>In article <3c0f558d.1210938@news.povray.org> , ken### [at] uniplanit (Angelo 
>'kENpEX' Pesce) wrote:
>
>> Mhm, why asm-optimization where not acceptable, and something that
>> involves translating stuff into native-code asm is acceptable now?
>
>It isn't that hard to understand, really...
>
>Lets say you have a function foo.  You want to write it in assembler.  So
>you end up with an assembler version of function foo that has to be in the
>source code, right?  For each and every platform you need this, and you need
>to clutter all the code with it.

so let's say that I have mhm affine.c code in the source tree... doing
this i'll "clutter" the source by adding an affine_486.s, affine_586,
affine_mac etc... ok, right...

>Contrary to this, a JIT compiler works on a provided intermediate data.  It
>is very clean to interface the intermediate format interpreter/VM with a
>platform specific optional compiler.  You do not need to clutter the source
>code of every platform.  And you force no platform dependent code into the
>platform independent code.

so let's say that I have an interpreter.c code in the source tree. jit
means that I need also, for some plattforms, a jit_x86, a jit_mac, a
jit_amiga etc... mhm no I feel stupid, I can't catch the difference

>Get the difference?
>Of course, somebody previously told you the answer to this already:  If you
>can improve an algorithm, from say O(n^2) to O(log n) you gain far more than
>one or two percent you would get from writing the O(n^2)  algorithm in
>assembler.  However, with assembler you introduce unnecessary and unreadable
>clutter for no good reason this way.

Of course. We agree. It's fine. But after U implemented something that
looks *final* why you can't do the second step?

>Get the difference now?
>If not, try to think about Donald E. Knuth's famous quote again.

Premature. Now I want to know, are U going to change ray-sphere
intersection? Do U see new techniques to do 3d vector cross product?
dot product? Matrix-vector product? I agree mabye it's premature to
optimize new algorithms or something that is still in research stage.
But how about the kernel math routines? Mabye I'm blind but I don't
think that optimizing such routines is asm (routines that are used
everywhere in the code and that could lead to a general speedup) is
premature not dumb... And if U think that using asm in basic 3d math
routines will not give U a reasonable speed up because the compiler
know how to do everything, well I can give U some of my code and you
can re-do some checks that I have already done...

>    Thorsten
>
>____________________________________________________
>Thorsten Froehlich, Duisburg, Germany
>e-mail: tho### [at] trfde
>
>Visit POV-Ray on the web: http://mac.povray.org

Post a reply to this message