|
d@0
|
1 <html lang="en">
|
|
d@0
|
2 <head>
|
|
d@0
|
3 <title>An improved replacement for MPI_Alltoall - FFTW 3.2alpha3</title>
|
|
d@0
|
4 <meta http-equiv="Content-Type" content="text/html">
|
|
d@0
|
5 <meta name="description" content="FFTW 3.2alpha3">
|
|
d@0
|
6 <meta name="generator" content="makeinfo 4.8">
|
|
d@0
|
7 <link title="Top" rel="start" href="index.html#Top">
|
|
d@0
|
8 <link rel="up" href="FFTW-MPI-Transposes.html#FFTW-MPI-Transposes" title="FFTW MPI Transposes">
|
|
d@0
|
9 <link rel="prev" href="Advanced-distributed_002dtranspose-interface.html#Advanced-distributed_002dtranspose-interface" title="Advanced distributed-transpose interface">
|
|
d@0
|
10 <link href="http://www.gnu.org/software/texinfo/" rel="generator-home" title="Texinfo Homepage">
|
|
d@0
|
11 <!--
|
|
d@0
|
12 This manual is for FFTW
|
|
d@0
|
13 (version 3.2alpha3, 14 August 2007).
|
|
d@0
|
14
|
|
d@0
|
15 Copyright (C) 2003 Matteo Frigo.
|
|
d@0
|
16
|
|
d@0
|
17 Copyright (C) 2003 Massachusetts Institute of Technology.
|
|
d@0
|
18
|
|
d@0
|
19 Permission is granted to make and distribute verbatim copies of
|
|
d@0
|
20 this manual provided the copyright notice and this permission
|
|
d@0
|
21 notice are preserved on all copies.
|
|
d@0
|
22
|
|
d@0
|
23 Permission is granted to copy and distribute modified versions of
|
|
d@0
|
24 this manual under the conditions for verbatim copying, provided
|
|
d@0
|
25 that the entire resulting derived work is distributed under the
|
|
d@0
|
26 terms of a permission notice identical to this one.
|
|
d@0
|
27
|
|
d@0
|
28 Permission is granted to copy and distribute translations of this
|
|
d@0
|
29 manual into another language, under the above conditions for
|
|
d@0
|
30 modified versions, except that this permission notice may be
|
|
d@0
|
31 stated in a translation approved by the Free Software Foundation.
|
|
d@0
|
32 -->
|
|
d@0
|
33 <meta http-equiv="Content-Style-Type" content="text/css">
|
|
d@0
|
34 <style type="text/css"><!--
|
|
d@0
|
35 pre.display { font-family:inherit }
|
|
d@0
|
36 pre.format { font-family:inherit }
|
|
d@0
|
37 pre.smalldisplay { font-family:inherit; font-size:smaller }
|
|
d@0
|
38 pre.smallformat { font-family:inherit; font-size:smaller }
|
|
d@0
|
39 pre.smallexample { font-size:smaller }
|
|
d@0
|
40 pre.smalllisp { font-size:smaller }
|
|
d@0
|
41 span.sc { font-variant:small-caps }
|
|
d@0
|
42 span.roman { font-family:serif; font-weight:normal; }
|
|
d@0
|
43 span.sansserif { font-family:sans-serif; font-weight:normal; }
|
|
d@0
|
44 --></style>
|
|
d@0
|
45 </head>
|
|
d@0
|
46 <body>
|
|
d@0
|
47 <div class="node">
|
|
d@0
|
48 <p>
|
|
d@0
|
49 <a name="An-improved-replacement-for-MPI_Alltoall"></a>
|
|
d@0
|
50 <a name="An-improved-replacement-for-MPI_005fAlltoall"></a>
|
|
d@0
|
51 Previous: <a rel="previous" accesskey="p" href="Advanced-distributed_002dtranspose-interface.html#Advanced-distributed_002dtranspose-interface">Advanced distributed-transpose interface</a>,
|
|
d@0
|
52 Up: <a rel="up" accesskey="u" href="FFTW-MPI-Transposes.html#FFTW-MPI-Transposes">FFTW MPI Transposes</a>
|
|
d@0
|
53 <hr>
|
|
d@0
|
54 </div>
|
|
d@0
|
55
|
|
d@0
|
56 <h4 class="subsection">6.7.3 An improved replacement for MPI_Alltoall</h4>
|
|
d@0
|
57
|
|
d@0
|
58 <p>We close this section by noting that FFTW's MPI transpose routines can
|
|
d@0
|
59 be thought of as a generalization for the <code>MPI_Alltoall</code> function
|
|
d@0
|
60 (albeit only for floating-point types), and in some circumstances can
|
|
d@0
|
61 function as an improved replacement.
|
|
d@0
|
62 <a name="index-MPI_005fAlltoall-376"></a>
|
|
d@0
|
63 <code>MPI_Alltoall</code> is defined by the MPI standard as:
|
|
d@0
|
64
|
|
d@0
|
65 <pre class="example"> int MPI_Alltoall(void *sendbuf, int sendcount, MPI_Datatype sendtype,
|
|
d@0
|
66 void *recvbuf, int recvcnt, MPI_Datatype recvtype,
|
|
d@0
|
67 MPI_Comm comm);
|
|
d@0
|
68 </pre>
|
|
d@0
|
69 <p>In particular, for <code>double*</code> arrays <code>in</code> and <code>out</code>,
|
|
d@0
|
70 consider the call:
|
|
d@0
|
71
|
|
d@0
|
72 <pre class="example"> MPI_Alltoall(in, howmany, MPI_DOUBLE, out, howmany MPI_DOUBLE, comm);
|
|
d@0
|
73 </pre>
|
|
d@0
|
74 <p>This is completely equivalent to:
|
|
d@0
|
75
|
|
d@0
|
76 <pre class="example"> MPI_Comm_size(comm, &P);
|
|
d@0
|
77 plan = fftw_mpi_plan_many_transpose(P, P, howmany, 1, 1, in, out, comm, FFTW_ESTIMATE);
|
|
d@0
|
78 fftw_execute(plan);
|
|
d@0
|
79 fftw_destroy_plan(plan);
|
|
d@0
|
80 </pre>
|
|
d@0
|
81 <p>That is, computing a P × P transpose on <code>P</code> processes,
|
|
d@0
|
82 with a block size of 1, is just a standard all-to-all communication.
|
|
d@0
|
83
|
|
d@0
|
84 <p>However, using the FFTW routine instead of <code>MPI_Alltoall</code> may
|
|
d@0
|
85 have certain advantages. First of all, FFTW's routine can operate
|
|
d@0
|
86 in-place <code>in == out</code>) whereas <code>MPI_Alltoall</code> can only
|
|
d@0
|
87 operate out-of-place.
|
|
d@0
|
88 <a name="index-in_002dplace-377"></a>
|
|
d@0
|
89 Second, even for out-of-place plans, FFTW's routine may be faster,
|
|
d@0
|
90 especially if you need to perform the all-to-all communication many
|
|
d@0
|
91 times and can afford to use <code>FFTW_MEASURE</code> or
|
|
d@0
|
92 <code>FFTW_PATIENT</code>. It should certainly be no slower, not including
|
|
d@0
|
93 the time to create the plan, since one of the possible algorithms that
|
|
d@0
|
94 FFTW uses for an out-of-place transpose <em>is</em> simply to call
|
|
d@0
|
95 <code>MPI_Alltoall</code>. However, FFTW also considers several other
|
|
d@0
|
96 possible algorithms that, depending on your MPI implementation and
|
|
d@0
|
97 your hardware, may be faster.
|
|
d@0
|
98 <a name="index-FFTW_005fMEASURE-378"></a><a name="index-FFTW_005fPATIENT-379"></a>
|
|
d@0
|
99 <!-- -->
|
|
d@0
|
100
|
|
d@0
|
101 </body></html>
|
|
d@0
|
102
|
