mpip · rainwoodman · Jul 7, 2015 · Sep 16, 2015 · Sep 16, 2015 · Jun 5, 2016
diff --git a/.gitmodules b/.gitmodules
@@ -0,0 +1,3 @@
+[submodule "fftw3"]
+	path = fftw3
+	url = https://github.com/rainwoodman/fftw3
diff --git a/Makefile.am b/Makefile.am
@@ -16,7 +16,7 @@ AM_DISTCHECK_CONFIGURE_FLAGS = \
   "LDFLAGS=@LDFLAGS@"
 
 # Traverse these subdirectories, the current one first (for config.h).
-SUBDIRS = kernel gcell util api .
+SUBDIRS = fftw3 kernel gcell util api .
 
 if ENABLE_TESTS
 SUBDIRS += tests
@@ -46,7 +46,7 @@ lib@PFFT_PREFIX@pfft@PREC_SUFFIX@@OPENMP_SUFFIX@_la_LIBADD = \
 	gcell/libgcell.la
 
 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)
+lib@PFFT_PREFIX@pfft@PREC_SUFFIX@@OPENMP_SUFFIX@_la_CFLAGS = $(OPENMP_CFLAGS) -I $(top_srcdir)/mpi -I $(top_srcdir)/api
 
 # Get Fortran compile rules that include preprocessing.
 include $(top_srcdir)/build-aux/fortran-rules.am
@@ -58,6 +58,8 @@ pfft1@PREC_SUFFIX@@OPENMP_SUFFIX@.pc: pfft.pc
 pkgconfigdir = $(libdir)/pkgconfig
 pkgconfig_DATA = pfft.pc
 
+DISTCHECK_CONFIGURE_FLAGS=--disable-shared --disable-doc --enable-mpi --enable-fortran
+
 #################################################################
 # Documentation
 #################################################################

diff --git a/api/Makefile.am b/api/Makefile.am
@@ -3,6 +3,10 @@ AM_CPPFLAGS = -I$(top_srcdir)/kernel
 
 # Directory of util.h
 AM_CPPFLAGS += -I$(top_srcdir)/util
+#
+# Directory of fftw3
+AM_CPPFLAGS += -I$(top_srcdir)/fftw3/api
+AM_CPPFLAGS += -I$(top_srcdir)/fftw3/mpi
 
 # OpenMP support
 AM_CFLAGS    = $(OPENMP_CFLAGS)

diff --git a/api/api-adv.c b/api/api-adv.c
@@ -22,11 +22,6 @@
 #include "ipfft.h"
 #include "util.h"
 
-static void calculate_3dto2d_blocks(
-    const INT *n, MPI_Comm comm_cart_3d,
-    INT *iblk);
-
-
 INT PX(local_size_many_dft)(
     int rnk_n, const INT *n, const INT *ni, const INT *no,
     INT howmany, const INT *iblock, const INT *oblock,
@@ -238,7 +233,7 @@ PX(gcplan) PX(plan_many_rgc)(
     )
 {
   int rnk_pm;
-  INT blk_3dto2d[3];
+  INT blk_for_remap[3];
   MPI_Comm *comms_pm;
   PX(gcplan) ths;
   MPI_Comm comm_cart = PX(assure_cart_comm)(comm);
@@ -256,13 +251,13 @@ PX(gcplan) PX(plan_many_rgc)(
   comms_pm = (MPI_Comm*) malloc(sizeof(MPI_Comm) * (size_t) rnk_pm);
   PX(split_cart_procmesh)(comm_cart, comms_pm);
 
-  if( PX(needs_3dto2d_remap)(rnk_n, comm_cart) ){
+  if( PX(needs_remap_nd)(rnk_n, comm_cart) ){
     /* 3d to 2d remap results in complicated blocks.
      * We ignore users input and use default block size. */
-    calculate_3dto2d_blocks(pn, comm_cart,
-        blk_3dto2d);
+    PX(remap_nd_calculate_blocks)(rnk_n, pn, comm_cart,
+        blk_for_remap);
     ths = PX(plan_rgc_internal)(
-        rnk_n, pn, howmany, blk_3dto2d, gc_below, gc_above,
+        rnk_n, pn, howmany, blk_for_remap, gc_below, gc_above,
         data, rnk_pm, comms_pm, comm_cart, gc_flags);
   } else
     ths = PX(plan_rgc_internal)(
@@ -279,19 +274,6 @@ PX(gcplan) PX(plan_many_rgc)(
   return ths;
 }
 
-static void calculate_3dto2d_blocks(
-    const INT *n, MPI_Comm comm_cart_3d,
-    INT *iblk
-    )
-{
-  int q0, q1, p0, p1;
-  INT mblk[3], oblk[3];
-
-  PX(get_procmesh_dims_2d)(comm_cart_3d, &p0, &p1, &q0, &q1);
-  PX(default_block_size_3dto2d)(n, p0, p1, q0, q1,
-      iblk, mblk, oblk);
-}
-
 PX(gcplan) PX(plan_many_cgc)(
     int rnk_n, const INT *n,
     INT howmany, const INT *block,

diff --git a/api/api-basic.c b/api/api-basic.c
@@ -177,7 +177,6 @@ static R check_array(
       case PFFTI_ARRAYTYPE_HERMITIAN_COMPLEX:
         err = cabs( ((C*)data)[k] - 0.5 * (d1 + conj(d2))); break;
     }
-
     if( err > maxerr )
       maxerr = err;
   }
@@ -258,7 +257,7 @@ static INT plain_index(
   INT k=0;
 
   for(INT t=0; t<rnk; t++)
-    k += k*n[t] + kvec[t];
+    k = k*n[t] + kvec[t];
 
   return k;
 }
@@ -1076,19 +1075,22 @@ static void PX(execute_full)(
     twiddle_input(ths, ths->in, ths->out, in, out);
   PFFT_FINISH_TIMING(ths->timer->itwiddle);
 
-  PFFT_START_TIMING(ths->comm_cart, ths->timer->remap_3dto2d[0]);
-  PX(execute_remap_3dto2d)(ths->remap_3dto2d[0], ths->in, ths->out, in, out);
-  PFFT_FINISH_TIMING(ths->timer->remap_3dto2d[0]);
+  PFFT_START_TIMING(ths->comm_cart, ths->timer->remap_nd[0]);
+  PX(execute_remap_nd)(ths->remap_nd[0], ths->in, ths->out, in, out);
+  PFFT_FINISH_TIMING(ths->timer->remap_nd[0]);
+
 
   execute_transposed(r, ths->serial_trafo, ths->global_remap,
       ths->timer->trafo, ths->timer->remap, ths->in, ths->out, in, out, ths->comm_cart);
 
   execute_transposed(r, &ths->serial_trafo[r+1], &ths->global_remap[r],
       &ths->timer->trafo[r+1], &ths->timer->remap[r], ths->in, ths->out, in, out, ths->comm_cart);
-
-  PFFT_START_TIMING(ths->comm_cart, ths->timer->remap_3dto2d[1]);
-  PX(execute_remap_3dto2d)(ths->remap_3dto2d[1], ths->in, ths->out, in, out);
-  PFFT_FINISH_TIMING(ths->timer->remap_3dto2d[1]);
+
+  PFFT_START_TIMING(ths->comm_cart, ths->timer->remap_nd[1]);
+
+  PX(execute_remap_nd)(ths->remap_nd[1], ths->in, ths->out, in, out);
+
+  PFFT_FINISH_TIMING(ths->timer->remap_nd[1]);
 
   /* twiddle outputs in order to get inputs shifted by n/2 */
   PFFT_START_TIMING(ths->comm_cart, ths->timer->otwiddle);

diff --git a/api/pfft.h b/api/pfft.h
@@ -499,8 +499,19 @@ BEGIN_C_DECLS
   PFFT_EXTERN PX(gctimer) PX(convert_vec2gctimer)(                                      \
       const double *times);                                                             \
   PFFT_EXTERN void PX(destroy_gctimer)(                                                 \
-      PX(gctimer) ths);
-
+      PX(gctimer) ths); \
+  \
+  PFFT_EXTERN R * PX(gather_array)(const int rnk_n, const int howmany, \
+    const R * local, \
+    const INT* local_start, \
+    const INT * local_n, \
+    const INT * global_n, \
+    MPI_Comm comm); \
+  void PX(print_array)(int rnk_n, \
+    int howmany, \
+    R * a, \
+    const INT * ni, \
+    MPI_Comm comm); \
 
 
 #define PFFT_MANGLE_DOUBLE(name) PFFT_CONCAT(pfft_, name)

diff --git a/bootstrap.sh b/bootstrap.sh
@@ -35,12 +35,12 @@
 alias libtoolize=$(type -p glibtoolize libtoolize | head -1)
 
 touch ChangeLog
+touch fftw3/ChangeLog
 
 echo "PLEASE IGNORE WARNINGS AND ERRORS"
-
 rm -rf autom4te.cache
 libtoolize
-autoreconf --verbose --install --force
+autoreconf --verbose --install --symlink --force
 
 rm -f config.cache
 

diff --git a/configure.ac b/configure.ac
@@ -31,7 +31,7 @@
 AC_PREREQ(2.64)
 
 # autoconf initialization
-AC_INIT([PFFT],[1.0.8-alpha],[michael.pippig@mathematik.tu-chemnitz.de],[pfft],[http://www.tu-chemnitz.de/~mpip/software/])
+AC_INIT([PFFT],[1.0.8-alpha3-fftw3],[michael.pippig@mathematik.tu-chemnitz.de],[pfft],[http://www.tu-chemnitz.de/~mpip/software/])
 m4_ifndef([AC_PACKAGE_URL],
 	  [AC_SUBST([PACKAGE_URL], [http://www.tu-chemnitz.de/~mpip/software/])])
 
@@ -263,6 +263,7 @@ SHARED_VERSION_INFO="0:0:0"
 # substitute SHARED_VERSION_INFO in generated Makefiles
 AC_SUBST(SHARED_VERSION_INFO)
 
+AC_DISABLE_SHARED dnl to hell with shared libraries
 
 ################################################################################
 # program checks
@@ -280,37 +281,6 @@ AC_SUBST(SHARED_VERSION_INFO)
 # May need sincos from libm.
 AC_CHECK_LIB([m], [sincos])
 
-# Check for FFTW3, MPI FFTW and threaded FFTW.
-if test "x$PRECISION" = "xs" ; then
-  AX_LIB_FFTW3F
-elif test "x$PRECISION" = "xl" ; then
-  AX_LIB_FFTW3L
-else
-  AX_LIB_FFTW3
-fi
-
-if test "x$ax_lib_fftw3" = "xno"; then
-  AC_MSG_ERROR([You do not seem to have the FFTW-3.3 library installed.]
-      [You can download it from http://www.fftw.org. If you have installed FFTW-3.3, ]
-      [make sure that this configure script can find it. See ./configure ]
-      [--help for more information.])
-fi
-
-if test "x$ax_lib_fftw3_mpi" = "xno"; then
-  AC_MSG_ERROR([You do not seem to have the MPI 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
-
-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
-
-
 ################################################################################
 # compiler options
 ################################################################################
@@ -482,4 +452,6 @@ AC_CONFIG_LINKS([tests/build_checks.sh:tests/build_checks.sh
                  tests/run_checks.sh:tests/run_checks.sh
                  tests/manual_c2c_3d.c:tests/manual_c2c_3d.c])
 
+AC_CONFIG_SUBDIRS([fftw3])
+
 AC_OUTPUT
diff --git a/fftw3 b/fftw3
diff --git a/gcell/Makefile.am b/gcell/Makefile.am
@@ -6,6 +6,10 @@ AM_CPPFLAGS += -I$(top_srcdir)/api
 
 # Directory of util.h
 AM_CPPFLAGS += -I$(top_srcdir)/util
+#
+# Directory of fftw3
+AM_CPPFLAGS += -I$(top_srcdir)/fftw3/api
+AM_CPPFLAGS += -I$(top_srcdir)/fftw3/mpi
 
 noinst_LTLIBRARIES = libgcell.la
 

diff --git a/kernel/Makefile.am b/kernel/Makefile.am
@@ -6,6 +6,10 @@ AM_CPPFLAGS += -I$(top_srcdir)/api
 
 # Directory of util.h
 AM_CPPFLAGS += -I$(top_srcdir)/util
+#
+# Directory of fftw
+AM_CPPFLAGS += -I$(top_srcdir)/fftw3/api
+AM_CPPFLAGS += -I$(top_srcdir)/fftw3/mpi
 
 noinst_LTLIBRARIES = libkernel.la
 
@@ -21,6 +25,8 @@ libkernel_la_SOURCES = \
 	partrafo.c \
 	sertrafo.c \
 	procmesh.c \
+	remap.c \
 	remap_3dto2d.c \
+	remap_2dto1d.c \
 	timer.c \
 	transpose.c