Goto Leads to Faster Code

pdoubleya writes "There's an article over at the NY Times (registration required) about Kazushige Goto, the author of the Goto Basic Linear Algebra Subroutines (BLAS, see the wiki); his BLAS implementation is used by 4 of the current 11 fastest computers in the world. Goto is known for painstaking effort in hand-optimizing his routines; in one case, "when computer scientists at the University at Buffalo added Goto BLAS to their Pentium-based supercomputer, the calculating power of the system jumped from 1.5 trillion to 2 trillion mathematical operations per second out of a theoretical limit of 3 trillion." To quote Jack Dongarra, from the University of Tennessee, "I tell them that if they want the fastest they should still turn to Mr. Goto."" Ever get the feeling someone wrote an article merely for the pun?

Comments

Re:30%+ Improvement

[Surt]

Considering the number of scientists who have been looking at this over a number of years, I think it really is a credit to Goto's work. Optimizing at this level is very challenging work on modern processors.

Re:I failed a coding test because of this guy

[CastrTroy]

Obviously this prof was wrong. If he wanted you to code it in as few lines as possible, then he should have expected everyone's code to be completely unreadable, goto's or not. If he wanted your code to be understandable, then he should have asked to make the code as clear as possible, by using as many lines as you may need.

Re:30%+ Improvement

[roystgnr]

Which is certainly good, but to me says more about the previous implementation than it does about Goto's work.

Yeah, that previous implementation must have totally sucked. I know all my linear algebra software is written around an assembly language core, hand tuned for each new version of a half dozen processors, and designed from the start to minimize TLB misses instead of just naively trying to fit a dataset into L1 or L2 cache. I don't know why those retards at the universities and national labs were ever using anything else!

(closing Slashdot, going back to working on my shamefully unoptimized C++ numerics code...)

Well Done Slashdotters

[Linker3000]

We have given birth to a new acronym: RPFH Read Past the F**king Headline.

Libgoto is fast but not open-source

[poszi]

Maybe I should not complain because the guy did a great job and his library is available free of charge but I hesitate to use this library because it is closed-source. I benchmarked it and found it fast and started to use it in my scientific codes. I once found a strange problem in a parallel code I was developing. The program crashed for one specific system I was calculating. It was something weird because it worked for many other systems I tested before. I spent a lot of time trying to find the bug in my program when finally I replaced libgoto with standard blas and the problem disapeared. I knew that the crash was when entering blas but I thought it is because I messed with the arrays that are used as parameters. If libgoto were open-source, I would be able to have a version with debugging info compiled and debug the program and the library. I would probably not fix the bug but I would likely figure out more quickly the problem is in the library and not in my code. After I had known the problem is in libgoto, I dowloaded a new version of libgoto and it worked so the bug has been fixed. There is no changelog on libgoto web page so I don't know what was the problem and how it affacted my previous caclulations.

Atlas is open-source and is a pretty good alternative. It is only a few percent slower than libgoto in most cases.