Thread: [Linuxsampler-devel] Filter code optimization, sinewave .GIG files request

SourceForge Headquarters 1320 Columbia Street Suite 310 San Diego, CA 92101 +1 (858) 422-6466

Christian and I talked about how speed up the (LP/HP/BP etc) filter code.
Currently it's suboptimal because UpdateFilter() does an if() and a %
at each sample.
The a better method is to use the sythesis matrix like for pitch
and volume.
some issues: currently LS does use two separate filter instances
in case of stereo samples, but this wastes CPU since AFAIK
in other samplers filters on stereo samples are applied equally
on both the left and right channel. So you still need to run this code:

            inline bq_t Apply(const bq_t x) {
                bq_t y;

                y = this->b0 * x + this->b1 * this->x1 + this->b2 * this->x2 +
                    this->a1 * this->y1 + this->a2 * this->y2;
                KillDenormal(y);
                this->x2 = this->x1;
                this->x1 = x;
                this->y2 = this->y1;
                this->y1 = y;

                return y;
            }

two times (for each of the stereo channels), but
the coefficients can be shared thus if the filter code gets implemented
into the event synthesis matrix then instead of transfering 10 coefficients
at time two times (a1,a2,b0,b1,b2) you need only 5 and the L1/L2 cache
will say a big thank you :)

I'm still curious if transfering 5 coefficients/sample from memory is
faster than placing a conditional in the event loop and check for filter
updates which recalculate these coefficients (filter events should be
relatively sparse).
As said more benchmarking is needed here, but in the meanwhile since
Christian's event matrix is already working fine I'd say he should go ahead
with his solution. (but Christian make sure to use shared coefficients for left
and right filters).
At a later stage to speed up things further we can write a SIMD version of the
filter (SSE etc) which is especially useful in the stereo case where full
parallelism can be achieved.

http://www.northernsounds.com/ubbthreads/showflat.php?Cat=&Number=162226&page=0&view=collapsed&sb=5&o=&fpart=1#162226
Even if I'm currently not contributing code (only some ideas) at least I do some
PR for LS :)

Rui, I told northernsounds.com users that you are writing a very nice GUI, don't
disappoint them :)

Mark: about the ML archives, you are right, for some strange reason the last
months are missing (while on other projects the archives are ok).
Anyway I think the SF.net archives are buggy anyway so we should probably
simply subscribe an other ML archiving site to

About initializing LS via the LSCP, this is what Christian wrote some time ago:

> The easiest way currently is to write a LSCP script and to send it via
> netcat to the running sampler:
>
>        cat yourscript.lscp | nc -t localhost 8888

you must construct your own yourscript.lscp
I'm not sure if it makes sense to ship some default lscp script for now.
perhaps some 1-2 channel example but it would need to be tweaked by the user
anyway. If you want after some experimenting you could contribute such a script
and some small README how to tweak it as the lscp PDF is a bit heavy to read :)

PS2: there is still this nasty aliasing bug in LS,
enable cubic interpolation (happens with linear too but let's use cubic
to rule out problems in the low quality interpolation code) load a piano sample
and play high notes at very low velocities, keep them sustained, you clearly
hear an aliasing noise floor which will fade away when the voice finishes
playing.
I tried to track it down (tried to remove the envelope section, did check
the mixing code) but I did not find the cause yet.
If some GSt user has some time, he could you produce a few .GIG files a non
looped pure sinewave at 0dB (full scale), 30secs of lenght, both a mono and a
stereo version. make a version with 440Hz root note and one with 3520Hz. (which
equals to 440Hz shifted up 3 octaves).
So basically 4 GIGs. In this case we can benchmark the quality of the
interpolator too. (take the 440Hz sample, play it at high pitches and check
the output, take the  3520Hz sample and play it at low pitches to see if the
wave's shape looks still good.

You can zip them together and mail them me at
sb...@ga...   (I have no space limits).
and then I'll make them available on the LS site so that other people
can perform tests too.

thanks,
Benno
http://www.linuxsampler.org

-------------------------------------------------
This mail sent through http://www.gardena.net

Thread: [Linuxsampler-devel] Filter code optimization, sinewave .GIG files request

linuxsampler-devel