1bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant% -*- mode: latex; TeX-master: "Vorbis_I_spec"; -*-
2bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant%!TEX root = Vorbis_I_spec.tex
3f5256e16dfc425c1d466f6308d4026d529ce9e0bHoward Hinnant% $Id$
4bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant\section{Probability Model and Codebooks} \label{vorbis:spec:codebook}
5b64f8b07c104c6cc986570ac8ee0ed16a9f23976Howard Hinnant
6b64f8b07c104c6cc986570ac8ee0ed16a9f23976Howard Hinnant\subsection{Overview}
7bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant
8bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard HinnantUnlike practically every other mainstream audio codec, Vorbis has no
9bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnantstatically configured probability model, instead packing all entropy
10bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnantdecoding configuration, VQ and Huffman, into the bitstream itself in
11bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnantthe third header, the codec setup header.  This packed configuration
12bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnantconsists of multiple 'codebooks', each containing a specific
13bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard HinnantHuffman-equivalent representation for decoding compressed codewords as
14bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnantwell as an optional lookup table of output vector values to which a
15bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnantdecoded Huffman value is applied as an offset, generating the final
16bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnantdecoded output corresponding to a given compressed codeword.
17bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant
18bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant\subsubsection{Bitwise operation}
19bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard HinnantThe codebook mechanism is built on top of the vorbis bitpacker. Both
20bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnantthe codebooks themselves and the codewords they decode are unrolled
21bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnantfrom a packet as a series of arbitrary-width values read from the
22bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnantstream according to \xref{vorbis:spec:bitpacking}.
23bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant
24bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant
25bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant
26bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant
27bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant\subsection{Packed codebook format}
28bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant
29bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard HinnantFor purposes of the examples below, we assume that the storage
30bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnantsystem's native byte width is eight bits.  This is not universally
31bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnanttrue; see \xref{vorbis:spec:bitpacking} for discussion
32bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnantrelating to non-eight-bit bytes.
33bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant
34bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant\subsubsection{codebook decode}
35bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant
36bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard HinnantA codebook begins with a 24 bit sync pattern, 0x564342:
37bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant
38bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant\begin{Verbatim}[commandchars=\\\{\}]
39bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnantbyte 0: [ 0 1 0 0 0 0 1 0 ] (0x42)
40bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnantbyte 1: [ 0 1 0 0 0 0 1 1 ] (0x43)
41bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnantbyte 2: [ 0 1 0 1 0 1 1 0 ] (0x56)
42bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant\end{Verbatim}
43bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant
44bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant16 bit \varname{[codebook_dimensions]} and 24 bit \varname{[codebook_entries]} fields:
45bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant
46bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant\begin{Verbatim}[commandchars=\\\{\}]
47bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant
48bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnantbyte 3: [ X X X X X X X X ]
49bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnantbyte 4: [ X X X X X X X X ] [codebook_dimensions] (16 bit unsigned)
50bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant
51bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnantbyte 5: [ X X X X X X X X ]
52bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnantbyte 6: [ X X X X X X X X ]
53bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnantbyte 7: [ X X X X X X X X ] [codebook_entries] (24 bit unsigned)
54bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant
55bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant\end{Verbatim}
56bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant
57bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard HinnantNext is the \varname{[ordered]} bit flag:
58bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant
59bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant\begin{Verbatim}[commandchars=\\\{\}]
60bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnant
61bc8d3f97eb5c958007f2713238472e0c1c8fe02Howard Hinnantbyte 8: [               X ] [ordered] (1 bit)
62
63\end{Verbatim}
64
65Each entry, numbering a
66total of \varname{[codebook_entries]}, is assigned a codeword length.
67We now read the list of codeword lengths and store these lengths in
68the array \varname{[codebook_codeword_lengths]}. Decode of lengths is
69according to whether the \varname{[ordered]} flag is set or unset.
70
71\begin{itemize}
72\item
73  If the \varname{[ordered]} flag is unset, the codeword list is not
74  length ordered and the decoder needs to read each codeword length
75  one-by-one.
76
77  The decoder first reads one additional bit flag, the
78  \varname{[sparse]} flag.  This flag determines whether or not the
79  codebook contains unused entries that are not to be included in the
80  codeword decode tree:
81
82\begin{Verbatim}[commandchars=\\\{\}]
83byte 8: [             X 1 ] [sparse] flag (1 bit)
84\end{Verbatim}
85
86  The decoder now performs for each of the \varname{[codebook_entries]}
87  codebook entries:
88
89\begin{Verbatim}[commandchars=\\\{\}]
90
91  1) if([sparse] is set) \{
92
93         2) [flag] = read one bit;
94         3) if([flag] is set) \{
95
96              4) [length] = read a five bit unsigned integer;
97              5) codeword length for this entry is [length]+1;
98
99            \} else \{
100
101              6) this entry is unused.  mark it as such.
102
103            \}
104
105     \} else the sparse flag is not set \{
106
107        7) [length] = read a five bit unsigned integer;
108        8) the codeword length for this entry is [length]+1;
109
110     \}
111
112\end{Verbatim}
113
114\item
115  If the \varname{[ordered]} flag is set, the codeword list for this
116  codebook is encoded in ascending length order.  Rather than reading
117  a length for every codeword, the encoder reads the number of
118  codewords per length.  That is, beginning at entry zero:
119
120\begin{Verbatim}[commandchars=\\\{\}]
121  1) [current_entry] = 0;
122  2) [current_length] = read a five bit unsigned integer and add 1;
123  3) [number] = read \link{vorbis:spec:ilog}{ilog}([codebook_entries] - [current_entry]) bits as an unsigned integer
124  4) set the entries [current_entry] through [current_entry]+[number]-1, inclusive,
125    of the [codebook_codeword_lengths] array to [current_length]
126  5) set [current_entry] to [number] + [current_entry]
127  6) increment [current_length] by 1
128  7) if [current_entry] is greater than [codebook_entries] ERROR CONDITION;
129    the decoder will not be able to read this stream.
130  8) if [current_entry] is less than [codebook_entries], repeat process starting at 3)
131  9) done.
132\end{Verbatim}
133
134\end{itemize}
135
136After all codeword lengths have been decoded, the decoder reads the
137vector lookup table.  Vorbis I supports three lookup types:
138\begin{enumerate}
139\item
140No lookup
141\item
142Implicitly populated value mapping (lattice VQ)
143\item
144Explicitly populated value mapping (tessellated or 'foam'
145VQ)
146\end{enumerate}
147
148
149The lookup table type is read as a four bit unsigned integer:
150\begin{Verbatim}[commandchars=\\\{\}]
151  1) [codebook_lookup_type] = read four bits as an unsigned integer
152\end{Verbatim}
153
154Codebook decode precedes according to \varname{[codebook_lookup_type]}:
155\begin{itemize}
156\item
157Lookup type zero indicates no lookup to be read.  Proceed past
158lookup decode.
159\item
160Lookup types one and two are similar, differing only in the
161number of lookup values to be read.  Lookup type one reads a list of
162values that are permuted in a set pattern to build a list of vectors,
163each vector of order \varname{[codebook_dimensions]} scalars.  Lookup
164type two builds the same vector list, but reads each scalar for each
165vector explicitly, rather than building vectors from a smaller list of
166possible scalar values.  Lookup decode proceeds as follows:
167
168\begin{Verbatim}[commandchars=\\\{\}]
169  1) [codebook_minimum_value] = \link{vorbis:spec:float32:unpack}{float32_unpack}( read 32 bits as an unsigned integer)
170  2) [codebook_delta_value] = \link{vorbis:spec:float32:unpack}{float32_unpack}( read 32 bits as an unsigned integer)
171  3) [codebook_value_bits] = read 4 bits as an unsigned integer and add 1
172  4) [codebook_sequence_p] = read 1 bit as a boolean flag
173
174  if ( [codebook_lookup_type] is 1 ) \{
175
176     5) [codebook_lookup_values] = \link{vorbis:spec:lookup1:values}{lookup1_values}(\varname{[codebook_entries]}, \varname{[codebook_dimensions]} )
177
178  \} else \{
179
180     6) [codebook_lookup_values] = \varname{[codebook_entries]} * \varname{[codebook_dimensions]}
181
182  \}
183
184  7) read a total of [codebook_lookup_values] unsigned integers of [codebook_value_bits] each;
185     store these in order in the array [codebook_multiplicands]
186\end{Verbatim}
187\item
188A \varname{[codebook_lookup_type]} of greater than two is reserved
189and indicates a stream that is not decodable by the specification in this
190document.
191
192\end{itemize}
193
194
195An 'end of packet' during any read operation in the above steps is
196considered an error condition rendering the stream undecodable.
197
198\paragraph{Huffman decision tree representation}
199
200The \varname{[codebook_codeword_lengths]} array and
201\varname{[codebook_entries]} value uniquely define the Huffman decision
202tree used for entropy decoding.
203
204Briefly, each used codebook entry (recall that length-unordered
205codebooks support unused codeword entries) is assigned, in order, the
206lowest valued unused binary Huffman codeword possible.  Assume the
207following codeword length list:
208
209\begin{Verbatim}[commandchars=\\\{\}]
210entry 0: length 2
211entry 1: length 4
212entry 2: length 4
213entry 3: length 4
214entry 4: length 4
215entry 5: length 2
216entry 6: length 3
217entry 7: length 3
218\end{Verbatim}
219
220Assigning codewords in order (lowest possible value of the appropriate
221length to highest) results in the following codeword list:
222
223\begin{Verbatim}[commandchars=\\\{\}]
224entry 0: length 2 codeword 00
225entry 1: length 4 codeword 0100
226entry 2: length 4 codeword 0101
227entry 3: length 4 codeword 0110
228entry 4: length 4 codeword 0111
229entry 5: length 2 codeword 10
230entry 6: length 3 codeword 110
231entry 7: length 3 codeword 111
232\end{Verbatim}
233
234
235\begin{note}
236Unlike most binary numerical values in this document, we
237intend the above codewords to be read and used bit by bit from left to
238right, thus the codeword '001' is the bit string 'zero, zero, one'.
239When determining 'lowest possible value' in the assignment definition
240above, the leftmost bit is the MSb.
241\end{note}
242
243It is clear that the codeword length list represents a Huffman
244decision tree with the entry numbers equivalent to the leaves numbered
245left-to-right:
246
247\begin{center}
248\includegraphics[width=10cm]{hufftree}
249\captionof{figure}{huffman tree illustration}
250\end{center}
251
252
253As we assign codewords in order, we see that each choice constructs a
254new leaf in the leftmost possible position.
255
256Note that it's possible to underspecify or overspecify a Huffman tree
257via the length list.  In the above example, if codeword seven were
258eliminated, it's clear that the tree is unfinished:
259
260\begin{center}
261\includegraphics[width=10cm]{hufftree-under}
262\captionof{figure}{underspecified huffman tree illustration}
263\end{center}
264
265
266Similarly, in the original codebook, it's clear that the tree is fully
267populated and a ninth codeword is impossible.  Both underspecified and
268overspecified trees are an error condition rendering the stream
269undecodable. Take special care that a codebook with a single used
270entry is handled properly; it consists of a single codework of zero
271bits and 'reading' a value out of such a codebook always returns the
272single used value and sinks zero bits.  
273
274Codebook entries marked 'unused' are simply skipped in the assigning
275process.  They have no codeword and do not appear in the decision
276tree, thus it's impossible for any bit pattern read from the stream to
277decode to that entry number.
278
279
280
281\paragraph{VQ lookup table vector representation}
282
283Unpacking the VQ lookup table vectors relies on the following values:
284\begin{programlisting}
285the [codebook_multiplicands] array
286[codebook_minimum_value]
287[codebook_delta_value]
288[codebook_sequence_p]
289[codebook_lookup_type]
290[codebook_entries]
291[codebook_dimensions]
292[codebook_lookup_values]
293\end{programlisting}
294
295\bigskip
296
297Decoding (unpacking) a specific vector in the vector lookup table
298proceeds according to \varname{[codebook_lookup_type]}.  The unpacked
299vector values are what a codebook would return during audio packet
300decode in a VQ context.
301
302\paragraph{Vector value decode: Lookup type 1}
303
304Lookup type one specifies a lattice VQ lookup table built
305algorithmically from a list of scalar values.  Calculate (unpack) the
306final values of a codebook entry vector from the entries in
307\varname{[codebook_multiplicands]} as follows (\varname{[value_vector]}
308is the output vector representing the vector of values for entry number
309\varname{[lookup_offset]} in this codebook):
310
311\begin{Verbatim}[commandchars=\\\{\}]
312  1) [last] = 0;
313  2) [index_divisor] = 1;
314  3) iterate [i] over the range 0 ... [codebook_dimensions]-1 (once for each scalar value in the value vector) \{
315
316       4) [multiplicand_offset] = ( [lookup_offset] divided by [index_divisor] using integer
317          division ) integer modulo [codebook_lookup_values]
318
319       5) vector [value_vector] element [i] =
320            ( [codebook_multiplicands] array element number [multiplicand_offset] ) *
321            [codebook_delta_value] + [codebook_minimum_value] + [last];
322
323       6) if ( [codebook_sequence_p] is set ) then set [last] = vector [value_vector] element [i]
324
325       7) [index_divisor] = [index_divisor] * [codebook_lookup_values]
326
327     \}
328
329  8) vector calculation completed.
330\end{Verbatim}
331
332
333
334\paragraph{Vector value decode: Lookup type 2}
335
336Lookup type two specifies a VQ lookup table in which each scalar in
337each vector is explicitly set by the \varname{[codebook_multiplicands]}
338array in a one-to-one mapping.  Calculate [unpack] the
339final values of a codebook entry vector from the entries in
340\varname{[codebook_multiplicands]} as follows (\varname{[value_vector]}
341is the output vector representing the vector of values for entry number
342\varname{[lookup_offset]} in this codebook):
343
344\begin{Verbatim}[commandchars=\\\{\}]
345  1) [last] = 0;
346  2) [multiplicand_offset] = [lookup_offset] * [codebook_dimensions]
347  3) iterate [i] over the range 0 ... [codebook_dimensions]-1 (once for each scalar value in the value vector) \{
348
349       4) vector [value_vector] element [i] =
350            ( [codebook_multiplicands] array element number [multiplicand_offset] ) *
351            [codebook_delta_value] + [codebook_minimum_value] + [last];
352
353       5) if ( [codebook_sequence_p] is set ) then set [last] = vector [value_vector] element [i]
354
355       6) increment [multiplicand_offset]
356
357     \}
358
359  7) vector calculation completed.
360\end{Verbatim}
361
362
363
364
365
366
367
368
369
370\subsection{Use of the codebook abstraction}
371
372The decoder uses the codebook abstraction much as it does the
373bit-unpacking convention; a specific codebook reads a
374codeword from the bitstream, decoding it into an entry number, and then
375returns that entry number to the decoder (when used in a scalar
376entropy coding context), or uses that entry number as an offset into
377the VQ lookup table, returning a vector of values (when used in a context
378desiring a VQ value). Scalar or VQ context is always explicit; any call
379to the codebook mechanism requests either a scalar entry number or a
380lookup vector.
381
382Note that VQ lookup type zero indicates that there is no lookup table;
383requesting decode using a codebook of lookup type 0 in any context
384expecting a vector return value (even in a case where a vector of
385dimension one) is forbidden.  If decoder setup or decode requests such
386an action, that is an error condition rendering the packet
387undecodable.
388
389Using a codebook to read from the packet bitstream consists first of
390reading and decoding the next codeword in the bitstream. The decoder
391reads bits until the accumulated bits match a codeword in the
392codebook.  This process can be though of as logically walking the
393Huffman decode tree by reading one bit at a time from the bitstream,
394and using the bit as a decision boolean to take the 0 branch (left in
395the above examples) or the 1 branch (right in the above examples).
396Walking the tree finishes when the decode process hits a leaf in the
397decision tree; the result is the entry number corresponding to that
398leaf.  Reading past the end of a packet propagates the 'end-of-stream'
399condition to the decoder.
400
401When used in a scalar context, the resulting codeword entry is the
402desired return value.
403
404When used in a VQ context, the codeword entry number is used as an
405offset into the VQ lookup table.  The value returned to the decoder is
406the vector of scalars corresponding to this offset.
407