Add documentation for SPV file format.

[pspp] / doc / dev / spv-file-format.texi

@node SPSS Viewer File Format
@chapter SPSS Viewer File Format

SPSS Viewer or @file{.spv} files, here called SPV files, are written
by SPSS 16 and later to represent the contents of its output editor.
This chapter documents the format, based on examination of a corpus of
about 500 files from a variety of sources.  This description is
detailed enough to read SPV files, but probably not enough to write
them.

SPSS 15 and earlier versions use a completely different output format
based on the Microsoft Compound Document Format.  This format is not
documented here.

An SPV file is a Zip archive that can be read with @command{zipinfo}
and @command{unzip} and similar programs.  The final member in the Zip
archive is a file named @file{META-INF/MANIFEST.MF}.  This structure
makes SPV files resemble Java ``JAR'' files (and ODF files), but
whereas a JAR manifest contains a sequence of colon-delimited
key/value pairs, an SPV manifest contains the string
@samp{allowPivoting=true}, without a new-line.  (This string may be
the best way to identify an SPV file; it is invariant across the
corpus.)

The rest of the members in an SPV file's Zip archive fall into two
categories: @dfn{structure} and @dfn{detail} members.  Structure
member names begin with @file{outputViewer@var{nnnnnnnnnn}}, where
each @var{n} is a decimal digit, and end with @file{.xml}, and often
include the string @file{_heading} in between.  Each of these members
represents some kind of output item (a table, a heading, a block of
text, etc.) or a group of them.  The member whose output goes at the
beginning of the document is numbered 0, the next member in the output
is numbered 1, and so on.

Structure members contain XML.  This XML is sometimes self-contained,
but it often references detail members in the Zip archive, which named
as follows:

@table @asis
@item @file{@var{prefix}_table.xml} and @file{@var{prefix}_tableData.bin}
@itemx @file{@var{prefix}_lightTableData.bin}
The structure of a table plus its data.  Older SPV files pair a
@file{@var{prefix}_table.xml} file that describes the table's
structure with a binary @file{@var{prefix}_tableData.bin} file that
gives its data.  Newer SPV files (the majority of those in the corpus)
instead include a single @file{@var{prefix}_lightTableData.bin} file
that incorporates both into a single binary format.

@item @file{@var{prefix}_warning.xml} and @file{@var{prefix}_warningData.bin}
@itemx @file{@var{prefix}_lightWarningData.bin}
Same format used for tables, with a different name.

@item @file{@var{prefix}_notes.xml} and @file{@var{prefix}_notesData.bin}
@itemx @file{@var{prefix}_lightNotesData.bin}
Same format used for tables, with a different name.

@item @file{@var{prefix}_chartData.bin} and @file{@var{prefix}_chart.xml}
The structure of a chart plus its data.  Charts do not have a
``light'' format.

@item @file{@var{prefix}_model.scf}
@itemx @file{@var{prefix}_pmml.scf}
Not yet investigated.  The corpus contains only one example of each.

@item @file{@var{prefix}_stats.xml}
Not yet investigated.  The corpus contains few examples.
@end table

The @file{@var{prefix}} in the names of the detail members is
typically an 11-digit decimal number that increases for each item,
tending to skip values.  Older SPV files use different naming
conventions.  Structure member refer to detail members by name, and so
their exact names do not matter to readers as long as they are unique.

@menu
* SPV Structure Member Format::
* SPV Light Detail Member Format::
* SPV Legacy Detail Member Binary Format::
@end menu

@node SPV Structure Member Format
@section Structure Member Format

Structure members' XML files claim conformance with a collection of
XML Schemas.  These schemas are distributed, under a nonfree license,
with SPSS binaries.  Fortunately, the schemas are not necessary to
understand the structure members.  To a degree, the schemas can even
be deceptive because they document elements and attributes that are
not in the corpus and do not document elements and attributes that are
commonly found there.

Structure members use a different XML namespace for each schema, but
these namespaces are not entirely consistent.  In some SPV files, for
example, the @code{viewer-tree} schema is associated with namespace
@indicateurl{http://xml.spss.com/spss/viewer-tree} and in others with
@indicateurl{http://xml.spss.com/spss/viewer/viewer-tree} (note the
additional @file{viewer/}).  Under either name, the schema URIs are
not resolvable to obtain the schemas themselves.

One may ignore all of the above in interpreting a structure member.
The actual XML has a simple and straightforward form that does not
require a reader to take schemas or namespaces into account.

The elements found in structure members are documented below.  For
each element, we note the possible parent elements and the element's
contents.  The contents are specified as pseudo-regular expressions
with the following conventions:

@table @asis
@item text
XML text content.

@item CDATA
XML CDATA content.

@item @code{element}
The named element.

@item (@dots{})
Grouping multiple elements.

@item [@var{x}]
An optional @var{x}.

@item @var{a} @math{|} @var{b}
A choice between @var{a} and @var{b}.

@item @var{x}*
Zero or more @var{x}.
@end table

@ifnottex
For a diagram illustrating the hierarchy of elements within an SPV
structure member, please refer to a PDF version of the manual.
@end ifnottex

@iftex
The following diagram shows the hierarchy of elements within an SPV
structure member.  Edges point from parent to child elements.
Unlabeled edges indicate that the child appears exactly once; edges
labeled with *, zero or more times; edges labeled with ?, zero or one
times.
@center @image{dev/spv-structure, 5in}
@end iftex

@menu
* SPV heading Element::
* SPV label Element::
* SPV container Element::
* SPV text Element (Inside @code{container})::
* SPV html Element::
* SPV table Element::
* SPV tableStructure Element::
* SPV dataPath Element::
* SPV pageSetup Element::
* SPV pageHeader and pageFooter Elements::
* SPV pageParagraph Element::
* SPV @code{text} Element (Inside @code{pageParagraph})::
@end menu

@node SPV heading Element
@subsection The @code{heading} Element

Parent: Document root or @code{heading} @*
Contents: [@code{pageSetup}] @code{label} (@code{container} @math{|} @code{heading})*

The root of a structure member is a @code{heading}, which represents a
section of output beginning with a title (the @code{label}) and
ordinarily followed by content containers or further nested
(sub)-sections of output.

The document root heading, only, may also contain a @code{pageSetup}
element.

The following attributes have been observed on both document root and
nested @code{heading} elements.

@defvr {Optional} creator-version
The version of the software that created this SPV file.  A string of
the form @code{xxyyzzww} represents software version xx.yy.zz.ww,
e.g.@: @code{21000001} is version 21.0.0.1.  Trailing pairs of zeros
are sometimes omitted, so that @code{21}, @code{210000}, and
@code{21000000} are all version 21.0.0.0 (and the corpus contains all
three of those forms).
@end defvr

@noindent
The following attributes have been observed on document root
@code{heading} elements only:

@defvr {Optional} @code{creator}
The directory in the file system of the software that created this SPV
file.
@end defvr

@defvr {Optional} @code{creation-date-time}
The date and time at which the SPV file was written, in a
locale-specific format, e.g. @code{Friday, May 16, 2014 6:47:37 PM
PDT} or @code{lunedì 17 marzo 2014 3.15.48 CET} or even @code{Friday,
December 5, 2014 5:00:19 o'clock PM EST}.
@end defvr

@defvr {Optional} @code{lockReader}
Whether a reader should be allowed to edit the output.  The possible
values are @code{true} and @code{false}, but the corpus only contains
@code{false}.
@end defvr

@defvr {Optional} @code{schemaLocation}
This is actually an XML Namespace attribute.  A reader may ignore it.
@end defvr

@noindent
The following attributes have been observed only on nested
@code{heading} elements:

@defvr {Required} @code{commandName}
The locale-invariant name of the command that produced the output,
e.g.@: @code{Frequencies}, @code{T-Test}, @code{Non Par Corr}.
@end defvr

@defvr {Optional} @code{visibility}
To what degree the output represented by the element is visible.  The
only observed value is @code{collapsed}.
@end defvr

@defvr {Optional} @code{locale}
The locale used for output, in Windows format, which is similar to the
format used in Unix with the underscore replaced by a hyphen, e.g.@:
@code{en-US}, @code{en-GB}, @code{el-GR}, @code{sr-Cryl-RS}.
@end defvr

@defvr {Optional} @code{olang}
The output language, e.g.@: @code{en}, @code{it}, @code{es},
@code{de}, @code{pt-BR}.
@end defvr

@node SPV label Element
@subsection The @code{label} Element

Parent: @code{heading} or @code{container} @*
Contents: text

Every @code{heading} and @code{container} holds a @code{label} as its
first child.  The root @code{heading} in a structure member always
contains the string ``Output''.  Otherwise, the text in @code{label}
describes what it labels, often by naming the statistical procedure
that was executed, e.g.@: ``Frequencies'' or ``T-Test''.  Labels are
often very generic, especially within a @code{container}, e.g.@:
``Title'' or ``Warnings'' or ``Notes''.  Label text is localized
according to the output language, e.g.@: in Italian a frequency table
procedure is labeled ``Frequenze''.

The corpus contains one example of an empty label, one that contains
no text.

This element has no attributes.

@node SPV container Element
@subsection The @code{container} Element

Parent: @code{heading} @*
Contents: @code{label} [@code{table} @math{|} @code{text}]

A @code{container} serves to label a @code{table} or a @code{text}
item.

This element has the following attributes.

@defvr {Required} @code{visibility}
Either @code{visible} or @code{hidden}, this indicates whether the
container's content is displayed.
@end defvr

@defvr {Optional} @code{text-align}
Presumably indicates the alignment of text within the container.  The
only observed value is @code{left}.  Observed with nested @code{table}
and @code{text} elements.
@end defvr

@defvr {Optional} @code{width}
The width of the container in the form @code{@var{n}px}, e.g.@:
@code{1097px}.
@end defvr

@node SPV text Element (Inside @code{container})
@subsection The @code{text} Element (Inside @code{container})

Parent: @code{container} @*
Contents: @code{html}

This @code{text} element is nested inside a @code{container}.  There
is a different @code{text} element that is nested inside a
@code{pageParagraph}.

This element has the following attributes.

@defvr {Required} @code{type}
One of @code{title}, @code{log}, or @code{text}.
@end defvr

@defvr {Optional} @code{commandName}
As on the @code{heading} element.  For output not specific to a
command, this is simply @code{log}.  The corpus contains one example
of where @code{commandName} is present but set to the empty string.
@end defvr

@defvr {Optional} @code{creator-version}
As on the @code{heading} element.
@end defvr

@node SPV html Element
@subsection The @code{html} Element

Parent: @code{text} @*
Contents: CDATA

The CDATA contains an HTML document.  In some cases, the document
starts with @code{<html>} and ends with @code{</html}; in others the
@code{html} element is implied.  Generally the HTML includes a
@code{head} element with a CSS stylesheet.  The HTML body often begins
with @code{<BR>}.  The actual content ranges from trivial to simple:
just discarding the CSS and tags yields readable results.

This element has the following attributes.

@defvr {Required} @code{lang}
This always contains @code{en} in the corpus.
@end defvr

@node SPV table Element
@subsection The @code{table} Element

Parent: @code{container} @*
Contents: @code{tableStructure}

This element has the following attributes.

@defvr {Required} @code{commandName}
As on the @code{heading} element.
@end defvr

@defvr {Required} @code{type}
One of @code{table}, @code{note}, or @code{warning}.
@end defvr

@defvr {Required} @code{subType}
The locale-invariant name for the particular kind of output that this
table represents in the procedure.  This can be the same as
@code{commandName} e.g.@: @code{Frequencies}, or different, e.g.@:
@code{Case Processing Summary}.  Generic subtypes @code{Notes} and
@code{Warnings} are often used.
@end defvr

@defvr {Required} @code{tableId}
A number that uniquely identifies the table within the SPV file,
typically a large negative number such as @code{-4147135649387905023}.
@end defvr

@defvr {Optional} @code{creator-version}
As on the @code{heading} element.  In the corpus, this is only present
for version 21 and up and always includes all 8 digits.
@end defvr

@node SPV tableStructure Element
@subsection The @code{tableStructure} Element

Parent: @code{table} @*
Contents: @code{dataPath}

This element has no attributes.

@node SPV dataPath Element
@subsection The @code{dataPath} Element

Parent: @code{tableStructure} @*
Contents: text

Contains the name of the Zip member that holds the table details,
e.g.@: @code{0000000001437_lightTableData.bin}.

This element has no attributes.

@node SPV pageSetup Element
@subsection The @code{pageSetup} Element

Parent: @code{heading} @*
Contents: @code{pageHeader} @code{pageFooter}

This element has the following attributes.

@defvr {Required} @code{initial-page-number}
Always @code{1}.
@end defvr

@defvr {Optional} @code{chart-size}
Always @code{as-is} or a localization (!) of it (e.g.@: @code{dimensione
attuale}, @code{Wie vorgegeben}).
@end defvr

@defvr {Optional} @code{margin-left}
@defvrx {Optional} @code{margin-right}
@defvrx {Optional} @code{margin-top}
@defvrx {Optional} @code{margin-bottom}
Margin sizes in the form @code{@var{size}in}, e.g.@: @code{0.25in}.
@end defvr

@defvr {Optional} @code{paper-height}
@defvrx {Optional} @code{paper-width}
Paper sizes in the form @code{@var{size}in}, e.g.@: @code{8.5in} by
@code{11in} for letter paper or @code{8.267in} by @code{11.692in} for
A4 paper.
@end defvr

@defvr {Optional} @code{reference-orientation}
Always @code{0deg}.
@end defvr

@defvr {Optional} @code{space-after}
Always @code{12pt}.
@end defvr

@node SPV pageHeader and pageFooter Elements
@subsection The @code{pageHeader} and @code{pageFooter} Elements

Parent: @code{pageSetup} @*
Contents: @code{pageParagraph}*

This element has no attributes.

@node SPV pageParagraph Element
@subsection The @code{pageParagraph} Element

Parent: @code{pageHeader} or @code{pageFooter} @*
Contents: @code{text}

Text to go at the top or bottom of a page, respectively.

This element has no attributes.

@node SPV @code{text} Element (Inside @code{pageParagraph})
@subsection The @code{text} Element (Inside @code{pageParagraph})

Parent: @code{pageParagraph} @*
Contents: [CDATA]

This @code{text} element is nested inside a @code{pageParagraph}.  There
is a different @code{text} element that is nested inside a
@code{container}.

The element is either empty, or contains CDATA that holds almost-XHTML
text: in the corpus, either an @code{html} or @code{p} element.  It is
@emph{almost}-XHTML because the @code{html} element designates the
default namespace as
@code{http://xml.spss.com/spss/viewer/viewer-tree} instead of an XHTML
namespace, and because the CDATA can contain substitution variables:
@code{&[Page]} for the page number and @code{&[PageTitle]} for the
page title.

Typical contents (indented for clarity):

@example
<html xmlns="http://xml.spss.com/spss/viewer/viewer-tree">
    <head></head>
    <body>
        <p style="text-align:right; margin-top: 0">Page &[Page]</p>
    </body>
</html>
@end example

This element has the following attributes.

@defvr {Required} @code{type}
Always @code{text}.
@end defvr

@node SPV Light Detail Member Format
@section Light Detail Member Format

This section describes the format of ``light'' detail member
@file{.bin} files.  These files have a binary format which we describe
here in terms of a context-free grammar using the following conventions:

@table @asis
@item NonTerminal
Names of nonterminals are written in CamelCaps style.

@item 00, 01, @dots{}, ff.
Bytes with fixed values are written in hexadecimal:

@item i0, i1, @dots{}, i9, i10, i11, @dots{}
32-bit integers with fixed values are written in decimal, prefixed by
@samp{i}.

@item byte
An arbitrary byte.

@item int
An arbitrary 32-bit integer.

@item double
An arbitrary 64-bit IEEE floating-point number.

@item string
A 32-bit integer followed by the specified number of bytes of
character data, encoded in UTF-8.

@item @var{x}?
@var{x} is optional, e.g.@: 00? is an optional zero byte.

@item @var{x}*@var{n}
@var{x} is repeated @var{n} times, e.g. byte*10 for ten arbitrary bytes.

@item @var{x}[@var{name}]
Gives @var{x} the specified @var{name}.  Names are useful in textual
explanations.  They are also used, also bracketed, to indicate counts,
e.g.@: int[@t{n}] byte*[@t{n}] for a 32-bit integer followed by the
specified number of arbitrary bytes.

@item @var{a} @math{|} @var{b}
Either @var{a} or @var{b}.

@item (@var{x})
Parentheses are used for grouping to make precedence clear, especially
in the presence of @math{|}, e.g.@: in 00 (01 @math{|} 02 @math{|} 03)
00.

@item count(@var{x})
A 32-bit integer that indicates the number of bytes in @var{x},
followed by @var{x} itself.

@item v1(@var{x})
In a version 1 @file{.bin} file, @var{x}; in version 3, nothing.
(The @file{.bin} file header indicates the version.)

@item v3(@var{x})
In a version 3 @file{.bin} file, @var{x}; in version 1, nothing.
@end table

All integer and floating-point values in this format use little-endian
byte order.

A ``light'' detail member @file{.bin} consists of a number of sections
concatenated together, terminated by a byte 01:

@format
LightMember @result{} Header Title Styles Dimensions Data 01
@end format

The first section is a 0x27-byte header:

@format
Header @result{}
    01 00
    (i1 @math{|} i3)[@t{version}]
    01 (00 @math{|} 01) byte*21 00 00
    int[@t{table-id}] byte*4
@end format

The Header includes @code{version}, a version number that affects the
interpretation of some of the other data in the member.  We will refer
to ``version 1'' and ``version 3'' members later on.  @code{table-id}
is a binary version of the @code{tableId} attribute in the structure
member that refers to the detail member.  For example, if
@code{tableId} is @code{-4154297861994971133}, then @code{table-id}
would be 0xdca00003.  The meaning of the other variable parts of the
header is not known.

@format
Title @result{}
    Value[@t{local}] 01?
    Value[@t{subtype}] 01? 31
    Value[@t{c}] 01? 00? 58
    (58 @math{|} 31 Value[@t{caption}])
    int[@t{n}] Footnote*[@t{n}]
Footnote @result{}
    Value[@t{text}]
    (58 @math{|} 31 Value[@t{marker}])
    byte*4
@end format

@format
Styles @result{}
    00 Font*8
    int[@t{n1}] byte*[@t{n1}]
    int[@t{n2}] byte*[@t{n2}]
    int[@t{n3}] byte*[@t{n3}]
    int[@t{n4}] int*[@t{n4}]
    string[@t{encoding}]
    (i0 @math{|} i-1) (00 @math{|} 01) 00 (00 @math{|} 01)
    int
    byte[@t{decimal}] byte[@t{grouping}]
    int[@t{n-ccs}] string*[@t{n-ccs}]
    v1(i0)
    v3(count(count(X5) count(X6)))

X5 @result{} byte*33 int[@t{n}] int*[@t{n}]
X6 @result{}
    01 00 (03 @math{|} 04) 00 00 00
    string[@t{command}] string[@t{subcommand}]
    string[@t{language}] string[@t{charset}] string[@t{locale}]
    (00 @math{|} 01) 00 (00 @math{|} 01) (00 @math{|} 01)
    int
    byte[@t{decimal}] byte[@t{grouping}]
    byte*8 01
    (string[@t{dataset}] string[@t{datafile}] i0 int i0)?
    int[@t{n-ccs}] string*[@t{n-ccs}]
    2e (00 @math{|} 01) (i2000000 i0)?
@end format

In every example in the corpus, @code{n1} is 240.  The meaning of the
bytes that follow it is unknown.

In every example in the corpus, @code{n2} is 18 and the bytes that
follow it are @code{00 00 00 01 00 00 00 00 00 00 00 00 00 02 00 00 00
00}.  The meaning of these bytes is unknown.

In every example in the corpus for version 1, @code{n3} is 16 and the
bytes that follow it are @code{00 00 00 01 00 00 00 01 00 00 00 00 01
01 01 01}.  In version 3, observed @code{n3} varies from 117 to 150,
and its bytes include a 1-byte count at offset 0x34.  When the count
is nonzero, a text string of that length at offset 0x35 is the name of
a ``TableLook'', e.g. ``Default'' or ``Academic''.

Observed values of @code{n4} vary from 0 to 17.  Out of 7060 examples
in the corpus, it is nonzero only 36 times.

@code{encoding} is a character encoding, usually a Windows code page
such as @code{en_US.windows-1252} or @code{it_IT.windows-1252}.  The
encoding string is itself encoded in US-ASCII.  The rest of the
character strings in the file use this encoding.

@code{decimal} is the decimal point character.  The observed values
are @samp{.} and @samp{,}.

@code{grouping} is the grouping character.  The observed values are
@samp{,}, @samp{.}, @samp{'}, @samp{ }, and zero (presumably
indicating that digits should not be grouped).

@code{n-ccs} is observed as either 0 or 5.  When it is 5, the
following strings are CCA through CCE format strings.  Most commonly
these are all @code{-,,,} but other strings occur.

@format
Font @result{}
    byte[@t{index}] 31 string[@t{typeface}]
    00 00
    (10 @math{|} 20 @math{|} 40 @math{|} 50 @math{|} 70 @math{|} 80)[@t{f1}]
    41
    (i0 @math{|} i1 @math{|} i2)[@t{f2}]
    00
    (i0 @math{|} i2 @math{|} i64173)[@t{f3}]
    (i0 @math{|} i1 @math{|} i2 @math{|} i3)[@t{f4}]
    string[@t{fgcolor}] string[@t{bgcolor}]
    i0 i0 00
    v3(int[@t{f5}] int[@t{f6}] int[@t{f7}] int[@t{f8}]))
@end format

Each Font, in order, represents the font style for a different
element: title, caption, footnote, row labels, column labels, corner
labels, data, and layers.

@code{index} is the 1-based index of the Font, i.e. 1 for the first
Font, through 8 for the final Font.

@code{typeface} is the string name of the font.  In the corpus, this
is @code{SansSerif} in over 99% of instances and @code{Times New
Roman} in the rest.

@code{fgcolor} and @code{bgcolor} are the foreground color and
background color, respectively.  In the corpus, these are always
@code{#000000} and @code{#ffffff}, respectively.

The meaning of the remaining data is unknown.  It seems likely to
include font sizes, horizontal and vertical alignment, attributes such
as bold or italic, and margins.  @code{f1} is @code{40} most of the
time.  @code{f2} is @code{i1} most of the time for the title and
@code{i0} most of the time for other fonts.

The table below lists the values observed in the corpus.  When a cell
contains a single value, then 99@math{+}% of the corpus contains that value.
When a cell contains a pair of values, then the first value is seen in
about two-third of the corpus and the second value in about the
remaining one-third.  In fonts that include multiple pairs, values are
correlated, that is, for font 3, f5 = 24, f6 = 24, f7 = 2 appears
about two-thirds of the time, as does the combination of f4 = 0, f6 =
10 for font 7.

@multitable {font} {40} {f2} {64173} {0/1} {24/11} {10/11} {2/3} {f8}
@headitem font @tab f1 @tab f2 @tab f3 @tab f4 @tab f5 @tab f6 @tab f7 @tab f8
@item 1 @tab 40 @tab 1 @tab     0 @tab   0 @tab 8 @tab 10/11 @tab   1 @tab 8
@item 2 @tab 40 @tab 0 @tab     2 @tab   1 @tab 8 @tab 10/11 @tab   1 @tab 1
@item 3 @tab 40 @tab 0 @tab     2 @tab 1 @tab 24/11 @tab 24/ 8 @tab 2/3 @tab 4
@item 4 @tab 40 @tab 0 @tab     2 @tab   3 @tab 8 @tab 10/11 @tab   1 @tab 1
@item 5 @tab 40 @tab 0 @tab     0 @tab   1 @tab 8 @tab 10/11 @tab   1 @tab 4
@item 6 @tab 40 @tab 0 @tab     2 @tab   1 @tab 8 @tab 10/11 @tab   1 @tab 4
@item 7 @tab 40 @tab 0 @tab 64173 @tab 0/1 @tab 8 @tab 10/11 @tab   1 @tab 1
@item 8 @tab 40 @tab 0 @tab     2 @tab   3 @tab 8 @tab 10/11 @tab   1  @tab 4
@end multitable

@format
Dimensions @result{} int[@t{n-dims}] Dimension*[@t{n-dims}]
Dimension @result{}
    value[@t{name}]
    byte[@t{d1}]
    (00 @math{|} 01 @math{|} 02)[@t{d2}]
    (i0 @math{|} i2)[@t{d3}]
    (00 @math{|} 01)[@t{d4}]
    (00 @math{|} 01)[@t{d5}]
    01
    int[@t{d6}]
    int[@t{n-categories}] category*[@t{n-categories}]
@end format

@code{name} is the name of the dimension, e.g. @code{Variables},
@code{Statistics}, or a variable name.

@code{d1} is usually 0 but many other values have been observed.

@code{d3} is 2 over 99% of the time.

@code{d5} is 0 over 99% of the time.

@code{d6} is either -1 or the 0-based index of the dimension, e.g.@: 0
for the first dimension, 1 for the second, and so on.  The latter is
the case 98% of the time in the corpus.

@format
Category @result{} value[@t{name}] (Terminal @math{|} Group)
Terminal @result{} 00 00 00 i2 int[@t{index}] i0
Group @result{}
    (00 @math{|} 01)[@t{merge}] 00 01 (i0 @math{|} i2)[@t{data}]
    i-1 int[@t{n-subcategories}] category*[@t{n-subcategories}]
@end format

@code{name} is the name of the category (or group).

A Category can represent a terminal category.  In that case,
@code{index} is a nonnegative integer less than @code{n-categories} in
the Dimension in which the Category is nested (directly or
indirectly).

Alternatively, a Category can represent a Group of nested categories.
Usually a Group has some nested content, so that
@code{n-subcategories} is positive, but a few Groups with
@code{n-subcategories} 0 has been observed.

If a Group's @code{merge} is 00, the most common value, then the group
is really a distinct group that should be represented as such in the
visual representation and user interface.  If @code{merge} is 01,
however, the categories in this group should be shown and treated as
if they were direct children of the group's containing group (or if it
has no parent group, then direct children of the dimension), and this
group's name is irrelevant and should not be displayed.  (Merged
groups can be nested!)

@code{data} appears to be i2 when all of the categories within a group
are terminal categories that directly represent data values for a
variable (e.g. in a frequency table or crosstabulation, a group of
values in a variable being tabulated) and i0 otherwise.

@format
Data @result{}
    int[@t{layers}] int[@t{rows}] int[@t{columns}] int*[@t{n-dimensions}]
    int[@t{n-data}] Datum*[@t{n-data}]
@end format

The values of @code{layers}, @code{rows}, and @code{columns} each
specifies the number of dimensions represented in layers or rows or
columns, respectively, and their values sum to the number of
dimensions.

The @code{n-dimensions} integers are a permutation of the 0-based
dimension numbers.  The first @code{layers} of them specify each of
the dimensions represented by layers, the next @code{rows} of them
specify the dimensions represented by rows, and the final
@code{columns} of them specify the dimensions represented by columns.
When there is more than one dimension of a given kind, the inner
dimensions are given first.

@format
Datum @result{} int64[@t{index}] v3(00?) Value
@end format

The format of a Datum varies slightly from version 1 to version 3: in
version 1 it allows for an extra optional 00 byte.

A Datum consists of an @code{index} and a Value.  Suppose there are
@math{d} dimensions and dimension @math{i}, @math{0 \le i < d}, has
@math{n_i} categories.  Consider the datum at coordinates @math{x_i},
@math{0 \le i < d}, and note that @math{0 \le x_i < n_i}.  Then the
index is calculated by the following algorithm:

@display
let @i{index} = 0
for each @math{i} from 0 to @math{d - 1}:
    @i{index} = @math{n_i \times} @i{index} @math{+} @math{x_i}
@end display

For example, suppose there are 3 dimensions with 3, 4, and 5
categories, respectively.  The datum at coordinates (1, 2, 3) has
index @math{5 \times (4 \times (3 \times 0 + 1) + 2) + 3 = 33}.

@format
Value @result{} 00? 00? 00? 00? RawValue
RawValue @result{}
    01 ValueMod int[@t{format}] double[@t{x}]
  @math{|} 02 ValueMod int[@t{format}] double[@t{x}]
    string[@t{varname}] string[@t{vallab}] (01 @math{|} 02 @math{|} 03)
  @math{|} 03 string[@t{local}] ValueMod string[@t{id}] string[@t{c}] (00 @math{|} 01)[@t{type}]
  @math{|} 04 ValueMod int[@t{format}] string[@t{vallab}] string[@t{varname}]
    (01 @math{|} 02 @math{|} 03) string[@t{s}]
  @math{|} 05 ValueMod string[@t{varname}] string[@t{varlabel}] (01 @math{|} 02 @math{|} 03)
  @math{|} ValueMod string[@t{format}] int[@t{n-args}] Arg*[@t{n-args}]
Arg @result{}
    i0 Value
  @math{|} int[@t{x}] i0 Value*[@t{x}@math{+}1]      /* @t{x} @math{>} 0 */
@end format

A Value boils down to a number or a string.  There are several
possibilities, which one can distinguish by the first nonzero byte in
the encoding.

@table @asis
@item 01
The numeric value @code{x}, intended to be presented to the user
formatted according to @code{format}, which is in the format described
for system files.  @xref{System File Output Formats}, for details.
Most commonly, @code{format} has width 40 (the maximum).

When @code{x} is the maximum negative double value @code{-DBL_MAX}, it
represents the system-missing value SYSMIS.  (HIGHEST and LOWEST have
not been observed.)  @xref{System File Format}, for more about these
special values.

@item 02
Similar to @code{01}, with the additional information that @code{x} is
a value of variable @code{varname} and has value label @code{vallab}.
Both @code{varname} and @code{vallab} can be the empty string, the
latter very commonly.

The meaning of the final byte is unknown.  Possibly it is connected to
whether the value or the label should be displayed.

@item 03
A text string, in two forms: @code{c} is in English, and sometimes
abbreviated or obscure, and @code{local} is localized to the user's
locale.  In an English-language locale, the two strings are often the
same, and in the cases where they differ, @code{local} is more
appropriate for a user interface, e.g.@: @code{c} of ``Not a PxP table
for MCN...'' versus @code{local} of ``Computed only for a PxP table,
where P must be greater than 1.''

@code{c} and @code{local} are always either both empty or both
nonempty.

@code{id} is a brief identifying string whose form seems to resemble a
programming language identifier, e.g.@: @code{cumulative_percent} or
@code{factor_14}.  It is not unique.

@code{type} is 00 for text taken from user input, such as syntax
fragment, expressions, file names, data set names, and 01 for fixed
text strings such as names of procedures or statistics.  In the former
case, @code{id} is always the empty string; in the latter case,
@code{id} is still sometimes empty.

@item 04
The string value @code{s}, intended to be presented to the user
formatted according to @code{format}.  The format for a string is not
too interesting, and the corpus contains many clearly invalid formats
like A16.39 or A255.127 or A134.1, so readers should probably ignore
the format entirely.

@code{s} is a value of variable @code{varname} and has value label
@code{vallab}.  @code{varname} is never empty but @code{vallab} is
commonly empty.

The meaning of the final byte is unknown.

@item 05
Variable @code{varname}, which is rarely observed as empty in the
corpus, with variable label @code{varlabel}, which is often empty.

The meaning of the final byte is unknown.

@item 31
@itemx 58
(These bytes begin a ValueMod.)  A format string, analogous to
@code{printf}, followed by one or more arguments, each of which has
one or more values.  The format string uses the following syntax:

@table @code
@item \%
@itemx \:
@itemx \[
@itemx \]
Each of these expands to the character following @samp{\\}, to escape
characters that have special meaning in format strings.  These are
effective inside and outside the @code{[@dots{}]}  syntax forms
described below.

@item \n
Expands to a new-line, inside or outside the @code{[@dots{}]} forms
described below.

@item ^@var{i}
Expands to a formatted version of argument @var{i}, which must have
only a single value.  For example, @code{^1} expands to the first
argument's @code{value}.

@item [:@var{a}:]@var{i}
Expands @var{a} for each of the values in @var{i}.  @var{a}
should contain one or more @code{^@var{j}} conversions, which are
drawn from the values for argument @var{i} in order.  Some examples
from the corpus:

@table @code
@item [:^1:]1
All of the values for the first argument, concatenated.

@item [:^1\n:]1
Expands to the values for the first argument, each followed by
a new-line.

@item [:^1 = ^2:]2
Expands to @code{@var{x} = @var{y}} where @var{x} is the second
argument's first value and @var{y} is its second value.  (This would
be used only if the argument has two values.  If there were more
values, the second and third values would be directly concatenated,
which would look funny.)
@end table

@item [@var{a}:@var{b}:]@var{i}
This extends the previous form so that the first values are expanded
using @var{a} and later values are expanded using @var{b}.  For an
unknown reason, within @var{a} the @code{^@var{j}} conversions are
instead written as @code{%@var{j}}.  Some examples from the corpus:

@table @code
@item [%1:*^1:]1
Expands to all of the values for the first argument, separated by
@samp{*}.

@item [%1 = %2:, ^1 = ^2:]1
Given appropriate values for the first argument, expands to @code{X =
1, Y = 2, Z = 3}.

@item [%1:, ^1:]1
Given appropriate values, expands to @code{1, 2, 3}.
@end table
@end table

The format string is localized to the user's locale.
@end table

@format
ValueMod @result{}
    31 i0 (i0 @math{|} i1 string[@t{subscript}])
    v1(00 (i1 @math{|} i2) 00 00 int 00 00)
    v3(count(FormatString
             (58 @math{|} 31 Style)
             (58
              @math{|} 31 i0 i0 i0 i0 01 00 (01 @math{|} 02 @math{|} 08)
                00 08 00 0a 00)))
  @math{|} 31 i1 int[@t{footnote-number}] Format
  @math{|} 31 i2 (00 @math{|} 01 @math{|} 02) 00 (i1 @math{|} i2 @math{|} i3) Format
  @math{|} 31 i3 00 00 01 00 i2 Format
  @math{|} 58

Style @result{} 01? 00? 00? 00? 01 string[@t{fgcolor}] string[@t{bgcolor}] string[@t{font}] byte
Format @result{} 00 00 count(FormatString (58 @math{|} 31 Style) 58)
FormatString @result{} count((i0 (58 @math{|} 31 string))?)
@end format

A ValueMod can specify special modifications to a Value:

@itemize @bullet
@item
The @code{footnote-number}, if present, specifies a footnote that the
Value references.  The footnote's marker is shown appended to the main
text of the Value, as a superscript.

@item
The @code{subscript}, if present, specifies a string to append to the
main text of the Value, as a subscript.  The subscript text is
normally a brief indicator, e.g.@: @samp{a} or @samp{a,b}, with its
meaning indicated by the table caption.  In this usage, subscripts are
similar to footnotes; one apparent difference is that a Value can only
reference one footnote but a subscript can list more than one letter.

@item
The Format, if present, is a format string for substitutions using the
syntax explained previously.  It appears to be an English-language
version of the localized format string in the Value in which the
Format is nested.

@item
The Style, if present, changes the style for this individual Value.
@end itemize

@node SPV Legacy Detail Member Binary Format
@section Legacy Detail Member Binary Format

Whereas the light binary format represents everything about a given
pivot table, the legacy binary format conceptually consists of a
number of named sources, each of which consists of a number of named
series, each of which is a 1-dimensional array of numbers or strings
or a mix.  Thus, the legacy binary file format is quite simple.

This section uses the same context-free grammar notation as in the
previous section, with the following additions:

@table @asis
@item vAF(@var{x})
In a version 0xaf legacy member, @var{x}; in other versions, nothing.
(The legacy member header indicates the version; see below.)

@item vB0(@var{x})
In a version 0xb0 legacy member, @var{x}; in other versions, nothing.
@end table

@format
LegacyBinary @result{}
    00 byte[@t{version}] int16[@t{n-sources}] int[@t{file-size}]
    Metadata*[@t{n-sources}] Data*[@t{n-sources}]
@end format

@code{version} is a version number that affects the interpretation of
some of the other data in the member.  Versions 0xaf and 0xb0 are
known.  We will refer to ``version 0xaf'' and ``version 0xb0'' members
later on.

A legacy member consists of @code{n-sources} data sources, each of
which has Metadata and Data.

@code{file-size} is the size of the file, in bytes.

@format
Metadata @result{}
    int[@t{per-series}] int[@t{n-series}] int[@t{ofs}]
    vAF(byte*32[@t{source-name}])
    vB0(byte*64[@t{source-name}] int[@t{x}])
@end format

A data source consists of @code{n-series} series of data, with
@code{per-series} data values per series.

@code{source-name} is a 32- or 64-byte string padded on the right with
zero bytes.  The names that appear in the corpus are very generic,
usually @code{tableData} or @code{source0}.

The @code{ofs} is the offset, in bytes, from the beginning of the file
to the start of this data source's Data.  This allows programs to skip
to the beginning of the data for a particular source; it is also
important to determine whether a source includes any string data (see
below).

The meaning of @code{x} in version 0xb0 is unknown.

@format
Data @result{} NumericData StringData?
NumericData @result{} NumericSeries*[@t{n-series}]
NumericSeries @result{} byte*288[@t{series-name}] double*[@t{per-series}]
@end format

Data follow the Metadata in the legacy binary format, with sources in
the same order.  Each NumericSeries begins with a @code{series-name}
that generally indicates its role in the pivot table, e.g.@: ``cell'',
``cellFormat'', ``dimension0categories'', ``dimension0group0''.  The
name is followed by the data, one double per element in the series.  A
double with the maximum negative double @code{-DBL_MAX} represents the
system-missing value SYSMIS.

@format
StringData @result{} i1 string[@t{source-name}] Pairs Labels

Pairs @result{} int[@t{n-string-series}] PairSeries*[@t{n-string-series}]
PairSeries @result{} string[@t{pair-series-name}] int[@t{n-pairs}] Pair*[@t{n-pairs}]
Pair @result{} int[@t{i}] int[@t{j}]

Labels @result{} int[@t{n-labels}] Label*[@t{n-labels}]
Label @result{} int[@t{frequency}] int[@t{s}]
@end format

A source may include a mix of numeric and string data values.  When a
source includes any string data, the data values that are strings are
set to SYSMIS in the NumericSeries, and StringData follows the
NumericData.  A source that contains no string data omits the
StringData.  To reliably determine whether a source includes
StringData, the reader should check whether the offset following the
Numericdata is the offset of the next series, as indicated by its
Metadata (or end of file, in the case of the last source in a file).

StringData repeats the name of the source.

The string data overlays the numeric data.  @code{n-string-series} is
the number of series within the source that include string data.  More
precisely, it is the 1-based index of the last series in the source
that includes any string data; thus, it would be 4 if there are 5
series and only the fourth one includes string data.

Each PairSeries consists a sequence of 0 or more pairs, each of which
maps from a 0-based index within the series @code{i} to a 0-based
label index @code{j}.  The pair @code{i} = 2, @code{j} = 3, for
example, would mean that the third data value (with value SYSMIS) is
to be replaced by the string of the fourth label.

The labels themselves follow the pairs.  The valuable part of each
label is the string @code{s}.  Each label also includes a
@code{frequency} that reports the number of pairs that reference it
(although this is not useful).