The data type interface provides a mechanism to describe the storage format of individual data points of a data set and is hopefully designed in such a way as to allow new features to be easily added without disrupting applications that use the data type interface. A dataset (the H5D interface) is composed of a collection or raw data points of homogeneous type organized according to the data space (the H5S interface).
A data type is a collection of data type properties, all of which can be stored on disk, and which when taken as a whole, provide complete information for data conversion to or from that data type. The interface provides functions to set and query properties of a data type.
A data point is an instance of a data type, which is an instance of a type class. We have defined a set of type classes and properties which can be extended at a later time. The atomic type classes are those which describe types which cannot be decomposed at the data type interface level; all other classes are compound.
The functions defined in this section operate on data types as
a whole. New data types can be created from scratch or copied
from existing data types. When a data type is no longer needed
its resources should be released by calling H5Tclose()
.
Data types come in two flavors: named data types and transient
data types. A named data type is stored in a file while the
transient flavor is independent of any file. Named data types
are always read-only, but transient types come in three
varieties: modifiable, read-only, and immutable. The difference
between read-only and immutable types is that immutable types
cannot be closed except when the entire library is closed (the
predefined types like H5T_NATIVE_INT
are immutable
transient types).
hid_t H5Tcreate (H5T_class_t class, size_t
size)
H5T_COMPOUND
to create a new empty compound data
type where size is the total size in bytes of an
instance of this data type. Other data types are created with
H5Tcopy()
. All functions that return data type
identifiers return a negative value for failure.
hid_t H5Topen (hid_t location, const char
*name)
H5Tclose()
to
release resources. The named data type returned by this
function is read-only or a negative value is returned for
failure. The location is either a file or group
handle.
herr_t H5Tcommit (hid_t location, const char
*name, hid_t type)
hbool_t H5Tcommitted (hid_t type)
H5Dget_type()
are able to share
the data type with other datasets in the same file.
hid_t H5Tcopy (hid_t type)
herr_t H5Tclose (hid_t type)
hbool_t H5Tequal (hid_t type1, hid_t
type2)
TRUE
, otherwise it returns FALSE
(an
error results in a negative return value).
herr_t H5Tlock (hid_t type)
H5close()
or by normal program termination).
An atomic type is a type which cannot be decomposed into smaller units at the API level. All atomic types have a common set of properties which are augmented by properties specific to a particular type class. Some of these properties also apply to compound data types, but we discuss them only as they apply to atomic data types here. The properties and the functions that query and set their values are:
H5T_class_t H5Tget_class (hid_t type)
H5T_INTEGER, H5T_FLOAT, H5T_TIME, H5T_STRING,
H5T_BITFIELD
, or H5T_OPAQUE
. This
property is read-only and is set when the datatype is
created or copied (see H5Tcreate()
,
H5Tcopy()
). If this function fails it returns
H5T_NO_CLASS
which has a negative value (all
other class constants are non-negative).
size_t H5Tget_size (hid_t type)
herr_t H5Tset_size (hid_t type, size_t
size)
offset
property is
decremented a bit at a time. If the offset reaches zero and
the significant part of the data still extends beyond the edge
of the data type then the precision
property is
decremented a bit at a time. Decreasing the size of a data
type may fail if the precesion must be decremented and the
data type is of the H5T_OPAQUE
class or the
H5T_FLOAT
bit fields would extend beyond the
significant part of the type. Adjusting the size of an
H5T_STRING
automatically adjusts the precision
as well. On error, H5Tget_size()
returns zero
which is never a valid size.
H5T_order_t H5Tget_order (hid_t type)
herr_t H5Tset_order (hid_t type, H5T_order_t
order)
H5T_ORDER_LE
. If the bytes are in the oposite
order then they are said to be big-endian or
H5T_ORDER_BE
. Some data types have the same byte
order on all machines and are H5T_ORDER_NONE
(like character strings). If H5Tget_order()
fails then it returns H5T_ORDER_ERROR
which is a
negative value (all successful return values are
non-negative).
size_t H5Tget_precision (hid_t type)
herr_t H5Tset_precision (hid_t type, size_t
precision)
short
on a Cray
is 32 significant bits in an eight-byte field. The
precision
property identifies the number of
significant bits of a datatype and the offset
property (defined below) identifies its location. The
size
property defined above represents the entire
size (in bytes) of the data type. If the precision is
decreased then padding bits are inserted on the MSB side of
the significant bits (this will fail for
H5T_FLOAT
types if it results in the sign,
mantissa, or exponent bit field extending beyond the edge of
the significant bit field). On the other hand, if the
precision is increased so that it "hangs over" the edge of the
total size then the offset
property is
decremented a bit at a time. If the offset
reaches zero and the significant bits still hang over the
edge, then the total size is increased a byte at a time. The
precision of an H5T_STRING
is read-only and is
always eight times the value returned by
H5Tget_size()
. H5Tget_precision()
returns zero on failure since zero is never a valid precision.
size_t H5Tget_offset (hid_t type)
herr_t H5Tset_offset (hid_t type, size_t
offset)
precision
property defines the number
of significant bits, the offset
property defines
the location of those bits within the entire datum. The bits
of the entire data are numbered beginning at zero at the least
significant bit of the least significant byte (the byte at the
lowest memory address for a little-endian type or the byte at
the highest address for a big-endian type). The
offset
property defines the bit location of the
least signficant bit of a bit field whose length is
precision
. If the offset is increased so the
significant bits "hang over" the edge of the datum, then the
size
property is automatically incremented. The
offset is a read-only property of an H5T_STRING
and is always zero. H5Tget_offset()
returns zero
on failure which is also a valid offset, but is guaranteed to
succeed if a call to H5Tget_precision()
succeeds
with the same arguments.
herr_t H5Tget_pad (hid_t type, H5T_pad_t
*lsb, H5T_pad_t *msb)
herr_t H5Tset_pad (hid_t type, H5T_pad_t
lsb, H5T_pad_t msb)
precision
and offset
properties
are called padding. Padding falls into two
categories: padding in the low-numbered bits is lsb
padding and padding in the high-numbered bits is msb
padding (bits are numbered according to the description for
the offset
property). Padding bits can always be
set to zero (H5T_PAD_ZERO
) or always set to one
(H5T_PAD_ONE
). The current pad types are returned
through arguments of H5Tget_pad()
either of which
may be null pointers.
Integer atomic types (class=H5T_INTEGER
)
describe integer number formats. Such types include the
following information which describes the type completely and
allows conversion between various integer atomic types.
H5T_sign_t H5Tget_sign (hid_t type)
herr_t H5Tset_sign (hid_t type, H5T_sign_t
sign)
H5T_SGN_2
) or unsigned
(H5T_SGN_NONE
). Whether data is signed or not
becomes important when converting between two integer data
types of differing sizes as it determines how values are
truncated and sign extended.
The library supports floating-point atomic types
(class=H5T_FLOAT
) as long as the bits of the
exponent are contiguous and stored as a biased positive number,
the bits of the mantissa are contiguous and stored as a positive
magnitude, and a sign bit exists which is set for negative
values. Properties specific to floating-point types are:
herr_t H5Tget_fields (hid_t type, size_t
*spos, size_t *epos, size_t
*esize, size_t *mpos, size_t
*msize)
herr_t H5Tset_fields (hid_t type, size_t
spos, size_t epos, size_t esize,
size_t mpos, size_t msize)
precision
and offset
properties). The sign bit is always of length one and none of
the fields are allowed to overlap. When expanding a
floating-point type one should set the precision first; when
decreasing the size one should set the field positions and
sizes first.
size_t H5Tget_ebias (hid_t type)
herr_t H5Tset_ebias (hid_t type, size_t
ebias)
ebias
larger than the true exponent.
H5Tget_ebias()
returns zero on failure which is
also a valid exponent bias, but the function is guaranteed to
succeed if H5Tget_precision()
succeeds when
called with the same arguments.
H5T_norm_t H5Tget_norm (hid_t type)
herr_t H5Tset_norm (hid_t type, H5T_norm_t
norm)
H5T_NORM_MSBSET
then the
mantissa is shifted left (if non-zero) until the first bit
after the radix point is set and the exponent is adjusted
accordingly. All bits of the mantissa after the radix
point are stored.
H5T_NORM_IMPLIED
then the
mantissa is shifted left (if non-zero) until the first bit
after the radix point is set and the exponent is adjusted
accordingly. The first bit after the radix point is not stored
since it's always set.
H5T_NORM_NONE
then the fractional
part of the mantissa is stored without normalizing it.
H5T_pad_t H5Tget_inpad (hid_t type)
herr_t H5Tset_inpad (hid_t type, H5T_pad_t
inpad)
H5T_PAD_ZERO
if the internal
padding should always be set to zero, or H5T_PAD_ONE
if it should always be set to one.
H5Tget_inpad()
returns H5T_PAD_ERROR
on failure which is a negative value (successful return is
always non-negative).
Dates and times (class=H5T_TIME
) are stored as
character strings in one of the ISO-8601 formats like
"1997-12-05 16:25:30"; as character strings using the
Unix asctime(3) format like "Thu Dec 05 16:25:30 1997";
as an integer value by juxtaposition of the year, month, and
day-of-month, hour, minute and second in decimal like
19971205162530; as an integer value in Unix time(2)
format; or other variations.
I'm deferring definition until later since they're probably not as important as the other data types.
Fixed-length character string types are used to store textual
information. The offset
property of a string is
always zero and the precision
property is eight
times as large as the value returned by
H5Tget_size()
(since precision is measured in bits
while size is measured in bytes). Both properties are
read-only.
H5T_cset_t H5Tget_cset (hid_t type)
herr_t H5Tset_cset (hid_t type, H5T_cset_t
cset)
H5T_CSET_ASCII
.
H5T_str_t H5Tget_strpad (hid_t type)
herr_t H5Tset_strpad (hid_t type, H5T_str_t
strpad)
H5T_STR_NULLTERM
H5T_STR_NULLPAD
H5T_STR_NULLPAD
string will truncate but not null terminate. Conversion
from a short value to a longer value will append null
characters as with H5T_STR_NULLTERM
.
H5T_STR_SPACEPAD
H5T_STR_NULLPAD
except the padding character
is a space instead of a null.
H5Tget_strpad()
returns
H5T_STR_ERROR
on failure, a negative value (all
successful return values are non-negative).
Converting a bit field (class=H5T_BITFIELD
) from
one type to another simply copies the significant bits. If the
destination is smaller than the source then bits are truncated.
Otherwise new bits are filled according to the msb
padding type.
Opaque atomic types (class=H5T_OPAQUE
) act like
bit fields except conversions which change the precision are not
allowed. However, padding can be added or removed from either
end and the bytes can be reordered. Opaque types can be used to
create novel data types not directly supported by the library,
but the application is responsible for data conversion of these
types.
A compound data type is similar to a struct
in C
or a common block in Fortran: it is a collection of one or more
atomic types or small arrays of such types. Each
member of a compound type has a name which is unique
within that type, and a byte offset that determines the first
byte (smallest byte address) of that member in a compound datum.
A compound data type has the following properties:
H5T_class_t H5Tget_class (hid_t type)
H5T_COMPOUND
. This property is read-only and is
defined when a data type is created or copied (see
H5Tcreate()
or H5Tcopy()
).
size_t H5Tget_size (hid_t type)
int H5Tget_nmembers (hid_t type)
H5Tget_nmembers()
returns -1 on failure.
char *H5Tget_member_name (hid_t type, int
membno)
malloc()
or the null pointer on failure. The
caller is responsible for freeing the memory returned by this
function.
size_t H5Tget_member_offset (hid_t type, int
membno)
H5Tget_member_dims()
succeeds when called with the same type and
membno arguments.
int H5Tget_member_dims (hid_t type, int
membno, int dims[4], int
perm[4])
hid_t H5Tget_member_type (hid_t type, int
membno)
H5Tclose()
on that type.
Properties of members of a compound data type are
defined when the member is added to the compound type (see
H5Tinsert()
) and cannot be subsequently modified.
This makes it imposible to define recursive data structures.
The library predefines a modest number of data types having
names like H5T_arch_base
where
arch is an architecture name and base is a
programming type name. New types can be derived from the
predifined types by copying the predefined type (see
H5Tcopy()
) and then modifying the result.
Architecture Name | Description |
---|---|
IEEE |
This architecture defines standard floating point types in various byte orders. |
STD |
This is an architecture that contains semi-standard datatypes like signed two's complement integers, unsigned integers, and bitfields in various byte orders. |
UNIX |
Types which are specific to Unix operating systems are
defined in this architecture. The only type currently
defined is the Unix date and time types
(time_t ). |
C |
Types which are specific to the C or Fortran
programming languages are defined in these
architectures. For instance, H5T_C_STRING
defines a base string type with null termination which
can be used to derive string types of other
lengths. |
NATIVE |
This architecture contains C-like data types for the
machine on which the library was compiled. The types
were actually defined by running the
H5detect program when the library was
compiled. In order to be portable, applications should
almost always use this architecture to describe things
in memory. |
CRAY |
Cray architectures. These are word-addressable, big-endian systems with non-IEEE floating point. |
INTEL |
All Intel and compatible CPU's including 80286, 80386, 80486, Pentium, Pentium-Pro, and Pentium-II. These are little-endian systems with IEEE floating-point. |
MIPS |
All MIPS CPU's commonly used in SGI systems. These are big-endian systems with IEEE floating-point. |
ALPHA |
All DEC Alpha CPU's, little-endian systems with IEEE floating-point. |
The base name of most types consists of a letter, a precision in bits, and an indication of the byte order. The letters are:
B | Bitfield |
D | Date and time |
F | Floating point |
I | Signed integer |
S | Character string |
U | Unsigned integer |
The byte order is a two-letter sequence:
BE | Big endian |
LE | Little endian |
VX | Vax order |
Example |
Description |
---|---|
H5T_IEEE_F64LE |
Eight-byte, little-endian, IEEE floating-point |
H5T_IEEE_F32BE |
Four-byte, big-endian, IEEE floating point |
H5T_STD_I32LE |
Four-byte, little-endian, signed two's complement integer |
H5T_STD_U16BE |
Two-byte, big-endian, unsigned integer |
H5T_UNIX_D32LE |
Four-byte, little-endian, time_t |
H5T_C_S1 |
One-byte, null-terminated string of eight-bit characters |
H5T_INTEL_B64 |
Eight-byte bit field on an Intel CPU |
H5T_CRAY_F64 |
Eight-byte Cray floating point |
The NATIVE
architecture has base names which don't
follow the same rules as the others. Instead, native type names
are similar to the C type names. Here are some examples:
Example |
Corresponding C Type |
---|---|
H5T_NATIVE_CHAR |
signed char |
H5T_NATIVE_UCHAR |
unsigned char |
H5T_NATIVE_SHORT |
short |
H5T_NATIVE_USHORT |
unsigned short |
H5T_NATIVE_INT |
int |
H5T_NATIVE_UINT |
unsigned |
H5T_NATIVE_LONG |
long |
H5T_NATIVE_ULONG |
unsigned long |
H5T_NATIVE_LLONG |
long long |
H5T_NATIVE_ULLONG |
unsigned long long |
H5T_NATIVE_FLOAT |
float |
H5T_NATIVE_DOUBLE |
double |
H5T_NATIVE_LDOUBLE |
long double |
To create a 128-bit, little-endian signed integer type one could use the following (increasing the precision of a type automatically increases the total size):
|
To create an 80-byte null terminated string type one might do this (the offset of a character string is always zero and the precision is adjusted automatically to match the size):
|
Unlike atomic data types which are derived from other atomic data types, compound data types are created from scratch. First, one creates an empty compound data type and specifies it's total size. Then members are added to the compound data type in any order.
Usually a C struct will be defined to hold a data point in memory, and the offsets of the members in memory will be the offsets of the struct members from the beginning of an instance of the struct.
HOFFSET(s,m)
offsetof(s,m)
stddef.h
does
exactly the same thing as the HOFFSET()
macro.
Each member must have a descriptive name which is the key used to uniquely identify the member within the compound data type. A member name in an HDF5 data type does not necessarily have to be the same as the name of the member in the C struct, although this is often the case. Nor does one need to define all members of the C struct in the HDF5 compound data type (or vice versa).
An HDF5 data type is created to describe complex
numbers whose type is defined by the
|
Member alignment is handled by the HOFFSET
macro. However, data stored on disk does not require alignment,
so unaligned versions of compound data structures can be created
to improve space efficiency on disk. These unaligned compound
data types can be created by computing offsets by hand to
eliminate inter-member padding, or the members can be packed by
calling H5Tpack()
(which modifies a data type
directly, so it is usually preceded by a call to
H5Tcopy()
):
This example shows how to create a disk version of a compound data type in order to store data on disk in as compact a form as possible. Packed compound data types should generally not be used to describe memory as they may violate alignment constraints for the architecture being used. Note also that using a packed data type for disk storage may involve a higher data conversion cost.
|
Compound data types that have a compound data type member can be handled two ways. This example shows that the compound data type can be flattened, resulting in a compound type with only atomic members.
|
However, when the
|
If a file has lots of datasets which have a common data type then the file could be made smaller by having all the datasets share a single data type. Instead of storing a copy of the data type in each dataset object header, a single data type is stored and the object headers point to it. The space savings is probably only significant for datasets with a compound data type since the simple data types can be described with just a few bytes anyway.
To create a bunch of datasets that share a single data type just create the datasets with a committed (named) data type.
To create two datasets that share a common data type one just commits the data type, giving it a name, and then uses that data type to create the datasets.
And to create two additional datasets later which share the same type as the first two datasets:
|
The library is capable of converting data from one type to another and does so automatically when reading or writing the raw data of a dataset. The data type interface does not provide functions to the application for changing data types directly, but the user is allowed a certain amount of control over the conversion process.
In order to insure that data conversion exceeds disk I/O rates, common data conversion paths can be hand-tuned and optimized for performance. If a hand-tuned conversion function is not available, then the library falls back to a slower but more general conversion function. Although conversion paths include data space conversion, only data type conversions are described here. Most applications will not be concerned with data type conversions since the library will contain hand-tuned conversion functions for many common conversion paths. In fact, if an application does define a conversion function which would be of general interest, we request that the function be submitted to the HDF5 development team for inclusion in the library (there might be less overhead involved with calling an internal conversion functions than calling an application-defined conversion function).
Note: The alpha version of the library does not contain a full set of conversions. It can convert from one integer format to another and one struct to another. It can also perform byte swapping when the source and destination types are otherwise the same.
A conversion path contains a source and destination data type and each path contains a hard conversion function and/or a soft conversion function. The only difference between hard and soft functions is the way in which the library chooses which function applies: A hard function applies to a specific conversion path while a soft function may apply to multiple paths. When both hard and soft functions apply to a conversion path, then the hard function is favored and when multiple soft functions apply, the one defined last is favored.
A data conversion function is of type
The conversion function is called with the source and
destination data types (src_type and
dst_type), path-constant data (cdata), the
number of instances of the data type to convert
(nelmts), a buffer which initially contains an array of
data having the source type and on return will contain an array
of data having the destination type (buffer), and a
temporary or background buffer (background). Functions
return a negative value on failure and some other value on
success.
The Whether a background buffer is supplied to a conversion
function, and whether the background buffer is initialized
depends on the value of Other fields of cdata can be read or written by
the conversion functions. Many of these contain
performance-measuring fields which can be printed by the
conversion function during the Once a conversion function is written it can be registered and
unregistered with these functions:
Here's an example application-level function that
converts Cray The background argument is ignored since
it's generally not applicable to atomic data types.
The convesion function described in the previous
example applies to more than one conversion path.
Instead of enumerating all possible paths, we register
it as a soft function and allow it to decide which
paths it can handle.
This causes it to be consulted for any conversion
from an integer type to another integer type. The
first argument is just a short identifier which will
be printed with the data type conversion statistics.
NOTE: The idea of a master soft list and being able to
query conversion functions for their abilities tries to overcome
problems we saw with AIO. Namely, that there was a dichotomy
between generic conversions and specific conversions that made
it very difficult to write a conversion function that operated
on, say, integers of any size and order as long as they don't
have zero padding. The AIO mechanism required such a function
to be explicitly registered (like
H5T_conv_t
which is defined as:
typedef herr_t (*H5T_conv_t)(hid_t src_type,
hid_t dest_type,
H5T_cdata_t *cdata,
size_t nelmts,
void *buffer,
void *background);
command
field of the cdata argument
determines what happens within the conversion function. It's
values can be:
H5T_CONV_INIT
priv
field of cdata (or private data can
be initialized later). It should also initialize the
need_bkg
field described below. The buf
and background pointers will be null pointers.
H5T_CONV_CONV
priv
field of
cdata if it wasn't initialize during the
H5T_CONV_INIT
command and then convert
nelmts instances of the src_type to the
dst_type. The buffer serves as both input
and output. The background buffer is supplied
according to the value of the need_bkg
field of
cdata (the values are described below).
H5T_CONV_FREE
cdata->priv
pointer) should be freed and
set to null. All other pointer arguments are null and the
nelmts argument is zero.
cdata->need_bkg
which the conversion function should have initialized during the
H5T_CONV_INIT command. It can have one of these values:
H5T_BKG_NONE
H5T_BKG_TEMP
H5T_BKG_YES
H5T_CONV_FREE
command which is issued whenever the function is removed from a
conversion path.
hbool_t recalc
unsigned long ncalls
H5T_CONV_CONV
. It is updated automatically by
the library.
unsigned long nelmts
herr_t H5Tregister_hard (const char *name,
hid_t src_type, hid_t dest_type,
H5T_conv_t func)
H5Tregister_hard()
, displacing any previous hard
conversion for those paths. The name is used only
for debugging but must be supplied.
herr_t H5Tregister_soft (const char *name,
H5T_class_t src_class, H5T_class_t dest_class,
H5T_conv_t func)
herr_t H5Tunregister (H5T_conv_t func)
H5Tunregister()
. The
function is removed from all conversion paths.
Example: A conversion
function
unsigned short
to any other
16-bit unsigned big-endian integer. A cray
short
is a big-endian value which has 32
bits of precision in the high-order bits of a 64-bit
word.
1 typedef struct {
2 size_t dst_size;
3 int direction;
4 } cray_ushort2be_t;
5
6 herr_t
7 cray_ushort2be (hid_t src, hid_t dst,
8 H5T_cdata_t *cdata,
9 size_t nelmts, void *buf,
10 const void *background)
11 {
12 unsigned char *src = (unsigned char *)buf;
13 unsigned char *dst = src;
14 cray_ushort2be_t *priv = NULL;
15
16 switch (cdata->command) {
17 case H5T_CONV_INIT:
18 /*
19 * We are being queried to see if we handle this
20 * conversion. We can handle conversion from
21 * Cray unsigned short to any other big-endian
22 * unsigned integer that doesn't have padding.
23 */
24 if (!H5Tequal (src, H5T_CRAY_USHORT) ||
25 H5T_ORDER_BE != H5Tget_order (dst) ||
26 H5T_SGN_NONE != H5Tget_signed (dst) ||
27 8*H5Tget_size (dst) != H5Tget_precision (dst)) {
28 return -1;
29 }
30
31 /*
32 * Initialize private data. If the destination size
33 * is larger than the source size, then we must
34 * process the elements from right to left.
35 */
36 cdata->priv = priv = malloc (sizeof(cray_ushort2be_t));
37 priv->dst_size = H5Tget_size (dst);
38 if (priv->dst_size>8) {
39 priv->direction = -1;
40 } else {
41 priv->direction = 1;
42 }
43 break;
44
45 case H5T_CONV_FREE:
46 /*
47 * Free private data.
48 */
49 free (cdata->priv);
50 cdata->priv = NULL;
51 break;
52
53 case H5T_CONV_CONV:
54 /*
55 * Convert each element, watch out for overlap src
56 * with dst on the left-most element of the buffer.
57 */
58 priv = (cray_ushort2be_t *)(cdata->priv);
59 if (priv->direction<0) {
60 src += (nelmts - 1) * 8;
61 dst += (nelmts - 1) * dst_size;
62 }
63 for (i=0; i<n; i++) {
64 if (src==dst && dst_size<4) {
65 for (j=0; j<dst_size; j++) {
66 dst[j] = src[j+4-dst_size];
67 }
68 } else {
69 for (j=0; j<4 && j<dst_size; j++) {
70 dst[dst_size-(j+1)] = src[3-j];
71 }
72 for (j=4; j<dst_size; j++) {
73 dst[dst_size-(j+1)] = 0;
74 }
75 }
76 src += 8 * direction;
77 dst += dst_size * direction;
78 }
79 break;
80
81 default:
82 /*
83 * Unknown command.
84 */
85 return -1;
86 }
87 return 0;
88 }
Example: Soft
Registration
H5Tregister_soft ("cus2be", H5T_INTEGER, H5T_INTEGER, cray_ushort2be);
H5Tregister_hard()
) for each an every possible
conversion path whether that conversion path was actually used
or not.
Robb Matzke
Quincey Koziol
Last modified: Fri Aug 7 10:24:52 EDT 1998