[Linuxsampler-devel] No records + some answers

[hilare <hil...@ya...>]

I am sending you this message again, without files,
linuxsampler-devel mail doesn't take the attached
files. I guess that some size limitation is involved
here ( 1M * 2 + 1.4M * 2 + 6M )


First, thank you for the informations.

I can confirm the bug on AC97, 44Khz setting leads to
resample inside the sound chip ( aka more xruns ). The
same is true for 24bits, the sound chip always
resamples it to 16bits, this is the reason for the
48Khz/16bits setting. Another bug on the AC97 prevents
any modification of the PCI priority, so the sound
chip is on the lowest level for PCI latency.

The topic on Linuxsampler bounded by memory speed was
very interesting. The Celeron use a 233Mhz speed while
the Pentium use a 400Mhz one, memtest reports them
respectively at 1800Mb/s and 2800Mb/s ( none of those
results is good, memory latency was reduced in
exchange for an overclock to the processor speed ). I
will try to decrease memory speed on the Pentium to
check if the processor can still assume the same
number of concurrent voices.

I will also try different memory allocations for
Linuxsampler, mainly because i want to use the disk
streaming, a dedicated disk is now assumed.

Back to your original post, you will find some
recording ( so with the sound of AC97, and Jconv ).

The original midi files are:

* OneBourbonOneScotchOneBeer.mid
* NI-BoogiePiece.mid
* NI-AHonkytonkPiece.mid

The NI came from Native Instrument and where demos for
Akoustic Piano. I had choose them because they are
short, so the .ogg files stay small.

Please note that the John Lee Hooker piece is modified
inside Rosegarden for instrument's setting ( i do not
follow some original instrument choices ).


The sound setting for the records was ( only
left/right audio channels per software, the sound chip
doesn't support more ):

Linuxsampler audio L/R ---> Jack playback L/R
     "       audio L/R ---> Jconv input L/R

Rosegarden   audio L/R ---> Jack playback L/R 
     "       audio L/R ---> Jconv input L/R

Jconv        audio L/R ---> Jack playback L/R

Jack       monitor L/R ---> Ardour track1 L/R

So you can ear the AC97 with Jconv.

A special track was created for
NI-Boogie-Jconv-only.ogg:

Jconv        audio L/R ---> Ardour track2 L/R

So you can ear only the result of Jconv.

To sum up, Jack audio input is always the addition of
native Linuxsampler + Jconv ( i will warn for the two
.ogg files that make an exception to this ). Which
means that Jconv is played live and never added to
post-recording.

The common setting are: 
- Piano PMI grandioso Bosendorfer,
 
- Pori Concert Hall binaural 7th row (
www.acoustics.hut.fi/projects/poririrs ) as IR for
convolution, 

- Linuxsampler 0.5.1.1 ( 160 voices ) over Jack 109.2
( Midi still carried by Alsa 1.0.14 ),

- jackd setting: realtime priority 30, Frames/period
128, Periods/buffers 2, Sample rate 48000 which gives
5,33ms latency, and two buffers are added 1 for input
latency and 1 for output latency. 


The results files are:

* JLHooker-OneBourbon.ogg

A multi instruments piece. Piano, bass, drum, trumpet
and accoustic guitar are played by Linuxsampler, while
the two electric guitars are soundfont played by
Rosegarden/DSSI/Fluidsynth and some Steve Harris
LADSPA plugins.

Jconv takes inputs from Linuxsampler and Rosegarden.



* NI-Boogie.ogg

"Standard" setting. Linuxsampler / piano + jconv.

* NI-Boogie-Jconv-only.ogg

Here only the output of Jconv was sent to Ardour. Not
very musical, this is only intended to show the work
and result of Jconv.


* NI-HonkyTonk-dry.ogg  

Linuxsampler / piano _without_ Jconv. So "dry",
whithout reverb.

* NI-HonkyTonk-Jconv-added.ogg

Standard setting. Linuxsampler / piano + jconv.





Best regards


--- Benno Senoner <ben...@go...>
wrote:

> As for Celeron , P4 difference, LS is mostly memory
> bound so probably if you
> push the polyphony up
> the RAM access speed will be the limiting factor.
> 
> 
> 
> Yes it is strange, LS runs stably on a slow 1GHz VIA
> although with lower
> polyphony.
> but keep in mind that an RT kernel is mandatory for
> low latencies.
> And often the sound card driver is important too. On
> some boxes with
> integrated audio chip with the default
> settings under linux performance sucks, you get
> droputs in jack even at 1024
> frames.
> I have not investigated the issue, probably one
> needs
> to pass some
> parameters to the drivers in order
> to achieve lower latencies. (I am not sure but I
> heard
> some consumer chips
> only support 48kHz in hw
> while 44.1kHz is emulated which is the cause of
> worse
> latencies, correct me
> if I am wrong).
> 
> 
> 
> During slow passages you have lots of sustained
> notes
> and probably the disk
> is not
> able to keep up with the amount of data that needs
> to
> be streamed.
> solutions are: a faster disk, larger sample preload
> sizes.
> 
> about the memory usage keep in mind that besides
> memory for preloading
> samples
> you need memory for the streaming buffers (one per
> voice).
> and as I already told some time ago, the streaming
> buffersizes are rounded
> up to a power of two.
> so a streaming buffer size of 35000 samples
> allocates
> 3 (24bit) x 35000 =
> 1050000 bytes rounded up to
> the next power of 2 = 2MBytes. so keep streaming
> bufffer size x 3 always a
> bit below of the next
> power of 2 in order to optimize memory usage.
> 
> About multi core, Alex Stone will certainly be able
> to
> explain the benefits
> he is getting on his dual core.
> (he told he will get a quad core soon so the numbers
> he will get out of his
> new box will be even more interesting
> as the quad has to share the memory amongst 4 cores)
> 






      ____________________________________________________________________________________
Looking for last minute shopping deals?  
Find them fast with Yahoo! Search.  http://tools.search.yahoo.com/newsearch/category.php?category=shopping