RawArray (RA) is a simple file format for storing n-dimensional arrays. RA stands for raw array and should be pronounced arr-ay, although it is not a coincidence that the mispronunciation rah (as in "raw" in some dialects) also makes sense.
RA was designed to be portable, fast, and storage efficient. For scientific applications in particular, it can allow the simple storage of large arrays without a separate header file to store the dimensions and type metadata.
The fundamental philosophy of RawArray is what you save is what you read.
In other words, the array after reading should look exactly like it did when you wrote it to disk.
I believe the world doesn't need another hierarchical data container. We already have one of those---it's called a filesystem. What is needed is a simple one-to-one mapping of data structures to disk files that preserves metadata and is fast and simple to read and write.
In addition to int, uint, and float of arbitrary sizes, RA also supports
(1) complex floats: which other common formats, such as HDF5, don't have.
(2) Booleans: both Boolean arrays with 8-bits per value and the compact BitArrays
(3) composite types: RA handles reading and writing these, but the encoding and decoding of those is left to the user, since only they can know the structure of their struct. Decoding can be as simple as a typecast, however, for types of fixed size.
As an aside, the RA format is technically recursive (or fractal?!). You could store an array of RA files in a RA file if you want by defining the file as a composite type.
Besides supplying support for the RA format to the Julia community, this package is intended to provide the definitive implementation of the RA format. New features will be implemented here first before being ported to the main RA repository. If something exists in this package that you would like in another language, feel free to implement it yourself and submit a pull request to the RA repository.
The file format is a simple concatenation of a header array and a data array. The header is made up of at least seven 64-bit unsigned integers. The array data is whatever you want it to be. Optionally text or binary metadata can be appended to the end of the file with no harmful effects, but this data is not saved or written back out by the library. It is up to you to keep track of it.
| offset (bytes) | object | type | meaning |
|---|---|---|---|
| HEADER | |||
| 0 | magic | UInt64 | magic number |
| 8 | flags | UInt64 | endianness, future options |
| 16 | eltype | UInt64 | element type code |
| 24 | elbyte | UInt64 | element size in bytes |
| 32 | size | UInt64 | data segment length in bytes |
| 40 | ndims | UInt64 | number of array dimensions |
| 48 | dims | Vector{UInt64} | array dimensions |
| 48 + 8 x ndims | data | Vector{UInt8} | ARRAY DATA |
| 48 + 8 x ndims + size | - | - | VOLATILE METADATA |
| code | type |
|---|---|
| 0 | user-defined |
| 1 | signed integer |
| 2 | unsigned integer |
| 3 | floating point (IEEE-754 standard) |
| 4 | complex float (pairs of IEEE floats) |
| 5 | Boolean |
The width of these types is defined separately in the elbyte field. For example,
- a 32-bit unsigned integer would be
eltype = 2,elbyte = 4; - a single-precision complex float (pairs of 32-bit floats) would be
eltype = 4,elbyte = 8; - a string would be
eltype = 2,elbyte = 1, andsizewould contain the length of the string.
The user-defined structure
struct foo {
char info[12];
uint32_t index;
double v[8];
}
contains a 12-byte string, a 4-byte int, and 8 8-byte floats, so the total size is 80 bytes. It would be coded as eltype = 0, elbyte = 80.
The data is written and read as the binary representation of the hardware you are on. Right now little endian is assumed, but big endian support can be added if there is interest.
The RA format is column major, so the first dimension will be the fastest varying one in memory. This decision was made because the majority of scientific languages are traditionally column major, and although C is technically row major it is actually agnostic in applications where multi-dimensional arrays are accessed through computed linear indices (e.g. CUDA). Of the supplied examples, all are column major except Python. In the case of Python, instead of reading the array into Python and reordering to non-optimal stride, we simply transpose the dimensions before writing and after reading. This means the array looks transposed in Python, but the same dimensions have the same strides in all languages. In other words, the last dimension of the array in Python will be the first one in Julia and Matlab.
To get a better handle on the format of an RA file, let's look inside one. If you are on a Unix system or have Cygwin installed on Windows, you can examine the contents of an RA file using command line tools. For this section, we will use the test.ra file provided in the examples/ subdirectory.
First, let's pretend you don't know the dimensionality of the array. Then
> od -t uL -N 48 test.ra
0000000 8746397786917265778 0
0000020 4 8
0000040 96 2
0000060
shows the dimension (2) as the second number on the third line. The command is extracting the first 48 bytes and formatting them as UInt64s. The ridiculous number listed first is the magic number indicating that this is an RA file. A slightly different command illuminates that:
> od -a -N 16 test.ra
0000000 r a w a r r a y nul nul nul nul nul nul nul nul
0000020
Armed with the knowledge that the array is 2D, we know that the header is 48 + 2*8 = 64 bytes long. The command to skip the header and view only the data would be:
> od -j 64 -f test.ra
0000100 0.000000e+00 -inf 1.000000e+00 -1.000000e+00
0000120 2.000000e+00 -5.000000e-01 3.000000e+00 -3.333333e-01
0000140 4.000000e+00 -2.500000e-01 5.000000e+00 -2.000000e-01
0000160 6.000000e+00 -1.666667e-01 7.000000e+00 -1.428571e-01
0000200 8.000000e+00 -1.250000e-01 9.000000e+00 -1.111111e-01
0000220 1.000000e+01 -1.000000e-01 1.100000e+01 -9.090909e-02
0000240
Here we are using -j to skip the first 64 bytes and -f to format the byte data as single-precision floats. Note od doesn't understand complex numbers, but the complex data is stored as real and imaginary float pairs that are contiguous on disk. This means that each line of the output is showing two complex numbers with columns 1 and 3 the real parts and columns 2 and 4 the imaginary parts. Notice that it correctly renders the negative infinity.
You can clone the git repository from inside Julia with
Pkg.add("RawArray")
To use RawArray, simply add the following line to your Julia script:
using RawArray
Now you can call raread and rawrite for Julia objects of type Array{T,N}.
See the test script test/runtests.jl for some examples of use.
A simple example of reading and writing a float array looks like this:
julia> using RawArray
julia> x = rand(8,8);
julia> rawrite(x, "test.ra")
julia> y = raread("test.ra")
julia> x == y
true
A test file called test/runtests.jl has been included, as well as a demo RA file in examples/test.ra. You can test the code on your machine at the command line by running julia runtests.jl. If the tests pass, you'll get a message saying so.
Notice the Julia version also contains a raquery() function that produces a YAML dump of the file header for easier parsing in other codes.
If you are storing integers, RawArray has compression through variable length integer encoding built in, so you can store your array with lossless compression:
julia> using RawArray
julia> n = rand(1:100, 8, 8);
julia> rawrite(n, "ints.ra", compress=true)
julia> m = raread("ints.ra")
julia> m == n
true
You can use this compression on floats if you have limited precision data, because you can then convert to integer for storage without losing any true precision. For example, assume you have data on the [0,1] real interval with three decimal digits of true precision. Converting to integer for compressed storage would look something like this:
julia> x = rand(3,3)
3×3 Array{Float64,2}:
0.269812 0.116996 0.415197
0.950308 0.583864 0.844317
0.0306206 0.558326 0.610574
julia> m = round(Int, x * 1000)
3×3 Array{Int64,2}:
270 117 415
950 584 844
31 558 611
julia> rawrite(m, "mydata.ra", compress=true)
julia> n = raread("mydata.ra")
3×3 Array{Int64,2}:
270 117 415
950 584 844
31 558 611
julia> y = n * 0.001
3×3 Array{Float64,2}:
0.27 0.117 0.415
0.95 0.584 0.844
0.031 0.558 0.611
To see what the potential size savings are, let's write a large, image-sized float array both as the original float and as a compressed Int array with three digits of precision:
julia> x = rand(512,512);
julia> rawrite(x,"x_float.ra")
julia> m = round(Int, x * 1000);
julia> rawrite(m, "x_int.ra", compress=true)
julia> sf = stat("x_float.ra").size
2097216
julia> si = stat("x_int.ra").size
507801
julia> sf / si
4.129995805443471
So an over 4x compression was achieved by this method that is very simple, fast, and internal to the RawArray package.
External compression libraries, like 7zip, can then be used to further compress the compressed int RA file:
shell> 7z a x_int.7z x_int.ra
7-Zip [64] 16.02 : Copyright (c) 1999-2016 Igor Pavlov : 2016-05-21
p7zip Version 16.02 (locale=utf8,Utf16=on,HugeFiles=on,64 bits,8 CPUs x64)
Scanning the drive:
1 file, 507801 bytes (496 KiB)
Creating archive: x_int.7z
Items to compress: 1
Files read from disk: 1
Archive size: 337078 bytes (330 KiB)
Everything is Ok
julia> siz = stat("x_int.7z").size
337078
julia> sf / siz
6.221752828722136
So you can see that the external compression algorithms are complementary to the variable length integer compression. The final compressed size was 337 kB, which for 512 x 512 floats works out to 10.3 bits per float. Or even smaller than the IEEE-754 half-precision float format that uses 16 bits per float.
For help, file an issue on the bug tracker or email one of the authors. Third party help is welcome and can be contributed through pull requests.
David S. Smith david.smith@gmail.com
This code comes with no warranty. Use at your own risk. If it breaks, let us know, and we'll try to help you fix it.