Re: fast bit rotation
Tommy wrote:
> Juergen Luethje wrote:
>>
>> Rolf wrote:
>>
>>> I programmed 'true euphorian' rotation functions rotl(objekt x) and
>>> rotr(object x), see include file ROTATION.E and test file ROTATION.EXW
>>> of attached zip file. To run it under DOS, rename *.EXW to *.EX.
>>>
>>> On my AMD K6-2, 400MHz I found (under NT 4.0) with ROTATION.EXW
>>> 1.91 micro seconds for rotr() and
>>> 1.63 micro seconds for rotl() per call.
>>
>> You can even gain speed by omitting 'pusha' and 'popa'. 
>>
>> I like the idea of wrapping the ASM functions by 'true euphorian'
>> functions, especially in order to be able to apply them to any Euphoria
>> object.
>>
>> However, the main disadvantage of Tommy's routines IMHO is the point,
>> that they use a fixed count of '1' for any rotation. A versatile library
>> routine IMHO should allow the user to specify the count as parameter, so
>> that it would look somehow like:
>> function rotate_left(value, count)
>
> Of course it only rotates a fixed count of '1': I didn't mean to create
> a versatile library routine, just the fastest possible solution to
> jiri's problem, which was integer rotation (left or right) of fixed
> count 1 (I think :D).
Sorry, I didn't want to offend you or something. I thought Jiri wanted
some versatile library routines.
> And I believe that if you want the highest
> performance possible, my routine (no real Euphoria-function) is perhaps
> a bit faster: no stack popping, just 1 poke, 2 moves, a rotation, a
> return and 1 peek.
> Haven't compared the performance of the 2 routines, but I don't need to.
> I know they both perform very well and I also know that maintainability,
> ease of use and readable code are much more important than that tiny bit
> of extra performance.
> So Rolf and Juergen, your routines are in general much better than what
> I pulled together in 5 minutes.
Please don't forget: I was somehow inspired by your code, and Rolf's
routines use yours as a basis.
> BTW: wouldn't it be nice if someone made a library of such very general
> high-performance routines?
I already started modifying your ASM routines, so that they take 'count'
as a second parameter. Then I'll wrap them by pure euphoria functions,
in order to be able to apply them to any Euphoria object -- very similar
to what Rolf did.
I'll do the same with my ASM routines, compare the speed, make some
tweaks here and there ...
I also plan do write some SHIFT functions.
> Or perhaps Robert could add some of those
> routines to machine.e or one of the general Euphoria-libraries? If you
> see how many people could use shifts, rotates and other routines like
> that.
That would be nice.
<snip>
Regards,
Juergen
--
/"\ ASCII ribbon campain | This message has been ROT-13 encrypted
\ / against HTML in | twice for higher security.
X e-mail and news, |
/ \ and unneeded MIME | http://home.arcor.de/luethje/prog/
|
Not Categorized, Please Help
|
|
