Re: [Faudiostream-devel] Hadamard matrix function

[Stéphane L. <le...@gr...>]

Le 9 déc. 2010 à 14:40, Orlarey Yann a écrit :

> Hi,
> 
> Sorry for my current high latency. I am now back from Asia, but still not in Grame. 
> Moreover I had not really the time to investigate the problem so far, therefore I don't 
> have a precise answer/ explanation to give. But anyway generating fast code is really 
> essential to Faust, so we clearly have a problem here that we will investigate and try to 
> solve.
> 
> My current hypothesis is that the code generated by Faust for the filters is not 
> efficient.

Well I tried the "don't use arrays for 1 ou 2 samples delay" optimization (that is using variables instead). It seems to help a bit (like, 2,5 % instead of 2,75%...) but not so much...

But looking at the cpp generated Faust, it seems so much different to Fons one... are we sure the complexity is comparable?


> So I will try alternative code generation to see if we can improve the 
> situation. Another problem to investigate, as pointed by stephane, is a bug in the vector 
> code generation that prevent these reverbs to run in (a potentially more efficient) vector 
> mode.

Yep. We should try to simplify the problem in the first step.

> 
> So many thanks to Fons for pointing the problem and to Julius for the zita-rev1 
> implementation in Faust. We will try to fill the performance gap...
> 
> Cheers
> 
> Yann
> 


Stéphane