yeah, I also get lots of glitches and artifacts with a 64 minimum window. I have to increase the delay up to 50ms so I get rid of them, which is kinda bad, even though it seems incredibly efficient. 

Brent is not using gardner's approach I assume. And what this does is that it increases the window over and over in the following partitions, this is why it gets so much lighter on the CPU

cheers 

Em qua, 9 de jan de 2019 às 07:47, Roman Haefeli <reduzent@gmail.com> escreveu:
Hi

It performs even better than William Brent's [convolve~] external, even
with small delays. When both set to 256, I get 9% load vs. 26%.

However, I get artefacts with a setting of 64 samples. When loading the
various IRs, the result sometimes sounds glitchy. Setting of 128 or
higher are always fine, regardless of IR.

Also with the church.wav and a setting of 64 you can hear the glitches
by feeding it a [phasor~ 5].

BTW: I can load many instances just fine (Pd 0.49 on Linux amd64)

Thanks for sharing!

Roman

On Tue, 2019-01-08 at 22:54 +0100, Philipp Schmalfuß wrote:
> Hi Alexandre,
>
> i made a pd-vanilla abstraction for real time convolution a while
> ago. 
> it is part of a collection of pd abstractions that i am planning to 
> share with the community, soon...
> it loosely follows gardner's approach 
> http://www.cs.ust.hk/mjg_lib/bibs/DPSu/DPSu.Files/Ga95.PDF
>
> with this, i get about 8-10% cpu-load with the church-IR and 64 
> samples min. fft-blocksize on an old lenovo t430 running linux. 
> however, i get the ugly clicks when i have more than one instance
> of 
> the abstraction running and on windows it causes pd-crashes, so i'm 
> not perfectly happy.
> i think it could be improved a lot by precomputing the IR like in
> your patch.
>
> https://we.tl/t-HYsWQww10V
>
> cheers!
>
> Quoting Alexandre Torres Porres <porres@gmail.com>:
>
> > Bug, for some reason, you may need to recreate the object so the
> > sound
> > comes out. I have no idea yet why...
> >
> > Em ter, 8 de jan de 2019 às 14:52, Alexandre Torres Porres <
> > porres@gmail.com>
> > escreveu:
> >
> > > oops, I hads uploaded the wrong file, here's the hopefully
> > > correct and
> > > last word on it
> > >
> > > https://www.dropbox.com/s/05xl7ml171noyjq/convolution~.zip?dl=0
> > >
> > > and my CPU load is actually at about 57%, not 50%
> > >
> > > The last file I uploaded was using a compiled object to perform
> > > the
> > > complex multiplication and that helped a little with the
> > > efficiency. I'm
> > > gonna use it for my non vanilla abstraction that I'm bringing
> > > into my ELSE
> > > library.
> > >
> > > cheers
> > >
> > >
> > >
> > > Em ter, 8 de jan de 2019 às 14:13, Alexandre Torres Porres <
> > > porres@gmail.com> escreveu:
> > >
> > > > Ok, here's the new deal...
> > > >
> > > > https://www.dropbox.com/s/l69gzv98g3th5d1/conv.rev~.zip?dl=0
> > > >
> > > > there are two subpatches for testing, one is light with a
> > > > relative big
> > > > window partition (1024) and a short Impulse Response (2 secs).
> > > >
> > > > The other is quite heavy, it's an 8 sec long IR with a window
> > > > size of
> > > > 512! This one takes just a bit over 50% of my CPU power, and
> > > > I'm on a last
> > > > generation macbook pro (2.6Ghz processor)... but I need to
> > > > increase the
> > > > Delay (msec) from 5 to 10 in the audio settings, otherwise I
> > > > get terrible
> > > > clicks!
> > > >
> > > > William Brent's convolve is ridiculously much more efficient,
> > > > the same
> > > > parameters take about 14% of my CPU power and I can use a delay
> > > > of 5 ms in
> > > > the audio settings.
> > > >
> > > > But anyway, this is useful for teaching and apps that implement
> > > > a light
> > > > convolution reverb (short IR/not too short window) need pure
> > > > vanilla
> > > > (libpd/camomille and stuff)
> > > >
> > > > Cheers!
> > > >
> > > > Em dom, 6 de jan de 2019 às 14:25, Alexandre Torres Porres <
> > > > porres@gmail.com> escreveu:
> > > >
> > > > > Meanwhile, I deleted the original file so people can't get it
> > > > > anymore :)
> > > > >
> > > > > Em dom, 6 de jan de 2019 às 14:16, Alexandre Torres Porres <
> > > > > porres@gmail.com> escreveu:
> > > > >
> > > > > > Hi, quick updates and developments over my weekend
> > > > > >
> > > > > >
> > > > > > > On Thursday, 3 January 2019, 04:19:50 GMT, Alexandre
> > > > > > > Torres Porres <
> > > > > > > porres@gmail.com> wrote:
> > > > > > >
> > > > > > > what you think, is it working?
> > > > > >
> > > > > >
> > > > > > So, the patch/algorithm was wrong and I've fixed
> > > > > >
> > > > > >
> > > > > > > Both objects on the help file take about 40% of my CPU
> > > > > > > power, but I'm
> > > > > > > on a wild machine
> > > > > > >
> > > > > >
> > > > > > I was able to do a few more things and make it much more
> > > > > > efficient
> > > > > >
> > > > > >
> > > > > > > I tried the idea of having each partition work with FFT
> > > > > > > saved on
> > > > > > > tables, so we wouldn't need to perform FFTs in different
> > > > > > > instances of
> > > > > > > clone, but that doesn't seem to be possible.
> > > > > >
> > > > > >
> > > > > > This is because things were wrong, like I said, now that
> > > > > > I've fixed it,
> > > > > > that was possible.
> > > > > >
> > > > > > But my current version is not vanilla anymore, as I'm
> > > > > > developing this
> > > > > > object to include it in my "ELSE" library. Once I'm done
> > > > > > I'll try to make
> > > > > > another vanilla compatible abstraction and re share it!
> > > > > >
> > > > > > Cheers
> > > > > >
>
>
>
>
>
> _______________________________________________
> Pd-list@lists.iem.at mailing list
> UNSUBSCRIBE and account-management ->
> https://lists.puredata.info/listinfo/pd-list
_______________________________________________
Pd-list@lists.iem.at mailing list
UNSUBSCRIBE and account-management -> https://lists.puredata.info/listinfo/pd-list