Khronos Data Format Specification License Information
Copyright (C) 2014-2019 The Khronos Group Inc. All Rights Reserved.
This specification is protected by copyright laws and contains material proprietary to the Khronos Group, Inc. It or any components may not be reproduced, republished, distributed, transmitted, displayed, broadcast, or otherwise exploited in any manner without the express prior written permission of Khronos Group. You may use this specification for implementing the functionality therein, without altering or removing any trademark, copyright or other notice from the specification, but the receipt or possession of this specification does not convey any rights to reproduce, disclose, or distribute its contents, or to manufacture, use, or sell anything that it may describe, in whole or in part.
This version of the Data Format Specification is published and copyrighted by Khronos, but is not a Khronos ratified specification. Accordingly, it does not fall within the scope of the Khronos IP policy, except to the extent that sections of it are normatively referenced in ratified Khronos specifications. Such references incorporate the referenced sections into the ratified specifications, and bring those sections into the scope of the policy for those specifications.
Khronos Group grants express permission to any current Promoter, Contributor or Adopter member of Khronos to copy and redistribute UNMODIFIED versions of this specification in any fashion, provided that NO CHARGE is made for the specification and the latest available update of the specification for any version of the API is used whenever possible. Such distributed specification may be reformatted AS LONG AS the contents of the specification are not changed in any way. The specification may be incorporated into a product that is sold as long as such product includes significant independent work developed by the seller. A link to the current version of this specification on the Khronos Group website should be included whenever possible with specification distributions.
Khronos Group makes no, and expressly disclaims any, representations or warranties, express or implied, regarding this specification, including, without limitation, any implied warranties of merchantability or fitness for a particular purpose or non-infringement of any intellectual property. Khronos Group makes no, and expressly disclaims any, warranties, express or implied, regarding the correctness, accuracy, completeness, timeliness, and reliability of the specification. Under no circumstances will the Khronos Group, or any of its Promoters, Contributors or Members or their respective partners, officers, directors, employees, agents, or representatives be liable for any damages, whether direct, indirect, special or consequential damages for lost revenues, lost profits, or otherwise, arising from or in connection with these materials.
Khronos, SYCL, SPIR, WebGL, EGL, COLLADA, StreamInput, OpenVX, OpenKCam, glTF, OpenKODE, OpenVG, OpenWF, OpenSL ES, OpenMAX, OpenMAX AL, OpenMAX IL and OpenMAX DL are trademarks and WebCL is a certification mark of the Khronos Group Inc. OpenCL is a trademark of Apple Inc. and OpenGL and OpenML are registered trademarks and the OpenGL ES and OpenGL SC logos are trademarks of Silicon Graphics International used under license by Khronos. All other product names, trademarks, and/or company names are used solely for identification and belong to their respective owners.
Revision History | ||
---|---|---|
Revision 0.1 | Jan 2015 | AG |
Initial sharing | ||
Revision 0.2 | Feb 2015 | AG |
Added clarification, tables, examples | ||
Revision 0.3 | Feb 2015 | AG |
Further cleanup | ||
Revision 0.4 | Apr 2015 | AG |
Channel ordering standardized | ||
Revision 0.5 | Apr 2015 | AG |
Typos and clarification | ||
Revision 1.0 rev 1 | May 2015 | AG |
Submission for 1.0 release | ||
Revision 1.0 rev 2 | Jun 2015 | AG |
Clarifications for 1.0 release | ||
Revision 1.0 rev 3 | Jul 2015 | AG |
Added KHR_DF_SAMPLE_DATATYPE_LINEAR | ||
Revision 1.0 rev 4 | Jul 2015 | AG |
Clarified KHR_DF_SAMPLE_DATATYPE_LINEAR | ||
Revision 1.0 rev 5 | Mar 2019 | AG |
Clarification and typography | ||
Revision 1.1 rev 1 | Nov 2015 | AG |
Added definitions of compressed texture formats | ||
Revision 1.1 rev 2 | Jan 2016 | AG |
Added definitions of floating point formats | ||
Revision 1.1 rev 3 | Feb 2016 | AG |
Fixed typo in sRGB conversion (thank you, Tom Grim!) | ||
Revision 1.1 rev 4 | Mar 2016 | AG |
Fixed typo/clarified sRGB in ASTC, typographical improvements | ||
Revision 1.1 rev 5 | Mar 2016 | AG |
Switch to official Khronos logo, removed scripts, restored title | ||
Revision 1.1 rev 6 | Jun 2016 | AG |
ASTC block footprint note, fixed credits/changelog/contents | ||
Revision 1.1 rev 7 | Sep 2016 | AG |
ASTC multi-point part and quint decode typo fixes | ||
Revision 1.1 rev 8 | Jun 2017 | AG |
ETC2 legibility and table typo fix | ||
Revision 1.1 rev 9 | Mar 2019 | AG |
Typo fixes and much reformatting | ||
Revision 1.2 rev 0 | Sep 2017 | AG |
Added color conversion formulae and extra options | ||
Revision 1.2 rev 1 | Mar 2019 | AG |
Typo fixes and much reformatting | ||
Revision 1.3 | Oct 2019 | AG/MC |
Updates for KTX2/glTF. BC6h and ASTC table fixes and typo fixes. More examples. | ||
Revision 1.3.1 | Mar 2020 | MC |
Fix swapped pre-multiplied alpha in DXT[2345] examples. Fix incorrect positions in BC7 bit layout table. Rewrite Unsized formats section. |
Table of Contents
Abstract
This document describes a data format specification for non-opaque (user-visible) representations of user data to be used by, and shared between, Khronos standards. The intent of this specification is to avoid replication of incompatible format descriptions between standards and to provide a definitive mechanism for describing data that avoids excluding useful information that may be ignored by other standards. Other APIs are expected to map internal formats to this standard scheme, allowing formats to be shared and compared. This document also acts as a reference for the memory layout of a number of common compressed texture formats, and describes conversion between a number of common color spaces.
Many APIs operate on bulk data — buffers, images, volumes, etc. — each composed of many elements with a fixed and often simple representation. Frequently, multiple alternative representations of data are supported: vertices can be represented with different numbers of dimensions, textures may have different bit depths and channel orders, and so on. Sometimes the representation of the data is highly specific to the application, but there are many types of data that are common to multiple APIs — and these can reasonably be described in a portable manner. In this standard, the term data format describes the representation of data.
It is typical for each API to define its own enumeration of the data formats on which it can operate. This causes a problem when multiple APIs are in use: the representations are likely to be incompatible, even where the capabilities intersect. When additional format-specific capabilities are added to an API which was designed without them, the description of the data representation often becomes inconsistent and disjoint. Concepts that are unimportant to the core design of an API may be represented simplistically or inaccurately, which can be a problem as the API is enhanced or when data is shared.
Some APIs do not have a strict definition of how to interpret their data. For example, a rendering API may treat all color channels of a texture identically, leaving the interpretation of each channel to the user’s choice of convention. This may be true even if color channels are given names that are associated with actual colors — in some APIs, nothing stops the user from storing the blue quantity in the red channel and the red quantity in the blue channel. Without enforcing a single data interpretation on such APIs, it is nonetheless often useful to offer a clear definition of the color interpretation convention that is in force, both for code maintenance and for communication with external APIs which do have a defined interpretation. Should the user wish to use an unconventional interpretation of the data, an appropriate descriptor can be defined that is specific to this choice, in order to simplify automated interpretation of the chosen representation and to provide concise documentation.
Where multiple APIs are in use, relying on an API-specific representation as an intermediary can cause loss of important information. For example, a camera API may associate color space information with a captured image, and a printer API may be able to operate with that color space, but if the data is passed through an intermediate compute API for processing and that API has no concept of a color space, the useful information may be discarded.
The intent of this standard is to provide a common, consistent, machine-readable way to describe those data formats which are amenable to non-proprietary representation. This standard provides a portable means of storing the most common descriptive information associated with data formats, and an extension mechanism that can be used when this common functionality must be supplemented.
While this standard is intended to support the description of many kinds of data, the most common class of bulk data used in Khronos standards represents color information. For this reason, the range of standard color representations used in Khronos standards is diverse, and a significant portion of this specification is devoted to color formats.
Later sections describe some of the common color space conversion operations and provide a description of the memory layout of a number of common texture compression formats.
This document describes a standard layout for a data structure that can be used to define the representation of simple, portable, bulk data. Using such a data structure has the following benefits:
The “bulk data” may be, for example:
The layout of proprietary data structures is beyond the remit of this specification, but the large number of ways to describe colors, vertices and other repeated data makes standardization useful. The widest variety of standard representations and the most common expected use of this API is to describe pixels or texels; as such the terms “texel” and “pixel” are used interchangeably in this specification when referring to elements of data, without intending to imply a restriction in use.
The data structure in this specification describes the elements in the bulk data in memory, not the layout of the whole. For example, it may describe the size, location and interpretation of color channels within a pixel, but is not responsible for determining the mapping between spatial coordinates and the location of pixels in memory. That is, two textures which share the same pixel layout can share the same descriptor as defined in this specification, but may have different sizes, line or plane strides, tiling or dimensionality; in common parlance, two images that describe (for example) color data in the same way but which are of different shapes or sizes are still described as having the same “format”.
An example pixel representation is described in Figure 1: a single 5:6:5-bit pixel composed of a blue channel in the low 5 bits, a green channel in the next 6 bits, and red channel in the top 5 bits of a 16-bit word as laid out in memory on a little-endian machine (see Table 89).
In bulk data, each element is interpreted first by addressing it in some form, then by interpreting the addressed values. Texels often represent a color (or other data) as a multi-dimensional set of values, each representing a channel. The bulk-data image or buffer then describes a number of these texels. Taking the simplest case of an array in the C programming language as an example, a developer might define the following structure to represent a color texel:
typedef struct _MyRGB { unsigned char red; unsigned char green; unsigned char blue; } MyRGB; MyRGB *myRGBarray = (MyRGB *) malloc(100 * sizeof(MyRGB));
To determine the location of, for example, the tenth element of
myRGBarray
, the compiler needs to know the base address of
the array and sizeof myRGB
.
Extracting the red
, green
and blue
components of
myRGBarray[9]
given its base address is, in a sense, orthogonal
to finding the base address of myRGBarray[9]
.
Note also that sizeof(MyRGB)
will often exceed the total size
of red
, green
and blue
due to padding; the difference
in address between one MyRGB
and the next can be described as
the pixel stride in bytes.
An alternative representation is a “structure of arrays”,
distinct from the “array of structures” myRGBarray
:
typedef struct _MyRGBSoA { unsigned char *red; unsigned char *green; unsigned char *blue; } MyRGBSoA; MyRGBSoA myRGBSoA; myRGBSoA.red = (unsigned char *) malloc(100); myRGBSoA.green = (unsigned char *) malloc(100); myRGBSoA.blue = (unsigned char *) malloc(100);
In this case, accessing a value requires the sizeof
each channel
element.
The best approach depends on the operations performed: calculations
on one whole MyRGB
a time likely favor MyRGB
, those processing
multiple values from a single channel may prefer MyRGBSoA
.
A “pixel” need not fill an entire byte — nor need pixel stride
be a whole number of bytes.
For example, a C++ std::vector<bool>
can be considered to be
a 1-D bulk data structure of individual bits.
The simplest way to represent two-dimensional data is in consecutive rows, each representing a one-dimensional array — as with a 2-D array in C. There may be padding after each row to achieve the required alignment: in some cases each row should begin at the start of a cache line, or rows may be deliberately offset to different cache lines to ensure that vertically-adjacent values can be cached concurrently. The offset from the start of one horizontal row to the next is a line stride or row stride (or just stride for short), and is necessarily at least the width of the row. If each row holds an whole number of pixels, row stride can be described either in bytes or pixels; it is rare not to start each row on a byte boundary. In a simple 2-D representation, the row stride and the offset from the start of the storage can be described as follows:
Figure 5 shows example coordinate byte offsets for a 13×4 buffer, padding row stride to a multiple of four elements.
Table 1. Order of byte storage in memory for coordinates in a linear 5×3 buffer, padded (italics) to 8×3
Byte | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 |
Coords | 0,0 | 1,0 | 2,0 | 3,0 | 4,0 | 5,0 | 6,0 | 7,0 | 0,1 | 1,1 | 2,1 | 3,1 | 4,1 | 5,1 | 6,1 | 7,1 | 0,2 | 1,2 | 2,2 | 3,2 | 4,2 | 5,2 | 6,2 | 7,2 |
By convention, this “linear” layout is in y-major order (with the x axis changing most quickly). This does not necessarily imply that an API which accesses the 2D data will do so in this orientation, and there are image layouts which have x and y swapped. The direction of the axes (particularly y) in relation to the image orientation also varies between representations.
Each contiguous region described in this way can be considered as a plane of data. In the same way that 1-D data can be described in structure-of-arrays form, two-dimensional data can also be implemented as multiple planes.
In early graphics, planes often contributed individual bits to the pixel. This allowed pixels of less than a byte, reducing bandwidth and storage; combining 1-bit planes allowed, for example, 64-color, 6-bit pixels. Storing six bits per pixel consecutively would make the number of memory bytes holding a pixel vary by alignment; a planar layout needs simpler hardware. This hardware could often process a subset of the planes — scrolling text may require blitting just one plane, and changes to the display controller start position for each plane independently provide cheap parallax effects.
In modern architectures, planes are typically byte-aligned and hold separate channels. One motivation is avoiding padding. Some common hardware can only access powers of two bytes for each pixel: for example, texels of three 1-byte channels may need padding to four bytes for alignment restrictions; separate planes of one byte per texel need no padding. Other benefits are supporting different downsampling of each channel (described below), and more efficient independent processing of each channel: many common operations (such as filtering and image compression) treat channels separately.
Note that, once all the data from any contributing channels have been combined, the interpretation of the resulting values is again orthogonal to addressing the data.
In many applications, the simple “linear” layout of memory has disadvantages. The primary concern is that data which is vertically adjacent in two-dimensional space may be widely separated in memory address. Many computer graphics rendering operations involve accessing the frame buffer in an order not aligned with its x-axis; for example, traversing horizontally across the width of a textured triangle while writing to the frame buffer will result in a texture access pattern which depends on the orientation of the texture relative to the triangle and the triangle relative to the frame buffer. If each texel access processed a different cache line, the resulting performance would be heavily compromised. Modern GPUs process multiple texels in parallel, meaning that many nearby texels in different orientations may need to be accessed quickly. Additionally, texture filtering operations typically read a 2×2 quad of texels, inherently requiring access to texels which would be distant in memory in the linear representation.
One solution to this is to divide the image into smaller rectangular tiles of texels (of tw × th) rather than horizontal rows. The texel ordering within the tiles may be treated as though each tile were an independent, smaller image, and the order of tiles in memory may be as in the linear layout:
This approach can be implemented efficiently in hardware, so long as
the tile dimensions are powers of two, by interleaving some bits from
the y-coordinate with the bits contributed by the x-coordinate.
If twb
= log2(tw) and thb
= log2(th):
pixelOffset = (x & ((1 << twb) - 1)) | ((y & ((1 << thb) - 1)) << twb) | ((x & ~((1 << twb) - 1)) << thb) | ((y & ~((1 << thb)-1)) * lineStride);
For example, if a linear 16×12 image calculates a pixel offset relative to the start of the buffer as:
pixelOffset = x | (y * 16);
a 16×12 image comprised of 4×4 tiles may calculate the pixel offset relative to the start of the buffer as:
pixelOffset = (x & 3) | ((y & 3) << 2) | ((x & ~3) << 2) | ((y & ~3) * 16);
Figure 8. 4×4 tiled order texel offsets from coordinates (consecutive values linked in red to show the pattern)
Textures which have dimensions that are not a multiple of the tile size require padding.
For so long as the size of a tile fits into an on-chip cache and can be filled efficiently by the memory subsystem, this approach has the benefit that only one in $1\over n$ transitions between vertically-adjacent lines requires accessing outside a tile of height n. The larger the tile, the greater the probability that vertically-adjacent accesses fall inside the same scan line. However, horizontally-adjacent texels that cross a tile boundary are made more distant in memory the larger the tile size, and tile sizes which do not conveniently fit the cache and bus sizes of the memory subsystem have inefficiencies; thus the tile size is an architectural trade-off.
While any non-linear representation is typically referred to as “tiling”, some hardware implementations actually use a more complex layout in order to provide further locality of reference. One such scheme is Morton order:
uint32_t mortonOffset(uint32_t x, uint32_t y, uint32_t width, uint32_t height) { const uint32_t minDim = (width <= height) ? width : height; uint32_t offset = 0, shift = 0, mask; // Tests xy bounds AND that width and height != 0 assert(x < width && y < height); // Smaller size must be a power of 2 assert((minDim & (minDim - 1)) == 0); // Larger size must be a multiple of the smaller assert(width % minDim == 0 && height % minDim == 0); for (mask = 1; mask < minDim; mask <<= 1) { offset |= (((y & mask) << 1) | (x & mask)) << shift; shift++; } // At least one of width and height will have run out of most-significant bits offset |= ((x | y) >> shift) << (shift * 2); return offset; }
Note that this particular implementation assumes that the smaller
of width
and height
is a power of two, and the larger is
a multiple of the smaller.
A practical hardware implementation of the calculation is likely much
more efficient than this C code would imply.
Another approach, with even more correlation between locality in x,y space and memory offset, is Hilbert curve order:
uint32_t h(uint32_t size, uint32_t x, uint32_t y) { const uint32_t z = x ^ y; uint32_t offset = 0; while (size >>= 1) { // Iterate in decreasing block size order // Accumulate preceding blocks of size * size offset += size * (((size & x) << 1) + (size & z)); y = z ^ x; // Transpose x and y if (!(size & y)) x = (size & x) ? ~y : y; // Conditionally swap/mirror } return offset; } uint32_t hilbertOffset(uint32_t x, uint32_t y, uint32_t width, uint32_t height) { const uint32_t minDim = (width <= height) ? width : height; // Tests xy bounds AND that width and height != 0 assert(x < width && y < height); // Smaller size must be a power of 2 assert((minDim & (minDim - 1)) == 0); // Larger size must be a multiple of the smaller assert(width % minDim == 0 && height % minDim == 0); if (width < height) return (width * width) * (y / width) + h(width, y % width, x); else return (height * height) * (x / height) + h(height, x % height, y); }
This code assumes that the smaller of width
and height
is a power of two, with the larger a multiple of the smaller.
Some implementations will mix these approaches — for example, having linear tiles arranged in Morton order, or Morton order texels within tiles which are themselves in linear order. Indeed, for non-square areas, the tiling, Morton and Hilbert orderings shown here can be considered as a linear row of square areas with edges of the shorter dimension.
In all these cases, the relationship between coordinates and the memory storage location of the buffer depends on the image size and, other than needing to know the amount of memory occupied by a single texel, is orthogonal to the “format”. Tiling schemes tend to be complemented by proprietary hardware that performs the coordinate-to-address mapping and depends on cache details, so many APIs expose only a linear layout to end-users, keeping the tiled representation opaque. The variety of possible mappings between coordinates and addresses mandates leaving the calculation to the application.
These storage approaches can be extended to additional dimensions, for example treating a 3-D image as being made up of 2-D “planes”, which are distinct from the concept of storing channels or bits of data in separate planes. In a linear layout, 3-dimensional textures can be accessed as:
Here, the plane stride is the offset between the start of one contiguous 2-D plane of data and the next. The plane stride is therefore height × row stride plus any additional padding.
Again, non-linear approaches can be used to increase the correlation between storage address and coordinates.
uint32_t tileOffset3D(uint32_t x, uint32_t y, uint32_t z, uint32_t twb, uint32_t thb, uint32_t tdb, uint32_t lineStride, uint32_t planeStride) { // twb = tile width bits (log2 of tile width) // thb = tile height bits (log2 of tile height) // tdb = tile depth bits (log2 of tile depth) return (x & ((1 << twb) - 1)) | ((y & ((1 << thb) - 1)) << twb) | ((z & ((1 << tdb) - 1)) << (twb + thb)) | ((x & ~((1 << twb) - 1)) << (thb + tdb)) | ((y & ~((1 << thb) - 1)) << tdb) * lineStride | ((z & ~((1 << tdb) - 1)) * planeStride); }
uint32_t mortonOffset3D(uint32_t x, uint32_t y, uint32_t z, uint32_t width, uint32_t height, uint32_t depth) { const uint32_t max = width | height | depth; uint32_t offset = 0, shift = 0, mask; for (mask = 1; max > mask; mask <<= 1) { if (width > mask) offset |= (x & mask) << shift++; if (height > mask) offset |= (y & mask) << shift++; if (depth > mask) offset |= (z & mask) << shift++; --shift; } return offset; }
There are multiple 3D variations on the Hilbert curve; one such is this:
uint32_t hilbertOffset3D(uint32_t x, uint32_t y, uint32_t z, uint32_t size) { // "Harmonious" 3D Hilbert curve for cube of power-of-two edge "size": // http://herman.haverkort.net/recursive_tilings_and_space-filling_curves uint32_t offset = 0; while (size >>= 1) { uint32_t tx = (size & x) > 0, ty = (size & y) > 0, tz = (size & z) > 0; switch (tx + 2 * ty + 4 * tz) { case 0: tx = z; ty = y; tz = x; break; case 1: tx = x; ty = z; tz = y; offset += size * size * size; break; case 2: tx = ~y; ty = ~x; tz = z; offset += size * size * size * 3; break; case 3: tx = z; ty = x; tz = y; offset += size * size * size * 2; break; case 4: tx = ~z; ty = y; tz = ~x; offset += size * size * size * 7; break; case 5: tx = x; ty = ~z; tz = ~y; offset += size * size * size * 6; break; case 6: tx = ~y; ty = ~x; tz = z; offset += size * size * size * 4; break; case 7: tx = ~z; ty = x; tz = ~y; offset += size * size * size * 5; break; } x = tx; y = ty; z = tz; } return offset; }
The examples provided so far have assumed that a unique value from each color channel is present at each access coordinate. However, some common representations break this assumption.
One reason for this variation comes from representing the image in the Y′CBCR color model, described in Section 15.1; in this description, the Y′ channel represents the intensity of light, with CB and CR channels describing how the color differs from a gray value of the same intensity in the blue and red axes. Since the human eye is more sensitive to high spatial frequencies in brightness than in the hue of a color, the CB and CR channels can be recorded at lower spatial resolution than the Y′ channel with little loss in visual quality, saving storage space and bandwidth.
In one common representation known colloquially as YUV 4:2:2, each horizontal pair of Y′ values has a single CB and CR value shared for the pair. For example, Figure 11 shows a 6-element 1-D buffer with one Y′ value for each element, but as shown in Figure 12 the CB and CR values are shared across pairs of elements.
In this case, we can say that the 2×1 coordinate region forms a texel block which contains two Y′ values, one CB value and one CR value; our bulk-data buffer or image is composed of a repeating pattern of these texel blocks.
Note that this representation raises a question: while we have assumed so far that accessing a value at texel coordinates will provide the value contained in the texel, how should the shared CB and CR values relate to the coordinates of the Y′ channel? Each texel coordinate represents a value in the coordinate space of the image or buffer, which can be considered as a sample of the value of a continuous surface at that location. The preferred means of reconstructing this surface is left to the application: since this specification only defines the values in the image and not how they are used, it is concerned with the sample values rather than the reconstruction algorithm or means of access. For example, in graphics APIs the coordinates used to access a 2-D texture may offset the sample locations of a texture by half a coordinate relative to the origin of sample space, and filtering between samples is typically used to implement antialiasing. However, to interpret the data correctly, any application still needs to know the theoretical location associated with the samples, so that information is within the remit of this specification.
Our Y′ samples should fall naturally on our native coordinates. However, the CB and CR sample locations (which are typically at the same location as each other) could be considered as located coincident with one or other of the Y′ values as shown in Figure 13, or could be defined as falling at the mid-point between them as in Figure 14. Different representations have chosen either of these alternatives — in some cases, choosing a different option for each coordinate axis. The application can choose how to treat these sample locations: in some cases it may suffice to duplicate CB and CR across the pair of Y′ values, in others bilinear or bicubic filtering may be more appropriate.
Traditional APIs have described the CB and CR as having 2×1 downsampling in this format (there are half as many samples available in the horizontal axis for these channels).
This concept can be extended to more dimensions. Commonly, a two-dimensional image stored in Y′CBCR format may store the CB and CR channels downsampled by a factor of two in each dimension (“2×2 downsampling”, also known for historical reasons as “4:2:0”). This approach is used in, for example, many JPEG images and MPEG video streams.
Because there are twice as many rows of Y′ data as there are CB and CR data, it is convenient to record the channels as separate planes as shown in Figure 15 (with 2×2 texel blocks outlined in red); in addition, image compression schemes often work with the channels independently, which is amenable to planar storage.
In this case, the CB and CR planes are half the width and half the height of the Y′ plane, and also have half the line stride. Therefore if we store one byte per channel, the offsets for each plane in a linear representation can be calculated as:
A description based on downsampling factors is sufficient in the case of common Y′CBCR layouts, but does not extend conveniently to all representations. For example, one common representation used in camera sensors is a Bayer pattern, in which there is only one of the red, green and blue channels at each sample location: one red and blue value per 2×2 texel block, and two green values offset diagonally, as in Figure 16.
A Bayer pattern can then be used to sample an image, as shown in Figure 17, and this sampled version can later be used to reconstruct the original image by relying on heuristic correlations between the channels. Technology for image processing continues to develop, so in many cases it is valuable to record the “raw” sensor data for later processing, and to pass the raw data unmodified to a range of algorithms; the choice of algorithm for reconstructing an image from samples is beyond the remit of this specification.
In the Bayer representation, the red and blue channels can be considered to be downsampled by a factor of two in each dimension. The two green channels per 2×2 repeating block mean that the “downsampling factor” for the green channel is effectively $\sqrt{2}$ in each direction.
More complex layouts are not uncommon. For example, the X-Trans sample arrangement developed by Fujifilm for their digital camera range, shown in Figure 18, consists of 6×6 texel blocks, with each sample, as in Bayer, corresponding to a single channel. In X-Trans, each block contains eight red, eight blue and twenty green samples; red and blue are “downsampled” by $\sqrt{3\over 2}$ and green is “downsampled” by $3\over\sqrt{5}$ .
Allowing for possible alternative orientations of the samples (such as whether the Bayer pattern starts with a row containing red or blue samples, and whether the first sample is green or red/blue), trying to encode these sample patterns implicitly is difficult.
The data structure defined in this specification is termed a data format descriptor. This is an extensible block of contiguous memory, with a defined layout. The size of the data format descriptor depends on its content, but is also stored in a field at the start of the descriptor, making it possible to copy the data structure without needing to interpret all possible contents.
The data format descriptor is divided into one or more descriptor blocks, each also consisting of contiguous data, as shown in Table 2. These descriptor blocks may, themselves, be of different sizes, depending on the data contained within. The size of a descriptor block is stored as part of its data structure, allowing an application to process a data format descriptor while skipping contained descriptor blocks that it does not need to understand. The data format descriptor mechanism is extensible by the addition of new descriptor blocks.
Table 2. Data format descriptor and descriptor blocks
Data format descriptor
|
The diversity of possible data makes a concise description that can support every possible format impractical. This document describes one type of descriptor block, a basic descriptor block, that is expected to be the first descriptor block inside the data format descriptor where present, and which is sufficient for a large number of common formats, particularly for pixels. Formats which cannot be described within this scheme can use additional descriptor blocks of other types as necessary.
Later sections of this specification describe several common color spaces, and provide a description of the in-memory representation of a number of common compressed texture formats.
Instead of an implicit representation, the descriptor specifies a texel block, which is described in terms of its finite extent in each of four dimensions, and which identifies the location of each sample within that region. The texel block is therefore a self-contained, repeating, axis-aligned pattern in the coordinate space of the image. This description conveniently corresponds to the concept of compressed texel blocks used in common block-based texture compression schemes, which similarly divide the image into (typically) small rectangular regions which are encoded collectively. The bounds of the texel block are chosen to be aligned to integer coordinates.
Although it is most common to consider one- and two-dimensional textures and texel blocks, it can be convenient to record additional dimensions; for example, the ASTC compressed texture format described in Section 23 can support compressed texel blocks with three dimensions. For convenience of encoding, this specification uses a four-dimensional space to define sample locations within a texel block — there is no meaning imposed on how those dimensions should be interpreted.
In many formats, all color channels refer to the same location, and the texel block defines a 1×1×1×1 region — that is, a single texel.
Tiling schemes for texel addressing can also be seen to break the image into rectangular (or higher-dimensional) sub-regions, and in many schemes these sub-regions are repeating and axis-aligned. Within reason, it is possible to define some coordinate-to-texel mapping in terms of a texel block; for example, instead of a simple tiled layout of 4×4 texels, it would be possible to describe the image in terms of a linear pattern of texel blocks, each of which contain 4×4 samples. In general, this is not a useful approach: it is very verbose to list each sample individually, it does not extend well to larger block sizes (or to infinite ranges in approaches such as Morton order), it does not handle special cases well (such as the “tail” of a mip chain), and encodings such as Hilbert order do not have a repeating mapping. In most contexts where these concepts exist, tiling is not considered to be part of a “format”. |
The description above has shown that the data contributing to a texel may be stored in separate locations in memory; for example, R, G and B stored in separate planes may need to be combined to produce a single texel. For the purposes of this specification, a plane is defined to be a contiguous region of bytes in memory that contributes to a single texel block.
This interpretation contradicts the traditional interpretation of downsampled channels: if two rows of Y′ data correspond to a single row of CB and CR (in a linear, non-tiled memory layout), the Y′ channel contribution to the texel block is not “a contiguous region of bytes”. Instead, each row of Y′ contributing to a texel block can be treated as a separate “plane”.
In linear layout, this can be represented by offsetting rows of Y′ data with odd y coordinates by the row stride of the original Y′ plane; each new Y′ plane’s stride is then double that of the original Y′ plane, as in Figure 19 (c.f. Figure 15). If the planes of a 6×4 Y′CBCR 4:2:0 texture are stored consecutively in memory with no padding, which might be described in a traditional API as Table 3, the representation used by this specification would be that shown in Table 4.
Table 3. Plane descriptors of a 6×4 Y′CBCR-format buffer in a conventional API
Plane | Byte offset | Byte stride | Downsample factor |
---|---|---|---|
Y′ | 0 | 6 | 1×1 |
CB | 24 | 3 | 2×2 |
CR | 30 | 3 | 2×2 |
Table 4. Plane descriptors of a 6×4 Y′CBCR-format buffer using this standard
Plane | Byte offset | Byte stride | Bytes per plane |
---|---|---|---|
Y′ plane 1 | 0 | 12 | 2 |
Y′ plane 2 | 6 | 12 | 2 |
CB | 24 | 3 | 1 |
CR | 30 | 3 | 1 |
There is no expectation that an API must actually use this representation in accessing the data: it is simple for an API with explicit support for the special case of integer chroma downsampling to detect interleaved planes and to deduce that they should be treated as a single plane of double vertical resolution. Many APIs will not support the full flexibility of formats supported by this specification, and will map to a more restrictive internal representation. |
The Khronos Basic Descriptor Block indicates a number of bytes contributing to the texel block from each of up to eight planes — if more than eight planes are required, an extension is needed. Eight planes are enough to encode, for example:
If a plane contributes less than a byte to a texel (or a non-integer number of bytes), the texel block size should be expanded to cover more texels until a whole number of bytes are covered — q.v. Table 92.
Interlaced Y′CBCR data may associate chroma channels with Y′ samples only from the same field, not the frame as a whole. This distinction is not made explicit, in part because a number of interlacing schemes exist. One suggested convention for interlaced data is that the field of a sample be encoded in the fourth sample coordinate (the first field as samplePosition3 = 0, the second field as samplePosition3 = 1, etc.) This interpretation is not mandated to allow other reasons for encoding four-dimensional texels, although it is consistent with the fourth dimension representing “time”. |
For the purposes of description, the bytes contributed by each plane are considered to be concatenated into a single contiguous logical bit stream. This “concatenation” of bits is purely conceptual for the purposes of determining the interpretation of the bits that contribute to the texel block, and is not expected to be the way that actual decoders would process the data.
That is, if planes zero and one contribute two bytes of data each and planes two and three contribute one byte of data each, this bit stream would consist of the two bytes from plane zero followed by the two bytes from plane one, then the byte from plane two, then the byte from plane three.
The data format descriptor then describes a series of samples that are contained within the texel block. Each sample represents a series of contiguous bits from the bit stream, interpreted in little-endian order, and is associated with a single channel and four-dimensional coordinate offset within the texel block; in formats for which only a single texel is being described, this coordinate offset will always be 0,0,0,0.
The descriptor block for a Y′CBCR 4:2:0 layout is shown in Table 96. Figure 20 shows this representation graphically: the three disjoint regions for each channels and the texel block covering them are shown on the left, represented in this specification as four planes of contiguous bytes. These are concatenated into a 48-bit logical bit stream (shown in blue, in top-to-bottom order); these samples then describe the contributions from these logical bits, with geometric location of the sample at the right.
Consecutive samples with the same channel, location and flags may describe additional bits that contribute to the same final value, and appear in increasing order. For example, a 16-bit big-endian word in memory can be described by one sample describing bits 0..7, coming from the higher byte address, followed by one sample describing bits 8..15, coming from the lower address. These samples comprise a single virtual sample, of 16 bits.
Figure 21 shows how multiple contributions to single value can be used to represent a channel which, due to a big-endian native word type (high bits stored in lower byte addresses), is not contiguous in a little-endian representation (low bits stored in lower byte addresses). Here, channels are comprised of consecutive bits from a 16-bit big-endian word; the bits of each channel cease to be contiguous when the memory bytes are interpreted in little-endian order for the logical bit stream. The bit contributions to each channel from each bit location are shown in superscript, as later in this specification; the channel contributions start at bit 0 with the first sample contributing to the value, and are deduced implicitly from the number of bits each sample contributes. This example assumes that we are describing a single texel, with the same 0,0,0,0 coordinates (not shown) for each sample. The descriptor block for this format is shown in Table 97.
In Figure 21, bit 0 of the logical bit stream corresponds to bit 3 of the virtual sample that describes the green channel; therefore the first channel to be encoded is green. Bits 0 to 2 of the green channel virtual sample correspond to bits 13..15 of the logical bit stream, so these are described by the first sample. The second sample continues describing bits 3..5 of the green channel virtual sample, in bits 0..2 of the logical bit stream. The next bit from the logical bit stream that has not yet been described is bit 3, which corresponds to bit 0 of the red channel. The third sample is therefore used to describe bits 3..7 of the logical bit stream as a 5-bit red channel. The last sample encodes the remaining 5-bit blue channel that forms bits 8..12 of the logical bit stream. Note that in the case where some bits of the data are ignored, they do not need to be covered by samples; bits may also appear repeatedly if they contribute to multiple samples.
The precision to which sample coordinates are specified depends on the size of the texel block: coordinates in a 256×256 texel block can be specified only to integer-coordinate granularity, but offsets within a texel block that is only a single coordinate wide are specified to the precision of $1\over 256$ of a coordinate; this approach allows large texel blocks, half-texel offsets for representations such as Y′CBCR 4:2:0, and precise fractional offsets for recording multisampled pattern locations.
The sequence of bits in the (virtual) sample defines a numerical range which may be interpreted as a fixed-point or floating-point value and signed or unsigned. Many common representations specify this range. For example, 8-bit RGB data may be interpreted in “unsigned normalized” form as meaning “0.0” (zero color contribution) when the channel bits are 0 and “1.0” (maximum color contribution) when the channel is 255. In Y′CBCR “narrow-range” encoding, there is head-room and foot-room to the encoding: “black” is encoded in the Y′ channel as the value 16, and “white” is encoded as 235, as shown in Section 16.1 encoding”. To allow the bit pattern of simply-encoded numerical encodings to be mapped to the desired values, each sample has an upper and lower value associated with it, usually representing 1.0 and either 0.0 or -1.0 depending on whether the sample is signed.
Note that it is typically part of the “format” to indicate the numbers which are being encoded; how the application chooses to process these numbers is not part of the “format”. For example, some APIs may have two separate “formats”, in which the 8-bit pattern 0x01 may be interpreted as either the float value 1.0 or the integer value 1; for the purposes of this specification, these formats are identical — “1” means “1”.
A sample has an associated color channel in the color model of the descriptor block — for example, if the descriptor block indicates an RGB color model, the sample’s channel field could specify the R channel. The descriptor block enumerates a range of common color models, color primary values and transfer functions, which apply to the samples.
There is some redundancy in the data format descriptor when it comes to the ordering of planes and samples. In the interests of simplifying comparison between formats, it is helpful to define a canonical ordering.
Planes should be described first in order of the first channel with which they are associated, then in increasing raster order of coordinates, then in increasing bit order (in a little-endian interpretation). For example:
The order of samples should be defined by the following rules:
Samples sharing a channel and location appear consecutively in increasing order of bits contributed to the virtual sample.
A minimum number of samples describes each sequence of contiguous virtual bit stream bits in a virtual sample value.
Samples that qualify an existing virtual sample immediately follow it.
Virtual samples are described in increasing order of the virtual bit stream bits to which they apply.
If the same bit in the virtual bit stream in increasing bit order is the first to be associated with more than one virtual sample, virtual samples are listed in channel number order.
Finally:
KHR_DF_TRANSFER_LINEAR
for the descriptor
as a whole if all samples are linear, or by
setting the sample’s KHR_DF_SAMPLE_DATATYPE_LINEAR
bit if other samples have multiple bits and a non-linear
transfer function is needed.
The data format descriptor describes only concepts which are conventionally part of the “format”. Additional information is required to access the image data:
For example, if texels are laid out in memory in linear order:
int numPlanes; char *planeBaseAddresses[8]; unsigned int strides[8][4]; // Note: The strides here are assumed to be in units of the // corresponding dimension of the texel block char *planeAddress(uint32_t descriptor, int plane, float coords[4]) { return planeBaseAddresses[plane] // Block stride + ((int) (coords[0] / descriptorSize(descriptor, 0))) * strides[plane][0] // Row stride + ((int) (coords[1] / descriptorSize(descriptor, 1))) * strides[plane][1] // Plane stride + ((int) (coords[2] / descriptorSize(descriptor, 2))) * strides[plane][2] // Volume(?) stride + ((int) (coords[3] / descriptorSize(descriptor, 3))) * strides[plane][3]; } decodeTexelGivenCoords(uint32_t *descriptor, float coords[4]) { char *addresses[8]; for (int i = 0; i < numPlanes; ++i) { addresses[i] = planeAddress(i, coords); } // processTexel reads bytesPlane[] bytes from each address and // decodes the concatenated data according to the descriptor processTexel(descriptor, addresses, coords); }
The processTexel
function would typically operate on the coordinates
having taken the remainder of dividing them by the texel block size.
For example, if the texel block size is 2×2, the block and row
stride provide the offset in bytes between adjacent blocks in the first
two dimensions.
processTexel
would then work with the data corresponding to the
2×2 region with coords[0]
and coords[1]
in the range [0..2).
For formats that describe a single texel, coords
can be considered
to be an integer.
Note that for more complex formats such as Bayer layouts, reconstructing
an R,G,B value at a location may require information from more than
one texel block, depending on the algorithm used, in a manner analogous
to sampling using bilinear filtering.
The stride values may be stored explicitly, or derived implicitly from the bytesPlane values and image size information, with application-specific knowledge about alignment constraints.
Despite being designed to balance size against flexibility, the data format container described here is too unwieldy to be expected to be used directly in most APIs, which will generally support only a subset of possible formats. The expectation is that APIs and users will define data descriptors in memory, but have API-specific names for the formats that the API supports. If these names are enumeration values, a mapping can be provided by having an array of pointers to the data descriptors, indexed by the enumeration. It may commonly be necessary to provide API-specific supplementary information in the same array structure, particularly where the API natively associates concepts with the data which is not uniquely associated with the content.
In this approach, it is likely that an API would predefine a number of common data formats which are natively supported. If there is a desire to support dynamic creation of data formats, this array could be made extensible with a manager assigning unique handles.
Even where an API supports only a fixed set of formats, it is flexible to provide a comparison with user-provided format descriptors in order to establish whether a format is compatible (and differs only in a manner irrelevant to the API).
Some APIs have the concept of a native data type for a format, if each channel is stored separately. Since this specification describes a number of bytes per plane and separately contiguous bit sequences, there is no such explicit concept. However, if a sample’s bitOffset and bitLength are byte-aligned and no further samples contribute to the same value, the bitLength trivially defines a little-endian native data type size. Big-endian data types can be identified by observing that in a big-endian format, a sequence of bits in the top bits of byte n may continue in the low bits of byte n - 1. Finally, “packed” formats consist of consecutive bit sequences per channel in either little- or big-endian order; little-endian sequences are a single stand-alone sample, and a big-endian sequence consists of a number of samples adjacent to byte boundaries in decreasing byte order (see Figure 21); the packed field size can typically be deduced from the bytesPlane0 value. There is no way to distinguish a logically “packed”, byte-aligned samples from distinct but consecutively-stored channels that have the same in-memory representation.
Data format: The interpretation of individual elements in bulk data. Examples include the channel ordering and bit positions in pixel data or the configuration of samples in a Bayer image. The format describes the elements, not the bulk data itself: an image’s size, stride, tiling, dimensionality, border control modes, and image reconstruction filter are not part of the format and are the responsibility of the application.
Data format descriptor: A contiguous block of memory containing information about how data is represented, in accordance with this specification. A data format descriptor is a container, within which can be found one or more descriptor blocks. This specification does not define where or how the the data format descriptor should be stored, only its content. For example, the descriptor may be directly prepended to the bulk data, perhaps as part of a file format header, or the descriptor may be stored in a CPU memory while the bulk data that it describes resides within GPU memory; this choice is application-specific.
(Data format) descriptor block: A contiguous block of memory with a defined layout, held within a data format descriptor. Each descriptor block has a common header that allows applications to identify and skip descriptor blocks that it does not understand, while continuing to process any other descriptor blocks that may be held in the data format descriptor.
Basic (data format) descriptor block: The initial form of descriptor block as described in this standard. Where present, it must be the first descriptor block held in the data format descriptor. This descriptor block can describe a large number of common formats and may be the only type of descriptor block that many portable applications will need to support.
Texel block: The units described by the Basic Data Format Descriptor: a repeating element within bulk data. In simple texture formats, a texel block may describe a single pixel. In formats where the bytes of each plane do not correspond uniquely to single pixels, as for example with subsampled channels, the texel block may cover several pixels. In a block-based compressed texture, the texel block typically describes the compression block unit. The basic descriptor block supports texel blocks of up to four dimensions.
Plane: In the Basic Data Format Descriptor, a plane describes a contiguous sequence of bytes that contribute to the texel block. The basic format descriptor block defines a texel block as being made of a number of concatenated bits which may come from different regions of memory, where each region is considered a separate plane. For common formats, it is sufficient to require that the contribution from each plane is an integer number of bytes. This specification places no requirements on the ordering of planes in memory — the plane locations are described outside the format. This allows support for multiplanar formats which have proprietary padding requirements that are hard to accommodate in a more terse representation.
Sample: In this standard, texel blocks are considered to be composed of contiguous bit patterns with a single channel or component type and a single spatial location. A typical ARGB pixel has four samples, one for each channel, held at the same coordinate. A texel block from a Bayer sensor might have a different location for different channels, and may have multiple samples representing the same channel at multiple locations. A Y′CBCR buffer with downsampled chroma may have more luma samples than chroma, each at different locations.
The data format descriptor consists of a contiguous area of memory, as shown in Table 5, divided into one or more descriptor blocks, which are tagged by the type of descriptor that they contain. The size of the data format descriptor varies according to its content.
Table 5. Data Format Descriptor layout
| totalSize |
Descriptor block | First descriptor |
Descriptor block | Second descriptor (optional) etc. |
The totalSize field, measured in bytes, allows the full format descriptor
to be copied without need for details of the descriptor to be interpreted.
totalSize includes its own uint32_t
, not just the following
descriptor blocks.
For example, we will see below that a four-sample Khronos Basic
Data Format Descriptor Block occupies 88 bytes; if there are no
other descriptor blocks in the data format descriptor, the totalSize
field would then indicate 88 + 4 bytes (for the totalSize field
itself) for a final value of 92.
For consistency of decode, each descriptor block should be aligned to a multiple of four bytes relative to the start of the descriptor; totalSize will therefore be a multiple of four.
This restriction was not present in versions of the Khronos Data Format Specification prior to version 1.3. |
The layout of the data structures described here are comprised solely
of 32-bit words, and for canonical communication between devices are
assumed to be stored with a little-endian representation.
For efficiency, applications may choose to convert the descriptor
to the native endianness of the underlying hardware where all
software using the data structure is prepared for this conversion.
Extensions which are composed of quantities other than 32-bit words
(for example if a data structure belonging to another standard is
incorporated directly) may define the expected impact of endianness
changes on the layout.
Since the environment is expected to know its own endianness, there
is no explicit means of automatically determining the endianness of a
descriptor, although it can be observed that it is highly unlikely
that a valid descriptor would be large enough for its size to need
to be represented in more than 16 bits — meaning that the endianness
of most descriptors can be deduced by which half of the uint32_t
totalSize field is non-zero.
To avoid expanding the size of the data structure, there is no “magic identifier” for a data format descriptor: applications are expected to know the type of the data structure being accessed, and to provide their own means of identifying a data format descriptor if one is embedded in a multi-purpose byte stream. |
Each descriptor block has the same prefix, shown in Table 6.
Table 6. Descriptor Block layout
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
descriptorType | vendorId | ||||||||||||||||||||||||||||||
descriptorBlockSize | versionNumber | ||||||||||||||||||||||||||||||
Format-specific data |
The vendorId is a 17-bit value uniquely assigned to organizations. If the organization has a 16-bit identifier assigned by the PCI SIG, this should occupy bits 0..15, with bit 16 set to 0. Other organizations should apply to Khronos of a unique identifier, which will be assigned consecutively starting with 65536. The identifier 0x1FFFF is reserved for internal use which is guaranteed not to clash with third-party implementations; this identifier should not be shipped in libraries to avoid conflicts with development code.
The descriptorType is a unique 15-bit identifier defined by the vendor to distinguish between potential data representations.
Prior to version 1.3 of the Khronos Data Format Specification, the vendorId field was 16-bit, and purely assigned through the auspices of this specification; the descriptorType was consequently also 16-bit. Since no vendor has requested an identifier and Khronos does not have a descriptor block with type 1, this change should not cause any ambiguity. This change is intended to allow consistency with the vendor IDs used by the Vulkan specification. |
The versionNumber is vendor-defined, and is intended to allow for backwards-compatible updates to existing descriptor blocks.
The DescriptorBlockSize indicates the size in bytes of this Descriptor Block, remembering that there may be multiple Descriptor Blocks within one container, as shown in Table 7. The descriptorBlockSize therefore gives the offset between the start of the current Descriptor Block and the start of the next — so the size includes the vendorId, descriptorType, versionNumber and descriptorBlockSize fields, which collectively contribute 8 bytes.
Having an explicit descriptorBlockSize allows implementations to skip a descriptor block whose format is unknown, allowing known data to be interpreted and unknown information to be ignored. Some descriptor block types may not be of a uniform size, and may vary according to the content within.
This specification initially describes only one type of stand-alone descriptor block, plus two extension blocks which modify the description in the first. Future revisions may define additional descriptor block types for additional applications — for example, to describe data with a large number of channels or pixels described in an arbitrary color space. Vendors can also implement proprietary descriptor blocks to hold vendor-specific information within the standard descriptor.
Unless otherwise specified, descriptor blocks can appear in any order, to make it easier to add and remove additional informative descriptor blocks to a preexisting data format descriptor as part of processing. Descriptor blocks that provide additional capabilities beyond a basic scheme (such as the descriptor block for supporting additional planes beyond the Khronos Basic Descriptor Block) should not be present unless their additional capabilities are needed; that is, redundancy should be resolved so as to minimize the number of descriptor blocks in the data format descriptor.
Table 7. Data format descriptor header and descriptor block headers for two descriptor blocks
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize | |||||||||||||||||||||||||||||||
First descriptor block | |||||||||||||||||||||||||||||||
descriptorType | vendorId | ||||||||||||||||||||||||||||||
descriptorBlockSize | versionNumber | ||||||||||||||||||||||||||||||
Descriptor body | |||||||||||||||||||||||||||||||
Second descriptor block | |||||||||||||||||||||||||||||||
descriptorType | vendorId | ||||||||||||||||||||||||||||||
descriptorBlockSize | versionNumber | ||||||||||||||||||||||||||||||
Descriptor body |
A basic descriptor block (Table 8) is designed to encode common metadata associated with bulk data — especially image or texture data. While this descriptor holds more information about the data interpretation than is needed by many applications, a comprehensive encoding reduces the risk of metadata needed by different APIs being lost in translation.
The format is described in terms of a repeating axis-aligned texel block composed of samples. Each sample contains a single channel of information with a single spatial offset within the texel block, and consists of an amount of contiguous data. This descriptor block consists of information about the interpretation of the texel block as a whole, supplemented by a description of a number of samples taken from one or more planes of contiguous memory. For example, a 24-bit red/green/blue format may be described as a 1×1 pixel region, in one plane of three samples, one describing each channel. A Y′CBCR 4:2:0 format may consist of a repeating 2×2 region: four Y′ samples and one sample each of CB and CR.
Table 8. Basic Data Format Descriptor layout
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
descriptorType = 0 | vendorId = 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize = 24 + 16 × #samples | versionNumber = 2 | ||||||||||||||||||||||||||||||
flags | transferFunction | colorPrimaries | colorModel | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
bytesPlane3 | bytesPlane2 | bytesPlane1 | bytesPlane0 | ||||||||||||||||||||||||||||
bytesPlane7 | bytesPlane6 | bytesPlane5 | bytesPlane4 | ||||||||||||||||||||||||||||
Sample information for the first sample | |||||||||||||||||||||||||||||||
Sample information for the second sample (optional), etc. |
The Basic Data Format Descriptor Block should be the first descriptor block in any data format descriptor of which it is a component.
The fields of the Basic Data Format Descriptor Block are described in the following sections.
To simplify code using the Basic Data Format Descriptor Block, the header
khr_df.h
provides enums of the following form for accessing descriptor
block fields:
Table 9. Field location information for field xxx
Word offset into basic descriptor block |
| … |
Word offset into descriptor |
| … |
Start bit within word |
| … |
Bit mask of value |
| … |
If the basic descriptor block is treated as a uint32_t
array bdb[]
,
field xxx can be accessed as follows:
xxx = KHR_DF_MASK_xxx
& (bdb
[KHR_DF_WORD_xxx
] >> KHR_DF_SHIFT_xxx
);
The macro KHR_DFDVAL
(BDB
, X
) is provided to perform this
calculation.
For example, KHR_DFDVAL
(bdb
, MODEL
) returns the value:
KHR_DF_MASK_MODEL
& (bdb
[KHR_DF_WORD_MODEL
] >> KHR_DF_SHIFT_MODEL
)
The vendorId for the Basic Data Format Descriptor Block is 0,
defined as KHR_DF_VENDORID_KHRONOS
in the enum khr_df_vendorid_e
.
Table 10. Field location information for vendorId
Word offset into basic descriptor block |
| 0 |
Word offset into descriptor |
| 1 |
Start bit within word |
| 0 |
Bit mask of value |
| 0x1FFFFU |
khr_df_vendorid_e
vendorId = KHR_DF_MASK_VENDORID
&
(bdb
[KHR_DF_WORD_VENDORID
] >> KHR_DF_SHIFT_VENDORID
);
The descriptorType for the Basic Data Format Descriptor Block is 0,
a value reserved in the enum of Khronos-specific descriptor types,
khr_df_khr_descriptortype_e
, as KHR_DF_KHR_DESCRIPTORTYPE_BASICFORMAT
.
Table 11. Field location information for descriptorType
Word offset into basic descriptor block |
| 0 |
Word offset into descriptor |
| 1 |
Start bit within word |
| 17 |
Bit mask of value |
| 0x7FFFU |
khr_df_descriptortype_e
descriptorType = KHR_DF_MASK_DESCRIPTORTYPE
&
(bdb
[KHR_DF_WORD_DESCRIPTORTYPE
] >> KHR_DF_SHIFT_DESCRIPTORTYPE
);
The versionNumber relating to the Basic Data Format Descriptor Block as described in this specification is 2.
The versionNumber is incremented to indicate an incompatible change in the descriptor. The addition of enumerant values, for example to represent more compressed texel formats, does not constitute an “incompatible change”, and implementations should be resilient against enumerants that have been added in later minor updates. |
Table 12. Field location information for versionNumber
Word offset into basic descriptor block |
| 1 |
Word offset into descriptor |
| 2 |
Start bit within word |
| 0 |
Bit mask of value |
| 0xFFFFU |
uint32_t
versionNumber = KHR_DF_MASK_VERSIONNUMBER
&
(bdb
[KHR_DF_WORD_VERSIONNUMBER
] >> KHR_DF_SHIFT_VERSIONNUMBER
);
The memory size of the Basic Data Format Descriptor Block depends on the number of samples contained within it. The memory requirements for this format are 24 bytes of shared data plus 16 bytes per sample. The descriptorBlockSize is measured in bytes.
Table 13. Field location information for descriptorBlockSize
Word offset into basic descriptor block |
| 1 |
Word offset into descriptor |
| 2 |
Start bit within word |
| 16 |
Bit mask of value |
| 0xFFFFU |
uint32_t
descriptorBlockSize = KHR_DF_MASK_DESCRIPTORBLOCKSIZE
&
(bdb
[KHR_DF_WORD_DESCRIPTORBLOCKSIZE
] >> KHR_DF_SHIFT_DESCRIPTORBLOCKSIZE
);
The colorModel determines the set of color (or other data) channels which may be encoded within the data, though there is no requirement that all of the possible channels from the colorModel be present. Most data fits into a small number of common color models, but compressed texture formats each have their own color model enumeration. Note that the data need not actually represent a color — this is just the most common type of content using this descriptor. Some standards use color container for this concept.
The available color models are described in the khr_df_model_e
enumeration,
and are represented as an unsigned 8-bit value.
Table 14. Field location information for colorModel
Word offset into basic descriptor block |
| 2 |
Word offset into descriptor |
| 3 |
Start bit within word |
| 0 |
Bit mask of value |
| 0xFF |
khr_df_model_e
colorModel = KHR_DF_MASK_MODEL
&
(bdb
[KHR_DF_WORD_MODEL
] >> KHR_DF_SHIFT_MODEL
);
Note that the numbering of the component channels is chosen such that those channel types which are common across multiple color models have the same enumeration value. That is, alpha is always encoded as channel ID 15, depth is always encoded as channel ID 14, and stencil is always encoded as channel ID 13. Luma/Luminance is always in channel ID 0. This numbering convention is intended to simplify code which can process a range of color models. Note that there is no guarantee that models which do not support these channels will not use this channel ID. Particularly, RGB formats do not have luma in channel 0, and a 16-channel undefined format is not obligated to represent alpha in any way in channel number 15.
The value of each enumerant is shown in parentheses following the enumerant name.
When the data format is unknown or does not fall into a predefined category, utilities which perform automatic conversion based on an interpretation of the data cannot operate on it. This format should be used when there is no expectation of portable interpretation of the data using only the basic descriptor block.
For portability reasons, it is recommended that pixel-like formats
with up to sixteen channels, but which cannot have those channels
described in the basic block, be represented with a basic descriptor
block with the appropriate number of samples from UNSPECIFIED
channels,
and then for the channel description to be stored in an extension block.
This allows software which understands only the basic descriptor to
be able to perform operations that depend only on channel location,
not channel interpretation (such as image cropping). For example, a
camera may store a raw format taken with a modified Bayer sensor,
with RGBW (red, green, blue and white) sensor sites, or RGBE
(red, green, blue and “emerald”). Rather than trying to encode the
exact color coordinates of each sample in the basic descriptor,
these formats could be represented by a four-channel UNSPECIFIED
model, with an extension block describing the interpretation of
each channel.
This color model represents additive colors of three channels, nominally red, green and blue, supplemented by channels for alpha, depth and stencil, as shown in Table 15. Note that in many formats, depth and stencil are stored in a completely independent buffer, but there are formats for which integrating depth and stencil with color data makes sense.
Table 15. Basic Data Format RGBSDA channels
Channel number | Name | Description | |
---|---|---|---|
0 | | Red | |
1 | | Green | |
2 | | Blue | |
13 | | Stencil | |
14 | | Depth | |
15 | | Alpha (opacity) |
Portable representation of additive colors with more than three primaries requires an extension to describe the full color space of the channels present. There is no practical way to do this portably without taking significantly more space.
This color model represents color differences with three channels, nominally luma (Y′) and two color-difference chroma channels, U (CB) and V (CR), supplemented by channels for alpha, depth and stencil, as shown in Table 16. These formats are distinguished by CB and CR being a delta between the Y′ channel and the blue and red channels respectively, rather than requiring a full color matrix. The conversion between Y′CBCR and RGB color spaces is defined in this case by the choice of value in the colorPrimaries field as described in Section 15.1.
Most single-channel luma/luminance monochrome data formats should
select |
Table 16. Basic Data Format YUVSDA channels
Channel number | Name | Description | |
---|---|---|---|
0 | | Y/Y′ (luma/luminance) | |
1 | | CB (alias for U) | |
1 | | U (alias for CB) | |
2 | | CR (alias for V) | |
2 | | V (alias for CR) | |
13 | | Stencil | |
14 | | Depth | |
15 | | Alpha (opacity) |
Terminology for this color model is often abused. This model is based on the idea of creating a representation of monochrome light intensity as a weighted average of color channels, then calculating color differences by subtracting two of the color channels from this monochrome value. Proper names vary for each variant of the ensuing numbers, but YUV is colloquially used for all of them. In the television standards from which this terminology is derived, Y′CBCR is more formally used to describe the representation of these color differences. See Section 15.1 for more detail. |
This color model represents color differences with three channels, nominally luma (Y) and two color-difference chroma channels, I and Q, supplemented by channels for alpha, depth and stencil, as shown in Table 17. This format is distinguished by I and Q each requiring all three additive channels to evaluate. I and Q are derived from CB and CR by a 33-degree rotation.
Table 17. Basic Data Format YIQSDA channels
Channel number | Name | Description | |
---|---|---|---|
0 | | Y (luma) | |
1 | | I (in-phase) | |
2 | | Q (quadrature) | |
13 | | Stencil | |
14 | | Depth | |
15 | | Alpha (opacity) |
This color model represents the ICC perceptually-uniform L*a*b* color space, combined with the option of an alpha channel, as shown in Table 18.
Table 18. Basic Data Format LABSDA channels
Channel number | Name | Description | |
---|---|---|---|
0 | | L* (luma) | |
1 | | a* | |
2 | | b* | |
13 | | Stencil | |
14 | | Depth | |
15 | | Alpha (opacity) |
This color model represents secondary (subtractive) colors and the combined key (black) channel, along with alpha, as shown in Table 19.
Table 19. Basic Data Format CMYKA channels
Channel number | Name | Description | |
---|---|---|---|
0 | | Cyan | |
1 | | Magenta | |
2 | | Yellow | |
3 | | Key/Black | |
15 | | Alpha (opacity) |
This “color model” represents channel data used for coordinate values, as shown in Table 20 — for example, as a representation of the surface normal in a bump map. Additional channels for higher-dimensional coordinates can be used by extending the channel number within the 4-bit limit of the channelType field.
Table 20. Basic Data Format XYZW channels
Channel number | Name | Description | |
---|---|---|---|
0 | | X | |
1 | | Y | |
2 | | Z | |
3 | | W |
This color model represents color differences with three channels, value (luminance or luma), saturation (distance from monochrome) and hue (dominant wavelength), supplemented by an alpha channel, as shown in Table 21. In this model, the hue relates to the angular offset on a color wheel.
Table 21. Basic Data Format angular HSVA channels
Channel number | Name | Description | |
---|---|---|---|
0 | | V (value) | |
1 | | S (saturation) | |
2 | | H (hue) | |
15 | | Alpha (opacity) |
This color model represents color differences with three channels, lightness (maximum intensity), saturation (distance from monochrome) and hue (dominant wavelength), supplemented by an alpha channel, as shown in Table 22. In this model, the hue relates to the angular offset on a color wheel.
Table 22. Basic Data Format angular HSLA channels
Channel number | Name | Description | |
---|---|---|---|
0 | | L (lightness) | |
1 | | S (saturation) | |
2 | | H (hue) | |
15 | | Alpha (opacity) |
This color model represents color differences with three channels, value (luminance or luma), saturation (distance from monochrome) and hue (dominant wavelength), supplemented by an alpha channel, as shown in Table 23. In this model, the hue is generated by interpolation between extremes on a color hexagon.
Table 23. Basic Data Format hexagonal HSVA channels
Channel number | Name | Description | |
---|---|---|---|
0 | | V (value) | |
1 | | S (saturation) | |
2 | | H (hue) | |
15 | | Alpha (opacity) |
This color model represents color differences with three channels, lightness (maximum intensity), saturation (distance from monochrome) and hue (dominant wavelength), supplemented by an alpha channel, as shown in Table 24. In this model, the hue is generated by interpolation between extremes on a color hexagon.
Table 24. Basic Data Format hexagonal HSLA channels
Channel number | Name | Description | |
---|---|---|---|
0 | | L (lightness) | |
1 | | S (saturation) | |
2 | | H (hue) | |
15 | | Alpha (opacity) |
This color model represents low-cost approximate color differences with three channels, nominally luma (Y) and two color-difference chroma channels, Cg (green/purple color difference) and Co (orange/cyan color difference), supplemented by a channel for alpha, as shown in Table 25.
Table 25. Basic Data Format YCoCgA channels
Channel number | Name | Description | |
---|---|---|---|
0 | | Y | |
1 | | Cg | |
2 | | Co | |
15 | | Alpha (opacity) |
This color model represents the “Constant luminance” $Y'_CC'_\mathit{BC}C'_\mathit{RC}$ color model defined as an optional representation in ITU-T BT.2020 and described in Section 15.2.
Table 26. Basic Data Format Y′CC′BCC′RC channels
Channel number | Name | Description | |
---|---|---|---|
0 | | $Y'_C$ (luminance) | |
1 | | $C'_\mathit{BC}$ | |
2 | | $C'_\mathit{RC}$ | |
13 | | Stencil | |
14 | | Depth | |
15 | | Alpha (opacity) |
This color model represents the “Constant intensity ICTCP color model” defined as an optional representation in ITU-T BT.2100 and described in Section 15.3.
Table 27. Basic Data Format ICTCP channels
Channel number | Name | Description | |
---|---|---|---|
0 | | I (intensity) | |
1 | | CT | |
2 | | CP | |
13 | | Stencil | |
14 | | Depth | |
15 | | Alpha (opacity) |
This color model represents channel data used to describe color coordinates in the CIE 1931 XYZ coordinate space, as shown in Table 28.
Table 28. Basic Data Format CIE XYZ channels
Channel number | Name | Description | |
---|---|---|---|
0 | | X | |
1 | | Y | |
2 | | Z |
This color model represents channel data used to describe chromaticity coordinates in the CIE 1931 xyY coordinate space, as shown in Table 29.
Table 29. Basic Data Format CIE xyY channels
Channel number | Name | Description | |
---|---|---|---|
0 | | x | |
1 | | y | |
2 | | Y |
A number of compressed formats are supported as part of khr_df_model_e
.
In general, these formats will have the texel block dimensions of the
compression block size. Most contain a single sample of channel type 0
at offset 0,0 — where further samples are required, they should also
be sited at 0,0. By convention, models which have multiple channels that
are disjoint in memory have these channel locations described
independently as separate samples; this can simplify some decoders.
The ASTC family of formats have a number of possible channels, and are distinguished by samples which reference some set of these channels. The texelBlockDimension fields determine the compression ratio for ASTC and PVRTC.
Compressed formats necessarily do not have an equivalent integer
representation in which to describe the sampleLower and
sampleUpper ranges — in particular, some have different ranges
on a block-by-block format.
Floating-point compressed formats have lower and upper limits specified
in floating point format, since this representation indicates the output
of compressed decoding.
Integer compressed formats with a lower and upper of 0 and UINT32_MAX
(for unsigned formats) or INT32_MIN
and INT32_MAX
(for signed
formats) are assumed to map the full representable range to 0..1 or -1..1
respectively.
If a format has a non-linear transfer function, any samples with channel
ID 15 (that is, the format has separate alpha encoding, for example
KHR_DF_BC2_ALPHA
) should set the KHR_DF_SAMPLE_DATATYPE_LINEAR
bit for that sample.
Example descriptors for compressed formats are provided after each model in this section.
This model represents the DXT1 or BC1 format, described in Section 18. Each compressed texel block consists of 4×4 texels in 8 bytes. A single sample with channel ID 0 indicates that the “special value” should be interpreted as black, as described in Section 18.1 — a descriptor block representing this is shown in Table 31. A single sample with channel ID 1 indicates that the “special value” should represent transparency, as described in Section 18.2 — a descriptor block representing this is shown in Table 32.
Enumerant names for these channel ids are listed in Table 30.
Table 30. BC1A channel names
Enumerant | Value |
---|---|
| 0 |
| |
| 1 |
| |
| |
|
Table 31. Example DXT1A descriptor with no punch-through alpha
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 44 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 1) = 40 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 8 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: |
Table 32. Example DXT1A descriptor with punch-through alpha
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 44 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 1) = 40 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 8 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: |
This model represents the DXT2/3 format, also known as BC2, and described in Section 18.3. Each compressed texel block consists of 4×4 texels in 16 bytes. The alpha premultiplication state (the distinction between DXT2 and DXT3) is recorded separately in the descriptor in the flags field. This model has two channels, recorded as separate samples: Sample 0 with channel ID 15 contains the alpha information. Sample 1 with channel ID 0 contains the color information.
Enumerant names for these channel ids are listed in Table 33.
Table 33. BC2 channel names
Enumerant | Value |
---|---|
| 0 |
| |
| |
| 15 |
| |
|
The alpha channel is 64 bits and at offset 0; the color channel is 64 bits and at offset 64. No attempt is made to describe the 16 alpha samples for this position independently, since understanding the other channels for any pixel requires the whole texel block.
Table 34. Example DXT2 descriptor (premultiplied alpha)
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 60 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 2) = 56 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 16 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Alpha sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Color sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 64 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: |
Table 35. Example DXT3 descriptor
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 60 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 2) = 56 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 16 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Alpha sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Color sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 64 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: |
This model represents the DXT4/5 format, also known as BC3, and described in Section 18.4. Each compressed texel block consists of 4×4 texels in 16 bytes. The alpha premultiplication state (the distinction between DXT4 and DXT5) is recorded separately in the descriptor in the flags field. This model has two channels, recorded as separate samples: Sample 0 with channel ID 15 contains the alpha information. Sample 1 with channel ID 0 contains the color information.
Enumerant names for these channel ids are listed in Table 36.
Table 36. BC3 channel names
Enumerant | Value |
---|---|
| 0 |
| |
| |
| 15 |
| |
|
The alpha channel is 64 bits and at offset 0; the color channel is 64 bits and at offset 64. No attempt is made to describe the 16 alpha samples for this position independently, since understanding the other channels for any pixel requires the whole texel block.
Table 37. Example DXT4 descriptor (premultiplied alpha)
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 60 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 2) = 56 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 16 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Alpha sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Color sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 64 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: |
Table 38. Example DXT5 descriptor
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 60 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 2) = 56 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 16 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Alpha sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Color sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 64 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: |
This model represents the Direct3D BC4 format for single-channel interpolated 8-bit data, as described in Section 19.1.
Each compressed texel block consists of 4×4 texels in 8 bytes. The model has a single channel of id 0 with offset 0 and length 64 bits.
The enumerant name for this channel id is listed in Table 39.
Table 40. Example BC4 unsigned descriptor
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 44 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 1) = 40 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 8 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: |
Table 41. Example BC4 signed descriptor
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 44 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 1) = 40 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 8 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information | ||||||||||||||||||||||||||
0 | 1 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: | |||||||||||||||||||||||||||||||
sampleUpper: |
This model represents the Direct3D BC5 format for dual-channel interpolated 8-bit data, as described in Section 19.3.
Each compressed texel block consists of 4×4 texels in 16 bytes. The model has two channels, 0 (red) and 1 (green), which should have their bit depths and offsets independently described: the red channel has offset 0 and length 64 bits and the green channel has offset 64 and length 64 bits.
Enumerant names for these channel ids are listed in Table 42.
Table 42. BC5 channel names
Enumerant | Value |
---|---|
| 0 |
| |
| 1 |
|
Table 43. Example BC5 unsigned descriptor
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 60 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 2) = 56 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 16 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Red sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Green sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 64 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: |
Table 44. Example BC5 signed descriptor
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 60 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 2) = 56 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 16 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Red sample information | ||||||||||||||||||||||||||
0 | 1 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: | |||||||||||||||||||||||||||||||
sampleUpper: | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Green sample information | ||||||||||||||||||||||||||
0 | 1 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 64 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: | |||||||||||||||||||||||||||||||
sampleUpper: |
A legacy variant of this format known as “ATI2n” or “3Dc” swaps the location of the two channels, and can be encoded as follows:
Table 45. Example ATI2n unsigned descriptor
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 60 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 2) = 56 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 16 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Red sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Green sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 64 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: |
This model represents the Direct3D BC6H format for RGB floating-point data, as described in Section 20.2.
Each compressed texel block consists of 4×4 texels in 16 bytes. The model has a single channel 0, representing all three channels, and occupying 128 bits.
The enumerant names for this channel id are listed in Table 46.
Table 47. Example BC6H signed descriptor
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 44 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 1) = 40 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 16 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information | ||||||||||||||||||||||||||
1 | 1 | 0 | 0 |
| bitLength: 127 (= “128”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0xBF800000U — -1.0f | |||||||||||||||||||||||||||||||
sampleUpper: 0x7F800000U — 1.0f |
Table 48. Example BC6H unsigned descriptor
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 44 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 1) = 40 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 16 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information | ||||||||||||||||||||||||||
1 | 0 | 0 | 0 |
| bitLength: 127 (= “128”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0xBF800000U — -1.0f | |||||||||||||||||||||||||||||||
sampleUpper: 0x7F800000U — 1.0f |
This model represents the Direct3D BC7 format for RGBA data, as described in Section 20.1.
Each compressed texel block consists of 4×4 texels in 16 bytes. The model has a single channel 0, representing all four channels, and occupying 128 bits.
The enumerant names for this channel id are listed in Table 49.
Table 50. Example BC7 descriptor
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 44 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 1) = 40 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 16 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 127 (= “128”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: |
This model represents the original Ericsson Texture Compression format, described in Section 21, with a guarantee that the format does not rely on the ETC2 extensions described in Section 22.
Each compressed texel block consists of 4×4 texels in 8 bytes. The model has a single channel 0, representing all three channels, and occupying 64 bits.
The enumerant names for this channel id are listed in Table 51.
Table 52. Example ETC1 descriptor
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 44 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 1) = 40 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 8 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: |
This model represents the updated Ericsson Texture Compression format, ETC2, and also the related R11 EAC and RG11 EAC formats. Each compressed texel block consists of 4×4 texels.
The enumerant names for these channel ids are listed in Table 53.
Table 53. ETC2 channel names
Enumerant | Value |
---|---|
| 0 |
| |
| 1 |
| |
| 2 |
| 15 |
|
Channel ID 0 represents red, and is used for the R11 EAC format, as described in Section 22.5; the texel block size in this format is 8 bytes, represented as a single 64-bit sample.
Table 54. Example R11 unsigned descriptor
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 44 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 1) = 40 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 8 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: |
Table 55. Example R11 signed descriptor
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 44 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 1) = 40 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 8 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information | ||||||||||||||||||||||||||
0 | 1 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: | |||||||||||||||||||||||||||||||
sampleUpper: |
Channel ID 1 represents green; the presence of samples for both red and green, in that order, indicates the RG11 EAC format as described in Section 22.6, which consists of a total of 16 bytes of data.
Table 56. Example RG11 unsigned descriptor
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 60 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 2) = 56 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 16 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Red sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Green sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 64 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: |
Table 57. Example RG11 signed descriptor
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 60 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 2) = 56 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 16 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Red sample information | ||||||||||||||||||||||||||
0 | 1 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: | |||||||||||||||||||||||||||||||
sampleUpper: | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Green sample information | ||||||||||||||||||||||||||
0 | 1 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 64 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: | |||||||||||||||||||||||||||||||
sampleUpper: |
Channel ID 2 represents RGB combined content, for the ETC2 format as described in Section 22.1. A single sample of ID 2 indicates RGB2 with no alpha, occupying 8 bytes.
Table 58. Example ETC2 descriptor (with no alpha)
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 44 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 1) = 40 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 8 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: |
Channel ID 15 indicates the presence of alpha. If the texel block size is 8 bytes and the RGB and alpha channels are co-sited, “punch through” alpha is supported as described in Section 22.9.
Table 59. Example ETC2 descriptor with punchthrough alpha
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 60 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 1) = 56 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 8 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Color sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Alpha sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: |
Finally, if the texel block size is 16 bytes and the alpha channel appears in the first 8 bytes, followed by 8 bytes for the RGB channel, 8-bit separate alpha is supported, as described in Section 22.3.
Table 60. Example ETC2 descriptor with separate alpha
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 60 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 1) = 56 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 16 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Alpha sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Color sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 64 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: |
This model represents Adaptive Scalable Texture Compression as a single
channel in a texel block of 16 bytes. ASTC HDR (high dynamic range) and
LDR (low dynamic range) modes are distinguished by the channelId containing
the flag KHR_DF_SAMPLE_DATATYPE_FLOAT
: an ASTC texture that is guaranteed by
the user to contain only LDR-encoded blocks should have the channelId
KHR_DF_SAMPLE_DATATYPE_FLOAT
bit clear, and an ASTC texture that may include
HDR-encoded blocks should have the channelId KHR_DF_SAMPLE_DATATYPE_FLOAT
bit set to 1.
ASTC supports a number of compression ratios defined by different texel block
sizes; these are selected by changing the texel block size fields in the data
format.
ASTC encoding is described in Section 23.
The single sample, of ID 0, has a size of 128 bits.
The enumerant name for this channel id is listed in Table 61.
Table 62. Example 4×4 ASTC LDR descriptor
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 44 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 1) = 40 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 16 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 127 (= “128”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: |
Table 63. Example 8×5 ASTC HDR descriptor
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 44 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 1) = 40 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 4 (= “5”) | 7 (= “8”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 16 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information | ||||||||||||||||||||||||||
1 | 1 | 0 | 0 |
| bitLength: 127 (= “128”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0xBF800000U — -1.0f | |||||||||||||||||||||||||||||||
sampleUpper: 0x7F800000U — 1.0f |
This model represents a subset of the original Ericsson Texture Compression format, described in Section 21.1, which is restricted in order to facilitate image compression.
Each compressed texel block consists of 4×4 texels in 8 bytes. The model has a single channel 0, representing all three channels, and occupying 64 bits.
The enumerant names for this channel id are listed in Table 64.
Table 65. Example ETC1S descriptor
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 44 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 1) = 40 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 8 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: |
This model represents the first generation of PowerVR Texture Compression as a single channel in a texel block of 8 bytes. 4-bit-per-pixel mode represents a 4×4 texel block; 2-bit-per-pixel mode represents an 8×4 texel block, and these can be distinguished by changing the texel block size fields in the data format. The single sample has a size of 64 bits.
The enumerant names for this channel id are listed in Table 66.
Table 67. Example PVRTC 4bpp descriptor
uint32_t bit | |||||||||||||||||||||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 44 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 1) = 40 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 8 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: |
Table 68. Example PVRTC 2bpp descriptor
uint32_t bit | |||||||||||||||||||||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 44 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 1) = 40 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 7 (= “8”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 8 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: |
This model represents the second generation of PowerVR Texture Compression as a single channel in a texel block of 8 bytes. 4-bit-per-pixel mode represents a 4×4 texel block; 2-bit-per-pixel mode represents an 8×4 texel block, and these can be distinguished by changing the texel block size fields in the data format. The single sample has a size of 64 bits.
The enumerant names for this channel id are listed in Table 69.
Table 69. PVRTC2 channel names
Enumerant | Value |
---|---|
| 0 |
|
Table 70. Example PVRTC2 4bpp descriptor
uint32_t bit | |||||||||||||||||||||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 44 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 1) = 40 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 3 (= “4”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 8 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: |
Table 71. Example PVRTC2 2bpp descriptor
uint32_t bit | |||||||||||||||||||||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 44 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 1) = 40 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 3 (= “4”) | 7 (= “8”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 8 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 63 (= “64”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: |
It is not sufficient to define a buffer as containing, for example,
additive primaries. Additional information is required to define
what “red” is provided by the “red” channel. A full definition of
primaries requires an extension which provides the full color space
of the data, but a subset of common primary spaces can be identified
by the khr_df_primaries_e
enumeration, represented as an unsigned 8-bit
integer.
More information about color primaries is provided in Section 14.
Table 72. Field location information for colorPrimaries
Word offset into basic descriptor block |
| 2 |
Word offset into descriptor |
| 3 |
Start bit within word |
| 8 |
Bit mask of value |
| 0xFF |
khr_df_primaries_e
colorPrimaries = KHR_DF_MASK_PRIMARIES
&
(bdb
[KHR_DF_WORD_PRIMARIES
] >> KHR_DF_SHIFT_PRIMARIES
);
This “set of primaries” identifies a data representation whose color representation is unknown or which does not fit into this list of common primaries. Having an “unspecified” value here precludes users of this data format from being able to perform automatic color conversion unless the primaries are defined in another way. Formats which require a proprietary color space — for example, raw data from a Bayer sensor that records the direct response of each filtered sample — can still indicate that samples represent “red”, “green” and “blue”, but should mark the primaries here as “unspecified” and provide a detailed description in an extension block.
This value represents the Color Primaries defined by the ITU-R BT.709 specification and described in Section 14.1, which are also shared by sRGB.
RGB data is distinguished between BT.709 and sRGB by the Transfer Function. Conversion to and from BT.709 Y′CBCR (YUV) representation uses the color conversion matrix defined in the BT.709 specification, and described in Section 15.1.1, except in the case of sYCC (which can be distinguished by the use of the sRGB transfer function), in which case conversion to and from BT.709 Y′CBCR representation uses the color conversion matrix defined in the BT.601 specification, and described in Section 15.1.2. This is the preferred set of color primaries used by HDTV and sRGB, and likely a sensible default set of color primaries for common rendering operations.
KHR_DF_PRIMARIES_SRGB
is provided as a synonym for KHR_DF_PRIMARIES_BT709
.
This value represents the Color Primaries defined in the ITU-R BT.601 specification for standard-definition television, particularly for 625-line signals, and described in Section 14.2. Conversion to and from BT.601 Y′CBCR (YUV) typically uses the color conversion matrix defined in the BT.601 specification and described in Section 15.1.2.
This value represents the Color Primaries defined in the ITU-R BT.601 specification for standard-definition television, particularly for 525-line signals, and described in Section 14.3. Conversion to and from BT.601 Y′CBCR (YUV) typically uses the color conversion matrix defined in the BT.601 specification and described in Section 15.1.2.
This value represents the Color Primaries defined in the ITU-R BT.2020 specification for ultra-high-definition television and described in Section 14.4. Conversion to and from BT.2020 Y′CBCR (YUV uses the color conversion matrix defined in the BT.2020 specification and described in Section 15.1.3.
This value represents the theoretical Color Primaries defined by the International Color Consortium for the ICC XYZ linear color space.
This value represents the Color Primaries defined for the Academy Color Encoding System and described in Section 14.7.
This value represents the Color Primaries defined for the Academy Color Encoding System compositor and described in Section 14.8.
This value represents the Color Primaries defined for the NTSC 1953 color television transmission standard and described in Section 14.5.
This value represents the Color Primaries defined for 525-line PAL signals, described in Section 14.6.
This value represents the Color Primaries defined for the Display P3 color space, described in Section 14.9.
This value represents the Color Primaries defined in Adobe RGB (1998), described in Section 14.10.
Many color representations contain a non-linear transfer function
which maps between a linear (intensity-based) representation and
a more perceptually-uniform encoding; more information is provided
in Section 13.
Common transfer functions are represented as an unsigned 8-bit integer
and encoded in the enumeration khr_df_transfer_e
.
A fully-flexible transfer function requires an extension with a full
color space definition.
Where the transfer function can be described as a simple power curve,
applying the function is commonly known as “gamma correction”.
The transfer function is applied to a sample only when the sample’s
KHR_DF_SAMPLE_DATATYPE_LINEAR
bit is 0; if this bit is 1, the
sample is represented linearly irrespective of the transferFunction.
When a color model contains more than one channel in a sample and
the transfer function should be applied only to a subset of those
channels, the convention of that model should be used when applying
the transfer function. For example, ASTC stores both alpha and RGB
data but is represented by a single sample; in ASTC, any sRGB transfer
function is not applied to the alpha channel of the ASTC texture.
In this case, the KHR_DF_SAMPLE_DATATYPE_LINEAR
bit being zero means
that the transfer function is “applied” to the ASTC sample in a way that
only affects the RGB channels. This is not a concern for most color
models, which explicitly store different channels in each sample.
If all the samples are linear, KHR_DF_TRANSFER_LINEAR
should be used.
In this case, no sample should have the KHR_DF_SAMPLE_DATATYPE_LINEAR
bit set.
If the samples encode a single bit, KHR_DF_TRANSFER_LINEAR
or
KHR_DF_TRANSFER_UNSPECIFIED
should be used, since there are no
intermediate values to which the transfer function should apply.
Table 73. Field location information for transferFunction
Word offset into basic descriptor block |
| 2 |
Word offset into descriptor |
| 3 |
Start bit within word |
| 16 |
Bit mask of value |
| 0xFF |
khr_df_transfer_e
transferFunction = KHR_DF_MASK_TRANSFER
&
(bdb
[KHR_DF_WORD_TRANSFER
] >> KHR_DF_SHIFT_TRANSFER
);
The enumerant value for each of the following transfer functions is shown in parentheses alongside the title.
This value should be used when the transfer function is unknown, or specified only in an extension block, precluding conversion of color spaces and correct filtering of the data values using only the information in the basic descriptor block.
This value represents a linear transfer function: for color data, there is a linear relationship between numerical pixel values and the intensity of additive colors. This transfer function allows for blending and filtering operations to be applied directly to the data values.
This value represents the non-linear transfer function defined in the sRGB specification for mapping between numerical pixel values and displayed light intensity, as described in Section 13.3.
Mapping from linear intensity to encoding | EOTF -1 | |
Mapping from encoding to linear intensity | EOTF |
Encoded values outside the range 0..1 use the extended formulae for EOTF and EOTF-1 described in Section 13.3.4.
This value represents the non-linear transfer function defined by the ITU and used in the BT.601, BT.709 and BT.2020 specifications for mapping between represented scene light intensity and numerical pixel values, as described in Section 13.2.
Mapping from linear intensity to encoding | OETF | |
Mapping from encoding to linear intensity | OETF -1 |
This value represents the non-linear transfer function defined by the original NTSC television broadcast specification for mapping between represented scene light intensity or display light intensity and numerical pixel values, as described in Section 13.8.
Mapping from linear intensity to encoding | EOTF -1 / OETF |
Mapping from encoding to linear intensity | EOTF / OETF -1 |
More recent formulations of this transfer functions, such as that defined in SMPTE 170M-2004, use the “ITU” formulation described above. |
This value represents a nonlinear Transfer Function between linear scene light intensity and nonlinear pixel values, used by some Sony video cameras to represent an increased dynamic range, and is described in Section 13.13.
Mapping from linear intensity to encoding | OETF |
Mapping from encoding to linear intensity | OETF -1 |
This value represents a nonlinear Transfer Function between linear scene light intensity and nonlinear pixel values, used by some Sony video cameras to represent a further increased dynamic range, and is described in Section 13.14.
Mapping from linear intensity to encoding | OETF |
Mapping from encoding to linear intensity | OETF -1 |
This value represents the nonlinear $\gamma = 2.4$ EOTF between encoded pixel values and linear image intensity defined in BT.1886 and described in Section 13.4.
Mapping from linear intensity to encoding | EOTF -1 | $\{R',G',B'\} = \{R,G,B\}^{2.4}$ |
Mapping from encoding to linear intensity | EOTF | $\{R,G,B\} = \{R',G',B'\}^{1\over{2.4}}$ |
This value represents the Hybrid Log Gamma OETF between linear scene light intensity and nonlinear pixel values, defined by the ITU in BT.2100 for high dynamic range television, and described in Section 13.5.
Mapping from linear intensity to encoding | OETF | |
Mapping from encoding to linear intensity | OETF -1 |
This value represents the Hybrid Log Gamma EOTF between nonlinear pixel values and linear image light intensity, defined by the ITU in BT.2100 for high dynamic range television, and described in Section 13.5.
Mapping from linear intensity to encoding | EOTF -1 | |
Mapping from encoding to linear intensity | EOTF |
This value represents the Perceptual Quantization EOTF between nonlinear pixel values and linear image light intensity, defined by the ITU in BT.2100 for high dynamic range television, and described in Section 13.6.
Mapping from linear intensity to encoding | EOTF -1 | |
Mapping from encoding to linear intensity | EOTF |
This value represents the Perceptual Quantization OETF between linear scene light intensity and nonlinear pixel values, defined by the ITU in BT.2100 for high dynamic range television, and described in Section 13.6.
Mapping from linear intensity to encoding | OETF | |
Mapping from encoding to linear intensity | OETF -1 |
This value represents the transfer function between nonlinear pixel values and linear image light intensity defined in DCI P3 and described in Section 13.7.
Mapping from linear intensity to encoding | EOTF -1 |
Mapping from encoding to linear intensity | EOTF |
This value represents the OETF between linear scene light intensity and nonlinear pixel values for legacy PAL systems described in Section 13.9.
Mapping from linear intensity to encoding | OETF |
Mapping from encoding to linear intensity | OETF -1 |
This value represents the EOTF between nonlinear pixel values and linear image light intensity for legacy 625-line PAL systems described in Section 13.10.
Mapping from linear intensity to encoding | EOTF -1 |
Mapping from encoding to linear intensity | EOTF |
This value represents the transfer function between linear scene light intensity and nonlinear pixel values associated with the legacy ST-240 (SMPTE240M) standard, described in Section 13.11.
Mapping from linear intensity to encoding | EOTF -1 / OETF |
Mapping from encoding to linear intensity | EOTF / OETF -1 |
This value represents the nonlinear transfer function between linear scene light intensity and nonlinear pixel values used in the ACEScc Academy Color Encoding System logarithmic encoding system for use within Color Grading Systems, S-2014-003, defined in ACES. This is described in Section 13.15.
Mapping from linear intensity to encoding | OETF |
Mapping from encoding to linear intensity | OETF -1 |
This value represents the nonlinear transfer function between linear scene light intensity and nonlinear pixel values used in the ACEScc Academy Color Encoding System quasi-logarithmic encoding system for use within Color Grading Systems, S-2016-001, defined in ACES. This is described in Section 13.16.
Mapping from linear intensity to encoding | OETF |
Mapping from encoding to linear intensity | OETF -1 |
This value represents the transfer function defined in the Adobe RGB (1998) specification and described in Section 13.12.
The format supports some configuration options in the form of
boolean flags; these are described in the enumeration
khr_df_flags_e
and represented in an unsigned 8-bit integer value.
Table 74. Field location information for flags
Word offset into basic descriptor block |
| 2 |
Word offset into descriptor |
| 3 |
Start bit within word |
| 24 |
Bit mask of value |
| 0xFF |
khr_df_flags_e
flags = KHR_DF_MASK_FLAGS
&
(bdb
[KHR_DF_WORD_FLAGS
] >> KHR_DF_SHIFT_FLAGS
);
If the KHR_DF_FLAG_ALPHA_PREMULTIPLIED
bit is set, any color
information in the data should be interpreted as having been
previously scaled/modulated by the alpha channel when performing
blending operations.
The value KHR_DF_FLAG_ALPHA_STRAIGHT
(= 0) is provided
to represent this flag not being set, which indicates that color
values in the data should be interpreted as needing to be scaled
by the alpha channel when performing blending operations.
This flag has no effect if there is no alpha channel in the format.
The texelBlockDimension fields define an integer bound on the range of coordinates covered by the repeating block described by the samples; that is, the texel block covers an integer range in each dimension of coordinate space. Four separate values, represented as unsigned 8-bit integers, are supported, corresponding to successive dimensions: the Basic Data Format Descriptor Block supports up to four dimensions of encoding within a texel block, supporting, for example, a texture with three spatial dimensions and one temporal dimension. Nothing stops the data structure as a whole from having higher dimensionality: for example, a two-dimensional texel block can be used as an element in a six-dimensional look-up table.
The value held in each of these fields is one fewer than the size of the block in that dimension — that is, a value of 0 represents a size of 1, a value of 1 represents a size of 2, etc. A texel block which covers fewer than four dimensions should have a size of 1 in each dimension that it lacks, and therefore the corresponding fields in the representation should be 0.
For example, a Y′CBCR 4:2:0 representation may use a Texel Block of 2×2 pixels in the nominal coordinate space, corresponding to the four Y′ samples, as shown in Table 75. The texel block dimensions in this case would be 2×2×1×1 (in the X, Y, Z and T dimensions, if the fourth dimension is interpreted as T). The texelBlockDimension[0..3] values would therefore be:
Table 75. Example Basic Data Format texelBlockDimension values for Y′CBCR 4:2:0
texelBlockDimension0 | 1 |
texelBlockDimension1 | 1 |
texelBlockDimension2 | 0 |
texelBlockDimension3 | 0 |
In the descriptor block examples in this specification, block dimensions larger than 1 (encoded as 0) are shown as the value to be stored in the texelBlockDimension field, but with the represented number in parentheses for clarity.
Table 76. Field location information for texelBlockDimension[0..3]
Word offset into basic descriptor block |
| 3 |
Word offset into descriptor |
| 4 |
Start bit within word |
| 0 |
| 8 | |
| 16 | |
| 24 | |
Bit mask of value |
| 0xFF |
uint32_t
texelBlockDimension0 = KHR_DF_MASK_TEXELBLOCKDIMENSION0
&
(bdb
[KHR_DF_WORD_TEXELBLOCKDIMENSION0
] >> KHR_DF_SHIFT_TEXELBLOCKDIMENSION0
);
uint32_t
texelBlockDimension1 = KHR_DF_MASK_TEXELBLOCKDIMENSION1
&
(bdb
[KHR_DF_WORD_TEXELBLOCKDIMENSION1
] >> KHR_DF_SHIFT_TEXELBLOCKDIMENSION1
);
uint32_t
texelBlockDimension2 = KHR_DF_MASK_TEXELBLOCKDIMENSION2
&
(bdb
[KHR_DF_WORD_TEXELBLOCKDIMENSION2
] >> KHR_DF_SHIFT_TEXELBLOCKDIMENSION2
);
uint32_t
texelBlockDimension3 = KHR_DF_MASK_TEXELBLOCKDIMENSION3
&
(bdb
[KHR_DF_WORD_TEXELBLOCKDIMENSION3
] >> KHR_DF_SHIFT_TEXELBLOCKDIMENSION3
);
The Basic Data Format Descriptor divides the image into a number of planes, each consisting of an integer number of consecutive bytes. The requirement that planes consist of consecutive data means that formats with distinct subsampled channels — such as Y′CBCR 4:2:0 — may require multiple planes to describe a channel. A typical Y′CBCR 4:2:0 image has two planes for the Y′ channel in this representation, offset by one line vertically.
The use of byte granularity to define planes is a choice to allow large texel blocks. A consequence of this is that formats which are not byte-aligned on each addressable unit, such as 1-bit-per-pixel formats, need to represent a texel block of multiple samples, covering multiple texels — as, for example, in Table 92.
A maximum of eight independent planes is supported in the Basic Data Format Descriptor. Formats which require more than eight planes — which are rare — require an extension.
The bytesPlane[0..7] fields each contain an unsigned 8-bit integer which represents the number of bytes which a plane contributes to the format. If the top bit of a bytesPlane[n] field is set, bits 6..0 of the bytesPlane[n+1] field form bits 13..7 of the number of bytes in the plane (and the next plane is described by field bytesPlane[n+2]). For example, if bytesPlane0 is 0xC0 and bytesPlane1 is 0x02, the first plane holds 0x40 + 128 × 0x02 = 0x140 bytes; bytesPlane2 then describes the number of bytes in the second plane.. Since only sixteen bits are used to encode a bit offset for each sample, 14 bits (two bytes excluding the top bits) are sufficient to encode any useful number of bytes — there is no need to “extend” the higher byte. Few formats are expected to require this “extension bit”, so for most of this specification, the number of bytes in a plane is considered to be synonymous with the bytesPlane value.
The first field which contains the value 0 indicates that only a subset of the 8 possible planes are present; that is, planes which are not present should be given the bytesPlane value of 0, and any bytesPlane values after the first 0 are ignored. If no bytesPlane value is zero, 8 planes are considered to exist.
Table 77. Field location information for bytesPlane[0..7]
Word offset into basic descriptor block |
| 4 |
| 5 | |
Word offset into descriptor |
| 5 |
| 6 | |
Start bit within word |
| 0 |
| 8 | |
| 16 | |
| 24 | |
| 0 | |
| 8 | |
| 16 | |
| 24 | |
Bit mask of value |
| 0xFF |
uint32_t
bytesPlane[0..7] = KHR_DF_MASK_BYTESPLANE[0..7]
&
(bdb
[KHR_DF_WORD_BYTESPLANE[0..7]
] >> KHR_DF_SHIFT_BYTESPLANE[0..7]
);
In versions of this specification prior to 1.3, there was no facility for the “extension bit”, and a bytesPlane0 value of 0 indicated a paletted format. The scheme for encoding paletted formats as of version 1.3 is described in Section 5.18. |
The layout and position of the information within each plane is determined by a number of samples, each consisting of a single channel of data and with a single corresponding position within the texel block, as shown in Table 78.
Table 78. Basic Data Format Descriptor Sample Information
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
F | S | E | L | channelType | bitLength | bitOffset | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
sampleLower | |||||||||||||||||||||||||||||||
sampleUpper |
Bits F, S, E and L are abbreviations for the following qualifier flags:
F |
| 0x80 |
S |
| 0x40 |
E |
| 0x20 |
L |
| 0x10 |
The sample information begins at word KHR_DF_WORD_SAMPLESTART
= 6, offset from the start of the basic descriptor block.
Each sample occupies KHR_DF_WORD_SAMPLEWORDS
= 4 32-bit words,
and is stored consecutively.
The bytes from the plane data contributing to the format are treated as though they have been concatenated into a bit stream, with the first byte of the lowest-numbered plane providing the lowest bits of the result. Each sample consists of a number of consecutive bits from this bit stream.
If the content for a channel cannot be represented in a single sample, for example because the data for a channel is non-consecutive within this bit stream, additional samples with the same coordinate position and channel number should follow from the first, in order increasing from the least significant bits from the channel data; the corresponding bits from the bit stream are concatenated in the increasing order of reference to provide the value representing the channel.
For example, some native big-endian formats may need to be supported with multiple samples in a channel, since the constituent bits may not be consecutive in a little-endian interpretation. There is an example, Table 97, in the list of example format descriptors provided.
See Section 3.4 for more information about the order in which samples should appear in the descriptor block.
The number of samples present in the format is determined by the descriptorBlockSize field:
numSamples =
(((KHR_DFDVAL
(BDB
, DESCRIPTORBLOCKSIZE
) >> 2) - KHR_DF_WORD_SAMPLESTART
)
/ KHR_DF_WORD_SAMPLEWORDS
);
The macro KHR_DFDSAMPLECOUNT
(BDB
) is provided to perform
this calculation.
There is no limit on the number of samples which may be present, other than the maximum size of the Data Format Descriptor Block. There is no requirement that samples should access unique parts of the bit-stream: formats such as combined intensity and alpha, or shared exponent formats, require that bits be reused. Nor is there a requirement that all the bits in a plane be used (a format may contain padding).
It is unusual but legal for a descriptor block to contain no
samples provided the color model is KHR_DF_MODEL_UNSPECIFIED
.
See Section 5.19 for details.
To simplify code using the Basic Data Format Descriptor Block, the
header khr_df.h
provides enums of the following form for accessing
sample fields:
Table 79. Field location information for sample field xxx
Word offset relative to start of sample |
| … |
Start bit within word |
| … |
Bit mask of value |
| … |
If the basic descriptor block is treated as a uint32_t
array bdb[]
,
sample field xxx can be accessed as follows:
xxx = KHR_DF_SAMPLEMASK_xxx
&
(bdb
[KHR_DF_WORD_SAMPLESTART
+ sample × KHR_DF_WORD_SAMPLEWORDS
+ KHR_DF_SAMPLEWORD_xxx
] >> KHR_DF_SAMPLESHIFT_xxx
);
The macro KHR_DFDSVAL
(BDB
, S
, X
) is provided to perform this
calculation.
For example, KHR_DFDSVAL
(bdb
, 2, CHANNELID
) returns the value:
KHR_DF_SAMPLEMASK_CHANNELID
&
(bdb
[KHR_DF_WORD_SAMPLESTART
+ 2 × KHR_DF_WORD_SAMPLEWORDS
+ KHR_DF_SAMPLEWORD_CHANNELID
] >> KHR_DF_SAMPLESHIFT_CHANNELID
)
The bitOffset field describes the offset of the least significant bit of this sample from the least significant bit of the least significant byte of the concatenated bit stream for the format. Typically the bitOffset of the first sample is therefore 0; a sample which begins at an offset of one byte relative to the data format would have a bitOffset of 8. The bitOffset is an unsigned 16-bit integer quantity.
In the special case that the bitOffset field contains the reserved value 0xFFFF, the sample contributes a constant value of the specified bit length, encoded in the sampleLower field. This mechanism notably supports values that are zero-extended.
Table 80. Field location information for sample bitOffset
Word offset relative to start of sample |
| 0 |
Start bit within word |
| 0 |
Bit mask of value |
| 0xFFFFU |
uint32_t
bitoffset = KHR_DF_SAMPLEMASK_BITOFFSET
&
((bdb
[KHR_DF_WORD_SAMPLESTART
+ (sample × KHR_DF_WORD_SAMPLEWORDS
)
+ KHR_DF_SAMPLEWORD_BITOFFSET
]) >> KHR_DF_SAMPLESHIFT_BITOFFSET
);
The bitLength field describes the number of consecutive bits from the concatenated bit stream that contribute to the sample. This field is an unsigned 8-bit integer quantity, and stores the number of bits contributed minus 1; thus a single-byte channel should have a bitLength field value of 7. If a bitLength of more than 256 is required, further samples should be added; the value for the sample is composed in increasing order from least to most significant bit as subsequent samples are processed.
Note that a large bitLength value means a sample can encode more bits than can be described in the sampleLower and sampleUpper fields. If the rules for expanding the sampleLower and sampleUpper values in this case produce the desired result, bitLength can exceed 32; otherwise the bitLength should be limited and multiple samples used to encode sampleLower and sampleUpper: there is no way to indicate a sample contribution of zero bits for a sample that exists only to expand upon sampleLower and sampleUpper.
Except in the case of a paletted texture (described in Section 5.18) or where the special bitOffset value 0xFFFF is used to indicate constant bits, the bitLength added to bitOffset should not be greater than eight times the total number of bytes contributed to the logical bit stream by the bytesPlane values.
In the descriptor block examples in this specification, bit lengths are shown as the value to be stored in the bitLength field, but with the represented number (without the -1 offset) in parentheses for clarity.
Table 81. Field location information for sample bitLength
Word offset relative to start of sample |
| 0 |
Start bit within word |
| 16 |
Bit mask of value |
| 0xFF |
uint32_t
bitLength = KHR_DF_SAMPLEMASK_BITLENGTH
&
((bdb
[KHR_DF_WORD_SAMPLESTART
+ (sample × KHR_DF_WORD_SAMPLEWORDS
)
+ KHR_DF_SAMPLEWORD_BITLENGTH
]) >> KHR_DF_SAMPLESHIFT_BITLENGTH
);
The channelType field is an unsigned 8-bit quantity.
The bottom four bits of the channelType indicates which channel is being described by this sample. The list of available channels is determined by the colorModel field of the Basic Data Format Descriptor Block, and the channelType field contains the number of the required channel within this list — see the colorModel field for the list of channels for each model.
Table 82. Field location information for sample channelType
Word offset relative to start of sample |
| 0 |
Start bit within word |
| 24 |
Bit mask of value |
| 0xF |
khr_df_model_channels_e
channelType = KHR_DF_SAMPLEMASK_CHANNELID
&
((bdb
[KHR_DF_WORD_SAMPLESTART
+ (sample × KHR_DF_WORD_SAMPLEWORDS
)
+ KHR_DF_SAMPLEWORD_CHANNELID
]) >> KHR_DF_SAMPLESHIFT_CHANNELID
);
The top four bits of the channelType are described by the
khr_df_sample_datatype_qualifiers_e
enumeration:
Table 83. Field location information for sample qualifiers
Word offset relative to start of sample |
| 0 |
Start bit within word |
| 24 |
Bit mask of value |
| 0xF0 |
khr_df_sample_datatype_qualifiers_e
qualifiers = KHR_DF_SAMPLEMASK_QUALIFIERS
&
((bdb
[KHR_DF_WORD_SAMPLESTART
+ (sample × KHR_DF_WORD_SAMPLEWORDS
)
+ KHR_DF_SAMPLEWORD_QUALIFIERS
]) >> KHR_DF_SAMPLESHIFT_QUALIFIERS
);
If the KHR_DF_SAMPLE_DATATYPE_EXPONENT
bit, shown as E in
Table 78, is not set, the sample contributes to a base value;
multiple samples with the same color channel (bottom four bits of
channelType) and same samplePos values are accumulated into
the virtual sample in increasing bit order from lowest to highest.
For samples in which the KHR_DF_SAMPLE_DATATYPE_EXPONENT
bit is not set:
KHR_DF_SAMPLE_DATATYPE_LINEAR
bit, shown as L in
Table 78, is not set, the final sample value (after any
modifier has been applied to the base value) is modified by the
transfer function defined in the transferFunction field of the
descriptor; if this bit is set, the sample is considered to
contain a linearly-encoded value irrespective of the format’s
transferFunction.
All samples referring to the same base value should have the same
value stored in the KHR_DF_SAMPLE_DATATYPE_LINEAR
bit.
KHR_DF_SAMPLE_DATATYPE_SIGNED
bit, shown as S in
Table 78, is set, the sample holds a signed value in two’s
complement form.
If this bit is not set, the sample holds an unsigned value.
It is possible to represent a sign/magnitude integer value by having a
sample of unsigned integer type with the same channel and sample location,
as a 1-bit signed sample.
KHR_DF_SAMPLE_DATATYPE_FLOAT
bit, shown as F in
Table 78, is set, the sample holds floating point data in a
conventional format of 10, 11 or 16 bits, as described in Section 10,
or of 32, or 64 bits as described in [IEEE 754].
Unless a genuine unsigned format is intended,
KHR_DF_SAMPLE_DATATYPE_SIGNED
(bit S) should also be set.
Less common floating point representations can be generated with
multiple samples and a combination of signed integer, unsigned integer
and exponent fields, as described above and in Section 10.4.
If the KHR_DF_SAMPLE_DATATYPE_EXPONENT
bit, shown as E in
Table 78, is set, the sample applies a modifier to the base
value, with the interpretation of the modifier determined according to
Table 84.
In this case, the virtual sample contains both a base value and
a modifier.
All samples contributing to a modifier for the same base value should contain the same L and F bits (it is not legal, for example, to define both a multiplier and a divisor). Samples which apply a modifier should directly follow the samples that describe the base value. If no samples have the E bit set for this channel and position, the base value directly represents the pixel value; it is not legal for a virtual sample to describe a modifier but no base value.
The same bits of the format may contribute to modifiers for more than one channel — this is commonly the case for high dynamic range formats with a shared exponent or divisor. The descriptor in this case should contain samples for each color channel in turn, with the description of the shared bits replicated for each channel, as shown in Table 98. |
Table 84. Qualifier interpretation when KHR_DF_SAMPLE_DATATYPE_EXPONENT
= 1
E | L | F | Interpretation | Formula |
1 | 0 | 0 | Exponent | $\mathit{base\ value}\times 2^\mathit{modifier}$ |
1 | 0 | 1 | Multiplier | $\mathit{base\ value}\times\mathit{modifier}$ |
1 | 1 | 0 | Divisor | $\mathit{base\ value}\over\mathit{modifier}$ |
1 | 1 | 1 | Power | $\mathit{base\ value}^\mathit{modifier}$ |
For samples in which the KHR_DF_SAMPLE_DATATYPE_EXPONENT
bit is set:
KHR_DF_SAMPLE_DATATYPE_LINEAR
and
KHR_DF_SAMPLE_DATATYPE_FLOAT
bits are clear, this modifier holds
an exponent (in integer form) describing a floating-point offset for
this channel.
For example, this would be used to describe the exponent of a custom
floating point format, as shown in Table 103,
or a shared exponent location in shared exponent formats (with the
exponent bits listed separately under each channel as shown in
Table 98).
If this modifer is used, the base value is considered to contain
mantissa information and the samples describing it would not normally
have the KHR_DF_SAMPLE_DATATYPE_FLOAT
bit set.
If the KHR_DF_SAMPLE_DATATYPE_SIGNED
bit (S) is also set,
the exponent is considered to be two’s complement — otherwise it is
treated as unsigned.
The bias of the exponent can be determined by the exponent’s
sampleLower value.
The presence or absence of an implicit leading digit in the mantissa of
a format with an exponent can be determined by the sampleUpper value
of the mantissa.
The use of the exponent is described in more detail in Section 10.4.
KHR_DF_SAMPLE_DATATYPE_FLOAT
bit is set and the
KHR_DF_SAMPLE_DATATYPE_LINEAR
bit is clear, this sample holds
a multiplier (in integer form) for this channel, such that the encoded
value is a product of this modifier value and the base value.
This approach is useful for encoding a shared multiplier as part of
a high dynamic range color image, for example.
KHR_DF_SAMPLE_DATATYPE_LINEAR
bit is set and the
KHR_DF_SAMPLE_DATATYPE_FLOAT
bit is clear, this sample holds
a divisor (in integer form) for this channel, such that the encoded
value is the base value divided by this modifier value.
This approach is useful for encoding a shared divisor as part of
a high dynamic range color image, for example.
KHR_DF_SAMPLE_DATATYPE_FLOAT
and
KHR_DF_SAMPLE_DATATYPE_LINEAR
are set, this sample holds
a power term (in integer form) for this channel, such that the encoded
value is the base value raised to the power of the modifier value.
This approach is useful for encoding a shared multiplier as part of
a high dynamic range color image, for example.
Note that in the multiplier, divisor and power cases, the sampleLower and sampleUpper values allow the modifier value to be represented in fixed-point terms, and the values may be signed depending on whether the S bit is set.
The sample has an associated location within the 4-dimensional space of the texel block. Therefore each sample has an offset relative to the 0,0 position of the texel block, represented as an 8-bit unsigned integer quantity.
Table 85. Field location information for sample samplePosition[0..3]
Word offset relative to start of sample |
| 1 |
Start bit within word |
| 0 |
| 8 | |
| 16 | |
| 24 | |
Bit mask of value |
| 0xF |
khr_df_model_channels_e
samplePosition0 = KHR_DF_SAMPLEMASK_SAMPLEPOSITION0
&
((bdb
[KHR_DF_WORD_SAMPLESTART
+ (sample × KHR_DF_WORD_SAMPLEWORDS
)
+ KHR_DF_SAMPLEWORD_SAMPLEPOSITION0
]) >> KHR_DF_SAMPLESHIFT_SAMPLEPOSITION0
);
khr_df_model_channels_e
samplePosition1 = KHR_DF_SAMPLEMASK_SAMPLEPOSITION1
&
((bdb
[KHR_DF_WORD_SAMPLESTART
+ (sample × KHR_DF_WORD_SAMPLEWORDS
)
+ KHR_DF_SAMPLEWORD_SAMPLEPOSITION1
]) >> KHR_DF_SAMPLESHIFT_SAMPLEPOSITION1
);
khr_df_model_channels_e
samplePosition2 = KHR_DF_SAMPLEMASK_SAMPLEPOSITION2
&
((bdb
[KHR_DF_WORD_SAMPLESTART
+ (sample × KHR_DF_WORD_SAMPLEWORDS
)
+ KHR_DF_SAMPLEWORD_SAMPLEPOSITION2
]) >> KHR_DF_SAMPLESHIFT_SAMPLEPOSITION2
);
khr_df_model_channels_e
samplePosition3 = KHR_DF_SAMPLEMASK_SAMPLEPOSITION3
&
((bdb
[KHR_DF_WORD_SAMPLESTART
+ (sample × KHR_DF_WORD_SAMPLEWORDS
)
+ KHR_DF_SAMPLEWORD_SAMPLEPOSITION3
]) >> KHR_DF_SAMPLESHIFT_SAMPLEPOSITION3
);
The interpretation of each samplePosition field depends on the corresponding texelBlockDimension value as follows:
For example, if texelBlockDimension0 is 1 (indicating a texel block width of two units) samplePosition0 is described in units of ${2\over 256} = {1\over 128}$ . That is, a samplePosition0 of 128 would encode an offset of “1.0” and a samplePosition0 of 64 would encode an offset of “0.5”. If texelBlockDimension0 is 5 (indicating a texel block width of six units), samplePosition0 is described in units of ${8\over 256} = {1\over 32}$ . That is, a samplePosition0 of 64 would encode an offset of “2.0” and a samplePosition0 of 24 would encode an offset of “0.75”.
The adjusted coordinateOffset must be less than the corresponding texel block dimension. That is, since coordinateOffset can represent fractional offsets, coordinateOffset < (texelBlockDimension + 1) for each dimension.
This approach allows the common situation of downsampled channels to have samples conceptually sited at the midpoint between full resolution samples. The direction of the sample offsets is determined by the coordinate addressing scheme used by the API. There is no limit on the dimensionality of the data, but if more than four dimensions need to be contained within a single texel block, an extension will be required.
It is legal, but unusual, to use the same bits to represent multiple samples at different coordinate locations.
Versions of this specification prior to 1.3 always recorded the sample position in a 7.1 fixed-point format (with half-coordinate granularity). This change does not affect the representation of single-coordinate texel blocks; that is, a samplePosition of “0” still represents “0.0”. |
The sampleLower and sampleUpper fields are used to define the mapping between the numerical value stored in the format and the conceptual numerical interpretation.
For unsigned formats, sampleLower typically represents the value which should be interpreted as zero (the black point). For signed formats, sampleLower typically represents “-1”. For color difference models such as Y′CBCR, sampleLower for chroma channels represents the lower extent of the color difference range (which corresponds to an encoding of -0.5 in numerical terms).
sampleUpper typically represents the value which should be interpreted as “1.0” (the “white point”). For color difference models such as Y′CBCR, sampleUpper for chroma channels represents the upper extent of the color difference range (which corresponds to an encoding of 0.5 in numerical terms).
Equation 1. Sample range conversion rules
For example, the BT.709 television broadcast standard dictates that the Y′ value stored in an 8-bit encoding should fall between the range 16 and 235, as described in Section 16.1 encoding”. In this case, sampleLower should contain the value 16 and sampleUpper 235.
Sensor data from a camera typically does not cover the full range of the bit depth used to represent it. sampleUpper can be used to specify an upper limit on sensor brightness — or to specify the value which should map to white on the display, which may be less than the full dynamic range of the captured image.
There is no guarantee or expectation that image data be guaranteed to fall between sampleLower and sampleUpper unless the users of a format agree that convention. For example, high dynamic range video formats may define “1.0” as a nominal brightness level substantially lower than the maximum, and coordinates may encode an arbitrary range. In some formats, the integer value should be interpreted directly as a number, in which case sampleUpper and sampleLower should hold “1” and either “0” or “-1” depending on whether the format is signed, respectively.
If the channel encoding is an integer format, the sampleLower and sampleUpper values are represented as 32-bit integers — signed or unsigned according to whether the channel encoding is signed. Signed negative values should be sign-extended if the channel has fewer than 32 bits, such that the value encoded in sampleLower and sampleUpper values are themselves negative if the encoded values are negative. If the channel encoding is a floating point value, the sampleLower and sampleUpper values are also described in floating point.
If the number of bits in the sample is greater than 32, sampleLower
and sampleUpper are converted to a data type of the actual number
of bits as follows:
A floating point value is converted to the native representation (for
example, a float
value is converted to double
in the normal way).
An integer value is expanded by preserving any sign bit and replicating
the top non-sign bit (for example, signed 0x80000000 is extended to
the 40-bit value 0x8000000000).
If these rules do not produce the desired result, it may be necessary
to describe the contribution to the channel in multiple samples of
no more than 32 bits.
In this case, the samples corresponding to lower bitOffset values
should occupy 32 bits, with any residual bits encoded in samples
corresponding to higher bitOffset values.
If multiple samples contribute to a single value, for example because the bits of a channel are non-contiguous in the logical bit stream, the sampleLower and sampleUpper fields of each sample are concatenated in increasing bit order to produce virtual sampleLower and sampleUpper values; each sample contributes a number of bits equal to the number of data bits that the sample describes. If this contribution is fewer than 32 bits and the value being encoded is signed, the sign bit of the final sampleLower and sampleUpper values should be used to pad the fields of the sample.
For example, if a signed 16-bit value with a minimum value of -32767 (0xFFFE) is described as two 8-bit values, the first sample should have a sampleLower value of 0xFFFFFFFE — of which the bottom 8 bits (0xFE) corresponds to the bottom 8 bits of the final minimum value and the upper 24 bits (0xFFFFFF00) are a result of replicating the sign of bit 15 of the final minimum value. The second sample should have a sampleLower value of 0xFFFFFFFF — of which the bottom 8 bits (0xFF) correspond to the top 8 bits of the final minimum value and the upper 24 bits (0xFFFFFF00) are a result of sign-extending this value. Only the sample corresponding to the top bits of the channel may have a sampleLower or sampleUpper occupying more bits than the input.
The sampleLower value for an exponent should represent the exponent bias — the value that should be subtracted from the encoded exponent to indicate that the mantissa’s sampleUpper value will represent 1.0. The sampleUpper value for an exponent should represent the largest conventional legal exponent value. If the encoded exponent exceeds this value, the encoded floating point value encodes either an infinity or a NaN value, depending on the mantissa. See Section 10.4 for more detail on this.
If the channel encoding is the mantissa of a custom floating point format (that is, the encoding is integer but the same sample location and channel is shared by a sample that encodes an exponent), the presence of an implicit “1” digit can be represented by setting the sampleUpper value to a value one larger than can be encoded in the available bits for the mantissa, as described in Section 10.4.
In OpenGL terminology, a “normalized” channel contains an integer value which is mapped to the range 0..1.0; a channel which is not normalized contains an integer value which is mapped to a floating point equivalent of the integer value. Similarly an “snorm” channel is a signed normalized value mapping from -1.0 to 1.0. Setting sampleLower to the minimum signed integer value representable in the channel (which is often the negative version of the maximum signed integer, for example -127 rather than -128 for an 8-bit value in order to allow the value 0.0 to be represented exactly) is equivalent to defining an “snorm” texture. Setting sampleUpper to the maximum signed integer value representable in the channel for a signed channel type is equivalent to defining an “snorm” texture. Setting sampleUpper to the maximum unsigned value representable in the channel for an unsigned channel type is equivalent to defining a “normalized” texture. Setting sampleUpper to “1” is equivalent to defining an “unnormalized” texture.
In the special case that the sample bitOffset field is 0xFFFF, only the bottom 16 bits of the sampleLower field indicate a contribution to the sample lower limit; the upper 16 bits are taken as a constant contribution to the interpreted value; in this case, the bitLength field of the sample must be no more than 16. For example, a 4-bit value which is interpreted as being zero-extended to eight bits before conversion may have four bits with the value 0 stored in bits 19..16 of sampleLower, indicated by a bitOffset of 0xFFFF.
These “virtual bits” may be needed to encode some numerical representations. For example, if an 8-bit integer encodes the value “-0.5” as 0 and “0.5” as 255 (in the manner of the color difference channel in Equation 1, but if we wish to apply this mapping to a channel other than color difference), Equation 1 suggests that sampleLower should hold -127.5 and sampleUpper should hold 382.5. This is impossible to encode, since sampleLower and sampleUpper are integers. However, if a virtual 0-bit is added below the encoded value (such that the effective range is 0..510), storing sampleLower as -255 and sampleUpper as 765 will have the desired effect. Note that in this case, the original 8-bits of data are unsigned, but the bounds must encode a negative value; this can be achieved by describing the data as a signed value that is guaranteed to be positive by another sample storing a virtual 0 bit in bit 9 (so the data is treated as a signed 10-bit quantity, of which the top and bottom bits are guaranteed to be 0). The cost of this flexibility is that multiple samples are needed per channel, which is why the common case of chroma channels which should map to the -0.5..0.5 range are treated specially.
The storage of the palette is considered to be outside the remit of this specification; however, the “format” describes both the encoding of the bits which index into the palette and the format of the entries in the palette itself.
The convention for encoding paletted formats was different in revisions of this specification prior to 1.3. |
If the bitOffset field of any of the samples equals eight times the total number of bytes indicated by the bytesPlane channels (that is, indicating an offset after the end of the logical bit stream), this sample indicates a palette entry. Samples with offsets within the range of the logical bit stream describe the index of the palette; sampleUpper is used to indicate the number of entries in the palette (which may be lower than could be addressed by the number of bits available). The index can be comprised of multiple samples, supporting non-contiguous bits.
The four bits encoding channelType indicate a choice of palette in the index samples. The palette entries indicate which palette they are associated with by encoding the corresponding palette id in the samplePosition0 field, with the other samplePosition fields set to 0. This approach allows both a simple palette in which each entry represents a complete color and per-channel look-up tables.
The index samples should be described first, as though the format were not paletted; samples for palette entries should then follow, sorted first by palette id, then by channel id.
The descriptor’s colorModel, colorPrimaries, transferFunction and flags apply to the palette entries. The texelBlockDimension values on the other hand refer to the storage of per-texel indices.
Table 94 shows a 240-entry palette with an index encoded in eight bits, and six bits for each of R, G and B per palette entry. Table 95 shows a 24-bit format with 256-entry palettes indexed by each of three 256-bit channels, with a 10-bit R, G and B palette associated with the corresponding channels.
The data format descriptor can be a convenient representation for describing some data which does not follow the constraint that texel blocks are of a fixed size. For example, it may be useful to use a descriptor to encode the color space of an image which is compressed with a variable-rate compressor (meaning there is no data-independent mapping between memory locations and corresponding pixels).
There are two ways to do this, each with its own uses. Only the color primaries, transfer function and flags may be encoded or these and the presence of color and alpha samples may be encoded.
In both cases an unsized format is indicated by setting all bytesPlane values to 0.
In the first case, KHR_DF_MODEL_UNSPECIFIED
is used in a
descriptor with no samples.
All texelBlockDimension values must be zero.
In the second case an appropriate color model is used,
e.g. KHR_DF_MODEL_RGBSDA
.
The presence of color and alpha samples is used to indicate the
presence of these samples in the image.
The texelBlockDimension values and bitOffset, bitLength and
samplePosition fields of the samples can all be set to 0 or the values
prior to variable-rate compression can be retained.
The basic descriptor block has been specified in terms of an array of
uint32_t
value.
C99 provides a more direct representation, but this relies on the bit
ordering of bitfields (which is implementation-defined) and flexible
array members, which are not supported in C++.
In the interests of portability, the following summary (which assumes that bitfields are encoded starting at bit 0) is therefore provided for information, but is not canonical:
typedef struct _DFDSampleType { uint32_t bitOffset: 16; uint32_t bitLength: 8; uint32_t channelType: 8; // Includes qualifiers uint32_t samplePosition0: 8; uint32_t samplePosition1: 8; uint32_t samplePosition2: 8; uint32_t samplePosition3: 8; uint32_t lower; uint32_t upper; } DFDSampleType; typedef struct _BasicDataFormatDescriptor { uint32_t vendorId: 17; uint32_t descriptorType: 15; uint32_t model: 8; uint32_t primaries: 8; uint32_t transfer: 8; uint32_t flags: 8; uint32_t texelBlockDimension0: 8; uint32_t texelBlockDimension1: 8; uint32_t texelBlockDimension2: 8; uint32_t texelBlockDimension3: 8; uint32_t bytesPlane0: 8; uint32_t bytesPlane1: 8; uint32_t bytesPlane2: 8; uint32_t bytesPlane3: 8; uint32_t bytesPlane4: 8; uint32_t bytesPlane5: 8; uint32_t bytesPlane6: 8; uint32_t bytesPlane7: 8; DFDSampleType samples[]; } BasicDataFormatDescriptor;
Some formats will require more channels than can be described in the Basic Format Descriptor, or may have more specific color requirements. For example, it is expected than an extension will be available which places an ICC color profile block into the descriptor block, allowing more color channels to be specified in more precise ways. This will significantly enlarge the space required for the descriptor, and is not expected to be needed for most common uses. A vendor may also use an extension block to associate metadata with the descriptor — for example, information required as part of hardware rendering. So long as software which uses the data format descriptor always uses the totalSize field to determine the size of the descriptor, this should be transparent to user code.
The extension mechanism is the preferred way to support even simple extensions such as additional color spaces transfer functions that can be supported by an additional enumeration. This approach improves compatibility with code which is unaware of the additional values. Simple extensions of this form that have cross-vendor support have a good chance of being incorporated more directly into future revisions of the specification, allowing application code to distinguish them by the versionId field.
If bit 13, KHR_DF_KHR_DESCRIPTORTYPE_NEEDED_FOR_WRITE_BIT
, of the
descriptorType field of an extension is set, an application must
understand the extension in order to write data coherently.
If this bit is clear, copying the bits which correspond to one texel to
another can be expected to result in a correct transfer of the texel even
if the application does not understand the extension.
If bit 14, KHR_DF_KHR_DESCRIPTORTYPE_NEEDED_FOR_DECODE_BIT
, of the
descriptorType field of an extension is set, an application must
understand the extension in order to decode the contents of the texel
coherently.
If this bit is clear, the data held in the extension can be considered to be
“informative” and that ignoring the extension will still result in correct
values to the extent specified by the basic descriptor block.
For example, an extension may associate an absolute brightness level with
a format, but software which does not have need of this concept can continue
processing the texel contents correctly.
As an example of the description of an extension, consider a single-channel 32-bit depth buffer, as shown in Table 86. A tiled renderer may wish to indicate that this buffer is “virtual”: it will be allocated real memory only if needed, and will otherwise exist only a subset at a time in an on-chip representation. Someone developing such a renderer may choose to add a vendor-specific extension (with ID 0x1FFFF to indicate development work and avoid the need for a vendor ID) which uses a boolean to establish whether this depth buffer exists only in virtual form. Note that the mere presence or absence of this extension within the data format descriptor itself forms a boolean, but for this example we will assume that an extension block is always present, and that a boolean is stored within. We will give the enumeration 32 bits, in order to simplify the possible addition of further extensions and pad to the alignment requirements.
In this example (which should not be taken as an implementation suggestion),
the data descriptor would first contain a descriptor block describing the
depth buffer format as conventionally described, followed by a second
descriptor block that contains only the enumeration.
The descriptor itself has a totalSize that includes both of these
descriptor blocks.
Note that KHR_DF_KHR_DESCRIPTORTYPE_NEEDED_FOR_WRITE_BIT
is not
set, indicating that depth data can be written without knowing about the
extension, and KHR_DF_KHR_DESCRIPTORTYPE_NEEDED_FOR_DECODE_BIT
is not
set, indicating that software can safely ignore the information about the
form of allocation while reading texel values.
Table 86. Example of a depth buffer with an extension to indicate a virtual allocation
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 56 — total size of the two blocks plus one 32-bit value | |||||||||||||||||||||||||||||||
Basic descriptor block | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 1) = 40 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags | transferFunction | colorPrimaries | colorModel | ||||||||||||||||||||||||||||
|
|
|
| ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 4 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for depth | ||||||||||||||||||||||||||
1 | 1 | 0 | 0 |
| bitLength: 31 (= “32”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0xBF800000U — -1.0f | |||||||||||||||||||||||||||||||
sampleUpper: 0x7F800000U — 1.0f | |||||||||||||||||||||||||||||||
Extension descriptor block | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0x1FFFF | ||||||||||||||||||||||||||||||
descriptorBlockSize: 8 + (4 × 1) = 12 | versionNumber: 0 | ||||||||||||||||||||||||||||||
Data specific to the extension follows | |||||||||||||||||||||||||||||||
1 — buffer is “virtual” |
It is possible for a vendor to use the extension block to store peripheral information required to access the image — plane base addresses, stride, etc. Since different implementations have different kinds of non-linear ordering and proprietary alignment requirements, this is not described as part of the standard. By many conventional definitions, this information is not part of the “format”, and particularly it ensures that an identical copy of the image will have a different descriptor block (because the addresses will have changed) and so a simple bitwise comparison of two descriptor blocks will disagree even though the “format” matches. Additionally, many APIs will use the format descriptor only for external communication, and have an internal representation that is more concise and less flexible. In this case, it is likely that address information will need to be represented separately from the format anyway. For these reasons, it is an implementation choice whether to store this information in an extension block, and how to do so, rather than being specified in this standard.
Under some relatively unusual circumstances, either the number of planes
described by the basic descriptor block or the number of bytes that can be
contributed to a texel may be insufficient to describe the memory layout of bulk
data.
For example, a format may describe 12-bit colors (4 bits each or red, green and
blue), and have the contributing bits stored as separate planes.
An extension descriptor block, with vendorId = KHR_DF_VENDORID_KHRONOS
and descriptorType = KHR_DF_DESCRIPTORTYPE_ADDITIONAL_PLANES
,
describes additional planes.
Table 87. Additional planes descriptor block
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
descriptorType: 0x6001 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 8 + (4 × number of planes) | versionNumber: 0 | ||||||||||||||||||||||||||||||
bytesPlane0 | |||||||||||||||||||||||||||||||
bytesPlane1 (optional) | |||||||||||||||||||||||||||||||
(etc.) |
If this descriptor block is present, the bytesPlane[0..7] fields of the basic descriptor block are ignored, and the number of bytes for plane n is taken directly from word n + 2 of the descriptor block. The number of planes described can be determined by the descriptor block size.
This extension both allows an arbitrary number of planes and makes it easy to specify planes that contribute a large number of bytes to the virtual bit stream. Note that the sample bitOffset field remains limited to 16 bits, so the total texel block memory footprint is limited.
Since knowing the bit stream contribution from all planes is
necessary when interpreting data, this descriptorType sets
|
This descriptor block should be used only if the Khronos Basic Descriptor Block is the first descriptor block in the data format descriptor, and cannot represent the format without extension.
The basic descriptor block allows texel blocks of up to four non-trivial
dimensions, and with a texel block size of up to 256 coordinate units,
with sample positions described in precision up to
$1\over 256$
of a coordinate.
Under rare circumstances, this may provide insufficient flexibility.
An extension descriptor block, with vendorId = KHR_DF_VENDORID_KHRONOS
and descriptorType = KHR_DF_DESCRIPTORTYPE_ADDITIONAL_DIMENSIONS
,
describes additional dimensions.
Note that in some cases where this solution might be useful, a texel block
is an inappropriate unit.
For example, this extension block allows the direct representation of
large texel tiling patterns, but it does so in a manner that is very
inefficient, having much more potential flexibility than is needed by the
users of the layout being described.
Table 88. Additional dimensions descriptor block
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
descriptorType: 0x6002 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 8+(4 × dimensions × (samples+1)) | versionNumber: 0 | ||||||||||||||||||||||||||||||
| |||||||||||||||||||||||||||||||
| |||||||||||||||||||||||||||||||
(etc.) | |||||||||||||||||||||||||||||||
|
| ||||||||||||||||||||||||||||||
|
| ||||||||||||||||||||||||||||||
(etc.) |
The fields texelBlockDimension[0..n] describe the size in coordinate units of the texel block in the corresponding dimension; as with the basic descriptor block, the value stored is the corresponding dimension minus 1 (so a stored value of “0” corresponds to a dimension of 1). The texelBlockDimension[0..7] fields of the basic descriptor block are ignored.
For each sample, the samplePos and samplePosDivisor fields store a numerator and denominator pair for the coordinate $\textit{offset} = {\textit{numerator}\over\textit{denominator}}$ for each dimension, with all dimensions for a sample described before describing the next sample. samplePos and samplePosDivisor should be minimized, leaving an irreducible fraction. The samplePos fields of the basic descriptor block are ignored. These fields are present even for samples corresponding to palette entries.
Since the number of samples in the texel block can be deduced from the size of the basic descriptor block, the number of dimensions that are to be described by the additional dimensions descriptor block can be deduced from its size: numDimensions = (descriptorBlockSize - 8) / (numSamples + 1).
Since knowing the mapping from coordinates to samples is
necessary when interpreting data, this descriptorType sets
|
This descriptor block should be used only if the Khronos Basic Descriptor Block is the first descriptor block in the data format descriptor, and cannot represent the format without extension.
While it is not expected that every new container will have a unique data descriptor or that analysis of the data format descriptor will be on a critical path in an application, it is still expected that comparison between formats may be time-sensitive. The data format descriptor is designed to allow relatively efficient queries for subsets of properties, to allow a large number of format descriptors to be stored, and to be amenable to hardware interpretation or processing in shaders. These goals preclude a text-based representation such as an XML schema.
Formats in FourCC.org do not describe in detail sufficient information for many APIs, and are sometimes inconsistent.
Enumerations are fast and easy to process, but are limited in that any software can only be aware of the enumeration values in place when it was defined. Software often behaves differently according to properties of a format, and must perform a look-up on the enumeration — if it knows what it is — in order to change behaviors. A descriptive format allows for more flexible software which can support a wide range of formats without needing each to be listed, and simplifies the programming of conditional behavior based on format properties.
Khronos supports multiple standards that have a range of internal data representations. There is no requirement that this standard be used specifically with other Khronos standards, but it is hoped that multiple Khronos standards may use this specification as part of a consistent approach to inter-standard operation.
While a library may not use all the data provided in the data format descriptor that is described within this standard, it is common for users of data — particularly pixel-like data — to have additional requirements. Capturing these requirements portably reduces the need for additional metadata to be associated with a proprietary descriptor. It is also common for additional functionality to be added retrospectively to existing libraries — for example, Y′CBCR support is often an afterthought in rendering APIs. Having a consistent and flexible representation in place from the start can reduce the pain of retrofitting this functionality.
Note that there is no expectation that the format descriptor from this standard be used directly, although it can be. The impact of providing a mapping between internal formats and format descriptors is expected to be low, but offers the opportunity both for simplified access from software outside the proprietary library and for reducing the effort needed to provide a complete, unambiguous and accurate description of a format in human-readable terms.
There is a trade-off between size and decoding effort. It is assumed that data which occupies the same 32-bit word may need to be tested concurrently, reducing the cost of comparisons. When transferring data formats, the packing reduces the overhead. Within these constraints, it is intended that most data can be extracted with low-cost operations, typically being byte-aligned (other than sample flags) and with the natural alignment applied to multi-byte quantities.
For simple ASCII content, there is no reason that plain text could not be described in some way, and this may be useful for image formats that contain comment sections. However, since many multilingual text representations do not have a fixed character size, this use is not seen as an obvious match for this standard.
Some common floating-point numeric representations are defined in [IEEE 754]. Additional floating point formats are defined in this section.
A 16-bit floating-point number has a 1-bit sign (S), a 5-bit exponent (E), and a 10-bit mantissa (M). The value V of a 16-bit floating-point number is determined by the following:
If the floating-point number is interpreted as an unsigned 16-bit integer N, then
An unsigned 11-bit floating-point number has no sign bit, a 5-bit exponent (E), and a 6-bit mantissa (M). The value V of an unsigned 11-bit floating-point number is determined by the following:
If the floating-point number is interpreted as an unsigned 11-bit integer N, then
An unsigned 10-bit floating-point number has no sign bit, a 5-bit exponent (E), and a 5-bit mantissa (M). The value V of an unsigned 10-bit floating-point number is determined by the following:
If the floating-point number is interpreted as an unsigned 10-bit integer N, then
Rather than attempting to enumerate every possible floating-point
format variation in this specification, the data format descriptor
can be used to describe the components of arbitrary floating-point
data, as follows. Note that non-standard floating point formats do
not use the KHR_DF_SAMPLE_DATATYPE_FLOAT
bit.
An example of use of the 16-bit floating point format described in Section 10.1 but described in terms of a custom floating point format is provided in Table 103. Note that this is provided for example only, and this particular format would be better described using the standard 16-bit floating point format as documented in Table 104.
The mantissa of a custom floating point format should be represented
as an integer channelType. If the mantissa represents a signed
quantity encoded in two’s complement, the KHR_DF_SAMPLE_DATATYPE_SIGNED
bit should be set. To encode a signed mantissa represented in
sign-magnitude format, the main part of the mantissa should be
represented as an unsigned integer quantity (with
KHR_DF_SAMPLE_DATATYPE_SIGNED
not set), and an additional one-bit
sample with KHR_DF_SAMPLE_DATATYPE_SIGNED
set should be used
to identify the sign bit. By convention, a sign bit should be encoded
in a later sample than the corresponding mantissa.
The sampleUpper and sampleLower values for the mantissa should be set to indicate the representation of 1.0 and 0.0 (for unsigned formats) or -1.0 (for signed formats) respectively when the exponent is in a 0 position after any bias has been corrected. If there is an implicit “1” bit, these values for the mantissa will exceed what can be represented in the number of available mantissa bits.
For example, the shared exponent formats shown in Table 98 does not have an implicit “1” bit, and therefore the sampleUpper values for the 9-bit mantissas are 256 — this being the mantissa value for 1.0 when the exponent is set to 0.
For the 16-bit signed floating point format described in Section 10.1, sampleUpper should be set to 1024, indicating the implicit “1” bit which is above the 10 bits representable in the mantissa. sampleLower should be 0 in this case, since the mantissa uses a sign-magnitude representation.
By convention, the sampleUpper and sampleLower values for a sign bit are 0 and -1 respectively.
The KHR_DF_SAMPLE_DATATYPE_EXPONENT
bit should be set in a sample
which contains the exponent of a custom floating point format.
The sampleLower for the exponent should indicate the exponent bias. That is, the mantissa should be scaled by two raised to the power of the stored exponent minus this sampleLower value.
The sampleUpper for the exponent indicates the maximum legal exponent value. Values above this are used to encode infinities and not-a-number (NaN) values. sampleUpper can therefore be used to indicate whether or not the format supports these encodings.
Floating point values encoded with an exponent of 0 (before bias) and a mantissa of 0 are used to represent the value 0. An explicit sign bit can distinguish between +0 and -0.
Floating point values encoded with an exponent of 0 (before bias) and a non-zero mantissa are assumed to indicate a denormalized number, if the format has an implicit “1” bit. That is, when the exponent is 0, the “1” bit becomes explicit and the exponent is considered to be the negative sample bias minus one.
Floating point values encoded with an exponent larger than the exponent’s sampleUpper value and with a mantissa of 0 are interpreted as representing +/- infinity, depending on the value of an explicit sign bit. Note that in some formats, no exponent above sampleUpper is possible — for example, Table 98.
Floating point values encoded with an exponent larger than the exponent’s sampleUpper value and with a mantissa of non-0 are interpreted as representing not-a-number (NaN).
Note that these interpretations are compatible with the corresponding numerical representations in [IEEE 754].
Given an optional sign bit S, a mantissa value of M and an exponent value of E, a format with an implicit “1” bit can be converted from its representation to a real value as follows:
If there is no implicit “1” bit (that is, the sampleUpper value of the mantissa is representable in the number of bits assigned to the mantissa), the value can be converted to a real value as follows:
A descriptor block for a format without an implicit “1” (and with the added complication of having the same exponent bits shared across multiple channels, which is why an implicit “1” bit does not make sense) is shown in Table 98. In the case of this particular example, the above equations simplify to:
Where:
Note that in general conversion from a real number to any representation may require rounding, truncation and special value management rules which are beyond the scope of a data format specification and may be documented in APIs which generate these formats.
Example data format descriptors for compressed formats can be found under the colorModel field in Section 5.6. |
Table 89. 565 RGB packed 16-bit format as written to memory by a little-endian architecture
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 76 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 3) = 72 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 2 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | First sample: low five bits blue | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 4 (= “5”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 31 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Second sample: middle six bits green | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 5 = (“6”) | bitOffset: 6 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 63 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Third sample: top five bits red | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 4 (= “5”) | bitOffset: 11 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 31 |
Table 90. Four co-sited 8-bit sRGB channels, assuming premultiplied alpha
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 92 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 4) = 88 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 4 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the first sample | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 7 (= “8”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 255 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the second sample | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 7 (= “8”) | bitOffset: 8 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 255 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the third sample | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 7 (= “8”) | bitOffset: 16 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 255 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the fourth sample | ||||||||||||||||||||||||||
0 | 0 | 0 | 1 |
| bitLength: 7 (= “8”) | bitOffset: 24 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 255 |
Table 91. A single 8-bit monochrome channel
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 44 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 1) = 40 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 1 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 8 | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 255 |
Table 92. A single 1-bit monochrome channel, as an 8×1 texel block to allow byte-alignment
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 156 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 8) = 142 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 7 (= “8”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 1 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the first sample | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 0 (= “1”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 1 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the second sample | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 0 (= “1”) | bitOffset: 1 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0x20 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 1 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the third sample | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 0 (= “1”) | bitOffset: 2 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0x40 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 1 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the fourth sample | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 0 (= “1”) | bitOffset: 3 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0x60 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 1 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the fifth sample | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 0 (= “1”) | bitOffset: 4 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0x80 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 1 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the sixth sample | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 0 (= “1”) | bitOffset: 5 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0xA0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 1 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the seventh sample | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 0 (= “1”) | bitOffset: 6 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0xC0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 1 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the eighth sample | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 0 (= “1”) | bitOffset: 7 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0xE0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 1 |
Table 93. 2×2 Bayer pattern: four 8-bit distributed sRGB channels, spread across two lines (as two planes)
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 92 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 4) = 88 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 1 (height = “2”) | 1 (width = “2”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 2 | bytesPlane0: 2 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the first sample | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 7 (= “8”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0x00 (y = “0”) | 0x00 (x = “0”) | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 255 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the second sample | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 7 (= “8”) | bitOffset: 8 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0x00 (y = “0”) | 0x80 (x = “1”) | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 255 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the third sample | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 7 (= “8”) | bitOffset: 16 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0x80 (y = “1”) | 0x00 (x = “0”) | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 255 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the fourth sample | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 7 (= “8”) | bitOffset: 24 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0x80 (y = “1”) | 0x80 (x = “1”) | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 255 |
Table 94. Simple paletted format: 8-bit index to 240 entries of 4-bit R, G, B channels
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 92 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 4) = 88 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 1 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the 8-bit index | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 | 0 | bitLength: 7 (= “8”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 239 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Red channel of the palette entries | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 3 (= “4”) | bitOffset: 8 (palette entry) | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 (palette id) | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 15 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Green channel of the palette entries | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 3 (= “4”) | bitOffset: 8 (palette entry) | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 (palette id) | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 15 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Blue channel of the palette entries | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 3 (= “4”) | bitOffset: 8 (palette entry) | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 (palette id) | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 15 |
Table 95. Paletted color look-up table format: three 8-bit indices into separate 256-entry 10-bit channels
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 124 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 6) = 120 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 4 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Red channel palette index | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 7 (= “8”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 255 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Green channel palette index | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 7 (= “8”) | bitOffset: 8 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 255 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Blue channel palette index | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 7 (= “8”) | bitOffset: 16 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 255 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Red palette entry | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 9 (= “10”) | bitOffset: 32 (palette entry) | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 ( | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 1023 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Green palette entry | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 9 (= “10”) | bitOffset: 32 (palette entry) | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 1 ( | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 1023 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Blue palette entry | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 9 (= “10”) | bitOffset: 32 (palette entry) | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 2 ( | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 1023 |
Table 96. Y′CBCR 4:2:0: BT.709 reduced-range data, with CB and CR aligned to the midpoint of the Y′ samples
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 124 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 6) = 120 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 1 (height = “2”) | 1 (width = “2”) | ||||||||||||||||||||||||||||
bytesPlane3: 1 | bytesPlane2: 1 | bytesPlane1: 2 | bytesPlane0: 2 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the first Y′ sample | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 7 (= “8”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0x00 (y = “0.0”) | 0x00 (x = “0.0”) | ||||||||||||||||||||||||||||
sampleLower: 16 | |||||||||||||||||||||||||||||||
sampleUpper: 235 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the second Y′ sample | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 7 (= “8”) | bitOffset: 8 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0x00 (y = “0.0”) | 0x80 (x = “1.0”) | ||||||||||||||||||||||||||||
sampleLower: 16 | |||||||||||||||||||||||||||||||
sampleUpper: 235 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the third Y′ sample | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 7 (= “8”) | bitOffset: 16 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0x80 (y = “1.0”) | 0x00 (x = “0.0”) | ||||||||||||||||||||||||||||
sampleLower: 16 | |||||||||||||||||||||||||||||||
sampleUpper: 235 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the fourth Y′ sample | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 7 (= “8”) | bitOffset: 24 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0x80 (y = “1.0”) | 0x80 (x = “1.0”) | ||||||||||||||||||||||||||||
sampleLower: 16 | |||||||||||||||||||||||||||||||
sampleUpper: 235 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the U sample | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 7 (= “8”) | bitOffset: 32 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0x40 (y = “0.5”) | 0x40 (x = “0.5”) | ||||||||||||||||||||||||||||
sampleLower: 16 | |||||||||||||||||||||||||||||||
sampleUpper: 240 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the V sample | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 7 (= “8”) | bitOffset: 40 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0x40 (y = “0.5”) | 0x40 (x = “0.5”) | ||||||||||||||||||||||||||||
sampleLower: 16 | |||||||||||||||||||||||||||||||
sampleUpper: 240 |
Table 97. 565 RGB packed 16-bit format as written to memory by a big-endian architecture
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 92 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 4) = 88 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 2 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Bit 0 is green, green bits 0..2 in bits 5..7 of the higher byte address (less-significant half of the big-endian word) | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 2 (= “3”) | bitOffset: 13 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 — lower 3 bits of 6-bit value 0 | |||||||||||||||||||||||||||||||
sampleUpper: 7 — lower 3 bits of 6-bit value 63 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Continuation: green bits 3..5 in bits 0..2 of the lower byte address (more-significant half of the big-endian word) | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 2 (= “3”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 — upper 3 bits of 6-bit value 0 | |||||||||||||||||||||||||||||||
sampleUpper: 7 — upper 3 bits of 6-bit value 63 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Next undescribed bit: red in bits 3..7 of the lower byte address (more-significant half of the big-endian word) | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 4 (= “5”) | bitOffset: 3 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 31 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Remaining bits: blue in bits 0..4 of the higher byte address (less-significant half of the big-endian word) | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 4 (= “5”) | bitOffset: 8 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 31 |
Table 98. R9G9B9E5 shared-exponent format
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 124 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 6) = 120 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 4 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the R mantissa | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 8 (= “9”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 256 — mantissa at 1.0 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the R exponent | ||||||||||||||||||||||||||
0 | 0 | 1 | 0 |
| bitLength: 4 (= “5”) | bitOffset: 27 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 15 — exponent bias | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the G mantissa | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 8 (= “9”) | bitOffset: 9 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 256 — mantissa at 1.0 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the G exponent | ||||||||||||||||||||||||||
0 | 0 | 1 | 0 |
| bitLength: 4 (= “5”) | bitOffset: 27 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 15 — exponent bias | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the B mantissa | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 8 (= “9”) | bitOffset: 18 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 256 — mantissa at 1.0 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information for the B exponent | ||||||||||||||||||||||||||
0 | 0 | 1 | 0 |
| bitLength: 4 (= “5”) | bitOffset: 27 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 15 — exponent bias |
Table 99. Acorn 256-color format (2 bits each independent RGB, 2 bits shared tint)
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 108 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 5) = 104 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 1 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | First sample: four bits of red starting at bit 0, including two shared bits 0..1 | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 3 (= “4”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 15 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Second sample: two bits of green “tint” shared with red | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 1 (= “2”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 3 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Third sample: two bits unique to green | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 1 (= “2”) | bitOffset: 4 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 3 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Fourth sample: two bits of blue “tint” shared with red | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 1 (= “2”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 3 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Fifth sample: two bits unique to blue | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 1 (= “2”) | bitOffset: 6 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 3 |
Table 100. V210 format (full-range Y′CBCR)
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 220 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 12) = 216 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 5 (width = “6”) | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 16 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | First sample: shared mid-sited U0/U1 | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 10 | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0x10 (x = “0.5”) | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 1023 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Second sample: Y′0 | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 9 (= “10”) | bitOffset: 10 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0x00 (x = “0.0”) | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 1023 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Third sample: shared mid-sited V0/V1 | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 9 (= “10”) | bitOffset: 20 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0x10 (x = “0.5”) | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 1023 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Fourth sample: Y′1 | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 9 (= “10”) | bitOffset: 32 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0x20 (x = “1.0”) | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 1023 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Fifth sample: shared mid-sited U2/U3 | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 9 (= “10”) | bitOffset: 42 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0x50 (x = “2.5”) | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 1023 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Sixth sample: Y′2 | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 9 (= “10”) | bitOffset: 52 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0x40 (x = “2.0”) | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 1023 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Seventh sample: shared mid-sited V2/V3 | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 9 (= “10”) | bitOffset: 64 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0x50 (x = “2.5”) | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 1023 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Eighth sample: Y′3 | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 9 (= “10”) | bitOffset: 74 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0x60 (x = “3.0”) | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 1023 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Ninth sample: shared mid-sited U4/U5 | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 9 (= “10”) | bitOffset: 84 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0x90 (x = “4.5”) | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 1023 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Tenth sample: Y′4 | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 9 (= “10”) | bitOffset: 96 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0x80 (x = “4.0”) | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 1023 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Eleventh sample: shared mid-sited V4/V5 | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 9 (= “10”) | bitOffset: 106 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0x90 (x = “4.5”) | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 1023 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Twelfth sample: Y′5 | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 9 (= “10”) | bitOffset: 116 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0xA0 (x = “5.0”) | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 1023 |
Table 101. Intensity-alpha format showing aliased samples
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 92 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 4) = 88 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 1 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | First sample: red | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 7 (= “8”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 255 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Second sample: green | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 7 (= “8”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 255 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Third sample: blue | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 7 (= “8”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 255 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Fourth sample: alpha | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 7 (= “8”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 255 |
Table 102. A 48-bit signed middle-endian red channel: three co-sited 16-bit little-endian words, high word first
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 76 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 3) = 72 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 6 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | First sample: bits 0..15 of the 48-bit value, in memory bytes 4..5 | ||||||||||||||||||||||||||
0 | 1 | 0 | 0 |
| bitLength: 15 (= “16”) | bitOffset: 32 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0xFFFF0000U — bottom 16 bits of 0x800000000000, sign-extended | |||||||||||||||||||||||||||||||
sampleUpper: 0x0000FFFFU — bottom 16 bits of 0x7FFFFFFFFFFF | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Second sample: bits 16..31 of the 48-bit value, in memory bytes 2..3 | ||||||||||||||||||||||||||
0 | 1 | 0 | 0 |
| bitLength: 15 (= “16”) | bitOffset: 16 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0xFFFF0000U — middle 16 bits of 0x800000000000, sign-extended | |||||||||||||||||||||||||||||||
sampleUpper: 0x0000FFFFU — middle 16 bits of 0x7FFFFFFFFFFF | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Third sample: bits 32..47 of the 48-bit value, in memory bytes 0..1 | ||||||||||||||||||||||||||
0 | 1 | 0 | 0 |
| bitLength: 15 (= “16”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0xFFFF8000U — top 16 bits of 0x800000000000, sign-extended | |||||||||||||||||||||||||||||||
sampleUpper: 0x00007FFFU — top 16 bits of 0x7FFFFFFFFFFF |
Table 103. A single 16-bit floating-point red value, described explicitly (example only!)
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 76 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 3) = 72 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 2 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | First sample: mantissa | ||||||||||||||||||||||||||
0 | 0 | 0 | 0 |
| bitLength: 9 (= “10”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0 | |||||||||||||||||||||||||||||||
sampleUpper: 1024 — implicit 1.0 bit since 1024 > 2bitLength - 1 | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Second sample: sign of mantissa | ||||||||||||||||||||||||||
0 | 1 | 0 | 0 |
| bitLength: 0 (= “1”) | bitOffset: 15 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 1 — 1-bit “-1” | |||||||||||||||||||||||||||||||
sampleUpper: 0 — 1-bit “0” | |||||||||||||||||||||||||||||||
F | S | E | L | channelType | Third sample: exponent | ||||||||||||||||||||||||||
0 | 0 | 1 | 0 |
| bitLength: 4 (= “5”) | bitOffset: 10 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 15 — bias | |||||||||||||||||||||||||||||||
sampleUpper: 30 — support for infinities because 30 < 2bitLength - 1 |
Table 104. A single 16-bit floating-point red value, described normally
| |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
totalSize: 44 | |||||||||||||||||||||||||||||||
descriptorType: 0 | vendorId: 0 | ||||||||||||||||||||||||||||||
descriptorBlockSize: 24 + (16 × 3) = 40 | versionNumber: 2 | ||||||||||||||||||||||||||||||
flags: | transferFunction: | colorPrimaries: | colorModel: | ||||||||||||||||||||||||||||
texelBlockDimension3 | texelBlockDimension2 | texelBlockDimension1 | texelBlockDimension0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
bytesPlane3: 0 | bytesPlane2: 0 | bytesPlane1: 0 | bytesPlane0: 2 | ||||||||||||||||||||||||||||
bytesPlane7: 0 | bytesPlane6: 0 | bytesPlane5: 0 | bytesPlane4: 0 | ||||||||||||||||||||||||||||
F | S | E | L | channelType | Sample information | ||||||||||||||||||||||||||
1 | 1 | 0 | 0 |
| bitLength: 15 (= “16”) | bitOffset: 0 | |||||||||||||||||||||||||
samplePosition3 | samplePosition2 | samplePosition1 | samplePosition0 | ||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | ||||||||||||||||||||||||||||
sampleLower: 0xBF800000U — IEEE 754 floating-point representation for -1.0f | |||||||||||||||||||||||||||||||
sampleUpper: 0x3F800000U — IEEE 754 floating-point representation for 1.0f |
A “color space” determines the meaning of decoded numerical color values: that is, it is distinct from the bit patterns, compression schemes and locations in memory used to store the data.
A color space consists of three basic components:
Transfer functions define the relationships between linear intensity and linear numbers in the encoding scheme. Since the human eye’s sensitivity to changes in intensity is non-linear, a non-linear encoding scheme typically allows improved visual quality at reduced storage cost.
Color primaries define the spectral response of a “pure color” in an additive color model - typically, what is meant by “red”, “green” and “blue” for a given system, and (allowing for the relative intensity of the primaries) consequently define the system’s white balance.
Color models describe the distinction between a color representation and additive colors. Since the human visual system treats differences in absolute intensity differently from differences in the spectrum composing a color, many formats benefit from transforming the color representation into one which can separate these aspects of color. Color models are frequently “named” by listing their component color channels.
Conversion between color representations may require a number of separate conversion operations:
Details of these conversion operations are described in the following chapters.
As described in the License Information at the start of this document, the Khronos Data Format Specification does not convey a right to implement the operations it describes. This is particularly true of the conversion formulae in the following sections, whose inclusion is purely informative. Please refer to the originating documents and the bodies responsible for the standards containing these formulae for the legal framework required for implementation. |
Common cases such as converting a Y′CBCR image encoded for 625-line BT.601 to a Y′CBCR image encoded for BT.709 can involve multiple costly operations. An example is shown in the following diagram, which represents sampling from a Y′CBCR texture in one color space, and the operations needed to generate a different set of Y′CBCR values representing the color of the sample position in a different color space:
In this diagram, non-linear luma Y′ channels are shown in black and white, color difference CB/CR channels are shown with the colors at the extremes of their range, and color primary channels are shown as the primary color and black. Linear representations are shown diagonally divided by a straight line; non-linear representations are shown with a gamma curve. The luma and color difference representation is discussed in Section 15.1. The interpretation of color primaries is discussed in Section 14. Non-linear transfer functions are described in Section 13. As described below, the diagram shows a 2×3 grid of input chroma texel values, corresponding to a 4×6 grid of luma texel values, since the chroma channels are stored at half the horizontal and half the vertical resolution of the luma channel (i.e. in “4:2:0” representation). Grayed-out texel values do not contribute to the final output, and are shown only to indicate relative alignment of the coordinates.
The stages numbered in Figure 22 show the following operations:
Range expansion to the correct range for the values in the color model (handled differently, for example, for “full” and “narrow” ranges); in this example, the result is to increase the effective dynamic range of the encoding: contrast and saturation are increased.
In this example, operations 1 and 2 can be combined into a single sparse matrix multiplication of the input channels, although actual implementations may wish to take advantage of the sparseness.
Application of a transfer function to convert from non-linear R′G′B′ to linear RGB, using the color primaries of the input representation. In this case, the conversion might be the EOTF -1 described in Section 13.2.
The separation of stages 4 and 5 is specific to the Y′CBCR to R′G′B′ color model conversion. Other representations such as Y′CC′BCC′RC and ICTCP have more complex interactions between the color model conversion and the transfer function.
If the output is to be written to a frame buffer with reduced-resolution chroma channels, chroma values for multiple samples need to be combined. Note that it is easy to introduce inadvertent chroma blurring in this operation if the source space chroma values are generated by interpolation.
In this example, generating the four linear RGB values required for linear interpolation at the magenta cross position requires six chroma samples. In the example shown, all four Y′ values fall between the same two chroma sample centers on the horizontal axis, and therefore recreation of these samples by linear blending on the horizontal axis only requires two horizontally-adjacent samples. However, the upper pair of Y′ values are sited above the sample position of the middle row of chroma sample centers, and therefore reconstruction of the corresponding chroma values requires interpolation between the upper four source chroma values. The lower pair of Y′ values are sited below the sample position of the middle row of chroma sample centers, and therefore reconstruction of the corresponding chroma values requires interpolation between the lower four source chroma values. In general, reconstructing four chroma values by interpolation may require four, six or nine source chroma values, depending on which samples are required. The worst case is reduced if chroma samples are aligned (“co-sited”) with the luma values, or if chroma channel reconstruction uses replication (nearest-neighbor filtering) rather than interpolation.
An approximation to the conversion described in Figure 22 is depicted in Figure 23:
A performance-optimized approximation to our example conversion may use the following steps:
Since stages 1 and 2 represent an affine matrix transform, linear interpolation of input values may equivalently be performed before these operations. This observation allows stages 1..4 to be combined into a single matrix transformation. |
Large areas of constant color will be correctly converted by this approximation. However, there are two sources of errors near color boundaries:
This approximation produces identical results to the more explicit sequence of operations in two cases:
As another example, the conversion from BT.709-encoded Y′CBCR to sRGB R′G′B′ may be considered to be a simple model conversion (to BT.709 R′G′B′ non-linear primaries using the “ITU” OETF), since sRGB shares the BT.709 color primaries and is defined as a complementary EOTF intended to be combined with BT.709’s OETF. This interpretation imposes a $\gamma \approx$ 1.1 OOTF. Matching the OOTF of a BT.709-BT.1886 system, for which $\gamma \approx$ 1.2, implies using the BT.1886 EOTF to convert to linear light, then the sRGB EOTF -1 to convert back to sRGB non-linear space. Encoding linear scene light with linear OOTF means applying the BT.709 OETF -1; if the sRGB R′G′B′ target is itself intended to represent a linear OOTF, then the {R′sRGB, G′sRGB, B′sRGB} should be calculated as:
The transfer function describes the mapping between a linear numerical representation and a non-linear encoding. The eye is more sensitive to relative light levels than absolute light levels. That is, if one image region is twice as bright as another, this will be more visible than if one region is 10% brighter than another, even if the absolute difference in brightness is the same in both cases. To make use of the eye’s non-linear response to light to provide better image quality with a limited number of quantization steps, it is common for color encodings to work with a non-linear representation which dedicates a disproportionate number of bits to darker colors compared with lighter colors. The typical effect of this encoding is that mid-tones are stored with a larger (nearer-to-white) numerical value than their actual brightness would suggest, and that mid-values in the non-linear encoding typically represent darker intensities than their fraction of the representation of white would suggest.
The behavior has historically been approximated by a power function with an exponent conventionally called $\gamma$ : {R,G,B}non-linear = {R,G,B}linear $\gamma$ . Hence this conversion is colloquially known as gamma correction.
Many practical transfer functions incorporate a small linear segment near 0, instead of being a pure power function. This linearity reduces the required resolution for representing the conversion, especially where results must be reversible, and also reduces the noise sensitivity of the function in an analog context. When combined with a linear segment, the power function has a different exponent from the pure power function that best approximates the resulting curve. |
A consequence of this non-linear encoding is that many image processing operations should not be applied directly to the raw non-linearly-encoded numbers, but require conversion back to a linear representation. For example, linear color gradients will appear distorted unless the encoding is adjusted to compensate for the encoding; CGI lighting calculations need linear intensity values for operation, and filtering operations require texel intensities converted to linear form.
In the following example, the checker patterns are filtered in the right-most square of each row by averaging the checker colors, emulating a view of the pattern from a distance at which the individual squares are no longer distinct. The intended effect can be seen by viewing the diagram from a distance, or deliberately out of focus. The output is interpreted using the sRGB EOTF, approximating the behavior of a CRT with uncorrected signals. The background represents 50% gray.
Incorrectly-applied transfer functions can also introduce color shifts, as demonstrated by the saturation change in the following examples:
Figure 27. R′, G′, B′ channels and combined color gradient with non-linear sRGB encoding in each channel
A standard for image representation typically defines one or both of two types of transfer functions:
Some color models derive chroma (color difference) channels wholly or partly from non-linear intensities. It is common for image representations which use these color models to use a reduced-resolution representation of the chroma channels, since the human eye is less sensitive to high resolution chroma errors than to errors in brightness. Despite the color shift introduced by interpolating non-linear values, these chroma channels are typically resampled directly in their native, non-linear representation. |
The opto-optical transfer function (OOTF) of a system is the relationship between linear input light intensity and displayed output light intensity: OOTF(input) = EOTF(OETF(input)). It is common for the OOTF to be non-linear. For example, a brightly-lit scene rendered on a display that is viewed in a dimly-lit room will typically appear washed-out and lacking contrast, despite mapping the full range of scene brightness to the full range supported by the display. A non-linear OOTF can compensate for this by reducing the intensity of mid-tones, which is why television standards typically assume a non-linear OOTF: logical scene light intensity is not proportional to logical display intensity.
In the following diagram, the upper pair of images are identical, as are the lower pair of images (which have mid-tones darkened but the same maximum brightness). Adaptation to the surround means that the top left and lower right images look similar.
In the context of a non-linear OOTF, an application should be aware of whether operations on the image are intended to reflect the representation of colors in the scene or whether the intent is to represent the output color accurately, at least when it comes to the transfer function applied. For example, an application could choose to convert lighting calculations from a linear to non-linear representation using the OETF (to match the appearance of lighting in the scene), but to perform image scaling operations using the EOTF in order to avoid introducing intensity shifts due to filtering. Working solely with the EOTF or OETF results in ignoring the intended OOTF of the system.
In practice, the OOTF is usually near enough to linear that this distinction is subtle and rarely worth making for computer graphics, especially since computer-generated images may be designed to be viewed in brighter conditions which would merit a linear OOTF, and since a lot of graphical content is inherently not photo-realistic (or of limited realism, so that the transfer functions are not the most important factor in suspending disbelief). For video and photographic content viewed in darker conditions, the non-linearity of the OOTF is significant. The effect of a non-linear OOTF is usually secondary to the benefits of using non-linear encoding to reduce quantization.
By convention, non-linearly-encoded values are distinguished from linearly-encoded values by the addition of a prime (′) symbol. For example, (R,G,B) may represent a linear set of red, green and blue components; (R′,G′,B′) would represent the non-linear encoding of each value. Typically the non-linear encoding is applied to additive primary colors; derived color differences may or may not retain the prime symbol.
Charles Poynton provides a further discussion on “Gamma” in http://www.poynton.com/PDFs/TIDV/Gamma.pdf.
“ITU” is used in this context as a shorthand for the OETF shared by the current BT.601, BT.709 and BT.2020 family of standard dynamic range digital television production standards. The same OETF is shared by SMPTE 170M. The ITU does define other transfer functions, for example the PQ and HLG transfer functions described below (originating in BT.2100) and the list of EOTFs listed in BT.470-6. |
The ITU-T BT.601, BT.709 and BT.2020 specifications (for standard definition television, HDTV and UHDTV respectively), and SMPTE 170M, which defines NTSC broadcasts, define an opto-electrical transfer function. The (OETF) conversion from linear (R,G,B) encoding to non-linear (R′,G′,B′) encoding is:
Where α = 1.0993 and β = 0.0181 for 12-bit encoding in the BT.2020 specification, and $\alpha = 1.099$ and $\beta = 0.018$ otherwise.
From this the inverse (OETF -1) transformation can be deduced:
δ can be deduced from α × β0.45 - (α - 1) ≈ 0.0812. Note that this is subtly different from 4.5 × β due to rounding. See the following section for the derivation of these values.
SMPTE 170M-2004, which defines the behavior of NTSC televisions, defines the EOTF of the “reference reproducer” as the OETF -1 function above, with δ explicitly written as 0.0812. Therefore the SMPTE 170M-2004 EOTF -1 equals the OETF given above. The “reference camera” of SMTPE 170M-2004 has an OETF function matching that of the ITU specifications. That is, the OOTF of the system described in SMPTE 170M-2004 provides a linear mapping of captured scene intensity to display intensity: the SMPTE 170M-2004 OETF is described as being chosen to result in a linear OOTF on a typical display. This is distinct from the current ITU specifications, which assume a non-linear OOTF. SMPTE 170M-2004 also represents a change from the “assumed gamma” of 2.2 associated with most NTSC display devices as described in ITU-T BT.470-6 and BT.2043, although these standards also define a linear OOTF.
This “ITU” OETF is closely approximated by a simple power function with an exponent of 0.5 (and therefore the OETF -1 is quite closely approximated by a simple power function with an exponent of 2.0); the linear segment and offset mean that the best match is not the exponent of 0.45 that forms part of the exact equation. ITU standards deliberately chose a different transfer curve from that of a typical CRT in order to introduce a non-linear OOTF, as a means to compensate for the typically dim conditions in which a television is viewed. ITU BT.2087 refers to the approximation of the OETF with a square root $\left(\gamma = {1\over{2}}\right)$ function.
The following graph shows the close relationship between the ITU OETF (shown in red) and a pure power function with $\gamma={1\over{2}}$ (in blue). The difference between the curves is shown in black. The largest difference between the curve values at the same point when quantized to 8 bits is 15, mostly due to the sharp vertical gradient near 0.
SMPTE 170M-2004 contains a note that the OETF is a more “technically correct” definition of the transfer function, and compares it to a “transfer gradient (gamma exponent) of 2.2” in previous specifications, and that the OETF in older documents is described as “1/2.2 (0.455…)”. While both versions define a linear OOTF, there is no explicit mention that curve has substantially changed; this might be due to conflation of the 0.455 exponent in older specifications with the 0.45 exponent in the new formulae. The ITU OETF is actually a closer match to a gamma exponent of $1\over{2.0}$ , as shown above; it is a relatively poor match to a gamma exponent of $1\over{2.2}$ ; the following graph shows the difference between the ITU OETF (shown in red) and a pure power function with $\gamma={1\over{2.2}}$ (in blue). The difference between the curves is shown in black. |
Using the 12-bit encoding values for α and β from Rec.2020, there is an overlap around a non-linear value of 0.08145. In other cases, the conversion from linear to non-linear representation with encoding introduces a discontinuity between (0.018 × 4.500) = 0.081 and (1.099 × 0.0180.45 - 0.099) ≈ 0.0812, corresponding to roughly a single level in a 12-bit range. SMPTE 170M-2004 provides formulae for both transformations and uses 0.0812 as a case selector for the non-linear-to-linear transformation.
The values of α and β in the ITU function were apparently chosen such that the linear segment and power segment meet at the same value and with the same derivative (that is, the linear segment meets the power segment at a tangent). The α and β values can be derived as follows:
At {R,G,B} = β, the linear and non-linear segments of the curve must calculate the same value:
Additionally, the derivatives of the linear and non-linear segments of the curve must match:
The derivative can be rearranged to give the equation:
Substituting this into the original equation results in the following:
This simplifies to:
This can be solved numerically (for example by Newton-Raphson iteration), and results in values of:
The sRGB specification defines an electro-optical transfer function. The EOTF conversion from non-linear $(R', G', B')$ encoding to linear $(R, G, B)$ encoding is:
The corresponding sRGB EOTF -1 conversion from linear $(R, G, B)$ encoding to non-linear $(R', G', B')$ encoding is:
The sRGB EOTF approximates a simple power function with an exponent of 2.2, which is intended to be consistent with legacy CRT content, particularly for NTSC devices, and to approximate the expected EOTF for BT.709 content, given the implicit OOTF used in production video content. sRGB is distinct from ITU-T BT.1886, which offers a (different) reference EOTF for flat panels used for HDTV and is also intended to complement BT.709; in addition to the change in EOTF, sRGB specifies a reference display maximum luminance of 80cd/m2, compared with 100cd/m2 for BT.1886. sRGB is also distinct from SMPTE 170M, which defines its EOTF as the inverse of its (and BT.709’s) OETF.
The following graph compares the sRGB EOTF (in red) and a pure power function with $\gamma=2.2$ (in blue); the area between the two curves is shown in black. The largest non-linear difference at the same linear value when quantized to 8 bits is 3.
The sRGB standard assumes a quantization scheme in which 0.0 is represented by the value 0 and 1.0 is represented by 255. Despite the goal of complementing ITU-T Rec. BT.709, this is different from the ITU “full-range” encoding scheme defined in ITU-T Rec. BT.2100, which represents 1.0 as a power of two (not $2^n-1$ ) and therefore cannot exactly represent 1.0. |
The following graph shows the relationship between the sRGB EOTF (shown in red) and the ITU OETF (shown in blue). The result of applying the two functions in turn, resulting in the OOTF of a combined ITU-sRGB system, is shown in black. Since the sRGB EOTF approximates a power function with $\gamma=2.2$ and the ITU OETF approximates a power function with $\gamma=2.0$ , also shown in green is the resulting OOTF corresponding to a power function with $\gamma={2.2\over{2.0}}=1.1$ .
The original sRGB specification was defined only in terms of positive values between 0 and 1. Subsequent standards, such as scRGB annex B, use the same transfer function but expand the range to incorporate values less than 0 and greater than 1.0. In these cases, when the input channel to the conversion is negative, the output should be the negative version of the conversion applied to the absolute value of the input. That is:
scRGB annex B changes the behavior of the $\{R,G,B\} = 0.0031308$ case compared with the sRGB specification. Since both calculations agree to seven decimal places, this is unlikely to be significant in most applications. scRGB annex B does not define the EOTF -1, so the formulae below are derived by extending the sRGB formulae. |
sYCC includes a hint that a 1cd/m2 level of flare should be assumed for the reference 80cd/m2 output, and that the black level should therefore be assumed to be ${1\over 80} = 0.0125$ . It notes that the non-linear sRGB { $R',G',B'$ } values can be corrected as follows: \begin{align*}
E_{sYCC} &= \begin{cases}
0.0125 - \left({1-0.0125\over 1.055^{2.4}}\right)
\times (0.055 - E'_{sRGB})^{2.4},
& E'_{sRGB} \leq -0.04045\ [\textrm{sic}]\\
0.0125 + {1-0.0125\over 12.92} \times
E'_{sRGB},
& -0.04045 \leq E'_{sRGB} \leq 0.04045 \\
0.0125 + \left({1-0.0125\over 1.055^{2.4}}\right)
\times (0.055 + E'_{sRGB})^{2.4},
& E'_{sRGB} > 0.04045
\end{cases}\\
E_{sYCC} &= (\textrm{linear}) \{R_{sYCC},G_{sYCC},B_{sYCC}\} \\
E'_{sRGB} &= (\textrm{non-linear}) \{R'_{sRGB},G'_{sRGB},B'_{sRGB}\}
\end{align*} This is equivalent to applying $E_{sYCC} = 0.0125 + {1\over 1-0.0125} \times E_{sRGB}$ to linear $\{R,G,B\}$ values. The resulting linear $E_{sYCC}$ values then need to be non-linearly encoded with the EOTF. |
Similar to the ITU transfer function, the EOTF -1 of the sRGB function can be written as:
Like the ITU transfer function above, the values of $\alpha$ and $\beta$ in the sRGB function appear to have been chosen such that the linear segment and power segment meet at the same value and with the same derivative (that is, the linear segment meets the power segment at a tangent). The $\alpha$ and $\beta$ values can be derived as follows:
At $\{R',G',B'\} = \beta$ , the linear and non-linear segments of the function must calculate the same value:
Additionally, the derivatives of the linear and non-linear segments of the function must match:
This formula can be rearranged to give $\alpha$ in terms of $\beta$ :
Substituting this into the formula for $\{R,G,B\}$ :
This equation simplifies to:
This can be further simplified to:
The value of $\beta$ can be found numerically (for example by Newton-Raphson iteration, with a derivative of $1.4-1.4\beta^{-{1\over{2.4}}}$ ), and results in values of:
Where $\delta$ is the value of the EOTF -1 at $\{R',G',B'\} = \beta$ .
These deduced values are appreciably different from those in the sRGB specification, which does not state the origin of its constants. The intersection point of the sRGB EOTF has less numerical stability (and more nearby local minima in curves being optimized) that the corresponding ITU function - it is sensitive to the start value used for numerical approximations. This may explain how different values were reached for the sRGB specification. However, the errors both in value and derivative at the point of selection between the linear and exponent segments are small in practice. |
The EOTF can be written with these derived values as:
Apple describes the Display P3 color space as using the sRGB transfer function. The profile viewer in Apple’s ColorSync utility reports that the EOTF is of the following form: $$f(x) = \begin{cases}
cx, & x < d \\
(ax+b)^\gamma, & x \geq d
\end{cases}$$ The reported figures for $\gamma=2.4,\ a=0.948,\ b=0.52$ and $c=0.077$ correspond to the equivalent values in the sRGB specification: \begin{align*}
{1\over{\alpha}} &\approx 0.948 = a\\
{{\alpha-1}\over\alpha} &\approx 0.52 = b\\
{1\over{12.92}} &\approx 0.077 = c
\end{align*} These values are correct to the reported precision both for the value $\alpha = 1.055$ in the sRGB specification and for the more precise $\alpha \approx 1.055010718947587$ derived above.
However, where the sRGB specification states that $\delta = 0.04045$ , the profile viewer reports a corresponding d = 0.039. The disparity can be explained if the profile values have been derived as described in this section: $$\delta \approx 0.039293370676848\approx 0.039 = d$$ Note that this value assumes a correspondingly corrected version of $\alpha$ rather than $a = 1.055$ .
The extra precision may be needed over the constants in the sRGB specification due to the use of additional bits of accuracy in the Display P3 representation, which may expose a discontinuity due to rounding with the original numbers, particularly in the gradient of the curve. However, this distinction is subtle: when calculated over a [0..1] range, the derived EOTF and EOTF -1 agree with the official sRGB formulae to greater than 16-bit precision. |
Without allowing for adjusting the $\alpha = 1.055$ constant in the sRGB formula, the power function cannot be made to intersect perfectly at a tangent to the linear segment with gradient of 12.92. However, the intersection point $\beta$ can be found by solving:
This equation can give us a slightly more precise pair of values for the original sRGB equation:
In practice this makes no measurable difference, but does suggest that the values of $\beta = 0.0031308$ in the sRGB specification may have been incorrectly rounded.
The BT.1886 standard for the “Reference electro-optical transfer function for flat panel displays used in HDTV studio production” is intended to represent a typical EOTF for CRTs and to document this to ensure consistency between other display technologies:
L = screen luminance in cd/m2 |
V = input signal normalized to [0..1] |
a = user gain (legacy “contrast”) |
b = black level lift (legacy “brightness”) |
$\gamma$ = 2.4 |
If LW is the screen luminance of maximum white and LB is the screen luminance of minimum black:
ITU BT.2087 proposes the use of a simple power function with a $\gamma = 2.4$ as an approximation to this EOTF for the purposes of color conversion, effectively assuming b = 0 and LB is pure black. The reference display described in BT.1886 has a maximum luminance level of 100cd/m2 (brighter than the equivalent sRGB reference display).
The following graph shows the relationship between the BT.1886 EOTF (shown in red) and the ITU OETF such as used for BT.709 (shown in blue). The result of applying the two functions in turn, resulting in the OOTF of a combined BT.709-BT.1886 system, is shown in black. Since the ITU OETF approximates a power function with $\gamma=2.0$ , also shown in green is the resulting OOTF corresponding to a power function with $\gamma={2.4\over{2.0}}=1.2$ .
BT.1886 also offers an alternative EOTF which may provide a better match to CRT measured luminance than the standard formula listed above: \begin{align*}
L &= \begin{cases}
k(V_C+b)^{(\alpha_1-\alpha_2)}(V+b)^{\alpha_2}, & V < V_C \\
k(V+b)^{\alpha_1}, & V_C \leq V
\end{cases}
\end{align*}
|
HLG (and PQ, below) are intended to allow a better encoding of high-dynamic-range content compared with the standard ITU OETF.
The BT.2100-2 Hybrid Log Gamma description defines the following OETF for linear scene light:
$E$ = the $R_S$ , $G_S$ or $B_S$ color component of linear scene light, normalized to [0..1] |
$E'$ = the resulting non-linear $R_S'$ , $G_S'$ or $B_S'$ non-linear scene light value in in the range [0..1] |
a = 0.17883277 |
b = $1 - 4\times a = 0.28466892$ |
c = $0.5 - a\times ln(4\times a) \approx 0.55991073$ |
BT.2100-0, in note 5b, defines these formulae equivalently, but slightly differently: $$E'_\mathit{norm} = \textrm{OETF}(E) = \begin{cases}
\sqrt{3E}, & 0 \leq E \leq {1\over{12}} \\
a \times \textrm{ln}(E - b_0) + c_0, & {1\over{12}} < E \leq 1
\end{cases}$$ This formulation in BT.2100-0 uses different constants for b and c (a is unmodified), as follows:
The BT.2100-0 variations can be derived from the BT.2100-2 numbers as: \begin{align*}
a \times \textrm{ln}((12 \times E) - b_1)+ c_1 & = a \times \textrm{ln}\left(12\times\left(E - {b_1\over{12}}\right)\right) + c_1\\
& = a \times \textrm{ln}\left(E - {{b_1}\over{12}}\right) + a\times\textrm{ln}(12) + c_1 \\
{{b_1}\over{12}} = {{0.28466892}\over{12}} &= 0.023772241 = b_0 \\
a\times\textrm{ln}(12) + c_1 = 0.17883277\times\textrm{ln}(12) + 0.55991073 &= 1.00429347 = c_0
\end{align*} |
The OETF -1 of normalized HLG is:
a, b and c are defined as for the normalized HLG OETF. BT.2100-0 again defines an equivalent formula without the $1\over{12}$ scale factor in the ${1\over 2}$ < $E' \leq 1$ term, using the modified b0 and c0 constants described in the note in the HLG OETF above.
BT.2100-0 describes the HLG OETF formulae with E “normalized” to the range [0..12], with the variant with the range normalized to [0..1] as an alternative. Only the variant normalized to the range [0..1] is described in the updated versions of the specification, BT.2100-1 and BT.2100-2.
$E'$ = the $R_S$ , $G_S$ or $B_S$ color component of linear scene light, normalized to [0..12] |
$E_S'$ = the resulting non-linear $R_S'$ , $G_S'$ or $B_S'$ value in in the range [0..1] |
a = 0.17883277 |
b = 0.28466892 |
c = 0.55991073 |
Note that these constants are the same as those used in the BT.2100-1 version of the normalized formulae.
The OETF -1 of “unnormalized” HLG (producing E in the range [0..12]) is:
a, b and c are defined as for the unnormalized HLG OETF.
BT.2100-0 describes this “unnormalized” version of the formulae, with the variant with the E normalized to [0..1] as an alternative. Only the variant with E normalized to [0..1] is described in the updated versions, BT.2100-1 and BT.2100-2.
HLG constants appear to have chosen a, b and c to meet the following constraints, which are easiest to express in terms of the unnormalized OETF -1:
The derivative of the $0 \leq E' \leq {1\over 2}$ term of the unnormalized OETF -1 has the same value as the derivative of the ${1\over 2}$ < $E' \leq 1$ term of the unnormalized OETF -1 at $E' = {1\over 2}$ :
The $0 \leq E' \leq {1\over 2}$ term of the unnormalized OETF -1 has the same value as the ${1\over 2}$ < $E'\leq 1$ term of the unnormalized OETF -1 at $E' = {1\over 2}$ :
At E' = 1, the ${1\over 2}$ < $E'$ term of the unnormalized OETF -1 = 12:
This last equation can be solved numerically to find:
With this precision, more accurate values of the other constants are:
The b = 0.28466892 official figure assumes the rounded a = 0.17883277 value as an input to the b = $1 - 4\times a$ relation.
No explanation for the choice of [0..12] range in the official version of the formula is explicitly offered in BT.2100-0 (it does not, for example, appear to relate to the BT.1886 OOTF $\gamma = 1.2$ combined with the $10\times$ ratio between the 1000cd/m2 of a standard HLG HDR TV and the 100cd/m2 of a standard dynamic range set). However, allowing for the difference in the maximum display brightness of HDR and SDR systems there is deliberate (scaled) compatibility between the HLG OETF and the BT.2020 OETF (which itself approximates a square root function) over much of the encodable dynamic range of a BT.2020 system. Since HDR content is intended to support accurate highlights more than to maintain a higher persistent screen brightness (many HDR displays can only support maximum brightness in a small area or over a small period without overheating), agreement over a significant chunk of the tone curve allows a simple adaptation between HDR and SDR devices: fed HLG-encoded content, an SDR display may represent darker tones accurately and simply under-represent highlights. The origins of both HLG and PQ are discussed in ITU-R BT.2390.
As graphed in ITU-R BT.2390, the “unnormalized” HLG OETF (red) is a good approximation to the standard dynamic range ITU transfer function (blue, output scaled by 0.5) up to $E \approx 1$ and $\textrm{OETF}(E) = E' \approx 0.5$ , with a smooth curve up to the maximum HLG representable scene light value of “12”: |
The OOTF of HLG is described as:
where RD, GD and BD describe the luminance of the displayed linear component in cd/m2 and RS, GS and BS describe each color component in scene linear light, scaled by camera exposure and normalized to the representable range.
BT.2100 notes that some legacy displays apply the $\gamma$ function to each channel separately, rather than to the luminance component. That is, $\{R_D,G_D,B_D\}=\alpha\times\{R_S,G_S,B_S\}^\gamma+\beta$ . This is an approximation to the official OOTF. |
YS is the normalized scene luminance, defined as:
$\alpha$ represents adjustable user gain (display “contrast”) representing LW, the nominal peak luminance of achromatic pixels.
Versions of BT.2100 prior to BT.2100-2 incorporated a $\beta$ black level offset (display “brightness”) representing the display luminance of black in cd/m2: \begin{align*}
R_D &= \alpha\times Y_S^{\gamma-1}\times R_S + \beta \\
G_D &= \alpha\times Y_S^{\gamma-1}\times G_S + \beta \\
B_D &= \alpha\times Y_S^{\gamma-1}\times B_S + \beta \\
\end{align*} $\alpha$ then represented the relative display “contrast”:
where LW is the nominal peak luminance of the display in cd/m2, and LB is the display luminance of black in cd/m2. That is, in older versions of BT.2100, $\alpha$ represented the difference between minimum and maximum brightness, whereas in BT.2100-2 $\alpha$ is independent of black level. |
$\gamma = 1.2$ for a nominal peak display luminance of 1000cd/m2. For displays with higher peak luminance or if peak luminance is reduced through a contrast control, $\gamma = 1.2 + 0.42\times \textrm{log}_{10}\left({L_W\over 1000}\right)$ .
For the purposes of general conversion, LW can be assumed to be 1000cd/m2, and LB can be approximated as 0, removing the constant offset from the above equations and meaning $\gamma=1.2$ .
The inverse OOTF for HLG can be defined as:
For processing without reference to a specific display, $\alpha$ can be assumed to be 1.0cd/m2.
Versions of BT.2100 prior to BT.2100-2 incorporated a $\beta$ term into the OOTF. Using this formula from the OOTF leads to the following relationship between YD and YS: \begin{align*}
Y_D &= 0.2627\times R_D + 0.6780\times G_D + 0.0593\times B_D \\
&= 0.2627\times(\alpha\times Y_S^{\gamma-1}\times R_S + \beta) +
0.6780\times(\alpha\times Y_S^{\gamma-1}\times G_S + \beta) +
0.0593\times(\alpha\times Y_S^{\gamma-1}\times B_S + \beta) \\
&= \alpha\times Y_S^{\gamma-1}\times(0.2627\times R_S
+ 0.6780\times G_S + 0.0593\times B_S)+\beta \\
&= \alpha\times Y_S^{\gamma}+\beta \\
\therefore Y_S =& \left({{Y_D-\beta}\over\alpha}\right)^{1\over\gamma}\\
Y_S^{1-\gamma} =& \left({{Y_D-\beta}\over\alpha}\right)^{(1-\gamma)/\gamma}
\end{align*} From this, the following relations can be derived: \begin{align*}
R_S &= {(R_D - \beta)\over{\alpha\times Y_S^{\gamma-1}}}
= Y_S^{1-\gamma}\times{{(R_D-\beta)}\over\alpha}
= \left({{Y_D - \beta}\over\alpha}\right)^{(1-\gamma)/\gamma} \times
\left({{R_D - \beta}\over{\alpha}}\right) \\
G_S &= {(G_D - \beta)\over{\alpha\times Y_S^{\gamma-1}}}
= Y_S^{1-\gamma}\times{{(G_D-\beta)}\over\alpha}
= \left({{Y_D - \beta}\over\alpha}\right)^{(1-\gamma)/\gamma} \times
\left({{G_D - \beta}\over{\alpha}}\right) \\
B_S &= {(B_D - \beta)\over{\alpha\times Y_S^{\gamma-1}}}
= Y_S^{1-\gamma}\times{{(B_D-\beta)}\over\alpha}
= \left({{Y_D - \beta}\over\alpha}\right)^{(1-\gamma)/\gamma} \times
\left({{B_D - \beta}\over{\alpha}}\right)
\end{align*} For processing without knowledge of the display, $\alpha$ can be treated as 1.0cd/m2 and $\beta$ can be considered to be 0.0cd/m2. This simplifies the equations as follows: \begin{align*}
Y_S &= Y_D^{1/\gamma} \\
Y_S^{1-\gamma} &= Y_D^{(1/\gamma)-1} \\
R_S&=Y_D^{(1/\gamma)-1}\times R_D \\
G_S&=Y_D^{(1/\gamma)-1}\times G_D \\
B_S&=Y_D^{(1/\gamma)-1}\times B_D
\end{align*} |
The EOTF of BT.2100 HLG is defined in terms of the OETF and OOTF defined above:
where LW is the nominal peak luminance of the display in cd/m2, and LB is the display luminance of black in cd/m2.
Versions of BT.2100 prior to BT.2100-2 incorporated the black level offset (display “brightness”) $\beta$ into the definition of the OOTF, such that: $$\{R_D,G_D,B_D\}=\textrm{OOTF}(\textrm{OETF}^{-1}(\{R_S',G_S',B_S'\}))$$ |
The EOTF -1 can be derived as:
Versions of BT.2100 prior to BT.2100-2 incorporated the black level offset (display “brightness”) $\beta$ into the definition of the OOTF, such that: $$\{R_S',G_S',B_S'\} = \textrm{OETF}(\textrm{OOTF}^{-1}(\{R_D,G_D,B_D\}))$$ |
Unlike BT.2100 HLG and other ITU broadcast standards, PQ is defined in terms of an EOTF (mapping from the encoded values to the display output), not an OETF (mapping from captured scene content to the encoded values). |
The BT.2100 Perceptual Quantization description defines the following EOTF:
$E'$ is a non-linear color channel $\{R',G',B'\}$ or $\{L',M',S'\}$ encoded as PQ in the range [0..1]. $F_D$ is the luminance of the displayed component in cd/m2 (where the luminance of an $\{R_D,G_D,B_D\}$ or $Y_D$ or $I_D$ component is considered to be the luminance of the color with all channels set to the same value as the component). When $R'=G'=B'$ the displayed pixel is monochromatic. $Y$ is a linear color value normalized to [0..1].
The corresponding EOTF -1 is:
The OOTF of PQ matches that of BT.1886's EOTF combined with BT.709's OETF:
where E is one of $\{R_S,G_S,B_S,Y_S,I_S\}$ , the linear representation of scene light scaled by camera exposure and in the range [0..1], G1886 is the EOTF described in BT.1886, and G709 is the OETF described in BT.709 with a scale factor of 59.5208 applied to E:
ITU-R BT.2390 explains the derivation of the scale factor:
PQ can encode 100 times the display brightness of a standard dynamic range (“SDR”) encoding (10000cd/m2 compared with the 100cd/m2 SDR reference display described in BT.1886). High dynamic range (HDR) displays are intended to represent the majority of scene content within a “standard” dynamic range, and exposure of a normalized SDR signal is chosen to provide suitable exposure. HDR displays offer extra capability for representation of small or transient highlights (few HDR displays can actually reach the maximum 10000cd/m2 encodable brightness, and few HDR displays can maintain their maximum intensity over a large area for an extended period without overheating). Therefore the behavior of HDR displays is intended to approximate a conventional standard dynamic range display for most of the image, while retaining the ability to encode extreme values.
As described in BT.2390, the OOTF of SDR is roughly $\gamma = 1.2$ (deviating from this curve more near a 0 value), so the maximum scene light intensity that can be represented is roughly $100^{1\over 1.2} \approx 46.42$ times that of a SDR encoding.
Using exact equations from BT.709 and BT.1886 to create the OOTF, rather than the $\gamma = 1.2$ approximation, the maximum representable scene brightness, if 1.0 is the maximum normalized SDR brightness is: \begin{align*}
\left({100^{1\over 2.4} + 0.099\over 1.099}\right)^{1\over 0.45} &\approx 59.5208
\end{align*} The other constants in the G709 formula are derived as follows: \begin{align*}
{0.018\over 59.5208} &\approx 0.0003024 \\
4.5\times 59.5208 &\approx 267.84
\end{align*} Note that these constants differ slightly if the more accurate $\alpha = 1.0993$ figure from BT.2020 is used instead of 1.099. |
The OETF of PQ is described in terms of the above OOTF:
The PQ OOTF -1 is:
where FD, display intensity, is one of $\{R_D,G_D,B_D,Y_D,I_D\}$ , and E is the corresponding normalized scene intensity.
DCI P3 defines a simple power function with an exponent of 2.6 (applied to scaled CIE XYZ values):
EOTF -1 | EOTF |
---|---|
$X' = \left({X\over{52.37}}\right)^{1\over{2.6}}$ $Y' = \left({Y\over{52.37}}\right)^{1\over{2.6}}$ $Z' = \left({Z\over{52.37}}\right)^{1\over{2.6}}$ | $X = X'^{2.6}\times 52.37$ $Y = Y'^{2.6}\times 52.37$ $Z = Z'^{2.6}\times 52.37$ |
This power function is applied directly to scaled CIE XYZ color coordinates: the “primaries” in DCI define the bounds of the gamut, but the actual color encoding uses XYZ coordinates. DCI scales the resulting non-linear values to the range [0..4095] prior to quantization, rounding to nearest.
“Display P3” uses the sRGB transfer function, modified in some implementations to have more accurate constants (see the section on the derivation of the sRGB constants). |
ITU-R BT.470-6, which has now been deprecated, lists a number of regional TV standard variants; an updated list of variant codes used by country is defined in ITU-R BT.2043. This standard, along with e-CFR title 47 section 73.682, documents a simple EOTF power function with $\gamma = 2.2$ for NTSC display devices.
EOTF -1 | EOTF |
---|---|
$R' = R^{1\over{2.2}}$ $G' = G^{1\over{2.2}}$ $B' = B^{1\over{2.2}}$ | $R = R'^{2.2}$ $G = G'^{2.2}$ $B = B'^{2.2}$ |
This value of $\gamma$ is also used for N/PAL signals in the Eastern Republic of Uruguay, and was also adopted by ST-240.
Combined with the reference in SMPTE 170M to a $\gamma = 2.2$ being used in “older documents”, this suggests a linear design OOTF for NTSC systems.
ITU-R BT.1700, which partly replaced BT.470, also describes an “assumed gamma of display device” of 2.2 for PAL and SECAM systems; this is distinct from the $\gamma = 2.8$ value listed in ITU-R BT.470-6. Combined with the ITU OETF which approximates $\gamma = {1\over{2.0}}$ , the PAL OOTF retains a $\gamma \approx 1.1$ when this value of $\gamma = 2.2$ is used for the EOTF, similar to the figure described under the legacy PAL EOTF.
In contrast, ITU-R BT.1700 also includes SMPTE 170m, which defines the assumed EOTF of the display device as being the inverse of the current ITU OETF. Hence the new NTSC formulation also assumes a linear OOTF.
ITU-R BT.472, “Video-frequency characteristics of a television system to be used for the international exchange of programmes between countries that have adopted 625-line colour or monochrome systems”, defines that the “gamma of the picture signal” should be “approximately 0.4”. The reciprocal of this value is 2.5.
That is, this standard defines an approximate OETF and OETF -1 for PAL content:
OETF | OETF -1 |
---|---|
$R' \approx R^{0.4}$ $G' \approx G^{0.4}$ $B' \approx B^{0.4}$ | $R \approx R'^{2.5}$ $G \approx G'^{2.5}$ $B \approx B'^{2.5}$ |
ITU-R BT.470-6, which has now been deprecated in favor of BT.1700, lists a number of regional TV standard variants; an updated list of variant codes used by country is defined in ITU-R BT.2043.
This specification describes a simple EOTF power function with $\gamma_{\textrm{EOTF}} = 2.8$ for most PAL and SECAM display devices:
EOTF -1 | EOTF |
---|---|
$R' = R^{1\over{2.8}}$ $G' = G^{1\over{2.8}}$ $B' = B^{1\over{2.8}}$ | $R = R'^{2.8}$ $G = G'^{2.8}$ $B = B'^{2.8}$ |
Poynton describes a $\gamma$ of 2.8 as being “unrealistically high” for actual CRT devices. |
Combined with the corresponding legacy EOTF with $\gamma_{\textrm{EOTF}} = 0.4$ , the described system OOTF is:
Or $\gamma_{\textrm{OOTF}} \approx 1.12$ .
The value of $\gamma_{\textrm{EOTF}} = 2.8$ is described in BT.470-6 as being chosen for “an overall system gamma” (OOTF power function exponent) of “approximately 1.2”; this suggests that the “approximately 0.4” exponent in BT.472-6 should be interpreted as nearer to ${1.2\over{2.8}} \approx 0.43$ , or at least that there was enough variation in early devices for precise formulae to be considered irrelevant.
The EOTF power function of $\gamma_{\textrm{EOTF}} = 2.2$ described in BT.1700 combines with the ITU OETF described in BT.601 (which approximates $\gamma_{\textrm{OETF}} \approx 0.5$ ) to give a similar system $\gamma_{\textrm{OOTF}} \approx 1.1$ . As described above, the ITU OETF combined with the BT.1886 EOTF results in a more strongly non-linear $\gamma_{\textrm{OOTF}} \approx {2.4\over{2.0}} = 1.2$ . |
The ST-240, formerly SMPTE240M, interim standard for HDTV defines the following OETF:
Like SMPTE170m, ST-240 defines a linear OOTF. Therefore the above relationship also holds for the EOTF -1.
The EOTF, and also OETF -1, is:
The Adobe RGB (1998) specification defines the following transfer function between nonlinear encoding and linear light intensity (notable for not including a linear component):
2.19921875 is obtained from $2{51\over{256}}$ or hexadecimal 2.33. Therefore the inverse transfer function between linear light intensity and nonlinear encoding is:
The Sony S-Log OETF is defined for each color channel as:
Linear camera input scaled by exposure t ranges from 0 to 10.0; y is the non-linear encoded value.
The OETF -1 is:
The encoded non-linear value t ranges from 0 to 1.09; Y is the linear scene light.
S-Log2 defines the following OETF:
x is the IRE in scene-linear space. y is the IRE in S-Log2 space.
The OETF -1 is:
x is the IRE in S-Log2 space. y is the IRE in scene-linear space.
A reflection is calculated by multiplying an IRE by 0.9.
ACES is scene-referred; therefore ACEScc defines an OETF.
For each linear color channel linAP1 transformed to the ACEScc primaries, the ACEScc non-linear encoding is:
ACES is scene-referred; therefore ACEScct defines an OETF.
For each linear color channel linAP1 transformed to the ACEScc primaries, the ACEScct non-linear encoding is:
Color primaries define the interpretation of each channel of the color model, particularly with respect to the RGB model. In the context of a typical display, color primaries describe the color of the red, green and blue phosphors or filters.
Primaries are typically defined using the CIE 1931 XYZ color space, which is a color space which preserves the linearity of light intensity. Consequently, the transform from linear-intensity (R, G, B) to (X, Y, Z) is a simple matrix multiplication. Conversion between two sets of (R, G, B) color primaries can be performed by converting to the (X, Y, Z) space and back.
The (X, Y, Z) space describes absolute intensity. Since most standards do not make a requirement about the absolute intensity of the display, color primaries are typically defined using the x and y components of the xyY color space, in which the Y channel represents linear luminance. xyY is related to XYZ via the following formulae:
This is relevant because, although the brightness of the display in a color space definition is typically undefined, the white point is known: the x and y coordinates in xyY color space which corresponds to equal amounts of R, G and B. This makes it possible to determine the relative intensities of these color primaries.
Many color standards use the CIE D65 standard illuminant as a white point. D65 is intended to represent average daylight, and has a color temperature of approximately 6500K. In CIE 1931 terms, this white point is defined in ITU standards as $x=0.3127,\ y=0.3290$ , but elsewhere given as $x=0.312713,\ y=0.329016$ . Different coordinates will affect the conversion matrices given below. The definition of the D65 white point is complicated by the constants in Planck’s Law (which is a component in calculating the white point from the color temperature) having been revised since D65 was standardized, such that the standard formula for calculating CIE coordinates from the color temperature do not agree with the D65 standard. The actual color temperature of D65 is nearer to $6500\times {1.4388\over 1.438} \approx 6503.6\textrm{K}$ . |
Assuming an arbitrary white luminance (Y value) of 1.0, it is possible to express the following identity for the X, Y and Z coordinates of each color channel R, G and B, and of the white point W:
The identities $X = Y{x\over{y}}$ and $Z = Y{{(1-x-y)}\over{y}}$ can be used to re-express the above terms in the xyY space:
This equation for WZ can be simplified to:
Since $R_Y + G_Y + B_Y = W_Y = 1$ , this further simplifies to:
The $R_Y+G_Y+B_Y$ term for WY can be multiplied by $R_x\over{R_y}$ and subtracted from the equation for WX:
Similarly, the $R_Y+G_Y+B_Y$ term can be multiplied by $1-R_x\over{R_y}$ and subtracted from the simplified WZ line:
Finally, the GY term can be eliminated by multiplying the former of these two equations by ${1-G_x\over{G_y}}-{1-R_x\over{R_y}}$ and subtracting it from the latter multiplied by ${G_x\over{G_y}}-{R_x\over{R_y}}$ , giving:
Thus:
This allows GY to be calculated by rearranging an earlier equation:
And finally:
These relative magnitudes allow the definition of vectors representing the color primaries in the XYZ space, which in turn provides a transformation between colors specified in terms of the color primaries and the XYZ space. Without an absolute magnitude the transformation to XYZ is incomplete, but sufficient to allow transformation to another set of color primaries.
The transform from the defined color primaries to XYZ space is:
The transform from XYZ space to the defined color primaries is therefore:
These transforms assume that the black point for the color space is at $(X, Y, Z) = (0, 0, 0)$ . If the black point is non-zero, these transforms require a translational component. In some color spaces the black point has the same color as the white point, in which case it is also possible to adjust the $(R, G, B)$ values outside the matrix. |
Real-world objects and light sources typically emit a full spectrum of wavelengths in varying intensities. Describing the color of an object in terms of three wavelengths is necessarily an approximation, and effective only because the human eye typically sees the world mostly in terms of combinations of three color cone cell responses — “tristimulus values”. The typical response of these “long”, “medium” and “short” wavelength-sensitive cones is the basis of the CIE XYZ color space, so a linear combination of primaries described in this color space can be mapped to a linear combination of other primaries. Two colors with a different spectrum but an identical appearance are known as metamers, and the tristimulus model of color relies on “metameric matching”: picking a linear combination of some set of primaries that will elicit the same cone cell response as the desired color.
Each of the cone cells can appear in mutated form which results in a change to its response to different frequencies of light. Since the genes for cone cells are encoded on the X chromosome, some women have more than one variant of the same cone cell, being “tetrachromats” — and there is evidence that such people can distinguish more colors than the general “trichromat” population, with only a subset of these colors describable with three color primaries. There is some evidence for people with two variants of more than one cone cell, thereby perceiving color in a five-dimensional space, although the processing performed by the human visual system is complex and it is not clear how independently these color dimensions are in practice. In the animal kingdom, it is not uncommon to have more than three types of cone cell, and mantis shrimp have up to sixteen kinds of photoreceptors. Note that rod cells, which dominate the human visual system in low light conditions, are typically ignored at higher illumination levels for the purposes of color matching.
The difference between a primary color representation and the full spectrum emitted by a real-world object is particularly visible for printed media. Printing colors are typically described assuming that they will reflect the spectrum of a theoretical illuminant; changing this light source may change the appearance of the printed object in a more complex manner than scaling the primary values. This can lead to “metameric failure”, where two colors may appear identical under some lighting conditions and different under others. This effect is present in nature, notably in the mineral alexandrite, which may appear reddish purple under incandescent light but green in daylight. Camera sensors and film have a related problem that the spectrum of filters does not perfectly match that of cone cells, which can lead some colors (particularly with sharp peaks in their spectra) to appear different. Inkjet printing often uses many more than three inks in order to maximize the representable gamut and control metamerism.
Adaptation to a different white point is often performed by a linear scaling of primaries, known as the “von Kries transform”. The Bradford color adaptation transform incorporates a slight nonlinear term to the blue component of colors to more accurately reflect visual behavior.
It is common for production rendering systems to represent each pixel with more than three channels. For example, the contributions from different lights may be recorded separately in order to allow later adjustment to color and relative intensity, specular and diffuse reflectance from a surface may be recorded separately (which is valuable, for example, in denoising algorithms), and virtual “channels” may be used to control shader effects. A larger number of wavelengths may also be used when rendering effects such as chromatic dispersion, which is important for rendering jewelry. Currently these are typically represented as multiple images or in a proprietary representation, although one could imagine a data format descriptor for such a pixel. |
ITU-T BT.709 (HDTV) defines the following chromaticity coordinates:
These chromaticity coordinates are also shared by sRGB and scRGB.
Therefore to convert from linear color values defined in terms of BT.709 color primaries to XYZ space the formulae in Section 14 result in the following matrix:
The inverse transformation, from the XYZ space to a color defined in terms of BT.709 color primaries, is:
sYCC lists a slightly different version of this matrix, possibly due to rounding errors. |
ITU-T Rec.601 defines different color primaries for 625-line systems (as used in most PAL systems) and for 525-line systems (as used in the SMPTE 170M-2004 standard for NTSC).
The following chromaticity coordinates are defined for 625-line “EBU” systems:
BT.470-6, which also describes these constants in a legacy context, approximates D65 as $x = 0.313,\ y = 0.329$ . |
Therefore to convert from linear color values defined in terms of BT.601 color primaries for 625-line systems to XYZ space the formulae in Section 14 result in the following matrix:
The inverse transformation, from the XYZ space to a color defined in terms of BT.601 “EBU” 625-line color primaries, is:
ITU-T Rec.601 defines different color primaries for 625-line systems (as used in most PAL systems) and for 525-line systems (as used in the SMPTE 170M-2004 standard for NTSC).
The following chromaticity coordinates are defined in BT.601 for 525-line digital systems and in SMPTE-170M:
Therefore to convert from linear color values defined in terms of BT.601 color primaries for 525-line systems to XYZ space the formulae in Section 14 result in the following matrix:
The inverse transformation, from the XYZ space to a color defined in terms of BT.601 525-line color primaries, is:
Analog 525-line PAL systems used a different white point, and therefore have a different conversion matrix. |
The following chromaticity coordinates are defined in BT.2020 for ultra-high-definition television:
The same primaries are used for BT.2100 for HDR TV.
Therefore to convert from linear color values defined in terms of BT.2020 color primaries to XYZ space the formulae in Section 14 result in the following matrix:
The inverse transformation, from the XYZ space to a color defined in terms of BT.2020 color primaries, is:
The following chromaticity coordinates are defined in ITU-R BT.470-6 and SMPTE 170m as a reference to the legacy NTSC standard:
These primaries apply to the 1953 revision of the NTSC standard. Modern NTSC systems, which reflect displays that are optimized for brightness over saturation, use the color primaries as described in Section 14.3. The white point used in the original NTSC 1953 specification is CIE Standard Illuminant C, 6774K, as distinct from the CIE Illuminant D65 used by most modern standards. BT.470-6 notes that SECAM systems may use these NTSC primaries and white point. Japanese legacy NTSC systems used the same primaries but with the white point set to D-white at 9300K. |
Therefore to convert from linear color values defined in terms of NTSC 1953 color primaries to XYZ space the formulae in Section 14 result in the following matrix:
The inverse transformation, from the XYZ space to a color defined in terms of NTSC 1953 color primaries, is:
ITU-R BT.1700 defines the following chromaticity coordinates for legacy 525-line PAL systems:
This matches the color primaries from SMPTE-170m analog NTSC and BT.601 525-line encoding, but the white point used is CIE Standard Illuminant C, 6774K, as distinct from the CIE Illuminant D65 white point used by most modern standards. |
Therefore to convert from linear color values defined in terms of PAL 525-line color primaries to XYZ space the formulae in Section 14 result in the following matrix:
The inverse transformation, from the XYZ space to a color defined in terms of PAL 525-line 1953 color primaries, is:
The following chromaticity coordinates are defined in SMPTE ST 2065-1
Therefore to convert from linear color values defined in terms of ACES color primaries to XYZ space the formulae in Section 14 result in the following matrix:
The inverse transformation, from the XYZ space to a color defined in terms of ACES color primaries, is:
The following chromaticity coordinates are defined in Academy S-2016-001 (ACEScct) and S-2014-003 (ACEScc), which share the same primaries:
Therefore to convert from linear color values defined in terms of ACEScc/ACEScct color primaries to XYZ space the formulae in Section 14 result in the following matrix:
The inverse transformation, from the XYZ space to a color defined in terms of ACEScc/ACEScct color primaries, is:
The following chromaticity coordinates are defined in Display P3:
The DCI P3 color space defines the bounds of its gamut using these primaries, but actual color data in DCI P3 is encoded using CIE XYZ coordinates. Display P3, on the other hand, uses these values as primaries in an RGB color space, with a D65 white point. |
Therefore to convert from linear color values defined in terms of Display P3 color primaries to XYZ space the formulae in Section 14 result in the following matrix:
The inverse transformation, from the XYZ space to a color defined in terms of DisplayP3 color primaries, is:
$$\left(\begin{array}{c} R_{\textrm{DisplayP3}} \\ G_{\textrm{DisplayP3}} \\ B_{\textrm{DisplayP3}}\end{array}\right) \approx \left(\begin{array}{ccc} 2.4934969119, & -0.9313836179, & -0.4027107845 \\ -0.8294889696, & 1.7626640603, & 0.0236246858 \\ 0.0358458302, & -0.0761723893, & 0.9568845240\end{array}\right) \left(\begin{array}{c}X \\ Y \\ Z\end{array}\right) $$These matrices differ from those given in SMPTE EG 432-1 due to the choice of a D65 white point in Display P3. The matrices in 432-1 can be reproduced by applying a white point of $W_x = 0.314,\ W_y = 0.351$ to the above primaries. |
The following chromaticity coordinates are defined in Adobe RGB (1998):
Therefore to convert from linear color values defined in terms of Adobe RGB (1998) color primaries to XYZ space the formulae in Section 14 result in the following matrix:
The inverse transformation, from the XYZ space to a color defined in terms of Adobe RGB (1998) color primaries, is:
Adobe RGB (1998) defines a reference display white brightness of 160cd/m2 and a black point 0.34731% of this brightness, or 0.5557cd/m2, for a contrast ratio of 287.9. The black point has the same color temperature as the white point, and this does not affect the above matrices.
Conversion from BT.709 to BT.601 625-line primaries can be performed using the matrices in Section 14.1 and Section 14.2 as follows:
Conversion from BT.601 625-line to BT.709 primaries can be performed using these matrices:
Conversion from BT.709 to BT.2020 primaries can be performed using the matrices in Section 14.4 and Section 14.1 as follows:
Conversion from BT.2020 primaries to BT.709 primaries can be performed with the following matrices:
The human eye is more sensitive to high-frequency changes in intensity (absolute quantity of light) than to high-frequency changes in the dominant wavelength and saturation of a color. Additionally the eye does not exhibit equal sensitivity to all wavelengths. Many image representations take advantage of these facts to distribute the number of bits used to represent a texel in a more perceptually-uniform manner than is achieved by representing the color primaries independently - for example by encoding the chroma information at a reduced spatial resolution.
Color models based on color differences are often referred to with incorrect or ambiguous terminology, the most common of which is YUV.
In the broadcast standards which define these models:
Color differences are calculated from the non-linear Y′ and color components as:
Note that, for R′, G′, B′ in the range [0..1]:
Using this terminology, the following conversion formulae can be derived:
For the inverse conversion:
The formula for G′ can be derived by substituting the formulae for R′ and B′ into the derivation of Y′:
The values chosen for KR and KB vary between standards.
The required color model conversion between $Y'C_BC_R$ and $R'G'B'$ can typically be deduced from other color space parameters: |
ITU Rec.709 defines KR = 0.2126 and KB = 0.0722.
That is, for conversion between (R′, G′, B′) defined in BT.709 color primaries and using the ITU transfer function:
Alternatively:
For the inverse conversion:
ITU Rec.601 defines KR = 0.299 and KB = 0.114.
That is, for conversion between (R′, G′, B′) defined in BT.601 EBU color primaries or BT.601 SMPTE color primaries, and using the ITU transfer function:
Alternatively:
For the inverse conversion:
ITU Rec.2020 and ITU Rec.2100 define KR = 0.2627 and KB = 0.0593.
That is, for conversion between (R′, G′, B′) defined in BT.2020 color primaries and using the ITU transfer function:
Alternatively:
For the inverse conversion:
ST240, formerly SMPTE 240M, defines KR = 0.212 and KB = 0.087.
That is, for conversion using the ST240 transfer function:
Alternatively:
For the inverse conversion:
ITU-T Rec. BT.2020 introduced a “constant luminance” color representation as an alternative representation to Y′CBCR:
This terminology follow’s BT.2020’s convention of describing the continuous values as $Y'_C$ , $C'_\mathit{BC}$ and $C'_\mathit{RC}$ ; BT.2020 uses $\mathit{DY}'_C$ , $\mathit{DC}'_\mathit{BC}$ and $\mathit{DC}'_\mathit{RC}$ to represent the quantized integer representations of the same values.
$Y'_C$ is derived from applying a non-linear transfer function to a combination of linear $\mathit{RGB}$ components and applying a non-linear transfer function to the result, but the $C'_\mathit{BC}$ and $C'_\mathit{RC}$ color differences still encode differences between non-linear values. |
The inverse transformation can be derived from the above:
Performing these calculations requires conversion between a linear representation and a non-linear transfer function during the transformation. This is distinct from the non-constant-luminance case, which is a simple matrix transform. |
ITU-T Rec. BT.2100 introduced a “constant intensity” color representation as an alternative representation to Y′CBCR:
Note that the suffix D indicates that PQ encoding is display-referred and the suffix S indicates that HLG encoding is scene-referred — that is, they refer to display and scene light respectively.
To invert this, it can be observed that:
The formulae in the previous sections are described in terms of operations on continuous values. These values are typically represented by quantized integers. There are standard encodings for representing some color models within a given bit depth range.
ITU broadcast standards typically reserve values at the ends of the representable integer range for rounding errors and for signal control data. The nominal range of representable values between these limits is represented by the following encodings, for bit depth n = {8, 10, 12}:
The dequantization formulae are therefore:
For consistency with $Y'_CC'_\mathit{BC}C'_\mathit{RC}$ , these formulae use the BT.2020 and BT.2100 terminology of prefixing a D to represent the digital quantized encoding of a numerical value.
That is, in “narrow range” encoding:
Value | Continuous encoding value | Quantized encoding |
---|---|---|
Black | {R′, G′, B′, Y′, $Y'_C$ , I} = 0.0 | {DR′, DG′, DB′, DY′, $\mathit{DY}'_C$ , DI} = $16 \times 2^{n-8}$ |
Peak brightness | {R′, G′, B′, Y′, $Y'_C$ , I} = 1.0 | {DR′, DG′, DB′, DY′, $\mathit{DY}'_C$ , DI} = $235 \times 2^{n-8}$ |
Minimum color difference value | { $C'_B$ , $C'_R$ , $C'_\mathit{BC}$ , $C'_\mathit{RC}$ , CT, CP} = -0.5 | { $\mathit{DC}'_B$ , $\mathit{DC}'_R$ , $\mathit{DC}'_\mathit{BC}$ , $\mathit{DC}'_\mathit{CR}$ , DCT, DCP} = $16 \times 2^{n-8}$ |
Maximum color difference value | { $C'_B$ , $C'_R$ , $C'_\mathit{BC}$ , $C'_\mathit{RC}$ , CT, CP} = 0.5 | { $\mathit{DC}'_B$ , $\mathit{DC}'_R$ , $\mathit{DC}'_\mathit{BC}$ , $\mathit{DC}'_\mathit{CR}$ , DCT, DCP} = $240 \times 2^{n-8}$ |
Achromatic colors | R′ = G′ = B′ { $C'_B$ , $C'_R$ , $C'_\mathit{BC}$ , $C'_\mathit{RC}$ , CT, CP} = 0.0 | { $\mathit{DC}'_B$ , $\mathit{DC}'_R$ , $\mathit{DC}'_\mathit{BC}$ , $\mathit{DC}'_\mathit{CR}$ , DCT, DCP} = $128 \times 2^{n-8}$ |
If, instead of the quantized values, the input is interpreted as fixed-point values in the range 0.0..1.0, as might be the case if the values were treated as unsigned normalized quantities in a computer graphics API, the following conversions can be applied instead:
ITU-T Rec. BT.2100-1 and the current Rec. T.871 JFIF specification define the following quantization scheme that does not incorporate any reserved head-room or foot-room, which is optional and described as “full range” in BT.2100, and integral to Rec. T.871.
Both these specifications modify a definition used in previous versions of their specifications, which is described below. |
For bit depth n = {8 (JFIF),10,12 (Rec.2100)}:
BT.2100-1 defines Round() as:
Note that a chroma channel value of exactly 0.5 corresponds to a quantized encoding of $2^n$ , and must therefore be clamped to the nominal peak value of $2^n-1$ . Narrow-range encoding does not have this problem. A chroma channel value of -0.5 corresponds to a quantized encoding of 1, which is the nominal minimum peak value.
In Rec. T.871 (which defines only n = 8), the corresponding formula is:
Allowing for the clamping at a chroma value of 0.5, these formulae are equivalent across the expected -0.5..0.5 range for chroma and 0.0..1.0 range for luma values.
The dequantization formulae are therefore:
That is, in “full range” encoding:
Value | Continuous encoding value | Quantized encoding |
---|---|---|
Black | {R′, G′, B′, Y′, $Y'_C$ , I} = 0.0 | {DR′, DG′, DB′, DY′, $\mathit{DY}'_C$ , DI} = 0 |
Peak brightness | {R′, G′, B′, Y′, $Y'_C$ , I} = 1.0 | {DR′, DG′, DB′, DY′, $\mathit{DY}'_C$ , DI} = 2n - 1 |
Minimum color difference value | { $C'_B$ , $C'_R$ , $C'_\mathit{BC}$ , $C'_\mathit{RC}$ , CT, CP} = -0.5 | { $\mathit{DC}'_B$ , $\mathit{DC}'_R$ , $\mathit{DC}'_\mathit{BC}$ , $\mathit{DC}'_\mathit{CR}$ , DCT, DCP} = 1 |
Maximum color difference value | { $C'_B$ , $C'_R$ , $C'_\mathit{BC}$ , $C'_\mathit{RC}$ , CT, CP} = 0.5 | { $\mathit{DC}'_B$ , $\mathit{DC}'_R$ , $\mathit{DC}'_\mathit{BC}$ , $\mathit{DC}'_\mathit{CR}$ , DCT, DCP} = $2^n - 1$ (clamped) |
Achromatic colors | R′ = G′ = B′ { $C'_B$ , $C'_R$ , $C'_\mathit{BC}$ , $C'_\mathit{RC}$ , CT, CP} = 0.0 | { $\mathit{DC}'_B$ , $\mathit{DC}'_R$ , $\mathit{DC}'_\mathit{BC}$ , $\mathit{DC}'_\mathit{CR}$ , DCT, DCP} = 2n-1 |
If, instead of the quantized values, the input is interpreted as fixed-point values in the range 0.0..1.0, as might be the case if the values were treated as unsigned normalized quantities in a computer graphics API, the following conversions can be applied instead:
ITU-T Rec. BT.2100-0 formalized an optional encoding scheme that does not incorporate any reserved head-room or foot-room. The legacy JFIF specification similarly used the full range of 8-bit channels to represent $Y'C_BC_R$ color. For bit depth n = {8 (JFIF),10,12 (Rec.2100)}:
The dequantization formulae are therefore:
These formulae map luma values of 1.0 and chroma values of 0.5 to $2^n$ , for bit depth $n$ . This has the effect that the maximum value (e.g. pure white) cannot be represented directly. Out-of-bounds values must be clamped to the largest representable value. |
ITU-R BT.2100-0 dictates that in 12-bit coding, the largest values encoded should be 4092 (“for consistency” with 10-bit encoding, with a maximum value of 1023). This slightly reduces the maximum intensity which can be expressed, and slightly reduces the saturation range. The achromatic color point is still 2048 in the 12-bit case, so no offset is applied in the transformation to compensate for this range reduction. BT.2100-1 removes this recommendation and lists 4095 as the nominal peak value. |
If, instead of the quantized values, the input is interpreted as fixed-point values in the range 0.0..1.0, as might be the case if the values were treated as unsigned normalized quantities in a computer graphics API, the following conversions can be applied instead:
That is, to match the behavior described in these specifications, the inputs to color model conversion should be expanded such that the maximum representable value is that defined by the quantization of these encodings $\left({255\over 256},\ {1023\over 1024}\ \textrm{or}\ {4095\over 4096}\right)$ , and the inverse operation should be applied to the result of the model conversion.
For example, a legacy shader-based JPEG decoder may read values in a normalized 0..1 range, where the in-memory value 0 represents 0.0 and the in-memory value 1 represents 1.0. The decoder should scale the Y′ value by a factor of $255\over 256$ to match the encoding in the JFIF3 document, and $C'_B$ and CR should be scaled by $255\over 256$ and offset by 0.5. After the model conversion matrix has been applied, the R′, G′ and B′ values should be scaled by $256\over 255$ , restoring the ability to represent pure white.
For computer graphics, a number of texture compression schemes exist which reduce both the overall texture memory footprint and the bandwidth requirements of using the textures. In this context, “texture compression” is distinct from “image compression” in that texture compression schemes are designed to allow efficient random access as part of texture sampling: “image compression” can further reduce image redundancy by considering the image as a whole, but doing so is impractical for efficient texture access operations. The common texture compression schemes are “block-based”, relying on similarities between nearby texel regions to describe “blocks” of nearby texels in a unit:
Example data format descriptors for compressed formats can be found under the colorModel field in Section 5.6. |
As can be seen above, the compression schemes have a number of features in common — particularly in having a number of endpoints described encoded in some of the bits of the texel block. For consistency and to make the terms more concise, the following descriptions use some slightly unusual terminology: The value Xnm refers to bit m (starting at 0) of the nth X value. For example, R13 would refer to bit 3 of red value 1 — R, G, B and A (capitalized and italicized) are generally used to refer to color channels. Similarly, R12..3 refers to bits 2..3 of red value 1. Although unusual, this terminology should be unambiguous (e.g. none of the formats require exponentiation of arguments).
This description is derived from the EXT_texture_compression_s3tc extension.
Compressed texture images stored using the S3TC compressed image formats are represented as a collection of 4×4 texel blocks, where each block contains 64 or 128 bits of texel data. The image is encoded as a normal 2D raster image in which each 4×4 block is treated as a single pixel. If an S3TC image has a width or height that is not a multiple of four, the data corresponding to texels outside the image are irrelevant and undefined.
When an S3TC image with a width of w, height of h, and block size of blocksize (8 or 16 bytes) is decoded, the corresponding image size (in bytes) is:
When decoding an S3TC image, the block containing the texel at offset (x, y) begins at an offset (in bytes) relative to the base of the image of:
The data corresponding to a specific texel (x, y) are extracted from a 4×4 texel block using a relative (x, y) value of
There are four distinct S3TC image formats:
Each 4×4 block of texels consists of 64 bits of RGB image data.
Each RGB image data block is encoded as a sequence of 8 bytes, called (in order of increasing address):
The 8 bytes of the block are decoded into three quantities:
color0 and color1 are 16-bit unsigned integers that are unpacked to RGB colors RGB0 and RGB1 as though they were 16-bit unsigned packed pixels with the R channel in the high 5 bits, G in the next 6 bits and B in the low 5 bits:
bits is a 32-bit unsigned integer, from which a two-bit control code is extracted for a texel at location (x, y) in the block using:
where bits[31] is the most significant and bits[0] is the least significant bit.
The RGB color for a texel at location (x, y) in the block is given in Table 105.
Table 105. Block decoding for BC1
Texel value | Condition |
---|---|
RGB0 | color0 > color1 and code(x, y) = 0 |
RGB1 | color0 > color1 and code(x, y) = 1 |
$(2\times \mathit{RGB}_0 + \mathit{RGB}_1)\over 3$ | color0 > color1 and code(x, y) = 2 |
$(\mathit{RGB}_0 + 2\times RGB_1)\over 3$ | color0 > color1 and code(x, y) = 3 |
RGB0 | color0 ≤ color1 and code(x, y) = 0 |
RGB1 | color0 ≤ color1 and code(x, y) = 1 |
$(\mathit{RGB}_0+\mathit{RGB}_1)\over 2$ | color0 ≤ color1 and code(x, y) = 2 |
BLACK | color0 ≤ color1 and code(x, y) = 3 |
Arithmetic operations are done per component, and BLACK refers to an RGB color where red, green, and blue are all zero.
Since this image has an RGB format, there is no alpha component and the image is considered fully opaque.
Each 4×4 block of texels consists of 64 bits of RGB image data and minimal alpha information. The RGB components of a texel are extracted in the same way as BC1 with no alpha.
The alpha component for a texel at location (x, y) in the block is given by Table 106.
The red, green, and blue components of any texels with a final alpha of 0 should be encoded as zero (black).
Figure 36 shows an example BC1 texel block: color0, encoded as $\left({{29}\over{31}}, {{60}\over{63}}, {{1}\over{31}}\right)$ , and color1, encoded as $\left({{20}\over{31}}, {{2}\over{63}}, {{30}\over{31}}\right)$ , are shown as circles. The interpolated values are shown as small diamonds. Since 29 > 20, there are two interpolated values, accessed when code(x, y) = 2 and code(x, y) = 3. Figure 37 shows the example BC1 texel block with the colors swapped: color0, encoded as $\left({{20}\over{31}}, {{2}\over{63}}, {{30}\over{31}}\right)$ , and color1, encoded as $\left({{29}\over{31}}, {{60}\over{63}}, {{1}\over{31}}\right)$ , are shown as circles. The interpolated value is shown as a small diamonds. Since 20 ≤ 29, there is one interpolated value for code(x, y) = 2, and code(x, y) = 3 represents (R, G, B) = (0, 0, 0). If the format is BC1 with alpha, code(x, y) = 3 is transparent (alpha = 0). If the format is BC1 with no alpha, code(x, y) = 3 represents opaque black. |
Each 4×4 block of texels consists of 64 bits of uncompressed alpha image data followed by 64 bits of RGB image data.
Each RGB image data block is encoded according to the BC1 formats, with the exception that the two code bits always use the non-transparent encodings. In other words, they are treated as though color0 > color1, regardless of the actual values of color0 and color1.
Each alpha image data block is encoded as a sequence of 8 bytes, called (in order of increasing address):
The 8 bytes of the block are decoded into one 64-bit integer:
alpha is a 64-bit unsigned integer, from which a four-bit alpha value is extracted for a texel at location (x, y) in the block using:
where bits[63] is the most significant and bits[0] is the least significant bit.
The alpha component for a texel at location (x, y) in the block is given by $\mathit{alpha}(x,y)\over 15$ .
Each 4×4 block of texels consists of 64 bits of compressed alpha image data followed by 64 bits of RGB image data.
Each RGB image data block is encoded according to the BC1 formats, with the exception that the two code bits always use the non-transparent encodings. In other words, they are treated as though color0 > color1, regardless of the actual values of color0 and color1.
Each alpha image data block is encoded as a sequence of 8 bytes, called (in order of increasing address):
The alpha0 and alpha1 are 8-bit unsigned bytes converted to alpha components by multiplying by $1\over 255$ .
The 6 bits bytes of the block are decoded into one 48-bit integer:
bits is a 48-bit unsigned integer, from which a three-bit control code is extracted for a texel at location (x, y) in the block using:
where bits[47] is the most-significant and bits[0] is the least-significant bit.
The alpha component for a texel at location (x, y) in the block is given by Table 107.
Table 107. Alpha encoding for BC3 blocks
Alpha value | Condition |
---|---|
alpha0 | code(x, y) = 0 |
alpha1 | code(x, y) = 1 |
$(6\times\mathit{alpha}_0 + 1\times\mathit{alpha}_1)\over 7$ | alpha0 > alpha1 and code(x, y) = 2 |
$(5\times\mathit{alpha}_0 + 2\times\mathit{alpha}_1)\over 7$ | alpha0 > alpha1 and code(x, y) = 3 |
$(4\times\mathit{alpha}_0 + 3\times\mathit{alpha}_1)\over 7$ | alpha0 > alpha1 and code(x, y) = 4 |
$(3\times\mathit{alpha}_0 + 4\times\mathit{alpha}_1)\over 7$ | alpha0 > alpha1 and code(x, y) = 5 |
$(2\times\mathit{alpha}_0 + 5\times\mathit{alpha}_1)\over 7$ | alpha0 > alpha1 and code(x, y) = 6 |
$(1\times\mathit{alpha}_0 + 6\times\mathit{alpha}_1)\over 7$ | alpha0 > alpha1 and code(x, y) = 7 |
$(4\times\mathit{alpha}_0 + 1\times\mathit{alpha}_1)\over 5$ | alpha0 ≤ alpha1 and code(x, y) = 2 |
$(3\times\mathit{alpha}_0 + 2\times\mathit{alpha}_1)\over 5$ | alpha0 ≤ alpha1 and code(x, y) = 3 |
$(2\times\mathit{alpha}_0 + 3\times\mathit{alpha}_1)\over 5$ | alpha0 ≤ alpha1 and code(x, y) = 4 |
$(1\times\mathit{alpha}_0 + 4\times\mathit{alpha}_1)\over 5$ | alpha0 ≤ alpha1 and code(x, y) = 5 |
0.0 | alpha0 ≤ alpha1 and code(x, y) = 6 |
1.0 | alpha0 ≤ alpha1 and code(x, y) = 7 |
This description is derived from the “RGTC Compressed Texture Image Formats” section of the OpenGL 4.4 specification.
Compressed texture images stored using the RGTC compressed image encodings are represented as a collection of 4×4 texel blocks, where each block contains 64 or 128 bits of texel data. The image is encoded as a normal 2D raster image in which each 4×4 block is treated as a single pixel. If an RGTC image has a width or height that is not a multiple of four, the data corresponding to texels outside the image are irrelevant and undefined.
When an RGTC image with a width of w, height of h, and block size of blocksize (8 or 16 bytes) is decoded, the corresponding image size (in bytes) is:
When decoding an RGTC image, the block containing the texel at offset $(x,y)$ begins at an offset (in bytes) relative to the base of the image of:
The data corresponding to a specific texel (x, y) are extracted from a 4×4 texel block using a relative (x, y) value of
There are four distinct RGTC image formats described in the following sections.
Each 4×4 block of texels consists of 64 bits of unsigned red image data.
Each red image data block is encoded as a sequence of 8 bytes, called (in order of increasing address):
The 6 bits{0..5} bytes of the block are decoded into a 48-bit bit vector:
red0 and red1 are 8-bit unsigned integers that are unpacked to red values RED0 and RED1 by multiplying by $1\over 255$ .
bits is a 48-bit unsigned integer, from which a three-bit control code is extracted for a texel at location (x, y) in the block using:
where bits[47] is the most-significant and bits[0] is the least-significant bit.
The red value R for a texel at location (x, y) in the block is given by Table 108.
Table 108. Block decoding for BC4
R value | Condition |
---|---|
RED0 | red0 > red1, code(x, y) = 0 |
RED1 | red0 > red1, code(x, y) = 1 |
${ 6 \times \mathit{RED}_0 + \mathit{RED}_1 } \over 7$ | red0 > red1, code(x, y) = 2 |
${ 5 \times \mathit{RED}_0 + 2 \times \mathit{RED}_1 } \over 7$ | red0 > red1, code(x, y) = 3 |
${ 4 \times \mathit{RED}_0 + 3 \times \mathit{RED}_1 } \over 7$ | red0 > red1, code(x, y) = 4 |
${ 3 \times \mathit{RED}_0 + 4 \times \mathit{RED}_1 } \over 7$ | red0 > red1, code(x, y) = 5 |
${ 2 \times \mathit{RED}_0 + 5 \times \mathit{RED}_1 } \over 7$ | red0 > red1, code(x, y) = 6 |
${ \mathit{RED}_0 + 6 \times \mathit{RED}_1 } \over 7$ | red0 > red1, code(x, y) = 7 |
RED0 | red0 ≤ red1, code(x, y) = 0 |
RED1 | red0 ≤ red1, code(x, y) = 1 |
${ 4 \times \mathit{RED}_0 + \mathit{RED}_1 } \over 5$ | red0 ≤ red1, code(x, y) = 2 |
${ 3 \times \mathit{RED}_0 + 2 \times \mathit{RED}_1 } \over 5$ | red0 ≤ red1, code(x, y) = 3 |
${ 2 \times \mathit{RED}_0 + 3 \times \mathit{RED}_1 } \over 5$ | red0 ≤ red1, code(x, y) = 4 |
${ \mathit{RED}_0 + 4 \times \mathit{RED}_1 } \over 5$ | red0 ≤ red1, code(x, y) = 5 |
REDmin | red0 ≤ red1, code(x, y) = 6 |
REDmax | red0 ≤ red1, code(x, y) = 7 |
REDmin and REDmax are 0.0 and 1.0 respectively.
Since the decoded texel has a red format, the resulting RGBA value for the texel is (R, 0, 0, 1).
Each 4×4 block of texels consists of 64 bits of signed red image data. The red values of a texel are extracted in the same way as BC4 unsigned except red0, red1, RED0, RED1, REDmin, and REDmax are signed values defined as follows:
red0 and red1 are 8-bit signed (two’s complement) integers.
CAVEAT: For signed red0 and red1 values: the expressions red0 > red1 and red0 ≤ red1 above are considered undefined (read: may vary by implementation) when red0 = -127 and red1 = -128. This is because if red0 were remapped to -127 prior to the comparison to reduce the latency of a hardware decompressor, the expressions would reverse their logic. Encoders for the signed red-green formats should avoid encoding blocks where red0 = -127 and red1 = -128.
Each 4×4 block of texels consists of 64 bits of compressed unsigned red image data followed by 64 bits of compressed unsigned green image data.
The first 64 bits of compressed red are decoded exactly like BC4 unsigned above. The second 64 bits of compressed green are decoded exactly like BC4 unsigned above except the decoded value R for this second block is considered the resulting green value G.
Since the decoded texel has a red-green format, the resulting RGBA value for the texel is (R, G, 0, 1).
Each 4×4 block of texels consists of 64 bits of compressed signed red image data followed by 64 bits of compressed signed green image data.
The first 64 bits of compressed red are decoded exactly like BC4 signed above. The second 64 bits of compressed green are decoded exactly like BC4 signed above except the decoded value R for this second block is considered the resulting green value G.
Since this image has a red-green format, the resulting RGBA value is (R, G, 0, 1).
This description is derived from the “BPTC Compressed Texture Image Formats” section of the OpenGL 4.4 specification. More information on BC7, BC7 modes and BC6h can be found in Microsoft’s online documentation.
Compressed texture images stored using the BPTC compressed image formats are represented as a collection of 4×4 texel blocks, each of which contains 128 bits of texel data stored in little-endian order. The image is encoded as a normal 2D raster image in which each 4×4 block is treated as a single pixel. If a BPTC image has a width or height that is not a multiple of four, the data corresponding to texels outside the image are irrelevant and undefined. When a BPTC image with width w, height h, and block size blocksize (16 bytes) is decoded, the corresponding image size (in bytes) is:
When decoding a BPTC image, the block containing the texel at offset (x, y) begins at an offset (in bytes) relative to the base of the image of:
The data corresponding to a specific texel (x, y) are extracted from a 4×4 texel block using a relative (x, y) value of:
There are two distinct BPTC image formats each of which has two variants. BC7 with or without an sRGB transform function used in the encoding of the RGB channels compresses 8-bit unsigned, normalized fixed-point data. BC6H in signed or unsigned form compresses high dynamic range floating-point values. The formats are similar, so the description of the BC6H format will reference significant sections of the BC7 description.
Each 4×4 block of texels consists of 128 bits of RGBA image data, of which the RGB channels may be encoded linearly or with the sRGB transfer function.
Each block contains enough information to select and decode a number of colors called endpoints, pairs of which forms subsets, then to interpolate between those endpoints in a variety of ways, and finally to remap the result into the final output by indexing into these interpolated values according to a partition layout which maps each relative coordinate to a subset.
Each block can contain data in one of eight modes. The mode is identified by the lowest bits of the lowest byte. It is encoded as zero or more zeros followed by a one. For example, using ‘x’ to indicate a bit not included in the mode number, mode 0 is encoded as xxxxxxx1 in the low byte in binary, mode 5 is xx100000, and mode 7 is 10000000. Encoding the low byte as zero is reserved and should not be used when encoding a BPTC texture; hardware decoders processing a texel block with a low byte of 0 should return 0 for all channels of all texels.
All further decoding is driven by the values derived from the mode listed in Table 109 and Table 110. The fields in the block are always in the same order for all modes. In increasing bit order after the mode, these fields are: partition pattern selection, rotation, index selection, color, alpha, per-endpoint P-bit, shared P-bit, primary indices, and secondary indices. The number of bits to be read in each field is determined directly from these tables, as shown in Table 111.
Per texel block, CB = 3(each of R, G, B)×2(endpoints)×NS(#subsets)×CB(bits/channel/endpoint). AB = 2(endpoints)×NS(#subsets)×AB(bits/endpoint). {IB,IB2} = 16(texels)×{IB,IB2}(#index bits/texel) - NS(1bit/subset). |
Table 109. Mode-dependent BPTC parameters
Mode | NS | PB | RB | ISB | CB | AB | EPB | SPB | IB | IB2 | M | CB | AB | EPB | SPB | IB | IB2 | |
Bits per… | …texel block | …channel/endpoint | …endpoint | …subset | …texel | Bits per texel block (total) | ||||||||||||
0 | 3 | 4 | 0 | 0 | 4 | 0 | 1 | 0 | 3 | 0 | 1 | 72 | 0 | 6 | 0 | 45 | 0 | |
1 | 2 | 6 | 0 | 0 | 6 | 0 | 0 | 1 | 3 | 0 | 2 | 72 | 0 | 0 | 2 | 46 | 0 | |
2 | 3 | 6 | 0 | 0 | 5 | 0 | 0 | 0 | 2 | 0 | 3 | 90 | 0 | 0 | 0 | 29 | 0 | |
3 | 2 | 6 | 0 | 0 | 7 | 0 | 1 | 0 | 2 | 0 | 4 | 84 | 0 | 4 | 0 | 30 | 0 | |
4 | 1 | 0 | 2 | 1 | 5 | 6 | 0 | 0 | 2 | 3 | 5 | 30 | 12 | 0 | 0 | 31 | 47 | |
5 | 1 | 0 | 2 | 0 | 7 | 8 | 0 | 0 | 2 | 2 | 6 | 42 | 16 | 0 | 0 | 31 | 31 | |
6 | 1 | 0 | 0 | 0 | 7 | 7 | 1 | 0 | 4 | 0 | 7 | 42 | 14 | 2 | 0 | 63 | 0 | |
7 | 2 | 6 | 0 | 0 | 5 | 5 | 1 | 0 | 2 | 0 | 8 | 60 | 20 | 4 | 0 | 30 | 0 |
Table 110. Full descriptions of the BPTC mode columns
M | Mode identifier bits |
NS | Number of subsets |
PB | Partition selection bits |
RB | Rotation bits |
ISB | Index selection bit |
CB | Color bits |
AB | Alpha bits |
EPB | Endpoint P-bits (all channels) |
SPB | Shared P-bits |
IB | Index bits |
IB2 | Secondary index bits |
Each block can be divided into between 1 and 3 groups of pixels called subsets, which have different endpoints. There are two endpoint colors per subset, grouped first by endpoint, then by subset, then by channel. For example, mode 1, with two subsets and six color bits, would have six bits of red for endpoint 0 of the first subset, then six bits of red for endpoint 1, then the two ends of the second subset, then green and blue stored similarly. If a block has any alpha bits, the alpha data follows the color data with the same organization. If not, alpha is overridden to 255. These bits are treated as the high bits of a fixed-point value in a byte for each color channel of the endpoints: {ER7..0, EG7..0, EB7..0, EA7..0} per endpoint. If the mode has shared P-bits, there are two endpoint bits, the lower of which applies to both endpoints of subset 0 and the upper of which applies to both endpoints of subset 1. If the mode has per-endpoint P-bits, then there are 2 × subsets P-bits stored in the same order as color and alpha. Both kinds of P-bits are added as a bit below the color data stored in the byte. So, for mode 1 with six red bits, the P-bit ends up in bit 1. For final scaling, the top bits of the value are replicated into any remaining bits in the byte. For the example of mode 1, bit 7 (which originated as bit 5 of the 6-bit encoded channel) would be replicated to bit 0. Table 112 and Table 113 show the origin of each endpoint color bit for each mode.
Table 111. Bit layout for BC7 modes (LSB..MSB)
Mode 0 | 0: M0 = 1 | 1..4: PB0..3 | ||||
5..8: R00..3 | 9..12: R10..3 | 13..16: R20..3 | 17..20: R30..3 | 21..24: R40..3 | 25..28: R50..3 | |
29..32: G00..3 | 33..36: G10..3 | 37..40: G20..3 | 41..44: G30..3 | 45..48: G40..3 | 49..52: G50..3 | |
53..56: B00..3 | 57..60: B10..3 | 61..64: B20..3 | 65..68: B30..3 | 69..72: B40..3 | 73..76: B50..3 | |
77: EPB00 | 78: EPB10 | 79: EPB20 | 80: EPB30 | 81: EPB40 | 82: EPB50 | |
83..127: IB0..44 | ||||||
Mode 1 | 0..1: M0..1 = 01 | 2..7: PB0..5 | ||||
8..13: R00..5 | 14..19: R10..5 | 20..25: R20..5 | 26..31: R30..5 | |||
32..37: G00..5 | 38..43: G10..5 | 44..49: G20..5 | 50..55: G30..5 | |||
56..61: B00..5 | 62..67: B10..5 | 68..73: B20..5 | 74..79: B30..5 | |||
80: SPB00 | 81: SPB10 | 82..127: IB0..45 | ||||
Mode 2 | 0..2: M0..2 = 001 | 3..8: PB0..5 | ||||
9..13: R00..4 | 14..18: R10..4 | 19..23: R20..4 | 24..28: R40..4 | 29..33: R40..4 | 34..38: R50..4 | |
39..43: G00..4 | 44..48: G10..4 | 49..53: G20..4 | 54..58: G40..4 | 59..63: G40..4 | 64..68: G50..4 | |
69..73: B00..4 | 74..78: B10..4 | 79..83: B20..4 | 84..88: B40..4 | 89..93: B40..4 | 94..98: B50..4 | |
99..127: IB0..28 | ||||||
Mode 3 | 0..3: M0..3 = 0001 | 4..9: PB0..5 | ||||
10..16: R00..6 | 17..23: R10..6 | 24..30: R20..6 | 31..37: R30..6 | |||
38..44: G00..6 | 45..51: G10..6 | 52..58: G20..6 | 59..65: G30..6 | |||
66..72: B00..6 | 73..79: B10..6 | 80..86: B20..6 | 87..93: B30..6 | |||
94: EPB00 | 95: EPB10 | 96: EPB20 | 97: EPB30 | 98..127: IB0..29 | ||
Mode 4 | 0..4: M0..4 = 00001 | 5..6: RB0..1 | 7: ISB0 | |||
8..12: R00..4 | 13..17: R10..4 | 18..22: G00..4 | 23..27: G10..4 | 28..32: B00..4 | 33..37: B10..4 | |
38..43: A00..5 | 44..49: A10..5 | 50..80: IB0..30 | 81..127: IB20..46 | |||
Mode 5 | 0..5: M0..5 = 000001 | 6..7: RB0..1 | ||||
8..14: R00..6 | 15..21: R10..6 | 22..28: G00..6 | 29..35: G10..6 | 36..42: B00..6 | 43..49: B10..6 | |
50..57: A00..7 | 58..65: A10..7 | 66..96: IB0..30 | 97..127: IB20..30 | |||
Mode 6 | 0..6: M0..6 = 0000001 | |||||
7..13: R00..6 | 14..20: R10..6 | 21..27: G00..6 | 28..34: G10..6 | 35..41: B00..6 | 42..48: B10..6 | |
49..55: A00..6 | 56..62: A10..6 | 63: EPB00 | 64: EPB10 | 65..127: IB0..62 | ||
Mode 7 | 0..7: M0..7 = 00000001 | 8..13: PB0..5 | ||||
14..18: R00..4 | 19..23: R10..4 | 24..28: R20..4 | 29..33: R30..4 | |||
34..38: G00..4 | 39..43: G10..4 | 44..48: G20..4 | 49..53: G30..4 | |||
54..58: B00..4 | 59..63: B10..4 | 64..68: B20..4 | 69..73: B30..4 | |||
74..78: A00..4 | 79..83: A10..4 | 84..88: A20..4 | 89..93: A30..4 | |||
94: EPB00 | 95: EPB10 | 96: EPB20 | 97: EPB30 | 98..127: IB0..29 |
Table 112. Bit sources for BC7 endpoints (modes 0..2, MSB..LSB per channel)
Mode 0 | |||||||||||||||||||||||||||||||
ER07..0 | EG07..0 | EB07..0 | EA07..0 | ||||||||||||||||||||||||||||
8 | 7 | 6 | 5 | 77 | 8 | 7 | 6 | 32 | 31 | 30 | 29 | 77 | 32 | 31 | 30 | 56 | 55 | 54 | 53 | 77 | 56 | 55 | 54 | 255 | |||||||
ER17..0 | EG17..0 | EB17..0 | EA17..0 | ||||||||||||||||||||||||||||
12 | 11 | 10 | 9 | 78 | 12 | 11 | 10 | 36 | 35 | 34 | 33 | 78 | 36 | 35 | 34 | 60 | 59 | 58 | 57 | 78 | 60 | 59 | 58 | 255 | |||||||
ER27..0 | EG27..0 | EB27..0 | EA27..0 | ||||||||||||||||||||||||||||
16 | 15 | 14 | 13 | 79 | 16 | 15 | 14 | 40 | 39 | 38 | 37 | 79 | 40 | 39 | 38 | 64 | 63 | 62 | 61 | 79 | 64 | 63 | 62 | 255 | |||||||
ER37..0 | EG37..0 | EB37..0 | EA37..0 | ||||||||||||||||||||||||||||
20 | 19 | 18 | 17 | 80 | 20 | 19 | 18 | 44 | 43 | 42 | 41 | 80 | 44 | 43 | 42 | 68 | 67 | 66 | 65 | 80 | 68 | 67 | 66 | 255 | |||||||
ER47..0 | EG47..0 | EB47..0 | EA47..0 | ||||||||||||||||||||||||||||
24 | 23 | 22 | 21 | 81 | 24 | 23 | 22 | 48 | 47 | 46 | 45 | 81 | 48 | 47 | 46 | 72 | 71 | 70 | 69 | 81 | 72 | 71 | 70 | 255 | |||||||
ER57..0 | EG57..0 | EB57..0 | EA57..0 | ||||||||||||||||||||||||||||
28 | 27 | 26 | 25 | 82 | 28 | 27 | 26 | 52 | 51 | 50 | 49 | 82 | 52 | 51 | 50 | 76 | 75 | 74 | 73 | 82 | 76 | 75 | 74 | 255 | |||||||
Mode 1 | |||||||||||||||||||||||||||||||
ER07..0 | EG07..0 | EB07..0 | EA07..0 | ||||||||||||||||||||||||||||
13 | 12 | 11 | 10 | 9 | 8 | 80 | 13 | 37 | 36 | 35 | 34 | 33 | 32 | 80 | 37 | 61 | 60 | 59 | 58 | 57 | 56 | 80 | 61 | 255 | |||||||
ER17..0 | EG17..0 | EB17..0 | EA17..0 | ||||||||||||||||||||||||||||
19 | 18 | 17 | 16 | 15 | 14 | 80 | 19 | 43 | 42 | 41 | 40 | 39 | 38 | 80 | 43 | 67 | 66 | 65 | 64 | 63 | 62 | 80 | 67 | 255 | |||||||
ER27..0 | EG27..0 | EB27..0 | EA27..0 | ||||||||||||||||||||||||||||
25 | 24 | 23 | 22 | 21 | 20 | 81 | 25 | 49 | 48 | 47 | 46 | 45 | 44 | 81 | 49 | 73 | 72 | 71 | 70 | 69 | 68 | 81 | 73 | 255 | |||||||
ER37..0 | EG37..0 | EB37..0 | EA37..0 | ||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 81 | 31 | 55 | 54 | 53 | 52 | 51 | 50 | 81 | 55 | 79 | 78 | 77 | 76 | 75 | 74 | 81 | 79 | 255 | |||||||
Mode 2 | |||||||||||||||||||||||||||||||
ER07..0 | EG07..0 | EB07..0 | EA07..0 | ||||||||||||||||||||||||||||
13 | 12 | 11 | 10 | 9 | 13 | 12 | 11 | 43 | 42 | 41 | 40 | 39 | 43 | 42 | 41 | 73 | 72 | 71 | 70 | 69 | 73 | 72 | 71 | 255 | |||||||
ER17..0 | EG17..0 | EB17..0 | EA17..0 | ||||||||||||||||||||||||||||
18 | 17 | 16 | 15 | 14 | 18 | 17 | 16 | 48 | 47 | 46 | 45 | 44 | 48 | 47 | 46 | 78 | 77 | 76 | 75 | 74 | 78 | 77 | 76 | 255 | |||||||
ER27..0 | EG27..0 | EB27..0 | EA27..0 | ||||||||||||||||||||||||||||
23 | 22 | 21 | 20 | 19 | 23 | 22 | 21 | 53 | 52 | 51 | 50 | 49 | 53 | 52 | 51 | 83 | 82 | 81 | 80 | 79 | 83 | 82 | 81 | 255 | |||||||
ER37..0 | EG37..0 | EB37..0 | EA37..0 | ||||||||||||||||||||||||||||
28 | 27 | 26 | 25 | 24 | 28 | 27 | 26 | 58 | 57 | 56 | 55 | 54 | 58 | 57 | 56 | 88 | 87 | 86 | 85 | 84 | 88 | 87 | 86 | 255 | |||||||
ER47..0 | EG47..0 | EB47..0 | EA47..0 | ||||||||||||||||||||||||||||
33 | 32 | 31 | 30 | 29 | 33 | 32 | 31 | 63 | 62 | 61 | 60 | 59 | 63 | 62 | 61 | 93 | 92 | 91 | 90 | 89 | 93 | 92 | 91 | 255 | |||||||
ER57..0 | EG57..0 | EB57..0 | EA57..0 | ||||||||||||||||||||||||||||
38 | 37 | 36 | 35 | 34 | 38 | 37 | 36 | 68 | 67 | 66 | 65 | 64 | 68 | 67 | 66 | 98 | 97 | 96 | 95 | 94 | 98 | 97 | 96 | 255 |
Table 113. Bit sources for BC7 endpoints (modes 3..7, MSB..LSB per channel)
Mode 3 | |||||||||||||||||||||||||||||||
ER07..0 | EG07..0 | EB07..0 | EA07..0 | ||||||||||||||||||||||||||||
16 | 15 | 14 | 13 | 12 | 11 | 10 | 94 | 44 | 43 | 42 | 41 | 40 | 39 | 38 | 94 | 72 | 71 | 70 | 69 | 68 | 67 | 66 | 94 | 255 | |||||||
ER17..0 | EG17..0 | EB17..0 | EA17..0 | ||||||||||||||||||||||||||||
23 | 22 | 21 | 20 | 19 | 18 | 17 | 95 | 51 | 50 | 49 | 48 | 47 | 46 | 45 | 95 | 79 | 78 | 77 | 76 | 75 | 74 | 73 | 95 | 255 | |||||||
ER27..0 | EG27..0 | EB27..0 | EA27..0 | ||||||||||||||||||||||||||||
30 | 29 | 28 | 27 | 26 | 25 | 24 | 96 | 58 | 57 | 56 | 55 | 54 | 53 | 52 | 96 | 86 | 85 | 84 | 83 | 82 | 81 | 80 | 96 | 255 | |||||||
ER37..0 | EG37..0 | EB37..0 | EA37..0 | ||||||||||||||||||||||||||||
37 | 36 | 35 | 34 | 33 | 32 | 31 | 97 | 65 | 64 | 63 | 62 | 61 | 60 | 59 | 97 | 93 | 92 | 91 | 90 | 89 | 88 | 87 | 97 | 255 | |||||||
Mode 4 | |||||||||||||||||||||||||||||||
ER07..0 | EG07..0 | EB07..0 | EA07..0 | ||||||||||||||||||||||||||||
12 | 11 | 10 | 9 | 8 | 12 | 11 | 10 | 22 | 21 | 20 | 19 | 18 | 22 | 21 | 20 | 32 | 31 | 30 | 29 | 28 | 32 | 31 | 30 | 43 | 42 | 41 | 40 | 39 | 38 | 43 | 42 |
ER17..0 | EG17..0 | EB17..0 | EA17..0 | ||||||||||||||||||||||||||||
17 | 16 | 15 | 14 | 13 | 17 | 16 | 15 | 27 | 26 | 25 | 24 | 23 | 27 | 26 | 25 | 37 | 36 | 35 | 34 | 33 | 37 | 36 | 35 | 49 | 48 | 47 | 46 | 45 | 44 | 49 | 48 |
Mode 5 | |||||||||||||||||||||||||||||||
ER07..0 | EG07..0 | EB07..0 | EA07..0 | ||||||||||||||||||||||||||||
14 | 13 | 12 | 11 | 10 | 9 | 8 | 14 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 28 | 42 | 41 | 40 | 39 | 38 | 37 | 36 | 42 | 57 | 56 | 55 | 54 | 53 | 52 | 51 | 50 |
ER17..0 | EG17..0 | EB17..0 | EA17..0 | ||||||||||||||||||||||||||||
21 | 20 | 19 | 18 | 17 | 16 | 15 | 21 | 35 | 34 | 33 | 32 | 31 | 30 | 29 | 35 | 49 | 48 | 47 | 46 | 45 | 44 | 43 | 49 | 65 | 64 | 63 | 62 | 61 | 60 | 59 | 58 |
Mode 6 | |||||||||||||||||||||||||||||||
ER07..0 | EG07..0 | EB07..0 | EA07..0 | ||||||||||||||||||||||||||||
13 | 12 | 11 | 10 | 9 | 8 | 7 | 63 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 63 | 41 | 40 | 39 | 38 | 37 | 36 | 35 | 63 | 55 | 54 | 53 | 52 | 51 | 50 | 49 | 63 |
ER17..0 | EG17..0 | EB17..0 | EA17..0 | ||||||||||||||||||||||||||||
20 | 19 | 18 | 17 | 16 | 15 | 14 | 64 | 34 | 33 | 32 | 31 | 30 | 29 | 28 | 64 | 48 | 47 | 46 | 45 | 44 | 43 | 42 | 64 | 62 | 61 | 60 | 59 | 58 | 57 | 56 | 64 |
Mode 7 | |||||||||||||||||||||||||||||||
ER07..0 | EG07..0 | EB07..0 | EA07..0 | ||||||||||||||||||||||||||||
18 | 17 | 16 | 15 | 14 | 94 | 18 | 17 | 38 | 37 | 36 | 35 | 34 | 94 | 38 | 37 | 58 | 57 | 56 | 55 | 54 | 94 | 58 | 57 | 78 | 77 | 76 | 75 | 74 | 94 | 78 | 77 |
ER17..0 | EG17..0 | EB17..0 | EA17..0 | ||||||||||||||||||||||||||||
23 | 22 | 21 | 20 | 19 | 95 | 23 | 22 | 43 | 42 | 41 | 40 | 39 | 95 | 43 | 42 | 63 | 62 | 61 | 60 | 59 | 95 | 63 | 62 | 83 | 82 | 81 | 80 | 79 | 95 | 83 | 82 |
ER27..0 | EG27..0 | EB27..0 | EA27..0 | ||||||||||||||||||||||||||||
28 | 27 | 26 | 25 | 24 | 96 | 28 | 27 | 48 | 47 | 46 | 45 | 44 | 96 | 48 | 47 | 68 | 67 | 66 | 65 | 64 | 96 | 68 | 67 | 88 | 87 | 86 | 85 | 84 | 96 | 88 | 87 |
ER37..0 | EG37..0 | EB37..0 | EA37..0 | ||||||||||||||||||||||||||||
33 | 32 | 31 | 30 | 29 | 97 | 33 | 32 | 53 | 52 | 51 | 50 | 49 | 97 | 53 | 52 | 73 | 72 | 71 | 70 | 69 | 97 | 73 | 72 | 93 | 92 | 91 | 90 | 89 | 97 | 93 | 92 |
A texel in a block with one subset is always considered to be in subset zero. Otherwise, a number encoded in the partition bits is used to look up a partition pattern in Table 114 or Table 115 for 2 subsets and 3 subsets respectively. This partition pattern is accessed by the relative x and y offsets within the block to determine the subset which defines the pixel at these coordinates.
The endpoint colors are interpolated using index values stored in the block. The index bits are stored in y-major order. That is, the bits for the index value corresponding to a relative (x, y) position of (0, 0) are stored in increasing order in the lowest index bits of the block (but see the next paragraph about anchor indices), the next bits of the block in increasing order store the index bits of (1, 0), followed by (2, 0) and (3, 0), then (0, 1) etc.
Each index has the number of bits indicated by the mode except for one special index per subset called the anchor index. Since the interpolation scheme between endpoints is symmetrical, we can save one bit on one index per subset by ordering the endpoints such that the highest bit for that index is guaranteed to be zero — and not storing that bit.
Each anchor index corresponds to an index in the corresponding partition number in Table 114 or Table 115, and are indicated in bold italics in those tables. In partition zero, the anchor index is always index zero — that is, at a relative position of (0,0) (as can be seen in Table 114 and Table 115, index 0 always corresponds to partition zero). In other partitions, the anchor index is specified by Table 118, Table 116, and Table 117.
In summary, the bit offset for index data with relative x,y coordinates within the texel block is: \begin{align*}
\textrm{index offset}_{x,y} &= \begin{cases}
0, & x = y = 0 \\
\textrm{IB} \times (x + 4\times y) - 1, & \textrm{NS} = 1,\ 0 < x + 4\times y \\
\textrm{IB} \times (x + 4\times y) - 1, & \textrm{NS} = 2,\ 0 < x + 4\times y \leq \textrm{anchor}_2[\mathit{part}] \\
\textrm{IB} \times (x + 4\times y) - 2, & \textrm{NS} = 2,\ \textrm{anchor}_2[\mathit{part}] < x + 4\times y \\
\textrm{IB} \times (x + 4\times y) - 1, & \textrm{NS} = 3,\ 0 < x + 4\times y \leq \textrm{anchor}_{3,2}[\mathit{part}],\ x + 4\times y \leq \textrm{anchor}_{3,2}[\mathit{part}]\\
\textrm{IB} \times (x + 4\times y) - 3, & \textrm{NS} = 3,\ x + 4\times y > \textrm{anchor}_{3,2}[\mathit{part}],\ x + 4\times y > \textrm{anchor}_{3,3}[\mathit{part}] \\
\textrm{IB} \times (x + 4\times y) - 2, & \textrm{NS} = 3,\ \textrm{otherwise} \\
\end{cases} \\
\end{align*} where anchor2 is Table 118, anchor3,2 is Table 116, anchor3,3 is Table 117, and part is encoded in the partition selection bits PB. |
If secondary index bits are present, they follow the primary index bits and are read in the same manner. The anchor index information is only used to determine the number of bits each index has when read from the block data.
The endpoint color and alpha values used for final interpolation are the decoded values corresponding to the applicable subset as selected above. The index value for interpolating color comes from the secondary index bits for the texel if the mode has an index selection bit and its value is one, and from the primary index bits otherwise. The alpha index comes from the secondary index bits if the block has a secondary index and the block either doesn’t have an index selection bit or that bit is zero, and from the primary index bits otherwise.
As an example of the texel decode process, consider a block encoded with mode 2 — that is, M0 = 0, M1 = 0, M2 = 1. This mode has three subsets, so Table 115 is used to determine which subset applies to each texel. Let us assume that this block has partition pattern 6 encoded in the partition selection bits, and that we wish to decode the texel at relative (x, y) offset (1, 2) — that is, index 9 in y-major order. We can see from Table 115 that this texel is partitioned into subset 1 (the second of three), and therefore by endpoints 2 and 3. Mode 2 stores two index bits per texel, except for index 0 (which is the anchor index for subset 0), index 15 (for subset 1, as indicated in Table 116) and index 3 (for subset 2, as indicated in Table 117). Index 9 is therefore stored in two bits starting at index bits offset 14 (for indices 1..2 and 4..8) plus 2 (for indices 0 and 3) — a total of 16 bit offset into the index bits or, as seen in Table 111, bits 115 and 116 of the block. These two bits are used to interpolate between endpoints 2 and 3 using Equation 2 with weights from the two-bit index row of Table 119, as described below. |
Table 114. Partition table for 2-subset BPTC, with the 4×4 block of values for each partition number
0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | |||||||||||||||||||||||||||||||
0 | 0 | 1 | 1 | 0 | 0 | 0 | 1 | 0 | 1 | 1 | 1 | 0 | 0 | 0 | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 1 | 0 | 0 | 0 | 1 | 0 | 0 | 0 | 0 | |||||||
0 | 0 | 1 | 1 | 0 | 0 | 0 | 1 | 0 | 1 | 1 | 1 | 0 | 0 | 1 | 1 | 0 | 0 | 0 | 1 | 0 | 1 | 1 | 1 | 0 | 0 | 1 | 1 | 0 | 0 | 0 | 1 | |||||||
0 | 0 | 1 | 1 | 0 | 0 | 0 | 1 | 0 | 1 | 1 | 1 | 0 | 0 | 1 | 1 | 0 | 0 | 0 | 1 | 0 | 1 | 1 | 1 | 0 | 1 | 1 | 1 | 0 | 0 | 1 | 1 | |||||||
0 | 0 | 1 | 1 | 0 | 0 | 0 | 1 | 0 | 1 | 1 | 1 | 0 | 1 | 1 | 1 | 0 | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 0 | 1 | 1 | 1 | |||||||
8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | |||||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | 0 | 0 | 1 | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | |||||||
0 | 0 | 0 | 0 | 0 | 1 | 1 | 1 | 0 | 0 | 0 | 1 | 0 | 0 | 0 | 0 | 0 | 1 | 1 | 1 | 0 | 0 | 0 | 0 | 1 | 1 | 1 | 1 | 0 | 0 | 0 | 0 | |||||||
0 | 0 | 0 | 1 | 1 | 1 | 1 | 1 | 0 | 1 | 1 | 1 | 0 | 0 | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 0 | 0 | 0 | 0 | |||||||
0 | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | |||||||
16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 | |||||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | 0 | 1 | 1 | 1 | 0 | 0 | 0 | 0 | 0 | 1 | 1 | 1 | 0 | 0 | 1 | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 1 | 1 | |||||||
1 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 1 | 0 | 0 | 0 | 1 | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 1 | |||||||
1 | 1 | 1 | 0 | 0 | 0 | 0 | 0 | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 0 | 0 | 0 | 0 | 1 | 1 | 0 | 0 | 1 | 0 | 0 | 0 | 0 | 0 | 1 | 1 | |||||||
1 | 1 | 1 | 1 | 0 | 0 | 0 | 0 | 1 | 1 | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 1 | 1 | 0 | 1 | 1 | 0 | 0 | 0 | 0 | 0 | 1 | |||||||
24 | 25 | 26 | 27 | 28 | 29 | 30 | 31 | |||||||||||||||||||||||||||||||
0 | 0 | 1 | 1 | 0 | 0 | 0 | 0 | 0 | 1 | 1 | 0 | 0 | 0 | 1 | 1 | 0 | 0 | 0 | 1 | 0 | 0 | 0 | 0 | 0 | 1 | 1 | 1 | 0 | 0 | 1 | 1 | |||||||
0 | 0 | 0 | 1 | 1 | 0 | 0 | 0 | 0 | 1 | 1 | 0 | 0 | 1 | 1 | 0 | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 0 | 0 | 0 | 1 | 1 | 0 | 0 | 1 | |||||||
0 | 0 | 0 | 1 | 1 | 0 | 0 | 0 | 0 | 1 | 1 | 0 | 0 | 1 | 1 | 0 | 1 | 1 | 1 | 0 | 1 | 1 | 1 | 1 | 1 | 0 | 0 | 0 | 1 | 0 | 0 | 1 | |||||||
0 | 0 | 0 | 0 | 1 | 1 | 0 | 0 | 0 | 1 | 1 | 0 | 1 | 1 | 0 | 0 | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 1 | 1 | 0 | 1 | 1 | 0 | 0 | |||||||
32 | 33 | 34 | 35 | 36 | 37 | 38 | 39 | |||||||||||||||||||||||||||||||
0 | 1 | 0 | 1 | 0 | 0 | 0 | 0 | 0 | 1 | 0 | 1 | 0 | 0 | 1 | 1 | 0 | 0 | 1 | 1 | 0 | 1 | 0 | 1 | 0 | 1 | 1 | 0 | 0 | 1 | 0 | 1 | |||||||
0 | 1 | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 0 | 1 | 0 | 0 | 0 | 1 | 1 | 1 | 1 | 0 | 0 | 0 | 1 | 0 | 1 | 1 | 0 | 0 | 1 | 1 | 0 | 1 | 0 | |||||||
0 | 1 | 0 | 1 | 0 | 0 | 0 | 0 | 0 | 1 | 0 | 1 | 1 | 1 | 0 | 0 | 0 | 0 | 1 | 1 | 1 | 0 | 1 | 0 | 0 | 1 | 1 | 0 | 1 | 0 | 1 | 0 | |||||||
0 | 1 | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 0 | 1 | 0 | 1 | 1 | 0 | 0 | 1 | 1 | 0 | 0 | 1 | 0 | 1 | 0 | 1 | 0 | 0 | 1 | 0 | 1 | 0 | 1 | |||||||
40 | 41 | 42 | 43 | 44 | 45 | 46 | 47 | |||||||||||||||||||||||||||||||
0 | 1 | 1 | 1 | 0 | 0 | 0 | 1 | 0 | 0 | 1 | 1 | 0 | 0 | 1 | 1 | 0 | 1 | 1 | 0 | 0 | 0 | 1 | 1 | 0 | 1 | 1 | 0 | 0 | 0 | 0 | 0 | |||||||
0 | 0 | 1 | 1 | 0 | 0 | 1 | 1 | 0 | 0 | 1 | 0 | 1 | 0 | 1 | 1 | 1 | 0 | 0 | 1 | 1 | 1 | 0 | 0 | 0 | 1 | 1 | 0 | 0 | 1 | 1 | 0 | |||||||
1 | 1 | 0 | 0 | 1 | 1 | 0 | 0 | 0 | 1 | 0 | 0 | 1 | 1 | 0 | 1 | 1 | 0 | 0 | 1 | 1 | 1 | 0 | 0 | 1 | 0 | 0 | 1 | 0 | 1 | 1 | 0 | |||||||
1 | 1 | 1 | 0 | 1 | 0 | 0 | 0 | 1 | 1 | 0 | 0 | 1 | 1 | 0 | 0 | 0 | 1 | 1 | 0 | 0 | 0 | 1 | 1 | 1 | 0 | 0 | 1 | 0 | 0 | 0 | 0 | |||||||
48 | 49 | 50 | 51 | 52 | 53 | 54 | 55 | |||||||||||||||||||||||||||||||
0 | 1 | 0 | 0 | 0 | 0 | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 1 | 0 | 0 | 0 | 1 | 1 | 0 | 1 | 1 | 0 | 0 | 0 | 1 | 1 | |||||||
1 | 1 | 1 | 0 | 0 | 1 | 1 | 1 | 0 | 0 | 1 | 0 | 0 | 1 | 0 | 0 | 1 | 1 | 0 | 0 | 0 | 1 | 1 | 0 | 0 | 0 | 1 | 1 | 1 | 0 | 0 | 1 | |||||||
0 | 1 | 0 | 0 | 0 | 0 | 1 | 0 | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 0 | 1 | 0 | 0 | 1 | 1 | 1 | 0 | 0 | 1 | 0 | 0 | 1 | 1 | 1 | 0 | 0 | |||||||
0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 0 | 0 | 1 | 0 | 0 | 0 | 0 | 1 | 1 | 1 | 0 | 0 | 1 | 1 | 1 | 0 | 0 | 0 | 1 | 1 | 0 | |||||||
56 | 57 | 58 | 59 | 60 | 61 | 62 | 63 | |||||||||||||||||||||||||||||||
0 | 1 | 1 | 0 | 0 | 1 | 1 | 0 | 0 | 1 | 1 | 1 | 0 | 0 | 0 | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 1 | 0 | 0 | 1 | 0 | 0 | 1 | 0 | 0 | |||||||
1 | 1 | 0 | 0 | 0 | 0 | 1 | 1 | 1 | 1 | 1 | 0 | 1 | 0 | 0 | 0 | 1 | 1 | 1 | 1 | 0 | 0 | 1 | 1 | 0 | 0 | 1 | 0 | 0 | 1 | 0 | 0 | |||||||
1 | 1 | 0 | 0 | 0 | 0 | 1 | 1 | 1 | 0 | 0 | 0 | 1 | 1 | 1 | 0 | 0 | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 0 | 0 | 1 | 1 | 1 | |||||||
1 | 0 | 0 | 1 | 1 | 0 | 0 | 1 | 0 | 0 | 0 | 1 | 0 | 1 | 1 | 1 | 0 | 0 | 1 | 1 | 0 | 0 | 0 | 0 | 1 | 1 | 1 | 0 | 0 | 1 | 1 | 1 |
Table 115. Partition table for 3-subset BPTC, with the 4×4 block of values for each partition number
0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | |||||||||||||||||||||||||||||||
0 | 0 | 1 | 1 | 0 | 0 | 0 | 1 | 0 | 0 | 0 | 0 | 0 | 2 | 2 | 2 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 1 | 0 | 0 | 2 | 2 | 0 | 0 | 1 | 1 | |||||||
0 | 0 | 1 | 1 | 0 | 0 | 1 | 1 | 2 | 0 | 0 | 1 | 0 | 0 | 2 | 2 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 1 | 0 | 0 | 2 | 2 | 0 | 0 | 1 | 1 | |||||||
0 | 2 | 2 | 1 | 2 | 2 | 1 | 1 | 2 | 2 | 1 | 1 | 0 | 0 | 1 | 1 | 1 | 1 | 2 | 2 | 0 | 0 | 2 | 2 | 1 | 1 | 1 | 1 | 2 | 2 | 1 | 1 | |||||||
2 | 2 | 2 | 2 | 2 | 2 | 2 | 1 | 2 | 2 | 1 | 1 | 0 | 1 | 1 | 1 | 1 | 1 | 2 | 2 | 0 | 0 | 2 | 2 | 1 | 1 | 1 | 1 | 2 | 2 | 1 | 1 | |||||||
8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | |||||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 2 | 0 | 1 | 1 | 2 | 0 | 1 | 2 | 2 | 0 | 0 | 1 | 1 | 0 | 0 | 1 | 1 | |||||||
0 | 0 | 0 | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 0 | 0 | 1 | 2 | 0 | 1 | 1 | 2 | 0 | 1 | 2 | 2 | 0 | 1 | 1 | 2 | 2 | 0 | 0 | 1 | |||||||
1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 2 | 2 | 2 | 2 | 0 | 0 | 1 | 2 | 0 | 1 | 1 | 2 | 0 | 1 | 2 | 2 | 1 | 1 | 2 | 2 | 2 | 2 | 0 | 0 | |||||||
2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 0 | 0 | 1 | 2 | 0 | 1 | 1 | 2 | 0 | 1 | 2 | 2 | 1 | 2 | 2 | 2 | 2 | 2 | 2 | 0 | |||||||
16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 | |||||||||||||||||||||||||||||||
0 | 0 | 0 | 1 | 0 | 1 | 1 | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 2 | 2 | 0 | 1 | 1 | 1 | 0 | 0 | 0 | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | |||||||
0 | 0 | 1 | 1 | 0 | 0 | 1 | 1 | 1 | 1 | 2 | 2 | 0 | 0 | 2 | 2 | 0 | 1 | 1 | 1 | 0 | 0 | 0 | 1 | 0 | 0 | 1 | 1 | 1 | 1 | 0 | 0 | |||||||
0 | 1 | 1 | 2 | 2 | 0 | 0 | 1 | 1 | 1 | 2 | 2 | 0 | 0 | 2 | 2 | 0 | 2 | 2 | 2 | 2 | 2 | 2 | 1 | 0 | 1 | 2 | 2 | 2 | 2 | 1 | 0 | |||||||
1 | 1 | 2 | 2 | 2 | 2 | 0 | 0 | 1 | 1 | 2 | 2 | 1 | 1 | 1 | 1 | 0 | 2 | 2 | 2 | 2 | 2 | 2 | 1 | 0 | 1 | 2 | 2 | 2 | 2 | 1 | 0 | |||||||
24 | 25 | 26 | 27 | 28 | 29 | 30 | 31 | |||||||||||||||||||||||||||||||
0 | 1 | 2 | 2 | 0 | 0 | 1 | 2 | 0 | 1 | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 2 | 2 | 0 | 1 | 1 | 0 | 0 | 0 | 1 | 1 | 0 | 0 | 0 | 0 | |||||||
0 | 1 | 2 | 2 | 0 | 0 | 1 | 2 | 1 | 2 | 2 | 1 | 0 | 1 | 1 | 0 | 1 | 1 | 0 | 2 | 0 | 1 | 1 | 0 | 0 | 1 | 2 | 2 | 2 | 0 | 0 | 0 | |||||||
0 | 0 | 1 | 1 | 1 | 1 | 2 | 2 | 1 | 2 | 2 | 1 | 1 | 2 | 2 | 1 | 1 | 1 | 0 | 2 | 2 | 0 | 0 | 2 | 0 | 1 | 2 | 2 | 2 | 2 | 1 | 1 | |||||||
0 | 0 | 0 | 0 | 2 | 2 | 2 | 2 | 0 | 1 | 1 | 0 | 1 | 2 | 2 | 1 | 0 | 0 | 2 | 2 | 2 | 2 | 2 | 2 | 0 | 0 | 1 | 1 | 2 | 2 | 2 | 1 | |||||||
32 | 33 | 34 | 35 | 36 | 37 | 38 | 39 | |||||||||||||||||||||||||||||||
0 | 0 | 0 | 0 | 0 | 2 | 2 | 2 | 0 | 0 | 1 | 1 | 0 | 1 | 2 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 2 | 0 | 0 | 1 | 2 | 0 | 0 | 0 | 1 | 1 | |||||||
0 | 0 | 0 | 2 | 0 | 0 | 2 | 2 | 0 | 0 | 1 | 2 | 0 | 1 | 2 | 0 | 1 | 1 | 1 | 1 | 1 | 2 | 0 | 1 | 2 | 0 | 1 | 2 | 2 | 2 | 0 | 0 | |||||||
1 | 1 | 2 | 2 | 0 | 0 | 1 | 2 | 0 | 0 | 2 | 2 | 0 | 1 | 2 | 0 | 2 | 2 | 2 | 2 | 2 | 0 | 1 | 2 | 1 | 2 | 0 | 1 | 1 | 1 | 2 | 2 | |||||||
1 | 2 | 2 | 2 | 0 | 0 | 1 | 1 | 0 | 2 | 2 | 2 | 0 | 1 | 2 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 2 | 0 | 0 | 1 | 2 | 0 | 0 | 0 | 1 | 1 | |||||||
40 | 41 | 42 | 43 | 44 | 45 | 46 | 47 | |||||||||||||||||||||||||||||||
0 | 0 | 1 | 1 | 0 | 1 | 0 | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 2 | 2 | 0 | 0 | 2 | 2 | 0 | 2 | 2 | 0 | 0 | 1 | 0 | 1 | 0 | 0 | 0 | 0 | |||||||
1 | 1 | 2 | 2 | 0 | 1 | 0 | 1 | 0 | 0 | 0 | 0 | 1 | 1 | 2 | 2 | 0 | 0 | 1 | 1 | 1 | 2 | 2 | 1 | 2 | 2 | 2 | 2 | 2 | 1 | 2 | 1 | |||||||
2 | 2 | 0 | 0 | 2 | 2 | 2 | 2 | 2 | 1 | 2 | 1 | 0 | 0 | 2 | 2 | 0 | 0 | 2 | 2 | 0 | 2 | 2 | 0 | 2 | 2 | 2 | 2 | 2 | 1 | 2 | 1 | |||||||
0 | 0 | 1 | 1 | 2 | 2 | 2 | 2 | 2 | 1 | 2 | 1 | 1 | 1 | 2 | 2 | 0 | 0 | 1 | 1 | 1 | 2 | 2 | 1 | 0 | 1 | 0 | 1 | 2 | 1 | 2 | 1 | |||||||
48 | 49 | 50 | 51 | 52 | 53 | 54 | 55 | |||||||||||||||||||||||||||||||
0 | 1 | 0 | 1 | 0 | 2 | 2 | 2 | 0 | 0 | 0 | 2 | 0 | 0 | 0 | 0 | 0 | 2 | 2 | 2 | 0 | 0 | 0 | 2 | 0 | 1 | 1 | 0 | 0 | 0 | 0 | 0 | |||||||
0 | 1 | 0 | 1 | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 2 | 2 | 1 | 1 | 2 | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 2 | 0 | 1 | 1 | 0 | 0 | 0 | 0 | 0 | |||||||
0 | 1 | 0 | 1 | 0 | 2 | 2 | 2 | 0 | 0 | 0 | 2 | 2 | 1 | 1 | 2 | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 2 | 0 | 1 | 1 | 0 | 2 | 1 | 1 | 2 | |||||||
2 | 2 | 2 | 2 | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 2 | 2 | 1 | 1 | 2 | 0 | 2 | 2 | 2 | 0 | 0 | 0 | 2 | 2 | 2 | 2 | 2 | 2 | 1 | 1 | 2 | |||||||
56 | 57 | 58 | 59 | 60 | 61 | 62 | 63 | |||||||||||||||||||||||||||||||
0 | 1 | 1 | 0 | 0 | 0 | 2 | 2 | 0 | 0 | 2 | 2 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 2 | 0 | 2 | 2 | 2 | 0 | 1 | 0 | 1 | 0 | 1 | 1 | 1 | |||||||
0 | 1 | 1 | 0 | 0 | 0 | 1 | 1 | 1 | 1 | 2 | 2 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 1 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 0 | 1 | 1 | |||||||
2 | 2 | 2 | 2 | 0 | 0 | 1 | 1 | 1 | 1 | 2 | 2 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 2 | 0 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 0 | 1 | |||||||
2 | 2 | 2 | 2 | 0 | 0 | 2 | 2 | 0 | 0 | 2 | 2 | 2 | 1 | 1 | 2 | 0 | 0 | 0 | 1 | 1 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 0 |
Table 116. BPTC anchor index values for the second subset of three-subset partitioning, by partition number
0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 |
3 | 3 | 15 | 15 | 8 | 3 | 15 | 15 |
8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 |
8 | 8 | 6 | 6 | 6 | 5 | 3 | 3 |
16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 |
3 | 3 | 8 | 15 | 3 | 3 | 6 | 10 |
24 | 25 | 26 | 27 | 28 | 29 | 30 | 31 |
5 | 8 | 8 | 6 | 8 | 5 | 15 | 15 |
32 | 33 | 34 | 35 | 36 | 37 | 38 | 39 |
8 | 15 | 3 | 5 | 6 | 10 | 8 | 15 |
40 | 41 | 42 | 43 | 44 | 45 | 46 | 47 |
15 | 3 | 15 | 5 | 15 | 15 | 15 | 15 |
48 | 49 | 50 | 51 | 52 | 53 | 54 | 55 |
3 | 15 | 5 | 5 | 5 | 8 | 5 | 10 |
56 | 57 | 58 | 59 | 60 | 61 | 62 | 63 |
5 | 10 | 8 | 13 | 15 | 12 | 3 | 3 |
Table 117. BPTC anchor index values for the third subset of three-subset partitioning, by partition number
0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 |
15 | 8 | 8 | 3 | 15 | 15 | 3 | 8 |
8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 |
15 | 15 | 15 | 15 | 15 | 15 | 15 | 8 |
16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 |
15 | 8 | 15 | 3 | 15 | 8 | 15 | 8 |
24 | 25 | 26 | 27 | 28 | 29 | 30 | 31 |
3 | 15 | 6 | 10 | 15 | 15 | 10 | 8 |
32 | 33 | 34 | 35 | 36 | 37 | 38 | 39 |
15 | 3 | 15 | 10 | 10 | 8 | 9 | 10 |
40 | 41 | 42 | 43 | 44 | 45 | 46 | 47 |
6 | 15 | 8 | 15 | 3 | 6 | 6 | 8 |
48 | 49 | 50 | 51 | 52 | 53 | 54 | 55 |
15 | 3 | 15 | 15 | 15 | 15 | 15 | 15 |
56 | 57 | 58 | 59 | 60 | 61 | 62 | 63 |
15 | 15 | 15 | 15 | 3 | 15 | 15 | 8 |
Table 118. BPTC anchor index values for the second subset of two-subset partitioning, by partition number
0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 |
15 | 15 | 15 | 15 | 15 | 15 | 15 | 15 |
8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 |
15 | 15 | 15 | 15 | 15 | 15 | 15 | 15 |
16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 |
15 | 2 | 8 | 2 | 2 | 8 | 8 | 15 |
24 | 25 | 26 | 27 | 28 | 29 | 30 | 31 |
2 | 8 | 2 | 2 | 8 | 8 | 2 | 2 |
32 | 33 | 34 | 35 | 36 | 37 | 38 | 39 |
15 | 15 | 6 | 8 | 2 | 8 | 15 | 15 |
40 | 41 | 42 | 43 | 44 | 45 | 46 | 47 |
2 | 8 | 2 | 2 | 2 | 15 | 15 | 6 |
48 | 49 | 50 | 51 | 52 | 53 | 54 | 55 |
6 | 2 | 6 | 8 | 15 | 15 | 2 | 2 |
56 | 57 | 58 | 59 | 60 | 61 | 62 | 63 |
15 | 15 | 15 | 15 | 15 | 2 | 2 | 15 |
Interpolation is always performed using a 6-bit interpolation factor. The effective interpolation factors for 2-, 3-, and 4-bit indices are given in Table 119.
Table 119. BPTC interpolation factors
2 | Index | 0 | 1 | 2 | 3 | |||||||||||||
Weight | 0 | 21 | 43 | 64 | ||||||||||||||
3 | Index | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | |||||||||
Weight | 0 | 9 | 18 | 27 | 37 | 46 | 55 | 64 | ||||||||||
4 | Index | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | |
Weight | 0 | 4 | 9 | 13 | 17 | 21 | 26 | 30 | 34 | 38 | 43 | 47 | 51 | 55 | 60 | 64 |
Given E0 and E1, unsigned integer endpoints [0 .. 255] for each channel and weight as an unsigned integer interpolation factor from Table 119:
Equation 2. BPTC endpoint interpolation formula
where $\gg$ performs a (truncating) bitwise right-shift, and interpolated value is an (unsigned) integer in the range [0..255]. The interpolation results in an RGBA color. If rotation bits are present, this interpolated color is remapped according to Table 120.
These 8-bit values should be interpreted as RGBA 8-bit normalized channels, either linearly encoded (by multiplying by $1\over 255$ ) or with the sRGB transfer function.
Each 4×4 block of texels consists of 128 bits of RGB data. The signed and unsigned formats are very similar and will be described together. In the description and pseudocode below, signed will be used as a condition which is true for the signed version of the format and false for the unsigned version of the format. Both formats only contain RGB data, so the returned alpha value is 1.0. If a block uses a reserved or invalid encoding, the return value is (0.0, 0.0, 0.0, 1.0).
Where BC7 encodes a fixed-point 8-bit value, BC6H encodes a 16-bit integer which will be interpreted as a 16-bit half float. Interpolation in BC6H is therefore nonlinear, but monotonic. |
Each block can contain data in one of 14 modes. The mode number is encoded in either the low two bits or the low five bits. If the low two bits are less than two, that is the mode number, otherwise the low five bits is the mode number. Mode numbers not listed in Table 121 (19, 23, 27, and 31) are reserved.
Table 121. Endpoint and partition parameters for BPTC block modes
Mode number | Transformed endpoints | Partition bits (PB) | Endpoint bits (EPB) | Delta bits | Mode | Endpoint | Delta | |
Bits per texel block | {R,G,B} bits per endpoint | Bits per texel block (total) | ||||||
0 | ✓ | 5 | {10, 10, 10} | {5, 5, 5} | 2 | 30 | 45 | |
1 | ✓ | 5 | {7, 7, 7} | {6, 6, 6} | 2 | 21 | 54 | |
2 | ✓ | 5 | {11, 11, 11} | {5, 4, 4} | 5 | 33 | 39 | |
6 | ✓ | 5 | {11, 11, 11} | {4, 5, 4} | 5 | 33 | 39 | |
10 | ✓ | 5 | {11, 11, 11} | {4, 4, 5} | 5 | 33 | 39 | |
14 | ✓ | 5 | {9, 9, 9} | {5, 5, 5} | 5 | 27 | 45 | |
18 | ✓ | 5 | {8, 8, 8} | {6, 5, 5} | 5 | 24 | 48 | |
22 | ✓ | 5 | {8, 8, 8} | {5, 6, 5} | 5 | 24 | 48 | |
26 | ✓ | 5 | {8, 8, 8} | {5, 5, 6} | 5 | 24 | 48 | |
30 | 5 | {6, 6, 6} | - | 5 | 72 | 0 | ||
3 | 0 | {10, 10, 10} | - | 5 | 60 | 0 | ||
7 | ✓ | 0 | {11, 11, 11} | {9, 9, 9} | 5 | 33 | 27 | |
11 | ✓ | 0 | {12, 12, 12} | {8, 8, 8} | 5 | 36 | 24 | |
15 | ✓ | 0 | {16, 16, 16} | {4, 4, 4} | 5 | 48 | 12 |
The data for the compressed blocks is stored in a different manner for each mode. The interpretation of bits for each mode are specified in Table 122. The descriptions are intended to be read from left to right with the LSB on the left. Each element is of the form va..b. If a ≥ b, this indicates extracting b - a + 1 bits from the block at that location and put them in the corresponding bits of the variable v. If a < b, then the bits are reversed. va is used as a shorthand for the one bit va..a. As an example, M1..0, G24 would move the low two bits from the block into the low two bits of mode number M, then the next bit of the block into bit 4 of G2. The resultant bit interpretations are shown explicitly in Table 123 and Table 124. The variable names given in the table will be referred to in the language below.
Subsets and indices work in much the same way as described for the BC7 formats above. If a float block has no partition bits, then it is a single-subset block. If it has partition bits, then it is a two-subset block. The partition number references the first half of Table 114.
Table 122. Block descriptions for BC6H block modes (LSB..MSB)
Mode Number | Block description |
---|---|
0 | M1..0, G24, B24, B34, R09..0, G09..0, B09..0, R14..0, G34, G23..0, G14..0, B30, G33..0, B14..0, B31, B23..0, R24..0, B32, R34..0, B33, PB4..0 |
1 | M1..0, G25, G34, G35, R06..0, B30, B31, B24, G06..0, B25, B32, G24, B06..0, B33, B35, B34, R15..0, G23..0, G15..0, G33..0, B15..0, B23..0, R25..0, R35..0, PB4..0 |
2 | M4..0, R09..0, G09..0, B09..0, R14..0, R010, G23..0, G13..0, G010, B30, G33..0, B13..0, B010, B31, B23..0, R24..0, B32, R34..0, B33, PB4..0 |
6 | M4..0, R09..0, G09..0, B09..0, R13..0, R010, G34, G23..0, G14..0, G010, G33..0, B13..0, B010, B31, B23..0, R23..0, B30, B32, R33..0, G24, B33, PB4..0 |
10 | M4..0, R09..0, G09..0, B09..0, R13..0, R010, B24, G23..0, G13..0, G010, B30, G33..0, B14..0, B010, B23..0, R23..0, B31, B32, R33..0, B34, B33, PB4..0 |
14 | M4..0, R08..0, B24, G08..0, G24, B08..0, B34, R14..0, G34, G23..0, G14..0, B30, G33..0, B14..0, B31, B23..0, R24..0, B32, R34..0, B33, PB4..0 |
18 | M4..0, R07..0, G34, B24, G07..0, B32, G24, B07..0, B33, B34, R15..0, G23..0, G14..0, B30, G33..0, B14..0, B31, B23..0, R25..0, R35..0, PB4..0 |
22 | M4..0, R07..0, B30, B24, G07..0, G25, G24, B07..0, G35, B34, R14..0, G34, G23..0, G15..0, G33..0, B14..0, B31, B23..0, R24..0, B32, R34..0, B33, PB4..0 |
26 | M4..0, R07..0, B31, B24, G07..0, B25, G24, B07..0, B35, B34, R14..0, G34, G23..0, G14..0, B30, G33..0, B15..0, B23..0, R24..0, B32, R34..0, B33, PB4..0 |
30 | M4..0, R05..0, G34, B30, B31, B24, G05..0, G25, B25, B32, G24, B05..0, G35, B33, B35, B34, R15..0, G23..0, G15..0, G33..0, B15..0, B23..0, R25..0, R35..0, PB4..0 |
3 | M4..0, R09..0, G09..0, B09..0, R19..0, G19..0, B19..0 |
7 | M4..0, R09..0, G09..0, B09..0, R18..0, R010, G18..0, G010, B18..0, B010 |
11 | M4..0, R09..0, G09..0, B09..0, R17..0, R010..11, G17..0, G010..11, B17..0, B010..11 |
15 | M4..0, R09..0, G09..0, B09..0, R13..0, R010..15, G13..0, G010..15, B13..0, B010..15 |
Indices are read in the same way as the BC7 formats including obeying the anchor values for index 0 and as needed by Table 118. That is, for modes with only one partition, the mode and endpoint data are followed by 63 bits of index data (four index bits IBx,y0..3 per texel, with one implicit bit for IBx,y3) starting at bit 65 with IB0,00. For modes with two partitions, the mode, endpoint and partition data are followed by 46 bits of index data (three per texel IBx,y0..2, with two implicit bits, one for partition 0 at IB0,02 and one IBx,y2 bit for partition 1 at an offset determined by the partition pattern selected) starting at bit 82 with IB0,00. In both cases, index bits are stored in y-major offset order by increasing little-endian bit number, with the bits for each index stored consecutively:
Table 123. Interpretation of lower bits for BC6H block modes
Mode | ||||||||||||||
Bit | 0 | 1 | 2 | 6 | 10 | 14 | 18 | 22 | 26 | 30 | 3 | 7 | 11 | 15 |
0 | M0: 0 | M0: 1 | M0: 0 | M0: 0 | M0: 0 | M0: 0 | M0: 0 | M0: 0 | M0: 0 | M0: 0 | M0: 1 | M0: 1 | M0: 1 | M0: 1 |
1 | M1: 0 | M1: 0 | M1: 1 | M1: 1 | M1: 1 | M1: 1 | M1: 1 | M1: 1 | M1: 1 | M1: 1 | M1: 1 | M1: 1 | M1: 1 | M1: 1 |
2 | G24 | G25 | M2: 0 | M2: 1 | M2: 0 | M2: 1 | M2: 0 | M2: 1 | M2: 0 | M2: 1 | M2: 0 | M2: 1 | M2: 0 | M2: 1 |
3 | B24 | G34 | M3: 0 | M3: 0 | M3: 1 | M3: 1 | M3: 0 | M3: 0 | M3: 1 | M3: 1 | M3: 0 | M3: 0 | M3: 1 | M3: 1 |
4 | B34 | G35 | M4: 0 | M4: 0 | M4: 0 | M4: 0 | M4: 1 | M4: 1 | M4: 1 | M4: 1 | M4: 0 | M4: 0 | M4: 0 | M4: 0 |
5 | R00 | R00 | R00 | R00 | R00 | R00 | R00 | R00 | R00 | R00 | R00 | R00 | R00 | R00 |
6 | R01 | R01 | R01 | R01 | R01 | R01 | R01 | R01 | R01 | R01 | R01 | R01 | R01 | R01 |
7 | R02 | R02 | R02 | R02 | R02 | R02 | R02 | R02 | R02 | R02 | R02 | R02 | R02 | R02 |
8 | R03 | R03 | R03 | R03 | R03 | R03 | R03 | R03 | R03 | R03 | R03 | R03 | R03 | R03 |
9 | R04 | R04 | R04 | R04 | R04 | R04 | R04 | R04 | R04 | R04 | R04 | R04 | R04 | R04 |
10 | R05 | R05 | R05 | R05 | R05 | R05 | R05 | R05 | R05 | R05 | R05 | R05 | R05 | R05 |
11 | R06 | R06 | R06 | R06 | R06 | R06 | R06 | R06 | R06 | G34 | R06 | R06 | R06 | R06 |
12 | R07 | B30 | R07 | R07 | R07 | R07 | R07 | R07 | R07 | B30 | R07 | R07 | R07 | R07 |
13 | R08 | B31 | R08 | R08 | R08 | R08 | G34 | B30 | B31 | B31 | R08 | R08 | R08 | R08 |
14 | R09 | B24 | R09 | R09 | R09 | B24 | B24 | B24 | B24 | B24 | R09 | R09 | R09 | R09 |
15 | G00 | G00 | G00 | G00 | G00 | G00 | G00 | G00 | G00 | G00 | G00 | G00 | G00 | G00 |
16 | G01 | G01 | G01 | G01 | G01 | G01 | G01 | G01 | G01 | G01 | G01 | G01 | G01 | G01 |
17 | G02 | G02 | G02 | G02 | G02 | G02 | G02 | G02 | G02 | G02 | G02 | G02 | G02 | G02 |
18 | G03 | G03 | G03 | G03 | G03 | G03 | G03 | G03 | G03 | G03 | G03 | G03 | G03 | G03 |
19 | G04 | G04 | G04 | G04 | G04 | G04 | G04 | G04 | G04 | G04 | G04 | G04 | G04 | G04 |
20 | G05 | G05 | G05 | G05 | G05 | G05 | G05 | G05 | G05 | G05 | G05 | G05 | G05 | G05 |
21 | G06 | G06 | G06 | G06 | G06 | G06 | G06 | G06 | G06 | G25 | G06 | G06 | G06 | G06 |
22 | G07 | B25 | G07 | G07 | G07 | G07 | G07 | G07 | G07 | B25 | G07 | G07 | G07 | G07 |
23 | G08 | B32 | G08 | G08 | G08 | G08 | B32 | G25 | B25 | B32 | G08 | G08 | G08 | G08 |
24 | G09 | G24 | G09 | G09 | G09 | G24 | G24 | G24 | G24 | G24 | G09 | G09 | G09 | G09 |
25 | B00 | B00 | B00 | B00 | B00 | B00 | B00 | B00 | B00 | B00 | B00 | B00 | B00 | B00 |
26 | B01 | B01 | B01 | B01 | B01 | B01 | B01 | B01 | B01 | B01 | B01 | B01 | B01 | B01 |
27 | B02 | B02 | B02 | B02 | B02 | B02 | B02 | B02 | B02 | B02 | B02 | B02 | B02 | B02 |
28 | B03 | B03 | B03 | B03 | B03 | B03 | B03 | B03 | B03 | B03 | B03 | B03 | B03 | B03 |
29 | B04 | B04 | B04 | B04 | B04 | B04 | B04 | B04 | B04 | B04 | B04 | B04 | B04 | B04 |
30 | B05 | B05 | B05 | B05 | B05 | B05 | B05 | B05 | B05 | B05 | B05 | B05 | B05 | B05 |
31 | B06 | B06 | B06 | B06 | B06 | B06 | B06 | B06 | B06 | G35 | B06 | B06 | B06 | B06 |
32 | B07 | B33 | B07 | B07 | B07 | B07 | B07 | B07 | B07 | B33 | B07 | B07 | B07 | B07 |
33 | B08 | B35 | B08 | B08 | B08 | B08 | B33 | G35 | B35 | B35 | B08 | B08 | B08 | B08 |
34 | B09 | B34 | B09 | B09 | B09 | B34 | B34 | B34 | B34 | B34 | B09 | B09 | B09 | B09 |
35 | R10 | R10 | R10 | R10 | R10 | R10 | R10 | R10 | R10 | R10 | R10 | R10 | R10 | R10 |
36 | R11 | R11 | R11 | R11 | R11 | R11 | R11 | R11 | R11 | R11 | R11 | R11 | R11 | R11 |
37 | R12 | R12 | R12 | R12 | R12 | R12 | R12 | R12 | R12 | R12 | R12 | R12 | R12 | R12 |
38 | R13 | R13 | R13 | R13 | R13 | R13 | R13 | R13 | R13 | R13 | R13 | R13 | R13 | R13 |
39 | R14 | R14 | R14 | R010 | R010 | R14 | R14 | R14 | R14 | R14 | R14 | R14 | R14 | R015 |
40 | G34 | R15 | R010 | G34 | B24 | G34 | R15 | G34 | G34 | R15 | R15 | R15 | R15 | R014 |
Table 124. Interpretation of upper bits for BC6H block modes
Mode | ||||||||||||||
Bit | 0 | 1 | 2 | 6 | 10 | 14 | 18 | 22 | 26 | 30 | 3 | 7 | 11 | 15 |
41 | G20 | G20 | G20 | G20 | G20 | G20 | G20 | G20 | G20 | G20 | R16 | R16 | R16 | R013 |
42 | G21 | G21 | G21 | G21 | G21 | G21 | G21 | G21 | G21 | G21 | R17 | R17 | R17 | R012 |
43 | G22 | G22 | G22 | G22 | G22 | G22 | G22 | G22 | G22 | G22 | R18 | R18 | R011 | R011 |
44 | G23 | G23 | G23 | G23 | G23 | G23 | G23 | G23 | G23 | G23 | R19 | R010 | R010 | R010 |
45 | G10 | G10 | G10 | G10 | G10 | G10 | G10 | G10 | G10 | G10 | G10 | G10 | G10 | G10 |
46 | G11 | G11 | G11 | G11 | G11 | G11 | G11 | G11 | G11 | G11 | G11 | G11 | G11 | G11 |
47 | G12 | G12 | G12 | G12 | G12 | G12 | G12 | G12 | G12 | G12 | G12 | G12 | G12 | G12 |
48 | G13 | G13 | G13 | G13 | G13 | G13 | G13 | G13 | G13 | G13 | G13 | G13 | G13 | G13 |
49 | G14 | G14 | G010 | G14 | G010 | G14 | G14 | G14 | G14 | G14 | G14 | G14 | G14 | G015 |
50 | B30 | G15 | B30 | G010 | B30 | B30 | B30 | G15 | B30 | G15 | G15 | G15 | G15 | G014 |
51 | G30 | G30 | G30 | G30 | G30 | G30 | G30 | G30 | G30 | G30 | G16 | G16 | G16 | G013 |
52 | G31 | G31 | G31 | G31 | G31 | G31 | G31 | G31 | G31 | G31 | G17 | G17 | G17 | G012 |
53 | G32 | G32 | G32 | G32 | G32 | G32 | G32 | G32 | G32 | G32 | G18 | G18 | G011 | G011 |
54 | G33 | G33 | G33 | G33 | G33 | G33 | G33 | G33 | G33 | G33 | G19 | G010 | G010 | G010 |
55 | B10 | B10 | B10 | B10 | B10 | B10 | B10 | B10 | B10 | B10 | B10 | B10 | B10 | B10 |
56 | B11 | B11 | B11 | B11 | B11 | B11 | B11 | B11 | B11 | B11 | B11 | B11 | B11 | B11 |
57 | B12 | B12 | B12 | B12 | B12 | B12 | B12 | B12 | B12 | B12 | B12 | B12 | B12 | B12 |
58 | B13 | B13 | B13 | B13 | B13 | B13 | B13 | B13 | B13 | B13 | B13 | B13 | B13 | B13 |
59 | B14 | B14 | B010 | B010 | B14 | B14 | B14 | B14 | B14 | B14 | B14 | B14 | B14 | B015 |
60 | B31 | B15 | B31 | B31 | B010 | B31 | B31 | B31 | B15 | B15 | B15 | B15 | B15 | B014 |
61 | B20 | B20 | B20 | B20 | B20 | B20 | B20 | B20 | B20 | B20 | B16 | B16 | B16 | B013 |
62 | B21 | B21 | B21 | B21 | B21 | B21 | B21 | B21 | B21 | B21 | B17 | B17 | B17 | B012 |
63 | B22 | B22 | B22 | B22 | B22 | B22 | B22 | B22 | B22 | B22 | B18 | B18 | B011 | B011 |
64 | B23 | B23 | B23 | B23 | B23 | B23 | B23 | B23 | B23 | B23 | B19 | B010 | B010 | B010 |
65 | R20 | R20 | R20 | R20 | R20 | R20 | R20 | R20 | R20 | R20 | IB0,00 | IB0,00 | IB0,00 | IB0,00 |
66 | R21 | R21 | R21 | R21 | R21 | R21 | R21 | R21 | R21 | R21 | IB0,01 | IB0,01 | IB0,01 | IB0,01 |
67 | R22 | R22 | R22 | R22 | R22 | R22 | R22 | R22 | R22 | R22 | IB0,02 | IB0,02 | IB0,02 | IB0,02 |
68 | R23 | R23 | R23 | R23 | R23 | R23 | R23 | R23 | R23 | R23 | IB1,00 | IB1,00 | IB1,00 | IB1,00 |
69 | R24 | R24 | R24 | B30 | B31 | R24 | R24 | R24 | R24 | R24 | IB1,01 | IB1,01 | IB1,01 | IB1,01 |
70 | B32 | R25 | B32 | B32 | B32 | B32 | R25 | B32 | B32 | R25 | IB1,02 | IB1,02 | IB1,02 | IB1,02 |
71 | R30 | R30 | R30 | R30 | R30 | R30 | R30 | R30 | R30 | R30 | IB1,03 | IB1,03 | IB1,03 | IB1,03 |
72 | R31 | R31 | R31 | R31 | R31 | R31 | R31 | R31 | R31 | R31 | IB2,00 | IB2,00 | IB2,00 | IB2,00 |
73 | R32 | R32 | R32 | R32 | R32 | R32 | R32 | R32 | R32 | R32 | IB2,01 | IB2,01 | IB2,01 | IB2,01 |
74 | R33 | R33 | R33 | R33 | R33 | R33 | R33 | R33 | R33 | R33 | IB2,02 | IB2,02 | IB2,02 | IB2,02 |
75 | R34 | R34 | R34 | G24 | B34 | R34 | R34 | R34 | R34 | R34 | IB2,03 | IB2,03 | IB2,03 | IB2,03 |
76 | B33 | R35 | B33 | B33 | B33 | B33 | R35 | B33 | B33 | R35 | IB3,00 | IB3,00 | IB3,00 | IB3,00 |
77 | PB0 | PB0 | PB0 | PB0 | PB0 | PB0 | PB0 | PB0 | PB0 | PB0 | IB3,01 | IB3,01 | IB3,01 | IB3,01 |
78 | PB1 | PB1 | PB1 | PB1 | PB1 | PB1 | PB1 | PB1 | PB1 | PB1 | IB3,02 | IB3,02 | IB3,02 | IB3,02 |
79 | PB2 | PB2 | PB2 | PB2 | PB2 | PB2 | PB2 | PB2 | PB2 | PB2 | IB3,03 | IB3,03 | IB3,03 | IB3,03 |
80 | PB3 | PB3 | PB3 | PB3 | PB3 | PB3 | PB3 | PB3 | PB3 | PB3 | IB0,10 | IB0,10 | IB0,10 | IB0,10 |
81 | PB4 | PB4 | PB4 | PB4 | PB4 | PB4 | PB4 | PB4 | PB4 | PB4 | IB0,11 | IB0,11 | IB0,11 | IB0,11 |
Table 123 and Table 124 show bits 0..81 for each mode. Since modes 3, 7, 11 and 15 each have only one partition, only the first index is an anchor index, and there is a fixed mapping between texels and index bits. These modes also have four index bits IBx,y0..3 per texel (except for the anchor index), and these pixel indices start at bit 65 with IB0,00. The interpretation of bits 82 and later is not tabulated. For modes with two partitions, the mapping from index bits IBx,y to coordinates depends on the choice of anchor index for the secondary partition (determined by the pattern selected by the partition bits PB4..0), and is therefore not uniquely defined by the mode — and not useful to tabulate in this form. |
In a single-subset blocks, the two endpoints are contained in R0, G0, B0 (collectively referred to as E0) and R1, G1, B1 (collectively E1). In a two-subset block, the endpoints for the second subset are in R2, G2, B2 and R3, G3, B3 (collectively E2 and E3 respectively). The values in E0 are sign-extended to the implementation’s internal integer representation if the format of the texture is signed. The values in E1 (and E2 and E3 if the block has two subsets) are sign-extended if the format of the texture is signed or if the block mode has transformed endpoints. If the mode has transformed endpoints, the values from E0 are used as a base to offset all other endpoints, wrapped at the number of endpoint bits. For example, R1 = (R0 + R1) & $((1 \ll \mathrm{EPB})-1)$ .
In BC7, all modes represent endpoint values independently. This means it is always possible to represent the endpoints nearest to the anchor indices by choosing the endpoint order appropriately. Since in BC6H transformed endpoints are represented as two’s complement offsets relative to the first endpoint, there is an asymmetry: it is possible to represent larger negative values in two’s complement than positive values, so E1, E2 and E3 can be more distant from E0 in a negative direction than positive in modes with transformed endpoints. This means that endpoints cannot necessarily be chosen independently of the anchor index in BC6H, since the order of endpoints cannot necessarily be reversed. In addition, E2 and E3 always depends on E0, so swapping E0 and E1 to suit the anchor bit for the first subset may make make the relative offsets of E2 and E3 unrepresentable in a given mode if they fall out of range. |
Next, the endpoints are unquantized to maximize the usage of the bits and to ensure that the negative ranges are oriented properly to interpolate as a two’s complement value. The following pseudocode assumes the computation uses sufficiently large intermediate values to avoid overflow. For the unsigned float format, we unquantize a value x to unq by:
if (EPB >= 15) unq = x; else if (x == 0) unq = 0; else if (x == ((1 << EPB)-1)) unq = 0xFFFF; else unq = ((x << 15) + 0x4000) >> (EPB-1);
The signed float unquantization is similar, but needs to worry about orienting the negative range:
s = 0; if (EPB >= 16) { unq = x; } else { if (x < 0) { s = 1; x = -x; } if (x == 0) unq = 0; else if (x >= ((1 << (EPB-1))-1)) unq = 0x7FFF; else unq = ((x << 15) + 0x4000) >> (EPB-1); if (s) unq = -unq; }
After the endpoints are unquantized, interpolation proceeds as in the fixed-point formats above using Equation 2, including the interpolation weight table, Table 119.
The interpolated values are passed through a final unquantization step. For the unsigned format, this limits the range of the integer representation to those bit sequences which, when interpreted as a 16-bit half float, represent [0.0..65504.0], where 65504.0 is the largest finite value representable in a half float. The bit pattern that represents 65504.0 is integer 0x7BFF, so the integer input range 0..0xFFFF can be mapped to this range by scaling the interpolated integer i by $31\over 64$ :
out = (i * 31) >> 6;
For the signed format, the final unquantization step limits the range of the integer representation to the bit sequences which, when interpreted as a 16-bit half float, represent the range [ $-\infty$ ..65504.0], where $-\infty$ is represented in half float as the bit pattern 0xFC00. The signed 16-bit integer range [-0x8000..0x7FFF] is remapped to this float representation by taking the absolute value of the interpolated value i, scaling it by $31\over 32$ , and restoring the sign bit:
out = i < 0 ? (((-i) * 31) >> 5) | 0x8000 : (i * 31) >> 5;
The resultant bit pattern should be interpreted as a 16-bit half float.
The ability to support $-\infty$ is considered “accidental” due to the asymmetry of two’s complement representation: in order to map integer 0x7FFF to 65504.0 and 0x0000 to 0.0, -0x7FFF maps to the largest finite negative value, -65504.0, represented as 0xFBFF. A two’s complement signed integer can also represent -0x8000; it happens that the same unquantization formula maps 0x8000 to 0xFC00, which is the half float bit pattern for $-\infty$ . Although decoders for BC6H should be bit-exact, encoders for this format are encouraged to map $-\infty$ to -65504.0 (and to map $\infty$ to 65504.0 and NaN values to 0.0) prior to encoding. |
This description is derived from the OES_compressed_ETC1_RGB8_texture OpenGL extension.
The texture is described as a number of 4×4 pixel blocks. If the texture (or a particular mip-level) is smaller than 4 pixels in any dimension (such as a 2×2 or a 8×1 texture), the texture is found in the upper left part of the block(s), and the rest of the pixels are not used. For instance, a texture of size 4×2 will be placed in the upper half of a 4×4 block, and the lower half of the pixels in the block will not be accessed.
Pixel a1 (see Figure 38) of the first block in memory will represent the texture coordinate (u=0, v=0). Pixel a2 in the second block in memory will be adjacent to pixel m1 in the first block, etc. until the width of the texture. Then pixel a3 in the following block (third block in memory for a 8×8 texture) will be adjacent to pixel d1 in the first block, etc. until the height of the texture. The data storage for an 8×8 texture using the first, second, third and fourth block if stored in that order in memory would have the texels encoded in the same order as a simple linear format as if the bytes describing the pixels came in the following memory order: a1 e1 i1 m1 a2 e2 i2 m2 b1 f1 i1 n1 b2 f2 i2 n2 c1 g1 k1 o1 c2 g2 k2 o2 d1 h1 l1 p1 d2 h2 l2 p2 a3 e3 i3 m3 a4 e4 i4 m4 b3 f3 i3 n3 b4 f4 i4 n4 c3 g3 k3 o3 c4 g4 k4 o4 d3 h3 l3 p3 d4 h4 l4 p4.
Note how pixel a2 in the second block is adjacent to pixel m1 in the first block.
The number of bits that represent a 4×4 texel block is 64 bits.
The data for a block is stored as a number of bytes, q0, q1, q2, q3, q4, q5, q6, q7, where byte q0 is located at the lowest memory address and q7 at the highest. The 64 bits specifying the block are then represented by the following 64 bit integer:
Each 64-bit word contains information about a 4×4 pixel block as shown in Figure 39.
There are two modes in ETC1: the ‘individual’ mode and the ‘differential’ mode. Which mode is active for a particular 4×4 block is controlled by bit 33, which we call diff bit. If diff bit = 0, the ‘individual’ mode is chosen, and if diff bit = 1, then the ‘differential’ mode is chosen. The bit layout for the two modes are different: The bit layout for the individual mode is shown in Table 125 part a and part c, and the bit layout for the differential mode is laid out in Table 125 part b and part c.
Table 125. Texel Data format for ETC1 compressed textures
63 | 62 | 61 | 60 | 59 | 58 | 57 | 56 | 55 | 54 | 53 | 52 | 51 | 50 | 49 | 48 |
Base color 1 R (4 bits) | Base color 2 R2 (4 bits) | Base color 1 G (4 bits) | Base color 2 G2 (4 bits) | ||||||||||||
47 | 46 | 45 | 44 | 43 | 42 | 41 | 40 | 39 | 38 | 37 | 36 | 35 | 34 | 33 | 32 |
Base color 1 B (4 bits) | Base color 2 B2 (4 bits) | Table codeword 1 | Table codeword 2 | diff bit | flip bit | ||||||||||
63 | 62 | 61 | 60 | 59 | 58 | 57 | 56 | 55 | 54 | 53 | 52 | 51 | 50 | 49 | 48 |
Base color R (5 bits) | Color delta Rd | Base color G (5 bits) | Color delta Gd | ||||||||||||
47 | 46 | 45 | 44 | 43 | 42 | 41 | 40 | 39 | 38 | 37 | 36 | 35 | 34 | 33 | 32 |
Base color B (5 bits) | Color delta Bd | Table codeword 1 | Table codeword 2 | diff bit | flip bit | ||||||||||
More significant pixel index bits | |||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 |
p1 | o1 | n1 | m1 | l1 | k1 | j1 | i1 | h1 | g1 | f1 | e1 | d1 | c1 | b1 | a1 |
Less significant pixel index bits | |||||||||||||||
15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
p0 | o0 | n0 | m0 | l0 | k0 | j0 | i0 | h0 | g0 | f0 | e0 | d0 | c0 | b0 | a0 |
In both modes, the 4×4 block is divided into two subblocks of either size 2×4 or 4×2. This is controlled by bit 32, which we call flip bit. If flip bit = 0, the block is divided into two 2×4 subblocks side-by-side, as shown in Figure 40. If flip bit = 1, the block is divided into two 4×2 subblocks on top of each other, as shown in Figure 41.
In both individual and differential mode, a base color for each subblock is stored, but the way they are stored is different in the two modes:
In the ‘individual’ mode (diff bit = 0), the base color for subblock 1 is derived from the codewords R (bits 63..60), G (bits 55..52) and B (bits 47..44), see section a of Table 125. These four bit values are extended to RGB:888 by replicating the four higher order bits in the four lower order bits. For instance, if R = 14 = 1110b, G = 3 = 0011b and B = 8 = 1000b, then the red component of the base color of subblock 1 becomes 11101110b = 238, and the green and blue components become 00110011b = 51 and 10001000b = 136. The base color for subblock 2 is decoded the same way, but using the 4-bit codewords R2 (bits 59..56), G2 (bits 51..48) and B2 (bits 43..40) instead. In summary, the base colors for the subblocks in the individual mode are:
In the ‘differential’ mode (diff bit = 1), the base color for subblock 1 is derived from the five-bit codewords R, G and B. These five-bit codewords are extended to eight bits by replicating the top three highest-order bits to the three lowest order bits. For instance, if R = 28 = 11100b, the resulting eight-bit red color component becomes 11100111b = 231. Likewise, if G = 4 = 00100b and B = 3 = 00011b, the green and blue components become 00100001b = 33 and 00011000b = 24 respectively. Thus, in this example, the base color for subblock 1 is (231, 33, 24). The five-bit representation for the base color of subblock 2 is obtained by modifying the five-bit codewords R, G and B by the codewords Rd, Gd and Bd. Each of Rd, Gd and Bd is a three-bit two’s-complement number that can hold values between -4 and +3. For instance, if R= 28 as above, an Rd = 100b = -4, then the five-bit representation for the red color component is 28+(-4) = 24 = 11000b, which is then extended to eight bits, to 11000110b = 198. Likewise, if G = 4, Gd = 2, B = 3 and Bd = 0, the base color of subblock 2 will be RGB = (198, 49, 24). In summary, the base colors for the subblocks in the differential mode are:
Note that these additions are not allowed to under- or overflow (go below zero or above 31). (The compression scheme can easily make sure they don’t.) For over- or underflowing values, the behavior is undefined for all pixels in the 4×4 block. Note also that the extension to eight bits is performed after the addition.
After obtaining the base color, the operations are the same for the two modes ‘individual’ and ‘differential’. First a table is chosen using the table codewords: For subblock 1, table codeword 1 is used (bits 39..37), and for subblock 2, table codeword 2 is used (bits 36..34), see Table 125. The table codeword is used to select one of eight modifier tables, see Table 126. For instance, if the table code word is 010b = 2, then the modifier table [-29, -9, 9, 29] is selected. Note that the values in Table 126 are valid for all textures and can therefore be hardcoded into the decompression unit.
Next, we identify which modifier value to use from the modifier table using the two ‘pixel index’ bits. The pixel index bits are unique for each pixel. For instance, the pixel index for pixel d (see Figure 39) can be found in bits 19 (most significant bit, MSB), and 3 (least significant bit, LSB), see section c of Table 125. Note that the pixel index for a particular texel is always stored in the same bit position, irrespectively of bits diff bit and flip bit. The pixel index bits are decoded using Table 127. If, for instance, the pixel index bits are 01b = 1, and the modifier table [-29, -9, 9, 29] is used, then the modifier value selected for that pixel is 29 (see Table 127). This modifier value is now used to additively modify the base color. For example, if we have the base color (231, 8, 16), we should add the modifier value 29 to all three components: (231+29, 8+29, 16+29) resulting in (260, 37, 45). These values are then clamped to [0..255], resulting in the color (255, 37, 45), and we are finished decoding the texel.
Table 126. Intensity modifier sets for ETC1 compressed textures
Table codeword | Modifier table | |||
0 | -8 | -2 | 2 | 8 |
1 | -17 | -5 | 5 | 17 |
2 | -29 | -9 | 9 | 29 |
3 | -42 | -13 | 13 | 42 |
4 | -60 | -18 | 18 | 60 |
5 | -80 | -24 | 24 | 80 |
6 | -106 | -33 | 33 | 106 |
7 | -183 | -47 | 47 | 183 |
Table 127. Mapping from pixel index values to modifier values for ETC1 compressed textures
Pixel index value | Resulting modifier value | |
MSB | LSB | |
1 | 1 | -b (large negative value) |
1 | 0 | -a (small negative value) |
0 | 0 | +a (small positive value) |
0 | 1 | +b (large positive value) |
ETC1 is a proper subset of ETC2. There are examples of “individual” and “differential” decoding in Section 22. |
The ETC1S format is a subset the ETC1, simplified to facilitate image compression. The blocks use differential encoding (diff bit = 1); Color deltas Rd = Gd = Bd = 0, so the two subblocks share base colors. The Table codeword for each subblock is identical. Finally, the flip bit is encoded as 0 — the subsets are identical anyway.
This description is derived from the “ETC Compressed Texture Image Formats” section of the OpenGL 4.4 specification.
The ETC formats form a family of related compressed texture image formats. They are designed to do different tasks, but also to be similar enough that hardware can be reused between them. Each one is described in detail below, but we will first give an overview of each format and describe how it is similar to others and the main differences.
RGB ETC2 is a format for compressing RGB data. It is a superset of the older ETC1 format. This means that an older ETC1 texture can be decoded using an ETC2-compliant decoder. The main difference is that the newer version contains three new modes; the ‘T-mode’ and the ‘H-mode’ which are good for sharp chrominance blocks and the ‘Planar’ mode which is good for smooth blocks.
RGB ETC2 with sRGB encoding is the same as linear RGB ETC2 with the difference that the values should be interpreted as being encoded with the sRGB transfer function instead of linear RGB-values.
RGBA ETC2 encodes RGBA 8-bit data. The RGB part is encoded exactly the same way as RGB ETC2. The alpha part is encoded separately.
RGBA ETC2 with sRGB encoding is the same as RGBA ETC2 but here the RGB values (but not the alpha value) should be interpreted as being encoded with the sRGB transfer function.
Unsigned R11 EAC is a one-channel unsigned format. It is similar to the alpha part of RGBA ETC2 but not exactly the same; it delivers higher precision. It is possible to make hardware that can decode both formats with minimal overhead.
Unsigned RG11 EAC is a two-channel unsigned format. Each channel is decoded exactly as Unsigned R11 EAC.
Signed R11 EAC is a one-channel signed format. This is good in situations when it is important to be able to preserve zero exactly, and still use both positive and negative values. It is designed to be similar enough to Unsigned R11 EAC so that hardware can decode both with minimal overhead, but it is not exactly the same. For example; the signed version does not add 0.5 to the base codeword, and the extension from 11 bits differ. For all details, see the corresponding sections.
Signed RG11 EAC is a two-channel signed format. Each channel is decoded exactly as signed R11 EAC.
RGB ETC2 with “punchthrough” alpha is very similar to RGB ETC2, but has the ability to represent “punchthrough” alpha (completely opaque or transparent). Each block can select to be completely opaque using one bit. To fit this bit, there is no individual mode in RGB ETC2 with punchthrough alpha. In other respects, the opaque blocks are decoded as in RGB ETC2. For the transparent blocks, one index is reserved to represent transparency, and the decoding of the RGB channels are also affected. For details, see the corresponding sections.
RGB ETC2 with punchthrough alpha and sRGB encoding is the same as linear RGB ETC2 with punchthrough alpha but the RGB channel values should be interpreted as being encoded with the sRGB transfer function.
A texture compressed using any of the ETC texture image formats is described as a number of 4×4 pixel blocks.
Pixel a1 (see Figure 42) of the first block in memory will represent the texture coordinate (u=0, v=0). Pixel a2 in the second block in memory will be adjacent to pixel m1 in the first block, etc. until the width of the texture. Then pixel a3 in the following block (third block in memory for an 8×8 texture) will be adjacent to pixel d1 in the first block, etc. until the height of the texture.
The data storage for an 8×8 texture using the first, second, third and fourth block if stored in that order in memory would have the texels encoded in the same order as a simple linear format as if the bytes describing the pixels came in the following memory order: a1 e1 i1 m1 a2 e2 i2 m2 b1 f1 i1 n1 b2 f2 i2 n2 c1 g1 k1 o1 c2 g2 k2 o2 d1 h1 l1 p1 d2 h2 l2 p2 a3 e3 i3 m3 a4 e4 i4 m4 b3 f3 i3 n3 b4 f4 i4 n4 c3 g3 k3 o3 c4 g4 k4 o4 d3 h3 l3 p3 d4 h4 l4 p4.
Note how pixel a3 in the third block is adjacent to pixel d1 in the first block.
If the width or height of the texture (or a particular mip-level) is not a multiple of four, then padding is added to ensure that the texture contains a whole number of 4×4 blocks in each dimension. The padding does not affect the texel coordinates. For example, the texel shown as a1 in Figure 42 always has coordinates (i=0, j=0). The values of padding texels are irrelevant, e.g., in a 3×3 texture, the texels marked as m1, n1, o1, d1, h1, l1 and p1 form padding and have no effect on the final texture image.
The number of bits that represent a 4×4 texel block is 64 bits if the format is RGB ETC2, RGB ETC2 with sRGB encoding, RGBA ETC2 with punchthrough alpha, or RGB ETC2 with punchthrough alpha and sRGB encoding.
In those cases the data for a block is stored as a number of bytes, {q0, q1, q2, q3, q4, q5, q6, q7}, where byte q0 is located at the lowest memory address and q7 at the highest. The 64 bits specifying the block are then represented by the following 64 bit integer:
The number of bits that represent a 4×4 texel block is 128 bits if the format is RGBA ETC2 with a linear or sRGB transfer function. In those cases the data for a block is stored as a number of bytes: {q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, q10, q11, q12, q13, q14, q15}, where byte q0 is located at the lowest memory address and q15 at the highest.
This is split into two 64-bit integers, one used for color channel decompression and one for alpha channel decompression:
For RGB ETC2, each 64-bit word contains information about a three-channel 4×4 pixel block as shown in Figure 43.
Table 128. Texel Data format for ETC2 compressed texture formats
The blocks are compressed using one of five different ‘modes’. Section a of Table 128 shows the bits used for determining the mode used in a given block. First, if the ‘differential bit’ marked D is set to 0, the ‘individual’ mode is used. Otherwise, the three 5-bit values R, G and B, and the three 3-bit values Rd, Gd and Bd are examined. R, G and B are treated as integers between 0 and 31 and Rd, Gd and Bd as two’s-complement integers between -4 and +3. First, R and Rd are added, and if the sum is not within the interval [0..31], the ‘T’ mode is selected. Otherwise, if the sum of G and Gd is outside the interval [0..31], the ‘H’ mode is selected. Otherwise, if the sum of B and Bd is outside of the interval [0..31], the ‘planar’ mode is selected. Finally, if the D bit is set to 1 and all of the aforementioned sums lie between 0 and 31, the ‘differential’ mode is selected.
The layout of the bits used to decode the ‘individual’ and ‘differential’ modes are shown in section b and section c of Table 128, respectively. Both of these modes share several characteristics. In both modes, the 4×4 block is split into two subblocks of either size 2×4 or 4×2. This is controlled by bit 32, which we dub the flip bit (FB in Table 128 (b) and (c)). If the flip bit is 0, the block is divided into two 2×4 subblocks side-by-side, as shown in Figure 44. If the flip bit is 1, the block is divided into two 4×2 subblocks on top of each other, as shown in Figure 45. In both modes, a base color for each subblock is stored, but the way they are stored is different in the two modes:
In the ‘individual’ mode, following the layout shown in section b of Table 128, the base color for subblock 1 is derived from the codewords R (bits 63..60), G (bits 55..52) and B (bits 47..44). These four bit values are extended to RGB:888 by replicating the four higher order bits in the four lower order bits. For instance, if R = 14 = 1110 binary (1110b for short), G = 3 = 0011b and B = 8 = 1000b, then the red component of the base color of subblock 1 becomes 11101110b = 238, and the green and blue components become 00110011b = 51 and 10001000b = 136. The base color for subblock 2 is decoded the same way, but using the 4-bit codewords R2 (bits 59..56), G2 (bits 51..48) and B2 (bits 43..40) instead. In summary, the base colors for the subblocks in the individual mode are:
In the ‘differential’ mode, following the layout shown in section c of Table 128, the base color for subblock 1 is derived from the five-bit codewords R, G and B. These five-bit codewords are extended to eight bits by replicating the top three highest-order bits to the three lowest-order bits. For instance, if R = 28 = 11100b, the resulting eight-bit red color component becomes 11100111b = 231. Likewise, if G = 4 = 00100b and B = 3 = 00011b, the green and blue components become 00100001b = 33 and 00011000b = 24 respectively. Thus, in this example, the base color for subblock 1 is (231, 33, 24). The five-bit representation for the base color of subblock 2 is obtained by modifying the five-bit codewords R, G and B by the codewords Rd, Gd and Bd. Each of Rd, Gd and Bd is a 3-bit two’s-complement number that can hold values between -4 and +3. For instance, if R = 28 as above, and Rd = 100b = y - 4, then the five bit representation for the red color component is 28+(-4) = 24 = 11000b, which is then extended to eight bits to 11000110b = 198. Likewise, if G = 4, Gd = 2, B = 3 and Bd = 0, the base color of subblock 2 will be RGB = 198, 49, 24. In summary, the base colors for the subblocks in the ‘differential’ mode are:
Note that these additions will not under- or overflow, or one of the alternative decompression modes would have been chosen instead of the ‘differential’ mode.
After obtaining the base color, the operations are the same for the two modes ‘individual’ and ‘differential’. First a table is chosen using the table codewords: For subblock 1, table codeword 1 is used (bits 39..37), and for subblock 2, table codeword 2 is used (bits 36..34), see section b or section c of Table 128. The table codeword is used to select one of eight modifier tables, see Table 129. For instance, if the table codeword is 010 binary = 2, then the modifier table [-29, -9, 9, 29] is selected for the corresponding sub-block. Note that the values in Table 129 are valid for all textures and can therefore be hardcoded into the decompression unit.
Table 129. ETC2 intensity modifier sets for ‘individual’ and ‘differential’ modes
Table codeword | Modifier table | |||
0 | -8 | -2 | 2 | 8 |
1 | -17 | -5 | 5 | 17 |
2 | -29 | -9 | 9 | 29 |
3 | -42 | -13 | 13 | 42 |
4 | -60 | -18 | 18 | 60 |
5 | -80 | -24 | 24 | 80 |
6 | -106 | -33 | 33 | 106 |
7 | -183 | -47 | 47 | 183 |
Table 130. Mapping from pixel index values to modifier values for RGB ETC2 compressed textures
Pixel index value | Resulting modifier value | |
MSB | LSB | |
1 | 1 | -b (large negative value) |
1 | 0 | -a (small negative value) |
0 | 0 | +a (small positive value) |
0 | 1 | +b (large positive value) |
Next, we identify which modifier value to use from the modifier table using the two pixel index bits. The pixel index bits are unique for each pixel. For instance, the pixel index for pixel d (see Figure 43) can be found in bits 19 (most significant bit, MSB), and 3 (least significant bit, LSB), see section f of Table 128. Note that the pixel index for a particular texel is always stored in the same bit position, irrespectively of bits diff bit and flip bit. The pixel index bits are decoded using Table 130. If, for instance, the pixel index bits are 01 binary = 1, and the modifier table [-29, -9, 9, 29] is used, then the modifier value selected for that pixel is 29 (see Table 130). This modifier value is now used to additively modify the base color. For example, if we have the base color (231, 8, 16), we should add the modifier value 29 to all three components: (231+29, 8+29, 16+29) resulting in (260, 37, 45). These values are then clamped to [0..255], resulting in the color (255, 37, 45), and we are finished decoding the texel.
Figure 46 shows an example ‘individual mode’ ETC2 block. The two base colors are shown as circles, and modifiers are applied to each channel to give the ‘paint colors’ selectable by each pixel index, shown as small diamonds. Since the same modifier is applied to each channel, each paint color for a subblock falls on a line (shown dashed) parallel to the grayscale (0, 0, 0) to (255, 255, 255) axis, unless the channels are modified by clamping to the range [0..255]. In this example, one base color is encoded as the 4-bit triple (4, 11, 9), which is expanded by extend4to8bits to (68, 187, 153). Modifier table 4 [-60, -18, 18, 60] is selected for this subblock, giving the following paint colors:
The other base color is encoded as the 4-bit triple (14, 3, 8), which is expanded by extend4to8bits to (238, 51, 136). Modifier table 0 [-8, -2, 2, 8] is selected for this subblock, giving the following paint colors for the subblock:
In this example, none of the paint colors are modified by the process of clipping the channels to the range [0..255]. Since there is no difference in the way the base colors are encoded in ‘individual mode’, either base color could correspond to either subblock. |
Figure 47 shows an example ‘differential mode’ ETC2 block. The two base colors are shown as circles; an arrow shows the base color of the second subblock (the upper left circle) derived from the first subblock’s base color (lower right circle). Modifiers to the base colors give ‘paint colors’ selectable by each pixel index, shown as small diamonds. Since the same modifier is applied to each channel, each paint color for a subblock falls on a line (shown dashed) parallel to the grayscale (0, 0, 0) to (255, 255, 255) axis, unless channels are modified by clamping to [0..255]. Here the first subblock’s base color is encoded as the 5-bit triple (29, 26, 8), and expanded by extend5to8bits to (239, 214, 66). Note that not every color representable in ‘individual mode’, exists in ‘differential mode’, or vice-versa.
The base color of subblock 2 is the five-bit representation of the base color of subblock 1 (29, 26, 8) plus a (Rd, Gd, Bd) offset of (-4, -3, +3), for a new base color of (25, 23, 11) - expanded by extend5to8bits to (206, 189, 90). The offset cannot exceed the range [0..31] (expanded to [0..255]): this would select the ‘T’, ‘H’ or ‘planar’ modes. For ‘differential mode’, the base colors must be similar in each channel. The two’s complement offset gives an asymmetry: we could not swap the subblocks of this example, since a Rd offset of +4 is unrepresentable.
In this example, modifier table 2 [-29, -9, 9, 29] is applied to subblock 1’s base color of (239, 214, 66):
The last row is clamped to (255, 243, 95), so subblock 1’s paint colors are not colinear in this example. With modifiers, all grays [0..255] are representable. Similarly, modifier table 3 [-42, -13, 13, 42] is applied to the base color of subblock 2, (206, 189, 90):
|
The ‘T’ and ‘H’ compression modes also share some characteristics: both use two base colors stored using 4 bits per channel decoded as in the individual mode. Unlike the ‘individual’ mode however, these bits are not stored sequentially, but in the layout shown in section d and section e of Table 128. To clarify, in the ‘T’ mode, the two colors are constructed as follows:
Here, $\ll$ denotes bit-wise left shift and $|$ denotes bit-wise OR. In the ‘H’ mode, the two colors are constructed as follows:
Both the ‘T’ and ‘H’ modes have four paint colors which are the colors that will be used in the decompressed block, but they are assigned in a different manner. In the ‘T’ mode, paint color 0 is simply the first base color, and paint color 2 is the second base color. To obtain the other paint colors, a ‘distance’ is first determined, which will be used to modify the luminance of one of the base colors. This is done by combining the values da and db shown in section d of Table 128 by (da $\ll$ 1) | db, and then using this value as an index into the small look-up table shown in Table 131. For example, if da is 10 binary and db is 1 binary, the distance index is 101 binary and the selected ‘distance’ d will be 32. Paint color 1 is then equal to the second base color with the ‘distance’ d added to each channel, and paint color 3 is the second base color with the ‘distance’ d subtracted.
Table 131. Distance table for ETC2 ‘T’ and ‘H’ modes
Distance index | Distance d |
---|---|
0 | 3 |
1 | 6 |
2 | 11 |
3 | 16 |
4 | 23 |
5 | 32 |
6 | 41 |
7 | 64 |
In summary, to determine the four paint colors for a ‘T’ block:
In both cases, the value of each channel is clamped to within [0..255].
Figure 48 shows an example ‘T-mode’ ETC2 block. The two base colors are shown as circles, and modifiers are applied to base color 2 to give the other two ‘paint colors’, shown as small diamonds. Since the same modifier is applied to each channel, base color 2 and the two paint colors derived from it fall on a line (shown dashed) parallel to the grayscale (0, 0, 0) to (255, 255, 255) axis, unless channels are modified by clamping to [0..255]. In this example, the first base color is defined as the triple of 4-bit RGB values (13, 1, 8), which is expanded by extend4to8bits to (221, 17, 136). This becomes paint color 0.
The second base color is encoded as the triple of 4-bit RGB values (4, 12, 13), which is expanded by extend4to8bits to (68, 204, 221).
Distance index 5 is used to select a distance value d of 32, which is added to and subtracted from the second base color, giving (100, 236, 253) as paint color 1 and (36, 172, 189) as paint color 3. On this occasion, the channels of these paint colors are not modified by the process of clamping them to [0..255]. |
A ‘distance’ value is computed for the ‘H’ mode as well, but doing so is slightly more complex. In order to construct the three-bit index into the distance table shown in Table 131, da and db shown in section e of Table 128 are used as the most significant bit and middle bit, respectively, but the least significant bit is computed as (base color 1 value $\geq$ base color 2 value), the ‘value’ of a color for the comparison being equal to (R $\ll$ 16) + (G $\ll$ 8) + B. Once the ‘distance’ d has been determined for an ‘H’ block, the four paint colors will be:
Again, all color components are clamped to within [0..255].
Figure 49 shows an example ‘H mode’ ETC2 block. The two base colors are shown as circles, and modifiers are applied to each channel to give the ‘paint colors’ selectable by each pixel index, shown as small diamonds. Since the same modifier is applied to each channel, each paint color falls on a line through the base color from which it is derived (shown dashed) parallel to the grayscale (0, 0, 0) to (255, 255, 255) axis, unless the channels are modified by clamping to the range [0..255]. In this example, the first base color is defined as the triple of 4-bit RGB values (13, 1, 8), as in the ‘T mode’ case above. This is expanded by extend4to8bits to (221, 17, 136).
The second base color is defined as the 4-bit triple (4, 12, 13), which expands to (68, 204, 221).
The block encodes a distance index of 5 (this means that base color 1 must be greater than base color 2), corresponding to a distance d of 32. This leads to the following paint colors:
The G channel of paint color 1 is clamped to 0, giving (189, 0, 104). This stops paint color 1 being colinear with paint color 0 and base color 1. |
Finally, in both the ‘T’ and ‘H’ modes, every pixel is assigned one of the four paint colors in the same way the four modifier values are distributed in ‘individual’ or ‘differential’ blocks. For example, to choose a paint color for pixel d, an index is constructed using bit 19 as most significant bit and bit 3 as least significant bit. Then, if a pixel has index 2, for example, it will be assigned paint color 2.
The final mode possible in an RGB ETC2-compressed block is the ‘planar’ mode. Here, three base colors are supplied and used to form a color plane used to determine the color of the individual pixels in the block.
All three base colors are stored in RGB:676 format, and stored in the manner shown in section g of Table 128. The two secondary colors are given the suffix ‘h’ and ‘v’, so that the red component of the three colors are R, Rh and Rv, for example. Some color channels are split into non-consecutive bit-ranges; for example B is reconstructed using B5 as the most-significant bit, B4..3 as the two following bits, and B2..0 as the three least-significant bits.
Once the bits for the base colors have been extracted, they must be extended to 8 bits per channel in a manner analogous to the method used for the base colors in other modes. For example, the 6-bit blue and red channels are extended by replicating the two most significant of the six bits to the two least significant of the final 8 bits.
With three base colors in RGB:888 format, the color of each pixel can then be determined as:
where x and y are values from 0 to 3 corresponding to the pixels coordinates within the block, x being in the u direction and y in the v direction. For example, the pixel g in Figure 43 would have x = 1 and y = 2.
These values are then rounded to the nearest integer (to the larger integer if there is a tie) and then clamped to a value between 0 and 255. Note that this is equivalent to
where clamp255(·) clamps the value to a number in the range [0..255] and where $\gg$ performs bit-wise right shift.
This specification gives the output for each compression mode in 8-bit integer colors between 0 and 255, and these values all need to be divided by 255 for the final floating point representation.
Figure 50 shows an example ‘planar mode’ ETC2 block. The three base colors are shown as circles, and the interpolated values are shown as small diamonds. In this example, the origin (R, G, B) is encoded as the 6-7-6-bit value (12, 64, 62), which is expanded to (48, 129, 251). The ‘horizontal’ (interpolated by x) base color (Rh, Gh, Bh) = (50, 5, 37) and ‘vertical’ (interpolated by y) base color (Rv, Gv, Bv) = (40, 112, 45) expand to (203, 10, 150) and (162, 225, 182) respectively.
The resulting texel colors are then:
|
Decompression of floating point sRGB values in RGB ETC2 with sRGB encoding follows that of floating point RGB values of linear RGB ETC2. The result is sRGB-encoded values between 0.0 and 1.0. The further conversion from an sRGB encoded component cs to a linear component cl is done according to the formulae in Section 13.3. Assume cs is the sRGB component in the range [0, 1].
Each 4×4 block of RGBA:8888 information is compressed to 128 bits. To decode a block, the two 64-bit integers int64bitAlpha and int64bitColor are calculated as described in Section 22.1. The RGB component is then decoded the same way as for RGB ETC2 (see Section 22.1), using int64bitColor as the int64bit codeword.
Table 132. Texel Data format for alpha part of RGBA ETC2 compressed textures
a) Bit layout in bits 63 through 48 | |||||||||||||||
63 | 62 | 61 | 60 | 59 | 58 | 57 | 56 | 55 | 54 | 53 | 52 | 51 | 50 | 49 | 48 |
base codeword | multiplier | table index | |||||||||||||
b) Bit layout in bits 47 through 0, with pixels as name in Figure 43, bits labeled from 0 being the LSB to 47 being the MSB | |||||||||||||||
47 | 46 | 45 | 44 | 43 | 42 | 41 | 40 | 39 | 38 | 37 | 36 | 35 | 34 | 33 | 32 |
aα2 | aα1 | aα0 | bα2 | bα1 | bα0 | cα2 | cα1 | cα0 | dα2 | dα1 | dα0 | eα2 | eα1 | eα0 | fα2 |
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 |
fα1 | fα0 | gα2 | gα1 | gα0 | hα2 | hα1 | hα0 | iα2 | iα1 | iα0 | jα2 | jα1 | jα0 | kα2 | kα1 |
15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
kα0 | lα2 | lα1 | lα0 | mα2 | mα1 | mα0 | nα2 | nα1 | nα0 | oα2 | oα1 | oα0 | pα2 | pα1 | pα0 |
The 64-bits in int64bitAlpha used to decompress the alpha channel are laid out as shown in Table 132. The information is split into two parts. The first 16 bits comprise a base codeword, a table codeword and a multiplier, which are used together to compute 8 pixel values to be used in the block. The remaining 48 bits are divided into 16 3-bit indices, which are used to select one of these 8 possible values for each pixel in the block.
The color pixel indices are stored in a..p order in increasing bit order in a big-endian word representation, with the low bit stored separately from the high bit. However, the alpha indices are stored in p..a order in increasing bit order in a big-endian word representation, with each bit of each alpha index stored consecutively. |
The decoded value of a pixel is a value between 0 and 255 and is calculated the following way:
Equation 3. ETC2 base
where clamp255(·) maps values outside the range [0..255] to 0.0 or 255.0.
The base codeword is stored in the first 8 bits (bits 63..56) as shown in Table 132 part (a). This is the first term in Equation 3.
Next, we want to obtain the modifier. Bits 51..48 in Table 132 part (a) form a 4-bit index used to select one of 16 pre-determined ‘modifier tables’, shown in Table 133.
Table 133. Intensity modifier sets for RGBA ETC2 alpha component
Table index | Modifier table | |||||||
---|---|---|---|---|---|---|---|---|
0 | -3 | -6 | -9 | -15 | 2 | 5 | 8 | 14 |
1 | -3 | -7 | -10 | -13 | 2 | 6 | 9 | 12 |
2 | -2 | -5 | -8 | -13 | 1 | 4 | 7 | 12 |
3 | -2 | -4 | -6 | -13 | 1 | 3 | 5 | 12 |
4 | -3 | -6 | -8 | -12 | 2 | 5 | 7 | 11 |
5 | -3 | -7 | -9 | -11 | 2 | 6 | 8 | 10 |
6 | -4 | -7 | -8 | -11 | 3 | 6 | 7 | 10 |
7 | -3 | -5 | -8 | -11 | 2 | 4 | 7 | 10 |
8 | -2 | -6 | -8 | -10 | 1 | 5 | 7 | 9 |
9 | -2 | -5 | -8 | -10 | 1 | 4 | 7 | 9 |
10 | -2 | -4 | -8 | -10 | 1 | 3 | 7 | 9 |
11 | -2 | -5 | -7 | -10 | 1 | 4 | 6 | 9 |
12 | -3 | -4 | -7 | -10 | 2 | 3 | 6 | 9 |
13 | -1 | -2 | -3 | -10 | 0 | 1 | 2 | 9 |
14 | -4 | -6 | -8 | -9 | 3 | 5 | 7 | 8 |
15 | -3 | -5 | -7 | -9 | 2 | 4 | 6 | 8 |
For example, a table index of 13 (1101 binary) means that we should use table [-1, -2 -3, -10, 0, 1, 2, 9]. To select which of these values we should use, we consult the pixel index of the pixel we want to decode. As shown in Table 132 part (b), bits 47..0 are used to store a 3-bit index for each pixel in the block, selecting one of the 8 possible values. Assume we are interested in pixel b. Its pixel index is stored in bits 44..42, with the most significant bit stored in 44 and the least significant bit stored in 42. If the pixel index is 011 binary = 3, this means we should take the value 3 from the left in the table, which is -10. This is now our modifier, which is the starting point of our second term in the addition.
In the next step we obtain the multiplier value; bits 55..52 form a four-bit multiplier between 0 and 15. This value should be multiplied with the modifier. An encoder is not allowed to produce a multiplier of zero, but the decoder should still be able to handle this case (and produce 0 × modifier = 0 in that case).
The modifier times the multiplier now provides the third and final term in the sum in Equation 3. The sum is calculated and the value is clamped to the interval [0..255]. The resulting value is the 8-bit output value.
For example, assume a base codeword of 103, a table index of 13, a pixel index of 3 and a multiplier of 2. We will then start with the base codeword 103 (01100111 binary). Next, a table index of 13 selects table [-1, -2, -3, -10, 0, 1, 2, 9], and using a pixel index of 3 will result in a modifier of -10. The multiplier is 2, forming -10 × 2 = -20. We now add this to the base value and get 103 - 20 = 83. After clamping we still get 83 = 01010011 binary. This is our 8-bit output value.
This specification gives the output for each channel in 8-bit integer values between 0 and 255, and these values all need to be divided by 255 to obtain the final floating point representation.
Note that hardware can be effectively shared between the alpha decoding part of this format and that of R11 EAC texture. For details on how to reuse hardware, see Section 22.5.
Decompression of floating point sRGB values in RGBA ETC2 with sRGB encoding follows that of floating point RGB values of linear RGBA ETC2. The result is sRGB values between 0.0 and 1.0. The further conversion from an sRGB encoded component cs to a linear component cl is according to the formula in Section 13.3. Assume cs is the sRGB component in the range [0, 1].
The alpha component of RGBA ETC2 with sRGB encoding is done in the same way as for linear RGBA ETC2.
The number of bits to represent a 4×4 texel block is 64 bits. if format is R11 EAC. In that case the data for a block is stored as a number of bytes, {q0, q1, q2, q3, q4, q5, q6, q7}, where byte q0 is located at the lowest memory address and q7 at the highest. The red component of the 4×4 block is then represented by the following 64-bit integer:
This 64-bit word contains information about a single-channel 4×4 pixel block as shown in Figure 43. The 64-bit word is split into two parts. The first 16 bits comprise a base codeword, a table codeword and a multiplier. The remaining 48 bits are divided into 16 3-bit indices, which are used to select one of the 8 possible values for each pixel in the block, as shown in Table 132.
The decoded value is calculated as:
Equation 4. Unsigned R11 EAC start
where clamp1(·) maps values outside the range [0.0, 1.0] to 0.0 or 1.0.
We will now go into detail how the decoding is done. The result will be an 11-bit fixed point number where 0 represents 0.0 and 2047 represents 1.0. This is the exact representation for the decoded value. However, some implementations may use, e.g., 16-bits of accuracy for filtering. In such a case the 11-bit value will be extended to 16 bits in a predefined way, which we will describe later.
To get a value between 0 and 2047 we must multiply Equation 4 by 2047.0:
where clamp2(·) clamps to the range [0.0, 2047.0].
Since $2047.0 \over 255.875$ is exactly 8.0, the above equation can be written as
Equation 5. Unsigned R11 EAC simple
The base codeword is stored in the first 8 bits as shown in Table 132 part (a). Bits 63..56 in each block represent an eight-bit integer (base codeword) which is multiplied by 8 by shifting three steps to the left. We can add 4 to this value without addition logic by just inserting 100 binary in the last three bits after the shift. For example, if base codeword is 129 = 10000001 binary (or 10000001b for short), the shifted value is 10000001000b and the shifted value including the +4 term is 10000001100b = 1036 = 129×8+4. Hence we have summed together the first two terms of the sum in Equation 5.
Next, we want to obtain the modifier. Bits 51..48 form a 4-bit index used to select one of 16 pre-determined ‘modifier tables’, shown in Table 133. For example, a table index of 13 (1101 binary) means that we should use table [-1, -2, -3, -10, 0, 1, 2, 9]. To select which of these values we should use, we consult the pixel index of the pixel we want to decode. Bits 47..0 are used to store a 3-bit index for each pixel in the block, selecting one of the 8 possible values. Assume we are interested in pixel b. Its pixel indices are stored in bit 44..42, with the most significant bit stored in 44 and the least significant bit stored in 42. If the pixel index is 011 binary = 3, this means we should take the value 3 from the left in the table, which is -10. This is now our modifier, which is the starting point of our second term in the sum.
In the next step we obtain the multiplier value; bits 55..52 form a four-bit multiplier between 0 and 15. We will later treat what happens if the multiplier value is zero, but if it is nonzero, it should be multiplied with the modifier. This product should then be shifted three steps to the left to implement the ×8 multiplication. The result now provides the third and final term in the sum in Equation 5. The sum is calculated and the result is clamped to a value in the interval [0..2047]. The resulting value is the 11-bit output value.
For example, assume a base codeword of 103, a table index of 13, a pixel index of 3 and a multiplier of 2. We will then first multiply the base codeword 103 (01100111b) by 8 by left-shifting it (0110111000b) and then add 4 resulting in 0110111100b = 828 = 103 × 8 + 4. Next, a table index of 13 selects table [-1, -2, -3, -10, 0, 1, 2, 9], and using a pixel index of 3 will result in a modifier of -10. The multiplier is nonzero, which means that we should multiply it with the modifier, forming -10 × 2 = -20 = 111111101100b. This value should in turn be multiplied by 8 by left-shifting it three steps: 111101100000b = -160. We now add this to the base value and get 828 - 160 = 668. After clamping we still get 668 = 01010011100b. This is our 11-bit output value, which represents the value ${668 \over 2047} = 0.32633121 \ldots$
If the multiplier value is zero (i.e., the multiplier bits 55..52 are all zero), we should set the multiplier to $1.0\over 8.0$ . Equation 5 can then be simplified to
Equation 6. Unsigned R11 EAC simpler
As an example, assume a base codeword of 103, a table index of 13, a pixel index of 3 and a multiplier value of 0. We treat the base codeword the same way, getting 828 = 103×8+4. The modifier is still -10. But the multiplier should now be $1 \over 8$ , which means that third term becomes $-10\times \left({1\over 8}\right)\times 8 = -10$ . The sum therefore becomes 828-10 = 818. After clamping we still get 818 = 01100110010b, and this is our 11-bit output value, and it represents ${818 \over 2047} = 0.39960918 \ldots$
Some OpenGL ES implementations may find it convenient to use 16-bit values for further processing. In this case, the 11-bit value should be extended using bit replication. An 11-bit value x is extended to 16 bits through $(x\ll 5) + (x \gg 6)$ . For example, the value 668 = 01010011100b should be extended to 0101001110001010b = 21386.
In general, the implementation may extend the value to any number of bits that is convenient for further processing, e.g., 32 bits. In these cases, bit replication should be used. On the other hand, an implementation is not allowed to truncate the 11-bit value to less than 11 bits.
Note that the method does not have the same reconstruction levels as the alpha part in the RGBA ETC2 format. For instance, for a base codeword of 255 and a table value of 0, the alpha part of the RGBA ETC2 format will represent a value of ${(255+0)\over 255.0} = 1.0$ exactly. In R11 EAC the same base codeword and table value will instead represent ${(255.5+0)\over 255.875} = 0.99853444 \ldots$ That said, it is still possible to decode the alpha part of the RGBA ETC2-format using R11 EAC hardware. This is done by truncating the 11-bit number to 8 bits. As an example, if base codeword = 255 and table value = 0, we get the 11-bit value (255×8+4+0) = 2044 = 1111111100b, which after truncation becomes the 8-bit value 11111111b = 255 which is exactly the correct value according to RGBA ETC2. Clamping has to be done to [0, 255] after truncation for RGBA ETC2 decoding. Care must also be taken to handle the case when the multiplier value is zero. In the 11-bit version, this means multiplying by $1 \over 8$ , but in the 8-bit version, it really means multiplication by 0. Thus, the decoder will have to know if it is an RGBA ETC2 texture or an R11 EAC texture to decode correctly, but the hardware can be 100% shared.
As stated above, a base codeword of 255 and a table value of 0 will represent a value of ${(255.5+0) \over 255.875} = 0.99853444 \ldots$ , and this does not reach 1.0 even though 255 is the highest possible base codeword. However, it is still possible to reach a pixel value of 1.0 since a modifier other than 0 can be used. Indeed, half of the modifiers will often produce a value of 1.0. As an example, assume we choose the base codeword 255, a multiplier of 1 and the modifier table [-3, -5, -7, -9, 2, 4, 6, 8]. Starting with Equation 5,
we get
which equals
or after clamping
which shows that several values can be 1.0, even though the base value does not reach 1.0. The same reasoning goes for 0.0.
The number of bits to represent a 4×4 texel block is 128 bits if the format is RG11 EAC. In that case the data for a block is stored as a number of bytes, {q0, q1, q2, q3, q4, q5, q6, q7, p0, p1, p2, p3, p4, p5, p6, p7} where byte q0 is located at the lowest memory address and p7 at the highest. The 128 bits specifying the block are then represented by the following two 64 bit integers:
The 64-bit word int64bit0 contains information about the red component of a two-channel 4×4 pixel block as shown in Figure 43, and the word int64bit1 contains information about the green component. Both 64-bit integers are decoded in the same way as R11 EAC described in Section 22.5.
The number of bits to represent a 4×4 texel block is 64 bits if the format is signed R11 EAC. In that case the data for a block is stored as a number of bytes, {q0, q1, q2, q3, q4, q5, q6, q7}, where byte q0 is located at the lowest memory address and q7 at the highest. The red component of the 4×4 block is then represented by the following 64 bit integer:
This 64-bit word contains information about a single-channel 4×4 pixel block as shown in Figure 43. The 64-bit word is split into two parts. The first 16 bits comprise a base codeword, a table codeword and a multiplier. The remaining 48 bits are divided into 16 3-bit indices, which are used to select one of the 8 possible values for each pixel in the block, as shown in Table 132.
The decoded value is calculated as
Equation 7. Signed R11 EAC start
where clamp1(·) maps values outside the range [-1.0, 1.0] to -1.0 or 1.0. We will now go into detail how the decoding is done. The result will be an 11-bit two’s-complement fixed point number where -1023 represents -1.0 and 1023 represents 1.0. This is the exact representation for the decoded value. However, some implementations may use, e.g., 16-bits of accuracy for filtering. In such a case the 11-bit value will be extended to 16 bits in a predefined way, which we will describe later.
To get a value between -1023 and 1023 we must multiply Equation 7 by 1023.0:
where clamp2(·) clamps to the range [-1023.0, 1023.0]. Since $1023.0\over 127.875$ is exactly 8, the above formula can be written as:
Equation 8. Signed R11 EAC simple
The base codeword is stored in the first 8 bits as shown in Table 132 part (a). It is a two’s-complement value in the range [-127, 127], and where the value -128 is not allowed; however, if it should occur anyway it must be treated as -127. The base codeword is then multiplied by 8 by shifting it left three steps. For example the value 65 = 01000001 binary (or 01000001b for short) is shifted to 01000001000b = 520 = 65×8.
Next, we want to obtain the modifier. Bits 51..48 form a 4-bit index used to select one of 16 pre-determined ‘modifier tables’, shown in Table 133. For example, a table index of 13 (1101 binary) means that we should use table [-1, -2, -3, -10, 0, 1, 2, 9]. To select which of these values we should use, we consult the pixel index of the pixel we want to decode. Bits 47..0 are used to store a 3-bit index for each pixel in the block, selecting one of the 8 possible values. Assume we are interested in pixel b. Its pixel indices are stored in bit 44..42, with the most significant bit stored in 44 and the least significant bit stored in 42. If the pixel index is 011 binary = 3, this means we should take the value 3 from the left in the table, which is -10. This is now our modifier, which is the starting point of our second term in the sum.
In the next step we obtain the multiplier value; bits 55..52 form a four-bit multiplier between 0 and 15. We will later treat what happens if the multiplier value is zero, but if it is nonzero, it should be multiplied with the modifier. This product should then be shifted three steps to the left to implement the ×8 multiplication. The result now provides the third and final term in the sum in Equation 8. The sum is calculated and the result is clamped to a value in the interval [-1023..1023]. The resulting value is the 11-bit output value.
For example, assume a a base codeword of 60, a table index of 13, a pixel index of 3 and a multiplier of 2. We start by multiplying the base codeword (00111100b) by 8 using bit shift, resulting in (00111100000b) = 480 = 60 × 8. Next, a table index of 13 selects table [-1, -2, -3, -10, 0, 1, 2, 9], and using a pixel index of 3 will result in a modifier of -10. The multiplier is nonzero, which means that we should multiply it with the modifier, forming -10×2 = -20 = 111111101100b. This value should in turn be multiplied by 8 by left-shifting it three steps: 111101100000b = -160. We now add this to the base value and get 480-160 = 320. After clamping we still get 320 = 00101000000b. This is our 11-bit output value, which represents the value ${320\over 1023} = 0.31280547\ldots$ .
If the multiplier value is zero (i.e., the multiplier bits 55..52 are all zero), we should set the multiplier to $1.0 \over 8.0$ . Equation 8 can then be simplified to:
Equation 9. Signed R11 EAC simpler
As an example, assume a base codeword of 65, a table index of 13, a pixel index of 3 and a multiplier value of 0. We treat the base codeword the same way, getting 480 = 60×8. The modifier is still -10. But the multiplier should now be $1 \over 8$ , which means that third term becomes $-10\times\left({1 \over 8}\right)\times 8 = -10$ . The sum therefore becomes 480-10 = 470. Clamping does not affect the value since it is already in the range [-1023, 1023], and the 11-bit output value is therefore 470 = 00111010110b. This represents ${470\over 1023} = 0.45943304 \dots$
Some OpenGL ES implementations may find it convenient to use two’s-complement 16-bit values for further processing. In this case, a positive 11-bit value should be extended using bit replication on all the bits except the sign bit. An 11-bit value x is extended to 16 bits through (x $\ll$ 5) + (x $\gg$ 5). Since the sign bit is zero for a positive value, no addition logic is needed for the bit replication in this case. For example, the value 470 = 00111010110b in the above example should be expanded to 0011101011001110b = 15054. A negative 11-bit value must first be made positive before bit replication, and then made negative again:
if (result11bit >= 0) { result16bit = (result11bit << 5) + (result11bit >> 5); } else { result11bit = -result11bit; result16bit = (result11bit << 5) + (result11bit >> 5); result16bit = -result16bit; }
Simply bit replicating a negative number without first making it positive will not give a correct result.
In general, the implementation may extend the value to any number of bits that is convenient for further processing, e.g., 32 bits. In these cases, bit replication according to the above should be used. On the other hand, an implementation is not allowed to truncate the 11-bit value to less than 11 bits.
Note that it is not possible to specify a base value of 1.0 or -1.0. The largest possible base codeword is +127, which represents ${127 \over 127.875} = 0.993\ldots$ . However, it is still possible to reach a pixel value of 1.0 or -1.0, since the base value is modified by the table before the pixel value is calculated. Indeed, half of the modifiers will often produce a value of 1.0. As an example, assume the base codeword is +127, the modifier table is [-3, -5, -7, -9, 2, 4, 6, 8] and the multiplier is one. Starting with Equation 7,
we get
which equals
or after clamping
This shows that it is indeed possible to arrive at the value 1.0. The same reasoning goes for -1.0.
Note also that Equation 8/Equation 9 are very similar to Equation 5/Equation 6 in the unsigned version EAC_R11. Apart from the +4, the clamping and the extension to bit sizes other than 11, the same decoding hardware can be shared between the two codecs.
The number of bits to represent a 4×4 texel block is 128 bits if the format is signed RG11 EAC. In that case the data for a block is stored as a number of bytes, {q0, q1, q2, q3, q4, q5, q6, q7, p0, p1, p2, p3, p4, p5, p6, p7} where byte q0 is located at the lowest memory address and p7 at the highest. The 128 bits specifying the block are then represented by the following two 64 bit integers:
The 64-bit word int64bit0 contains information about the red component of a two-channel 4×4 pixel block as shown in Figure 43, and the word int64bit1 contains information about the green component. Both 64-bit integers are decoded in the same way as signed R11 EAC described in Section 22.8.
For RGB ETC2 with punchthrough alpha, each 64-bit word contains information about a four-channel 4×4 pixel block as shown in Figure 43.
The blocks are compressed using one of four different ‘modes’. Table 134 part (a) shows the bits used for determining the mode used in a given block.
Table 134. Texel Data format for punchthrough alpha ETC2 compressed texture formats
a) Location of bits for mode selection | |||||||||||||||||||||||||||||||
63 | 62 | 61 | 60 | 59 | 58 | 57 | 56 | 55 | 54 | 53 | 52 | 51 | 50 | 49 | 48 | 47 | 46 | 45 | 44 | 43 | 42 | 41 | 40 | 39 | 38 | 37 | 36 | 35 | 34 | 33 | 32 |
R | Rd | G | Gd | B | Bd | …… | Op | . | |||||||||||||||||||||||
b) Bit layout for bits 63 through 32 for ‘differential’ mode | |||||||||||||||||||||||||||||||
63 | 62 | 61 | 60 | 59 | 58 | 57 | 56 | 55 | 54 | 53 | 52 | 51 | 50 | 49 | 48 | 47 | 46 | 45 | 44 | 43 | 42 | 41 | 40 | 39 | 38 | 37 | 36 | 35 | 34 | 33 | 32 |
R | Rd | G | Gd | B | Bd | table1 | table2 | Op | FB | ||||||||||||||||||||||
c) Bit layout for bits 63 through 32 for ‘T’ mode | |||||||||||||||||||||||||||||||
63 | 62 | 61 | 60 | 59 | 58 | 57 | 56 | 55 | 54 | 53 | 52 | 51 | 50 | 49 | 48 | 47 | 46 | 45 | 44 | 43 | 42 | 41 | 40 | 39 | 38 | 37 | 36 | 35 | 34 | 33 | 32 |
… | R3..2 | . | R1..0 | G | B | R2 | G2 | B2 | da | Op | db | ||||||||||||||||||||
d) Bit layout for bits 63 through 32 for ‘H’ mode | |||||||||||||||||||||||||||||||
63 | 62 | 61 | 60 | 59 | 58 | 57 | 56 | 55 | 54 | 53 | 52 | 51 | 50 | 49 | 48 | 47 | 46 | 45 | 44 | 43 | 42 | 41 | 40 | 39 | 38 | 37 | 36 | 35 | 34 | 33 | 32 |
. | R | G3..1 | … | G0 | B3 | . | B2..0 | R2 | G2 | B2 | da | Op | db | ||||||||||||||||||
e) Bit layout for bits 31 through 0 for ‘differential’, ‘T’ and ‘H’ modes | |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
p1 | o1 | n1 | m1 | l1 | k1 | j1 | i1 | h1 | g1 | f1 | e1 | d1 | c1 | b1 | a1 | p0 | o0 | n0 | m0 | l0 | k0 | j0 | i0 | h0 | g0 | f0 | e0 | d0 | c0 | b0 | a0 |
f) Bit layout for bits 63 through 0 for ‘planar’ mode | |||||||||||||||||||||||||||||||
63 | 62 | 61 | 60 | 59 | 58 | 57 | 56 | 55 | 54 | 53 | 52 | 51 | 50 | 49 | 48 | 47 | 46 | 45 | 44 | 43 | 42 | 41 | 40 | 39 | 38 | 37 | 36 | 35 | 34 | 33 | 32 |
. | R | G6 | . | G5..0 | B5 | … | B4..3 | . | B2..0 | Rh5..1 | 1 | Rh0 | |||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
Gh | Bh | Rv | Gv | Bv |
To determine the mode, the three 5-bit values R, G and B, and the three 3-bit values Rd, Gd and Bd are examined. R, G and B are treated as integers between 0 and 31 and Rd, Gd and Bd as two’s-complement integers between -4 and +3. First, R and Rd are added, and if the sum is not within the interval [0..31], the ‘T’ mode is selected. Otherwise, if the sum of G and Gd is outside the interval [0..31], the ‘H’ mode is selected. Otherwise, if the sum of B and Bd is outside of the interval [0..31], the ‘planar’ mode is selected. Finally, if all of the aforementioned sums lie between 0 and 31, the ‘differential’ mode is selected.
The layout of the bits used to decode the ‘differential’ mode is shown in Table 134 part (b). In this mode, the 4×4 block is split into two subblocks of either size 2×4 or 4×2. This is controlled by bit 32, which we dub the flip bit (FB in Table 134 (b) and (c)). If the flip bit is 0, the block is divided into two 2×4 subblocks side-by-side, as shown in Figure 44. If the flip bit is 1, the block is divided into two 4×2 subblocks on top of each other, as shown in Figure 45. For each subblock, a base color is stored.
In the ‘differential’ mode, following the layout shown in Table 134 part (b), the base color for subblock 1 is derived from the five-bit codewords R, G and B. These five-bit codewords are extended to eight bits by replicating the top three highest-order bits to the three lowest-order bits. For instance, if R = 28 = 11100 binary (11100b for short), the resulting eight-bit red color component becomes 11100111b = 231. Likewise, if G = 4 = 00100b and B = 3 = 00011b, the green and blue components become 00100001b = 33 and 00011000b = 24 respectively. Thus, in this example, the base color for subblock 1 is (231, 33, 24). The five bit representation for the base color of subblock 2 is obtained by modifying the 5-bit codewords R, G and B by the codewords Rd, Gd and Bd. Each of Rd, Gd and Bd is a 3-bit two’s-complement number that can hold values between -4 and +3. For instance, if R = 28 as above, and Rd = 100b = -4, then the five bit representation for the red color component is 28+(-4)=24 = 11000b, which is then extended to eight bits to 11000110b = 198. Likewise, if G = 4, Gd = 2, B = 3 and Bd = 0, the base color of subblock 2 will be RGB = (198, 49, 24). In summary, the base colors for the subblocks in the differential mode are:
Note that these additions will not under- or overflow, or one of the alternative decompression modes would have been chosen instead of the ‘differential’ mode.
Table 135. ETC2 intensity modifier sets for the ‘differential’ if ‘opaque’ (Op) is set
Table codeword | Modifier table | |||
---|---|---|---|---|
0 | -8 | -2 | 2 | 8 |
1 | -17 | -5 | 5 | 17 |
2 | -29 | -9 | 9 | 29 |
3 | -42 | -13 | 13 | 42 |
4 | -60 | -18 | 18 | 60 |
5 | -80 | -24 | 24 | 80 |
6 | -106 | -33 | 33 | 106 |
7 | -183 | -47 | 47 | 183 |
Table 136. ETC2 intensity modifier sets for the ‘differential’ if ‘opaque’ (Op) is unset
Table codeword | Modifier table | |||
---|---|---|---|---|
0 | -8 | 0 | 0 | 8 |
1 | -17 | 0 | 0 | 17 |
2 | -29 | 0 | 0 | 29 |
3 | -42 | 0 | 0 | 42 |
4 | -60 | 0 | 0 | 60 |
5 | -80 | 0 | 0 | 80 |
6 | -106 | 0 | 0 | 106 |
7 | -183 | 0 | 0 | 183 |
After obtaining the base color, a table is chosen using the table codewords: For subblock 1, table codeword 1 is used (bits 39..37), and for subblock 2, table codeword 2 is used (bits 36..34), see Table 134 part (b). The table codeword is used to select one of eight modifier tables. If the ‘opaque’-bit (bit 33) is set, Table 135 is used. If it is unset, Table 136 is used. For instance, if the ‘opaque’-bit is 1 and the table codeword is 010 binary = 2, then the modifier table [-29, -9, 9, 29] is selected for the corresponding sub-block. Note that the values in Table 135 and Table 136 are valid for all textures and can therefore be hardcoded into the decompression unit.
Next, we identify which modifier value to use from the modifier table using the two pixel index bits. The pixel index bits are unique for each pixel. For instance, the pixel index for pixel d (see Figure 43) can be found in bits 19 (most significant bit, MSB), and 3 (least significant bit, LSB), see Table 134 part (e). Note that the pixel index for a particular texel is always stored in the same bit position, irrespectively of the flip bit.
If the ‘opaque’-bit (bit 33) is set, the pixel index bits are decoded using Table 137. If the ‘opaque’-bit is unset, Table 138 will be used instead. If, for instance, the ‘opaque’-bit is 1, and the pixel index bits are 01 binary = 1, and the modifier table [-29, -9, 9, 29] is used, then the modifier value selected for that pixel is 29 (see Table 137). This modifier value is now used to additively modify the base color. For example, if we have the base color (231, 8, 16), we should add the modifier value 29 to all three components: (231+29, 8+29, 16+29) resulting in (260, 37, 45). These values are then clamped to [0..255], resulting in the color (255, 37, 45).
Table 137. ETC2 mapping from pixel index values to modifier values when ‘opaque’ bit is set
Pixel index value | Resulting modifier value | |
msb | lsb | |
1 | 1 | -b (large negative value) |
1 | 0 | -a (small negative value) |
0 | 0 | +a (small positive value) |
0 | 1 | +b (large positive value) |
Table 138. ETC2 mapping from pixel index values to modifier values when ‘opaque’ bit is unset
Pixel index value | Resulting modifier value | |
msb | lsb | |
1 | 1 | -b (large negative value) |
1 | 0 | 0 (zero) |
0 | 0 | 0 (zero) |
0 | 1 | +b (large positive value) |
The alpha component is decoded using the ‘opaque’-bit, which is positioned in bit 33 (see Table 134 part (b)). If the ‘opaque’-bit is set, alpha is always 255. However, if the ‘opaque’-bit is zero, the alpha-value depends on the pixel indices; if MSB==1 and LSB==0, the alpha value will be zero, otherwise it will be 255. Finally, if the alpha value equals 0, the red, green and blue components will also be zero.
if (opaque == 0 && MSB == 1 && LSB == 0) { red = 0; green = 0; blue = 0; alpha = 0; } else { alpha = 255; }
Hence paint color 2 will equal RGBA = (0, 0, 0, 0) if opaque = 0.
In the example above, assume that the ‘opaque’-bit was instead 0. Then, since the MSB = 0 and LSB 1, alpha will be 255, and the final decoded RGBA-tuple will be (255, 37, 45, 255).
The ‘T’ and ‘H’ compression modes share some characteristics: both use two base colors stored using 4 bits per channel. These bits are not stored sequentially, but in the layout shown in Table 134 part (c) and Table 134 part (d). To clarify, in the ‘T’ mode, the two colors are constructed as follows:
In the ‘H’ mode, the two colors are constructed as follows:
The function extend4to8bits(·) just replicates the four bits twice. This is equivalent to multiplying by 17. As an example, extend4to8bits(1101b) equals 11011101b = 221.
Both the ‘T’ and ‘H’ modes have four paint colors which are the colors that will be used in the decompressed block, but they are assigned in a different manner. In the ‘T’ mode, paint color 0 is simply the first base color, and paint color 2 is the second base color. To obtain the other paint colors, a ‘distance’ is first determined, which will be used to modify the luminance of one of the base colors. This is done by combining the values da and db shown in Table 134 part (c) by (da $\ll$ 1) | db, and then using this value as an index into the small look-up table shown in Table 131. For example, if da is 10 binary and db is 1 binary, the index is 101 binary and the selected distance d will be 32. Paint color 1 is then equal to the second base color with the ‘distance’ d added to each channel, and paint color 3 is the second base color with the ‘distance’ d subtracted. In summary, to determine the four paint colors for a ‘T’ block:
In both cases, the value of each channel is clamped to within [0..255].
Just as for the differential mode, the RGB channels are set to zero if alpha is zero, and the alpha component is calculated the same way:
if (opaque == 0 && MSB == 1 && LSB == 0) { red = 0; green = 0; blue = 0; alpha = 0; } else { alpha = 255; }
A ‘distance’ value is computed for the ‘H’ mode as well, but doing so is slightly more complex. In order to construct the three-bit index into the distance table shown in Table 131, da and db shown in Table 134 part (d) are used as the most significant bit and middle bit, respectively, but the least significant bit is computed as (base color 1 value $\geq$ base color 2 value), the ‘value’ of a color for the comparison being equal to (R $\ll$ 16) + (G $\ll$ 8) + B. Once the ‘distance’ d has been determined for an ‘H’ block, the four paint colors will be:
Yet again, RGB is zeroed if alpha is 0 and the alpha component is determined the same way:
if (opaque == 0 && MSB == 1 && LSB == 0) { red = 0; green = 0; blue = 0; alpha = 0; } else { alpha = 255; }
Hence paint color 2 will have R = G = B = alpha = 0 if opaque = 0.
Again, all color components are clamped to within [0..255]. Finally, in both the ‘T’ and ‘H’ modes, every pixel is assigned one of the four paint colors in the same way the four modifier values are distributed in ‘individual’ or ‘differential’ blocks. For example, to choose a paint color for pixel d, an index is constructed using bit 19 as most significant bit and bit 3 as least significant bit. Then, if a pixel has index 2, for example, it will be assigned paint color 2.
The final mode possible in an RGB ETC2 with punchthrough alpha — compressed block is the ‘planar’ mode. In this mode, the ‘opaque’-bit must be 1 (a valid encoder should not produce an ‘opaque’-bit equal to 0 in the planar mode), but should the ‘opaque’-bit anyway be 0 the decoder should treat it as if it were 1. In the ‘planar’ mode, three base colors are supplied and used to form a color plane used to determine the color of the individual pixels in the block.
All three base colors are stored in RGB:676 format, and stored in the manner shown in Table 134 part (f). The two secondary colors are given the suffix ‘h’ and ‘v’, so that the red component of the three colors are R, Rh and Rv, for example. Some color channels are split into non-consecutive bit-ranges; for example B is reconstructed using B5 as the most-significant bit, B4..3 as the two following bits, and B2..0 as the three least-significant bits.
Once the bits for the base colors have been extracted, they must be extended to 8 bits per channel in a manner analogous to the method used for the base colors in other modes. For example, the 6-bit blue and red channels are extended by replicating the two most significant of the six bits to the two least significant of the final 8 bits.
With three base colors in RGB:888 format, the color of each pixel can then be determined as:
where x and y are values from 0 to 3 corresponding to the pixels coordinates within the block, x being in the u direction and y in the v direction. For example, the pixel g in Figure 43 would have x = 1 and y = 2.
These values are then rounded to the nearest integer (to the larger integer if there is a tie) and then clamped to a value between 0 and 255. Note that this is equivalent to
where clamp255(·) clamps the value to a number in the range [0..255].
Note that the alpha component is always 255 in the planar mode.
This specification gives the output for each compression mode in 8-bit integer colors between 0 and 255, and these values all need to be divided by 255 for the final floating point representation.
Decompression of floating point sRGB values in RGB ETC2 with sRGB encoding and punchthrough alpha follows that of floating point RGB values of RGB ETC2 with punchthrough alpha. The result is sRGB values between 0.0 and 1.0. The further conversion from an sRGB encoded component, cs, to a linear component, cl, is according to the formula in Section 13.3. Assume cs is the sRGB component in the range [0, 1]. Note that the alpha component is not gamma corrected, and hence does not use this formula.
This description is derived from the Khronos OES_texture_compression_astc OpenGL extension.
ASTC stands for Adaptive Scalable Texture Compression. The ASTC formats form a family of related compressed texture image formats. They are all derived from a common set of definitions.
ASTC textures may be either 2D or 3D.
ASTC textures may be encoded using either high or low dynamic range. Low dynamic range images may optionally be specified using the sRGB transfer function for the RGB channels.
Two sub-profiles (“LDR Profile” and “HDR Profile”) may be implemented, which support only 2D images at low or high dynamic range respectively.
ASTC textures may be encoded as 1, 2, 3 or 4 components, but they are all decoded into RGBA. ASTC has a variable block size.
The design goals for the format are as follows:
ASTC is a block-based lossy compression format. The compressed image is divided into a number of blocks of uniform size, which makes it possible to quickly determine which block a given texel resides in.
Each block has a fixed memory footprint of 128 bits, but these bits can represent varying numbers of texels (the block “footprint”).
The term “block footprint” in ASTC refers to the same concept as “compressed texel block dimensions” elsewhere in the Data Format Specification. |
Block footprint sizes are not confined to powers-of-two, and are also not confined to be square. They may be 2D, in which case the block dimensions range from 4 to 12 texels, or 3D, in which case the block dimensions range from 3 to 6 texels.
Decoding one texel requires only the data from a single block. This simplifies cache design, reduces bandwidth and improves encoder throughput.
To understand how the blocks are stored and decoded, it is useful to start with a simple example, and then introduce additional features.
The simplest block encoding starts by defining two color “endpoints”. The endpoints define two colors, and a number of additional colors are generated by interpolating between them. We can define these colors using 1, 2, 3, or 4 components (usually corresponding to R, RG, RGB and RGBA textures), and using low or high dynamic range.
We then store a color interpolant weight for each texel in the image, which specifies how to calculate the color to use. From this, a weighted average of the two endpoint colors is used to generate the intermediate color, which is the returned color for this texel.
There are several different ways of specifying the endpoint colors, and the weights, but once they have been defined, calculation of the texel colors proceeds identically for all of them. Each block is free to choose whichever encoding scheme best represents its color endpoints, within the constraint that all the data fits within the 128 bit block.
For blocks which have a large number of texels (e.g. a 12×12 block), there is not enough space to explicitly store a weight for every texel. In this case, a sparser grid with fewer weights is stored, and interpolation is used to determine the effective weight to be used for each texel position. This allows very low bit rates to be used with acceptable quality. This can also be used to more efficiently encode blocks with low detail, or with strong vertical or horizontal features.
For blocks which have a mixture of disparate colors, a single line in the color space is not a good fit to the colors of the pixels in the original image. It is therefore possible to partition the texels into multiple sets, the pixels within each set having similar colors. For each of these “partitions”, we specify separate endpoint pairs, and choose which pair of endpoints to use for a particular texel by looking up the partition index from a partitioning pattern table. In ASTC, this partition table is actually implemented as a function.
The endpoint encoding for each partition is independent.
For blocks which have uncorrelated channels — for example an image with a transparency mask, or an image used as a normal map — it may be necessary to specify two weights for each texel. Interpolation between the components of the endpoint colors can then proceed independently for each “plane” of the image. The assignment of channels to planes is selectable.
Since each of the above options is independent, it is possible to specify any combination of channels, endpoint color encoding, weight encoding, interpolation, multiple partitions and single or dual planes.
Since these values are specified per block, it is important that they are represented with the minimum possible number of bits. As a result, these values are packed together in ways which can be difficult to read, but which are nevertheless highly amenable to hardware decode.
All of the values used as weights and color endpoint values can be specified with a variable number of bits. The encoding scheme used allows a fine-grained tradeoff between weight bits and color endpoint bits using “integer sequence encoding”. This can pack adjacent values together, allowing us to use fractional numbers of bits per value.
Finally, a block may be just a single color. This is a so-called “void extent block” and has a special coding which also allows it to identify nearby regions of single color. This may be used to short-circuit fetching of what would be identical blocks, and further reduce memory bandwidth.
The decoding process for LDR content can be simplified if it is known in advance that sRGB output is required. This selection is therefore included as part of the global configuration.
The two modes differ in various ways, as shown in Table 139.
Table 139. ASTC differences between LDR and HDR modes
Operation | LDR mode | HDR mode |
---|---|---|
Returned Value | Determined by decoding mode | |
sRGB compatible | Yes | No |
LDR endpoint decoding precision | 16 bits, or 8 bits for sRGB | 16 bits |
HDR endpoint mode results | Error color | As decoded |
Error results | Error color | Vector of NaNs (0xFFFF) |
The type of the values returned by the decoding process is determined by the decoding mode as shown in Table 140.
Table 140. ASTC decoding modes
Decode mode | LDR mode | HDR mode |
---|---|---|
decode_float16 | Vector of FP16 values | |
decode_unorm8 | Vector of 8-bit unsigned normalized values | invalid |
decode_rgb9e5 | Vector using a shared exponent format |
Using the decode_unorm8 decoding mode in HDR mode gives undefined results.
For sRGB, the decoding mode is ignored, ad the decoding always returns a vector of 8-bit unsigned normalized values.
The error color is opaque fully-saturated magenta (R,G,B,A) = (0xFF, 0x00, 0xFF, 0xFF). This has been chosen as it is much more noticeable than black or white, and occurs far less often in valid images.
For linear RGB decode, the error color may be either opaque fully-saturated magenta (R,G,B,A) = (1.0, 0.0, 1.0, 1.0) or a vector of four NaNs (R,G,B,A) = (NaN, NaN, NaN, NaN). In the latter case, the recommended NaN value returned is 0xFFFF.
When using the decode_rgb9e5 decoding mode in HDR mode, error results will return the error color because NaN cannot be represented.
The error color is returned as an informative response to invalid conditions, including invalid block encodings or use of reserved endpoint modes.
Future, forward-compatible extensions to ASTC may define valid interpretations of these conditions, which will decode to some other color. Therefore, encoders and applications must not rely on invalid encodings as a way of generating the error color.
The global configuration data for the format are as follows:
The data specified per block are as follows:
To decode one texel:
(Optimization: If within known void-extent, immediately return single color) Find block containing texel Read block mode If void-extent block, store void extent and immediately return single color For each plane in image If block mode requires infill Find and decode stored weights adjacent to texel, unquantize and interpolate Else Find and decode weight for texel, and unquantize Read number of partitions If number of partitions > 1 Read partition table pattern index Look up partition number from pattern Read color endpoint mode and endpoint data for selected partition Unquantize color endpoints Interpolate color endpoints using weight (or weights in dual-plane mode) Return interpolated color
The block footprint is a global setting for any given texture, and is therefore not encoded in the individual blocks.
For 2D textures, the block footprint’s width and height are selectable from a number of predefined sizes, namely 4, 5, 6, 8, 10 and 12 pixels.
For square and nearly-square blocks, this gives the bit rates in Table 141.
Table 141. ASTC 2D footprint and bit rates
Footprint | Bit Rate | Increment | |
Width | Height | ||
4 | 4 | 8.00 | 125% |
5 | 4 | 6.40 | 125% |
5 | 5 | 5.12 | 120% |
6 | 5 | 4.27 | 120% |
6 | 6 | 3.56 | 114% |
8 | 5 | 3.20 | 120% |
8 | 6 | 2.67 | 105% |
10 | 5 | 2.56 | 120% |
10 | 6 | 2.13 | 107% |
8 | 8 | 2.00 | 125% |
10 | 8 | 1.60 | 125% |
10 | 10 | 1.28 | 120% |
12 | 10 | 1.07 | 120% |
12 | 12 | 0.89 |
The “Increment” column indicates the ratio of bit rate against the next lower available rate. A consistent value in this column indicates an even spread of bit rates.
For 3D textures, the block footprint’s width, height and depth are selectable from a number of predefined sizes, namely 3, 4, 5, and 6 pixels.
For cubic and near-cubic blocks, this gives the bit rates in Table 142.
Table 142. ASTC 3D footprint and bit rates
Block Footprint | Bit Rate | Increment | ||
Width | Height | Depth | ||
3 | 3 | 3 | 4.74 | 133% |
4 | 3 | 3 | 3.56 | 133% |
4 | 4 | 3 | 2.67 | 133% |
4 | 4 | 4 | 2.00 | 125% |
5 | 4 | 4 | 1.60 | 125% |
5 | 5 | 4 | 1.28 | 125% |
5 | 5 | 5 | 1.02 | 120% |
6 | 5 | 5 | 0.85 | 120% |
6 | 6 | 5 | 0.71 | 120% |
6 | 6 | 6 | 0.59 |
The full profile supports only those block footprints listed in Table 141 and Table 142. Other block sizes are not supported.
For images which are not an integer multiple of the block size, additional texels are added to the edges with maximum X and Y (and Z for 3D textures). These texels may be any color, as they will not be accessed.
Although these are not all powers of two, it is possible to calculate block addresses and pixel addresses within the block, for legal image sizes, without undue complexity.
Given an image which is W × H × D pixels in size, with block size w × h × d, the size of the image in blocks is:
For a 3D image built from 2D slices, each 2D slice is a single texel thick, so that for an image which is W × H × D pixels in size, with block size w × h, the size of the image in blocks is:
Each block in the image is stored as a single 128-bit block in memory. These blocks are laid out in raster order, starting with the block at (0, 0, 0), then ordered sequentially by X, Y and finally Z (if present). They are aligned to 128-bit boundaries in memory.
The bits in the block are labeled in little-endian order — the byte at the lowest address contains bits 0..7. Bit 0 is the least significant bit in the byte.
Each block has the same basic layout, shown in Table 143.
Table 143. ASTC block layout
127 | 126 | 125 | 124 | 123 | 122 | 121 | 120 | 119 | 118 | 117 | 116 | 115 | 114 | 113 | 112 |
Texel weight data (variable width) | Fill direction $\rightarrow$ | ||||||||||||||
111 | 110 | 109 | 108 | 107 | 106 | 105 | 104 | 103 | 102 | 101 | 100 | 99 | 98 | 97 | 96 |
Texel weight data | |||||||||||||||
95 | 94 | 93 | 92 | 91 | 90 | 89 | 88 | 87 | 86 | 85 | 84 | 83 | 82 | 81 | 80 |
Texel weight data | |||||||||||||||
79 | 78 | 77 | 76 | 75 | 74 | 73 | 72 | 71 | 70 | 69 | 68 | 67 | 66 | 65 | 64 |
Texel weight data | |||||||||||||||
63 | 62 | 61 | 60 | 59 | 58 | 57 | 56 | 55 | 54 | 53 | 52 | 51 | 50 | 49 | 48 |
More config data | |||||||||||||||
47 | 46 | 45 | 44 | 43 | 42 | 41 | 40 | 39 | 38 | 37 | 36 | 35 | 34 | 33 | 32 |
$\leftarrow$ Fill direction | Color endpoint data | ||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 |
Extra configuration data | |||||||||||||||
15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
Extra | Part | Block mode |
Since the size of the “texel weight data” field is variable, the positions shown for the “more config data” field and “color endpoint data” field are only representative and not fixed.
The “Block mode” field specifies how the Texel Weight Data is encoded.
The “Part” field specifies the number of partitions, minus one. If dual plane mode is enabled, the number of partitions must be 3 or fewer. If 4 partitions are specified, the error value is returned for all texels in the block.
The size and layout of the extra configuration data depends on the number of partitions, and the number of planes in the image, as shown in Table 144 (only the bottom 32 bits are shown).
Table 144. ASTC single-partition block layout
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 |
Color endpoint data | CEM | ||||||||||||||
15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
CEM | 0 | 0 | Block mode |
CEM is the color endpoint mode field, which determines how the Color Endpoint Data is encoded.
If dual-plane mode is active, the color component selector bits appear directly below the weight bits, as shown in Table 145.
Table 145. ASTC multi-partition block layout
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 |
CEM | Partition index | ||||||||||||||
15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
Partition index | Part | Block mode |
The Partition Index field specifies which partition layout to use. CEM is the first 6 bits of color endpoint mode information for the various partitions. For modes which require more than 6 bits of CEM data, the additional bits appear at a variable position directly beneath the texel weight data.
If dual-plane mode is active, the color component selector bits then appear directly below the additional CEM bits.
The final special case is that if bits [8..0] of the block are “111111100”, then the block is a void-extent block, which has a separate encoding described in Section 23.23.
The block mode field specifies the width, height and depth of the grid of weights, what range of values they use, and whether dual weight planes are present. Since some these are not represented using powers of two (there are 12 possible weight widths, for example), and not all combinations are allowed, this is not a simple bit packing. However, it can be unpacked quickly in hardware.
The weight ranges are encoded using a 3-bit range value ρ, which is interpreted together with a low/high-precision bit P, as shown in Table 146. Each weight value is encoded using the specified number of Trits, Quints and Bits. The details of this encoding can be found in Section 23.12.
Table 146. ASTC weight range encodings
ρ2..0 | Low-precision range (P=0) | High-precision range (P=1) | ||||||
Weight range | Trits | Quints | Bits | Weight range | Trits | Quints | Bits | |
000 | Invalid | Invalid | ||||||
001 | Invalid | Invalid | ||||||
010 | 0..1 | 1 | 0..9 | 1 | 1 | |||
011 | 0..2 | 1 | 0..11 | 1 | 2 | |||
100 | 0..3 | 2 | 0..15 | 4 | ||||
101 | 0..4 | 1 | 0..19 | 1 | 2 | |||
110 | 0..5 | 1 | 1 | 0..23 | 1 | 3 | ||
111 | 0..7 | 3 | 0..31 | 5 |
For 2D blocks, the Block Mode field is laid out as shown in Table 147.
Table 147. ASTC 2D block mode layout, weight grid width and height
10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 | Wwidth | Wheight | Notes | ||||
DP | P | W | H | ρ0 | 0 | 0 | ρ2 | ρ1 | W+4 | H+2 | |||||||
DP | P | W | H | ρ0 | 0 | 1 | ρ2 | ρ1 | W+8 | H+2 | |||||||
DP | P | H | W | ρ0 | 1 | 0 | ρ2 | ρ1 | W+2 | H+8 | |||||||
DP | P | 0 | H | W | ρ0 | 1 | 1 | ρ2 | ρ1 | W+2 | H+6 | ||||||
DP | P | 1 | W | H | ρ0 | 1 | 1 | ρ2 | ρ1 | W+2 | H+2 | ||||||
DP | P | 0 | 0 | H | ρ0 | ρ2 | ρ1 | 0 | 0 | 12 | H+2 | ||||||
DP | P | 0 | 1 | W | ρ0 | ρ2 | ρ1 | 0 | 0 | W+2 | 12 | ||||||
DP | P | 1 | 1 | 0 | 0 | ρ0 | ρ2 | ρ1 | 0 | 0 | 6 | 10 | |||||
DP | P | 1 | 1 | 0 | 1 | ρ0 | ρ2 | ρ1 | 0 | 0 | 10 | 6 | |||||
H | 1 | 0 | W | ρ0 | ρ2 | ρ1 | 0 | 0 | W+6 | H+6 | DP=0, P=0 | ||||||
x | x | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 0 | 0 | - | - | Void-extent | ||||
x | x | 1 | 1 | 1 | x | x | x | x | 0 | 0 | - | - | Reserved* | ||||
x | x | x | x | x | x | x | 0 | 0 | 0 | 0 | - | - | Reserved |
Note that, due to the encoding of the ρ field, as described in the previous page, bits ρ2 and ρ1 cannot both be zero, which disambiguates the first five rows from the rest of the table.
Bit positions with a value of x are ignored for purposes of determining if a block is a void-extent block or reserved, but may have defined encodings for specific void-extent blocks.
The penultimate row of Table 147 is reserved only if bits [5..2] are not all 1, in which case it encodes a void-extent block (as shown in the previous row).
For 3D blocks, the Block Mode field is laid out as shown in Table 148.
Table 148. ASTC 3D block mode layout, weight grid width, height and depth
10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 | Wwidth | Wheight | Wdepth | Notes | |||||
DP | P | H | W | ρ0 | D | ρ2 | ρ1 | W+2 | H+2 | D+2 | |||||||||
H | 0 | 0 | D | ρ0 | ρ2 | ρ1 | 0 | 0 | 6 | H+2 | D+2 | DP=0, P=0 | |||||||
D | 0 | 1 | W | ρ0 | ρ2 | ρ1 | 0 | 0 | W+2 | 6 | D+2 | DP=0, P=0 | |||||||
H | 1 | 0 | W | ρ0 | ρ2 | ρ1 | 0 | 0 | W+2 | H+2 | 6 | DP=0, P=0 | |||||||
DP | P | 1 | 1 | 0 | 0 | ρ0 | ρ2 | ρ1 | 0 | 0 | 6 | 2 | 2 | ||||||
DP | P | 1 | 1 | 0 | 1 | ρ0 | ρ2 | ρ1 | 0 | 0 | 2 | 6 | 2 | ||||||
DP | P | 1 | 1 | 1 | 0 | ρ0 | ρ2 | ρ1 | 0 | 0 | 2 | 2 | 6 | ||||||
x | x | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 0 | 0 | - | - | - | Void-extent | |||||
x | x | 1 | 1 | 1 | 1 | x | x | x | 0 | 0 | - | - | - | Reserved* | |||||
x | x | x | x | x | x | x | 0 | 0 | 0 | 0 | - | - | - | Reserved |
The DP bit is set to indicate dual-plane mode. In this mode, the maximum allowed number of partitions is 3.
The penultimate row of Table 148 is reserved only if bits [4..2] are not all 1, in which case it encodes a void-extent block (as shown in the previous row).
The size of the weight grid in each dimension must be less than or equal to the corresponding dimension of the block footprint. If the grid size is greater than the footprint dimension in any axis, then this is an illegal block encoding and all texels will decode to the error color.
In single-partition mode, the Color Endpoint Mode (CEM) field stores one of 16 possible values. Each of these specifies how many raw data values are encoded, and how to convert these raw values into two RGBA color endpoints. They can be summarized as shown in Table 149.
Table 149. ASTC color endpoint modes
CEM | Description | Class |
---|---|---|
0 | LDR Luminance, direct | 0 |
1 | LDR Luminance, base+offset | 0 |
2 | HDR Luminance, large range | 0 |
3 | HDR Luminance, small range | 0 |
4 | LDR Luminance+Alpha, direct | 1 |
5 | LDR Luminance+Alpha, base+offset | 1 |
6 | LDR RGB, base+scale | 1 |
7 | HDR RGB, base+scale | 1 |
8 | LDR RGB, direct | 2 |
9 | LDR RGB, base+offset | 2 |
10 | LDR RGB, base+scale plus two A | 2 |
11 | HDR RGB, direct | 2 |
12 | LDR RGBA, direct | 3 |
13 | LDR RGBA, base+offset | 3 |
14 | HDR RGB, direct + LDR Alpha | 3 |
15 | HDR RGB, direct + HDR Alpha | 3 |
In multi-partition mode, the CEM field is of variable width, from 6 to 14 bits. The lowest 2 bits of the CEM field specify how the endpoint mode for each partition is calculated as shown in Table 150.
Table 150. ASTC Multi-Partition Color Endpoint Modes
Value | Meaning |
---|---|
00 | All color endpoint pairs are of the same type; a full 4-bit CEM is stored in block bits [28..25] and is used for all partitions |
01 | All endpoint pairs are of class 0 or 1 |
10 | All endpoint pairs are of class 1 or 2 |
11 | All endpoint pairs are of class 2 or 3 |
If the CEM selector value in bits [24..23] is not 00, then data layout is as shown in Table 151 and Table 152.
Table 151. ASTC multi-partition color endpoint mode layout
Part | n | m | l | k | j | i | h | g | |||
2 | … | Weight | M1 | … | |||||||
3 | … | Weight | M2 | M1 | M0 | … | |||||
4 | … | Weight | M3 | M2 | M1 | M0 | … |
Table 152. ASTC multi-partition color endpoint mode layout (2)
Part | 28 | 27 | 26 | 25 | 24 | 23 |
2 | M0 | C1 | C0 | CEM | ||
3 | M0 | C2 | C1 | C0 | CEM | |
4 | C3 | C2 | C1 | C0 | CEM |
In this view, each partition i has two fields. Ci is the class selector bit, choosing between the two possible CEM classes (0 indicates the lower of the two classes), and Mi is a two-bit field specifying the low bits of the color endpoint mode within that class. The additional bits appear at a variable bit position, immediately below the texel weight data.
The ranges used for the data values are not explicitly specified. Instead, they are derived from the number of available bits remaining after the configuration data and weight data have been specified.
Details of the decoding procedure for Color Endpoints can be found in Section 23.13.
Both the weight data and the endpoint color data are variable width, and are specified using a sequence of integer values. The range of each value in a sequence (e.g. a color weight) is constrained.
Since it is often the case that the most efficient range for these values is not a power of two, each value sequence is encoded using a technique known as “integer sequence encoding”. This allows efficient, hardware-friendly packing and unpacking of values with non-power-of-two ranges.
In a sequence, each value has an identical range. The range is specified in one of the forms shown in Table 153 and Table 154.
Table 153. ASTC range forms
Value range | MSB encoding | LSB encoding |
---|---|---|
$0 \dots 2^n-1$ | - | n-bit value m (n ≤ 8) |
$0 \dots (3 \times 2^n)-1$ | Base-3 “trit” value t | n-bit value m (n ≤ 6) |
$0 \dots (5 \times 2^n)-1$ | Base-5 “quint” value q | n-bit value m (n ≤ 5) |
Table 154. ASTC encoding for different ranges
Value range | Value | Block | Packed block size |
---|---|---|---|
$0 \dots 2^n-1$ | m | 1 | n |
$0 \dots (3 \times 2^n)-1$ | $t \times 2^n + m$ | 5 | 8 + 5 × n |
$0 \dots (5 \times 2^n)-1$ | $q \times 2^n + m$ | 3 | 7 + 3 × n |
Since 35 is 243, it is possible to pack five trits into 8 bits (which has 256 possible values), so a trit can effectively be encoded as 1.6 bits. Similarly, since 53 is 125, it is possible to pack three quints into 7 bits (which has 128 possible values), so a quint can be encoded as 2.33 bits.
The encoding scheme packs the trits or quints, and then interleaves the n additional bits in positions that satisfy the requirements of an arbitrary-length stream. This makes it possible to correctly specify lists of values whose length is not an integer multiple of 3 or 5 values. It also makes it possible to easily select a value at random within the stream.
If there are insufficient bits in the stream to fill the final block, then unused (higher-order) bits are assumed to be 0 when decoding.
To decode the bits for value number i in a sequence of bits b, both indexed from 0, perform the following:
If the range is encoded as n bits per value, then the value is bits $b^{i\times n + n - 1 .. i \times n}$ — a simple multiplexing operation.
If the range is encoded using a trit, then each block contains 5 values (v0 to v4), each of which contains a trit (t0 to t4) and a corresponding LSB value (m0 to m4). The first bit of the packed block is bit $\left\lfloor {i\over 5} \right\rfloor \times (8 + 5 \times n)$ . The bits in the block are packed as shown in Table 155 (in this example, n is 4).
Table 155. ASTC trit-based packing
27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | ||||
T7 | m4 | T6 | T5 | m3 | T4 | ||||||||||
15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
m2 | T3 | T2 | m1 | T1 | T0 | m0 |
The five trits t0 to t4 are obtained by bit manipulations of the 8 bits T7..0 as follows:
if T[4:2] = 111 C = { T[7:5], T[1:0] }; t4 = t3 = 2 else C = T[4:0] if T[6:5] = 11 t4 = 2; t3 = T[7] else t4 = T[7]; t3 = T[6:5] if C[1:0] = 11 t2 = 2; t1 = C[4]; t0 = { C[3], C[2]&~C[3] } else if C[3:2] = 11 t2 = 2; t1 = 2; t0 = C[1:0] else t2 = C[4]; t1 = C[3:2]; t0 = { C[1], C[0]&~C[1] }
If the range is encoded using a quint, then each block contains 3 values (v0 to v2), each of which contains a quint (q0 to q2) and a corresponding LSB value (m0 to m2). The first bit of the packed block is bit $ \left\lfloor {i\over 3} \right\rfloor \times (7+3 \times n)$ .
The bits in the block are packed as described in Table 156 and Table 157 (in this example, n is 4).
The three quints q0 to q2 are obtained by bit manipulations of the 7 bits Q6..0 as follows:
if Q[2:1] = 11 and Q[6:5] = 00 q2 = { Q[0], Q[4]&~Q[0], Q[3]&~Q[0] }; q1 = q0 = 4 else if Q[2:1] = 11 q2 = 4; C = { Q[4:3], ~Q[6:5], Q[0] } else q2 = Q[6:5]; C = Q[4:0] if C[2:0] = 101 q1 = 4; q0 = C[4:3] else q1 = C[4:3]; q0 = C[2:0]
Both these procedures ensure a valid decoding for all 128 possible values (even though a few are duplicates). They can also be implemented efficiently in software using small tables.
Encoding methods are not specified here, although table-based mechanisms work well.
Each color endpoint is specified as a sequence of integers in a given range. These values are packed using integer sequence encoding, as a stream of bits stored from just above the configuration data, and growing upwards.
Once unpacked, the values must be unquantized from their storage range, returning them to a standard range of 0..255.
For bit-only representations, this is simple bit replication from the most significant bit of the value.
For trit or quint-based representations, this involves a set of bit manipulations and adjustments to avoid the expense of full-width multipliers. This procedure ensures correct scaling, but scrambles the order of the decoded values relative to the encoded values. This must be compensated for using a table in the encoder.
The initial inputs to the procedure are denoted A
(9 bits), B
(9 bits), C
(9 bits) and D
(3 bits), and are decoded using the
range as described in Table 158.
Table 158. ASTC color unquantization parameters
Range | #Trits | #Quints | #Bits | Bit layout | A | B | C | D |
---|---|---|---|---|---|---|---|---|
0..5 | 1 | 1 | a | aaaaaaaaa | 000000000 | 204 | Trit value | |
0..9 | 1 | 1 | a | aaaaaaaaa | 000000000 | 113 | Quint value | |
0..11 | 1 | 2 | ba | aaaaaaaaa | b000b0bb0 | 93 | Trit value | |
0..19 | 1 | 2 | ba | aaaaaaaaa | b0000bb00 | 54 | Quint value | |
0..23 | 1 | 3 | cba | aaaaaaaaa | cb000cbcb | 44 | Trit value | |
0..39 | 1 | 3 | cba | aaaaaaaaa | cb0000cbc | 26 | Quint value | |
0..47 | 1 | 4 | dcba | aaaaaaaaa | dcb000dcb | 22 | Trit value | |
0..79 | 1 | 4 | dcba | aaaaaaaaa | dcb0000dc | 13 | Quint value | |
0..95 | 1 | 5 | edcba | aaaaaaaaa | edcb000ed | 11 | Trit value | |
0..159 | 1 | 5 | edcba | aaaaaaaaa | edcb0000e | 6 | Quint value | |
0..191 | 1 | 6 | fedcba | aaaaaaaaa | fedcb000f | 5 | Trit value |
These are then processed as follows:
unq = D * C + B; unq = unq ^ A; unq = (A & 0x80) | (unq >> 2);
Note that the multiply in the first line is nearly trivial as it only needs to multiply by 0, 1, 2, 3 or 4.
The decoding method used depends on the Color Endpoint Mode (CEM) field, which specifies how many values are used to represent the endpoint.
The CEM field also specifies how to take the n unquantized color endpoint values v0 to vn-1 and convert them into two RGBA color endpoints e0 and e1.
The HDR Modes are more complex and do not fit neatly into this section. They are documented in following section.
The methods can be summarized as shown in Table 159.
Table 159. ASTC LDR color endpoint modes
CEM | Range | Description | n |
---|---|---|---|
0 | LDR | Luminance, direct | 2 |
1 | LDR | Luminance, base+offset | 2 |
2 | HDR | Luminance, large range | 2 |
3 | HDR | Luminance, small range | 2 |
4 | LDR | Luminance+Alpha, direct | 4 |
5 | LDR | Luminance+Alpha, base+offset | 4 |
6 | LDR | RGB, base+scale | 4 |
7 | HDR | RGB, base+scale | 4 |
8 | LDR | RGB, direct | 6 |
9 | LDR | RGB, base+offset | 6 |
10 | LDR | RGB, base+scale plus two A | 6 |
11 | HDR | RGB | 6 |
12 | LDR | RGBA, direct | 8 |
13 | LDR | RGBA, base+offset | 8 |
14 | HDR | RGB + LDR Alpha | 8 |
15 | HDR | RGB + HDR Alpha | 8 |
Mode 14 is special in that the alpha values are interpolated linearly, but the color components are interpolated logarithmically. This is the only endpoint format with mixed-mode operation, and will return the error value if encountered in LDR mode.
Decode the different LDR endpoint modes as follows:
L0 = (v0>>2)|(v1&0xC0); L1=L0+(v1&0x3F); if (L1>0xFF) { L1=0xFF; } e0=(L0,L0,L0,0xFF); e1=(L1,L1,L1,0xFF);
bit_transfer_signed(v1,v0); bit_transfer_signed(v3,v2); e0=(v0,v0,v0,v2); e1=(v0+v1,v0+v1,v0+v1,v2+v3); clamp_unorm8(e0); clamp_unorm8(e1);
s0= v0+v2+v4; s1= v1+v3+v5; if (s1>=s0){e0=(v0,v2,v4,0xFF); e1=(v1,v3,v5,0xFF); } else { e0=blue_contract(v1,v3,v5,0xFF); e1=blue_contract(v0,v2,v4,0xFF); }
bit_transfer_signed(v1,v0); bit_transfer_signed(v3,v2); bit_transfer_signed(v5,v4); if(v1+v3+v5 >= 0) { e0=(v0,v2,v4,0xFF); e1=(v0+v1,v2+v3,v4+v5,0xFF); } else { e0=blue_contract(v0+v1,v2+v3,v4+v5,0xFF); e1=blue_contract(v0,v2,v4,0xFF); } clamp_unorm8(e0); clamp_unorm8(e1);
e0=(v0*v3>>8,v1*v3>>8,v2*v3>>8, v4); e1=(v0,v1,v2, v5);
s0= v0+v2+v4; s1= v1+v3+v5; if (s1>=s0){e0=(v0,v2,v4,v6); e1=(v1,v3,v5,v7); } else { e0=blue_contract(v1,v3,v5,v7); e1=blue_contract(v0,v2,v4,v6); }
bit_transfer_signed(v1,v0); bit_transfer_signed(v3,v2); bit_transfer_signed(v5,v4); bit_transfer_signed(v7,v6); if(v1+v3+v5>=0) { e0=(v0,v2,v4,v6); e1=(v0+v1,v2+v3,v4+v5,v6+v7); } else { e0=blue_contract(v0+v1,v2+v3,v4+v5,v6+v7); e1=blue_contract(v0,v2,v4,v6); } clamp_unorm8(e0); clamp_unorm8(e1);
The bit_transfer_signed()
procedure transfers a bit from one value (a)
to another (b). Initially, both a and b are in the range 0..255.
After calling this procedure, a's range becomes -32..31, and b remains
in the range 0..255. Note that, as is often the case, this is easier to
express in hardware than in C:
bit_transfer_signed(int& a, int& b) { b >>= 1; b |= a & 0x80; a >>= 1; a &= 0x3F; if( (a&0x20)!=0 ) a-=0x40; }
The blue_contract()
procedure is used to give additional precision to
RGB colors near gray:
color blue_contract( int r, int g, int b, int a ) { color c; c.r = (r+b) >> 1; c.g = (g+b) >> 1; c.b = b; c.a = a; return c; }
The clamp_unorm8()
procedure is used to clamp a color into
8-bit unsigned normalized fixed-point range:
void clamp_unorm8(color c) { if(c.r < 0) {c.r=0;} else if(c.r > 255) {c.r=255;} if(c.g < 0) {c.g=0;} else if(c.g > 255) {c.g=255;} if(c.b < 0) {c.b=0;} else if(c.b > 255) {c.b=255;} if(c.a < 0) {c.a=0;} else if(c.a > 255) {c.a=255;} }
For HDR endpoint modes, color values are represented in a 12-bit pseudo-logarithmic representation.
Mode 2 represents luminance-only data with a large range. It encodes using two values (v0, v1). The complete decoding procedure is as follows:
if(v1 >= v0) { y0 = (v0 << 4); y1 = (v1 << 4); } else { y0 = (v1 << 4) + 8; y1 = (v0 << 4) - 8; } // Construct RGBA result (0x780 is 1.0f) e0 = (y0, y0, y0, 0x780); e1 = (y1, y1, y1, 0x780);
Mode 3 represents luminance-only data with a small range. It packs the bits for a base luminance value, together with an offset, into two values (v0, v1), according to Table 160.
The bit field marked as X allocates different bits to L or d depending on the value of the mode bit M.
The complete decoding procedure is as follows:
// Check mode bit and extract. if((v0&0x80) !=0) { y0 = ((v1 & 0xE0) << 4) | ((v0 & 0x7F) << 2); d = (v1 & 0x1F) << 2; } else { y0 = ((v1 & 0xF0) << 4) | ((v0 & 0x7F) << 1); d = (v1 & 0x0F) << 1; } // Add delta and clamp y1 = y0 + d; if(y1 > 0xFFF) { y1 = 0xFFF; } // Construct RGBA result (0x780 is 1.0f) e0 = (y0, y0, y0, 0x780); e1 = (y1, y1, y1, 0x780);
Mode 7 packs the bits for a base RGB value, a scale factor, and some mode bits into the four values (v0, v1, v2, v3), as shown in Table 161.
Table 161. ASTC HDR mode 7 value layout
Value | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
v0 | M3..2 | R5..0 | ||||||
v1 | M1 | X0 | X1 | G4..0 | ||||
v2 | M0 | X2 | X3 | B4..0 | ||||
v3 | X4 | X5 | X6 | S4..0 |
The mode bits M0 to M3 are a packed representation of an endpoint bit mode, together with the major component index. For modes 0 to 4, the component (red, green, or blue) with the largest magnitude is identified, and the values swizzled to ensure that it is decoded from the red channel.
The endpoint bit mode is used to determine the number of bits assigned to each component of the endpoint, and the destination of each of the extra bits X0 to X6, as shown in Table 162.
Table 162. ASTC HDR mode 7 endpoint bit mode
Number of bits | Destination of extra bits | |||||||||||
Mode | R | G | B | S | X0 | X1 | X2 | X3 | X4 | X5 | X6 | |
0 | 11 | 5 | 5 | 7 | R9 | R8 | R7 | R10 | R6 | S6 | S5 | |
1 | 11 | 6 | 6 | 5 | R8 | G5 | R7 | B5 | R6 | R10 | R9 | |
2 | 10 | 5 | 5 | 8 | R9 | R8 | R7 | R6 | S7 | S6 | S5 | |
3 | 9 | 6 | 6 | 7 | R8 | G5 | R7 | B5 | R6 | S6 | S5 | |
4 | 8 | 7 | 7 | 6 | G6 | G5 | B6 | B5 | R6 | R7 | S5 | |
5 | 7 | 7 | 7 | 7 | G6 | G5 | B6 | B5 | R6 | S6 | S5 |
As noted before, this appears complex when expressed in C, but much easier to achieve in hardware: bit masking, extraction, shifting and assignment usually ends up as a single wire or multiplexer.
The complete decoding procedure is as follows:
// Extract mode bits and unpack to major component and mode. int majcomp; int mode; int modeval = ((v0&0xC0)>>6) | ((v1&0x80)>>5) | ((v2&0x80)>>4); if( (modeval & 0xC ) != 0xC ) { majcomp = modeval >> 2; mode = modeval & 3; } else if( modeval != 0xF ) { majcomp = modeval & 3; mode = 4; } else { majcomp = 0; mode = 5; } // Extract low-order bits of r, g, b, and s. int red = v0 & 0x3f; int green = v1 & 0x1f; int blue = v2 & 0x1f; int scale = v3 & 0x1f; // Extract high-order bits, which may be assigned depending on mode int x0 = (v1 >> 6) & 1; int x1 = (v1 >> 5) & 1; int x2 = (v2 >> 6) & 1; int x3 = (v2 >> 5) & 1; int x4 = (v3 >> 7) & 1; int x5 = (v3 >> 6) & 1; int x6 = (v3 >> 5) & 1; // Now move the high-order xs into the right place. int ohm = 1 << mode; if( ohm & 0x30 ) green |= x0 << 6; if( ohm & 0x3A ) green |= x1 << 5; if( ohm & 0x30 ) blue |= x2 << 6; if( ohm & 0x3A ) blue |= x3 << 5; if( ohm & 0x3D ) scale |= x6 << 5; if( ohm & 0x2D ) scale |= x5 << 6; if( ohm & 0x04 ) scale |= x4 << 7; if( ohm & 0x3B ) red |= x4 << 6; if( ohm & 0x04 ) red |= x3 << 6; if( ohm & 0x10 ) red |= x5 << 7; if( ohm & 0x0F ) red |= x2 << 7; if( ohm & 0x05 ) red |= x1 << 8; if( ohm & 0x0A ) red |= x0 << 8; if( ohm & 0x05 ) red |= x0 << 9; if( ohm & 0x02 ) red |= x6 << 9; if( ohm & 0x01 ) red |= x3 << 10; if( ohm & 0x02 ) red |= x5 << 10; // Shift the bits to the top of the 12-bit result. static const int shamts[6] = { 1,1,2,3,4,5 }; int shamt = shamts[mode]; red <<= shamt; green <<= shamt; blue <<= shamt; scale <<= shamt; // Minor components are stored as differences if( mode != 5 ) { green = red - green; blue = red - blue; } // Swizzle major component into place if( majcomp == 1 ) swap( red, green ); if( majcomp == 2 ) swap( red, blue ); // Clamp output values, set alpha to 1.0 e1.r = clamp( red, 0, 0xFFF ); e1.g = clamp( green, 0, 0xFFF ); e1.b = clamp( blue, 0, 0xFFF ); e1.alpha = 0x780; e0.r = clamp( red - scale, 0, 0xFFF ); e0.g = clamp( green - scale, 0, 0xFFF ); e0.b = clamp( blue - scale, 0, 0xFFF ); e0.alpha = 0x780;
Mode 11 specifies two RGB values, which it calculates from a number of bitfields (a, b0, b1, c, d0 and d1) which are packed together with some mode bits into the six values (v0, v1, v2, v3, v4, v5) as shown in Table 163.
Table 163. ASTC HDR mode 11 value layout
Value | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
v0 | a7..0 | |||||||
v1 | m0 | a8 | c5..0 | |||||
v2 | m1 | X0 | b05..0 | |||||
v3 | m2 | X1 | b15..0 | |||||
v4 | mj0 | X2 | X4 | d04..0 | ||||
v5 | mj1 | X3 | X5 | d14..0 |
If the major component bits mj1..0 are both 1, then the RGB values are specified directly by Table 164.
Table 164. ASTC HDR mode 11 direct value layout
Value | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
v0 | R011..4 | |||||||
v1 | R111..4 | |||||||
v2 | G011..4 | |||||||
v3 | G111..4 | |||||||
v4 | 1 | B011..5 | ||||||
v5 | 1 | B111..5 |
The mode bits m2..0 specify the bit allocation for the different values, and the destinations of the extra bits X0 to X5 as shown in Table 165.
Table 165. ASTC HDR mode 11 endpoint bit mode
Number of bits | Destination of extra bits | ||||||||||
Mode | a | b | c | d | X0 | X1 | X2 | X3 | X4 | X5 | |
0 | 9 | 7 | 6 | 7 | b06 | b16 | d06 | d16 | d05 | d15 | |
1 | 9 | 8 | 6 | 6 | b06 | b16 | b07 | b17 | d05 | d15 | |
2 | 10 | 6 | 7 | 7 | a9 | c6 | d06 | d16 | d05 | d15 | |
3 | 10 | 7 | 7 | 6 | b06 | b16 | a9 | c6 | d05 | d15 | |
4 | 11 | 8 | 6 | 5 | b06 | b16 | b07 | b17 | a9 | a10 | |
5 | 11 | 6 | 7 | 6 | a9 | a10 | c7 | c6 | d05 | d15 | |
6 | 12 | 7 | 7 | 5 | b06 | b16 | a11 | c6 | a9 | a10 | |
7 | 12 | 6 | 7 | 6 | a9 | a10 | a11 | c6 | d05 | d15 |
The complete decoding procedure is as follows:
// Find major component int majcomp = ((v4 & 0x80) >> 7) | ((v5 & 0x80) >> 6); // Deal with simple case first if( majcomp == 3 ) { e0 = (v0 << 4, v2 << 4, (v4 & 0x7f) << 5, 0x780); e1 = (v1 << 4, v3 << 4, (v5 & 0x7f) << 5, 0x780); return; } // Decode mode, parameters. int mode = ((v1&0x80)>>7) | ((v2&0x80)>>6) | ((v3&0x80)>>5); int va = v0 | ((v1 & 0x40) << 2); int vb0 = v2 & 0x3f; int vb1 = v3 & 0x3f; int vc = v1 & 0x3f; int vd0 = v4 & 0x7f; int vd1 = v5 & 0x7f; // Assign top bits of vd0, vd1. static const int dbitstab[8] = {7,6,7,6,5,6,5,6}; vd0 = signextend( vd0, dbitstab[mode] ); vd1 = signextend( vd1, dbitstab[mode] ); // Extract and place extra bits int x0 = (v2 >> 6) & 1; int x1 = (v3 >> 6) & 1; int x2 = (v4 >> 6) & 1; int x3 = (v5 >> 6) & 1; int x4 = (v4 >> 5) & 1; int x5 = (v5 >> 5) & 1; int ohm = 1 << mode; if( ohm & 0xA4 ) va |= x0 << 9; if( ohm & 0x08 ) va |= x2 << 9; if( ohm & 0x50 ) va |= x4 << 9; if( ohm & 0x50 ) va |= x5 << 10; if( ohm & 0xA0 ) va |= x1 << 10; if( ohm & 0xC0 ) va |= x2 << 11; if( ohm & 0x04 ) vc |= x1 << 6; if( ohm & 0xE8 ) vc |= x3 << 6; if( ohm & 0x20 ) vc |= x2 << 7; if( ohm & 0x5B ) vb0 |= x0 << 6; if( ohm & 0x5B ) vb1 |= x1 << 6; if( ohm & 0x12 ) vb0 |= x2 << 7; if( ohm & 0x12 ) vb1 |= x3 << 7; // Now shift up so that major component is at top of 12-bit value int shamt = (modeval >> 1) ^ 3; va <<= shamt; vb0 <<= shamt; vb1 <<= shamt; vc <<= shamt; vd0 <<= shamt; vd1 <<= shamt; e1.r = clamp( va, 0, 0xFFF ); e1.g = clamp( va - vb0, 0, 0xFFF ); e1.b = clamp( va - vb1, 0, 0xFFF ); e1.alpha = 0x780; e0.r = clamp( va - vc, 0, 0xFFF ); e0.g = clamp( va - vb0 - vc - vd0, 0, 0xFFF ); e0.b = clamp( va - vb1 - vc - vd1, 0, 0xFFF ); e0.alpha = 0x780; if( majcomp == 1 ) { swap( e0.r, e0.g ); swap( e1.r, e1.g ); } else if( majcomp == 2 ) { swap( e0.r, e0.b ); swap( e1.r, e1.b ); }
Mode 14 specifies two RGBA values, using the eight values (v0, v1, v2, v3, v4, v5, v6, v7). First, the RGB values are decoded from (v0..v5) using the method from Mode 11, then the alpha values are filled in from v6 and v7:
// Decode RGB as for mode 11 (e0,e1) = decode_mode_11(v0,v1,v2,v3,v4,v5) // Now fill in the alphas e0.alpha = v6; e1.alpha = v7;
Note that in this mode, the alpha values are interpreted (and interpolated) as 8-bit unsigned normalized values, as in the LDR modes. This is the only mode that exhibits this behavior.
Mode 15 specifies two RGBA values, using the eight values (v0, v1, v2, v3, v4, v5, v6, v7). First, the RGB values are decoded from (v0..v5) using the method from Mode 11. The alpha values are stored in values v6 and v7 as a mode and two values which are interpreted according to the mode M, as shown in Table 166.
The alpha values are decoded from v6 and v7 as follows:
// Decode RGB as for mode 11 (e0,e1) = decode_mode_11(v0,v1,v2,v3,v4,v5) // Extract mode bits mode = ((v6 >> 7) & 1) | ((v7 >> 6) & 2); v6 &= 0x7F; v7 &= 0x7F; if(mode==3) { // Directly specify alphas e0.alpha = v6 << 5; e1.alpha = v7 << 5; } else { // Transfer bits from v7 to v6 and sign extend v7. v6 |= (v7 << (mode+1))) & 0x780; v7 &= (0x3F >> mode); v7 ^= 0x20 >> mode; v7 -= 0x20 >> mode; v6 <<= (4-mode); v7 <<= (4-mode); // Add delta and clamp v7 += v6; v7 = clamp(v7, 0, 0xFFF); e0.alpha = v6; e1.alpha = v7; }
Note that in this mode, the alpha values are interpreted (and interpolated) as 12-bit HDR values, and are interpolated as for any other HDR component.
The weight information is stored as a stream of bits, growing downwards from the most significant bit in the block. Bit n in the stream is thus bit 127-n in the block.
For each location in the weight grid, a value (in the specified range) is packed into the stream. These are ordered in a raster pattern starting from location (0,0,0), with the X dimension increasing fastest, and the Z dimension increasing slowest. If dual-plane mode is selected, both weights are emitted together for each location, plane 0 first, then plane 1.
Each weight plane is specified as a sequence of integers in a given range. These values are packed using integer sequence encoding.
Once unpacked, the values must be unquantized from their storage range, returning them to a standard range of 0..64. The procedure for doing so is similar to the color endpoint unquantization.
First, we unquantize the actual stored weight values to the range 0..63.
For bit-only representations, this is simple bit replication from the most significant bit of the value.
For trit or quint-based representations, this involves a set of bit manipulations and adjustments to avoid the expense of full-width multipliers.
For representations with no additional bits, the results are as shown in Table 167.
For other values, we calculate the initial inputs to a bit manipulation
procedure. These are denoted A
(7 bits), B
(7 bits), C
(7 bits),
and D
(3 bits) and are decoded using the range as shown in
Table 168.
Table 168. ASTC weight unquantization parameters
Range | #Trits | #Quints | #Bits | Bit layout | A | B | C | D |
---|---|---|---|---|---|---|---|---|
0..5 | 1 | 1 | a | aaaaaaa | 0000000 | 50 | Trit value | |
0..9 | 1 | 1 | a | aaaaaaa | 0000000 | 28 | Quint value | |
0..11 | 1 | 2 | ba | aaaaaaa | b000b0b | 23 | Trit value | |
0..19 | 1 | 2 | ba | aaaaaaa | b0000b0 | 13 | Quint value | |
0..23 | 1 | 3 | cba | aaaaaaa | cb000cb | 11 | Trit value |
These are then processed as follows:
unq = D * C + B; unq = unq ^ A; unq = (A & 0x20) | (unq >> 2);
Note that the multiply in the first line is nearly trivial as it only needs to multiply by 0, 1, 2, 3 or 4.
As a final step, for all types of value, the range is expanded from 0..63 up to 0..64 as follows:
if (unq > 32) { unq += 1; }
This allows the implementation to use 64 as a divisor during interpolation, which is much easier than using 63.
After unquantization, the weights are subject to weight selection and infill. The infill method is used to calculate the weight for a texel position, based on the weights in the stored weight grid array (which may be a different size). The procedure below must be followed exactly, to ensure bit exact results.
The block size is specified as three dimensions along the s, t and r axes (Bs, Bt, Br). Texel coordinates within the block (bs, bt, br) can have values from 0 to one less than the block dimension in that axis. For each block dimension, we compute scale factors (Ds, Dt, Dr):
Since the block dimensions are constrained, these are easily looked up in a table. These scale factors are then used to scale the (bs, bt, br) coordinates to a homogeneous coordinate (cs, ct, cr):
cs = Ds * bs; ct = Dt * bt; cr = Dr * br;
This homogeneous coordinate (cs, ct, cr) is then scaled again to give a coordinate (gs, gt, gr) in the weight-grid space. The weight-grid is of size (Wwidth, Wheight, Wdepth), as specified in the block mode field (Table 147 and Table 148):
gs = (cs*(Wwidth-1)+32) >> 6; gt = (ct*(Wheight-1)+32) >> 6; gr = (cr*(Wdepth-1)+32) >> 6;
The resulting coordinates may be in the range 0..176. These are interpreted as 4:4 unsigned fixed point numbers in the range 0.0 .. 11.0.
If we label the integral parts of these (js, jt, jr) and the fractional parts (fs, ft, fr), then:
js = gs >> 4; fs = gs & 0x0F; jt = gt >> 4; ft = gt & 0x0F; jr = gr >> 4; fr = gr & 0x0F;
These values are then used to interpolate between the stored weights. This process differs for 2D and 3D.
For 2D, bilinear interpolation is used:
v0 = js + jt*Wwidth; p00 = decode_weight(v0); p01 = decode_weight(v0 + 1); p10 = decode_weight(v0 + Wwidth); p11 = decode_weight(v0 + Wwidth + 1);
The function decode_weight(n)
decodes the nth weight in the stored
weight stream.
The values p00 to p11 are the weights at the corner of the square
in which the texel position resides. These are then weighted using the
fractional position to produce the effective weight i as follows:
w11 = (fs*ft+8) >> 4; w10 = ft - w11; w01 = fs - w11; w00 = 16 - fs - ft + w11; i = (p00*w00 + p01*w01 + p10*w10 + p11*w11 + 8) >> 4;
For 3D, simplex interpolation is used as it is cheaper than a naïve trilinear interpolation. First, we pick some parameters for the interpolation based on comparisons of the fractional parts of the texel position as shown in Table 169.
Table 169. ASTC simplex interpolation parameters
fs > ft | ft > fr | fs > fr | s1 | s2 | w0 | w1 | w2 | w3 |
---|---|---|---|---|---|---|---|---|
True | True | True | 1 | Wwidth | 16 - fs | fs - ft | ft - fr | fr |
False | True | True | Wwidth | 1 | 16 - ft | ft - fs | fs - fr | fr |
True | False | True | 1 | Wwidth × Wheight | 16 - fs | fs - fr | fr - ft | ft |
True | False | False | Wwidth × Wheight | 1 | 16 - fr | fr - fs | fs - ft | ft |
False | True | False | Wwidth | Wwidth × Wheight | 16 - ft | ft - fr | fr - fs | fs |
False | False | False | Wwidth × Wheight | Wwidth | 16 - fr | fr - ft | ft - fs | fs |
Italicized test results are implied by the others. The effective weight i is then calculated as:
v0 = js + jt*Wwidth + jr*Wwidth*Wheight; p0 = decode_index(v0); p1 = decode_index(v0 + s1); p2 = decode_index(v0 + s1 + s2); p3 = decode_index(v0 + Wwidth*Wheight + Wwidth + 1); i = (p0*w0 + p1*w1 + p2*w2 + p3*w3 + 8) >> 4;
Once the effective weight i for the texel has been calculated, the color endpoints are interpolated and expanded.
For LDR endpoint modes, each color component C is calculated from the corresponding 8-bit endpoint components C0 and C1 as follows:
If sRGB conversion is not enabled, or for the alpha channel in any case, C0 and C1 are first expanded to 16 bits by bit replication:
C0 = (C0 << 8) | C0; C1 = (C1 << 8) | C1;
If sRGB conversion is enabled, C0 and C1 for the R, G, and B channels are expanded to 16 bits differently, as follows:
C0 = (C0 << 8) | 0x80; C1 = (C1 << 8) | 0x80;
C0 and C1 are then interpolated to produce a UNORM16 result C:
C = floor( (C0*(64-i) + C1*i + 32)/64 )
If sRGB conversion is not enabled and the decoding mode is decode_float16, then if C = 65535 the final result is 1.0 (0x3C00); otherwise C is divided by 65536 and the infinite-precision result of the division is converted to FP16 with round-to-zero semantics.
If sRGB conversion is not enabled and the decoding mode is decode_unorm8, then the top 8 bits of the interpolation result for the R, G, B and A channels are used as the final result.
If sRGB conversion is not enabled and the decoding mode is decode_rgb9e5, then the final result is a combination of the (UNORM16) values of C for the three color components (Cr, Cg and Cb) computed as follows:
int lz = clz17(Cr | Cg | Cb | 1); if (Cr == 65535) { Cr = 65536; lz = 0; } if (Cg == 65535) { Cg = 65536; lz = 0; } if (Cb == 65535) { Cb = 65536; lz = 0; } Cr <<= lz; Cg <<= lz; Cb <<= lz; Cr = (Cr >> 8) & 0x1FF; Cg = (Cg >> 8) & 0x1FF; Cb = (Cb >> 8) & 0x1FF; uint32_t exponent = 16 - lz; uint32_t texel = (exponent << 27) | (Cb << 18) | Cg << 9) | Cr;
The clz17()
function counts leading zeroes in a 17-bit value.
if sRGB conversion is enabled, then the decoding mode is ignored and the top 8 bits of the interpolation result for the R, G and B channels are passed to the external sRGB conversion block and used as the final result. The A channel uses the decode_float16 decoding mode.
For HDR endpoint modes, color values are represented in a 12-bit pseudo-logarithmic representation, and interpolation occurs in a piecewise-approximate logarithmic manner as follows:
In LDR mode, the error result is returned.
In HDR mode, the color components from each endpoint, C0 and C1, are initially shifted left 4 bits to become 16-bit integer values and these are interpolated in the same way as LDR. The 16-bit value C is then decomposed into the top five bits, E, and the bottom 11 bits M, which are then processed and recombined with E to form the final value Cf:
C = floor( (C0*(64-i) + C1*i + 32)/64 ) E = (C & 0xF800) >> 11; M = C & 0x7FF; if (M < 512) { Mt = 3*M; } else if (M >= 1536) { Mt = 5*M - 2048; } else { Mt = 4*M - 512; } Cf = (E<<10) + (Mt>>3);
This interpolation is a considerably closer approximation to a logarithmic space than simple 16-bit interpolation.
This final value Cf is interpreted as an IEEE FP16 value. If the result is +Inf or NaN, it is converted to the bit pattern 0x7BFF, which is the largest representable finite value.
If the decoding mode is decode_rgb9e5, then the final result is a combination fo the (IEEE FP16) values of Cf for the three color components (Cr, Cg and Cb) computed as follows:
if( Cr > 0x7c00 ) Cr = 0; else if( Cr == 0x7c00 ) Cr = 0x7bff; if( Cg > 0x7c00 ) Cg = 0; else if( Cg == 0x7c00 ) Cg = 0x7bff; if( Cb > 0x7c00 ) Cb = 0; else if( Cb == 0x7c00 ) Cb = 0x7bff; int Re = (Cr >> 10) & 0x1F; int Ge = (Cg >> 10) & 0x1F; int Be = (Cb >> 10) & 0x1F; int Rex = Re == 0 ? 1 : Re; int Gex = Ge == 0 ? 1 : Ge; int Bex = Be == 0 ? 1 : Be; int Xm = ((Cr | Cg | Cb) & 0x200) >> 9; int Xe = Re | Ge | Be; uint32_t rshift, gshift, bshift, expo; if (Xe == 0) { expo = rshift = gshift = bshift = Xm; } else if (Re >= Ge && Re >= Be) { expo = Rex + 1; rshift = 2; gshift = Rex - Gex + 2; bshift = Rex - Bex + 2; } else if (Ge >= Be) { expo = Gex + 1; rshift = Gex - Rex + 2; gshift = 2; bshift = Gex - Bex + 2; } else { expo = Bex + 1; rshift = Bex - Rex + 2; gshift = Bex - Gex + 2; bshift = 2; } int Rm = (Cr & 0x3FF) | (Re == 0 ? 0 : 0x400); int Gm = (Cg & 0x3FF) | (Ge == 0 ? 0 : 0x400); int Bm = (Cb & 0x3FF) | (Be == 0 ? 0 : 0x400); Rm = (Rm >> rshift) & 0x1FF; Gm = (Gm >> gshift) & 0x1FF; Bm = (Bm >> bshift) & 0x1FF; uint32_t texel = (expo << 27) | (Bm << 18) | (Gm << 9) | (Rm << 0);
If dual-plane mode is disabled, all of the endpoint components are interpolated using the same weight value.
If dual-plane mode is enabled, two weights are stored with each texel. One component is then selected to use the second weight for interpolation, instead of the first weight. The first weight is then used for all other components.
The component to treat specially is indicated using the 2-bit Color Component Selector (CCS) field as shown in Table 170.
Table 170. ASTC dual plane color component selector values
Value | Weight 0 | Weight 1 |
---|---|---|
0 | GBA | R |
1 | RBA | G |
2 | RGA | B |
3 | RGB | A |
The CCS bits are stored at a variable position directly below the weight bits and any additional CEM bits.
When multiple partitions are active, each texel position is assigned a
partition index. This partition index is calculated using a seed (the
partition pattern index), the texel’s x, y, z position within
the block, and the number of partitions.
An additional argument, small_block
, is set to 1 if the number of
texels in the block is less than 31, otherwise it is set to 0.
This function is specified in terms of x, y and z in order to support 3D textures. For 2D textures and texture slices, z will always be 0.
The full partition selection algorithm is as follows:
int select_partition(int seed, int x, int y, int z, int partitioncount, int small_block) { if( small_block ){ x <<= 1; y <<= 1; z <<= 1; } seed += (partitioncount-1) * 1024; uint32_t rnum = hash52(seed); uint8_t seed1 = rnum & 0xF; uint8_t seed2 = (rnum >> 4) & 0xF; uint8_t seed3 = (rnum >> 8) & 0xF; uint8_t seed4 = (rnum >> 12) & 0xF; uint8_t seed5 = (rnum >> 16) & 0xF; uint8_t seed6 = (rnum >> 20) & 0xF; uint8_t seed7 = (rnum >> 24) & 0xF; uint8_t seed8 = (rnum >> 28) & 0xF; uint8_t seed9 = (rnum >> 18) & 0xF; uint8_t seed10 = (rnum >> 22) & 0xF; uint8_t seed11 = (rnum >> 26) & 0xF; uint8_t seed12 = ((rnum >> 30) | (rnum << 2)) & 0xF; seed1 *= seed1; seed2 *= seed2; seed3 *= seed3; seed4 *= seed4; seed5 *= seed5; seed6 *= seed6; seed7 *= seed7; seed8 *= seed8; seed9 *= seed9; seed10 *= seed10; seed11 *= seed11; seed12 *= seed12; int sh1, sh2, sh3; if( seed & 1 ) { sh1 = (seed&2 ? 4:5); sh2 = (partitioncount==3 ? 6:5); } else { sh1 = (partitioncount==3 ? 6:5); sh2 = (seed&2 ? 4:5); } sh3 = (seed & 0x10) ? sh1 : sh2: seed1 >>= sh1; seed2 >>= sh2; seed3 >>= sh1; seed4 >>= sh2; seed5 >>= sh1; seed6 >>= sh2; seed7 >>= sh1; seed8 >>= sh2; seed9 >>= sh3; seed10 >>= sh3; seed11 >>= sh3; seed12 >>= sh3; int a = seed1*x + seed2*y + seed11*z + (rnum >> 14); int b = seed3*x + seed4*y + seed12*z + (rnum >> 10); int c = seed5*x + seed6*y + seed9 *z + (rnum >> 6); int d = seed7*x + seed8*y + seed10*z + (rnum >> 2); a &= 0x3F; b &= 0x3F; c &= 0x3F; d &= 0x3F; if( partitioncount < 4 ) d = 0; if( partitioncount < 3 ) c = 0; if( a >= b && a >= c && a >= d ) return 0; else if( b >= c && b >= d ) return 1; else if( c >= d ) return 2; else return 3; }
As has been observed before, the bit selections are much easier to express in hardware than in C.
The seed is expanded using a hash function hash52()
, which is defined
as follows:
uint32_t hash52( uint32_t p ) { p ^= p >> 15; p -= p << 17; p += p << 7; p += p << 4; p ^= p >> 5; p += p << 16; p ^= p >> 7; p ^= p >> 3; p ^= p << 6; p ^= p >> 17; return p; }
This assumes that all operations act on 32-bit values
The size of the data used to represent color endpoints is not explicitly specified. Instead, it is determined from the block mode and number of partitions as follows:
config_bits = 17; if(num_partitions>1) if(single_CEM) config_bits = 29; else config_bits = 25 + 3*num_partitions; num_weights = Wwidth * Wheight * Wdepth; // size of weight grid if(dual_plane) config_bits += 2; num_weights *= 2; weight_bits = ceil(num_weights*8*trits_in_weight_range/5) + ceil(num_weights*7*quints_in_weight_range/3) + num_weights*bits_in_weight_range; remaining_bits = 128 - config_bits - weight_bits; num_CEM_pairs = base_CEM_class+1 + count_bits(extra_CEM_bits);
The CEM value range is then looked up from a table indexed by remaining
bits and num_CEM_pairs
.
This table is initialized such that the range is as large as possible,
consistent with the constraint that the number of bits required to
encode num_CEM_pairs
pairs of values is not more than the number of
remaining bits.
An equivalent iterative algorithm would be:
num_CEM_values = num_CEM_pairs*2; for(range = each possible CEM range in descending order of size) { CEM_bits = ceil(num_CEM_values*8*trits_in_CEM_range/5) + ceil(num_CEM_values*7*quints_in_CEM_range/3) + num_CEM_values*bits_in_CEM_range; if(CEM_bits <= remaining_bits) break; } return range;
In cases where this procedure results in unallocated bits, these bits are not read by the decoding process and can have any value.
A void-extent block is a block encoded with a single color. It also specifies some additional information about the extent of the single-color area beyond this block, which can optionally be used by a decoder to reduce or prevent redundant block fetches.
In the HDR case, if the decoding mode is decode_rgb9e5, then any negative color component values are set to 0 before conversion to the shared exponent format (as described in Section 23.19).
The layout of a 2D Void-Extent block is as shown in Table 171.
Table 171. ASTC 2D void-extent block layout overview
127 | 126 | 125 | 124 | 123 | 122 | 121 | 120 | 119 | 118 | 117 | 116 | 115 | 114 | 113 | 112 |
Block color A component15..0 | |||||||||||||||
111 | 110 | 109 | 108 | 107 | 106 | 105 | 104 | 103 | 102 | 101 | 100 | 99 | 98 | 97 | 96 |
Block color B component15..0 | |||||||||||||||
95 | 94 | 93 | 92 | 91 | 90 | 89 | 88 | 87 | 86 | 85 | 84 | 83 | 82 | 81 | 80 |
Block color G component15..0 | |||||||||||||||
79 | 78 | 77 | 76 | 75 | 74 | 73 | 72 | 71 | 70 | 69 | 68 | 67 | 66 | 65 | 64 |
Block color R component15..0 | |||||||||||||||
63 | 62 | 61 | 60 | 59 | 58 | 57 | 56 | 55 | 54 | 53 | 52 | 51 | 50 | 49 | 48 |
Void-extent maximum t coordinate12..0 | Min t coord12..10 | ||||||||||||||
47 | 46 | 45 | 44 | 43 | 42 | 41 | 40 | 39 | 38 | 37 | 36 | 35 | 34 | 33 | 32 |
Void-extent minimum t coordinate9..0 | Void-extent maximum s coordinate12..7 | ||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 |
Void-extent maximum s coordinate5..0 | Void-extent minimum s coordinate12..4 | ||||||||||||||
15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
Minimum s coordinate3..0 | 1 | 1 | D | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
The layout of a 3D Void-Extent block is as shown in Table 172.
Table 172. ASTC 3D void-extent block layout overview
127 | 126 | 125 | 124 | 123 | 122 | 121 | 120 | 119 | 118 | 117 | 116 | 115 | 114 | 113 | 112 |
Block color A component15..0 | |||||||||||||||
111 | 110 | 109 | 108 | 107 | 106 | 105 | 104 | 103 | 102 | 101 | 100 | 99 | 98 | 97 | 96 |
Block color B component15..0 | |||||||||||||||
95 | 94 | 93 | 92 | 91 | 90 | 89 | 88 | 87 | 86 | 85 | 84 | 83 | 82 | 81 | 80 |
Block color G component15..0 | |||||||||||||||
79 | 78 | 77 | 76 | 75 | 74 | 73 | 72 | 71 | 70 | 69 | 68 | 67 | 66 | 65 | 64 |
Block color R component15..0 | |||||||||||||||
63 | 62 | 61 | 60 | 59 | 58 | 57 | 56 | 55 | 54 | 53 | 52 | 51 | 50 | 49 | 48 |
Void-extent maximum r coordinate8..0 | Void-extent minimum r coordinate8..2 | ||||||||||||||
47 | 46 | 45 | 44 | 43 | 42 | 41 | 40 | 39 | 38 | 37 | 36 | 35 | 34 | 33 | 32 |
Min r coord1..0 | Void-extent maximum t coordinate8..0 | Void-extent min t coordinate8..4 | |||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 |
Minimum t coordinate3..0 | Void-extent minimum s coordinate8..0 | Min s coord8..6 | |||||||||||||
15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
Void-extent minimum s coordinate5..0 | D | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 0 | 0 |
Bit 9 is the Dynamic Range flag, which indicates the format in which colors are stored. A 0 value indicates LDR, in which case the color components are stored as UNORM16 values. A 1 indicates HDR, in which case the color components are stored as FP16 values.
The reason for the storage of UNORM16 values in the LDR case is due to the possibility that the value will need to be passed on to sRGB conversion. By storing the color value in the format which comes out of the interpolator, before the conversion to FP16, we avoid having to have separate versions for sRGB and linear modes.
If a void-extent block with HDR values is decoded in LDR mode, then the result will be the error color, opaque magenta, for all texels within the block.
In the HDR case, if the color component values are infinity or NaN, this will result in undefined behavior. As usual, this must not lead to an API’s interruption or termination.
Bits 10 and 11 are reserved and must be 1.
The minimum and maximum coordinate values are treated as unsigned integers and then normalized into the range 0..1 (by dividing by 213-1 or 29-1, for 2D and 3D respectively). The maximum values for each dimension must be greater than the corresponding minimum values, unless they are all all-1s.
If all the coordinates are all-1s, then the void extent is ignored, and the block is simply a constant-color block.
The existence of single-color blocks with void extents must not produce results different from those obtained if these single-color blocks are defined without void-extents. Any situation in which the results would differ is invalid. Results from invalid void extents are undefined.
If a void-extent appears in a MIPmap level other than the most detailed one, then the extent will apply to all of the more detailed levels too. This allows decoders to avoid sampling more detailed MIPmaps.
If the more detailed MIPmap level is not a constant color in this region, then the block may be marked as constant color, but without a void extent, as detailed above.
If a void-extent extends to the edge of a texture, then filtered texture colors may not be the same color as that specified in the block, due to texture border colors, wrapping, or cube face wrapping.
Care must be taken when updating or extracting partial image data that void-extents in the image do not become invalid.
In ASTC, there is a variety of ways to encode an illegal block. Decoders are required to recognize all illegal blocks and emit the standard error color value upon encountering an illegal block.
Here is a comprehensive list of situations that represent illegal block encodings:
Note also that, in LDR mode, a block which has both HDR and LDR endpoint modes assigned to different partitions is not an error block. Only those texels which belong to the HDR partition will result in the error color. Texels belonging to a LDR partition will be decoded as normal.
In order to ease verification and accelerate adoption, an LDR-only subset of the full ASTC specification has been made available.
Implementations of this LDR Profile must satisfy the following requirements:
The feature subset for the LDR profile is:
In order to ease verification and accelerate adoption, a second subset of the full ASTC specification has been made available, known as the HDR profile.
Implementations of the HDR Profile must satisfy the following requirements:
The feature subset for the HDR profile is:
This description is derived from the PVRTC Texture Compression User Guide, part of the PowerVR Documentation.
PVRTC is PowerVR’s family of proprietary texture compression schemes, providing compression of color data at 4 or 2 bits per pixel (4/2bpp).
There are two generations of PVRTC: PVRTC1 and PVRTC2. Both support 4bpp and 2bpp compression ratios. They are broadly similar, but PVRTC2 adds additional features to the format. Both primarily use an interpolation and modulation scheme to compress texture data wherein texel values are encoded as two low-resolution images, ALow and BLow, along with a full-resolution, low bit-precision modulation signal, MSig, to combine those images.
More information on the specifics of the PVRTC1 compression technique can be found in the “Graphics Hardware 2003” paper Texture Compression using Low-Frequency Signal Modulation.
In PVRTC, images are described in terms of 64-bit little-endian words, each of which contains a pixel from each of the low-resolution images, ALow and BLow, and a subset of the modulation data, MSig, corresponding to either a 4×4 or 8×4 set of pixels. Unlike traditional block-based formats, PVRTC uses adjacent data-words to reconstruct the original image.
If combined with this encoding scheme, the sRGB EOTF is applied to the unquantized version of the R′, G′ and B′ channels (that is, $\textit{R}_\textit{out} = \textit{EOTF}_{\textit{sRGB}}\left({\textit{R}'\over 255}\right)$ , etc.) at the end of the texel decode process, but the A channel is interpreted linearly.
For PVRTC1 4bpp, each 64-bit word WX,Y at block coordinates (X,Y) contains the modulation information for a 4×4 block of texels beginning at image pixels $(4\times\mathit{X}, 4\times\mathit{Y})$ , and a color sample for each low-resolution image which influences an overlapping 7×7 texel region of the final output, with each sample centered on output pixel $((4\times\mathit{X})+2, (4\times\mathit{Y})+2)$ . Nearly every texel location requires data from a set of 2×2 data words in order to be decoded: the low-resolution encoded images are bilinearly interpolated to, in effect, generate a pair of full-resolution images prior to modulation, and this interpolation requires examining multiple data words in the compressed image data. At least one word of PVRTC data must exist in each mip level of a texture. If a texture (or a particular mip-level) is smaller than the size of the block of texels described by one word in any dimension, reconstruction of an image occurs as if fetching the upper-left most texels from the region covered by those words. PVRTC1 images must have dimensions that are power of two values.
Each data word contains two colors and a set of modulation data describing how to interpolate between them for each pixel location in the original image. A flag M describes how the modulation data should be interpreted, and each color contains a bit to describe whether it contains alpha data.
Table 173. Texel Data format for PVRTC1 4bpp compressed texture formats
Bits 64..32: Color data and flags | |||||||||||||||||||||||||||||||
63 | 62 | 61 | 60 | 59 | 58 | 57 | 56 | 55 | 54 | 53 | 52 | 51 | 50 | 49 | 48 | 47 | 46 | 45 | 44 | 43 | 42 | 41 | 40 | 39 | 38 | 37 | 36 | 35 | 34 | 33 | 32 |
Color B | Color A | M | |||||||||||||||||||||||||||||
Bits 31..0: Modulation data bits [1..0] for pixel offset (x, y) | |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
3,3 | 2,3 | 1,3 | 0,3 | 3,2 | 2,2 | 1,2 | 0,2 | 3,1 | 2,1 | 1,1 | 0,1 | 3,0 | 2,0 | 1,0 | 0,0 |
Data words in PVRTC1 formats are stored in a reflected Morton order, as shown in the figure below, where each cell corresponds to the index of a 64-bit word WX,Y:
Expressing each of the X and Y indices as an array of bits, the index of a particular PVRTC word can be found by interleaving the bits of the X and Y indices as follows:
Let Xb be the number of bits used to express X — i.e. $\mathit{Xb} = \mathrm{log}_2\left(\left\lceil{\mathit{width}\over 4}\right\rceil\right)$ .
Let Yb be the number of bits used to express Y — i.e. $\mathit{Yb} = \mathrm{log}_2\left(\left\lceil{\mathit{height}\over 4}\right\rceil\right)$ .
Then:
That is, to form the word offset, bits of X and Y are interleaved, with bits from Y in the lower of each interleaved pair. These bits are interleaved up to the number of bits needed to represent the largest possible X or Y, for whichever of width and height is smaller. Any remaining bits from X or Y, for whichever of width and height is larger, are appended to the offset bit pattern.
For example, Figure 51 represents a 64×16-texel image represented by 16×4 = 64 words of 64 bits. The largest possible X value in this example is ${64\over 4} - 1$ = 15; the largest possible Y value in this example is ${16\over 4} - 1$ = 3. The bottom four bits of the word offset are composed by interleaving the bottom two bits of Y and the bottom two bits of X (with Y in the lowest bit). Bits 5..4 of the word offset are derived from bits 3..2 of X, since no further bits are required to represent Y.
In this case, the word at X=13, Y=2, W13,2=54 is constructed as follows:
Where wordWidth and wordHeight are the image width and height in units of 4×4 areas encoded by words:
the word offset for $\mathit{X} = \left\lfloor{\mathit{x}\over 4}\right\rfloor$ and $\mathit{Y} = \left\lfloor{\mathit{y}\over 4}\right\rfloor$ can be calculated iteratively as follows:
uint32_t reflectedMortonOffset(const uint32_t X, const uint32_t Y, const uint32_t wordWidth, const uint32_t wordHeight) { const uint32_t minDim = (wordWidth <= wordHeight) ? wordWidth : wordHeight; uint32_t offset = 0, shift = 0, mask; // Tests XY bounds AND that Width and Height != 0 assert(X < wordWidth && Y < wordHeight); // Must be (non-zero) powers of 2 assert((wordWidth & (wordWidth - 1)) == 0); assert((wordHeight & (wordHeight - 1)) == 0); for (mask = 1; mask < minDim; mask <<= 1) { offset |= (((Y & mask) | ((X & mask) << 1))) << shift; shift++; } // At least one of X or Y will have run out of MSBs offset |= ((X | Y) >> shift) << (shift * 2); return offset; }
Each data word encodes a color sample value from each of the two low-resolution images, with the (X, Y) location of the block corresponding to the (XLow, YLow) location of the colors in the low-resolution images. The image colors for a given pixel location (x, y) are reconstructed using the words containing the four nearest color samples: $W_{X_\textit{Low},Y_\textit{Low}}$ , $W_{X_\textit{Low}+1,Y_\textit{Low}}$ , $W_{X_\textit{Low},Y_\textit{Low}+1}$ and $W_{X_\textit{Low}+1,Y_\textit{Low}+1}$ , where XLow and YLow are derived as follows:
Figure 52 shows a grid of pixels with (x = 0, y = 0) at top left. Each word $W_{X_\mathit{Mod},Y_\mathit{Mod}}$ holds modulation values for a 4×4 texel region $\mathbf{M}_{\mathbf{X}_\mathbf{Mod},\mathbf{Y}_\mathbf{Mod}}$ , as described in Section 24.2.5, where $X_\mathit{Mod} = \left\lfloor{x\over 4}\right\rfloor$ and $Y_\mathit{Mod} = \left\lfloor{y\over 4}\right\rfloor$ . For $X_\mathit{Low} = \left\lfloor{{x - 2}\over 4}\right\rfloor$ and $Y_\mathit{Low} = \left\lfloor{{y - 2}\over 4}\right\rfloor$ , color reconstruction for the pixels shaded in Figure 52 requires data from the words $W_{X_\mathit{Low},Y_\mathit{Low}}$ through $W_{X_\mathit{Low}+1,Y_\mathit{Low}+1}$ ; the pixels for which these these words hold modulation values are shown as $\mathbf{M}_{\mathbf{X}_\mathbf{Mod},\mathbf{Y}_\mathbf{Mod}}$ through $\mathbf{M}_{\mathbf{X}_\mathbf{Mod}+1,\mathbf{Y}_\mathbf{Mod}+1}$ , outlined in red. All pixels within the region contained by the dashed outline have the same values for $X_\mathit{Low}$ and $Y_\mathit{Low}$ . The remaining shaded pixels have different calculated $X_\mathit{Low}$ and/or $Y_\mathit{Low}$ values, but due to Equation 10, no contribution is required from additional words. |
The texture data words are wrapped toroidally, such that the “nearest” sample may exist on the opposite side of the image.
For example, sampling a pixel in any corner of the image results in the words in all four corners being examined — or sampling a pixel at the bottom of the image will result in words from the top of the image being examined, as shown in Figure 53. In this example, the nearest samples “below” the shaded pixels in regions $\mathbf{M}_{\mathbf{X}_\mathbf{Mod},\mathbf{Y}_\mathbf{Mod}}$ and $\mathbf{M}_{\mathbf{X}_\mathbf{Mod}+1,\mathbf{Y}_\mathbf{Mod}}$ the row of words at the top of the image, and the nearest samples “above” the shaded pixels in regions $\mathbf{M}_{\mathbf{X}_\mathbf{Mod},\mathbf{Y}_\mathbf{Mod}+1}$ and $\mathbf{M}_{\mathbf{X}_\mathbf{Mod}+1,\mathbf{Y}_\mathbf{Mod}+1}$ are in words $W_{X_\mathit{Low},Y_\mathit{Low}}$ and $W_{X_\mathit{Low}+1,Y_\mathit{Low}}$ at the bottom of the image. |
The layout of color data in PVRTC word depends on the value of the opacity flag Op, stored in the most-significant bit of the color. If Op is 1, then the color is treated as opaque (no alpha data), and if Op is 0, the color has alpha data and may be translucent. The exact data layout of each color is described below:
Table 175. Data layout of color segments in a PVRTC1 word
Color B — opaque color mode (opacity flag Op = 1) | |||||||||||||||
63 | 62 | 61 | 60 | 59 | 58 | 57 | 56 | 55 | 54 | 53 | 52 | 51 | 50 | 49 | 48 |
1 | Red | Green | Blue | ||||||||||||
Color B — translucent color mode (opacity flag Op = 0) | |||||||||||||||
63 | 62 | 61 | 60 | 59 | 58 | 57 | 56 | 55 | 54 | 53 | 52 | 51 | 50 | 49 | 48 |
0 | Alpha | Red | Green | Blue | |||||||||||
Color A — opaque color mode (opacity flag Op = 1) | |||||||||||||||
47 | 46 | 45 | 44 | 43 | 42 | 41 | 40 | 39 | 38 | 37 | 36 | 35 | 34 | 33 | 32 |
1 | Red | Green | Blue | M | |||||||||||
Color A — translucent color mode (opacity flag Op = 0) | |||||||||||||||
47 | 46 | 45 | 44 | 43 | 42 | 41 | 40 | 39 | 38 | 37 | 36 | 35 | 34 | 33 | 32 |
0 | Alpha | Red | Green | Blue | M |
The Color A values and Color B values for each word are bilinearly interpolated by a factor of 4 in both dimensions, resulting in a pair of color values from two virtual images, Image A and Image B. This upscale operation is performed by treating each color as an ARGB:4555 format, and generating results in an ARGB:8888 format. Any R, G, or B channel in each color with fewer than 5 bits is initially expanded via bit replication:
The 3-bit alpha channel values are expanded to 4 bits by zero padding:
If the color is opaque, then the alpha channel is expanded to the 4-bit value, 0b1111.
For each channel C of each color (Color A and Color B), the interpolation proceeds as follows:
For low-resolution image color channel $\textbf{C}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ stored in word $\textit{W}_{\textit{X}_\textit{Low},\textit{Y}_\textit{Low}}$ :
Using relative coordinates:
Equation 10. PVRTC1 4bpp color channel interpolation
The colors of Image A and Image B at ( $x = (4 \times X_\mathit{Low} + 2)$ , $y = (4 \times Y_\mathit{Low} + 2)$ ) are exactly the corresponding colors that word $W_{X_\mathit{Low}, Y_\mathit{Low}}$ encodes. Texels in the same column as a texel block such that $x_r = 0$ are not influenced by the color samples from words $W_{X_\mathit{Low}+1, Y_\mathit{Low}}$ and $W_{X_\mathit{Low}+1, Y_\mathit{Low}+1}$ . Texels in the same row as a texel block such that $y_r = 0$ are not influenced by the color samples from words $W_{X_\mathit{Low}, Y_\mathit{Low}+1}$ and $W_{X_\mathit{Low}+1, Y_\mathit{Low}+1}$ . Therefore a single color sample at $C_{x,y}$ influences the interpolated color of all texels in the 7×7 region from $C_{x-3,y-3}$ to $C_{x+3,y+3}$ centered on the color sample.
Any 2×2 quad of texel values can be evaluated from a single set of four adjacent texel blocks. This means that the number of texel blocks accessed during bilinear filtering is no worse than the worst case of the self-contained texel blocks of other schemes. |
For the red, green and blue channels, Cx,y is a 5.4-bit fixed-point value whose bit pattern can be converted to an 8-bit normalized value, i.e. UNORM, as $\textbf{Image \{A,B\}}\{\textit{R,G,B}\}_{x,y} = \left\lfloor{{C_{x,y}}\over 2}\right\rfloor+\left\lfloor{{C_{x,y}}\over 64}\right\rfloor$ .
For the alpha channel, Cx,y is a 4.4-bit fixed-point value whose bit pattern can be converted to an 8-bit normalized value as $\textbf{Image \{A,B\}}\{\textit{A}\}_{x,y} = C_{x,y}+\left\lfloor{{C_{x,y}}\over 16}\right\rfloor$ .
The final image is created by linearly interpolating between the Image A and Image B texels, using the modulation data for each pixel to determine the weighting. The modulation information is retrieved from word $W_{X_\textit{Mod},Y_\textit{Mod}}$ where $X_\textit{Mod} = \left\lfloor{x\over 4}\right\rfloor$ and $Y_\textit{Mod} = \left\lfloor{y\over 4}\right\rfloor$ . The weight for the interpolation is derived from the 2 bits of the modulation data corresponding to the relevant pixel offset (xoffset = x - (4 × XMod), yoffset = y - (4 × YMod)), depending on the value of modulation flag M.
Table 176. Modulation weights for PVRTC1 4bpp
Standard bilinear (M = 0) | Punch-through (M = 1) | ||||
Modulation bits | Weight | Modulation bits | Weight | Alpha | |
00 | 0 | 00 | 0 | Normal | |
01 | 3 | 01 | 4 | ||
10 | 5 | 10 | “Punch-through” | ||
11 | 8 | 11 | 8 | Normal |
Equation 11. PVRTC image modulation
If punch-through mode is selected, and the modulation bits for a given pixel have a value of 0b10, the alpha value of the resulting color is 0x00. This is irrespective of the presence or values of any alpha channel in the input colors.
For punch-through pixels, the RGB components are 50:50 blends of the corresponding pixels in the upscaled images. For this reason, with PVRTC1 4bpp, it is advised to not use pre-multiplied alpha textures, and to change the color of fully transparent areas to the average of the local neighborhood. PVRTexTool provides “alpha bleed” functionality to modify fully-transparent areas appropriately. |
PVRTC1 2bpp has the same broad data layout as PVRTC1 4bpp, but instead uses an 8×4 bilinear upscale.
The inputs to the Morton order encoding for 2bpp mode are:
Let Xb be the number of bits used to express X — i.e. $\mathit{Xb} = \mathrm{log}_2\left(\left\lceil{\mathit{width}\over 8}\right\rceil\right)$ .
Let Yb be the number of bits used to express Y — i.e. $\mathit{Yb} = \mathrm{log}_2\left(\left\lceil{\mathit{height}\over 4}\right\rceil\right)$ .
For each channel C of each color (Color A and Color B), the interpolation proceeds as follows:
For low-resolution image color channel $\textbf{C}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ stored in word $\textit{W}_{\textit{X}_\textit{Low},\textit{Y}_\textit{Low}}$ :
Using relative coordinates:
Equation 12. PVRTC1 2bpp color channel interpolation
For the red, green and blue channels, Cx,y is a 5.5-bit fixed-point value whose bit pattern can be converted to an 8-bit normalized value as $\textbf{Image \{A,B\}}\{\textit{R,G,B}\}_{x,y} = \left\lfloor{{C_{x,y}}\over 4}\right\rfloor+\left\lfloor{{C_{x,y}}\over 128}\right\rfloor$ .
For the alpha channel, Cx,y is a 4.5-bit fixed-point value whose bit pattern can be converted to an 8-bit normalized value as $\textbf{Image \{A,B\}}\{\textit{A}\}_{x,y} = \left\lfloor{{C_{x,y}}\over 2}\right\rfloor+\left\lfloor{{C_{x,y}}\over 32}\right\rfloor$ .
The modulation data, retrieved from word WX,Y where $X_\mathit{Mod} = \left\lfloor{x\over 8}\right\rfloor$ and $Y_\mathit{Mod} = \left\lfloor{y\over 4}\right\rfloor$ , are interpreted differently in order to accommodate the additional pixels. Each word holds the modulation data which corresponds to pixels that have offsets (xoffset = x - 8 × XMod, yoffset = y - 4 × YMod).
In PVRTC1 2bpp, rather than changing the weight values as in PVRTC1 4bpp, the modulation flag M in bit 32 affects the modulation data layout. Optional flags in bit 0 and bit 20 of the modulation data may further affect the layout:
Table 177. Texel Data format for PVRTC1 2bpp compressed texture formats
Bits 64..32: Color data and flags | |||||||||||||||||||||||||||||||
Identical to PVRTC1 4bpp | |||||||||||||||||||||||||||||||
Bits 31..0: Modulation data: direct encoding, 1 bit per pixel (modulation flag M = 0) for pixel offset (x, y) | |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
7,3 | 6,3 | 5,3 | 4,3 | 3,3 | 2,3 | 1,3 | 0,3 | 7,2 | 6,2 | 5,2 | 4,2 | 3,2 | 2,2 | 1,2 | 0,2 | 7,1 | 6,1 | 5,1 | 4,1 | 3,1 | 2,1 | 1,1 | 0,1 | 7,0 | 6,0 | 5,0 | 4,0 | 3,0 | 2,0 | 1,0 | 0,0 |
Bits 31..0: Modulation data, checkerboard-interpolated encoding, samples for pixel offset (x, y), bits[1..0] (modulation flag M = 1, bit 0 flag = 0) | |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
7,3 | 5,3 | 3,3 | 1,3 | 6,2 | 4,2 | 2,2 | 0,2 | 7,1 | 5,1 | 3,1 | 1,1 | 6,0 | 4,0 | 2,0 | 0,0 | 0 | |||||||||||||||
Bits 31..0: Modulation data, horizontally- or vertically-interpolated encoding, samples for pixel offset (x, y) (modulation flag M = 1, bit 0 flag = 1) | |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
7,3 | 5,3 | 3,3 | 1,3 | 6,2 | 4,2 | F | 2,2 | 0,2 | 7,1 | 5,1 | 3,1 | 1,1 | 6,0 | 4,0 | 2,0 | 0,0 | 1 |
If the modulation flag M is set to 0, each pixel only has a single bit of modulation data. The selected color is Image A for modulation data bit = 0, and Image B for modulation data bit = 1.
Table 178. Modulation modes for PVRTC1 2bpp
Modulation flag value M | Mode |
0 | Standard Bilinear, 1bpp modulation |
1 | Punch-through, interpolated modulation |
If the modulation flag M is set to 1, the pixels with 2-bit stored values have modulation weights equal to those of PVRTC1 4bpp for modulation mode 1, as shown in Table 176.
The modulation data for the pixel at 0,0 (and when bit0 = 1, the pixel at 2,4) is treated as if the value was duplicated — so the single-bit 0 encoding becomes 0b00, and 1 becomes 0b11, such that the texels always correspond to Image A or Image B.
For pixels without stored modulation data, bit0 and flag F in bit 20 determine how they are reconstructed:
where md(x,y) is the modulation data for texel offset (x, y) and w() is the weighting described in Table 176 for M = 1.
If an adjacent pixel’s modulation value is not present in the current word, the value is obtained from the adjacent PVRTC word which does contain that pixel’s modulation data, with the location wrapping to the other side of the image if necessary.
This weight is then applied to Equation 11.
PVRTC2 is the second revision of the PVRTC compression scheme, and shares most of its layout and interpretation. The PVRTC2 sections document the differences between the formats.
PVRTC2 images may have logical dimensions of any size greater than zero. If the logical size of the image is not a multiple of the interpolation size, the stored format uses the next-larger multiple. Any padding texels added by this resize are not accessed when reconstructing the image.
PVRTC2 words are not laid out in a reflected Morton order, instead they are laid out in a standard linear order, as with all other formats.
The layout of PVRTC2 data words is very similar to that of PVRTC1, though there are two main differences:
Table 179. Texel Data format for PVRTC2 4bpp compressed texture formats
Bits 64..32: Color data and flags | |||||||||||||||||||||||||||||||
63 | 62 | 61 | 60 | 59 | 58 | 57 | 56 | 55 | 54 | 53 | 52 | 51 | 50 | 49 | 48 | 47 | 46 | 45 | 44 | 43 | 42 | 41 | 40 | 39 | 38 | 37 | 36 | 35 | 34 | 33 | 32 |
Op | Color B | H | Color A | M | |||||||||||||||||||||||||||
Bits 31..0: Modulation data bits [1..0] for pixel offset (x, y) — identical to PVRTC1 4bpp | |||||||||||||||||||||||||||||||
31 | 30 | 29 | 28 | 27 | 26 | 25 | 24 | 23 | 22 | 21 | 20 | 19 | 18 | 17 | 16 | 15 | 14 | 13 | 12 | 11 | 10 | 9 | 8 | 7 | 6 | 5 | 4 | 3 | 2 | 1 | 0 |
3,3 | 2,3 | 1,3 | 0,3 | 3,2 | 2,2 | 1,2 | 0,2 | 3,1 | 2,1 | 1,1 | 0,1 | 3,0 | 2,0 | 1,0 | 0,0 |
Table 180. Data layout of color segments in a PVRTC2 word
Color B — opaque color mode (opacity flag Op = 1) | |||||||||||||||
63 | 62 | 61 | 60 | 59 | 58 | 57 | 56 | 55 | 54 | 53 | 52 | 51 | 50 | 49 | 48 |
1 | Red | Green | Blue | ||||||||||||
Color B — translucent color mode (opacity flag Op = 0) | |||||||||||||||
63 | 62 | 61 | 60 | 59 | 58 | 57 | 56 | 55 | 54 | 53 | 52 | 51 | 50 | 49 | 48 |
0 | Alpha | Red | Green | Blue | |||||||||||
Color A — opaque color mode (opacity flag Op = 1) | |||||||||||||||
47 | 46 | 45 | 44 | 43 | 42 | 41 | 40 | 39 | 38 | 37 | 36 | 35 | 34 | 33 | 32 |
H | Red | Green | Blue | M | |||||||||||
Color A — translucent color mode (opacity flag Op = 0) | |||||||||||||||
47 | 46 | 45 | 44 | 43 | 42 | 41 | 40 | 39 | 38 | 37 | 36 | 35 | 34 | 33 | 32 |
H | Alpha | Red | Green | Blue | M |
There is one change to the interpretation of the color data relative to PVRTC1: as there is only one opacity flag for each PVRTC2 data word, to allow a local area to span the full alpha gamut when in translucent mode, Color B's 3-bit alpha channel is expanded to four bits as A2A1A01, instead of A2A1A00.
The bilinear interpolation scheme of PVRTC1 usually works well as most areas of natural image textures are reasonably ‘continuous’. However at the boundaries of non-tiling textures, around sub-textures in texture atlases, or in some areas of hand-drawn graphics, this assumption can break down. To address these issues, PVRTC2 includes the hard transition flag, H.
Since it can be assumed that such discontinuities are more likely to be centered on boundaries of multiples of 4×4 texel regions, the hard transition flag H changes the behavior of the entire red-dotted region shown in Figure 54. This is a subset of the logical 4×4 pixel regions $\textbf{M}_{\textbf{X}_\textbf{Mod},\textbf{Y}_\textbf{Mod}}$ through $\textbf{M}_{\textbf{X}_\textbf{Mod}+1,\textbf{Y}_\textbf{Mod}+1}$ that correspond to the modulation data stored in 64-bit data words $W_{X_\mathit{Mod},Y_\mathit{Mod}}$ through $W_{X_\mathit{Mod}+1,Y_\mathit{Mod}+1}$ . The flag H for this hard transition region is stored in $W_{X_\mathit{Mod},Y_\mathit{Mod}}^{47}$ .
The hard transition region is further subdivided into four smaller subregions, shown with the dotted 2×2 texel outlines in Figure 54, where it intersects the pixel regions $\textbf{M}_{\textbf{X}_\textbf{Mod},\textbf{Y}_\textbf{Mod}}$ through $\textbf{M}_{\textbf{X}_\textbf{Mod}+1,\textbf{Y}_\textbf{Mod}+1}$ . The hard transition flag H, coupled with the relevant modulation flag M for the texel subregion, determines how the colors for each reconstructed pixel in the subregion are evaluated, as summarized in the table below.
Table 181. Modulation modes for PVRTC2 4bpp
Modulation flag M | Hard transition flag H | Mode |
0 | 0 | Standard bilinear |
1 | 0 | Punch-through alpha |
0 | 1 | Non-interpolated |
1 | 1 | Local palette |
In ‘standard bilinear’ the modulation behaves as described for PVRTC1 4bpp.
In ‘punch-through alpha’, the modulation behaves in almost the same manner as the equivalent mode as for PVRTC1 4bpp except that for texels marked as ‘punch-through’; i.e. using the 2-bit encoding 0b10, the output texel is set to transparent black, which may be better suited to pre-multiplied texture formats.
The remaining two modes are described below.
In the non-interpolated mode, the A and B base colors are not bilinearly interpolated in the affected regions. Instead, the colors for the word encapsulating each particular pixel are used directly, in the sense that the stored colors are expanded, where necessary, via bit replication to an ARGB:4555 format and then, again, by bit replication to ARGB:8888. The modulation encodings are interpreted in the same manner as for the “standard bilinear” weights, and the colors blended, as before, with Equation 11.
In local palette mode, the hard transition region is no longer reconstructed by interpolating the upscaled images. Instead, the eight distinct colors from each surrounding word make up a local palette from which colors are selected.
Denoting Color B and Color A from words $W_{X_\textit{Low},Y_\textit{Low}}$ through $W_{X_\textit{Low}+1,Y_\textit{Low}+1}$ as described above, the following colors are available: $\textbf{A}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ , $\textbf{B}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ , $\textbf{A}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}}$ , $\textbf{B}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}}$ , $\textbf{A}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}+1}$ , $\textbf{B}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}+1}$ , $\textbf{A}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}+1}$ and $\textbf{B}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}+1}$ .
Whilst 8 distinct colors exist in each of those four words, only two bits of modulation data are available for each pixel. Subsequently, each pixel at offset $(x_r = (x - 2) - (4 \times X_\textit{Low}), y_r = (y - 2) - (4 \times Y_\textit{Low}))$ relative to start of the hard transition region has access to a subset of the palette as follows:
Table 182. Color mappings in local palette mode for PVRTC2 4bpp
Modulation bits | 0,0 | 1,0 | 2,0 | 3,0 |
0 | $\textbf{A}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ | $\textbf{A}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ | $\textbf{A}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ | $\textbf{A}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ |
1 | $\left\lfloor{{5\times \textbf{A}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}+3\times \textbf{B}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}}\over 8}\right\rfloor$ | $\textbf{B}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ | $\textbf{B}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ | $\textbf{B}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ |
2 | $\left\lfloor{{3\times \textbf{A}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}+5\times \textbf{B}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}}\over 8}\right\rfloor$ | $\textbf{A}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}}$ | $\textbf{A}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}}$ | $\textbf{A}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}}$ |
3 | $\textbf{B}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ | $\textbf{B}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}}$ | $\textbf{B}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}}$ | $\textbf{B}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}}$ |
Modulation bits | 0,1 | 1,1 | 2,1 | 3,1 |
0 | $\textbf{A}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ | $\textbf{A}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ | $\textbf{A}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ | $\textbf{A}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}+1}$ |
1 | $\textbf{B}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ | $\textbf{B}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ | $\textbf{B}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ | $\textbf{B}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ |
2 | $\textbf{A}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}+1}$ | $\textbf{A}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}}$ | $\textbf{A}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}}$ | $\textbf{A}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}}$ |
3 | $\textbf{B}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}+1}$ | $\textbf{B}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}+1}$ | $\textbf{B}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}}$ | $\textbf{B}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}}$ |
Modulation bits | 0,2 | 1,2 | 2,2 | 3,2 |
0 | $\textbf{A}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ | $\textbf{A}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ | $\textbf{A}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ | $\textbf{A}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}+1}$ |
1 | $\textbf{B}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ | $\textbf{B}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ | $\textbf{B}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}+1}$ | $\textbf{B}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}+1}$ |
2 | $\textbf{A}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}+1}$ | $\textbf{A}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}+1}$ | $\textbf{A}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}+1}$ | $\textbf{A}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}}$ |
3 | $\textbf{B}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}+1}$ | $\textbf{B}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}+1}$ | $\textbf{B}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}}$ | $\textbf{B}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}}$ |
Modulation bits | 0,3 | 1,3 | 2,3 | 3,3 |
0 | $\textbf{A}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ | $\textbf{A}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ | $\textbf{A}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}+1}$ | $\textbf{A}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}+1}$ |
1 | $\textbf{B}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}}$ | $\textbf{B}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}+1}$ | $\textbf{B}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}+1}$ | $\textbf{B}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}+1}$ |
2 | $\textbf{A}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}+1}$ | $\textbf{A}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}+1}$ | $\textbf{A}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}+1}$ | $\textbf{A}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}+1}$ |
3 | $\textbf{B}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}+1}$ | $\textbf{B}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}+1}$ | $\textbf{B}_{\textbf{X}_\textbf{Low},\textbf{Y}_\textbf{Low}+1}$ | $\textbf{B}_{\textbf{X}_\textbf{Low}+1,\textbf{Y}_\textbf{Low}}$ |
The entry for offset 0,0 is interpolated as per PVRTC’s standard bilinear filtering mode. It will thus only use colors from word $W_{X_\textit{Low},Y_\textit{Low}}$ .
The local palette mode shares with the other modes of PVRTC2 the property that the column $x_r = 0$ has no contribution from words $W_{X_\mathit{Low}+1, Y_\mathit{Low}}$ and $W_{X_\mathit{Low}+1, Y_\mathit{Low}+1}$ , and that row $y_r = 0$ has no contribution from words $W_{X_\mathit{Low}, Y_\mathit{Low}+1}$ and $W_{X_\mathit{Low}+1, Y_\mathit{Low}+1}$ . Therefore any 2×2 quad of texel values, required for example by bilinear filtering, can be evaluated from a single set of four adjacent texel blocks. |
The stored color values are first expanded, where necessary, to ARGB:4555 (via bit replication for R, G, or B and via padding for Alpha) and then from 4555 to ARGB:8888, again by bit replication. For example,
PVRTC2 2bpp data layout has color data laid out identically to PVRTC2 4bpp, and modulation data equivalent to PVRTC1 2bpp. The only difference between the PVRTC1 and PVRTC2 variants is the addition of the hard transition flag H and the single opacity flag Op, as specified in the PVRTC2 4bpp format.
Table 183. Texel Data format for PVRTC2 2bpp compressed texture formats
Bits 63-32: Color data and flags |
Identical to PVRTC2 4bpp |
Bits 31-0: Modulation data |
Identical to PVRTC1 2bpp |
Color values are interpreted in the same manner as for the PVRTC2 4bpp format.
The 2bpp variation of PVRTC2 is slightly simpler than the 4bpp, in that it only uses the non-interpolated mode — no local palette mode exists.
Table 184. Modulation modes for PVRTC2 2bpp
Modulation flag M | Hard transition H | Mode |
0 | 0 | Standard bilinear, 1bpp modulation |
1 | 0 | Standard bilinear, interpolated modulation |
0 | 1 | Non-interpolated, 1bpp modulation |
1 | 1 | Non-interpolated, interpolated modulation |
If the hard transition flag H for PVRTC2 2bpp is equal to 0, the format is interpreted in the same manner as the PVRTC1 2bpp format.
When the hard transition flag H for PVRTC2 2bpp is equal to 1, the initial bilinear interpolation of block colors across the hard transition region is skipped as for PVRTC2 4bpp. Subsequently, the format is interpreted in the same way as the PVRTC1 2bpp format.
When interpreting modulation data in the interpolated modulation mode, the hard transition flag has no effect on this — the modulation values are always interpolated across PVRTC words. |
IEEE754-2008 - IEEE standard for floating-point arithmetic
IEEE Std 754-2008 http://dx.doi.org/10.1109/IEEESTD.2008.4610935, August, 2008.
CIE Colorimetry - Part 3: CIE tristimulus values
http://cie.co.at/index.php?i_ca_id=823
ITU-R BT.601 Studio encoding parameters of digital television for standard 4:3 and wide-screen 16:9 aspect ratios
http://www.itu.int/rec/R-REC-BT.601/en
ITU-R BT.709 Parameter values for the HDTV standards for production and international programme exchange
https://www.itu.int/rec/R-REC-BT.709/en
ITU-R BT.2020 Parameter values for ultra-high definition television systems for production and international programme exchange
http://www.itu.int/rec/R-REC-BT.2020/en
ITU-R BT.2100 Image parameter values for high dynamic range television for use in production and international programme exchange
http://www.itu.int/rec/R-REC-BT.2100/en
JPEG File Interchange Format (JFIF)
https://www.itu.int/rec/T-REC-T.871/en
Legacy version:
https://www.w3.org/Graphics/JPEG/jfif3.pdf
ITU-R BT.1886: Reference electro-optical transfer function for flat panel displays used in HDTV studio production
https://www.itu.int/rec/R-REC-BT.1886/en
ITU-R BT.2087 Colour conversion from Recommendation ITU-R BT.709 to Recommendation ITU-R BT.2020
http://www.itu.int/rec/R-REC-BT.2087/en
ITU-R BT.2390-1 High dynamic range television for production and international programme exchange
https://www.itu.int/pub/R-REP-BT.2390-5-2018
ITU-R BT.470 Conventional analogue television systems
https://www.itu.int/rec/R-REC-BT.470/en
BT.470-6 contains descriptions of analog broadcast systems. BT.470-7 deprecates this description in favor of BT.1700. |
Although this specification is written in English, the countries in Appendix 1 appear to be listed in alphabetical order as they would have been written in French. |
ITU-R BT.472-3: Video-frequency characteristics of a television system to be used for the international exchange of programmes between countries that have adopted 625-line colour or monochrome systems
http://www.itu.int/rec/R-REC-BT.472/en
ITU-R BT.1700: Characteristics of composite video signals for conventional analogue television systems
https://www.itu.int/rec/R-REC-BT.1700/en
This specification includes SMTPE170M-2004 (which describes NTSC) along with PAL and NTSC. |
ITU-R BT.2043: Analogue television systems currently in use throughout the world
https://www.itu.int/dms_pub/itu-r/opb/rep/R-REP-BT.2043-2004-PDF-E.pdf
Although this specification is written in English, the countries appear to be listed in alphabetical order as they would have been written in French. |
SMPTE 170m Composite analog video signal — NTSC for studio applications
The latest version is available from https://www.smpte.org/ and from the IEEE at:
http://ieeexplore.ieee.org/document/7291416/.
SMPTE 170m-2004 is freely available as part of ITU-R BT.1700 at:
https://www.itu.int/rec/R-REC-BT.1700/en
FCC 73.682 - TV transmission standards
ST 240:1999 - SMPTE Standard - For Television — 1125-Line High-Definition Production Systems — Signal Parameters
Formerly known as SMPTE240M - interim HDTV standard prior to international agreement on BT.709.
Available from the SMPTE and via IEEE:
http://ieeexplore.ieee.org/document/7291461/
IEC/4WD 61966-2-1: Colour measurement and management in multimedia systems and equipment - part 2-1: default RGB colour space - sRGB
https://webstore.iec.ch/publication/6169 (specification)
http://www.w3.org/Graphics/Color/srgb
IEC 61966-2-2:2003: Multimedia systems and equipment — Colour measurement and management — Part 2-2: Colour management — Extended RGB colour space — scRGB
https://www.iso.org/standard/35876.html
http://webstore.iec.ch/webstore/webstore.nsf/artnum/029678
http://www.color.org/chardata/rgb/scrgb.xalter
https://webstore.iec.ch/preview/info_iec61966-2-2%7Bed1.0%7Den.pdf
A working draft is freely available at https://web.archive.org/web/20110725185444/http://www.colour.org/tc8-05/Docs/colorspace/61966-2-2NPa.pdf.
http://www.color.org/sycc.pdf - sYCC (the $Y'C_BC_R$ variant of sRGB)
https://webstore.iec.ch/corrigenda/iec61966-2-2-cor1%7Bed1.0%7Den.pdf - Annex B: Non-linear encoding for scRGB : scRGB-nl and its YCC Transformation: scYCC-nl
The latest version is available from https://www.smpte.org/
https://developer.apple.com/reference/coregraphics/cgcolorspace/1408916-displayp3 describes Apple’s Display P3 color space.
http://www.oscars.org/science-technology/sci-tech-projects/aces/aces-documentation
The international standard for ACES, SMPTE ST 2065-1:2012 - Academy Color Encoding Specification (ACES), is available from the SMPTE, and also from the IEEE.
TB-2014-004: Informative Notes on SMPTE ST 2065-1 – Academy Color Encoding Specification (ACES) is freely available and contains a draft of the international standard.
ACEScc — A Logarithmic Encoding of ACES Data for use within Color Grading Systems
ACEScct — A Quasi-Logarithmic Encoding of ACES Data for use within Color Grading Systems
https://pro.sony.com/bbsccms/assets/files/mkt/cinema/solutions/slog_manual.pdf - S-Log description
https://pro.sony.com/bbsccms/assets/files/micro/dmpc/training/S-Log2_Technical_PaperV1_0.pdf - S-Log2 description
http://www.sony.co.uk/pro/support/attachment/1237494271390/1237494271406/technical-summary-for-s-gamut3-cine-s-log3-and-s-gamut3-s-log3.pdf - S-Log3 description
Frank Brill
Mark Callow
Sean Ellis
Simon Fenney
Jan-Harald Fredriksen
Andrew Garrard
Fabien Giesen
Courtney Goeltzenleuchter
Jonas Gustavsson
Chris Hebert
Tobias Hector
Alexey Knyazev
Daniel Koch
Jon Leech
Thierry Lepley
Tommaso Maestri
Kathleen Mattson
Hans-Peter Nilsson
XianQuan Ooi
Alon Or-bach
Jan Outters
Erik Rainey
Daniel Rakos
Donald Scorgie
Graham Sellers
David Sena
Stuart Smith
Alex Walters
Eric Werness
David Wilkinson