1 || ddxyz[ZZ] > 1 || npme > 0))
+ {
+ gmx_fatal(FARGS,
+ "The -dd or -npme option request a parallel simulation, "
+#if !GMX_MPI
+ "but %s was compiled without threads or MPI enabled"
+#else
+#if GMX_THREAD_MPI
+ "but the number of MPI-threads (option -ntmpi) is not set or is 1"
+#else
+ "but %s was not started through mpirun/mpiexec or only one rank was requested through mpirun/mpiexec"
+#endif
+#endif
+ , output_env_get_program_display_name(oenv)
+ );
+ }
+
+ if (bRerunMD &&
+ (EI_ENERGY_MINIMIZATION(inputrec->eI) || eiNM == inputrec->eI))
+ {
+ gmx_fatal(FARGS, "The .mdp file specified an energy mininization or normal mode algorithm, and these are not compatible with mdrun -rerun");
+ }
+
+ if (can_use_allvsall(inputrec, TRUE, cr, fplog) && DOMAINDECOMP(cr))
+ {
+ gmx_fatal(FARGS, "All-vs-all loops do not work with domain decomposition, use a single MPI rank");
+ }
+
+ if (!(EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype)))
+ {
+ if (npme > 0)
+ {
+ gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
+ "PME-only ranks are requested, but the system does not use PME for electrostatics or LJ");
+ }
+
+ npme = 0;
+ }
+
+ if (bUseGPU && npme < 0)
+ {
+ /* With GPUs we don't automatically use PME-only ranks. PME ranks can
+ * improve performance with many threads per GPU, since our OpenMP
+ * scaling is bad, but it's difficult to automate the setup.
+ */
+ npme = 0;
+ }
+
+#ifdef GMX_FAHCORE
+ if (MASTER(cr))
+ {
+ fcRegisterSteps(inputrec->nsteps, inputrec->init_step);
+ }
+#endif
+
+ /* NMR restraints must be initialized before load_checkpoint,
+ * since with time averaging the history is added to t_state.
+ * For proper consistency check we therefore need to extend
+ * t_state here.
+ * So the PME-only nodes (if present) will also initialize
+ * the distance restraints.
+ */
+ snew(fcd, 1);
+
+ /* This needs to be called before read_checkpoint to extend the state */
+ init_disres(fplog, mtop, inputrec, cr, fcd, state, repl_ex_nst > 0);
+
+ init_orires(fplog, mtop, state->x, inputrec, cr, &(fcd->orires),
+ state);
+
+ if (inputrecDeform(inputrec))
+ {
+ /* Store the deform reference box before reading the checkpoint */
+ if (SIMMASTER(cr))
+ {
+ copy_mat(state->box, box);
+ }
+ if (PAR(cr))
+ {
+ gmx_bcast(sizeof(box), box, cr);
+ }
+ /* Because we do not have the update struct available yet
+ * in which the reference values should be stored,
+ * we store them temporarily in static variables.
+ * This should be thread safe, since they are only written once
+ * and with identical values.
+ */
+ tMPI_Thread_mutex_lock(&deform_init_box_mutex);
+ deform_init_init_step_tpx = inputrec->init_step;
+ copy_mat(box, deform_init_box_tpx);
+ tMPI_Thread_mutex_unlock(&deform_init_box_mutex);
+ }
+
+ if (Flags & MD_STARTFROMCPT)
+ {
+ /* Check if checkpoint file exists before doing continuation.
+ * This way we can use identical input options for the first and subsequent runs...
+ */
+ gmx_bool bReadEkin;
+
+ load_checkpoint(opt2fn_master("-cpi", nfile, fnm, cr), &fplog,
+ cr, ddxyz, &npme,
+ inputrec, state, &bReadEkin,
+ (Flags & MD_APPENDFILES),
+ (Flags & MD_APPENDFILESSET));
+
+ if (bReadEkin)
+ {
+ Flags |= MD_READ_EKIN;
+ }
+ }
+
+ if (MASTER(cr) && (Flags & MD_APPENDFILES))
+ {
+ gmx_log_open(ftp2fn(efLOG, nfile, fnm), cr,
+ Flags, &fplog);
+ }
+
+ /* override nsteps with value from cmdline */
+ override_nsteps_cmdline(fplog, nsteps_cmdline, inputrec, cr);
+
+ if (SIMMASTER(cr))
+ {
+ copy_mat(state->box, box);
+ }
+
+ if (PAR(cr))
+ {
+ gmx_bcast(sizeof(box), box, cr);
+ }
+
+ // TODO This should move to do_md(), because it only makes sense
+ // with dynamical integrators, but there is no test coverage and
+ // it interacts with constraints, somehow.
+ /* Essential dynamics */
+ if (opt2bSet("-ei", nfile, fnm))
+ {
+ /* Open input and output files, allocate space for ED data structure */
+ ed = ed_open(mtop->natoms, &state->edsamstate, nfile, fnm, Flags, oenv, cr);
+ }
+
+ if (PAR(cr) && !(EI_TPI(inputrec->eI) ||
+ inputrec->eI == eiNM))
+ {
+ cr->dd = init_domain_decomposition(fplog, cr, Flags, ddxyz, npme,
+ dd_rank_order,
+ rdd, rconstr,
+ dddlb_opt, dlb_scale,
+ ddcsx, ddcsy, ddcsz,
+ mtop, inputrec,
+ box, state->x,
+ &ddbox, &npme_major, &npme_minor);
+ }
+ else
+ {
+ /* PME, if used, is done on all nodes with 1D decomposition */
+ cr->npmenodes = 0;
+ cr->duty = (DUTY_PP | DUTY_PME);
+ npme_major = 1;
+ npme_minor = 1;
+
+ if (inputrec->ePBC == epbcSCREW)
+ {
+ gmx_fatal(FARGS,
+ "pbc=%s is only implemented with domain decomposition",
+ epbc_names[inputrec->ePBC]);
+ }
+ }
+
+ if (PAR(cr))
+ {
+ /* After possible communicator splitting in make_dd_communicators.
+ * we can set up the intra/inter node communication.
+ */
+ gmx_setup_nodecomm(fplog, cr);
+ }
+
+ /* Initialize per-physical-node MPI process/thread ID and counters. */
+ gmx_init_intranode_counters(cr);
+#if GMX_MPI
+ if (MULTISIM(cr))
+ {
+ md_print_info(cr, fplog,
+ "This is simulation %d out of %d running as a composite GROMACS\n"
+ "multi-simulation job. Setup for this simulation:\n\n",
+ cr->ms->sim, cr->ms->nsim);
+ }
+ md_print_info(cr, fplog, "Using %d MPI %s\n",
+ cr->nnodes,
+#if GMX_THREAD_MPI
+ cr->nnodes == 1 ? "thread" : "threads"
+#else
+ cr->nnodes == 1 ? "process" : "processes"
+#endif
+ );
+ fflush(stderr);
+#endif
+
+ /* Check and update hw_opt for the cut-off scheme */
+ check_and_update_hw_opt_2(hw_opt, inputrec->cutoff_scheme);
+
+ /* Check and update hw_opt for the number of MPI ranks */
+ check_and_update_hw_opt_3(hw_opt);
+
+ gmx_omp_nthreads_init(fplog, cr,
+ hwinfo->nthreads_hw_avail,
+ hw_opt->nthreads_omp,
+ hw_opt->nthreads_omp_pme,
+ (cr->duty & DUTY_PP) == 0,
+ inputrec->cutoff_scheme == ecutsVERLET);
+
+#ifndef NDEBUG
+ if (EI_TPI(inputrec->eI) &&
+ inputrec->cutoff_scheme == ecutsVERLET)
+ {
+ gmx_feenableexcept();
+ }
+#endif
+
+ if (bUseGPU)
+ {
+ /* Select GPU id's to use */
+ gmx_select_gpu_ids(fplog, cr, &hwinfo->gpu_info, bForceUseGPU,
+ &hw_opt->gpu_opt);
+ }
+ else
+ {
+ /* Ignore (potentially) manually selected GPUs */
+ hw_opt->gpu_opt.n_dev_use = 0;
+ }
+
+ /* check consistency across ranks of things like SIMD
+ * support and number of GPUs selected */
+ gmx_check_hw_runconf_consistency(fplog, hwinfo, cr, hw_opt, bUseGPU);
+
+ /* Now that we know the setup is consistent, check for efficiency */
+ check_resource_division_efficiency(hwinfo, hw_opt, Flags & MD_NTOMPSET,
+ cr, fplog);
+
+ if (DOMAINDECOMP(cr))
+ {
+ /* When we share GPUs over ranks, we need to know this for the DLB */
+ dd_setup_dlb_resource_sharing(cr, hwinfo, hw_opt);
+ }
+
+ /* getting number of PP/PME threads
+ PME: env variable should be read only on one node to make sure it is
+ identical everywhere;
+ */
+ nthreads_pme = gmx_omp_nthreads_get(emntPME);
+
+ wcycle = wallcycle_init(fplog, resetstep, cr);
+
+ if (PAR(cr))
+ {
+ /* Master synchronizes its value of reset_counters with all nodes
+ * including PME only nodes */
+ reset_counters = wcycle_get_reset_counters(wcycle);
+ gmx_bcast_sim(sizeof(reset_counters), &reset_counters, cr);
+ wcycle_set_reset_counters(wcycle, reset_counters);
+ }
+
+ snew(nrnb, 1);
+ if (cr->duty & DUTY_PP)
+ {
+ bcast_state(cr, state);
+
+ /* Initiate forcerecord */
+ fr = mk_forcerec();
+ fr->hwinfo = hwinfo;
+ fr->gpu_opt = &hw_opt->gpu_opt;
+ init_forcerec(fplog, fr, fcd, inputrec, mtop, cr, box,
+ opt2fn("-table", nfile, fnm),
+ opt2fn("-tablep", nfile, fnm),
+ opt2fn("-tableb", nfile, fnm),
+ nbpu_opt,
+ FALSE,
+ pforce);
+
+ /* Initialize QM-MM */
+ if (fr->bQMMM)
+ {
+ init_QMMMrec(cr, mtop, inputrec, fr);
+ }
+
+ /* Initialize the mdatoms structure.
+ * mdatoms is not filled with atom data,
+ * as this can not be done now with domain decomposition.
+ */
+ mdatoms = init_mdatoms(fplog, mtop, inputrec->efep != efepNO);
+
+ /* Initialize the virtual site communication */
+ vsite = init_vsite(mtop, cr, FALSE);
+
+ calc_shifts(box, fr->shift_vec);
+
+ /* With periodic molecules the charge groups should be whole at start up
+ * and the virtual sites should not be far from their proper positions.
+ */
+ if (!inputrec->bContinuation && MASTER(cr) &&
+ !(inputrec->ePBC != epbcNONE && inputrec->bPeriodicMols))
+ {
+ /* Make molecules whole at start of run */
+ if (fr->ePBC != epbcNONE)
+ {
+ do_pbc_first_mtop(fplog, inputrec->ePBC, box, mtop, state->x);
+ }
+ if (vsite)
+ {
+ /* Correct initial vsite positions are required
+ * for the initial distribution in the domain decomposition
+ * and for the initial shell prediction.
+ */
+ construct_vsites_mtop(vsite, mtop, state->x);
+ }
+ }
+
+ if (EEL_PME(fr->eeltype) || EVDW_PME(fr->vdwtype))
+ {
+ ewaldcoeff_q = fr->ewaldcoeff_q;
+ ewaldcoeff_lj = fr->ewaldcoeff_lj;
+ pmedata = &fr->pmedata;
+ }
+ else
+ {
+ pmedata = NULL;
+ }
+ }
+ else
+ {
+ /* This is a PME only node */
+
+ /* We don't need the state */
+ done_state(state);
+
+ ewaldcoeff_q = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
+ ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
+ snew(pmedata, 1);
+ }
+
+ if (hw_opt->thread_affinity != threadaffOFF)
+ {
+ /* Before setting affinity, check whether the affinity has changed
+ * - which indicates that probably the OpenMP library has changed it
+ * since we first checked).
+ */
+ gmx_check_thread_affinity_set(fplog, cr,
+ hw_opt, hwinfo->nthreads_hw_avail, TRUE);
+
+ /* Set the CPU affinity */
+ gmx_set_thread_affinity(fplog, cr, hw_opt, hwinfo);
+ }
+
+ /* Initiate PME if necessary,
+ * either on all nodes or on dedicated PME nodes only. */
+ if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
+ {
+ if (mdatoms)
+ {
+ nChargePerturbed = mdatoms->nChargePerturbed;
+ if (EVDW_PME(inputrec->vdwtype))
+ {
+ nTypePerturbed = mdatoms->nTypePerturbed;
+ }
+ }
+ if (cr->npmenodes > 0)
+ {
+ /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
+ gmx_bcast_sim(sizeof(nChargePerturbed), &nChargePerturbed, cr);
+ gmx_bcast_sim(sizeof(nTypePerturbed), &nTypePerturbed, cr);
+ }
+
+ if (cr->duty & DUTY_PME)
+ {
+ status = gmx_pme_init(pmedata, cr, npme_major, npme_minor, inputrec,
+ mtop ? mtop->natoms : 0, nChargePerturbed, nTypePerturbed,
+ (Flags & MD_REPRODUCIBLE), nthreads_pme);
+ if (status != 0)
+ {
+ gmx_fatal(FARGS, "Error %d initializing PME", status);
+ }
+ }
+ }
+
+
+ if (EI_DYNAMICS(inputrec->eI))
+ {
+ /* Turn on signal handling on all nodes */
+ /*
+ * (A user signal from the PME nodes (if any)
+ * is communicated to the PP nodes.
+ */
+ signal_handler_install();
+ }
+
+ if (cr->duty & DUTY_PP)
+ {
+ /* Assumes uniform use of the number of OpenMP threads */
+ walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntDefault));
+
+ if (inputrec->bPull)
+ {
+ /* Initialize pull code */
+ inputrec->pull_work =
+ init_pull(fplog, inputrec->pull, inputrec, nfile, fnm,
+ mtop, cr, oenv, inputrec->fepvals->init_lambda,
+ EI_DYNAMICS(inputrec->eI) && MASTER(cr), Flags);
+ }
+
+ if (inputrec->bRot)
+ {
+ /* Initialize enforced rotation code */
+ init_rot(fplog, inputrec, nfile, fnm, cr, state->x, state->box, mtop, oenv,
+ bVerbose, Flags);
+ }
+
+ constr = init_constraints(fplog, mtop, inputrec, ed, state, cr);
+
+ if (DOMAINDECOMP(cr))
+ {
+ GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
+ /* This call is not included in init_domain_decomposition mainly
+ * because fr->cginfo_mb is set later.
+ */
+ dd_init_bondeds(fplog, cr->dd, mtop, vsite, inputrec,
+ Flags & MD_DDBONDCHECK, fr->cginfo_mb);
+ }
+
+ /* Now do whatever the user wants us to do (how flexible...) */
+ my_integrator(inputrec->eI) (fplog, cr, nfile, fnm,
+ oenv, bVerbose,
+ nstglobalcomm,
+ vsite, constr,
+ nstepout, inputrec, mtop,
+ fcd, state,
+ mdatoms, nrnb, wcycle, ed, fr,
+ repl_ex_nst, repl_ex_nex, repl_ex_seed,
+ cpt_period, max_hours,
+ imdport,
+ Flags,
+ walltime_accounting);
+
+ if (inputrec->bRot)
+ {
+ finish_rot(inputrec->rot);
+ }
+
+ if (inputrec->bPull)
+ {
+ finish_pull(inputrec->pull_work);
+ }
+
+ }
+ else
+ {
+ GMX_RELEASE_ASSERT(pmedata, "pmedata was NULL while cr->duty was not DUTY_PP");
+ /* do PME only */
+ walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntPME));
+ gmx_pmeonly(*pmedata, cr, nrnb, wcycle, walltime_accounting, ewaldcoeff_q, ewaldcoeff_lj, inputrec);
+ }
+
+ wallcycle_stop(wcycle, ewcRUN);
+
+ /* Finish up, write some stuff
+ * if rerunMD, don't write last frame again
+ */
+ finish_run(fplog, cr,
+ inputrec, nrnb, wcycle, walltime_accounting,
+ fr ? fr->nbv : NULL,
+ EI_DYNAMICS(inputrec->eI) && !MULTISIM(cr));
+
+
+ /* Free GPU memory and context */
+ free_gpu_resources(fr, cr, &hwinfo->gpu_info, fr ? fr->gpu_opt : NULL);
+
+ gmx_hardware_info_free(hwinfo);
+
+ /* Does what it says */
+ print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
+ walltime_accounting_destroy(walltime_accounting);
+
+ /* Close logfile already here if we were appending to it */
+ if (MASTER(cr) && (Flags & MD_APPENDFILES))
+ {
+ gmx_log_close(fplog);
+ }
+
+ rc = (int)gmx_get_stop_condition();
+
+ done_ed(&ed);
+
+#if 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;
+}
+
+} // namespace gmx