[alsa-devel] Reasonable sound hardware parameters?

[Lorenz Kolb]

> Message: 4
> Date: Mon, 28 May 2007 10:52:18 +0200 (CEST)
> From: Jaroslav Kysela <perex at suse.cz>
> Subject: Re: [alsa-devel] Reasonable sound hardware parameters?
> To: Joachim F?rster <mls.JOFT at gmx.de>
> Cc: alsa-devel at alsa-project.org
> Message-ID: <Pine.LNX.4.61.0705281045310.7894 at tm8103.perex-int.cz>
> Content-Type: text/plain; charset="iso8859-2"
> 
> On Sat, 26 May 2007, Joachim F?rster wrote:
> 
> >
> > First question: To implement mmap on such kind of device IO memory, 
> > I have to use snd_pcm_lib_mmap_iomem() as mmap() callback - like it 
> > is used in the rme32 driver, right?
> 
> Yes. rme32 is quite similar.
> 
> > Second question: Do the following parameters make sense regarding 
> > the size of the HW ring buffer and period sizes:
> >
> > .buffer_bytes_max=16*1024;
> > .period_bytes_max=8*1024;
> > .period_bytes_min=?? resonable value ??
> 
> 32 or 64
> 
> > .periods_min=2;
> > .periods_max=?? resonable value ??
> 
> 16
> 
> > Furthermore, the sound controller will fire an interrupt after one 
> > period is played. Are there any arguments against these values? Any 
> > comments?
> 
> Yes, I think that the access through a PCI I/O window to a PCI device 
> is much slower than PCI DMA transfers with bursts. I would really 
> consider to design hardware with real DMA. Of course, if it's only 
> experimental stuff, then don't worry.
> 
> 						Jaroslav
> 
> -----
> Jaroslav Kysela <perex at suse.cz>
> Linux Kernel Sound Maintainer
> ALSA Project, SUSE Labs

Hi Jaroslav,

First of all: thanks for Your answer.
Second: We do not have a PCI Bus. It is an embedded system with an OCP-Bus
(aka OPB).

The ML403 holds a SoC with PPC and FPGA.

So the only difference between DMA with bursts and a dedicated memory (as
far as we do see it) is the amount of time the bus gets occupied.
With dedicated memory the system will (hopefully ;-)) do a burst write each
period. So what happens in DMA case (except the more complex hardware as the
DMA controller has to be written (in VHDL) as well and even more memory is
located at an more complex bus (PLB - processor local bus): the bus will be
occupied twice: (1) CPU writes data to memory. --> start DMA --> (2)
Controller fetches data from memory.
While (2) is in progress the poor CPU is "disconnected" from the bus.
In an embedded system with a not too powerful CPU (PPC 405 @300MHz) this
might be quite inefficient.
Or did we miss something?

Regards,

Lorenz
--
Lorenz Kolb <alsa-dev at lkmail.de>
Embedded specialist
ESIC-Solutions