Hacker News Reader

Comment by nyanpasu64 | Hacker News Reader

https://www.hackerfactor.com/blog/index.php?/archives/895-Co...

> on the first retina iPads, decoding PNGs was a huge portion of the total launch times for some apps, while one of the two cores sat completely idle.

Could they distribute decoding PNG files to a thread pool, instead of making multithreaded PNG files? Or would this fail for single large PNG files?

View on HN

Well, one thing that does work for decoding regular old PNGs is separating the entropy decoding (in PNG, DEFLATE) from the prediction (what PNG calls filtering). You can create a pipeline where one thread is working on the entropy coding and handing off completed rows to a thread that performs the prediction decode. This doesn't get a 2x speedup because usually DEFLATE is slower than the prediction step, so it ends up bottlenecked on zlib, but it helps a good bit. (I implemented this in Rust at one point with the intention to ship in Firefox, but it never did.)

Of course, this only scales to 2 CPUs. Beyond that you will need to do some sort of splitting of the input to achieve wins (since no matter how much you optimize the filtering, you're still bottlenecked on DEFLATE, which is inherently serial).

How I read nyanpasu64; I think that the point was that a web page will often have multiple png files -- often more than you have cores, so the gains of decoding a single specially-formatted png might disappear in practice.

iDOT seems like overengineering to me: Why not break up the PNG and tile them in an SVG? This should be just as fast, about the same size, and you'd never have run the risk of inventing your own format and implementing it badly[1]. Users would probably be happier with all their SVGs being faster than having specially-crafted PNGs accelerated (if given the choice)

[1]: https://cve.mitre.org/cgi-bin/cvename.cgi?name=CVE-2016-1811

I’m a bit confused, are you suggesting they convert the PNGs into an SVG? How is that possible?

I'm not entirely sure how to begin to answer this question. Forgive me if I start too early.

An SVG is a list of compositing instructions for some (graphical) artefact. These instructions are given so that their dependencies (the things that must be composed before other things) are stated directly.

If you have familiar with Adobe's Photoshop, you might imagine a SVG as a set of nested layers: An SVG "renderer" will simply compose these layers together in order to have some pixels to display.

Now, to give a clear example of what I'm referring to, I am going to show you a simplified SVG that composes four png files together in tiles:

    <svg width="100" height="100">
      <image x="0" y="0" width="50" height="50" href="data:image/png,xxx" />
      <image x="50" y="0" width="50" height="50" href="data:image/png,xxx" />
      <image x="0" y="50" width="50" height="50" href="data:image/png,xxx" />
      <image x="50" y="50" width="50" height="50" href="data:image/png,xxx" />
    </svg>

That data:image/png,xxx stanza is the SVG-encoding of a PNG (a slight simplification: the format actually belongs to a number of different standards, that the SVG specification leverages).

That is to say, I'm not exactly suggesting converting a PNG to an SVG: I am also suggesting breaking apart the (large!) PNG into several component PNG files (tiles) so that they can be decoded independently. Note carefully my example, how none of the tiles overlap. A decoder can (trivially) determine these instructions are independent, and so process them independently.

Being able to decode parts of the resulting image independently is what the iDOT metadata makes possible: It is essentially a different encoding of the x/y/width/height information in the above, the difference is that SVG already existed.

Amen!

I assume iOS may have a problem of the application doing:

    for each path {
       decode(path)
    }

and the OS can't parallelize that, because it learns about a new file only after decoding the previous one.

Decompressing zlib/DEFLATE streams, in the general case, is an inherently serial task. It cannot be meaningfully parallelised.

The "trick" used by Apple, and a few other encoders, is to flush the zlib state every so often (ZLIB_FULL_FLUSH), which means that subsequent data is both byte alligned, and does not make any backreferences to before the sync.

If you know where these sync points are (Apple encodes this information in their non-standard "iDOT" chunk) then you can start decompressing from that point, in an isolated thread.

There are ongoing discussions as to how this metadata could be standardised: https://github.com/w3c/PNG-spec/issues/54

Technically speaking you can also speculatively decode Huffman coded bitstreams because the use of canonical Huffman tree means that the end-of-block symbol is almost likely the longest code and has no 0 bits, so you can start decoding at a long string of 1 bits. Of course this doesn't solve an issue of shared LZSS window across multiple blocks.

This is hilarious to me, but makes complete sense. Do you know if this has ever been done anywhere?

There is indeed a hilarious backstory. Back in time one of my friends was looking for a task for the Distributed Computing course where you write a program for the Cell microprocessor, and I half-jokingly suggested this, not realizing that it won't work at all because of the shared window. The friend eventually came to realize that issue and couldn't finish the project in time, but still got a good grade solely because of the novelty of the task (!). (According to that friend virtually everyone else did the parallel JPEG stuffs.)

This doesn't surprise me. The Cell architecture wrought more insane hacks than good designs.

I think that's what restart markers are for in JPEG from the original spec. IIRC in JPEG you can have markers that are marked as "crucial for proper decode" and programs should reject the image if they don't understand them. Maybe the latter was from PNG I forget.

It would fail for single large splash-screen PNG files, and that's where they were looking to shave milliseconds.

You'd also have to be kicking off PNG decodes in a way that is compatible with offloading them to a threadpool, and lots of software likely just loads images synchronously and moves on with its day.

Reply on news.ycombinator.com