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${CMAKE_SOURCE_DIR}/include" )
  13
  14         if (TEST_ATOMICS)
  15             message(STATUS "Atomics found")
  16             set(${VARIABLE} TRUE CACHE INTERNAL "Whether atomic operations for thread-MPI were found")
  17         else (TEST_ATOMICS)
  18             if (TEST_TMPI_ATOMICS_ONLY)
  19                 message(WARNING "Atomic operations not found for this CPU+compiler combination. Atomic operations 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.")
  20             else (TEST_TMPI_ATOMICS_ONLY)
  21                 message(WARNING "Atomic operations not found for this
  22             CPU+compiler combination. Thread support will be unbearably slow: disable threads. Atomic operations 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.")
  23             endif (TEST_TMPI_ATOMICS_ONLY)
  24             set(${VARIABLE} FALSE CACHE INTERNAL "Whether atomic operations for thread-MPI were found")
  25         endif(TEST_ATOMICS)
  26     endif(NOT DEFINED TMPI_ATOMICS)
  27 ENDMACRO(TEST_TMPI_ATOMICS VARIABLE)
  28
  29 MACRO(TMPI_MAKE_CXX_LIB)
  30     set(TMPI_CXX_LIB 1)
  31     # the C++ library
  32     set(THREAD_MPI_CXX_SRC
  33         thread_mpi/system_error.cpp )
  34 ENDMACRO(TMPI_MAKE_CXX_LIB)
  35
  36 test_tmpi_atomics(TMPI_ATOMICS)
  37
  38 # do we want to only the atomics of tMPI (with GPU + MPI)
  39 if(NOT TEST_TMPI_ATOMICS_ONLY)
  40 include(FindThreads)
  41 if (CMAKE_USE_PTHREADS_INIT)
  42     check_include_files(pthread.h    HAVE_PTHREAD_H)
  43     set(THREAD_PTHREADS 1)
  44     #add_definitions(-DTHREAD_PTHREADS)
  45     set(THREAD_MPI_SRC
  46         thread_mpi/alltoall.c      thread_mpi/p2p_protocol.c
  47         thread_mpi/barrier.c       thread_mpi/p2p_send_recv.c
  48         thread_mpi/bcast.c         thread_mpi/p2p_wait.c
  49         thread_mpi/collective.c    thread_mpi/profile.c
  50         thread_mpi/comm.c          thread_mpi/pthreads.c
  51         thread_mpi/errhandler.c    thread_mpi/reduce.c
  52         thread_mpi/event.c         thread_mpi/reduce_fast.c
  53         thread_mpi/gather.c        thread_mpi/scatter.c
  54         thread_mpi/group.c         thread_mpi/tmpi_init.c
  55         thread_mpi/topology.c      thread_mpi/list.c
  56         thread_mpi/type.c          thread_mpi/lock.c
  57         thread_mpi/numa_malloc.c   thread_mpi/once.c
  58         thread_mpi/scan.c          thread_mpi/tmpi_malloc.c)
  59     set(THREAD_LIB ${CMAKE_THREAD_LIBS_INIT})
  60 else (CMAKE_USE_PTHREADS_INIT)
  61     if (CMAKE_USE_WIN32_THREADS_INIT)
  62         set(THREAD_WINDOWS 1)
  63         #add_definitions(-DTHREAD_WINDOWS)
  64         set(THREAD_MPI_SRC
  65             thread_mpi/alltoall.c      thread_mpi/p2p_protocol.c
  66             thread_mpi/barrier.c       thread_mpi/p2p_send_recv.c
  67             thread_mpi/bcast.c         thread_mpi/p2p_wait.c
  68             thread_mpi/collective.c    thread_mpi/profile.c
  69             thread_mpi/comm.c
  70             thread_mpi/errhandler.c    thread_mpi/reduce.c
  71             thread_mpi/event.c         thread_mpi/reduce_fast.c
  72             thread_mpi/gather.c        thread_mpi/scatter.c
  73             thread_mpi/group.c         thread_mpi/tmpi_init.c
  74             thread_mpi/topology.c      thread_mpi/list.c
  75             thread_mpi/type.c          thread_mpi/lock.c
  76             thread_mpi/winthreads.c    thread_mpi/once.c
  77             thread_mpi/numa_malloc.c
  78             thread_mpi/scan.c          thread_mpi/tmpi_malloc.c)
  79         set(THREAD_LIBRARY )
  80     endif (CMAKE_USE_WIN32_THREADS_INIT)
  81 endif (CMAKE_USE_PTHREADS_INIT)
  82
  83
  84 # the spin-waiting option
  85 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)
  86 mark_as_advanced(THREAD_MPI_WAIT_FOR_NO_ONE)
  87 if (THREAD_MPI_WAIT_FOR_NO_ONE)
  88     add_definitions(-DTMPI_WAIT_FOR_NO_ONE)
  89 else (THREAD_MPI_WAIT_FOR_NO_ONE)
  90     add_definitions()
  91 endif (THREAD_MPI_WAIT_FOR_NO_ONE)
  92
  93
  94 # the copy buffer option
  95 option(THREAD_MPI_COPY_BUFFER "Use an intermediate copy buffer for small message sizes, to allow blocking sends to return quickly." ON)
  96 mark_as_advanced(THREAD_MPI_COPY_BUFFER)
  97 if (THREAD_MPI_COPY_BUFFER)
  98     add_definitions()
  99 else (THREAD_MPI_COPY_BUFFER)
 100     add_definitions(-DTMPI_NO_COPY_BUFFER)
 101 endif (THREAD_MPI_COPY_BUFFER)
 102
 103
 104 # the profiling option
 105 option(THREAD_MPI_PROFILING "Turn on simple MPI profiling." OFF)
 106 mark_as_advanced(THREAD_MPI_PROFILING)
 107 if (THREAD_MPI_PROFILING)
 108     add_definitions(-DTMPI_PROFILE)
 109 else (THREAD_MPI_PROFILING)
 110     add_definitions()
 111 endif (THREAD_MPI_PROFILING)
 112
 113 include(CheckCSourceCompiles)
 114
 115 # option to set affinity
 116 option(THREAD_MPI_SET_AFFINITY "Set thread affinity to a core if number of threads equal to number of hardware threads." ON)
 117 mark_as_advanced(THREAD_MPI_SET_AFFINITY)
 118 if (THREAD_MPI_SET_AFFINITY)
 119     add_definitions(-DTMPI_SET_AFFINITY)
 120 else (THREAD_MPI_SET_AFFINITY)
 121     add_definitions()
 122 endif (THREAD_MPI_SET_AFFINITY)
 123
 124 include(CheckFunctionExists)
 125 if (THREAD_PTHREADS)
 126     set(CMAKE_REQUIRED_LIBRARIES ${CMAKE_THREAD_LIBS_INIT})
 127     # check for sched_setaffinity
 128     check_c_source_compiles(
 129         "#define _GNU_SOURCE
 130 #include <pthread.h>
 131 #include <stdlib.h>
 132 #include <stdio.h>
 133 #include <errno.h>
 134 int main(void) { cpu_set_t set;
 135     CPU_ZERO(&set);
 136     CPU_SET(0, &set);
 137     pthread_setaffinity_np(pthread_self(), sizeof(set), &set);
 138     return 0;
 139 }"
 140         PTHREAD_SETAFFINITY
 141     )
 142     if (PTHREAD_SETAFFINITY)
 143         set(HAVE_PTHREAD_SETAFFINITY 1)
 144     endif (PTHREAD_SETAFFINITY)
 145 endif (THREAD_PTHREADS)
 146
 147
 148 # this runs on POSIX systems
 149 check_include_files(unistd.h        HAVE_UNISTD_H)
 150 check_include_files(sched.h         HAVE_SCHED_H)
 151 check_include_files(sys/time.h      HAVE_SYS_TIME_H)
 152 check_function_exists(sysconf       HAVE_SYSCONF)
 153 # this runs on windows
 154 #check_include_files(windows.h          HAVE_WINDOWS_H)
 155
 156 endif(NOT TEST_TMPI_ATOMICS_ONLY)