Slashdot Mirror
Computer Program Fixes Old Code Faster Than Expert Engineers
An anonymous reader writes: Less than two weeks after one group of MIT researchers unveiled a system capable of repairing software bugs automatically, a different group has demonstrated another system called Helium, which "revamps and fine-tunes code without ever needing the original source, in a matter of hours or even minutes." The process works like this: "The team started with a simple building block of programming that's nevertheless extremely difficult to analyze: binary code that has been stripped of debug symbols, which represents the only piece of code that is available for proprietary software such as Photoshop. ... With Helium, the researchers are able to lift these kernels from a stripped binary and restructure them as high-level representations that are readable in Halide, a CSAIL-designed programming language geared towards image-processing. ... From there, the Helium system then replaces the original bit-rotted components with the re-optimized ones. The net result: Helium can improve the performance of certain Photoshop filters by 75 percent, and the performance of less optimized programs such as Microsoft Windows' IrfanView by 400 to 500 percent." Their full academic paper (PDF) is available online.
Unemployment is a sort of nirvana.
I just want to know why when they improve the speed of something by 75-500% it's considered research but when I do it it's considered a waste of time? Companies just don't appreciate efficiency.
This kind of application has been ported to about half a dozen different operating systems. i use it on a regular basis but for those not familiar I find it works best with legacy PHP. the "rm" command (or as its known in windows "del") uses an advanced algorythm to determine the target codes age, total number of lines, and any optimizations that are required in order to make it fully web 4.0 SOAP, ROPE, and DOPE compliant. It then quickly converts the code to the updated version by changing the physical structure of the space it occupies on the disk. My coworkers always seem amazed when they see me break out my legacy optimizer. "What have you done?!" they exclaim, baffled at my efficiency. "holy crap do you know how important that was?!?" they remark in awe
Good people go to bed earlier.
The paper is not about automatic code optimisation in general. This program only fixes image-related algorithms (which most likely involve simple & repetitive modifications).
In fact, there are quite a few references to Adobe in the associated links and the paper is signed by two Adobe guys. Thus, I guess that it is mostly focused on Adobe products, like Photoshop; and might not even be able to deal with any image-processing algorithm.
PS: is it because of the summer-period lack of news or Slashdot is planning to make the move to pure sensationalism? Because a big proportion of the titles lately are quite unclear (to say it softly).
Custom Solvers 2.0 = Alvaro Carballo Garcia = varocarbas.
$ Helium -i Helium -o Helium2
Program ran in 656.76s
$ Helium2
Welcome to SkyNET
When you read the actual paper, it seems to analyse the code, and take advantage of vectorization in modern hardware to do the equivalent act in a more parallel way increasing performance. So it strips off local optimization, analyses the action, then recreates it in a form better suited to the processor its compiled for, resulting in a speedup.
Sort of reminds me of Transmeta, the processor Linus was involved with, that ran microcode, analysed the execution of the x86 code, and did little mini compiles of microcode to simulate that. aka 'Code Morphing',
https://en.wikipedia.org/wiki/Transmeta
It did well on Benchmarks (and filters) because of the repetitive nature of them, it could recreate those loops more efficiently.
Interesting, pity its all full of boastful fluff, and vague bitrot claims, it sounds like a proper paper otherwise. Of course for most software if it has value, it is a competition and its worth re-optimizing for each new platform otherwise you lose market to competitors. So actual programmers would rework those parts of the code rather than a general tool hack.
So this would only be a stopgap for dead end software you're locked into that isn't being maintained.
Actually, copyright is violated when you copy the work without permission and the purpose for copying the work does not otherwise make it exempt from infringement. If you don't distribute at all, and you do not profit from it in any way, then nobody else is necessarily even going to know you copied it, let alone want to prosecute you for it, but that does not necessarily mean it is necessarily legal (some jurisdictions even have private copying as exempt from infringement anyways, or for example, in Canada, have private-use copying exempt only so long as a mechanism for compensating at least some percentage private-use copies exists, such as levies on blank media where such copied content is considered by said controlling agencies to be likely to be copied onto).
Distribution of unauthorized copies is usually taken as gold-bar standard proof of infringement, since it establishes, in an objectively verifiable way, a non-private-use purpose for making those copies which can be presented by the copyright holder as a basis for the claim against the infringer. The action of distributing unauthorized copies is generally considered illegal as well, a crime related to infringement and no less than at least evidence of conspiracy to commit copyright infringement, but is not technically copyright infringement, not is it absolutely required to commit infringement... at most it may only be required to be held accountable for infringement.
File under 'M' for 'Manic ranting'
If you look at how their system works, you can see that it is limited to certain types of algorithm, and even then it gives absolutely zero correctness guarantees. What they want to do is abstract and complex. What they're actually doing can be described in a 20-page paper. Your job is safe for right now. Strong AI still doesn't exist.
Anyway, even if they automate some parts of your job, the part of your job that isn't automated will expand to fill that time. I'm not sure if there's a real law for that but it's a heuristic observation shared by many. Partial automation doesn't lead to declining workforces, it leads to expanded capability and larger workloads than could be imagined before. Even if they perfected this program it still couldn't write a new kernel, let alone a new application.