fixes so that MPI code compiles when MPICC is a C++ compiler, even if…

… the serial code is compiled with a C compiler; thanks to Kyle Spyksma for the bug report
squidruge · Aug 5, 2011 · 6fbb063 · 6fbb063
1 parent d7feb2d
commit 6fbb063
Show file tree

Hide file tree

Showing 10 changed files with 46 additions and 10 deletions.
diff --git a/api/api.h b/api/api.h
@@ -51,6 +51,12 @@
 
 #include "fftw3.h"
 #include "ifftw.h"
+#include "rdft.h"
+
+#ifdef __cplusplus
+extern "C"
+{
+#endif /* __cplusplus */
 
 /* the API ``plan'' contains both the kernel plan and problem */
 struct X(plan_s) {
@@ -95,4 +101,10 @@ void X(mapflags)(planner *, unsigned);
 
 apiplan *X(mkapiplan)(int sign, unsigned flags, problem *prb);
 
+rdft_kind *X(map_r2r_kind)(int rank, const X(r2r_kind) * kind);
+
+#ifdef __cplusplus
+}  /* extern "C" */
+#endif /* __cplusplus */
+
 #endif				/* __API_H__ */
diff --git a/api/plan-guru-r2r.h b/api/plan-guru-r2r.h
@@ -21,8 +21,6 @@
 #include "api.h"
 #include "rdft.h"
 
-rdft_kind *X(map_r2r_kind)(int rank, const X(r2r_kind) * kind);
-
 X(plan) XGURU(r2r)(int rank, const IODIM *dims,
 			 int howmany_rank,
 			 const IODIM *howmany_dims,

diff --git a/api/plan-many-r2r.c b/api/plan-many-r2r.c
@@ -21,8 +21,6 @@
 #include "api.h"
 #include "rdft.h"
 
-rdft_kind *X(map_r2r_kind)(int rank, const X(r2r_kind) * kind);
-
 #define N0(nembed)((nembed) ? (nembed) : n)
 
 X(plan) X(plan_many_r2r)(int rank, const int *n,

diff --git a/dft/dft.h b/dft/dft.h
@@ -25,6 +25,11 @@
 #include "ifftw.h"
 #include "codelet-dft.h"
 
+#ifdef __cplusplus
+extern "C"
+{
+#endif /* __cplusplus */
+
 /* problem.c: */
 typedef struct {
      problem super;
@@ -76,4 +81,8 @@ void X(ct_genericbuf_register)(planner *p);
 /* configurations */
 void X(dft_conf_standard)(planner *p);
 
+#ifdef __cplusplus
+}  /* extern "C" */
+#endif /* __cplusplus */
+
 #endif /* __DFT_H__ */
diff --git a/kernel/ifftw.h b/kernel/ifftw.h
@@ -41,6 +41,11 @@
 # include <inttypes.h>           /* uintptr_t, maybe */
 #endif
 
+#ifdef __cplusplus
+extern "C"
+{
+#endif /* __cplusplus */
+
 /* Windows annoyances -- since tests/hook.c uses some internal
    FFTW functions, we need to given them the dllexport attribute
    under Windows when compiling as a DLL (see api/fftw3.h). */
@@ -1120,4 +1125,8 @@ static __inline__ E FNMS(E a, E b, E c)
 #define FNMS(a, b, c) ((c) - ((a) * (b)))
 #endif
 
+#ifdef __cplusplus
+}  /* extern "C" */
+#endif /* __cplusplus */
+
 #endif /* __IFFTW_H__ */
diff --git a/mpi/api.c b/mpi/api.c
@@ -172,7 +172,7 @@ static dtensor *default_sz(int rnk, const XM(ddim) *dims0, int n_pes,
 	block sizes weren't specified (i.e. 0), then set the
 	unspecified blocks so as to use as many processes as
 	possible with as few distributed dimensions as possible. */
-     for (k = IB; k <= OB; ++k) {
+     FORALL_BLOCK_KIND(k) {
 	  INT nb = XM(num_blocks_total)(sz, k);
 	  INT np = n_pes / nb;
 	  for (i = 0; i < rnk && np > 1; ++i)
@@ -635,8 +635,6 @@ X(plan) XM(plan_dft_3d)(ptrdiff_t nx, ptrdiff_t ny, ptrdiff_t nz,
 /*************************************************************************/
 /* R2R API */
 
-extern rdft_kind *X(map_r2r_kind)(int rank, const X(r2r_kind) * kind);
-
 X(plan) XM(plan_guru_r2r)(int rnk, const XM(ddim) *dims0,
 			  ptrdiff_t howmany,
 			  R *in, R *out,

diff --git a/mpi/dtensor.c b/mpi/dtensor.c
@@ -100,7 +100,7 @@ dtensor *XM(dtensor_canonical)(const dtensor *sz, int compress)
      for (i = rnk = 0; i < sz->rnk; ++i) {
 	  if (!compress || sz->dims[i].n > 1) {
                x->dims[rnk].n = sz->dims[i].n;
-	       for (k = IB; k <= OB; ++k) {
+	       FORALL_BLOCK_KIND(k) {
 		    if (XM(num_blocks)(sz->dims[i].n, sz->dims[i].b[k]) == 1)
 			 x->dims[rnk].b[k] = sz->dims[i].n;
 		    else

diff --git a/mpi/ifftw-mpi.h b/mpi/ifftw-mpi.h
@@ -63,6 +63,9 @@ typedef struct {
      INT b[2]; /* b[IB], b[OB] */
 } ddim;
 
+/* Loop over k in {IB, OB}.  Note: need explicit casts for C++. */
+#define FORALL_BLOCK_KIND(k) for (k = IB; k <= OB; k = (block_kind) (((int) k) + 1))
+
 /* unlike tensors in the serial FFTW, the ordering of the dtensor
    dimensions matters - both the array and the block layout are
    row-major order. */
@@ -136,7 +139,7 @@ typedef enum {
 
 /* skipping SQUARE_AFTER since it doesn't seem to offer any advantage
    over SQUARE_BEFORE */
-#define FORALL_REARRANGE(rearrange) for (rearrange = CONTIG; rearrange <= SQUARE_MIDDLE; ++rearrange)
+#define FORALL_REARRANGE(rearrange) for (rearrange = CONTIG; rearrange <= SQUARE_MIDDLE; rearrange = (rearrangement) (((int) rearrange) + 1))
 
 int XM(rearrange_applicable)(rearrangement rearrange, 
 			     ddim dim0, INT vn, int n_pes);

diff --git a/mpi/mpi-bench.c b/mpi/mpi-bench.c
@@ -8,8 +8,8 @@
 #include <math.h>
 #include <stdio.h>
 #include <string.h>
-#include "fftw-bench.h"
 #include "fftw3-mpi.h"
+#include "fftw-bench.h"
 
 #if defined(BENCHFFT_SINGLE)
 #  define BENCH_MPI_TYPE MPI_FLOAT

diff --git a/rdft/rdft.h b/rdft/rdft.h
@@ -24,6 +24,11 @@
 #include "ifftw.h"
 #include "codelet-rdft.h"
 
+#ifdef __cplusplus
+extern "C"
+{
+#endif /* __cplusplus */
+
 /* problem.c: */
 typedef struct {
      problem super;
@@ -164,4 +169,8 @@ void X(rdft2_rank_geq2_register)(planner *p);
 /* configurations */
 void X(rdft_conf_standard)(planner *p);
 
+#ifdef __cplusplus
+}  /* extern "C" */
+#endif /* __cplusplus */
+
 #endif /* __RDFT_H__ */