WSJT-X/portaudio-v19/pa_unix_oss/low_latency_tip.txt

95 lines
3.0 KiB
Plaintext
Raw Normal View History

From: "Benno Senoner" <sbenno@gardena.net>
To: <music-dsp@shoko.calarts.edu>
Subject: Re: [music-dsp] coding realtime guitar efx on a "pc"
Date: Saturday, June 30, 2001 8:19 AM
Andr<EFBFBD>,
you are solving your problem the wrong way:
you need to use a single threaded solution which does this:
- set the audio I/O parameters to fragnum=4 fragsize=128 bytes (=32samples) if
you use stereo or fragsize=64 bytes (=32 samples) if you use mono.
(do not forget to activate fulltuplex with using the _TRIGGER_ stuff)
(you need to frist deactivate audio and then start the trigger after the DAC is
prefilled (see below))
This will give you a total input to output latency of 4x32 samples
= 128 samples which at 44.1kHz correspond to 2.9msec latency.
now set your process to SCHED_FIFO (see man sched_setscheduler)
after the initialization your code should do more than less this:
- write() 4 x 32 samples to the audio fd in order to prefill the DAC.
Without this you will get dropouts.
while(1) {
read() 32 samples from ADC
perform_dsp_stuff() on the 32 samples
write() 32 samples to DAC
}
If you use a low latency kernel and pay attention to all the stuff above, then
you will get rock solid 3msec latencies (plus eventual converter latencies but
these are in the 1-2msec range AFAIK).
Using multiple threads , pipes etc, only complicates your life and often makes
it impossible to achieve these low latences.
Realtime/audio programming is not an easy task , this is why people often
fail to get the desired results even if their hardware is low-latency capable.
The problem is that the final latency depends on the hardware you use,
the application and the operating system.
cheers,
Benno.
http://www.linuxaudiodev.org The Home of Linux Audio Development
On Sat, 30 Jun 2001, you wrote:
> On 2001-06-29 21:38 +0200, Benno Senoner wrote:
>
> > OSS/Free refuses to use a low # of frags ?
> >
> > That's a myth.
>
> I hope it is. :-)
>
> The fact is that ioctl(SNDCTL_DSP_SETFRAGMENT) succeeds with
> values as low a 0x10007 (one 128-B fragment) but the latency is
> still high enough to be clearly noticeable, which suggests that
> it's *way* above 2/3 ms. This is on an otherwise idle machine
> equipped with a SB PCI 128.
>
> But maybe it's me who's doing something wrong. I've been careful
> to flush stdio buffers or use unbuffered I/O (write(2)) but I
> may have let something else through.
>
> For example, since the signal processing and the I/O are done by
> two different vanilla processes communicating via pipes, it may
> be a scheduling granularity problem (E.G. the kernel giving the
> I/O process a time slice every 20 ms).
>
> --
> Andr<64> Majorel <amajorel@teaser.fr>
> http://www.teaser.fr/~amajorel/
>
> dupswapdrop -- the music-dsp mailing list and website: subscription info,
> FAQ, source code archive, list archive, book reviews, dsp links
> http://shoko.calarts.edu/musicdsp/
--
dupswapdrop -- the music-dsp mailing list and website: subscription info,
FAQ, source code archive, list archive, book reviews, dsp links
http://shoko.calarts.edu/musicdsp/