Hi,
I'm getting a slow-but-steady climbing CPU when running some synthesis patches. I (like the poster below) have sub-modules that can be switch~ed on and off. An archive search turned up the following un-replied post:
Hi all
Okay. I'm stumped.
Recently, I've noticed that my cpu meter has been steadily rising. I
think I may have found the culprit.A while back I thought I'd put a power switch (just a switch~ object)
on my poor-man's granular delay line, as my performance patches are
getting a bit bulky.Try this (if you have the time):
http://www.sideshowmedia.ca/cputest.zip
cputest.pd: (needs expr, but I think that's all)
- turn on audio in pd, but leave the power (p toggle) in "lois"
off. Watch the cpu meter climb (over a period of 5 minutes or so -
the toggle prints the cputime every 10 seconds). On my computer (G4
Powerbook, HCS extended 0.38.4) it starts around 5% and hits 10%
after 5 minutes.
- Do the same with "lois" on. For me, cpu hovers around 13-14%. No
increase.
- Repeat step 1. After 10-15 minutes, cputime is higher than with
"lois" on. Switch "lois" on, and cpu drops. Huh?
I have no idea what's going on, but am I missing something
important? Thanks in advance.cheers dafydd
Does anybody have any insight on this? I don't remember this happening until fairly recently, but don't know if it dates to when I started switch~ing modules. I'm currently running 0.39.2-extended RC1 on a MacBook Pro.
Phil Stone
Hi Phil,
just to keep this in people's minds... I have exact same problem described in these emails with my Particle Chamber granular synthesis patch. It has 32 "switch~ed" voices. Using one instance seems to be OK, but two or more leads to exponentially-growing CPU usage that eventually makes PD unresponsive to the GUI (and eventually would probably crash PD if I let it go on...) I guess this could relate to the "isn't the GUI supposed to have lower priority than process?" thread, since this particular problem makes PD a bit unusable for the kind of performances I want to be doing with it. The fact that PD lacks a usable voice-allocation method (i.e. CPU resources can be allocated and de-allocated to a voice, something which switch~ has a bug with, and which Nqpoly~ and so on does not do at all) means that I must start looking into SuperCollider to work with my polyphonic granular synthesis stuff. Sad but true...
d.
Phil Stone wrote:
Hi,
I'm getting a slow-but-steady climbing CPU when running some synthesis patches. I (like the poster below) have sub-modules that can be switch~ed on and off. An archive search turned up the following un-replied post:
Hi all
Okay. I'm stumped.
Recently, I've noticed that my cpu meter has been steadily rising. I
think I may have found the culprit.A while back I thought I'd put a power switch (just a switch~ object)
on my poor-man's granular delay line, as my performance patches are
getting a bit bulky.Try this (if you have the time):
http://www.sideshowmedia.ca/cputest.zip
cputest.pd: (needs expr, but I think that's all)
- turn on audio in pd, but leave the power (p toggle) in "lois"
off. Watch the cpu meter climb (over a period of 5 minutes or so -
the toggle prints the cputime every 10 seconds). On my computer (G4
Powerbook, HCS extended 0.38.4) it starts around 5% and hits 10%
after 5 minutes.
- Do the same with "lois" on. For me, cpu hovers around 13-14%. No
increase.
- Repeat step 1. After 10-15 minutes, cputime is higher than with
"lois" on. Switch "lois" on, and cpu drops. Huh?
I have no idea what's going on, but am I missing something
important? Thanks in advance.cheers dafydd
Does anybody have any insight on this? I don't remember this happening until fairly recently, but don't know if it dates to when I started switch~ing modules. I'm currently running 0.39.2-extended RC1 on a MacBook Pro.
Phil Stone
PD-list@iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list
On Wed, 2007-06-06 at 13:27 +0300, Derek Holzer wrote:
Hi Phil,
just to keep this in people's minds... I have exact same problem described in these emails with my Particle Chamber granular synthesis patch. It has 32 "switch~ed" voices. Using one instance seems to be OK, but two or more leads to exponentially-growing CPU usage that eventually makes PD unresponsive to the GUI (and eventually would probably crash PD if I let it go on...) I guess this could relate to the "isn't the GUI supposed to have lower priority than process?" thread, since this particular problem makes PD a bit unusable for the kind of performances I want to be doing with it. The fact that PD lacks a usable voice-allocation method (i.e. CPU resources can be allocated and de-allocated to a voice, something which switch~ has a bug with, and which Nqpoly~ and so on does not do at all) means that I must start looking into SuperCollider to work with my polyphonic granular synthesis stuff. Sad but true...
this is definitely not a general problem of pd. the most netpd-instruments do switch~ing all the time and i never noticed an increasing cpu usage over time on my and on many other computers.
that is why i think it would be interesting to know, if this problem is os or pd-distro specific. maybe it would make sense to collect information about systems, where this is happening and where not, in order to isolate the source of the problem. you might also provide a little testpatch, that triggers to problem on your machine, so that everyone can easily test as well.
here:
ubuntu dapper pd-vanilla 0.40.2 (compiled from source) no increased cpu usage due to [switch~]ing
Telefonate ohne weitere Kosten vom PC zum PC: http://messenger.yahoo.de
Hi Roman,
Roman Haefeli wrote:
this is definitely not a general problem of pd. the most netpd-instruments do switch~ing all the time and i never noticed an increasing cpu usage over time on my and on many other computers.
you might also provide a little testpatch, that triggers to problem on your machine, so that everyone can easily test as well.
One test patch came up earlier in thread:
http://lists.puredata.info/pipermail/pd-list/2007-04/049372.html
here:
ubuntu dapper pd-vanilla 0.40.2 (compiled from source) no increased cpu usage due to [switch~]ing
Here:
Mac Powerbook 1.67 GHz PowerPC G4
OSX 10.4.9, PD Extended 0.39.2 and Gentoo Linux, PD 0.39.2 self-compiled
both exhibit exponential CPU climbing, will test more with mentioned test patch and report back
best, d.
Ooops! File not found! Dafydd...could you re-post your test patch please?
http://www.sideshowmedia.ca/cputest.zip
thx, d.
Derek Holzer wrote:
One test patch came up earlier in thread:
http://lists.puredata.info/pipermail/pd-list/2007-04/049372.html
Oh gosh...
I think I can remember what it involved, but I don't think I have the patch anymore.
Let me see if I can dig it up.
cheers dafydd
On 6/6/07, Derek Holzer derek@umatic.nl wrote:
Ooops! File not found! Dafydd...could you re-post your test patch please?
http://www.sideshowmedia.ca/cputest.zip
thx, d.
Derek Holzer wrote:
One test patch came up earlier in thread:
http://lists.puredata.info/pipermail/pd-list/2007-04/049372.html
-- derek holzer ::: http://www.umatic.nl ::: http://blog.myspace.com/macumbista ---Oblique Strategy # 104: "Listen in total darkness, or in a very large room, very quietly"
Hi Derek,
I'm not positive about this, but I think I've traced the CPU-climb to the [moog~] filter. It has an odd behavior that can be broken into three phases:
- no audio greater than zero amplitude has passed through it yet (since
patch load) 2) audio of some amplitude >0 passes through 3) audio passing through goes back to zero amplitude
At 3) (or, more accurately, several seconds after), its CPU usage climbs fairly dramatically, only to drop down again when the amplitude of the audio goes up again. Something in the filter model doesn't like zero amplitude, but only after it's first processed something greater than zero.
I meant to test this more rigorously, but got sidetracked. At any rate, when I switch~ the [moog~] filter out, CPU usage holds steady. To be clear, I don't think this relates to [switch~]. I'm on a Macbook Pro, currently running 0.39rc2.
Phil
Derek Holzer wrote:
Hi Phil,
just to keep this in people's minds... I have exact same problem described in these emails with my Particle Chamber granular synthesis patch. It has 32 "switch~ed" voices. Using one instance seems to be OK, but two or more leads to exponentially-growing CPU usage that eventually makes PD unresponsive to the GUI (and eventually would probably crash PD if I let it go on...) I guess this could relate to the "isn't the GUI supposed to have lower priority than process?" thread, since this particular problem makes PD a bit unusable for the kind of performances I want to be doing with it. The fact that PD lacks a usable voice-allocation method (i.e. CPU resources can be allocated and de-allocated to a voice, something which switch~ has a bug with, and which Nqpoly~ and so on does not do at all) means that I must start looking into SuperCollider to work with my polyphonic granular synthesis stuff. Sad but true...
d.
Phil Stone wrote:
Hi,
I'm getting a slow-but-steady climbing CPU when running some synthesis patches. I (like the poster below) have sub-modules that can be switch~ed on and off. An archive search turned up the following un-replied post:
Hi all
Okay. I'm stumped.
Recently, I've noticed that my cpu meter has been steadily rising.
I think I may have found the culprit.A while back I thought I'd put a power switch (just a switch~ object) on my poor-man's granular delay line, as my performance patches are getting a bit bulky.
Try this (if you have the time):
http://www.sideshowmedia.ca/cputest.zip
cputest.pd: (needs expr, but I think that's all)
- turn on audio in pd, but leave the power (p toggle) in "lois"
off. Watch the cpu meter climb (over a period of 5 minutes or so -
the toggle prints the cputime every 10 seconds). On my computer (G4 Powerbook, HCS extended 0.38.4) it starts around 5% and hits 10% after 5 minutes.
- Do the same with "lois" on. For me, cpu hovers around 13-14%.
No increase.
- Repeat step 1. After 10-15 minutes, cputime is higher than with
"lois" on. Switch "lois" on, and cpu drops. Huh?
I have no idea what's going on, but am I missing something
important? Thanks in advance.cheers dafydd
Does anybody have any insight on this? I don't remember this happening until fairly recently, but don't know if it dates to when I started switch~ing modules. I'm currently running 0.39.2-extended RC1 on a MacBook Pro.
Phil Stone
PD-list@iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list
This could almost be the denormal problem again. Is it an Intel processor on the MacBook? Intels are notorious for poor handling of denormal numbers, which occur when the mantissa (decimal places) of a number representing an audio signal becomes too long (such as happens with reverb tails, etc). Check the archives for more than just a bit of discussion on this, with some solutions (?) and certainly some workarounds.
Sounds like I need to look deeper into my own patches as well, to find this CPU climbing business in my granulators. I find it most weird that one instance has no issues, but two or more cause problems. It's like they interfere with each other. I rigorously checked all send/receive pairs and such, so I don't know what it could be at this point. Except grossly inconvenient!
d.
Phil Stone wrote:
Hi Derek,
I'm not positive about this, but I think I've traced the CPU-climb to the [moog~] filter. It has an odd behavior that can be broken into three phases:
- no audio greater than zero amplitude has passed through it yet (since
patch load) 2) audio of some amplitude >0 passes through 3) audio passing through goes back to zero amplitude
At 3) (or, more accurately, several seconds after), its CPU usage climbs fairly dramatically, only to drop down again when the amplitude of the audio goes up again. Something in the filter model doesn't like zero amplitude, but only after it's first processed something greater than zero.
I meant to test this more rigorously, but got sidetracked. At any rate, when I switch~ the [moog~] filter out, CPU usage holds steady. To be clear, I don't think this relates to [switch~]. I'm on a Macbook Pro, currently running 0.39rc2.
Derek Holzer wrote:
This could almost be the denormal problem again. Is it an Intel processor on the MacBook? Intels are notorious for poor handling of denormal numbers, which occur when the mantissa (decimal places) of a number representing an audio signal becomes too long (such as happens with reverb tails, etc). Check the archives for more than just a bit of discussion on this, with some solutions (?) and certainly some workarounds.
Some solutions are presented here, with rationales as when each solution is best used:
http://www.musicdsp.org/files/denormal.pdf
In my exponential decay envelopes, I add a small number and subtract it again, which forces denormals (which are very close to zero) to be exactly zero.
Example: if you have 4 significant digits:
denoraml = 0.00000001234 denormal + 0.0001234 => 0.00012341234 truncate to 4 significant digits => 0.0001234 truncated - 0.0001234 => 0 (exact)
This has to be done *within* the feedback loop, as that is where the denormals cause bigger problems.
Other methods may be more appropriate for different scenarios...
BTW: it isn't just Intel processors, on AMD athlon-xp I reduced CPU usage a fair bit too (although I hear Intel Pentium-4 is particularly bad).
Claude
Derek Holzer wrote:
This could almost be the denormal problem again. Is it an Intel processor on the MacBook? Intels are notorious for poor handling of denormal numbers, which occur when the mantissa (decimal places) of a number representing an audio signal becomes too long (such as happens with reverb tails, etc). Check the archives for more than just a bit of discussion on this, with some solutions (?) and certainly some workarounds.
Yes, it's an Intel Core Duo in the MacBook. When I first noticed this pattern, I thought it sounded like denormalling, but I dismissed the idea, thinking that Core Duo's couldn't *still* have this problem.
How wrong I was. Your reply prompted me to do a little web searching.
See: http://charm.cs.uiuc.edu/subnormal/ and notice how abysmally the
Core Duo performs in denormal situations -- worse than previous Intels,
if this benchmark is to be believed.
Thank you (and others who mentioned this) for pushing me past the "can't be" stage. I'll look into workarounds.
Phil
Hi folks
Okay - I think this is more or less the same patch, with a couple of minor changes:
http://sideshowmedia.ca/cputest.zip
Oddly, I can't reproduce my original problem, although some fascinating new things have shown up. Perhaps it was a PowerPC issue? (I'm now on a MacBook)
Try switching on & off the main dsp and test_lois. On my computer (extended test 6 - delays not working in later tests) cputime goes up slightly when I turn dsp off.
I never figured out exactly why the climb was happening, but I found that it happened when I turned off dsp in the abstraction, but left the non-dsp elements running, ie continually trying to pass messages to line~ and vd~ objects etc. Switching off the mechanics when switching off dsp solved it for me.
I hope this helps! I'd love to hear if this test patch climbs on a PowerPC still.
cheers dafydd
On 6/6/07, Phil Stone pkstone@ucdavis.edu wrote:
Hi Derek,
I'm not positive about this, but I think I've traced the CPU-climb to the [moog~] filter. It has an odd behavior that can be broken into three phases:
- no audio greater than zero amplitude has passed through it yet (since
patch load) 2) audio of some amplitude >0 passes through 3) audio passing through goes back to zero amplitude
At 3) (or, more accurately, several seconds after), its CPU usage climbs fairly dramatically, only to drop down again when the amplitude of the audio goes up again. Something in the filter model doesn't like zero amplitude, but only after it's first processed something greater than zero.
I meant to test this more rigorously, but got sidetracked. At any rate, when I switch~ the [moog~] filter out, CPU usage holds steady. To be clear, I don't think this relates to [switch~]. I'm on a Macbook Pro, currently running 0.39rc2.
Phil
Derek Holzer wrote:
Hi Phil,
just to keep this in people's minds... I have exact same problem described in these emails with my Particle Chamber granular synthesis patch. It has 32 "switch~ed" voices. Using one instance seems to be OK, but two or more leads to exponentially-growing CPU usage that eventually makes PD unresponsive to the GUI (and eventually would probably crash PD if I let it go on...) I guess this could relate to the "isn't the GUI supposed to have lower priority than process?" thread, since this particular problem makes PD a bit unusable for the kind of performances I want to be doing with it. The fact that PD lacks a usable voice-allocation method (i.e. CPU resources can be allocated and de-allocated to a voice, something which switch~ has a bug with, and which Nqpoly~ and so on does not do at all) means that I must start looking into SuperCollider to work with my polyphonic granular synthesis stuff. Sad but true...
d.
Phil Stone wrote:
Hi,
I'm getting a slow-but-steady climbing CPU when running some synthesis patches. I (like the poster below) have sub-modules that can be switch~ed on and off. An archive search turned up the following un-replied post:
Hi all
Okay. I'm stumped.
Recently, I've noticed that my cpu meter has been steadily rising. I think I may have found the culprit.
A while back I thought I'd put a power switch (just a switch~ object) on my poor-man's granular delay line, as my performance patches are getting a bit bulky.
Try this (if you have the time):
http://www.sideshowmedia.ca/cputest.zip
cputest.pd: (needs expr, but I think that's all)
- turn on audio in pd, but leave the power (p toggle) in "lois"
off. Watch the cpu meter climb (over a period of 5 minutes or so - the toggle prints the cputime every 10 seconds). On my computer (G4 Powerbook, HCS extended 0.38.4) it starts around 5% and hits 10% after 5 minutes.
- Do the same with "lois" on. For me, cpu hovers around 13-14%.
No increase.
- Repeat step 1. After 10-15 minutes, cputime is higher than with
"lois" on. Switch "lois" on, and cpu drops. Huh?
I have no idea what's going on, but am I missing something important? Thanks in advance.
cheers dafydd
Does anybody have any insight on this? I don't remember this happening until fairly recently, but don't know if it dates to when I started switch~ing modules. I'm currently running 0.39.2-extended RC1 on a MacBook Pro.
Phil Stone
PD-list@iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list
PD-list@iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list
No problems here on a G4 PowerBook, but this delay is pretty rad, do you have any more?!
~Kyle
On 6/6/07, Dafydd Hughes dafydd61@gmail.com wrote:
Hi folks
Okay - I think this is more or less the same patch, with a couple of minor changes:
http://sideshowmedia.ca/cputest.zip
Oddly, I can't reproduce my original problem, although some fascinating new things have shown up. Perhaps it was a PowerPC issue? (I'm now on a MacBook)
Try switching on & off the main dsp and test_lois. On my computer (extended test 6 - delays not working in later tests) cputime goes up slightly when I turn dsp off.
I never figured out exactly why the climb was happening, but I found that it happened when I turned off dsp in the abstraction, but left the non-dsp elements running, ie continually trying to pass messages to line~ and vd~ objects etc. Switching off the mechanics when switching off dsp solved it for me.
I hope this helps! I'd love to hear if this test patch climbs on a PowerPC still.
cheers dafydd
On 6/6/07, Phil Stone pkstone@ucdavis.edu wrote:
Hi Derek,
I'm not positive about this, but I think I've traced the CPU-climb to the [moog~] filter. It has an odd behavior that can be broken into three phases:
- no audio greater than zero amplitude has passed through it yet (since
patch load) 2) audio of some amplitude >0 passes through 3) audio passing through goes back to zero amplitude
At 3) (or, more accurately, several seconds after), its CPU usage climbs fairly dramatically, only to drop down again when the amplitude of the audio goes up again. Something in the filter model doesn't like zero amplitude, but only after it's first processed something greater than zero.
I meant to test this more rigorously, but got sidetracked. At any rate, when I switch~ the [moog~] filter out, CPU usage holds steady. To be clear, I don't think this relates to [switch~]. I'm on a Macbook Pro, currently running 0.39rc2.
Phil
Derek Holzer wrote:
Hi Phil,
just to keep this in people's minds... I have exact same problem described in these emails with my Particle Chamber granular synthesis patch. It has 32 "switch~ed" voices. Using one instance seems to be OK, but two or more leads to exponentially-growing CPU usage that eventually makes PD unresponsive to the GUI (and eventually would probably crash PD if I let it go on...) I guess this could relate to the "isn't the GUI supposed to have lower priority than process?" thread, since this particular problem makes PD a bit unusable for the kind of performances I want to be doing with it. The fact that PD lacks a usable voice-allocation method (i.e. CPU resources can be allocated and de-allocated to a voice, something which switch~ has a bug with, and which Nqpoly~ and so on does not do at all) means that I must start looking into SuperCollider to work with my polyphonic granular synthesis stuff. Sad but true...
d.
Phil Stone wrote:
Hi,
I'm getting a slow-but-steady climbing CPU when running some synthesis patches. I (like the poster below) have sub-modules that can be switch~ed on and off. An archive search turned up the following un-replied post:
Hi all
Okay. I'm stumped.
Recently, I've noticed that my cpu meter has been steadily rising. I think I may have found the culprit.
A while back I thought I'd put a power switch (just a switch~ object) on my poor-man's granular delay line, as my performance patches are getting a bit bulky.
Try this (if you have the time):
http://www.sideshowmedia.ca/cputest.zip
cputest.pd: (needs expr, but I think that's all)
- turn on audio in pd, but leave the power (p toggle) in "lois"
off. Watch the cpu meter climb (over a period of 5 minutes or so - the toggle prints the cputime every 10 seconds). On my computer (G4 Powerbook, HCS extended 0.38.4) it starts around 5% and hits 10% after 5 minutes.
- Do the same with "lois" on. For me, cpu hovers around 13-14%.
No increase.
- Repeat step 1. After 10-15 minutes, cputime is higher than with
"lois" on. Switch "lois" on, and cpu drops. Huh?
I have no idea what's going on, but am I missing something important? Thanks in advance.
cheers dafydd
Does anybody have any insight on this? I don't remember this happening until fairly recently, but don't know if it dates to when I started switch~ing modules. I'm currently running 0.39.2-extended RC1 on a MacBook Pro.
Phil Stone
PD-list@iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list
PD-list@iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list
-- www.sideshowmedia.ca skype: chickeninthegrass
PD-list@iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list
i'm getting a similar problem in a large modular patch that does a LOT of switch~ing
cpu usage just gradually climbs. also i just switched from a powerbook running osx, to a toshiba running ubuntu studio, so i doubt it is platform specific.
this thread has given me a few ideas for how to debug it. will report back if i have any success.
Ha! Thanks, Kyle. Actually, it's mostly based on David Golightly's granular delay, which is a really elegant piece of work, and without which I never would have been able to try this. I ended up rewriting it without the scheduler, and changed some other stuff too. I'm halfway through another, but here's the most recent, which should be way more efficient and talks OSC. Disclaimer: it's still pretty old and has lots of my dumb patching!
http://sideshowmedia.ca/zips/lois.zip
So strange that weirdness isn't getting duplicated. I wonder what's up...
cheers dafydd
On 6/6/07, Kyle Klipowicz kyleklip@gmail.com wrote:
No problems here on a G4 PowerBook, but this delay is pretty rad, do you have any more?!
~Kyle
On 6/6/07, Dafydd Hughes dafydd61@gmail.com wrote:
Hi folks
Okay - I think this is more or less the same patch, with a couple of minor changes:
http://sideshowmedia.ca/cputest.zip
Oddly, I can't reproduce my original problem, although some fascinating new things have shown up. Perhaps it was a PowerPC issue? (I'm now on a MacBook)
Try switching on & off the main dsp and test_lois. On my computer (extended test 6 - delays not working in later tests) cputime goes up slightly when I turn dsp off.
I never figured out exactly why the climb was happening, but I found that it happened when I turned off dsp in the abstraction, but left the non-dsp elements running, ie continually trying to pass messages to line~ and vd~ objects etc. Switching off the mechanics when switching off dsp solved it for me.
I hope this helps! I'd love to hear if this test patch climbs on a PowerPC still.
cheers dafydd
On 6/6/07, Phil Stone pkstone@ucdavis.edu wrote:
Hi Derek,
I'm not positive about this, but I think I've traced the CPU-climb to the [moog~] filter. It has an odd behavior that can be broken into three phases:
- no audio greater than zero amplitude has passed through it yet (since
patch load) 2) audio of some amplitude >0 passes through 3) audio passing through goes back to zero amplitude
At 3) (or, more accurately, several seconds after), its CPU usage climbs fairly dramatically, only to drop down again when the amplitude of the audio goes up again. Something in the filter model doesn't like zero amplitude, but only after it's first processed something greater than zero.
I meant to test this more rigorously, but got sidetracked. At any rate, when I switch~ the [moog~] filter out, CPU usage holds steady. To be clear, I don't think this relates to [switch~]. I'm on a Macbook Pro, currently running 0.39rc2.
Phil
Derek Holzer wrote:
Hi Phil,
just to keep this in people's minds... I have exact same problem described in these emails with my Particle Chamber granular synthesis patch. It has 32 "switch~ed" voices. Using one instance seems to be OK, but two or more leads to exponentially-growing CPU usage that eventually makes PD unresponsive to the GUI (and eventually would probably crash PD if I let it go on...) I guess this could relate to the "isn't the GUI supposed to have lower priority than process?" thread, since this particular problem makes PD a bit unusable for the kind of performances I want to be doing with it. The fact that PD lacks a usable voice-allocation method (i.e. CPU resources can be allocated and de-allocated to a voice, something which switch~ has a bug with, and which Nqpoly~ and so on does not do at all) means that I must start looking into SuperCollider to work with my polyphonic granular synthesis stuff. Sad but true...
d.
Phil Stone wrote:
Hi,
I'm getting a slow-but-steady climbing CPU when running some synthesis patches. I (like the poster below) have sub-modules that can be switch~ed on and off. An archive search turned up the following un-replied post:
Hi all
Okay. I'm stumped.
Recently, I've noticed that my cpu meter has been steadily rising. I think I may have found the culprit.
A while back I thought I'd put a power switch (just a switch~ object) on my poor-man's granular delay line, as my performance patches are getting a bit bulky.
Try this (if you have the time):
http://www.sideshowmedia.ca/cputest.zip
cputest.pd: (needs expr, but I think that's all)
- turn on audio in pd, but leave the power (p toggle) in "lois"
off. Watch the cpu meter climb (over a period of 5 minutes or so - the toggle prints the cputime every 10 seconds). On my computer (G4 Powerbook, HCS extended 0.38.4) it starts around 5% and hits 10% after 5 minutes.
- Do the same with "lois" on. For me, cpu hovers around 13-14%.
No increase.
- Repeat step 1. After 10-15 minutes, cputime is higher than with
"lois" on. Switch "lois" on, and cpu drops. Huh?
I have no idea what's going on, but am I missing something important? Thanks in advance.
cheers dafydd
Does anybody have any insight on this? I don't remember this happening until fairly recently, but don't know if it dates to when I started switch~ing modules. I'm currently running 0.39.2-extended RC1 on a MacBook Pro.
Phil Stone
PD-list@iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list
PD-list@iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list
-- www.sideshowmedia.ca skype: chickeninthegrass
PD-list@iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list
--
---- -----
Dafydd Hughes wrote:
I never figured out exactly why the climb was happening, but I found that it happened when I turned off dsp in the abstraction, but left the non-dsp elements running, ie continually trying to pass messages to line~ and vd~ objects etc. Switching off the mechanics when switching off dsp solved it for me.
Yes, sending many messages to a vline~ in a switch~'d off subpatch causes CPU usage to rise dramatically. I nearly froze my box with Pd in RT mode when I set the metro period to 1ms and switch~'d the subpatch off, with the subpatch switch~'d on the cpu usage remains stable. See attached test patch...
Why this is happening:
Sending a message to vline~ creates a t_vseg, which are stored in a sorted linear linked list, which means the time taken to add each new line segment would be O(n), where n is the number of existing line segments.
If the subpatch containing vline~ is switch~'d off, the t_vseg's aren't removed by the dsp perform routine, which means that CPU usage rises quadratically if messages are sent to the vline~ at a constant rate (I think, it's been a while since I analyzed algorithmic time complexity).
A solution:
Use a better data structure than a linked list; a balanced tree would reduce time complexity to O(log n) instead of O(n) (if I remember correctly), admittedly at the cost of harder implementation.
Workaround:
Don't send messages to a switch~'d off vline~.
Claude
http://claudiusmaximus.goto10.org
#N canvas 0 0 450 300 10; #N canvas 0 0 450 300 $0-subpatch 0; #X obj 28 17 inlet; #X obj 26 270 outlet~; #X obj 208 269 switch~; #X obj 208 19 inlet; #X obj 29 48 t b b; #X obj 59 74 random 100; #X obj 24 94 random 100; #X obj 23 121 pack f f; #X obj 24 164 vline~; #X connect 0 0 4 0; #X connect 3 0 2 0; #X connect 4 0 6 0; #X connect 4 1 5 0; #X connect 5 0 7 1; #X connect 6 0 7 0; #X connect 7 0 8 0; #X connect 8 0 1 0; #X restore 68 117 pd $0-subpatch; #X obj 161 67 tgl 15 0 empty empty empty 17 7 0 10 -262144 -1 -1 0 1; #X obj 68 68 tgl 15 0 empty empty empty 17 7 0 10 -262144 -1 -1 0 1 ; #X obj 197 67 tgl 15 0 empty empty empty 17 7 0 10 -262144 -1 -1 0 1; #X msg 197 90 ; pd dsp $1; #X obj 302 88 metro 1000; #X obj 302 111 t b b; #X obj 302 133 cputime; #X obj 302 64 tgl 15 0 empty empty empty 17 7 0 10 -262144 -1 -1 0 1; #X floatatom 302 165 5 0 0 0 - - -; #X floatatom 119 67 5 0 0 0 - - -; #X obj 68 91 metro 10; #X text 14 205 cputime climbs when many messages are sent to a vline~ in a switched off subpatch...; #X connect 1 0 0 1; #X connect 2 0 11 0; #X connect 3 0 4 0; #X connect 5 0 6 0; #X connect 6 0 7 0; #X connect 6 1 7 1; #X connect 7 0 9 0; #X connect 8 0 5 0; #X connect 10 0 11 1; #X connect 11 0 0 0;
So just use this for now:
[X] [inlet messages] | | [t f f]_____ |____ | | | [switch] [spigot] | [vline~]
~Kyle
On 6/12/07, Claude Heiland-Allen claudiusmaximus@goto10.org wrote:
Dafydd Hughes wrote:
I never figured out exactly why the climb was happening, but I found that it happened when I turned off dsp in the abstraction, but left the non-dsp elements running, ie continually trying to pass messages to line~ and vd~ objects etc. Switching off the mechanics when switching off dsp solved it for me.
Yes, sending many messages to a vline~ in a switch~'d off subpatch causes CPU usage to rise dramatically. I nearly froze my box with Pd in RT mode when I set the metro period to 1ms and switch~'d the subpatch off, with the subpatch switch~'d on the cpu usage remains stable. See attached test patch...
Why this is happening:
Sending a message to vline~ creates a t_vseg, which are stored in a sorted linear linked list, which means the time taken to add each new line segment would be O(n), where n is the number of existing line segments.
If the subpatch containing vline~ is switch~'d off, the t_vseg's aren't removed by the dsp perform routine, which means that CPU usage rises quadratically if messages are sent to the vline~ at a constant rate (I think, it's been a while since I analyzed algorithmic time complexity).
A solution:
Use a better data structure than a linked list; a balanced tree would reduce time complexity to O(log n) instead of O(n) (if I remember correctly), admittedly at the cost of harder implementation.
Workaround:
Don't send messages to a switch~'d off vline~.
Claude
http://claudiusmaximus.goto10.org
#N canvas 0 0 450 300 10; #N canvas 0 0 450 300 $0-subpatch 0; #X obj 28 17 inlet; #X obj 26 270 outlet~; #X obj 208 269 switch~; #X obj 208 19 inlet; #X obj 29 48 t b b; #X obj 59 74 random 100; #X obj 24 94 random 100; #X obj 23 121 pack f f; #X obj 24 164 vline~; #X connect 0 0 4 0; #X connect 3 0 2 0; #X connect 4 0 6 0; #X connect 4 1 5 0; #X connect 5 0 7 1; #X connect 6 0 7 0; #X connect 7 0 8 0; #X connect 8 0 1 0; #X restore 68 117 pd $0-subpatch; #X obj 161 67 tgl 15 0 empty empty empty 17 7 0 10 -262144 -1 -1 0 1; #X obj 68 68 tgl 15 0 empty empty empty 17 7 0 10 -262144 -1 -1 0 1 ; #X obj 197 67 tgl 15 0 empty empty empty 17 7 0 10 -262144 -1 -1 0 1; #X msg 197 90 ; pd dsp $1; #X obj 302 88 metro 1000; #X obj 302 111 t b b; #X obj 302 133 cputime; #X obj 302 64 tgl 15 0 empty empty empty 17 7 0 10 -262144 -1 -1 0 1; #X floatatom 302 165 5 0 0 0 - - -; #X floatatom 119 67 5 0 0 0 - - -; #X obj 68 91 metro 10; #X text 14 205 cputime climbs when many messages are sent to a vline~ in a switched off subpatch...; #X connect 1 0 0 1; #X connect 2 0 11 0; #X connect 3 0 4 0; #X connect 5 0 6 0; #X connect 6 0 7 0; #X connect 6 1 7 1; #X connect 7 0 9 0; #X connect 8 0 5 0; #X connect 10 0 11 1; #X connect 11 0 0 0;
PD-list@iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list
Hallo, Claude Heiland-Allen hat gesagt: // Claude Heiland-Allen wrote:
A solution:
Use a better data structure than a linked list; a balanced tree would reduce time complexity to O(log n) instead of O(n) (if I remember correctly), admittedly at the cost of harder implementation.
Workaround:
Don't send messages to a switch~'d off vline~.
I think, people should be careful with sending messages to switched-off dsp-objects anyway and consider switching off message calculation inside their subpatches using [spigot] as well.
I suspect, other objects besides [vline~] may show similar behaviour. At least I've encountered the cpu-climbing in switched-off subpatches in some of my patches as well, that definitely weren't using [vline~]. I didn't yet find out which objects or combination of objects were causing this, but blocking message calculation managed to fix it.
Ciao
Frank Barknecht _ ______footils.org_ __goto10.org__
-
Claude Heiland-Allen -
Dafydd Hughes -
dafydd hughes -
Derek Holzer -
Frank Barknecht -
hard off -
Kyle Klipowicz -
Phil Stone -
Roman Haefeli