configure.ac

-AC_INIT([nanox], [0.14], [pm-tools@bsc.es])
+AC_INIT([nanox], [0.14.1], [pm-tools@bsc.es])
 AC_LANG([C++])
 AC_CONFIG_SRCDIR([src/core/system.cpp])
 AC_DEFINE_UNQUOTED([NANOX_CONFIGURE_ARGS], ["$0 $@"], [Configure line])
@@ -22,7 +22,7 @@ AM_CONDITIONAL([SUPPORTED_SILENT_RULES], test x$supported_silent_rules = x1)
  
 # Nanos version check
 # Make sure build_version is double-quoted!
-AS_IF([test -d "$srcdir/.git"],[
+AS_IF([test -e "$srcdir/.git"],[
    # Git directory found
    AC_PATH_PROG([GIT], [git], [])
    AC_MSG_CHECKING([revision of git repository])

m4/ax_check_extrae.m4

@@ -10,6 +10,8 @@
 
 AC_DEFUN([AX_CHECK_EXTRAE],
 [
+   AC_REQUIRE([AC_PROG_AWK])
+
    MPITRACE_HOME=""
    MPITRACE_INC=""
    MPITRACE_LIB=""
@@ -25,15 +27,28 @@ AC_DEFUN([AX_CHECK_EXTRAE],
 
    AS_IF([test "x$with_extrae" != xno], [
       # Extrae 2.4 or higher is needed
-      # Header extrae_version.h exists until Extrae 3.4
+      # Header extrae_version.h exists until Extrae 3.4 and again since 3.5.3
       # API Extrae_get_version exists since Extrae 3.4
 
       AC_MSG_CHECKING([for Extrae version])
 
-      extrae_version_h=$with_extrae/include/extrae_version.h
-      AS_IF([test -e $extrae_version_h], [
-         extrae_version_string=$(sed -ne '/^\s*#\s*define\s\+EXTRAE_VERSION .*$/p' $extrae_version_h)
-         extrae_version=$(echo $extrae_version_string | sed -e 's/#define EXTRAE_VERSION EXTRAE_VERSION_NUMBER(\([0-9]*\),\([0-9]*\),\([0-9]*\).*$/\1\2\3/')
+      extrae_version_h="$with_extrae/include/extrae_version.h"
+      AS_IF([test -e "$extrae_version_h"], [
+         # Obtain version through Extrae header
+         extrae_version=$($AWK -v FS='[[(),]]' \
+            '/EXTRAE_VERSION_NUMBER\([[0-9]],[[0-9]],[[0-9]]\)/{print $[2]"."$[3]"."$[4];}' \
+            $extrae_version_h)
+         AS_IF([test -z "$extrae_version"], [
+            extrae_version=$($AWK \
+               '
+               BEGIN{major=""; minor=""; micro="";}
+               /EXTRAE_MAJOR [[0-9]]/{major=$[3];}
+               /EXTRAE_MINOR [[0-9]]/{minor=$[3];}
+               /EXTRAE_MICRO [[0-9]]/{micro=$[3];}
+               END{if (length(major)>0 && length(minor)>0 && length(micro)>0)
+                     print major"."minor"."micro;}
+               ' $extrae_version_h)
+         ])
       ], [
          # Obtain version through Extrae API
          AC_LANG_PUSH([C])
@@ -45,7 +60,7 @@ AC_DEFUN([AX_CHECK_EXTRAE],
                {
                   unsigned major, minor, revision;
                   Extrae_get_version(&major, &minor, &revision);
-                  printf("%1d%1d%1d\n", major, minor, revision);
+                  printf("%1d.%1d.%1d\n", major, minor, revision);
                   return 0;
                }
             ]])
@@ -67,15 +82,15 @@ AC_DEFUN([AX_CHECK_EXTRAE],
       ])
 
       AC_MSG_CHECKING([if Extrae library is compatible])
-      AS_IF([test "$extrae_version" -ge 240],[
-            MPITRACE_HOME="$withval"
-            MPITRACE_INC="$withval/include"
-            MPITRACE_LIB="$withval/lib"
-            AS_IF([test -d "$MPITRACE_HOME/lib64"],[
-               MPITRACE_LIB="$MPITRACE_HOME/lib64"
-            ])
-            MPITRACE_BIN="$withval/bin"
-            AC_MSG_RESULT([yes])
+      AX_COMPARE_VERSION([$extrae_version],[ge],[2.4], [
+         MPITRACE_HOME="$withval"
+         MPITRACE_INC="$withval/include"
+         MPITRACE_LIB="$withval/lib"
+         AS_IF([test -d "$MPITRACE_HOME/lib64"],[
+            MPITRACE_LIB="$MPITRACE_HOME/lib64"
+         ])
+         MPITRACE_BIN="$withval/bin"
+         AC_MSG_RESULT([yes])
       ],[
          AC_MSG_ERROR([no (Extrae >= 2.4 needed)])
       ])

m4/ax_check_hwloc.m4

@@ -55,16 +55,15 @@ AS_IF([test "x$with_hwloc" != xyes dnl
          -o "x$with_hwloc" != xno dnl
          -o "x$with_hwloc" != x dnl
 ],[
-  hwlocinc="-I $with_hwloc/include"
-  hwloc_h="$with_hwloc/include/hwloc.h"
-  AS_IF([test -d $with_hwloc/lib64],
-    [hwloclib="-L$with_hwloc/lib64 -Wl,-rpath,$with_hwloc/lib64"],
-    [hwloclib="-L$with_hwloc/lib -Wl,-rpath,$with_hwloc/lib"])dnl
-])dnl
-
-AS_IF([test "x$with_hwloc_include" != x],[
-  hwlocinc="-isystem $with_hwloc_include"
-  hwloc_h="$with_hwloc_include/hwloc.h"
+  AS_IF([test -d $with_hwloc/include], [
+    hwlocinc="-I $with_hwloc/include"
+    hwloc_h="$with_hwloc/include/hwloc.h"
+  ])
+  AS_IF([test -d $with_hwloc/lib64], [
+    hwloclib="-L$with_hwloc/lib64 -Wl,-rpath,$with_hwloc/lib64"
+  ], [test -d $with_hwloc/lib], [
+    hwloclib="-L$with_hwloc/lib -Wl,-rpath,$with_hwloc/lib"
+  ])dnl
 ])dnl
 
 AS_IF([test "x$with_hwloc_lib" != x],[
@@ -73,7 +72,7 @@ AS_IF([test "x$with_hwloc_lib" != x],[
 
 # This condition is satisfied even if $with_hwloc="yes" 
 # This happens when user leaves --with-value alone
-AS_IF([test "x$with_hwloc$with_hwloc_include$with_hwloc_lib" != x],[
+AS_IF([test "x$with_hwloc$with_hwloc_lib" != x],[
   AC_LANG_PUSH([C++])
 
   #tests if provided headers and libraries are usable and correct

src/arch/smp/smpthread.cpp

@@ -255,3 +255,15 @@ void SMPMultiThread::addThreadsFromPEs(unsigned int representingPEsCount, PE **r
       _threads.push_back( &( representingPEs[ i ]->startWorker( this ) ) );
    }
 }
+
+void SMPMultiThread::initializeDependent( void )
+{
+   BaseThread *tmpMyThread = myThread;
+   for ( unsigned int i = 0; i < _threads.size(); i++ ) {
+      //Change myThread so calls to myThread->... or getMythreadSafe()->...
+      //    work as expected and do not try call parent multithread (this)
+      myThread = _threads[ i ];
+      myThread->initializeDependent();
+   }
+   myThread = tmpMyThread;
+}

src/arch/smp/smpthread.hpp

@@ -145,6 +145,7 @@ namespace ext {
 
          void addThreadsFromPEs(unsigned int representingPEsCount, PE **representingPEs);
          virtual bool canBlock() { return false;}
+         virtual void initializeDependent( void );
    };
 } // namespace ext
 } // namespace nanos

src/core/basethread_decl.hpp

@@ -202,7 +202,6 @@ namespace nanos {
          ThreadTeam             *_nextTeam;      //!< If thread has no team, which team should it join
 
       private:
-         virtual void initializeDependent () = 0;
          virtual void runDependent () = 0;
 
          // These must be called through the Scheduler interface
@@ -238,6 +237,7 @@ namespace nanos {
          virtual void lock ();
          virtual void unlock ();
 
+         virtual void initializeDependent () = 0;
          virtual void start () = 0;
          virtual void finish ();
          void run();

src/core/instrumentation_decl.hpp

@@ -586,7 +586,7 @@ namespace nanos {
             /* 43 */ registerEventKey("dep-address", "Dependence address", true, EVENT_DEVELOPER );
             /* 44 */ registerEventKey("copy-data-in","WD id that is copying data in", true, EVENT_DEVELOPER );
             /* 45 */ registerEventKey("cache-copy-data-in","WD id that is copying data in", true, EVENT_DEVELOPER );
-            /* 46 */ registerEventKey("cache-copy-data-out","WD id that is copying data in", true, EVENT_DEVELOPER );
+            /* 46 */ registerEventKey("cache-copy-data-out","WD id that is copying data out", true, EVENT_DEVELOPER );
             /* 47 */ registerEventKey("sched-affinity-constraint","Constraint used by affinity scheduler", true, EVENT_DEVELOPER );
 
             /* 48*/ registerEventKey("in-mpi-runtime","Inside MPI runtime", true, EVENT_DEVELOPER );

src/pms/openmp/nanos_omp.h

@@ -45,7 +45,7 @@ NANOS_API_DECL(nanos_err_t, nanos_omp_set_implicit, ( nanos_wd_t uwd ));
 NANOS_API_DECL(int, nanos_omp_get_max_threads, ( void ));
 NANOS_API_DECL(int, nanos_omp_get_num_threads, ( void ));
 NANOS_API_DECL(int, nanos_omp_get_thread_num, ( void ));
-NANOS_API_DECL(int, nanos_omp_set_num_threads, ( int nthreads ));
+NANOS_API_DECL(void, nanos_omp_set_num_threads, ( int nthreads ));
 
 NANOS_API_DECL(nanos_ws_t, nanos_omp_find_worksharing, ( nanos_omp_sched_t kind ));
 NANOS_API_DECL(nanos_err_t, nanos_omp_get_schedule, ( nanos_omp_sched_t *kind, int *modifier ));

src/pms/openmp/omp_api.cpp

@@ -56,8 +56,8 @@ extern "C"
       omp_set_num_threads(*nthreads);
    }
 
-   int nanos_omp_set_num_threads ( int nthreads ) __attribute__ ((alias ("omp_set_num_threads")));
-   int nanos_omp_set_num_threads_ ( int nthreads ) __attribute__ ((alias ("omp_set_num_threads")));
+   void nanos_omp_set_num_threads ( int nthreads ) __attribute__ ((alias ("omp_set_num_threads")));
+   void nanos_omp_set_num_threads_ ( int nthreads ) __attribute__ ((alias ("omp_set_num_threads")));
 
    NANOS_API_DEF(int, omp_get_thread_num, ( void ))
    {

src/support/hwloc.cpp

@@ -70,8 +70,10 @@ void Hwloc::loadHwloc ()
       fatal_cond0( res != 0, "Could not load hwloc topology xml file." );
    }
 
+#if (HWLOC_API_VERSION >> 16) == 1
    // Enable GPU detection
    hwloc_topology_set_flags( _hwlocTopology, HWLOC_TOPOLOGY_FLAG_IO_DEVICES );
+#endif
 
    // Perform the topology detection.
    hwloc_topology_load( _hwlocTopology );