rename transpose-inplace to transpose-pairwise, as the algorithm is n…

…ot restricted to inplace operation

[empty commit message]

mpi/Makefile.am

@@ -12,7 +12,7 @@ CC=@MPICC@
 
 EXTRADIST = testsched.c
 
-TRANSPOSE_SRC = transpose-alltoall.c transpose-inplace.c transpose-radix2.c transpose-problem.c transpose-solve.c mpi-transpose.h
+TRANSPOSE_SRC = transpose-alltoall.c transpose-pairwise.c transpose-radix2.c transpose-problem.c transpose-solve.c mpi-transpose.h
 DFT_SRC = dft-serial.c dft-rank-geq2.c dft-rank-geq2-transposed.c dft-rank1.c dft-rank1-bigvec.c dft-problem.c dft-solve.c mpi-dft.h
 
 libfftw3@PREC_SUFFIX@_mpi_la_SOURCES = api.c any-true.c block.c choose-radix.c conf.c dtensor.c fftw3-mpi.h ifftw-mpi.h $(TRANSPOSE_SRC) $(DFT_SRC)

mpi/conf.c

@@ -24,7 +24,7 @@
 
 static const solvtab s =
 {
-     SOLVTAB(XM(transpose_inplace_register)),
+     SOLVTAB(XM(transpose_pairwise_register)),
      SOLVTAB(XM(transpose_alltoall_register)),
      SOLVTAB(XM(transpose_radix2_register)),
      SOLVTAB(XM(dft_rank_geq2_register)),

mpi/mpi-transpose.h

@@ -51,6 +51,6 @@ typedef plan_rdft plan_mpi_transpose;
   (type *)X(mkplan_rdft)(sizeof(type), adt, apply)
 
 /* various solvers */
-void XM(transpose_inplace_register)(planner *p);
+void XM(transpose_pairwise_register)(planner *p);
 void XM(transpose_alltoall_register)(planner *p);
 void XM(transpose_radix2_register)(planner *p);

mpi/testsched.c

@@ -26,7 +26,7 @@
    It is not used in FFTW 3, but I keep it around in case we ever want
    to play with this again or to change algorithms.  In particular, I
    used it to implement and test the fill1_comm_sched routine in
-   transpose-inplace.c, which allows us to create a schedule for one
+   transpose-pairwise.c, which allows us to create a schedule for one
    process at a time and is much more compact than the FFTW 2 code.
 
    Note that the scheduling algorithm is somewhat modified from that

mpi/transpose-inplace.c → mpi/transpose-pairwise.c

@@ -18,12 +18,12 @@
  *
  */
 
-/* Plans for distributed in-place transposes (or out-of-place, while
-   preserving the input).  These require a slightly tricky sequence of
-   exchanges between processors.  The amount of buffer space required
-   is proportional to the local size divided by the number of
-   processors (i.e. to total array size divided by the number of
-   processors squared). */
+/* Distributed transposes using a sequence of carefully scheduled
+   pairwise exchanges.  This has the advantage that it can be done
+   in-place, or out-of-place while preserving the input, using buffer
+   space proportional to the local size divided by the number of
+   processes (i.e. to the total array size divided by the number of
+   processes squared). */
 
 #include "mpi-transpose.h"
 #include <string.h>
@@ -154,7 +154,7 @@ static int applicable(const S *ego, const problem *p_,
 {
      const problem_mpi_transpose *p = (const problem_mpi_transpose *) p_;
      /* Note: this is *not* UGLY for out-of-place, destroy-input plans;
-	the planner often prefers transpose-inplace to transpose-alltoall,
+	the planner often prefers transpose-pairwise to transpose-alltoall,
 	at least with LAM MPI on my machine. */
      return (1
 	     && (!ego->preserve_input || (!NO_DESTROY_INPUTP(plnr)
@@ -186,7 +186,7 @@ static void destroy(plan *ego_)
 static void print(const plan *ego_, printer *p)
 {
      const P *ego = (const P *) ego_;
-     p->print(p, "(mpi-transpose-inplace%s%(%p%)%(%p%)%(%p%)%(%p%))", 
+     p->print(p, "(mpi-transpose-pairwise%s%(%p%)%(%p%)%(%p%)%(%p%))", 
 	      ego->preserve_input==2 ?"/p":"",
 	      ego->cld1, ego->cld2, ego->cld2rest, ego->cld3);
 }
@@ -441,7 +441,7 @@ static solver *mksolver(int preserve_input)
      return &(slv->super);
 }
 
-void XM(transpose_inplace_register)(planner *p)
+void XM(transpose_pairwise_register)(planner *p)
 {
      int preserve_input;
      for (preserve_input = 0; preserve_input <= 1; ++preserve_input)