Building head & numerics project

[Bulat Ziganshin]

Hello Creighton,

Thursday, May 4, 2006, 10:00:55 PM, you wrote:

> So I'm interested in working on the project on improving numerics
> performance in GHC proposed on the SoC page here
> http://hackage.haskell.org/trac/summer-of-code/ticket/13

which states "GHC's performance for double and float intensive code is
not as good as it could be. Find out why and improve it. Requires GHC
backend hacking. Must be very Haskell literate or have knowledge of
code generators."

i can't agree with this statement - source of problem is well known,
it's an inefficient STG-to-asm translation. and we know two ways to
solve this problem - either translate STG to idiomatic C code, as jhc
does and then rely on perfect GCC optimization, or make STG-to-asm
translation more efficient. last time it was discussed here in Feb'06

in general, it's too complex problem, otherwise Simons may already work
on it, because current ghc-generated code is, say, 3 times slower than
it could be. this sort of optimization will change performance for any
program, not only numeric-intensive ones. one possible rather simple
sub-project may be dealing with "leaf" functions with strict arguments
and results - i'm not sure but it's possible that implementing this
optimization will make significant shift exactly in numeric intensive
code. to decide this, it will be great to see this "numeric intensive
code" for beginning

-- 
Best regards,
 Bulat                            mailto:Bulat.Ziganshin at gmail.com