Name

    KHR_texture_compression_astc_hdr

Name Strings

    GL_KHR_texture_compression_astc_hdr
    GL_KHR_texture_compression_astc_ldr

Contact

    Sean Ellis (sean.ellis 'at' arm.com)
    Jon Leech (oddhack 'at' sonic.net)

Contributors

    Sean Ellis, ARM
    Jorn Nystad, ARM
    Tom Olson, ARM
    Andy Pomianowski, AMD
    Cass Everitt, NVIDIA
    Walter Donovan, NVIDIA
    Robert Simpson, Qualcomm
    Maurice Ribble, Qualcomm
    Larry Seiler, Intel
    Daniel Koch, NVIDIA
    Anthony Wood, Imagination Technologies
    Jon Leech
    Andrew Garrard, Samsung

IP Status

    No known issues.

Notice

    Copyright (c) 2012-2016 The Khronos Group Inc. Copyright terms at
            http://www.khronos.org/registry/speccopyright.html

Status

    Complete.
    Approved by the ARB on 2012/06/18.
    Approved by the OpenGL ES WG on 2012/06/15.
    Ratified by the Khronos Board of Promoters on 2012/07/27 (LDR profile).
    Ratified by the Khronos Board of Promoters on 2013/09/27 (HDR profile).

Version

    Version 8, June 8, 2017

Number

    ARB Extension #118
    OpenGL ES Extension #117

Dependencies

    Written based on the wording of the OpenGL ES 3.1 (April 29, 2015)
    Specification

    May be implemented against any version of OpenGL or OpenGL ES supporting
    compressed textures.

    Some of the functionality of these extensions is not supported if the
    underlying implementation does not support cube map array textures.


Overview

    Adaptive Scalable Texture Compression (ASTC) is a new texture
    compression technology that offers unprecendented flexibility, while
    producing better or comparable results than existing texture
    compressions at all bit rates. It includes support for 2D and
    slice-based 3D textures, with low and high dynamic range, at bitrates
    from below 1 bit/pixel up to 8 bits/pixel in fine steps.

    The goal of these extensions is to support the full 2D profile of the
    ASTC texture compression specification, and allow construction of 3D
    textures from multiple compressed 2D slices.

    ASTC-compressed textures are handled in OpenGL ES and OpenGL by adding
    new supported formats to the existing commands for defining and updating
    compressed textures, and defining the interaction of the ASTC formats
    with each texture target.

New Procedures and Functions

    None

New Tokens

    Accepted by the <internalformat> parameter of CompressedTexImage2D,
    CompressedTexSubImage2D, CompressedTexImage3D, CompressedTexSubImage3D,
    TexStorage2D, TextureStorage2D, TexStorage3D, and TextureStorage3D:

    COMPRESSED_RGBA_ASTC_4x4_KHR            0x93B0
    COMPRESSED_RGBA_ASTC_5x4_KHR            0x93B1
    COMPRESSED_RGBA_ASTC_5x5_KHR            0x93B2
    COMPRESSED_RGBA_ASTC_6x5_KHR            0x93B3
    COMPRESSED_RGBA_ASTC_6x6_KHR            0x93B4
    COMPRESSED_RGBA_ASTC_8x5_KHR            0x93B5
    COMPRESSED_RGBA_ASTC_8x6_KHR            0x93B6
    COMPRESSED_RGBA_ASTC_8x8_KHR            0x93B7
    COMPRESSED_RGBA_ASTC_10x5_KHR           0x93B8
    COMPRESSED_RGBA_ASTC_10x6_KHR           0x93B9
    COMPRESSED_RGBA_ASTC_10x8_KHR           0x93BA
    COMPRESSED_RGBA_ASTC_10x10_KHR          0x93BB
    COMPRESSED_RGBA_ASTC_12x10_KHR          0x93BC
    COMPRESSED_RGBA_ASTC_12x12_KHR          0x93BD

    COMPRESSED_SRGB8_ALPHA8_ASTC_4x4_KHR    0x93D0
    COMPRESSED_SRGB8_ALPHA8_ASTC_5x4_KHR    0x93D1
    COMPRESSED_SRGB8_ALPHA8_ASTC_5x5_KHR    0x93D2
    COMPRESSED_SRGB8_ALPHA8_ASTC_6x5_KHR    0x93D3
    COMPRESSED_SRGB8_ALPHA8_ASTC_6x6_KHR    0x93D4
    COMPRESSED_SRGB8_ALPHA8_ASTC_8x5_KHR    0x93D5
    COMPRESSED_SRGB8_ALPHA8_ASTC_8x6_KHR    0x93D6
    COMPRESSED_SRGB8_ALPHA8_ASTC_8x8_KHR    0x93D7
    COMPRESSED_SRGB8_ALPHA8_ASTC_10x5_KHR   0x93D8
    COMPRESSED_SRGB8_ALPHA8_ASTC_10x6_KHR   0x93D9
    COMPRESSED_SRGB8_ALPHA8_ASTC_10x8_KHR   0x93DA
    COMPRESSED_SRGB8_ALPHA8_ASTC_10x10_KHR  0x93DB
    COMPRESSED_SRGB8_ALPHA8_ASTC_12x10_KHR  0x93DC
    COMPRESSED_SRGB8_ALPHA8_ASTC_12x12_KHR  0x93DD

    If extension "EXT_texture_storage" is supported, these tokens are also
    accepted by TexStorage2DEXT, TextureStorage2DEXT, TexStorage3DEXT and
    TextureStorage3DEXT.

Additions to Chapter 8 of the OpenGL ES 3.1 Specification (Textures and Samplers)

    Add to Section 8.7 Compressed Texture Images:

    Modify table 8.19 (Compressed internal formats) to add all the ASTC
    format tokens in the New Tokens section. The "Base Internal Format"
    column is RGBA for all ASTC formats.

    Add a new column "Block Width x Height", which is 4x4 for all non-ASTC
    formats in the table, and matches the size in the token name for ASTC
    formats (e.g. COMPRESSED_SRGB8_ALPHA8_ASTC_10x8_KHR has a block size of
    10 x 8).

    Add a second new column "3D Tex." which is empty for all non-ASTC
    formats. If only the LDR profile is supported by the implementation,
    this column is also empty for all ASTC formats. If both the LDR and HDR
    profiles are supported, this column is checked for all ASTC formats.

    Add a third new column "Cube Map Array Tex." which is empty for all
    non-ASTC formats, and checked for all ASTC formats.

    Append to the table caption:

   "The "Block Size" column specifies the compressed block size of the
    format. Modifying compressed images along aligned block boundaries is
    possible, as described in this section. The "3D Tex." and "Cube Map
    Array Tex." columns determine if 3D images composed of compressed 2D
    slices, and cube map array textures respectively can be specified using
    CompressedTexImage3D."

    Append to the paragraph at the bottom of p. 168:

   "If <internalformat> is one of the specific ... supports only
    two-dimensional images. However, if the "3D Tex." column of table 8.19
    is checked, CompressedTexImage3D will accept a three-dimensional image
    specified as an array of compressed data consisting of multiple rows of
    compressed blocks laid out as described in section 8.5."

    Modify the second and third errors in the Errors section for
    CompressedTexImage[2d]D on p. 169, and add a new error:

   "An INVALID_VALUE error is generated by

      * CompressedTexImage2D if <target> is
        one of the cube map face targets from table 8.21, and
      * CompressedTexImage3D if <target> is TEXTURE_CUBE_MAP_ARRAY,

    and <width> and <height> are not equal.

    An INVALID_OPERATION error is generated by CompressedTexImage3D if
    <internalformat> is one of the the formats in table 8.19 and <target> is
    not TEXTURE_2D_ARRAY, TEXTURE_CUBE_MAP_ARRAY, or TEXTURE_3D.

    An INVALID_OPERATION error is generated by CompressedTexImage3D if
    <internalformat> is TEXTURE_CUBE_MAP_ARRAY and the "Cube Map Array"
    column of table 8.19 is *not* checked, or if <internalformat> is
    TEXTURE_3D and the "3D Tex." column of table 8.19 is *not* checked"

    Modify the fifth and sixth paragraphs on p. 170:

   "Since these specific compressed formats are easily edited along texel
    block boundaries, the limitations on subimage location and size are
    relaxed for CompressedTexSubImage2D and CompressedTexSubImage3D.

    The block width and height varies for different formats, as described in
    table 8.19. The contents of any block of texels of a compressed texture
    image in these specific compressed formats that does not intersect the
    area being modified are preserved during CompressedTexSubImage* calls."

    Modify the second error in the Errors section for
    CompressedTexSubImage[23]D on p. 170, and add a new error:

   "An INVALID_OPERATION error is generated by CompressedTexSubImage3D if
    <format> is one of the formats in table 8.19 and <target> is not
    TEXTURE_2D_ARRAY, TEXTURE_CUBE_MAP_ARRAY, or TEXTURE_3D.

    An INVALID_OPERATION error is generated by CompressedTexSubImage3D if
    <format> is TEXTURE_CUBE_MAP_ARRAY and the "Cube Map Array" column of
    table 8.19 is *not* checked, or if <format> is TEXTURE_3D and the "3D
    Tex." column of table 8.19 is *not* checked"

    Modify the final error in the same section, on p. 171:

   "An INVALID_OPERATION error is generated if format is one of the formats
    in table 8.19 and any of the following conditions occurs. The block
    width and height refer to the values in the corresponding column of the
    table.

      * <width> is not a multiple of the format's block width, and <width> +
        <xoffset> is not equal to the value of TEXTURE_WIDTH.
      * height is not a multiple of the format's block height, and <height>
        + <yoffset> is not equal to the value of TEXTURE_HEIGHT.
      * <xoffset> or <yoffset> is not a multiple of the block width or
        height, respectively."

    Modify table 8.24 (sRGB texture internal formats) to add all of the
    COMPRESSED_SRGB8_ALPHA8_ASTC_*_KHR formats defined above.

Additions to Appendix C of the OpenGL ES 3.1 Specification (Compressed
Texture Image Formats)

    Add a new sub-section on ASTC image formats, as follows:

   "C.2 ASTC Compressed Texture Image Formats
    =========================================

    C.2.1   What is ASTC?
    ---------------------

    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 encoded using either high or low dynamic range,
    corresponding to the "HDR profile" and "LDR profile". Support for the
    HDR profile is indicated by the "GL_KHR_texture_compression_astc_hdr"
    extension string, and support for the LDR profile is indicated by the
    "GL_KHR_texture_compression_astc_ldr" extension string.

    The LDR profile supports two-dimensional images for texture targets
    TEXTURE_2D. TEXTURE_2D_ARRAY, the six texture cube map face targets, and
    TEXTURE_CUBE_MAP_ARRAY. These images may optionally be specified using
    the sRGB color space for the RGB channels.

    The HDR profile is a superset of the LDR profile, and also supports
    texture target TEXTURE_3D for images made up of multiple two-dimensional
    slices of compressed data. HDR images may be a mix of low and high
    dynamic range data. If the HDR profile is supported, the LDR profile and
    its extension string must also be supported.

    ASTC textures may be encoded as 1, 2, 3 or 4 components, but they are
    all decoded into RGBA.

    Different ASTC formats have different block sizes, specified as part of
    the name of the format token passed to CompressedImage2D and its related
    functions, and in table 8.19.

    Additional ASTC formats (the "Full profile") exist which support 3D data
    specified as compressed 3D blocks. However, such formats are not defined
    by either the LDR or HDR profiles, and are not described in this
    specification.

    C.2.2   Design Goals
    --------------------

    The design goals for the format are as follows:

    * Random access. This is a must for any texture compression format.
    * Bit exact decode. This is a must for conformance testing and
      reproducibility.
    * Suitable for mobile use. The format should be suitable for both
      desktop and mobile GPU environments. It should be low bandwidth
      and low in area.
    * Flexible choice of bit rate. Current formats only offer a few bit
      rates, leaving content developers with only coarse control over
      the size/quality tradeoff.
    * Scalable and long-lived. The format should support existing R, RG,
      RGB and RGBA image types, and also have high "headroom", allowing
      continuing use for several years and the ability to innovate in
      encoders. Part of this is the choice to include HDR and 3D.
    * Feature orthogonality. The choices for the various features of the
      format are all orthogonal to each other. This has three effects:
      first, it allows a large, flexible configuration space; second,
      it makes that space easier to understand; and third, it makes
      verification easier.
    * Best in class at given bit rate. It should beat or match the current
      best in class for peak signal-to-noise ratio (PSNR) at all bit rates.
    * Fast decode. Texel throughput for a cached texture should be one
      texel decode per clock cycle per decoder. Parallel decoding of several
      texels from the same block should be possible at incremental cost.
    * Low bandwidth. The encoding scheme should ensure that memory access
      is kept to a minimum, cache reuse is high and memory bandwidth for
      the format is low.
    * Low area. It must occupy comparable die size to competing formats.

    C.2.3   Basic Concepts
    ----------------------

    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").

    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.

    C.2.4   Block Encoding
    ----------------------

    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 12x12 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.

    C.2.5   LDR and HDR Modes
    -------------------------

    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.

    -----------------------------------------------------------------------------
    Operation           LDR Mode                    HDR Mode
    -----------------------------------------------------------------------------
    Returned value      Vector of FP16 values,      Vector of FP16 values
                        or Vector of UNORM8 values.

    sRGB compatible     Yes                         No

    LDR endpoint        16 bits, or                 16 bits
    decoding precision  8 bits for sRGB

    HDR endpoint mode   Error color                 As decoded
    results

    Error results       Error color                 Vector of NaNs (0xFFFF)
    -----------------------------------------------------------------------------
          Table C.2.1 - Differences Between LDR and HDR Modes

    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.

    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 KHR_texture_compression_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.

    C.2.6   Configuration Summary
    -----------------------------

    The global configuration data for the format is as follows:

    *   Block dimension (always 2D for both LDR and HDR profiles)
    *   Block footprint size
    *   sRGB output enabled or not

    The data specified per block is as follows:

    *   Texel weight grid size
    *   Texel weight range
    *   Texel weight values
    *   Number of partitions
    *   Partition pattern index
    *   Color endpoint modes (includes LDR or HDR selection)
    *   Color endpoint data
    *   Number of planes
    *   Plane-to-channel assignment

    C.2.7   Decode Procedure
    ------------------------

    To decode one texel:

    Find block containing texel
    Read block mode
    If void-extent block, store void extent and immediately return single
        color (optimization)

    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

    C.2.8   Block Determination and Bit Rates
    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 following bit rates:

        -------------------------------------
         Footprint
        Width Height    Bit Rate    Increment
        -------------------------------------
        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
        -------------------------------------
        Table C.2.2 - 2D Footprint and Bit Rates

    The block footprint is shown as <width>x<height> in the format name. For
    example, the format COMPRESSED_RGBA_ASTC_8x6_KHR specifies an image with
    a block width of 8 texels, and a block height of 6 texels.

    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.

    The HDR profile supports only those block footprints listed in Table
    C.2.2. 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. 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 a 2D image which is W x H pixels in size, with block size
    w x h, the size of the image in blocks is:

        Bw = ceiling(W/w)
        Bh = ceiling(H/h)

    For a 3D image, each 2D slice is a single texel thick, so that for an
    image which is W x H x D pixels in size, with block size w x h, the size
    of the image in blocks is:

        Bw = ceiling(W/w)
        Bh = ceiling(H/h)
        Bd = D

    C.2.9   Block Layout
    --------------------

    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, as shown in figure C.1.

    127 126 125 124 123 122 121 120 119 118 117 116 115 114 113 112
     --------------------------------------------------------------
    | Texel Weight Data (variable width)        Fill direction ->
     --------------------------------------------------------------

    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
     --------------------------------------------------------------
          <-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                                 |
     --------------------------------------------------------------

        Figure C.1 - Block Layout Overview

    Dotted partition lines indicate that the split position is 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
    figures C.2 and C.3 (only the bottom 32 bits are shown):

    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                   |
    --------------------------------------------------------------

        Figure C.2 - Single-partition Block Layout

    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.

    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 |              Block Mode                   |
    --------------------------------------------------------------

        Figure C.3 - Multi-partition Block Layout

    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 C.2.22.

    C.2.10  Block Mode
    ------------------

    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 value R, which is interpreted
    together with a precision bit H, as follows:

        Low Precision Range (H=0)           High Precision Range (H=1)
    R   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
    -------------------------------------------------------------------------
    Table C.2.7 - Weight Range Encodings

    Each weight value is encoded using the specified number of Trits, Quints
    and Bits. The details of this encoding can be found in Section C.3.12 -
    Integer Sequence Encoding.

    For 2D blocks, the Block Mode field is laid out as follows:

    -------------------------------------------------------------------------
    10  9   8   7   6   5   4   3   2   1   0   Width Height Notes
    -------------------------------------------------------------------------
    D   H     B       A     R0  0   0   R2  R1  B+4   A+2
    D   H     B       A     R0  0   1   R2  R1  B+8   A+2
    D   H     B       A     R0  1   0   R2  R1  A+2   B+8
    D   H   0   B     A     R0  1   1   R2  R1  A+2   B+6
    D   H   1   B     A     R0  1   1   R2  R1  B+2   A+2
    D   H   0   0     A     R0  R2  R1  0   0   12    A+2
    D   H   0   1     A     R0  R2  R1  0   0   A+2   12
    D   H   1   1   0   0   R0  R2  R1  0   0   6     10
    D   H   1   1   0   1   R0  R2  R1  0   0   10    6
      B     1   0     A     R0  R2  R1  0   0   A+6   B+6   D=0, H=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
    -------------------------------------------------------------------------
    Table C.2.8 - 2D Block Mode Layout

    Note that, due to the encoding of the R field, as described in the
    previous page, bits R2 and R1 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 the table 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).

    The D bit is set to indicate dual-plane mode. In this mode, the maximum
    allowed number of partitions is 3.

    The penultimate row of the table 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 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.

    C.2.11  Color Endpoint Mode
    ---------------------------

    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 follows:

    ---------------------------------------------
    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
    ---------------------------------------------
    Table C.2.10 - Color Endpoint Modes.
      [[ If the HDR profile is not implemented, remove from table C.2.10
         all rows whose description starts with "HDR", and add to the
         caption: ]]
    Modes not described in the CEM column are reserved for HDR modes, and
    will generate errors in an unextended OpenGL ES implementation.

    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:

    ----------------------------------------------------
    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.
    ----------------------------------------------------
    Table C.2.11 - Multi-Partition Color Endpoint Modes

    If the CEM selector value in bits [24:23] is not 00,
    then data layout is as follows:

    ---------------------------------------------------
    Part            n   m   l   k   j   i   h   g
            ------------------------------------------
    2   ... Weight :  M1   :                         ...
            ------------------------------------------
    3   ... Weight :  M2   :  M1   :M0 :             ...
            ------------------------------------------
    4   ... Weight :  M3   :  M2   :  M1   :  M0   : ...
            ------------------------------------------

    Part    28  27  26  25  24  23
            ----------------------
    2      |  M0   |C1 |C0 | CEM  |
            ----------------------
    3      |M0 |C2 |C1 |C0 | CEM  |
            ----------------------
    4      |C3 |C2 |C1 |C0 | CEM  |
            ----------------------
    ---------------------------------------------------
    Figure C.4 - Multi-Partition Color Endpoint Modes

    In this view, each partition i has two fields. C<i> is the class
    selector bit, choosing between the two possible CEM classes (0 indicates
    the lower of the two classes), and M<i> 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 C.2.13.

    C.2.12  Integer Sequence Encoding
    ---------------------------------

    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 following forms:

    Value range      MSB encoding   LSB encoding Value       Block Packed
                                                                   block size
    -----------      ------------   ------------ ----------- ----- ----------
    0 .. 2^n-1       -              n bit value  m           1     n
                                    m (n <= 8)
    0 .. (3 * 2^n)-1 Base-3 "trit"  n bit value  t * 2^n + m 5     8 + 5*n
                     value t        m (n <= 6)
    0 .. (5 * 2^n)-1 Base-5 "quint" n bit value  q * 2^n + m 3     7 + 3*n
                     value q        m (n <= 5)
    -------------------------------------------
    Table C.2.13 -Encoding for Different Ranges

    Since 3^5 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 5^3 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*n+n-1:i*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
    floor(i/5)*(8+5*n). The bits in the block are packed as follows
    (in this example, n is 4):

                    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       |
     --------------------------------------------------------------

    Figure C.5 - Trit-based Packing

    The five trits t0 to t4 are obtained by bit manipulations of the 8 bits
    T[7:0] 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
    floor(i/3)*(7+3*n).

    The bits in the block are packed as follows (in this example, n is 4):

                                                        18  17  16
                                                        -----------
                                                       |Q6  Q5 | m2
                                                        -----------
    15  14  13  12  11  10  9   8   7   6   5   4   3   2   1   0
    ---------------------------------------------------------------
      m2       |Q4  Q3 |     m1        |Q2  Q1  Q0 |      m0       |
    ---------------------------------------------------------------

    Figure C.6 - Quint-based Packing

    The three quints q0 to q2 are obtained by bit manipulations of the 7 bits
    Q[6:0] 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 = T[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.

    C.2.13  Endpoint Unquantization
    -------------------------------

    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 follows:

    ---------------------------------------------------------------
    Range   T Q B   Bits    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
    ---------------------------------------------------------------
    Table C.2.16 - Color Unquantization Parameters

    These are then processed as follows:

        T = D * C + B;
        T = T ^ A;
        T = (A & 0x80) | (T >> 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.

    C.2.14  LDR Endpoint Decoding
    -----------------------------
    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 v[n-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 follows.

    -------------------------------------------------
    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
    -------------------------------------------------
    Table C.2.17 -Color Endpoint Modes
      [[ If the HDR profile is not implemented, remove from table C.2.17
         all rows whose description starts with "HDR", and add to the
         caption: ]]
    Modes not described are reserved, as described in table C.2.10.

      [[ HDR profile only ]]
    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:

    Mode 0  LDR Luminance, direct

        e0=(v0,v0,v0,0xFF); e1=(v1,v1,v1,0xFF);

    Mode 1  LDR Luminance, base+offset

        L0 = (v0>>2)|(v1&0xC0); L1=L0+(v1&0x3F);
        if (L1>0xFF) { L1=0xFF; }
        e0=(L0,L0,L0,0xFF); e1=(L1,L1,L1,0xFF);

    Mode 4  LDR Luminance+Alpha,direct

        e0=(v0,v0,v0,v2);
        e1=(v1,v1,v1,v3);

    Mode 5  LDR Luminance+Alpha, base+offset

        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);

    Mode 6  LDR RGB, base+scale

        e0=(v0*v3>>8,v1*v3>>8,v2*v3>>8, 0xFF);
        e1=(v0,v1,v2,0xFF);

    Mode 8  LDR RGB, Direct

        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); }

    Mode 9  LDR RGB, base+offset

        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);

    Mode 10 LDR RGB, base+scale plus two A

        e0=(v0*v3>>8,v1*v3>>8,v2*v3>>8, v4);
        e1=(v0,v1,v2, v5);

    Mode 12 LDR RGBA, direct

        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); }

    Mode 13 LDR RGBA, base+offset

        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 grey:

        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 the UNORM8 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;}
        }

      [[ If the HDR profile is not implemented, do not include section
         C.2.15 ]]

    C.2.15 HDR Endpoint Decoding
    -------------------------

    For HDR endpoint modes, color values are represented in a 12-bit
    pseudo-logarithmic representation.

    HDR Endpoint Mode 2

    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);

    HDR Endpoint Mode 3

    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):

    Value   7   6   5   4   3   2   1   0
    -----   ------------------------------
    v0     |M  |         L[6:0]           |
            ------------------------------
    v1     |    X[3:0]     |   d[3:0]     |
            ------------------------------

    Table C.2.18 - HDR Mode 3 Value Layout

    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);

    HDR Endpoint Mode 7

    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):

    Value   7   6   5   4   3   2   1   0
    -----   ------------------------------
    v0     |M[3:2] |       R[5:0]         |
    -----   ------------------------------
    v1     |M1 |X0 |X1 |      G[4:0]      |
    -----   ------------------------------
    v2     |M0 |X2 |X3 |      B[4:0]      |
    -----   ------------------------------
    v3     |X4 |X5 |X6 |      S[4:0]      |
    -----   ------------------------------
    Table C.2.19 - HDR Mode 7 Value Layout

    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 follows:

    ------------------------------------------------------
            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
    ------------------------------------------------------
    Table C.2.20 - Endpoint Bit Mode

    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 modeval = ((v0&0xC0)>>6) | ((v1&0x80)>>5) | ((v2&0x80)>>4);

        int majcomp;
        int mode;

        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;

    HDR Endpoint Mode 11

    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):

    Value   7   6   5   4   3   2   1   0
    -----   ------------------------------
    v0     |            a[7:0]            |
    -----   ------------------------------
    v1     |m0 |a8 |      c[5:0]          |
    -----   ------------------------------
    v2     |m1 |X0 |     b0[5:0]          |
    -----   ------------------------------
    v3     |m2 |X1 |     b1[5:0]          |
    -----   ------------------------------
    v4     |mj0|X2 |X4 |     d0[4:0]      |
    -----   ------------------------------
    v5     |mj1|X3 |X5 |     d1[4:0]      |
    -----   ------------------------------
    Table C.2.21 - HDR Mode 11 Value Layout

    If the major component bits mj[1:0 ] are both 1, then the RGB values
    are specified directly

    Value   7   6   5   4   3   2   1   0
    -----   ------------------------------
    v0     |          R0[11:4]            |
    -----   ------------------------------
    v1     |          R1[11:4]            |
    -----   ------------------------------
    v2     |          G0[11:4]            |
    -----   ------------------------------
    v3     |          G1[11:4]            |
    -----   ------------------------------
    v4     | 1 |        B0[11:5]          |
    -----   ------------------------------
    v5     | 1 |        B1[11:5]          |
    -----   ------------------------------
    Table C.2.22 - HDR Mode 11 Value Layout

    The mode bits m[2:0] specify the bit allocation for the different
    values, and the destinations of the extra bits X0 to X5:

    -------------------------------------------------------------------------
            Number of bits      Destination of extra bits
    Mode    a   b   c   d       X0      X1      X2      X3      X4      X5
    -------------------------------------------------------------------------
    0       9   7   6   7       b0[6]   b1[6]   d0[6]   d1[6]   d0[5]   d1[5]
    1       9   8   6   6       b0[6]   b1[6]   b0[7]   b1[7]   d0[5]   d1[5]
    2       10  6   7   7       a[9]    c[6]    d0[6]   d1[6]   d0[5]   d1[5]
    3       10  7   7   6       b0[6]   b1[6]   a[9]    c[6]    d0[5]   d1[5]
    4       11  8   6   5       b0[6]   b1[6]   b0[7]   b1[7]   a[9]    a[10]
    5       11  6   7   6       a[9]    a[10]   c[7]    c[6]    d0[5]   d1[5]
    6       12  7   7   5       b0[6]   b1[6]   a[11]   c[6]    a[9]    a[10]
    7       12  6   7   6       a[9]    a[10]   a[11]   c[6]    d0[5]   d1[5]
    -------------------------------------------------------------------------
    Table C.2.23 - Endpoint Bit Mode

    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 ); }

    HDR Endpoint Mode 14

    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 behaviour.

    HDR Endpoint Mode 15

    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:

    Value   7   6   5   4   3   2   1   0
    -----   ------------------------------
    v6     |M0 |        A[6:0]            |
    -----   ------------------------------
    v7     |M1 |        B[6:0]            |
    -----   ------------------------------
    Table C.2.24 - HDR Mode 15 Alpha Value Layout

    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.

    C.2.16  Weight Decoding
    -----------------------
    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.

    C.2.17  Weight Unquantization
    -----------------------------

    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 follows:

    Range   0   1   2   3   4
    --------------------------
    0..2    0   32  63  -   -
    0..4    0   16  32  47  63
    --------------------------
    Table C.2.25 - Weight Unquantization Values

    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 follows:

    Range   T Q B   Bits    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
    -------------------------------------------------------
    Table C.2.26 - Weight Unquantization Parameters

    These are then processed as follows:

        T = D * C + B;
        T = T ^ A;
        T = (A & 0x20) | (T >> 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 (T > 32) { T += 1; }

    This allows the implementation to use 64 as a divisor during inter-
    polation, which is much easier than using 63.

    C.2.18  Weight Infill
    ---------------------

    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 two dimensions along the s and t
    axes (Bs, Bt). Texel coordinates within the block (s,t) 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)

        Ds = floor( (1024 + floor(Bs/2)) / (Bs-1) );
        Dt = floor( (1024 + floor(Bt/2)) / (Bt-1) );

    Since the block dimensions are constrained, these are easily looked up
    in a table. These scale factors are then used to scale the (s,t)
    coordinates to a homogeneous coordinate (cs, ct):

        cs = Ds * s;
        ct = Dt * t;

    This homogeneous coordinate (cs, ct) is then scaled again to give
    a coordinate (gs, gt) in the weight-grid space . The weight-grid is
    of size (N, M), as specified in the block mode field:

        gs = (cs*(N-1)+32) >> 6;
        gt = (ct*(M-1)+32) >> 6;

    The resulting coordinates may be in the range 0..176. These are inter-
    preted as 4:4 unsigned fixed point numbers in the range 0.0 .. 11.0.

    If we label the integral parts of these (js, jt) and the fractional
    parts (fs, ft), then:

        js = gs >> 4; fs = gs & 0x0F;
        jt = gt >> 4; ft = gt & 0x0F;

    These values are then used to bilinearly interpolate between the stored
    weights.

        v0 = js + jt*N;
        p00 = decode_weight(v0);
        p01 = decode_weight(v0 + 1);
        p10 = decode_weight(v0 + N);
        p11 = decode_weight(v0 + N + 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;

    C.2.19  Weight Application
    --------------------------
    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 enabled, the top 8 bits of the interpolation
    result for the R, G and B channels are passed to the external sRGB
    conversion block. Otherwise, if C = 65535, then 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.

    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.

    C.2.20  Dual-Plane Decoding
    ---------------------------
    If dual-plane mode is disabled, all of the endpoint components are inter-
    polated 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 follows:

    Value   Weight 0  Weight 1
    --------------------------
    0         GBA        R
    1         RBA        G
    2         RGA        B
    3         RGB        A
    --------------------------
    Table C.2.28 - Dual Plane Color Component Selector Values

    The CCS bits are stored at a variable position directly below the weight
    bits and any additional CEM bits.

    C.2.21  Partition Pattern Generation
    ------------------------------------

    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

    C.2.22  Data Size Determination
    -------------------------------

    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 = M * N * Q; // 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.

    C.2.23  Void-Extent Blocks
    --------------------------

    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.

    The layout of a 2D Void-Extent block is as follows:

    127 126 125 124 123 122 121 120 119 118 117 116 115 114 113 112
     ---------------------------------------------------------------
    |                 Block color A component                       |
     ---------------------------------------------------------------

    111 110 109 108 107 106 105 104 103 102 101 100 99  98  97  96
    ----------------------------------------------------------------
    |                 Block color B component                       |
    ----------------------------------------------------------------

    95  94  93  92  91  90  89  88  87  86  85  84  83  82  81  80
    ----------------------------------------------------------------
    |                 Block color G component                       |
    ----------------------------------------------------------------
    79  78  77  76  75  74  73  72  71  70  69  68  67  66  65  64
    ----------------------------------------------------------------
    |                 Block color R component                       |
    ----------------------------------------------------------------

    63  62  61  60  59  58  57  56  55  54  53  52  51  50  49  48
    ----------------------------------------------------------------
    |    Void-extent maximum T coordinate              |    Min T   |
    ----------------------------------------------------------------

    47  46  45  44  43  42  41  40  39  38  37  36  35  34  33  32
    ----------------------------------------------------------------
    Void-extent minimum T coordinate       |   Void-extent max S    |
    ----------------------------------------------------------------

    31  30  29  28  27  26  25  24  23  22  21  20  19  18  17  16
    ----------------------------------------------------------------
    Void-extent max S coord    |  Void-extent minimum S coordinate  |
    ----------------------------------------------------------------
    15  14  13  12  11  10   9   8   7   6   5   4   3   2   1   0
    ----------------------------------------------------------------
    Min S coord    | 1 | 1 | D | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 0 | 0  |
    ----------------------------------------------------------------
    -------------------------------------------------
    Figure C.7 - 2D Void-Extent Block Layout Overview

    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 GL
    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 2^13-1
    or 2^9-1, 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.

    C.2.24  Illegal Encodings
    -------------------------

    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:

    *   The block mode specified is one of the modes explicitly listed
        as Reserved.
    *   A 2D void-extent block that has any of the reserved bits not
        set to 1.
    *   A block mode has been specified that would require more than
        64 weights total.
    *   A block mode has been specified that would require more than
        96 bits for integer sequence encoding of the weight grid.
    *   A block mode has been specifed that would require fewer than
        24 bits for integer sequence encoding of the weight grid.
    *   The size of the weight grid exceeds the size of the block footprint
        in any dimension.
    *   Color endpoint modes have been specified such that the color
        integer sequence encoding would require more than 18 integers.
    *   The number of bits available for color endpoint encoding after all
        the other fields have been counted is less than ceil(13C/5) where C
        is the number of color endpoint integers (this would restrict color
        integers to a range smaller than 0..5, which is not supported).
    *   Dual weight mode is enabled for a block with 4 partitions.
    *   Void-Extent blocks where the low coordinate for some texture axis
        is greater than or equal to the high coordinate.

    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.

    C.2.25  LDR PROFILE SUPPORT
    ---------------------------

    Implementations of the LDR Profile must satisfy the following requirements:

    *   All textures with valid encodings for LDR Profile must decode
        identically using either a LDR Profile, HDR Profile, or Full Profile
        decoder.
    *   All features included only in the HDR Profile or Full Profile must be
        treated as reserved in the LDR Profile, and return the error color on
        decoding.
    *   Any sequence of API calls valid for the LDR Profile must also be valid
        for the HDR Profile or Full Profile and return identical results when
        given a texture encoded for the LDR Profile.

    The feature subset for the LDR profile is:

    *   2D textures only, including 2D, 2D array, cube map face,
        and cube map array texture targets.
    *   Only those block sizes listed in Table C.2.2 are supported.
    *   LDR operation mode only.
    *   Only LDR endpoint formats must be supported, namely formats
        0, 1, 4, 5, 6, 8, 9, 10, 12, 13.
    *   Decoding from a HDR endpoint results in the error color.
    *   Interpolation returns UNORM8 results when used in conjunction
        with sRGB.
    *   LDR void extent blocks must be supported, but void extents
        may not be checked."

    If only the LDR profile is supported, read this extension by striking
    all descriptions of HDR modes and decoding algorithms. The extension
    documents how to modify the document for some particularly tricky cases,
    but the general rule is as described in this paragraph.

Interactions with immutable-format texture images

    ASTC texture formats are supported by immutable-format textures only if
    such textures are supported by the underlying implementation (e.g.
    OpenGL 4.1 or later, OpenGL ES 3.0 or later, or earlier versions
    supporting the GL_EXT_texture_storage extension). Otherwise, remove all
    references to the Tex*Storage* commands from this specification.

Interactions with texture cube map arrays

    ASTC textures are supported for the TEXTURE_CUBE_MAP_ARRAY target only
    when cube map arrays are supported by the underlying implementation
    (e.g. OpenGL 4.0 or later, or an OpenGL or OpenGL ES version supporting
    an extension defining cube map arrays). Otherwise, remove all references
    to texture cube map arrays from this specification.

Interactions with OpenGL (all versions)

    ASTC is not supported for 1D textures and texture rectangles, and does
    not support non-zero borders.

    Add the following error conditions to CompressedTexImage*D:

    "An INVALID_ENUM error is generated by CompressedTexImage1D if
    <internalformat> is one of the ASTC formats.

    An INVALID_OPERATION error is generated by CompressedTexImage2D
    and CompressedTexImage3D if <internalformat> is one of the ASTC
    formats and <border> is non-zero."

    Add the following error conditions to CompressedTexSubImage*D:

    "An INVALID_ENUM error is generated by CompressedTex*SubImage1D
    if the internal format of the texture is one of the ASTC formats.

    An INVALID_OPERATION error is generated by CompressedTex*SubImage2D
    if the internal format of the texture is one of the ASTC formats
    and <border> is non-zero."

    Add the following error conditions to TexStorage1D and TextureStorage1D:

    "An INVALID_ENUM error is generated by TexStorage1D and TextureStorage1D
    if <format> is one of the ASTC formats."

    Add the following error conditions to TexStorage2D and TextureStorage2D
    for versions of OpenGL that support texture rectangles:

    "An INVALID_OPERATON error is generated by TexStorage2D and
    TextureStorage2D if <format> is one of the ASTC formats and <target>
    is TEXTURE_RECTANGLE.

Interactions with OpenGL 4.2

    OpenGL 4.2 supports the feature that compressed textures can be
    compressed online, by passing the compressed texture format enum as
    the internal format when uploading a texture using TexImage1D,
    TexImage2D or TexImage3D (see Section 3.9.3, Texture Image
    Specification, subsection Encoding of Special Internal Formats).

    Due to the complexity of the ASTC compression algorithm, it is not
    usually suitable for online use, and therefore ASTC support will be
    limited to pre-compressed textures only. Where on-device compression
    is required, a domain-specific limited compressor will typically
    be used, and this is therefore not suitable for implementation in
    the driver.

    In particular, the ASTC format specifiers will not be added to
    Table 3.14, and thus will not be accepted by the TexImage*D
    functions, and will not be returned by the (already deprecated)
    COMPRESSED_TEXTURE_FORMATS query.

Issues

 1) Three-dimensional block ASTC formats (e.g. formats whose block depth
    is greater than one) are not supported by these extensions.

 2) The first release of the extension was not clear about the
    restrictions of the LDR profile and did not document interactions
    with cube map array textures.

    RESOLVED. This extension has been rewritten to be based on OpenGL ES
    3.1, to clearly document LDR restrictions, and to add cube map array
    texture interactions.

Revision History

    Revision 8, June 8, 2017 - Added missing interactions with OpenGL.

    Revision 7, July 14, 2016 - Clarified definition of 2D void-extent
    blocks.

    Revision 6, March 8, 2016 - Clarified that sRGB transform is not
    applied to Alpha channel.

    Revision 5, September 15, 2015 - fix typo in third paragraph of section
    8.7.

    Revision 4, June 24, 2015 - minor cleanup from feedback. Move Issues and
    Interactions sections to the end of the document. Merge some language
    from OpenGL ES specification edits and rename some tables to figures,
    due to how they're generated in the core specifications. Include a
    description of the "Cube Map Array Texture" column added to table 3.19
    and expand the description of how to read this document when supporting
    only the LDR profile (Bug 13921).

    Revision 3, May 28, 2015 - rebase extension on OpenGL ES 3.1. Clarify
    texture formats and targets supported by LDR and HDR profiles. Add cube
    map array targets and an Interactions section defining when they are
    supported. Add an Interactions section for immutable-format textures
    (Bug 13921).

    Revision 2, April 28, 2015 - added CompressedTex{Sub,}Image3D to
    commands accepting ASTC format tokens in the New Tokens section (Bug
    10183).