diff --git a/patches/gromacs-2019.1.config b/patches/gromacs-2019.1.config
new file mode 100644
index 0000000000000000000000000000000000000000..4348d93fd2f9c9db2567ee332f7ba123c58b7abc
--- /dev/null
+++ b/patches/gromacs-2019.1.config
@@ -0,0 +1,33 @@
+
+
+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.
+
+On clusters you may want to patch gromacs using the static version of plumed, in this case
+building gromacs can result in multiple errors. One possible solution is to configure gromacs
+with these additional options:
+
+cmake -DBUILD_SHARED_LIBS=OFF -DGMX_PREFER_STATIC_LIBS=ON
+
+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
+}
+
+plumed_before_patch(){
+ plumed_patch_info
+}
+
diff --git a/patches/gromacs-2019.1.diff/src/gromacs/CMakeLists.txt b/patches/gromacs-2019.1.diff/src/gromacs/CMakeLists.txt
new file mode 100644
index 0000000000000000000000000000000000000000..80f7d30c13f328603ca75f3b37b4d1b0e97eb22c
--- /dev/null
+++ b/patches/gromacs-2019.1.diff/src/gromacs/CMakeLists.txt
@@ -0,0 +1,446 @@
+#
+# This file is part of the GROMACS molecular simulation package.
+#
+# Copyright (c) 2010,2011,2012,2013,2014,2015,2016,2017,2018,2019, 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)
+
+if (GMX_CLANG_CUDA)
+ include(gmxClangCudaUtils)
+endif()
+
+set_property(GLOBAL PROPERTY GMX_LIBGROMACS_SOURCES)
+set_property(GLOBAL PROPERTY GMX_LIBGROMACS_GPU_IMPL_SOURCES)
+set_property(GLOBAL PROPERTY GMX_INSTALLED_HEADERS)
+set_property(GLOBAL PROPERTY GMX_AVX_512_SOURCE)
+
+add_library(libgromacs_external OBJECT "")
+if(CMAKE_COMPILER_IS_GNUCXX)
+ # Keep quiet about e.g. linearalgebra module
+ target_compile_options(libgromacs_external PRIVATE ${CXXFLAGS_NO_STRINGOP_TRUNCATION})
+endif()
+
+add_library(libgromacs_generated OBJECT "")
+if (BUILD_SHARED_LIBS)
+ set_target_properties(libgromacs_external PROPERTIES POSITION_INDEPENDENT_CODE true)
+ set_target_properties(libgromacs_generated PROPERTIES POSITION_INDEPENDENT_CODE true)
+endif()
+
+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 ()
+
+# TODO Reconsider this, as the CUDA driver API is probably a simpler
+# approach, at least for the build system. See Redmine #2530
+function (gmx_compile_cpp_as_cuda)
+ _gmx_add_files_to_property(GMX_LIBGROMACS_GPU_IMPL_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 ${ARGN}
+ DESTINATION "${CMAKE_INSTALL_INCLUDEDIR}/${_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()
+
+add_subdirectory(gmxlib)
+add_subdirectory(mdlib)
+add_subdirectory(applied-forces)
+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(mdrun)
+add_subdirectory(mdrunutility)
+add_subdirectory(mdtypes)
+add_subdirectory(onlinehelp)
+add_subdirectory(options)
+add_subdirectory(pbcutil)
+add_subdirectory(random)
+add_subdirectory(restraint)
+add_subdirectory(tables)
+add_subdirectory(taskassignment)
+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(awh)
+add_subdirectory(simd)
+add_subdirectory(imd)
+add_subdirectory(compat)
+add_subdirectory(mimic)
+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(energyanalysis)
+ 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)
+target_sources(libgromacs_external PRIVATE ${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.cpp)
+gmx_configure_version_file(
+ utility/baseversion-gen.cpp.cmakein ${GENERATED_VERSION_FILE}
+ REMOTE_HASH
+ EXTRA_VARS
+ GMX_SOURCE_DOI
+ )
+list(APPEND LIBGROMACS_SOURCES ${GENERATED_VERSION_FILE}
+ $<TARGET_OBJECTS:libgromacs_external>
+ $<TARGET_OBJECTS:libgromacs_generated>)
+
+# Mark some shared GPU implementation files to compile with CUDA if needed
+if (GMX_USE_CUDA)
+ get_property(LIBGROMACS_GPU_IMPL_SOURCES GLOBAL PROPERTY GMX_LIBGROMACS_GPU_IMPL_SOURCES)
+ set_source_files_properties(${LIBGROMACS_GPU_IMPL_SOURCES} PROPERTIES CUDA_SOURCE_PROPERTY_FORMAT OBJ)
+endif()
+
+# set up CUDA compilation with clang
+if (GMX_CLANG_CUDA)
+ foreach (_file ${LIBGROMACS_SOURCES})
+ get_filename_component(_ext ${_file} EXT)
+ get_source_file_property(_cuda_source_format ${_file} CUDA_SOURCE_PROPERTY_FORMAT)
+ if ("${_ext}" STREQUAL ".cu" OR _cuda_source_format)
+ gmx_compile_cuda_file_with_clang(${_file})
+ endif()
+ endforeach()
+endif()
+
+if (GMX_USE_CUDA)
+ # Work around FindCUDA that prevents using target_link_libraries()
+ # with keywords otherwise...
+ set(CUDA_LIBRARIES PRIVATE ${CUDA_LIBRARIES})
+ if (NOT GMX_CLANG_CUDA)
+ cuda_add_library(libgromacs ${LIBGROMACS_SOURCES})
+ else()
+ add_library(libgromacs ${LIBGROMACS_SOURCES})
+ endif()
+ target_link_libraries(libgromacs PRIVATE ${CUDA_CUFFT_LIBRARIES})
+else()
+ add_library(libgromacs ${LIBGROMACS_SOURCES})
+endif()
+
+if (GMX_USE_OPENCL)
+ option(GMX_EXTERNAL_CLFFT "True if an external clFFT is required to be used" FALSE)
+ mark_as_advanced(GMX_EXTERNAL_CLFFT)
+
+ # Default to using clFFT found on the system
+ # switch to quiet at the second run.
+ if (DEFINED clFFT_LIBRARY)
+ set (clFFT_FIND_QUIETLY TRUE)
+ endif()
+ find_package(clFFT)
+ if (NOT clFFT_FOUND)
+ if (GMX_EXTERNAL_CLFFT)
+ message(FATAL_ERROR "Did not find required external clFFT library, consider setting clFFT_ROOT_DIR")
+ endif()
+
+ if(MSVC)
+ message(FATAL_ERROR
+"An OpenCL build was requested with Visual Studio compiler, but GROMACS
+requires clFFT, which was not found on your system. GROMACS does bundle
+clFFT to help with building for OpenCL, but that clFFT has not yet been
+ported to the more recent versions of that compiler that GROMACS itself
+requires. Thus for now, OpenCL is not available with MSVC and the internal
+build of clFFT in GROMACS 2019. Either change compiler, try installing
+a clFFT package, or use the latest GROMACS 2018 point release.")
+ endif()
+
+ # Fall back on the internal version
+ set (_clFFT_dir ../external/clFFT/src)
+ add_subdirectory(${_clFFT_dir} clFFT-build)
+ target_sources(libgromacs PRIVATE
+ $<TARGET_OBJECTS:clFFT>
+ )
+ target_include_directories(libgromacs SYSTEM PRIVATE ${_clFFT_dir}/include)
+ # Use the magic variable for how to link any library needed for
+ # dlopen, etc. which is -ldl where needed, and empty otherwise
+ # (e.g. Windows, BSD, Mac).
+ target_link_libraries(libgromacs PRIVATE "${CMAKE_DL_LIBS}")
+ else()
+ target_link_libraries(libgromacs PRIVATE clFFT)
+ endif()
+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). Same issue exists for nonbonded kernels
+# so we supress them for all generated files.
+include(CheckCXXCompilerFlag)
+check_cxx_compiler_flag("-Wno-unused -Wno-unused-parameter" HAS_NO_UNUSED)
+check_cxx_compiler_flag(-Wno-missing-declarations HAS_NO_MISSING_DECL)
+check_cxx_compiler_flag(-Wno-missing-prototypes HAS_NO_MISSING_PROTO)
+check_cxx_compiler_flag(/wd4101 HAS_NO_MSVC_UNUSED)
+if (NOT MSVC)
+ check_cxx_compiler_flag(-wd1419 HAS_DECL_IN_SOURCE)
+endif()
+if (HAS_NO_UNUSED)
+ target_compile_options(libgromacs_generated PRIVATE "-Wno-unused;-Wno-unused-parameter")
+endif()
+if (HAS_NO_MISSING_DECL)
+ target_compile_options(libgromacs_generated PRIVATE "-Wno-missing-declarations")
+endif()
+# TODO The group scheme kernels don't use proper function prototype
+# declarations, and clang warns about such use, which we suppress
+# rather than fix. We would prefer to use no suppressions. However
+# other compilers do not support such a warning suppression for C++
+# source files, and issue warnings about that. Remove the use of
+# -Wno-missing-prototypes here and above when the group scheme is
+# removed.
+if (HAS_NO_MISSING_PROTO AND "${CMAKE_CXX_COMPILER_ID}" MATCHES "Clang")
+ target_compile_options(libgromacs_generated PRIVATE "-Wno-missing-prototypes")
+endif()
+if (HAS_NO_MSVC_UNUSED)
+ target_compile_options(libgromacs_generated PRIVATE "/wd4101")
+endif()
+if (HAS_DECL_IN_SOURCE)
+ target_compile_options(libgromacs_generated PRIVATE "-wd1419")
+endif()
+
+if(SIMD_AVX_512_CXX_SUPPORTED AND NOT ("${GMX_SIMD_ACTIVE}" STREQUAL "AVX_512_KNL"))
+ # Since we might be overriding -march=core-avx2, add a flag so we don't warn for this specific file.
+ # On KNL this can cause illegal instruction because the compiler might use non KNL AVX instructions
+ # with the SIMD_AVX_512_CXX_FLAGS flags.
+ set_source_files_properties(hardware/identifyavx512fmaunits.cpp PROPERTIES COMPILE_FLAGS "${SIMD_AVX_512_CXX_FLAGS} ${CXX_NO_UNUSED_OPTION_WARNING_FLAGS}")
+endif()
+
+gmx_setup_tng_for_libgromacs()
+
+target_link_libraries(libgromacs
+ PRIVATE
+ ${EXTRAE_LIBRARIES}
+ ${GMX_EXTRA_LIBRARIES}
+ ${GMX_COMMON_LIBRARIES}
+ ${FFT_LIBRARIES} ${LINEAR_ALGEBRA_LIBRARIES}
+ ${THREAD_LIB} ${GMX_SHARED_LINKER_FLAGS}
+ ${OpenCL_LIBRARIES}
+ ${GMX_STDLIB_LIBRARIES}
+ PUBLIC
+ ${GMX_PUBLIC_LIBRARIES}
+ ${PLUMED_LOAD}
+ )
+set_target_properties(libgromacs PROPERTIES
+ OUTPUT_NAME "gromacs${GMX_LIBS_SUFFIX}"
+ SOVERSION ${LIBRARY_SOVERSION_MAJOR}
+ VERSION ${LIBRARY_VERSION}
+ COMPILE_FLAGS "${OpenMP_C_FLAGS}")
+
+gmx_manage_lmfit()
+target_link_libraries(libgromacs PRIVATE lmfit)
+
+if (CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND CMAKE_CXX_COMPILER_VERSION MATCHES "^6\.0")
+ target_compile_options(libgromacs PRIVATE $<$<COMPILE_LANGUAGE:CXX>:-Weverything ${IGNORED_CLANG_ALL_WARNINGS}>)
+endif()
+if (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
+ target_compile_options(libgromacs PRIVATE $<$<COMPILE_LANGUAGE:CXX>:/analyze /analyze:stacksize 70000
+ #Control flow warnings are disabled because the commond line output is insufficient. There is no tool
+ #to convert the xml report to e.g. HTML and even in Visual Studio the viewer doesn't work with cmake support.
+ /wd6001 #unitialized memory
+ /wd6011 #derefencing NULL
+ /wd6053 #prior call not zero-terminate
+ /wd6054 #might not be zero-terminated
+ /wd6385 #reading invalid data
+ /wd6386 #buffer overrun
+ /wd6387 #could be '0'
+ /wd28199 #uninitialized memory
+ # For compile time constant (e.g. templates) the following warnings have flase postives
+ /wd6239 #(<non-zero> && <expr>)
+ /wd6240 #(<expr> && <non-zero>)
+ /wd6294 #Ill-defined for-loop
+ /wd6326 #comparison of constant with other constant
+ /wd28020 #expression involving paramter is not true
+ # Misc
+ /wd6330 #incorrect type to function (warns for char (instead of unsigned) for isspace/isalpha/isdigit/..))
+ /wd6993 #OpenMP ignored
+ #TODO
+ /wd6031 #return value ignored (important - mostly warnigns about sscanf)
+ /wd6244 #hides declaration (known issue - we ingore similar warnings for other compilers)
+ /wd6246 #hides declaration
+ >
+ )
+endif()
+
+if (GMX_CLANG_TIDY)
+ set_target_properties(libgromacs PROPERTIES CXX_CLANG_TIDY
+ "${CLANG_TIDY_EXE};-warnings-as-errors=*")
+endif()
+
+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 ${CMAKE_INSTALL_LIBDIR}
+ RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
+ ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
+ 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
+ ${CMAKE_INSTALL_LIBDIR} COMPONENT libraries)
+ endif()
+ endforeach()
+ else()
+ message(WARNING "INSTALL_CUDART_LIB only makes sense when configuring for CUDA support")
+ endif()
+endif()
+
+if(GMX_USE_OPENCL)
+ # Install the utility headers
+ file(GLOB OPENCL_INSTALLED_FILES
+ gpu_utils/vectype_ops.clh
+ gpu_utils/device_utils.clh
+ )
+ install(FILES ${OPENCL_INSTALLED_FILES}
+ DESTINATION ${GMX_INSTALL_OCLDIR}/gromacs/gpu_utils
+ COMPONENT libraries)
+ file(GLOB OPENCL_INSTALLED_FILES
+ pbcutil/ishift.h
+ )
+ install(FILES ${OPENCL_INSTALLED_FILES}
+ DESTINATION ${GMX_INSTALL_OCLDIR}/gromacs/pbcutil
+ COMPONENT libraries)
+
+ # Install the NB source and headers
+ file(GLOB OPENCL_INSTALLED_FILES
+ mdlib/nbnxn_consts.h
+ )
+ install(FILES ${OPENCL_INSTALLED_FILES}
+ DESTINATION ${GMX_INSTALL_OCLDIR}/gromacs/mdlib
+ COMPONENT libraries)
+ file(GLOB OPENCL_INSTALLED_FILES
+ mdlib/nbnxn_ocl/nbnxn_ocl_kernels.cl
+ mdlib/nbnxn_ocl/nbnxn_ocl_kernel.clh
+ mdlib/nbnxn_ocl/nbnxn_ocl_kernel_pruneonly.clh
+ mdlib/nbnxn_ocl/nbnxn_ocl_kernels.clh
+ mdlib/nbnxn_ocl/nbnxn_ocl_kernels_fastgen.clh
+ mdlib/nbnxn_ocl/nbnxn_ocl_kernels_fastgen_add_twincut.clh
+ mdlib/nbnxn_ocl/nbnxn_ocl_kernel_utils.clh
+ mdlib/nbnxn_ocl/nbnxn_ocl_consts.h
+ )
+ install(FILES ${OPENCL_INSTALLED_FILES}
+ DESTINATION ${GMX_INSTALL_OCLDIR}/gromacs/mdlib/nbnxn_ocl
+ COMPONENT libraries)
+
+ # Install the PME source and headers
+ file(GLOB OPENCL_INSTALLED_FILES
+ ewald/pme-spread.clh
+ ewald/pme-solve.clh
+ ewald/pme-gather.clh
+ ewald/pme-gpu-utils.clh
+ ewald/pme-program.cl
+ ewald/pme-gpu-types.h
+ )
+ install(FILES ${OPENCL_INSTALLED_FILES}
+ DESTINATION ${GMX_INSTALL_OCLDIR}/gromacs/ewald
+ COMPONENT libraries)
+endif()
diff --git a/patches/gromacs-2019.1.diff/src/gromacs/CMakeLists.txt.preplumed b/patches/gromacs-2019.1.diff/src/gromacs/CMakeLists.txt.preplumed
new file mode 100644
index 0000000000000000000000000000000000000000..f94af553a91047274f879856836782856d62abcb
--- /dev/null
+++ b/patches/gromacs-2019.1.diff/src/gromacs/CMakeLists.txt.preplumed
@@ -0,0 +1,445 @@
+#
+# This file is part of the GROMACS molecular simulation package.
+#
+# Copyright (c) 2010,2011,2012,2013,2014,2015,2016,2017,2018,2019, 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)
+
+if (GMX_CLANG_CUDA)
+ include(gmxClangCudaUtils)
+endif()
+
+set_property(GLOBAL PROPERTY GMX_LIBGROMACS_SOURCES)
+set_property(GLOBAL PROPERTY GMX_LIBGROMACS_GPU_IMPL_SOURCES)
+set_property(GLOBAL PROPERTY GMX_INSTALLED_HEADERS)
+set_property(GLOBAL PROPERTY GMX_AVX_512_SOURCE)
+
+add_library(libgromacs_external OBJECT "")
+if(CMAKE_COMPILER_IS_GNUCXX)
+ # Keep quiet about e.g. linearalgebra module
+ target_compile_options(libgromacs_external PRIVATE ${CXXFLAGS_NO_STRINGOP_TRUNCATION})
+endif()
+
+add_library(libgromacs_generated OBJECT "")
+if (BUILD_SHARED_LIBS)
+ set_target_properties(libgromacs_external PROPERTIES POSITION_INDEPENDENT_CODE true)
+ set_target_properties(libgromacs_generated PROPERTIES POSITION_INDEPENDENT_CODE true)
+endif()
+
+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 ()
+
+# TODO Reconsider this, as the CUDA driver API is probably a simpler
+# approach, at least for the build system. See Redmine #2530
+function (gmx_compile_cpp_as_cuda)
+ _gmx_add_files_to_property(GMX_LIBGROMACS_GPU_IMPL_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 ${ARGN}
+ DESTINATION "${CMAKE_INSTALL_INCLUDEDIR}/${_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()
+
+add_subdirectory(gmxlib)
+add_subdirectory(mdlib)
+add_subdirectory(applied-forces)
+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(mdrun)
+add_subdirectory(mdrunutility)
+add_subdirectory(mdtypes)
+add_subdirectory(onlinehelp)
+add_subdirectory(options)
+add_subdirectory(pbcutil)
+add_subdirectory(random)
+add_subdirectory(restraint)
+add_subdirectory(tables)
+add_subdirectory(taskassignment)
+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(awh)
+add_subdirectory(simd)
+add_subdirectory(imd)
+add_subdirectory(compat)
+add_subdirectory(mimic)
+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(energyanalysis)
+ 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)
+target_sources(libgromacs_external PRIVATE ${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.cpp)
+gmx_configure_version_file(
+ utility/baseversion-gen.cpp.cmakein ${GENERATED_VERSION_FILE}
+ REMOTE_HASH
+ EXTRA_VARS
+ GMX_SOURCE_DOI
+ )
+list(APPEND LIBGROMACS_SOURCES ${GENERATED_VERSION_FILE}
+ $<TARGET_OBJECTS:libgromacs_external>
+ $<TARGET_OBJECTS:libgromacs_generated>)
+
+# Mark some shared GPU implementation files to compile with CUDA if needed
+if (GMX_USE_CUDA)
+ get_property(LIBGROMACS_GPU_IMPL_SOURCES GLOBAL PROPERTY GMX_LIBGROMACS_GPU_IMPL_SOURCES)
+ set_source_files_properties(${LIBGROMACS_GPU_IMPL_SOURCES} PROPERTIES CUDA_SOURCE_PROPERTY_FORMAT OBJ)
+endif()
+
+# set up CUDA compilation with clang
+if (GMX_CLANG_CUDA)
+ foreach (_file ${LIBGROMACS_SOURCES})
+ get_filename_component(_ext ${_file} EXT)
+ get_source_file_property(_cuda_source_format ${_file} CUDA_SOURCE_PROPERTY_FORMAT)
+ if ("${_ext}" STREQUAL ".cu" OR _cuda_source_format)
+ gmx_compile_cuda_file_with_clang(${_file})
+ endif()
+ endforeach()
+endif()
+
+if (GMX_USE_CUDA)
+ # Work around FindCUDA that prevents using target_link_libraries()
+ # with keywords otherwise...
+ set(CUDA_LIBRARIES PRIVATE ${CUDA_LIBRARIES})
+ if (NOT GMX_CLANG_CUDA)
+ cuda_add_library(libgromacs ${LIBGROMACS_SOURCES})
+ else()
+ add_library(libgromacs ${LIBGROMACS_SOURCES})
+ endif()
+ target_link_libraries(libgromacs PRIVATE ${CUDA_CUFFT_LIBRARIES})
+else()
+ add_library(libgromacs ${LIBGROMACS_SOURCES})
+endif()
+
+if (GMX_USE_OPENCL)
+ option(GMX_EXTERNAL_CLFFT "True if an external clFFT is required to be used" FALSE)
+ mark_as_advanced(GMX_EXTERNAL_CLFFT)
+
+ # Default to using clFFT found on the system
+ # switch to quiet at the second run.
+ if (DEFINED clFFT_LIBRARY)
+ set (clFFT_FIND_QUIETLY TRUE)
+ endif()
+ find_package(clFFT)
+ if (NOT clFFT_FOUND)
+ if (GMX_EXTERNAL_CLFFT)
+ message(FATAL_ERROR "Did not find required external clFFT library, consider setting clFFT_ROOT_DIR")
+ endif()
+
+ if(MSVC)
+ message(FATAL_ERROR
+"An OpenCL build was requested with Visual Studio compiler, but GROMACS
+requires clFFT, which was not found on your system. GROMACS does bundle
+clFFT to help with building for OpenCL, but that clFFT has not yet been
+ported to the more recent versions of that compiler that GROMACS itself
+requires. Thus for now, OpenCL is not available with MSVC and the internal
+build of clFFT in GROMACS 2019. Either change compiler, try installing
+a clFFT package, or use the latest GROMACS 2018 point release.")
+ endif()
+
+ # Fall back on the internal version
+ set (_clFFT_dir ../external/clFFT/src)
+ add_subdirectory(${_clFFT_dir} clFFT-build)
+ target_sources(libgromacs PRIVATE
+ $<TARGET_OBJECTS:clFFT>
+ )
+ target_include_directories(libgromacs SYSTEM PRIVATE ${_clFFT_dir}/include)
+ # Use the magic variable for how to link any library needed for
+ # dlopen, etc. which is -ldl where needed, and empty otherwise
+ # (e.g. Windows, BSD, Mac).
+ target_link_libraries(libgromacs PRIVATE "${CMAKE_DL_LIBS}")
+ else()
+ target_link_libraries(libgromacs PRIVATE clFFT)
+ endif()
+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). Same issue exists for nonbonded kernels
+# so we supress them for all generated files.
+include(CheckCXXCompilerFlag)
+check_cxx_compiler_flag("-Wno-unused -Wno-unused-parameter" HAS_NO_UNUSED)
+check_cxx_compiler_flag(-Wno-missing-declarations HAS_NO_MISSING_DECL)
+check_cxx_compiler_flag(-Wno-missing-prototypes HAS_NO_MISSING_PROTO)
+check_cxx_compiler_flag(/wd4101 HAS_NO_MSVC_UNUSED)
+if (NOT MSVC)
+ check_cxx_compiler_flag(-wd1419 HAS_DECL_IN_SOURCE)
+endif()
+if (HAS_NO_UNUSED)
+ target_compile_options(libgromacs_generated PRIVATE "-Wno-unused;-Wno-unused-parameter")
+endif()
+if (HAS_NO_MISSING_DECL)
+ target_compile_options(libgromacs_generated PRIVATE "-Wno-missing-declarations")
+endif()
+# TODO The group scheme kernels don't use proper function prototype
+# declarations, and clang warns about such use, which we suppress
+# rather than fix. We would prefer to use no suppressions. However
+# other compilers do not support such a warning suppression for C++
+# source files, and issue warnings about that. Remove the use of
+# -Wno-missing-prototypes here and above when the group scheme is
+# removed.
+if (HAS_NO_MISSING_PROTO AND "${CMAKE_CXX_COMPILER_ID}" MATCHES "Clang")
+ target_compile_options(libgromacs_generated PRIVATE "-Wno-missing-prototypes")
+endif()
+if (HAS_NO_MSVC_UNUSED)
+ target_compile_options(libgromacs_generated PRIVATE "/wd4101")
+endif()
+if (HAS_DECL_IN_SOURCE)
+ target_compile_options(libgromacs_generated PRIVATE "-wd1419")
+endif()
+
+if(SIMD_AVX_512_CXX_SUPPORTED AND NOT ("${GMX_SIMD_ACTIVE}" STREQUAL "AVX_512_KNL"))
+ # Since we might be overriding -march=core-avx2, add a flag so we don't warn for this specific file.
+ # On KNL this can cause illegal instruction because the compiler might use non KNL AVX instructions
+ # with the SIMD_AVX_512_CXX_FLAGS flags.
+ set_source_files_properties(hardware/identifyavx512fmaunits.cpp PROPERTIES COMPILE_FLAGS "${SIMD_AVX_512_CXX_FLAGS} ${CXX_NO_UNUSED_OPTION_WARNING_FLAGS}")
+endif()
+
+gmx_setup_tng_for_libgromacs()
+
+target_link_libraries(libgromacs
+ PRIVATE
+ ${EXTRAE_LIBRARIES}
+ ${GMX_EXTRA_LIBRARIES}
+ ${GMX_COMMON_LIBRARIES}
+ ${FFT_LIBRARIES} ${LINEAR_ALGEBRA_LIBRARIES}
+ ${THREAD_LIB} ${GMX_SHARED_LINKER_FLAGS}
+ ${OpenCL_LIBRARIES}
+ ${GMX_STDLIB_LIBRARIES}
+ PUBLIC
+ ${GMX_PUBLIC_LIBRARIES}
+ )
+set_target_properties(libgromacs PROPERTIES
+ OUTPUT_NAME "gromacs${GMX_LIBS_SUFFIX}"
+ SOVERSION ${LIBRARY_SOVERSION_MAJOR}
+ VERSION ${LIBRARY_VERSION}
+ COMPILE_FLAGS "${OpenMP_C_FLAGS}")
+
+gmx_manage_lmfit()
+target_link_libraries(libgromacs PRIVATE lmfit)
+
+if (CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND CMAKE_CXX_COMPILER_VERSION MATCHES "^6\.0")
+ target_compile_options(libgromacs PRIVATE $<$<COMPILE_LANGUAGE:CXX>:-Weverything ${IGNORED_CLANG_ALL_WARNINGS}>)
+endif()
+if (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
+ target_compile_options(libgromacs PRIVATE $<$<COMPILE_LANGUAGE:CXX>:/analyze /analyze:stacksize 70000
+ #Control flow warnings are disabled because the commond line output is insufficient. There is no tool
+ #to convert the xml report to e.g. HTML and even in Visual Studio the viewer doesn't work with cmake support.
+ /wd6001 #unitialized memory
+ /wd6011 #derefencing NULL
+ /wd6053 #prior call not zero-terminate
+ /wd6054 #might not be zero-terminated
+ /wd6385 #reading invalid data
+ /wd6386 #buffer overrun
+ /wd6387 #could be '0'
+ /wd28199 #uninitialized memory
+ # For compile time constant (e.g. templates) the following warnings have flase postives
+ /wd6239 #(<non-zero> && <expr>)
+ /wd6240 #(<expr> && <non-zero>)
+ /wd6294 #Ill-defined for-loop
+ /wd6326 #comparison of constant with other constant
+ /wd28020 #expression involving paramter is not true
+ # Misc
+ /wd6330 #incorrect type to function (warns for char (instead of unsigned) for isspace/isalpha/isdigit/..))
+ /wd6993 #OpenMP ignored
+ #TODO
+ /wd6031 #return value ignored (important - mostly warnigns about sscanf)
+ /wd6244 #hides declaration (known issue - we ingore similar warnings for other compilers)
+ /wd6246 #hides declaration
+ >
+ )
+endif()
+
+if (GMX_CLANG_TIDY)
+ set_target_properties(libgromacs PROPERTIES CXX_CLANG_TIDY
+ "${CLANG_TIDY_EXE};-warnings-as-errors=*")
+endif()
+
+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 ${CMAKE_INSTALL_LIBDIR}
+ RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
+ ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
+ 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
+ ${CMAKE_INSTALL_LIBDIR} COMPONENT libraries)
+ endif()
+ endforeach()
+ else()
+ message(WARNING "INSTALL_CUDART_LIB only makes sense when configuring for CUDA support")
+ endif()
+endif()
+
+if(GMX_USE_OPENCL)
+ # Install the utility headers
+ file(GLOB OPENCL_INSTALLED_FILES
+ gpu_utils/vectype_ops.clh
+ gpu_utils/device_utils.clh
+ )
+ install(FILES ${OPENCL_INSTALLED_FILES}
+ DESTINATION ${GMX_INSTALL_OCLDIR}/gromacs/gpu_utils
+ COMPONENT libraries)
+ file(GLOB OPENCL_INSTALLED_FILES
+ pbcutil/ishift.h
+ )
+ install(FILES ${OPENCL_INSTALLED_FILES}
+ DESTINATION ${GMX_INSTALL_OCLDIR}/gromacs/pbcutil
+ COMPONENT libraries)
+
+ # Install the NB source and headers
+ file(GLOB OPENCL_INSTALLED_FILES
+ mdlib/nbnxn_consts.h
+ )
+ install(FILES ${OPENCL_INSTALLED_FILES}
+ DESTINATION ${GMX_INSTALL_OCLDIR}/gromacs/mdlib
+ COMPONENT libraries)
+ file(GLOB OPENCL_INSTALLED_FILES
+ mdlib/nbnxn_ocl/nbnxn_ocl_kernels.cl
+ mdlib/nbnxn_ocl/nbnxn_ocl_kernel.clh
+ mdlib/nbnxn_ocl/nbnxn_ocl_kernel_pruneonly.clh
+ mdlib/nbnxn_ocl/nbnxn_ocl_kernels.clh
+ mdlib/nbnxn_ocl/nbnxn_ocl_kernels_fastgen.clh
+ mdlib/nbnxn_ocl/nbnxn_ocl_kernels_fastgen_add_twincut.clh
+ mdlib/nbnxn_ocl/nbnxn_ocl_kernel_utils.clh
+ mdlib/nbnxn_ocl/nbnxn_ocl_consts.h
+ )
+ install(FILES ${OPENCL_INSTALLED_FILES}
+ DESTINATION ${GMX_INSTALL_OCLDIR}/gromacs/mdlib/nbnxn_ocl
+ COMPONENT libraries)
+
+ # Install the PME source and headers
+ file(GLOB OPENCL_INSTALLED_FILES
+ ewald/pme-spread.clh
+ ewald/pme-solve.clh
+ ewald/pme-gather.clh
+ ewald/pme-gpu-utils.clh
+ ewald/pme-program.cl
+ ewald/pme-gpu-types.h
+ )
+ install(FILES ${OPENCL_INSTALLED_FILES}
+ DESTINATION ${GMX_INSTALL_OCLDIR}/gromacs/ewald
+ COMPONENT libraries)
+endif()
diff --git a/patches/gromacs-2019.1.diff/src/gromacs/mdlib/force.cpp b/patches/gromacs-2019.1.diff/src/gromacs/mdlib/force.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..32bbd0cc08ab73c71ffa8cbfba8f0cdb32d8224e
--- /dev/null
+++ b/patches/gromacs-2019.1.diff/src/gromacs/mdlib/force.cpp
@@ -0,0 +1,879 @@
+/*
+ * 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) 2013,2014,2015,2016,2017,2018,2019, 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 <cassert>
+#include <cmath>
+#include <cstring>
+
+#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/force_flags.h"
+#include "gromacs/mdlib/forcerec-threading.h"
+#include "gromacs/mdlib/mdrun.h"
+#include "gromacs/mdlib/ns.h"
+#include "gromacs/mdlib/qmmm.h"
+#include "gromacs/mdlib/rf_util.h"
+#include "gromacs/mdlib/wall.h"
+#include "gromacs/mdtypes/commrec.h"
+#include "gromacs/mdtypes/enerdata.h"
+#include "gromacs/mdtypes/forceoutput.h"
+#include "gromacs/mdtypes/forcerec.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,
+ const gmx_groups_t *groups,
+ gmx_localtop_t *top,
+ const t_mdatoms *md,
+ const 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 clearEwaldThreadOutput(ewald_corr_thread_t *ewc_t)
+{
+ ewc_t->Vcorr_q = 0;
+ ewc_t->Vcorr_lj = 0;
+ ewc_t->dvdl[efptCOUL] = 0;
+ ewc_t->dvdl[efptVDW] = 0;
+ clear_mat(ewc_t->vir_q);
+ clear_mat(ewc_t->vir_lj);
+}
+
+static void reduceEwaldThreadOuput(int nthreads, ewald_corr_thread_t *ewc_t)
+{
+ ewald_corr_thread_t &dest = ewc_t[0];
+
+ for (int t = 1; t < nthreads; t++)
+ {
+ dest.Vcorr_q += ewc_t[t].Vcorr_q;
+ dest.Vcorr_lj += ewc_t[t].Vcorr_lj;
+ dest.dvdl[efptCOUL] += ewc_t[t].dvdl[efptCOUL];
+ dest.dvdl[efptVDW] += ewc_t[t].dvdl[efptVDW];
+ m_add(dest.vir_q, ewc_t[t].vir_q, dest.vir_q);
+ m_add(dest.vir_lj, ewc_t[t].vir_lj, dest.vir_lj);
+ }
+}
+
+void do_force_lowlevel(t_forcerec *fr,
+ const t_inputrec *ir,
+ const t_idef *idef,
+ const t_commrec *cr,
+ const gmx_multisim_t *ms,
+ t_nrnb *nrnb,
+ gmx_wallcycle_t wcycle,
+ const t_mdatoms *md,
+ rvec x[],
+ history_t *hist,
+ rvec *forceForUseWithShiftForces,
+ gmx::ForceWithVirial *forceWithVirial,
+ gmx_enerdata_t *enerd,
+ t_fcdata *fcd,
+ matrix box,
+ t_lambda *fepvals,
+ real *lambda,
+ const t_graph *graph,
+ const t_blocka *excl,
+ rvec mu_tot[],
+ int flags,
+ float *cycles_pme)
+{
+ int i, j;
+ int donb_flags;
+ int pme_flags;
+ 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;
+ }
+
+ /* do QMMM first if requested */
+ if (fr->bQMMM)
+ {
+ enerd->term[F_EQM] = calculate_QMMM(cr, forceForUseWithShiftForces, 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,
+ forceWithVirial, lambda[efptVDW],
+ enerd->grpp.ener[egLJSR], nrnb);
+ enerd->dvdl_lin[efptVDW] += dvdl_walls;
+ }
+
+ /* 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, forceForUseWithShiftForces, 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, forceForUseWithShiftForces, 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);
+ }
+
+#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->ic->eeltype) || EEL_FULL(fr->ic->eeltype) || EVDW_PME(fr->ic->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, x, hist,
+ forceForUseWithShiftForces, forceWithVirial,
+ fr, &pbc, graph, enerd, nrnb, lambda, md, fcd,
+ DOMAINDECOMP(cr) ? cr->dd->globalAtomIndices.data() : nullptr,
+ flags);
+
+
+ *cycles_pme = 0;
+
+ /* Do long-range electrostatics and/or LJ-PME, including related short-range
+ * corrections.
+ */
+ if (EEL_FULL(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
+ {
+ int status = 0;
+ real Vlr_q = 0, Vlr_lj = 0;
+
+ /* We reduce all virial, dV/dlambda and energy contributions, except
+ * for the reciprocal energies (Vlr_q, Vlr_lj) into the same struct.
+ */
+ ewald_corr_thread_t &ewaldOutput = fr->ewc_t[0];
+ clearEwaldThreadOutput(&ewaldOutput);
+
+ if (EEL_PME_EWALD(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
+ {
+ /* 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
+ {
+ ewald_corr_thread_t &ewc_t = fr->ewc_t[t];
+ if (t > 0)
+ {
+ clearEwaldThreadOutput(&ewc_t);
+ }
+
+ /* 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 forceWithVirial->force_ array.
+ */
+ ewald_LRcorrection(md->homenr, cr, nthreads, t, fr, ir,
+ md->chargeA, md->chargeB,
+ md->sqrt_c6A, md->sqrt_c6B,
+ md->sigmaA, md->sigmaB,
+ md->sigma3A, md->sigma3B,
+ (md->nChargePerturbed != 0) || (md->nTypePerturbed != 0),
+ ir->cutoff_scheme != ecutsVERLET,
+ excl, x, box, mu_tot,
+ ir->ewald_geometry,
+ ir->epsilon_surface,
+ as_rvec_array(forceWithVirial->force_.data()),
+ ewc_t.vir_q, ewc_t.vir_lj,
+ &ewc_t.Vcorr_q, &ewc_t.Vcorr_lj,
+ lambda[efptCOUL], lambda[efptVDW],
+ &ewc_t.dvdl[efptCOUL], &ewc_t.dvdl[efptVDW]);
+ }
+ GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+ }
+ if (nthreads > 1)
+ {
+ reduceEwaldThreadOuput(nthreads, fr->ewc_t);
+ }
+ wallcycle_sub_stop(wcycle, ewcsEWALD_CORRECTION);
+ }
+
+ if (EEL_PME_EWALD(fr->ic->eeltype) && fr->n_tpi == 0)
+ {
+ /* This is not in a subcounter because it takes a
+ negligible and constant-sized amount of time */
+ ewaldOutput.Vcorr_q +=
+ ewald_charge_correction(cr, fr, lambda[efptCOUL], box,
+ &ewaldOutput.dvdl[efptCOUL],
+ ewaldOutput.vir_q);
+ }
+
+ if ((EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype)) &&
+ thisRankHasDuty(cr, DUTY_PME) && (pme_run_mode(fr->pmedata) == PmeRunMode::CPU))
+ {
+ /* 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 (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;
+ }
+
+ /* With domain decomposition we close the CPU side load
+ * balancing region here, because PME does global
+ * communication that acts as a global barrier.
+ */
+ if (DOMAINDECOMP(cr))
+ {
+ ddCloseBalanceRegionCpu(cr->dd);
+ }
+
+ wallcycle_start(wcycle, ewcPMEMESH);
+ status = gmx_pme_do(fr->pmedata,
+ 0, md->homenr - fr->n_tpi,
+ x,
+ as_rvec_array(forceWithVirial->force_.data()),
+ md->chargeA, md->chargeB,
+ md->sqrt_c6A, md->sqrt_c6B,
+ md->sigmaA, md->sigmaB,
+ box, cr,
+ DOMAINDECOMP(cr) ? dd_pme_maxshift_x(cr->dd) : 0,
+ DOMAINDECOMP(cr) ? dd_pme_maxshift_y(cr->dd) : 0,
+ nrnb, wcycle,
+ ewaldOutput.vir_q, ewaldOutput.vir_lj,
+ &Vlr_q, &Vlr_lj,
+ lambda[efptCOUL], lambda[efptVDW],
+ &ewaldOutput.dvdl[efptCOUL],
+ &ewaldOutput.dvdl[efptVDW],
+ 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->ic->eeltype) && EEL_PME_EWALD(fr->ic->eeltype))
+ {
+ Vlr_q = do_ewald(ir, x, as_rvec_array(forceWithVirial->force_.data()),
+ md->chargeA, md->chargeB,
+ box, cr, md->homenr,
+ ewaldOutput.vir_q, fr->ic->ewaldcoeff_q,
+ lambda[efptCOUL], &ewaldOutput.dvdl[efptCOUL],
+ fr->ewald_table);
+ }
+
+ /* Note that with separate PME nodes we get the real energies later */
+ // TODO it would be simpler if we just accumulated a single
+ // long-range virial contribution.
+ forceWithVirial->addVirialContribution(ewaldOutput.vir_q);
+ forceWithVirial->addVirialContribution(ewaldOutput.vir_lj);
+ enerd->dvdl_lin[efptCOUL] += ewaldOutput.dvdl[efptCOUL];
+ enerd->dvdl_lin[efptVDW] += ewaldOutput.dvdl[efptVDW];
+ enerd->term[F_COUL_RECIP] = Vlr_q + ewaldOutput.Vcorr_q;
+ enerd->term[F_LJ_RECIP] = Vlr_lj + ewaldOutput.Vcorr_lj;
+
+ if (debug)
+ {
+ fprintf(debug, "Vlr_q = %g, Vcorr_q = %g, Vlr_corr_q = %g\n",
+ Vlr_q, ewaldOutput.Vcorr_q, enerd->term[F_COUL_RECIP]);
+ pr_rvecs(debug, 0, "vir_el_recip after corr", ewaldOutput.vir_q, 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, ewaldOutput.Vcorr_lj, enerd->term[F_LJ_RECIP]);
+ pr_rvecs(debug, 0, "vir_lj_recip after corr", ewaldOutput.vir_lj, 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->ic->eeltype))
+ {
+ real dvdl_rf_excl = 0;
+ enerd->term[F_RF_EXCL] =
+ RF_excl_correction(fr, graph, md, excl, DOMAINDECOMP(cr),
+ x, forceForUseWithShiftForces,
+ fr->fshift, &pbc, lambda[efptCOUL], &dvdl_rf_excl);
+
+ enerd->dvdl_lin[efptCOUL] += dvdl_rf_excl;
+ }
+ }
+
+ 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, const 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]);
+
+/* 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, gmx::ArrayRef<const real> lambda, t_lambda *fepvals)
+{
+ int index;
+
+ enerd->dvdl_lin[efptVDW] += enerd->term[F_DVDL_VDW]; /* include dispersion correction */
+ enerd->term[F_DVDL] = 0.0;
+ for (int 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]);
+ }
+ }
+ }
+
+ 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];
+ }
+
+ for (int 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 */
+
+ double &enerpart_lambda = enerd->enerpart_lambda[i + 1];
+
+ for (gmx::index j = 0; j < lambda.size(); j++)
+ {
+ /* Note that this loop is over all dhdl components, not just the separated ones */
+ const double dlam = fepvals->all_lambda[j][i] - lambda[j];
+
+ enerpart_lambda += dlam*enerd->dvdl_lin[j];
+
+ /* Constraints can not be evaluated at foreign lambdas, so we add
+ * a linear extrapolation. This is an approximation, but usually
+ * quite accurate since constraints change little between lambdas.
+ */
+ if ((j == efptBONDED && fepvals->separate_dvdl[efptBONDED]) ||
+ (j == efptFEP && !fepvals->separate_dvdl[efptBONDED]))
+ {
+ enerpart_lambda += dlam*enerd->term[F_DVDL_CONSTR];
+ }
+
+ if (j == efptMASS && !fepvals->separate_dvdl[j])
+ {
+ enerpart_lambda += dlam*enerd->term[F_DKDL];
+ }
+
+ if (debug)
+ {
+ fprintf(debug, "enerdiff lam %g: (%15s), non-linear %f linear %f*%f\n",
+ fepvals->all_lambda[j][i], efpt_names[j],
+ enerpart_lambda - enerd->enerpart_lambda[0],
+ dlam, enerd->dvdl_lin[j]);
+ }
+ }
+ }
+
+ /* The constrain contribution is now included in other terms, so clear it */
+ enerd->term[F_DVDL_CONSTR] = 0;
+}
+
+
+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-2019.1.diff/src/gromacs/mdlib/force.cpp.preplumed b/patches/gromacs-2019.1.diff/src/gromacs/mdlib/force.cpp.preplumed
new file mode 100644
index 0000000000000000000000000000000000000000..8a7615365140ca43108e1a445b360f36696a93b3
--- /dev/null
+++ b/patches/gromacs-2019.1.diff/src/gromacs/mdlib/force.cpp.preplumed
@@ -0,0 +1,866 @@
+/*
+ * 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) 2013,2014,2015,2016,2017,2018,2019, 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 <cassert>
+#include <cmath>
+#include <cstring>
+
+#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/force_flags.h"
+#include "gromacs/mdlib/forcerec-threading.h"
+#include "gromacs/mdlib/mdrun.h"
+#include "gromacs/mdlib/ns.h"
+#include "gromacs/mdlib/qmmm.h"
+#include "gromacs/mdlib/rf_util.h"
+#include "gromacs/mdlib/wall.h"
+#include "gromacs/mdtypes/commrec.h"
+#include "gromacs/mdtypes/enerdata.h"
+#include "gromacs/mdtypes/forceoutput.h"
+#include "gromacs/mdtypes/forcerec.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,
+ const gmx_groups_t *groups,
+ gmx_localtop_t *top,
+ const t_mdatoms *md,
+ const 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 clearEwaldThreadOutput(ewald_corr_thread_t *ewc_t)
+{
+ ewc_t->Vcorr_q = 0;
+ ewc_t->Vcorr_lj = 0;
+ ewc_t->dvdl[efptCOUL] = 0;
+ ewc_t->dvdl[efptVDW] = 0;
+ clear_mat(ewc_t->vir_q);
+ clear_mat(ewc_t->vir_lj);
+}
+
+static void reduceEwaldThreadOuput(int nthreads, ewald_corr_thread_t *ewc_t)
+{
+ ewald_corr_thread_t &dest = ewc_t[0];
+
+ for (int t = 1; t < nthreads; t++)
+ {
+ dest.Vcorr_q += ewc_t[t].Vcorr_q;
+ dest.Vcorr_lj += ewc_t[t].Vcorr_lj;
+ dest.dvdl[efptCOUL] += ewc_t[t].dvdl[efptCOUL];
+ dest.dvdl[efptVDW] += ewc_t[t].dvdl[efptVDW];
+ m_add(dest.vir_q, ewc_t[t].vir_q, dest.vir_q);
+ m_add(dest.vir_lj, ewc_t[t].vir_lj, dest.vir_lj);
+ }
+}
+
+void do_force_lowlevel(t_forcerec *fr,
+ const t_inputrec *ir,
+ const t_idef *idef,
+ const t_commrec *cr,
+ const gmx_multisim_t *ms,
+ t_nrnb *nrnb,
+ gmx_wallcycle_t wcycle,
+ const t_mdatoms *md,
+ rvec x[],
+ history_t *hist,
+ rvec *forceForUseWithShiftForces,
+ gmx::ForceWithVirial *forceWithVirial,
+ gmx_enerdata_t *enerd,
+ t_fcdata *fcd,
+ matrix box,
+ t_lambda *fepvals,
+ real *lambda,
+ const t_graph *graph,
+ const t_blocka *excl,
+ rvec mu_tot[],
+ int flags,
+ float *cycles_pme)
+{
+ int i, j;
+ int donb_flags;
+ int pme_flags;
+ 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;
+ }
+
+ /* do QMMM first if requested */
+ if (fr->bQMMM)
+ {
+ enerd->term[F_EQM] = calculate_QMMM(cr, forceForUseWithShiftForces, 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,
+ forceWithVirial, lambda[efptVDW],
+ enerd->grpp.ener[egLJSR], nrnb);
+ enerd->dvdl_lin[efptVDW] += dvdl_walls;
+ }
+
+ /* 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, forceForUseWithShiftForces, 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, forceForUseWithShiftForces, 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);
+ }
+
+#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->ic->eeltype) || EEL_FULL(fr->ic->eeltype) || EVDW_PME(fr->ic->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, x, hist,
+ forceForUseWithShiftForces, forceWithVirial,
+ fr, &pbc, graph, enerd, nrnb, lambda, md, fcd,
+ DOMAINDECOMP(cr) ? cr->dd->globalAtomIndices.data() : nullptr,
+ flags);
+
+
+ *cycles_pme = 0;
+
+ /* Do long-range electrostatics and/or LJ-PME, including related short-range
+ * corrections.
+ */
+ if (EEL_FULL(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
+ {
+ int status = 0;
+ real Vlr_q = 0, Vlr_lj = 0;
+
+ /* We reduce all virial, dV/dlambda and energy contributions, except
+ * for the reciprocal energies (Vlr_q, Vlr_lj) into the same struct.
+ */
+ ewald_corr_thread_t &ewaldOutput = fr->ewc_t[0];
+ clearEwaldThreadOutput(&ewaldOutput);
+
+ if (EEL_PME_EWALD(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
+ {
+ /* 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
+ {
+ ewald_corr_thread_t &ewc_t = fr->ewc_t[t];
+ if (t > 0)
+ {
+ clearEwaldThreadOutput(&ewc_t);
+ }
+
+ /* 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 forceWithVirial->force_ array.
+ */
+ ewald_LRcorrection(md->homenr, cr, nthreads, t, fr, ir,
+ md->chargeA, md->chargeB,
+ md->sqrt_c6A, md->sqrt_c6B,
+ md->sigmaA, md->sigmaB,
+ md->sigma3A, md->sigma3B,
+ (md->nChargePerturbed != 0) || (md->nTypePerturbed != 0),
+ ir->cutoff_scheme != ecutsVERLET,
+ excl, x, box, mu_tot,
+ ir->ewald_geometry,
+ ir->epsilon_surface,
+ as_rvec_array(forceWithVirial->force_.data()),
+ ewc_t.vir_q, ewc_t.vir_lj,
+ &ewc_t.Vcorr_q, &ewc_t.Vcorr_lj,
+ lambda[efptCOUL], lambda[efptVDW],
+ &ewc_t.dvdl[efptCOUL], &ewc_t.dvdl[efptVDW]);
+ }
+ GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+ }
+ if (nthreads > 1)
+ {
+ reduceEwaldThreadOuput(nthreads, fr->ewc_t);
+ }
+ wallcycle_sub_stop(wcycle, ewcsEWALD_CORRECTION);
+ }
+
+ if (EEL_PME_EWALD(fr->ic->eeltype) && fr->n_tpi == 0)
+ {
+ /* This is not in a subcounter because it takes a
+ negligible and constant-sized amount of time */
+ ewaldOutput.Vcorr_q +=
+ ewald_charge_correction(cr, fr, lambda[efptCOUL], box,
+ &ewaldOutput.dvdl[efptCOUL],
+ ewaldOutput.vir_q);
+ }
+
+ if ((EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype)) &&
+ thisRankHasDuty(cr, DUTY_PME) && (pme_run_mode(fr->pmedata) == PmeRunMode::CPU))
+ {
+ /* 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 (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;
+ }
+
+ /* With domain decomposition we close the CPU side load
+ * balancing region here, because PME does global
+ * communication that acts as a global barrier.
+ */
+ if (DOMAINDECOMP(cr))
+ {
+ ddCloseBalanceRegionCpu(cr->dd);
+ }
+
+ wallcycle_start(wcycle, ewcPMEMESH);
+ status = gmx_pme_do(fr->pmedata,
+ 0, md->homenr - fr->n_tpi,
+ x,
+ as_rvec_array(forceWithVirial->force_.data()),
+ md->chargeA, md->chargeB,
+ md->sqrt_c6A, md->sqrt_c6B,
+ md->sigmaA, md->sigmaB,
+ box, cr,
+ DOMAINDECOMP(cr) ? dd_pme_maxshift_x(cr->dd) : 0,
+ DOMAINDECOMP(cr) ? dd_pme_maxshift_y(cr->dd) : 0,
+ nrnb, wcycle,
+ ewaldOutput.vir_q, ewaldOutput.vir_lj,
+ &Vlr_q, &Vlr_lj,
+ lambda[efptCOUL], lambda[efptVDW],
+ &ewaldOutput.dvdl[efptCOUL],
+ &ewaldOutput.dvdl[efptVDW],
+ 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->ic->eeltype) && EEL_PME_EWALD(fr->ic->eeltype))
+ {
+ Vlr_q = do_ewald(ir, x, as_rvec_array(forceWithVirial->force_.data()),
+ md->chargeA, md->chargeB,
+ box, cr, md->homenr,
+ ewaldOutput.vir_q, fr->ic->ewaldcoeff_q,
+ lambda[efptCOUL], &ewaldOutput.dvdl[efptCOUL],
+ fr->ewald_table);
+ }
+
+ /* Note that with separate PME nodes we get the real energies later */
+ // TODO it would be simpler if we just accumulated a single
+ // long-range virial contribution.
+ forceWithVirial->addVirialContribution(ewaldOutput.vir_q);
+ forceWithVirial->addVirialContribution(ewaldOutput.vir_lj);
+ enerd->dvdl_lin[efptCOUL] += ewaldOutput.dvdl[efptCOUL];
+ enerd->dvdl_lin[efptVDW] += ewaldOutput.dvdl[efptVDW];
+ enerd->term[F_COUL_RECIP] = Vlr_q + ewaldOutput.Vcorr_q;
+ enerd->term[F_LJ_RECIP] = Vlr_lj + ewaldOutput.Vcorr_lj;
+
+ if (debug)
+ {
+ fprintf(debug, "Vlr_q = %g, Vcorr_q = %g, Vlr_corr_q = %g\n",
+ Vlr_q, ewaldOutput.Vcorr_q, enerd->term[F_COUL_RECIP]);
+ pr_rvecs(debug, 0, "vir_el_recip after corr", ewaldOutput.vir_q, 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, ewaldOutput.Vcorr_lj, enerd->term[F_LJ_RECIP]);
+ pr_rvecs(debug, 0, "vir_lj_recip after corr", ewaldOutput.vir_lj, 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->ic->eeltype))
+ {
+ real dvdl_rf_excl = 0;
+ enerd->term[F_RF_EXCL] =
+ RF_excl_correction(fr, graph, md, excl, DOMAINDECOMP(cr),
+ x, forceForUseWithShiftForces,
+ fr->fshift, &pbc, lambda[efptCOUL], &dvdl_rf_excl);
+
+ enerd->dvdl_lin[efptCOUL] += dvdl_rf_excl;
+ }
+ }
+
+ 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, const 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]);
+
+/* 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, gmx::ArrayRef<const real> lambda, t_lambda *fepvals)
+{
+ int index;
+
+ enerd->dvdl_lin[efptVDW] += enerd->term[F_DVDL_VDW]; /* include dispersion correction */
+ enerd->term[F_DVDL] = 0.0;
+ for (int 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]);
+ }
+ }
+ }
+
+ 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];
+ }
+
+ for (int 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 */
+
+ double &enerpart_lambda = enerd->enerpart_lambda[i + 1];
+
+ for (gmx::index j = 0; j < lambda.size(); j++)
+ {
+ /* Note that this loop is over all dhdl components, not just the separated ones */
+ const double dlam = fepvals->all_lambda[j][i] - lambda[j];
+
+ enerpart_lambda += dlam*enerd->dvdl_lin[j];
+
+ /* Constraints can not be evaluated at foreign lambdas, so we add
+ * a linear extrapolation. This is an approximation, but usually
+ * quite accurate since constraints change little between lambdas.
+ */
+ if ((j == efptBONDED && fepvals->separate_dvdl[efptBONDED]) ||
+ (j == efptFEP && !fepvals->separate_dvdl[efptBONDED]))
+ {
+ enerpart_lambda += dlam*enerd->term[F_DVDL_CONSTR];
+ }
+
+ if (j == efptMASS && !fepvals->separate_dvdl[j])
+ {
+ enerpart_lambda += dlam*enerd->term[F_DKDL];
+ }
+
+ if (debug)
+ {
+ fprintf(debug, "enerdiff lam %g: (%15s), non-linear %f linear %f*%f\n",
+ fepvals->all_lambda[j][i], efpt_names[j],
+ enerpart_lambda - enerd->enerpart_lambda[0],
+ dlam, enerd->dvdl_lin[j]);
+ }
+ }
+ }
+
+ /* The constrain contribution is now included in other terms, so clear it */
+ enerd->term[F_DVDL_CONSTR] = 0;
+}
+
+
+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-2019.1.diff/src/gromacs/mdrun/legacymdrunoptions.cpp b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/legacymdrunoptions.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..4d9c29a1d5ad78296794ffa6fd2f8a6599aa560c
--- /dev/null
+++ b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/legacymdrunoptions.cpp
@@ -0,0 +1,263 @@
+/*
+ * 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) 2011,2012,2013,2014,2015,2016,2017,2018, 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 declares helper functionality for legacy option handling for 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 "legacymdrunoptions.h"
+
+#include "config.h"
+
+#include <cstring>
+
+#include "gromacs/gmxlib/network.h"
+#include "gromacs/math/functions.h"
+#include "gromacs/mdrun/multisim.h"
+#include "gromacs/mdrunutility/handlerestart.h"
+#include "gromacs/mdtypes/commrec.h"
+#include "gromacs/utility/arraysize.h"
+#include "gromacs/utility/fatalerror.h"
+
+/* PLUMED */
+#include "../../../Plumed.h"
+extern int plumedswitch;
+extern plumed plumedmain;
+extern void(*plumedcmd)(plumed,const char*,const void*);
+/* END PLUMED */
+
+namespace gmx
+{
+
+/*! \brief Return whether the command-line parameter 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], "-multidir") == 0)
+ {
+ return true;
+ }
+ }
+ return false;
+}
+
+int LegacyMdrunOptions::updateFromCommandLine(int argc, char **argv, ArrayRef<const char *> desc)
+{
+ unsigned long PCA_Flags = PCA_CAN_SET_DEFFNM;
+ // With -multidir, the working directory still needs to be
+ // changed, so we can't check for the existence of files 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;
+ }
+
+ if (!parse_common_args(&argc, argv, PCA_Flags,
+ static_cast<int>(filenames.size()), filenames.data(), asize(pa), pa,
+ static_cast<int>(desc.size()), desc.data(), 0, nullptr, &oenv))
+ {
+ return 0;
+ }
+
+ // Handle the options that permits the user to either declare
+ // which compatible GPUs are availble for use, or to select a GPU
+ // task assignment. Either could be in an environment variable (so
+ // that there is a way to customize it, when using MPI in
+ // heterogeneous contexts).
+ {
+ // TODO Argument parsing can't handle std::string. We should
+ // fix that by changing the parsing, once more of the roles of
+ // handling, validating and implementing defaults for user
+ // command-line options have been seperated.
+ hw_opt.gpuIdsAvailable = gpuIdsAvailable;
+ hw_opt.userGpuTaskAssignment = userGpuTaskAssignment;
+
+ const char *env = getenv("GMX_GPU_ID");
+ if (env != nullptr)
+ {
+ if (!hw_opt.gpuIdsAvailable.empty())
+ {
+ gmx_fatal(FARGS, "GMX_GPU_ID and -gpu_id can not be used at the same time");
+ }
+ hw_opt.gpuIdsAvailable = env;
+ }
+
+ env = getenv("GMX_GPUTASKS");
+ if (env != nullptr)
+ {
+ if (!hw_opt.userGpuTaskAssignment.empty())
+ {
+ gmx_fatal(FARGS, "GMX_GPUTASKS and -gputasks can not be used at the same time");
+ }
+ hw_opt.userGpuTaskAssignment = env;
+ }
+
+ if (!hw_opt.gpuIdsAvailable.empty() && !hw_opt.userGpuTaskAssignment.empty())
+ {
+ gmx_fatal(FARGS, "-gpu_id and -gputasks cannot be used at the same time");
+ }
+ }
+
+ hw_opt.thread_affinity = nenum(thread_aff_opt_choices);
+
+ // now check for a multi-simulation
+ ArrayRef<const std::string> multidir = opt2fnsIfOptionSet("-multidir",
+ static_cast<int>(filenames.size()),
+ filenames.data());
+
+ if (replExParams.exchangeInterval != 0 && multidir.size() < 2)
+ {
+ gmx_fatal(FARGS, "Need at least two replicas for replica exchange (use option -multidir)");
+ }
+
+ if (replExParams.numExchanges < 0)
+ {
+ gmx_fatal(FARGS, "Replica exchange number of exchanges needs to be positive");
+ }
+
+ ms = init_multisystem(MPI_COMM_WORLD, multidir);
+
+ /* Prepare the intra-simulation communication */
+ // TODO consolidate this with init_commrec, after changing the
+ // relative ordering of init_commrec and init_multisystem
+#if GMX_MPI
+ if (ms != nullptr)
+ {
+ cr->nnodes = cr->nnodes / ms->nsim;
+ MPI_Comm_split(MPI_COMM_WORLD, ms->sim, cr->sim_nodeid, &cr->mpi_comm_mysim);
+ cr->mpi_comm_mygroup = cr->mpi_comm_mysim;
+ MPI_Comm_rank(cr->mpi_comm_mysim, &cr->sim_nodeid);
+ MPI_Comm_rank(cr->mpi_comm_mygroup, &cr->nodeid);
+ }
+#endif
+
+ if (!opt2bSet("-cpi",
+ static_cast<int>(filenames.size()), filenames.data()))
+ {
+ // If we are not starting from a checkpoint we never allow files to be appended
+ // to, since that has caused a ton of strange behaviour and bugs in the past.
+ if (opt2parg_bSet("-append", asize(pa), pa))
+ {
+ // If the user explicitly used the -append option, explain that it is not possible.
+ gmx_fatal(FARGS, "GROMACS can only append to files when restarting from a checkpoint.");
+ }
+ else
+ {
+ // If the user did not say anything explicit, just disable appending.
+ bTryToAppendFiles = FALSE;
+ }
+ }
+
+ ContinuationOptions &continuationOptions = mdrunOptions.continuationOptions;
+
+ continuationOptions.appendFilesOptionSet = opt2parg_bSet("-append", asize(pa), pa);
+
+ handleRestart(cr, ms, bTryToAppendFiles,
+ static_cast<int>(filenames.size()),
+ filenames.data(),
+ &continuationOptions.appendFiles,
+ &continuationOptions.startedFromCheckpoint);
+
+ mdrunOptions.rerun = opt2bSet("-rerun",
+ static_cast<int>(filenames.size()),
+ filenames.data());
+ mdrunOptions.ntompOptionIsSet = opt2parg_bSet("-ntomp", asize(pa), pa);
+
+ domdecOptions.rankOrder = static_cast<DdRankOrder>(nenum(ddrank_opt_choices));
+ domdecOptions.dlbOption = static_cast<DlbOption>(nenum(dddlb_opt_choices));
+ domdecOptions.numCells[XX] = roundToInt(realddxyz[XX]);
+ domdecOptions.numCells[YY] = roundToInt(realddxyz[YY]);
+ domdecOptions.numCells[ZZ] = roundToInt(realddxyz[ZZ]);
+
+ /* PLUMED */
+ plumedswitch=0;
+ if (opt2bSet("-plumed", static_cast<int>(filenames.size()), filenames.data())) 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,static_cast<int>(filenames.size()), filenames.data()));
+ plumedswitch=1;
+ }
+ /* PLUMED HREX*/
+ if(getenv("PLUMED_HREX")) plumed_hrex=1;
+ if(plumed_hrex){
+ if(!plumedswitch) gmx_fatal(FARGS,"-hrex (or PLUMED_HREX) requires -plumed");
+ if(replExParams.exchangeInterval==0) gmx_fatal(FARGS,"-hrex (or PLUMED_HREX) replica exchange");
+ if(replExParams.numExchanges!=0) gmx_fatal(FARGS,"-hrex (or PLUMED_HREX) not compatible with -nex");
+ }
+ /* END PLUMED HREX */
+
+ /* END PLUMED */
+
+ return 1;
+}
+
+LegacyMdrunOptions::~LegacyMdrunOptions()
+{
+ if (GMX_LIB_MPI)
+ {
+ done_commrec(cr);
+ }
+ done_multisim(ms);
+}
+
+} // namespace gmx
diff --git a/patches/gromacs-2019.1.diff/src/gromacs/mdrun/legacymdrunoptions.cpp.preplumed b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/legacymdrunoptions.cpp.preplumed
new file mode 100644
index 0000000000000000000000000000000000000000..de685e90348bfff377e9fe6cf20efa9fc2d516d6
--- /dev/null
+++ b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/legacymdrunoptions.cpp.preplumed
@@ -0,0 +1,222 @@
+/*
+ * 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) 2011,2012,2013,2014,2015,2016,2017,2018, 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 declares helper functionality for legacy option handling for 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 "legacymdrunoptions.h"
+
+#include "config.h"
+
+#include <cstring>
+
+#include "gromacs/gmxlib/network.h"
+#include "gromacs/math/functions.h"
+#include "gromacs/mdrun/multisim.h"
+#include "gromacs/mdrunutility/handlerestart.h"
+#include "gromacs/mdtypes/commrec.h"
+#include "gromacs/utility/arraysize.h"
+#include "gromacs/utility/fatalerror.h"
+
+namespace gmx
+{
+
+/*! \brief Return whether the command-line parameter 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], "-multidir") == 0)
+ {
+ return true;
+ }
+ }
+ return false;
+}
+
+int LegacyMdrunOptions::updateFromCommandLine(int argc, char **argv, ArrayRef<const char *> desc)
+{
+ unsigned long PCA_Flags = PCA_CAN_SET_DEFFNM;
+ // With -multidir, the working directory still needs to be
+ // changed, so we can't check for the existence of files 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;
+ }
+
+ if (!parse_common_args(&argc, argv, PCA_Flags,
+ static_cast<int>(filenames.size()), filenames.data(), asize(pa), pa,
+ static_cast<int>(desc.size()), desc.data(), 0, nullptr, &oenv))
+ {
+ return 0;
+ }
+
+ // Handle the options that permits the user to either declare
+ // which compatible GPUs are availble for use, or to select a GPU
+ // task assignment. Either could be in an environment variable (so
+ // that there is a way to customize it, when using MPI in
+ // heterogeneous contexts).
+ {
+ // TODO Argument parsing can't handle std::string. We should
+ // fix that by changing the parsing, once more of the roles of
+ // handling, validating and implementing defaults for user
+ // command-line options have been seperated.
+ hw_opt.gpuIdsAvailable = gpuIdsAvailable;
+ hw_opt.userGpuTaskAssignment = userGpuTaskAssignment;
+
+ const char *env = getenv("GMX_GPU_ID");
+ if (env != nullptr)
+ {
+ if (!hw_opt.gpuIdsAvailable.empty())
+ {
+ gmx_fatal(FARGS, "GMX_GPU_ID and -gpu_id can not be used at the same time");
+ }
+ hw_opt.gpuIdsAvailable = env;
+ }
+
+ env = getenv("GMX_GPUTASKS");
+ if (env != nullptr)
+ {
+ if (!hw_opt.userGpuTaskAssignment.empty())
+ {
+ gmx_fatal(FARGS, "GMX_GPUTASKS and -gputasks can not be used at the same time");
+ }
+ hw_opt.userGpuTaskAssignment = env;
+ }
+
+ if (!hw_opt.gpuIdsAvailable.empty() && !hw_opt.userGpuTaskAssignment.empty())
+ {
+ gmx_fatal(FARGS, "-gpu_id and -gputasks cannot be used at the same time");
+ }
+ }
+
+ hw_opt.thread_affinity = nenum(thread_aff_opt_choices);
+
+ // now check for a multi-simulation
+ ArrayRef<const std::string> multidir = opt2fnsIfOptionSet("-multidir",
+ static_cast<int>(filenames.size()),
+ filenames.data());
+
+ if (replExParams.exchangeInterval != 0 && multidir.size() < 2)
+ {
+ gmx_fatal(FARGS, "Need at least two replicas for replica exchange (use option -multidir)");
+ }
+
+ if (replExParams.numExchanges < 0)
+ {
+ gmx_fatal(FARGS, "Replica exchange number of exchanges needs to be positive");
+ }
+
+ ms = init_multisystem(MPI_COMM_WORLD, multidir);
+
+ /* Prepare the intra-simulation communication */
+ // TODO consolidate this with init_commrec, after changing the
+ // relative ordering of init_commrec and init_multisystem
+#if GMX_MPI
+ if (ms != nullptr)
+ {
+ cr->nnodes = cr->nnodes / ms->nsim;
+ MPI_Comm_split(MPI_COMM_WORLD, ms->sim, cr->sim_nodeid, &cr->mpi_comm_mysim);
+ cr->mpi_comm_mygroup = cr->mpi_comm_mysim;
+ MPI_Comm_rank(cr->mpi_comm_mysim, &cr->sim_nodeid);
+ MPI_Comm_rank(cr->mpi_comm_mygroup, &cr->nodeid);
+ }
+#endif
+
+ if (!opt2bSet("-cpi",
+ static_cast<int>(filenames.size()), filenames.data()))
+ {
+ // If we are not starting from a checkpoint we never allow files to be appended
+ // to, since that has caused a ton of strange behaviour and bugs in the past.
+ if (opt2parg_bSet("-append", asize(pa), pa))
+ {
+ // If the user explicitly used the -append option, explain that it is not possible.
+ gmx_fatal(FARGS, "GROMACS can only append to files when restarting from a checkpoint.");
+ }
+ else
+ {
+ // If the user did not say anything explicit, just disable appending.
+ bTryToAppendFiles = FALSE;
+ }
+ }
+
+ ContinuationOptions &continuationOptions = mdrunOptions.continuationOptions;
+
+ continuationOptions.appendFilesOptionSet = opt2parg_bSet("-append", asize(pa), pa);
+
+ handleRestart(cr, ms, bTryToAppendFiles,
+ static_cast<int>(filenames.size()),
+ filenames.data(),
+ &continuationOptions.appendFiles,
+ &continuationOptions.startedFromCheckpoint);
+
+ mdrunOptions.rerun = opt2bSet("-rerun",
+ static_cast<int>(filenames.size()),
+ filenames.data());
+ mdrunOptions.ntompOptionIsSet = opt2parg_bSet("-ntomp", asize(pa), pa);
+
+ domdecOptions.rankOrder = static_cast<DdRankOrder>(nenum(ddrank_opt_choices));
+ domdecOptions.dlbOption = static_cast<DlbOption>(nenum(dddlb_opt_choices));
+ domdecOptions.numCells[XX] = roundToInt(realddxyz[XX]);
+ domdecOptions.numCells[YY] = roundToInt(realddxyz[YY]);
+ domdecOptions.numCells[ZZ] = roundToInt(realddxyz[ZZ]);
+
+ return 1;
+}
+
+LegacyMdrunOptions::~LegacyMdrunOptions()
+{
+ if (GMX_LIB_MPI)
+ {
+ done_commrec(cr);
+ }
+ done_multisim(ms);
+}
+
+} // namespace gmx
diff --git a/patches/gromacs-2019.1.diff/src/gromacs/mdrun/legacymdrunoptions.h b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/legacymdrunoptions.h
new file mode 100644
index 0000000000000000000000000000000000000000..88520e58be55f24710cedde0e53bd016d7301c35
--- /dev/null
+++ b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/legacymdrunoptions.h
@@ -0,0 +1,301 @@
+/*
+ * 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) 2011,2012,2013,2014,2015,2016,2017,2018, 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.
+ */
+/*! \libinternal \file
+ *
+ * \brief This file declares helper functionality for legacy option handling for 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
+ * \inlibraryapi
+ */
+#ifndef GMX_MDRUN_LEGACYMDRUNOPTIONS_H
+#define GMX_MDRUN_LEGACYMDRUNOPTIONS_H
+
+#include "gromacs/commandline/filenm.h"
+#include "gromacs/commandline/pargs.h"
+#include "gromacs/domdec/domdec.h"
+#include "gromacs/hardware/hw_info.h"
+#include "gromacs/mdlib/mdrun.h"
+#include "gromacs/mdrun/logging.h"
+
+#include "replicaexchange.h"
+
+/* PLUMED */
+#include "../../../Plumed.h"
+extern int plumedswitch;
+extern plumed plumedmain;
+extern void(*plumedcmd)(plumed,const char*,const void*);
+/* END PLUMED */
+
+/* PLUMED HREX */
+int plumed_hrex;
+/* END PLUMED HREX */
+
+struct gmx_multisim_t;
+
+namespace gmx
+{
+
+/*! \libinternal
+ * \brief This class provides the same command-line option
+ * functionality to both CLI and API sessions.
+ *
+ * This class should not exist, but is necessary now to introduce
+ * support for the CLI and API without duplicating code. It should be
+ * eliminated following the TODOs below.
+ *
+ * \todo Modules in mdrun should acquire proper option handling so
+ * that all of these declarations and defaults are local to the
+ * modules.
+ *
+ * \todo Contextual aspects, such as working directory, MPI
+ * environment, and environment variable handling are more properly
+ * the role of SimulationContext, and should be moved there */
+class LegacyMdrunOptions
+{
+ public:
+ //! Ongoing collection of mdrun options
+ MdrunOptions mdrunOptions;
+ //! Options for the domain decomposition.
+ DomdecOptions domdecOptions;
+ //! Parallelism-related user options.
+ gmx_hw_opt_t hw_opt;
+ //! Command-line override for the duration of a neighbor list with the Verlet scheme.
+ int nstlist_cmdline = 0;
+ //! Parameters for replica-exchange simulations.
+ ReplicaExchangeParameters replExParams;
+
+ //! Filename options to fill from command-line argument values.
+ std::vector<t_filenm> filenames =
+ {{{ efTPR, nullptr, nullptr, ffREAD },
+ { efTRN, "-o", nullptr, ffWRITE },
+ { efCOMPRESSED, "-x", nullptr, ffOPTWR },
+ { efCPT, "-cpi", nullptr, ffOPTRD | ffALLOW_MISSING },
+ { efCPT, "-cpo", nullptr, 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", ffOPTRDMULT },
+ { 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", nullptr, ffOPTRDMULT},
+ { efXVG, "-awh", "awhinit", ffOPTRD },
+ { 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 }}};
+
+ //! Print a warning if any force is larger than this (in kJ/mol nm).
+ real pforce = -1;
+
+ /*! \brief Output context for writing text files
+ *
+ * \todo Clarify initialization, ownership, and lifetime. */
+ gmx_output_env_t *oenv = nullptr;
+
+ //! Handle to file used for logging.
+ LogFilePtr logFileGuard = nullptr;
+
+ /*! \brief Command line options, defaults, docs and storage for them to fill. */
+ /*! \{ */
+ rvec realddxyz = {0, 0, 0};
+ const char *ddrank_opt_choices[static_cast<int>(DdRankOrder::nr)+1] =
+ { nullptr, "interleave", "pp_pme", "cartesian", nullptr };
+ const char *dddlb_opt_choices[static_cast<int>(DlbOption::nr)+1] =
+ { nullptr, "auto", "no", "yes", nullptr };
+ const char *thread_aff_opt_choices[threadaffNR+1] =
+ { nullptr, "auto", "on", "off", nullptr };
+ const char *nbpu_opt_choices[5] =
+ { nullptr, "auto", "cpu", "gpu", nullptr };
+ const char *pme_opt_choices[5] =
+ { nullptr, "auto", "cpu", "gpu", nullptr };
+ const char *pme_fft_opt_choices[5] =
+ { nullptr, "auto", "cpu", "gpu", nullptr };
+ const char *bonded_opt_choices[5] =
+ { nullptr, "auto", "cpu", "gpu", nullptr };
+ gmx_bool bTryToAppendFiles = TRUE;
+ const char *gpuIdsAvailable = "";
+ const char *userGpuTaskAssignment = "";
+
+ ImdOptions &imdOptions = mdrunOptions.imdOptions;
+
+ t_pargs pa[49] = {
+
+ { "-dd", FALSE, etRVEC, {&realddxyz},
+ "Domain decomposition grid, 0 is optimize" },
+ { "-ddorder", FALSE, etENUM, {ddrank_opt_choices},
+ "DD rank order" },
+ { "-npme", FALSE, etINT, {&domdecOptions.numPmeRanks},
+ "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 ranks 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_choices},
+ "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, {&gpuIdsAvailable},
+ "List of unique GPU device IDs available to use" },
+ { "-gputasks", FALSE, etSTR, {&userGpuTaskAssignment},
+ "List of GPU device IDs, mapping each PP task on each node to a device" },
+ { "-ddcheck", FALSE, etBOOL, {&domdecOptions.checkBondedInteractions},
+ "Check for all bonded interactions with DD" },
+ { "-ddbondcomm", FALSE, etBOOL, {&domdecOptions.useBondedCommunication},
+ "HIDDENUse special bonded atom communication when [TT]-rdd[tt] > cut-off" },
+ { "-rdd", FALSE, etREAL, {&domdecOptions.minimumCommunicationRange},
+ "The maximum distance for bonded interactions with DD (nm), 0 is determine from initial coordinates" },
+ { "-rcon", FALSE, etREAL, {&domdecOptions.constraintCommunicationRange},
+ "Maximum distance for P-LINCS (nm), 0 is estimate" },
+ { "-dlb", FALSE, etENUM, {dddlb_opt_choices},
+ "Dynamic load balancing (with DD)" },
+ { "-dds", FALSE, etREAL, {&domdecOptions.dlbScaling},
+ "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, {&domdecOptions.cellSizeX},
+ "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, {&domdecOptions.cellSizeY},
+ "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, {&domdecOptions.cellSizeZ},
+ "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, {&mdrunOptions.globalCommunicationInterval},
+ "Global communication frequency" },
+ { "-nb", FALSE, etENUM, {nbpu_opt_choices},
+ "Calculate non-bonded interactions on" },
+ { "-nstlist", FALSE, etINT, {&nstlist_cmdline},
+ "Set nstlist when using a Verlet buffer tolerance (0 is guess)" },
+ { "-tunepme", FALSE, etBOOL, {&mdrunOptions.tunePme},
+ "Optimize PME load between PP/PME ranks or GPU/CPU (only with the Verlet cut-off scheme)" },
+ { "-pme", FALSE, etENUM, {pme_opt_choices},
+ "Perform PME calculations on" },
+ { "-pmefft", FALSE, etENUM, {pme_fft_opt_choices},
+ "Perform PME FFT calculations on" },
+ { "-bonded", FALSE, etENUM, {bonded_opt_choices},
+ "Perform bonded calculations on" },
+ { "-v", FALSE, etBOOL, {&mdrunOptions.verbose},
+ "Be loud and noisy" },
+ { "-pforce", FALSE, etREAL, {&pforce},
+ "Print all forces larger than this (kJ/mol nm)" },
+ { "-reprod", FALSE, etBOOL, {&mdrunOptions.reproducible},
+ "Try to avoid optimizations that affect binary reproducibility" },
+ { "-cpt", FALSE, etREAL, {&mdrunOptions.checkpointOptions.period},
+ "Checkpoint interval (minutes)" },
+ { "-cpnum", FALSE, etBOOL, {&mdrunOptions.checkpointOptions.keepAndNumberCheckpointFiles},
+ "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, {&mdrunOptions.numStepsCommandline},
+ "Run this number of steps, overrides .mdp file option (-1 means infinite, -2 means use mdp option, smaller is invalid)" },
+ { "-maxh", FALSE, etREAL, {&mdrunOptions.maximumHoursToRun},
+ "Terminate after 0.99 times this time (hours)" },
+ { "-replex", FALSE, etINT, {&replExParams.exchangeInterval},
+ "Attempt replica exchange periodically with this period (steps)" },
+ { "-nex", FALSE, etINT, {&replExParams.numExchanges},
+ "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, {&replExParams.randomSeed},
+ "Seed for replica exchange, -1 is generate a seed" },
+ { "-hrex", FALSE, etBOOL, {&plumed_hrex}, /* PLUMED HREX */
+ "Enable hamiltonian replica exchange" },
+ { "-imdport", FALSE, etINT, {&imdOptions.port},
+ "HIDDENIMD listening port" },
+ { "-imdwait", FALSE, etBOOL, {&imdOptions.wait},
+ "HIDDENPause the simulation while no IMD client is connected" },
+ { "-imdterm", FALSE, etBOOL, {&imdOptions.terminatable},
+ "HIDDENAllow termination of the simulation from IMD client" },
+ { "-imdpull", FALSE, etBOOL, {&imdOptions.pull},
+ "HIDDENAllow pulling in the simulation from IMD client" },
+ { "-rerunvsite", FALSE, etBOOL, {&mdrunOptions.rerunConstructVsites},
+ "HIDDENRecalculate virtual site coordinates with [TT]-rerun[tt]" },
+ { "-confout", FALSE, etBOOL, {&mdrunOptions.writeConfout},
+ "HIDDENWrite the last configuration with [TT]-c[tt] and force checkpointing at the last step" },
+ { "-stepout", FALSE, etINT, {&mdrunOptions.verboseStepPrintInterval},
+ "HIDDENFrequency of writing the remaining wall clock time for the run" },
+ { "-resetstep", FALSE, etINT, {&mdrunOptions.timingOptions.resetStep},
+ "HIDDENReset cycle counters after these many time steps" },
+ { "-resethway", FALSE, etBOOL, {&mdrunOptions.timingOptions.resetHalfway},
+ "HIDDENReset the cycle counters after half the number of steps or halfway [TT]-maxh[tt]" }
+ };
+ /*! \} */
+
+ //! Handle to communication object.
+ t_commrec *cr = nullptr;
+ //! Multi-simulation object.
+ gmx_multisim_t *ms = nullptr;
+
+ //! Parses the command-line input and prepares to start mdrun.
+ int updateFromCommandLine(int argc, char **argv, ArrayRef<const char *> desc);
+
+ ~LegacyMdrunOptions();
+};
+
+} // end namespace gmx
+
+#endif
diff --git a/patches/gromacs-2019.1.diff/src/gromacs/mdrun/legacymdrunoptions.h.preplumed b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/legacymdrunoptions.h.preplumed
new file mode 100644
index 0000000000000000000000000000000000000000..637cd2a6b93dabe9fd632d92e30794922c3a8fec
--- /dev/null
+++ b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/legacymdrunoptions.h.preplumed
@@ -0,0 +1,287 @@
+/*
+ * 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) 2011,2012,2013,2014,2015,2016,2017,2018, 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.
+ */
+/*! \libinternal \file
+ *
+ * \brief This file declares helper functionality for legacy option handling for 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
+ * \inlibraryapi
+ */
+#ifndef GMX_MDRUN_LEGACYMDRUNOPTIONS_H
+#define GMX_MDRUN_LEGACYMDRUNOPTIONS_H
+
+#include "gromacs/commandline/filenm.h"
+#include "gromacs/commandline/pargs.h"
+#include "gromacs/domdec/domdec.h"
+#include "gromacs/hardware/hw_info.h"
+#include "gromacs/mdlib/mdrun.h"
+#include "gromacs/mdrun/logging.h"
+
+#include "replicaexchange.h"
+
+struct gmx_multisim_t;
+
+namespace gmx
+{
+
+/*! \libinternal
+ * \brief This class provides the same command-line option
+ * functionality to both CLI and API sessions.
+ *
+ * This class should not exist, but is necessary now to introduce
+ * support for the CLI and API without duplicating code. It should be
+ * eliminated following the TODOs below.
+ *
+ * \todo Modules in mdrun should acquire proper option handling so
+ * that all of these declarations and defaults are local to the
+ * modules.
+ *
+ * \todo Contextual aspects, such as working directory, MPI
+ * environment, and environment variable handling are more properly
+ * the role of SimulationContext, and should be moved there */
+class LegacyMdrunOptions
+{
+ public:
+ //! Ongoing collection of mdrun options
+ MdrunOptions mdrunOptions;
+ //! Options for the domain decomposition.
+ DomdecOptions domdecOptions;
+ //! Parallelism-related user options.
+ gmx_hw_opt_t hw_opt;
+ //! Command-line override for the duration of a neighbor list with the Verlet scheme.
+ int nstlist_cmdline = 0;
+ //! Parameters for replica-exchange simulations.
+ ReplicaExchangeParameters replExParams;
+
+ //! Filename options to fill from command-line argument values.
+ std::vector<t_filenm> filenames =
+ {{{ efTPR, nullptr, nullptr, ffREAD },
+ { efTRN, "-o", nullptr, ffWRITE },
+ { efCOMPRESSED, "-x", nullptr, ffOPTWR },
+ { efCPT, "-cpi", nullptr, ffOPTRD | ffALLOW_MISSING },
+ { efCPT, "-cpo", nullptr, 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", ffOPTRDMULT },
+ { 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", nullptr, ffOPTRDMULT},
+ { efXVG, "-awh", "awhinit", ffOPTRD },
+ { efDAT, "-membed", "membed", ffOPTRD },
+ { efTOP, "-mp", "membed", ffOPTRD },
+ { efNDX, "-mn", "membed", ffOPTRD },
+ { efXVG, "-if", "imdforces", ffOPTWR },
+ { efXVG, "-swap", "swapions", ffOPTWR }}};
+
+ //! Print a warning if any force is larger than this (in kJ/mol nm).
+ real pforce = -1;
+
+ /*! \brief Output context for writing text files
+ *
+ * \todo Clarify initialization, ownership, and lifetime. */
+ gmx_output_env_t *oenv = nullptr;
+
+ //! Handle to file used for logging.
+ LogFilePtr logFileGuard = nullptr;
+
+ /*! \brief Command line options, defaults, docs and storage for them to fill. */
+ /*! \{ */
+ rvec realddxyz = {0, 0, 0};
+ const char *ddrank_opt_choices[static_cast<int>(DdRankOrder::nr)+1] =
+ { nullptr, "interleave", "pp_pme", "cartesian", nullptr };
+ const char *dddlb_opt_choices[static_cast<int>(DlbOption::nr)+1] =
+ { nullptr, "auto", "no", "yes", nullptr };
+ const char *thread_aff_opt_choices[threadaffNR+1] =
+ { nullptr, "auto", "on", "off", nullptr };
+ const char *nbpu_opt_choices[5] =
+ { nullptr, "auto", "cpu", "gpu", nullptr };
+ const char *pme_opt_choices[5] =
+ { nullptr, "auto", "cpu", "gpu", nullptr };
+ const char *pme_fft_opt_choices[5] =
+ { nullptr, "auto", "cpu", "gpu", nullptr };
+ const char *bonded_opt_choices[5] =
+ { nullptr, "auto", "cpu", "gpu", nullptr };
+ gmx_bool bTryToAppendFiles = TRUE;
+ const char *gpuIdsAvailable = "";
+ const char *userGpuTaskAssignment = "";
+
+ ImdOptions &imdOptions = mdrunOptions.imdOptions;
+
+ t_pargs pa[48] = {
+
+ { "-dd", FALSE, etRVEC, {&realddxyz},
+ "Domain decomposition grid, 0 is optimize" },
+ { "-ddorder", FALSE, etENUM, {ddrank_opt_choices},
+ "DD rank order" },
+ { "-npme", FALSE, etINT, {&domdecOptions.numPmeRanks},
+ "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 ranks 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_choices},
+ "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, {&gpuIdsAvailable},
+ "List of unique GPU device IDs available to use" },
+ { "-gputasks", FALSE, etSTR, {&userGpuTaskAssignment},
+ "List of GPU device IDs, mapping each PP task on each node to a device" },
+ { "-ddcheck", FALSE, etBOOL, {&domdecOptions.checkBondedInteractions},
+ "Check for all bonded interactions with DD" },
+ { "-ddbondcomm", FALSE, etBOOL, {&domdecOptions.useBondedCommunication},
+ "HIDDENUse special bonded atom communication when [TT]-rdd[tt] > cut-off" },
+ { "-rdd", FALSE, etREAL, {&domdecOptions.minimumCommunicationRange},
+ "The maximum distance for bonded interactions with DD (nm), 0 is determine from initial coordinates" },
+ { "-rcon", FALSE, etREAL, {&domdecOptions.constraintCommunicationRange},
+ "Maximum distance for P-LINCS (nm), 0 is estimate" },
+ { "-dlb", FALSE, etENUM, {dddlb_opt_choices},
+ "Dynamic load balancing (with DD)" },
+ { "-dds", FALSE, etREAL, {&domdecOptions.dlbScaling},
+ "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, {&domdecOptions.cellSizeX},
+ "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, {&domdecOptions.cellSizeY},
+ "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, {&domdecOptions.cellSizeZ},
+ "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, {&mdrunOptions.globalCommunicationInterval},
+ "Global communication frequency" },
+ { "-nb", FALSE, etENUM, {nbpu_opt_choices},
+ "Calculate non-bonded interactions on" },
+ { "-nstlist", FALSE, etINT, {&nstlist_cmdline},
+ "Set nstlist when using a Verlet buffer tolerance (0 is guess)" },
+ { "-tunepme", FALSE, etBOOL, {&mdrunOptions.tunePme},
+ "Optimize PME load between PP/PME ranks or GPU/CPU (only with the Verlet cut-off scheme)" },
+ { "-pme", FALSE, etENUM, {pme_opt_choices},
+ "Perform PME calculations on" },
+ { "-pmefft", FALSE, etENUM, {pme_fft_opt_choices},
+ "Perform PME FFT calculations on" },
+ { "-bonded", FALSE, etENUM, {bonded_opt_choices},
+ "Perform bonded calculations on" },
+ { "-v", FALSE, etBOOL, {&mdrunOptions.verbose},
+ "Be loud and noisy" },
+ { "-pforce", FALSE, etREAL, {&pforce},
+ "Print all forces larger than this (kJ/mol nm)" },
+ { "-reprod", FALSE, etBOOL, {&mdrunOptions.reproducible},
+ "Try to avoid optimizations that affect binary reproducibility" },
+ { "-cpt", FALSE, etREAL, {&mdrunOptions.checkpointOptions.period},
+ "Checkpoint interval (minutes)" },
+ { "-cpnum", FALSE, etBOOL, {&mdrunOptions.checkpointOptions.keepAndNumberCheckpointFiles},
+ "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, {&mdrunOptions.numStepsCommandline},
+ "Run this number of steps, overrides .mdp file option (-1 means infinite, -2 means use mdp option, smaller is invalid)" },
+ { "-maxh", FALSE, etREAL, {&mdrunOptions.maximumHoursToRun},
+ "Terminate after 0.99 times this time (hours)" },
+ { "-replex", FALSE, etINT, {&replExParams.exchangeInterval},
+ "Attempt replica exchange periodically with this period (steps)" },
+ { "-nex", FALSE, etINT, {&replExParams.numExchanges},
+ "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, {&replExParams.randomSeed},
+ "Seed for replica exchange, -1 is generate a seed" },
+ { "-imdport", FALSE, etINT, {&imdOptions.port},
+ "HIDDENIMD listening port" },
+ { "-imdwait", FALSE, etBOOL, {&imdOptions.wait},
+ "HIDDENPause the simulation while no IMD client is connected" },
+ { "-imdterm", FALSE, etBOOL, {&imdOptions.terminatable},
+ "HIDDENAllow termination of the simulation from IMD client" },
+ { "-imdpull", FALSE, etBOOL, {&imdOptions.pull},
+ "HIDDENAllow pulling in the simulation from IMD client" },
+ { "-rerunvsite", FALSE, etBOOL, {&mdrunOptions.rerunConstructVsites},
+ "HIDDENRecalculate virtual site coordinates with [TT]-rerun[tt]" },
+ { "-confout", FALSE, etBOOL, {&mdrunOptions.writeConfout},
+ "HIDDENWrite the last configuration with [TT]-c[tt] and force checkpointing at the last step" },
+ { "-stepout", FALSE, etINT, {&mdrunOptions.verboseStepPrintInterval},
+ "HIDDENFrequency of writing the remaining wall clock time for the run" },
+ { "-resetstep", FALSE, etINT, {&mdrunOptions.timingOptions.resetStep},
+ "HIDDENReset cycle counters after these many time steps" },
+ { "-resethway", FALSE, etBOOL, {&mdrunOptions.timingOptions.resetHalfway},
+ "HIDDENReset the cycle counters after half the number of steps or halfway [TT]-maxh[tt]" }
+ };
+ /*! \} */
+
+ //! Handle to communication object.
+ t_commrec *cr = nullptr;
+ //! Multi-simulation object.
+ gmx_multisim_t *ms = nullptr;
+
+ //! Parses the command-line input and prepares to start mdrun.
+ int updateFromCommandLine(int argc, char **argv, ArrayRef<const char *> desc);
+
+ ~LegacyMdrunOptions();
+};
+
+} // end namespace gmx
+
+#endif
diff --git a/patches/gromacs-2019.1.diff/src/gromacs/mdrun/md.cpp b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/md.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..887526a8d1dff7bdb05905ed784c1e500086501c
--- /dev/null
+++ b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/md.cpp
@@ -0,0 +1,1728 @@
+/*
+ * 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) 2011,2012,2013,2014,2015,2016,2017,2018, 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 integrator for normal molecular dynamics simulations
+ *
+ * \author David van der Spoel <david.vanderspoel@icm.uu.se>
+ * \ingroup module_mdrun
+ */
+#include "gmxpre.h"
+
+#include <cinttypes>
+#include <cmath>
+#include <cstdio>
+#include <cstdlib>
+
+#include <algorithm>
+#include <memory>
+
+#include "gromacs/awh/awh.h"
+#include "gromacs/commandline/filenm.h"
+#include "gromacs/compat/make_unique.h"
+#include "gromacs/domdec/collect.h"
+#include "gromacs/domdec/domdec.h"
+#include "gromacs/domdec/domdec_network.h"
+#include "gromacs/domdec/domdec_struct.h"
+#include "gromacs/domdec/partition.h"
+#include "gromacs/essentialdynamics/edsam.h"
+#include "gromacs/ewald/pme.h"
+#include "gromacs/ewald/pme-load-balancing.h"
+#include "gromacs/fileio/trxio.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/checkpointhandler.h"
+#include "gromacs/mdlib/compute_io.h"
+#include "gromacs/mdlib/constr.h"
+#include "gromacs/mdlib/ebin.h"
+#include "gromacs/mdlib/expanded.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/mdsetup.h"
+#include "gromacs/mdlib/membed.h"
+#include "gromacs/mdlib/nb_verlet.h"
+#include "gromacs/mdlib/nbnxn_gpu_data_mgmt.h"
+#include "gromacs/mdlib/ns.h"
+#include "gromacs/mdlib/resethandler.h"
+#include "gromacs/mdlib/shellfc.h"
+#include "gromacs/mdlib/sighandler.h"
+#include "gromacs/mdlib/sim_util.h"
+#include "gromacs/mdlib/simulationsignal.h"
+#include "gromacs/mdlib/stophandler.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/awh-history.h"
+#include "gromacs/mdtypes/awh-params.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/observableshistory.h"
+#include "gromacs/mdtypes/pullhistory.h"
+#include "gromacs/mdtypes/state.h"
+#include "gromacs/pbcutil/mshift.h"
+#include "gromacs/pbcutil/pbc.h"
+#include "gromacs/pulling/output.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/logger.h"
+#include "gromacs/utility/real.h"
+#include "gromacs/utility/smalloc.h"
+
+#include "integrator.h"
+#include "replicaexchange.h"
+
+#if GMX_FAHCORE
+#include "corewrap.h"
+#endif
+
+/* PLUMED */
+#include "../../../Plumed.h"
+extern int plumedswitch;
+extern plumed plumedmain;
+/* END PLUMED */
+
+/* PLUMED HREX */
+extern int plumed_hrex;
+/* END PLUMED HREX */
+
+using gmx::SimulationSignaller;
+
+void gmx::Integrator::do_md()
+{
+ // TODO Historically, the EM and MD "integrators" used different
+ // names for the t_inputrec *parameter, but these must have the
+ // same name, now that it's a member of a struct. We use this ir
+ // alias to avoid a large ripple of nearly useless changes.
+ // t_inputrec is being replaced by IMdpOptionsProvider, so this
+ // will go away eventually.
+ t_inputrec *ir = inputrec;
+ gmx_mdoutf *outf = nullptr;
+ int64_t step, step_rel;
+ double t, t0, lam0[efptNR];
+ gmx_bool bGStatEveryStep, bGStat, bCalcVir, bCalcEnerStep, bCalcEner;
+ gmx_bool bNS, bNStList, bSimAnn, bStopCM,
+ bFirstStep, bInitStep, bLastStep = FALSE;
+ gmx_bool bDoDHDL = FALSE, bDoFEP = FALSE, bDoExpanded = FALSE;
+ gmx_bool do_ene, do_log, do_verbose;
+ gmx_bool bMasterState;
+ int force_flags, cglo_flags;
+ tensor force_vir, shake_vir, total_vir, tmp_vir, pres;
+ int i, m;
+ rvec mu_tot;
+ t_vcm *vcm;
+ matrix parrinellorahmanMu, M;
+ gmx_repl_ex_t repl_ex = nullptr;
+ gmx_localtop_t *top;
+ t_mdebin *mdebin = nullptr;
+ gmx_enerdata_t *enerd;
+ PaddedVector<gmx::RVec> f {};
+ gmx_global_stat_t gstat;
+ gmx_update_t *upd = nullptr;
+ t_graph *graph = nullptr;
+ gmx_groups_t *groups;
+ gmx_ekindata_t *ekind;
+ gmx_shellfc_t *shellfc;
+ gmx_bool bSumEkinhOld, bDoReplEx, bExchanged, bNeedRepartition;
+ gmx_bool bTemp, bPres, bTrotter;
+ real dvdl_constr;
+ rvec *cbuf = nullptr;
+ 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];
+
+ /* PME load balancing data for GPU kernels */
+ pme_load_balancing_t *pme_loadbal = nullptr;
+ gmx_bool bPMETune = FALSE;
+ gmx_bool bPMETunePrinting = FALSE;
+
+ /* Interactive MD */
+ gmx_bool bIMDstep = FALSE;
+
+ /* PLUMED */
+ int plumedNeedsEnergy=0;
+ int plumedWantsToStop=0;
+ matrix plumed_vir;
+ /* END PLUMED */
+
+ /* 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;
+
+ SimulationSignals signals;
+ // Most global communnication stages don't propagate mdrun
+ // signals, and will use this object to achieve that.
+ SimulationSignaller nullSignaller(nullptr, nullptr, nullptr, false, false);
+
+ if (!mdrunOptions.writeConfout)
+ {
+ // This is on by default, and the main known use case for
+ // turning it off is for convenience in benchmarking, which is
+ // something that should not show up in the general user
+ // interface.
+ GMX_LOG(mdlog.info).asParagraph().
+ appendText("The -noconfout functionality is deprecated, and may be removed in a future version.");
+ }
+
+ /* 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)));
+
+ const bool bRerunMD = false;
+ int nstglobalcomm = mdrunOptions.globalCommunicationInterval;
+
+ nstglobalcomm = check_nstglobalcomm(mdlog, nstglobalcomm, ir, cr);
+ bGStatEveryStep = (nstglobalcomm == 1);
+
+ groups = &top_global->groups;
+
+ std::unique_ptr<EssentialDynamics> ed = nullptr;
+ if (opt2bSet("-ei", nfile, fnm) || observablesHistory->edsamHistory != nullptr)
+ {
+ /* Initialize essential dynamics sampling */
+ ed = init_edsam(mdlog,
+ opt2fn_null("-ei", nfile, fnm), opt2fn("-eo", nfile, fnm),
+ top_global,
+ ir, cr, constr,
+ state_global, observablesHistory,
+ oenv, mdrunOptions.continuationOptions.appendFiles);
+ }
+
+ /* Initial values */
+ init_md(fplog, cr, outputProvider, ir, oenv, mdrunOptions,
+ &t, &t0, state_global, lam0,
+ nrnb, top_global, &upd, deform,
+ nfile, fnm, &outf, &mdebin,
+ force_vir, shake_vir, total_vir, pres, mu_tot, &bSimAnn, &vcm, wcycle);
+
+ /* Energy terms and groups */
+ snew(enerd, 1);
+ init_enerdata(top_global->groups.grps[egcENER].nr, ir->fepvals->n_lambda,
+ enerd);
+
+ /* 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, constr ? constr->numFlexibleConstraints() : 0,
+ ir->nstcalcenergy, DOMAINDECOMP(cr));
+
+ {
+ 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);
+ }
+ }
+
+ /* Set up interactive MD (IMD) */
+ init_IMD(ir, cr, ms, top_global, fplog, ir->nstcalcenergy,
+ MASTER(cr) ? state_global->x.rvec_array() : nullptr,
+ nfile, fnm, oenv, mdrunOptions);
+
+ // Local state only becomes valid now.
+ std::unique_ptr<t_state> stateInstance;
+ t_state * state;
+
+ if (DOMAINDECOMP(cr))
+ {
+ top = dd_init_local_top(top_global);
+
+ stateInstance = compat::make_unique<t_state>();
+ state = stateInstance.get();
+ dd_init_local_state(cr->dd, state_global, state);
+
+ /* Distribute the charge groups over the nodes from the master node */
+ dd_partition_system(fplog, mdlog, ir->init_step, cr, TRUE, 1,
+ state_global, top_global, ir,
+ state, &f, mdAtoms, top, fr,
+ vsite, constr,
+ nrnb, nullptr, FALSE);
+ shouldCheckNumberOfBondedInteractions = true;
+ update_realloc(upd, state->natoms);
+ }
+ else
+ {
+ state_change_natoms(state_global, state_global->natoms);
+ f.resizeWithPadding(state_global->natoms);
+ /* Copy the pointer to the global state */
+ state = state_global;
+
+ snew(top, 1);
+ mdAlgorithmsSetupAtomData(cr, ir, top_global, top, fr,
+ &graph, mdAtoms, constr, vsite, shellfc);
+
+ update_realloc(upd, state->natoms);
+ }
+
+ auto mdatoms = mdAtoms->mdatoms();
+
+ // NOTE: The global state is no longer used at this point.
+ // But state_global is still used as temporary storage space for writing
+ // the global state to file and potentially for replica exchange.
+ // (Global topology should persist.)
+
+ update_mdatoms(mdatoms, state->lambda[efptMASS]);
+
+ const ContinuationOptions &continuationOptions = mdrunOptions.continuationOptions;
+ bool startingFromCheckpoint = continuationOptions.startedFromCheckpoint;
+
+ if (ir->bExpanded)
+ {
+ /* Check nstexpanded here, because the grompp check was broken */
+ if (ir->expandedvals->nstexpanded % ir->nstcalcenergy != 0)
+ {
+ gmx_fatal(FARGS, "With expanded ensemble, nstexpanded should be a multiple of nstcalcenergy");
+ }
+ init_expanded_ensemble(startingFromCheckpoint, ir, state->dfhist);
+ }
+
+ if (MASTER(cr))
+ {
+ if (startingFromCheckpoint)
+ {
+ /* Update mdebin with energy history if appending to output files */
+ if (continuationOptions.appendFiles)
+ {
+ /* If no history is available (because a checkpoint is from before
+ * it was written) make a new one later, otherwise restore it.
+ */
+ if (observablesHistory->energyHistory)
+ {
+ restore_energyhistory_from_state(mdebin, observablesHistory->energyHistory.get());
+ }
+ }
+ else if (observablesHistory->energyHistory)
+ {
+ /* We might have read an energy history from checkpoint.
+ * As we are not appending, we want to restart the statistics.
+ * Free the allocated memory and reset the counts.
+ */
+ observablesHistory->energyHistory = {};
+ /* We might have read a pull history from checkpoint.
+ * We will still want to keep the statistics, so that the files
+ * can be joined and still be meaningful.
+ * This means that observablesHistory->pullHistory
+ * should not be reset.
+ */
+ }
+ }
+ if (!observablesHistory->energyHistory)
+ {
+ observablesHistory->energyHistory = compat::make_unique<energyhistory_t>();
+ }
+ if (!observablesHistory->pullHistory)
+ {
+ observablesHistory->pullHistory = compat::make_unique<PullHistory>();
+ }
+ /* Set the initial energy history in state by updating once */
+ update_energyhistory(observablesHistory->energyHistory.get(), mdebin);
+ }
+
+ preparePrevStepPullCom(ir, mdatoms, state, state_global, cr, startingFromCheckpoint);
+
+ // TODO: Remove this by converting AWH into a ForceProvider
+ auto awh = prepareAwhModule(fplog, *ir, state_global, cr, ms, startingFromCheckpoint,
+ shellfc != nullptr,
+ opt2fn("-awh", nfile, fnm), ir->pull_work);
+
+ const bool useReplicaExchange = (replExParams.exchangeInterval > 0);
+ if (useReplicaExchange && MASTER(cr))
+ {
+ repl_ex = init_replica_exchange(fplog, ms, top_global->natoms, ir,
+ replExParams);
+ }
+ /* PME tuning is only supported in the Verlet scheme, with PME for
+ * Coulomb. It is not supported with only LJ PME. */
+ bPMETune = (mdrunOptions.tunePme && EEL_PME(fr->ic->eeltype) &&
+ !mdrunOptions.reproducible && ir->cutoff_scheme != ecutsGROUP);
+ if (bPMETune)
+ {
+ pme_loadbal_init(&pme_loadbal, cr, mdlog, *ir, state->box,
+ *fr->ic, *fr->nbv->listParams, fr->pmedata, use_GPU(fr->nbv),
+ &bPMETunePrinting);
+ }
+
+ if (!ir->bContinuation)
+ {
+ if (state->flags & (1 << estV))
+ {
+ auto v = makeArrayRef(state->v);
+ /* Set the velocities of vsites, shells and frozen atoms to zero */
+ for (i = 0; i < mdatoms->homenr; i++)
+ {
+ if (mdatoms->ptype[i] == eptVSite ||
+ mdatoms->ptype[i] == eptShell)
+ {
+ clear_rvec(v[i]);
+ }
+ else if (mdatoms->cFREEZE)
+ {
+ for (m = 0; m < DIM; m++)
+ {
+ if (ir->opts.nFreeze[mdatoms->cFREEZE[i]][m])
+ {
+ v[i][m] = 0;
+ }
+ }
+ }
+ }
+ }
+
+ if (constr)
+ {
+ /* Constrain the initial coordinates and velocities */
+ do_constrain_first(fplog, constr, ir, mdatoms, state);
+ }
+ if (vsite)
+ {
+ /* Construct the virtual sites for the initial configuration */
+ construct_vsites(vsite, state->x.rvec_array(), ir->delta_t, nullptr,
+ 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 (useReplicaExchange)
+ {
+ nstfep = gmx_greatest_common_divisor(replExParams.exchangeInterval, 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 (continuationOptions.haveReadEkin)
+ {
+ restore_ekinstate_from_state(cr, ekind, &state_global->ekinstate);
+ }
+
+ cglo_flags = (CGLO_INITIALIZATION | CGLO_TEMPERATURE | CGLO_GSTAT
+ | (EI_VV(ir->eI) ? CGLO_PRESSURE : 0)
+ | (EI_VV(ir->eI) ? CGLO_CONSTRAINT : 0)
+ | (continuationOptions.haveReadEkin ? CGLO_READEKIN : 0));
+
+ bSumEkinhOld = FALSE;
+ /* To minimize communication, compute_globals computes the COM velocity
+ * and the kinetic energy for the velocities without COM motion removed.
+ * Thus to get the kinetic energy without the COM contribution, we need
+ * to call compute_globals twice.
+ */
+ for (int cgloIteration = 0; cgloIteration < (bStopCM ? 2 : 1); cgloIteration++)
+ {
+ int cglo_flags_iteration = cglo_flags;
+ if (bStopCM && cgloIteration == 0)
+ {
+ cglo_flags_iteration |= CGLO_STOPCM;
+ cglo_flags_iteration &= ~CGLO_TEMPERATURE;
+ }
+ compute_globals(fplog, gstat, cr, ir, fr, ekind, state, mdatoms, nrnb, vcm,
+ nullptr, enerd, force_vir, shake_vir, total_vir, pres, mu_tot,
+ constr, &nullSignaller, state->box,
+ &totalNumberOfBondedInteractions, &bSumEkinhOld, cglo_flags_iteration
+ | (shouldCheckNumberOfBondedInteractions ? CGLO_CHECK_NUMBER_OF_BONDED_INTERACTIONS : 0));
+ }
+ checkNumberOfBondedInteractions(mdlog, 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,
+ nullptr, enerd, force_vir, shake_vir, total_vir, pres, mu_tot,
+ constr, &nullSignaller, state->box,
+ nullptr, &bSumEkinhOld,
+ cglo_flags & ~CGLO_PRESSURE);
+ }
+
+ /* Calculate the initial half step temperature, and save the ekinh_old */
+ if (!continuationOptions.startedFromCheckpoint)
+ {
+ for (i = 0; (i < ir->opts.ngtc); i++)
+ {
+ copy_mat(ekind->tcstat[i].ekinh, ekind->tcstat[i].ekinh_old);
+ }
+ }
+
+ /* 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 (!ir->bContinuation)
+ {
+ if (constr && ir->eConstrAlg == econtLINCS)
+ {
+ fprintf(fplog,
+ "RMS relative constraint deviation after constraining: %.2e\n",
+ constr->rmsd());
+ }
+ if (EI_STATE_VELOCITY(ir->eI))
+ {
+ real temp = enerd->term[F_TEMP];
+ if (ir->eI != eiVV)
+ {
+ /* Result of Ekin averaged over velocities of -half
+ * and +half step, while we only have -half step here.
+ */
+ temp *= 2;
+ }
+ fprintf(fplog, "Initial temperature: %g K\n", temp);
+ }
+ }
+
+ 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( (continuationOptions.startedFromCheckpoint) ) plumed_cmd(plumedmain,"setRestart",&res);
+ }
+
+ if(ms && ms->nsim>1) {
+ if(MASTER(cr)) plumed_cmd(plumedmain,"GREX setMPIIntercomm",&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=ir->delta_t;
+ plumed_cmd(plumedmain,"setTimestep",&real_delta_t);
+ plumed_cmd(plumedmain,"init",NULL);
+
+ if(PAR(cr)){
+ if(DOMAINDECOMP(cr)) {
+ int nat_home = dd_numHomeAtoms(*cr->dd);
+ plumed_cmd(plumedmain,"setAtomsNlocal",&nat_home);
+ plumed_cmd(plumedmain,"setAtomsGatindex",cr->dd->globalAtomIndices.data());
+ }
+ }
+ }
+ /* END PLUMED */
+
+ walltime_accounting_start_time(walltime_accounting);
+ wallcycle_start(wcycle, ewcRUN);
+ print_start(fplog, cr, walltime_accounting, "mdrun");
+
+#if GMX_FAHCORE
+ /* safest point to do file checkpointing is here. More general point would be immediately before integrator call */
+ int 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
+ *
+ ************************************************************/
+
+ 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;
+
+ bool simulationsShareState = false;
+ int nstSignalComm = nstglobalcomm;
+ {
+ // TODO This implementation of ensemble orientation restraints is nasty because
+ // a user can't just do multi-sim with single-sim orientation restraints.
+ bool usingEnsembleRestraints = (fcd->disres.nsystems > 1) || ((ms != nullptr) && (fcd->orires.nr != 0));
+ bool awhUsesMultiSim = (ir->bDoAwh && ir->awhParams->shareBiasMultisim && (ms != nullptr));
+
+ // Replica exchange, ensemble restraints and AWH need all
+ // simulations to remain synchronized, so they need
+ // checkpoints and stop conditions to act on the same step, so
+ // the propagation of such signals must take place between
+ // simulations, not just within simulations.
+ // TODO: Make algorithm initializers set these flags.
+ simulationsShareState = useReplicaExchange || usingEnsembleRestraints || awhUsesMultiSim || (plumedswitch && ms); // PLUMED hack, if we have multiple sim and plumed we usually want them to be in sync
+
+ if (simulationsShareState)
+ {
+ // Inter-simulation signal communication does not need to happen
+ // often, so we use a minimum of 200 steps to reduce overhead.
+ const int c_minimumInterSimulationSignallingInterval = 200;
+ nstSignalComm = ((c_minimumInterSimulationSignallingInterval + nstglobalcomm - 1)/nstglobalcomm)*nstglobalcomm;
+ }
+ }
+
+ auto stopHandler = stopHandlerBuilder->getStopHandlerMD(
+ compat::not_null<SimulationSignal*>(&signals[eglsSTOPCOND]), simulationsShareState,
+ MASTER(cr), ir->nstlist, mdrunOptions.reproducible, nstSignalComm,
+ mdrunOptions.maximumHoursToRun, ir->nstlist == 0, fplog, step, bNS, walltime_accounting);
+
+ auto checkpointHandler = compat::make_unique<CheckpointHandler>(
+ compat::make_not_null<SimulationSignal*>(&signals[eglsCHKPT]),
+ simulationsShareState, ir->nstlist == 0, MASTER(cr),
+ mdrunOptions.writeConfout, mdrunOptions.checkpointOptions.period);
+
+ const bool resetCountersIsLocal = true;
+ auto resetHandler = compat::make_unique<ResetHandler>(
+ compat::make_not_null<SimulationSignal*>(&signals[eglsRESETCOUNTERS]), !resetCountersIsLocal,
+ ir->nsteps, MASTER(cr), mdrunOptions.timingOptions.resetHalfway,
+ mdrunOptions.maximumHoursToRun, mdlog, wcycle, walltime_accounting);
+
+ DdOpenBalanceRegionBeforeForceComputation ddOpenBalanceRegion = (DOMAINDECOMP(cr) ? DdOpenBalanceRegionBeforeForceComputation::yes : DdOpenBalanceRegionBeforeForceComputation::no);
+ DdCloseBalanceRegionAfterForceComputation ddCloseBalanceRegion = (DOMAINDECOMP(cr) ? DdCloseBalanceRegionAfterForceComputation::yes : DdCloseBalanceRegionAfterForceComputation::no);
+
+ step = ir->init_step;
+ step_rel = 0;
+
+ // TODO extract this to new multi-simulation module
+ if (MASTER(cr) && isMultiSim(ms) && !useReplicaExchange)
+ {
+ if (!multisim_int_all_are_equal(ms, ir->nsteps))
+ {
+ GMX_LOG(mdlog.warning).appendText(
+ "Note: The number of steps is not consistent across multi simulations,\n"
+ "but we are proceeding anyway!");
+ }
+ if (!multisim_int_all_are_equal(ms, ir->init_step))
+ {
+ GMX_LOG(mdlog.warning).appendText(
+ "Note: The initial step is not consistent across multi simulations,\n"
+ "but we are proceeding anyway!");
+ }
+ }
+
+ /* and stop now if we should */
+ bLastStep = (bLastStep || (ir->nsteps >= 0 && step_rel > ir->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,
+ (mdrunOptions.verbose && MASTER(cr)) ? stderr : nullptr,
+ fplog, mdlog,
+ *ir, fr, *state,
+ wcycle,
+ step, step_rel,
+ &bPMETunePrinting);
+ }
+
+ wallcycle_start(wcycle, ewcSTEP);
+
+ bLastStep = (step_rel == ir->nsteps);
+ t = t0 + step*ir->delta_t;
+
+ // TODO Refactor this, so that nstfep does not need a default value of zero
+ if (ir->efep != efepNO || ir->bSimTemp)
+ {
+ /* find and set the current lambdas */
+ setCurrentLambdasLocal(step, ir->fepvals, lam0, state);
+
+ 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 = (useReplicaExchange && (step > 0) && !bLastStep &&
+ do_per_step(step, replExParams.exchangeInterval));
+
+ if (bSimAnn)
+ {
+ update_annealing_target_temp(ir, t, upd);
+ }
+
+ /* Stop Center of Mass motion */
+ bStopCM = (ir->comm_mode != ecmNO && do_per_step(step, ir->nstcomm));
+
+ /* Determine whether or not to do Neighbour Searching */
+ bNS = (bFirstStep || bNStList || bExchanged || bNeedRepartition);
+
+ bLastStep = bLastStep || stopHandler->stoppingAfterCurrentStep(bNS);
+
+ /* 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;
+ do_verbose = mdrunOptions.verbose &&
+ (step % mdrunOptions.verboseStepPrintInterval == 0 || bFirstStep || bLastStep);
+
+ if (bNS && !(bFirstStep && ir->bContinuation))
+ {
+ 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, mdlog, 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->natoms);
+
+ /* PLUMED */
+ if(plumedswitch){
+ int nat_home = dd_numHomeAtoms(*cr->dd);
+ plumed_cmd(plumedmain,"setAtomsNlocal",&nat_home);
+ plumed_cmd(plumedmain,"setAtomsGatindex",cr->dd->globalAtomIndices.data());
+ }
+ /* 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 (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, nullptr, nullptr, nullptr, nullptr, mu_tot,
+ constr, &nullSignaller, state->box,
+ &totalNumberOfBondedInteractions, &bSumEkinhOld,
+ CGLO_GSTAT | CGLO_TEMPERATURE | CGLO_CHECK_NUMBER_OF_BONDED_INTERACTIONS);
+ checkNumberOfBondedInteractions(mdlog, cr, totalNumberOfBondedInteractions,
+ top_global, top, state,
+ &shouldCheckNumberOfBondedInteractions);
+ }
+ clear_mat(force_vir);
+
+ /* PLUMED HREX */
+ gmx_bool bHREX = bDoReplEx && plumed_hrex;
+
+ if (plumedswitch && bHREX) {
+ gmx_enerdata_t *hrex_enerd;
+ snew(hrex_enerd, 1);
+ init_enerdata(top_global->groups.grps[egcENER].nr,ir->fepvals->n_lambda,hrex_enerd);
+ int repl = -1;
+ int nrepl = -1;
+ if (MASTER(cr)){
+ repl = replica_exchange_get_repl(repl_ex);
+ nrepl = replica_exchange_get_nrepl(repl_ex);
+ }
+
+ if (DOMAINDECOMP(cr)) {
+ dd_collect_state(cr->dd,state,state_global);
+ } else {
+ copy_state_serial(state, state_global);
+ }
+
+ if(MASTER(cr)){
+ if(repl%2==step/replExParams.exchangeInterval%2){
+ if(repl-1>=0) exchange_state(ms,repl-1,state_global);
+ }else{
+ if(repl+1<nrepl) exchange_state(ms,repl+1,state_global);
+ }
+ }
+ if (!DOMAINDECOMP(cr)) {
+ copy_state_serial(state_global, state);
+ }
+ if(PAR(cr)){
+ if (DOMAINDECOMP(cr)) {
+ dd_partition_system(fplog,mdlog,step,cr,TRUE,1,
+ state_global,top_global,ir,
+ state,&f,mdAtoms,top,fr,vsite,constr,
+ nrnb,wcycle,FALSE);
+ }
+ }
+ do_force(fplog, cr, ms, ir, awh.get(), enforcedRotation,
+ step, nrnb, wcycle, top, groups,
+ state->box, state->x.arrayRefWithPadding(), &state->hist,
+ f.arrayRefWithPadding(), force_vir, mdatoms, hrex_enerd, fcd,
+ state->lambda, graph,
+ fr, ppForceWorkload, vsite, mu_tot, t, ed ? ed->getLegacyED() : nullptr,
+ GMX_FORCE_STATECHANGED |
+ GMX_FORCE_DYNAMICBOX |
+ GMX_FORCE_ALLFORCES |
+ GMX_FORCE_VIRIAL |
+ GMX_FORCE_ENERGY |
+ GMX_FORCE_DHDL |
+ GMX_FORCE_NS,
+ ddOpenBalanceRegion, ddCloseBalanceRegion);
+
+ plumed_cmd(plumedmain,"GREX cacheLocalUSwap",&hrex_enerd->term[F_EPOT]);
+ sfree(hrex_enerd);
+
+ /* exchange back */
+ if (DOMAINDECOMP(cr)) {
+ dd_collect_state(cr->dd,state,state_global);
+ } else {
+ copy_state_serial(state, state_global);
+ }
+
+ if(MASTER(cr)){
+ if(repl%2==step/replExParams.exchangeInterval%2){
+ if(repl-1>=0) exchange_state(ms,repl-1,state_global);
+ }else{
+ if(repl+1<nrepl) exchange_state(ms,repl+1,state_global);
+ }
+ }
+
+ if (!DOMAINDECOMP(cr)) {
+ copy_state_serial(state_global, state);
+ }
+ if(PAR(cr)){
+ if (DOMAINDECOMP(cr)) {
+ dd_partition_system(fplog,mdlog,step,cr,TRUE,1,
+ state_global,top_global,ir,
+ state,&f,mdAtoms,top,fr,vsite,constr,
+ nrnb,wcycle,FALSE);
+ int nat_home = dd_numHomeAtoms(*cr->dd);
+ plumed_cmd(plumedmain,"setAtomsNlocal",&nat_home);
+ plumed_cmd(plumedmain,"setAtomsGatindex",cr->dd->globalAtomIndices.data());
+ }
+ }
+ }
+ /* END PLUMED HREX */
+
+ checkpointHandler->decideIfCheckpointingThisStep(bNS, bFirstStep, bLastStep);
+
+ /* Determine the energy and pressure:
+ * at nstcalcenergy steps and at energy output steps (set below).
+ */
+ if (EI_VV(ir->eI) && (!bInitStep))
+ {
+ bCalcEnerStep = do_per_step(step, 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);
+
+ 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)) ? 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, ms, mdrunOptions.verbose,
+ enforcedRotation, step,
+ ir, bNS, force_flags, top,
+ constr, enerd, fcd,
+ state, f.arrayRefWithPadding(), force_vir, mdatoms,
+ nrnb, wcycle, graph, groups,
+ shellfc, fr, ppForceWorkload, t, mu_tot,
+ vsite,
+ ddOpenBalanceRegion, ddCloseBalanceRegion);
+ }
+ else
+ {
+ /* The AWH history need to be saved _before_ doing force calculations where the AWH bias is updated
+ (or the AWH update will be performed twice for one step when continuing). It would be best to
+ call this update function from do_md_trajectory_writing but that would occur after do_force.
+ One would have to divide the update_awh function into one function applying the AWH force
+ and one doing the AWH bias update. The update AWH bias function could then be called after
+ do_md_trajectory_writing (then containing update_awh_history).
+ The checkpointing will in the future probably moved to the start of the md loop which will
+ rid of this issue. */
+ if (awh && checkpointHandler->isCheckpointingStep() && MASTER(cr))
+ {
+ awh->updateHistory(state_global->awhHistory.get());
+ }
+
+ /* 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){
+ int pversion=0;
+ plumed_cmd(plumedmain,"getApiVersion",&pversion);
+ 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);
+ int checkp=0; if(checkpointHandler->isCheckpointingStep()) checkp=1;
+ if(pversion>3) plumed_cmd(plumedmain,"doCheckPoint",&checkp);
+ plumed_cmd(plumedmain,"setForces",&f[0][0]);
+ plumed_cmd(plumedmain,"isEnergyNeeded",&plumedNeedsEnergy);
+ if(plumedNeedsEnergy) force_flags |= GMX_FORCE_ENERGY | GMX_FORCE_VIRIAL;
+ clear_mat(plumed_vir);
+ plumed_cmd(plumedmain,"setVirial",&plumed_vir[0][0]);
+ }
+ /* END PLUMED */
+ do_force(fplog, cr, ms, ir, awh.get(), enforcedRotation,
+ step, nrnb, wcycle, top, groups,
+ state->box, state->x.arrayRefWithPadding(), &state->hist,
+ f.arrayRefWithPadding(), force_vir, mdatoms, enerd, fcd,
+ state->lambda, graph,
+ fr, ppForceWorkload, vsite, mu_tot, t, ed ? ed->getLegacyED() : nullptr,
+ (bNS ? GMX_FORCE_NS : 0) | force_flags,
+ ddOpenBalanceRegion, ddCloseBalanceRegion);
+ /* 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(bDoReplEx) plumed_cmd(plumedmain,"GREX savePositions",NULL);
+ if(plumedWantsToStop) ir->nsteps=step_rel+1;
+ if(bHREX) plumed_cmd(plumedmain,"GREX cacheLocalUNow",&enerd->term[F_EPOT]);
+ }
+ /* END PLUMED */
+ }
+
+ if (EI_VV(ir->eI) && !startingFromCheckpoint)
+ /* ############### START FIRST UPDATE HALF-STEP FOR VV METHODS############### */
+ {
+ rvec *vbuf = nullptr;
+
+ 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.rvec_array(), 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(step, ir, mdatoms, state, f.arrayRefWithPadding(), fcd,
+ ekind, M, upd, etrtVELOCITY1,
+ cr, constr);
+
+ wallcycle_stop(wcycle, ewcUPDATE);
+ constrain_velocities(step, nullptr,
+ state,
+ shake_vir,
+ constr,
+ bCalcVir, do_log, do_ene);
+ wallcycle_start(wcycle, ewcUPDATE);
+ /* 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, &nullSignaller, state->box,
+ &totalNumberOfBondedInteractions, &bSumEkinhOld,
+ (bGStat ? CGLO_GSTAT : 0)
+ | (bCalcEner ? CGLO_ENERGY : 0)
+ | (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(mdlog, 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);
+
+ /* TODO This is only needed when we're about to write
+ * a checkpoint, because we use it after the restart
+ * (in a kludge?). But what should we be doing if
+ * startingFromCheckpoint or bInitStep are true? */
+ if (inputrecNptTrotter(ir) || inputrecNphTrotter(ir))
+ {
+ 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, nullptr, (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, nullptr, nullptr, nullptr, nullptr, mu_tot,
+ constr, &nullSignaller, state->box,
+ nullptr, &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.rvec_array(), 0, state->natoms);
+ sfree(vbuf);
+ }
+ wallcycle_stop(wcycle, ewcUPDATE);
+ }
+
+ /* compute the conserved quantity */
+ if (EI_VV(ir->eI))
+ {
+ saved_conserved_quantity = NPT_energy(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)
+ {
+ 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.rvec_array(), mdatoms);
+ /* history is maintained in state->dfhist, but state_global is what is sent to trajectory and log output */
+ if (MASTER(cr))
+ {
+ 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, observablesHistory,
+ top_global, fr,
+ outf, mdebin, ekind, f,
+ checkpointHandler->isCheckpointingStep(),
+ bRerunMD, bLastStep,
+ mdrunOptions.writeConfout,
+ 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.rvec_array(), ir, t, wcycle);
+
+ /* kludge -- virial is lost with restart for MTTK NPT control. Must reload (saved earlier). */
+ if (startingFromCheckpoint && (inputrecNptTrotter(ir) || inputrecNphTrotter(ir)))
+ {
+ copy_mat(state->svir_prev, shake_vir);
+ copy_mat(state->fvir_prev, force_vir);
+ }
+
+ stopHandler->setSignal();
+ resetHandler->setSignal(walltime_accounting);
+
+ if (bGStat || !PAR(cr))
+ {
+ /* In parallel we only have to check for checkpointing in steps
+ * where we do global communication,
+ * otherwise the other nodes don't know.
+ */
+ checkpointHandler->setSignal(walltime_accounting);
+ }
+
+ /* ######### 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->v, upd, constr);
+ /* if we have constraints, we have to remove the kinetic energy parallel to the bonds */
+ if (constr && bIfRandomize)
+ {
+ constrain_velocities(step, nullptr,
+ state,
+ tmp_vir,
+ constr,
+ bCalcVir, do_log, do_ene);
+ }
+ }
+ /* 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;
+
+ 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_before_coordinates(fplog, step, ir, state,
+ parrinellorahmanMu, M,
+ bInitStep);
+ }
+
+ if (EI_VV(ir->eI))
+ {
+ /* velocity half-step update */
+ update_coords(step, ir, mdatoms, state, f.arrayRefWithPadding(), 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(as_rvec_array(state->x.data()), cbuf, 0, state->natoms);
+ }
+
+ update_coords(step, ir, mdatoms, state, f.arrayRefWithPadding(), fcd,
+ ekind, M, upd, etrtPOSITION, cr, constr);
+ wallcycle_stop(wcycle, ewcUPDATE);
+
+ constrain_coordinates(step, &dvdl_constr, state,
+ shake_vir,
+ upd, constr,
+ bCalcVir, do_log, do_ene);
+ update_sd_second_half(step, &dvdl_constr, ir, mdatoms, state,
+ cr, nrnb, wcycle, upd, constr, do_log, do_ene);
+ finish_update(ir, mdatoms,
+ state, graph,
+ nrnb, wcycle, upd, constr);
+
+ if (ir->bPull && ir->pull->bSetPbcRefToPrevStepCOM)
+ {
+ updatePrevStepPullCom(ir->pull_work, state);
+ }
+
+ 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, &nullSignaller, lastbox,
+ nullptr, &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, as_rvec_array(state->x.data()), 0, state->natoms);
+
+ update_coords(step, ir, mdatoms, state, f.arrayRefWithPadding(), fcd,
+ ekind, M, upd, etrtPOSITION, cr, constr);
+ wallcycle_stop(wcycle, ewcUPDATE);
+
+ /* do we need an extra constraint here? just need to copy out of as_rvec_array(state->v.data()) 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*/
+ finish_update(ir, mdatoms,
+ state, graph,
+ nrnb, wcycle, upd, nullptr);
+ }
+ 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;
+ }
+
+ if (vsite != nullptr)
+ {
+ wallcycle_start(wcycle, ewcVSITECONSTR);
+ if (graph != nullptr)
+ {
+ shift_self(graph, state->box, state->x.rvec_array());
+ }
+ construct_vsites(vsite, state->x.rvec_array(), ir->delta_t, state->v.rvec_array(),
+ top->idef.iparams, top->idef.il,
+ fr->ePBC, fr->bMolPBC, cr, state->box);
+
+ if (graph != nullptr)
+ {
+ unshift_self(graph, state->box, state->x.rvec_array());
+ }
+ 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.
+ */
+ {
+ // Organize to do inter-simulation signalling on steps if
+ // and when algorithms require it.
+ bool doInterSimSignal = (simulationsShareState && do_per_step(step, nstSignalComm));
+
+ if (bGStat || (!EI_VV(ir->eI) && do_per_step(step+1, nstglobalcomm)) || doInterSimSignal)
+ {
+ // Since we're already communicating at this step, we
+ // can propagate intra-simulation signals. Note that
+ // check_nstglobalcomm has the responsibility for
+ // choosing the value of nstglobalcomm that is one way
+ // bGStat becomes true, so we can't get into a
+ // situation where e.g. checkpointing can't be
+ // signalled.
+ bool doIntraSimSignal = true;
+ SimulationSignaller signaller(&signals, cr, ms, doInterSimSignal, doIntraSimSignal);
+
+ compute_globals(fplog, gstat, cr, ir, fr, ekind, state, mdatoms, nrnb, vcm,
+ wcycle, enerd, force_vir, shake_vir, total_vir, pres, mu_tot,
+ constr, &signaller,
+ lastbox,
+ &totalNumberOfBondedInteractions, &bSumEkinhOld,
+ (bGStat ? CGLO_GSTAT : 0)
+ | (!EI_VV(ir->eI) && bCalcEner ? CGLO_ENERGY : 0)
+ | (!EI_VV(ir->eI) && bStopCM ? CGLO_STOPCM : 0)
+ | (!EI_VV(ir->eI) ? CGLO_TEMPERATURE : 0)
+ | (!EI_VV(ir->eI) ? CGLO_PRESSURE : 0)
+ | CGLO_CONSTRAINT
+ | (shouldCheckNumberOfBondedInteractions ? CGLO_CHECK_NUMBER_OF_BONDED_INTERACTIONS : 0)
+ );
+ checkNumberOfBondedInteractions(mdlog, 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))
+ {
+ /* 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_pcouple_after_coordinates(fplog, step, ir, mdatoms,
+ pres, force_vir, shake_vir,
+ parrinellorahmanMu,
+ state, 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;
+ }
+
+ if (bCalcEner)
+ {
+ /* ######### 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 (integratorHasConservedEnergyQuantity(ir))
+ {
+ 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] + NPT_energy(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 : nullptr,
+ 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 : nullptr,
+ step, t,
+ eprNORMAL, mdebin, fcd, groups, &(ir->opts), awh.get());
+
+ 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 (isMasterSimMasterRank(ms, 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,
+ as_rvec_array(state->x.data()),
+ state->box,
+ MASTER(cr) && mdrunOptions.verbose,
+ bRerunMD);
+
+ if (bNeedRepartition && DOMAINDECOMP(cr))
+ {
+ dd_collect_state(cr->dd, state, state_global);
+ }
+ }
+
+ /* Replica exchange */
+ bExchanged = FALSE;
+ if (bDoReplEx)
+ {
+ bExchanged = replica_exchange(fplog, cr, ms, repl_ex,
+ state_global, enerd,
+ state, step, t);
+ }
+
+ if ( (bExchanged || bNeedRepartition) && DOMAINDECOMP(cr) )
+ {
+ dd_partition_system(fplog, mdlog, 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->natoms);
+ }
+
+ 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 != nullptr) && (!bLastStep) )
+ {
+ rescale_membed(step_rel, membed, as_rvec_array(state_global->x.data()));
+ }
+
+ cycles = wallcycle_stop(wcycle, ewcSTEP);
+ if (DOMAINDECOMP(cr) && wcycle)
+ {
+ dd_cycles_add(cr->dd, cycles, ddCyclStep);
+ }
+
+ /* increase the MD step number */
+ step++;
+ step_rel++;
+
+ resetHandler->resetCounters(
+ step, step_rel, mdlog, fplog, cr, (use_GPU(fr->nbv) ? fr->nbv : nullptr),
+ nrnb, fr->pmedata, pme_loadbal, wcycle, walltime_accounting);
+
+ /* 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_time(walltime_accounting);
+
+ if (!thisRankHasDuty(cr, DUTY_PME))
+ {
+ /* Tell the PME only node to finish */
+ gmx_pme_send_finish(cr);
+ }
+
+ if (MASTER(cr))
+ {
+ if (ir->nstcalcenergy > 0)
+ {
+ print_ebin(mdoutf_get_fp_ene(outf), FALSE, FALSE, FALSE, fplog, step, t,
+ eprAVER, mdebin, fcd, groups, &(ir->opts), awh.get());
+ }
+ }
+ done_mdebin(mdebin);
+ done_mdoutf(outf);
+
+ if (bPMETune)
+ {
+ pme_loadbal_done(pme_loadbal, fplog, mdlog, use_GPU(fr->nbv));
+ }
+
+ done_shellfc(fplog, shellfc, step_rel);
+
+ if (useReplicaExchange && MASTER(cr))
+ {
+ print_replica_exchange_statistics(fplog, repl_ex);
+ }
+
+ // Clean up swapcoords
+ if (ir->eSwapCoords != eswapNO)
+ {
+ finish_swapcoords(ir->swap);
+ }
+
+ /* IMD cleanup, if bIMD is TRUE. */
+ IMD_finalize(ir->bIMD, ir->imd);
+
+ walltime_accounting_set_nsteps_done(walltime_accounting, step_rel);
+
+ destroy_enerdata(enerd);
+ sfree(enerd);
+ sfree(top);
+}
diff --git a/patches/gromacs-2019.1.diff/src/gromacs/mdrun/md.cpp.preplumed b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/md.cpp.preplumed
new file mode 100644
index 0000000000000000000000000000000000000000..a49a15df53a4246970f4f4a00e542f1c9a7cedbe
--- /dev/null
+++ b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/md.cpp.preplumed
@@ -0,0 +1,1531 @@
+/*
+ * 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) 2011,2012,2013,2014,2015,2016,2017,2018, 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 integrator for normal molecular dynamics simulations
+ *
+ * \author David van der Spoel <david.vanderspoel@icm.uu.se>
+ * \ingroup module_mdrun
+ */
+#include "gmxpre.h"
+
+#include <cinttypes>
+#include <cmath>
+#include <cstdio>
+#include <cstdlib>
+
+#include <algorithm>
+#include <memory>
+
+#include "gromacs/awh/awh.h"
+#include "gromacs/commandline/filenm.h"
+#include "gromacs/compat/make_unique.h"
+#include "gromacs/domdec/collect.h"
+#include "gromacs/domdec/domdec.h"
+#include "gromacs/domdec/domdec_network.h"
+#include "gromacs/domdec/domdec_struct.h"
+#include "gromacs/domdec/partition.h"
+#include "gromacs/essentialdynamics/edsam.h"
+#include "gromacs/ewald/pme.h"
+#include "gromacs/ewald/pme-load-balancing.h"
+#include "gromacs/fileio/trxio.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/checkpointhandler.h"
+#include "gromacs/mdlib/compute_io.h"
+#include "gromacs/mdlib/constr.h"
+#include "gromacs/mdlib/ebin.h"
+#include "gromacs/mdlib/expanded.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/mdsetup.h"
+#include "gromacs/mdlib/membed.h"
+#include "gromacs/mdlib/nb_verlet.h"
+#include "gromacs/mdlib/nbnxn_gpu_data_mgmt.h"
+#include "gromacs/mdlib/ns.h"
+#include "gromacs/mdlib/resethandler.h"
+#include "gromacs/mdlib/shellfc.h"
+#include "gromacs/mdlib/sighandler.h"
+#include "gromacs/mdlib/sim_util.h"
+#include "gromacs/mdlib/simulationsignal.h"
+#include "gromacs/mdlib/stophandler.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/awh-history.h"
+#include "gromacs/mdtypes/awh-params.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/observableshistory.h"
+#include "gromacs/mdtypes/pullhistory.h"
+#include "gromacs/mdtypes/state.h"
+#include "gromacs/pbcutil/mshift.h"
+#include "gromacs/pbcutil/pbc.h"
+#include "gromacs/pulling/output.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/logger.h"
+#include "gromacs/utility/real.h"
+#include "gromacs/utility/smalloc.h"
+
+#include "integrator.h"
+#include "replicaexchange.h"
+
+#if GMX_FAHCORE
+#include "corewrap.h"
+#endif
+
+using gmx::SimulationSignaller;
+
+void gmx::Integrator::do_md()
+{
+ // TODO Historically, the EM and MD "integrators" used different
+ // names for the t_inputrec *parameter, but these must have the
+ // same name, now that it's a member of a struct. We use this ir
+ // alias to avoid a large ripple of nearly useless changes.
+ // t_inputrec is being replaced by IMdpOptionsProvider, so this
+ // will go away eventually.
+ t_inputrec *ir = inputrec;
+ gmx_mdoutf *outf = nullptr;
+ int64_t step, step_rel;
+ double t, t0, lam0[efptNR];
+ gmx_bool bGStatEveryStep, bGStat, bCalcVir, bCalcEnerStep, bCalcEner;
+ gmx_bool bNS, bNStList, bSimAnn, bStopCM,
+ bFirstStep, bInitStep, bLastStep = FALSE;
+ gmx_bool bDoDHDL = FALSE, bDoFEP = FALSE, bDoExpanded = FALSE;
+ gmx_bool do_ene, do_log, do_verbose;
+ gmx_bool bMasterState;
+ int force_flags, cglo_flags;
+ tensor force_vir, shake_vir, total_vir, tmp_vir, pres;
+ int i, m;
+ rvec mu_tot;
+ t_vcm *vcm;
+ matrix parrinellorahmanMu, M;
+ gmx_repl_ex_t repl_ex = nullptr;
+ gmx_localtop_t *top;
+ t_mdebin *mdebin = nullptr;
+ gmx_enerdata_t *enerd;
+ PaddedVector<gmx::RVec> f {};
+ gmx_global_stat_t gstat;
+ gmx_update_t *upd = nullptr;
+ t_graph *graph = nullptr;
+ gmx_groups_t *groups;
+ gmx_ekindata_t *ekind;
+ gmx_shellfc_t *shellfc;
+ gmx_bool bSumEkinhOld, bDoReplEx, bExchanged, bNeedRepartition;
+ gmx_bool bTemp, bPres, bTrotter;
+ real dvdl_constr;
+ rvec *cbuf = nullptr;
+ 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];
+
+ /* PME load balancing data for GPU kernels */
+ pme_load_balancing_t *pme_loadbal = nullptr;
+ gmx_bool bPMETune = FALSE;
+ gmx_bool bPMETunePrinting = FALSE;
+
+ /* Interactive MD */
+ gmx_bool bIMDstep = FALSE;
+
+ /* 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;
+
+ SimulationSignals signals;
+ // Most global communnication stages don't propagate mdrun
+ // signals, and will use this object to achieve that.
+ SimulationSignaller nullSignaller(nullptr, nullptr, nullptr, false, false);
+
+ if (!mdrunOptions.writeConfout)
+ {
+ // This is on by default, and the main known use case for
+ // turning it off is for convenience in benchmarking, which is
+ // something that should not show up in the general user
+ // interface.
+ GMX_LOG(mdlog.info).asParagraph().
+ appendText("The -noconfout functionality is deprecated, and may be removed in a future version.");
+ }
+
+ /* 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)));
+
+ const bool bRerunMD = false;
+ int nstglobalcomm = mdrunOptions.globalCommunicationInterval;
+
+ nstglobalcomm = check_nstglobalcomm(mdlog, nstglobalcomm, ir, cr);
+ bGStatEveryStep = (nstglobalcomm == 1);
+
+ groups = &top_global->groups;
+
+ std::unique_ptr<EssentialDynamics> ed = nullptr;
+ if (opt2bSet("-ei", nfile, fnm) || observablesHistory->edsamHistory != nullptr)
+ {
+ /* Initialize essential dynamics sampling */
+ ed = init_edsam(mdlog,
+ opt2fn_null("-ei", nfile, fnm), opt2fn("-eo", nfile, fnm),
+ top_global,
+ ir, cr, constr,
+ state_global, observablesHistory,
+ oenv, mdrunOptions.continuationOptions.appendFiles);
+ }
+
+ /* Initial values */
+ init_md(fplog, cr, outputProvider, ir, oenv, mdrunOptions,
+ &t, &t0, state_global, lam0,
+ nrnb, top_global, &upd, deform,
+ nfile, fnm, &outf, &mdebin,
+ force_vir, shake_vir, total_vir, pres, mu_tot, &bSimAnn, &vcm, wcycle);
+
+ /* Energy terms and groups */
+ snew(enerd, 1);
+ init_enerdata(top_global->groups.grps[egcENER].nr, ir->fepvals->n_lambda,
+ enerd);
+
+ /* 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, constr ? constr->numFlexibleConstraints() : 0,
+ ir->nstcalcenergy, DOMAINDECOMP(cr));
+
+ {
+ 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);
+ }
+ }
+
+ /* Set up interactive MD (IMD) */
+ init_IMD(ir, cr, ms, top_global, fplog, ir->nstcalcenergy,
+ MASTER(cr) ? state_global->x.rvec_array() : nullptr,
+ nfile, fnm, oenv, mdrunOptions);
+
+ // Local state only becomes valid now.
+ std::unique_ptr<t_state> stateInstance;
+ t_state * state;
+
+ if (DOMAINDECOMP(cr))
+ {
+ top = dd_init_local_top(top_global);
+
+ stateInstance = compat::make_unique<t_state>();
+ state = stateInstance.get();
+ dd_init_local_state(cr->dd, state_global, state);
+
+ /* Distribute the charge groups over the nodes from the master node */
+ dd_partition_system(fplog, mdlog, ir->init_step, cr, TRUE, 1,
+ state_global, top_global, ir,
+ state, &f, mdAtoms, top, fr,
+ vsite, constr,
+ nrnb, nullptr, FALSE);
+ shouldCheckNumberOfBondedInteractions = true;
+ update_realloc(upd, state->natoms);
+ }
+ else
+ {
+ state_change_natoms(state_global, state_global->natoms);
+ f.resizeWithPadding(state_global->natoms);
+ /* Copy the pointer to the global state */
+ state = state_global;
+
+ snew(top, 1);
+ mdAlgorithmsSetupAtomData(cr, ir, top_global, top, fr,
+ &graph, mdAtoms, constr, vsite, shellfc);
+
+ update_realloc(upd, state->natoms);
+ }
+
+ auto mdatoms = mdAtoms->mdatoms();
+
+ // NOTE: The global state is no longer used at this point.
+ // But state_global is still used as temporary storage space for writing
+ // the global state to file and potentially for replica exchange.
+ // (Global topology should persist.)
+
+ update_mdatoms(mdatoms, state->lambda[efptMASS]);
+
+ const ContinuationOptions &continuationOptions = mdrunOptions.continuationOptions;
+ bool startingFromCheckpoint = continuationOptions.startedFromCheckpoint;
+
+ if (ir->bExpanded)
+ {
+ /* Check nstexpanded here, because the grompp check was broken */
+ if (ir->expandedvals->nstexpanded % ir->nstcalcenergy != 0)
+ {
+ gmx_fatal(FARGS, "With expanded ensemble, nstexpanded should be a multiple of nstcalcenergy");
+ }
+ init_expanded_ensemble(startingFromCheckpoint, ir, state->dfhist);
+ }
+
+ if (MASTER(cr))
+ {
+ if (startingFromCheckpoint)
+ {
+ /* Update mdebin with energy history if appending to output files */
+ if (continuationOptions.appendFiles)
+ {
+ /* If no history is available (because a checkpoint is from before
+ * it was written) make a new one later, otherwise restore it.
+ */
+ if (observablesHistory->energyHistory)
+ {
+ restore_energyhistory_from_state(mdebin, observablesHistory->energyHistory.get());
+ }
+ }
+ else if (observablesHistory->energyHistory)
+ {
+ /* We might have read an energy history from checkpoint.
+ * As we are not appending, we want to restart the statistics.
+ * Free the allocated memory and reset the counts.
+ */
+ observablesHistory->energyHistory = {};
+ /* We might have read a pull history from checkpoint.
+ * We will still want to keep the statistics, so that the files
+ * can be joined and still be meaningful.
+ * This means that observablesHistory->pullHistory
+ * should not be reset.
+ */
+ }
+ }
+ if (!observablesHistory->energyHistory)
+ {
+ observablesHistory->energyHistory = compat::make_unique<energyhistory_t>();
+ }
+ if (!observablesHistory->pullHistory)
+ {
+ observablesHistory->pullHistory = compat::make_unique<PullHistory>();
+ }
+ /* Set the initial energy history in state by updating once */
+ update_energyhistory(observablesHistory->energyHistory.get(), mdebin);
+ }
+
+ preparePrevStepPullCom(ir, mdatoms, state, state_global, cr, startingFromCheckpoint);
+
+ // TODO: Remove this by converting AWH into a ForceProvider
+ auto awh = prepareAwhModule(fplog, *ir, state_global, cr, ms, startingFromCheckpoint,
+ shellfc != nullptr,
+ opt2fn("-awh", nfile, fnm), ir->pull_work);
+
+ const bool useReplicaExchange = (replExParams.exchangeInterval > 0);
+ if (useReplicaExchange && MASTER(cr))
+ {
+ repl_ex = init_replica_exchange(fplog, ms, top_global->natoms, ir,
+ replExParams);
+ }
+ /* PME tuning is only supported in the Verlet scheme, with PME for
+ * Coulomb. It is not supported with only LJ PME. */
+ bPMETune = (mdrunOptions.tunePme && EEL_PME(fr->ic->eeltype) &&
+ !mdrunOptions.reproducible && ir->cutoff_scheme != ecutsGROUP);
+ if (bPMETune)
+ {
+ pme_loadbal_init(&pme_loadbal, cr, mdlog, *ir, state->box,
+ *fr->ic, *fr->nbv->listParams, fr->pmedata, use_GPU(fr->nbv),
+ &bPMETunePrinting);
+ }
+
+ if (!ir->bContinuation)
+ {
+ if (state->flags & (1 << estV))
+ {
+ auto v = makeArrayRef(state->v);
+ /* Set the velocities of vsites, shells and frozen atoms to zero */
+ for (i = 0; i < mdatoms->homenr; i++)
+ {
+ if (mdatoms->ptype[i] == eptVSite ||
+ mdatoms->ptype[i] == eptShell)
+ {
+ clear_rvec(v[i]);
+ }
+ else if (mdatoms->cFREEZE)
+ {
+ for (m = 0; m < DIM; m++)
+ {
+ if (ir->opts.nFreeze[mdatoms->cFREEZE[i]][m])
+ {
+ v[i][m] = 0;
+ }
+ }
+ }
+ }
+ }
+
+ if (constr)
+ {
+ /* Constrain the initial coordinates and velocities */
+ do_constrain_first(fplog, constr, ir, mdatoms, state);
+ }
+ if (vsite)
+ {
+ /* Construct the virtual sites for the initial configuration */
+ construct_vsites(vsite, state->x.rvec_array(), ir->delta_t, nullptr,
+ 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 (useReplicaExchange)
+ {
+ nstfep = gmx_greatest_common_divisor(replExParams.exchangeInterval, 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 (continuationOptions.haveReadEkin)
+ {
+ restore_ekinstate_from_state(cr, ekind, &state_global->ekinstate);
+ }
+
+ cglo_flags = (CGLO_INITIALIZATION | CGLO_TEMPERATURE | CGLO_GSTAT
+ | (EI_VV(ir->eI) ? CGLO_PRESSURE : 0)
+ | (EI_VV(ir->eI) ? CGLO_CONSTRAINT : 0)
+ | (continuationOptions.haveReadEkin ? CGLO_READEKIN : 0));
+
+ bSumEkinhOld = FALSE;
+ /* To minimize communication, compute_globals computes the COM velocity
+ * and the kinetic energy for the velocities without COM motion removed.
+ * Thus to get the kinetic energy without the COM contribution, we need
+ * to call compute_globals twice.
+ */
+ for (int cgloIteration = 0; cgloIteration < (bStopCM ? 2 : 1); cgloIteration++)
+ {
+ int cglo_flags_iteration = cglo_flags;
+ if (bStopCM && cgloIteration == 0)
+ {
+ cglo_flags_iteration |= CGLO_STOPCM;
+ cglo_flags_iteration &= ~CGLO_TEMPERATURE;
+ }
+ compute_globals(fplog, gstat, cr, ir, fr, ekind, state, mdatoms, nrnb, vcm,
+ nullptr, enerd, force_vir, shake_vir, total_vir, pres, mu_tot,
+ constr, &nullSignaller, state->box,
+ &totalNumberOfBondedInteractions, &bSumEkinhOld, cglo_flags_iteration
+ | (shouldCheckNumberOfBondedInteractions ? CGLO_CHECK_NUMBER_OF_BONDED_INTERACTIONS : 0));
+ }
+ checkNumberOfBondedInteractions(mdlog, 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,
+ nullptr, enerd, force_vir, shake_vir, total_vir, pres, mu_tot,
+ constr, &nullSignaller, state->box,
+ nullptr, &bSumEkinhOld,
+ cglo_flags & ~CGLO_PRESSURE);
+ }
+
+ /* Calculate the initial half step temperature, and save the ekinh_old */
+ if (!continuationOptions.startedFromCheckpoint)
+ {
+ for (i = 0; (i < ir->opts.ngtc); i++)
+ {
+ copy_mat(ekind->tcstat[i].ekinh, ekind->tcstat[i].ekinh_old);
+ }
+ }
+
+ /* 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 (!ir->bContinuation)
+ {
+ if (constr && ir->eConstrAlg == econtLINCS)
+ {
+ fprintf(fplog,
+ "RMS relative constraint deviation after constraining: %.2e\n",
+ constr->rmsd());
+ }
+ if (EI_STATE_VELOCITY(ir->eI))
+ {
+ real temp = enerd->term[F_TEMP];
+ if (ir->eI != eiVV)
+ {
+ /* Result of Ekin averaged over velocities of -half
+ * and +half step, while we only have -half step here.
+ */
+ temp *= 2;
+ }
+ fprintf(fplog, "Initial temperature: %g K\n", temp);
+ }
+ }
+
+ 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_time(walltime_accounting);
+ wallcycle_start(wcycle, ewcRUN);
+ print_start(fplog, cr, walltime_accounting, "mdrun");
+
+#if GMX_FAHCORE
+ /* safest point to do file checkpointing is here. More general point would be immediately before integrator call */
+ int 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
+ *
+ ************************************************************/
+
+ 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;
+
+ bool simulationsShareState = false;
+ int nstSignalComm = nstglobalcomm;
+ {
+ // TODO This implementation of ensemble orientation restraints is nasty because
+ // a user can't just do multi-sim with single-sim orientation restraints.
+ bool usingEnsembleRestraints = (fcd->disres.nsystems > 1) || ((ms != nullptr) && (fcd->orires.nr != 0));
+ bool awhUsesMultiSim = (ir->bDoAwh && ir->awhParams->shareBiasMultisim && (ms != nullptr));
+
+ // Replica exchange, ensemble restraints and AWH need all
+ // simulations to remain synchronized, so they need
+ // checkpoints and stop conditions to act on the same step, so
+ // the propagation of such signals must take place between
+ // simulations, not just within simulations.
+ // TODO: Make algorithm initializers set these flags.
+ simulationsShareState = useReplicaExchange || usingEnsembleRestraints || awhUsesMultiSim;
+
+ if (simulationsShareState)
+ {
+ // Inter-simulation signal communication does not need to happen
+ // often, so we use a minimum of 200 steps to reduce overhead.
+ const int c_minimumInterSimulationSignallingInterval = 200;
+ nstSignalComm = ((c_minimumInterSimulationSignallingInterval + nstglobalcomm - 1)/nstglobalcomm)*nstglobalcomm;
+ }
+ }
+
+ auto stopHandler = stopHandlerBuilder->getStopHandlerMD(
+ compat::not_null<SimulationSignal*>(&signals[eglsSTOPCOND]), simulationsShareState,
+ MASTER(cr), ir->nstlist, mdrunOptions.reproducible, nstSignalComm,
+ mdrunOptions.maximumHoursToRun, ir->nstlist == 0, fplog, step, bNS, walltime_accounting);
+
+ auto checkpointHandler = compat::make_unique<CheckpointHandler>(
+ compat::make_not_null<SimulationSignal*>(&signals[eglsCHKPT]),
+ simulationsShareState, ir->nstlist == 0, MASTER(cr),
+ mdrunOptions.writeConfout, mdrunOptions.checkpointOptions.period);
+
+ const bool resetCountersIsLocal = true;
+ auto resetHandler = compat::make_unique<ResetHandler>(
+ compat::make_not_null<SimulationSignal*>(&signals[eglsRESETCOUNTERS]), !resetCountersIsLocal,
+ ir->nsteps, MASTER(cr), mdrunOptions.timingOptions.resetHalfway,
+ mdrunOptions.maximumHoursToRun, mdlog, wcycle, walltime_accounting);
+
+ DdOpenBalanceRegionBeforeForceComputation ddOpenBalanceRegion = (DOMAINDECOMP(cr) ? DdOpenBalanceRegionBeforeForceComputation::yes : DdOpenBalanceRegionBeforeForceComputation::no);
+ DdCloseBalanceRegionAfterForceComputation ddCloseBalanceRegion = (DOMAINDECOMP(cr) ? DdCloseBalanceRegionAfterForceComputation::yes : DdCloseBalanceRegionAfterForceComputation::no);
+
+ step = ir->init_step;
+ step_rel = 0;
+
+ // TODO extract this to new multi-simulation module
+ if (MASTER(cr) && isMultiSim(ms) && !useReplicaExchange)
+ {
+ if (!multisim_int_all_are_equal(ms, ir->nsteps))
+ {
+ GMX_LOG(mdlog.warning).appendText(
+ "Note: The number of steps is not consistent across multi simulations,\n"
+ "but we are proceeding anyway!");
+ }
+ if (!multisim_int_all_are_equal(ms, ir->init_step))
+ {
+ GMX_LOG(mdlog.warning).appendText(
+ "Note: The initial step is not consistent across multi simulations,\n"
+ "but we are proceeding anyway!");
+ }
+ }
+
+ /* and stop now if we should */
+ bLastStep = (bLastStep || (ir->nsteps >= 0 && step_rel > ir->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,
+ (mdrunOptions.verbose && MASTER(cr)) ? stderr : nullptr,
+ fplog, mdlog,
+ *ir, fr, *state,
+ wcycle,
+ step, step_rel,
+ &bPMETunePrinting);
+ }
+
+ wallcycle_start(wcycle, ewcSTEP);
+
+ bLastStep = (step_rel == ir->nsteps);
+ t = t0 + step*ir->delta_t;
+
+ // TODO Refactor this, so that nstfep does not need a default value of zero
+ if (ir->efep != efepNO || ir->bSimTemp)
+ {
+ /* find and set the current lambdas */
+ setCurrentLambdasLocal(step, ir->fepvals, lam0, state);
+
+ 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 = (useReplicaExchange && (step > 0) && !bLastStep &&
+ do_per_step(step, replExParams.exchangeInterval));
+
+ if (bSimAnn)
+ {
+ update_annealing_target_temp(ir, t, upd);
+ }
+
+ /* Stop Center of Mass motion */
+ bStopCM = (ir->comm_mode != ecmNO && do_per_step(step, ir->nstcomm));
+
+ /* Determine whether or not to do Neighbour Searching */
+ bNS = (bFirstStep || bNStList || bExchanged || bNeedRepartition);
+
+ bLastStep = bLastStep || stopHandler->stoppingAfterCurrentStep(bNS);
+
+ /* 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;
+ do_verbose = mdrunOptions.verbose &&
+ (step % mdrunOptions.verboseStepPrintInterval == 0 || bFirstStep || bLastStep);
+
+ if (bNS && !(bFirstStep && ir->bContinuation))
+ {
+ 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, mdlog, 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->natoms);
+ }
+ }
+
+ 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 (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, nullptr, nullptr, nullptr, nullptr, mu_tot,
+ constr, &nullSignaller, state->box,
+ &totalNumberOfBondedInteractions, &bSumEkinhOld,
+ CGLO_GSTAT | CGLO_TEMPERATURE | CGLO_CHECK_NUMBER_OF_BONDED_INTERACTIONS);
+ checkNumberOfBondedInteractions(mdlog, cr, totalNumberOfBondedInteractions,
+ top_global, top, state,
+ &shouldCheckNumberOfBondedInteractions);
+ }
+ clear_mat(force_vir);
+
+ checkpointHandler->decideIfCheckpointingThisStep(bNS, bFirstStep, bLastStep);
+
+ /* Determine the energy and pressure:
+ * at nstcalcenergy steps and at energy output steps (set below).
+ */
+ if (EI_VV(ir->eI) && (!bInitStep))
+ {
+ bCalcEnerStep = do_per_step(step, 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);
+
+ 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)) ? 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, ms, mdrunOptions.verbose,
+ enforcedRotation, step,
+ ir, bNS, force_flags, top,
+ constr, enerd, fcd,
+ state, f.arrayRefWithPadding(), force_vir, mdatoms,
+ nrnb, wcycle, graph, groups,
+ shellfc, fr, ppForceWorkload, t, mu_tot,
+ vsite,
+ ddOpenBalanceRegion, ddCloseBalanceRegion);
+ }
+ else
+ {
+ /* The AWH history need to be saved _before_ doing force calculations where the AWH bias is updated
+ (or the AWH update will be performed twice for one step when continuing). It would be best to
+ call this update function from do_md_trajectory_writing but that would occur after do_force.
+ One would have to divide the update_awh function into one function applying the AWH force
+ and one doing the AWH bias update. The update AWH bias function could then be called after
+ do_md_trajectory_writing (then containing update_awh_history).
+ The checkpointing will in the future probably moved to the start of the md loop which will
+ rid of this issue. */
+ if (awh && checkpointHandler->isCheckpointingStep() && MASTER(cr))
+ {
+ awh->updateHistory(state_global->awhHistory.get());
+ }
+
+ /* 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, ms, ir, awh.get(), enforcedRotation,
+ step, nrnb, wcycle, top, groups,
+ state->box, state->x.arrayRefWithPadding(), &state->hist,
+ f.arrayRefWithPadding(), force_vir, mdatoms, enerd, fcd,
+ state->lambda, graph,
+ fr, ppForceWorkload, vsite, mu_tot, t, ed ? ed->getLegacyED() : nullptr,
+ (bNS ? GMX_FORCE_NS : 0) | force_flags,
+ ddOpenBalanceRegion, ddCloseBalanceRegion);
+ }
+
+ if (EI_VV(ir->eI) && !startingFromCheckpoint)
+ /* ############### START FIRST UPDATE HALF-STEP FOR VV METHODS############### */
+ {
+ rvec *vbuf = nullptr;
+
+ 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.rvec_array(), 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(step, ir, mdatoms, state, f.arrayRefWithPadding(), fcd,
+ ekind, M, upd, etrtVELOCITY1,
+ cr, constr);
+
+ wallcycle_stop(wcycle, ewcUPDATE);
+ constrain_velocities(step, nullptr,
+ state,
+ shake_vir,
+ constr,
+ bCalcVir, do_log, do_ene);
+ wallcycle_start(wcycle, ewcUPDATE);
+ /* 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, &nullSignaller, state->box,
+ &totalNumberOfBondedInteractions, &bSumEkinhOld,
+ (bGStat ? CGLO_GSTAT : 0)
+ | (bCalcEner ? CGLO_ENERGY : 0)
+ | (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(mdlog, 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);
+
+ /* TODO This is only needed when we're about to write
+ * a checkpoint, because we use it after the restart
+ * (in a kludge?). But what should we be doing if
+ * startingFromCheckpoint or bInitStep are true? */
+ if (inputrecNptTrotter(ir) || inputrecNphTrotter(ir))
+ {
+ 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, nullptr, (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, nullptr, nullptr, nullptr, nullptr, mu_tot,
+ constr, &nullSignaller, state->box,
+ nullptr, &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.rvec_array(), 0, state->natoms);
+ sfree(vbuf);
+ }
+ wallcycle_stop(wcycle, ewcUPDATE);
+ }
+
+ /* compute the conserved quantity */
+ if (EI_VV(ir->eI))
+ {
+ saved_conserved_quantity = NPT_energy(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)
+ {
+ 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.rvec_array(), mdatoms);
+ /* history is maintained in state->dfhist, but state_global is what is sent to trajectory and log output */
+ if (MASTER(cr))
+ {
+ 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, observablesHistory,
+ top_global, fr,
+ outf, mdebin, ekind, f,
+ checkpointHandler->isCheckpointingStep(),
+ bRerunMD, bLastStep,
+ mdrunOptions.writeConfout,
+ 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.rvec_array(), ir, t, wcycle);
+
+ /* kludge -- virial is lost with restart for MTTK NPT control. Must reload (saved earlier). */
+ if (startingFromCheckpoint && (inputrecNptTrotter(ir) || inputrecNphTrotter(ir)))
+ {
+ copy_mat(state->svir_prev, shake_vir);
+ copy_mat(state->fvir_prev, force_vir);
+ }
+
+ stopHandler->setSignal();
+ resetHandler->setSignal(walltime_accounting);
+
+ if (bGStat || !PAR(cr))
+ {
+ /* In parallel we only have to check for checkpointing in steps
+ * where we do global communication,
+ * otherwise the other nodes don't know.
+ */
+ checkpointHandler->setSignal(walltime_accounting);
+ }
+
+ /* ######### 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->v, upd, constr);
+ /* if we have constraints, we have to remove the kinetic energy parallel to the bonds */
+ if (constr && bIfRandomize)
+ {
+ constrain_velocities(step, nullptr,
+ state,
+ tmp_vir,
+ constr,
+ bCalcVir, do_log, do_ene);
+ }
+ }
+ /* 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;
+
+ 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_before_coordinates(fplog, step, ir, state,
+ parrinellorahmanMu, M,
+ bInitStep);
+ }
+
+ if (EI_VV(ir->eI))
+ {
+ /* velocity half-step update */
+ update_coords(step, ir, mdatoms, state, f.arrayRefWithPadding(), 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(as_rvec_array(state->x.data()), cbuf, 0, state->natoms);
+ }
+
+ update_coords(step, ir, mdatoms, state, f.arrayRefWithPadding(), fcd,
+ ekind, M, upd, etrtPOSITION, cr, constr);
+ wallcycle_stop(wcycle, ewcUPDATE);
+
+ constrain_coordinates(step, &dvdl_constr, state,
+ shake_vir,
+ upd, constr,
+ bCalcVir, do_log, do_ene);
+ update_sd_second_half(step, &dvdl_constr, ir, mdatoms, state,
+ cr, nrnb, wcycle, upd, constr, do_log, do_ene);
+ finish_update(ir, mdatoms,
+ state, graph,
+ nrnb, wcycle, upd, constr);
+
+ if (ir->bPull && ir->pull->bSetPbcRefToPrevStepCOM)
+ {
+ updatePrevStepPullCom(ir->pull_work, state);
+ }
+
+ 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, &nullSignaller, lastbox,
+ nullptr, &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, as_rvec_array(state->x.data()), 0, state->natoms);
+
+ update_coords(step, ir, mdatoms, state, f.arrayRefWithPadding(), fcd,
+ ekind, M, upd, etrtPOSITION, cr, constr);
+ wallcycle_stop(wcycle, ewcUPDATE);
+
+ /* do we need an extra constraint here? just need to copy out of as_rvec_array(state->v.data()) 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*/
+ finish_update(ir, mdatoms,
+ state, graph,
+ nrnb, wcycle, upd, nullptr);
+ }
+ 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;
+ }
+
+ if (vsite != nullptr)
+ {
+ wallcycle_start(wcycle, ewcVSITECONSTR);
+ if (graph != nullptr)
+ {
+ shift_self(graph, state->box, state->x.rvec_array());
+ }
+ construct_vsites(vsite, state->x.rvec_array(), ir->delta_t, state->v.rvec_array(),
+ top->idef.iparams, top->idef.il,
+ fr->ePBC, fr->bMolPBC, cr, state->box);
+
+ if (graph != nullptr)
+ {
+ unshift_self(graph, state->box, state->x.rvec_array());
+ }
+ 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.
+ */
+ {
+ // Organize to do inter-simulation signalling on steps if
+ // and when algorithms require it.
+ bool doInterSimSignal = (simulationsShareState && do_per_step(step, nstSignalComm));
+
+ if (bGStat || (!EI_VV(ir->eI) && do_per_step(step+1, nstglobalcomm)) || doInterSimSignal)
+ {
+ // Since we're already communicating at this step, we
+ // can propagate intra-simulation signals. Note that
+ // check_nstglobalcomm has the responsibility for
+ // choosing the value of nstglobalcomm that is one way
+ // bGStat becomes true, so we can't get into a
+ // situation where e.g. checkpointing can't be
+ // signalled.
+ bool doIntraSimSignal = true;
+ SimulationSignaller signaller(&signals, cr, ms, doInterSimSignal, doIntraSimSignal);
+
+ compute_globals(fplog, gstat, cr, ir, fr, ekind, state, mdatoms, nrnb, vcm,
+ wcycle, enerd, force_vir, shake_vir, total_vir, pres, mu_tot,
+ constr, &signaller,
+ lastbox,
+ &totalNumberOfBondedInteractions, &bSumEkinhOld,
+ (bGStat ? CGLO_GSTAT : 0)
+ | (!EI_VV(ir->eI) && bCalcEner ? CGLO_ENERGY : 0)
+ | (!EI_VV(ir->eI) && bStopCM ? CGLO_STOPCM : 0)
+ | (!EI_VV(ir->eI) ? CGLO_TEMPERATURE : 0)
+ | (!EI_VV(ir->eI) ? CGLO_PRESSURE : 0)
+ | CGLO_CONSTRAINT
+ | (shouldCheckNumberOfBondedInteractions ? CGLO_CHECK_NUMBER_OF_BONDED_INTERACTIONS : 0)
+ );
+ checkNumberOfBondedInteractions(mdlog, 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))
+ {
+ /* 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_pcouple_after_coordinates(fplog, step, ir, mdatoms,
+ pres, force_vir, shake_vir,
+ parrinellorahmanMu,
+ state, 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;
+ }
+
+ if (bCalcEner)
+ {
+ /* ######### 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 (integratorHasConservedEnergyQuantity(ir))
+ {
+ 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] + NPT_energy(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 : nullptr,
+ 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 : nullptr,
+ step, t,
+ eprNORMAL, mdebin, fcd, groups, &(ir->opts), awh.get());
+
+ 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 (isMasterSimMasterRank(ms, 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,
+ as_rvec_array(state->x.data()),
+ state->box,
+ MASTER(cr) && mdrunOptions.verbose,
+ bRerunMD);
+
+ if (bNeedRepartition && DOMAINDECOMP(cr))
+ {
+ dd_collect_state(cr->dd, state, state_global);
+ }
+ }
+
+ /* Replica exchange */
+ bExchanged = FALSE;
+ if (bDoReplEx)
+ {
+ bExchanged = replica_exchange(fplog, cr, ms, repl_ex,
+ state_global, enerd,
+ state, step, t);
+ }
+
+ if ( (bExchanged || bNeedRepartition) && DOMAINDECOMP(cr) )
+ {
+ dd_partition_system(fplog, mdlog, 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->natoms);
+ }
+
+ 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 != nullptr) && (!bLastStep) )
+ {
+ rescale_membed(step_rel, membed, as_rvec_array(state_global->x.data()));
+ }
+
+ cycles = wallcycle_stop(wcycle, ewcSTEP);
+ if (DOMAINDECOMP(cr) && wcycle)
+ {
+ dd_cycles_add(cr->dd, cycles, ddCyclStep);
+ }
+
+ /* increase the MD step number */
+ step++;
+ step_rel++;
+
+ resetHandler->resetCounters(
+ step, step_rel, mdlog, fplog, cr, (use_GPU(fr->nbv) ? fr->nbv : nullptr),
+ nrnb, fr->pmedata, pme_loadbal, wcycle, walltime_accounting);
+
+ /* 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_time(walltime_accounting);
+
+ if (!thisRankHasDuty(cr, DUTY_PME))
+ {
+ /* Tell the PME only node to finish */
+ gmx_pme_send_finish(cr);
+ }
+
+ if (MASTER(cr))
+ {
+ if (ir->nstcalcenergy > 0)
+ {
+ print_ebin(mdoutf_get_fp_ene(outf), FALSE, FALSE, FALSE, fplog, step, t,
+ eprAVER, mdebin, fcd, groups, &(ir->opts), awh.get());
+ }
+ }
+ done_mdebin(mdebin);
+ done_mdoutf(outf);
+
+ if (bPMETune)
+ {
+ pme_loadbal_done(pme_loadbal, fplog, mdlog, use_GPU(fr->nbv));
+ }
+
+ done_shellfc(fplog, shellfc, step_rel);
+
+ if (useReplicaExchange && MASTER(cr))
+ {
+ print_replica_exchange_statistics(fplog, repl_ex);
+ }
+
+ // Clean up swapcoords
+ if (ir->eSwapCoords != eswapNO)
+ {
+ finish_swapcoords(ir->swap);
+ }
+
+ /* IMD cleanup, if bIMD is TRUE. */
+ IMD_finalize(ir->bIMD, ir->imd);
+
+ walltime_accounting_set_nsteps_done(walltime_accounting, step_rel);
+
+ destroy_enerdata(enerd);
+ sfree(enerd);
+ sfree(top);
+}
diff --git a/patches/gromacs-2019.1.diff/src/gromacs/mdrun/minimize.cpp b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/minimize.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..62a5a07a08daec87adf60e3947b0c3ecb006b4ad
--- /dev/null
+++ b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/minimize.cpp
@@ -0,0 +1,3064 @@
+/*
+ * 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) 2013,2014,2015,2016,2017,2018, 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_mdrun
+ */
+#include "gmxpre.h"
+
+#include "config.h"
+
+#include <cmath>
+#include <cstring>
+#include <ctime>
+
+#include <algorithm>
+#include <vector>
+
+#include "gromacs/commandline/filenm.h"
+#include "gromacs/domdec/collect.h"
+#include "gromacs/domdec/domdec.h"
+#include "gromacs/domdec/domdec_struct.h"
+#include "gromacs/domdec/partition.h"
+#include "gromacs/ewald/pme.h"
+#include "gromacs/fileio/confio.h"
+#include "gromacs/fileio/mtxio.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/mdsetup.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/mdtypes/state.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/topology/topology.h"
+#include "gromacs/utility/cstringutil.h"
+#include "gromacs/utility/exceptions.h"
+#include "gromacs/utility/fatalerror.h"
+#include "gromacs/utility/logger.h"
+#include "gromacs/utility/smalloc.h"
+
+#include "integrator.h"
+
+/* PLUMED */
+#include "../../../Plumed.h"
+extern int plumedswitch;
+extern plumed plumedmain;
+extern void(*plumedcmd)(plumed,const char*,const void*);
+/* END PLUMED */
+
+//! Utility structure for manipulating states during EM
+typedef struct {
+ //! Copy of the global state
+ t_state s;
+ //! Force array
+ PaddedVector<gmx::RVec> f;
+ //! Potential energy
+ real epot;
+ //! Norm of the force
+ real fnorm;
+ //! Maximum force
+ real fmax;
+ //! Direction
+ int a_fmax;
+} em_state_t;
+
+//! Print the EM starting conditions
+static void print_em_start(FILE *fplog,
+ const t_commrec *cr,
+ gmx_walltime_accounting_t walltime_accounting,
+ gmx_wallcycle_t wcycle,
+ const char *name)
+{
+ walltime_accounting_start_time(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_time(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,
+ real fmax,
+ gmx_bool bLastStep,
+ gmx_bool bConstrain)
+{
+ constexpr bool realIsDouble = GMX_DOUBLE;
+ char buffer[2048];
+
+ if (!std::isfinite(fmax))
+ {
+ sprintf(buffer,
+ "\nEnergy minimization has stopped because the force "
+ "on at least one atom is not finite. This usually means "
+ "atoms are overlapping. Modify the input coordinates to "
+ "remove atom overlap or use soft-core potentials with "
+ "the free energy code to avoid infinite forces.\n%s",
+ !realIsDouble ?
+ "You could also be lucky that switching to double precision "
+ "is sufficient to obtain finite forces.\n" :
+ "");
+ }
+ else 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,
+ !realIsDouble ?
+ "\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), stderr);
+ 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,
+ int64_t count, gmx_bool bDone, int64_t nsteps,
+ const em_state_t *ems, double sqrtNumAtoms)
+{
+ 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", ems->epot);
+ fprintf(fp, "Maximum force = %21.14e on atom %d\n", ems->fmax, ems->a_fmax + 1);
+ fprintf(fp, "Norm of force = %21.14e\n", ems->fnorm/sqrtNumAtoms);
+#else
+ fprintf(fp, "Potential Energy = %14.7e\n", ems->epot);
+ fprintf(fp, "Maximum force = %14.7e on atom %d\n", ems->fmax, ems->a_fmax + 1);
+ fprintf(fp, "Norm of force = %14.7e\n", ems->fnorm/sqrtNumAtoms);
+#endif
+}
+
+//! Compute the norm and max of the force array in parallel
+static void get_f_norm_max(const t_commrec *cr,
+ t_grpopts *opts, t_mdatoms *mdatoms, const 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->globalAtomIndices[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 = gmx::roundToInt(sum[2*i+1]);
+ }
+ }
+ 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(const t_commrec *cr,
+ t_grpopts *opts, t_mdatoms *mdatoms,
+ em_state_t *ems)
+{
+ get_f_norm_max(cr, opts, mdatoms, ems->f.rvec_array(),
+ &ems->fnorm, &ems->fmax, &ems->a_fmax);
+}
+
+//! Initialize the energy minimization
+static void init_em(FILE *fplog,
+ const gmx::MDLogger &mdlog,
+ const char *title,
+ const t_commrec *cr,
+ const gmx_multisim_t *ms,
+ gmx::IMDOutputProvider *outputProvider,
+ t_inputrec *ir,
+ const MdrunOptions &mdrunOptions,
+ t_state *state_global, gmx_mtop_t *top_global,
+ em_state_t *ems, gmx_localtop_t **top,
+ t_nrnb *nrnb, rvec mu_tot,
+ t_forcerec *fr, gmx_enerdata_t **enerd,
+ t_graph **graph, gmx::MDAtoms *mdAtoms, gmx_global_stat_t *gstat,
+ gmx_vsite_t *vsite, gmx::Constraints *constr, gmx_shellfc_t **shellfc,
+ int nfile, const t_filenm fnm[],
+ gmx_mdoutf_t *outf, t_mdebin **mdebin,
+ gmx_wallcycle_t wcycle)
+{
+ real dvdl_constr;
+
+ if (fplog)
+ {
+ fprintf(fplog, "Initiating %s\n", title);
+ }
+
+ if (MASTER(cr))
+ {
+ state_global->ngtc = 0;
+
+ /* Initialize lambda variables */
+ initialize_lambdas(fplog, ir, &(state_global->fep_state), state_global->lambda, nullptr);
+ }
+
+ init_nrnb(nrnb);
+
+ /* Interactive molecular dynamics */
+ init_IMD(ir, cr, ms, top_global, fplog, 1,
+ MASTER(cr) ? state_global->x.rvec_array() : nullptr,
+ nfile, fnm, nullptr, mdrunOptions);
+
+ if (ir->eI == eiNM)
+ {
+ GMX_ASSERT(shellfc != nullptr, "With NM we always support shells");
+
+ *shellfc = init_shell_flexcon(stdout,
+ top_global,
+ constr ? constr->numFlexibleConstraints() : 0,
+ ir->nstcalcenergy,
+ DOMAINDECOMP(cr));
+ }
+ else
+ {
+ GMX_ASSERT(EI_ENERGY_MINIMIZATION(ir->eI), "This else currently only handles energy minimizers, consider if your algorithm needs shell/flexible-constraint support");
+
+ /* With energy minimization, shells and flexible constraints are
+ * automatically minimized when treated like normal DOFS.
+ */
+ if (shellfc != nullptr)
+ {
+ *shellfc = nullptr;
+ }
+ }
+
+ auto mdatoms = mdAtoms->mdatoms();
+ if (DOMAINDECOMP(cr))
+ {
+ *top = dd_init_local_top(top_global);
+
+ dd_init_local_state(cr->dd, state_global, &ems->s);
+
+ /* Distribute the charge groups over the nodes from the master node */
+ dd_partition_system(fplog, mdlog, ir->init_step, cr, TRUE, 1,
+ state_global, top_global, ir,
+ &ems->s, &ems->f, mdAtoms, *top,
+ fr, vsite, constr,
+ nrnb, nullptr, FALSE);
+ dd_store_state(cr->dd, &ems->s);
+
+ *graph = nullptr;
+ }
+ else
+ {
+ state_change_natoms(state_global, state_global->natoms);
+ /* Just copy the state */
+ ems->s = *state_global;
+ state_change_natoms(&ems->s, ems->s.natoms);
+ ems->f.resizeWithPadding(ems->s.natoms);
+
+ snew(*top, 1);
+ mdAlgorithmsSetupAtomData(cr, ir, top_global, *top, fr,
+ graph, mdAtoms,
+ constr, vsite, shellfc ? *shellfc : nullptr);
+
+ if (vsite)
+ {
+ set_vsite_top(vsite, *top, mdatoms);
+ }
+ }
+
+ update_mdatoms(mdAtoms->mdatoms(), ems->s.lambda[efptMASS]);
+
+ if (constr)
+ {
+ // TODO how should this cross-module support dependency be managed?
+ 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 (!ir->bContinuation)
+ {
+ /* Constrain the starting coordinates */
+ dvdl_constr = 0;
+ constr->apply(TRUE, TRUE,
+ -1, 0, 1.0,
+ ems->s.x.rvec_array(),
+ ems->s.x.rvec_array(),
+ nullptr,
+ ems->s.box,
+ ems->s.lambda[efptFEP], &dvdl_constr,
+ nullptr, nullptr, gmx::ConstraintVariable::Positions);
+ }
+ }
+
+ if (PAR(cr))
+ {
+ *gstat = global_stat_init(ir);
+ }
+ else
+ {
+ *gstat = nullptr;
+ }
+
+ *outf = init_mdoutf(fplog, nfile, fnm, mdrunOptions, cr, outputProvider, ir, top_global, nullptr, wcycle);
+
+ snew(*enerd, 1);
+ init_enerdata(top_global->groups.grps[egcENER].nr, ir->fepvals->n_lambda,
+ *enerd);
+
+ if (mdebin != nullptr)
+ {
+ /* Init bin for energy stuff */
+ *mdebin = init_mdebin(mdoutf_get_fp_ene(*outf), top_global, ir, nullptr);
+ }
+
+ clear_rvec(mu_tot);
+ calc_shifts(ems->s.box, fr->shift_vec);
+
+ /* PLUMED */
+ if(plumedswitch){
+ if(ms && ms->nsim>1) {
+ if(MASTER(cr)) (*plumedcmd) (plumedmain,"GREX setMPIIntercomm",&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)) {
+ int nat_home = dd_numHomeAtoms(*cr->dd);
+ (*plumedcmd) (plumedmain,"setAtomsNlocal",&nat_home);
+ (*plumedcmd) (plumedmain,"setAtomsGatindex",cr->dd->globalAtomIndices.data());
+
+ }
+ }
+ }
+ /* END PLUMED */
+}
+
+//! Finalize the minimization
+static void finish_em(const t_commrec *cr, gmx_mdoutf_t outf,
+ gmx_walltime_accounting_t walltime_accounting,
+ gmx_wallcycle_t wcycle)
+{
+ if (!thisRankHasDuty(cr, 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;
+}
+
+//! Save the EM trajectory
+static void write_em_traj(FILE *fplog, const t_commrec *cr,
+ gmx_mdoutf_t outf,
+ gmx_bool bX, gmx_bool bF, const char *confout,
+ gmx_mtop_t *top_global,
+ t_inputrec *ir, int64_t step,
+ em_state_t *state,
+ t_state *state_global,
+ ObservablesHistory *observablesHistory)
+{
+ int 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, static_cast<double>(step),
+ &state->s, state_global, observablesHistory,
+ state->f);
+
+ if (confout != nullptr)
+ {
+ if (DOMAINDECOMP(cr))
+ {
+ /* If bX=true, x was collected to state_global in the call above */
+ if (!bX)
+ {
+ gmx::ArrayRef<gmx::RVec> globalXRef = MASTER(cr) ? makeArrayRef(state_global->x) : gmx::EmptyArrayRef();
+ dd_collect_vec(cr->dd, &state->s, makeArrayRef(state->s.x), globalXRef);
+ }
+ }
+ else
+ {
+ /* Copy the local state pointer */
+ state_global = &state->s;
+ }
+
+ if (MASTER(cr))
+ {
+ if (ir->ePBC != epbcNONE && !ir->bPeriodicMols && DOMAINDECOMP(cr))
+ {
+ /* Make molecules whole only for confout writing */
+ do_pbc_mtop(fplog, ir->ePBC, state->s.box, top_global,
+ state_global->x.rvec_array());
+ }
+
+ write_sto_conf_mtop(confout,
+ *top_global->name, top_global,
+ state_global->x.rvec_array(), nullptr, ir->ePBC, state->s.box);
+ }
+ }
+}
+
+//! \brief Do one minimization step
+//
+// \returns true when the step succeeded, false when a constraint error occurred
+static bool do_em_step(const t_commrec *cr,
+ t_inputrec *ir, t_mdatoms *md,
+ em_state_t *ems1, real a, const PaddedVector<gmx::RVec> *force,
+ em_state_t *ems2,
+ gmx::Constraints *constr,
+ int64_t count)
+
+{
+ t_state *s1, *s2;
+ int start, end;
+ real dvdl_constr;
+ int nthreads gmx_unused;
+
+ bool validStep = true;
+
+ 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->natoms != s1->natoms)
+ {
+ state_change_natoms(s2, s1->natoms);
+ ems2->f.resizeWithPadding(s2->natoms);
+ }
+ if (DOMAINDECOMP(cr) && s2->cg_gl.size() != s1->cg_gl.size())
+ {
+ s2->cg_gl.resize(s1->cg_gl.size());
+ }
+
+ copy_mat(s1->box, s2->box);
+ /* Copy free energy state */
+ s2->lambda = s1->lambda;
+ copy_mat(s1->box, s2->box);
+
+ start = 0;
+ end = md->homenr;
+
+ nthreads = gmx_omp_nthreads_get(emntUpdate);
+#pragma omp parallel num_threads(nthreads)
+ {
+ const rvec *x1 = s1->x.rvec_array();
+ rvec *x2 = s2->x.rvec_array();
+ const rvec *f = force->rvec_array();
+
+ int gf = 0;
+#pragma omp for schedule(static) nowait
+ for (int i = start; i < end; i++)
+ {
+ try
+ {
+ if (md->cFREEZE)
+ {
+ gf = md->cFREEZE[i];
+ }
+ for (int 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 */
+ const rvec *p1 = s1->cg_p.rvec_array();
+ rvec *p2 = s2->cg_p.rvec_array();
+#pragma omp for schedule(static) nowait
+ for (int i = start; i < end; i++)
+ {
+ // Trivial OpenMP block that does not throw
+ copy_rvec(p1[i], p2[i]);
+ }
+ }
+
+ if (DOMAINDECOMP(cr))
+ {
+ s2->ddp_count = s1->ddp_count;
+
+ /* OpenMP does not supported unsigned loop variables */
+#pragma omp for schedule(static) nowait
+ for (int i = 0; i < static_cast<int>(s2->cg_gl.size()); i++)
+ {
+ s2->cg_gl[i] = s1->cg_gl[i];
+ }
+ s2->ddp_count_cg_gl = s1->ddp_count_cg_gl;
+ }
+ }
+
+ if (constr)
+ {
+ dvdl_constr = 0;
+ validStep =
+ constr->apply(TRUE, TRUE,
+ count, 0, 1.0,
+ s1->x.rvec_array(), s2->x.rvec_array(),
+ nullptr, s2->box,
+ s2->lambda[efptBONDED], &dvdl_constr,
+ nullptr, nullptr, gmx::ConstraintVariable::Positions);
+
+ if (cr->nnodes > 1)
+ {
+ /* This global reduction will affect performance at high
+ * parallelization, but we can not really avoid it.
+ * But usually EM is not run at high parallelization.
+ */
+ int reductionBuffer = static_cast<int>(!validStep);
+ gmx_sumi(1, &reductionBuffer, cr);
+ validStep = (reductionBuffer == 0);
+ }
+
+ // We should move this check to the different minimizers
+ if (!validStep && ir->eI != eiSteep)
+ {
+ gmx_fatal(FARGS, "The coordinates could not be constrained. Minimizer '%s' can not handle constraint failures, use minimizer '%s' before using '%s'.",
+ EI(ir->eI), EI(eiSteep), EI(ir->eI));
+ }
+ }
+
+ return validStep;
+}
+
+//! Prepare EM for using domain decomposition parallellization
+static void em_dd_partition_system(FILE *fplog,
+ const gmx::MDLogger &mdlog,
+ int step, const t_commrec *cr,
+ gmx_mtop_t *top_global, t_inputrec *ir,
+ em_state_t *ems, gmx_localtop_t *top,
+ gmx::MDAtoms *mdAtoms, t_forcerec *fr,
+ gmx_vsite_t *vsite, gmx::Constraints *constr,
+ t_nrnb *nrnb, gmx_wallcycle_t wcycle)
+{
+ /* Repartition the domain decomposition */
+ dd_partition_system(fplog, mdlog, step, cr, FALSE, 1,
+ nullptr, top_global, ir,
+ &ems->s, &ems->f,
+ mdAtoms, top, fr, vsite, constr,
+ nrnb, wcycle, FALSE);
+ dd_store_state(cr->dd, &ems->s);
+}
+
+namespace
+{
+
+/*! \brief Class to handle the work of setting and doing an energy evaluation.
+ *
+ * This class is a mere aggregate of parameters to pass to evaluate an
+ * energy, so that future changes to names and types of them consume
+ * less time when refactoring other code.
+ *
+ * Aggregate initialization is used, for which the chief risk is that
+ * if a member is added at the end and not all initializer lists are
+ * updated, then the member will be value initialized, which will
+ * typically mean initialization to zero.
+ *
+ * We only want to construct one of these with an initializer list, so
+ * we explicitly delete the default constructor. */
+class EnergyEvaluator
+{
+ public:
+ //! We only intend to construct such objects with an initializer list.
+#if __GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 9)
+ // Aspects of the C++11 spec changed after GCC 4.8.5, and
+ // compilation of the initializer list construction in
+ // runner.cpp fails in GCC 4.8.5.
+ EnergyEvaluator() = delete;
+#endif
+ /*! \brief Evaluates an energy on the state in \c ems.
+ *
+ * \todo In practice, the same objects mu_tot, vir, and pres
+ * are always passed to this function, so we would rather have
+ * them as data members. However, their C-array types are
+ * unsuited for aggregate initialization. When the types
+ * improve, the call signature of this method can be reduced.
+ */
+ void run(em_state_t *ems, rvec mu_tot,
+ tensor vir, tensor pres,
+ int64_t count, gmx_bool bFirst);
+ //! Handles logging (deprecated).
+ FILE *fplog;
+ //! Handles logging.
+ const gmx::MDLogger &mdlog;
+ //! Handles communication.
+ const t_commrec *cr;
+ //! Coordinates multi-simulations.
+ const gmx_multisim_t *ms;
+ //! Holds the simulation topology.
+ gmx_mtop_t *top_global;
+ //! Holds the domain topology.
+ gmx_localtop_t *top;
+ //! User input options.
+ t_inputrec *inputrec;
+ //! Manages flop accounting.
+ t_nrnb *nrnb;
+ //! Manages wall cycle accounting.
+ gmx_wallcycle_t wcycle;
+ //! Coordinates global reduction.
+ gmx_global_stat_t gstat;
+ //! Handles virtual sites.
+ gmx_vsite_t *vsite;
+ //! Handles constraints.
+ gmx::Constraints *constr;
+ //! Handles strange things.
+ t_fcdata *fcd;
+ //! Molecular graph for SHAKE.
+ t_graph *graph;
+ //! Per-atom data for this domain.
+ gmx::MDAtoms *mdAtoms;
+ //! Handles how to calculate the forces.
+ t_forcerec *fr;
+ //! Schedule of force-calculation work each step for this task.
+ gmx::PpForceWorkload *ppForceWorkload;
+ //! Stores the computed energies.
+ gmx_enerdata_t *enerd;
+};
+
+void
+EnergyEvaluator::run(em_state_t *ems, rvec mu_tot,
+ tensor vir, tensor pres,
+ 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.rvec_array(), 1, nullptr,
+ 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, mdlog, 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->mdatoms()->massT[0]);
+ (*plumedcmd) (plumedmain,"setCharges",&mdAtoms->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, ms, inputrec, nullptr, nullptr,
+ count, nrnb, wcycle, top, &top_global->groups,
+ ems->s.box, ems->s.x.arrayRefWithPadding(), &ems->s.hist,
+ ems->f.arrayRefWithPadding(), force_vir, mdAtoms->mdatoms(), enerd, fcd,
+ ems->s.lambda, graph, fr, ppForceWorkload, vsite, mu_tot, t, nullptr,
+ GMX_FORCE_STATECHANGED | GMX_FORCE_ALLFORCES |
+ GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY |
+ (bNS ? GMX_FORCE_NS : 0),
+ DOMAINDECOMP(cr) ?
+ DdOpenBalanceRegionBeforeForceComputation::yes :
+ DdOpenBalanceRegionBeforeForceComputation::no,
+ DOMAINDECOMP(cr) ?
+ DdCloseBalanceRegionAfterForceComputation::yes :
+ DdCloseBalanceRegionAfterForceComputation::no);
+ /* 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, nullptr, nullptr, nullptr, 1, &terminate,
+ nullptr, 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 */
+ dvdl_constr = 0;
+ rvec *f_rvec = ems->f.rvec_array();
+ constr->apply(FALSE, FALSE,
+ count, 0, 1.0,
+ ems->s.x.rvec_array(), f_rvec, f_rvec,
+ ems->s.box,
+ ems->s.lambda[efptBONDED], &dvdl_constr,
+ nullptr, &shake_vir, gmx::ConstraintVariable::ForceDispl);
+ enerd->term[F_DVDL_CONSTR] += dvdl_constr;
+ m_add(force_vir, shake_vir, vir);
+ }
+ 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->mdatoms(), ems);
+ }
+}
+
+} // namespace
+
+//! Parallel utility summing energies and forces
+static double reorder_partsum(const t_commrec *cr, t_grpopts *opts, t_mdatoms *mdatoms,
+ gmx_mtop_t *top_global,
+ em_state_t *s_min, em_state_t *s_b)
+{
+ 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");
+ }
+
+ const rvec *fm = s_min->f.rvec_array();
+ const rvec *fb = s_b->f.rvec_array();
+
+ 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.
+ */
+ rvec *fmg;
+ snew(fmg, top_global->natoms);
+
+ ncg = s_min->s.cg_gl.size();
+ cg_gl = s_min->s.cg_gl.data();
+ 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.cg_gl.size();
+ cg_gl = s_b->s.cg_gl.data();
+ 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(const t_commrec *cr, t_grpopts *opts, t_mdatoms *mdatoms,
+ gmx_mtop_t *top_global,
+ em_state_t *s_min, em_state_t *s_b)
+{
+ double sum;
+
+ /* 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))
+ {
+ const rvec *fm = s_min->f.rvec_array();
+ const rvec *fb = s_b->f.rvec_array();
+ sum = 0;
+ int gf = 0;
+ /* This part of code can be incorrect with DD,
+ * since the atom ordering in s_b and s_min might differ.
+ */
+ for (int i = 0; i < mdatoms->homenr; i++)
+ {
+ if (mdatoms->cFREEZE)
+ {
+ gf = mdatoms->cFREEZE[i];
+ }
+ for (int 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
+{
+
+void
+Integrator::do_cg()
+{
+ const char *CG = "Polak-Ribiere Conjugate Gradients";
+
+ gmx_localtop_t *top;
+ gmx_enerdata_t *enerd;
+ gmx_global_stat_t gstat;
+ t_graph *graph;
+ double tmp, minstep;
+ 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 m, step, nminstep;
+ auto mdatoms = mdAtoms->mdatoms();
+
+ GMX_LOG(mdlog.info).asParagraph().
+ appendText("Note that activating conjugate gradient energy minimization via the "
+ "integrator .mdp option and the command gmx mdrun may "
+ "be available in a different form in a future version of GROMACS, "
+ "e.g. gmx minimize and an .mdp option.");
+
+ step = 0;
+
+ if (MASTER(cr))
+ {
+ // In CG, the state is extended with a search direction
+ state_global->flags |= (1<<estCGP);
+
+ // Ensure the extra per-atom state array gets allocated
+ state_change_natoms(state_global, state_global->natoms);
+
+ // Initialize the search direction to zero
+ for (RVec &cg_p : state_global->cg_p)
+ {
+ cg_p = { 0, 0, 0 };
+ }
+ }
+
+ /* Create 4 states on the stack and extract pointers that we will swap */
+ em_state_t s0 {}, s1 {}, s2 {}, s3 {};
+ em_state_t *s_min = &s0;
+ em_state_t *s_a = &s1;
+ em_state_t *s_b = &s2;
+ em_state_t *s_c = &s3;
+
+ /* Init em and store the local state in s_min */
+ init_em(fplog, mdlog, CG, cr, ms, outputProvider, inputrec, mdrunOptions,
+ state_global, top_global, s_min, &top,
+ nrnb, mu_tot, fr, &enerd, &graph, mdAtoms, &gstat,
+ vsite, constr, nullptr,
+ nfile, fnm, &outf, &mdebin, 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);
+ }
+
+ EnergyEvaluator energyEvaluator {
+ fplog, mdlog, cr, ms,
+ top_global, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph,
+ mdAtoms, fr, ppForceWorkload, enerd
+ };
+ /* 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
+ */
+ energyEvaluator.run(s_min, mu_tot, vir, pres, -1, TRUE);
+
+ if (MASTER(cr))
+ {
+ /* Copy stuff to the energy bin for easy printing etc. */
+ matrix nullBox = {};
+ upd_mdebin(mdebin, FALSE, FALSE, static_cast<double>(step),
+ mdatoms->tmass, enerd, nullptr, nullptr, nullptr, nullBox,
+ nullptr, nullptr, vir, pres, nullptr, 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), nullptr);
+ }
+
+ /* 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 */
+ rvec *pm = s_min->s.cg_p.rvec_array();
+ const rvec *sfm = s_min->f.rvec_array();
+ double gpa = 0;
+ int gf = 0;
+ for (int 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])
+ {
+ pm[i][m] = sfm[i][m] + beta*pm[i][m];
+ gpa -= pm[i][m]*sfm[i][m];
+ /* f is negative gradient, thus the sign */
+ }
+ else
+ {
+ pm[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, pm, &pnorm, nullptr, nullptr);
+
+ /* 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;
+ auto s_min_x = makeArrayRef(s_min->s.x);
+ for (int i = 0; i < mdatoms->homenr; i++)
+ {
+ for (m = 0; m < DIM; m++)
+ {
+ tmp = fabs(s_min_x[i][m]);
+ if (tmp < 1.0)
+ {
+ tmp = 1.0;
+ }
+ tmp = pm[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*top_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, nullptr,
+ top_global, inputrec, step,
+ s_min, state_global, observablesHistory);
+
+ /* 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, mdlog, 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, s_min, c, &s_min->s.cg_p, s_c,
+ constr, -1);
+
+ neval++;
+ /* Calculate energy for the trial step */
+ energyEvaluator.run(s_c, mu_tot, vir, pres, -1, FALSE);
+
+ /* Calc derivative along line */
+ const rvec *pc = s_c->s.cg_p.rvec_array();
+ const rvec *sfc = s_c->f.rvec_array();
+ double gpc = 0;
+ for (int i = 0; i < mdatoms->homenr; i++)
+ {
+ for (m = 0; m < DIM; m++)
+ {
+ gpc -= pc[i][m]*sfc[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 = std::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.
+ */
+ double gpb;
+ 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, mdlog, -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, s_min, b, &s_min->s.cg_p, s_b,
+ constr, -1);
+
+ neval++;
+ /* Calculate energy for the trial step */
+ energyEvaluator.run(s_b, mu_tot, 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.
+ */
+ const rvec *pb = s_b->s.cg_p.rvec_array();
+ const rvec *sfb = s_b->f.rvec_array();
+ gpb = 0;
+ for (int i = 0; i < mdatoms->homenr; i++)
+ {
+ for (m = 0; m < DIM; m++)
+ {
+ gpb -= pb[i][m]*sfb[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 (std::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 (mdrunOptions.verbose)
+ {
+ 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 */
+ matrix nullBox = {};
+ upd_mdebin(mdebin, FALSE, FALSE, static_cast<double>(step),
+ mdatoms->tmass, enerd, nullptr, nullptr, nullptr, nullBox,
+ nullptr, nullptr, vir, pres, nullptr, 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 : nullptr, step, step, eprNORMAL,
+ mdebin, fcd, &(top_global->groups), &(inputrec->opts), nullptr);
+ }
+
+ /* Send energies and positions to the IMD client if bIMD is TRUE. */
+ if (MASTER(cr) && do_IMD(inputrec->bIMD, step, cr, TRUE, state_global->box, state_global->x.rvec_array(), inputrec, 0, wcycle))
+ {
+ 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(fplog, inputrec->em_tol, s_min->fmax,
+ 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 : nullptr, step, step, eprNORMAL,
+ mdebin, fcd, &(top_global->groups), &(inputrec->opts), nullptr);
+ }
+ }
+
+ /* 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).
+ */
+ /* Note that with 0 < nstfout != nstxout we can end up with two frames
+ * in the trajectory with the same step number.
+ */
+ 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, observablesHistory);
+
+
+ if (MASTER(cr))
+ {
+ double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
+ print_converged(stderr, CG, inputrec->em_tol, step, converged, number_steps,
+ s_min, sqrtNumAtoms);
+ print_converged(fplog, CG, inputrec->em_tol, step, converged, number_steps,
+ s_min, 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);
+}
+
+
+void
+Integrator::do_lbfgs()
+{
+ static const char *LBFGS = "Low-Memory BFGS Minimizer";
+ em_state_t ems;
+ gmx_localtop_t *top;
+ gmx_enerdata_t *enerd;
+ 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, *p, *s, **dx, **dg;
+ real a, b, c, maxdelta, delta;
+ real diag, Epot0;
+ real dgdx, dgdg, sq, yr, beta;
+ t_mdebin *mdebin;
+ gmx_bool converged;
+ rvec mu_tot;
+ 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, gf, step;
+ int mdof_flags;
+ auto mdatoms = mdAtoms->mdatoms();
+
+ GMX_LOG(mdlog.info).asParagraph().
+ appendText("Note that activating L-BFGS energy minimization via the "
+ "integrator .mdp option and the command gmx mdrun may "
+ "be available in a different form in a future version of GROMACS, "
+ "e.g. gmx minimize and an .mdp option.");
+
+ if (PAR(cr))
+ {
+ gmx_fatal(FARGS, "L-BFGS minimization only supports a single rank");
+ }
+
+ if (nullptr != 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;
+
+ snew(frozen, n);
+
+ snew(p, 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, mdlog, LBFGS, cr, ms, outputProvider, inputrec, mdrunOptions,
+ state_global, top_global, &ems, &top,
+ nrnb, mu_tot, fr, &enerd, &graph, mdAtoms, &gstat,
+ vsite, constr, nullptr,
+ nfile, fnm, &outf, &mdebin, wcycle);
+
+ start = 0;
+ end = mdatoms->homenr;
+
+ /* We need 4 working states */
+ em_state_t s0 {}, s1 {}, s2 {}, s3 {};
+ em_state_t *sa = &s0;
+ em_state_t *sb = &s1;
+ em_state_t *sc = &s2;
+ em_state_t *last = &s3;
+ /* Initialize by copying the state from ems (we could skip x and f here) */
+ *sa = ems;
+ *sb = ems;
+ *sc = ems;
+
+ /* 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] != 0);
+ }
+ }
+ 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.rvec_array(), 1, nullptr,
+ 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++;
+ EnergyEvaluator energyEvaluator {
+ fplog, mdlog, cr, ms,
+ top_global, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph,
+ mdAtoms, fr, ppForceWorkload, enerd
+ };
+ energyEvaluator.run(&ems, mu_tot, vir, pres, -1, TRUE);
+
+ if (MASTER(cr))
+ {
+ /* Copy stuff to the energy bin for easy printing etc. */
+ matrix nullBox = {};
+ upd_mdebin(mdebin, FALSE, FALSE, static_cast<double>(step),
+ mdatoms->tmass, enerd, nullptr, nullptr, nullptr, nullBox,
+ nullptr, nullptr, vir, pres, nullptr, 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), nullptr);
+ }
+
+ /* 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", ems.fmax, ems.a_fmax + 1);
+ fprintf(stderr, " F-Norm = %12.5e\n", ems.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", ems.fmax, ems.a_fmax + 1);
+ fprintf(fplog, " F-Norm = %12.5e\n", ems.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.
+ real *fInit = static_cast<real *>(ems.f.rvec_array()[0]);
+ for (i = 0; i < n; i++)
+ {
+ if (!frozen[i])
+ {
+ dx[point][i] = fInit[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/ems.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, static_cast<real>(step), &ems.s, state_global, observablesHistory, ems.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];
+
+ real *xx = static_cast<real *>(ems.s.x.rvec_array()[0]);
+ real *ff = static_cast<real *>(ems.f.rvec_array()[0]);
+
+ // 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
+ *last = ems;
+ real *lastx = static_cast<real *>(last->s.x.data()[0]);
+ real *lastf = static_cast<real *>(last->f.data()[0]);
+ Epot0 = ems.epot;
+
+ *sa = ems;
+
+ /* 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
+ 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
+ real *xc = static_cast<real *>(sc->s.x.rvec_array()[0]);
+ for (i = 0; i < n; i++)
+ {
+ xc[i] = lastx[i] + c*s[i];
+ }
+
+ neval++;
+ // Calculate energy for the trial step in position C
+ energyEvaluator.run(sc, mu_tot, vir, pres, step, FALSE);
+
+ // Calc line gradient in position C
+ real *fc = static_cast<real *>(sc->f.rvec_array()[0]);
+ 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 = std::sqrt(GMX_REAL_EPS)*fabs(sa->epot);
+
+ // 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.
+ foundlower = sc->epot < sa->epot || (gpc < 0 && sc->epot < (sa->epot + tmp));
+ // If true, 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.
+
+ // If false, 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.
+
+ 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.
+ real fnorm = 0;
+ 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
+ real *xb = static_cast<real *>(sb->s.x.rvec_array()[0]);
+ for (i = 0; i < n; i++)
+ {
+ xb[i] = lastx[i] + b*s[i];
+ }
+
+ neval++;
+ // Calculate energy for the trial step in point B
+ energyEvaluator.run(sb, mu_tot, vir, pres, step, FALSE);
+ fnorm = sb->fnorm;
+
+ // Calculate gradient in point B
+ real *fb = static_cast<real *>(sb->f.rvec_array()[0]);
+ 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 */
+ c = b;
+ /* swap states b and c */
+ swap_em_state(&sb, &sc);
+ }
+ else
+ {
+ /* Replace a endpoint with b */
+ a = b;
+ /* swap states a and b */
+ swap_em_state(&sa, &sb);
+ }
+
+ /*
+ * 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 ((sb->epot > sa->epot || sb->epot > sc->epot) && (nminstep < 20));
+
+ if (std::fabs(sb->epot - 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 (sc->epot < sa->epot)
+ {
+ /* Use state C */
+ ems = *sc;
+ step_taken = c;
+ }
+ else
+ {
+ /* Use state A */
+ ems = *sa;
+ step_taken = a;
+ }
+
+ }
+ else
+ {
+ /* found lower */
+ /* Use state C */
+ ems = *sc;
+ 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;
+ }
+ }
+
+ /* Print it if necessary */
+ if (MASTER(cr))
+ {
+ if (mdrunOptions.verbose)
+ {
+ 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, ems.epot, ems.fnorm/sqrtNumAtoms, ems.fmax, ems.a_fmax + 1);
+ fflush(stderr);
+ }
+ /* Store the new (lower) energies */
+ matrix nullBox = {};
+ upd_mdebin(mdebin, FALSE, FALSE, static_cast<double>(step),
+ mdatoms->tmass, enerd, nullptr, nullptr, nullptr, nullBox,
+ nullptr, nullptr, vir, pres, nullptr, 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 : nullptr, step, step, eprNORMAL,
+ mdebin, fcd, &(top_global->groups), &(inputrec->opts), nullptr);
+ }
+
+ /* 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.rvec_array(), inputrec, 0, wcycle) && MASTER(cr))
+ {
+ IMD_send_positions(inputrec->imd);
+ }
+
+ // Reset stepsize in we are doing more iterations
+ stepsize = 1.0/ems.fnorm;
+
+ /* Stop when the maximum force lies below tolerance.
+ * If we have reached machine precision, converged is already set to true.
+ */
+ converged = converged || (ems.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 (ems.fmax > inputrec->em_tol)
+ {
+ if (MASTER(cr))
+ {
+ warn_step(fplog, inputrec->em_tol, ems.fmax,
+ 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 : nullptr, step, step, eprNORMAL,
+ mdebin, fcd, &(top_global->groups), &(inputrec->opts), nullptr);
+ }
+
+ /* 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, observablesHistory);
+
+ if (MASTER(cr))
+ {
+ double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
+ print_converged(stderr, LBFGS, inputrec->em_tol, step, converged,
+ number_steps, &ems, sqrtNumAtoms);
+ print_converged(fplog, LBFGS, inputrec->em_tol, step, converged,
+ number_steps, &ems, 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);
+}
+
+void
+Integrator::do_steep()
+{
+ const char *SD = "Steepest Descents";
+ gmx_localtop_t *top;
+ gmx_enerdata_t *enerd;
+ gmx_global_stat_t gstat;
+ t_graph *graph;
+ real stepsize;
+ real ustep;
+ 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;
+ auto mdatoms = mdAtoms->mdatoms();
+
+ GMX_LOG(mdlog.info).asParagraph().
+ appendText("Note that activating steepest-descent energy minimization via the "
+ "integrator .mdp option and the command gmx mdrun may "
+ "be available in a different form in a future version of GROMACS, "
+ "e.g. gmx minimize and an .mdp option.");
+
+ /* Create 2 states on the stack and extract pointers that we will swap */
+ em_state_t s0 {}, s1 {};
+ em_state_t *s_min = &s0;
+ em_state_t *s_try = &s1;
+
+ /* Init em and store the local state in s_try */
+ init_em(fplog, mdlog, SD, cr, ms, outputProvider, inputrec, mdrunOptions,
+ state_global, top_global, s_try, &top,
+ nrnb, mu_tot, fr, &enerd, &graph, mdAtoms, &gstat,
+ vsite, constr, nullptr,
+ nfile, fnm, &outf, &mdebin, 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);
+ }
+ EnergyEvaluator energyEvaluator {
+ fplog, mdlog, cr, ms,
+ top_global, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph,
+ mdAtoms, fr, ppForceWorkload, enerd
+ };
+
+ /**** 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 */
+ bool validStep = true;
+ if (count > 0)
+ {
+ validStep =
+ do_em_step(cr, inputrec, mdatoms,
+ s_min, stepsize, &s_min->f, s_try,
+ constr, count);
+ }
+
+ if (validStep)
+ {
+ energyEvaluator.run(s_try, mu_tot, vir, pres, count, count == 0);
+ }
+ else
+ {
+ // Signal constraint error during stepping with energy=inf
+ s_try->epot = std::numeric_limits<real>::infinity();
+ }
+
+ 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 (mdrunOptions.verbose)
+ {
+ 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 */
+ matrix nullBox = {};
+ upd_mdebin(mdebin, FALSE, FALSE, static_cast<double>(count),
+ mdatoms->tmass, enerd, nullptr, nullptr, nullptr,
+ nullBox, nullptr, nullptr, vir, pres, nullptr, 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), nullptr);
+ 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, nullptr,
+ top_global, inputrec, count,
+ s_min, state_global, observablesHistory);
+ }
+ 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, mdlog, 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(fplog, inputrec->em_tol, s_min->fmax,
+ count == nsteps, constr != nullptr);
+ }
+ bAbort = TRUE;
+ }
+
+ /* Send IMD energies and positions, if bIMD is TRUE. */
+ if (do_IMD(inputrec->bIMD, count, cr, TRUE, state_global->box,
+ MASTER(cr) ? state_global->x.rvec_array() : nullptr,
+ 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 != 0, ftp2fn(efSTO, nfile, fnm),
+ top_global, inputrec, count,
+ s_min, state_global, observablesHistory);
+
+ if (MASTER(cr))
+ {
+ double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
+
+ print_converged(stderr, SD, inputrec->em_tol, count, bDone, nsteps,
+ s_min, sqrtNumAtoms);
+ print_converged(fplog, SD, inputrec->em_tol, count, bDone, nsteps,
+ s_min, sqrtNumAtoms);
+ }
+
+ 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);
+}
+
+void
+Integrator::do_nm()
+{
+ const char *NM = "Normal Mode Analysis";
+ gmx_mdoutf_t outf;
+ int nnodes, node;
+ gmx_localtop_t *top;
+ gmx_enerdata_t *enerd;
+ gmx_global_stat_t gstat;
+ t_graph *graph;
+ tensor vir, pres;
+ rvec mu_tot;
+ rvec *dfdx;
+ gmx_bool bSparse; /* use sparse matrix storage format */
+ size_t sz;
+ gmx_sparsematrix_t * sparse_matrix = nullptr;
+ real * full_matrix = nullptr;
+
+ /* added with respect to mdrun */
+ int row, col;
+ real der_range = 10.0*std::sqrt(GMX_REAL_EPS);
+ real x_min;
+ bool bIsMaster = MASTER(cr);
+ auto mdatoms = mdAtoms->mdatoms();
+
+ GMX_LOG(mdlog.info).asParagraph().
+ appendText("Note that activating normal-mode analysis via the integrator "
+ ".mdp option and the command gmx mdrun may "
+ "be available in a different form in a future version of GROMACS, "
+ "e.g. gmx normal-modes.");
+
+ if (constr != nullptr)
+ {
+ gmx_fatal(FARGS, "Constraints present with Normal Mode Analysis, this combination is not supported");
+ }
+
+ gmx_shellfc_t *shellfc;
+
+ em_state_t state_work {};
+
+ /* Init em and store the local state in state_minimum */
+ init_em(fplog, mdlog, NM, cr, ms, outputProvider, inputrec, mdrunOptions,
+ state_global, top_global, &state_work, &top,
+ nrnb, mu_tot, fr, &enerd, &graph, mdAtoms, &gstat,
+ vsite, constr, &shellfc,
+ nfile, fnm, &outf, nullptr, wcycle);
+
+ std::vector<int> atom_index = get_atom_index(top_global);
+ std::vector<gmx::RVec> fneg(atom_index.size(), {0, 0, 0});
+ 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->ic->eeltype) || fr->rlist == 0.0)
+ {
+ GMX_LOG(mdlog.warning).appendText("Non-cutoff electrostatics used, forcing full Hessian format.");
+ bSparse = FALSE;
+ }
+ else if (atom_index.size() < 1000)
+ {
+ GMX_LOG(mdlog.warning).appendTextFormatted("Small system size (N=%zu), using full Hessian format.",
+ atom_index.size());
+ bSparse = FALSE;
+ }
+ else
+ {
+ GMX_LOG(mdlog.warning).appendText("Using compressed symmetric sparse Hessian format.");
+ 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);
+
+
+ /* 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%" PRId64 " steps.\n\n",
+ *(top_global->name), inputrec->nsteps);
+ }
+
+ nnodes = cr->nnodes;
+
+ /* Make evaluate_energy do a single node force calculation */
+ cr->nnodes = 1;
+ EnergyEvaluator energyEvaluator {
+ fplog, mdlog, cr, ms,
+ top_global, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph,
+ mdAtoms, fr, ppForceWorkload, enerd
+ };
+ energyEvaluator.run(&state_work, mu_tot, 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);
+
+ GMX_LOG(mdlog.warning).appendTextFormatted("Maximum force:%12.5e", state_work.fmax);
+ if (state_work.fmax > 1.0e-3)
+ {
+ GMX_LOG(mdlog.warning).appendText(
+ "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.");
+ }
+
+ /***********************************************************
+ *
+ * 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;
+ auto state_work_x = makeArrayRef(state_work.s.x);
+ auto state_work_f = makeArrayRef(state_work.f);
+ 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++)
+ {
+ int64_t step = 0;
+ int force_flags = GMX_FORCE_STATECHANGED | GMX_FORCE_ALLFORCES;
+ double t = 0;
+
+ x_min = state_work_x[atom][d];
+
+ for (unsigned int dx = 0; (dx < 2); dx++)
+ {
+ if (dx == 0)
+ {
+ state_work_x[atom][d] = x_min - der_range;
+ }
+ else
+ {
+ state_work_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 */
+ relax_shell_flexcon(fplog,
+ cr,
+ ms,
+ mdrunOptions.verbose,
+ nullptr,
+ step,
+ inputrec,
+ bNS,
+ force_flags,
+ top,
+ constr,
+ enerd,
+ fcd,
+ &state_work.s,
+ state_work.f.arrayRefWithPadding(),
+ vir,
+ mdatoms,
+ nrnb,
+ wcycle,
+ graph,
+ &top_global->groups,
+ shellfc,
+ fr,
+ ppForceWorkload,
+ t,
+ mu_tot,
+ vsite,
+ DdOpenBalanceRegionBeforeForceComputation::no,
+ DdCloseBalanceRegionAfterForceComputation::no);
+ bNS = false;
+ step++;
+ }
+ else
+ {
+ energyEvaluator.run(&state_work, mu_tot, vir, pres, aid*2+dx, FALSE);
+ }
+
+ cr->nnodes = nnodes;
+
+ if (dx == 0)
+ {
+ std::copy(state_work_f.begin(), state_work_f.begin()+atom_index.size(), fneg.begin());
+ }
+ }
+
+ /* x is restored to original */
+ state_work_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 && aid+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 = (aid + 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 (mdrunOptions.verbose && fplog)
+ {
+ fflush(fplog);
+ }
+ }
+ /* write progress */
+ if (bIsMaster && mdrunOptions.verbose)
+ {
+ 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);
+}
+
+} // namespace gmx
diff --git a/patches/gromacs-2019.1.diff/src/gromacs/mdrun/minimize.cpp.preplumed b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/minimize.cpp.preplumed
new file mode 100644
index 0000000000000000000000000000000000000000..eb8e910468fd99643d8d9f5aa061caa7a10c13ab
--- /dev/null
+++ b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/minimize.cpp.preplumed
@@ -0,0 +1,2988 @@
+/*
+ * 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) 2013,2014,2015,2016,2017,2018, 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_mdrun
+ */
+#include "gmxpre.h"
+
+#include "config.h"
+
+#include <cmath>
+#include <cstring>
+#include <ctime>
+
+#include <algorithm>
+#include <vector>
+
+#include "gromacs/commandline/filenm.h"
+#include "gromacs/domdec/collect.h"
+#include "gromacs/domdec/domdec.h"
+#include "gromacs/domdec/domdec_struct.h"
+#include "gromacs/domdec/partition.h"
+#include "gromacs/ewald/pme.h"
+#include "gromacs/fileio/confio.h"
+#include "gromacs/fileio/mtxio.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/mdsetup.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/mdtypes/state.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/topology/topology.h"
+#include "gromacs/utility/cstringutil.h"
+#include "gromacs/utility/exceptions.h"
+#include "gromacs/utility/fatalerror.h"
+#include "gromacs/utility/logger.h"
+#include "gromacs/utility/smalloc.h"
+
+#include "integrator.h"
+
+//! Utility structure for manipulating states during EM
+typedef struct {
+ //! Copy of the global state
+ t_state s;
+ //! Force array
+ PaddedVector<gmx::RVec> f;
+ //! Potential energy
+ real epot;
+ //! Norm of the force
+ real fnorm;
+ //! Maximum force
+ real fmax;
+ //! Direction
+ int a_fmax;
+} em_state_t;
+
+//! Print the EM starting conditions
+static void print_em_start(FILE *fplog,
+ const t_commrec *cr,
+ gmx_walltime_accounting_t walltime_accounting,
+ gmx_wallcycle_t wcycle,
+ const char *name)
+{
+ walltime_accounting_start_time(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_time(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,
+ real fmax,
+ gmx_bool bLastStep,
+ gmx_bool bConstrain)
+{
+ constexpr bool realIsDouble = GMX_DOUBLE;
+ char buffer[2048];
+
+ if (!std::isfinite(fmax))
+ {
+ sprintf(buffer,
+ "\nEnergy minimization has stopped because the force "
+ "on at least one atom is not finite. This usually means "
+ "atoms are overlapping. Modify the input coordinates to "
+ "remove atom overlap or use soft-core potentials with "
+ "the free energy code to avoid infinite forces.\n%s",
+ !realIsDouble ?
+ "You could also be lucky that switching to double precision "
+ "is sufficient to obtain finite forces.\n" :
+ "");
+ }
+ else 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,
+ !realIsDouble ?
+ "\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), stderr);
+ 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,
+ int64_t count, gmx_bool bDone, int64_t nsteps,
+ const em_state_t *ems, double sqrtNumAtoms)
+{
+ 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", ems->epot);
+ fprintf(fp, "Maximum force = %21.14e on atom %d\n", ems->fmax, ems->a_fmax + 1);
+ fprintf(fp, "Norm of force = %21.14e\n", ems->fnorm/sqrtNumAtoms);
+#else
+ fprintf(fp, "Potential Energy = %14.7e\n", ems->epot);
+ fprintf(fp, "Maximum force = %14.7e on atom %d\n", ems->fmax, ems->a_fmax + 1);
+ fprintf(fp, "Norm of force = %14.7e\n", ems->fnorm/sqrtNumAtoms);
+#endif
+}
+
+//! Compute the norm and max of the force array in parallel
+static void get_f_norm_max(const t_commrec *cr,
+ t_grpopts *opts, t_mdatoms *mdatoms, const 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->globalAtomIndices[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 = gmx::roundToInt(sum[2*i+1]);
+ }
+ }
+ 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(const t_commrec *cr,
+ t_grpopts *opts, t_mdatoms *mdatoms,
+ em_state_t *ems)
+{
+ get_f_norm_max(cr, opts, mdatoms, ems->f.rvec_array(),
+ &ems->fnorm, &ems->fmax, &ems->a_fmax);
+}
+
+//! Initialize the energy minimization
+static void init_em(FILE *fplog,
+ const gmx::MDLogger &mdlog,
+ const char *title,
+ const t_commrec *cr,
+ const gmx_multisim_t *ms,
+ gmx::IMDOutputProvider *outputProvider,
+ t_inputrec *ir,
+ const MdrunOptions &mdrunOptions,
+ t_state *state_global, gmx_mtop_t *top_global,
+ em_state_t *ems, gmx_localtop_t **top,
+ t_nrnb *nrnb, rvec mu_tot,
+ t_forcerec *fr, gmx_enerdata_t **enerd,
+ t_graph **graph, gmx::MDAtoms *mdAtoms, gmx_global_stat_t *gstat,
+ gmx_vsite_t *vsite, gmx::Constraints *constr, gmx_shellfc_t **shellfc,
+ int nfile, const t_filenm fnm[],
+ gmx_mdoutf_t *outf, t_mdebin **mdebin,
+ gmx_wallcycle_t wcycle)
+{
+ real dvdl_constr;
+
+ if (fplog)
+ {
+ fprintf(fplog, "Initiating %s\n", title);
+ }
+
+ if (MASTER(cr))
+ {
+ state_global->ngtc = 0;
+
+ /* Initialize lambda variables */
+ initialize_lambdas(fplog, ir, &(state_global->fep_state), state_global->lambda, nullptr);
+ }
+
+ init_nrnb(nrnb);
+
+ /* Interactive molecular dynamics */
+ init_IMD(ir, cr, ms, top_global, fplog, 1,
+ MASTER(cr) ? state_global->x.rvec_array() : nullptr,
+ nfile, fnm, nullptr, mdrunOptions);
+
+ if (ir->eI == eiNM)
+ {
+ GMX_ASSERT(shellfc != nullptr, "With NM we always support shells");
+
+ *shellfc = init_shell_flexcon(stdout,
+ top_global,
+ constr ? constr->numFlexibleConstraints() : 0,
+ ir->nstcalcenergy,
+ DOMAINDECOMP(cr));
+ }
+ else
+ {
+ GMX_ASSERT(EI_ENERGY_MINIMIZATION(ir->eI), "This else currently only handles energy minimizers, consider if your algorithm needs shell/flexible-constraint support");
+
+ /* With energy minimization, shells and flexible constraints are
+ * automatically minimized when treated like normal DOFS.
+ */
+ if (shellfc != nullptr)
+ {
+ *shellfc = nullptr;
+ }
+ }
+
+ auto mdatoms = mdAtoms->mdatoms();
+ if (DOMAINDECOMP(cr))
+ {
+ *top = dd_init_local_top(top_global);
+
+ dd_init_local_state(cr->dd, state_global, &ems->s);
+
+ /* Distribute the charge groups over the nodes from the master node */
+ dd_partition_system(fplog, mdlog, ir->init_step, cr, TRUE, 1,
+ state_global, top_global, ir,
+ &ems->s, &ems->f, mdAtoms, *top,
+ fr, vsite, constr,
+ nrnb, nullptr, FALSE);
+ dd_store_state(cr->dd, &ems->s);
+
+ *graph = nullptr;
+ }
+ else
+ {
+ state_change_natoms(state_global, state_global->natoms);
+ /* Just copy the state */
+ ems->s = *state_global;
+ state_change_natoms(&ems->s, ems->s.natoms);
+ ems->f.resizeWithPadding(ems->s.natoms);
+
+ snew(*top, 1);
+ mdAlgorithmsSetupAtomData(cr, ir, top_global, *top, fr,
+ graph, mdAtoms,
+ constr, vsite, shellfc ? *shellfc : nullptr);
+
+ if (vsite)
+ {
+ set_vsite_top(vsite, *top, mdatoms);
+ }
+ }
+
+ update_mdatoms(mdAtoms->mdatoms(), ems->s.lambda[efptMASS]);
+
+ if (constr)
+ {
+ // TODO how should this cross-module support dependency be managed?
+ 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 (!ir->bContinuation)
+ {
+ /* Constrain the starting coordinates */
+ dvdl_constr = 0;
+ constr->apply(TRUE, TRUE,
+ -1, 0, 1.0,
+ ems->s.x.rvec_array(),
+ ems->s.x.rvec_array(),
+ nullptr,
+ ems->s.box,
+ ems->s.lambda[efptFEP], &dvdl_constr,
+ nullptr, nullptr, gmx::ConstraintVariable::Positions);
+ }
+ }
+
+ if (PAR(cr))
+ {
+ *gstat = global_stat_init(ir);
+ }
+ else
+ {
+ *gstat = nullptr;
+ }
+
+ *outf = init_mdoutf(fplog, nfile, fnm, mdrunOptions, cr, outputProvider, ir, top_global, nullptr, wcycle);
+
+ snew(*enerd, 1);
+ init_enerdata(top_global->groups.grps[egcENER].nr, ir->fepvals->n_lambda,
+ *enerd);
+
+ if (mdebin != nullptr)
+ {
+ /* Init bin for energy stuff */
+ *mdebin = init_mdebin(mdoutf_get_fp_ene(*outf), top_global, ir, nullptr);
+ }
+
+ clear_rvec(mu_tot);
+ calc_shifts(ems->s.box, fr->shift_vec);
+}
+
+//! Finalize the minimization
+static void finish_em(const t_commrec *cr, gmx_mdoutf_t outf,
+ gmx_walltime_accounting_t walltime_accounting,
+ gmx_wallcycle_t wcycle)
+{
+ if (!thisRankHasDuty(cr, 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;
+}
+
+//! Save the EM trajectory
+static void write_em_traj(FILE *fplog, const t_commrec *cr,
+ gmx_mdoutf_t outf,
+ gmx_bool bX, gmx_bool bF, const char *confout,
+ gmx_mtop_t *top_global,
+ t_inputrec *ir, int64_t step,
+ em_state_t *state,
+ t_state *state_global,
+ ObservablesHistory *observablesHistory)
+{
+ int 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, static_cast<double>(step),
+ &state->s, state_global, observablesHistory,
+ state->f);
+
+ if (confout != nullptr)
+ {
+ if (DOMAINDECOMP(cr))
+ {
+ /* If bX=true, x was collected to state_global in the call above */
+ if (!bX)
+ {
+ gmx::ArrayRef<gmx::RVec> globalXRef = MASTER(cr) ? makeArrayRef(state_global->x) : gmx::EmptyArrayRef();
+ dd_collect_vec(cr->dd, &state->s, makeArrayRef(state->s.x), globalXRef);
+ }
+ }
+ else
+ {
+ /* Copy the local state pointer */
+ state_global = &state->s;
+ }
+
+ if (MASTER(cr))
+ {
+ if (ir->ePBC != epbcNONE && !ir->bPeriodicMols && DOMAINDECOMP(cr))
+ {
+ /* Make molecules whole only for confout writing */
+ do_pbc_mtop(fplog, ir->ePBC, state->s.box, top_global,
+ state_global->x.rvec_array());
+ }
+
+ write_sto_conf_mtop(confout,
+ *top_global->name, top_global,
+ state_global->x.rvec_array(), nullptr, ir->ePBC, state->s.box);
+ }
+ }
+}
+
+//! \brief Do one minimization step
+//
+// \returns true when the step succeeded, false when a constraint error occurred
+static bool do_em_step(const t_commrec *cr,
+ t_inputrec *ir, t_mdatoms *md,
+ em_state_t *ems1, real a, const PaddedVector<gmx::RVec> *force,
+ em_state_t *ems2,
+ gmx::Constraints *constr,
+ int64_t count)
+
+{
+ t_state *s1, *s2;
+ int start, end;
+ real dvdl_constr;
+ int nthreads gmx_unused;
+
+ bool validStep = true;
+
+ 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->natoms != s1->natoms)
+ {
+ state_change_natoms(s2, s1->natoms);
+ ems2->f.resizeWithPadding(s2->natoms);
+ }
+ if (DOMAINDECOMP(cr) && s2->cg_gl.size() != s1->cg_gl.size())
+ {
+ s2->cg_gl.resize(s1->cg_gl.size());
+ }
+
+ copy_mat(s1->box, s2->box);
+ /* Copy free energy state */
+ s2->lambda = s1->lambda;
+ copy_mat(s1->box, s2->box);
+
+ start = 0;
+ end = md->homenr;
+
+ nthreads = gmx_omp_nthreads_get(emntUpdate);
+#pragma omp parallel num_threads(nthreads)
+ {
+ const rvec *x1 = s1->x.rvec_array();
+ rvec *x2 = s2->x.rvec_array();
+ const rvec *f = force->rvec_array();
+
+ int gf = 0;
+#pragma omp for schedule(static) nowait
+ for (int i = start; i < end; i++)
+ {
+ try
+ {
+ if (md->cFREEZE)
+ {
+ gf = md->cFREEZE[i];
+ }
+ for (int 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 */
+ const rvec *p1 = s1->cg_p.rvec_array();
+ rvec *p2 = s2->cg_p.rvec_array();
+#pragma omp for schedule(static) nowait
+ for (int i = start; i < end; i++)
+ {
+ // Trivial OpenMP block that does not throw
+ copy_rvec(p1[i], p2[i]);
+ }
+ }
+
+ if (DOMAINDECOMP(cr))
+ {
+ s2->ddp_count = s1->ddp_count;
+
+ /* OpenMP does not supported unsigned loop variables */
+#pragma omp for schedule(static) nowait
+ for (int i = 0; i < static_cast<int>(s2->cg_gl.size()); i++)
+ {
+ s2->cg_gl[i] = s1->cg_gl[i];
+ }
+ s2->ddp_count_cg_gl = s1->ddp_count_cg_gl;
+ }
+ }
+
+ if (constr)
+ {
+ dvdl_constr = 0;
+ validStep =
+ constr->apply(TRUE, TRUE,
+ count, 0, 1.0,
+ s1->x.rvec_array(), s2->x.rvec_array(),
+ nullptr, s2->box,
+ s2->lambda[efptBONDED], &dvdl_constr,
+ nullptr, nullptr, gmx::ConstraintVariable::Positions);
+
+ if (cr->nnodes > 1)
+ {
+ /* This global reduction will affect performance at high
+ * parallelization, but we can not really avoid it.
+ * But usually EM is not run at high parallelization.
+ */
+ int reductionBuffer = static_cast<int>(!validStep);
+ gmx_sumi(1, &reductionBuffer, cr);
+ validStep = (reductionBuffer == 0);
+ }
+
+ // We should move this check to the different minimizers
+ if (!validStep && ir->eI != eiSteep)
+ {
+ gmx_fatal(FARGS, "The coordinates could not be constrained. Minimizer '%s' can not handle constraint failures, use minimizer '%s' before using '%s'.",
+ EI(ir->eI), EI(eiSteep), EI(ir->eI));
+ }
+ }
+
+ return validStep;
+}
+
+//! Prepare EM for using domain decomposition parallellization
+static void em_dd_partition_system(FILE *fplog,
+ const gmx::MDLogger &mdlog,
+ int step, const t_commrec *cr,
+ gmx_mtop_t *top_global, t_inputrec *ir,
+ em_state_t *ems, gmx_localtop_t *top,
+ gmx::MDAtoms *mdAtoms, t_forcerec *fr,
+ gmx_vsite_t *vsite, gmx::Constraints *constr,
+ t_nrnb *nrnb, gmx_wallcycle_t wcycle)
+{
+ /* Repartition the domain decomposition */
+ dd_partition_system(fplog, mdlog, step, cr, FALSE, 1,
+ nullptr, top_global, ir,
+ &ems->s, &ems->f,
+ mdAtoms, top, fr, vsite, constr,
+ nrnb, wcycle, FALSE);
+ dd_store_state(cr->dd, &ems->s);
+}
+
+namespace
+{
+
+/*! \brief Class to handle the work of setting and doing an energy evaluation.
+ *
+ * This class is a mere aggregate of parameters to pass to evaluate an
+ * energy, so that future changes to names and types of them consume
+ * less time when refactoring other code.
+ *
+ * Aggregate initialization is used, for which the chief risk is that
+ * if a member is added at the end and not all initializer lists are
+ * updated, then the member will be value initialized, which will
+ * typically mean initialization to zero.
+ *
+ * We only want to construct one of these with an initializer list, so
+ * we explicitly delete the default constructor. */
+class EnergyEvaluator
+{
+ public:
+ //! We only intend to construct such objects with an initializer list.
+#if __GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 9)
+ // Aspects of the C++11 spec changed after GCC 4.8.5, and
+ // compilation of the initializer list construction in
+ // runner.cpp fails in GCC 4.8.5.
+ EnergyEvaluator() = delete;
+#endif
+ /*! \brief Evaluates an energy on the state in \c ems.
+ *
+ * \todo In practice, the same objects mu_tot, vir, and pres
+ * are always passed to this function, so we would rather have
+ * them as data members. However, their C-array types are
+ * unsuited for aggregate initialization. When the types
+ * improve, the call signature of this method can be reduced.
+ */
+ void run(em_state_t *ems, rvec mu_tot,
+ tensor vir, tensor pres,
+ int64_t count, gmx_bool bFirst);
+ //! Handles logging (deprecated).
+ FILE *fplog;
+ //! Handles logging.
+ const gmx::MDLogger &mdlog;
+ //! Handles communication.
+ const t_commrec *cr;
+ //! Coordinates multi-simulations.
+ const gmx_multisim_t *ms;
+ //! Holds the simulation topology.
+ gmx_mtop_t *top_global;
+ //! Holds the domain topology.
+ gmx_localtop_t *top;
+ //! User input options.
+ t_inputrec *inputrec;
+ //! Manages flop accounting.
+ t_nrnb *nrnb;
+ //! Manages wall cycle accounting.
+ gmx_wallcycle_t wcycle;
+ //! Coordinates global reduction.
+ gmx_global_stat_t gstat;
+ //! Handles virtual sites.
+ gmx_vsite_t *vsite;
+ //! Handles constraints.
+ gmx::Constraints *constr;
+ //! Handles strange things.
+ t_fcdata *fcd;
+ //! Molecular graph for SHAKE.
+ t_graph *graph;
+ //! Per-atom data for this domain.
+ gmx::MDAtoms *mdAtoms;
+ //! Handles how to calculate the forces.
+ t_forcerec *fr;
+ //! Schedule of force-calculation work each step for this task.
+ gmx::PpForceWorkload *ppForceWorkload;
+ //! Stores the computed energies.
+ gmx_enerdata_t *enerd;
+};
+
+void
+EnergyEvaluator::run(em_state_t *ems, rvec mu_tot,
+ tensor vir, tensor pres,
+ 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.rvec_array(), 1, nullptr,
+ 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, mdlog, 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, ms, inputrec, nullptr, nullptr,
+ count, nrnb, wcycle, top, &top_global->groups,
+ ems->s.box, ems->s.x.arrayRefWithPadding(), &ems->s.hist,
+ ems->f.arrayRefWithPadding(), force_vir, mdAtoms->mdatoms(), enerd, fcd,
+ ems->s.lambda, graph, fr, ppForceWorkload, vsite, mu_tot, t, nullptr,
+ GMX_FORCE_STATECHANGED | GMX_FORCE_ALLFORCES |
+ GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY |
+ (bNS ? GMX_FORCE_NS : 0),
+ DOMAINDECOMP(cr) ?
+ DdOpenBalanceRegionBeforeForceComputation::yes :
+ DdOpenBalanceRegionBeforeForceComputation::no,
+ DOMAINDECOMP(cr) ?
+ DdCloseBalanceRegionAfterForceComputation::yes :
+ DdCloseBalanceRegionAfterForceComputation::no);
+
+ /* 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, nullptr, nullptr, nullptr, 1, &terminate,
+ nullptr, 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 */
+ dvdl_constr = 0;
+ rvec *f_rvec = ems->f.rvec_array();
+ constr->apply(FALSE, FALSE,
+ count, 0, 1.0,
+ ems->s.x.rvec_array(), f_rvec, f_rvec,
+ ems->s.box,
+ ems->s.lambda[efptBONDED], &dvdl_constr,
+ nullptr, &shake_vir, gmx::ConstraintVariable::ForceDispl);
+ enerd->term[F_DVDL_CONSTR] += dvdl_constr;
+ m_add(force_vir, shake_vir, vir);
+ }
+ 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->mdatoms(), ems);
+ }
+}
+
+} // namespace
+
+//! Parallel utility summing energies and forces
+static double reorder_partsum(const t_commrec *cr, t_grpopts *opts, t_mdatoms *mdatoms,
+ gmx_mtop_t *top_global,
+ em_state_t *s_min, em_state_t *s_b)
+{
+ 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");
+ }
+
+ const rvec *fm = s_min->f.rvec_array();
+ const rvec *fb = s_b->f.rvec_array();
+
+ 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.
+ */
+ rvec *fmg;
+ snew(fmg, top_global->natoms);
+
+ ncg = s_min->s.cg_gl.size();
+ cg_gl = s_min->s.cg_gl.data();
+ 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.cg_gl.size();
+ cg_gl = s_b->s.cg_gl.data();
+ 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(const t_commrec *cr, t_grpopts *opts, t_mdatoms *mdatoms,
+ gmx_mtop_t *top_global,
+ em_state_t *s_min, em_state_t *s_b)
+{
+ double sum;
+
+ /* 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))
+ {
+ const rvec *fm = s_min->f.rvec_array();
+ const rvec *fb = s_b->f.rvec_array();
+ sum = 0;
+ int gf = 0;
+ /* This part of code can be incorrect with DD,
+ * since the atom ordering in s_b and s_min might differ.
+ */
+ for (int i = 0; i < mdatoms->homenr; i++)
+ {
+ if (mdatoms->cFREEZE)
+ {
+ gf = mdatoms->cFREEZE[i];
+ }
+ for (int 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
+{
+
+void
+Integrator::do_cg()
+{
+ const char *CG = "Polak-Ribiere Conjugate Gradients";
+
+ gmx_localtop_t *top;
+ gmx_enerdata_t *enerd;
+ gmx_global_stat_t gstat;
+ t_graph *graph;
+ double tmp, minstep;
+ 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 m, step, nminstep;
+ auto mdatoms = mdAtoms->mdatoms();
+
+ GMX_LOG(mdlog.info).asParagraph().
+ appendText("Note that activating conjugate gradient energy minimization via the "
+ "integrator .mdp option and the command gmx mdrun may "
+ "be available in a different form in a future version of GROMACS, "
+ "e.g. gmx minimize and an .mdp option.");
+
+ step = 0;
+
+ if (MASTER(cr))
+ {
+ // In CG, the state is extended with a search direction
+ state_global->flags |= (1<<estCGP);
+
+ // Ensure the extra per-atom state array gets allocated
+ state_change_natoms(state_global, state_global->natoms);
+
+ // Initialize the search direction to zero
+ for (RVec &cg_p : state_global->cg_p)
+ {
+ cg_p = { 0, 0, 0 };
+ }
+ }
+
+ /* Create 4 states on the stack and extract pointers that we will swap */
+ em_state_t s0 {}, s1 {}, s2 {}, s3 {};
+ em_state_t *s_min = &s0;
+ em_state_t *s_a = &s1;
+ em_state_t *s_b = &s2;
+ em_state_t *s_c = &s3;
+
+ /* Init em and store the local state in s_min */
+ init_em(fplog, mdlog, CG, cr, ms, outputProvider, inputrec, mdrunOptions,
+ state_global, top_global, s_min, &top,
+ nrnb, mu_tot, fr, &enerd, &graph, mdAtoms, &gstat,
+ vsite, constr, nullptr,
+ nfile, fnm, &outf, &mdebin, 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);
+ }
+
+ EnergyEvaluator energyEvaluator {
+ fplog, mdlog, cr, ms,
+ top_global, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph,
+ mdAtoms, fr, ppForceWorkload, enerd
+ };
+ /* 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
+ */
+ energyEvaluator.run(s_min, mu_tot, vir, pres, -1, TRUE);
+
+ if (MASTER(cr))
+ {
+ /* Copy stuff to the energy bin for easy printing etc. */
+ matrix nullBox = {};
+ upd_mdebin(mdebin, FALSE, FALSE, static_cast<double>(step),
+ mdatoms->tmass, enerd, nullptr, nullptr, nullptr, nullBox,
+ nullptr, nullptr, vir, pres, nullptr, 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), nullptr);
+ }
+
+ /* 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 */
+ rvec *pm = s_min->s.cg_p.rvec_array();
+ const rvec *sfm = s_min->f.rvec_array();
+ double gpa = 0;
+ int gf = 0;
+ for (int 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])
+ {
+ pm[i][m] = sfm[i][m] + beta*pm[i][m];
+ gpa -= pm[i][m]*sfm[i][m];
+ /* f is negative gradient, thus the sign */
+ }
+ else
+ {
+ pm[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, pm, &pnorm, nullptr, nullptr);
+
+ /* 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;
+ auto s_min_x = makeArrayRef(s_min->s.x);
+ for (int i = 0; i < mdatoms->homenr; i++)
+ {
+ for (m = 0; m < DIM; m++)
+ {
+ tmp = fabs(s_min_x[i][m]);
+ if (tmp < 1.0)
+ {
+ tmp = 1.0;
+ }
+ tmp = pm[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*top_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, nullptr,
+ top_global, inputrec, step,
+ s_min, state_global, observablesHistory);
+
+ /* 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, mdlog, 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, s_min, c, &s_min->s.cg_p, s_c,
+ constr, -1);
+
+ neval++;
+ /* Calculate energy for the trial step */
+ energyEvaluator.run(s_c, mu_tot, vir, pres, -1, FALSE);
+
+ /* Calc derivative along line */
+ const rvec *pc = s_c->s.cg_p.rvec_array();
+ const rvec *sfc = s_c->f.rvec_array();
+ double gpc = 0;
+ for (int i = 0; i < mdatoms->homenr; i++)
+ {
+ for (m = 0; m < DIM; m++)
+ {
+ gpc -= pc[i][m]*sfc[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 = std::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.
+ */
+ double gpb;
+ 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, mdlog, -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, s_min, b, &s_min->s.cg_p, s_b,
+ constr, -1);
+
+ neval++;
+ /* Calculate energy for the trial step */
+ energyEvaluator.run(s_b, mu_tot, 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.
+ */
+ const rvec *pb = s_b->s.cg_p.rvec_array();
+ const rvec *sfb = s_b->f.rvec_array();
+ gpb = 0;
+ for (int i = 0; i < mdatoms->homenr; i++)
+ {
+ for (m = 0; m < DIM; m++)
+ {
+ gpb -= pb[i][m]*sfb[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 (std::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 (mdrunOptions.verbose)
+ {
+ 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 */
+ matrix nullBox = {};
+ upd_mdebin(mdebin, FALSE, FALSE, static_cast<double>(step),
+ mdatoms->tmass, enerd, nullptr, nullptr, nullptr, nullBox,
+ nullptr, nullptr, vir, pres, nullptr, 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 : nullptr, step, step, eprNORMAL,
+ mdebin, fcd, &(top_global->groups), &(inputrec->opts), nullptr);
+ }
+
+ /* Send energies and positions to the IMD client if bIMD is TRUE. */
+ if (MASTER(cr) && do_IMD(inputrec->bIMD, step, cr, TRUE, state_global->box, state_global->x.rvec_array(), inputrec, 0, wcycle))
+ {
+ 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(fplog, inputrec->em_tol, s_min->fmax,
+ 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 : nullptr, step, step, eprNORMAL,
+ mdebin, fcd, &(top_global->groups), &(inputrec->opts), nullptr);
+ }
+ }
+
+ /* 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).
+ */
+ /* Note that with 0 < nstfout != nstxout we can end up with two frames
+ * in the trajectory with the same step number.
+ */
+ 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, observablesHistory);
+
+
+ if (MASTER(cr))
+ {
+ double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
+ print_converged(stderr, CG, inputrec->em_tol, step, converged, number_steps,
+ s_min, sqrtNumAtoms);
+ print_converged(fplog, CG, inputrec->em_tol, step, converged, number_steps,
+ s_min, 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);
+}
+
+
+void
+Integrator::do_lbfgs()
+{
+ static const char *LBFGS = "Low-Memory BFGS Minimizer";
+ em_state_t ems;
+ gmx_localtop_t *top;
+ gmx_enerdata_t *enerd;
+ 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, *p, *s, **dx, **dg;
+ real a, b, c, maxdelta, delta;
+ real diag, Epot0;
+ real dgdx, dgdg, sq, yr, beta;
+ t_mdebin *mdebin;
+ gmx_bool converged;
+ rvec mu_tot;
+ 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, gf, step;
+ int mdof_flags;
+ auto mdatoms = mdAtoms->mdatoms();
+
+ GMX_LOG(mdlog.info).asParagraph().
+ appendText("Note that activating L-BFGS energy minimization via the "
+ "integrator .mdp option and the command gmx mdrun may "
+ "be available in a different form in a future version of GROMACS, "
+ "e.g. gmx minimize and an .mdp option.");
+
+ if (PAR(cr))
+ {
+ gmx_fatal(FARGS, "L-BFGS minimization only supports a single rank");
+ }
+
+ if (nullptr != 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;
+
+ snew(frozen, n);
+
+ snew(p, 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, mdlog, LBFGS, cr, ms, outputProvider, inputrec, mdrunOptions,
+ state_global, top_global, &ems, &top,
+ nrnb, mu_tot, fr, &enerd, &graph, mdAtoms, &gstat,
+ vsite, constr, nullptr,
+ nfile, fnm, &outf, &mdebin, wcycle);
+
+ start = 0;
+ end = mdatoms->homenr;
+
+ /* We need 4 working states */
+ em_state_t s0 {}, s1 {}, s2 {}, s3 {};
+ em_state_t *sa = &s0;
+ em_state_t *sb = &s1;
+ em_state_t *sc = &s2;
+ em_state_t *last = &s3;
+ /* Initialize by copying the state from ems (we could skip x and f here) */
+ *sa = ems;
+ *sb = ems;
+ *sc = ems;
+
+ /* 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] != 0);
+ }
+ }
+ 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.rvec_array(), 1, nullptr,
+ 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++;
+ EnergyEvaluator energyEvaluator {
+ fplog, mdlog, cr, ms,
+ top_global, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph,
+ mdAtoms, fr, ppForceWorkload, enerd
+ };
+ energyEvaluator.run(&ems, mu_tot, vir, pres, -1, TRUE);
+
+ if (MASTER(cr))
+ {
+ /* Copy stuff to the energy bin for easy printing etc. */
+ matrix nullBox = {};
+ upd_mdebin(mdebin, FALSE, FALSE, static_cast<double>(step),
+ mdatoms->tmass, enerd, nullptr, nullptr, nullptr, nullBox,
+ nullptr, nullptr, vir, pres, nullptr, 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), nullptr);
+ }
+
+ /* 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", ems.fmax, ems.a_fmax + 1);
+ fprintf(stderr, " F-Norm = %12.5e\n", ems.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", ems.fmax, ems.a_fmax + 1);
+ fprintf(fplog, " F-Norm = %12.5e\n", ems.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.
+ real *fInit = static_cast<real *>(ems.f.rvec_array()[0]);
+ for (i = 0; i < n; i++)
+ {
+ if (!frozen[i])
+ {
+ dx[point][i] = fInit[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/ems.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, static_cast<real>(step), &ems.s, state_global, observablesHistory, ems.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];
+
+ real *xx = static_cast<real *>(ems.s.x.rvec_array()[0]);
+ real *ff = static_cast<real *>(ems.f.rvec_array()[0]);
+
+ // 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
+ *last = ems;
+ real *lastx = static_cast<real *>(last->s.x.data()[0]);
+ real *lastf = static_cast<real *>(last->f.data()[0]);
+ Epot0 = ems.epot;
+
+ *sa = ems;
+
+ /* 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
+ 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
+ real *xc = static_cast<real *>(sc->s.x.rvec_array()[0]);
+ for (i = 0; i < n; i++)
+ {
+ xc[i] = lastx[i] + c*s[i];
+ }
+
+ neval++;
+ // Calculate energy for the trial step in position C
+ energyEvaluator.run(sc, mu_tot, vir, pres, step, FALSE);
+
+ // Calc line gradient in position C
+ real *fc = static_cast<real *>(sc->f.rvec_array()[0]);
+ 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 = std::sqrt(GMX_REAL_EPS)*fabs(sa->epot);
+
+ // 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.
+ foundlower = sc->epot < sa->epot || (gpc < 0 && sc->epot < (sa->epot + tmp));
+ // If true, 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.
+
+ // If false, 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.
+
+ 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.
+ real fnorm = 0;
+ 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
+ real *xb = static_cast<real *>(sb->s.x.rvec_array()[0]);
+ for (i = 0; i < n; i++)
+ {
+ xb[i] = lastx[i] + b*s[i];
+ }
+
+ neval++;
+ // Calculate energy for the trial step in point B
+ energyEvaluator.run(sb, mu_tot, vir, pres, step, FALSE);
+ fnorm = sb->fnorm;
+
+ // Calculate gradient in point B
+ real *fb = static_cast<real *>(sb->f.rvec_array()[0]);
+ 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 */
+ c = b;
+ /* swap states b and c */
+ swap_em_state(&sb, &sc);
+ }
+ else
+ {
+ /* Replace a endpoint with b */
+ a = b;
+ /* swap states a and b */
+ swap_em_state(&sa, &sb);
+ }
+
+ /*
+ * 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 ((sb->epot > sa->epot || sb->epot > sc->epot) && (nminstep < 20));
+
+ if (std::fabs(sb->epot - 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 (sc->epot < sa->epot)
+ {
+ /* Use state C */
+ ems = *sc;
+ step_taken = c;
+ }
+ else
+ {
+ /* Use state A */
+ ems = *sa;
+ step_taken = a;
+ }
+
+ }
+ else
+ {
+ /* found lower */
+ /* Use state C */
+ ems = *sc;
+ 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;
+ }
+ }
+
+ /* Print it if necessary */
+ if (MASTER(cr))
+ {
+ if (mdrunOptions.verbose)
+ {
+ 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, ems.epot, ems.fnorm/sqrtNumAtoms, ems.fmax, ems.a_fmax + 1);
+ fflush(stderr);
+ }
+ /* Store the new (lower) energies */
+ matrix nullBox = {};
+ upd_mdebin(mdebin, FALSE, FALSE, static_cast<double>(step),
+ mdatoms->tmass, enerd, nullptr, nullptr, nullptr, nullBox,
+ nullptr, nullptr, vir, pres, nullptr, 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 : nullptr, step, step, eprNORMAL,
+ mdebin, fcd, &(top_global->groups), &(inputrec->opts), nullptr);
+ }
+
+ /* 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.rvec_array(), inputrec, 0, wcycle) && MASTER(cr))
+ {
+ IMD_send_positions(inputrec->imd);
+ }
+
+ // Reset stepsize in we are doing more iterations
+ stepsize = 1.0/ems.fnorm;
+
+ /* Stop when the maximum force lies below tolerance.
+ * If we have reached machine precision, converged is already set to true.
+ */
+ converged = converged || (ems.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 (ems.fmax > inputrec->em_tol)
+ {
+ if (MASTER(cr))
+ {
+ warn_step(fplog, inputrec->em_tol, ems.fmax,
+ 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 : nullptr, step, step, eprNORMAL,
+ mdebin, fcd, &(top_global->groups), &(inputrec->opts), nullptr);
+ }
+
+ /* 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, observablesHistory);
+
+ if (MASTER(cr))
+ {
+ double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
+ print_converged(stderr, LBFGS, inputrec->em_tol, step, converged,
+ number_steps, &ems, sqrtNumAtoms);
+ print_converged(fplog, LBFGS, inputrec->em_tol, step, converged,
+ number_steps, &ems, 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);
+}
+
+void
+Integrator::do_steep()
+{
+ const char *SD = "Steepest Descents";
+ gmx_localtop_t *top;
+ gmx_enerdata_t *enerd;
+ gmx_global_stat_t gstat;
+ t_graph *graph;
+ real stepsize;
+ real ustep;
+ 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;
+ auto mdatoms = mdAtoms->mdatoms();
+
+ GMX_LOG(mdlog.info).asParagraph().
+ appendText("Note that activating steepest-descent energy minimization via the "
+ "integrator .mdp option and the command gmx mdrun may "
+ "be available in a different form in a future version of GROMACS, "
+ "e.g. gmx minimize and an .mdp option.");
+
+ /* Create 2 states on the stack and extract pointers that we will swap */
+ em_state_t s0 {}, s1 {};
+ em_state_t *s_min = &s0;
+ em_state_t *s_try = &s1;
+
+ /* Init em and store the local state in s_try */
+ init_em(fplog, mdlog, SD, cr, ms, outputProvider, inputrec, mdrunOptions,
+ state_global, top_global, s_try, &top,
+ nrnb, mu_tot, fr, &enerd, &graph, mdAtoms, &gstat,
+ vsite, constr, nullptr,
+ nfile, fnm, &outf, &mdebin, 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);
+ }
+ EnergyEvaluator energyEvaluator {
+ fplog, mdlog, cr, ms,
+ top_global, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph,
+ mdAtoms, fr, ppForceWorkload, enerd
+ };
+
+ /**** 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 */
+ bool validStep = true;
+ if (count > 0)
+ {
+ validStep =
+ do_em_step(cr, inputrec, mdatoms,
+ s_min, stepsize, &s_min->f, s_try,
+ constr, count);
+ }
+
+ if (validStep)
+ {
+ energyEvaluator.run(s_try, mu_tot, vir, pres, count, count == 0);
+ }
+ else
+ {
+ // Signal constraint error during stepping with energy=inf
+ s_try->epot = std::numeric_limits<real>::infinity();
+ }
+
+ 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 (mdrunOptions.verbose)
+ {
+ 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 */
+ matrix nullBox = {};
+ upd_mdebin(mdebin, FALSE, FALSE, static_cast<double>(count),
+ mdatoms->tmass, enerd, nullptr, nullptr, nullptr,
+ nullBox, nullptr, nullptr, vir, pres, nullptr, 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), nullptr);
+ 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, nullptr,
+ top_global, inputrec, count,
+ s_min, state_global, observablesHistory);
+ }
+ 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, mdlog, 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(fplog, inputrec->em_tol, s_min->fmax,
+ count == nsteps, constr != nullptr);
+ }
+ bAbort = TRUE;
+ }
+
+ /* Send IMD energies and positions, if bIMD is TRUE. */
+ if (do_IMD(inputrec->bIMD, count, cr, TRUE, state_global->box,
+ MASTER(cr) ? state_global->x.rvec_array() : nullptr,
+ 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 != 0, ftp2fn(efSTO, nfile, fnm),
+ top_global, inputrec, count,
+ s_min, state_global, observablesHistory);
+
+ if (MASTER(cr))
+ {
+ double sqrtNumAtoms = sqrt(static_cast<double>(state_global->natoms));
+
+ print_converged(stderr, SD, inputrec->em_tol, count, bDone, nsteps,
+ s_min, sqrtNumAtoms);
+ print_converged(fplog, SD, inputrec->em_tol, count, bDone, nsteps,
+ s_min, sqrtNumAtoms);
+ }
+
+ 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);
+}
+
+void
+Integrator::do_nm()
+{
+ const char *NM = "Normal Mode Analysis";
+ gmx_mdoutf_t outf;
+ int nnodes, node;
+ gmx_localtop_t *top;
+ gmx_enerdata_t *enerd;
+ gmx_global_stat_t gstat;
+ t_graph *graph;
+ tensor vir, pres;
+ rvec mu_tot;
+ rvec *dfdx;
+ gmx_bool bSparse; /* use sparse matrix storage format */
+ size_t sz;
+ gmx_sparsematrix_t * sparse_matrix = nullptr;
+ real * full_matrix = nullptr;
+
+ /* added with respect to mdrun */
+ int row, col;
+ real der_range = 10.0*std::sqrt(GMX_REAL_EPS);
+ real x_min;
+ bool bIsMaster = MASTER(cr);
+ auto mdatoms = mdAtoms->mdatoms();
+
+ GMX_LOG(mdlog.info).asParagraph().
+ appendText("Note that activating normal-mode analysis via the integrator "
+ ".mdp option and the command gmx mdrun may "
+ "be available in a different form in a future version of GROMACS, "
+ "e.g. gmx normal-modes.");
+
+ if (constr != nullptr)
+ {
+ gmx_fatal(FARGS, "Constraints present with Normal Mode Analysis, this combination is not supported");
+ }
+
+ gmx_shellfc_t *shellfc;
+
+ em_state_t state_work {};
+
+ /* Init em and store the local state in state_minimum */
+ init_em(fplog, mdlog, NM, cr, ms, outputProvider, inputrec, mdrunOptions,
+ state_global, top_global, &state_work, &top,
+ nrnb, mu_tot, fr, &enerd, &graph, mdAtoms, &gstat,
+ vsite, constr, &shellfc,
+ nfile, fnm, &outf, nullptr, wcycle);
+
+ std::vector<int> atom_index = get_atom_index(top_global);
+ std::vector<gmx::RVec> fneg(atom_index.size(), {0, 0, 0});
+ 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->ic->eeltype) || fr->rlist == 0.0)
+ {
+ GMX_LOG(mdlog.warning).appendText("Non-cutoff electrostatics used, forcing full Hessian format.");
+ bSparse = FALSE;
+ }
+ else if (atom_index.size() < 1000)
+ {
+ GMX_LOG(mdlog.warning).appendTextFormatted("Small system size (N=%zu), using full Hessian format.",
+ atom_index.size());
+ bSparse = FALSE;
+ }
+ else
+ {
+ GMX_LOG(mdlog.warning).appendText("Using compressed symmetric sparse Hessian format.");
+ 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);
+
+
+ /* 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%" PRId64 " steps.\n\n",
+ *(top_global->name), inputrec->nsteps);
+ }
+
+ nnodes = cr->nnodes;
+
+ /* Make evaluate_energy do a single node force calculation */
+ cr->nnodes = 1;
+ EnergyEvaluator energyEvaluator {
+ fplog, mdlog, cr, ms,
+ top_global, top,
+ inputrec, nrnb, wcycle, gstat,
+ vsite, constr, fcd, graph,
+ mdAtoms, fr, ppForceWorkload, enerd
+ };
+ energyEvaluator.run(&state_work, mu_tot, 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);
+
+ GMX_LOG(mdlog.warning).appendTextFormatted("Maximum force:%12.5e", state_work.fmax);
+ if (state_work.fmax > 1.0e-3)
+ {
+ GMX_LOG(mdlog.warning).appendText(
+ "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.");
+ }
+
+ /***********************************************************
+ *
+ * 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;
+ auto state_work_x = makeArrayRef(state_work.s.x);
+ auto state_work_f = makeArrayRef(state_work.f);
+ 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++)
+ {
+ int64_t step = 0;
+ int force_flags = GMX_FORCE_STATECHANGED | GMX_FORCE_ALLFORCES;
+ double t = 0;
+
+ x_min = state_work_x[atom][d];
+
+ for (unsigned int dx = 0; (dx < 2); dx++)
+ {
+ if (dx == 0)
+ {
+ state_work_x[atom][d] = x_min - der_range;
+ }
+ else
+ {
+ state_work_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 */
+ relax_shell_flexcon(fplog,
+ cr,
+ ms,
+ mdrunOptions.verbose,
+ nullptr,
+ step,
+ inputrec,
+ bNS,
+ force_flags,
+ top,
+ constr,
+ enerd,
+ fcd,
+ &state_work.s,
+ state_work.f.arrayRefWithPadding(),
+ vir,
+ mdatoms,
+ nrnb,
+ wcycle,
+ graph,
+ &top_global->groups,
+ shellfc,
+ fr,
+ ppForceWorkload,
+ t,
+ mu_tot,
+ vsite,
+ DdOpenBalanceRegionBeforeForceComputation::no,
+ DdCloseBalanceRegionAfterForceComputation::no);
+ bNS = false;
+ step++;
+ }
+ else
+ {
+ energyEvaluator.run(&state_work, mu_tot, vir, pres, aid*2+dx, FALSE);
+ }
+
+ cr->nnodes = nnodes;
+
+ if (dx == 0)
+ {
+ std::copy(state_work_f.begin(), state_work_f.begin()+atom_index.size(), fneg.begin());
+ }
+ }
+
+ /* x is restored to original */
+ state_work_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 && aid+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 = (aid + 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 (mdrunOptions.verbose && fplog)
+ {
+ fflush(fplog);
+ }
+ }
+ /* write progress */
+ if (bIsMaster && mdrunOptions.verbose)
+ {
+ 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);
+}
+
+} // namespace gmx
diff --git a/patches/gromacs-2019.1.diff/src/gromacs/mdrun/replicaexchange.cpp b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/replicaexchange.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..f0306c64f05d0c66387e05b8fe68c4c96fef437c
--- /dev/null
+++ b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/replicaexchange.cpp
@@ -0,0 +1,1487 @@
+/*
+ * 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) 2011,2012,2013,2014,2015,2016,2017,2018, 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 replica exchange routines.
+ *
+ * \author David van der Spoel <david.vanderspoel@icm.uu.se>
+ * \author Mark Abraham <mark.j.abraham@gmail.com>
+ * \ingroup module_mdrun
+ */
+#include "gmxpre.h"
+
+#include "replicaexchange.h"
+
+#include "config.h"
+
+#include <cmath>
+
+#include <random>
+
+#include "gromacs/domdec/collect.h"
+#include "gromacs/gmxlib/network.h"
+#include "gromacs/math/units.h"
+#include "gromacs/math/vec.h"
+#include "gromacs/mdrun/multisim.h"
+#include "gromacs/mdtypes/commrec.h"
+#include "gromacs/mdtypes/enerdata.h"
+#include "gromacs/mdtypes/inputrec.h"
+#include "gromacs/mdtypes/md_enums.h"
+#include "gromacs/mdtypes/state.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 */
+
+/* PLUMED HREX */
+extern int plumed_hrex;
+/* END PLUMED HREX */
+
+//! Helps cut off probability values.
+constexpr int c_probabilityCutoff = 100;
+
+/* we don't bother evaluating if events are more rare than exp(-100) = 3.7x10^-44 */
+
+//! Rank in the multisimulation
+#define MSRANK(ms, nodeid) (nodeid)
+
+//! Enum for replica exchange flavours
+enum {
+ ereTEMP, ereLAMBDA, ereENDSINGLE, ereTL, ereNR
+};
+/*! \brief Strings describing replica exchange flavours.
+ *
+ * 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 */
+static const char *erename[ereNR] = { "temperature", "lambda", "end_single_marker", "temperature and lambda"};
+
+//! Working data for replica exchange.
+struct gmx_repl_ex
+{
+ //! Replica ID
+ int repl;
+ //! Total number of replica
+ int nrepl;
+ //! Temperature
+ real temp;
+ //! Replica exchange type from ere enum
+ int type;
+ //! Quantity, e.g. temperature or lambda; first index is ere, second index is replica ID
+ real **q;
+ //! Use constant pressure and temperature
+ gmx_bool bNPT;
+ //! Replica pressures
+ real *pres;
+ //! Replica indices
+ int *ind;
+ //! Used for keeping track of all the replica swaps
+ int *allswaps;
+ //! Replica exchange interval (number of steps)
+ int nst;
+ //! Number of exchanges per interval
+ int nex;
+ //! Random seed
+ int seed;
+ //! Number of even and odd replica change attempts
+ int nattempt[2];
+ //! Sum of probabilities
+ real *prob_sum;
+ //! Number of moves between replicas i and j
+ int **nmoves;
+ //! i-th element of the array is the number of exchanges between replica i-1 and i
+ int *nexchange;
+
+ /*! \brief 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;
+ //! \}
+};
+
+// TODO We should add Doxygen here some time.
+//! \cond
+
+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,
+ int numAtomsInSystem,
+ const t_inputrec *ir,
+ const ReplicaExchangeParameters &replExParams)
+{
+ real pres;
+ int i, j;
+ struct gmx_repl_ex *re;
+ gmx_bool bTemp;
+ gmx_bool bLambda = FALSE;
+
+ fprintf(fplog, "\nInitializing Replica Exchange\n");
+
+ if (!isMultiSim(ms) || ms->nsim == 1)
+ {
+ gmx_fatal(FARGS, "Nothing to exchange with only one replica, maybe you forgot to set the -multidir 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 isMultiSim(ms). 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);
+
+ /* We only check that the number of atoms in the systms match.
+ * This, of course, do not guarantee that the systems are the same,
+ * but it does guarantee that we can perform replica exchange.
+ */
+ check_multi_int(fplog, ms, numAtomsInSystem, "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);
+ const int nst = replExParams.exchangeInterval;
+ 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, static_cast<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]);
+ }
+ 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", static_cast<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", static_cast<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 (replExParams.randomSeed == -1)
+ {
+ if (isMasterSim(ms))
+ {
+ re->seed = static_cast<int>(gmx::makeRandomSeed());
+ }
+ else
+ {
+ re->seed = 0;
+ }
+ gmx_sumi_sim(1, &(re->seed), ms);
+ }
+ else
+ {
+ re->seed = replExParams.randomSeed;
+ }
+ 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 = replExParams.numExchanges;
+ 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);
+ }
+}
+
+/* PLUMED HREX */
+void exchange_state(const gmx_multisim_t *ms, int b, t_state *state)
+/* END PLUMED HREX */
+{
+ /* 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.data(), ngtc);
+ exchange_doubles(ms, b, state->nosehoover_vxi.data(), ngtc);
+ exchange_doubles(ms, b, state->nhpres_xi.data(), nnhpres);
+ exchange_doubles(ms, b, state->nhpres_vxi.data(), nnhpres);
+ exchange_doubles(ms, b, state->therm_integral.data(), state->ngtc);
+ exchange_doubles(ms, b, &state->baros_integral, 1);
+ exchange_rvecs(ms, b, state->x.rvec_array(), state->natoms);
+ exchange_rvecs(ms, b, state->v.rvec_array(), state->natoms);
+}
+
+/* PLUMED HREX */
+void copy_state_serial(const t_state *src, t_state *dest)
+/* END PLUMED HREX */
+{
+ if (dest != src)
+ {
+ /* Currently the local state is always a pointer to the global
+ * in serial, so we should never end up here.
+ * TODO: Implement a (trivial) t_state copy once converted to C++.
+ */
+ GMX_RELEASE_ASSERT(false, "State copying is currently not implemented in replica exchange");
+ }
+}
+
+static void scale_velocities(gmx::ArrayRef<gmx::RVec> velocities, real fac)
+{
+ for (auto &v : velocities)
+ {
+ v *= fac;
+ }
+}
+
+static void print_transition_matrix(FILE *fplog, int n, int **nmoves, const 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, const gmx_bool *bEx)
+{
+ int i;
+
+ fprintf(fplog, "Repl %2s %2d", leg, ind[0]);
+ for (i = 1; i < n; i++)
+ {
+ fprintf(fplog, " %c %2d", (bEx != nullptr && 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);
+ }
+/* PLUMED HREX */
+/* this is necessary because with plumed HREX the energy contribution is
+ already taken into account */
+ if(plumed_hrex) delta=0.0;
+/* END PLUMED HREX */
+ 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,
+ const gmx_enerdata_t *enerd,
+ real vol,
+ 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 %" 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[static_cast<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];
+ }
+
+ rng.restart( step, 0 );
+
+ /* PLUMED */
+ int plumed_test_exchange_pattern=0;
+ if(plumed_test_exchange_pattern && plumed_hrex) gmx_fatal(FARGS,"hrex not compatible with ad hoc exchange patterns");
+ /* END PLUMED */
+
+ if (bMultiEx)
+ {
+ /* multiple random switch exchange */
+ int nself = 0;
+
+
+ for (i = 0; i < re->nex + nself; i++)
+ {
+ // For now this is superfluous, but just in case we ever add more
+ // calls in different branches it is safer to always reset the distribution.
+ uniformNreplDist.reset();
+
+ /* 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 > c_probabilityCutoff)
+ {
+ prob[0] = 0;
+ }
+ else
+ {
+ prob[0] = exp(-delta);
+ }
+ // roll a number to determine if accepted. For now it is superfluous to
+ // reset, but just in case we ever add more calls in different branches
+ // it is safer to always reset the distribution.
+ uniformRealDist.reset();
+ 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 > c_probabilityCutoff)
+ {
+ prob[i] = 0;
+ }
+ else
+ {
+ prob[i] = exp(-delta);
+ }
+ // roll a number to determine if accepted. For now it is superfluous to
+ // reset, but just in case we ever add more calls in different branches
+ // it is safer to always reset the distribution.
+ uniformRealDist.reset();
+ 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,
+ const gmx_multisim_t *ms, struct gmx_repl_ex *re,
+ t_state *state, const gmx_enerdata_t *enerd,
+ t_state *state_local, 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, 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_serial(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(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->v,
+ std::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_serial(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, nullptr);
+ print_prob(fplog, "", re->nrepl, re->prob);
+
+ fprintf(fplog, "Repl number of exchanges:\n");
+ print_ind(fplog, "", re->nrepl, re->ind, nullptr);
+ 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] = (static_cast<real>(re->nexchange[i]))/re->nattempt[i%2];
+ }
+ }
+ print_ind(fplog, "", re->nrepl, re->ind, nullptr);
+ print_prob(fplog, "", re->nrepl, re->prob);
+
+ fprintf(fplog, "\n");
+ }
+ /* print the transition matrix */
+ print_transition_matrix(fplog, re->nrepl, re->nmoves, re->nattempt);
+}
+
+/* PLUMED HREX */
+int replica_exchange_get_repl(const gmx_repl_ex_t re){
+ return re->repl;
+};
+
+int replica_exchange_get_nrepl(const gmx_repl_ex_t re){
+ return re->nrepl;
+};
+/* END PLUMED HREX */
+//! \endcond
diff --git a/patches/gromacs-2019.1.diff/src/gromacs/mdrun/replicaexchange.cpp.preplumed b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/replicaexchange.cpp.preplumed
new file mode 100644
index 0000000000000000000000000000000000000000..a633d688c67c1755effb063557635296875bc876
--- /dev/null
+++ b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/replicaexchange.cpp.preplumed
@@ -0,0 +1,1383 @@
+/*
+ * 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) 2011,2012,2013,2014,2015,2016,2017,2018, 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 replica exchange routines.
+ *
+ * \author David van der Spoel <david.vanderspoel@icm.uu.se>
+ * \author Mark Abraham <mark.j.abraham@gmail.com>
+ * \ingroup module_mdrun
+ */
+#include "gmxpre.h"
+
+#include "replicaexchange.h"
+
+#include "config.h"
+
+#include <cmath>
+
+#include <random>
+
+#include "gromacs/domdec/collect.h"
+#include "gromacs/gmxlib/network.h"
+#include "gromacs/math/units.h"
+#include "gromacs/math/vec.h"
+#include "gromacs/mdrun/multisim.h"
+#include "gromacs/mdtypes/commrec.h"
+#include "gromacs/mdtypes/enerdata.h"
+#include "gromacs/mdtypes/inputrec.h"
+#include "gromacs/mdtypes/md_enums.h"
+#include "gromacs/mdtypes/state.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"
+
+//! Helps cut off probability values.
+constexpr int c_probabilityCutoff = 100;
+
+/* we don't bother evaluating if events are more rare than exp(-100) = 3.7x10^-44 */
+
+//! Rank in the multisimulation
+#define MSRANK(ms, nodeid) (nodeid)
+
+//! Enum for replica exchange flavours
+enum {
+ ereTEMP, ereLAMBDA, ereENDSINGLE, ereTL, ereNR
+};
+/*! \brief Strings describing replica exchange flavours.
+ *
+ * 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 */
+static const char *erename[ereNR] = { "temperature", "lambda", "end_single_marker", "temperature and lambda"};
+
+//! Working data for replica exchange.
+struct gmx_repl_ex
+{
+ //! Replica ID
+ int repl;
+ //! Total number of replica
+ int nrepl;
+ //! Temperature
+ real temp;
+ //! Replica exchange type from ere enum
+ int type;
+ //! Quantity, e.g. temperature or lambda; first index is ere, second index is replica ID
+ real **q;
+ //! Use constant pressure and temperature
+ gmx_bool bNPT;
+ //! Replica pressures
+ real *pres;
+ //! Replica indices
+ int *ind;
+ //! Used for keeping track of all the replica swaps
+ int *allswaps;
+ //! Replica exchange interval (number of steps)
+ int nst;
+ //! Number of exchanges per interval
+ int nex;
+ //! Random seed
+ int seed;
+ //! Number of even and odd replica change attempts
+ int nattempt[2];
+ //! Sum of probabilities
+ real *prob_sum;
+ //! Number of moves between replicas i and j
+ int **nmoves;
+ //! i-th element of the array is the number of exchanges between replica i-1 and i
+ int *nexchange;
+
+ /*! \brief 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;
+ //! \}
+};
+
+// TODO We should add Doxygen here some time.
+//! \cond
+
+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,
+ int numAtomsInSystem,
+ const t_inputrec *ir,
+ const ReplicaExchangeParameters &replExParams)
+{
+ real pres;
+ int i, j;
+ struct gmx_repl_ex *re;
+ gmx_bool bTemp;
+ gmx_bool bLambda = FALSE;
+
+ fprintf(fplog, "\nInitializing Replica Exchange\n");
+
+ if (!isMultiSim(ms) || ms->nsim == 1)
+ {
+ gmx_fatal(FARGS, "Nothing to exchange with only one replica, maybe you forgot to set the -multidir 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 isMultiSim(ms). 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);
+
+ /* We only check that the number of atoms in the systms match.
+ * This, of course, do not guarantee that the systems are the same,
+ * but it does guarantee that we can perform replica exchange.
+ */
+ check_multi_int(fplog, ms, numAtomsInSystem, "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);
+ const int nst = replExParams.exchangeInterval;
+ 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, static_cast<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]);
+ }
+ 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", static_cast<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", static_cast<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 (replExParams.randomSeed == -1)
+ {
+ if (isMasterSim(ms))
+ {
+ re->seed = static_cast<int>(gmx::makeRandomSeed());
+ }
+ else
+ {
+ re->seed = 0;
+ }
+ gmx_sumi_sim(1, &(re->seed), ms);
+ }
+ else
+ {
+ re->seed = replExParams.randomSeed;
+ }
+ 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 = replExParams.numExchanges;
+ 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.data(), ngtc);
+ exchange_doubles(ms, b, state->nosehoover_vxi.data(), ngtc);
+ exchange_doubles(ms, b, state->nhpres_xi.data(), nnhpres);
+ exchange_doubles(ms, b, state->nhpres_vxi.data(), nnhpres);
+ exchange_doubles(ms, b, state->therm_integral.data(), state->ngtc);
+ exchange_doubles(ms, b, &state->baros_integral, 1);
+ exchange_rvecs(ms, b, state->x.rvec_array(), state->natoms);
+ exchange_rvecs(ms, b, state->v.rvec_array(), state->natoms);
+}
+
+static void copy_state_serial(const t_state *src, t_state *dest)
+{
+ if (dest != src)
+ {
+ /* Currently the local state is always a pointer to the global
+ * in serial, so we should never end up here.
+ * TODO: Implement a (trivial) t_state copy once converted to C++.
+ */
+ GMX_RELEASE_ASSERT(false, "State copying is currently not implemented in replica exchange");
+ }
+}
+
+static void scale_velocities(gmx::ArrayRef<gmx::RVec> velocities, real fac)
+{
+ for (auto &v : velocities)
+ {
+ v *= fac;
+ }
+}
+
+static void print_transition_matrix(FILE *fplog, int n, int **nmoves, const 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, const gmx_bool *bEx)
+{
+ int i;
+
+ fprintf(fplog, "Repl %2s %2d", leg, ind[0]);
+ for (i = 1; i < n; i++)
+ {
+ fprintf(fplog, " %c %2d", (bEx != nullptr && 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,
+ const gmx_enerdata_t *enerd,
+ real vol,
+ 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 %" 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[static_cast<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];
+ }
+
+ rng.restart( step, 0 );
+
+ if (bMultiEx)
+ {
+ /* multiple random switch exchange */
+ int nself = 0;
+
+
+ for (i = 0; i < re->nex + nself; i++)
+ {
+ // For now this is superfluous, but just in case we ever add more
+ // calls in different branches it is safer to always reset the distribution.
+ uniformNreplDist.reset();
+
+ /* 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 > c_probabilityCutoff)
+ {
+ prob[0] = 0;
+ }
+ else
+ {
+ prob[0] = exp(-delta);
+ }
+ // roll a number to determine if accepted. For now it is superfluous to
+ // reset, but just in case we ever add more calls in different branches
+ // it is safer to always reset the distribution.
+ uniformRealDist.reset();
+ 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 > c_probabilityCutoff)
+ {
+ prob[i] = 0;
+ }
+ else
+ {
+ prob[i] = exp(-delta);
+ }
+ // roll a number to determine if accepted. For now it is superfluous to
+ // reset, but just in case we ever add more calls in different branches
+ // it is safer to always reset the distribution.
+ uniformRealDist.reset();
+ 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,
+ const gmx_multisim_t *ms, struct gmx_repl_ex *re,
+ t_state *state, const gmx_enerdata_t *enerd,
+ t_state *state_local, 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, 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_serial(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(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->v,
+ std::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_serial(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, nullptr);
+ print_prob(fplog, "", re->nrepl, re->prob);
+
+ fprintf(fplog, "Repl number of exchanges:\n");
+ print_ind(fplog, "", re->nrepl, re->ind, nullptr);
+ 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] = (static_cast<real>(re->nexchange[i]))/re->nattempt[i%2];
+ }
+ }
+ print_ind(fplog, "", re->nrepl, re->ind, nullptr);
+ print_prob(fplog, "", re->nrepl, re->prob);
+
+ fprintf(fplog, "\n");
+ }
+ /* print the transition matrix */
+ print_transition_matrix(fplog, re->nrepl, re->nmoves, re->nattempt);
+}
+
+//! \endcond
diff --git a/patches/gromacs-2019.1.diff/src/gromacs/mdrun/replicaexchange.h b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/replicaexchange.h
new file mode 100644
index 0000000000000000000000000000000000000000..bff7c22ff9ac47719ec2a283253792f5dfec4e1d
--- /dev/null
+++ b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/replicaexchange.h
@@ -0,0 +1,116 @@
+/*
+ * 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) 2011,2012,2013,2014,2015,2017,2018, 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.
+ */
+/*! \libinternal \file
+ *
+ * \brief Declares the routines for replica exchange.
+ *
+ * \author David van der Spoel <david.vanderspoel@icm.uu.se>
+ * \author Mark Abraham <mark.j.abraham@gmail.com>
+ *
+ * \ingroup module_mdrun
+ */
+#ifndef GMX_MDRUN_REPLICAEXCHANGE_H
+#define GMX_MDRUN_REPLICAEXCHANGE_H
+
+#include <cstdio>
+
+#include "gromacs/utility/basedefinitions.h"
+#include "gromacs/utility/real.h"
+
+struct gmx_enerdata_t;
+struct gmx_multisim_t;
+struct t_commrec;
+struct t_inputrec;
+class t_state;
+
+/*! \libinternal
+ * \brief The parameters for the replica exchange algorithm. */
+struct ReplicaExchangeParameters
+{
+ //! Interval in steps at which to attempt exchanges, 0 means no replica exchange.
+ int exchangeInterval = 0;
+ //! The number of exchanges to attempt at an exchange step.
+ int numExchanges = 0;
+ //! The random seed, -1 means generate a seed.
+ int randomSeed = -1;
+};
+
+//! Abstract type for replica exchange
+typedef struct gmx_repl_ex *gmx_repl_ex_t;
+
+/*! \brief Setup function.
+ *
+ * Should only be called on the master ranks */
+gmx_repl_ex_t
+init_replica_exchange(FILE *fplog,
+ const gmx_multisim_t *ms,
+ int numAtomsInSystem,
+ const t_inputrec *ir,
+ const ReplicaExchangeParameters &replExParams);
+
+/*! \brief Attempts replica exchange.
+ *
+ * Should be called on all ranks. When running each replica in
+ * parallel, this routine collects the state on the master rank before
+ * exchange. With domain decomposition, the global state after
+ * exchange is stored in state and still needs to be redistributed
+ * over the ranks.
+ *
+ * \returns TRUE if the state has been exchanged.
+ */
+gmx_bool replica_exchange(FILE *fplog,
+ const t_commrec *cr,
+ const gmx_multisim_t *ms,
+ gmx_repl_ex_t re,
+ t_state *state, const gmx_enerdata_t *enerd,
+ t_state *state_local,
+ int64_t step, real time);
+
+/*! \brief Prints replica exchange statistics to the log file.
+ *
+ * Should only be called on the master ranks */
+void print_replica_exchange_statistics(FILE *fplog, gmx_repl_ex_t re);
+
+/* PLUMED HREX */
+extern int replica_exchange_get_repl(const gmx_repl_ex_t re);
+extern int replica_exchange_get_nrepl(const gmx_repl_ex_t re);
+extern void pd_collect_state(const t_commrec *cr, t_state *state);
+extern void exchange_state(const gmx_multisim_t *ms, int b, t_state *state);
+extern void copy_state_serial(const t_state *src, t_state *dest);
+/* END PLUMED HREX */
+
+#endif
diff --git a/patches/gromacs-2019.1.diff/src/gromacs/mdrun/replicaexchange.h.preplumed b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/replicaexchange.h.preplumed
new file mode 100644
index 0000000000000000000000000000000000000000..8f4211febe707fc6d474d2a5f840c2a2c79ca0de
--- /dev/null
+++ b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/replicaexchange.h.preplumed
@@ -0,0 +1,108 @@
+/*
+ * 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) 2011,2012,2013,2014,2015,2017,2018, 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.
+ */
+/*! \libinternal \file
+ *
+ * \brief Declares the routines for replica exchange.
+ *
+ * \author David van der Spoel <david.vanderspoel@icm.uu.se>
+ * \author Mark Abraham <mark.j.abraham@gmail.com>
+ *
+ * \ingroup module_mdrun
+ */
+#ifndef GMX_MDRUN_REPLICAEXCHANGE_H
+#define GMX_MDRUN_REPLICAEXCHANGE_H
+
+#include <cstdio>
+
+#include "gromacs/utility/basedefinitions.h"
+#include "gromacs/utility/real.h"
+
+struct gmx_enerdata_t;
+struct gmx_multisim_t;
+struct t_commrec;
+struct t_inputrec;
+class t_state;
+
+/*! \libinternal
+ * \brief The parameters for the replica exchange algorithm. */
+struct ReplicaExchangeParameters
+{
+ //! Interval in steps at which to attempt exchanges, 0 means no replica exchange.
+ int exchangeInterval = 0;
+ //! The number of exchanges to attempt at an exchange step.
+ int numExchanges = 0;
+ //! The random seed, -1 means generate a seed.
+ int randomSeed = -1;
+};
+
+//! Abstract type for replica exchange
+typedef struct gmx_repl_ex *gmx_repl_ex_t;
+
+/*! \brief Setup function.
+ *
+ * Should only be called on the master ranks */
+gmx_repl_ex_t
+init_replica_exchange(FILE *fplog,
+ const gmx_multisim_t *ms,
+ int numAtomsInSystem,
+ const t_inputrec *ir,
+ const ReplicaExchangeParameters &replExParams);
+
+/*! \brief Attempts replica exchange.
+ *
+ * Should be called on all ranks. When running each replica in
+ * parallel, this routine collects the state on the master rank before
+ * exchange. With domain decomposition, the global state after
+ * exchange is stored in state and still needs to be redistributed
+ * over the ranks.
+ *
+ * \returns TRUE if the state has been exchanged.
+ */
+gmx_bool replica_exchange(FILE *fplog,
+ const t_commrec *cr,
+ const gmx_multisim_t *ms,
+ gmx_repl_ex_t re,
+ t_state *state, const gmx_enerdata_t *enerd,
+ t_state *state_local,
+ int64_t step, real time);
+
+/*! \brief Prints replica exchange statistics to the log file.
+ *
+ * Should only be called on the master ranks */
+void print_replica_exchange_statistics(FILE *fplog, gmx_repl_ex_t re);
+
+#endif
diff --git a/patches/gromacs-2019.1.diff/src/gromacs/mdrun/runner.cpp b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/runner.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..7d206adf79a142aeb3bddc42b5e8e69c81dff63b
--- /dev/null
+++ b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/runner.cpp
@@ -0,0 +1,1956 @@
+/*
+ * 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) 2011,2012,2013,2014,2015,2016,2017,2018,2019, 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_mdrun
+ */
+#include "gmxpre.h"
+
+#include "runner.h"
+
+#include "config.h"
+
+#include <cassert>
+#include <cinttypes>
+#include <csignal>
+#include <cstdlib>
+#include <cstring>
+
+#include <algorithm>
+
+#include "gromacs/commandline/filenm.h"
+#include "gromacs/compat/make_unique.h"
+#include "gromacs/domdec/domdec.h"
+#include "gromacs/domdec/domdec_struct.h"
+#include "gromacs/domdec/localatomsetmanager.h"
+#include "gromacs/ewald/ewald-utils.h"
+#include "gromacs/ewald/pme.h"
+#include "gromacs/ewald/pme-gpu-program.h"
+#include "gromacs/fileio/checkpoint.h"
+#include "gromacs/fileio/gmxfio.h"
+#include "gromacs/fileio/oenv.h"
+#include "gromacs/fileio/tpxio.h"
+#include "gromacs/gmxlib/network.h"
+#include "gromacs/gmxlib/nrnb.h"
+#include "gromacs/gpu_utils/clfftinitializer.h"
+#include "gromacs/gpu_utils/gpu_utils.h"
+#include "gromacs/hardware/cpuinfo.h"
+#include "gromacs/hardware/detecthardware.h"
+#include "gromacs/hardware/printhardware.h"
+#include "gromacs/listed-forces/disre.h"
+#include "gromacs/listed-forces/gpubonded.h"
+#include "gromacs/listed-forces/orires.h"
+#include "gromacs/math/functions.h"
+#include "gromacs/math/utilities.h"
+#include "gromacs/math/vec.h"
+#include "gromacs/mdlib/boxdeformation.h"
+#include "gromacs/mdlib/calc_verletbuf.h"
+#include "gromacs/mdlib/forcerec.h"
+#include "gromacs/mdlib/gmx_omp_nthreads.h"
+#include "gromacs/mdlib/makeconstraints.h"
+#include "gromacs/mdlib/md_support.h"
+#include "gromacs/mdlib/mdatoms.h"
+#include "gromacs/mdlib/mdrun.h"
+#include "gromacs/mdlib/membed.h"
+#include "gromacs/mdlib/nb_verlet.h"
+#include "gromacs/mdlib/nbnxn_gpu_data_mgmt.h"
+#include "gromacs/mdlib/nbnxn_search.h"
+#include "gromacs/mdlib/nbnxn_tuning.h"
+#include "gromacs/mdlib/ppforceworkload.h"
+#include "gromacs/mdlib/qmmm.h"
+#include "gromacs/mdlib/sighandler.h"
+#include "gromacs/mdlib/sim_util.h"
+#include "gromacs/mdlib/stophandler.h"
+#include "gromacs/mdrun/legacymdrunoptions.h"
+#include "gromacs/mdrun/logging.h"
+#include "gromacs/mdrun/multisim.h"
+#include "gromacs/mdrun/simulationcontext.h"
+#include "gromacs/mdrunutility/mdmodules.h"
+#include "gromacs/mdrunutility/threadaffinity.h"
+#include "gromacs/mdtypes/commrec.h"
+#include "gromacs/mdtypes/fcdata.h"
+#include "gromacs/mdtypes/inputrec.h"
+#include "gromacs/mdtypes/md_enums.h"
+#include "gromacs/mdtypes/observableshistory.h"
+#include "gromacs/mdtypes/state.h"
+#include "gromacs/pbcutil/pbc.h"
+#include "gromacs/pulling/output.h"
+#include "gromacs/pulling/pull.h"
+#include "gromacs/pulling/pull_rotation.h"
+#include "gromacs/restraint/manager.h"
+#include "gromacs/restraint/restraintmdmodule.h"
+#include "gromacs/restraint/restraintpotential.h"
+#include "gromacs/swap/swapcoords.h"
+#include "gromacs/taskassignment/decidegpuusage.h"
+#include "gromacs/taskassignment/resourcedivision.h"
+#include "gromacs/taskassignment/taskassignment.h"
+#include "gromacs/taskassignment/usergpuids.h"
+#include "gromacs/timing/wallcycle.h"
+#include "gromacs/topology/mtop_util.h"
+#include "gromacs/trajectory/trajectoryframe.h"
+#include "gromacs/utility/basenetwork.h"
+#include "gromacs/utility/cstringutil.h"
+#include "gromacs/utility/exceptions.h"
+#include "gromacs/utility/fatalerror.h"
+#include "gromacs/utility/filestream.h"
+#include "gromacs/utility/gmxassert.h"
+#include "gromacs/utility/gmxmpi.h"
+#include "gromacs/utility/logger.h"
+#include "gromacs/utility/loggerbuilder.h"
+#include "gromacs/utility/physicalnodecommunicator.h"
+#include "gromacs/utility/pleasecite.h"
+#include "gromacs/utility/programcontext.h"
+#include "gromacs/utility/smalloc.h"
+#include "gromacs/utility/stringutil.h"
+
+#include "integrator.h"
+#include "replicaexchange.h"
+
+#if GMX_FAHCORE
+#include "corewrap.h"
+#endif
+
+/* PLUMED */
+#include "../../../Plumed.h"
+extern int plumedswitch;
+extern plumed plumedmain;
+/* END PLUMED */
+namespace gmx
+{
+
+/*! \brief Barrier for safe simultaneous thread access to mdrunner data
+ *
+ * Used to ensure that the master thread does not modify mdrunner during copy
+ * on the spawned threads. */
+static void threadMpiMdrunnerAccessBarrier()
+{
+#if GMX_THREAD_MPI
+ MPI_Barrier(MPI_COMM_WORLD);
+#endif
+}
+
+Mdrunner Mdrunner::cloneOnSpawnedThread() const
+{
+ auto newRunner = Mdrunner();
+
+ // All runners in the same process share a restraint manager resource because it is
+ // part of the interface to the client code, which is associated only with the
+ // original thread. Handles to the same resources can be obtained by copy.
+ {
+ newRunner.restraintManager_ = compat::make_unique<RestraintManager>(*restraintManager_);
+ }
+
+ // Copy original cr pointer before master thread can pass the thread barrier
+ newRunner.cr = reinitialize_commrec_for_this_thread(cr);
+
+ // Copy members of master runner.
+ // \todo Replace with builder when Simulation context and/or runner phases are better defined.
+ // Ref https://redmine.gromacs.org/issues/2587 and https://redmine.gromacs.org/issues/2375
+ newRunner.hw_opt = hw_opt;
+ newRunner.filenames = filenames;
+
+ newRunner.oenv = oenv;
+ newRunner.mdrunOptions = mdrunOptions;
+ newRunner.domdecOptions = domdecOptions;
+ newRunner.nbpu_opt = nbpu_opt;
+ newRunner.pme_opt = pme_opt;
+ newRunner.pme_fft_opt = pme_fft_opt;
+ newRunner.bonded_opt = bonded_opt;
+ newRunner.nstlist_cmdline = nstlist_cmdline;
+ newRunner.replExParams = replExParams;
+ newRunner.pforce = pforce;
+ newRunner.ms = ms;
+ newRunner.stopHandlerBuilder_ = compat::make_unique<StopHandlerBuilder>(*stopHandlerBuilder_);
+
+ threadMpiMdrunnerAccessBarrier();
+
+ GMX_RELEASE_ASSERT(!MASTER(newRunner.cr), "cloneOnSpawnedThread should only be called on spawned threads");
+
+ return newRunner;
+}
+
+/*! \brief The callback used for running on spawned threads.
+ *
+ * Obtains the pointer to the master mdrunner object from the one
+ * argument permitted to the thread-launch API call, copies it to make
+ * a new runner for this thread, reinitializes necessary data, and
+ * proceeds to the simulation. */
+static void mdrunner_start_fn(const void *arg)
+{
+ try
+ {
+ auto masterMdrunner = reinterpret_cast<const gmx::Mdrunner *>(arg);
+ /* copy the arg list to make sure that it's thread-local. This
+ doesn't copy pointed-to items, of course; fnm, cr and fplog
+ are reset in the call below, all others should be const. */
+ gmx::Mdrunner mdrunner = masterMdrunner->cloneOnSpawnedThread();
+ mdrunner.mdrunner();
+ }
+ GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+}
+
+
+/*! \brief Start thread-MPI threads.
+ *
+ * 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 are the same as for
+ * mdrunner(). */
+t_commrec *Mdrunner::spawnThreads(int numThreadsToLaunch) const
+{
+
+ /* first check whether we even need to start tMPI */
+ if (numThreadsToLaunch < 2)
+ {
+ return cr;
+ }
+
+#if GMX_THREAD_MPI
+ /* now spawn new threads that start mdrunner_start_fn(), while
+ the main thread returns, we set thread affinity later */
+ if (tMPI_Init_fn(TRUE, numThreadsToLaunch, TMPI_AFFINITY_NONE,
+ mdrunner_start_fn, static_cast<const void*>(this)) != TMPI_SUCCESS)
+ {
+ GMX_THROW(gmx::InternalError("Failed to spawn thread-MPI threads"));
+ }
+
+ threadMpiMdrunnerAccessBarrier();
+#else
+ GMX_UNUSED_VALUE(mdrunner_start_fn);
+#endif
+
+ return reinitialize_commrec_for_this_thread(cr);
+}
+
+} // namespace gmx
+
+/*! \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,
+ const matrix box,
+ bool makeGpuPairList,
+ 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 */
+
+ /* 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.
+ */
+ ListSetupType listType = (makeGpuPairList ? ListSetupType::Gpu : ListSetupType::CpuSimdWhenSupported);
+ VerletbufListSetup listSetup = verletbufGetSafeListSetup(listType);
+
+ real rlist_new;
+ calc_verlet_buffer_size(mtop, det(box), ir, ir->nstlist, ir->nstlist - 1, -1, &listSetup, nullptr, &rlist_new);
+
+ if (rlist_new != ir->rlist)
+ {
+ if (fplog != nullptr)
+ {
+ fprintf(fplog, "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
+ ir->rlist, rlist_new,
+ listSetup.cluster_size_i, listSetup.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 */
+ increaseNstlist(fplog, cr, ir, nstlist_cmdline, mtop, box, makeGpuPairList, cpuinfo);
+ }
+}
+
+/*! \brief Override the nslist value in inputrec
+ *
+ * with value passed on the command line (if any)
+ */
+static void override_nsteps_cmdline(const gmx::MDLogger &mdlog,
+ int64_t nsteps_cmdline,
+ t_inputrec *ir)
+{
+ assert(ir);
+
+ /* 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));
+ }
+
+ GMX_LOG(mdlog.warning).asParagraph().appendText(sbuf_msg);
+ }
+ else if (nsteps_cmdline < -2)
+ {
+ gmx_fatal(FARGS, "Invalid nsteps value passed on the command line: %" PRId64,
+ nsteps_cmdline);
+ }
+ /* Do nothing if nsteps_cmdline == -2 */
+}
+
+namespace gmx
+{
+
+/*! \brief Return whether GPU acceleration of nonbondeds is supported with the given settings.
+ *
+ * If not, and if a warning may be issued, logs a warning about
+ * falling back to CPU code. With thread-MPI, only the first
+ * call to this function should have \c issueWarning true. */
+static bool gpuAccelerationOfNonbondedIsUseful(const MDLogger &mdlog,
+ const t_inputrec *ir,
+ bool issueWarning)
+{
+ if (ir->opts.ngener - ir->nwall > 1)
+ {
+ /* The GPU code does not support more than one energy group.
+ * If the user requested GPUs explicitly, a fatal error is given later.
+ */
+ if (issueWarning)
+ {
+ GMX_LOG(mdlog.warning).asParagraph()
+ .appendText("Multiple energy groups is not implemented for GPUs, falling back to the CPU. "
+ "For better performance, run on the GPU without energy groups and then do "
+ "gmx mdrun -rerun option on the trajectory with an energy group .tpr file.");
+ }
+ return false;
+ }
+ return true;
+}
+
+//! Initializes the logger for mdrun.
+static gmx::LoggerOwner buildLogger(FILE *fplog, const t_commrec *cr)
+{
+ gmx::LoggerBuilder builder;
+ if (fplog != nullptr)
+ {
+ builder.addTargetFile(gmx::MDLogger::LogLevel::Info, fplog);
+ }
+ if (cr == nullptr || SIMMASTER(cr))
+ {
+ builder.addTargetStream(gmx::MDLogger::LogLevel::Warning,
+ &gmx::TextOutputFile::standardError());
+ }
+ return builder.build();
+}
+
+//! Make a TaskTarget from an mdrun argument string.
+static TaskTarget findTaskTarget(const char *optionString)
+{
+ TaskTarget returnValue = TaskTarget::Auto;
+
+ if (strncmp(optionString, "auto", 3) == 0)
+ {
+ returnValue = TaskTarget::Auto;
+ }
+ else if (strncmp(optionString, "cpu", 3) == 0)
+ {
+ returnValue = TaskTarget::Cpu;
+ }
+ else if (strncmp(optionString, "gpu", 3) == 0)
+ {
+ returnValue = TaskTarget::Gpu;
+ }
+ else
+ {
+ GMX_ASSERT(false, "Option string should have been checked for sanity already");
+ }
+
+ return returnValue;
+}
+
+int Mdrunner::mdrunner()
+{
+ matrix box;
+ t_nrnb *nrnb;
+ t_forcerec *fr = nullptr;
+ t_fcdata *fcd = nullptr;
+ real ewaldcoeff_q = 0;
+ real ewaldcoeff_lj = 0;
+ int nChargePerturbed = -1, nTypePerturbed = 0;
+ gmx_wallcycle_t wcycle;
+ gmx_walltime_accounting_t walltime_accounting = nullptr;
+ int rc;
+ int64_t reset_counters;
+ int nthreads_pme = 1;
+ gmx_membed_t * membed = nullptr;
+ gmx_hw_info_t *hwinfo = nullptr;
+
+ /* CAUTION: threads may be started later on in this function, so
+ cr doesn't reflect the final parallel state right now */
+ std::unique_ptr<gmx::MDModules> mdModules(new gmx::MDModules);
+ t_inputrec inputrecInstance;
+ t_inputrec *inputrec = &inputrecInstance;
+ gmx_mtop_t mtop;
+
+ bool doMembed = opt2bSet("-membed", filenames.size(), filenames.data());
+ bool doRerun = mdrunOptions.rerun;
+
+ // Handle task-assignment related user options.
+ EmulateGpuNonbonded emulateGpuNonbonded = (getenv("GMX_EMULATE_GPU") != nullptr ?
+ EmulateGpuNonbonded::Yes : EmulateGpuNonbonded::No);
+ std::vector<int> gpuIdsAvailable;
+ try
+ {
+ gpuIdsAvailable = parseUserGpuIds(hw_opt.gpuIdsAvailable);
+ // TODO We could put the GPU IDs into a std::map to find
+ // duplicates, but for the small numbers of IDs involved, this
+ // code is simple and fast.
+ for (size_t i = 0; i != gpuIdsAvailable.size(); ++i)
+ {
+ for (size_t j = i+1; j != gpuIdsAvailable.size(); ++j)
+ {
+ if (gpuIdsAvailable[i] == gpuIdsAvailable[j])
+ {
+ GMX_THROW(InvalidInputError(formatString("The string of available GPU device IDs '%s' may not contain duplicate device IDs", hw_opt.gpuIdsAvailable.c_str())));
+ }
+ }
+ }
+ }
+ GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+
+ std::vector<int> userGpuTaskAssignment;
+ try
+ {
+ userGpuTaskAssignment = parseUserGpuIds(hw_opt.userGpuTaskAssignment);
+ }
+ GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+ auto nonbondedTarget = findTaskTarget(nbpu_opt);
+ auto pmeTarget = findTaskTarget(pme_opt);
+ auto pmeFftTarget = findTaskTarget(pme_fft_opt);
+ auto bondedTarget = findTaskTarget(bonded_opt);
+ PmeRunMode pmeRunMode = PmeRunMode::None;
+
+ // Here we assume that SIMMASTER(cr) does not change even after the
+ // threads are started.
+
+ FILE *fplog = nullptr;
+ // If we are appending, we don't write log output because we need
+ // to check that the old log file matches what the checkpoint file
+ // expects. Otherwise, we should start to write log output now if
+ // there is a file ready for it.
+ if (logFileHandle != nullptr && !mdrunOptions.continuationOptions.appendFiles)
+ {
+ fplog = gmx_fio_getfp(logFileHandle);
+ }
+ gmx::LoggerOwner logOwner(buildLogger(fplog, cr));
+ gmx::MDLogger mdlog(logOwner.logger());
+
+ // TODO The thread-MPI master rank makes a working
+ // PhysicalNodeCommunicator here, but it gets rebuilt by all ranks
+ // after the threads have been launched. This works because no use
+ // is made of that communicator until after the execution paths
+ // have rejoined. But it is likely that we can improve the way
+ // this is expressed, e.g. by expressly running detection only the
+ // master rank for thread-MPI, rather than relying on the mutex
+ // and reference count.
+ PhysicalNodeCommunicator physicalNodeComm(MPI_COMM_WORLD, gmx_physicalnode_id_hash());
+ hwinfo = gmx_detect_hardware(mdlog, physicalNodeComm);
+
+ gmx_print_detected_hardware(fplog, cr, ms, mdlog, hwinfo);
+
+ std::vector<int> gpuIdsToUse;
+ auto compatibleGpus = getCompatibleGpus(hwinfo->gpu_info);
+ if (gpuIdsAvailable.empty())
+ {
+ gpuIdsToUse = compatibleGpus;
+ }
+ else
+ {
+ for (const auto &availableGpuId : gpuIdsAvailable)
+ {
+ bool availableGpuIsCompatible = false;
+ for (const auto &compatibleGpuId : compatibleGpus)
+ {
+ if (availableGpuId == compatibleGpuId)
+ {
+ availableGpuIsCompatible = true;
+ break;
+ }
+ }
+ if (!availableGpuIsCompatible)
+ {
+ gmx_fatal(FARGS, "You limited the set of compatible GPUs to a set that included ID #%d, but that ID is not for a compatible GPU. List only compatible GPUs.", availableGpuId);
+ }
+ gpuIdsToUse.push_back(availableGpuId);
+ }
+ }
+
+ if (fplog != nullptr)
+ {
+ /* 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");
+ writeSourceDoi(fplog);
+ }
+
+ std::unique_ptr<t_state> globalState;
+
+ if (SIMMASTER(cr))
+ {
+ /* Only the master rank has the global state */
+ globalState = compat::make_unique<t_state>();
+
+ /* Read (nearly) all data required for the simulation */
+ read_tpx_state(ftp2fn(efTPR, filenames.size(), filenames.data()), inputrec, globalState.get(), &mtop);
+
+ /* In rerun, set velocities to zero if present */
+ if (doRerun && ((globalState->flags & (1 << estV)) != 0))
+ {
+ // rerun does not use velocities
+ GMX_LOG(mdlog.info).asParagraph().appendText(
+ "Rerun trajectory contains velocities. Rerun does only evaluate "
+ "potential energy and forces. The velocities will be ignored.");
+ for (int i = 0; i < globalState->natoms; i++)
+ {
+ clear_rvec(globalState->v[i]);
+ }
+ globalState->flags &= ~(1 << estV);
+ }
+
+ if (inputrec->cutoff_scheme != ecutsVERLET)
+ {
+ if (nstlist_cmdline > 0)
+ {
+ gmx_fatal(FARGS, "Can not set nstlist with the group cut-off scheme");
+ }
+
+ if (!compatibleGpus.empty())
+ {
+ GMX_LOG(mdlog.warning).asParagraph().appendText(
+ "NOTE: GPU(s) found, but the current simulation can not use GPUs\n"
+ " To use a GPU, set the mdp option: cutoff-scheme = Verlet");
+ }
+ }
+ }
+
+ /* Check and update the hardware options for internal consistency */
+ check_and_update_hw_opt_1(mdlog, &hw_opt, cr, domdecOptions.numPmeRanks);
+
+ /* Early check for externally set process affinity. */
+ gmx_check_thread_affinity_set(mdlog, cr,
+ &hw_opt, hwinfo->nthreads_hw_avail, FALSE);
+
+ if (GMX_THREAD_MPI && SIMMASTER(cr))
+ {
+ if (domdecOptions.numPmeRanks > 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);
+
+ bool useGpuForNonbonded = false;
+ bool useGpuForPme = false;
+ try
+ {
+ // If the user specified the number of ranks, then we must
+ // respect that, but in default mode, we need to allow for
+ // the number of GPUs to choose the number of ranks.
+ auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
+ useGpuForNonbonded = decideWhetherToUseGpusForNonbondedWithThreadMpi
+ (nonbondedTarget, gpuIdsToUse, userGpuTaskAssignment, emulateGpuNonbonded,
+ canUseGpuForNonbonded,
+ inputrec->cutoff_scheme == ecutsVERLET,
+ gpuAccelerationOfNonbondedIsUseful(mdlog, inputrec, GMX_THREAD_MPI),
+ hw_opt.nthreads_tmpi);
+ useGpuForPme = decideWhetherToUseGpusForPmeWithThreadMpi
+ (useGpuForNonbonded, pmeTarget, gpuIdsToUse, userGpuTaskAssignment,
+ *hwinfo, *inputrec, mtop, hw_opt.nthreads_tmpi, domdecOptions.numPmeRanks);
+
+ }
+ GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+
+ /* Determine how many thread-MPI ranks to start.
+ *
+ * TODO Over-writing the user-supplied value here does
+ * prevent any possible subsequent checks from working
+ * correctly. */
+ hw_opt.nthreads_tmpi = get_nthreads_mpi(hwinfo,
+ &hw_opt,
+ gpuIdsToUse,
+ useGpuForNonbonded,
+ useGpuForPme,
+ inputrec, &mtop,
+ mdlog,
+ doMembed);
+
+ // Now start the threads for thread MPI.
+ cr = spawnThreads(hw_opt.nthreads_tmpi);
+ /* The main thread continues here with a new cr. We don't deallocate
+ the old cr because other threads may still be reading it. */
+ // TODO Both master and spawned threads call dup_tfn and
+ // reinitialize_commrec_for_this_thread. Find a way to express
+ // this better.
+ physicalNodeComm = PhysicalNodeCommunicator(MPI_COMM_WORLD, gmx_physicalnode_id_hash());
+ }
+ // END OF CAUTION: cr and physicalNodeComm are now reliable
+
+ if (PAR(cr))
+ {
+ /* now broadcast everything to the non-master nodes/threads: */
+ init_parallel(cr, inputrec, &mtop);
+ }
+
+ // Now each rank knows the inputrec that SIMMASTER read and used,
+ // and (if applicable) cr->nnodes has been assigned the number of
+ // thread-MPI ranks that have been chosen. The ranks can now all
+ // run the task-deciding functions and will agree on the result
+ // without needing to communicate.
+ //
+ // TODO Should we do the communication in debug mode to support
+ // having an assertion?
+ //
+ // Note that these variables describe only their own node.
+ //
+ // Note that when bonded interactions run on a GPU they always run
+ // alongside a nonbonded task, so do not influence task assignment
+ // even though they affect the force calculation workload.
+ bool useGpuForNonbonded = false;
+ bool useGpuForPme = false;
+ bool useGpuForBonded = false;
+ try
+ {
+ // It's possible that there are different numbers of GPUs on
+ // different nodes, which is the user's responsibilty to
+ // handle. If unsuitable, we will notice that during task
+ // assignment.
+ bool gpusWereDetected = hwinfo->ngpu_compatible_tot > 0;
+ bool usingVerletScheme = inputrec->cutoff_scheme == ecutsVERLET;
+ auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
+ useGpuForNonbonded = decideWhetherToUseGpusForNonbonded(nonbondedTarget, userGpuTaskAssignment,
+ emulateGpuNonbonded,
+ canUseGpuForNonbonded,
+ usingVerletScheme,
+ gpuAccelerationOfNonbondedIsUseful(mdlog, inputrec, !GMX_THREAD_MPI),
+ gpusWereDetected);
+ useGpuForPme = decideWhetherToUseGpusForPme(useGpuForNonbonded, pmeTarget, userGpuTaskAssignment,
+ *hwinfo, *inputrec, mtop,
+ cr->nnodes, domdecOptions.numPmeRanks,
+ gpusWereDetected);
+ auto canUseGpuForBonded = buildSupportsGpuBondeds(nullptr) && inputSupportsGpuBondeds(*inputrec, mtop, nullptr);
+ useGpuForBonded =
+ decideWhetherToUseGpusForBonded(useGpuForNonbonded, useGpuForPme, usingVerletScheme,
+ bondedTarget, canUseGpuForBonded,
+ EVDW_PME(inputrec->vdwtype),
+ EEL_PME_EWALD(inputrec->coulombtype),
+ domdecOptions.numPmeRanks, gpusWereDetected);
+
+ pmeRunMode = (useGpuForPme ? PmeRunMode::GPU : PmeRunMode::CPU);
+ if (pmeRunMode == PmeRunMode::GPU)
+ {
+ if (pmeFftTarget == TaskTarget::Cpu)
+ {
+ pmeRunMode = PmeRunMode::Mixed;
+ }
+ }
+ else if (pmeFftTarget == TaskTarget::Gpu)
+ {
+ gmx_fatal(FARGS, "Assigning FFTs to GPU requires PME to be assigned to GPU as well. With PME on CPU you should not be using -pmefft.");
+ }
+ }
+ GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+
+ // Build restraints.
+ // TODO: hide restraint implementation details from Mdrunner.
+ // There is nothing unique about restraints at this point as far as the
+ // Mdrunner is concerned. The Mdrunner should just be getting a sequence of
+ // factory functions from the SimulationContext on which to call mdModules->add().
+ // TODO: capture all restraints into a single RestraintModule, passed to the runner builder.
+ for (auto && restraint : restraintManager_->getRestraints())
+ {
+ auto module = RestraintMDModule::create(restraint,
+ restraint->sites());
+ mdModules->add(std::move(module));
+ }
+
+ // TODO: Error handling
+ mdModules->assignOptionsToModules(*inputrec->params, nullptr);
+
+ if (fplog != nullptr)
+ {
+ pr_inputrec(fplog, 0, "Input Parameters", inputrec, FALSE);
+ fprintf(fplog, "\n");
+ }
+
+ if (SIMMASTER(cr))
+ {
+ /* now make sure the state is initialized and propagated */
+ set_state_entries(globalState.get(), inputrec);
+ }
+
+ /* NM and TPI parallelize over force/energy calculations, not atoms,
+ * so we need to initialize and broadcast the global state.
+ */
+ if (inputrec->eI == eiNM || inputrec->eI == eiTPI)
+ {
+ if (!MASTER(cr))
+ {
+ globalState = compat::make_unique<t_state>();
+ }
+ broadcastStateWithoutDynamics(cr, globalState.get());
+ }
+
+ /* A parallel command line option consistency check that we can
+ only do after any threads have started. */
+ if (!PAR(cr) && (domdecOptions.numCells[XX] > 1 ||
+ domdecOptions.numCells[YY] > 1 ||
+ domdecOptions.numCells[ZZ] > 1 ||
+ domdecOptions.numPmeRanks > 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", output_env_get_program_display_name(oenv));
+#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", output_env_get_program_display_name(oenv));
+#endif
+#endif
+ }
+
+ if (doRerun &&
+ (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 (domdecOptions.numPmeRanks > 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");
+ }
+
+ domdecOptions.numPmeRanks = 0;
+ }
+
+ if (useGpuForNonbonded && domdecOptions.numPmeRanks < 0)
+ {
+ /* With NB GPUs we don't automatically use PME-only CPU 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.
+ */
+ domdecOptions.numPmeRanks = 0;
+ }
+ if (useGpuForPme)
+ {
+ if (domdecOptions.numPmeRanks < 0)
+ {
+ domdecOptions.numPmeRanks = 0;
+ // TODO possibly print a note that one can opt-in for a separate PME GPU rank?
+ }
+ else
+ {
+ GMX_RELEASE_ASSERT(domdecOptions.numPmeRanks <= 1, "PME GPU decomposition is not supported");
+ }
+ }
+
+#if 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, ms, fcd, globalState.get(), replExParams.exchangeInterval > 0);
+
+ init_orires(fplog, &mtop, inputrec, cr, ms, globalState.get(), &(fcd->orires));
+
+ auto deform = prepareBoxDeformation(globalState->box, cr, *inputrec);
+
+ ObservablesHistory observablesHistory = {};
+
+ ContinuationOptions &continuationOptions = mdrunOptions.continuationOptions;
+
+ if (continuationOptions.startedFromCheckpoint)
+ {
+ /* 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", filenames.size(), filenames.data(), cr),
+ logFileHandle,
+ cr, domdecOptions.numCells,
+ inputrec, globalState.get(),
+ &bReadEkin, &observablesHistory,
+ continuationOptions.appendFiles,
+ continuationOptions.appendFilesOptionSet,
+ mdrunOptions.reproducible);
+
+ if (bReadEkin)
+ {
+ continuationOptions.haveReadEkin = true;
+ }
+
+ if (continuationOptions.appendFiles && logFileHandle)
+ {
+ // Now we can start normal logging to the truncated log file.
+ fplog = gmx_fio_getfp(logFileHandle);
+ prepareLogAppending(fplog);
+ logOwner = buildLogger(fplog, cr);
+ mdlog = logOwner.logger();
+ }
+ }
+
+ if (mdrunOptions.numStepsCommandline > -2)
+ {
+ GMX_LOG(mdlog.info).asParagraph().
+ appendText("The -nsteps functionality is deprecated, and may be removed in a future version. "
+ "Consider using gmx convert-tpr -nsteps or changing the appropriate .mdp file field.");
+ }
+ /* override nsteps with value set on the commamdline */
+ override_nsteps_cmdline(mdlog, mdrunOptions.numStepsCommandline, inputrec);
+
+ if (SIMMASTER(cr))
+ {
+ copy_mat(globalState->box, box);
+ }
+
+ if (PAR(cr))
+ {
+ gmx_bcast(sizeof(box), box, cr);
+ }
+
+ /* Update rlist and nstlist. */
+ if (inputrec->cutoff_scheme == ecutsVERLET)
+ {
+ prepare_verlet_scheme(fplog, cr, inputrec, nstlist_cmdline, &mtop, box,
+ useGpuForNonbonded || (emulateGpuNonbonded == EmulateGpuNonbonded::Yes), *hwinfo->cpuInfo);
+ }
+
+ LocalAtomSetManager atomSets;
+
+ if (PAR(cr) && !(EI_TPI(inputrec->eI) ||
+ inputrec->eI == eiNM))
+ {
+ cr->dd = init_domain_decomposition(mdlog, cr, domdecOptions, mdrunOptions,
+ &mtop, inputrec,
+ box, positionsFromStatePointer(globalState.get()),
+ &atomSets);
+ // Note that local state still does not exist yet.
+ }
+ else
+ {
+ /* PME, if used, is done on all nodes with 1D decomposition */
+ cr->npmenodes = 0;
+ cr->duty = (DUTY_PP | DUTY_PME);
+
+ if (inputrec->ePBC == epbcSCREW)
+ {
+ gmx_fatal(FARGS,
+ "pbc=screw is only implemented with domain decomposition");
+ }
+ }
+
+ 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);
+ }
+
+#if GMX_MPI
+ if (isMultiSim(ms))
+ {
+ GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
+ "This is simulation %d out of %d running as a composite GROMACS\n"
+ "multi-simulation job. Setup for this simulation:\n",
+ ms->sim, ms->nsim);
+ }
+ GMX_LOG(mdlog.warning).appendTextFormatted(
+ "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 the number of OpenMP threads requested */
+ checkAndUpdateRequestedNumOpenmpThreads(&hw_opt, *hwinfo, cr, ms, physicalNodeComm.size_,
+ pmeRunMode, mtop);
+
+ gmx_omp_nthreads_init(mdlog, cr,
+ hwinfo->nthreads_hw_avail,
+ physicalNodeComm.size_,
+ hw_opt.nthreads_omp,
+ hw_opt.nthreads_omp_pme,
+ !thisRankHasDuty(cr, DUTY_PP),
+ inputrec->cutoff_scheme == ecutsVERLET);
+
+ // Enable FP exception detection for the Verlet scheme, but not in
+ // Release mode and not for compilers with known buggy FP
+ // exception support (clang with any optimization) or suspected
+ // buggy FP exception support (gcc 7.* with optimization).
+#if !defined NDEBUG && \
+ !((defined __clang__ || (defined(__GNUC__) && !defined(__ICC) && __GNUC__ == 7)) \
+ && defined __OPTIMIZE__)
+ const bool bEnableFPE = inputrec->cutoff_scheme == ecutsVERLET;
+#else
+ const bool bEnableFPE = false;
+#endif
+ //FIXME - reconcile with gmx_feenableexcept() call from CommandLineModuleManager::run()
+ if (bEnableFPE)
+ {
+ gmx_feenableexcept();
+ }
+
+ // Build a data structure that expresses which kinds of non-bonded
+ // task are handled by this rank.
+ //
+ // TODO Later, this might become a loop over all registered modules
+ // relevant to the mdp inputs, to find those that have such tasks.
+ //
+ // TODO This could move before init_domain_decomposition() as part
+ // of refactoring that separates the responsibility for duty
+ // assignment from setup for communication between tasks, and
+ // setup for tasks handled with a domain (ie including short-ranged
+ // tasks, bonded tasks, etc.).
+ //
+ // Note that in general useGpuForNonbonded, etc. can have a value
+ // that is inconsistent with the presence of actual GPUs on any
+ // rank, and that is not known to be a problem until the
+ // duty of the ranks on a node become known.
+ //
+ // TODO Later we might need the concept of computeTasksOnThisRank,
+ // from which we construct gpuTasksOnThisRank.
+ //
+ // Currently the DD code assigns duty to ranks that can
+ // include PP work that currently can be executed on a single
+ // GPU, if present and compatible. This has to be coordinated
+ // across PP ranks on a node, with possible multiple devices
+ // or sharing devices on a node, either from the user
+ // selection, or automatically.
+ auto haveGpus = !gpuIdsToUse.empty();
+ std::vector<GpuTask> gpuTasksOnThisRank;
+ if (thisRankHasDuty(cr, DUTY_PP))
+ {
+ if (useGpuForNonbonded)
+ {
+ // Note that any bonded tasks on a GPU always accompany a
+ // non-bonded task.
+ if (haveGpus)
+ {
+ gpuTasksOnThisRank.push_back(GpuTask::Nonbonded);
+ }
+ else if (nonbondedTarget == TaskTarget::Gpu)
+ {
+ gmx_fatal(FARGS, "Cannot run short-ranged nonbonded interactions on a GPU because there is none detected.");
+ }
+ else if (bondedTarget == TaskTarget::Gpu)
+ {
+ gmx_fatal(FARGS, "Cannot run bonded interactions on a GPU because there is none detected.");
+ }
+ }
+ }
+ // TODO cr->duty & DUTY_PME should imply that a PME algorithm is active, but currently does not.
+ if (EEL_PME(inputrec->coulombtype) && (thisRankHasDuty(cr, DUTY_PME)))
+ {
+ if (useGpuForPme)
+ {
+ if (haveGpus)
+ {
+ gpuTasksOnThisRank.push_back(GpuTask::Pme);
+ }
+ else if (pmeTarget == TaskTarget::Gpu)
+ {
+ gmx_fatal(FARGS, "Cannot run PME on a GPU because there is none detected.");
+ }
+ }
+ }
+
+ GpuTaskAssignment gpuTaskAssignment;
+ try
+ {
+ // Produce the task assignment for this rank.
+ gpuTaskAssignment = runTaskAssignment(gpuIdsToUse, userGpuTaskAssignment, *hwinfo,
+ mdlog, cr, ms, physicalNodeComm, gpuTasksOnThisRank,
+ useGpuForBonded, pmeRunMode);
+ }
+ GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+
+ /* Prevent other ranks from continuing after an issue was found
+ * and reported as a fatal error.
+ *
+ * TODO This function implements a barrier so that MPI runtimes
+ * can organize an orderly shutdown if one of the ranks has had to
+ * issue a fatal error in various code already run. When we have
+ * MPI-aware error handling and reporting, this should be
+ * improved. */
+#if GMX_MPI
+ if (PAR(cr))
+ {
+ MPI_Barrier(cr->mpi_comm_mysim);
+ }
+ if (isMultiSim(ms))
+ {
+ if (SIMMASTER(cr))
+ {
+ MPI_Barrier(ms->mpi_comm_masters);
+ }
+ /* We need another barrier to prevent non-master ranks from contiuing
+ * when an error occured in a different simulation.
+ */
+ MPI_Barrier(cr->mpi_comm_mysim);
+ }
+#endif
+
+ /* Now that we know the setup is consistent, check for efficiency */
+ check_resource_division_efficiency(hwinfo, !gpuTaskAssignment.empty(), mdrunOptions.ntompOptionIsSet,
+ cr, mdlog);
+
+ gmx_device_info_t *nonbondedDeviceInfo = nullptr;
+
+ if (thisRankHasDuty(cr, DUTY_PP))
+ {
+ // This works because only one task of each type is currently permitted.
+ auto nbGpuTaskMapping = std::find_if(gpuTaskAssignment.begin(), gpuTaskAssignment.end(),
+ hasTaskType<GpuTask::Nonbonded>);
+ if (nbGpuTaskMapping != gpuTaskAssignment.end())
+ {
+ int nonbondedDeviceId = nbGpuTaskMapping->deviceId_;
+ nonbondedDeviceInfo = getDeviceInfo(hwinfo->gpu_info, nonbondedDeviceId);
+ init_gpu(nonbondedDeviceInfo);
+
+ if (DOMAINDECOMP(cr))
+ {
+ /* When we share GPUs over ranks, we need to know this for the DLB */
+ dd_setup_dlb_resource_sharing(cr, nonbondedDeviceId);
+ }
+
+ }
+ }
+
+ std::unique_ptr<ClfftInitializer> initializedClfftLibrary;
+
+ gmx_device_info_t *pmeDeviceInfo = nullptr;
+ // Later, this program could contain kernels that might be later
+ // re-used as auto-tuning progresses, or subsequent simulations
+ // are invoked.
+ PmeGpuProgramStorage pmeGpuProgram;
+ // This works because only one task of each type is currently permitted.
+ auto pmeGpuTaskMapping = std::find_if(gpuTaskAssignment.begin(), gpuTaskAssignment.end(), hasTaskType<GpuTask::Pme>);
+ const bool thisRankHasPmeGpuTask = (pmeGpuTaskMapping != gpuTaskAssignment.end());
+ if (thisRankHasPmeGpuTask)
+ {
+ pmeDeviceInfo = getDeviceInfo(hwinfo->gpu_info, pmeGpuTaskMapping->deviceId_);
+ init_gpu(pmeDeviceInfo);
+ pmeGpuProgram = buildPmeGpuProgram(pmeDeviceInfo);
+ // TODO It would be nice to move this logic into the factory
+ // function. See Redmine #2535.
+ bool isMasterThread = !GMX_THREAD_MPI || MASTER(cr);
+ if (pmeRunMode == PmeRunMode::GPU && !initializedClfftLibrary && isMasterThread)
+ {
+ initializedClfftLibrary = initializeClfftLibrary();
+ }
+ }
+
+ /* 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);
+
+ int numThreadsOnThisRank;
+ /* threads on this MPI process or TMPI thread */
+ if (thisRankHasDuty(cr, DUTY_PP))
+ {
+ numThreadsOnThisRank = gmx_omp_nthreads_get(emntNonbonded);
+ }
+ else
+ {
+ numThreadsOnThisRank = nthreads_pme;
+ }
+
+ checkHardwareOversubscription(numThreadsOnThisRank, cr->nodeid,
+ *hwinfo->hardwareTopology,
+ physicalNodeComm, mdlog);
+
+ 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(mdlog, cr,
+ &hw_opt, hwinfo->nthreads_hw_avail, TRUE);
+
+ int numThreadsOnThisNode, intraNodeThreadOffset;
+ analyzeThreadsOnThisNode(physicalNodeComm, numThreadsOnThisRank, &numThreadsOnThisNode,
+ &intraNodeThreadOffset);
+
+ /* Set the CPU affinity */
+ gmx_set_thread_affinity(mdlog, cr, &hw_opt, *hwinfo->hardwareTopology,
+ numThreadsOnThisRank, numThreadsOnThisNode,
+ intraNodeThreadOffset, nullptr);
+ }
+
+ if (mdrunOptions.timingOptions.resetStep > -1)
+ {
+ GMX_LOG(mdlog.info).asParagraph().
+ appendText("The -resetstep functionality is deprecated, and may be removed in a future version.");
+ }
+ wcycle = wallcycle_init(fplog, mdrunOptions.timingOptions.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);
+ }
+
+ // Membrane embedding must be initialized before we call init_forcerec()
+ if (doMembed)
+ {
+ 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, filenames.size(), filenames.data(), &mtop, inputrec, globalState.get(), cr,
+ &mdrunOptions
+ .checkpointOptions.period);
+ }
+
+ std::unique_ptr<MDAtoms> mdAtoms;
+ std::unique_ptr<gmx_vsite_t> vsite;
+
+ snew(nrnb, 1);
+ if (thisRankHasDuty(cr, DUTY_PP))
+ {
+ /* Initiate forcerecord */
+ fr = mk_forcerec();
+ fr->forceProviders = mdModules->initForceProviders();
+ init_forcerec(fplog, mdlog, fr, fcd,
+ inputrec, &mtop, cr, box,
+ opt2fn("-table", filenames.size(), filenames.data()),
+ opt2fn("-tablep", filenames.size(), filenames.data()),
+ opt2fns("-tableb", filenames.size(), filenames.data()),
+ *hwinfo, nonbondedDeviceInfo,
+ useGpuForBonded,
+ FALSE,
+ pforce);
+
+ /* Initialize the mdAtoms structure.
+ * mdAtoms is not filled with atom data,
+ * as this can not be done now with domain decomposition.
+ */
+ mdAtoms = makeMDAtoms(fplog, mtop, *inputrec, thisRankHasPmeGpuTask);
+ if (globalState && thisRankHasPmeGpuTask)
+ {
+ // The pinning of coordinates in the global state object works, because we only use
+ // PME on GPU without DD or on a separate PME rank, and because the local state pointer
+ // points to the global state object without DD.
+ // FIXME: MD and EM separately set up the local state - this should happen in the same function,
+ // which should also perform the pinning.
+ changePinningPolicy(&globalState->x, pme_get_pinning_policy());
+ }
+
+ /* Initialize the virtual site communication */
+ vsite = initVsite(mtop, cr);
+
+ calc_shifts(box, fr->shift_vec);
+
+ /* With periodic molecules the charge groups should be whole at start up
+ * and the virtual sites should not be far from their proper positions.
+ */
+ if (!inputrec->bContinuation && MASTER(cr) &&
+ !(inputrec->ePBC != epbcNONE && inputrec->bPeriodicMols))
+ {
+ /* Make molecules whole at start of run */
+ if (fr->ePBC != epbcNONE)
+ {
+ do_pbc_first_mtop(fplog, inputrec->ePBC, box, &mtop, globalState->x.rvec_array());
+ }
+ if (vsite)
+ {
+ /* Correct initial vsite positions are required
+ * for the initial distribution in the domain decomposition
+ * and for the initial shell prediction.
+ */
+ constructVsitesGlobal(mtop, globalState->x);
+ }
+ }
+
+ if (EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
+ {
+ ewaldcoeff_q = fr->ic->ewaldcoeff_q;
+ ewaldcoeff_lj = fr->ic->ewaldcoeff_lj;
+ }
+ }
+ else
+ {
+ /* This is a PME only node */
+
+ GMX_ASSERT(globalState == nullptr, "We don't need the state on a PME only rank and expect it to be unitialized");
+
+ ewaldcoeff_q = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
+ ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
+ }
+
+ gmx_pme_t *sepPmeData = nullptr;
+ // This reference hides the fact that PME data is owned by runner on PME-only ranks and by forcerec on other ranks
+ GMX_ASSERT(thisRankHasDuty(cr, DUTY_PP) == (fr != nullptr), "Double-checking that only PME-only ranks have no forcerec");
+ gmx_pme_t * &pmedata = fr ? fr->pmedata : sepPmeData;
+
+ /* 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 && mdAtoms->mdatoms())
+ {
+ nChargePerturbed = mdAtoms->mdatoms()->nChargePerturbed;
+ if (EVDW_PME(inputrec->vdwtype))
+ {
+ nTypePerturbed = mdAtoms->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 (thisRankHasDuty(cr, DUTY_PME))
+ {
+ try
+ {
+ pmedata = gmx_pme_init(cr,
+ getNumPmeDomains(cr->dd),
+ inputrec,
+ mtop.natoms, nChargePerturbed != 0, nTypePerturbed != 0,
+ mdrunOptions.reproducible,
+ ewaldcoeff_q, ewaldcoeff_lj,
+ nthreads_pme,
+ pmeRunMode, nullptr,
+ pmeDeviceInfo, pmeGpuProgram.get(), mdlog);
+ }
+ GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+ }
+ }
+
+
+ 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 (thisRankHasDuty(cr, 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,
+ &mtop, cr, &atomSets, inputrec->fepvals->init_lambda);
+ if (inputrec->pull->bXOutAverage || inputrec->pull->bFOutAverage)
+ {
+ initPullHistory(inputrec->pull_work, &observablesHistory);
+ }
+ if (EI_DYNAMICS(inputrec->eI) && MASTER(cr))
+ {
+ init_pull_output_files(inputrec->pull_work,
+ filenames.size(), filenames.data(), oenv,
+ continuationOptions);
+ }
+ }
+
+ std::unique_ptr<EnforcedRotation> enforcedRotation;
+ if (inputrec->bRot)
+ {
+ /* Initialize enforced rotation code */
+ enforcedRotation = init_rot(fplog,
+ inputrec,
+ filenames.size(),
+ filenames.data(),
+ cr,
+ &atomSets,
+ globalState.get(),
+ &mtop,
+ oenv,
+ mdrunOptions);
+ }
+
+ if (inputrec->eSwapCoords != eswapNO)
+ {
+ /* Initialize ion swapping code */
+ init_swapcoords(fplog, inputrec, opt2fn_master("-swap", filenames.size(), filenames.data(), cr),
+ &mtop, globalState.get(), &observablesHistory,
+ cr, &atomSets, oenv, mdrunOptions);
+ }
+
+ /* Let makeConstraints know whether we have essential dynamics constraints.
+ * TODO: inputrec should tell us whether we use an algorithm, not a file option or the checkpoint
+ */
+ bool doEssentialDynamics = (opt2fn_null("-ei", filenames.size(), filenames.data()) != nullptr
+ || observablesHistory.edsamHistory);
+ auto constr = makeConstraints(mtop, *inputrec, doEssentialDynamics,
+ fplog, *mdAtoms->mdatoms(),
+ cr, ms, nrnb, wcycle, fr->bMolPBC);
+
+ 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.get(), inputrec,
+ domdecOptions.checkBondedInteractions,
+ fr->cginfo_mb);
+ }
+
+ // TODO This is not the right place to manage the lifetime of
+ // this data structure, but currently it's the easiest way to
+ // make it work. Later, it should probably be made/updated
+ // after the workload for the lifetime of a PP domain is
+ // understood.
+ PpForceWorkload ppForceWorkload;
+
+ GMX_ASSERT(stopHandlerBuilder_, "Runner must provide StopHandlerBuilder to integrator.");
+
+ /* PLUMED */
+ if(plumedswitch){
+ /* detect plumed API version */
+ int pversion=0;
+ plumed_cmd(plumedmain,"getApiVersion",&pversion);
+ if(pversion>5) {
+ int nth = gmx_omp_nthreads_get(emntDefault);
+ if(pversion>5) plumed_cmd(plumedmain,"setNumOMPthreads",&nth);
+ }
+ }
+ /* END PLUMED */
+
+ /* Now do whatever the user wants us to do (how flexible...) */
+ Integrator integrator {
+ fplog, cr, ms, mdlog, static_cast<int>(filenames.size()), filenames.data(),
+ oenv,
+ mdrunOptions,
+ vsite.get(), constr.get(),
+ enforcedRotation ? enforcedRotation->getLegacyEnfrot() : nullptr,
+ deform.get(),
+ mdModules->outputProvider(),
+ inputrec, &mtop,
+ fcd,
+ globalState.get(),
+ &observablesHistory,
+ mdAtoms.get(), nrnb, wcycle, fr,
+ &ppForceWorkload,
+ replExParams,
+ membed,
+ walltime_accounting,
+ std::move(stopHandlerBuilder_)
+ };
+ integrator.run(inputrec->eI, doRerun);
+
+ 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, inputrec, pmeRunMode);
+ }
+
+ wallcycle_stop(wcycle, ewcRUN);
+
+ /* Finish up, write some stuff
+ * if rerunMD, don't write last frame again
+ */
+ finish_run(fplog, mdlog, cr,
+ inputrec, nrnb, wcycle, walltime_accounting,
+ fr ? fr->nbv : nullptr,
+ pmedata,
+ EI_DYNAMICS(inputrec->eI) && !isMultiSim(ms));
+
+ // Free PME data
+ if (pmedata)
+ {
+ gmx_pme_destroy(pmedata);
+ pmedata = nullptr;
+ }
+
+ // FIXME: this is only here to manually unpin mdAtoms->chargeA_ and state->x,
+ // before we destroy the GPU context(s) in free_gpu_resources().
+ // Pinned buffers are associated with contexts in CUDA.
+ // As soon as we destroy GPU contexts after mdrunner() exits, these lines should go.
+ mdAtoms.reset(nullptr);
+ globalState.reset(nullptr);
+ mdModules.reset(nullptr); // destruct force providers here as they might also use the GPU
+
+ /* Free GPU memory and set a physical node tMPI barrier (which should eventually go away) */
+ free_gpu_resources(fr, physicalNodeComm);
+ free_gpu(nonbondedDeviceInfo);
+ free_gpu(pmeDeviceInfo);
+ done_forcerec(fr, mtop.molblock.size(), mtop.groups.grps[egcENER].nr);
+ sfree(fcd);
+
+ if (doMembed)
+ {
+ free_membed(membed);
+ }
+
+ gmx_hardware_info_free();
+
+ /* Does what it says */
+ print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
+ walltime_accounting_destroy(walltime_accounting);
+ sfree(nrnb);
+
+ /* PLUMED */
+ if(plumedswitch){
+ plumed_finalize(plumedmain);
+ }
+ /* END PLUMED */
+
+ // Ensure log file content is written
+ if (logFileHandle)
+ {
+ gmx_fio_flush(logFileHandle);
+ }
+
+ /* Reset FPEs (important for unit tests) by disabling them. Assumes no
+ * exceptions were enabled before function was called. */
+ if (bEnableFPE)
+ {
+ gmx_fedisableexcept();
+ }
+
+ rc = static_cast<int>(gmx_get_stop_condition());
+
+#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))
+ {
+ done_commrec(cr);
+ tMPI_Finalize();
+ }
+#endif
+
+ return rc;
+}
+
+Mdrunner::~Mdrunner()
+{
+ // Clean up of the Manager.
+ // This will end up getting called on every thread-MPI rank, which is unnecessary,
+ // but okay as long as threads synchronize some time before adding or accessing
+ // a new set of restraints.
+ if (restraintManager_)
+ {
+ restraintManager_->clear();
+ GMX_ASSERT(restraintManager_->countRestraints() == 0,
+ "restraints added during runner life time should be cleared at runner destruction.");
+ }
+};
+
+void Mdrunner::addPotential(std::shared_ptr<gmx::IRestraintPotential> puller,
+ std::string name)
+{
+ GMX_ASSERT(restraintManager_, "Mdrunner must have a restraint manager.");
+ // Not sure if this should be logged through the md logger or something else,
+ // but it is helpful to have some sort of INFO level message sent somewhere.
+ // std::cout << "Registering restraint named " << name << std::endl;
+
+ // When multiple restraints are used, it may be wasteful to register them separately.
+ // Maybe instead register an entire Restraint Manager as a force provider.
+ restraintManager_->addToSpec(std::move(puller),
+ std::move(name));
+}
+
+Mdrunner::Mdrunner(Mdrunner &&) noexcept = default;
+
+//NOLINTNEXTLINE(performance-noexcept-move-constructor) working around GCC bug 58265
+Mdrunner &Mdrunner::operator=(Mdrunner && /*handle*/) noexcept(BUGFREE_NOEXCEPT_STRING) = default;
+
+class Mdrunner::BuilderImplementation
+{
+ public:
+ BuilderImplementation() = delete;
+ explicit BuilderImplementation(SimulationContext* context);
+ ~BuilderImplementation();
+
+ BuilderImplementation &setExtraMdrunOptions(const MdrunOptions &options,
+ real forceWarningThreshold);
+
+ void addDomdec(const DomdecOptions &options);
+
+ void addVerletList(int nstlist);
+
+ void addReplicaExchange(const ReplicaExchangeParameters ¶ms);
+
+ void addMultiSim(gmx_multisim_t* multisim);
+
+ void addNonBonded(const char* nbpu_opt);
+
+ void addPME(const char* pme_opt_, const char* pme_fft_opt_);
+
+ void addBondedTaskAssignment(const char* bonded_opt);
+
+ void addHardwareOptions(const gmx_hw_opt_t &hardwareOptions);
+
+ void addFilenames(ArrayRef <const t_filenm> filenames);
+
+ void addOutputEnvironment(gmx_output_env_t* outputEnvironment);
+
+ void addLogFile(t_fileio *logFileHandle);
+
+ void addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder);
+
+ Mdrunner build();
+
+ private:
+ // Default parameters copied from runner.h
+ // \todo Clarify source(s) of default parameters.
+
+ const char* nbpu_opt_ = nullptr;
+ const char* pme_opt_ = nullptr;
+ const char* pme_fft_opt_ = nullptr;
+ const char *bonded_opt_ = nullptr;
+
+ MdrunOptions mdrunOptions_;
+
+ DomdecOptions domdecOptions_;
+
+ ReplicaExchangeParameters replicaExchangeParameters_;
+
+ //! Command-line override for the duration of a neighbor list with the Verlet scheme.
+ int nstlist_ = 0;
+
+ //! Non-owning multisim communicator handle.
+ std::unique_ptr<gmx_multisim_t*> multisim_ = nullptr;
+
+ //! Print a warning if any force is larger than this (in kJ/mol nm).
+ real forceWarningThreshold_ = -1;
+
+ /*! \brief Non-owning pointer to SimulationContext (owned and managed by client)
+ *
+ * \internal
+ * \todo Establish robust protocol to make sure resources remain valid.
+ * SimulationContext will likely be separated into multiple layers for
+ * different levels of access and different phases of execution. Ref
+ * https://redmine.gromacs.org/issues/2375
+ * https://redmine.gromacs.org/issues/2587
+ */
+ SimulationContext* context_ = nullptr;
+
+ //! \brief Parallelism information.
+ gmx_hw_opt_t hardwareOptions_;
+
+ //! filename options for simulation.
+ ArrayRef<const t_filenm> filenames_;
+
+ /*! \brief Handle to output environment.
+ *
+ * \todo gmx_output_env_t needs lifetime management.
+ */
+ gmx_output_env_t* outputEnvironment_ = nullptr;
+
+ /*! \brief Non-owning handle to MD log file.
+ *
+ * \todo Context should own output facilities for client.
+ * \todo Improve log file handle management.
+ * \internal
+ * Code managing the FILE* relies on the ability to set it to
+ * nullptr to check whether the filehandle is valid.
+ */
+ t_fileio* logFileHandle_ = nullptr;
+
+ /*!
+ * \brief Builder for simulation stop signal handler.
+ */
+ std::unique_ptr<StopHandlerBuilder> stopHandlerBuilder_ = nullptr;
+};
+
+Mdrunner::BuilderImplementation::BuilderImplementation(SimulationContext* context) :
+ context_(context)
+{
+ GMX_ASSERT(context_, "Bug found. It should not be possible to construct builder without a valid context.");
+}
+
+Mdrunner::BuilderImplementation::~BuilderImplementation() = default;
+
+Mdrunner::BuilderImplementation &
+Mdrunner::BuilderImplementation::setExtraMdrunOptions(const MdrunOptions &options,
+ real forceWarningThreshold)
+{
+ mdrunOptions_ = options;
+ forceWarningThreshold_ = forceWarningThreshold;
+ return *this;
+}
+
+void Mdrunner::BuilderImplementation::addDomdec(const DomdecOptions &options)
+{
+ domdecOptions_ = options;
+}
+
+void Mdrunner::BuilderImplementation::addVerletList(int nstlist)
+{
+ nstlist_ = nstlist;
+}
+
+void Mdrunner::BuilderImplementation::addReplicaExchange(const ReplicaExchangeParameters ¶ms)
+{
+ replicaExchangeParameters_ = params;
+}
+
+void Mdrunner::BuilderImplementation::addMultiSim(gmx_multisim_t* multisim)
+{
+ multisim_ = compat::make_unique<gmx_multisim_t*>(multisim);
+}
+
+Mdrunner Mdrunner::BuilderImplementation::build()
+{
+ auto newRunner = Mdrunner();
+
+ GMX_ASSERT(context_, "Bug found. It should not be possible to call build() without a valid context.");
+
+ newRunner.mdrunOptions = mdrunOptions_;
+ newRunner.domdecOptions = domdecOptions_;
+
+ // \todo determine an invariant to check or confirm that all gmx_hw_opt_t objects are valid
+ newRunner.hw_opt = hardwareOptions_;
+
+ // No invariant to check. This parameter exists to optionally override other behavior.
+ newRunner.nstlist_cmdline = nstlist_;
+
+ newRunner.replExParams = replicaExchangeParameters_;
+
+ newRunner.filenames = filenames_;
+
+ GMX_ASSERT(context_->communicationRecord_, "SimulationContext communications not initialized.");
+ newRunner.cr = context_->communicationRecord_;
+
+ if (multisim_)
+ {
+ // nullptr is a valid value for the multisim handle, so we don't check the pointed-to pointer.
+ newRunner.ms = *multisim_;
+ }
+ else
+ {
+ GMX_THROW(gmx::APIError("MdrunnerBuilder::addMultiSim() is required before build()"));
+ }
+
+ // \todo Clarify ownership and lifetime management for gmx_output_env_t
+ // \todo Update sanity checking when output environment has clearly specified invariants.
+ // Initialization and default values for oenv are not well specified in the current version.
+ if (outputEnvironment_)
+ {
+ newRunner.oenv = outputEnvironment_;
+ }
+ else
+ {
+ GMX_THROW(gmx::APIError("MdrunnerBuilder::addOutputEnvironment() is required before build()"));
+ }
+
+ newRunner.logFileHandle = logFileHandle_;
+
+ if (nbpu_opt_)
+ {
+ newRunner.nbpu_opt = nbpu_opt_;
+ }
+ else
+ {
+ GMX_THROW(gmx::APIError("MdrunnerBuilder::addNonBonded() is required before build()"));
+ }
+
+ if (pme_opt_ && pme_fft_opt_)
+ {
+ newRunner.pme_opt = pme_opt_;
+ newRunner.pme_fft_opt = pme_fft_opt_;
+ }
+ else
+ {
+ GMX_THROW(gmx::APIError("MdrunnerBuilder::addElectrostatics() is required before build()"));
+ }
+
+ if (bonded_opt_)
+ {
+ newRunner.bonded_opt = bonded_opt_;
+ }
+ else
+ {
+ GMX_THROW(gmx::APIError("MdrunnerBuilder::addBondedTaskAssignment() is required before build()"));
+ }
+
+ newRunner.restraintManager_ = compat::make_unique<gmx::RestraintManager>();
+
+ if (stopHandlerBuilder_)
+ {
+ newRunner.stopHandlerBuilder_ = std::move(stopHandlerBuilder_);
+ }
+ else
+ {
+ newRunner.stopHandlerBuilder_ = compat::make_unique<StopHandlerBuilder>();
+ }
+
+ return newRunner;
+}
+
+void Mdrunner::BuilderImplementation::addNonBonded(const char* nbpu_opt)
+{
+ nbpu_opt_ = nbpu_opt;
+}
+
+void Mdrunner::BuilderImplementation::addPME(const char* pme_opt,
+ const char* pme_fft_opt)
+{
+ pme_opt_ = pme_opt;
+ pme_fft_opt_ = pme_fft_opt;
+}
+
+void Mdrunner::BuilderImplementation::addBondedTaskAssignment(const char* bonded_opt)
+{
+ bonded_opt_ = bonded_opt;
+}
+
+void Mdrunner::BuilderImplementation::addHardwareOptions(const gmx_hw_opt_t &hardwareOptions)
+{
+ hardwareOptions_ = hardwareOptions;
+}
+
+void Mdrunner::BuilderImplementation::addFilenames(ArrayRef<const t_filenm> filenames)
+{
+ filenames_ = filenames;
+}
+
+void Mdrunner::BuilderImplementation::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
+{
+ outputEnvironment_ = outputEnvironment;
+}
+
+void Mdrunner::BuilderImplementation::addLogFile(t_fileio *logFileHandle)
+{
+ logFileHandle_ = logFileHandle;
+}
+
+void Mdrunner::BuilderImplementation::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
+{
+ stopHandlerBuilder_ = std::move(builder);
+}
+
+MdrunnerBuilder::MdrunnerBuilder(compat::not_null<SimulationContext*> context) :
+ impl_ {gmx::compat::make_unique<Mdrunner::BuilderImplementation>(context)}
+{
+}
+
+MdrunnerBuilder::~MdrunnerBuilder() = default;
+
+MdrunnerBuilder &MdrunnerBuilder::addSimulationMethod(const MdrunOptions &options,
+ real forceWarningThreshold)
+{
+ impl_->setExtraMdrunOptions(options, forceWarningThreshold);
+ return *this;
+}
+
+MdrunnerBuilder &MdrunnerBuilder::addDomainDecomposition(const DomdecOptions &options)
+{
+ impl_->addDomdec(options);
+ return *this;
+}
+
+MdrunnerBuilder &MdrunnerBuilder::addNeighborList(int nstlist)
+{
+ impl_->addVerletList(nstlist);
+ return *this;
+}
+
+MdrunnerBuilder &MdrunnerBuilder::addReplicaExchange(const ReplicaExchangeParameters ¶ms)
+{
+ impl_->addReplicaExchange(params);
+ return *this;
+}
+
+MdrunnerBuilder &MdrunnerBuilder::addMultiSim(gmx_multisim_t* multisim)
+{
+ impl_->addMultiSim(multisim);
+ return *this;
+}
+
+MdrunnerBuilder &MdrunnerBuilder::addNonBonded(const char* nbpu_opt)
+{
+ impl_->addNonBonded(nbpu_opt);
+ return *this;
+}
+
+MdrunnerBuilder &MdrunnerBuilder::addElectrostatics(const char* pme_opt,
+ const char* pme_fft_opt)
+{
+ // The builder method may become more general in the future, but in this version,
+ // parameters for PME electrostatics are both required and the only parameters
+ // available.
+ if (pme_opt && pme_fft_opt)
+ {
+ impl_->addPME(pme_opt, pme_fft_opt);
+ }
+ else
+ {
+ GMX_THROW(gmx::InvalidInputError("addElectrostatics() arguments must be non-null pointers."));
+ }
+ return *this;
+}
+
+MdrunnerBuilder &MdrunnerBuilder::addBondedTaskAssignment(const char* bonded_opt)
+{
+ impl_->addBondedTaskAssignment(bonded_opt);
+ return *this;
+}
+
+Mdrunner MdrunnerBuilder::build()
+{
+ return impl_->build();
+}
+
+MdrunnerBuilder &MdrunnerBuilder::addHardwareOptions(const gmx_hw_opt_t &hardwareOptions)
+{
+ impl_->addHardwareOptions(hardwareOptions);
+ return *this;
+}
+
+MdrunnerBuilder &MdrunnerBuilder::addFilenames(ArrayRef<const t_filenm> filenames)
+{
+ impl_->addFilenames(filenames);
+ return *this;
+}
+
+MdrunnerBuilder &MdrunnerBuilder::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
+{
+ impl_->addOutputEnvironment(outputEnvironment);
+ return *this;
+}
+
+MdrunnerBuilder &MdrunnerBuilder::addLogFile(t_fileio *logFileHandle)
+{
+ impl_->addLogFile(logFileHandle);
+ return *this;
+}
+
+MdrunnerBuilder &MdrunnerBuilder::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
+{
+ impl_->addStopHandlerBuilder(std::move(builder));
+ return *this;
+}
+
+MdrunnerBuilder::MdrunnerBuilder(MdrunnerBuilder &&) noexcept = default;
+
+MdrunnerBuilder &MdrunnerBuilder::operator=(MdrunnerBuilder &&) noexcept = default;
+
+} // namespace gmx
diff --git a/patches/gromacs-2019.1.diff/src/gromacs/mdrun/runner.cpp.preplumed b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/runner.cpp.preplumed
new file mode 100644
index 0000000000000000000000000000000000000000..ae158a4984deefd1d61ca9e2a912143db1c75e05
--- /dev/null
+++ b/patches/gromacs-2019.1.diff/src/gromacs/mdrun/runner.cpp.preplumed
@@ -0,0 +1,1932 @@
+/*
+ * 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) 2011,2012,2013,2014,2015,2016,2017,2018,2019, 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_mdrun
+ */
+#include "gmxpre.h"
+
+#include "runner.h"
+
+#include "config.h"
+
+#include <cassert>
+#include <cinttypes>
+#include <csignal>
+#include <cstdlib>
+#include <cstring>
+
+#include <algorithm>
+
+#include "gromacs/commandline/filenm.h"
+#include "gromacs/compat/make_unique.h"
+#include "gromacs/domdec/domdec.h"
+#include "gromacs/domdec/domdec_struct.h"
+#include "gromacs/domdec/localatomsetmanager.h"
+#include "gromacs/ewald/ewald-utils.h"
+#include "gromacs/ewald/pme.h"
+#include "gromacs/ewald/pme-gpu-program.h"
+#include "gromacs/fileio/checkpoint.h"
+#include "gromacs/fileio/gmxfio.h"
+#include "gromacs/fileio/oenv.h"
+#include "gromacs/fileio/tpxio.h"
+#include "gromacs/gmxlib/network.h"
+#include "gromacs/gmxlib/nrnb.h"
+#include "gromacs/gpu_utils/clfftinitializer.h"
+#include "gromacs/gpu_utils/gpu_utils.h"
+#include "gromacs/hardware/cpuinfo.h"
+#include "gromacs/hardware/detecthardware.h"
+#include "gromacs/hardware/printhardware.h"
+#include "gromacs/listed-forces/disre.h"
+#include "gromacs/listed-forces/gpubonded.h"
+#include "gromacs/listed-forces/orires.h"
+#include "gromacs/math/functions.h"
+#include "gromacs/math/utilities.h"
+#include "gromacs/math/vec.h"
+#include "gromacs/mdlib/boxdeformation.h"
+#include "gromacs/mdlib/calc_verletbuf.h"
+#include "gromacs/mdlib/forcerec.h"
+#include "gromacs/mdlib/gmx_omp_nthreads.h"
+#include "gromacs/mdlib/makeconstraints.h"
+#include "gromacs/mdlib/md_support.h"
+#include "gromacs/mdlib/mdatoms.h"
+#include "gromacs/mdlib/mdrun.h"
+#include "gromacs/mdlib/membed.h"
+#include "gromacs/mdlib/nb_verlet.h"
+#include "gromacs/mdlib/nbnxn_gpu_data_mgmt.h"
+#include "gromacs/mdlib/nbnxn_search.h"
+#include "gromacs/mdlib/nbnxn_tuning.h"
+#include "gromacs/mdlib/ppforceworkload.h"
+#include "gromacs/mdlib/qmmm.h"
+#include "gromacs/mdlib/sighandler.h"
+#include "gromacs/mdlib/sim_util.h"
+#include "gromacs/mdlib/stophandler.h"
+#include "gromacs/mdrun/legacymdrunoptions.h"
+#include "gromacs/mdrun/logging.h"
+#include "gromacs/mdrun/multisim.h"
+#include "gromacs/mdrun/simulationcontext.h"
+#include "gromacs/mdrunutility/mdmodules.h"
+#include "gromacs/mdrunutility/threadaffinity.h"
+#include "gromacs/mdtypes/commrec.h"
+#include "gromacs/mdtypes/fcdata.h"
+#include "gromacs/mdtypes/inputrec.h"
+#include "gromacs/mdtypes/md_enums.h"
+#include "gromacs/mdtypes/observableshistory.h"
+#include "gromacs/mdtypes/state.h"
+#include "gromacs/pbcutil/pbc.h"
+#include "gromacs/pulling/output.h"
+#include "gromacs/pulling/pull.h"
+#include "gromacs/pulling/pull_rotation.h"
+#include "gromacs/restraint/manager.h"
+#include "gromacs/restraint/restraintmdmodule.h"
+#include "gromacs/restraint/restraintpotential.h"
+#include "gromacs/swap/swapcoords.h"
+#include "gromacs/taskassignment/decidegpuusage.h"
+#include "gromacs/taskassignment/resourcedivision.h"
+#include "gromacs/taskassignment/taskassignment.h"
+#include "gromacs/taskassignment/usergpuids.h"
+#include "gromacs/timing/wallcycle.h"
+#include "gromacs/topology/mtop_util.h"
+#include "gromacs/trajectory/trajectoryframe.h"
+#include "gromacs/utility/basenetwork.h"
+#include "gromacs/utility/cstringutil.h"
+#include "gromacs/utility/exceptions.h"
+#include "gromacs/utility/fatalerror.h"
+#include "gromacs/utility/filestream.h"
+#include "gromacs/utility/gmxassert.h"
+#include "gromacs/utility/gmxmpi.h"
+#include "gromacs/utility/logger.h"
+#include "gromacs/utility/loggerbuilder.h"
+#include "gromacs/utility/physicalnodecommunicator.h"
+#include "gromacs/utility/pleasecite.h"
+#include "gromacs/utility/programcontext.h"
+#include "gromacs/utility/smalloc.h"
+#include "gromacs/utility/stringutil.h"
+
+#include "integrator.h"
+#include "replicaexchange.h"
+
+#if GMX_FAHCORE
+#include "corewrap.h"
+#endif
+
+namespace gmx
+{
+
+/*! \brief Barrier for safe simultaneous thread access to mdrunner data
+ *
+ * Used to ensure that the master thread does not modify mdrunner during copy
+ * on the spawned threads. */
+static void threadMpiMdrunnerAccessBarrier()
+{
+#if GMX_THREAD_MPI
+ MPI_Barrier(MPI_COMM_WORLD);
+#endif
+}
+
+Mdrunner Mdrunner::cloneOnSpawnedThread() const
+{
+ auto newRunner = Mdrunner();
+
+ // All runners in the same process share a restraint manager resource because it is
+ // part of the interface to the client code, which is associated only with the
+ // original thread. Handles to the same resources can be obtained by copy.
+ {
+ newRunner.restraintManager_ = compat::make_unique<RestraintManager>(*restraintManager_);
+ }
+
+ // Copy original cr pointer before master thread can pass the thread barrier
+ newRunner.cr = reinitialize_commrec_for_this_thread(cr);
+
+ // Copy members of master runner.
+ // \todo Replace with builder when Simulation context and/or runner phases are better defined.
+ // Ref https://redmine.gromacs.org/issues/2587 and https://redmine.gromacs.org/issues/2375
+ newRunner.hw_opt = hw_opt;
+ newRunner.filenames = filenames;
+
+ newRunner.oenv = oenv;
+ newRunner.mdrunOptions = mdrunOptions;
+ newRunner.domdecOptions = domdecOptions;
+ newRunner.nbpu_opt = nbpu_opt;
+ newRunner.pme_opt = pme_opt;
+ newRunner.pme_fft_opt = pme_fft_opt;
+ newRunner.bonded_opt = bonded_opt;
+ newRunner.nstlist_cmdline = nstlist_cmdline;
+ newRunner.replExParams = replExParams;
+ newRunner.pforce = pforce;
+ newRunner.ms = ms;
+ newRunner.stopHandlerBuilder_ = compat::make_unique<StopHandlerBuilder>(*stopHandlerBuilder_);
+
+ threadMpiMdrunnerAccessBarrier();
+
+ GMX_RELEASE_ASSERT(!MASTER(newRunner.cr), "cloneOnSpawnedThread should only be called on spawned threads");
+
+ return newRunner;
+}
+
+/*! \brief The callback used for running on spawned threads.
+ *
+ * Obtains the pointer to the master mdrunner object from the one
+ * argument permitted to the thread-launch API call, copies it to make
+ * a new runner for this thread, reinitializes necessary data, and
+ * proceeds to the simulation. */
+static void mdrunner_start_fn(const void *arg)
+{
+ try
+ {
+ auto masterMdrunner = reinterpret_cast<const gmx::Mdrunner *>(arg);
+ /* copy the arg list to make sure that it's thread-local. This
+ doesn't copy pointed-to items, of course; fnm, cr and fplog
+ are reset in the call below, all others should be const. */
+ gmx::Mdrunner mdrunner = masterMdrunner->cloneOnSpawnedThread();
+ mdrunner.mdrunner();
+ }
+ GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+}
+
+
+/*! \brief Start thread-MPI threads.
+ *
+ * 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 are the same as for
+ * mdrunner(). */
+t_commrec *Mdrunner::spawnThreads(int numThreadsToLaunch) const
+{
+
+ /* first check whether we even need to start tMPI */
+ if (numThreadsToLaunch < 2)
+ {
+ return cr;
+ }
+
+#if GMX_THREAD_MPI
+ /* now spawn new threads that start mdrunner_start_fn(), while
+ the main thread returns, we set thread affinity later */
+ if (tMPI_Init_fn(TRUE, numThreadsToLaunch, TMPI_AFFINITY_NONE,
+ mdrunner_start_fn, static_cast<const void*>(this)) != TMPI_SUCCESS)
+ {
+ GMX_THROW(gmx::InternalError("Failed to spawn thread-MPI threads"));
+ }
+
+ threadMpiMdrunnerAccessBarrier();
+#else
+ GMX_UNUSED_VALUE(mdrunner_start_fn);
+#endif
+
+ return reinitialize_commrec_for_this_thread(cr);
+}
+
+} // namespace gmx
+
+/*! \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,
+ const matrix box,
+ bool makeGpuPairList,
+ 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 */
+
+ /* 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.
+ */
+ ListSetupType listType = (makeGpuPairList ? ListSetupType::Gpu : ListSetupType::CpuSimdWhenSupported);
+ VerletbufListSetup listSetup = verletbufGetSafeListSetup(listType);
+
+ real rlist_new;
+ calc_verlet_buffer_size(mtop, det(box), ir, ir->nstlist, ir->nstlist - 1, -1, &listSetup, nullptr, &rlist_new);
+
+ if (rlist_new != ir->rlist)
+ {
+ if (fplog != nullptr)
+ {
+ fprintf(fplog, "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
+ ir->rlist, rlist_new,
+ listSetup.cluster_size_i, listSetup.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 */
+ increaseNstlist(fplog, cr, ir, nstlist_cmdline, mtop, box, makeGpuPairList, cpuinfo);
+ }
+}
+
+/*! \brief Override the nslist value in inputrec
+ *
+ * with value passed on the command line (if any)
+ */
+static void override_nsteps_cmdline(const gmx::MDLogger &mdlog,
+ int64_t nsteps_cmdline,
+ t_inputrec *ir)
+{
+ assert(ir);
+
+ /* 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));
+ }
+
+ GMX_LOG(mdlog.warning).asParagraph().appendText(sbuf_msg);
+ }
+ else if (nsteps_cmdline < -2)
+ {
+ gmx_fatal(FARGS, "Invalid nsteps value passed on the command line: %" PRId64,
+ nsteps_cmdline);
+ }
+ /* Do nothing if nsteps_cmdline == -2 */
+}
+
+namespace gmx
+{
+
+/*! \brief Return whether GPU acceleration of nonbondeds is supported with the given settings.
+ *
+ * If not, and if a warning may be issued, logs a warning about
+ * falling back to CPU code. With thread-MPI, only the first
+ * call to this function should have \c issueWarning true. */
+static bool gpuAccelerationOfNonbondedIsUseful(const MDLogger &mdlog,
+ const t_inputrec *ir,
+ bool issueWarning)
+{
+ if (ir->opts.ngener - ir->nwall > 1)
+ {
+ /* The GPU code does not support more than one energy group.
+ * If the user requested GPUs explicitly, a fatal error is given later.
+ */
+ if (issueWarning)
+ {
+ GMX_LOG(mdlog.warning).asParagraph()
+ .appendText("Multiple energy groups is not implemented for GPUs, falling back to the CPU. "
+ "For better performance, run on the GPU without energy groups and then do "
+ "gmx mdrun -rerun option on the trajectory with an energy group .tpr file.");
+ }
+ return false;
+ }
+ return true;
+}
+
+//! Initializes the logger for mdrun.
+static gmx::LoggerOwner buildLogger(FILE *fplog, const t_commrec *cr)
+{
+ gmx::LoggerBuilder builder;
+ if (fplog != nullptr)
+ {
+ builder.addTargetFile(gmx::MDLogger::LogLevel::Info, fplog);
+ }
+ if (cr == nullptr || SIMMASTER(cr))
+ {
+ builder.addTargetStream(gmx::MDLogger::LogLevel::Warning,
+ &gmx::TextOutputFile::standardError());
+ }
+ return builder.build();
+}
+
+//! Make a TaskTarget from an mdrun argument string.
+static TaskTarget findTaskTarget(const char *optionString)
+{
+ TaskTarget returnValue = TaskTarget::Auto;
+
+ if (strncmp(optionString, "auto", 3) == 0)
+ {
+ returnValue = TaskTarget::Auto;
+ }
+ else if (strncmp(optionString, "cpu", 3) == 0)
+ {
+ returnValue = TaskTarget::Cpu;
+ }
+ else if (strncmp(optionString, "gpu", 3) == 0)
+ {
+ returnValue = TaskTarget::Gpu;
+ }
+ else
+ {
+ GMX_ASSERT(false, "Option string should have been checked for sanity already");
+ }
+
+ return returnValue;
+}
+
+int Mdrunner::mdrunner()
+{
+ matrix box;
+ t_nrnb *nrnb;
+ t_forcerec *fr = nullptr;
+ t_fcdata *fcd = nullptr;
+ real ewaldcoeff_q = 0;
+ real ewaldcoeff_lj = 0;
+ int nChargePerturbed = -1, nTypePerturbed = 0;
+ gmx_wallcycle_t wcycle;
+ gmx_walltime_accounting_t walltime_accounting = nullptr;
+ int rc;
+ int64_t reset_counters;
+ int nthreads_pme = 1;
+ gmx_membed_t * membed = nullptr;
+ gmx_hw_info_t *hwinfo = nullptr;
+
+ /* CAUTION: threads may be started later on in this function, so
+ cr doesn't reflect the final parallel state right now */
+ std::unique_ptr<gmx::MDModules> mdModules(new gmx::MDModules);
+ t_inputrec inputrecInstance;
+ t_inputrec *inputrec = &inputrecInstance;
+ gmx_mtop_t mtop;
+
+ bool doMembed = opt2bSet("-membed", filenames.size(), filenames.data());
+ bool doRerun = mdrunOptions.rerun;
+
+ // Handle task-assignment related user options.
+ EmulateGpuNonbonded emulateGpuNonbonded = (getenv("GMX_EMULATE_GPU") != nullptr ?
+ EmulateGpuNonbonded::Yes : EmulateGpuNonbonded::No);
+ std::vector<int> gpuIdsAvailable;
+ try
+ {
+ gpuIdsAvailable = parseUserGpuIds(hw_opt.gpuIdsAvailable);
+ // TODO We could put the GPU IDs into a std::map to find
+ // duplicates, but for the small numbers of IDs involved, this
+ // code is simple and fast.
+ for (size_t i = 0; i != gpuIdsAvailable.size(); ++i)
+ {
+ for (size_t j = i+1; j != gpuIdsAvailable.size(); ++j)
+ {
+ if (gpuIdsAvailable[i] == gpuIdsAvailable[j])
+ {
+ GMX_THROW(InvalidInputError(formatString("The string of available GPU device IDs '%s' may not contain duplicate device IDs", hw_opt.gpuIdsAvailable.c_str())));
+ }
+ }
+ }
+ }
+ GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+
+ std::vector<int> userGpuTaskAssignment;
+ try
+ {
+ userGpuTaskAssignment = parseUserGpuIds(hw_opt.userGpuTaskAssignment);
+ }
+ GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+ auto nonbondedTarget = findTaskTarget(nbpu_opt);
+ auto pmeTarget = findTaskTarget(pme_opt);
+ auto pmeFftTarget = findTaskTarget(pme_fft_opt);
+ auto bondedTarget = findTaskTarget(bonded_opt);
+ PmeRunMode pmeRunMode = PmeRunMode::None;
+
+ // Here we assume that SIMMASTER(cr) does not change even after the
+ // threads are started.
+
+ FILE *fplog = nullptr;
+ // If we are appending, we don't write log output because we need
+ // to check that the old log file matches what the checkpoint file
+ // expects. Otherwise, we should start to write log output now if
+ // there is a file ready for it.
+ if (logFileHandle != nullptr && !mdrunOptions.continuationOptions.appendFiles)
+ {
+ fplog = gmx_fio_getfp(logFileHandle);
+ }
+ gmx::LoggerOwner logOwner(buildLogger(fplog, cr));
+ gmx::MDLogger mdlog(logOwner.logger());
+
+ // TODO The thread-MPI master rank makes a working
+ // PhysicalNodeCommunicator here, but it gets rebuilt by all ranks
+ // after the threads have been launched. This works because no use
+ // is made of that communicator until after the execution paths
+ // have rejoined. But it is likely that we can improve the way
+ // this is expressed, e.g. by expressly running detection only the
+ // master rank for thread-MPI, rather than relying on the mutex
+ // and reference count.
+ PhysicalNodeCommunicator physicalNodeComm(MPI_COMM_WORLD, gmx_physicalnode_id_hash());
+ hwinfo = gmx_detect_hardware(mdlog, physicalNodeComm);
+
+ gmx_print_detected_hardware(fplog, cr, ms, mdlog, hwinfo);
+
+ std::vector<int> gpuIdsToUse;
+ auto compatibleGpus = getCompatibleGpus(hwinfo->gpu_info);
+ if (gpuIdsAvailable.empty())
+ {
+ gpuIdsToUse = compatibleGpus;
+ }
+ else
+ {
+ for (const auto &availableGpuId : gpuIdsAvailable)
+ {
+ bool availableGpuIsCompatible = false;
+ for (const auto &compatibleGpuId : compatibleGpus)
+ {
+ if (availableGpuId == compatibleGpuId)
+ {
+ availableGpuIsCompatible = true;
+ break;
+ }
+ }
+ if (!availableGpuIsCompatible)
+ {
+ gmx_fatal(FARGS, "You limited the set of compatible GPUs to a set that included ID #%d, but that ID is not for a compatible GPU. List only compatible GPUs.", availableGpuId);
+ }
+ gpuIdsToUse.push_back(availableGpuId);
+ }
+ }
+
+ if (fplog != nullptr)
+ {
+ /* 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");
+ writeSourceDoi(fplog);
+ }
+
+ std::unique_ptr<t_state> globalState;
+
+ if (SIMMASTER(cr))
+ {
+ /* Only the master rank has the global state */
+ globalState = compat::make_unique<t_state>();
+
+ /* Read (nearly) all data required for the simulation */
+ read_tpx_state(ftp2fn(efTPR, filenames.size(), filenames.data()), inputrec, globalState.get(), &mtop);
+
+ /* In rerun, set velocities to zero if present */
+ if (doRerun && ((globalState->flags & (1 << estV)) != 0))
+ {
+ // rerun does not use velocities
+ GMX_LOG(mdlog.info).asParagraph().appendText(
+ "Rerun trajectory contains velocities. Rerun does only evaluate "
+ "potential energy and forces. The velocities will be ignored.");
+ for (int i = 0; i < globalState->natoms; i++)
+ {
+ clear_rvec(globalState->v[i]);
+ }
+ globalState->flags &= ~(1 << estV);
+ }
+
+ if (inputrec->cutoff_scheme != ecutsVERLET)
+ {
+ if (nstlist_cmdline > 0)
+ {
+ gmx_fatal(FARGS, "Can not set nstlist with the group cut-off scheme");
+ }
+
+ if (!compatibleGpus.empty())
+ {
+ GMX_LOG(mdlog.warning).asParagraph().appendText(
+ "NOTE: GPU(s) found, but the current simulation can not use GPUs\n"
+ " To use a GPU, set the mdp option: cutoff-scheme = Verlet");
+ }
+ }
+ }
+
+ /* Check and update the hardware options for internal consistency */
+ check_and_update_hw_opt_1(mdlog, &hw_opt, cr, domdecOptions.numPmeRanks);
+
+ /* Early check for externally set process affinity. */
+ gmx_check_thread_affinity_set(mdlog, cr,
+ &hw_opt, hwinfo->nthreads_hw_avail, FALSE);
+
+ if (GMX_THREAD_MPI && SIMMASTER(cr))
+ {
+ if (domdecOptions.numPmeRanks > 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);
+
+ bool useGpuForNonbonded = false;
+ bool useGpuForPme = false;
+ try
+ {
+ // If the user specified the number of ranks, then we must
+ // respect that, but in default mode, we need to allow for
+ // the number of GPUs to choose the number of ranks.
+ auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
+ useGpuForNonbonded = decideWhetherToUseGpusForNonbondedWithThreadMpi
+ (nonbondedTarget, gpuIdsToUse, userGpuTaskAssignment, emulateGpuNonbonded,
+ canUseGpuForNonbonded,
+ inputrec->cutoff_scheme == ecutsVERLET,
+ gpuAccelerationOfNonbondedIsUseful(mdlog, inputrec, GMX_THREAD_MPI),
+ hw_opt.nthreads_tmpi);
+ useGpuForPme = decideWhetherToUseGpusForPmeWithThreadMpi
+ (useGpuForNonbonded, pmeTarget, gpuIdsToUse, userGpuTaskAssignment,
+ *hwinfo, *inputrec, mtop, hw_opt.nthreads_tmpi, domdecOptions.numPmeRanks);
+
+ }
+ GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+
+ /* Determine how many thread-MPI ranks to start.
+ *
+ * TODO Over-writing the user-supplied value here does
+ * prevent any possible subsequent checks from working
+ * correctly. */
+ hw_opt.nthreads_tmpi = get_nthreads_mpi(hwinfo,
+ &hw_opt,
+ gpuIdsToUse,
+ useGpuForNonbonded,
+ useGpuForPme,
+ inputrec, &mtop,
+ mdlog,
+ doMembed);
+
+ // Now start the threads for thread MPI.
+ cr = spawnThreads(hw_opt.nthreads_tmpi);
+ /* The main thread continues here with a new cr. We don't deallocate
+ the old cr because other threads may still be reading it. */
+ // TODO Both master and spawned threads call dup_tfn and
+ // reinitialize_commrec_for_this_thread. Find a way to express
+ // this better.
+ physicalNodeComm = PhysicalNodeCommunicator(MPI_COMM_WORLD, gmx_physicalnode_id_hash());
+ }
+ // END OF CAUTION: cr and physicalNodeComm are now reliable
+
+ if (PAR(cr))
+ {
+ /* now broadcast everything to the non-master nodes/threads: */
+ init_parallel(cr, inputrec, &mtop);
+ }
+
+ // Now each rank knows the inputrec that SIMMASTER read and used,
+ // and (if applicable) cr->nnodes has been assigned the number of
+ // thread-MPI ranks that have been chosen. The ranks can now all
+ // run the task-deciding functions and will agree on the result
+ // without needing to communicate.
+ //
+ // TODO Should we do the communication in debug mode to support
+ // having an assertion?
+ //
+ // Note that these variables describe only their own node.
+ //
+ // Note that when bonded interactions run on a GPU they always run
+ // alongside a nonbonded task, so do not influence task assignment
+ // even though they affect the force calculation workload.
+ bool useGpuForNonbonded = false;
+ bool useGpuForPme = false;
+ bool useGpuForBonded = false;
+ try
+ {
+ // It's possible that there are different numbers of GPUs on
+ // different nodes, which is the user's responsibilty to
+ // handle. If unsuitable, we will notice that during task
+ // assignment.
+ bool gpusWereDetected = hwinfo->ngpu_compatible_tot > 0;
+ bool usingVerletScheme = inputrec->cutoff_scheme == ecutsVERLET;
+ auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
+ useGpuForNonbonded = decideWhetherToUseGpusForNonbonded(nonbondedTarget, userGpuTaskAssignment,
+ emulateGpuNonbonded,
+ canUseGpuForNonbonded,
+ usingVerletScheme,
+ gpuAccelerationOfNonbondedIsUseful(mdlog, inputrec, !GMX_THREAD_MPI),
+ gpusWereDetected);
+ useGpuForPme = decideWhetherToUseGpusForPme(useGpuForNonbonded, pmeTarget, userGpuTaskAssignment,
+ *hwinfo, *inputrec, mtop,
+ cr->nnodes, domdecOptions.numPmeRanks,
+ gpusWereDetected);
+ auto canUseGpuForBonded = buildSupportsGpuBondeds(nullptr) && inputSupportsGpuBondeds(*inputrec, mtop, nullptr);
+ useGpuForBonded =
+ decideWhetherToUseGpusForBonded(useGpuForNonbonded, useGpuForPme, usingVerletScheme,
+ bondedTarget, canUseGpuForBonded,
+ EVDW_PME(inputrec->vdwtype),
+ EEL_PME_EWALD(inputrec->coulombtype),
+ domdecOptions.numPmeRanks, gpusWereDetected);
+
+ pmeRunMode = (useGpuForPme ? PmeRunMode::GPU : PmeRunMode::CPU);
+ if (pmeRunMode == PmeRunMode::GPU)
+ {
+ if (pmeFftTarget == TaskTarget::Cpu)
+ {
+ pmeRunMode = PmeRunMode::Mixed;
+ }
+ }
+ else if (pmeFftTarget == TaskTarget::Gpu)
+ {
+ gmx_fatal(FARGS, "Assigning FFTs to GPU requires PME to be assigned to GPU as well. With PME on CPU you should not be using -pmefft.");
+ }
+ }
+ GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+
+ // Build restraints.
+ // TODO: hide restraint implementation details from Mdrunner.
+ // There is nothing unique about restraints at this point as far as the
+ // Mdrunner is concerned. The Mdrunner should just be getting a sequence of
+ // factory functions from the SimulationContext on which to call mdModules->add().
+ // TODO: capture all restraints into a single RestraintModule, passed to the runner builder.
+ for (auto && restraint : restraintManager_->getRestraints())
+ {
+ auto module = RestraintMDModule::create(restraint,
+ restraint->sites());
+ mdModules->add(std::move(module));
+ }
+
+ // TODO: Error handling
+ mdModules->assignOptionsToModules(*inputrec->params, nullptr);
+
+ if (fplog != nullptr)
+ {
+ pr_inputrec(fplog, 0, "Input Parameters", inputrec, FALSE);
+ fprintf(fplog, "\n");
+ }
+
+ if (SIMMASTER(cr))
+ {
+ /* now make sure the state is initialized and propagated */
+ set_state_entries(globalState.get(), inputrec);
+ }
+
+ /* NM and TPI parallelize over force/energy calculations, not atoms,
+ * so we need to initialize and broadcast the global state.
+ */
+ if (inputrec->eI == eiNM || inputrec->eI == eiTPI)
+ {
+ if (!MASTER(cr))
+ {
+ globalState = compat::make_unique<t_state>();
+ }
+ broadcastStateWithoutDynamics(cr, globalState.get());
+ }
+
+ /* A parallel command line option consistency check that we can
+ only do after any threads have started. */
+ if (!PAR(cr) && (domdecOptions.numCells[XX] > 1 ||
+ domdecOptions.numCells[YY] > 1 ||
+ domdecOptions.numCells[ZZ] > 1 ||
+ domdecOptions.numPmeRanks > 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", output_env_get_program_display_name(oenv));
+#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", output_env_get_program_display_name(oenv));
+#endif
+#endif
+ }
+
+ if (doRerun &&
+ (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 (domdecOptions.numPmeRanks > 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");
+ }
+
+ domdecOptions.numPmeRanks = 0;
+ }
+
+ if (useGpuForNonbonded && domdecOptions.numPmeRanks < 0)
+ {
+ /* With NB GPUs we don't automatically use PME-only CPU 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.
+ */
+ domdecOptions.numPmeRanks = 0;
+ }
+ if (useGpuForPme)
+ {
+ if (domdecOptions.numPmeRanks < 0)
+ {
+ domdecOptions.numPmeRanks = 0;
+ // TODO possibly print a note that one can opt-in for a separate PME GPU rank?
+ }
+ else
+ {
+ GMX_RELEASE_ASSERT(domdecOptions.numPmeRanks <= 1, "PME GPU decomposition is not supported");
+ }
+ }
+
+#if 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, ms, fcd, globalState.get(), replExParams.exchangeInterval > 0);
+
+ init_orires(fplog, &mtop, inputrec, cr, ms, globalState.get(), &(fcd->orires));
+
+ auto deform = prepareBoxDeformation(globalState->box, cr, *inputrec);
+
+ ObservablesHistory observablesHistory = {};
+
+ ContinuationOptions &continuationOptions = mdrunOptions.continuationOptions;
+
+ if (continuationOptions.startedFromCheckpoint)
+ {
+ /* 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", filenames.size(), filenames.data(), cr),
+ logFileHandle,
+ cr, domdecOptions.numCells,
+ inputrec, globalState.get(),
+ &bReadEkin, &observablesHistory,
+ continuationOptions.appendFiles,
+ continuationOptions.appendFilesOptionSet,
+ mdrunOptions.reproducible);
+
+ if (bReadEkin)
+ {
+ continuationOptions.haveReadEkin = true;
+ }
+
+ if (continuationOptions.appendFiles && logFileHandle)
+ {
+ // Now we can start normal logging to the truncated log file.
+ fplog = gmx_fio_getfp(logFileHandle);
+ prepareLogAppending(fplog);
+ logOwner = buildLogger(fplog, cr);
+ mdlog = logOwner.logger();
+ }
+ }
+
+ if (mdrunOptions.numStepsCommandline > -2)
+ {
+ GMX_LOG(mdlog.info).asParagraph().
+ appendText("The -nsteps functionality is deprecated, and may be removed in a future version. "
+ "Consider using gmx convert-tpr -nsteps or changing the appropriate .mdp file field.");
+ }
+ /* override nsteps with value set on the commamdline */
+ override_nsteps_cmdline(mdlog, mdrunOptions.numStepsCommandline, inputrec);
+
+ if (SIMMASTER(cr))
+ {
+ copy_mat(globalState->box, box);
+ }
+
+ if (PAR(cr))
+ {
+ gmx_bcast(sizeof(box), box, cr);
+ }
+
+ /* Update rlist and nstlist. */
+ if (inputrec->cutoff_scheme == ecutsVERLET)
+ {
+ prepare_verlet_scheme(fplog, cr, inputrec, nstlist_cmdline, &mtop, box,
+ useGpuForNonbonded || (emulateGpuNonbonded == EmulateGpuNonbonded::Yes), *hwinfo->cpuInfo);
+ }
+
+ LocalAtomSetManager atomSets;
+
+ if (PAR(cr) && !(EI_TPI(inputrec->eI) ||
+ inputrec->eI == eiNM))
+ {
+ cr->dd = init_domain_decomposition(mdlog, cr, domdecOptions, mdrunOptions,
+ &mtop, inputrec,
+ box, positionsFromStatePointer(globalState.get()),
+ &atomSets);
+ // Note that local state still does not exist yet.
+ }
+ else
+ {
+ /* PME, if used, is done on all nodes with 1D decomposition */
+ cr->npmenodes = 0;
+ cr->duty = (DUTY_PP | DUTY_PME);
+
+ if (inputrec->ePBC == epbcSCREW)
+ {
+ gmx_fatal(FARGS,
+ "pbc=screw is only implemented with domain decomposition");
+ }
+ }
+
+ 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);
+ }
+
+#if GMX_MPI
+ if (isMultiSim(ms))
+ {
+ GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
+ "This is simulation %d out of %d running as a composite GROMACS\n"
+ "multi-simulation job. Setup for this simulation:\n",
+ ms->sim, ms->nsim);
+ }
+ GMX_LOG(mdlog.warning).appendTextFormatted(
+ "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 the number of OpenMP threads requested */
+ checkAndUpdateRequestedNumOpenmpThreads(&hw_opt, *hwinfo, cr, ms, physicalNodeComm.size_,
+ pmeRunMode, mtop);
+
+ gmx_omp_nthreads_init(mdlog, cr,
+ hwinfo->nthreads_hw_avail,
+ physicalNodeComm.size_,
+ hw_opt.nthreads_omp,
+ hw_opt.nthreads_omp_pme,
+ !thisRankHasDuty(cr, DUTY_PP),
+ inputrec->cutoff_scheme == ecutsVERLET);
+
+ // Enable FP exception detection for the Verlet scheme, but not in
+ // Release mode and not for compilers with known buggy FP
+ // exception support (clang with any optimization) or suspected
+ // buggy FP exception support (gcc 7.* with optimization).
+#if !defined NDEBUG && \
+ !((defined __clang__ || (defined(__GNUC__) && !defined(__ICC) && __GNUC__ == 7)) \
+ && defined __OPTIMIZE__)
+ const bool bEnableFPE = inputrec->cutoff_scheme == ecutsVERLET;
+#else
+ const bool bEnableFPE = false;
+#endif
+ //FIXME - reconcile with gmx_feenableexcept() call from CommandLineModuleManager::run()
+ if (bEnableFPE)
+ {
+ gmx_feenableexcept();
+ }
+
+ // Build a data structure that expresses which kinds of non-bonded
+ // task are handled by this rank.
+ //
+ // TODO Later, this might become a loop over all registered modules
+ // relevant to the mdp inputs, to find those that have such tasks.
+ //
+ // TODO This could move before init_domain_decomposition() as part
+ // of refactoring that separates the responsibility for duty
+ // assignment from setup for communication between tasks, and
+ // setup for tasks handled with a domain (ie including short-ranged
+ // tasks, bonded tasks, etc.).
+ //
+ // Note that in general useGpuForNonbonded, etc. can have a value
+ // that is inconsistent with the presence of actual GPUs on any
+ // rank, and that is not known to be a problem until the
+ // duty of the ranks on a node become known.
+ //
+ // TODO Later we might need the concept of computeTasksOnThisRank,
+ // from which we construct gpuTasksOnThisRank.
+ //
+ // Currently the DD code assigns duty to ranks that can
+ // include PP work that currently can be executed on a single
+ // GPU, if present and compatible. This has to be coordinated
+ // across PP ranks on a node, with possible multiple devices
+ // or sharing devices on a node, either from the user
+ // selection, or automatically.
+ auto haveGpus = !gpuIdsToUse.empty();
+ std::vector<GpuTask> gpuTasksOnThisRank;
+ if (thisRankHasDuty(cr, DUTY_PP))
+ {
+ if (useGpuForNonbonded)
+ {
+ // Note that any bonded tasks on a GPU always accompany a
+ // non-bonded task.
+ if (haveGpus)
+ {
+ gpuTasksOnThisRank.push_back(GpuTask::Nonbonded);
+ }
+ else if (nonbondedTarget == TaskTarget::Gpu)
+ {
+ gmx_fatal(FARGS, "Cannot run short-ranged nonbonded interactions on a GPU because there is none detected.");
+ }
+ else if (bondedTarget == TaskTarget::Gpu)
+ {
+ gmx_fatal(FARGS, "Cannot run bonded interactions on a GPU because there is none detected.");
+ }
+ }
+ }
+ // TODO cr->duty & DUTY_PME should imply that a PME algorithm is active, but currently does not.
+ if (EEL_PME(inputrec->coulombtype) && (thisRankHasDuty(cr, DUTY_PME)))
+ {
+ if (useGpuForPme)
+ {
+ if (haveGpus)
+ {
+ gpuTasksOnThisRank.push_back(GpuTask::Pme);
+ }
+ else if (pmeTarget == TaskTarget::Gpu)
+ {
+ gmx_fatal(FARGS, "Cannot run PME on a GPU because there is none detected.");
+ }
+ }
+ }
+
+ GpuTaskAssignment gpuTaskAssignment;
+ try
+ {
+ // Produce the task assignment for this rank.
+ gpuTaskAssignment = runTaskAssignment(gpuIdsToUse, userGpuTaskAssignment, *hwinfo,
+ mdlog, cr, ms, physicalNodeComm, gpuTasksOnThisRank,
+ useGpuForBonded, pmeRunMode);
+ }
+ GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+
+ /* Prevent other ranks from continuing after an issue was found
+ * and reported as a fatal error.
+ *
+ * TODO This function implements a barrier so that MPI runtimes
+ * can organize an orderly shutdown if one of the ranks has had to
+ * issue a fatal error in various code already run. When we have
+ * MPI-aware error handling and reporting, this should be
+ * improved. */
+#if GMX_MPI
+ if (PAR(cr))
+ {
+ MPI_Barrier(cr->mpi_comm_mysim);
+ }
+ if (isMultiSim(ms))
+ {
+ if (SIMMASTER(cr))
+ {
+ MPI_Barrier(ms->mpi_comm_masters);
+ }
+ /* We need another barrier to prevent non-master ranks from contiuing
+ * when an error occured in a different simulation.
+ */
+ MPI_Barrier(cr->mpi_comm_mysim);
+ }
+#endif
+
+ /* Now that we know the setup is consistent, check for efficiency */
+ check_resource_division_efficiency(hwinfo, !gpuTaskAssignment.empty(), mdrunOptions.ntompOptionIsSet,
+ cr, mdlog);
+
+ gmx_device_info_t *nonbondedDeviceInfo = nullptr;
+
+ if (thisRankHasDuty(cr, DUTY_PP))
+ {
+ // This works because only one task of each type is currently permitted.
+ auto nbGpuTaskMapping = std::find_if(gpuTaskAssignment.begin(), gpuTaskAssignment.end(),
+ hasTaskType<GpuTask::Nonbonded>);
+ if (nbGpuTaskMapping != gpuTaskAssignment.end())
+ {
+ int nonbondedDeviceId = nbGpuTaskMapping->deviceId_;
+ nonbondedDeviceInfo = getDeviceInfo(hwinfo->gpu_info, nonbondedDeviceId);
+ init_gpu(nonbondedDeviceInfo);
+
+ if (DOMAINDECOMP(cr))
+ {
+ /* When we share GPUs over ranks, we need to know this for the DLB */
+ dd_setup_dlb_resource_sharing(cr, nonbondedDeviceId);
+ }
+
+ }
+ }
+
+ std::unique_ptr<ClfftInitializer> initializedClfftLibrary;
+
+ gmx_device_info_t *pmeDeviceInfo = nullptr;
+ // Later, this program could contain kernels that might be later
+ // re-used as auto-tuning progresses, or subsequent simulations
+ // are invoked.
+ PmeGpuProgramStorage pmeGpuProgram;
+ // This works because only one task of each type is currently permitted.
+ auto pmeGpuTaskMapping = std::find_if(gpuTaskAssignment.begin(), gpuTaskAssignment.end(), hasTaskType<GpuTask::Pme>);
+ const bool thisRankHasPmeGpuTask = (pmeGpuTaskMapping != gpuTaskAssignment.end());
+ if (thisRankHasPmeGpuTask)
+ {
+ pmeDeviceInfo = getDeviceInfo(hwinfo->gpu_info, pmeGpuTaskMapping->deviceId_);
+ init_gpu(pmeDeviceInfo);
+ pmeGpuProgram = buildPmeGpuProgram(pmeDeviceInfo);
+ // TODO It would be nice to move this logic into the factory
+ // function. See Redmine #2535.
+ bool isMasterThread = !GMX_THREAD_MPI || MASTER(cr);
+ if (pmeRunMode == PmeRunMode::GPU && !initializedClfftLibrary && isMasterThread)
+ {
+ initializedClfftLibrary = initializeClfftLibrary();
+ }
+ }
+
+ /* 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);
+
+ int numThreadsOnThisRank;
+ /* threads on this MPI process or TMPI thread */
+ if (thisRankHasDuty(cr, DUTY_PP))
+ {
+ numThreadsOnThisRank = gmx_omp_nthreads_get(emntNonbonded);
+ }
+ else
+ {
+ numThreadsOnThisRank = nthreads_pme;
+ }
+
+ checkHardwareOversubscription(numThreadsOnThisRank, cr->nodeid,
+ *hwinfo->hardwareTopology,
+ physicalNodeComm, mdlog);
+
+ 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(mdlog, cr,
+ &hw_opt, hwinfo->nthreads_hw_avail, TRUE);
+
+ int numThreadsOnThisNode, intraNodeThreadOffset;
+ analyzeThreadsOnThisNode(physicalNodeComm, numThreadsOnThisRank, &numThreadsOnThisNode,
+ &intraNodeThreadOffset);
+
+ /* Set the CPU affinity */
+ gmx_set_thread_affinity(mdlog, cr, &hw_opt, *hwinfo->hardwareTopology,
+ numThreadsOnThisRank, numThreadsOnThisNode,
+ intraNodeThreadOffset, nullptr);
+ }
+
+ if (mdrunOptions.timingOptions.resetStep > -1)
+ {
+ GMX_LOG(mdlog.info).asParagraph().
+ appendText("The -resetstep functionality is deprecated, and may be removed in a future version.");
+ }
+ wcycle = wallcycle_init(fplog, mdrunOptions.timingOptions.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);
+ }
+
+ // Membrane embedding must be initialized before we call init_forcerec()
+ if (doMembed)
+ {
+ 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, filenames.size(), filenames.data(), &mtop, inputrec, globalState.get(), cr,
+ &mdrunOptions
+ .checkpointOptions.period);
+ }
+
+ std::unique_ptr<MDAtoms> mdAtoms;
+ std::unique_ptr<gmx_vsite_t> vsite;
+
+ snew(nrnb, 1);
+ if (thisRankHasDuty(cr, DUTY_PP))
+ {
+ /* Initiate forcerecord */
+ fr = mk_forcerec();
+ fr->forceProviders = mdModules->initForceProviders();
+ init_forcerec(fplog, mdlog, fr, fcd,
+ inputrec, &mtop, cr, box,
+ opt2fn("-table", filenames.size(), filenames.data()),
+ opt2fn("-tablep", filenames.size(), filenames.data()),
+ opt2fns("-tableb", filenames.size(), filenames.data()),
+ *hwinfo, nonbondedDeviceInfo,
+ useGpuForBonded,
+ FALSE,
+ pforce);
+
+ /* Initialize the mdAtoms structure.
+ * mdAtoms is not filled with atom data,
+ * as this can not be done now with domain decomposition.
+ */
+ mdAtoms = makeMDAtoms(fplog, mtop, *inputrec, thisRankHasPmeGpuTask);
+ if (globalState && thisRankHasPmeGpuTask)
+ {
+ // The pinning of coordinates in the global state object works, because we only use
+ // PME on GPU without DD or on a separate PME rank, and because the local state pointer
+ // points to the global state object without DD.
+ // FIXME: MD and EM separately set up the local state - this should happen in the same function,
+ // which should also perform the pinning.
+ changePinningPolicy(&globalState->x, pme_get_pinning_policy());
+ }
+
+ /* Initialize the virtual site communication */
+ vsite = initVsite(mtop, cr);
+
+ calc_shifts(box, fr->shift_vec);
+
+ /* With periodic molecules the charge groups should be whole at start up
+ * and the virtual sites should not be far from their proper positions.
+ */
+ if (!inputrec->bContinuation && MASTER(cr) &&
+ !(inputrec->ePBC != epbcNONE && inputrec->bPeriodicMols))
+ {
+ /* Make molecules whole at start of run */
+ if (fr->ePBC != epbcNONE)
+ {
+ do_pbc_first_mtop(fplog, inputrec->ePBC, box, &mtop, globalState->x.rvec_array());
+ }
+ if (vsite)
+ {
+ /* Correct initial vsite positions are required
+ * for the initial distribution in the domain decomposition
+ * and for the initial shell prediction.
+ */
+ constructVsitesGlobal(mtop, globalState->x);
+ }
+ }
+
+ if (EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
+ {
+ ewaldcoeff_q = fr->ic->ewaldcoeff_q;
+ ewaldcoeff_lj = fr->ic->ewaldcoeff_lj;
+ }
+ }
+ else
+ {
+ /* This is a PME only node */
+
+ GMX_ASSERT(globalState == nullptr, "We don't need the state on a PME only rank and expect it to be unitialized");
+
+ ewaldcoeff_q = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
+ ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
+ }
+
+ gmx_pme_t *sepPmeData = nullptr;
+ // This reference hides the fact that PME data is owned by runner on PME-only ranks and by forcerec on other ranks
+ GMX_ASSERT(thisRankHasDuty(cr, DUTY_PP) == (fr != nullptr), "Double-checking that only PME-only ranks have no forcerec");
+ gmx_pme_t * &pmedata = fr ? fr->pmedata : sepPmeData;
+
+ /* 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 && mdAtoms->mdatoms())
+ {
+ nChargePerturbed = mdAtoms->mdatoms()->nChargePerturbed;
+ if (EVDW_PME(inputrec->vdwtype))
+ {
+ nTypePerturbed = mdAtoms->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 (thisRankHasDuty(cr, DUTY_PME))
+ {
+ try
+ {
+ pmedata = gmx_pme_init(cr,
+ getNumPmeDomains(cr->dd),
+ inputrec,
+ mtop.natoms, nChargePerturbed != 0, nTypePerturbed != 0,
+ mdrunOptions.reproducible,
+ ewaldcoeff_q, ewaldcoeff_lj,
+ nthreads_pme,
+ pmeRunMode, nullptr,
+ pmeDeviceInfo, pmeGpuProgram.get(), mdlog);
+ }
+ GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
+ }
+ }
+
+
+ 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 (thisRankHasDuty(cr, 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,
+ &mtop, cr, &atomSets, inputrec->fepvals->init_lambda);
+ if (inputrec->pull->bXOutAverage || inputrec->pull->bFOutAverage)
+ {
+ initPullHistory(inputrec->pull_work, &observablesHistory);
+ }
+ if (EI_DYNAMICS(inputrec->eI) && MASTER(cr))
+ {
+ init_pull_output_files(inputrec->pull_work,
+ filenames.size(), filenames.data(), oenv,
+ continuationOptions);
+ }
+ }
+
+ std::unique_ptr<EnforcedRotation> enforcedRotation;
+ if (inputrec->bRot)
+ {
+ /* Initialize enforced rotation code */
+ enforcedRotation = init_rot(fplog,
+ inputrec,
+ filenames.size(),
+ filenames.data(),
+ cr,
+ &atomSets,
+ globalState.get(),
+ &mtop,
+ oenv,
+ mdrunOptions);
+ }
+
+ if (inputrec->eSwapCoords != eswapNO)
+ {
+ /* Initialize ion swapping code */
+ init_swapcoords(fplog, inputrec, opt2fn_master("-swap", filenames.size(), filenames.data(), cr),
+ &mtop, globalState.get(), &observablesHistory,
+ cr, &atomSets, oenv, mdrunOptions);
+ }
+
+ /* Let makeConstraints know whether we have essential dynamics constraints.
+ * TODO: inputrec should tell us whether we use an algorithm, not a file option or the checkpoint
+ */
+ bool doEssentialDynamics = (opt2fn_null("-ei", filenames.size(), filenames.data()) != nullptr
+ || observablesHistory.edsamHistory);
+ auto constr = makeConstraints(mtop, *inputrec, doEssentialDynamics,
+ fplog, *mdAtoms->mdatoms(),
+ cr, ms, nrnb, wcycle, fr->bMolPBC);
+
+ 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.get(), inputrec,
+ domdecOptions.checkBondedInteractions,
+ fr->cginfo_mb);
+ }
+
+ // TODO This is not the right place to manage the lifetime of
+ // this data structure, but currently it's the easiest way to
+ // make it work. Later, it should probably be made/updated
+ // after the workload for the lifetime of a PP domain is
+ // understood.
+ PpForceWorkload ppForceWorkload;
+
+ GMX_ASSERT(stopHandlerBuilder_, "Runner must provide StopHandlerBuilder to integrator.");
+ /* Now do whatever the user wants us to do (how flexible...) */
+ Integrator integrator {
+ fplog, cr, ms, mdlog, static_cast<int>(filenames.size()), filenames.data(),
+ oenv,
+ mdrunOptions,
+ vsite.get(), constr.get(),
+ enforcedRotation ? enforcedRotation->getLegacyEnfrot() : nullptr,
+ deform.get(),
+ mdModules->outputProvider(),
+ inputrec, &mtop,
+ fcd,
+ globalState.get(),
+ &observablesHistory,
+ mdAtoms.get(), nrnb, wcycle, fr,
+ &ppForceWorkload,
+ replExParams,
+ membed,
+ walltime_accounting,
+ std::move(stopHandlerBuilder_)
+ };
+ integrator.run(inputrec->eI, doRerun);
+
+ 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, inputrec, pmeRunMode);
+ }
+
+ wallcycle_stop(wcycle, ewcRUN);
+
+ /* Finish up, write some stuff
+ * if rerunMD, don't write last frame again
+ */
+ finish_run(fplog, mdlog, cr,
+ inputrec, nrnb, wcycle, walltime_accounting,
+ fr ? fr->nbv : nullptr,
+ pmedata,
+ EI_DYNAMICS(inputrec->eI) && !isMultiSim(ms));
+
+ // Free PME data
+ if (pmedata)
+ {
+ gmx_pme_destroy(pmedata);
+ pmedata = nullptr;
+ }
+
+ // FIXME: this is only here to manually unpin mdAtoms->chargeA_ and state->x,
+ // before we destroy the GPU context(s) in free_gpu_resources().
+ // Pinned buffers are associated with contexts in CUDA.
+ // As soon as we destroy GPU contexts after mdrunner() exits, these lines should go.
+ mdAtoms.reset(nullptr);
+ globalState.reset(nullptr);
+ mdModules.reset(nullptr); // destruct force providers here as they might also use the GPU
+
+ /* Free GPU memory and set a physical node tMPI barrier (which should eventually go away) */
+ free_gpu_resources(fr, physicalNodeComm);
+ free_gpu(nonbondedDeviceInfo);
+ free_gpu(pmeDeviceInfo);
+ done_forcerec(fr, mtop.molblock.size(), mtop.groups.grps[egcENER].nr);
+ sfree(fcd);
+
+ if (doMembed)
+ {
+ free_membed(membed);
+ }
+
+ gmx_hardware_info_free();
+
+ /* Does what it says */
+ print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
+ walltime_accounting_destroy(walltime_accounting);
+ sfree(nrnb);
+
+ // Ensure log file content is written
+ if (logFileHandle)
+ {
+ gmx_fio_flush(logFileHandle);
+ }
+
+ /* Reset FPEs (important for unit tests) by disabling them. Assumes no
+ * exceptions were enabled before function was called. */
+ if (bEnableFPE)
+ {
+ gmx_fedisableexcept();
+ }
+
+ rc = static_cast<int>(gmx_get_stop_condition());
+
+#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))
+ {
+ done_commrec(cr);
+ tMPI_Finalize();
+ }
+#endif
+
+ return rc;
+}
+
+Mdrunner::~Mdrunner()
+{
+ // Clean up of the Manager.
+ // This will end up getting called on every thread-MPI rank, which is unnecessary,
+ // but okay as long as threads synchronize some time before adding or accessing
+ // a new set of restraints.
+ if (restraintManager_)
+ {
+ restraintManager_->clear();
+ GMX_ASSERT(restraintManager_->countRestraints() == 0,
+ "restraints added during runner life time should be cleared at runner destruction.");
+ }
+};
+
+void Mdrunner::addPotential(std::shared_ptr<gmx::IRestraintPotential> puller,
+ std::string name)
+{
+ GMX_ASSERT(restraintManager_, "Mdrunner must have a restraint manager.");
+ // Not sure if this should be logged through the md logger or something else,
+ // but it is helpful to have some sort of INFO level message sent somewhere.
+ // std::cout << "Registering restraint named " << name << std::endl;
+
+ // When multiple restraints are used, it may be wasteful to register them separately.
+ // Maybe instead register an entire Restraint Manager as a force provider.
+ restraintManager_->addToSpec(std::move(puller),
+ std::move(name));
+}
+
+Mdrunner::Mdrunner(Mdrunner &&) noexcept = default;
+
+//NOLINTNEXTLINE(performance-noexcept-move-constructor) working around GCC bug 58265
+Mdrunner &Mdrunner::operator=(Mdrunner && /*handle*/) noexcept(BUGFREE_NOEXCEPT_STRING) = default;
+
+class Mdrunner::BuilderImplementation
+{
+ public:
+ BuilderImplementation() = delete;
+ explicit BuilderImplementation(SimulationContext* context);
+ ~BuilderImplementation();
+
+ BuilderImplementation &setExtraMdrunOptions(const MdrunOptions &options,
+ real forceWarningThreshold);
+
+ void addDomdec(const DomdecOptions &options);
+
+ void addVerletList(int nstlist);
+
+ void addReplicaExchange(const ReplicaExchangeParameters ¶ms);
+
+ void addMultiSim(gmx_multisim_t* multisim);
+
+ void addNonBonded(const char* nbpu_opt);
+
+ void addPME(const char* pme_opt_, const char* pme_fft_opt_);
+
+ void addBondedTaskAssignment(const char* bonded_opt);
+
+ void addHardwareOptions(const gmx_hw_opt_t &hardwareOptions);
+
+ void addFilenames(ArrayRef <const t_filenm> filenames);
+
+ void addOutputEnvironment(gmx_output_env_t* outputEnvironment);
+
+ void addLogFile(t_fileio *logFileHandle);
+
+ void addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder);
+
+ Mdrunner build();
+
+ private:
+ // Default parameters copied from runner.h
+ // \todo Clarify source(s) of default parameters.
+
+ const char* nbpu_opt_ = nullptr;
+ const char* pme_opt_ = nullptr;
+ const char* pme_fft_opt_ = nullptr;
+ const char *bonded_opt_ = nullptr;
+
+ MdrunOptions mdrunOptions_;
+
+ DomdecOptions domdecOptions_;
+
+ ReplicaExchangeParameters replicaExchangeParameters_;
+
+ //! Command-line override for the duration of a neighbor list with the Verlet scheme.
+ int nstlist_ = 0;
+
+ //! Non-owning multisim communicator handle.
+ std::unique_ptr<gmx_multisim_t*> multisim_ = nullptr;
+
+ //! Print a warning if any force is larger than this (in kJ/mol nm).
+ real forceWarningThreshold_ = -1;
+
+ /*! \brief Non-owning pointer to SimulationContext (owned and managed by client)
+ *
+ * \internal
+ * \todo Establish robust protocol to make sure resources remain valid.
+ * SimulationContext will likely be separated into multiple layers for
+ * different levels of access and different phases of execution. Ref
+ * https://redmine.gromacs.org/issues/2375
+ * https://redmine.gromacs.org/issues/2587
+ */
+ SimulationContext* context_ = nullptr;
+
+ //! \brief Parallelism information.
+ gmx_hw_opt_t hardwareOptions_;
+
+ //! filename options for simulation.
+ ArrayRef<const t_filenm> filenames_;
+
+ /*! \brief Handle to output environment.
+ *
+ * \todo gmx_output_env_t needs lifetime management.
+ */
+ gmx_output_env_t* outputEnvironment_ = nullptr;
+
+ /*! \brief Non-owning handle to MD log file.
+ *
+ * \todo Context should own output facilities for client.
+ * \todo Improve log file handle management.
+ * \internal
+ * Code managing the FILE* relies on the ability to set it to
+ * nullptr to check whether the filehandle is valid.
+ */
+ t_fileio* logFileHandle_ = nullptr;
+
+ /*!
+ * \brief Builder for simulation stop signal handler.
+ */
+ std::unique_ptr<StopHandlerBuilder> stopHandlerBuilder_ = nullptr;
+};
+
+Mdrunner::BuilderImplementation::BuilderImplementation(SimulationContext* context) :
+ context_(context)
+{
+ GMX_ASSERT(context_, "Bug found. It should not be possible to construct builder without a valid context.");
+}
+
+Mdrunner::BuilderImplementation::~BuilderImplementation() = default;
+
+Mdrunner::BuilderImplementation &
+Mdrunner::BuilderImplementation::setExtraMdrunOptions(const MdrunOptions &options,
+ real forceWarningThreshold)
+{
+ mdrunOptions_ = options;
+ forceWarningThreshold_ = forceWarningThreshold;
+ return *this;
+}
+
+void Mdrunner::BuilderImplementation::addDomdec(const DomdecOptions &options)
+{
+ domdecOptions_ = options;
+}
+
+void Mdrunner::BuilderImplementation::addVerletList(int nstlist)
+{
+ nstlist_ = nstlist;
+}
+
+void Mdrunner::BuilderImplementation::addReplicaExchange(const ReplicaExchangeParameters ¶ms)
+{
+ replicaExchangeParameters_ = params;
+}
+
+void Mdrunner::BuilderImplementation::addMultiSim(gmx_multisim_t* multisim)
+{
+ multisim_ = compat::make_unique<gmx_multisim_t*>(multisim);
+}
+
+Mdrunner Mdrunner::BuilderImplementation::build()
+{
+ auto newRunner = Mdrunner();
+
+ GMX_ASSERT(context_, "Bug found. It should not be possible to call build() without a valid context.");
+
+ newRunner.mdrunOptions = mdrunOptions_;
+ newRunner.domdecOptions = domdecOptions_;
+
+ // \todo determine an invariant to check or confirm that all gmx_hw_opt_t objects are valid
+ newRunner.hw_opt = hardwareOptions_;
+
+ // No invariant to check. This parameter exists to optionally override other behavior.
+ newRunner.nstlist_cmdline = nstlist_;
+
+ newRunner.replExParams = replicaExchangeParameters_;
+
+ newRunner.filenames = filenames_;
+
+ GMX_ASSERT(context_->communicationRecord_, "SimulationContext communications not initialized.");
+ newRunner.cr = context_->communicationRecord_;
+
+ if (multisim_)
+ {
+ // nullptr is a valid value for the multisim handle, so we don't check the pointed-to pointer.
+ newRunner.ms = *multisim_;
+ }
+ else
+ {
+ GMX_THROW(gmx::APIError("MdrunnerBuilder::addMultiSim() is required before build()"));
+ }
+
+ // \todo Clarify ownership and lifetime management for gmx_output_env_t
+ // \todo Update sanity checking when output environment has clearly specified invariants.
+ // Initialization and default values for oenv are not well specified in the current version.
+ if (outputEnvironment_)
+ {
+ newRunner.oenv = outputEnvironment_;
+ }
+ else
+ {
+ GMX_THROW(gmx::APIError("MdrunnerBuilder::addOutputEnvironment() is required before build()"));
+ }
+
+ newRunner.logFileHandle = logFileHandle_;
+
+ if (nbpu_opt_)
+ {
+ newRunner.nbpu_opt = nbpu_opt_;
+ }
+ else
+ {
+ GMX_THROW(gmx::APIError("MdrunnerBuilder::addNonBonded() is required before build()"));
+ }
+
+ if (pme_opt_ && pme_fft_opt_)
+ {
+ newRunner.pme_opt = pme_opt_;
+ newRunner.pme_fft_opt = pme_fft_opt_;
+ }
+ else
+ {
+ GMX_THROW(gmx::APIError("MdrunnerBuilder::addElectrostatics() is required before build()"));
+ }
+
+ if (bonded_opt_)
+ {
+ newRunner.bonded_opt = bonded_opt_;
+ }
+ else
+ {
+ GMX_THROW(gmx::APIError("MdrunnerBuilder::addBondedTaskAssignment() is required before build()"));
+ }
+
+ newRunner.restraintManager_ = compat::make_unique<gmx::RestraintManager>();
+
+ if (stopHandlerBuilder_)
+ {
+ newRunner.stopHandlerBuilder_ = std::move(stopHandlerBuilder_);
+ }
+ else
+ {
+ newRunner.stopHandlerBuilder_ = compat::make_unique<StopHandlerBuilder>();
+ }
+
+ return newRunner;
+}
+
+void Mdrunner::BuilderImplementation::addNonBonded(const char* nbpu_opt)
+{
+ nbpu_opt_ = nbpu_opt;
+}
+
+void Mdrunner::BuilderImplementation::addPME(const char* pme_opt,
+ const char* pme_fft_opt)
+{
+ pme_opt_ = pme_opt;
+ pme_fft_opt_ = pme_fft_opt;
+}
+
+void Mdrunner::BuilderImplementation::addBondedTaskAssignment(const char* bonded_opt)
+{
+ bonded_opt_ = bonded_opt;
+}
+
+void Mdrunner::BuilderImplementation::addHardwareOptions(const gmx_hw_opt_t &hardwareOptions)
+{
+ hardwareOptions_ = hardwareOptions;
+}
+
+void Mdrunner::BuilderImplementation::addFilenames(ArrayRef<const t_filenm> filenames)
+{
+ filenames_ = filenames;
+}
+
+void Mdrunner::BuilderImplementation::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
+{
+ outputEnvironment_ = outputEnvironment;
+}
+
+void Mdrunner::BuilderImplementation::addLogFile(t_fileio *logFileHandle)
+{
+ logFileHandle_ = logFileHandle;
+}
+
+void Mdrunner::BuilderImplementation::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
+{
+ stopHandlerBuilder_ = std::move(builder);
+}
+
+MdrunnerBuilder::MdrunnerBuilder(compat::not_null<SimulationContext*> context) :
+ impl_ {gmx::compat::make_unique<Mdrunner::BuilderImplementation>(context)}
+{
+}
+
+MdrunnerBuilder::~MdrunnerBuilder() = default;
+
+MdrunnerBuilder &MdrunnerBuilder::addSimulationMethod(const MdrunOptions &options,
+ real forceWarningThreshold)
+{
+ impl_->setExtraMdrunOptions(options, forceWarningThreshold);
+ return *this;
+}
+
+MdrunnerBuilder &MdrunnerBuilder::addDomainDecomposition(const DomdecOptions &options)
+{
+ impl_->addDomdec(options);
+ return *this;
+}
+
+MdrunnerBuilder &MdrunnerBuilder::addNeighborList(int nstlist)
+{
+ impl_->addVerletList(nstlist);
+ return *this;
+}
+
+MdrunnerBuilder &MdrunnerBuilder::addReplicaExchange(const ReplicaExchangeParameters ¶ms)
+{
+ impl_->addReplicaExchange(params);
+ return *this;
+}
+
+MdrunnerBuilder &MdrunnerBuilder::addMultiSim(gmx_multisim_t* multisim)
+{
+ impl_->addMultiSim(multisim);
+ return *this;
+}
+
+MdrunnerBuilder &MdrunnerBuilder::addNonBonded(const char* nbpu_opt)
+{
+ impl_->addNonBonded(nbpu_opt);
+ return *this;
+}
+
+MdrunnerBuilder &MdrunnerBuilder::addElectrostatics(const char* pme_opt,
+ const char* pme_fft_opt)
+{
+ // The builder method may become more general in the future, but in this version,
+ // parameters for PME electrostatics are both required and the only parameters
+ // available.
+ if (pme_opt && pme_fft_opt)
+ {
+ impl_->addPME(pme_opt, pme_fft_opt);
+ }
+ else
+ {
+ GMX_THROW(gmx::InvalidInputError("addElectrostatics() arguments must be non-null pointers."));
+ }
+ return *this;
+}
+
+MdrunnerBuilder &MdrunnerBuilder::addBondedTaskAssignment(const char* bonded_opt)
+{
+ impl_->addBondedTaskAssignment(bonded_opt);
+ return *this;
+}
+
+Mdrunner MdrunnerBuilder::build()
+{
+ return impl_->build();
+}
+
+MdrunnerBuilder &MdrunnerBuilder::addHardwareOptions(const gmx_hw_opt_t &hardwareOptions)
+{
+ impl_->addHardwareOptions(hardwareOptions);
+ return *this;
+}
+
+MdrunnerBuilder &MdrunnerBuilder::addFilenames(ArrayRef<const t_filenm> filenames)
+{
+ impl_->addFilenames(filenames);
+ return *this;
+}
+
+MdrunnerBuilder &MdrunnerBuilder::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
+{
+ impl_->addOutputEnvironment(outputEnvironment);
+ return *this;
+}
+
+MdrunnerBuilder &MdrunnerBuilder::addLogFile(t_fileio *logFileHandle)
+{
+ impl_->addLogFile(logFileHandle);
+ return *this;
+}
+
+MdrunnerBuilder &MdrunnerBuilder::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
+{
+ impl_->addStopHandlerBuilder(std::move(builder));
+ return *this;
+}
+
+MdrunnerBuilder::MdrunnerBuilder(MdrunnerBuilder &&) noexcept = default;
+
+MdrunnerBuilder &MdrunnerBuilder::operator=(MdrunnerBuilder &&) noexcept = default;
+
+} // namespace gmx