1 <!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Strict//EN" "http://www.w3.org/TR/xhtml1/DTD/xhtml1-strict.dtd">
5 <meta http-equiv="Content-Type" content="text/html; charset=iso-8859-15"/>
6 <title>Ogg Documentation</title>
8 <style type="text/css">
10 margin: 0 18px 0 18px;
12 font-family: Verdana, Arial, Helvetica, sans-serif;
26 margin: 30px 0 16px 0;
33 h1, h1 a, h2, h2 a, h3, h3 a {
36 margin: 1.3em 0 8px 0;
70 <a href="http://www.xiph.org/"><img src="fish_xiph_org.png" alt="Fish Logo and Xiph.org"/></a>
73 <h1>Ogg logical bitstream framing</h1>
75 <h2>Ogg bitstreams</h2>
77 <p>The Ogg transport bitstream is designed to provide framing, error
78 protection and seeking structure for higher-level codec streams that
79 consist of raw, unencapsulated data packets, such as the Vorbis audio
80 codec or Theora video codec.</p>
82 <h2>Application example: Vorbis</h2>
84 <p>Vorbis encodes short-time blocks of PCM data into raw packets of
85 bit-packed data. These raw packets may be used directly by transport
86 mechanisms that provide their own framing and packet-separation
87 mechanisms (such as UDP datagrams). For stream based storage (such as
88 files) and transport (such as TCP streams or pipes), Vorbis uses the
89 Ogg bitstream format to provide framing/sync, sync recapture
90 after error, landmarks during seeking, and enough information to
91 properly separate data back into packets at the original packet
92 boundaries without relying on decoding to find packet boundaries.</p>
94 <h2>Design constraints for Ogg bitstreams</h2>
97 <li>True streaming; we must not need to seek to build a 100%
98 complete bitstream.</li>
99 <li>Use no more than approximately 1-2% of bitstream bandwidth for
100 packet boundary marking, high-level framing, sync and seeking.</li>
101 <li>Specification of absolute position within the original sample
103 <li>Simple mechanism to ease limited editing, such as a simplified
104 concatenation mechanism.</li>
105 <li>Detection of corruption, recapture after error and direct, random
106 access to data at arbitrary positions in the bitstream.</li>
109 <h2>Logical and Physical Bitstreams</h2>
111 <p>A <em>logical</em> Ogg bitstream is a contiguous stream of
112 sequential pages belonging only to the logical bitstream. A
113 <em>physical</em> Ogg bitstream is constructed from one or more
114 than one logical Ogg bitstream (the simplest physical bitstream
115 is simply a single logical bitstream). We describe below the exact
116 formatting of an Ogg logical bitstream. Combining logical
117 bitstreams into more complex physical bitstreams is described in the
118 <a href="oggstream.html">Ogg bitstream overview</a>. The exact
119 mapping of raw Vorbis packets into a valid Ogg Vorbis physical
120 bitstream is described in the Vorbis I Specification.</p>
122 <h2>Bitstream structure</h2>
124 <p>An Ogg stream is structured by dividing incoming packets into
125 segments of up to 255 bytes and then wrapping a group of contiguous
126 packet segments into a variable length page preceded by a page
127 header. Both the header size and page size are variable; the page
128 header contains sizing information and checksum data to determine
129 header/page size and data integrity.</p>
131 <p>The bitstream is captured (or recaptured) by looking for the beginning
132 of a page, specifically the capture pattern. Once the capture pattern
133 is found, the decoder verifies page sync and integrity by computing
134 and comparing the checksum. At that point, the decoder can extract the
135 packets themselves.</p>
137 <h3>Packet segmentation</h3>
139 <p>Packets are logically divided into multiple segments before encoding
140 into a page. Note that the segmentation and fragmentation process is a
141 logical one; it's used to compute page header values and the original
142 page data need not be disturbed, even when a packet spans page
145 <p>The raw packet is logically divided into [n] 255 byte segments and a
146 last fractional segment of < 255 bytes. A packet size may well
147 consist only of the trailing fractional segment, and a fractional
148 segment may be zero length. These values, called "lacing values" are
149 then saved and placed into the header segment table.</p>
151 <p>An example should make the basic concept clear:</p>
156 ___________________________________________
157 |______________packet data__________________| 753 bytes
159 lacing values for page header segment table: 255,255,243
163 <p>We simply add the lacing values for the total size; the last lacing
164 value for a packet is always the value that is less than 255. Note
165 that this encoding both avoids imposing a maximum packet size as well
166 as imposing minimum overhead on small packets (as opposed to, eg,
167 simply using two bytes at the head of every packet and having a max
168 packet size of 32k. Small packets (<255, the typical case) are
169 penalized with twice the segmentation overhead). Using the lacing
170 values as suggested, small packets see the minimum possible
171 byte-aligned overhead (1 byte) and large packets, over 512 bytes or
172 so, see a fairly constant ~.5% overhead on encoding space.</p>
174 <p>Note that a lacing value of 255 implies that a second lacing value
175 follows in the packet, and a value of < 255 marks the end of the
176 packet after that many additional bytes. A packet of 255 bytes (or a
177 multiple of 255 bytes) is terminated by a lacing value of 0:</p>
181 _______________________________
182 |________packet data____________| 255 bytes
184 lacing values: 255, 0
187 <p>Note also that a 'nil' (zero length) packet is not an error; it
188 consists of nothing more than a lacing value of zero in the header.</p>
190 <h3>Packets spanning pages</h3>
192 <p>Packets are not restricted to beginning and ending within a page,
193 although individual segments are, by definition, required to do so.
194 Packets are not restricted to a maximum size, although excessively
195 large packets in the data stream are discouraged.</p>
197 <p>After segmenting a packet, the encoder may decide not to place all the
198 resulting segments into the current page; to do so, the encoder places
199 the lacing values of the segments it wishes to belong to the current
200 page into the current segment table, then finishes the page. The next
201 page is begun with the first value in the segment table belonging to
202 the next packet segment, thus continuing the packet (data in the
203 packet body must also correspond properly to the lacing values in the
204 spanned pages. The segment data in the first packet corresponding to
205 the lacing values of the first page belong in that page; packet
206 segments listed in the segment table of the following page must begin
207 the page body of the subsequent page).</p>
209 <p>The last mechanic to spanning a page boundary is to set the header
210 flag in the new page to indicate that the first lacing value in the
211 segment table continues rather than begins a packet; a header flag of
212 0x01 is set to indicate a continued packet. Although mandatory, it
213 is not actually algorithmically necessary; one could inspect the
214 preceding segment table to determine if the packet is new or
215 continued. Adding the information to the packet_header flag allows a
216 simpler design (with no overhead) that needs only inspect the current
217 page header after frame capture. This also allows faster error
218 recovery in the event that the packet originates in a corrupt
219 preceding page, implying that the previous page's segment table
220 cannot be trusted.</p>
222 <p>Note that a packet can span an arbitrary number of pages; the above
223 spanning process is repeated for each spanned page boundary. Also a
224 'zero termination' on a packet size that is an even multiple of 255
225 must appear even if the lacing value appears in the next page as a
226 zero-length continuation of the current packet. The header flag
227 should be set to 0x01 to indicate that the packet spanned, even though
228 the span is a nil case as far as data is concerned.</p>
230 <p>The encoding looks odd, but is properly optimized for speed and the
231 expected case of the majority of packets being between 50 and 200
232 bytes (note that it is designed such that packets of wildly different
233 sizes can be handled within the model; placing packet size
234 restrictions on the encoder would have only slightly simplified design
235 in page generation and increased overall encoder complexity).</p>
237 <p>The main point behind tracking individual packets (and packet
238 segments) is to allow more flexible encoding tricks that requiring
239 explicit knowledge of packet size. An example is simple bandwidth
240 limiting, implemented by simply truncating packets in the nominal case
241 if the packet is arranged so that the least sensitive portion of the
244 <a name="page_header"></a>
247 <p>The headering mechanism is designed to avoid copying and re-assembly
248 of the packet data (ie, making the packet segmentation process a
249 logical one); the header can be generated directly from incoming
250 packet data. The encoder buffers packet data until it finishes a
251 complete page at which point it writes the header followed by the
252 buffered packet segments.</p>
254 <h4>capture_pattern</h4>
256 <p>A header begins with a capture pattern that simplifies identifying
257 pages; once the decoder has found the capture pattern it can do a more
258 intensive job of verifying that it has in fact found a page boundary
259 (as opposed to an inadvertent coincidence in the byte stream).</p>
270 <h4>stream_structure_version</h4>
272 <p>The capture pattern is followed by the stream structure revision:</p>
280 <h4>header_type_flag</h4>
282 <p>The header type flag identifies this page's context in the bitstream:</p>
287 5 bitflags: 0x01: unset = fresh packet
288 set = continued packet
289 0x02: unset = not first page of logical bitstream
290 set = first page of logical bitstream (bos)
291 0x04: unset = not last page of logical bitstream
292 set = last page of logical bitstream (eos)
295 <h4>absolute granule position</h4>
297 <p>(This is packed in the same way the rest of Ogg data is packed; LSb
298 of LSB first. Note that the 'position' data specifies a 'sample'
299 number (eg, in a CD quality sample is four octets, 16 bits for left
300 and 16 bits for right; in video it would likely be the frame number.
301 It is up to the specific codec in use to define the semantic meaning
302 of the granule position value). The position specified is the total
303 samples encoded after including all packets finished on this page
304 (packets begun on this page but continuing on to the next page do not
305 count). The rationale here is that the position specified in the
306 frame header of the last page tells how long the data coded by the
307 bitstream is. A truncated stream will still return the proper number
308 of samples that can be decoded fully.</p>
310 <p>A special value of '-1' (in two's complement) indicates that no packets
311 finish on this page.</p>
326 <h4>stream serial number</h4>
328 <p>Ogg allows for separate logical bitstreams to be mixed at page
329 granularity in a physical bitstream. The most common case would be
330 sequential arrangement, but it is possible to interleave pages for
331 two separate bitstreams to be decoded concurrently. The serial
332 number is the means by which pages physical pages are associated with
333 a particular logical stream. Each logical stream must have a unique
334 serial number within a physical stream:</p>
345 <h4>page sequence no</h4>
347 <p>Page counter; lets us know if a page is lost (useful where packets
348 span page boundaries).</p>
359 <h4>page checksum</h4>
361 <p>32 bit CRC value (direct algorithm, initial val and final XOR = 0,
362 generator polynomial=0x04c11db7). The value is computed over the
363 entire header (with the CRC field in the header set to zero) and then
364 continued over the page. The CRC field is then filled with the
367 <p>(A thorough discussion of CRC algorithms can be found in <a
368 href="http://www.ross.net/crc/download/crc_v3.txt">"A
369 Painless Guide to CRC Error Detection Algorithms"</a> by Ross
370 Williams <a href="mailto:ross@ross.net">ross@ross.net</a>.)</p>
381 <h4>page_segments</h4>
383 <p>The number of segment entries to appear in the segment table. The
384 maximum number of 255 segments (255 bytes each) sets the maximum
385 possible physical page size at 65307 bytes or just under 64kB (thus
386 we know that a header corrupted so as destroy sizing/alignment
387 information will not cause a runaway bitstream. We'll read in the
388 page according to the corrupted size information that's guaranteed to
389 be a reasonable size regardless, notice the checksum mismatch, drop
390 sync and then look for recapture).</p>
398 <h4>segment_table (containing packet lacing values)</h4>
400 <p>The lacing values for each packet segment physically appearing in
401 this page are listed in contiguous order.</p>
408 n 0x00-0xff (0-255, n=page_segments+26)
411 <p>Total page size is calculated directly from the known header size and
412 lacing values in the segment table. Packet data segments follow
413 immediately after the header.</p>
415 <p>Page headers typically impose a flat .25-.5% space overhead assuming
416 nominal ~8k page sizes. The segmentation table needed for exact
417 packet recovery in the streaming layer adds approximately .5-1%
418 nominal assuming expected encoder behavior in the 44.1kHz, 128kbps
419 stereo encodings.</p>
422 The Xiph Fish Logo is a
423 trademark (™) of Xiph.Org.<br/>
425 These pages © 1994 - 2005 Xiph.Org. All rights reserved.