It's a somewhat known tradeoff, you can streamline and make the format friendlier to do memcpy which this library targets, the more memcpys you do, the faster it is overall to decode. On highly compressible data lz4, snappy become faster. Snappy on level 2 has faster decompression speed
But you have to pay the price that you need a slower encoding, because finding matches, putting restrictions on match lengths, putting things in different streams have costs you need to pay upfront.
Anyway good work, there is probably a need for that.
// Currently own Google's snappy and do compression at Google
P.S. if you want better snappy's results, compile with clang.
I haven't implemented any ARM-specific dispatch yet (there are currently no NEON paths for vector ops either, and we instead trust the compiler to autovectorize), and will do so for upcoming versions.
So, I couldn't see it in the readme, apologies if I missed it but why?
It's a very significant speedup in decompression speed (albeit with a compression speed slowdown as a trade-off), but what's the insight that makes it faster? What was the idea or approach behind it?
So it's essentially "LZ4 unshackled". I've made several modifications to the LZ4 format, most of which are in service of eliminating branches/making them more predictable, and making decompression very friendly to out-of-order cores by hiding false data dependencies behind a rarely taken branch (similar to this: https://news.ycombinator.com/item?id=48889148).
Some concrete changes in the format are:
- match length per block is capped to 32
- distance to a match must be >= a fixed constant
- unlike lz4, tokens and literals have separate streams
- format of the token byte has been changed
Now, this format allows our decompressor's hot loop to be very simple (in terms of the number of branches it has). This simplicity in turn allows our compressor to create a compressed stream that is friendly to the (small number of) branches in the decompressor.
The experimental compression modes (see readme) attempt to exploit this even further (but are even slower at compression). I define a "cost", which is a linear function of the branches induced by a compressed stream (this function serves as a proxy for decompression time), and then do a DP to minimise this cost.
Almost none. Once again, simplicity comes to our rescue here. The decompressor is simple and a naive safe version I implemented but haven't merged into main yet (see: https://encode.su/threads/4514-misa77-ridiculously-fast-deco...) is only ~5% slower than the current unsafe version (and can very likely be made faster).
Almost double the decompression speed, *and* a higher compression ratio than LZ4? That's very promising.
I don't see much in your README that talks about how a developer would integrate misa into their code. I might suggest some basic code samples to help a developer integrate misa decode into their project.
Interesting, but if you are not robust to corrupted/malicious data, it is really in a different class of algorithm and it is hard to compare speeds directly.
From memory, 2505 MB/sec also sounds on the low side for LZ4 on a modern CPU?
In short, my decompressor is very simple, and a naive safe version of the decompressor is only about 5% slower than the current unsafe one (and I will add this safe version in v0.3.0).
As for the raw throughput numbers being low here, it's because Intel Turbo (frequency boost) was disabled for stability, and the CPU was running at a fixed frequency of 2.1 GHz (I've confirmed that the relative performance scales similarly even with Turbo enabled).
> It offers particularly high decompression throughput on highly compressible files.
You know, you would expect that this is commonly observed. You'd expect, say, simple RLE (run length encoding) to be like this. It's more expensive to produce output requiring the processing of a large number of RLE opcodes that produce short sequences than one opcode that produces one long sequence.
this is super interesting! im excited to give this a look this afternoon, since I specifically have wanted faster throughout for decompressing maps in a game engine.
misa77 primarily targets textual data (ie. byte-aligned data formats where each byte corresponds to a symbol), so I hadn't tested it on game assets much until now.
After seeing your comment, I pulled some random assets from Pathfinder WoTR (in fact, Unity compresses them with lz4hc) and DOS2. The gains are much more modest here due to asset data being mostly floats, but level 0 performs decently nevertheless.
You can look at the LZ4 decompression implementation in ClickHouse: https://presentations.clickhouse.com/2018-highload-siberia/ (the presentation is quite old, but the implementation was updated recently). It uses the same LZ4 format, just decompresses faster.
Just tested a few minutes ago on the ClickBench dataset. Overall quite good, but, depending on particular columns, most of the time slower - e.g., ~2.0 vs ~2.8 GB/sec on Graviton 4 machine in AWS.
But you have to pay the price that you need a slower encoding, because finding matches, putting restrictions on match lengths, putting things in different streams have costs you need to pay upfront.
Anyway good work, there is probably a need for that.
// Currently own Google's snappy and do compression at Google
P.S. if you want better snappy's results, compile with clang.
P.S.S you can optimize aarch64 speed by movemasks from shrn instruction. https://developer.arm.com/community/arm-community-blogs/b/se...
I haven't implemented any ARM-specific dispatch yet (there are currently no NEON paths for vector ops either, and we instead trust the compiler to autovectorize), and will do so for upcoming versions.
- misa77's format may change unexpectedly as it's still v0.x.y.
- The decoder assumes that the input is a valid misa77 stream. Invalid input is UB and I offer no guarantees for whatever misa77 does in this case.
- It's been through some local fuzzing but is not hardened, so treat it as experimental.
It's a very significant speedup in decompression speed (albeit with a compression speed slowdown as a trade-off), but what's the insight that makes it faster? What was the idea or approach behind it?
Some concrete changes in the format are:
Now, this format allows our decompressor's hot loop to be very simple (in terms of the number of branches it has). This simplicity in turn allows our compressor to create a compressed stream that is friendly to the (small number of) branches in the decompressor.The experimental compression modes (see readme) attempt to exploit this even further (but are even slower at compression). I define a "cost", which is a linear function of the branches induced by a compressed stream (this function serves as a proxy for decompression time), and then do a DP to minimise this cost.
And i do not mean this in a flippant way, as how to harden with speed in mind might alter how to design the format and the codec.
I don't see much in your README that talks about how a developer would integrate misa into their code. I might suggest some basic code samples to help a developer integrate misa decode into their project.
Congrats, looks like a cool project.
From memory, 2505 MB/sec also sounds on the low side for LZ4 on a modern CPU?
In short, my decompressor is very simple, and a naive safe version of the decompressor is only about 5% slower than the current unsafe one (and I will add this safe version in v0.3.0).
As for the raw throughput numbers being low here, it's because Intel Turbo (frequency boost) was disabled for stability, and the CPU was running at a fixed frequency of 2.1 GHz (I've confirmed that the relative performance scales similarly even with Turbo enabled).
You know, you would expect that this is commonly observed. You'd expect, say, simple RLE (run length encoding) to be like this. It's more expensive to produce output requiring the processing of a large number of RLE opcodes that produce short sequences than one opcode that produces one long sequence.
misa77 primarily targets textual data (ie. byte-aligned data formats where each byte corresponds to a symbol), so I hadn't tested it on game assets much until now.
After seeing your comment, I pulled some random assets from Pathfinder WoTR (in fact, Unity compresses them with lz4hc) and DOS2. The gains are much more modest here due to asset data being mostly floats, but level 0 performs decently nevertheless.
Results on a map asset (WoTR):
Results on equipment asset (WoTR): Results on texture asset (DOS2): Note: the benchmarking setup is identical to the intel x86-64 one described in the readme.For reference, to test ARM64, I tested v0.1.0 on an M3 mac with this fork of lzbench: https://github.com/welcome-to-the-sunny-side/lzbench/tree/ad...
Here, lz4's decompression speed was far slower than misa77 and zxc. Results are here: https://github.com/welcome-to-the-sunny-side/misa77/blob/mai...