Theora

Theora Specication
Xiph.Org Foundation March 16, 2011
Contents
1 Introduction 1.1 VP3 and Theora . . . . . . . . . . . 1.2 Video Formats . . . . . . . . . . . . 1.3 Classication . . . . . . . . . . . . . 1.4 Assumptions . . . . . . . . . . . . . 1.5 Codec Setup and Probability Model 1.6 Format Conformance . . . . . . . . . 2 Coded Video Structure 2.1 Frame Layout . . . . . . . . . 2.2 Picture Region . . . . . . . . 2.3 Blocks and Super Blocks . . . 2.4 Macro Blocks . . . . . . . . . 2.5 Coding Modes and Prediction 2.6 DCT Coecients . . . . . . . 3 Decoding Overview 3.1 Decoder Conguration . . . . 3.1.1 Global Conguration . 3.1.2 Quantization Matrices 3.1.3 Human Codebooks . 3.2 High-Level Decode Process . 3.2.1 Decoder Setup . . . . 3.2.2 Decode Procedure . . 1 1 1 2 2 2 3 5 5 6 7 9 11 11 13 13 13 13 14 15 15 15
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
4 Video Formats 21 4.1 Color Space Conventions . . . . . . . . . . . . . . . . . . . . . . . 21 4.2 Color Space Conversions and Parameters . . . . . . . . . . . . . 22 4.3 Available Color Spaces . . . . . . . . . . . . . . . . . . . . . . . . 24 4.3.1 Rec. 470M (Rec. ITU-R BT.470-6 System M/NTSC with Rec. ITU-R BT.601-5) . . . . . . . . . . . . . . . . . . . . 24 4.3.2 Rec. 470BG (Rec. ITU-R BT.470-6 Systems B and G with Rec. ITU-R BT.601-5) . . . . . . . . . . . . . . . . . . . . 25 4.4 Pixel Formats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 i
ii 4.4.1 4.4.2 4.4.3 4.4.4 4:4:4 Subsampling . 4:2:2 Subsampling . 4:2:0 Subsampling . Subsampling and the . . . . . . . . . . . . . . . . . . . . . . . . . . . Picture Region . . . . . . . . . . . . . . . . . . . . . . . . . . . .
CONTENTS . . . . . . . . . . . . . . . . . . . . 27 27 28 29 33 33 33 33 34 34 34 35 36 37 37 39 40 40 47 47 48 49 51 51 52 56 59 60 63 63 65 65 67 69 71 74 74 76 81 85 85 88 96
5 Bitpacking Convention 5.1 Overview . . . . . . . . . . . . . 5.1.1 Octets and Bytes . . . . . 5.1.2 Words and Byte Order . . 5.1.3 Bit Order . . . . . . . . . 5.2 Coding Bits into Bytes . . . . . . 5.2.1 Signedness . . . . . . . . 5.2.2 Encoding Example . . . . 5.2.3 Decoding Example . . . . 5.2.4 End-of-Packet Alignment 5.2.5 Reading Zero Bit Integers
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
6 Bitstream Headers 6.1 Common Header Decode . . . . . . . . . . 6.2 Identication Header Decode . . . . . . . 6.3 Comment Header . . . . . . . . . . . . . . 6.3.1 Comment Length Decode . . . . . 6.3.2 Comment Header Decode . . . . . 6.3.3 User Comment Format . . . . . . . 6.4 Setup Header . . . . . . . . . . . . . . . . 6.4.1 Loop Filter Limit Table Decode . . 6.4.2 Quantization Parameters Decode . 6.4.3 Computing a Quantization Matrix 6.4.4 DCT Token Human Tables . . . . 6.4.5 Setup Header Decode . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
7 Frame Decode 7.1 Frame Header Decode . . . . . . . . . . . . 7.2 Run-Length Encoded Bit Strings . . . . . . 7.2.1 Long-Run Bit String Decode . . . . 7.2.2 Short-Run Bit String Decode . . . . 7.3 Coded Block Flags Decode . . . . . . . . . 7.4 Macro Block Coding Modes . . . . . . . . . 7.5 Motion Vectors . . . . . . . . . . . . . . . . 7.5.1 Motion Vector Decode . . . . . . . . 7.5.2 Macro Block Motion Vector Decode 7.6 Block-Level qi Decode . . . . . . . . . . . . 7.7 DCT Coecients . . . . . . . . . . . . . . . 7.7.1 EOB Token Decode . . . . . . . . . 7.7.2 Coecient Token Decode . . . . . . 7.7.3 DCT Coecient Decode . . . . . . .
. . . . . . . . . . . . . .
. . . . . . . . . . . . . .
. . . . . . . . . . . . . .
. . . . . . . . . . . . . .
. . . . . . . . . . . . . .
. . . . . . . . . . . . . .
. . . . . . . . . . . . . .
. . . . . . . . . . . . . .
. . . . . . . . . . . . . .
. . . . . . . . . . . . . .
. . . . . . . . . . . . . .
. . . . . . . . . . . . . .
CONTENTS Undoing DC Prediction . . . . . . . . . . . . . . . . . . . 7.8.1 Computing the DC Predictor . . . . . . . . . . . . 7.8.2 Inverting the DC Prediction Process . . . . . . . . 7.9 Reconstruction . . . . . . . . . . . . . . . . . . . . . . . . 7.9.1 Predictors . . . . . . . . . . . . . . . . . . . . . . . 7.9.1.1 The Intra Predictor . . . . . . . . . . . . 7.9.1.2 The Whole-Pixel Predictor . . . . . . . . 7.9.1.3 The Half-Pixel Predictor . . . . . . . . . 7.9.2 Dequantization . . . . . . . . . . . . . . . . . . . . 7.9.3 The Inverse DCT . . . . . . . . . . . . . . . . . . . 7.9.3.1 The 1D Inverse DCT . . . . . . . . . . . 7.9.3.2 The 2D Inverse DCT . . . . . . . . . . . 7.9.3.3 The 1D Forward DCT (Non-Normative) . 7.9.4 The Complete Reconstruction Algorithm . . . . . 7.10 Loop Filtering . . . . . . . . . . . . . . . . . . . . . . . . . 7.10.1 Horizontal Filter . . . . . . . . . . . . . . . . . . . 7.10.2 Vertical Filter . . . . . . . . . . . . . . . . . . . . . 7.10.3 Complete Loop Filter . . . . . . . . . . . . . . . . 7.11 Complete Frame Decode . . . . . . . . . . . . . . . . . . . A Ogg Bitstream Encapsulation A.1 Overview . . . . . . . . . . . . . A.1.1 MIME type . . . . . . . . A.2 Embedding in a logical bitstream A.2.1 Headers . . . . . . . . . . A.2.2 Frame data . . . . . . . . A.2.3 Granule position . . . . . A.3 Multiplexed stream mapping . . A.3.1 Chained streams . . . . . A.3.2 Grouped streams . . . . . B VP3 B.1 VP3 Compatibility . . . . B.2 Loop Filter Limit Values . B.3 Quantization Parameters . B.4 Human Tables . . . . . . C Colophon 7.8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
iii 98 99 103 105 105 106 106 108 110 113 113 118 119 122 129 130 131 133 137 145 145 145 146 146 146 146 147 147 147 149 149 150 150 151 193
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
iv
CONTENTS
List of Figures
2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 4.1 4.2 4.3 4.4 4.5 4.6 4.7 6.1 6.2 6.3 6.4 6.5 7.1 7.2 7.3 Location of frame and picture regions . . . . . . . . . . . . Subdivision of a frame into blocks and super blocks . . . . . Raster ordering of n m blocks . . . . . . . . . . . . . . . . Hilbert curve ordering of blocks within a super block . . . . Subdivision of a frame into macro blocks . . . . . . . . . . . Hilbert curve ordering of macro blocks within a super block Example of reference frames for an inter frame . . . . . . . Zig-zag order . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 7 8 9 10 10 11 12 27 28 29 30 31 31 31 39 41 47 48 51
Pixels encoded 4:4:4 . . . . . . . . . . . . . . . . . . . . . . . . . Pixels encoded 4:2:2 . . . . . . . . . . . . . . . . . . . . . . . . . Pixels encoded 4:2:0 . . . . . . . . . . . . . . . . . . . . . . . . . Pixel correspondence between color planes with even picture oset and even picture size . . . . . . . . . . . . . . . . . . . . . . . Pixel correspondence with even picture oset and odd picture size Pixel correspondence with odd picture oset and odd picture size Pixel correspondence with odd picture oset and even picture size Common Header Packet Layout Identication Header Packet . . Length encoded string layout . Comment Header Layout . . . Setup Header structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Signal Flow Graph for the 1D Inverse DCT . . . . . . . . . . . . 115 Signal Flow Graph for the 1D Forward DCT . . . . . . . . . . . 121 The loop lter response function. . . . . . . . . . . . . . . . . . . 129
vi
LIST OF FIGURES
List of Tables
2.1 3.1 4.1 4.2 6.3 6.4 6.5 6.6 6.18 7.3 7.7 7.11 7.18 7.19 7.23 7.33 7.38 7.42 7.46 7.47 7.65 7.75 7.85 7.89 Color Plane Indices . . . . . . . . . . . . . . . . . . . . . . . . . . Quantization Type Indices . . . . . . . . . . . . . . . . . . . . . . Rec. 470M Parameters . . . . . . . . . . . . . . . . . . . . . . . . Rec. 470BG Parameters . . . . . . . . . . . . . . . . . . . . . . . Enumerated List of Color Spaces . . . . . . . . Enumerated List of Pixel Formats . . . . . . . Number of Super Blocks for each Pixel Format Number of Blocks for each Pixel Format . . . . Minimum Quantization Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 14 25 26 45 46 46 47 58 64 67 68 73 73 83 86 90 98 101 103 115 127 135 142
Frame Type Values . . . . . . . . . . . . . . . . . . . . . . . . Human Codes for Long Run Lengths . . . . . . . . . . . . . Human Codes for Short Run Lengths . . . . . . . . . . . . . Macro Block Coding Modes . . . . . . . . . . . . . . . . . . . Macro Block Mode Schemes . . . . . . . . . . . . . . . . . . . Human Codes for Motion Vector Components . . . . . . . . EOB Token Summary . . . . . . . . . . . . . . . . . . . . . . Coecient Token Summary . . . . . . . . . . . . . . . . . . . Human Table Groups . . . . . . . . . . . . . . . . . . . . . . Reference Frames for Each Coding Mode . . . . . . . . . . . . Weights and Divisors for Each Set of Available DC Predictors 16-bit Approximations of Sines and Cosines . . . . . . . . . . Reference Planes and Sizes for Each r and pli . . . . . . . . Reconstructed Planes and Sizes for Each pli . . . . . . . . . . Width and Height of Chroma Planes for each Pixel Format .
vii
viii
LIST OF TABLES
Notation and Conventions

All parameters either passed in or out of a decoding procedure are given in bold face. The prex b indicates that the following value is to be interpreted as a binary number (base 2). Example: The value b1110100 is equal to the decimal value 116. The prex 0x indicates the the following value is to be interpreted as a hexadecimal number (base 16). Example: The value 0x74 is equal to the decimal value 116. All arithmetic dened by this specication is exact. However, any real numbers that do arise will always be converted back to integers again in short order. The entire specication can be implemented using only normal integer operations. All operations are to be implemented with suciently large integers so that overow cannot occur. Where the result of a computation is to be truncated to a xed-sized binary representation, this will be explicitly noted. The size given for all variables is the maximum number of bits needed to store any value in that variable. Intermediate computations involving that variable may require more bits. The following operators are dened: |a| The absolute value of a number a. |a| = a, a, a<0 a0
a b Multiplication of a number a by a number b.

a b
Exact division of a number a by a number b, producing a potentially noninteger result.
a The largest integer less than or equal to a real number a. a The smallest integer greater than or equal to a real number a. ix
x a//b Integer division of a by b. a//b =

a b a b
, ,
a<0 a0
a%b The remainder from the integer division of a by b. a%b = a |b| a |b|
Note that with this denition, the result is always non-negative and less than |b|. a << b The value obtained by left-shifting the twos complement integer a by b bits. For purposes of this specication, overow is ignored, and so this is equivalent to integer multiplication of a by 2b . a >> b The value obtained by right-shifting the twos complement integer a by b bits, lling in the leftmost bits of the new value with 0 if a is non-negative and 1 if a is negative. This is not equivalent to integer division of a by 2b . Instead, a >> b = a . 2b
round(a) Rounds a number a to the nearest integer, with ties rounded away from 0. round(a) = a a+
1 2 1 2
a0 a>0
sign(a) Returns the sign of a given number. 1 0 sign(a) = 1
a<0 a=0 a>0
ilog(a) The minimum number of bits required to store a positive integer a in twos complement notation, or 0 for a non-positive integer a. ilog(a) = Examples:

0, log2 a + 1,
a0 a>0
ilog(1) = 0 ilog(0) = 0 ilog(1) = 1 ilog(2) = 2

ilog(3) = 2 ilog(4) = 3 ilog(7) = 3
xi
min(a, b) The minimum of two numbers a and b. max(a, b) The maximum of two numbers a and b.
xii
Key words
The key words MUST, MUST NOT, REQUIRED, SHALL, SHALL NOT, SHOULD, SHOULD NOT, RECOMMENDED, MAY, and OPTIONAL in this document are to be intrepreted as described in RFC 2119 [Bra97]. Where such assertions are placed on the contents of a Theora bitstream itself, implementations should be prepared to encounter bitstreams that do not follow these requirements. An applications behavior in the presecence of such nonconforming bitstreams is not dened by this specication, but any reasonable method of handling them MAY be used. By way of example, applications MAY discard the current frame, retain the current output thus far, or attempt to continue on by assuming some default values for the erroneous bits. When such an error occurs in the bitstream headers, an application MAY refuse to decode the entire stream. An application SHOULD NOT allow such non-conformant bitstreams to overow buers and potentially execute arbitrary code, as this represents a serious security risk. An application MUST, however, ensure any bits marked as reserved have the value zero, and refuse to decode the stream if they do not. These are used as place holders for future bitstream features with which the current bitstream is forward-compatible. Such features may not increment the bitstream version number, and can only be recognized by checking the value of these reserved bits.
xiii
xiv
Key words
Chapter 1
Introduction
Theora is a general purpose, lossy video codec. It is based on the VP3 video codec produced by On2 Technologies (http://www.on2.com/). On2 donated the VP3.1 source code to the Xiph.Org Foundation and released it under a BSD-like license. On2 also made an irrevocable, royalty-free license grant for any patent claims it might have over the software and any derivatives. No formal specication exists for the VP3 format beyond this source code, however Mike Melanson maintains a detailed description [Mel04]. Portions of this specication were adopted from that text with permission.
1.1
VP3 and Theora
Theora contains a superset of the features that were available in the original VP3 codec. Content encoded with VP3.1 can be losslessly transcoded into the Theora format. Theora content cannot, in general, be losslessly transcoded into the VP3 format. If a feature is not available in the original VP3 format, this is mentioned when that feature is dened. A complete list of these features appears in Appendix B.1.
1.2
Video Formats
Theora currently supports progressive video data of arbitrary dimensions at a constant frame rate in one of several Y Cb Cr color spaces. The precise denition the supported color spaces appears in Section 4.3. Three dierent chroma subsampling formats are supported: 4:2:0, 4:2:2, and 4:4:4. The precise details of each of these formats and their sampling locations are described in Section 4.4. The Theora format does not support interlaced material, variable frame rates, bit-depths larger than 8 bits per component, nor alternate color spaces such as RGB or arbitrary multi-channel spaces. Black and white content can be eciently encoded, however, because the uniform chroma planes compress well. Support for interlaced material is planned for a future version. 1
CHAPTER 1. INTRODUCTION Note: Infrequently changing frame ratesas when lm and video sequences are cut togethercan be supported in the Ogg container format by chaining several Theora streams together.
Support for increased bit depths or additional color spaces is not planned.
1.3
Classication
Theora is a block-based lossy transform codec that utilizes an 8 8 Type-II Discrete Cosine Transform and block-based motion compensation. This places it in the same class of codecs as MPEG-1, -2, -4, and H.263. The details of how individual blocks are organized and how DCT coecients are stored in the bitstream dier substantially from these codecs, however. Theora supports only intra frames (I frames in MPEG) and inter frames (P frames in MPEG). There is no equivalent to the bi-predictive frames (B frames) found in MPEG codecs.
1.4
Assumptions
The Theora codec design assumes a complex, psychovisually-aware encoder and a simple, low-complexity decoder. Theora provides none of its own framing, synchronization, or protection against transmission errors. An encoder is solely a method of accepting input video frames and compressing these frames into raw, unformatted packets. The decoder then accepts these raw packets in sequence, decodes them, and synthesizes a fascimile of the original video frames. Theora is a free-form variable bit rate (VBR) codec, and packets have no minimum size, maximum size, or xed/expected size. Theora packets are thus intended to be used with a transport mechanism that provides free-form framing, synchronization, positioning, and error correction in accordance with these design assumptions, such as Ogg (for le transport) or RTP (for network multicast). For the purposes of a few examples in this document, we will assume that Theora is embedded in an Ogg stream specically, although this is by no means a requirement or fundamental assumption in the Theora design. The specication for embedding Theora into an Ogg transport stream is given in Appendix A.
1.5
Codec Setup and Probability Model
Theoras heritage is the proprietary commerical codec VP3, and it retains a fair amount of inexibility when compared to Vorbis [Xip02], the rst Xiph.Org codec, which began as a research codec. However, to provide additional scope for encoder improvement, Theora adopts some of the congurable aspects of
1.6. FORMAT CONFORMANCE
decoder setup that are present in Vorbis. This conguration data is not available in VP3, which uses hardcoded values instead. Theora makes the same controversial design decision that Vorbis made to include the entire probability model for the DCT coecients and all the quantization parameters in the bitstream headers. This is often several hundred elds. It is therefore impossible to decode any frame in the stream without having previously fetched the codec info and codec setup headers. Note: Theora can initiate decode at an arbitrary intra-frame packet within a bitstream so long as the codec has been initialized with the setup headers. Thus, Theora headers are both required for decode to begin and relatively large as bitstream headers go. The header size is unbounded, although as a rule-of-thumb less than 16kB is recommended, and Xiph.Orgs reference encoder follows this suggestion. Our own design work indicates that the primary liability of the required header is in mindshare; it is an unusual design and thus causes some amount of complaint among engineers as this runs against current design trends and points out limitations in some existing software/interface designs. However, we nd that it does not fundamentally limit Theoras suitable application space.
1.6
Format Conformance
The Theora format is well-dened by its decode specication; any encoder that produces packets that are correctly decoded by an implementation following this specication may be considered a proper Theora encoder. A decoder must faithfully and completely implement the specication dened herein to be considered a conformant Theora decoder. A decoder need not be implemented strictly as described, but the actual decoder process MUST be entirely mathematically equivalent to the described process. Where appropriate, a non-normative description of encoder processes is included. These sections will be marked as such, and a proper Theora encoder is not bound to follow them.
CHAPTER 1. INTRODUCTION
Chapter 2
Coded Video Structure

Theoras encoding and decoding process is based on 8 8 blocks of pixels. This sections describes how a video frame is laid out, divided into blocks, and how those blocks are organized.
2.1
Frame Layout
A video frame in Theora is a two-dimensional array of pixels. Theora, like VP3, uses a right-handed coordinate system, with the origin in the lower-left corner of the frame. This is contrary to many video formats which use a left-handed coordinate system with the origin in the upper-left corner of the frame. Theora divides the pixel array up into three separate color planes, one for each of the Y , Cb , and Cr components of the pixel. The Y plane is also called the luma plane, and the Cb and Cr planes are also called the chroma planes. Each plane is assigned a numerical value, as shown in Table 2.1. Index 0 1 2 Color Plane Y Cb Cr
Table 2.1: Color Plane Indices In some pixel formats, the chroma planes are subsampled by a factor of two in one or both directions. This means that the width or height of the chroma planes may be half that of the total frame width and height. The luma plane is never subsampled. 5
CHAPTER 2. CODED VIDEO STRUCTURE
2.2
Picture Region
An encoded video frame in Theora is required to have a width and height that are multiples of sixteen, making an integral number of blocks even when the chroma planes are subsampled. However, inside a frame a smaller picture region may be dened to present material whose dimensions are not a multiple of sixteen pixels, as shown in Figure 2.1. The picture region can be oset from the lowerleft corner of the frame by up to 255 pixels in each direction, and may have an arbitrary width and height, provided that it is contained entirely within the coded frame. It is this picture region that contains the actual video data. The portions of the frame which lie outside the picture region may contain arbitrary image data, so the frame must be cropped to the picture region before display. The picture region plays no other role in the decode process, which operates on the entire video frame.
Frame Width
Frame
Picture Height
Picture
Picture Y Offset
(0,0) Picture X Offset Picture Width
Figure 2.1: Location of frame and picture regions
Frame Height
2.3. BLOCKS AND SUPER BLOCKS
2.3
Blocks and Super Blocks
Each color plane is subdivided into blocks of 8 8 pixels. Blocks are grouped into 4 4 arrays called super blocks as shown in Figure 2.2. Each color plane has its own set of blocks and super blocks. If the chroma planes are subsampled, they are still divided into 8 8 blocks of pixels; there are just fewer blocks than in the luma plane. The boundaries of blocks and super blocks in the luma plane do not necessarily coincide with those of the chroma planes, if the chroma planes have been subsampled.
Frame
Super Block (4x4)
... ...
Block 8x8 (0,0)
Figure 2.2: Subdivision of a frame into blocks and super blocks Blocks are accessed in two dierent orders in the various decoder processes. The rst is raster order, illustrated in Figure 2.3. This accesses each block in row-major order, starting in the lower left of the frame and continuing along the bottom row of the entire frame, followed by the next row up, starting on the left edge of the frame, etc. The second is coded order. In coded order, blocks are accessed by super block. Within each frame, super blocks are traversed in raster order, similar to
(n1)*m
n*m1
...
m 2m1
...
m1
Figure 2.3: Raster ordering of n m blocks raster order for blocks. Within each super block, however, blocks are accessed in a Hilbert curve pattern, illustrated in Figure 2.4. If a color plane does not contain a complete super block on the top or right sides, the same ordering is still used, simply with any blocks outside the frame boundary ommitted. To illustrate this ordering, consider a frame that is 240 pixels wide and 48 pixels high. Each row of the luma plane has 30 blocks and 8 super blocks, and there are 6 rows of blocks and two rows of super blocks. When accessed in coded order, each block in the luma plane is assigned the following indices: 123 120 5 4 3 0 122 121 6 7 2 1 125 126 9 8 13 14 124 127 10 11 12 15 ... ... ... ... ... ... 179 176 117 116 115 112 178 177 118 119 114 113
Here the index values specify the order in which the blocks would be accessed. The indices of the blocks are numbered continuously from one color plane to the next. They do not reset to zero at the start of each plane. Instead, the numbering increases continuously from the Y plane to the Cb plane to the Cr
2.4. MACRO BLOCKS
10
11
13
12
14
15
Figure 2.4: Hilbert curve ordering of blocks within a super block plane. The implication is that the blocks from all planes are treated as a unit during the various processing steps. Although blocks are sometimes accessed in raster order, in this document the index associated with a block is always its index in coded order.
2.4
Macro Blocks
A macro block contains a 2 2 array of blocks in the luma plane and the colocated blocks in the chroma planes, as shown in Figure 2.5. Thus macro blocks can represent anywhere from six to twelve blocks, depending on how the chroma planes are subsampled. This is in contrast to super blocks, which only contain blocks from a single color plane. Macro blocks contain information about coding mode and motion vectors for the corresponding blocks in all color planes. Macro blocks are also accessed in a coded order. This coded order proceeds by examining each super block in the luma plane in raster order, and traversing the four macro blocks inside using a smaller Hilbert curve, as shown in Figure 2.6. If the luma plane does not contain a complete super block on the top or right sides, the same ordering is still used, with any macro blocks outside the frame boundary simply omitted. Because the frame size is constrained to be a multiple of 16, there are never any partial macro blocks. Unlike blocks, macro blocks need never be accessed in a pure raster order.
10
Frame: chroma plane
Frame: chroma plane
Frame: luma plane
Macroblock
Block Block 8x8 (0,0)
Figure 2.5: Subdivision of a frame into macro blocks
Figure 2.6: Hilbert curve ordering of macro blocks within a super block
2.5. CODING MODES AND PREDICTION
11
Using the same frame size as the example above, there are 15 macro blocks in each row and 3 rows of macro blocks. The macro blocks are assigned the following indices:
30 1 0
31 2 3
32 5 4
33 6 7
42 25 24
43 26 27
44 29 28
2.5
Coding Modes and Prediction
Each block is coded using one of a small, xed set of coding modes that dene how the block is predicted from previous frames. A block is predicted using one of two reference frames, selected according to the coding mode. A reference frame is the fully decoded version of a previous frame in the stream. The rst available reference frame is the previous intra frame, called the golden frame. The second available reference frame is the previous frame, whether it was an intra frame or an inter frame. If the previous frame was an intra frame, then both reference frames are the same. See Figure 2.7 for an illustration of the reference frames used for an intra frame that does not follow an intra frame.
Intra
Inter
Inter
Inter
Inter
Inter
Inter
...
golden frame reference frame current frame
Figure 2.7: Example of reference frames for an inter frame Two coding modes in particular are worth mentioning here. The INTRA mode is used for blocks that are not predicted from either reference frame. This is the only coding mode allowed in intra frames. The INTER NOMV coding mode uses the co-located contents of the block in the previous frame as the predictor. This is the default coding mode.
2.6
DCT Coecients
A residual is added to the predicted contents of a block to form the nal reconstruction. The residual is stored as a set of quantized coecients from an
12
integer approximation of a two-dimensional Type II Discrete Cosine Transform. The DCT takes an 8 8 array of pixel values as input and returns an 8 8 array of coecient values. The natural ordering of these coecients is dened to be row-major order, from lowest to highest frequency. They are also often indexed in zig-zag order, as shown in Figure 2.8. c 0 1 2 3 r 4 5 6 7 0 1 01 2 3 9 10 20 21 4 8 11 19 22 34 2 3 56 7 12 18 23 33 37 13 17 24 32 38 47 4 5 1415 16 25 31 39 46 50 26 30 40 45 51 56 6 7 2728 29 41 44 52 55 59 42 43 53 54 60 61
3536
4849
5758
6263
Figure 2.8: Zig-zag order
Note: the row and column indices refer to frequency number and not pixel locations. The frequency numbers are dened independently of the memory organization of the pixels. They have been written from top to bottom here to follow conventional notation, despite the right-handed coordinate system Theora uses for pixel locations. Many implementations of the DCT operate in-place. That is, they return DCT coecients in the same memory buer that the initial pixel values were stored in. Due to the right-handed coordinate system used for pixel locations in Theora, one must note carefully how both pixel values and DCT coecients are organized in memory in such a system. DCT coecient (0, 0) is called the DC coecient. All the other coecients are called AC coecients.
Chapter 3
Decoding Overview
This section provides a high level description of the Theora codecs construction. A bit-by-bit specication appears beginning in Section 5. The later sections assume a high-level understanding of the Theora decode process, which is provided below.
3.1
Decoder Conguration
Decoder setup consists of conguration of the quantization matrices and the Human codebooks for the DCT coecients, and a table of limit values for the deblocking lter. The remainder of the decoding pipeline is not congurable.
3.1.1
Global Conguration
The global codec conguration consists of a few video related elds, such as frame rate, frame size, picture size and oset, aspect ratio, color space, pixel format, and a version number. The version number is divided into a major version, a minor version, amd a minor revision number. For the format dened in this specication, these are 3, 2, and 1, respectively, in reference to Theoras origin as a successor to the VP3.1 format.
3.1.2
Quantization Matrices
Theora allows up to 384 dierent quantization matrices to be dened, one for each quantization type, color plane (Y , Cb , or Cr ), and quantization index, qi , which ranges from zero to 63, inclusive. There are currently two quantization types dened, which depend on the coding mode of the block being dequantized, as shown in Table 3.1. The quantization index, on the other hand, nominally represents a progressive range of quality levels, from low quality near zero to high quality near 63. However, the interpretation is arbitrary, and it is possible, for example, to partition the scale into two completely separate ranges with 32 levels each that are 13
14 Quantization Type 0 1
CHAPTER 3. DECODING OVERVIEW Usage INTRA-mode blocks Blocks in any other mode.
Table 3.1: Quantization Type Indices meant to represent dierent classes of source material, or any other arrangement that suits the encoders requirements. Each quantization matrix is an 88 matrix of 16-bit values, which is used to quantize the output of the 8 8 DCT. Quantization matrices are specied using three components: a base matrix and two scale values. The rst scale value is the DC scale, which is applied to the DC component of the base matrix. The second scale value is the AC scale, which is applied to all the other components of the base matrix. There are 64 DC scale values and 64 AC scale values, one for each qi value. There are 64 elements in each base matrix, one for each DCT coecient. They are stored in natural order (cf. Section 2.6). There is a separate set of base matrices for each quantization type and each color plane, with up to 64 possible base matrices in each set, one for each qi value. Typically the bitstream contains matrices for only a sparse subset of the possible qi values. The base matrices for the remainder of the qi values are computed using linear interpolation. This conguration allows the encoder to adjust the quantization matrices to approximate the complex, non-linear response of the human visual system to dierent quantization errors. Finally, because the in-loop deblocking lter strength depends on the strength of the quantization matrices dened in this header, a table of 64 loop lter limit values is dened, one for each qi value. The precise specication of how all of this information is decoded appears in Section 6.4.1 and Section 6.4.2.
3.1.3
Human Codebooks
Theora uses 80 congurable binary Human codes to represent the 32 tokens used to encode DCT coecients. Each of the 32 token values has a dierent semantic meaning and is used to represent single coecient values, zero runs, combinations of the two, and End-Of-Block markers. The 80 codes are divided up into ve groups of 16, with each group corresponding to a set of DCT coecient indices. The rst group corresponds to the DC coecient, while the remaining four groups correspond to dierent subsets of the AC coecients. Within each frame, two pairs of 4-bit codebook indices are stored. The rst pair selects which codebooks to use from the DC coecient group for the Y coecients and the Cb and Cr coecients. The second pair selects which codebooks to use from all four of the AC coecient groups for the Y coecients and the Cb and Cr coecients.
3.2. HIGH-LEVEL DECODE PROCESS
15
The precise specication of how the codebooks are decoded appears in Section 6.4.4.
3.2
3.2.1
High-Level Decode Process

Decoder Setup
Before decoding can begin, a decoder MUST be initialized using the bitstream headers corresponding to the stream to be decoded. Theora uses three header packets; all are required, in order, by this specication. Once set up, decode may begin at any intra-frame packetor even inter-frame packets, provided the appropriate decoded reference frames have already been decoded and cached belonging to the Theora stream. In Theora I, all packets after the three initial headers are intra-frame or inter-frame packets. The header packets are, in order, the identication header, the comment header, and the setup header. Identication Header The identication header identies the stream as Theora, provides a version number, and denes the characteristics of the video stream such as frame size. A complete description of the identication header appears in Section 6.2. Comment Header The comment header includes user text comments (tags) and a vendor string for the application/library that produced the stream. The format of the comment header is the same as that used in the Vorbis I and Speex codecs, with slight modications due to the use of a dierent bit packing mechanism. A complete description of how the comment header is coded appears in Section 6.3, along with a suggested set of tags. Setup Header The setup header includes extensive codec setup information, including the complete set of quantization matrices and Human codebooks needed to decode the DCT coecients. A complete description of the setup header appears in Section 6.4.
3.2.2
Decode Procedure
The decoding and synthesis procedure for all video packets is fundamentally the same, with some steps omitted for intra frames.
Decode packet type ag. Decode frame header.
16
CHAPTER 3. DECODING OVERVIEW

Decode coded block information (inter frames only). Decode macro block mode information (inter frames only). Decode motion vectors (inter frames only). Decode block-level qi information. Decode DC coecient for each coded block. Decode 1st AC coecient for each coded block. Decode 2nd AC coecient for each coded block. ... Decode 63rd AC coecient for each coded block. Perform DC coecient prediction. Reconstruct coded blocks. Copy uncoded bocks. Perform loop ltering.
Note: clever rearrangement of the steps in this process is possible. As an example, in a memory-constrained environment, one can make multiple passes through the DCT coecients to avoid buering them all in memory. On the rst pass, the starting location of each coecient is identied, and then 64 separate get pointers are used to read in the 64 DCT coecients required to reconstruct each coded block in sequence. This operation produces entirely equivalent output and is naturally perfectly legal. It may even be a benet in non-memory-constrained environments due to a reduced cache footprint. Theora makes equivalence easy to check by dening all decoding operations in terms of exact integer operations. No oating-point math is required, and in particular, the implementation of the iDCT transform MUST be followed precisely. This prevents the decoder mismatch problem commonly associated with codecs that provide a less rigorous transform specication. Such a mismatch problem would be devastating to Theora, since a single rounding error in one frame could propagate throughout the entire succeeding frame due to DC prediction.
3.2. HIGH-LEVEL DECODE PROCESS Packet Type Decode
17
Theora uses four packet types. The rst three packet types mark each of the three Theora headers described above. The fourth packet type marks a video packet. All other packet types are reserved; packets marked with a reserved type should be ignored. Additionally, zero-length packets are treated as if they were an inter frame with no blocks coded. That is, as a duplicate frame. Frame Header Decode The frame header contains some global information about the current frame. The rst is the frame type eld, which species if this is an intra frame or an inter frame. Inter frames predict their contents from previously decoded reference frames. Intra frames can be independently decoded with no established reference frames. The next piece of information in the frame header is the list of qi values allowed in the frame. Theora allows from one to three dierent qi values to be used in a single frame, each of which selects a set of six quantization matrices, one for each quantization type (inter or intra), and one for each color plane. The rst qi value is always used when dequantizing DC coecients. The qi value used when dequantizing AC coecients, however, can vary from block to block. VP3, in contrast, only allows a single qi value per frame for both the DC and AC coecients. Coded Block Information This stage determines which blocks in the frame are coded and which are uncoded. A coded block list is constructed which lists all the coded blocks in coded order. For intra frames, every block is coded, and so no data needs to be read from the packet. Macro Block Mode Information For intra frames, every block is coded in INTRA mode, and this stage is skipped. In inter frames a coded macro block list is constructed from the coded block list. Any macro block which has at least one of its luma blocks coded is considered coded; all other macro blocks are uncoded, even if they contain coded chroma blocks. A coding mode is decoded for each coded macro block, and assigned to all its constituent coded blocks. All coded chroma blocks in uncoded macro blocks are assigned the INTER NOMV coding mode. Motion Vectors Intra frames are coded entirely in INTRA mode, and so this stage is skipped. Some inter coding modes, however, require one or more motion vectors to be
18
specied for each macro block. These are decoded in this stage, and an appropriate motion vector is assigned to each coded block in the macro block. Block-Level qi Information If a frame allows multiple qi values, the qi value assigned to each block is decoded here. Frames that use only a single qi value have nothing to decode. DCT Coecients Finally, the quantized DCT coecients are decoded. A list of DCT coecients in zig-zag order for a single block is represented by a list of tokens. A token can take on one of 32 dierent values, each with a dierent semantic meaning. A single token can represent a single DCT coecient, a run of zero coecients within a single block, a combination of a run of zero coecients followed by a single non-zero coecient, an End-Of-Block marker, or a run of EOB markers. EOB markers signify that the remainder of the block is one long zero run. Unlike JPEG and MPEG, there is no requirement for each block to end with a special marker. If non-EOB tokens yield values for all 64 of the coecients in a block, then no EOB marker occurs. Each token is associated with a specic token index in a block. For singlecoecient tokens, this index is the zig-zag index of the token in the block. For zero-run tokens, this index is the zig-zag index of the rst coecient in the run. For combination tokens, the index is again the zig-zag index of the rst coecient in the zero run. For EOB markers, which signify that the remainder of the block is one long zero run, the index is the zig-zag index of the rst zero coecient in that run. For EOB runs, the token index is that of the rst EOB marker in the run. Due to zero runs and EOB markers, a block does not have to have a token for every zig-zag index. Tokens are grouped in the stream by token index, not by the block they originate from. This means that for each zig-zag index in turn, the tokens with that index from all the coded blocks are coded in coded block order. When decoding, a current token index is maintained for each coded block. This index is advanced by the number of coecients that are added to the block as each token is decoded. After fully decoding all the tokens with token index ti , the current token index of every coded block will be ti or greater. If an EOB run of n blocks is decoded at token index ti , then it ends the next n blocks in coded block order whose current token index is equal to ti , but not greater. If there are fewer than n blocks with a current token index of ti , then the decoder goes through the coded block list again from the start, ending blocks with a current token index of ti + 1, and so on, until n blocks have been ended. Tokens are read by parsing a Human code that depends on ti and the color plane of the next coded block whose current token index is equal to ti , but not greater. The Human codebooks are selected on a per-frame basis from the 80 codebooks dened in the setup header. Many tokens have a xed number of
3.2. HIGH-LEVEL DECODE PROCESS
19
extra bits associated with them. These bits are read from the packet immediately after the token is decoded. These are used to dene things such as coecient magnitude, sign, and the length of runs. DC Prediction After the coecients for each block are decoded, the quantized DC value of each block is adjusted based on the DC values of its neighbors. This adjustment is performed by scanning the blocks in raster order, not coded block order. Reconstruction Finally, using the coding mode, motion vector (if applicable), quantized coecient list, and qi value dened for each block, all the coded blocks are reconstructed. The DCT coecients are dequantized, an inverse DCT transform is applied, and the predictor is formed from the coding mode and motion vector and added to the result. Loop Filtering To complete the reconstructed frame, an in-loop deblocking lter is applied to the edges of all coded blocks.
20
Chapter 4
Video Formats
This section gives a precise description of the video formats that Theora is capable of storing. The Theora bitstream is capable of handling video at any arbitrary resolution up to 1048560 1048560. Such video would require almost three terabytes of storage per frame for uncompressed data, so compliant decoders MAY refuse to decode images with sizes beyond their capabilities. The remainder of this section talks about two specic aspects of the video format: the color space and the pixel format. The rst describes how color is represented and how to transform that color representation into a device independent color space such as CIE XY Z (1931). The second describes the various schemes for sampling the color values in time and space.
4.1
Color Space Conventions
There are a large number of dierent color standards used in digital video. Since Theora is a lossy codec, it restricts itself to only a few of them to simplify playback. Unlike the alternate method of describing all the parameters of the color model, this allows a few dedicated routines for color conversion to be written and heavily optimized in a decoder. More exible conversion functions should instead be specied in an encoder, where additional computational complexity is more easily tolerated. The color spaces were selected to give a fair representation of color standards in use around the world today. Most of the standards that do not exactly match one of these can be converted to one fairly easily. All Theora color spaces are Y Cb Cr color spaces with one luma channel and two chroma channels. Each channel contains 8-bit discrete values in the range 0 . . . 255, which represent non-linear gamma pre-corrected signals. The Theora identication header contains an 8-bit value that describes the color space. This merely selects one of the color spaces available from an enumerated list. Currently, only two color spaces are dened, with a third possibility that indicates the color space is unknown. 21
22
CHAPTER 4. VIDEO FORMATS
4.2
Color Space Conversions and Parameters
The parameters which describe the conversions between each color space are listed below. These are the parameters needed to map colors from the encoded Y Cb Cr representation to the device-independent color space CIE XY Z (1931). These parameters dene abstract mathematical conversion functions which are innitely precise. The accuracy and precision with which the conversions are performed in a real system is determined by the quality of output desired and the available processing power. Exact decoder output is dened by this specication only in the original Y Cb Cr space. Y Cb Cr to Y Pb Pr : This conversion takes 8-bit discrete values in the range [0 . . . 255] and maps them to real values in the range [0 . . . 1] for Y and [ 1 . . . 1 ] for Pb and 2 2 Pr . Because some values may fall outside the oset and excursion dened for each channel in the Y Cb Cr space, the results may fall outside these ranges in Y Pb Pr space. No clamping should be done at this stage. Yin OsetY ExcursionY Cb OsetCb Pb = ExcursionCb Cr OsetCr Pr = ExcursionCr
Yout =
(4.1) (4.2) (4.3)
Parameters: OsetY,Cb ,Cr , ExcursionY,Cb ,Cr . Y Pb Pr to R G B : This conversion takes the one luma and two chroma channel representation and maps it to the non-linear R G B space used to drive actual output devices. Values should be clamped into the range [0 . . . 1] after this stage.
R = Y + 2(1 Kr )Pr (1 Kb )Kb (1 Kr )Kr Pb 2 Pr 1 Kb Kr 1 Kb Kr B = Y + 2(1 Kb )Pb G =Y 2 Parameters: Kb , Kr . R G B to RGB (Output device gamma correction):
(4.4) (4.5) (4.6)
4.2. COLOR SPACE CONVERSIONS AND PARAMETERS
23
This conversion takes the non-linear R G B voltage levels and maps them to linear light levels produced by the actual output device. Note that this conversion is only that of the output device, and its inverse is not that used by the input device. Because a dim viewing environment is assumed in most television standards, the overall gamma between the input and output devices is usually around 1.1 to 1.2, and not a strict 1.0. For calibration with actual output devices, the model L = (E + ) (4.7)
should be used, with the free parameter and held xed to the value specied in this document. The conversion function presented here is an idealized version with = 0.
R=R G=G B=B Parameters: . RGB to R G B (Input device gamma correction):
(4.8) (4.9) (4.10)
This conversion takes linear light levels and maps them to the non-linear voltage levels produced in the actual input device. This information is merely informative. It is not required for building a decoder or for converting between the various formats and the actual output capabilities of a particular device. A linear segment is introduced on the low end to reduce noise in dark areas of the image. The rest of the scale is adjusted so that the power segment of the curve intersects the linear segment with the proper slope, and so that it still maps 0 to 0 and 1 to 1. R, 0R< (1 + )R , R 1 G, 0G< (1 + )G , G 1 B, 0B< (1 + )B , B 1
R = G = B = Parameters: , , , . RGB to CIE XY Z (1931):
(4.11) (4.12) (4.13)
24
CHAPTER 4. VIDEO FORMATS This conversion maps a device-dependent linear RGB space to the deviceindependent linear CIE XY Z space. The parameters are the CIE chromaticity coordinates of the three primariesred, green, and blueas well as the chromaticity coordinates of the white point of the device. This is how hardware manufacturers and standards typically describe a particular RGB space. The math required to convert these parameters into a useful transformation matrix is reproduced below.
xg yg
F =
xr yr
xb yb
(4.14)
1
1xr yr yr
1
1xg yg yg
1
1xb yb yb
xw sr yw sg = F 1 1 1xw yw sb y w sr R X Y = F sg G sb B Z Parameters: xr , xg , xb , xw , yr , yg , yb , yw .
(4.15)
(4.16)
4.3
Available Color Spaces
These are the color spaces currently dened for use by Theora video. Each one has a short name, with which it is referred to in this document, and a more detailed specication of the standards from which its parameters are derived. Some standards do not specify all the parameters necessary. For these unspecied parameters, this document serves as the denition of what should be used when encoding or decoding Theora video.
4.3.1
Rec. 470M (Rec. ITU-R BT.470-6 System M/NTSC with Rec. ITU-R BT.601-5)
This color space is used by broadcast television and DVDs in much of the Americas, Japan, Korea, and the Union of Myanmar [ITU98]. This color space may also be used for System M/PAL (Brazil), with an appropriate conversion supplied by the encoder to compensate for the dierent gamma value. See Section 4.3.2 for an appropriate gamma value to assume for M/PAL input. In the US, studio monitors are adjusted to a D65 white point (xw , yw = 0.313, 0.329). In Japan, studio monitors are adjusted to a D white of 9300K (xw , yw = 0.285, 0.293). Rec. 470 does not specify a digital encoding of the color signals. For Theora, Rec. ITU-R BT.601-5 [ITU95] is used, starting from the R G B signals specied by Rec. 470.
4.3. AVAILABLE COLOR SPACES
25
Rec. 470 does not specify an input gamma function. For Theora, the Rec. 709 [ITU02] input function is assumed. This is the same as that specied by SMPTE 170M [SMP94], which claims to reect modern practice in the creation of NTSC signals circa 1994. The parameters for all the color transformations dened in Section 4.2 are given in Table 4.1.
OsetY,Cb ,Cr = (16, 128, 128) ExcursionY,Cb ,Cr = (219, 224, 224) Kr = 0.299 Kb = 0.114 = 2.2 = 0.45 = 4.5 = 0.018 = 0.099 xr , yr = 0.67, 0.33 xg , yg = 0.21, 0.71 xb , yb = 0.14, 0.08 (Illuminant C) xw , yw = 0.310, 0.316
Table 4.1: Rec. 470M Parameters
4.3.2
Rec. 470BG (Rec. ITU-R BT.470-6 Systems B and G with Rec. ITU-R BT.601-5)
This color space is used by the PAL and SECAM systems in much of the rest of the world [ITU98] This can be used directly by systems (B, B1, D, D1, G, H, I, K, N)/PAL and (B, D, G, H, K, K1, L)/SECAM. Note: the Rec. 470BG chromaticity values are dierent from those specied in Rec. 470M. When PAL and SECAM systems were rst designed, they were based upon the same primaries as NTSC. However, as methods of making color picture tubes have changed, the primaries used have changed as well. The U.S. recommends using correction circuitry to approximate the existing, standard NTSC primaries. Current PAL and SECAM systems have standardized on primaries in accord with more recent technology.
26
Rec. 470 provisionally permits the use of the NTSC chromaticity values (given in Section 4.3.1) with legacy PAL and SECAM equipment. In Theora, material must be decoded assuming the new PAL and SECAM primaries. Material intended for display on old legacy devices should be converted by the decoder. The ocial Rec. 470BG species a gamma value of = 2.8. However, in practice this value is unrealistically high [Poy97]. Rec. 470BG states that the overall system gamma should be approximately = 1.2. Since most cameras pre-correct with a gamma value of = 0.45, this suggests an output device gamma of approximately = 2.67. This is the value recommended for use with PAL systems in Theora. Rec. 470 does not specify a digital encoding of the color signals. For Theora, Rec. ITU-R BT.601-5 [ITU95] is used, starting from the R G B signals specied by Rec. 470. Rec. 470 does not specify an input gamma function. For Theora, the Rec 709 [ITU02] input function is assumed. The parameters for all the color transformations dened in Section 4.2 are given in Table 4.2.
OsetY,Cb ,Cr = (16, 128, 128) ExcursionY,Cb ,Cr = (219, 224, 224) Kr = 0.299 Kb = 0.114 = 2.67 = 0.45 = 4.5 = 0.018 = 0.099 xr , yr = 0.64, 0.33 xg , yg = 0.29, 0.60 xb , yb = 0.15, 0.06 (D65) xw , yw = 0.313, 0.329
Table 4.2: Rec. 470BG Parameters
4.4
Pixel Formats
Theora supports several dierent pixel formats, each of which uses dierent subsampling for the chroma planes relative to the luma plane. A decoder may
4.4. PIXEL FORMATS
27
need to recover a full resolution chroma plane with samples co-sited with the luma plane in order to convert to RGB for display or perform other processing. Decoders can assume that the chroma signal satises the Nyquist-Shannon sampling theorem. The ideal low-pass reconstruction lter this implies is not practical, but any suitable approximation can be used, depending on the available computing power. Decoders MAY simply use a box lter, assigning to each luma sample the chroma sample closest to it. Encoders would not go wrong in assuming that this will be the most common approach.
4.4.1
4:4:4 Subsampling
All three color planes are stored at full resolutioneach pixel has a Y , a Cb and a Cr value (see Figure 4.1). The samples in the dierent planes are all at co-located sites.
Frame: chroma plane Cr
Frame: chroma plane Cb
Frame: luma planeY
Pixels
(0,0)
Figure 4.1: Pixels encoded 4:4:4
4.4.2
4:2:2 Subsampling
The Cb and Cr planes are stored with half the horizontal resolution of the Y plane. Thus, each of these planes has half the number of horizontal blocks as the
28
luma plane (see Figure 4.2). Similarly, they have half the number of horizontal super blocks, rounded up. Macro blocks are dened across color planes, and so their number does not change, but each macro block contains half as many chroma blocks. The chroma samples are vertically aligned with the luma samples, but horizontally centered between two luma samples. Thus, each luma sample has a unique closest chroma sample. A horizontal phase shift may be required to produce signals which use dierent horizontal chroma sampling locations for compatibility with dierent systems.
Frame: luma planeY
Pixels
(0,0)
4.4.3
4:2:0 Subsampling
The Cb and Cr planes are stored with half the horizontal and half the vertical resolution of the Y plane. Thus, each of these planes has half the number of horizontal blocks and half the number of vertical blocks as the luma plane, for a total of one quarter the number of blocks (see Figure 4.3). Similarly, they have half the number of horizontal super blocks and half the number of vertical super blocks, rounded up. Macro blocks are dened across color planes, and so their
4.4. PIXEL FORMATS
29
number does not change, but each macro block contains within it one quarter as many chroma blocks. The chroma samples are vertically and horizontally centered between four luma samples. Thus, each luma sample has a unique closest chroma sample. This is the same sub-sampling pattern used with JPEG, MJPEG, and MPEG1, and was inherited from VP3. A horizontal or vertical phase shift may be required to produce signals which use dierent chroma sampling locations for compatibility with dierent systems.
Frame: luma planeY
Pixels
(0,0)
4.4.4
Subsampling and the Picture Region
Although the frame size must be an integral number of macro blocks, and thus both the number of pixels and the number of blocks in each direction must be even, no such requirement is made of the picture region. Thus, when using subsampled pixel formats, careful attention must be paid to which chroma samples correspond to which luma samples. As mentioned above, for each pixel format, there is a unique chroma sample that is the closest to each luma sample. When cropping the chroma planes to the picture region, all the chroma samples corresponding to a luma sample in
30
the cropped picture region must be included. Thus, when dividing the width or height of the picture region by two to obtain the size of the subsampled chroma planes, they must be rounded up. Furthermore, the sampling locations are dened relative to the frame, not the picture region. When using the 4:2:2 and 4:2:0 formats, the locations of chroma samples relative to the luma samples depends on whether or not the X oset of the picture region is odd. If the oset is even, each column of chroma samples corresponds to two columns of luma samples (see Figure 4.4 for an example). The only exception is if the width is odd, in which case the last column corresponds to only one column of luma samples (see Figure 4.5). If the oset is odd, then the rst column of chroma samples corresponds to only one column of luma samples, while the remaining columns each correspond to two (see Figure 4.6). In this case, if the width is even, the last column again corresponds to only one column of luma samples (see Figure 4.7). A similar process is followed with the rows of a picture region of odd height encoded in the 4:2:0 format. If the Y oset is even, each row of chroma samples corresponds to two rows of luma samples (see Figure 4.4), except with an odd height, where the last row corresponds to one row of chroma luna samples only (see Figure 4.5). If the oset is odd, then it is the rst row of chroma samples which corresponds to only one row of luma samples, while the remaining rows each correspond to two (Figure 4.6), except with an even height, where the last row also corresponds to one (Figure 4.7). Encoders should be aware of these dierences in the subsampling when using an even or odd oset. In the typical case, with an even width and height, where one expects two rows or columns of luma samples for every row or column of chroma samples, the encoder must take care to ensure that the osets used are both even.
Frame: chroma (4:2:2 case) Frame: luma
... ...
(0,0)
Frame: chroma (4:2:0 case)
... ...
Pixels
...
(0,0)
...
(0,0)
Figure 4.4: Pixel correspondence between color planes with even picture oset and even picture size
4.4. PIXEL FORMATS

31
... ...
(0,0)
... ...
Pixels
...
(0,0)
...
(0,0)
Figure 4.5: Pixel correspondence with even picture oset and odd picture size
... ...
(0,0)
... ...
Pixels
...
(0,0)
...
(0,0)
Figure 4.6: Pixel correspondence with odd picture oset and odd picture size
... ...
(0,0)
... ...
Pixels
...
(0,0)
...
(0,0)
Figure 4.7: Pixel correspondence with odd picture oset and even picture size
32
Chapter 5
Bitpacking Convention
5.1 Overview
The Theora codec uses relatively unstructured raw packets containing binary integer elds of arbitrary width. Logically, each packet is a bitstream in which bits are written one-by-one by the encoder and then read one-by-one in the same order by the decoder. Most current binary storage arrangements group bits into a native storage unit of eight bits (octets), sixteen bits, thirty-two bits, or less commonly other xed sizes. The Theora bitpacking convention species the correct mapping of the logical packet bitstream into an actual representation in xed-width units.
5.1.1
Octets and Bytes
In most contemporary architectures, a byte is synonymous with an octect, that is, eight bits. For purposes of the bitpacking convention, a byte implies the smallest native integer storage representation oered by a platform. Modern le systems invariably oer bytes as the fundamental atom of storage. The most ubiquitous architectures today consider a byte to be an octet. Note, however, that the Theora bitpacking convention is still well dened for any native byte size; an implementation can use the native bit-width of a given storage system. This document assumes that a byte is one octet for purposes of example only.
5.1.2
Words and Byte Order
A word is an integer size that is a grouped multiple of the byte size. Most architectures consider a word to be a group of two, four, or eight bytes. Each byte in the word can be ranked by order of signicance, e.g. the signicance of the bits in each byte when storing a binary integer in the word. Several byte orderings are possible in a word. The common ones are 33
34
CHAPTER 5. BITPACKING CONVENTION

Big-endian: in which the most signicant byte comes rst, e.g. 3-2-1-0, Little-endian: in which the least signicant byte comes rst, e.g. 0-1-2-3, and Mixed-endian: one of the less-common orderings that cannot be put into the above two categories, e.g. 3-1-2-0 or 0-2-1-3.
The Theora bitpacking convention species storage and bitstream manipulation at the byte, not word, level. Thus host word ordering is of a concern only during optimization, when writing code that operates on a word of storage at a time rather than a byte. Logically, bytes are always encoded and decoded in order from byte zero through byte n.
5.1.3
Bit Order
A byte has a well-dened least signicant bit (LSb), which is the only bit set when the byte is storing the twos complement integer value +1. A bytes most signicant bit (MSb) is at the opposite end. Bits in a byte are numbered from zero at the LSb to n for the MSb, where n = 7 in an octet.
5.2
Coding Bits into Bytes
The Theora codec needs to encode arbitrary bit-width integers from zero to 32 bits wide into packets. These integer elds are not aligned to the boundaries of the byte representation; the next eld is read at the bit position immediately after the end of the previous eld. The decoder logically unpacks integers by rst reading the MSb of a binary integer from the logical bitstream, followed by the next most signicant bit, etc., until the required number of bits have been read. When unpacking the bytes into bits, the decoder begins by reading the MSb of the integer to be read from the most signicant unread bit position of the source byte, followed by the next-most signicant bit position of the destination integer, and so on up to the requested number of bits. Note that this diers from the Vorbis I codec, which begins decoding with the LSb of the source integer, reading it from the LSb of the source byte. When all the bits of the current source byte are read, decoding continues with the MSb of the next byte. Any unlled bits in the last byte of the packet MUST be cleared to zero by the encoder.
5.2.1
Signedness
The binary integers decoded by the above process may be either signed or unsigned. This varies from integer to integer, and this specication indicates how each value should be interpreted as it is read. That is, depending on context, the three bit binary pattern b111 can be taken to represent either 7 as an unsigned integer or 1 as a signed, twos complement integer.
5.2. CODING BITS INTO BYTES
35
5.2.2
Encoding Example
The following example shows the state of an (8-bit) byte stream after several binary integers are encoded, including the location of the put pointer for the next bit to write to and the total length of the stream in bytes. Encode the 4 bit unsigned integer value 12 (b1100) into an empty byte stream. 7 1 0 0 0 0 6 1 0 0 0 0 5 0 0 0 0 0 4 0 0 0 0 . . . 0 0 0 0 0 byte stream length: 1 byte 3 0 0 0 0 2 0 0 0 0 1 0 0 0 0 0 0 0 0 0
byte byte byte byte . . .
0 1 2 3
byte n
Continue by encoding the 3 bit signed integer value -1 (b111). 0 0 0 0 0 0
0 1 2 3
7 1 0 0 0 0
6 1 0 0 0 0
5 0 0 0 0 0
4 0 0 0 0 . . . 0
3 1 0 0 0 0
2 1 0 0 0 0
1 1 0 0 0 0
byte n
byte stream length: 1 byte
Continue by encoding the 7 bit integer value 17 (b0010001). 1 1 0 0 0 0
0 1 2 3
7 1 0 0 0 0
6 1 1 0 0 0
5 0 0 0 0 0
4 0 0 0 0 . . . 0
3 1 0 0 0 0
2 1 1 0 0 0
0 0 0 0 0 0
byte n
byte stream length: 2 bytes
Continue by encoding the 13 bit integer value 6969 (b11011 00111001).
36 4 0 0 0 0 . . . 0 0 0 0
0 1 2 3
7 1 0 0 0
6 1 1 1 0
5 0 0 1 1
3 1 0 0 0 0
2 1 1 1 0 0
1 1 1 1 0 0
0 0 1 1 0 0
byte stream length: 4 bytes
byte n
5.2.3
Decoding Example
The following example shows the state of the (8-bit) byte stream encoded in the previous example after several binary integers are decoded, including the location of the get pointer for the next bit to read. Read a two bit unsigned integer from the example encoded above. 7 6 5 4 3 2 1 0 byte 0 1 1 0 0 1 1 1 0 byte 1 0 1 0 0 0 1 1 1 byte 2 0 1 1 0 0 1 1 1 byte 3 0 0 1 0 0 0 0 0 byte stream length: 4 bytes Value read: 3 (b11). Read another two bit unsigned integer from the example encoded above. 7 6 5 4 3 2 1 0 byte 0 1 1 0 0 1 1 1 0 byte 1 0 1 0 0 0 1 1 1 byte 2 0 1 1 0 0 1 1 1 byte 3 0 0 1 0 0 0 0 0 byte stream length: 4 bytes Value read: 0 (b00). Two things are worth noting here.
Although these four bits were originally written as a single four-bit integer, reading some other combination of bit-widths from the bitstream is well dened. No articial alignment boundaries are maintained in the bitstream. The rst value is the integer 3 only because the context stated we were reading an unsigned integer. Had the context stated we were reading a signed integer, the returned value would have been the integer 1.
5.2. CODING BITS INTO BYTES
37
5.2.4
End-of-Packet Alignment
The typical use of bitpacking is to produce many independent byte-aligned packets which are embedded into a larger byte-aligned container structure, such as an Ogg transport bitstream. Externally, each bitstream encoded as a byte stream MUST begin and end on a byte boundary. Often, the encoded packet bitstream is not an integer number of bytes, and so there is unused space in the last byte of a packet. When a Theora encoder produces packets for embedding in a byte-aligned container, unused space in the last byte of a packet is always zeroed during the encoding process. Thus, should this unused space be read, it will return binary zeroes. There is no marker pattern or stung bits that will allow the decoder to obtain the exact size, in bits, of the original bitstream. This knowledge is not required for decoding. Attempting to read past the end of an encoded packet results in an end-ofpacket condition. Any further read operations after an end-of-packet condition shall also return end-of-packet. Unlike Vorbis, Theora does not use truncated packets as a normal mode of operation. Therefore if a decoder encounters the end-of-packet condition during normal decoding, it may attempt to use the bits that were read to recover as much of encoded data as possible, signal a warning or error, or both.
5.2.5
Reading Zero Bit Integers
Reading a zero bit integer returns the value 0 and does not increment the stream pointer. Reading to the end of the packet, but not past the end, so that an end-of-packet condition is not triggered, and then reading a zero bit integer shall succeed, returning 0, and not trigger an end-of-packet condition. Reading a zero bit integer after a previous read sets the end-of-packet condition shall fail, also returning end-of-packet.
38
Chapter 6
Bitstream Headers
A Theora bitstream begins with three header packets. The header packets are, in order, the identication header, the comment header, and the setup header. All are required for decode compliance. An end-of-packet condition encountered while decoding the identication or setup header packets renders the stream undecodable. An end-of-packet condition encountered while decode the comment header is a non-fatal error condition, and MAY be ignored by a decoder.
VP3 Compatibility VP3 relies on the headers provided by its container, usually either AVI or Quicktime. As such, several parameters available in these headers are not available to VP3 streams. These are indicated as they appear in the sections below.
0 1 2 3 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | header type | t | h | e | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | o | r | a | data... | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | ... header-specific data ... | | ... | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ Figure 6.1: Common Header Packet Layout 39
40
CHAPTER 6. BITSTREAM HEADERS
6.1
None.
Common Header Decode
Input parameters:
Output parameters:
Name HEADERTYPE
Type Integer
Size (bits) 8
Signed? No
Description and restrictions The type of the header being decoded.
Variables used: None. Each header packet begins with the same header elds, which are decoded as follows: 1. Read an 8-bit unsigned integer as HEADERTYPE. If the most signicant bit of this integer is not set, then stop. This is not a header packet. 2. Read 6 8-bit unsigned integers. If these do not have the values 0x74, 0x68, 0x65, 0x6F, 0x72, and 0x61, respectively, then stop. This stream is not decodable by this specication. These values correspond to the ASCII values of the characters t, h, e, o, r, and a. Decode continues according to HEADERTYPE. The identication header is type 0x80, the comment header is type 0x81, and the setup header is type 0x82. These packets must occur in the order: identication, comment, setup. All header packets have the most signicant bit of the type eldwhich is the initial bit in the packetset. This distinguishes them from video data packets in which the rst bit is unset. Packets with other header types (0x830xFF) are reserved and MUST be ignored.
6.2
None.
Identication Header Decode
Input parameters:
6.2. IDENTIFICATION HEADER DECODE
41
0 1 2 3 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | 0x80 | t | h | e | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | o | r | a | VMAJ | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | VMIN | VREV | FMBW | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | FMBH | PICW... | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | ...PICW | PICH | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | PICX | PICY | FRN... | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | ...FRN | FRD... | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | ...FRD | PARN... | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | ...PARN | PARD | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | CS | NOMBR | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | QUAL | KFGSHIFT| PF| Res | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ Figure 6.2: Identication Header Packet
42 Output parameters:
6.2. IDENTIFICATION HEADER DECODE Name VMAJ VMIN VREV FMBW FMBH NSBS NBS NMBS PICW PICH PICX PICY FRN FRD PARN PARD CS PF NOMBR QUAL KFGSHIFT Type Integer Integer Integer Integer Integer Integer Integer Integer Integer Integer Integer Integer Integer Integer Integer Integer Integer Integer Integer Integer Integer Size (bits) 8 8 8 16 16 32 36 32 20 20 8 8 32 32 24 24 8 2 24 6 5 Signed? No No No No No No No No No No No No No No No No No No No No No Description and restrictions
43
The major version number. The minor version number. The version revision number. The width of the frame in macro blocks. The height of the frame in macro blocks. The total number of super blocks in a frame. The total number of blocks in a frame. The total number of macro blocks in a frame. The width of the picture region in pixels. The height of the picture region in pixels. The X oset of the picture region in pixels. The Y oset of the picture region in pixels. The frame-rate numerator. The frame-rate denominator. The pixel aspect-ratio numerator. The pixel aspect-ratio denominator. The color space. The pixel format. The nominal bitrate of the stream, in bits per second. The quality hint. The amount to shift the key frame number by in the granule position.
Variables used: None. The identication header is a short header with only a few elds used to declare the stream denitively as Theora and provide detailed information about the format of the fully decoded video data. The identication header is decoded as follows:
44
CHAPTER 6. BITSTREAM HEADERS 1. Decode the common header elds according to the procedure described in Section 6.1. If HEADERTYPE returned by this procedure is not 0x80, then stop. This packet is not the identication header. 2. Read an 8-bit unsigned integer as VMAJ. If VMAJ is not 3, then stop. This stream is not decodable according to this specication. 3. Read an 8-bit unsigned integer as VMIN. If VMIN is not 2, then stop. This stream is not decodable according to this specication. 4. Read an 8-bit unsigned integer as VREV. If VREV is greater than 1, then this stream may contain optional features or interpretational changes documented in a future version of this specication. Regardless of the value of VREV, the stream is decodable according to this specication. 5. Read a 16-bit unsigned integer as FMBW. This MUST be greater than zero. This species the width of the coded frame in macro blocks. The actual width of the frame in pixels is FMBW 16. 6. Read a 16-bit unsigned integer as FMBH. This MUST be greater than zero. This species the height of the coded frame in macro blocks. The actual height of the frame in pixels is FMBH 16. 7. Read a 24-bit unsigned integer as PICW. This MUST be no greater than (FMBW 16). Note that 24 bits are read, even though only 20 bits are sucient to specify any value of the picture width. This is done to preserve octet alignment in this header, to allow for a simplied parser implementation. 8. Read a 24-bit unsigned integer as PICH. This MUST be no greater than (FMBH 16). Together with PICW, this species the size of the displayable picture region within the coded frame. See Figure 2.1. Again, 24 bits are read instead of 20. 9. Read an 8-bit unsigned integer as PICX. This MUST be no greater than (FMBW 16 PICX).
10. Read an 8-bit unsigned integer as PICY. This MUST be no greater than (FMBH 16 PICY). Together with PICX, this species the location of the lower-left corner of the displayable picture region. See Figure 2.1. 11. Read a 32-bit unsigned integer as FRN. This MUST be greater than zero. 12. Read a 32-bit unsigned integer as FRD. This MUST be greater than zero. Theora is a xed-frame rate video codec. Frames are sampled at the constant rate of FRN frames per second. The presentation time of FRD the rst frame is at zero seconds. No mechanism is provided to specify a non-zero oset for the initial frame.
6.2. IDENTIFICATION HEADER DECODE 13. Read a 24-bit unsigned integer as PARN.
45
14. Read a 24-bit unsigned integer as PARD. Together with PARN, these specify the aspect ratio of the pixels within a frame, dened as the ratio of the physical width of a pixel to its physical height. This is given by the ratio PARN : PARD. If either of these elds are zero, this indicates that pixel aspect ratio information was not available to the encoder. In this case it MAY be specied by the application via an external means, or a default value of 1 : 1 MAY be used. 15. Read an 8-bit unsigned integer as CS. This is a value from an enumerated list of the available color spaces, given in Table 6.3. The Undened value indicates that color space information was not available to the encoder. It MAY be specied by the application via an external means. If a reserved value is given, a decoder MAY refuse to decode the stream. Value 0 1 2 3 . . . 255 Color Space Undened. Rec. 470M (see Section 4.3.1). Rec. 470BG (see Section 4.3.2). Reserved.
Table 6.3: Enumerated List of Color Spaces
16. Read a 24-bit unsigned integer as NOMBR signifying a rate in bits per second. Rates equal to or greater than 224 1 bits per second are represented as 224 1. The NOMBR eld is used only as a hint. For pure VBR streams, this value may be considerably o. The eld MAY be set to zero to indicate that the encoder did not care to speculate. 17. Read a 6-bit unsigned integer as QUAL. This value is used to provide a hint as to the relative quality of the stream when compared to others produced by the same encoder. Larger values indicate higher quality. This can be used, for example, to select among several streams containing the same material encoded with dierent settings. 18. Read a 5-bit unsigned integer as KFGSHIFT. The KFGSHIFT is used to partition the granule position associated with each packet into two dierent parts. The frame number of the last key frame, starting from zero, is stored in the upper 64KFGSHIFT bits, while the lower KFGSHIFT bits contain the number of frames since the last keyframe. Complete details on the granule position mapping are specied in Section REF.
46
19. Read a 2-bit unsigned integer as PF. The PF eld contains a value from an enumerated list of the available pixel formats, given in Table 6.4. If the reserved value 1 is given, stop. This stream is not decodable according to this specication. Value 0 1 2 3 Pixel Format 4:2:0 (see Section 4.4.3). Reserved. 4:2:2 (see Section 4.4.2). 4:4:4 (see Section 4.4.1).
Table 6.4: Enumerated List of Pixel Formats 20. Read a 3-bit unsigned integer. These bits are reserved. If this value is not zero, then stop. This stream is not decodable according to this specication. 21. Assign NSBS a value according to PF, as given by Table 6.5. 22. Assign NBS a value according to PF, as given by Table 6.6. 23. Assign NMBS the value (FMBW FMBH).
VP3 Compatibility VP3 does not correctly handle frame sizes that are not a multiple of 16. Thus, PICW and PICH should be set to the frame width and height in pixels, respectively, and PICX and PICY should be set to zero. VP3 headers do not specify a color space. VP3 only supports the 4:2:0 pixel format.
PF 0 2 3
NSBS ((FMBW + 1)//2) ((FMBH + 1)//2) + 2 ((FMBW + 3)//4) ((FMBH + 3)//4) ((FMBW + 1)//2) ((FMBH + 1)//2) + 2 ((FMBW + 3)//4) ((FMBH + 1)//2) 3 ((FMBW + 1)//2) ((FMBH + 1)//2)
Table 6.5: Number of Super Blocks for each Pixel Format
6.3. COMMENT HEADER PF 0 2 3 NBS 6 FMBW FMBH 8 FMBW FMBH 12 FMBW FMBH
47
Table 6.6: Number of Blocks for each Pixel Format 4 byte length UTF-8 encoded string ...
Figure 6.3: Length encoded string layout
6.3
Comment Header
The Theora comment header is the second of three header packets that begin a Theora stream. It is meant for short text comments, not aribtrary metadata; arbitrary metadata belongs in a separate logical stream that provides greater structure and machine parseability. The comment eld is meant to be used much like someone jotting a quick note on the label of a video. It should be a little information to remember the disc or tape by and explain it to others; a short, to-the-point text note that can be more than a couple words, but isnt going to be more than a short paragraph. The essentials, in other words, whatever they turn out to be, e.g.: The comment header is stored as a logical list of eight-bit clean vectors; the number of vectors is bounded at 232 1 and the length of each vector is limited to 232 1 bytes. The vector length is encoded; the vector contents themselves are not null terminated. In addition to the vector list, there is a single vector for a vendor name, also eight-bit clean with a length encoded in 32 bits.
6.3.1
None.
Comment Length Decode
Input parameters:
Output parameters:
Name LEN
Type Integer
Size (bits) 32
Signed? No
Description and restrictions A single 32-bit length value.
Variables used:
Name LEN0 LEN1 LEN2
Type Integer Integer Integer
Size (bits) 8 8 8
Signed? No No No
Description and restrictions The rst octet of the string length. The second octet of the string length. The third octet of the string length.
48
CHAPTER 6. BITSTREAM HEADERS vendor string number of comments comment string comment string ... Figure 6.4: Comment Header Layout 1. Read an 8-bit unsigned integer as LEN0. 2. Read an 8-bit unsigned integer as LEN1. 3. Read an 8-bit unsigned integer as LEN2. 4. Read an 8-bit unsigned integer as LEN3. 5. Assign LEN the value (LEN0+(LEN1 << 8)+(LEN2 << 16)+(LEN3 << 24)). This construction is used so that on platforms with 8-bit bytes, the memory organization of the comment header is identical with that of Vorbis I, allowing for common parsing code despite the dierent bit packing conventions.
6.3.2
None.
Comment Header Decode
Input parameters:
Output parameters:
Name VENDOR NCOMMENTS COMMENTS
Type String Integer
Size (bits) 32
Signed?
Description and restrictions The vendor string. The number of user comments. A list of NCOMMENTS user comment values.
No
String Array
Variables used:
Name ci
Type Integer
Size (bits) 32
Signed? No
Description and restrictions The index of the current user comment.
The complete comment header is decoded as follows: 1. Decode the common header elds according to the procedure described in Section 6.1. If HEADERTYPE returned by this procedure is not 0x81, then stop. This packet is not the comment header.
6.3. COMMENT HEADER
49
2. Decode the length of the vendor string using the procedure given in Section 6.3.1 into LEN. 3. Read LEN 8-bit unsigned integers. 4. Set the string VENDOR to the contents of these octets. 5. Decode the number of user comments using the procedure given in Section 6.3.1 into LEN. 6. Assign NCOMMENTS the value stored in LEN. 7. For each consecutive value of ci from 0 to (NCOMMENTS 1), inclusive: (a) Decode the length of the current user comment using the procedure given in Section 6.3.1 into LEN. (b) Read LEN 8-bit unsigned integers. (c) Set the string COMMENTS[ci ] to the contents of these octets. The comment header comprises the entirety of the second header packet. Unlike the rst header packet, it is not generally the only packet on the second page and may span multiple pages. The length of the comment header packet is (practically) unbounded. The comment header packet is not optional; it must be present in the stream even if it is logically empty.
6.3.3
User Comment Format
The user comment vectors are structured similarly to a UNIX environment variable. That is, comment elds consist of a eld name and a corresponding value and look like: COMMENTS[0] COMMENTS[1] = = TITLE=the look of Theora DIRECTOR=me
The eld name is case-insensitive and MUST consist of ASCII characters 0x20 through 0x7D, 0x3D (=) excluded. ASCII 0x41 through 0x5A inclusive (characters AZ) are to be considered equivalent to ASCII 0x61 through 0x7A inclusive (characters az). An entirely empty eld nameone that is zero characters longis not disallowed. The eld name is immediately followed by ASCII 0x3D (=); this equals sign is used to terminate the eld name. The data immediately after 0x3D until the end of the vector is the eight-bit clean value of the eld contents encoded as a UTF-8 string [Yer96]. Field names MUST NOT be internationalized; this is a concession to simplicity, not an attempt to exclude the majority of the world that doesnt speak English. Applications MAY wish to present internationalized versions of the standard eld names listed below to the user, but they are not to be stored in
50
the bitstream. Field contents, however, use the UTF-8 character encoding to allow easy representation of any language. Individual vendors MAY use non-standard eld names within reason. The proper use of comment elds as human-readable notes has already been explained. Abuse will be discouraged. There is no vendor-specic prex to non-standard eld names. Vendors SHOULD make some eort to avoid arbitrarily polluting the common namespace. Xiph.Org and other bodies will generally collect and rationalize the more useful tags to help with standardization. Field names are not restricted to occur only once within a comment header. Field Names Below is a proposed, minimal list of standard eld names with a description of their intended use. No eld names are mandatory; a comment header may contain one or more, all, or none of the names in this list. TITLE: Video name. ARTIST: Filmmaker or other creator name. VERSION: Subtitle, remix info, or other text distinguishing versions of a video. DATE: Date associated with the video. Implementations SHOULD attempt to parse this eld as an ISO 8601 date for machine interpretation and conversion. LOCATION: Location associated with the video. This is usually the lming location for non-ction works. COPYRIGHT: Copyright statement. LICENSE: Copyright and other licensing information. Implementations wishing to do automatic parsing of e.g of distribution terms SHOULD look here for a URL uniquely dening the license. If no instance of this eld is present, or if no instance contains a parseable URL, and implementation MAY look in the COPYRIGHT eld for such a URL. ORGANIZATION: Studio name, Publisher, or other organization involved in the creation of the video. DIRECTOR: Director or Filmmaker credit, similar to ARTIST. PRODUCER: Producer credit for the video. COMPOSER: Music credit for the video. ACTOR: Acting credit for the video.
6.4. SETUP HEADER common header block loop lter table resolution loop lter table scale table resolution AC scale table DC scale table number of base matricies base quatization matricies ... quant range interpolation table DCT token Human tables Figure 6.5: Setup Header structure
51
TAG: subject or category tag, keyword, or other content classication labels. The value of each instance of this eld SHOULD be treated as a single label, with multiple instances of the eld for multiple tags. The value of a single eld SHOULD NOT be parsed into multiple tags based on some internal delimeter. DESCRIPTION: General description, summary, or blurb.
6.4
Setup Header
The Theora setup header contains the limit values used to drive the loop lter, the base matrices and scale values used to build the dequantization tables, and the Human tables used to unpack the DCT tokens. Because the contents of this header are specic to Theora, no concessions have been made to keep the elds octet-aligned for easy parsing.
6.4.1
None.
Loop Filter Limit Table Decode
Input parameters:
Output parameters:
Name LFLIMS
Type Integer array
Size (bits) 7
Signed? No
Description and restrictions A 64-element array of loop lter limit values.
52 Name Type Size (bits)
CHAPTER 6. BITSTREAM HEADERS Signed? Description and restrictions
Variables used:
Name qi NBITS
Type Integer Integer
Size (bits) 6 3
Signed? No No
Description and restrictions The quantization index. The size of values being read in the current table.
This procedure decodes the table of loop lter limit values used to drive the loop lter, which is described in Section 6.4.1. It is decoded as follows: 1. Read a 3-bit unsigned integer as NBITS. 2. For each consecutive value of qi from 0 to 63, inclusive: (a) Read an NBITS-bit unsigned integer as LFLIMS[qi ]. VP3 Compatibility The loop lter limit values are hardcoded in VP3. The values used are given in Appendix B.2.
6.4.2
None.
Quantization Parameters Decode
Input parameters:
Output parameters:
Name ACSCALE
Type Integer array
Size (bits) 16
Signed? No
Description and restrictions A 64-element array of scale values for AC coecients for each qi value.
6.4. SETUP HEADER Name DCSCALE Type Integer array Integer 2D Integer array 2D Integer array Size (bits) 16 Signed? No
53 Description and restrictions A 64-element array of scale values for the DC coecient for each qi value. The number of base matrices. A NBMS 64 array containing the base matrices. A 2 3 array containing the number of quant ranges for a given qti and pli , respectively. This is at most 63. A 2 3 63 array of the sizes of each quant range for a given qti and pli , respectively. Only the rst NQRS[qti ][pli ] values are used. A 2 3 64 array of the bmi s used for each quant range for a given qti and pli , respectively. Only the rst (NQRS[qti ][pli ] + 1) values are used.
NBMS BMS NQRS
10 8 6
No No No
QRSIZES
3D Integer array
No
QRBMIS
3D Integer array
No
Variables used:
Name qti qtj pli plj qi ci bmi qri NBITS NEWQR
Type Integer Integer Integer Integer Integer Integer Integer Integer Integer Integer
Size (bits) 1 1 2 2 6 6 9 6 5 1
Signed? No No No No No No No No No No
Description and restrictions A quantization type index. See Table 3.1. A quantization type index. A color plane index. See Table 2.1. A color plane index. The quantization index. The DCT coecient index. The base matrix index. The quant range index. The size of elds to read. Flag that indicates a new set of quant ranges will be dened.
54 Name RPQR Type Integer Size (bits) 1
CHAPTER 6. BITSTREAM HEADERS Signed? No Description and restrictions Flag that indicates the quant ranges to copy will come from the same color plane.
The AC scale and DC scale values are dened in two simple tables with 64 values each, one for each qi value. The same scale values are used for every quantization type and color plane. The base matrices for all quantization types and color planes are stored in a single table. These are then referenced by index in several sets of quant ranges. The purpose of the quant ranges is to specify which base matrices are used for which qi values. A set of quant ranges is dened for each quantization type and color plane. To save space in the header, bit ags allow a set of quant ranges to be copied from a previously dened set instead of being specied explicitly. Every set except the rst one can be copied from the immediately preceding set. Similarly, if the quantization type is not 0, the set can be copied from the set dened for the same color plane for the preceding quantization type. This formulation allows compact representation of, for example, the same set of quant ranges in both chroma channels, as is done in the original VP3, or the same set of quant ranges in INTRA and INTER modes. Each quant range is dened by a size and two base matrix indices, one for each end of the range. The base matrix for the end of one range is used as the start of the next range, so that for n ranges, n + 1 base matrices are specied. The base matrices for the qi values between the two endpoints of the range are generated by linear interpolation. The location of the endpoints of each range is encoded by their size. The qi value for the left end-point is the sum of the sizes of all preceding ranges, and the qi value for the right end-point adds the size of the current range. Thus the sum of the sizes of all the ranges MUST be 63, so that the last range falls on the last possible qi value. The complete set of quantization parameters are decoded as follows: 1. Read a 4-bit unsigned integer. Assign NBITS the value read, plus one. 2. For each consecutive value of qi from 0 to 63, inclusive: (a) Read an NBITS-bit unsigned integer as ACSCALE[qi ]. 3. Read a 4-bit unsigned integer. Assign NBITS the value read, plus one. 4. For each consecutive value of qi from 0 to 63, inclusive: (a) Read an NBITS-bit unsigned integer as DCSCALE[qi ].
6.4. SETUP HEADER
55
5. Read a 9-bit unsigned integer. Assign NBMS the value decoded, plus one. NBMS MUST be no greater than 384. 6. For each consecutive value of bmi from 0 to (NBMS 1), inclusive: (a) For each consecutive value of ci from 0 to 63, inclusive: i. Read an 8-bit unsigned integer as BMS[bmi ][ci ]. 7. For each consecutive value of qti from 0 to 1, inclusive: (a) For each consecutive value of pli from 0 to 2, inclusive: i. If qti > 0 or pli > 0, read a 1-bit unsigned integer as NEWQR. ii. Else, assign NEWQR the value one. iii. If NEWQR is zero, then we are copying a previously dened set of quant ranges. In that case: A. If qti > 0, read a 1-bit unsigned integer as RPQR. B. Else, assign RPQR the value zero. C. If RPQR is one, assign qtj the value (qti 1) and assign plj the value pli . This selects the set of quant ranges dened for the same color plane as this one, but for the previous quantization type. D. Else assign qtj the value (3 qti + pli 1)//3 and assign plj the value (pli + 2)%3. This selects the most recent set of quant ranges dened. E. Assign NQRS[qti ][pli ] the value NQRS[qtj ][plj ]. F. Assign QRSIZES[qti ][pli ] the values in QRSIZES[qtj ][plj ]. G. Assign QRBMIS[qti ][pli ] the values in QRBMIS[qtj ][plj ]. iv. Else, NEWQR is one, which indicates that we are dening a new set of quant ranges. In that case: A. Assign qri the value zero. B. Assign qi the value zero. C. Read an ilog(NBMS 1)-bit unsigned integer as QRBMIS[qti ][pli ][qri ]. If this is greater than or equal to NBMS, stop. The stream is undecodable. D. Read an ilog(62 qi )-bit unsigned integer. Assign QRSIZES[qti ][pli ][qri ] the value read, plus one. E. Assign qi the value qi + QRSIZES[qti ][pli ][qri ]. F. Assign qri the value qri + 1. G. Read an ilog(NBMS 1)-bit unsigned integer as QRBMIS[qti ][pli ][qri ]. H. If qi is less than 63, go back to step 7(a)ivD. I. If qi is greater than 63, stop. The stream is undecodable. J. Assign NQRS[qti ][pli ] the value qri .
56 VP3 Compatibility
The quantization parameters are hardcoded in VP3. The values used are given in Appendix B.3.
6.4.3
Computing a Quantization Matrix
Input parameters:
Name ACSCALE
Type Integer array Integer array 2D Integer array 2D Integer array
Size (bits) 16
Signed? No
Description and restrictions A 64-element array of scale values for AC coecients for each qi value. A 64-element array of scale values for the DC coecient for each qi value. A NBMS 64 array containing the base matrices. A 2 3 array containing the number of quant ranges for a given qti and pli , respectively. This is at most 63. A 2 3 63 array of the sizes of each quant range for a given qti and pli , respectively. Only the rst NQRS[qti ][pli ] values are used. A 2 3 64 array of the bmi s used for each quant range for a given qti and pli , respectively. Only the rst (NQRS[qti ][pli ] + 1) values are used. A quantization type index. See Table 3.1. A color plane index. See Table 2.1. The quantization index.
DCSCALE
16
No
BMS NQRS
8 6
No No
QRSIZES
3D Integer array
No
QRBMIS
3D Integer array
No
qti pli qi
Integer Integer Integer
1 2 6
No No No
6.4. SETUP HEADER Output parameters:
57
Name QMAT
Type Integer array
Size (bits) 16
Signed? No
Description and restrictions A 64-element array of quantization values for each DCT coecient in natural order.
Variables used:
Name ci bmi bmj qri QISTART QIEND BM QMIN QSCALE
Type Integer Integer Integer Integer Integer Integer Integer array Integer Integer
Size (bits) 6 9 9 6 6 6 8 16 16
Signed? No No No No No No No No No
Description and restrictions The DCT coecient index. The base matrix index. The base matrix index. The quant range index. The left end-point of the qi range. The right end-point of the qi range. A 64-element array containing the interpolated base matrix. The minimum quantization value allowed for the current coecient. The current scale value.
The following procedure can be used to generate a single quantization matrix for a given quantization type, color plane, and qi value, given the quantization parameters decoded in Section 6.4.2. Note that the product of the scale value and the base matrix value is in units of 100ths of a pixel value, and thus is divided by 100 to return it to units of a single pixel value. This value is then scaled by four, to match the scaling of the DCT output, which is also a factor of four larger than the orthonormal version of the transform. 1. Assign qri the index of a quant range such that
qri1
qi
qrj =0
QRSIZES[qti ][pli ][qrj ],
58 and qi
qrj =0
qri
QRSIZES[qti ][pli ][qrj ],
where summation from 0 to 1 is dened to be zero. If there is more than one such value of qri , i.e., if qi lies on the boundary between two quant ranges, then the output will be the same regardless of which one is chosen. 2. Assign QISTART the value
qri1
QRSIZES[qti ][pli ][qrj ].

qrj =0
3. Assign QIEND the value

qri
QRSIZES[qti ][pli ][qrj ].

qrj =0
4. Assign bmi the value QRBMIS[qti ][pli ][qri ]. 5. Assign bmj the value QRBMIS[qti ][pli ][qri + 1]. 6. For each consecutive value of ci from 0 to 63, inclusive: (a) Assign BM[ci ] the value (2 (QIEND qi ) BMS[bmi ][ci ] + 2 (qi QISTART) BMS[bmj ][ci ] + QRSIZES[qti ][pli ][qri ])//(2 QRSIZES[qti ][pli ][qri ]) (b) Assign QMIN the value given by Table 6.18 according to qti and ci . Coecient ci ci ci ci =0 >0 =0 >0 0 0 1 1 qti (Intra) (Intra) (Inter) (Inter) QMIN 16 8 32 16
Table 6.18: Minimum Quantization Values (c) If ci equals zero, assign QSCALE the value DCSCALE[qi ]. (d) Else, assign QSCALE the value ACSCALE[qi ]. (e) Assign QMAT[ci ] the value max(QMIN, min((QSCALE BM[ci ]//100) 4, 4096)).
6.4. SETUP HEADER
59
6.4.4
None.
DCT Token Human Tables
Input parameters:
Output parameters:
Name HTS
Type
Size (bits)
Signed?
Description and restrictions An 80-element array of Human tables with up to 32 entries each.
Human table array
Variables used:
Name HBITS TOKEN ISLEAF
Type Bit string Integer Integer
Size (bits) 32 5 1
Signed? No No No
Description and restrictions A string of up to 32 bits. A single DCT token value. Flag that indicates if the current node of the tree being decoded is a leaf node.
The Human tables used to decode DCT tokens are stored in the setup header in the form of a binary tree. This enforces the requirements that the code be fullso that any sequence of bits will produce a valid sequence of tokensand that the code be prex-free so that there is no ambiguity when decoding. One more restriction is placed on the tables that is not explicitly enforced by the bitstream syntax, but nevertheless must be obeyed by compliant encoders. There must be no more than 32 entries in a single table. Note that this restriction along with the fullness requirement limit the maximum size of a single Human code to 32 bits. It is probably a good idea to enforce this latter consequence explicitly when implementing the decoding procedure as a recursive algorithm, so as to prevent a possible stack overow given an invalid bitstream. Although there are 32 dierent DCT tokens, and thus a normal table will have exactly 32 entries, this is not explicitly required. It is allowable to use a Human code that omits somebut not allof the possible token values. It is also allowable, if not particularly useful, to specify multiple codes for the same
60
token value in a single table. Note also that token values may appear in the tree in any order. In particular, it is not safe to assume that token value zero (which ends a single block), has a Human code of all zeros. The tree is decoded as follows: 1. For each consecutive value of hti from 0 to 79, inclusive: (a) Set HBITS to the empty string. (b) If HBITS is longer than 32 bits in length, stop. The stream is undecodable. (c) Read a 1-bit unsigned integer as ISLEAF. (d) If ISLEAF is one: i. If the number of entries in table HTS[hti ] is already 32, stop. The stream is undecodable. ii. Read a 5-bit unsigned integer as TOKEN. iii. Add the pair (HBITS, TOKEN) to Human table HTS[hti ]. (e) Otherwise: i. Add a 0 to the end of HBITS. ii. Decode the 0 sub-tree using this procedure, starting from step 1b. iii. Remove the 0 from the end of HBITS and add a 1 to the end of HBITS. iv. Decode the 1 sub-tree using this procedure, starting from step 1b. v. Remove the 1 from the end of HBITS. VP3 Compatibility The DCT token Human tables are hardcoded in VP3. The values used are given in Appendix B.4.
6.4.5
None.
Setup Header Decode
Input parameters:
Output parameters:
Name LFLIMS
Type Integer array
Size (bits) 7
Signed? No
Description and restrictions A 64-element array of loop lter limit values.
6.4. SETUP HEADER Name ACSCALE Type Integer array Integer array Integer 2D Integer array 2D Integer array Size (bits) 16 Signed? No
61 Description and restrictions A 64-element array of scale values for AC coecients for each qi value. A 64-element array of scale values for the DC coecient for each qi value. The number of base matrices. A NBMS 64 array containing the base matrices. A 2 3 array containing the number of quant ranges for a given qti and pli , respectively. This is at most 63. A 2 3 63 array of the sizes of each quant range for a given qti and pli , respectively. Only the rst NQRS[qti ][pli ] values will be used. A 2 3 64 array of the bmi s used for each quant range for a given qti and pli , respectively. Only the rst (NQRS[qti ][pli ] + 1) values will be used. An 80-element array of Human tables with up to 32 entries each.
DCSCALE
16
No
NBMS BMS NQRS
10 8 6
No No No
QRSIZES
3D Integer array
No
QRBMIS
3D Integer array
No
HTS
Human table array
Variables used: None. The complete setup header is decoded as follows: 1. Decode the common header elds according to the procedure described in Section 6.1. If HEADERTYPE returned by this procedure is not 0x82, then stop. This packet is not the setup header. 2. Decode the loop lter limit value table using the procedure given in Section 6.4.1 into LFLIMS. 3. Decode the quantization parameters using the procedure given in Section 6.4.2. The results are stored in ACSCALE, DCSCALE, NBMS, BMS, NQRS, QRSIZES, and QRBMIS.
62
CHAPTER 6. BITSTREAM HEADERS 4. Decode the DCT token Human tables using the procedure given in Section 6.4.4 into HTS.
Chapter 7
Frame Decode
This section describes the complete procedure necessary to decode a single frame. This begins with the frame header, followed by coded block ags, macro block modes, motion vectors, block-level qi values, and nally the DCT residual tokens, which are used to reconstruct the frame.
7.1
None.
Frame Header Decode
Input parameters:
Output parameters:
Name FTYPE NQIS QIS
Type Integer Integer Integer array
Size (bits) 1 2 6
Signed? No No No
Description and restrictions The frame type. The number of qi values. An NQIS-element array of qi values.
Variables used:
63
64 Name MOREQIS Type Integer Size (bits) 1 Signed? No
CHAPTER 7. FRAME DECODE Description and restrictions A ag indicating there are more qi values to be decoded.
The frame header selects which type of frame is being decoded, intra or inter, and contains the list of qi values that will be used in this frame. The rst qi value will be used for all DC coecients in all blocks. This is done to ensure that DC prediction, which is done in the quantized domain, works as expected. The AC coecients, however, can be dequantized using any qi value on the list, selected on a block-by-block basis. 1. Read a 1-bit unsigned integer. If the value read is not zero, stop. This is not a data packet. 2. Read a 1-bit unsigned integer as FTYPE. This is the type of frame being decoded, as given in Table 7.3. If this is the rst frame being decoded, this MUST be zero. FTYPE 0 1 Frame Type Intra frame Inter frame
Table 7.3: Frame Type Values 3. Read in a 6-bit unsigned integer as QIS[0]. 4. Read a 1-bit unsigned integer as MOREQIS. 5. If MOREQIS is zero, set NQIS to 1. 6. Otherwise: (a) Read in a 6-bit unsigned integer as QIS[1]. (b) Read a 1-bit unsigned integer as MOREQIS. (c) If MOREQIS is zero, set NQIS to 2. (d) Otherwise: i. Read in a 6-bit unsigned integer as QIS[2]. ii. Set NQIS to 3. 7. If FTYPE is 0, read a 3-bit unsigned integer. These bits are reserved. If this value is not zero, stop. This frame is not decodable according to this specication.
7.2. RUN-LENGTH ENCODED BIT STRINGS VP3 Compatibility
65
The precise format of the frame header is substantially dierent in Theora than in VP3. The original VP3 format includes a larger number of unused, reserved bits that are required to be zero. The original VP3 frame header also can contain only a single qi value, because VP3 does not support block-level qi values and uses the same qi value for all the coecients in a frame.
7.2
Run-Length Encoded Bit Strings
Two variations of run-length encoding are used to store sequences of bits for the block coded ags and the block-level qi values. The procedures to decode these bit sequences are specied in the following two sections.
7.2.1
Long-Run Bit String Decode
Input parameters:
Name NBITS
Type Integer
Size (bits) 36
Signed? No
Description and restrictions The number of bits to decode.
Output parameters:
Name BITS
Type Bit string
Size (bits)
Signed?
Description and restrictions The decoded bits.
Variables used:
Name LEN BIT RLEN
Size (bits) 36 1 13
Signed? No No No
Description and restrictions The number of bits decoded so far. The value associated with the current run. The length of the current run.
66 Name RBITS RSTART ROFFS Type Integer Integer Integer Size (bits) 4 6 12 Signed? No No No
CHAPTER 7. FRAME DECODE Description and restrictions The number of extra bits needed to decode the run length. The start of the possible run-length values for a given Human code. The oset from RSTART of the runlength.
There is no practical limit to the number of consecutive 0s and 1s that can be decoded with this procedure. In reality, the run length is limited by the number of blocks in a single frame, because more will never be requested. A separate procedure described in Section 7.2.2 is used when there is a known limit on the maximum size of the runs. For the rst run, a single bit value is read, and then a Human-coded representation of a run length is decoded, and that many copies of the bit value are appended to the bit string. For each consecutive run, the value of the bit is toggled instead of being read from the bitstream. The only exception is if the length of the previous run was 4129, the maximum possible length encodable by the Human-coded representation. In this case another bit value is read from the stream, to allow for consecutive runs of 0s or 1s longer than this maximum. Note that in both casesfor the rst run and after a run of length 4129if no more bits are needed, then no bit value is read. The complete decoding procedure is as follows: 1. Assign LEN the value 0. 2. Assign BITS the empty string. 3. If LEN equals NBITS, return the completely decoded string BITS. 4. Read a 1-bit unsigned integer as BIT. 5. Read a bit at a time until one of the Human codes given in Table 7.7 is recognized. 6. Assign RSTART and RBITS the values given in Table 7.7 according to the Human code read. 7. Read an RBITS-bit unsigned integer as ROFFS. 8. Assign RLEN the value (RSTART + ROFFS). 9. Append RLEN copies of BIT to BITS. 10. Add RLEN to the value LEN. LEN MUST be less than or equal to NBITS.
7.2. RUN-LENGTH ENCODED BIT STRINGS Human Code b0 b10 b110 b1110 b11110 b111110 b111111 RSTART 1 2 4 6 10 18 34 RBITS 0 1 1 2 3 4 12 Run Lengths 1 2...3 4...5 6...9 10 . . . 17 18 . . . 33 34 . . . 4129
67
Table 7.7: Human Codes for Long Run Lengths 11. If LEN equals NBITS, return the completely decoded string BITS. 12. If RLEN equals 4129, read a 1-bit unsigned integer as BIT. 13. Otherwise, assign BIT the value (1 BIT). 14. Continue decoding runs from step 5. VP3 Compatibility VP3 does not read a new bit value after decoding a run length of 4129. This limits the maximum number of consecutive 0s or 1s to 4129 in VP3-compatible streams. For reasonable video sizes of 1920 1080 or less in 4:2:0 formatthe only pixel format VP3 supportsthis does not pose any problems because runs longer than 4129 are not needed.
7.2.2
Short-Run Bit String Decode
Input parameters:
Name NBITS
Type Integer
Size (bits) 36
Signed? No
Description and restrictions The number of bits to decode.
Output parameters:
Name BITS
Type Bit string
Size (bits)
Signed?
Description and restrictions The decoded bits.
Variables used:
Name
Type
Size (bits)
Signed?
Description and restrictions
68 Name RLEN RBITS RSTART ROFFS Type Integer Integer Integer Integer Size (bits) 13 4 6 12 Signed? No No No No
CHAPTER 7. FRAME DECODE Description and restrictions The length of the current run. The number of extra bits needed to decode the run length. The start of the possible run-length values for a given Human code. The oset from RSTART of the runlength.
This procedure is similar to the procedure outlined in Section 7.2.1, except that the maximum number of consecutive 0s or 1s is limited to 30. This is the maximum run length needed when encoding a bit for each of the 16 blocks in a super block when it is known that not all the bits in a super block are the same. The complete decoding procedure is as follows: 1. Assign LEN the value 0. 2. Assign BITS the empty string. 3. If LEN equals NBITS, return the completely decoded string BITS. 4. Read a 1-bit unsigned integer as BIT. 5. Read a bit at a time until one of the Human codes given in Table 7.11 is recognized. Human Code b0 b10 b110 b1110 b11110 b11111 RSTART 1 3 5 7 11 15 RBITS 1 1 1 2 2 4 Run Lengths 1...2 3...4 5...6 7 . . . 10 11 . . . 14 15 . . . 30
Table 7.11: Human Codes for Short Run Lengths 6. Assign RSTART and RBITS the values given in Table 7.11 according to the Human code read. 7. Read an RBITS-bit unsigned integer as ROFFS. 8. Assign RLEN the value (RSTART + ROFFS). 9. Append RLEN copies of BIT to BITS.
7.3. CODED BLOCK FLAGS DECODE
69
10. Add RLEN to the value LEN. LEN MUST be less than or equal to NBITS. 11. If LEN equals NBITS, return the completely decoded string BITS. 12. Assign BIT the value (1 BIT). 13. Continue decoding runs from step 5.
7.3
Coded Block Flags Decode
Input parameters:
Name FTYPE NSBS NBS
Size (bits) 1 32 36
Signed? No No No
Description and restrictions The frame type. The total number of super blocks in a frame. The total number of blocks in a frame.
Output parameters:
Name BCODED
Type Integer Array
Size (bits) 1
Signed? No
Description and restrictions An NBS-element array of ags indicating which blocks are coded.
Variables used:
Name NBITS BITS
Type Integer Bit string
Size (bits) 36
Signed? No
Description and restrictions The length of a bit string to decode. A decoded set of ags.
70 Name SBPCODED Type Integer Array Size (bits) 1
CHAPTER 7. FRAME DECODE Signed? No Description and restrictions An NSBS-element array of ags indicating whether or not each super block is partially coded. An NSBS-element array of ags indicating whether or not each non-partially coded super block is fully coded. The index of the current super block. The index of the current block in coded order.
SBFCODED
Integer Array
No
sbi bi
Integer Integer
32 36
No No
This procedure determines which blocks are coded in a given frame. In an intra frame, it marks all blocks coded. In an inter frame, however, any or all of the blocks may remain uncoded. The output is a list of bit ags, one for each block, marking it coded or not coded. It is important to note that ags are still decoded for any blocks which lie entirely outside the picture region, even though they are not displayed. Encoders MAY choose to code such blocks. Decoders MUST faithfully reconstruct such blocks, because their contents can be used for predictors in future frames. Flags are not decoded for portions of a super block which lie outside the full frame, as there are no blocks in those regions. The complete procedure is as follows: 1. If FTYPE is zero (intra frame): (a) For each consecutive value of bi from 0 to (NBS1), assign BCODED[bi ] the value one. 2. Otherwise (inter frame): (a) Assign NBITS the value NSBS. (b) Read an NBITS-bit bit string into BITS, using the procedure described in Section 7.2.1. This represents the list of partially coded super blocks. (c) For each consecutive value of sbi from 0 to (NSBS 1), remove the bit at the head of the string BITS and assign it to SBPCODED[sbi ]. (d) Assign NBITS the total number of super blocks such that SBPCODED[sbi ] equals zero. (e) Read an NBITS-bit bit string into BITS, using the procedure described in Section 7.2.1. This represents the list of fully coded super blocks.
7.4. MACRO BLOCK CODING MODES
71
(f) For each consecutive value of sbi from 0 to (NSBS 1) such that SBPCODED[sbi ] equals zero, remove the bit at the head of the string BITS and assign it to SBFCODED[sbi ]. (g) Assign NBITS the number of blocks contained in super blocks where SBPCODED[sbi ] equals one. Note that this might not be equal to 16 times the number of partially coded super blocks, since super blocks which overlap the edge of the frame will have fewer than 16 blocks in them. (h) Read an NBITS-bit bit string into BITS, using the procedure described in Section 7.2.2. (i) For each block in coded orderindexed by bi : i. Assign sbi the index of the super block containing block bi . ii. If SBPCODED[sbi ] is zero, assign BCODED[bi ] the value SBFCODED[sbi ]. iii. Otherwise, remove the bit at the head of the string BITS and assign it to BCODED[bi ].
7.4
Macro Block Coding Modes
Input parameters:
Name FTYPE NMBS NBS BCODED
Type Integer Integer Integer Integer Array
Size (bits) 1 32 36 1
Signed? No No No No
Description and restrictions The frame type. The total number of macro blocks in a frame. The total number of blocks in a frame. An NBS-element array of ags indicating which blocks are coded.
Output parameters:
72 Name Type Size (bits) Size (bits) 3
CHAPTER 7. FRAME DECODE Signed? Description and restrictions
Name MBMODES
Type Integer Array
Signed? No
Description and restrictions An NMBS-element array of coding modes for each macro block.
Variables used:
Name MSCHEME MALPHABET mbi bi mi
Type Integer Integer array Integer Integer Integer
Size (bits) 3 3 32 36 3
Signed? No No No No No
Description and restrictions The mode coding scheme. The list of modes corresponding to each Human code. The index of the current macro block. The index of the current block in coded order. The index of a Human code from Table 7.19, starting from 0.
In an intra frame, every macro block marked as coded in INTRA mode. In an inter frame, however, a macro block can be coded in one of eight coding modes, given in Table 7.18. All of the blocks in all color planes contained in a macro block will be assigned the coding mode of that macro block. An important thing to note is that a coding mode is only stored in the bitstream for a macro block if it has at least one luma block coded. A macro block that contains coded blocks in the chroma planes, but not in the luma plane, MUST be coded in INTER NOMV mode. Thus, no coding mode needs to be decoded for such a macro block. Coding modes are encoded using one of eight dierent schemes. Schemes 0 through 6 use the same simple Human code to represent the mode numbers, as given in Table 7.19. The dierence in the schemes is the mode number assigned to each code. Scheme 0 uses an assignment specied in the bitstream, while schemes 16 use a xed assignment, also given in Table 7.19. Scheme 7 simply codes each mode directly in the bitstream using three bits. 1. If FTYPE is 0 (intra frame):
7.4. MACRO BLOCK CODING MODES
73
Index 0 1 2 3 4 5 6 7
Coding Mode INTER NOMV INTRA INTER MV INTER MV LAST INTER MV LAST2 INTER GOLDEN NOMV INTER GOLDEN MV INTER MV FOUR
Table 7.18: Macro Block Coding Modes
Scheme Human Code b0 b10 b110 b1110 b11110 b111110 b1111110 b1111111
1 3 4 2 0 1 5 6 7
2 3 4 0 2 1 5 6 7
3 3 2 4 0 1 5 6 7
4 3 2 0 4 1 5 6 7
5 0 3 4 2 1 5 6 7
6 0 5 3 4 2 1 6 7
7 mi 0 1 2 3 4 5 6 7
Coding Mode
Table 7.19: Macro Block Mode Schemes
74
CHAPTER 7. FRAME DECODE (a) For each consecutive value of mbi from 0 to (NMBS 1), inclusive, assign MBMODES[mbi ] the value 1 (INTRA). 2. Otherwise (inter frame): (a) Read a 3-bit unsigned integer as MSCHEME. (b) If MSCHEME is 0: i. For each consecutive value of MODE from 0 to 7, inclusive: A. Read a 3-bit unsigned integer as mi . B. Assign MALPHABET[mi ] the value MODE. (c) Otherwise, if MSCHEME is not 7, assign the entries of MALPHABET the values in the corresponding column of Table 7.19. (d) For each consecutive macro block in coded order (cf. Section 2.4) indexed by mbi : i. If a block bi in the luma plane of macro block mbi exists such that BCODED[bi ] is 1: A. If MSCHEME is not 7, read one bit at a time until one of the Human codes in Table 7.19 is recognized, and assign MBMODES[mbi ] the value MALPHABET[mi ], where mi is the index of the Human code decoded. B. Otherwise, read a 3-bit unsigned integer as MBMODES[mbi ]. ii. Otherwise, if no luma-plane blocks in the macro block are coded, assign MBMODES[mbi ] the value 0 (INTER NOMV).
7.5
Motion Vectors
In an intra frame, no motion vectors are used, and so motion vector decoding is skipped. In an inter frame, however, many of the inter coding modes require a motion vector in order to specify an oset into the reference frame from which to predict a block. These procedures assigns such a motion vector to every block.
7.5.1
Motion Vector Decode
Input parameters:
Name MVMODE
Type Integer
Size (bits) 1
Signed? No
Description and restrictions The motion method. vector decoding
7.5. MOTION VECTORS Output parameters:
75
Name MVX MVY
Size (bits) 6 6
Signed? Yes Yes
Description and restrictions The X component of the motion vector. The Y component of the motion vector.
Variables used:
Name MVSIGN
Type Integer
Size (bits) 1
Signed? No
Description and restrictions The sign of the motion vector component just decoded.
The individual components of a motion vector can be coded using one of two methods. The rst uses a variable length Human code, given in Table 7.23. The second encodes the magnitude of the component directly in 5 bits, and the sign in one bit. Note that in this case there are two representations for the value zero. For compatibility with VP3, a sign bit is read even if the magnitude read is zero. One scheme is chosen and used for the entire frame. Each component can take on integer values from 31 . . . 31, inclusive, at half-pixel resolution, i.e. 15.5 . . . 15.5 pixels in the luma plane. For each subsampled axis in the chroma planes, the corresponding motion vector component is interpreted as being at quarter-pixel resolution, i.e. 7.75 . . . 7.75 pixels. The precise details of how these vectors are used to compute predictors for each block are described in Section 7.9.1. A single motion vector is decoded is follows: 1. If MVMODE is 0: (a) Read 1 bit at a time until one of the Human codes in Table 7.23 is recognized, and assign the value to MVX. (b) Read 1 bit at a time until one of the Human codes in Table 7.23 is recognized, and assign the value to MVY. 2. Otherwise: (a) Read a 5-bit unsigned integer as MVX. (b) Read a 1-bit unsigned integer as MVSIGN.
76
CHAPTER 7. FRAME DECODE (c) If MVSIGN is 1, assign MVX the value MVX. (d) Read a 5-bit unsigned integer as MVY. (e) Read a 1-bit unsigned integer as MVSIGN. (f) If MVSIGN is 1, assign MVY the value MVY.
7.5.2
Macro Block Motion Vector Decode
Input parameters:
Name PF NMBS MBMODES
Type Integer Integer Integer Array Integer Integer Array
Size (bits) 2 32 3
Signed? No No No
Description and restrictions The pixel format. The total number of macro blocks in a frame. An NMBS-element array of coding modes for each macro block. The total number of blocks in a frame. An NBS-element array of ags indicating which blocks are coded.
NBS BCODED
36 1
No No
Output parameters:
Name MVECTS
Type Array of 2D Integer Vectors
Size (bits) 6
Signed? Yes
Description and restrictions An NBS-element array of motion vectors for each block.
Variables used:
7.5. MOTION VECTORS Name LAST1 LAST2 MVX MVY mbi A B C D E Type 2D Integer Vector 2D Integer Vector Integer Integer Integer Integer Integer Integer Integer Integer Size (bits) 6 6 6 6 32 36 36 36 36 36 Signed? Yes Yes Yes Yes No No No No No No Description and restrictions The last motion vector.
77
The second to last motion vector. The X component of a motion vector. The Y component of a motion vector. The index of the current macro block. The index of the lower-left luma block in the macro block. The index of the lower-right luma block in the macro block. The index of the upper-left luma block in the macro block. The index of the upper-right luma block in the macro block. The index of a chroma block in the macro block, depending on the pixel format. The index of a chroma block in the macro block, depending on the pixel format. The index of a chroma block in the macro block, depending on the pixel format. The index of a chroma block in the macro block, depending on the pixel format. The index of a chroma block in the macro block, depending on the pixel format. The index of a chroma block in the macro block, depending on the pixel format. The index of a chroma block in the macro block, depending on the pixel format. The index of a chroma block in the macro block, depending on the pixel format.
Integer
36
No
Integer
36
No
Integer
36
No
Integer
36
No
Integer
36
No
Integer
36
No
Integer
36
No
78
CHAPTER 7. FRAME DECODE
Motion vectors are stored for each macro block. In every mode except for INTER MV FOUR, every block in all the color planes are assigned the same motion vector. In INTER MV FOUR mode, all four blocks in the luma plane are assigned their own motion vector, and motion vectors for blocks in the chroma planes are computed from these, using averaging appropriate to the pixel format. For INTER MV and INTER GOLDEN MV modes, a single motion vector is decoded and applied to each block. For INTER MV FOUR macro blocks, a motion vector is decoded for each coded luma block, in raster order, not coded order. Uncoded luma blocks receive the default (0, 0) vector for the purposes of computing the chroma motion vectors. None of the remaining macro block coding modes require decoding motion vectors from the stream. INTRA mode does not use a motion-compensated predictor, and thus does not require a motion vector. Both the INTER NOMV and the INTER GOLDEN NOMV modes use the default vector (0, 0) for each block. This also includes all macro blocks with no coded luma blocks, as they are coded in INTER NOMV mode by denition. The modes INTER MV LAST and INTER MV LAST2 use the motion vector from the last macro block (in coded order) and the second to last macro block, respectively, that contained a motion vector pointing to the previous frame. Thus no explicit motion vector needs to be decoded for these modes. Macro blocks coded in INTRA mode or one of the GOLDEN modes are not considered in this process. For macro blocks coded in INTER MV FOUR mode, this process uses the vector from the last coded luma block in the macro block, again in raster order, not coded order. This is usually the upper-right block. If an insucient number of macro blocks have been coded in one of the INTER modes, then the (0, 0) vector is used instead. The motion vectors are decoded from the stream as follows: 1. Assign LAST1 and LAST2 both the value (0, 0). 2. Read a 1-bit unsigned integer as MVMODE. Note that this value is read even if no macro blocks require a motion vector to be decoded. 3. For each consecutive value of mbi from 0 to (NMBS 1): (a) If MBMODES[mbi ] is 7 (INTER MV FOUR): i. Let A, B, C, and D be the indices in coded order bi of the luma blocks in macro block mbi , arranged into raster order. Thus, A is the index in coded order of the block in the lower left, B the lower right, C the upper left, and D the upper right. ii. If BCODED[A] is non-zero: A. Decode a single motion vector into MVX and MVY using the procedure described in Section 7.5.1. B. Assign MVECTS[A] the value (MVX, MVY). iii. Otherwise, assign MVECTS[A] the value (0, 0).
7.5. MOTION VECTORS iv. If BCODED[B] is non-zero:
79
A. Decode a single motion vector into MVX and MVY using the procedure described in Section 7.5.1. B. Assign MVECTS[B] the value (MVX, MVY). v. Otherwise assign MVECTS[B] the value (0, 0). vi. If BCODED[C] is non-zero: A. Decode a single motion vector into MVX and MVY using the procedure described in Section 7.5.1. B. Assign MVECTS[C] the value (MVX, MVY). vii. Otherwise assign MVECTS[C] the value (0, 0). viii. If BCODED[D] is non-zero: A. Decode a single motion vector into MVX and MVY using the procedure described in Section 7.5.1. B. Assign MVECTS[D] the value (MVX, MVY). ix. Otherwise, assign MVECTS[D] the value (0, 0). x. If PF is 0 (4:2:0): A. Let E and F be the index in coded order of the one block in the macro block from the Cb and Cr planes, respectively. B. Assign MVECTS[E] and MVECTS[F] the value MVECTS[A]x + MVECTS[B]x + (round MVECTS[C]x + MVECTS[D]x , 4 MVECTS[A]y + MVECTS[B]y + round xi. If PF is 2 (4:2:2): A. Let E and F be the indices in coded order of the bottom and top blocks in the macro block from the Cb plane, respectively, and G and H be the indices in coded order of the bottom and top blocks in the Cr plane, respectively. B. Assign MVECTS[E] and MVECTS[G] the value MVECTS[A]x + MVECTS[B]x , 2 MVECTS[A]y + MVECTS[B]y round ) 2 MVECTS[C]y + MVECTS[D]y ) 4
(round
80
CHAPTER 7. FRAME DECODE C. Assign MVECTS[F] and MVECTS[H] the value (round MVECTS[C]x + MVECTS[D]x , 2 MVECTS[C]y + MVECTS[D]y ) round 2
xii. If PF is 3 (4:4:4): A. Let E, F, G, and H be the indices bi in coded order of the Cb plane blocks in macro block mbi , arranged into raster order, and I, J, K, and L be the indices bi in coded order of the Cr plane blocks in macro block mbi , arranged into raster order. B. Assign MVECTS[E] and MVECTS[I] the value MVECTS[A]. C. Assign MVECTS[F] and MVECTS[J] the value MVECTS[B]. D. Assign MVECTS[G] and MVECTS[K] the value MVECTS[C]. E. Assign MVECTS[H] and MVECTS[L] the value MVECTS[D]. xiii. Assign LAST2 the value LAST1. xiv. Assign LAST1 the value (MVX, MVY). This is the value of the motion vector decoded from the last coded luma block in raster order. There must always be at least one, since macro blocks with no coded luma blocks must use mode 0: INTER NOMV. (b) Otherwise, if MBMODES[mbi ] is 6 (INTER GOLDEN MV), decode a single motion vector into MVX and MVY using the procedure described in Section 7.5.1. (c) Otherwise, if MBMODES[mbi ] is 4 (INTER MV LAST2): i. Assign (MVX, MVY) the value LAST2. ii. Assign LAST2 the value LAST1. iii. Assign LAST1 the value (MVX, MVY). (d) Otherwise, if MBMODES[mbi ] is 3 (INTER MV LAST), assign (MVX, MVY) the value LAST1. (e) Otherwise, if MBMODES[mbi ] is 2 (INTER MV): i. Decode a single motion vector into MVX and MVY using the procedure described in Section 7.5.1. ii. Assign LAST2 the value LAST1. iii. Assign LAST1 the value (MVX, MVY). (f) Otherwise (MBMODES[mbi ] is 5: INTER GOLDEN NOMV, 1: INTRA, or 0: INTER NOMV), assign MVX and MVY the value zero. (g) If MBMODES[mbi ] is not 7 (not INTER MV FOUR), then for each coded block bi in macro block mbi : i. Assign MVECTS[bi ] the value (MVX, MVY).
7.6. BLOCK-LEVEL QI DECODE VP3 Compatibility
81
Unless all four luma blocks in the macro block are coded, the VP3 encoder does not select mode INTER MV FOUR. Theora removes this restriction by treating the motion vector for an uncoded luma block as the default (0, 0) vector. This is consistent with the premise that the block has not changed since the previous frame and that chroma information can be largely ignored when estimating motion. No modication is required for INTER MV FOUR macro blocks in VP3 streams to be decoded correctly by a Theora decoder. However, regardless of how many of the luma blocks are actually coded, the VP3 decoder always reads four motion vectors from the stream for INTER MV FOUR mode. The motion vectors read are used to calculate the motion vectors for the chroma blocks, but are otherwise ignored. Thus, care should be taken when creating Theora streams meant to be backwards compatible with VP3 to only use INTER MV FOUR mode when all four luma blocks are coded.
7.6
Block-Level qi Decode
Input parameters:
Name NBS BCODED
Type Integer Integer Array Integer
Size (bits) 36 1
Signed? No No
Description and restrictions The total number of blocks in a frame. An NBS-element array of ags indicating which blocks are coded. The number of qi values.
NQIS
No
Output parameters:
Name QIIS
Type Integer Array
Size (bits) 2
Signed? No
Description and restrictions An NBS-element array of qii values for each block.
82 Variables used:
Name NBITS BITS bi qii
Type Integer Bit string Integer Integer
Size (bits) 36 36 2
Signed? No No No
Description and restrictions The length of a bit string to decode. A decoded set of ags. The index of the current block in coded order. The index of qi value in the list of qi values dened for this frame.
This procedure selects the qi value to be used for dequantizing the AC coecients of each block. DC coecients all use the same qi value, so as to avoid interference with the DC prediction mechanism, which occurs in the quantized domain. The value is actually represented by an index qii into the list of qi values dened for the frame. The decoder makes multiple passes through the list of coded blocks, one for each qi value except the last one. In each pass, an RLEcoded bitmask is decoded to divide the blocks into two groups: those that use the current qi value in the list, and those that use a value from later in the list. Each subsequent pass is restricted to the blocks in the second group. 1. For each value of bi from 0 to (NBS 1), assign QIIS[bi ] the value zero. 2. For each consecutive value of qii from 0 to (NQIS 2): (a) Assign NBITS be the number of blocks bi such that BCODED[bi ] is non-zero and QIIS[bi ] equals qii . (b) Read an NBITS-bit bit string into BITS, using the procedure described in Section 7.2.1. This represents the list of blocks that use qi value qii or higher. (c) For each consecutive value of bi from 0 to (NBS 1) such that BCODED[bi ] is non-zero and QIIS[bi ] equals qii : i. Remove the bit at the head of the string BITS and add its value to QIIS[bi ]. VP3 Compatibility For VP3 compatible streams, only one qi value can be specied in the frame header, so the main loop of the above procedure, which would iterate from 0 to 1, is never executed. Thus, no bits are read, and each block uses the one qi value dened for the frame.
7.6. BLOCK-LEVEL QI DECODE
83
Human Code b000 b001 b0110 b1000 b101000 b101010 b101100 b101110 b1100000 b1100010 b1100100 b1100110 b1101000 b1101010 b1101100 b1101110 b11100000 b11100010 b11100100 b11100110 b11101000 b11101010 b11101100 b11101110 b11110000 b11110010 b11110100 b11110110 b11111000 b11111010 b11111100 b11111110
Value 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31
Human Code b010 b0111 b1001 b101001 b101011 b101101 b101111 b1100001 b1100011 b1100101 b1100111 b1101001 b1101011 b1101101 b1101111 b11100001 b11100011 b11100101 b11100111 b11101001 b11101011 b11101101 b11101111 b11110001 b11110011 b11110101 b11110111 b11111001 b11111011 b11111101 b11111111
Value 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31
Table 7.23: Human Codes for Motion Vector Components
84
7.7. DCT COEFFICIENTS
85
7.7
DCT Coecients
The quantized DCT coecients are decoded by making 64 passes through the list of coded blocks, one for each token index in zig-zag order. For the DC tokens, two Human tables are chosen from among the rst 16, one for the luma plane and one for the chroma planes. The AC tokens, however, are divided into four dierent groups. Again, two 4-bit indices are decoded, one for the luma plane, and one for the chroma planes, but these select the codebooks for all four groups. AC coecients in group one use codebooks 16 . . . 31, while group two uses 32 . . . 47, etc. Note that this second set of indices is decoded even if there are no non-zero AC coecients in the frame. Tokens are divided into two major types: EOB tokens, which ll the remainder of one or more blocks with zeros, and coecient tokens, which ll in one or more coecients within a single block. A decoding procedure for the rst is given in Section 7.7.1, and for the second in Section 7.7.2. The decoding procedure for the complete set of quantized coecients is given in Section 7.7.3.
7.7.1
EOB Token Decode
Input parameters:
Name TOKEN
Type Integer
Size (bits) 5
Signed? No
Description and restrictions The token being decoded. This must be in the range 0 . . . 6. The total number of blocks in a frame. An NBS-element array of the current token index for each block. An NBS-element array of the coecient count for each block. An NBS 64 array of quantized DCT coecient values for each block in zig-zag order. The index of the current block in coded order. The current token index.
NBS TIS
Integer Integer Array Integer Array 2D Integer Array
36 7
No No
NCOEFFS
No
COEFFS
16
Yes
bi ti
Integer Integer
36 6
No No
Name TIS
Type Integer Array 2D Integer Array Integer
Size (bits) 7
Signed? No
Description and restrictions An NBS-element array of the current token index for each block. An NBS64 array of quantized DCT coecient values for each block in zig-zag order. The remaining length of the current EOB run.
COEFFS
16
Yes
EOBS
36
No
Variables used:
Name bj tj
Size (bits) 36 6
Signed? No No
Description and restrictions Another index of a block in coded order. Another token index.
A summary of the EOB tokens is given in Table 7.33. An important thing to note is that token 6 does not add an oset to the decoded run value, even though in general it should only be used for runs of size 32 or longer. If a value of zero is decoded for this run, it is treated as an EOB run the size of the remaining coded blocks. Token Value 0 1 2 3 4 5 6 Extra Bits 0 0 0 2 3 4 12 EOB Run Lengths 1 2 3 4...7 8 . . . 15 16 . . . 31 1 . . . 4095, or all remaining blocks
Table 7.33: EOB Token Summary There is no restriction that one EOB token cannot be immediately followed
7.7. DCT COEFFICIENTS
87
by another, so no special cases are necessary to extend the range of the maximum run length as were required in Section 7.2.1. Indeed, depending on the lengths of the Human codes, it may even cheaper to encode, by way of example, an EOB run of length 31 followed by an EOB run of length 1 than to encode an EOB run of length 32 directly. There is also no restriction that an EOB run stop at the end of a color plane or a token index. The run MUST, however, end at or before the end of the frame. 1. If TOKEN is 0, assign EOBS the value 1. 2. Otherwise, if TOKEN is 1, assign EOBS the value 2. 3. Otherwise, if TOKEN is 2, assign EOBS the value 3. 4. Otherwise, if TOKEN is 3: (a) Read a 2-bit unsigned integer as EOBS. (b) Assign EOBS the value (EOBS + 4). 5. Otherwise, if TOKEN is 4: (a) Read a 3-bit unsigned integer as EOBS. (b) Assign EOBS the value (EOBS + 8). 6. Otherwise, if TOKEN is 5: (a) Read a 4-bit unsigned integer as EOBS. (b) Assign EOBS the value (EOBS + 16). 7. Otherwise, TOKEN is 6: (a) Read a 12-bit unsigned integer as EOBS. (b) If EOBS is zero, assign EOBS to be the number of coded blocks bj such that TIS[bj ] is less than 64. 8. For each value of tj from ti to 63, assign COEFFS[bi ][tj ] the value zero. 9. Assign NCOEFFS[bi ] the value TIS[bi ]. 10. Assign TIS[bi ] the value 64. 11. Assign EOBS the value (EOBS 1). VP3 Compatibility The VP3 encoder does not use the special interpretation of a zero-length EOB run, though its decoder does support it. That may be due more to a happy accident in the way the decoder was written than intentional design, however, and other VP3 implementations might not reproduce it faithfully. For backwards compatibility, it may be wise to avoid it, especially as for most frame sizes there are fewer than 4095 blocks, making it unnecessary.
88
7.7.2
Coecient Token Decode
Input parameters:
Name TOKEN NBS TIS
Type Integer Integer Integer Array 2D Integer Array Integer Integer
Size (bits) 5 36 7
Signed? No No No
Description and restrictions The token being decoded. This must be in the range 7 . . . 31. The total number of blocks in a frame. An NBS-element array of the current token index for each block. An NBS64 array of quantized DCT coecient values for each block in zig-zag order. The index of the current block in coded order. The current token index.
COEFFS
16
Yes
bi ti
36 6
No No
Output parameters:
Name TIS
Type Integer Array Integer Array 2D Integer Array
Size (bits) 7
Signed? No
Description and restrictions An NBS-element array of the current token index for each block. An NBS-element array of the coecient count for each block. An NBS 64 array of quantized DCT coecient values for each block in zig-zag order.
NCOEFFS
No
COEFFS
16
Yes
Variables used:
7.7. DCT COEFFICIENTS Name SIGN MAG RLEN tj Type Integer Integer Integer Integer Size (bits) 1 10 6 6 Signed? No No No No Description and restrictions
89
A ag indicating the sign of the current coecient. The magnitude of the current coecient. The length of the current zero run. Another token index.
Each of these tokens decodes one or more coecients in the current block. A summary of the meanings of the token values is presented in Table 7.38. There are often several dierent ways to tokenize a given coecient list. Which one is optimal depends on the exact lengths of the Human codes used to represent each token. Note that we do not update the coecient count for the block if we decode a pure zero run. For tokens which represent more than one coecient, they MUST NOT bring the total number of coecients in the block to more than 64. Care should be taken in a decoder to check for this, as otherwise it may permit buer overows from invalidly formed packets. Note: One way to achieve this eciently is to combine the inverse zig-zag mapping (described later in Section 7.9.2) with coecient decode, and use a table look-up to map zig-zag indices greater than 63 to a safe location. 1. If TOKEN is 7: (a) Read in a 3-bit unsigned integer as RLEN. (b) Assign RLEN the value (RLEN + 1). (c) For each value of tj from ti to (ti +RLEN1), assign COEFFS[bi ][tj ] the value zero. (d) Assign TIS[bi ] the value TIS[bi ] + RLEN. 2. Otherwise, if TOKEN is 8: (a) Read in a 6-bit unsigned integer as RLEN. (b) Assign RLEN the value (RLEN + 1). (c) For each value of tj from ti to (ti +RLEN1), assign COEFFS[bi ][tj ] the value zero. (d) Assign TIS[bi ] the value TIS[bi ] + RLEN. 3. Otherwise, if TOKEN is 9: (a) Assign COEFFS[bi ][ti ] the value 1.
90
Token Value 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31
Extra Bits 3 6 0 0 0 0 1 1 1 1 2 3 4 5 6 10 1 1 1 1 1 3 4 2 3
Number of Coecients 1...8 1 . . . 64 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 3 4 5 6 7 . . . 10 11 . . . 18 2 3...4
Description Short zero run. Zero run. 1. 1. 2. 2. 3. 4. 5. 6. 7 . . . 8. 9 . . . 12. 13 . . . 20. 21 . . . 36. 37 . . . 68. 69 . . . 580. One zero followed by 1. Two zeros followed by 1. Three zeros followed by 1. Four zeros followed by 1. Five zeros followed by 1. 6 . . . 9 zeros followed by 1. 10 . . . 17 zeros followed by 1. One zero followed by 2 . . . 3. 2 . . . 3 zeros followed by 2 . . . 3.
Table 7.38: Coecient Token Summary
7.7. DCT COEFFICIENTS (b) Assign TIS[bi ] the value TIS[bi ] + 1. (c) Assign NCOEFFS[bi ] the value TIS[bi ]. 4. Otherwise, if TOKEN is 10: (a) Assign COEFFS[bi ][ti ] the value 1. (b) Assign TIS[bi ] the value TIS[bi ] + 1. (c) Assign NCOEFFS[bi ] the value TIS[bi ]. 5. Otherwise, if TOKEN is 11: (a) Assign COEFFS[bi ][ti ] the value 2. (b) Assign TIS[bi ] the value TIS[bi ] + 1. (c) Assign NCOEFFS[bi ] the value TIS[bi ]. 6. Otherwise, if TOKEN is 12: (a) Assign COEFFS[bi ][ti ] the value 2. (b) Assign TIS[bi ] the value TIS[bi ] + 1. (c) Assign NCOEFFS[bi ] the value TIS[bi ]. 7. Otherwise, if TOKEN is 13: (a) Read a 1-bit unsigned integer as SIGN. (b) If SIGN is zero, assign COEFFS[bi ][ti ] the value 3. (c) Otherwise, assign COEFFS[bi ][ti ] the value 3. (d) Assign TIS[bi ] the value TIS[bi ] + 1. (e) Assign NCOEFFS[bi ] the value TIS[bi ]. 8. Otherwise, if TOKEN is 14: (a) Read a 1-bit unsigned integer as SIGN. (b) If SIGN is zero, assign COEFFS[bi ][ti ] the value 4. (c) Otherwise, assign COEFFS[bi ][ti ] the value 4. (d) Assign TIS[bi ] the value TIS[bi ] + 1. (e) Assign NCOEFFS[bi ] the value TIS[bi ]. 9. Otherwise, if TOKEN is 15: (a) Read a 1-bit unsigned integer as SIGN. (b) If SIGN is zero, assign COEFFS[bi ][ti ] the value 5. (c) Otherwise, assign COEFFS[bi ][ti ] the value 5. (d) Assign TIS[bi ] the value TIS[bi ] + 1. (e) Assign NCOEFFS[bi ] the value TIS[bi ].
91
92 10. Otherwise, if TOKEN is 16:
(a) Read a 1-bit unsigned integer as SIGN. (b) If SIGN is zero, assign COEFFS[bi ][ti ] the value 6. (c) Otherwise, assign COEFFS[bi ][ti ] the value 6. (d) Assign TIS[bi ] the value TIS[bi ] + 1. (e) Assign NCOEFFS[bi ] the value TIS[bi ]. 11. Otherwise, if TOKEN is 17: (a) Read a 1-bit unsigned integer as SIGN. (b) Read a 1-bit unsigned integer as MAG. (c) Assign MAG the value (MAG + 7). (d) If SIGN is zero, assign COEFFS[bi ][ti ] the value MAG. (e) Otherwise, assign COEFFS[bi ][ti ] the value MAG. (f) Assign TIS[bi ] the value TIS[bi ] + 1. (g) Assign NCOEFFS[bi ] the value TIS[bi ]. 12. Otherwise, if TOKEN is 18: (a) Read a 1-bit unsigned integer as SIGN. (b) Read a 2-bit unsigned integer as MAG. (c) Assign MAG the value (MAG + 9). (d) If SIGN is zero, assign COEFFS[bi ][ti ] the value MAG. (e) Otherwise, assign COEFFS[bi ][ti ] the value MAG. (f) Assign TIS[bi ] the value TIS[bi ] + 1. (g) Assign NCOEFFS[bi ] the value TIS[bi ]. 13. Otherwise, if TOKEN is 19: (a) Read a 1-bit unsigned integer as SIGN. (b) Read a 3-bit unsigned integer as MAG. (c) Assign MAG the value (MAG + 13). (d) If SIGN is zero, assign COEFFS[bi ][ti ] the value MAG. (e) Otherwise, assign COEFFS[bi ][ti ] the value MAG. (f) Assign TIS[bi ] the value TIS[bi ] + 1. (g) Assign NCOEFFS[bi ] the value TIS[bi ]. 14. Otherwise, if TOKEN is 20: (a) Read a 1-bit unsigned integer as SIGN. (b) Read a 4-bit unsigned integer as MAG.
7.7. DCT COEFFICIENTS (c) Assign MAG the value (MAG + 21). (d) If SIGN is zero, assign COEFFS[bi ][ti ] the value MAG. (e) Otherwise, assign COEFFS[bi ][ti ] the value MAG. (f) Assign TIS[bi ] the value TIS[bi ] + 1. (g) Assign NCOEFFS[bi ] the value TIS[bi ]. 15. Otherwise, if TOKEN is 21: (a) Read a 1-bit unsigned integer as SIGN. (b) Read a 5-bit unsigned integer as MAG. (c) Assign MAG the value (MAG + 37). (d) If SIGN is zero, assign COEFFS[bi ][ti ] the value MAG. (e) Otherwise, assign COEFFS[bi ][ti ] the value MAG. (f) Assign TIS[bi ] the value TIS[bi ] + 1. (g) Assign NCOEFFS[bi ] the value TIS[bi ]. 16. Otherwise, if TOKEN is 22: (a) Read a 1-bit unsigned integer as SIGN. (b) Read a 9-bit unsigned integer as MAG. (c) Assign MAG the value (MAG + 69). (d) If SIGN is zero, assign COEFFS[bi ][ti ] the value MAG. (e) Otherwise, assign COEFFS[bi ][ti ] the value MAG. (f) Assign TIS[bi ] the value TIS[bi ] + 1. (g) Assign NCOEFFS[bi ] the value TIS[bi ]. 17. Otherwise, if TOKEN is 23: (a) Assign COEFFS[bi ][ti ] the value zero. (b) Read a 1-bit unsigned integer as SIGN. (c) If SIGN is zero, assign COEFFS[bi ][ti + 1] the value 1. (d) Otherwise, assign COEFFS[bi ][ti + 1] the value 1. (e) Assign TIS[bi ] the value TIS[bi ] + 2. (f) Assign NCOEFFS[bi ] the value TIS[bi ]. 18. Otherwise, if TOKEN is 24:
93
(a) For each value of tj from ti to (ti + 1), assign COEFFS[bi ][tj ] the value zero. (b) Read a 1-bit unsigned integer as SIGN. (c) If SIGN is zero, assign COEFFS[bi ][ti + 2] the value 1.
94
CHAPTER 7. FRAME DECODE (d) Otherwise, assign COEFFS[bi ][ti + 2] the value 1. (e) Assign TIS[bi ] the value TIS[bi ] + 3. (f) Assign NCOEFFS[bi ] the value TIS[bi ].
19. Otherwise, if TOKEN is 25: (a) For each value of tj from ti to (ti + 2), assign COEFFS[bi ][tj ] the value zero. (b) Read a 1-bit unsigned integer as SIGN. (c) If SIGN is zero, assign COEFFS[bi ][ti + 3] the value 1. (d) Otherwise, assign COEFFS[bi ][ti + 3] the value 1. (e) Assign TIS[bi ] the value TIS[bi ] + 4. (f) Assign NCOEFFS[bi ] the value TIS[bi ]. 20. Otherwise, if TOKEN is 26: (a) For each value of tj from ti to (ti + 3), assign COEFFS[bi ][tj ] the value zero. (b) Read a 1-bit unsigned integer as SIGN. (c) If SIGN is zero, assign COEFFS[bi ][ti + 4] the value 1. (d) Otherwise, assign COEFFS[bi ][ti + 4] the value 1. (e) Assign TIS[bi ] the value TIS[bi ] + 5. (f) Assign NCOEFFS[bi ] the value TIS[bi ]. 21. Otherwise, if TOKEN is 27: (a) For each value of tj from ti to (ti + 4), assign COEFFS[bi ][tj ] the value zero. (b) Read a 1-bit unsigned integer as SIGN. (c) If SIGN is zero, assign COEFFS[bi ][ti + 5] the value 1. (d) Otherwise, assign COEFFS[bi ][ti + 5] the value 1. (e) Assign TIS[bi ] the value TIS[bi ] + 6. (f) Assign NCOEFFS[bi ] the value TIS[bi ]. 22. Otherwise, if TOKEN is 28: (a) Read a 1-bit unsigned integer as SIGN. (b) Read a 2-bit unsigned integer as RLEN. (c) Assign RLEN the value (RLEN + 6). (d) For each value of tj from ti to (ti +RLEN1), assign COEFFS[bi ][tj ] the value zero. (e) If SIGN is zero, assign COEFFS[bi ][ti + RLEN] the value 1.
7.7. DCT COEFFICIENTS (f) Otherwise, assign COEFFS[bi ][ti + RLEN] the value 1. (g) Assign TIS[bi ] the value TIS[bi ] + RLEN + 1. (h) Assign NCOEFFS[bi ] the value TIS[bi ]. 23. Otherwise, if TOKEN is 29: (a) Read a 1-bit unsigned integer as SIGN. (b) Read a 3-bit unsigned integer as RLEN. (c) Assign RLEN the value (RLEN + 10).
95
(d) For each value of tj from ti to (ti +RLEN1), assign COEFFS[bi ][tj ] the value zero. (e) If SIGN is zero, assign COEFFS[bi ][ti + RLEN] the value 1. (f) Otherwise, assign COEFFS[bi ][ti + RLEN] the value 1. (g) Assign TIS[bi ] the value TIS[bi ]+RLEN+1. Assign NCOEFFS[bi ] the value TIS[bi ]. 24. Otherwise, if TOKEN is 30: (a) Assign COEFFS[bi ][ti ] the value zero. (b) Read a 1-bit unsigned integer as SIGN. (c) Read a 1-bit unsigned integer as MAG. (d) Assign MAG the value (MAG + 2). (e) If SIGN is zero, assign COEFFS[bi ][ti + 1] the value MAG. (f) Otherwise, assign COEFFS[bi ][ti + 1] the value MAG. (g) Assign TIS[bi ] the value TIS[bi ] + 2. Assign NCOEFFS[bi ] the value TIS[bi ]. 25. Otherwise, if TOKEN is 31: (a) Read a 1-bit unsigned integer as SIGN. (b) Read a 1-bit unsigned integer as MAG. (c) Assign MAG the value (MAG + 2). (d) Read a 1-bit unsigned integer as RLEN. (e) Assign RLEN the value (RLEN + 2). (f) For each value of tj from ti to (ti +RLEN1), assign COEFFS[bi ][tj ] the value zero. (g) If SIGN is zero, assign COEFFS[bi ][ti + RLEN] the value MAG. (h) Otherwise, assign COEFFS[bi ][ti + RLEN] the value MAG. (i) Assign TIS[bi ] the value TIS[bi ]+RLEN+1. Assign NCOEFFS[bi ] the value TIS[bi ].
96
7.7.3
DCT Coecient Decode
Input parameters:
Name NBS BCODED
Type Integer Integer Array Integer
Size (bits) 36 1
Signed? No No
Description and restrictions The total number of blocks in a frame. An NBS-element array of ags indicating which blocks are coded. The total number of macro blocks in a frame. An 80-element array of Human tables with up to 32 entries each.
NMBS HTS
32
No
Human table array
Output parameters:
Name COEFFS
Type 2D Integer Array
Size (bits) 16
Signed? Yes
Description and restrictions An NBS 64 array of quantized DCT coecient values for each block in zig-zag order. An NBS-element array of the coecient count for each block.
NCOEFFS
Integer Array
No
Variables used:
Name NLBS TIS
Type Integer Integer Array
Size (bits) 34 7
Signed? No No
Description and restrictions The number of blocks in the luma plane. An NBS-element array of the current token index for each block.
7.7. DCT COEFFICIENTS Name EOBS TOKEN HG cbi bi bj ti tj hti L Type Integer Integer Integer Integer Integer Integer Integer Integer Integer Size (bits) 36 5 3 36 36 36 6 6 4 Signed? No No No No No No No No No Description and restrictions
97
hti C
Integer
No
hti
Integer
No
The remaining length of the current EOB run. The current token being decoded. The current Human table group. The index of the current block in the coded block list. The index of the current block in coded order. Another index of a block in coded order. The current token index. Another token index. The index of the current Human table to use for the luma plane within a group. The index of the current Human table to use for the chroma planes within a group. The index of the current Human table to use.
This procedure puts the above two procedures to work to decode the entire set of DCT coecients for the frame. At the end of this procedure, EOBS MUST be zero, and TIS[bi ] MUST be 64 for every coded bi . Note that we update the coecient count of every block before continuing an EOB run or decoding a token, despite the fact that it is already up to date unless the previous token was a pure zero run. This is done intentionally to mimic the VP3 accounting rules. Thus the only time the coecient count does not include the coecients in a pure zero run is when when that run reaches all the way to coecient 63. Note, however, that regardless of the coecient count, any additional coecients are still set to zero. The only use of the count is in determining if a special case of the inverse DCT can be used in Section 7.9.3.2. 1. Assign NLBS the value (NMBS 4). 2. For each consecutive value of bi from 0 to (NBS 1), assign TIS[bi ] the value zero. 3. Assign EOBS the value 0. 4. For each consecutive value of ti from 0 to 63: (a) If ti is 0 or 1: i. Read a 4-bit unsigned integer as hti L .
98
CHAPTER 7. FRAME DECODE ii. Read a 4-bit unsigned integer as hti C . (b) For each consecutive value of bi from 0 to (NBS 1) for which BCODED[bi ] is non-zero and TIS[bi ] equals ti : i. Assign NCOEFFS[bi ] the value ti . ii. If EOBS is greater than zero: A. For each value of tj from ti to 63, assign COEFFS[bi ][tj ] the value zero. B. Assign TIS[bi ] the value 64. C. Assign EOBS the value (EOBS 1). iii. Otherwise: A. Assign HG a value based on ti from Table 7.42. ti 0 1...5 6 . . . 14 15 . . . 27 28 . . . 63 HG 0 1 2 3 4
Table 7.42: Human Table Groups B. If bi is less than NLBS, assign hti the value (16 HG + hti L ). C. Otherwise, assign hti the value (16 HG + hti C ). D. Read one bit at a time until one of the codes in HTS[hti ] is recognized, and assign the value to TOKEN. E. If TOKEN is less than 7, expand an EOB token using the procedure given in Section 7.7.1 to update TIS[bi ], COEFFS[bi ], and EOBS. F. Otherwise, expand a coecient token using the procedure given in Section 7.7.2 to update TIS[bi ], COEFFS[bi ], and NCOEFFS[bi ].
7.8
Undoing DC Prediction
The actual value of a DC coecient decoded by Section 7.7 is the residual from a predicted value computed by the encoder. This prediction is only applied to DC coecients. Quantized AC coecients are encoded directly. This section describes how to undo this prediction to recover the original DC coecients. The predicted DC value for a block is computed from the DC values of its immediate neighbors which precede the block in raster order. Thus, reversing this prediction must procede in raster order, instead of coded order. Note that this step comes before dequantizing the coecients. For this reason, DC coecients are all quantized with the same qi value, regardless of
7.8. UNDOING DC PREDICTION
99
the block-level qi values decoded in Section 7.6. Those qi values are applied only to the AC coecients.
7.8.1
Computing the DC Predictor
Input parameters:
Name BCODED
Type Integer Array Integer Array Integer Array
Size (bits) 1
Signed? No
Description and restrictions An NBS-element array of ags indicating which blocks are coded. An NMBS-element array of coding modes for each macro block. A 3-element array containing the most recently decoded DC value, one for inter mode and for each reference frame. An NBS64 array of quantized DCT coecient values for each block in zig-zag order. The index of the current block in coded order.
MBMODES
No
LASTDC
16
Yes
COEFFS
2D Integer Array
16
Yes
bi
Integer
36
No
Output parameters:
Name DCPRED
Type Integer
Size (bits) 16
Signed? Yes
Description and restrictions The predicted DC value for the current block.
Variables used:
100 Name P Type Integer Array Integer Array Integer Array Integer Integer Integer Integer Integer Size (bits) 1 Signed? No
CHAPTER 7. FRAME DECODE Description and restrictions A 4-element array indicating which neighbors can be used for DC prediction. A 4-element array containing the coded-order block index of the current blocks neighbors. A 4-element array of the weights to apply to each neighboring DC value. The valud to divide the weighted sum by. The index of a neighboring block in coded order. The index of the macro block containing block bi . The index of the macro block containing block bj . The index of the reference frame indicated by the coding mode for macro block mbi .
PBI
36
No
W PDIV bj mbi mbi r
7 8 36 32 32 2
Yes No No No No No
This procedure outlines how a predictor is formed for a single block. The predictor is computed as a weighted sum of the neighboring DC values from coded blocks which use the same reference frame. This latter condition is determined only by checking the coding mode for the block. Even if the golden frame and the previous frame are in fact the same, e.g. for the rst inter frame after an intra frame, they are still treated as being dierent for the purposes of DC prediction. The weighted sum is divided by a power of two, with truncation towards zero, and the result is checked for outranging if necessary. If there are no neighboring coded blocks which use the same reference frame as the current block, then the most recent DC value of any block that used that reference frame is used instead. If no such block exists, then the predictor is set to zero. 1. Assign mbi the index of the macro block containing block bi . 2. Assign r the value of the Reference Frame Index column of Table 7.46 corresponding to MBMODES[mbi ]. 3. If block bi is not along the left edge of the coded frame: (a) Assign bj the coded-order index of block bi s left neighbor, i.e., in the same row but one column to the left. (b) If BCODED[bj ] is not zero:
7.8. UNDOING DC PREDICTION Coding Mode 0 1 2 3 4 5 6 7 (INTER NOMV) (INTRA) (INTER MV) (INTER MV LAST) (INTER MV LAST2) (INTER GOLDEN NOMV) (INTER GOLDEN MV) (INTER MV FOUR) Reference Frame Index 1 0 1 1 1 2 2 1 (Previous) (None) (Previous) (Previous) (Previous) (Golden) (Golden) (Previous)
101
Table 7.46: Reference Frames for Each Coding Mode i. Assign mbj the index of the macro block containing block bj . ii. If the value of the Reference Frame Index column of Table 7.46 corresonding to MBMODES[mbj ] equals r : A. Assign P[0] the value 1. B. Assign PBI[0] the value bj . iii. Otherwise, assign P[0] the value zero. (c) Otherwise, assign P[0] the value zero. 4. Otherwise, assign P[0] the value zero. 5. If block bi is not along the left edge nor the bottom edge of the coded frame: (a) Assign bj the coded-order index of block bi s lower-left neighbor, i.e., one row down and one column to the left. (b) If BCODED[bj ] is not zero: i. Assign mbj the index of the macro block containing block bj . ii. If the value of the Reference Frame Index column of Table 7.46 corresonding to MBMODES[mbj ] equals r : A. Assign P[1] the value 1. B. Assign PBI[1] the value bj . iii. Otherwise, assign P[1] the value zero. (c) Otherwise, assign P[1] the value zero. 6. Otherwise, assign P[1] the value zero. 7. If block bi is not along the the bottom edge of the coded frame: (a) Assign bj the coded-order index of block bi s lower neighbor, i.e., in the same column but one row down. (b) If BCODED[bj ] is not zero:
102
CHAPTER 7. FRAME DECODE i. Assign mbj the index of the macro block containing block bj . ii. If the value of the Reference Frame Index column of Table 7.46 corresonding to MBMODES[mbj ] equals r : A. Assign P[2] the value 1. B. Assign PBI[2] the value bj . iii. Otherwise, assign P[2] the value zero. (c) Otherwise, assign P[2] the value zero.
8. Otherwise, assign P[2] the value zero. 9. If block bi is not along the right edge nor the bottom edge of the coded frame: (a) Assign bj the coded-order index of block bi s lower-right neighbor, i.e., one row down and one column to the right. (b) If BCODED[bj ] is not zero: i. Assign mbj the index of the macro block containing block bj . ii. If the value of the Reference Frame Index column of Table 7.46 corresonding to MBMODES[mbj ] equals r : A. Assign P[3] the value 1. B. Assign PBI[3] the value bj . iii. Otherwise, assign P[3] the value zero. (c) Otherwise, assign P[3] the value zero. 10. Otherwise, assign P[3] the value zero. 11. If none of the values P[0], P[1], P[2], nor P[3] are non-zero, then assign DCPRED the value LASTDC[r ]. 12. Otherwise: (a) Assign the array W and the variable PDIV the values from the row of Table 7.47 corresonding to the values of each P[i ]. (b) Assign DCPRED the value zero. (c) If P[0] is non-zero, assign DCPRED the value (DCPRED + W[0] COEFFS[PBI[0]][0]). (d) If P[1] is non-zero, assign DCPRED the value (DCPRED + W[1] COEFFS[PBI[1]][0]). (e) If P[2] is non-zero, assign DCPRED the value (DCPRED + W[2] COEFFS[PBI[2]][0]). (f) If P[3] is non-zero, assign DCPRED the value (DCPRED + W[3] COEFFS[PBI[3]][0]). (g) Assign DCPRED the value (DCPRED//PDIV).
7.8. UNDOING DC PREDICTION P[0] (L) 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 P[1] (DL) 0 1 1 0 0 1 1 0 0 1 1 0 0 1 1 P[2] (D) 0 0 0 1 1 1 1 0 0 0 0 1 1 1 1 P[3] (DR) 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 W[3] (L) 1 0 1 0 1 0 29 0 75 0 75 0 75 0 29 W[1] (DL) 0 1 0 0 0 0 26 0 0 1 0 0 0 3 26 W[2] (D) 0 0 0 1 1 1 29 0 0 0 0 1 0 10 29 W[3] (DR) 0 0 0 0 0 0 0 1 53 1 53 0 53 3 0
103 PDIV 1 1 1 1 2 1 32 1 128 2 128 1 128 16 32
Table 7.47: Weights and Divisors for Each Set of Available DC Predictors (h) If P[0], P[1], and P[2] are all non-zero: i. If |DCPREDCOEFFS[PBI[2]][0]| is greater than 128, assign DCPRED the value COEFFS[PBI[2]][0]. ii. Otherwise, if |DCPRED COEFFS[PBI[0]][0]| is greater than 128, assign DCPRED the value COEFFS[PBI[0]][0]. iii. Otherwise, if |DCPRED COEFFS[PBI[1]][0]| is greater than 128, assign DCPRED the value COEFFS[PBI[1]][0].
7.8.2
Inverting the DC Prediction Process
Input parameters:
Name BCODED
Type Integer Array Integer Array
Size (bits) 1
Signed? No
Description and restrictions An NBS-element array of ags indicating which blocks are coded. An NMBS-element array of coding modes for each macro block.
MBMODES
No
104 Name COEFFS Type 2D Integer Array Size (bits) 16
CHAPTER 7. FRAME DECODE Signed? Yes Description and restrictions An NBS64 array of quantized DCT coecient values for each block in zig-zag order.
Output parameters:
Name COEFFS
Size (bits) 16
Signed? Yes
Description and restrictions An NBS64 array of quantized DCT coecient values for each block in zig-zag order. The DC value of each block will be updated.
Variables used:
Name LASTDC
Type Integer Array
Size (bits) 16
Signed? Yes
Description and restrictions A 3-element array containing the most recently decoded DC value, one for inter mode and for each reference frame. The predicted DC value for the current block. The actual DC value for the current block. The index of the current block in coded order. The index of the macro block containing block bi . The index of the reference frame indicated by the coding mode for macro block mbi . A color plane index.
DCPRED DC bi mbi r
Integer Integer Integer Integer Integer
11 17 36 32 2
Yes Yes No No No
pli
Integer
No
7.9. RECONSTRUCTION
105
This procedure describes the complete process of undoing the DC prediction to recover the original DC values. Because it is possible to add a value as large as 580 to the predicted DC coecient value at every block, which will then be used to increase the predictor for the next block, the reconstructed DC value could overow a 16-bit integer. This is handled by truncating the result to a 16-bit signed representation, simply throwing away any higher bits in the twos complement representation of the number. 1. For each consecutive value of pli from 0 to 2: (a) Assign LASTDC[0] the value zero. (b) Assign LASTDC[1] the value zero. (c) Assign LASTDC[2] the value zero. (d) For each block of color plane pli in raster order, with coded-order index bi : i. If BCODED[bi ] is non-zero: A. Compute the value DCPRED using the procedure outlined in Section 7.8.1. B. Assign DC the value (COEFFS[bi ][0] + DCPRED). C. Truncate DC to a 16-bit representation by dropping any higher-order bits. D. Assign COEFFS[bi ][0] the value DC. E. Assign mbi the index of the macro block containing block bi . F. Assign r the value of the Reference Frame Index column of Table 7.46 corresponding to MBMODES[mbi ]. G. Assign LASTDC[r ] the value DC.
7.9
Reconstruction
At this stage, the complete contents of the data packet have been decoded. All that remains is to reconstruct the contents of the new frame. This is applied on a block by block basis, and as each block is independent, the order they are processed in does not matter.
7.9.1
Predictors
For each block, a predictor is formed based on its coding mode and motion vector. There are three basic types of predictors: the intra predictor, the wholepixel predictor, and the half-pixel predictor. The former is used for all blocks coded in INTRA mode, while all other blocks use one of the latter two. The whole-pixel predictor is used if the fractional part of both motion vector components is zero, otherwise the half-pixel predictor is used.
106 7.9.1.1 The Intra Predictor
Input parameters: None. Output parameters:
Name PRED
Size (bits) 8
Signed? No
Description and restrictions An 8 8 array of predictor values to use for INTRA coded blocks.
Variables used:
Name bx by
Size (bits) 3 3
Signed? No No
Description and restrictions The horizontal pixel index in the block. The vertical pixel index in the block.
The intra predictor is nothing more than the constant value 128. This is applied for the sole purpose of centering the range of possible DC values for INTRA blocks around zero. 1. For each value of by from 0 to 7, inclusive: (a) For each value of bx from 0 to 7, inclusive: i. Assign PRED[by][bx ] the value 128. 7.9.1.2 The Whole-Pixel Predictor
Input parameters:
Name RPW
Type Integer
Size (bits) 20
Signed? No
Description and restrictions The width of the current plane of the reference frame in pixels.
7.9. RECONSTRUCTION Name RPH REFP Type Integer 2D Integer Array Integer Size (bits) 20 8 Signed? No No Description and restrictions
107
BX
20
No
BY
Integer
20
No
MVX
Integer
No
MVY
Integer
No
The height of the current plane of the reference frame in pixels. A RPH RPW array containing the contents of the current plane of the reference frame. The horizontal pixel index of the lower-left corner of the current block. The vertical pixel index of the lower-left corner of the current block. The horizontal component of the block motion vector. This is always a whole-pixel value. The vertical component of the block motion vector. This is always a whole-pixel value.
Output parameters:
Name PRED
Size (bits) 8
Signed? No
Description and restrictions An 8 8 array of predictor values to use for INTER coded blocks.
Variables used:
Name bx by rx ry
Type Integer Integer Integer Integer
Size (bits) 3 3 20 20
Signed? Yes Yes No No
Description and restrictions The horizontal pixel index in the block. The vertical pixel index in the block. The horizontal pixel index in the reference frame. The vertical pixel index in the reference frame.
108
The whole pixel predictor simply copies verbatim the contents of the reference frame pointed to by the blocks motion vector. If the vector points outside the reference frame, then the closest value on the edge of the reference frame is used instead. In practice, this is usually implemented by expanding the size of the reference frame by 8 or 16 pixels on each sidedepending on whether or not the corresponding axis is subsampled in the current planeand copying the border pixels into this region. 1. For each value of by from 0 to 7, inclusive: (a) Assign ry the value (BY + MVY + by). (b) If ry is greater than (RPH 1), assign ry the value (RPH 1). (c) If ry is less than zero, assign ry the value zero. (d) For each value of bx from 0 to 7, inclusive: i. ii. iii. iv. 7.9.1.3 Assign rx the value (BX + MVX + bx ). If rx is greater than (RPW1), assign rx the value (RPW1). If rx is less than zero, assign rx the value zero. Assign PRED[by][bx ] the value REFP[ry][rx ].
The Half-Pixel Predictor
Input parameters:
Name RPW RPH REFP
Type Integer Integer 2D Integer Array Integer
Size (bits) 20 20 8
Signed? No No No
Description and restrictions The width of the current plane of the reference frame in pixels. The height of the current plane of the reference frame in pixels. A RPH RPW array containing the contents of the current plane of the reference frame. The horizontal pixel index of the lower-left corner of the current block. The vertical pixel index of the lower-left corner of the current block. The horizontal component of the rst whole-pixel motion vector. The vertical component of the rst whole-pixel motion vector.
BX
20
No
BY
Integer
20
No
MVX MVY
Integer Integer
5 5
No No
7.9. RECONSTRUCTION Name MVX2 MVY2 Type Integer Integer Size (bits) 5 5 Signed? No No Description and restrictions
109
The horizontal component of the second whole-pixel motion vector. The vertical component of the second whole-pixel motion vector.
Output parameters:
Name PRED
Size (bits) 8
Signed? No
Description and restrictions An 8 8 array of predictor values to use for INTER coded blocks.
Variables used:
Name bx by rx1 ry1 rx2 ry2
Type Integer Integer Integer Integer Integer Integer
Size (bits) 3 3 20 20 20 20
Signed? Yes Yes No No No No
Description and restrictions The horizontal pixel index in the block. The vertical pixel index in the block. The rst horizontal pixel index in the reference frame. The rst vertical pixel index in the reference frame. The second horizontal pixel index in the reference frame. The second vertical pixel index in the reference frame.
If one or both of the components of the block motion vector is not a wholepixel value, then the half-pixel predictor is used. The half-pixel predictor converts the fractional motion vector into two whole-pixel motion vectors. The rst is formed by truncating the values of each component towards zero, and the second is formed by truncating them away from zero. The contributions from the reference frame at the locations pointed to by each vector are averaged, truncating towards negative innity. Only two samples from the reference frame contribute to each predictor
110
value, even if both components of the motion vector have non-zero fractional components. Motion vector components with quarter-pixel accuracy in the chroma planes are treated exactly the same as those with half-pixel accuracy. Any non-zero fractional part gets rounded one way in the rst vector, and the other way in the second. 1. For each value of by from 0 to 7, inclusive: (a) Assign ry1 the value (BY + MVY1 + by). (b) If ry1 is greater than (RPH 1), assign ry1 the value (RPH 1). (c) If ry1 is less than zero, assign ry1 the value zero. (d) Assign ry2 the value (BY + MVY2 + by). (e) If ry2 is greater than (RPH 1), assign ry2 the value (RPH 1). (f) If ry2 is less than zero, assign ry2 the value zero. (g) For each value of bx from 0 to 7, inclusive: i. Assign rx1 the value (BX + MVX1 + bx ). ii. If rx1 is greater than (RPW 1), assign rx1 1). iii. If rx1 is less than zero, assign rx1 the value iv. Assign rx2 the value (BX + MVX2 + bx ). v. If rx2 is greater than (RPW 1), assign rx2 1). vi. If rx2 is less than zero, assign rx2 the value vii. Assign PRED[by][bx ] the value the value (RPW zero. the value (RPW zero.
(REFP[ry1 ][rx1 ] + REFP[ry2 ][rx2 ]) >> 1.
7.9.2
Dequantization
Input parameters:
Name COEFFS
Size (bits) 16
Signed? Yes
Description and restrictions An NBS 64 array of quantized DCT coecient values for each block in zig-zag order. A 64-element array of scale values for AC coecients for each qi value.
ACSCALE
Integer array
16
No
7.9. RECONSTRUCTION Name DCSCALE Type Integer array 2D Integer array 2D Integer array Size (bits) 16 Signed? No
111 Description and restrictions A 64-element array of scale values for the DC coecient for each qi value. A NBMS 64 array containing the base matrices. A 2 3 array containing the number of quant ranges for a given qti and pli , respectively. This is at most 63. A 2 3 63 array of the sizes of each quant range for a given qti and pli , respectively. Only the rst NQRS[qti ][pli ] values are used. A 2 3 64 array of the bmi s used for each quant range for a given qti and pli , respectively. Only the rst (NQRS[qti ][pli ] + 1) values are used. A quantization type index. See Table 3.1. A color plane index. See Table 2.1. The quantization index of the DC coecient. The quantization index of the AC coecients. The index of the current block in coded order.
BMS NQRS
8 6
No No
QRSIZES
3D Integer array
No
QRBMIS
3D Integer array
No
qti pli qi0 qi bi
1 2 6 6 36
No No No No No
Output parameters:
Name DQC
Type Integer Array
Size (bits) 14
Signed? Yes
Description and restrictions A 64-element array of dequantized DCT coecients in natural order (cf. Section 2.6).
112 Variables used:
Name QMAT
Type Integer array Integer Integer Integer
Size (bits) 16
Signed? No
Description and restrictions A 64-element array of quantization values for each DCT coecient in natural order. The DCT coecient index in natural order. The DCT coecient index in zig-zag order. A single dequantized coecient.
ci zzi C
6 6 29
No No Yes
This procedure takes the quantized DCT coecient values in zig-zag order for a single blockafter DC prediction has been undoneand returns the dequantized values in natural order. If large coecient values are decoded for coarsely quantized coecients, the resulting dequantized value can be signicantly larger than 16 bits. Such a coecient is truncated to a signed 16-bit representation by discarding the higher-order bits of its twos-complement representation. Although this procedure recomputes the quantization matrices from the parameters in the setup header for each block, there are at most six dierent ones used for each color plane. An ecient implementation could compute them once in advance. 1. Using ACSCALE, DCSCALE, BMS, NQRS, QRSIZES, QRBMIS, qti , pli , and qi0 , use the procedure given in Section 6.4.3 to compute the DC quantization matrix QMAT. 2. Assign C the value COEFFS[bi ][0] QMAT[0]. 3. Truncate C to a 16-bit representation by dropping any higher-order bits. 4. Assign DQC[0] the value C. 5. Using ACSCALE, DCSCALE, BMS, NQRS, QRSIZES, QRBMIS, qti , pli , and qi , use the procedure given in Section 6.4.3 to compute the AC quantization matrix QMAT. 6. For each value of ci from 1 to 63, inclusive: (a) Assign zzi the index in zig-zag order corresponding to ci . E.g., the value at row (ci //8) and column (ci %8) in Figure 2.8 (b) Assign C the value COEFFS[bi ][zzi ] QMAT[ci ].
7.9. RECONSTRUCTION
113
(c) Truncate C to a 16-bit representation by dropping any higher-order bits. (d) Assign DQC[ci ] the value C.
7.9.3
The Inverse DCT
The 2D inverse DCT is separated into two applications of the 1D inverse DCT. The transform is rst applied to each row, and then applied to each column of the result. Each application of the 1D inverse DCT scales the values by a factor of two relative to the orthonormal version of the transform, for a total scale factor of four for the 2D transform. It is assumed that a similar scale factor is applied during the forward DCT used in the encoder, so that a division by 16 is required after the transform has been applied in both directions. The inclusion of this scale factor allows the integerized transform to operate with increased precision. All divisions throughout the transform are implemented with right shifts. Only the nal division by 16 is rounded, with ties rounded towards positive innity. All intermediate values are truncated to a 32-bit signed representation by discarding any higher-order bits in their twos complement representation. The nal output of each 1D transform is truncated to a 16-bit signed value in the same manner. In practice, if the high word of a 16 16 bit multiplication can be obtained directly, 16 bits is sucient for every calculation except scaling by C4. Thus we truncate to 16 bits before that multiplication to allow an implementation entirely in 16-bit registers. Implementations using larger registers must sign-extend the 16-bit value to maintain compatibility. Note that if 16-bit register are used, overow in the additions and subtractions should be handled using unsaturated arithmetic. That is, the high-order bits should be discarded and the low-order bits retained, instead of clamping the result to the maximum or minimum value. This allows the maximum exibility in re-ordering these instructions without deviating from this specication. The 1D transform can only overow if input coecients larger than 6201 are present. However, the result of applying the 2D forward transform on pixel values in the range 255 . . . 255 can be as large as 8157 due to the scale factor of four that is applied, and quantization errors could make this even larger. Therefore, the coecients cannot simply be clamped into a valid range before the transform. 7.9.3.1 The 1D Inverse DCT
Input parameters:
114 Name Y Type Integer Array Size (bits) 16 Signed? Yes
CHAPTER 7. FRAME DECODE Description and restrictions An 8-element array of DCT coecients.
Output parameters:
Name X
Type Integer Array
Size (bits) 16
Signed? Yes
Description and restrictions An 8-element array of output values.
Variables used:
Name T R
Type Integer Array Integer
Size (bits) 32 32
Signed? Yes Yes
Description and restrictions An 8-element array containing the current value of each signal line. A temporary value.
A compliant decoder MUST use the exact implementation of the inverse DCT dened in this specication. Some operations may be re-ordered, but the result must be precisely equivalent. This is a design decision that limits some avenues of decoder optimization, but prevents any drift in the prediction loop. The full inverse DCT described in this section is only used when there is at least one non-zero AC coecient. Otherwise, a special DC-only case is used, which is described below in step 2(d)vii of Section 7.9.4. The DC-only special case is not exactly equivalent to applying the full inverse transform, because it avoids the intermediate multiplications and truncation, and it MUST be used in place of the full transform to ensure bit-exact decode. Theora uses a 16-bit integerized approximation of of the 8-point 1D inverse DCT based on the Chen factorization [CSF77]. It requires 16 multiplications and 26 additions and subtractions. A signal ow graph of the transformation is presented in Figure 7.1. This graph provides a good visualization of which parts of the transform are parallelizable. Time increases from left to right. Each signal line is involved in an operation where the line is marked with a
7.9. RECONSTRUCTION
C4 0 C4 4 C6 S6 C6 S6 C7 S7 5 C3 S3 C3 S3 S7 7 C7 C4 C4
115
0
Figure 7.1: Signal Flow Graph for the 1D Inverse DCT dot or a circled plus sign . The constants Ci and Sj are the 16-bit integer i approximations of cos( 16 ) and sin( j ) listed in Table 7.65. When they appear 16 next to a signal line, the value on that line is scaled by the given constant. A circled minus sign next to a signal line indicates that the value on that line is negated. Operations on a single signal path through the graph cannot be reordered, but operations on dierent paths may be, or may be executed in parallel. Different graphs may be obtainable using the associative, commutative, and distributive properties of unsaturated arithmetic. The column of numbers on the left represents an initial permutation of the input DCT coecients. The column on the right represents the unpermuted output. One can be obtained by bit-reversing the 3-bit binary representation of the other. Ci C1 C2 C3 C4 C5 C6 C7 Sj S7 S6 S5 S4 S3 S2 S1 Value 64277 60547 54491 46341 36410 25080 12785
Table 7.65: 16-bit Approximations of Sines and Cosines
116 1. Assign T[0] the value Y[0] + Y[4].
2. Truncate T[0] to a 16-bit signed representation by dropping any higherorder bits. 3. Assign T[0] the value C4 T[0] >> 16. 4. Assign T[1] the value Y[0] Y[4]. 5. Truncate T[1] to a 16-bit signed representation by dropping any higherorder bits. 6. Assign T[1] the value C4 T[1] >> 16. 7. Assign T[2] the value (C6 Y[2] >> 16) (S6 Y[6] >> 16). 8. Assign T[3] the value (S6 Y[2] >> 16) + (C6 Y[6] >> 16). 9. Assign T[4] the value (C7 Y[1] >> 16) (S7 Y[7] >> 16). 10. Assign T[5] the value (C3 Y[5] >> 16) (S3 Y[3] >> 16). 11. Assign T[6] the value (S3 Y[5] >> 16) + (C3 Y[3] >> 16). 12. Assign T[7] the value (S7 Y[1] >> 16) + (C7 Y[7] >> 16). 13. Assign R the value T[4] + T[5]. 14. Assign T[5] the value T[4] T[5]. 15. Truncate T[5] to a 16-bit signed representation by dropping any higherorder bits. 16. Assign T[5] the value C4 T[5] >> 16. 17. Assign T[4] the value R. 18. Assign R the value T[7] + T[6]. 19. Assign T[6] the value T[7] T[6]. 20. Truncate T[6] to a 16-bit signed representation by dropping any higherorder bits. 21. Assign T[6] the value C4 T[6] >> 16. 22. Assign T[7] the value R. 23. Assign R the value T[0] + T[3]. 24. Assign T[3] the value T[0] T[3]. 25. Assign T[0] the value R. 26. Assign R the value T[1] + T[2]
7.9. RECONSTRUCTION 27. Assign T[2] the value T[1] T[2] 28. Assign T[1] the value R. 29. Assign R the value T[6] + T[5]. 30. Assign T[5] the value T[6] T[5]. 31. Assign T[6] the value R. 32. Assign R the value T[0] + T[7].
117
33. Truncate R to a 16-bit signed representation by dropping any higher-order bits. 34. Assign X[0] the value R. 35. Assign R the value T[1] + T[6]. 36. Truncate R to a 16-bit signed representation by dropping any higher-order bits. 37. Assign X[1] the value R. 38. Assign R the value T[2] + T[5]. 39. Truncate R to a 16-bit signed representation by dropping any higher-order bits. 40. Assign X[2] the value R. 41. Assign R the value T[3] + T[4]. 42. Truncate R to a 16-bit signed representation by dropping any higher-order bits. 43. Assign X[3] the value R. 44. Assign R the value T[3] T[4]. 45. Truncate R to a 16-bit signed representation by dropping any higher-order bits. 46. Assign X[4] the value R. 47. Assign R the value T[2] T[5]. 48. Truncate R to a 16-bit signed representation by dropping any higher-order bits. 49. Assign X[5] the value R. 50. Assign R the value T[1] T[6].
118
51. Truncate R to a 16-bit signed representation by dropping any higher-order bits. 52. Assign X[6] the value R. 53. Assign R the value T[0] T[7]. 54. Truncate R to a 16-bit signed representation by dropping any higher-order bits. 55. Assign X[7] the value R. 7.9.3.2 The 2D Inverse DCT
Input parameters:
Name DQC
Type Integer Array
Size (bits) 14
Signed? Yes
Description and restrictions A 64-element array of dequantized DCT coecients in natural order (cf. Section 2.6).
Output parameters:
Name RES
Size (bits) 16
Signed? Yes
Description and restrictions An 8 8 array containing the decoded residual for the current block.
Variables used:
Name ci ri Y
Type Integer Integer Integer Array
Size (bits) 3 3 16
Signed? No No Yes
Description and restrictions The column index. The row index. An 8-element array of 1D iDCT input values.
7.9. RECONSTRUCTION Name X Type Integer Array Size (bits) 16 Signed? Yes Description and restrictions
119
An 8-element array of 1D iDCT output values.
This procedure applies the 1D inverse DCT transform 16 times to a block of dequantized coecients: once for each of the 8 rows, and once for each of the 8 columns of the result. Note that the coordinate system used for the columns is the same right-handed coordinate system used by the rest of Theora. Thus, the column is indexed from bottom to top, not top to bottom. The nal values are divided by sixteen, rounding with ties rounded towards postive innity. 1. For each value of ri from 0 to 7: (a) For each value of ci from 0 to 7: i. Assign Y[ci ] the value DQC[ri 8 + ci ]. (b) Compute X, the 1D inverse DCT of Y using the procedure described in Section 7.9.3.1. (c) For each value of ci from 0 to 7: i. Assign RES[ri ][ci ] the value X[ci ]. 2. For each value of ci from 0 to 7: (a) For each value of ri from 0 to 7: i. Assign Y[ri ] the value RES[ri ][ci ]. (b) Compute X, the 1D inverse DCT of Y using the procedure described in Section 7.9.3.1. (c) For each value of ri from 0 to 7: i. Assign RES[ri ][ci ] the value (X[ri ] + 8) >> 4. 7.9.3.3 The 1D Forward DCT (Non-Normative)
Input parameters:
Name X
Type Integer Array
Size (bits) 14
Signed? Yes
Description and restrictions An 8-element array of input values.
Name Y
Type Integer Array
Size (bits) 16
Signed? Yes
Description and restrictions An 8-element array of DCT coecients.
Variables used:
Name T R
Type Integer Array Integer
Size (bits) 16 16
Signed? Yes Yes
Description and restrictions An 8-element array containing the current value of each signal line. A temporary value.
The forward transform used in the encoder is not mandated by this standard as the inverse one is. Precise equivalence in the inverse transform alone is all that is required to guarantee that there is no mismatch in the prediction loop between encoder and any compliant decoder implementation. However, a forward transform is provided here as a convenience for implementing an encoder. This is the version of the transform used by Xiph.Orgs Theora encoder, which is the same as that used by VP3. Like the inverse DCT, it is rst applied to each row, and then applied to each column of the result. The signal ow graph for the forward transform is given in Figure 7.2. It is largely the reverse of the ow graph given for the inverse DCT. It is important to note that the signs on the constants in the rotations have changed, and the C4 scale factors on one of the lower butteries now appear on the opposite side. The column of numbers on the left represents the unpermuted input, and the column on the right the permuted output DCT coecients. A proper division by 216 is done after the multiplications instead of a shift in the forward transform. This can be implemented quickly by adding an oset of 0xFFFF if the number is negative, and then shifting as before. This slightly increases the computational complexity of the transform. Unlike the inverse DCT, 16-bit registers and a 16 16 32 bit multiply are sucient to avoid any overow, so long as the input is in the range 6270 . . . 6270, which is larger than required. 1. Assign T[0] the value X[0] + X[7].
7.9. RECONSTRUCTION
C4 0 C4 1 C6 S6 C6 S6
121
0
C7 S7 C4 C4 S7 C3 S3 C3 S3
C7
Figure 7.2: Signal Flow Graph for the 1D Forward DCT 2. Assign T[1] the value X[1] + X[6]. 3. Assign T[2] the value X[2] + X[5]. 4. Assign T[3] the value X[3] + X[4]. 5. Assign T[4] the value X[3] X[4]. 6. Assign T[5] the value X[2] X[5]. 7. Assign T[6] the value X[1] X[6]. 8. Assign T[7] the value X[0] X[7]. 9. Assign R the value T[0] + T[3]. 10. Assign T[3] the value T[0] T[3]. 11. Assign T[0] the value R. 12. Assign R the value T[1] + T[2]. 13. Assign T[2] the value T[1] T[2]. 14. Assign T[1] the value R. 15. Assign R the value T[6] T[5]. 16. Assign T[6] the value (C4 (T[6] + T[5]))//16. 17. Assign T[5] the value (C4 R)//16.
122 18. Assign R the value T[4] + T[5]. 19. Assign T[5] the value T[4] T[5]. 20. Assign T[4] the value R. 21. Assign R the value T[7] + T[6]. 22. Assign T[6] the value T[7] T[6]. 23. Assign T[7] the value R.
24. Assign Y[0] the value (C4 (T[0] + T[1]))//16. 25. Assign Y[4] the value (C4 (T[0] T[1]))//16. 26. Assign Y[2] the value ((S6 T[3])//16) + ((C6 T[2])//16). 27. Assign Y[6] the value ((C6 T[3])//16) ((S6 T[2])//16). 28. Assign Y[1] the value ((S7 T[7])//16) + ((C7 T[4])//16). 29. Assign Y[5] the value ((S3 T[6])//16) + ((C3 T[5])//16). 30. Assign Y[3] the value ((C3 T[6])//16) ((S3 T[5])//16). 31. Assign Y[7] the value ((C7 T[7])//16) ((S7 T[4])//16).
7.9.4
The Complete Reconstruction Algorithm
Input parameters:
Name ACSCALE
Type Integer array Integer array
Size (bits) 16
Signed? No
Description and restrictions A 64-element array of scale values for AC coefcients for each qi value. A 64-element array of scale values for the DC coecient for each qi value. A NBMS64 array containing the base matrices. A 2 3 array containing the number of quant ranges for a given qti and pli , respectively. This is at most 63.
DCSCALE
16
No
BMS NQRS
2D Integer array 2D Integer array
8 6
No No
7.9. RECONSTRUCTION Name QRSIZES Type 3D Integer array Size (bits) 6 Signed? No
123 Description and restrictions A 2 3 63 array of the sizes of each quant range for a given qti and pli , respectively. Only the rst NQRS[qti ][pli ] values are used. A 2 3 64 array of the bmi s used for each quant range for a given qti and pli , respectively. Only the rst (NQRS[qti ][pli ] + 1) values are used. The width of the Y plane of the reference frames in pixels. The height of the Y plane of the reference frames in pixels. The width of the Cb and Cr planes of the reference frames in pixels. The height of the Cb and Cr planes of the reference frames in pixels. A RPYH RPYW array containing the contents of the Y plane of the golden reference frame. A RPCH RPCW array containing the contents of the Cb plane of the golden reference frame. A RPCH RPCW array containing the contents of the Cr plane of the golden reference frame.
QRBMIS
3D Integer array
No
RPYW
Integer
20
No
RPYH
Integer
20
No
RPCW
Integer
20
No
RPCH
Integer
20
No
GOLDREFY
2D Integer Array
No
GOLDREFCB
2D Integer Array
No
GOLDREFCR
2D Integer Array
No
124 Name PREVREFY Type 2D Integer Array Size (bits) 8
CHAPTER 7. FRAME DECODE Signed? No Description and restrictions A RPYH RPYW array containing the contents of the Y plane of the previous reference frame. A RPCH RPCW array containing the contents of the Cb plane of the previous reference frame. A RPCH RPCW array containing the contents of the Cr plane of the previous reference frame. The total number of blocks in a frame. An NBS-element array of ags indicating which blocks are coded. An NMBS-element array of coding modes for each macro block. An NBS-element array of motion vectors for each block. An NBS 64 array of quantized DCT coecient values for each block in zig-zag order. An NBS-element array of the coecient count for each block. An NQIS-element array of qi values. An NBS-element array of qii values for each block.
PREVREFCB
2D Integer Array
No
PREVREFCR
2D Integer Array
No
NBS BCODED
Integer Integer Array Integer Array Array of 2D Integer Vectors 2D Integer Array
36 1
No No
MBMODES
No
MVECTS
Yes
COEFFS
16
Yes
NCOEFFS
Integer Array Integer array Integer Array
No
QIS QIIS
6 2
No No
Output parameters:
7.9. RECONSTRUCTION Name RECY Type 2D Integer Array 2D Integer Array 2D Integer Array Size (bits) 8 Signed? No
125 Description and restrictions A RPYH RPYW array containing the contents of the Y plane of the reconstructed frame. A RPCH RPCW array containing the contents of the Cb plane of the reconstructed frame. A RPCH RPCW array containing the contents of the Cr plane of the reconstructed frame.
RECCB
No
RECCR
No
Variables used:
Name RPW
Type Integer
Size (bits) 20
Signed? No
Description and restrictions The width of the current plane of the current reference frame in pixels. The height of the current plane of the current reference frame in pixels. A RPH RPW array containing the contents of the current plane of the current reference frame. The horizontal pixel index of the lower-left corner of the current block. The vertical pixel index of the lower-left corner of the current block. The horizontal component of the rst whole-pixel motion vector. The vertical component of the rst whole-pixel motion vector. The horizontal component of the second whole-pixel motion vector. The vertical component of the second whole-pixel motion vector. An 8 8 array of predictor values to use for the current block.
RPH
Integer
20
No
REFP
2D Integer Array Integer
No
BX
20
No
BY
Integer
20
No
MVX MVY MVX2 MVY2 PRED
Integer Integer Integer Integer 2D Integer Array
5 5 5 5 8
No No No No No
126 Name RES Type 2D Integer Array Integer array Integer Integer Integer Integer Integer Integer Size (bits) 16 Signed? Yes
CHAPTER 7. FRAME DECODE Description and restrictions An 8 8 array containing the decoded residual for the current block. A 64-element array of quantization values for each DCT coecient in natural order. The dequantized DC coecient of a block. A reconstructed pixel value. The index of the current block in coded order. The index of the macro block containing block bi . The color plane index of the current block. The index of the reference frame indicated by the coding mode for macro block mbi . The horizontal pixel index in the block. The vertical pixel index in the block. A quantization type index. See Table 3.1. The quantization index of the DC coecient. The quantization index of the AC coecients.
QMAT
16
No
DC P bi mbi pli r
29 17 36 32 2 2
Yes Yes No No No No
bx by qti qi0 qi
3 3 1 6 6
No No No No No
This section takes the decoded packet data and uses the previously dened procedures to reconstruct each block of the current frame. For coded blocks, a predictor is formed using the coding mode and, if applicable, the motion vector, and then the residual is computed from the quantized DCT coecients. For uncoded blocks, the contents of the co-located block are copied from the previous frame and the residual is cleared to zero. Then the predictor and residual are added, and the result clamped to the range 0 . . . 255 and stored in the current frame. In the special case that a block contains only a DC coecient, the dequantization and inverse DCT transform is skipped. Instead the constant pixel value for the entire block is computed in one step. Note that the truncation of intermediate operations is omitted and the nal rounding is slightly dierent in this case. The check for whether or not the block contains only a DC coecient
7.9. RECONSTRUCTION
127
is based on the coecient count returned from the token decode procedure of Section 7.7, and not by checking to see if the remaining coecient values are zero. Also note that even when the coecient count indicates the block contains zero coecients, the DC coecient is still processed, as undoing DC prediction might have made it non-zero. After this procedure, the frame is completely reconstructed, but before it can be used as a reference frame, a loop lter must be run over it to help reduce blocking artifacts. This is detailed in Section 7.10. 1. Assign qi0 the value QIS[0]. 2. For each value of bi from 0 to (NBS 1): (a) Assign pli the index of the color plane block bi belongs to. (b) Assign BX the horizontal pixel index of the lower-left corner of block bi . (c) Assign BY the vertical pixel index of the lower-left corner of block bi . (d) If BCODED[bi ] is non-zero: i. ii. iii. iv. Assign mbi the index of the macro block containing block bi . If MBMODES[mbi ] is 1 (INTRA), assign qti the value 0. Otherwise, assign qti the value 1. Assign r the value of the Reference Frame Index column of Table 7.46 corresponding to MBMODES[mbi ]. v. If r is zero, compute PRED using the procedure given in Section 7.9.1.1. vi. Otherwise: A. Assign REFP, RPW, and RPH the values given in Table 7.75 corresponding to current value of r and pli . r 1 1 1 2 2 2 pli 0 1 2 0 1 2 REFP PREVREFY PREVREFCB PREVREFCR GOLDREFY GOLDREFCB GOLDREFCR RPW RPYW RPCW RPCW RPYW RPCW RPCW RPH RPYH RPCH RPCH RPYH RPCH RPCH
Table 7.75: Reference Planes and Sizes for Each r and pli B. Assign MVX the value |MVECTS[bi ]x | sign(MVECTS[bi ]x ).
128 C. Assign MVY the value
|MVECTS[bi ]y | sign(MVECTS[bi ]y ). D. Assign MVX2 the value |MVECTS[bi ]x | sign(MVECTS[bi ]x ). E. Assign MVY2 the value |MVECTS[bi ]y | sign(MVECTS[bi ]y ). F. If MVX equals MVX2 and MVY equals MVY2, use the values REFP, RPW, RPH, BX, BY, MVX, and MVY, compute PRED using the procedure given in Section 7.9.1.2. G. Otherwise, use the values REFP, RPW, RPH, BX, BY, MVX, MVY, MVX2, and MVY2 to compute PRED using the procedure given in Section 7.9.1.3. vii. If NCOEFFS[bi ] is less than 2: A. Using ACSCALE, DCSCALE, BMS, NQRS, QRSIZES, QRBMIS, qti , pli , and qi0 , use the procedure given in Section 6.4.3 to compute the DC quantization matrix QMAT. B. Assign DC the value (COEFFS[bi ][0] QMAT[0] + 15) >> 5. C. Truncate DC to a 16-bit signed representation by dropping any higher-order bits. D. For each value of by from 0 to 7, and each value of bx from 0 to 7, assign RES[by][bx ] the value DC. viii. Otherwise: A. Assign qi the value QIS[QIIS[bi ]]. B. Using ACSCALE, DCSCALE, BMS, NQRS, QRSIZES, QRBMIS, qti , pli , qi0 , and qi , compute DQC using the procedure given in Section 7.9.2. C. Using DQC, compute RES using the procedure given in Section 7.9.3.2. (e) Otherwise: i. Assign r the value 1. ii. Assign REFP, RPW, and RPH the values given in Table 7.75 corresponding to current value of r and pli . iii. Assign MVX the value 0. iv. Assign MVY the value 0.
7.10. LOOP FILTERING
129
v. Using the values REFP, RPW, RPH, BX, BY, MVX, and MVY, compute PRED using the procedure given in Section 7.9.1.2. This is simply a copy of the co-located block in the previous reference frame. vi. For each value of by from 0 to 7, and each value of bx from 0 to 7, assign RES[by][bx ] the value 0. (f) For each value of by from 0 to 7, and each value of bx from 0 to 7: i. ii. iii. iv. v. Assign P the value (PRED[by][bx ] + RES[by][bx ]). If P is greater than 255, assign P the value 255. If P is less than 0, assign P the value 0. If pli equals 0, assign RECY[BY + by][BX + bx ] the value P. Otherwise, if pli equals 1, assign RECB[BY + by][BX + bx ] the value P. vi. Otherwise, pli equals 2, so assign RECR[BY + by][BX + bx ] the value P.
7.10
Loop Filtering
lflim (L,L)
(2L,0) (2L,0) R
(L,L)
Figure 7.3: The loop lter response function. The loop lter is a simple deblocking lter that is based on running a small edge detecting lter over the coded block edges and adjusting the pixel values by
130
a tapered response. The lter response is modulated by the following non-linear function: R 2 L 0, R 2 L, 2 L < R L R, L < R < L lim(R, L) = R + 2 L, L R < 2 L 0, 2LR Here L is a limiting value equal to LFLIMS[qi0 ]. It denes the peaks of the function, illustrated in Figure 7.3. LFLIMS is an array of values specied in the setup header and is indexed by qi0 , the rst quantization index for the frame, the one used for all the DC coecients. Larger values of L indicate a stronger lter.
7.10.1
Horizontal Filter
Input parameters:
Name RECP
Type 2D Integer Array Integer
Size (bits) 8
Signed? No
Description and restrictions A RPH RPW array containing the contents of a plane of the reconstructed frame. The horizontal pixel index of the lower-left corner of the area to be ltered. The vertical pixel index of the lower-left corner of the area to be ltered. The loop lter limit value.
FX
20
No
FY
Integer
20
No
Integer
No
Output parameters:
Name RECP
Size (bits) 8
Signed? No
Description and restrictions A RPH RPW array containing the contents of a plane of the reconstructed frame.
7.10. LOOP FILTERING Variables used:
131
Name R P by
Size (bits) 9 9 20
Signed? Yes Yes No
Description and restrictions The edge detector response. A ltered pixel value. The vertical pixel index in the block.
This procedure applies a 4-tap horizontal lter to each row of a vertical block edge. 1. For each value of by from 0 to 7: (a) Assign R the value (RECP[FY + by][FX] 3 RECP[FY + by][FX + 1]+ 3 RECP[FY + by][FX + 2] RECP[FY + by][FX + 3] + 4) >> 3 (b) Assign P the value (RECP[FY + by][FX + 1] + lim(R, L)). (c) If P is less than zero, assign RECP[FY + by][FX + 1] the value zero. (d) Otherwise, if P is greater than 255, assign RECP[FY + by][FX + 1] the value 255. (e) Otherwise, assign RECP[FY + by][FX + 1] the value P. (f) Assign P the value (RECP[FY + by][FX + 2] lim(R, L)). (g) If P is less than zero, assign RECP[FY + by][FX + 2] the value zero. (h) Otherwise, if P is greater than 255, assign RECP[FY + by][FX + 2] the value 255. (i) Otherwise, assign RECP[FY + by][FX + 2] the value P.
7.10.2
Vertical Filter
Input parameters:
132 Name RECP Type 2D Integer Array Integer Size (bits) 8 Signed? No
CHAPTER 7. FRAME DECODE Description and restrictions A RPH RPW array containing the contents of a plane of the reconstructed frame. The horizontal pixel index of the lower-left corner of the area to be ltered. The vertical pixel index of the lower-left corner of the area to be ltered. The loop lter limit value.
FX
20
No
FY
Integer
20
No
Integer
No
Output parameters:
Name RECP
Size (bits) 8
Signed? No
Description and restrictions A RPH RPW array containing the contents of a plane of the reconstructed frame.
Variables used:
Name R P bx
Size (bits) 9 9 20
Signed? Yes Yes No
Description and restrictions The edge detector response. A ltered pixel value. The horizontal pixel index in the block.
This procedure applies a 4-tap vertical lter to each column of a horizontal block edge. 1. For each value of bx from 0 to 7: (a) Assign R the value (RECP[FY][FX + bx ] 3 RECP[FY + 1][FX + bx ]+ 3 RECP[FY + 2][FX + bx ] RECP[FY + 3][FX + bx ] + 4) >> 3
7.10. LOOP FILTERING (b) Assign P the value (RECP[FY + 1][FX + bx ] + lim(R, L)).
133
(c) If P is less than zero, assign RECP[FY + 1][FX + bx ] the value zero. (d) Otherwise, if P is greater than 255, assign RECP[FY + 1][FX + bx ] the value 255. (e) Otherwise, assign RECP[FY + 1][FX + bx ] the value P. (f) Assign P the value (RECP[FY + 2][FX + bx ] lim(R, L)). (g) If P is less than zero, assign RECP[FY + 2][FX + bx ] the value zero. (h) Otherwise, if P is greater than 255, assign RECP[FY + 2][FX + bx ] the value 255. (i) Otherwise, assign RECP[FY + 2][FX + bx ] the value P.
7.10.3
Complete Loop Filter
Input parameters:
Name LFLIMS RPYW
Type Integer array Integer
Size (bits) 7 20
Signed? No No
Description and restrictions A 64-element array of loop lter limit values. The width of the Y plane of the reconstruced frame in pixels. The height of the Y plane of the reconstruced frame in pixels. The width of the Cb and Cr planes of the reconstruced frame in pixels. The height of the Cb and Cr planes of the reconstruced frame in pixels. The total number of blocks in a frame. An NBS-element array of ags indicating which blocks are coded. An NQIS-element array of qi values. A RPYH RPYW array containing the contents of the Y plane of the reconstructed frame.
RPYH
Integer
20
No
RPCW
Integer
20
No
RPCH
Integer
20
No
NBS BCODED
Integer Integer Array Integer array 2D Integer Array
36 1
No No
QIS RECY
6 8
No No
134 Name RECCB Type 2D Integer Array Size (bits) 8
CHAPTER 7. FRAME DECODE Signed? No Description and restrictions A RPCH RPCW array containing the contents of the Cb plane of the reconstructed frame. A RPCH RPCW array containing the contents of the Cr plane of the reconstructed frame.
RECCR
2D Integer Array
No
Output parameters:
Name RECY
Type 2D Integer Array 2D Integer Array 2D Integer Array
Size (bits) 8
Signed? No
Description and restrictions A RPYH RPYW array containing the contents of the Y plane of the reconstructed frame. A RPCH RPCW array containing the contents of the Cb plane of the reconstructed frame. A RPCH RPCW array containing the contents of the Cr plane of the reconstructed frame.
RECCB
No
RECCR
No
Variables used:
Name RPW RPH RECP
Type Integer Integer 2D Integer Array
Size (bits) 20 20 8
Signed? No No No
Description and restrictions The width of the current plane of the reconstructed frame in pixels. The height of the current plane of the reconstructed frame in pixels. A RPH RPW array containing the contents of the current plane of the reconstruced frame.
7.10. LOOP FILTERING Name BX Type Integer Size (bits) 20 Signed? No Description and restrictions
135
BY
Integer
20
No
FX
Integer
20
No
FY
Integer
20
No
L bi bj pli
Integer Integer Integer Integer
7 36 36 2
No No No No
The horizontal pixel index of the lower-left corner of the current block. The vertical pixel index of the lower-left corner of the current block. The horizontal pixel index of the lower-left corner of the area to be ltered. The vertical pixel index of the lower-left corner of the area to be ltered. The loop lter limit value. The index of the current block in coded order. The index of a neighboring block in coded order. The color plane index of the current block.
This procedure denes the order that the various block edges are ltered. Because each application of one of the two lters above destructively modies the contents of the reconstructed image, the precise output obtained diers depending on the order that horizontal and vertical lters are applied to the edges of a single block. The order dened here conforms to that used by VP3. 1. Assign L the value LFLIMS[QIS[0]]. 2. For each block in raster order, with coded-order index bi : (a) If BCODED[bi ] is non-zero: i. Assign pli the index of the color plane block bi belongs to. ii. Assign RECP, RPW, and RPH the values given in Table 7.85 corresponding to the value of pli . pli 0 1 2 RECP RECY RECCB RECCR RPW RPYW RPCW RPCW RPH RPYH RPCH RPCH
Table 7.85: Reconstructed Planes and Sizes for Each pli
136
CHAPTER 7. FRAME DECODE iii. Assign BX the horizontal pixel index of the lower-left corner of the block bi . iv. Assign BY the vertical pixel index of the lower-left corner of the block bi . v. If BX is greater than zero: A. Assign FX the value (BX 2). B. Assign FY the value BY. C. Using RECP, FX, FY, and L, apply the horizontal block lter to the left edge of block bi with the procedure described in Section 7.10.1. vi. If BY is greater than zero: A. Assign FX the value BX. B. Assign FY the value (BY 2) C. Using RECP, FX, FY, and L, apply the vertical block lter to the bottom edge of block bi with the procedure described in Section 7.10.2. vii. If (BX + 8) is less than RPW and BCODED[bj ] is zero, where bj is the coded-order index of the block adjacent to bi on the right: A. Assign FX the value (BX + 6). B. Assign FY the value BY. C. Using RECP, FX, FY, and L, apply the horizontal block lter to the right edge of block bi with the procedure described in Section 7.10.1. viii. If (BY + 8) is less than RPH and BCODED[bj ] is zero, where bj is the coded-order index of the block adjacent to bi above: A. Assign FX the value BX. B. Assign FY the value (BY + 6) C. Using RECP, FX, FY, and L, apply the vertical block lter to the top edge of block bi with the procedure described in Section 7.10.2.
VP3 Compatibility The original VP3 decoder implemented unrestricted motion vectors by enlarging the reconstructed frame buers and repeating the pixels on its edges into the padding region. However, for the previous reference frame this padding ocurred before the loop lter was applied, but for the golden reference frame it occurred afterwards. This means that for the previous reference frame, the padding values were required to be stored separately from the main image values. Furthermore, even if the previous and golden reference frames were in fact the same frame, they could have dierent padding values. Finally, the encoder did not apply the loop
7.11. COMPLETE FRAME DECODE
137
lter at all, which resulted in artifacts, particularly in near-static scenes, due to prediction-loop mismatch. This last can only be considered a bug in the VP3 encoder. Given all these things, Theora now uniformly applies the loop lter before the reference frames are padded. This means it is possible to use the same buer for the previous and golden reference frames when they do indeed refer to the same frame. It also means that on architectures where memory bandwidth is limited, it is possible to avoid storing padding values, and simply clamp the motion vectors applied to each pixel as described in Sections 7.9.1.2 and 7.9.1.3. This means that the predicted pixel values along the edges of the frame might dier slightly between VP3 and Theora, but since the VP3 encoder did not apply the loop lter in the rst place, this is not likely to impose any serious compatibility issues.
7.11
Complete Frame Decode
Input parameters:
Name FMBW FMBH NSBS NBS NMBS FRN FRD PARN PARD CS PF NOMBR
Type Integer Integer Integer Integer Integer Integer Integer Integer Integer Integer Integer Integer
Size (bits) 16 16 32 36 32 32 32 24 24 8 2 24
Signed? No No No No No No No No No No No No
Description and restrictions The width of the frame in macro blocks. The height of the frame in macro blocks. The total number of super blocks in a frame. The total number of blocks in a frame. The total number of macro blocks in a frame. The frame-rate numerator. The frame-rate denominator. The pixel aspect-ratio numerator. The pixel aspect-ratio denominator. The color space. The pixel format. The nominal bitrate of the stream, in bits per second.
138 Name QUAL KFGSHIFT Type Integer Integer Size (bits) 6 5
CHAPTER 7. FRAME DECODE Signed? No No Description and restrictions The quality hint. The amount to shift the key frame number by in the granule position. A 64-element array of loop lter limit values. A 64-element array of scale values for AC coefcients for each qi value. A 64-element array of scale values for the DC coecient for each qi value. The number of base matrices. A NBMS64 array containing the base matrices. A 2 3 array containing the number of quant ranges for a given qti and pli , respectively. This is at most 63. A 2 3 63 array of the sizes of each quant range for a given qti and pli , respectively. Only the rst NQRS[qti ][pli ] values will be used. A 2 3 64 array of the bmi s used for each quant range for a given qti and pli , respectively. Only the rst (NQRS[qti ][pli ] + 1) values will be used. An 80-element array of Human tables with up to 32 entries each. A RPYH RPYW array containing the contents of the Y plane of the golden reference frame.
LFLIMS ACSCALE
Integer array Integer array Integer array
7 16
No No
DCSCALE
16
No
NBMS BMS NQRS
Integer 2D Integer array 2D Integer array
10 8 6
No No No
QRSIZES
3D Integer array
No
QRBMIS
3D Integer array
No
HTS
Human table array
GOLDREFY
2D Integer Array
No
7.11. COMPLETE FRAME DECODE Name GOLDREFCB Type 2D Integer Array Size (bits) 8 Signed? No
139 Description and restrictions A RPCH RPCW array containing the contents of the Cb plane of the golden reference frame. A RPCH RPCW array containing the contents of the Cr plane of the golden reference frame. A RPYH RPYW array containing the contents of the Y plane of the previous reference frame. A RPCH RPCW array containing the contents of the Cb plane of the previous reference frame. A RPCH RPCW array containing the contents of the Cr plane of the previous reference frame.
GOLDREFCR
2D Integer Array
No
PREVREFY
2D Integer Array
No
PREVREFCB
2D Integer Array
No
PREVREFCR
2D Integer Array
No
Output parameters:
Name RECY
Size (bits) 8
Signed? No
Description and restrictions A RPYH RPYW array containing the contents of the Y plane of the reconstructed frame. A RPCH RPCW array containing the contents of the Cb plane of the reconstructed frame.
RECCB
2D Integer Array
No
140 Name RECCR Type 2D Integer Array Size (bits) 8
CHAPTER 7. FRAME DECODE Signed? No Description and restrictions A RPCH RPCW array containing the contents of the Cr plane of the reconstructed frame. A RPYH RPYW array containing the contents of the Y plane of the golden reference frame. A RPCH RPCW array containing the contents of the Cb plane of the golden reference frame. A RPCH RPCW array containing the contents of the Cr plane of the golden reference frame. A RPYH RPYW array containing the contents of the Y plane of the previous reference frame. A RPCH RPCW array containing the contents of the Cb plane of the previous reference frame. A RPCH RPCW array containing the contents of the Cr plane of the previous reference frame.
GOLDREFY
2D Integer Array
No
GOLDREFCB
2D Integer Array
No
GOLDREFCR
2D Integer Array
No
PREVREFY
2D Integer Array
No
PREVREFCB
2D Integer Array
No
PREVREFCR
2D Integer Array
No
Variables used:
Name FTYPE
Type Integer
Size (bits) 1
Signed? No
Description and restrictions The frame type.
7.11. COMPLETE FRAME DECODE Name NQIS QIS BCODED Type Integer Integer array Integer Array Integer Array Array of 2D Integer Vectors Integer Array 2D Integer Array Integer Array Integer Integer Integer Size (bits) 2 6 1 Signed? No No No
141 Description and restrictions The number of qi values. An NQIS-element array of qi values. An NBS-element array of ags indicating which blocks are coded. An NMBS-element array of coding modes for each macro block. An NBS-element array of motion vectors for each block. An NBS-element array of qii values for each block. An NBS 64 array of quantized DCT coecient values for each block in zig-zag order. An NBS-element array of the coecient count for each block. The width of the Y plane of the reference frames in pixels. The height of the Y plane of the reference frames in pixels. The width of the Cb and Cr planes of the reference frames in pixels. The height of the Cb and Cr planes of the reference frames in pixels. The index of the current block in coded order.
MBMODES
No
MVECTS
Yes
QIIS COEFFS
2 16
No Yes
NCOEFFS
No
RPYW RPYH RPCW
20 20 20
No No No
RPCH
Integer
20
No
bi
Integer
36
No
This procedure uses all the procedures dened in the previous section of this chapter to decode and reconstruct a complete frame. It takes as input values decoded from the headers, as well as the current reference frames. As output, it gives the uncropped, reconstructed frame. This should be cropped to picture region before display. As a special case, a 0-byte packet is treated exactly like an inter frame with no coded blocks. 1. If the size of the data packet is non-zero: (a) Decode the frame header values FTYPE, NQIS, and QIS using the procedure given in Section 7.1.
142 PF 0 2 3 RPCW 8 FMBW 8 FMBW 16 FMBW
CHAPTER 7. FRAME DECODE RPCH 8 FMBH 16 FMBH 16 FMBH
Table 7.89: Width and Height of Chroma Planes for each Pixel Format (b) Using FTYPE, NSBS, and NBS, decode the list of coded block ags into BCODED using the procedure given in Section 7.3. (c) Using FTYPE, NMBS, NBS, and BCODED, decode the macro block coding modes into MBMODES using the procedure given in Section 7.4. (d) If FTYPE is non-zero (inter frame), using PF, NMBS, MBMODES, NBS, and BCODED, decode the motion vectors into MVECTS using the procedure given in Section 7.5.1. (e) Using NBS, BCODED, and NQIS, decode the block-level qi values into QIIS using the procedure given in Section 7.6. (f) Using NBS, NMBS, BCODED, and HTS, decode the DCT coecients into NCOEFFS and NCOEFFS using the procedure given in Section 7.7.3. (g) Using BCODED and MBMODES, undo the DC prediction on the DC coecients stored in COEFFS using the procedure given in Section 7.8.2. 2. Otherwise: (a) Assign FTYPE the value 1 (inter frame). (b) Assign NQIS the value 1. (c) Assign QIS[0] the value 63. (d) For each value of bi from 0 to (NBS 1), assign BCODED[bi ] the value zero. 3. Assign RPYW and RPYH the values (16 FMBW) and (16 FMBH), respectively. 4. Assign RPCW and RPCH the values from the row of Table 7.89 corresponding to PF. 5. Using ACSCALE, DCSCALE, BMS, NQRS, QRSIZES, QRBMIS, NBS, BCODED, MBMODES, MVECTS, COEFFS, NCOEFFS, QIS, QIIS, RPYW, RPYH, RPCW, RPCH, GOLDREFY, GOLDREFCB, GOLDREFCR, PREVREFY, PREVREFCB, and PREVREFCR, reconstruct the complete frame into RECY, RECCB, and RECCR using the procedure given in Section 7.9.4.
7.11. COMPLETE FRAME DECODE
143
6. Using LFLIMS, RPYW, RPYH, RPCW, RPCH, NBS, BCODED, and QIS, apply the loop lter to the reconstructed frame in RECY, RECCB, and RECCR using the procedure given in Section 7.10.3. 7. If FTYPE is zero (intra frame), assign GOLDREFY, GOLDREFCB, and GOLDREFCR the values RECY, RECCB, and RECCR, respectively. 8. Assign PREVREFY, PREVREFCB, and PREVREFCR the values RECY, RECCB, and RECCR, respectively.
144
Appendix A
Ogg Bitstream Encapsulation

A.1 Overview
This document species the embedding or encapsulation of Theora packets in an Ogg transport stream. Ogg is a stream oriented wrapper for coded, linear time-based data. It provides syncronization, multiplexing, framing, error detection and seeking landmarks for the decoder and complements the raw packet format used by the Theora codec. This document assumes familiarity with the details of the Ogg standard. The Xiph.Org documentation provides an overview of the Ogg transport stream format at http://www.xiph.org/ogg/doc/oggstream.html and a detailed description at http://www.xiph.org/ogg/doc/framing.html. The format is also dened in RFC 3533 [Pfe03]. While Theora packets can be embedded in a wide variety of media containers and streaming mechanisms, the Xiph.Org Foundation recommends Ogg as the native format for Theora video in le-oriented storage and transmission contexts.
A.1.1
MIME type
The generic MIME type of any Ogg le is application/ogg. The specic MIME type for the Ogg Theora prole documented here is video/ogg. This is the MIME type recommended for les conforming to this appendix. The recommended lename extension is .ogv. Outside of an encapsulation, the mime type video/theora may be used to refer specically to the Theora compressed video stream. 145
146
APPENDIX A. OGG BITSTREAM ENCAPSULATION
A.2
Embedding in a logical bitstream
Ogg separates the concept of a logical bitstream consisting of the framing of a particular sequence of packets and complete within itself from the physical bitstream which may consist either of a single logical bitstream or a number of logical bitstreams multiplexed together. This section species the embedding of Theora packets in a logical Ogg bitstream. The mapping of Ogg Theora logical bitstreams into a multiplexed physical Ogg stream is described in the next section.
A.2.1
Headers
The initial identication header packet appears by itself in a single Ogg page. This page denes the start of the logical stream and MUST have the beginning of stream ag set. The second and third header packets (comment metadata and decoder setup data) can together span one or more Ogg pages. If there are additional nonnormative header packets, they MUST be included in this sequence of pages as well. The comment header packet MUST begin the second Ogg page in the logical bitstream, and there MUST be a page break between the last header packet and the rst frame data packet. These two page break requirements facilitate stream identication and simplify header acquisition for seeking and live streaming applications. All header pages MUST have their granule position eld set to zero.
A.2.2
Frame data
The rst frame data packet in a logical bitstream MUST begin a new Ogg page. All other data packets are placed one at a time into Ogg pages until the end of the stream. Packets can span pages and multiple packets can be placed within any one page. The last page in the logical bitstream SHOULD have its end of stream ag set to indicate complete transmission of the available video. Frame data pages MUST be marked with a granule position corresponding to the end of the display interval of the last frame/packet that nishes in that page. See the next section for details.
A.2.3
Granule position
Data packets are marked by a granulepos derived from the count of decodable frames after that packet is processed. The eld itself is divided into two sections, the width of the less signicant section being given by the KFGSHIFT parameter decoded from the identication header (Section 6.2). The more signicant portion of the eld gives the count of coded frames after the coding of the last keyframe in stream, and the less signicant portion gives the count of frames since the last keyframe. Thus a stream would begin with a split granulepos of
A.3. MULTIPLEXED STREAM MAPPING
147
1|0 (a keyframe), followed by 1|1, 1|2, 1|3, etc. Around a keyframe in the middle of the stream the granulepos sequence might be 1234|35, 1234|36, 1234|37, 1271|0 (for the keyframe), 1271|1, and so on. In this way the granulepos eld increased monotonically as required by the Ogg format, but contains information necessary to eciently nd the previous keyframe to continue decoding after a seek. Prior to bitstream version 3.2.1, data packets were marked by a granulepos derived from the index of the frame being decoded, rather than the count. That is they marked the beginning of the display interval of a frame rather than the end. Such streams have the VREV eld of the identication header set to 0 instead of 1. They can be interpreted according to the description above by adding 1 to the more signication eld of the split granulepos when VREV is less than 1.
A.3
Multiplexed stream mapping
Applications supporting Ogg Theora must support Theora bitstreams multiplexed with compressed audio data in the Vorbis I and Speex formats, and should support Ogg-encapsulated MNG graphics for overlays. Multiple audio and video bitstreams may be multiplexed together. How playback of multiple/alternate streams is handled is up to the application. Some conventions based on included metadata aide interoperability in this respect.
A.3.1
Chained streams
Ogg Theora decoders and playback applications MUST support both grouped streams (multiplexed concurrent logical streams) and chained streams (sequential concatenation of independent physical bitstreams). The number and codec data types of multiplexed streams and the decoder parameters for those stream types that re-occur can all change at a chaining boundary. A playback application MUST be prepared to handle such changes and SHOULD do so smoothly with the minimum possible visible disruption. The specication of grouped streams below applies independently to each segment of a chained bitstream.
A.3.2
Grouped streams
At the beginning of a multiplexed stream, the beginning of stream pages for each logical bitstream will be grouped together. Within these, the rst page to occur MUST be the Theora page. This facilitates identication of Ogg Theora les among other Ogg-encapsulated content. A playback application must nevertheless handle streams where this arrangement is not correct. If there is more than one Theora logical stream, the rst page should be from the primary stream. That is, the best choice for the stream a generic player should begin displaying without special user direction. If there is more than
148
APPENDIX A. OGG BITSTREAM ENCAPSULATION
one audio stream, or of any other stream type, the identication page of the primary stream of that type should be placed before the others. After the beginning of stream pages, the header pages of each of the logical streams MUST be grouped together before any data pages occur. After all the header pages have been placed, the data pages are multiplexed together. They should be placed in the stream in increasing order by the time equivalents of their granule position elds. This facilitates seeking while limiting the buering requirements of the playback demultiplexer.
Appendix B
VP3
B.1 VP3 Compatibility
This section lists all of the encoder and decoder issues that may aect VP3 compatibly. Each is described in more detail in the text itself. This list is provided merely for reference.
Bitstream headers (Section 6).
Identication header (Section 6.2).

* Non-multiple of 16 picture sizes. * Standardized color spaces. * Support for 4 : 4 : 4 and 4 : 2 : 2 pixel formats.
Setup header
* Loop lter limit values (Section 6.4.1). * Quantization parameters (Section 6.4.2). * Human tables (Section 6.4.4). Frame header format (Section 7.1). Extended long-run bit strings (Section 7.2.1). INTER MV FOUR handling of uncoded blocks (Section 7.5.2). Block-level qi values (Section 7.6). Zero-length EOB runs (Section 7.7.1). Unrestricted motion vector padding and the loop lter (Section 7.10.3).
149
150
APPENDIX B. VP3
B.2
Loop Filter Limit Values
The hard-coded loop lter limit values used in VP3 are dened as follows: LFLIMS = {30, 25, 20, 20, 15, 15, 14, 14, 13, 13, 12, 12, 11, 11, 10, 10, 9, 9, 8, 8, 7, 7, 7, 7, 6, 6, 6, 6, 5, 5, 5, 5, 4, 4, 4, 4, 3, 3, 3, 3, 2, 2, 2, 2, 2, 2, 2, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}
B.3
Quantization Parameters
The hard-coded quantization parameters used by VP3 are dened as follows:
ACSCALE = {500, 450, 400, 370, 340, 310, 285, 265, 245, 225, 210, 195, 185, 180, 170, 160, 150, 145, 135, 130, 125, 115, 110, 107, 100, 96, 93, 89, 85, 82, 75, 74, 70, 68, 64, 60, 57, 56, 52, 50, 49, 45, 44, 43, 40, 38, 37, 35, 33, 32, 30, 29, 28, 25, 24, 22, 21, 19, 18, 17, 15, 13, 12, 10 } DCSCALE = {220, 200, 190, 180, 170, 170, 160, 160, 150, 150, 140, 140, 130, 130, 120, 120, 110, 110, 100, 100, 90, 90, 90, 80, 80, 80, 70, 70, 70, 60, 60, 60, 60, 50, 50, 50, 50, 40, 40, 40, 40, 40, 30, 30, 30, 30, 30, 30, 30, 20, 20, 20, 20, 20, 20, 20, 20, 10, 10, 10, 10, 10, 10, 10 }
VP3 denes only a single quantization range for each quantization type and color plane, and the base matrix used is constant throughout the range. There are three base matrices dened. The rst is used for the Y channel of INTRA mode blocks, and the second for both the Cb and Cr channels of INTRA mode blocks. The last is used for INTER mode blocks of all channels.
B.4. HUFFMAN TABLES
151
BMS = { {16, 12, 14, 14, 18, 24, 49, 72, {17, 18, 24, 47, 99, 99, 99, 99, {16, 16, 16, 20, 24, 28, 32, 40,
11, 12, 13, 17, 22, 35, 64, 92, 18, 21, 26, 66, 99, 99, 99, 99, 16, 16, 20, 24, 28, 32, 40, 48,
10, 14, 16, 22, 37, 55, 78, 95, 24, 26, 56, 99, 99, 99, 99, 99, 16, 20, 24, 28, 32, 40, 48, 64,
16, 24, 40, 19, 26, 58, 24, 40, 57, 29, 51, 87, 58, 68, 109, 64, 81, 104, 87, 103, 121, 98, 112, 100, 47, 99, 99, 66, 99, 99, 99, 99, 99, 99, 99, 99, 99, 99, 99, 99, 99, 99, 99, 99, 99, 99, 99, 99, 20, 24, 28, 24, 28, 32, 28, 32, 40, 32, 40, 48, 40, 48, 64, 48, 64, 64, 64, 64, 64, 64, 64, 96,
51, 61, 60, 55, 69, 56, 80, 62, 103, 77, 113, 92, 120, 101, 103, 99 }, 99, 99, 99, 99, 99, 99, 99, 99, 99, 99, 99, 99, 99, 99, 99, 99 }, 32, 40, 40, 48, 48, 64, 64, 64, 64, 64, 64, 96, 96, 128, 128, 128 } }
The remaining parameters simply assign these matrices to the proper quant ranges. NQRS ={{1, 1, 1}, {1, 1, 1}} QRSIZES ={{{63}, {63}, {63}}, {{63}, {63}, {63}}} QRBMIS ={{{0, 0}, {1, 1}, {1, 1}}, {{2, 2}, {2, 2}, {2, 2}}}
B.4
Human Tables
The following tables contain the hard-coded Human codes used by VP3. There are 80 tables in all, each with a Human code for all 32 token values. The tokens are sorted by the most signicant bits of their Human code. This is the same order in which they will be decoded from the setup header.
152 Human Code b000 b001 b01000 b010010 b0100110 b01001110 b010011110 b010011111000 b010011111001 b01001111101 b0100111111 b0101 b0110 b0111 b1000 b1001 b101000 b101001 b10101 b1011000 b10110010 b101100110 b1011001110 b10110011110 b10110011111 b101101 b10111 b1100 b11010 b11011 b1110 b1111 Token Value 20 19 7 30 1 3 28 26 8 6 29 14 12 17 13 21 24 23 16 31 25 2 4 5 27 0 22 18 15 11 10 9
APPENDIX B. VP3 Human Code b000 b001 b0100 b0101 b0110 b0111 b10000 b100010 b10001100 b1000110100000 b1000110100001 b100011010001 b10001101001 b10001101010 b10001101011 b100011011 b1000111 b1001 b1010 b10110 b10111 b1100 b1101 b1110 b11110 b1111100 b1111101 b1111110 b11111110 b1111111100 b1111111101 b111111111 Token Value 20 19 11 12 14 17 0 7 3 29 8 26 6 5 28 25 1 13 21 16 22 18 10 9 15 30 23 24 31 4 27 2
VP3.1 Human Table Number 0
B.4. HUFFMAN TABLES Human Code b000 b001 b0100 b0101 b0110 b0111 b100000 b1000010 b10000110 b10000111 b10001 b1001 b1010 b10110 b10111 b1100 b11010 b11011000 b1101100100 b11011001010 b11011001011 b11011001100 b110110011010 b11011001101100 b11011001101101 b1101100110111 b1101100111 b1101101 b1101110 b1101111 b1110 b1111 Token Value 20 19 11 12 14 17 1 30 2 3 16 21 13 0 22 18 15 31 25 27 6 5 26 29 8 28 4 24 7 23 10 9 Human Code b0000 b0001 b0010 b00110 b001110 b001111 b010 b011 b1000 b1001 b1010 b1011000 b101100100 b1011001010 b1011001011 b1011001100000 b1011001100001 b101100110001 b10110011001 b1011001101 b101100111 b101101 b10111 b1100 b1101 b1110 b1111000 b1111001 b1111010 b11110110 b11110111 b11111
153 Token Value 0 17 14 21 7 23 10 9 11 12 20 3 25 6 5 29 8 28 26 27 4 1 16 18 13 19 22 30 24 31 2 15
155 Token Value 10 3 19 24 7 5 21 6 16 14 23 2 1 11 12 20 4 25 28 8 29 22 26 27 31 30 18 17 15 13 0 9
157 Token Value 0 4 21 30 15 25 29 7 6 5 22 8 23 26 19 16 2 13 1 14 12 11 20 31 17 3 18 27 28 24 10 9
158 Human Code b00 b010 b01100 b01101 b011100 b011101 b011110 b01111100 b01111101 b0111111 b1000 b1001 b101000 b10100100 b1010010100 b1010010101 b101001011 b1010011 b101010 b101011 b10110 b101110 b101111 b11000 b11001 b1101000 b1101001 b1101010 b11010110 b11010111 b11011 b111 Token Value 9 0 14 3 26 18 17 8 21 30 12 11 15 7 6 5 22 4 28 27 24 25 2 1 23 29 19 16 31 20 13 10 Human Code b00 b010 b01100 b01101 b01110 b011110 b0111110 b01111110 b01111111 b1000 b1001 b101000 b101001 b101010 b101011 b10110 b101110 b1011110000 b1011110001 b101111001 b10111101 b1011111 b11000 b1100100 b1100101 b110011000 b110011001 b11001101 b1100111 b11010 b11011 b111
APPENDIX B. VP3 Token Value 9 0 2 14 24 17 29 21 5 12 11 28 4 15 27 23 25 6 22 8 30 19 3 16 26 7 31 20 18 13 1 10
159 Token Value 10 9 12 11 15 5 30 29 28 22 8 7 31 21 26 19 16 3 2 4 18 24 17 6 25 13 14 27 20 23 1 0
161 Token Value 28 27 22 8 6 29 5 4 2 21 1 15 23 7 11 17 12 19 25 26 3 20 18 14 31 24 30 10 9 13 16 0
163 Token Value 0 10 9 3 27 16 13 31 17 4 28 11 12 24 7 25 26 2 1 14 23 19 20 8 22 21 29 6 5 18 15 30
165 Token Value 12 11 10 9 23 7 14 3 6 22 21 5 20 4 18 1 24 15 29 28 26 8 19 16 27 13 30 25 17 2 31 0
167 Token Value 0 10 9 24 29 17 22 21 18 20 19 8 16 15 27 12 23 1 11 13 7 14 3 2 26 28 5 4 6 31 30 25
169 Token Value 30 18 16 9 10 31 26 1 0 14 17 24 23 11 12 13 15 5 6 22 21 4 20 3 19 25 28 27 2 29 8 7
171 Token Value 23 1 13 15 8 18 20 19 22 21 17 16 6 7 27 0 11 12 9 10 2 30 26 4 25 31 5 29 14 3 28 24
173 Token Value 23 10 9 0 27 14 2 24 13 26 30 29 1 15 4 19 22 21 5 20 6 18 17 16 31 28 25 7 8 3 12 11
175 Token Value 3 25 14 6 15 16 18 17 20 19 22 21 8 13 29 4 10 0 9 26 27 12 11 5 7 31 30 1 24 2 23 28
177 Token Value 8 17 15 31 7 25 0 9 10 13 14 28 4 19 20 6 22 21 5 18 16 2 3 27 12 11 1 29 26 24 30 23
179 Token Value 0 4 24 5 1 25 26 31 27 23 10 9 12 11 2 7 30 22 21 8 16 15 14 18 17 20 19 29 6 28 13 3
181 Token Value 12 11 2 26 1 9 10 3 30 14 15 18 6 20 19 5 22 21 17 16 4 25 13 8 0 28 23 29 31 7 27 24
183 Token Value 10 9 0 3 12 11 24 23 27 5 25 6 8 7 22 21 31 30 18 17 20 19 14 13 16 15 26 29 2 28 4 1
185 Token Value 3 18 17 20 19 22 21 6 16 15 14 5 31 7 30 25 12 11 9 10 1 28 0 23 2 4 8 13 29 24 26 27
187 Token Value 23 3 4 1 2 0 24 26 17 16 19 18 13 7 15 14 6 30 21 20 31 22 25 8 27 29 12 11 28 5 10 9
189 Token Value 12 11 25 13 30 7 28 3 24 4 27 23 29 0 1 26 31 5 16 6 18 17 15 20 19 22 21 14 8 2 9 10
191 Token Value 10 9 0 3 12 11 24 23 27 5 25 6 8 7 22 21 31 30 18 17 20 19 14 13 16 15 26 29 2 28 4 1
192
APPENDIX B. VP3
Appendix C
Colophon
Ogg is a Xiph.Org Foundation eort to protect essential tenets of Internet multimedia from corporate hostage-taking; Open Source is the nets greatest tool to keep everyone honest. See About the Xiph.Org Foundation for details. Ogg Theora is the rst Ogg video codec. Anyone may freely use and distribute the Ogg and Theora specications, whether in private, public, or corporate capacity. However, the Xiph.Org Foundation and the Ogg project reserve the right to set the Ogg Theora specication and certify specication compliance. Xiph.Orgs Theora software codec implementation is distributed under a BSD-like license. This does not restrict third parties from distributing independent implementations of Theora software under other licenses. These pages are Copyright 2004-2007 Xiph.Org Foundation. All rights reserved. Ogg, Theora, Vorbis, Xiph.Org Foundation and their logos are trademarks () of the Xiph.Org Foundation. A This document is set in L TEX.
193
194
APPENDIX C. COLOPHON
Bibliography
[Bra97] Scott Bradner. RFC 2119: Key words for use in RFCs to Indicate Requirement Levels, March 1997. http://www.ietf.org/rfc/rfc2119. txt. [CSF77] Wen-Hsiung Chen, C. Harrison Smith, and S. C. Fralick. A fast computational algorithm for the discrete cosine transform. IEEE Transactions on Communications, COM-25(9):10041011, September 1977. [ITU95] International Telecommunications Union, 1211 Geneva 20, Switzerland. Reccomendation ITU-R BT.601-5: Studio Encoding Parameters of Digital Television for Standard 4:3 and Wide-Screen 16:9 Aspect Ratios, 1982, revised edition, 1995. [ITU98] International Telecommunications Union, 1211 Geneva 20, Switzerland. Reccomendation ITU-R BT.470-6: Conventional Television Systems, 1970, revised edition, 1998. [ITU02] International Telecommunications Union, 1211 Geneva 20, Switzerland. Recommendation ITU-R BT.709-5: Parameter values for the HDTV standards for production and international programme exchange, 1990, revised edition, 2002. [Mel04] Mike Melanson. VP3 bitstream format and decoding process. http: //www.multimedia.cx/vp3-format.txt, March 2004. [Pfe03] Silvia Pfeier. RFC 3533: The Ogg Encapsulation Format Version 0, May 2003. http://www.ietf.org/rfc/rfc3533.txt. [Poy97] Charles Poynton. Frequently-asked questions about gamma. http: //www.poynton.com/GammaFAQ.html, February 1997. [SMP94] Society of Motion Pciture and Television Engineers. SMPTE-170M: Television Composite Analog Video Signal NTSC for Studio Applications, 1994. [Xip02] Xiph.Org Foundation. Vorbis I specication, 2002. http://www.xiph. org/ogg/vorbis/doc/. 195
196
BIBLIOGRAPHY
[Yer96] Francois Yergeau. RFC 2044: UTF-8, a transformation format of Unicode and ISO 10646, October 1996. http://www.ietf.org/rfc/ rfc2044.txt.

Theora

Uploaded by

Copyright:

Available Formats

Theora

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Theora

Uploaded by

Copyright:

Available Formats

Theora Specication

Xiph.Org Foundation March 16, 2011

Notation and Conventions

a b Multiplication of a number a by a number b.

Exact division of a number a by a number b, producing a potentially noninteger result.

x a//b Integer division of a by b. a//b =

Notation and Conventions

sign(a) Returns the sign of a given number. 1 0 sign(a) = 1

a<0 a=0 a>0

ilog(1) = 0 ilog(0) = 0 ilog(1) = 1 ilog(2) = 2

Notation and Conventions

Notation and Conventions

VP3 and Theora

Codec Setup and Probability Model

1.6. FORMAT CONFORMANCE

Coded Video Structure

CHAPTER 2. CODED VIDEO STRUCTURE

(0,0) Picture X Offset Picture Width

Figure 2.1: Location of frame and picture regions

2.3. BLOCKS AND SUPER BLOCKS

Blocks and Super Blocks

Super Block (4x4)

CHAPTER 2. CODED VIDEO STRUCTURE

2.4. MACRO BLOCKS

CHAPTER 2. CODED VIDEO STRUCTURE

Frame: chroma plane

Frame: chroma plane

Frame: luma plane

Block Block 8x8 (0,0)

Figure 2.5: Subdivision of a frame into macro blocks

2.5. CODING MODES AND PREDICTION

Coding Modes and Prediction

CHAPTER 2. CODED VIDEO STRUCTURE

Figure 2.8: Zig-zag order

3.2. HIGH-LEVEL DECODE PROCESS

High-Level Decode Process

CHAPTER 3. DECODING OVERVIEW

3.2. HIGH-LEVEL DECODE PROCESS Packet Type Decode

CHAPTER 3. DECODING OVERVIEW

3.2. HIGH-LEVEL DECODE PROCESS

CHAPTER 3. DECODING OVERVIEW

Color Space Conventions

CHAPTER 4. VIDEO FORMATS

Color Space Conversions and Parameters

(4.1) (4.2) (4.3)

(4.4) (4.5) (4.6)

4.2. COLOR SPACE CONVERSIONS AND PARAMETERS

R=R G=G B=B Parameters: . RGB to R G B (Input device gamma correction):

(4.8) (4.9) (4.10)

R = G = B = Parameters: , , , . RGB to CIE XY Z (1931):

(4.11) (4.12) (4.13)

Available Color Spaces

4.3. AVAILABLE COLOR SPACES

Table 4.1: Rec. 470M Parameters

CHAPTER 4. VIDEO FORMATS

Table 4.2: Rec. 470BG Parameters

4.4. PIXEL FORMATS

Frame: chroma plane Cb

Frame: luma planeY