[PD] Efficient FIFO in Pd: How?

[Matt Barber]

Your [fifo-list] is very much like [list-fifo] from list-abs, which suffers
from the poor [list] performance. The message-box one also suffers from
having to use lists to dequeue.

I've thought of a way to do it using [list fromsymbol], [list tosymbol],
and an elaborate array storage scheme which would store the length of each
list at the beginning of the list like you do in the others, and then the
length of each symbol before each symbol (or -1 to store a float).
Enqueueing would cost one [list-drip] plus the encoding and storing per
list, but since the array is random-access, it should be O(n) over all with
the size of the fifo.

Retrieving would do it in reverse, which would cost one [list prepend]X[t
a] loop plus reading and decoding per list, which should also be O(n) with
the size of the fifo.

You'd maintain a write and a read index into the array; [array size] the
array to 2^24. What to do when you reach the end of the array is the tricky
part. If you want to be able to do a ringbuffer kind of thing, you could
just maintain another array to prewrite the list encoding into. By doing
this, you'll know the entire length of what needs to be written, and if
that exceeds what you have in your table, you could write an instruction
(-100, say), to start at the beginning again, and then reset your write
index to 0. You'd need to test to make sure that an incoming write doesn't
allow your "write head" to lap your "read head".

If you needed REALLY arbitrary size, if you got to the end of one array,
you'd dynamically allocate a new one, and write an instruction to move
reading to the next array when your read index reaches it.

I have no idea if this would work (it should), or what the constant time
complexity factors are here; my guess is that it's slower than the others
for short fifos. [array set] and [array get] are faster than you'd think,
though, allowing you to write chunks at a time, unlike [tabwrite].

If you make it I'll tweak it (for style) to include in my array-abs
collection; otherwise I'll add it to my list and make it later.

Matt

On Sat, Nov 21, 2015 at 7:52 PM, Roman Haefeli <reduzent at gmail.com> wrote:

> On Sun, 2015-11-22 at 00:47 +0100, Christof Ressi wrote:
> > Just in case you don't know: There's the [fifop] object in zexy which
> works with lists and also lets you set the priority. I'm sure it's also
> faster than anything one can do with Pd objects.
>
> Thanks, somehow I forgot that zexy has that. It is indeed fast (as in
> linear fast). I'm going to use it.
>
> Just out of curiosity, I wonder if there is a O(n) vanilla
> implementation, too.
>
> Roman
>
>
> > > Gesendet: Sonntag, 22. November 2015 um 00:24 Uhr
> > > Von: "Roman Haefeli" <reduzent at gmail.com>
> > > An: Pd-List <pd-list at lists.iem.at>
> > > Betreff: [PD] Efficient FIFO in Pd: How?
> > >
> > > Hi all
> > >
> > > I'm looking for a way to implement a FIFO buffer in Pd. I'd like it to
> > > be able enqueue messages with an arbitrary number of elements in order
> > > to retrieve them in the same order later. I came up with three
> different
> > > methods, but they are all either slow at enqueuing or at dequeuing or
> > > even both. It seems for a quick implementation you need something like
> a
> > > ring buffer, but I can't think of any method that would work for
> > > symbolic elements, too (as opposed to numbers).
> > >
> > > Attached patch comparison.pd compares the methods. Isn't there really a
> > > O(n) implementation as opposed to O(n^2)?
> > >
> > > Roman
> > > _______________________________________________
> > > Pd-list at lists.iem.at mailing list
> > > UNSUBSCRIBE and account-management ->
> http://lists.puredata.info/listinfo/pd-list
> > >
>
>
> _______________________________________________
> Pd-list at lists.iem.at mailing list
> UNSUBSCRIBE and account-management ->
> http://lists.puredata.info/listinfo/pd-list
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.puredata.info/pipermail/pd-list/attachments/20151122/9f2e6412/attachment-0001.html>