Statistiques
| Révision :

root / man / man3 / HPL_pdpanrlN.3

Historique | Voir | Annoter | Télécharger (3,01 ko)

1 
.TH HPL_pdpanrlN 3 "September 10, 2008" "HPL 2.0" "HPL Library Functions"
2 
.SH NAME
3 
HPL_pdpanrlN \- Right-looking panel factorization.
4 
.SH SYNOPSIS
5 
\fB\&#include "hpl.h"\fR
6
7 
\fB\&void\fR
8 
\fB\&HPL_pdpanrlN(\fR
9 
\fB\&HPL_T_panel *\fR
10 
\fI\&PANEL\fR,
11 
\fB\&const int\fR
12 
\fI\&M\fR,
13 
\fB\&const int\fR
14 
\fI\&N\fR,
15 
\fB\&const int\fR
16 
\fI\&ICOFF\fR,
17 
\fB\&double *\fR
18 
\fI\&WORK\fR
19 
\fB\&);\fR
20 
.SH DESCRIPTION
21 
\fB\&HPL_pdpanrlN\fR
22 
factorizes  a panel of columns  that is a sub-array of a
23 
larger one-dimensional panel A using the Right-looking variant of the
24 
usual one-dimensional algorithm.  The lower triangular N0-by-N0 upper
25 
block of the panel is stored in no-transpose form (i.e. just like the
26 
input matrix itself).
27
28 
Bi-directional  exchange  is  used  to  perform  the  swap::broadcast
29 
operations  at once  for one column in the panel.  This  results in a
30 
lower number of slightly larger  messages than usual.  On P processes
31 
and assuming bi-directional links,  the running time of this function
32 
can be approximated by (when N is equal to N0):
33
34 
   N0 * log_2( P ) * ( lat + ( 2*N0 + 4 ) / bdwth ) +
35 
   N0^2 * ( M - N0/3 ) * gam2-3
36
37 
where M is the local number of rows of  the panel, lat and bdwth  are
38 
the latency and bandwidth of the network for  double  precision  real
39 
words, and  gam2-3  is  an estimate of the  Level 2 and Level 3  BLAS
40 
rate of execution. The  recursive  algorithm  allows indeed to almost
41 
achieve  Level 3 BLAS  performance  in the panel factorization.  On a
42 
large  number of modern machines,  this  operation is however latency
43 
bound,  meaning  that its cost can  be estimated  by only the latency
44 
portion N0 * log_2(P) * lat.  Mono-directional links will double this
45 
communication cost.
46
47 
Note that  one  iteration of the the main loop is unrolled. The local
48 
computation of the absolute value max of the next column is performed
49 
just after its update by the current column. This allows to bring the
50 
current column only  once through  cache at each  step.  The  current
51 
implementation  does not perform  any blocking  for  this sequence of
52 
BLAS operations, however the design allows for plugging in an optimal
53 
(machine-specific) specialized  BLAS-like kernel.  This idea has been
54 
suggested to us by Fred Gustavson, IBM T.J. Watson Research Center.
55 
.SH ARGUMENTS
56 
.TP 8
57 
PANEL   (local input/output)    HPL_T_panel *
58 
On entry,  PANEL  points to the data structure containing the
59 
panel information.
60 
.TP 8
61 
M       (local input)           const int
62 
On entry,  M specifies the local number of rows of sub(A).
63 
.TP 8
64 
N       (local input)           const int
65 
On entry,  N specifies the local number of columns of sub(A).
66 
.TP 8
67 
ICOFF   (global input)          const int
68 
On entry, ICOFF specifies the row and column offset of sub(A)
69 
in A.
70 
.TP 8
71 
WORK    (local workspace)       double *
72 
On entry, WORK  is a workarray of size at least 2*(4+2*N0).
73 
.SH SEE ALSO
74 
.BR HPL_dlocmax \ (3),
75 
.BR HPL_dlocswpN \ (3),
76 
.BR HPL_dlocswpT \ (3),
77 
.BR HPL_pdmxswp \ (3),
78 
.BR HPL_pdpancrN \ (3),
79 
.BR HPL_pdpancrT \ (3),
80 
.BR HPL_pdpanllN \ (3),
81 
.BR HPL_pdpanllT \ (3),
82 
.BR HPL_pdpanrlT \ (3).