simplify OpenMP support

mpip · May 5, 2015 · bb2e9f1 · bb2e9f1
2 parents ad98d63 + d51ff30
commit bb2e9f1
Show file tree

Hide file tree

Showing 11 changed files with 91 additions and 86 deletions.
diff --git a/Makefile.am b/Makefile.am
@@ -28,50 +28,31 @@ endif
 
 EXTRA_DIST = bootstrap.sh CONVENTIONS pfft.pc.in
 
-LIBPFFT_OMP_LA =
-if ENABLE_OPENMP
-LIBPFFT_OMP_LA += lib@PFFT_PREFIX@pfft@PREC_SUFFIX@_omp.la
-endif
-
 # Libraries that are built and installed.
 # set library name according to precision
 if ENABLE_LIBRARY_INSTALL
-lib_LTLIBRARIES = lib@PFFT_PREFIX@pfft@PREC_SUFFIX@.la $(LIBPFFT_OMP_LA)
+lib_LTLIBRARIES = lib@PFFT_PREFIX@pfft@PREC_SUFFIX@@OPENMP_SUFFIX@.la
 else
-noinst_LTLIBRARIES = lib@PFFT_PREFIX@pfft@PREC_SUFFIX@.la $(LIBPFFT_OMP_LA)
+noinst_LTLIBRARIES = lib@PFFT_PREFIX@pfft@PREC_SUFFIX@@OPENMP_SUFFIX@.la
 endif
 
 
 # Libtool convenience libraries.
-lib@PFFT_PREFIX@pfft@PREC_SUFFIX@_la_SOURCES =
-lib@PFFT_PREFIX@pfft@PREC_SUFFIX@_la_LIBADD = \
+lib@PFFT_PREFIX@pfft@PREC_SUFFIX@@OPENMP_SUFFIX@_la_SOURCES =
+lib@PFFT_PREFIX@pfft@PREC_SUFFIX@@OPENMP_SUFFIX@_la_LIBADD = \
 	kernel/libkernel.la \
 	util/libutil.la \
 	api/libapi.la \
 	gcell/libgcell.la
 
-lib@PFFT_PREFIX@pfft@PREC_SUFFIX@_la_LDFLAGS = -no-undefined -version-info @SHARED_VERSION_INFO@
-
-
-if ENABLE_OPENMP
-# Libtool convenience libraries.
-lib@PFFT_PREFIX@pfft@PREC_SUFFIX@_omp_la_SOURCES =
-lib@PFFT_PREFIX@pfft@PREC_SUFFIX@_omp_la_LIBADD = \
-	kernel/libkernel.la \
-	util/libutil.la \
-	api/libapi_omp.la \
-	gcell/libgcell.la
-
-lib@PFFT_PREFIX@pfft@PREC_SUFFIX@_omp_la_LDFLAGS = -no-undefined -version-info @SHARED_VERSION_INFO@
-lib@PFFT_PREFIX@pfft@PREC_SUFFIX@_omp_la_CFLAGS = $(OPENMP_CFLAGS)
-endif
-
+lib@PFFT_PREFIX@pfft@PREC_SUFFIX@@OPENMP_SUFFIX@_la_LDFLAGS = -no-undefined -version-info @SHARED_VERSION_INFO@
+lib@PFFT_PREFIX@pfft@PREC_SUFFIX@@OPENMP_SUFFIX@_la_CFLAGS = $(OPENMP_CFLAGS)
 
 # Get Fortran compile rules that include preprocessing.
 include $(top_srcdir)/build-aux/fortran-rules.am
 
-pfft1@PREC_SUFFIX@.pc: pfft.pc
-	cp -f pfft.pc pfft1@PREC_SUFFIX@.pc
+pfft1@PREC_SUFFIX@@OPENMP_SUFFIX@.pc: pfft.pc
+	cp -f pfft.pc pfft1@PREC_SUFFIX@@OPENMP_SUFFIX@.pc
 
 
 #################################################################

diff --git a/api/Makefile.am b/api/Makefile.am
@@ -4,6 +4,10 @@ AM_CPPFLAGS = -I$(top_srcdir)/kernel
 # Directory of util.h
 AM_CPPFLAGS += -I$(top_srcdir)/util
 
+# OpenMP support
+AM_CFLAGS    = $(OPENMP_CFLAGS)
+AM_CPPFLAGS += $(OPENMP_CFLAGS)
+
 # Headers that are installed.
 if ENABLE_HEADER_INSTALL
 include_HEADERS = pfft.h
@@ -13,12 +17,7 @@ else
 noinst_HEADERS = pfft.h
 endif
 
-LIBAPI_OMP_LA =
-if ENABLE_OPENMP
-LIBAPI_OMP_LA += libapi_omp.la
-endif
-
-noinst_LTLIBRARIES = libapi.la $(LIBAPI_OMP_LA)
+noinst_LTLIBRARIES = libapi.la
 
 EXTRA_DIST = f03-api.sh f03-wrap.sh genf03-api.pl genf03-wrap.pl build-f03-api.sh
 
@@ -35,19 +34,3 @@ libapi_la_SOURCES = \
 	pfftl.f03.in \
 	f03-wrap.c
 
-if ENABLE_OPENMP
-libapi_omp_la_CFLAGS = $(OPENMP_CFLAGS) 
-
-# Group local sources into convenience lib.
-libapi_omp_la_SOURCES = \
-        api-basic.c \
-        api-adv.c \
-	pfft.f.in \
-	fortran-api.c \
-	fortran-mangling.h \
-	fortran-prototypes.h \
-	fortran-wrappers.h \
-	pfft.f03.in \
-	pfftl.f03.in \
-	f03-wrap.c
-endif
diff --git a/api/api-basic.c b/api/api-basic.c
@@ -190,13 +190,39 @@ static R check_array(
 
 
 
+#ifdef _OPENMP
+#include <omp.h>
+static int _nthreads;
+#endif
+
+void PX(plan_with_nthreads) (int nthreads){
+#ifdef _OPENMP
+  _nthreads = nthreads;
+  X(plan_with_nthreads)(nthreads);
+#endif
+}
+
+int PX(get_nthreads)() {
+#ifdef _OPENMP
+  return _nthreads;
+#else
+  return 1;
+#endif
+}
 
 /* wrappers for fftw init and cleanup */
 void PX(init) (void){
+#ifdef _OPENMP
+  X(init_threads)();
+  PX(plan_with_nthreads)(omp_get_max_threads());
+#endif
   XM(init)();
 }
 
 void PX(cleanup) (void){
+#ifdef _OPENMP
+  X(cleanup_threads());
+#endif
   XM(cleanup)();
 }
 
@@ -1277,12 +1303,3 @@ static void execute_transposed(
   PFFT_FINISH_TIMING(timer_trafo[t]);
 }
 
-#if _OPENMP
-void PX(plan_with_nthreads)(
-    int nthreads
-    )
-{
-  X(plan_with_nthreads)(nthreads);
-}
-#endif
-
diff --git a/api/pfft.h b/api/pfft.h
@@ -76,6 +76,8 @@ BEGIN_C_DECLS
                                                                                         \
   PFFT_EXTERN void PX(init)(void);                                                      \
   PFFT_EXTERN void PX(cleanup)(void);                                                   \
+  PFFT_EXTERN void PX(plan_with_nthreads)(int nthreads);                                \
+  PFFT_EXTERN int  PX(get_nthreads)(void);                                              \
                                                                                         \
   PFFT_EXTERN void *PX(malloc)(size_t n);                                               \
   PFFT_EXTERN R *PX(alloc_real)(size_t n);                                              \
@@ -380,8 +382,6 @@ BEGIN_C_DECLS
       int neededArgs, unsigned type,                                                    \
       void *parameter);                                                                 \
                                                                                         \
-  PFFT_EXTERN void PX(plan_with_nthreads)(int nthreads);                                \
-                                                                                        \
                                                                                         \
   PFFT_EXTERN void PX(reset_timer)(                                                     \
       PX(plan) ths);                                                                    \

diff --git a/configure.ac b/configure.ac
@@ -58,12 +58,6 @@ AM_INIT_AUTOMAKE([1.9.6 foreign])
 # options for customizing the build process
 ################################################################################
 
-# multithreaded code
-AC_ARG_ENABLE(threads, 
-  [AS_HELP_STRING([--enable-threads], [enable multithreaded code])],
-  enable_threads=$enableval, enable_threads=no)
-AM_CONDITIONAL([HAVE_THREADS], [test "x$enable_threads" = "xyes"])
-
 # debug mode
 AC_ARG_ENABLE(debug, 
   [AS_HELP_STRING([--enable-debug], [compile with extra runtime checks for debugging])],
@@ -229,6 +223,27 @@ fi
 AC_LANG_POP([Fortran])
 
 
+################################################################################
+# Parallel characteristics
+################################################################################
+
+# disable OpenMP on default as long as it does not work perfectly
+: ${enable_openmp=no}
+AC_OPENMP
+
+if test "$enable_openmp" = "yes"; then
+  OPENMP_SUFFIX="_omp"
+else
+  OPENMP_SUFFIX=
+fi
+AC_SUBST(OPENMP_SUFFIX) # append _omp to library pfft name
+AM_CONDITIONAL(ENABLE_OPENMP, test "x$enable_openmp" = "xyes")
+
+AC_LANG_PUSH([Fortran])
+: ${enable_openmp=no}
+AC_OPENMP
+AC_LANG_POP([Fortran])
+
 ################################################################################
 # Libtool initialization
 ################################################################################
@@ -265,15 +280,6 @@ AC_SUBST(SHARED_VERSION_INFO)
 # May need sincos from libm.
 AC_CHECK_LIB([m], [sincos])
 
-# Check for OpenMP.
-AC_ARG_ENABLE(openmp, [AS_HELP_STRING([--enable-openmp],[use OpenMP directives for parallelism])],
-  [enable_openmp=$enableval], [enable_openmp=no])
-if test "$enable_openmp" = "yes"; then
-  AX_OPENMP([],[AC_MSG_ERROR([Do not know how to enable OpenMP])])
-fi
-AM_CONDITIONAL(ENABLE_OPENMP, test "x$enable_openmp" = "xyes")
-AC_SUBST(OPENMP_CFLAGS)
-
 # Check for FFTW3, MPI FFTW and threaded FFTW.
 if test "x$PRECISION" = "xs" ; then
   AX_LIB_FFTW3F
@@ -284,8 +290,8 @@ else
 fi
 
 if test "x$ax_lib_fftw3" = "xno"; then
-  AC_MSG_ERROR([You do not seem to have the FFTW-3.3 library installed. You can ]
-      [download it from http://www.fftw.org. If you have installed FFTW-3.3, ]
+  AC_MSG_ERROR([You do not seem to have the FFTW-3.3 library installed.]
+      [You can download it from http://www.fftw.org. If you have installed FFTW-3.3, ]
       [make sure that this configure script can find it. See ./configure ]
       [--help for more information.])
 fi
@@ -297,8 +303,11 @@ if test "x$ax_lib_fftw3_mpi" = "xno"; then
       [for more information.])
 fi
 
-if  test "x$enable_threads" = "xyes" -a "x$ax_lib_fftw3_threads" = "xno"; then
-  AC_MSG_ERROR([You do not seem to have the threaded FFTW-3.3 library installed.])
+if  test "x$enable_openmp" = "xyes" -a "x$ax_lib_fftw3_openmp" = "xno"; then
+  AC_MSG_ERROR([You do not seem to have the OpenMP part of the FFTW-3.3 library installed.]
+      [You can download it from http://www.fftw.org. If you have installed FFTW-3.3, ]
+      [make sure that this configure script can find it. See ./configure --help]
+      [for more information.])
 fi
 
 

diff --git a/m4/ax_lib_fftw3.m4 b/m4/ax_lib_fftw3.m4
@@ -124,6 +124,7 @@ AC_DEFUN([_AX_LIB_FFTW3_CHECK],[
   fftw3_LDFLAGS=
   fftw3_PREFIX=
   fftw3_LIBS=
+  fftw3_openmp_LIBS=
   fftw3_threads_LIBS=
   fftw3_mpi_LIBS=
 
@@ -169,6 +170,7 @@ AC_DEFUN([_AX_LIB_FFTW3_CHECK],[
 
   ax_lib_fftw3=no
   ax_lib_fftw3_threads=no
+  ax_lib_fftw3_openmp=no
   ax_lib_fftw3_mpi=no
 
   fftw3_PREFIX="$ax_with_fftw3_prefix"
@@ -190,11 +192,11 @@ AC_DEFUN([_AX_LIB_FFTW3_CHECK],[
         ax_lib_fftw3_threads=no
         ;;
     esac
-    
+
     ax_lib_fftw3_openmp=yes
     case $ac_configure_args in
       *--enable-openmp*)
-        fftw3_openmp_LIBS="-l${ax_with_fftw3_prefix}fftw3${ax_type_suffix}_omp"
+        fftw3_threads_LIBS="-l${ax_with_fftw3_prefix}fftw3${ax_type_suffix}_omp"
         ;;
       *)
         ax_lib_fftw3_openmp=no

diff --git a/tests/Makefile.am b/tests/Makefile.am
@@ -20,8 +20,11 @@ AM_CPPFLAGS = -I$(top_srcdir)/kernel
 # Directory of pfft.h
 AM_CPPFLAGS += -I$(top_srcdir)/api
 
+# OpenMP support
+AM_CFLAGS = $(OPENMP_CFLAGS)
+
 # Libraries to add to all programs that are built.
-LDADD = $(top_builddir)/lib@PFFT_PREFIX@pfft@PREC_SUFFIX@.la $(fftw3_mpi_LIBS) $(fftw3_LIBS)
+LDADD = $(top_builddir)/lib@PFFT_PREFIX@pfft@PREC_SUFFIX@@OPENMP_SUFFIX@.la $(fftw3_openmp_LIBS) $(fftw3_mpi_LIBS) $(fftw3_LIBS)
 
 
 EXTRA_DIST = \

diff --git a/tests/bench_c2c.c b/tests/bench_c2c.c
@@ -438,6 +438,9 @@ static void init_parameters(
 
   pfft_printf(MPI_COMM_WORLD, "******************************************************************************************************\n");
   pfft_printf(MPI_COMM_WORLD, "* Computation of loops=%d parallel forward and backward FFTs (change with -pfft_loops *)\n", *loops);
+#ifdef _OPENMP
+  pfft_printf(MPI_COMM_WORLD, "* of with %d threads per process (change with OMP_NUM_THREADS=* before the command line) \n", pfft_get_nthreads());
+#endif
   pfft_printf(MPI_COMM_WORLD, "* for n[0] x n[1] x n[2] = %td x %td x %td Fourier coefficients (change with -pfft_n * * *)\n", n[0], n[1], n[2]);
   pfft_printf(MPI_COMM_WORLD, "* on  np[0] x np[1] x np[2] = %td x %td x %td processes (change with -pfft_np * * *)\n", np[0], np[1], np[2]);
 

diff --git a/tests/f03/Makefile.am b/tests/f03/Makefile.am
@@ -4,8 +4,11 @@ include $(top_srcdir)/build-aux/fortran-rules.am
 # Directory of pnfft.f (which is build first)
 AM_FCCPPFLAGS = -I$(top_builddir)/api
 
+# OpenMP support
+AM_FCFLAGS = $(OPENMP_FCFLAGS)
+
 # Libraries to add to all programs that are built.
-LDADD = $(top_builddir)/lib@PFFT_PREFIX@pfft@PREC_SUFFIX@.la $(fftw3_mpi_LIBS) $(fftw3_LIBS)
+LDADD = $(top_builddir)/lib@PFFT_PREFIX@pfft@PREC_SUFFIX@@OPENMP_SUFFIX@.la $(fftw3_openmp_LIBS) $(fftw3_mpi_LIBS) $(fftw3_LIBS)
 
 # noinst_LTLIBRARIES = libtests.la
 

diff --git a/tests/fortran/Makefile.am b/tests/fortran/Makefile.am
@@ -4,8 +4,11 @@ include $(top_srcdir)/build-aux/fortran-rules.am
 # Directory of pfft.f (which is build first)
 AM_FCCPPFLAGS = -I$(top_builddir)/api
 
+# OpenMP support
+AM_FCFLAGS = $(OPENMP_FCFLAGS)
+
 # Libraries to add to all programs that are built.
-LDADD = $(top_builddir)/lib@PFFT_PREFIX@pfft@PREC_SUFFIX@.la $(fftw3_mpi_LIBS) $(fftw3_LIBS)
+LDADD = $(top_builddir)/lib@PFFT_PREFIX@pfft@PREC_SUFFIX@@OPENMP_SUFFIX@.la $(fftw3_openmp_LIBS) $(fftw3_mpi_LIBS) $(fftw3_LIBS)
 
 # noinst_LTLIBRARIES = libtests.la
 

diff --git a/tests/openmp/Makefile.am b/tests/openmp/Makefile.am
@@ -4,11 +4,12 @@ AM_CPPFLAGS = -I$(top_srcdir)/kernel
 # Directory of pfft.h
 AM_CPPFLAGS += -I$(top_srcdir)/api
 
+# OpenMP support
+AM_CFLAGS = $(OPENMP_CFLAGS)
+
 # Libraries to add to all programs that are built.
 LDADD = $(top_builddir)/lib@PFFT_PREFIX@pfft@PREC_SUFFIX@_omp.la $(fftw3_openmp_LIBS) $(fftw3_mpi_LIBS) $(fftw3_LIBS)
 
-AM_CFLAGS = $(OPENMP_CFLAGS)
-
 # These programs are built by 'make check' and may be tested afterwards.
 check_PROGRAMS = \
 	simple_check_c2c simple_check_c2c_transposed\