[PD] How's Pd limited?
jancsika at yahoo.com
Wed Feb 24 18:16:27 CET 2016
I think one of the big limitations is a difficulty in turning "hot" code "cool".
For example, suppose the [hungry~] abstraction is at the heart of your patch but it consumes a lot of CPU. What do you do?
Typically the process involves only two steps:1. make esoteric changes that marginally decrease the CPU load2. give up and port [hungry~] to a C or C++ external
#1 decreases readability, and #2 decreases portability (and hopefully readability as well).
Parallelization may be a means to address this, but it is a means and not an end. In any case the first place to start is
to profile CPU usage and patch performance, as well as signal and object performance within the patch. Pd needs tools
to accurately measure which classes and abstractions are responsible when a patch runs hot.
Desiredata apparently added some functionality to do that but it was apparently buggy and didn't get a lot of testing. Anyhow,
these tools are crucial to a sensible discussion of parallelization-- without them we can only measure object performance with
very blunt tools.
On Wednesday, February 24, 2016 10:53 AM, peiman khosravi <peimankhosravi at gmail.com> wrote:
One great advantage of maxmsp is gen, which gives you sample-level patching with the possibility of a one-sample delay.
On Tuesday, February 23, 2016, Samuel Burt <composer.samuel.burt at gmail.com> wrote:
One thing I attempted and couldn't find a solution for was the following, mostly owing to the limitation of interfacing with a 64 sample block size.
I wanted to have a directory of hundreds of audio recordings. Each one would be a single wavelength from an interesting sound, like a bass clarinet, marimba, harpsichord, tambourine, etc. Each would begin and end at a zero crossing so you could chain them together to make complex timbres. They could be chained in sequence, randomized, or loaded in meta-data-matched chunks. I ran into a problem figuring out how to trigger the next sound based on the ending of the last sound in a sample accurate way. Sound file loading or even buffer playback triggering waits until the start of the next block size before it updates. If you have a waveform that lasts 205 samples (64+64+64+13), you have a gap of 51 silent samples before the next waveform would start. Not only do you not get the continuous sound you want, this winds up creating a periodic pattern with a frequency of 689 Hz (44100/64).
David, I like your idea "what (if anything) someone tried to do in Pd, but couldn't given its limitations". I think this could be a wonderful challenge if we could have a monthly thread like this where the best minds among us come up with solutions to some of the hardest conceptual challenges in Pd.
I'm still struggling with loading dozens of files, audio dropouts, and other similar problems. Someone else expressed frustration about Pd's single-threaded status. I too have feared upgrading my computer based on the limitations of current multicore processors (although realistically I think we can all look at the "turbo-boost" level or whatever Intel calls it to determine where our processor might run with a demanding patch. I understand the fact that you can't run your audio process on multiple cores, because it is a linear process. It would be great if the GUI could run on a second core, a process that loads audio into memory could run on third core, while GEM could automatically run on a fourth core. I don't have any concept of how feasible that would be, though. Does the GUI in pd-l2orc run on a separate core?
Date: Tue, 23 Feb 2016 09:01:06 -0800
From: david medine <dmedine at ucsd.edu>
One thing I'd be interested in knowing about is what (if anything)
someone tried to do in Pd, but couldn't given its limitations (apart
from look/feel/convenience issues).
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...
More information about the Pd-list