\par
\section{Prototypes and descriptions of {\tt GPart} methods}
\label{section:GPart:proto}
\par
This section contains brief descriptions including prototypes
of all methods that belong to the {\tt GPart} object.
% There are four families: 
% basics, 
% IO,
% initializers 
% and
% utilities.
There are no IO methods.
\par
\subsection{Basic methods}
\label{subsection:GPart:proto:basics}
\par
As usual, there are four basic methods to support object creation,
setting default fields, clearing any allocated data, and free'ing
the object.
\par
%=======================================================================
\begin{enumerate}
%-----------------------------------------------------------------------
\item
\begin{verbatim}
GPart * GPart_new ( void ) ;
\end{verbatim}
\index{GPart_new@{\tt GPart\_new()}}
This method simply allocates storage for the {\tt GPart} structure 
and then sets the default fields by a call to 
{\tt GPart\_setDefaultFields()}.
%-----------------------------------------------------------------------
\item
\begin{verbatim}
void GPart_setDefaultFields ( GPart *gpart ) ;
\end{verbatim}
\index{GPart_setDefaultFields@{\tt GPart\_setDefaultFields()}}
This method sets the structure's fields to default values:
{\tt id = -1}, {\tt nvtx = nvbnd = ncomp = 0},
{\tt g} = {\tt par} = {\tt fch} = {\tt sib} = {\tt NULL},
and the default fields for {\tt compidsIV}, {\tt cweightsIV}
and {\tt vtxMapIV} are set via calls to {\tt
IV\_setDefaultFields()}.
\par \noindent {\it Error checking:}
If {\tt gpart} is {\tt NULL},
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\item
\begin{verbatim}
void GPart_clearData ( GPart *gpart ) ;
\end{verbatim}
\index{GPart_clearData@{\tt GPart\_clearData()}}
The {\tt IV\_clearData()} method is called for the
{\tt compidsIV}, {\tt cweightsIV} and
{\tt vtxMapIV} objects.
The structure's fields are then set
with a call to {\tt GPart\_setDefaultFields()}.
Note, storage for the {\tt Graph} object {\tt gpart->graph} 
is {\bf not} free'd.
The {\tt GPart} object does not own its {\tt Graph} object,
it only uses it.
\par \noindent {\it Error checking:}
If {\tt gpart} is {\tt NULL},
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\item
\begin{verbatim}
void GPart_free ( GPart *gpart ) ;
\end{verbatim}
\index{GPart_free@{\tt GPart\_free()}}
This method releases any storage by a call to 
{\tt GPart\_clearData()} then free's the storage for the 
structure with a call to {\tt free()}.
\par \noindent {\it Error checking:}
If {\tt gpart} is {\tt NULL},
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\end{enumerate}
\par
\subsection{Initializer methods}
\label{subsection:GPart:proto:initializers}
\par
There are two initializer methods.
\par
%=======================================================================
\begin{enumerate}
%-----------------------------------------------------------------------
\item
\begin{verbatim}
void GPart_init ( GPart *gpart, Graph *graph ) ;
\end{verbatim}
\index{GPart_init@{\tt GPart\_init()}}
This method initializes the {\tt Gpart} object given a {\tt Graph}
object as input.
Any previous data is cleared with a call to 
{\tt GPart\_clearData()}.
The {\tt graph}, {\tt nvtx}, {\tt nvbnd} fields are set.
The {\tt compidsIV} and {\tt cweightsIV} {\tt IV} objects are 
initialized.
The remaining fields are not changed from their default values.
\par \noindent {\it Error checking:}
If {\tt gpart} or {\tt g} is {\tt NULL},
or if $\mbox{\tt g->nvtx} \le 0$,
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\item
\begin{verbatim}
void GPart_setMessageInfo ( GPart *gpart, int msglvl, FILE *msgFile ) ;
\end{verbatim}
\index{GPart_setMessageInfo@{\tt GPart\_setMessageInfo()}}
This method sets the {\tt msglvl} and {\tt msgFile} fields.
\par \noindent {\it Error checking:}
If {\tt gpart} is {\tt NULL},
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\end{enumerate}
\par
\subsection{Utility methods}
\label{subsection:GPart:proto:utilities}
\par
%=======================================================================
\begin{enumerate}
%-----------------------------------------------------------------------
\item
\begin{verbatim}
void GPart_setCweights ( GPart *gpart ) ;
\end{verbatim}
\index{GPart_setCweights@{\tt GPart\_setCweights()}}
This method sets the component weights vector {\tt cweightsIV}.
We assume that the {\tt compidsIV} vector has been set prior to
entering this method.
The weight of a component is not simply the sum of the weights of
the vertices with that component's id.
We accept the separator or multisector vertices (those {\tt v} with
{\tt compids[v] == 0}) but then find the connected components of
the remaining vertices, renumbering the {\tt compidsIV} vector
where necessary.
Thus, {\tt ncomp} and {\tt compidsIV} may be updated, 
and {\tt cweightsIV} is set.
\par \noindent {\it Error checking:}
If {\tt gpart} is {\tt NULL},
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\item
\begin{verbatim}
int GPart_sizeOf ( GPart *gpart ) ;
\end{verbatim}
\index{GPart_sizeOf@{\tt GPart\_sizeOf()}}
This method returns the number of bytes owned by the object.
This includes the structure itself, the {\tt compidsIV},
{\tt cweightsIV} and {\tt vtxMapIV} arrays (if present),
but {\tt not} the {\tt Graph} object.
\par \noindent {\it Error checking:}
If {\tt gpart} is {\tt NULL},
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\item
\begin{verbatim}
int GPart_validVtxSep ( GPart *gpart ) ;
\end{verbatim}
\index{GPart_validVtxSep@{\tt GPart\_validVtxSep()}}
This method returns 1 if the partition defined by the {\tt
compidsIV} vector has a valid vertex separator and zero otherwise.
When there is a valid vertex separator, there are no
adjacent vertices not in the multisector that belong to different
components (as defined by the {\tt compidsIV} vector).
\par \noindent {\it Error checking:}
If {\tt gpart} is {\tt NULL},
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\item
\begin{verbatim}
void GPart_split ( GPart *gpart ) ;
\end{verbatim}
\index{GPart_split@{\tt GPart\_split()}}
This method is used to split a subgraph during the nested
dissection process that builds a tree of {\tt GPart} objects.
We first generate a valid partition via the
{\tt GPart\_setCweights()} method,
and
then split the graph into its component subgraphs. 
Each subgraph is
assigned to a new child {\tt GPart} object.
The {\tt Graph} object for each subgraph is formed from the parent
graph using the {\tt Graph\_subGraph()} method.
This means that the storage for the adjacency lists of the subgraph 
is taken from the storage for the parent graph --- the lists are
mapped into the local ordering via the {\tt vtxMap} vector.
After {\tt GPart\_split(gpart)} is called, the adjacency lists for the
vertices in {\tt gpart->g} are no longer valid.
\par \noindent {\it Error checking:}
If {\tt gpart} or {\tt g} is {\tt NULL},
or if {\tt gpart->fch} is not {\tt NULL}
(meaning that the subgraph has already been split),
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\item
\begin{verbatim}
int GPart_vtxIsAdjToOneDomain ( GPart *gpart, int v, int *pdomid ) ;
\end{verbatim}
\index{GPart_vtxIsAdjToOneDomain@{\tt GPart\_vtxIsAdjToOneDomain()}}
This method determines whether the vertex {\tt v} is
adjacent to just one domain or not. 
We use this method to make a separator or multisector minimal.
If the vertex is adjacent to only one domain, the return value is
{\tt 1} and {\tt *pdomid} is set to the domain's id.
If a vertex is adjacent to zero or two or more domains, the return
value is zero.
If a vertex belongs to a domain, it is considered adjacent to that
domain.
\par \noindent {\it Error checking:}
If {\tt gpart}, {\tt g} or {\tt domid} is {\tt NULL},
or if {\tt v} is out of range
(i.e., ${\tt v} < 0$ or ${\tt nvtx} \le {\tt v}$),
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\item
\begin{verbatim}
IV * GPart_bndWeightsIV ( GPart *gpart ) ;
\end{verbatim}
\index{GPart_bndWeightsIV@{\tt GPart\_bndWeightsIV()}}
This method returns an {\tt IV} object that contains the weights of
the vertices on the boundaries of the components.
\par \noindent {\it Error checking:}
If {\tt gpart} or {\tt g} is {\tt NULL},
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\end{enumerate}
\par
\subsection{Domain decomposition methods}
\label{subsection:GPart:proto:domain-decomposition}
\par
There are presently two methods that create a domain decomposition 
of a graph or a subgraph.
\par
%=======================================================================
\begin{enumerate}
%-----------------------------------------------------------------------
\item
\begin{verbatim}
void GPart_DDviaFishnet ( GPart *gpart, double frac, int minweight,
                          int maxweight, int seed ) ;
\end{verbatim}
\index{GPart_DDviaFishnet@{\tt GPart\_DDviaFishnet()}}
This method generates a domain decomposition of a graph using the
{\it fishnet} algorithm (see \cite{ash97-DDSEP} for details).
On return, the {\tt compidsIV} vector is filled with component ids
and {\tt ncomp} is set with the number of domains.
The {\tt frac} parameter governs the exclusion of nodes of high
degree from the domain sets.
We have found this to be useful for some graphs.
Nodes of very high degree (relative to the average or mean degree)
can severely distort a domain decomposition.
We have found that setting {\tt frac} to four works well in
practice.
The {\tt minweight} and {\tt maxweight} parameters are the minimum
target weight and maximum target weight for a domain.
The {\tt seed} parameter is used to insert a degree of randomness
into the algorithm.
This allows us to make several runs and take the best partition.
\par \noindent {\it Error checking:}
If {\tt gpart} or {\tt g} is {\tt NULL},
or if ${\tt freeze} \le 0.0$,
or if ${\tt minweight} < 0$,
or if ${\tt maxweight} < 0$,
or if ${\tt minweight} \ge {\tt maxweight}$,
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\item
\begin{verbatim}
void GPart_DDviaProjection ( GPart *gpart, IV *DDmapIV ) ;
\end{verbatim}
\index{GPart_DDviaProjection@{\tt GPart\_DDviaProjection()}}
This method generates a domain decomposition for a subgraph by
projecting an existing domain decoposition for the original graph
onto the subgraph.
Using this method (as opposed to generating a domain decomposition
for each subgraph) can typically save 15\% of the overall time
to find the graph decomposition, though the resulting partition is
usually not as good.
\par \noindent {\it Error checking:}
If {\tt gpart} or {\tt DDmapIV} is {\tt NULL},
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\end{enumerate}
\par
\subsection{Methods to generate a 2-set partition}
\label{subsection:GPart:proto:2-set}
\par
These two methods are used to generate a 2-set partition
$[S,B,W]$ from a domain decomposition.
\par
%=======================================================================
\begin{enumerate}
%-----------------------------------------------------------------------
\item
\begin{verbatim}
double GPart_TwoSetViaBKL ( GPart *gpart, double alpha, int seed,
                            double cpus[] ) ;
\end{verbatim}
\index{GPart_TwoSetViaBKL@{\tt GPart\_TwoSetViaBKL()}}
This method takes a domain decomposition $\{\Phi, \Omega_1,
\ldots, \Omega_m\}$ defined by the {\tt compidsIV} vector
and generates a two set partition $[S,B,W]$.
We first compute the map from vertices to domains and segments
(the segments partition the interface nodes $\Phi$).
We then construct the bipartite graph that represents the
connectivity of the domains and segments.
Each segment is an ``edge'' that connects two ``adjacent'' domains.
This allows us to use a variant of the Kernighan-Lin algorithm to
find an ``edge'' separator formed of segments, 
which is really a vertex separator, a subset of $\Phi$.
The {\tt alpha} parameter is used in the cost function evaluation
for the partition,
$\displaystyle
\mbox{cost}([S,B,W]) = |S|\left(1 + \alpha
\frac{\max\{|B|,|W|\}}
{\min\{|B|,|W|\}}
\right)
$.
The {\tt seed} parameter is used to randomize the algorithm.
One can make several runs with different seeds and chose the best
partition.
The {\tt cpus[]} array is used to store execution times for
segments of the algorithm:
{\tt cpus[0]} stores the time to compute the domain/segment map;
{\tt cpus[2]} stores the time to create the domain/segment bipartite
graph;
{\tt cpus[3]} stores the time to find the bisector using the block
Kernighan-Lin algorithm.
\par \noindent {\it Error checking:}
If {\tt gpart} or {\tt cpus} is {\tt NULL},
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\item
\begin{verbatim}
IV * GPart_domSegMap ( GPart *gpart, int *pndom, int *pnseg ) ;
\end{verbatim}
\index{GPart_domSegMap@{\tt GPart\_domSegMap()}}
This method takes a domain decomposition as defined by the {\tt
compidsIV} vector and generates the map from the vertices to the
domains and segments that are used in the Block Kernighan-Lin
procedure to find an initial separator.
The map is returned in an {\tt IV} object, and the numbers of
domains and segments are set in the {\tt pndom} and {\tt pnseg}
addresses.
This method is called by {\tt GPart\_TwoSetViaBKL}.
\par \noindent {\it Error checking:}
If {\tt gpart}, {\tt g}, {\tt pndom} or {\tt pnseg} is {\tt NULL},
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\end{enumerate}
\par
\subsection{Methods to improve a 2-set partition}
\label{subsection:GPart:proto:improve}
\par
These methods are used to improve a 2-set partition $[S,B,W]$.
They hinge on finding a {\it wide separator} $Y$ and constructing a
better separator ${\widehat S} \subseteq Y$.
The {\tt alpha} parameter is used in the cost function
$\displaystyle
\mbox{cost}([S,B,W]) = |S|\left(1 + \alpha
\frac{\max\{|B|,|W|\}}
{\min\{|B|,|W|\}}
\right)
$.
\par
%=======================================================================
\begin{enumerate}
%-----------------------------------------------------------------------
\item
\begin{verbatim}
IV * GPart_identifyWideSep ( GPart *gpart, int nlayer1, int nlayer2 ) ;
\end{verbatim}
\index{GPart_identifyWideSep@{\tt GPart\_identifyWideSep()}}
This method takes a 2-set partition $[S,B,W]$ and identifies a {\it wide
separator} $Y$ that contains $S$.
The portions of $B$ and $W$ that are included in $Y$ are specified
using the {\tt nlayer1} and {\tt nlayer2} parameters.
If both are zero, then $Y$ is simply $S$.
If {\tt nlayer1 > 0}, then $Y$ contains all vertices 
in the first component whose distance is {\tt nlayer1} or less from
$S$, and similarly for {\tt nlayer2 > 0}.
The vertices in $Y$ are placed in an {\tt IV} object which is then
returned.
\par \noindent {\it Error checking:}
If {\tt gpart} or {\tt g} is {\tt NULL},
or if ${\tt nlevel1} < 0$ or ${\tt nlevel2} < 0$,
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\item
\begin{verbatim}
IV * GPart_makeYCmap ( GPart *gpart, IV *YVmapIV ) ;
\end{verbatim}
\index{GPart_makeYCmap@{\tt GPart\_makeYCmap()}}
This method contructs and returns an {\tt IV} object that is the
blueprint used to form the network.
The wide separator $Y$ can be partitioned into four disjoint sets
(though some may be empty):
\begin{eqnarray*}
Y_0 & = & \{y \in Y\ |\ 
            y \notin Adj(B \setminus Y)
%           y \cap Adj(B \setminus Y) = \emptyset
            \mbox{\ and\ }
            y \notin Adj(W \setminus Y)
%           y \cap Adj(W \setminus Y) = \emptyset 
\} \\
Y_1 & = & \{y \in Y\ |\ 
            y \in Adj(B \setminus Y)
%           y \cap Adj(B \setminus Y) \ne \emptyset
            \mbox{\ and\ }
            y \notin Adj(W \setminus Y)
%           y \cap Adj(W \setminus Y) = \emptyset 
\} \\
Y_2 & = & \{y \in Y\ |\ 
            y \notin Adj(B \setminus Y)
%           y \cap Adj(B \setminus Y) = \emptyset
            \mbox{\ and\ }
            y \in Adj(W \setminus Y)
%           y \cap Adj(W \setminus Y) \ne \emptyset 
\} \\
Y_3 & = & \{y \in Y\ |\ 
            y \in Adj(B \setminus Y)
%           y \cap Adj(B \setminus Y) \ne \emptyset
            \mbox{\ and\ }
            y \in Adj(W \setminus Y)
%           y \cap Adj(W \setminus Y) \ne \emptyset 
\}
\end{eqnarray*}
The {\tt YVmapIV} object contains the list of vertices in the wide
separator $Y$.
The {\tt IV} object that is returned, (called {\tt YCmapIV} in
the calling method) contains the subscripts of the $Y_0$, $Y_1$,
$Y_2$ or $Y_3$ sets that contains each vertex.
\par \noindent {\it Error checking:}
If {\tt gpart}, {\tt g} or {\tt YVmapIV} is {\tt NULL},
or if ${\tt nvtx} \le 0$,
or if {\tt YVmapIV} is empty,
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\item
\begin{verbatim}
void * GPart_smoothBy2layers ( GPart *gpart, int bipartite, float alpha ) ;
\end{verbatim}
\index{GPart_smoothBy2layers@{\tt GPart\_smoothBy2layers()}}
This method forms the wide separator $Y$ from two layers of
vertices, either $Y_B = S \cup (Adj(S) \cap B)$ 
or $Y_W = S \cup (Adj(S) \cap W)$.
(If $|B| \ge |W|$, we first look at $Y_B$ and if no improvement can
be made we look at $Y_W$, and the reverse if $|W| > |B|$.)
The {\tt bipartite} parameter defines the type of network problem
we solve.
The network induced by the wide separator $Y$ need not be bipartite,
and will not be bipartite if $Y_0 \ne \emptyset$ or $Y_3 \ne \emptyset$,
($Y_0$ and $Y_3$ are defined immediately above).
The $Y_3$ set is not important, since it must be included in any
separator ${\widehat S} \subseteq Y$.
When $Y_0$ is not empty, it forms a layer {\it between} $Y_1$ and $Y_2$,
and so the network is not bipartite.
We can force the network to be bipartite 
(set {\tt bipartite} to {\tt 1}) by moving all nodes in $Y_0$ to
the appropriate $Y_1$ or $Y_2$.
When the graph is unit-weight and the network is bipartite, 
we can use the Dulmage-Mendelsohn decomposition to find one 
or more separators of minimum weight.
In general, forcing a non-bipartite network to be bipartite results
in possibly a larger separator, so we do not recommend this option.
The capability is there to compare the Dulmage-Mendelsohn
decomposition smoothers with the more general (and robust) max
flow smoothers.
\par \noindent {\it Error checking:}
If {\tt gpart} is {\tt NULL},
or if ${\tt alpha} < 0.0$,
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\item
\begin{verbatim}
float * GPart_smoothYSep ( GPart *gpart, IV *YVmapIV, 
                           IV *YCmapIV, float alpha ) ;
\end{verbatim}
\index{GPart_smoothYSep@{\tt GPart\_smoothYSep()}}
This methods takes as input a 2-set partition $[S,B,W]$
(defined by {\tt gpart->compidsIV}), a wide separator $Y$
(defined by {\tt YVmapIV}) and a 
$\langle Y_0, Y_1, Y_2, Y_3 \rangle$ partition of $Y$
(defined by {\tt YCmapIV}) and attempts to find a better partition.
A max flow problem is solved on a network induced by
$\langle Y_0, Y_1, Y_2, Y_3 \rangle$.
Two min-cuts and the partitions they induce are examined and the
better partition is accepted if better than $[S,B,W]$.
The parameter {\tt alpha} is used in the partition's cost function,
and the cost of the best partition is returned.
\par \noindent {\it Error checking:}
If {\tt gpart}, {\tt YVmapIV} or {\tt YCmapIV} is {\tt NULL},
or if ${\tt alpha} < 0.0$,
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\item
\begin{verbatim}
float * GPart_smoothBisector ( GPart *gpart, int nlayer, float alpha ) ;
\end{verbatim}
\index{GPart_smoothBisector@{\tt GPart\_smoothBisector()}}
This method takes a two-set partition $[S,B,W]$ as defined by the
{\tt compidsIV} vector and improves it (if possible). 
The methods returns the cost of 
a (possibly) new two-set partition
$[{\widehat S}, {\widehat B}, {\widehat W}]$ 
defined by the {\tt compidsIV} vector.
The wide separator $Y$ that is constructed is {\it centered} around
$S$, i.e., $Y$ includes all nodes in $B$ and $W$ that are {\tt
nlayer} distance or less from $S$.
This method calls {\tt GPart\_smoothYSep()}.
\par \noindent {\it Error checking:}
If {\tt gpart} is {\tt NULL},
or if ${\tt nlevel} < 0$,
or if ${\tt alpha} < 0.0$,
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\end{enumerate}
\par
\subsection{Recursive Bisection method}
\label{subsection:GPart:proto:RB}
\par
There is presently one method to construct the domain/separator tree.
\par
%=======================================================================
\begin{enumerate}
%-----------------------------------------------------------------------
\item
\begin{verbatim}
DSTree * GPart_RBviaDDsep ( GPart *gpart, DDsepInfo *info ) ;
\end{verbatim}
\index{GPart_RBviaDDsep@{\tt GPart\_RBviaDDsep()}}
This method performs a recursive bisection of the graph using the
{\tt DDSEP} algorithm and returns a {\tt DSTree} object that represents
the domain/separator tree and the map from vertices to domains and
separators.
The {\tt DDsepInfo} structure contains all the parameters to the
different steps of the {\tt DDSEP} algorithm (the fishnet method to find
the domain decomposition, the Block Kernighan-Lin method to find an
initial separator, and solves a max flow problem to
improve the separator).
An attempt is made to split a subgraph if the weight of the
internal vertices of the subgraph exceeds {\tt info->maxcompweight}.
The cpu times for the different segments of the algorithm are
accumulated in fields of the {\tt DDsepInfo} object.
\par \noindent {\it Error checking:}
If {\tt gpart} or {\tt info} is {\tt NULL},
or if ${\tt nvtx} \le 0$,
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\end{enumerate}
\par
\subsection{{\tt DDsepInfo} methods}
\label{subsection:GPart:proto:DDsepInfo}
\par
The {\tt DDsepInfo} {\it helper}-object is used during the
{\tt DDSEP} recursive bisection process.
It stores the necessary input parameters for the different stages
of the {\tt DDSEP} algorithm and collects statistics about the
resulting partition.
\par
%=======================================================================
\begin{enumerate}
%-----------------------------------------------------------------------
\item
\begin{verbatim}
DDsepInfo * DDsepInfo_new ( void ) ;
\end{verbatim}
\index{DDepInfo_new@{\tt DDsepInfo\_new()}}
This method simply allocates storage for the {\tt DDsepInfo} structure 
and then sets the default fields by a call to 
{\tt DDsepInfo\_setDefaultFields()}.
%-----------------------------------------------------------------------
\item
\begin{verbatim}
void DDsepInfo_setDefaultFields ( DDsepInfo *info ) ;
\end{verbatim}
\index{DDepInfo_setDefaultFields@{\tt DDsepInfo\_setDefaultFields()}}
This method checks to see whether {\tt info} is {\tt NULL}.
If so, an error message is printed and the program exits.
Otherwise, the structure's fields are set to the following 
default values.
\begin{verbatim}
info->seed          =   1 ; info->cpuDD     =    0.0 ;
info->minweight     =  40 ; info->cpuMap    =    0.0 ;
info->maxweight     =  80 ; info->cpuBPG    =    0.0 ;
info->frac          = 4.0 ; info->cpuBKL    =    0.0 ;
info->alpha         = 1.0 ; info->cpuSmooth =    0.0 ;
info->maxcompweight = 500 ; info->cpuSplit  =    0.0 ;
info->ntreeobj      =   0 ; info->cpuTotal  =    0.0 ;
info->DDoption      =   1 ; info->msglvl    =      0 ;
info->nlayer        =   3 ; info->msgFile   = stdout ;
\end{verbatim}
\par \noindent {\it Error checking:}
If {\tt info} is {\tt NULL},
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\item
\begin{verbatim}
void DDsepInfo_clearData ( DDsepInfo *info ) ;
\end{verbatim}
\index{DDepInfo_clearData@{\tt DDsepInfo\_clearData()}}
This method checks to see whether {\tt info} is {\tt NULL}.
{\tt DDsepInfo\_setDefaultFields()} is
called to set the default values.
\par \noindent {\it Error checking:}
If {\tt info} is {\tt NULL},
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\item
\begin{verbatim}
void DDsepInfo_free ( DDsepInfo *info ) ;
\end{verbatim}
\index{DDepInfo_free@{\tt DDsepInfo\_free()}}
This method checks to see whether {\tt info} is {\tt NULL}.
If so, an error message is printed and the program exits.
Otherwise, it releases any storage by a call to 
{\tt DDsepInfo\_clearData()} then free's the storage for the 
structure with a call to {\tt free()}.
\par \noindent {\it Error checking:}
If {\tt info} is {\tt NULL},
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\item
\begin{verbatim}
void DDsepInfo_writeCpuTimes ( DDsepInfo *info, FILE *msgFile ) ;
\end{verbatim}
\index{DDepInfo_writeCpuTimes@{\tt DDsepInfo\_writeCpuTimes()}}
This method writes a breakdown of the CPU times in a meaningful
format.
Here is sample output.
\begin{verbatim}
 CPU breakdown for graph partition
               raw CPU   per cent
 misc       :      1.61    1.2%
 Split      :     24.68   17.7%
 find DD    :     12.13    8.7%
 DomSeg Map :     13.09    9.4%
 DomSeg BPG :      4.66    3.3%
 BKL        :      5.68    4.1%
 Smooth     :     77.83   55.7%
 Total      :    139.67  100.0%
\end{verbatim}
\par \noindent {\it Error checking:}
If {\tt info} or {\tt msgFile} is {\tt NULL},
an error message is printed and the program exits.
%-----------------------------------------------------------------------
\end{enumerate}