webKNOSSOS wrapper is a file format designed for large-scale, three-dimensional voxel datasets. It was optimized for high-speed access to data subvolumes, and supports multi-channel data and dataset compression.

This repository contains reference implementations for the webKNOSSOS wrapper format. Code is available for

• C/C++
• Scala
• MATLAB
• Rust
• Python

The Python implementation is a binding around the C library and demonstrates how wk-wrap files can be read and written from within other programming languages.

Each file contains the data for a cube with side-length (CLEN) of FILE_CLEN (MUST be a power of two; e.g., 1024) voxels. Within each file, the data is split into smaller, non-overlapping cubes (called "blocks") with a side-length of BLOCK_CLEN (MUST be a power of two; e.g., 32) voxels.

To enable fast access to subvolumes of the voxel cube, blocks are stored in Morton order. That is,

  block index           0         1         2         3         4         5
  block coordinates (0, 0, 0) (1, 0, 0) (0, 1, 0) (1, 1, 0) (0, 0, 1) (1, 0, 1)
         6         7         8         9        10        11        12     ...
     (0, 1, 1) (1, 1, 1) (2, 0, 0) (3, 0, 0) (2, 1, 0) (3, 1, 0) (2, 0, 1) ...

For further information, see the Wikipedia entry on the Z-order curve.

Each wk-wrap file begins with a file header. Depending on the content of this header, additional meta data MAY follow. The content of the file header and the optional meta data MUST be sufficient to determine the offset and size (in bytes) of each encoded block.

Each wk-wrap file MUST begin with the following header:

• version contains the wk-wrap format version as unsigned byte. At the time of writing, the only valid version number is 0x01.

• perDimLog2 contains two 4-bit values (nibbles). The lower nibble (perDimLog2 & 0x0F) contains voxelsPerBlockDimLog2, i.e., the log2 of the number of voxels per block dimension. The higher nibble ((perDimLog2 & 0xF0) >> 4) contains blocksPerFileDimLog2, i.e., the log2 of the number of blocks per file dimension. Files and blocks are three-dimensional.

• blockType determines how the individual blocks were encoded. Valid values are: 0x01 for RAW encoding, 0x02 for LZ4 compressed, and 0x03 for the high- compression version of LZ4.

• voxelType encodes the data type of the voxel values. Valid values are

  value of voxelType	0x01	0x02	0x03	0x04	0x05	0x06
  data type	uint8	uint16	uint32	uint64	float	double

• voxelSize is an uint8 of the number of bytes per voxel. If the wk-wrap file contains a single value per voxel, then voxelSize is equal to the byte size of the data type. If the wk-wrap file, however, contains multiple channels (e.g., three 8-bit values for RGB), voxelSize is a multiple of the data type size (specified as byte count). In the example of three 8-bit values, voxelSize would be 3.

• dataOffset contains the absolute address of the first byte of the first block (relative to the beginning of the file) as unsigned 64-bit integer.

Except when noted otherwise, multi-byte voxel values are stored in little-endian order. That is, bytes are stored in order of increasing significance.

Within raw blocks, the voxel values are stored in Fortran order. That is,

  voxel index              X + Y * BLOCK_CLEN + Z * BLOCK_CLEN * BLOCK_CLEN
  voxel coordinates                         (X, Y, Z)

In wk-wrap version 0x01, the block with index 0 begins immediately after the fixed header. The bytes of subsequent blocks are immediately following each other (i.e., no padding).

If the file header indicates that the blocks were compressed using LZ4 (by having a blockType value of either 0x02 or 0x03), the file header is immediately followed by the jump table.

The jump table is an array of N unsigned 64-bit integers, where N is the number of blocks in the file. The n-th entry of the jump table contains the absolute address (relative to the beginning of the file) of the first byte after the data of block n.

Note that

• the data of block n begins at address jumpTable[n - 1]
• the data of block n is jumpTable[n] - jumpTable[n - 1] bytes long

The value of jumpTable[-1] is defined as dataOffset. For this reason, it is convenient to build an extended jump table with the N + 1 unsigned 64-bit integers starting at the position of the dataOffset field.

Decompression is identical for the standard and the high-compression variants of LZ4. For a wk-wrap reader, the difference between blockType 0x02 and 0x03 is only semantic.

Decompression must produce valid raw blocks.

• Max Planck Institute for Brain Research
  • Alessandro Motta
  • Manuel Berning
• scalable minds
  • Johannes Frohnhofen
  • Tom Bocklisch
  • Norman Rzepka

MIT