Merge release-4-6 into master
[alexxy/gromacs.git] / cmake / ThreadMPI.cmake
1
2 include(CheckIncludeFiles)
3 include(CheckFunctionExists)
4 #include(CheckCSourceCompiles)
5
6 #option(THREAD_PTHREADS "Use posix threads" ON)
7
8 MACRO(TEST_TMPI_ATOMICS VARIABLE)
9     if (NOT DEFINED TMPI_ATOMICS)
10         try_compile(TEST_ATOMICS "${CMAKE_BINARY_DIR}"
11                 "${CMAKE_SOURCE_DIR}/cmake/TestAtomics.c"
12                 COMPILE_DEFINITIONS "-I${CMAKE_SOURCE_DIR}/src/gromacs/legacyheaders" )
13
14         if (TEST_ATOMICS)
15             message(STATUS "Atomics found")
16             set(${VARIABLE} CACHE INTERNAL 1)
17         else (TEST_ATOMICS)
18             message(WARNING "Atomics not found for this compiler+cpu combination. Thread support will be unbearably slow: disable threads. Atomics should work on all but the most obscure CPU+compiler combinations; if your system is not obscure -- like, for example, x86 with gcc --  please contact the developers.")
19             set(${VARIABLE} CACHE INTERNAL 0)
20         endif(TEST_ATOMICS)
21     endif(NOT DEFINED TMPI_ATOMICS)
22 ENDMACRO(TEST_TMPI_ATOMICS VARIABLE)
23
24 MACRO(TMPI_MAKE_CXX_LIB)
25     set(TMPI_CXX_LIB 1)
26 ENDMACRO(TMPI_MAKE_CXX_LIB)
27
28 MACRO(TMPI_GET_SOURCE_LIST SRC_VARIABLE)
29     foreach (_option IN ITEMS ${ARGN})
30         if (_option STREQUAL "CXX")
31             set(TMPI_CXX_LIB 1)
32         elseif (_option STREQUAL "NOMPI")
33             set(TMPI_NO_MPI_LIB 1)
34         else ()
35             message(FATAL_ERROR "Unknown thread_mpi option '${_option}'")
36         endif ()
37     endforeach ()
38     set(${SRC_VARIABLE}
39         thread_mpi/errhandler.c
40         thread_mpi/tmpi_malloc.c)
41     if (THREAD_PTHREADS)
42         list(APPEND ${SRC_VARIABLE} thread_mpi/pthreads.c)
43     elseif (THREAD_WINDOWS)
44         list(APPEND ${SRC_VARIABLE} thread_mpi/winthreads.c)
45     endif (THREAD_PTHREADS)
46     if (TMPI_CXX_LIB)
47         list(APPEND ${SRC_VARIABLE} thread_mpi/system_error.cpp)
48     endif (TMPI_CXX_LIB)
49     if (NOT TMPI_NO_MPI_LIB)
50         list(APPEND ${SRC_VARIABLE}
51              thread_mpi/alltoall.c      thread_mpi/p2p_protocol.c
52              thread_mpi/barrier.c       thread_mpi/p2p_send_recv.c
53              thread_mpi/bcast.c         thread_mpi/p2p_wait.c
54              thread_mpi/collective.c    thread_mpi/profile.c
55              thread_mpi/comm.c          thread_mpi/reduce.c
56              thread_mpi/event.c         thread_mpi/reduce_fast.c
57              thread_mpi/gather.c        thread_mpi/scatter.c
58              thread_mpi/group.c         thread_mpi/tmpi_init.c
59              thread_mpi/topology.c      thread_mpi/list.c
60              thread_mpi/type.c          thread_mpi/lock.c
61              thread_mpi/numa_malloc.c   thread_mpi/once.c
62              thread_mpi/scan.c)
63     endif()
64 ENDMACRO(TMPI_GET_SOURCE_LIST)
65
66 include(FindThreads)
67 if (CMAKE_USE_PTHREADS_INIT)
68     check_include_files(pthread.h    HAVE_PTHREAD_H)
69     set(THREAD_PTHREADS 1)
70     #add_definitions(-DTHREAD_PTHREADS)
71     set(THREAD_LIB ${CMAKE_THREAD_LIBS_INIT})
72 else (CMAKE_USE_PTHREADS_INIT)
73     if (CMAKE_USE_WIN32_THREADS_INIT)
74         set(THREAD_WINDOWS 1)
75         #add_definitions(-DTHREAD_WINDOWS)
76         set(THREAD_LIB)
77     endif (CMAKE_USE_WIN32_THREADS_INIT)
78 endif (CMAKE_USE_PTHREADS_INIT)
79
80
81 # the spin-waiting option
82 option(THREAD_MPI_WAIT_FOR_NO_ONE "Use busy waits without yielding to the OS scheduler. Turning this on might improve performance (very) slightly at the cost of very poor performance if the threads are competing for CPU time." OFF)
83 mark_as_advanced(THREAD_MPI_WAIT_FOR_NO_ONE)
84 if (THREAD_MPI_WAIT_FOR_NO_ONE)
85     add_definitions(-DTMPI_WAIT_FOR_NO_ONE)
86 else (THREAD_MPI_WAIT_FOR_NO_ONE)
87     add_definitions()
88 endif (THREAD_MPI_WAIT_FOR_NO_ONE)
89
90
91 # the copy buffer option
92 option(THREAD_MPI_COPY_BUFFER "Use an intermediate copy buffer for small message sizes, to allow blocking sends to return quickly." ON)
93 mark_as_advanced(THREAD_MPI_COPY_BUFFER)
94 if (THREAD_MPI_COPY_BUFFER)
95     add_definitions()
96 else (THREAD_MPI_COPY_BUFFER)
97     add_definitions(-DTMPI_NO_COPY_BUFFER)
98 endif (THREAD_MPI_COPY_BUFFER)
99
100
101 # the profiling option
102 option(THREAD_MPI_PROFILING "Turn on simple MPI profiling." OFF)
103 mark_as_advanced(THREAD_MPI_PROFILING)
104 if (THREAD_MPI_PROFILING)
105     add_definitions(-DTMPI_PROFILE)
106 else (THREAD_MPI_PROFILING)
107     add_definitions()
108 endif (THREAD_MPI_PROFILING)
109
110 include(CheckCSourceCompiles)
111
112 # option to set affinity 
113 option(THREAD_MPI_SET_AFFINITY "Set thread affinity to a core if number of threads equal to number of hardware threads." ON)
114 mark_as_advanced(THREAD_MPI_SET_AFFINITY)
115 if (THREAD_MPI_SET_AFFINITY)
116     add_definitions(-DTMPI_SET_AFFINITY)
117 else (THREAD_MPI_SET_AFFINITY)
118     add_definitions()
119 endif (THREAD_MPI_SET_AFFINITY)
120
121 include(CheckFunctionExists)
122 if (THREAD_PTHREADS)
123     set(CMAKE_REQUIRED_LIBRARIES ${CMAKE_THREAD_LIBS_INIT})
124     # check for sched_setaffinity
125     check_c_source_compiles(
126         "#define _GNU_SOURCE
127 #include <pthread.h>
128 #include <stdlib.h>
129 #include <stdio.h>
130 #include <errno.h>
131 int main(void) { cpu_set_t set;
132     CPU_ZERO(&set);
133     CPU_SET(0, &set);
134     pthread_setaffinity_np(pthread_self(), sizeof(set), &set);
135     return 0;
136 }"
137         PTHREAD_SETAFFINITY
138     )
139     if (PTHREAD_SETAFFINITY)
140         set(HAVE_PTHREAD_SETAFFINITY 1)
141     endif (PTHREAD_SETAFFINITY)
142 endif (THREAD_PTHREADS)
143
144
145 # this runs on POSIX systems
146 check_include_files(unistd.h        HAVE_UNISTD_H)
147 check_include_files(sched.h         HAVE_SCHED_H)
148 check_include_files(sys/time.h      HAVE_SYS_TIME_H)
149 check_function_exists(sysconf       HAVE_SYSCONF)
150 # this runs on windows
151 #check_include_files(windows.h          HAVE_WINDOWS_H)
152
153
154 test_tmpi_atomics(TMPI_ATOMICS)