[WORK IN PROGRESS: Spec should be complete, but needs to be converted to markdown with more external links added.]

File: basis_spec.txt
Version 1.01

1.0 Introduction
----------------

The Basis Universal GPU texture codec supports reading and writing ".basis" files. 
The .basis file format supports ETC1S or UASTC 4x4 texture data.

* ETC1S is a simplified subset of ETC1.

The mode is always differential (diff bit=1), the Rd, Gd, and Bd color deltas 
are always (0,0,0), and the flip bit is always set. ETC1S texture data is fully 
compliant with all existing software and hardware ETC1 decoders. Existing encoders 
can be easily modified to limit their output to ETC1S.

* UASTC 4x4 is a 19 mode subset of the ASTC texture format. Its specification is 
[here](https://github.com/BinomialLLC/basis_universal/wiki/UASTC-Texture-Specification). UASTC texture data can always be losslessly transcoded to ASTC.

2.0 High-Level File Structure
-----------------------------

A .basis file consists of multiple sections. Apart from the header, which must always
be at the start of the file, the other sections may appear in any order. 

Here's the high level organization of a typical .basis file:

* The file header
* Optional ETC1S compressed endpoint/selector codebooks
* Optional ETC1S Huffman table information
* A required "slice" description array describing the resolutions and file offset/compressed sizes of each texture slice present in the file
* 1 or more slices containing ETC1S or UASTC compressed texture data. 
* For future expansion, the format supports an "extended" header which may be located anywhere in the file. This section contains .PNG-like chunked data. 

3.0 File Enums
--------------

// basis_file_header::m_tex_type

enum basis_texture_type
{
  cBASISTexType2D = 0,     
  cBASISTexType2DArray = 1,   
  cBASISTexTypeCubemapArray = 2, 
  cBASISTexTypeVideoFrames = 3, 
  cBASISTexTypeVolume = 4,  
  cBASISTexTypeTotal
};

// basis_slice_desc::flags

enum basis_slice_desc_flags
{
  cSliceDescFlagsHasAlpha = 1,
  cSliceDescFlagsFrameIsIFrame = 2   
};

// basis_file_header::m_tex_format 

enum basis_tex_format
{
  cETC1S = 0,
  cUASTC4x4 = 1
};

// basis_file_header::m_flags 

enum basis_header_flags
{
  cBASISHeaderFlagETC1S = 1.
  cBASISHeaderFlagYFlipped = 2,
  cBASISHeaderFlagHasAlphaSlices = 4
};

4.0 File Structures
-------------------

All individual members in all file structures are byte aligned and little endian. The structs 
have no padding (i.e. they are declared with #pragma pack(1)).

4.1 "basis_file_header" structure
---------------------------------

The file header must always be at the beginning of the file.

struct basis_file_header
{
  uint16      m_sig;              // 2 byte file signature
  uint16      m_ver;              // File version
  uint16      m_header_size;      // Header size in bytes, sizeof(basis_file_header) or 0x4D
  uint16      m_header_crc16;     // CRC16/genibus of the remaining header data

  uint32      m_data_size;        // The total size of all data after the header
  uint16      m_data_crc16;       // The CRC16 of all data after the header

  uint24      m_total_slices;     // The number of compressed slices 
  uint24      m_total_images;     // The total # of images

  byte        m_tex_format;       // enum basis_tex_format
  uint16      m_flags;            // enum basis_header_flags
  byte        m_tex_type;         // enum basis_texture_type
  uint24      m_us_per_frame;     // Video: microseconds per frame

  uint32      m_reserved;         // For future use
  uint32      m_userdata0;        // For client use
  uint32      m_userdata1;        // For client use

  uint16      m_total_endpoints;          // ETC1S: The number of endpoints in the endpoint codebook 
  uint32      m_endpoint_cb_file_ofs;     // ETC1S: The compressed endpoint codebook's file offset relative to the header
  uint24      m_endpoint_cb_file_size;    // ETC1S: The compressed endpoint codebook's size in bytes

  uint16      m_total_selectors;          // ETC1S: The number of selectors in the selector codebook 
  uint32      m_selector_cb_file_ofs;     // ETC1S: The compressed selector codebook's file offset relative to the header
  uint24      m_selector_cb_file_size;    // ETC1S: The compressed selector codebook's size in bytes

  uint32      m_tables_file_ofs;          // ETC1S: The file offset of the compressed Huffman codelength tables.
  uint32      m_tables_file_size;         // ETC1S: The file size in bytes of the compressed Huffman codelength tables.

  uint32      m_slice_desc_file_ofs;      // The file offset to the slice description array, usually follows the header
  uint32      m_extended_file_ofs;        // The file offset of the "extended" header and compressed data, for future use
  uint32      m_extended_file_size;       // The file size in bytes of the "extended" header and compressed data, for future use
};

4.1.1 Details:

* m_sig is always 'B' * 256 + 's', or 0x4273.
* m_ver is currently always 0x10.
* m_header_size is sizeof(basis_file_header). It's always 0x4D.
* m_header_crc16 is the CRC-16 of the remaining header data. See the "CRC-16" section for more information.
* m_data_size, m_data_crc16: The size of all data following the header, and its CRC-16.
* m_total_slices: The total number of slices, from [1,2^24-1]
* m_total_images: The total number of images (where one image can contain multiple mipmap levels, and each mipmap level is a different slice).
* m_tex_format: basis_tex_format. Either cETC1S (0), or cUASTC4x4 (1).
* m_flags: A combination of flags from the basis_header_flags enum.
* m_tex_type: The texture type, from enum basis_texture_type
* m_us_per_frame: Microseconds per frame, only valid for cBASISTexTypeVideoFrames texture types.
* m_total_endpoints, m_endpoint_cb_file_ofs, m_endpoint_cb_file_size: Information about the compressed ETC1S endpoint codebook: The total # of entries, the offset to the compressed data, and the compressed data's size.
* m_total_selectors, m_selector_cb_file_ofs, m_selector_cb_file_size: Information about the compressed ETC1S selector codebook: The total # of entries, the offset to the compressed data, and the compressed data's size.
* m_tables_file_ofs, m_tables_file_size: The file offset and size of the compressed Huffman tables for ETC1S format files. 
* m_slice_desc_file_ofs: The file offset to the array of slice description structures. There will be m_total_slices structures at this file offset.
* m_extended_file_ofs, m_extended_file_size: The "extended" header, for future expansion. Currently unused.

4.2 "basis_slice_desc" structure
--------------------------------

struct basis_slice_desc
{
    uint24 m_image_index;  
    uint8 m_level_index;   
    uint8 m_flags;         

    uint16 m_orig_width;   
    uint16 m_orig_height;  

    uint16 m_num_blocks_x; 
    uint16 m_num_blocks_y; 

    uint32 m_file_ofs;     
    uint32 m_file_size;    

    uint16 m_slice_data_crc16; 
};

4.2.1 Details:

* m_image_index: The index of the source image provided to the encoder (will always appear in order from first to last, first image index is 0, no skipping allowed)
* m_level_index: The mipmap level index (mipmaps will always appear from largest to smallest)
* m_flags: enum basis_slice_desc_flags
* m_orig_width: The original image width (may not be a multiple of 4 pixels)
* m_orig_height: The original image height (may not be a multiple of 4 pixels)
* m_num_blocks_x: The slice's block X dimensions. Each block is 4x4 pixels. The slice's pixel resolution may or may not be a power of 2.
* m_num_blocks_y: The slice's block Y dimensions. 
* m_file_ofs: Offset from the header to the start of the slice's data
* m_file_size: The size of the compressed slice data in bytes
* m_slice_data_crc16: The CRC16 of the compressed slice data, for extra-paranoid use cases

5.0 CRC-16 Function
-------------------

.basis files use CRC-16/genibus(aka CRC-16 EPC, CRC-16 I-CODE, CRC-16 DARC) format CRC-16's. 

Here's an example function in C++:

uint16_t crc16(const void* r, size_t size, uint16_t crc)
{
  crc = ~crc;
  const uint8_t* p = static_cast<const uint8_t*>(r);
  for ( ; size; --size)
  {
    const uint16_t q = *p++ ^ (crc >> 8);
    uint16_t k = (q >> 4) ^ q;
    crc = (((crc << 8) ^ k) ^ (k << 5)) ^ (k << 12);
  }

  return static_cast<uint16_t>(~crc);
}

This function is called with 0 in the final "crc" parameter when computing CRC-16's of file data.

6.0 Compressed Huffman Tables
-----------------------------

ETC1S format .basis files rely heavily on static [canonical Huffman
prefix coding](https://en.wikipedia.org/wiki/Canonical_Huffman_code).  Multiple
Huffman tables are used by each compressed section. Huffman codes are stored in
each output byte in LSB to MSB order. (This is opposite of the JPEG format,
which stores the codes in MSB to LSB order.)

Huffman coding in .basis is compatible with the canonical Huffman methods used
by Deflate encoders/decoders. Section 3.2.2 of [Deflate - RFC
1951](https://tools.ietf.org/html/rfc1951), which describes how to compute the
value of each Huffman code given an array of symbol codelengths. This document
assumes familiarity with how Huffman coding works in Deflate.

First, some enums:

enum
{
    // Max supported Huffman code size is 16-bits
    cHuffmanMaxSupportedCodeSize = 16, 

    // The maximum number of symbols  is 2^14
    cHuffmanMaxSymsLog2 = 14, 
    cHuffmanMaxSyms = 1 << cHuffmanMaxSymsLog2,

    // Small zero runs may range from 3-10 entries
    cHuffmanSmallZeroRunSizeMin = 3, 
    cHuffmanSmallZeroRunSizeMax = 10, 
    cHuffmanSmallZeroRunExtraBits = 3,

    // Big zero runs may range from 11-138 entries
    cHuffmanBigZeroRunSizeMin = 11, 
    cHuffmanBigZeroRunSizeMax = 138, 
    cHuffmanBigZeroRunExtraBits = 7,

    // Small non-zero runs may range from 3-6 entries
    cHuffmanSmallRepeatSizeMin = 3, 
    cHuffmanSmallRepeatSizeMax = 6, 
    cHuffmanSmallRepeatExtraBits = 2,

    // Big non-zero run may range from 7-134 entries
    cHuffmanBigRepeatSizeMin = 7, 
    cHuffmanBigRepeatSizeMax = 134, 
    cHuffmanBigRepeatExtraBits = 7,

    // There are a maximum of 21 symbols in a compressed Huffman code length table.
    cHuffmanTotalCodelengthCodes = 21, 
    
    // Symbols [0,16] indicate code sizes. Other symbols indicate zero runs or repeats:
    cHuffmanSmallZeroRunCode = 17, 
    cHuffmanBigZeroRunCode = 18, 
    cHuffmanSmallRepeatCode = 19, 
    cHuffmanBigRepeatCode = 20
};

A .basis Huffman table consists of 1 to cHuffmanMaxSyms symbols. Each compressed
Huffman table is described by an array of symbol code lengths in bits.

The table's symbol code lengths are themselves RLE+Huffman coded, just like
Deflate. (Note this can be confusing to developers unfamiliar with Deflate.)
Each table begins with a small fixed header:

    14 bits: total_used_syms [1, cHuffmanMaxSyms]
    5 bits: num_codelength_codes [1, cHuffmanTotalCodelengthCodes]
    
Next, the code lengths for the small Huffman table which is used to send the compressed codelengths (and RLE/repeat codes) are sent uncompressed but in a reordered manner:
    
    3*num_codelength_codes bits: Code size of each Huffman symbol for the compressed Huffman codelength table.
    
    These code lengths are sent in this order (to help reduce the number that must be sent):
    
    { 
        cHuffmanSmallZeroRunCode, cHuffmanBigZeroRunCode, cHuffmanSmallRepeatCode, cHuffmanBigRepeatCode, 
        0, 8, 7, 9, 6, 0xA, 5, 0xB, 4, 0xC, 3, 0xD, 2, 0xE, 1, 0xF, 0x10 
    };
            
A canonical Huffman decoding table (of up to 21 symbols) should be built from
these code lengths. Immediately following this data are the Huffman symbols
(sometimes intermixed with raw bits) which describe how to unpack the
codelengths of each symbol in the Huffman table:

    - Symbols [0,16] indicate a specific symbol code length in bits.
    
    - Symbol cHuffmanSmallZeroRunCode (17) indicates a short run of symbols with 0 bit code lengths.
      cHuffmanSmallZeroRunExtraBits (3) bits are sent after this symbol, which indicates the run's size after adding the minimum size (cHuffmanSmallZeroRunSizeMin).
      
    - Symbol cHuffmanBigZeroRunCode (18) indicates a long run of symbols with 0 bit code lengths. 
      cHuffmanBigZeroRunExtraBits (7) bits are sent after this symbol, which indicates the run's size after adding the minimum size (cHuffmanBigZeroRunSizeMin)

    - Symbol cHuffmanSmallRepeatCode (19) indicates a short run of symbols that repeat the previous symbol's code length.
      cHuffmanSmallRepeatExtraBits (2) bits are sent after this symbol, which indicates the number of times to repeat the previous symbol's code length, 
      after adding the minimum size (cHuffmanSmallRepeatSizeMin).
      Cannot be the first symbol, and the previous symbol cannot have a code length of 0.
      
    - Symbol cHuffmanBigRepeatCode (20) indicates a short run of symbols that repeat the previous symbol's code length.
      cHuffmanBigRepeatExtraBits (7) bits are sent after this symbol, which indicates the number of times to repeat the previous symbol's code length,
      after adding the minimum size (cHuffmanBigRepeatSizeMin).
      Cannot be the first symbol, and the previous symbol cannot have a code length of 0.
      
There should be exactly total_used_syms code lengths stored in the compressed Huffman table. If not the stream is either corrupted or invalid.

After all the symbol codelengths are uncompressed, the symbol codes can be computed and the canonical Huffman decoding tables can be built.

7.0 ETC1S Endpoint Codebooks
----------------------------

The endpoint codebook section starts at file offset
basis_file_header::m_endpoint_cb_file_ofs and is m_endpoint_cb_file_size bytes
long. The endpoint codebook will have basis_file_header::m_total_endpoints total
entries.

At the beginning of the compressed endpoint codebook section are four compressed
Huffman tables, stored using the procedure outlined in section 6.0. The Huffman tables
appear in this order:

    1. color5_delta_model0
    2. color5_delta_model1
    3. color5_delta_model2
    4. inten_delta_model

Following the data for these Huffman tables is a single 1-bit code which
indicates if the color endpoint codebook is grayscale or not. 

Immediately following this code is the compressed color endpoint codebook data. 
A simple form of DPCM (Delta Pulse Code Modulation) coding is used to send the
ETC1S intensity table indices and color values. Here is the procedure to decode
the endpoint codebook:

    const int COLOR5_PAL0_PREV_HI = 9, COLOR5_PAL0_DELTA_LO = -9, COLOR5_PAL0_DELTA_HI = 31;
    const int COLOR5_PAL1_PREV_HI = 21, COLOR5_PAL1_DELTA_LO = -21, COLOR5_PAL1_DELTA_HI = 21;
    const int COLOR5_PAL2_PREV_HI = 31, COLOR5_PAL2_DELTA_LO = -31, COLOR5_PAL2_DELTA_HI = 9;

    // Assume previous endpoint color is (16, 16, 16), and the previous intensity is 0.
    color32 prev_color5(16, 16, 16, 0);
    uint32_t prev_inten = 0;

    // For each endpoint codebook entry
    for (uint32_t i = 0; i < num_endpoints; i++)
    {
        // Decode the intensity delta Huffman code
        uint32_t inten_delta = decode_huffman(inten_delta_model);
        endpoints[i].m_inten5 = static_cast<uint8_t>((inten_delta + prev_inten) & 7);
        prev_inten = endpoints[i].m_inten5;

        // Now decode the endpoint entry's color or intensity value
        for (uint32_t c = 0; c < (endpoints_are_grayscale ? 1U : 3U); c++)
        {
            // The Huffman table we used to decode the delta depends on the previous color's value
            int delta;
            if (prev_color5[c] <= basist::COLOR5_PAL0_PREV_HI)
                delta = decode_huffman(color5_delta_model0);
            else if (prev_color5[c] <= basist::COLOR5_PAL1_PREV_HI)
                delta = decode_huffman(color5_delta_model1);
            else
                delta = decode_huffman(color5_delta_model2);

            // Apply the delta
            int v = (prev_color5[c] + delta) & 31;

            endpoints[i].m_color5[c] = static_cast<uint8_t>(v);

            prev_color5[c] = static_cast<uint8_t>(v);
        }

        // If the endpoints are grayscale, set G and B to match R.
        if (endpoints_are_grayscale)
        {
            endpoints[i].m_color5[1] = endpoints[i].m_color5[0];
            endpoints[i].m_color5[2] = endpoints[i].m_color5[0];
        }
    }

The rest of the section's data (if any) can be ignored.

8.0 ETC1S Selector Codebooks
----------------------------

The selector codebook section starts at file offset
basis_file_header::m_selector_cb_file_ofs and is m_selector_cb_file_size bytes
long. The selector codebook will have basis_file_header::m_total_selectors total
entries.

The first bit of this section indicates if "global" selector codebooks are used.
Basis Universal doesn't currently utilize global selector codebooks, so this bit
should always be 0.

The second bit of this section indicates if "hybrid" global/local selector
codebooks are used. Hybrid codebooks are not supported either, so this bit
should always be 0.

The third bit indicates of the selector codebook has been sent in raw form
(uncompressed). If it's set, each selector is sent as four 8-bit bytes. Each
byte corresponds to four 2-bit ETC1S selectors. The first selector of each group
of 4 selectors starts at the LSB (least significant bit) of each byte, and is
2-bits wide.

If the third bit is 0, the selectors have been DPCM coded with Huffman coding. 
The "delta_selector_pal_model" Huffman table will immediately follow the third
bit, and is stored using the procedure outlined in section 6.0.

Here is the DPCM decoding procedure for selector codebooks:

        uint8_t prev_bytes[4] = { 0, 0, 0, 0 };

        for (uint32_t i = 0; i < num_selectors; i++)
        {
            if (!i)
            {
                // First selector is sent raw
                for (uint32_t j = 0; j < 4; j++)
                {
                    uint32_t cur_byte = get_bits(8);
                    prev_bytes[j] = static_cast<uint8_t>(cur_byte);

                    for (uint32_t k = 0; k < 4; k++)
                        selectors[i].set_selector(k, j, (cur_byte >> (k * 2)) & 3);
                }
                selectors[i].init_flags();
                continue;
            }

            // Subsequent selectors are sent with a simple form of byte-wise DPCM coding.
            for (uint32_t j = 0; j < 4; j++)
            {
                int delta_byte = decode_huffman(delta_selector_pal_model);

                uint32_t cur_byte = delta_byte ^ prev_bytes[j];
                prev_bytes[j] = static_cast<uint8_t>(cur_byte);

                for (uint32_t k = 0; k < 4; k++)
                    selectors[i].set_selector(k, j, (cur_byte >> (k * 2)) & 3);
            }
        }

Any bytes in this section following the selector codebook bits can be safely ignored.

9.0 ETC1S Compressed Slice Decoding Huffman Tables
--------------------------------------------------

Each ETC1S slice is compressed with four Huffman tables stored using the
procedural outlined in section 6.0. These Huffman tables are stored at file
offset basis_file_header::m_tables_file_ofs. This section will be 
basis_file_header::m_tables_file_size bytes long.

The following four Huffman tables are sent, in this order:

    1. endpoint_pred_model
    2. delta_endpoint_model
    3. selector_model
    4. selector_history_buf_rle_model

Following the last Huffman table are 13-bits indicating the size of the selector
history buffer. Any remaining bits may be safely ignored.

10. ETC1S Slice Decoding
------------------------

ETC1S slices consist of a compressed 2D array of ETC1S blocks, always compressed
in top-down/left-right raster order. For texture video, the previous slice's
already decoded contents may be referred to when blocks are encoded using
Conditional Replenishment (also known as "skip blocks"). 

Each ETC1S block is encoded by using references to the color endpoint codebook
and the selector codebook. Sections 10.1 and 10.2 describe the helper procedures
using by the decoder, and section 10.3 describes how the array of ETC1S blocks
is actually decoded.

10.1 Approximate Move to Front Routines
---------------------------------------

An approximate Move to Front (MTF) approach is used to efficiently encode the
selector codebook references. Here is the C++ example class for approximate MTF
decoding:

    class approx_move_to_front
    {
    public:
        approx_move_to_front(uint32_t n)
        {
            init(n);
        }

        void init(uint32_t n)
        {
            m_values.resize(n);
            m_rover = n / 2;
        }

        size_t size() const { return m_values.size(); }

        const int& operator[] (uint32_t index) const { return m_values[index]; }
              int operator[] (uint32_t index)        { return m_values[index]; }

        void add(int new_value)
        {
            m_values[m_rover++] = new_value;
            if (m_rover == m_values.size())
                m_rover = (uint32_t)m_values.size() / 2;
        }

        void use(uint32_t index)
        {
            if (index)
            {
                int x = m_values[index / 2];
                int y = m_values[index];
                m_values[index / 2] = y;
                m_values[index] = x;
            }
        }
        
    private:
        std::vector<int> m_values;
        uint32_t m_rover;
    };

10.2 VLC Decoding Procedure
---------------------------

ETC1S slice decoding utilizes a simple Variable Length Coding (VLC) scheme that
sends raw bits using variable-size chunks. Here is the VLC decoding procedure:

    uint32_t decode_vlc(uint32_t chunk_bits)
    {
        assert(chunk_bits);

        const uint32_t chunk_size = 1 << chunk_bits;
        const uint32_t chunk_mask = chunk_size - 1;
                
        uint32_t v = 0;
        uint32_t ofs = 0;

        for ( ; ; )
        {
            uint32_t s = get_bits(chunk_bits + 1);
            v |= ((s & chunk_mask) << ofs);
            ofs += chunk_bits;

            if ((s & chunk_size) == 0)
                break;
            
            if (ofs >= 32)
            {
                assert(0);
                break;
            }
        }

        return v;
    }

10.3 ETC1S Slice Block Decoding
-------------------------------

Each slice has a corresponding "basis_slice_desc" structure, described in section
4.2. The slice's dimensions in ETC1S blocks are stored in
basis_slice_desc::m_num_blocks_x and basis_slice_desc::m_num_blocks_y. Each
slice is located at file offset basis_slice_desc::m_file_ofs, and is
basis_slice_desc::m_file_size bytes long.

The decoder iterates through all the slice blocks in top-down, left-right raster
order. Each block is represented by an index into the color endpoint codebook
and another index into the selector endpoint codebook. The endpoint codebook
contains each ETC1S block's base RGB color and intensity table information, and
the selector codebook contains the 4x4 texel selector entry (which are 2-bits
each) information. This is all the information needed to fully represent the
texels within each block.

The decoding procedure loops over all the blocks in raster order, and decodes
the endpoint and selector indices used to represent each block. The decoding
procedural is complex enough that commented code is best used to describe it.

Here's the slice decoding procedure. This block of code shows the block loop,
and how endpoint codebook indices are decoded. The next block of code shows how
selector codebook indices are decoded.

    // Constants used by the decoder
    const uint32_t ENDPOINT_PRED_TOTAL_SYMBOLS = (4 * 4 * 4 * 4) + 1;
    const uint32_t ENDPOINT_PRED_REPEAT_LAST_SYMBOL = ENDPOINT_PRED_TOTAL_SYMBOLS - 1;
    const uint32_t ENDPOINT_PRED_MIN_REPEAT_COUNT = 3;
    const uint32_t ENDPOINT_PRED_COUNT_VLC_BITS = 4;

    const uint32_t NUM_ENDPOINT_PREDS = 3;
    const uint32_t CR_ENDPOINT_PRED_INDEX = NUM_ENDPOINT_PREDS - 1;
    const uint32_t NO_ENDPOINT_PRED_INDEX = 3;
    
    // Endpoint/selector codebooks - decoded previously. See sections 7.0 and 8.0.
    endpoint endpoints[endpoint_codebook_size];
    selector selectors[selector_codebook_size]; 
    
    // Array of per-block values used for endpoint index prediction (enough for 2 rows).
    struct block_preds
    {
        uint16_t m_endpoint_index;
        uint8_t m_pred_bits;
    };
    block_preds block_endpoint_preds[2][num_blocks_x];
    
    // State used during block decoding
    uint32_t cur_pred_bits = 0;
    int prev_endpoint_pred_sym = 0;
    int endpoint_pred_repeat_count = 0;
    uint32_t prev_endpoint_index = 0;

    // This array is only used for texture video. It holds the previous frame's endpoint and selector indices (each 16-bits, for 32-bits total).
    uint32_t prev_frame_indices[block_x][block_y]; 
    
    // Selector history buffer - See section 10.1.
    approx_move_to_front selector_history_buf;

    // Loop over all slice blocks in raster order
    for (uint32_t block_y = 0; block_y < num_blocks_y; block_y++)
    {
        // The index into the block_endpoint_preds array
        const uint32_t cur_block_endpoint_pred_array = block_y & 1;

        for (uint32_t block_x = 0; block_x < num_blocks_x; block_x++)
        {
            // Check if we're at the start of a 2x2 block group.
            if ((block_x & 1) == 0)
            {
                // Are we on an even or odd row of blocks?
                if ((block_y & 1) == 0)
                {
                    // We're on an even row and column of blocks. Decode the combined endpoint index predictor symbols for 2x2 blocks.
                    // This symbol tells the decoder how the endpoints are decoded for each block in a 2x2 group of blocks.
                                        
                    // Are we in an RLE run?
                    if (endpoint_pred_repeat_count)
                    {
                        // Inside a run of endpoint predictor symbols.
                        endpoint_pred_repeat_count--;
                        cur_pred_bits = prev_endpoint_pred_sym;
                    }
                    else
                    {
                        // Decode the endpoint prediction symbol, using the "endpoint pred" Huffman table (see section 9.0).
                        cur_pred_bits = decode_huffman(m_endpoint_pred_model);
                        if (cur_pred_bits == ENDPOINT_PRED_REPEAT_LAST_SYMBOL)
                        {
                            // It's a run of symbols, so decode the count using VLC decoding (see section 10.2)
                            endpoint_pred_repeat_count = decode_vlc(ENDPOINT_PRED_COUNT_VLC_BITS) + ENDPOINT_PRED_MIN_REPEAT_COUNT - 1;

                            cur_pred_bits = prev_endpoint_pred_sym;
                        }
                        else
                        {
                            // It's not a run of symbols
                            prev_endpoint_pred_sym = cur_pred_bits;
                        }
                    }

                    // The symbol has enough endpoint prediction information for 4 blocks (2 bits per block), so 8 bits total. 
                    // Remember the prediction information we should use for the next row of 2 blocks beneath the current block.
                    block_endpoint_preds[cur_block_endpoint_pred_array ^ 1][block_x].m_pred_bits = (uint8_t)(cur_pred_bits >> 4);
                }
                else
                {
                    // We're on an odd row of blocks, so use the endpoint prediction information we previously stored on the previous even row.
                    cur_pred_bits = block_endpoint_preds[cur_block_endpoint_pred_array][block_x].m_pred_bits;
                }
            }
            
            // Decode the current block's endpoint and selector indices.
            uint32_t endpoint_index, selector_index = 0;

            // Get the 2-bit endpoint prediction index for this block.
            const uint32_t pred = cur_pred_bits & 3;

            // Get the next block's endpoint prediction bits ready.
            cur_pred_bits >>= 2;            
            
            // Now check to see if we should reuse a previously encoded block's endpoints.
            if (pred == 0)
            {
                // Reuse the left block's endpoint index
                assert(block_x > 0);
                endpoint_index = prev_endpoint_index;
            }
            else if (pred == 1)
            {
                // Reuse the upper block's endpoint index
                assert(block_y > 0)
                endpoint_index = block_endpoint_preds[cur_block_endpoint_pred_array ^ 1][block_x].m_endpoint_index;
            }
            else if (pred == 2)
            {
                if (is_video)
                {
                    // If it's texture video, reuse the previous frame's endpoint index, at this block.
                    assert(pred == CR_ENDPOINT_PRED_INDEX);
                    endpoint_index = prev_frame_indices[block_x][block_y];
                    selector_index = endpoint_index >> 16;
                    endpoint_index &= 0xFFFFU;
                }
                else
                {
                    // Reuse the upper left block's endpoint index.
                    assert((block_x > 0) && (block_y > 0));
                    endpoint_index = block_endpoint_preds[cur_block_endpoint_pred_array ^ 1][block_x - 1].m_endpoint_index;
                }
            }
            else
            {
                // We need to decode and apply a DPCM encoded delta to the previously used endpoint index.
                // This uses the delta endpoint Huffman table (see section 9.0).
                const uint32_t delta_sym = decode_huffman(delta_endpoint_model);

                endpoint_index = delta_sym + prev_endpoint_index;
                
                // Wrap around if the index goes beyond the end of the endpoint codebook
                if (endpoint_index >= endpoints.size())
                    endpoint_index -= (int)endpoints.size();
            }

            // Remember the endpoint index we used on this block, so the next row can potentially reuse the index.
            block_endpoint_preds[cur_block_endpoint_pred_array][block_x].m_endpoint_index = (uint16_t)endpoint_index;

            // Remember the endpoint index used
            prev_endpoint_index = endpoint_index;
            
            // Now we have fully decoded the ETC1S endpoint codebook index, in endpoint_index. 
            
            // Now decode the selector index (see the next block of code, below).
            < selector decoding - see below >
            
        } // block_x
    } // block_y

The compressed format allows the encoder to reuse the endpoint index used by
the previous block, the block immediately above the current block, or the
block to the upper left (if the file is not texture video). Alternately, the
encoder can send a Huffman coded DPCM encoded index relative to the
previously used endpoint index.

Which type of prediction was used by the encoder is controlled by the "endpoint
pred" (endpoint prediction) indices, which are sent with Huffman coding (using
the "endpoint_pred_model" table described in Section 9.0) once every 2x2 blocks.

For texture video, the endpoint prediction symbol normally used to refer to the
upper left block (endpoint pred index 2) instead indicates that both the
endpoint and selector indices from the previous frame's block should be reused
on the current frame's block. The endpoint pred indices are RLE coded, so this
allows the encoder to efficiently skip over a large number of unchanged blocks
in a video sequence.

The code to decode the selector codebook index immediately follows the code above for decoding the endpoint indices:

    const uint32_t MAX_SELECTOR_HISTORY_BUF_SIZE = 64;
    const uint32_t SELECTOR_HISTORY_BUF_RLE_COUNT_THRESH = 3;
    const uint32_t SELECTOR_HISTORY_BUF_RLE_COUNT_BITS = 6;
    const uint32_t SELECTOR_HISTORY_BUF_RLE_COUNT_TOTAL = (1 << SELECTOR_HISTORY_BUF_RLE_COUNT_BITS);

    // Decode selector index, unless it's texture video and the endpoint predictor indicated that the 
    // block's endpoints were reused from the previous frame.
    if ((!is_video) || (pred != CR_ENDPOINT_PRED_INDEX))
    {
        int selector_sym;

        // Are we in a selector RLE run?
        if (cur_selector_rle_count > 0)
        {
            // Handle selector RLE run.
            cur_selector_rle_count--;

            selector_sym = (int)selectors.size();
        }
        else
        {
            // Decode the selector symbol, using the selector Huffman table (see section 9.0).
            selector_sym = decode_huffman(m_selector_model);

            // Is it a run?
            if (selector_sym == static_cast<int>(SELECTOR_HISTORY_BUF_RLE_SYMBOL_INDEX))
            {
                // Decode the selector run's size, using the selector history buf RLE Huffman table (see section 9.0).
                int run_sym = decode_huffman(selector_history_buf_rle_model);

                // Is it a very long run?
                if (run_sym == (SELECTOR_HISTORY_BUF_RLE_COUNT_TOTAL - 1))
                    cur_selector_rle_count = decode_vlc(7) + SELECTOR_HISTORY_BUF_RLE_COUNT_THRESH;
                else
                    cur_selector_rle_count = run_sym + SELECTOR_HISTORY_BUF_RLE_COUNT_THRESH;

                selector_sym = (int)selectors.size();

                cur_selector_rle_count--;
            }
        }

        // Is it a reference into the selector history buffer?
        if (selector_sym >= (int)selectors.size())
        {
            assert(m_selector_history_buf_size > 0);

            // Compute the history buffer index
            int history_buf_index = selector_sym - (int)selectors.size();

            if (history_buf_index < selector_history_buf.size());

            // Access the history buffer
            selector_index = selector_history_buf[history_buf_index];

            // Update the history buffer
            if (history_buf_index != 0)
                selector_history_buf.use(history_buf_index);
        }
        else
        {
            // It's an index into the selector codebook
            selector_index = selector_sym;

            // Add it to the selector history buffer
            if (m_selector_history_buf_size)
                selector_history_buf.add(selector_index);
        }
    }
        
    // For texture video, remember the endpoint and selector indices used by the block on this frame, for later reuse on the next frame.
    if (is_video)
        prev_frame_indices[block_x + block_y * num_blocks_x] = endpoint_index | (selector_index << 16);

    // The block is fully decoded here. The codebook indices are endpoint_index and selector_index.
    // Make sure they are valid
    assert((endpoint_index < endpoints.size()) && (selector_index < selectors.size()));

At this point, the decoder has decoded each block's  endpoint and selector codebook indices.
It can now fetch the actual ETC1S endpoints/selectors from the codebooks and write out ETC1S
texture data, or it can transcode the ETC1S data to another texture format.

11.0 Example Bitstreams
-----------------------

This section will include several example .basis file bitstreams, along with their decoded equivalents, which should be helpful for new decoder verification.