[LAD] Aeolus

[hermann meyer]

Am 20.09.2013 17:58, schrieb Fons Adriaensen:
> On Thu, Sep 19, 2013 at 07:07:37PM +0200, hermann meyer wrote:
>   
>> I've attached the diff, for your review.
>> Some of the changes are not necessary needed, but handy for our use case.
>> Ignore please the #include "gx_compiler.h" and __rt_func
> Thanks. It would seem possible to merge the two versions, but
> I have a few questions/remarks.
It will be very welcome, if you merge ffmpeg support into zita-convolver
and we will switch to "official" version as soon you release it. :-)

> 1. You shouldn't have removed the 'options'. In the current release
>     they are used for things not related to the FFT library, and other
>     options may be added in the future. It has no performance penalty.
>
> 2. You have changed the gain normalisation from 0.5 / partition to
>     1.0 / partition. Why ? The 0.5 is not related to anything specific
>     FFTW. It's there because the FFT size is twice the partition size.
>     The result should be that convolution with a single impulse of 1.0
>     amplitude reproduces the input without a gain change.
we experience a reduced gain ,  with 1.0 / partition for ffmpeg FFT
and 0.5 / partition for fftw3, we get comparable results
when switching between ffmpeg and fftw3.

> 3. If av_malloc() can ever fail, this should raise an exception.
>
> 4. Assuming that the 'swap' operation that is part of the av FFT
>     only changes the order of complete complex values (and not of
>     their components), all the swaps can be removed. The only thing
>     that happens in the F-domain is that values with the same index
>     are multiplied. Their order is not important.
>
> Ciao,
>