Skip to content

spiraldb/streamvbyte-zig

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Zig StreamVByte

A Zig port of Stream VByte.

This project started as an experiment to explore the feasibility and DevX of implementing compression codecs in Zig.

In particular,

  • Leveraging comptime to avoid writing (or generating) repetitive code.
  • Using the Zig @Vector API for portable SIMD.

Comptime Lookup Tables

Stream VByte leverages lookup tables (LUTs) for pre-computing shuffle masks as well as the lengths of compressed quads.

Zig's comptime feature allows us to generate static LUTs at compile-time by defining them in regular Zig. This makes it possible for the reader to understand the origin of what are otherwise typically opaque "magic" values.

Zig SIMD

Zig leans on LLVM for its @Vector SIMD API. While Zig offers a @shuffle builtin to operate on vectors, LLVM (and therefore Zig) require the shuffle mask to be comptime known.

Since the shuffle mask is a lookup based on runtime control bits, it isn't possible to use the @shuffle builtin.

It is worth noting at this point that we use Zig 0.11.0.

We were able to generate functions for each of the shuffle masks, and then store these function pointers in a LUT. But since these functions cannot be inlined, the overhead was far too much.

We also tried to @cImport the relevant instrinsics headers. But found that Zig is currently unable to inline callconv(.C) extern functions.

This left us to implement a shuffle function ourselves using pshufb (and tbl.16b) with Zig's inline ASM. Whilst this isn't ideal, it doesn't seem unreasonable for these instructions to be supported by the builtin @shuffle operator when the operand is only runtime known.

Benchmarks

We spent most of our time playing around with shuffle LUTs, so the remainder of the code isn't particularly heavily optimized. Nor have we run these benchmarks on a variety of architectures. Take everything with a handful of salt. Especially given how much variance there is in the "benchmarks".

$ zig build test -Doptimize=ReleaseFast
Zig StreamVByte
	Encode 10000000 ints between 0 and 10000 in mean 3467410ns
	=> 2883 million ints per second

	Decode 10000000 ints between 0 and 10000 in mean 1681129ns
	=> 5948 million ints per second

Original C StreamVByte
	Encode 10000000 ints between 0 and 10000 in mean 2065012ns
	=> 4842 million ints per second

	Decode 10000000 ints between 0 and 10000 in mean 1669633ns
	=> 5989 million ints per second

The source code is released under the MIT license.

Languages