Re: [LAD] Is -ffast-math safe for audio?

Am 22.11.18 um 23:44 schrieb Robin Gareus:
> On 11/22/2018 09:27 PM, Hermann Meyer wrote:
>>
>> However, problems with NAN's and INF's when use -ffinite-math-only
>> occurs only when we save preset values to file, and only sometimes then.
> A shot in the dark..
>
> Serializing a float in most parts of the world uses comma as decimal
> separator. e.g. "0,5"
>
> Reading this in the C-locale (dot as decimal separator) with c/c++
> standard library functions returns 0. This may lead to a division by
> zero later.
>
> Cheers!
> robin

that happen during write, not during read. During read, nan's and inf's
will be replaced by standard values when detected, but it mean's that
original settings been lost.

The values been saved as 0.5 with a dot, the comma is used as separator
between the entry's.

(and no error with entry separation ever happen)

It could happen as well to values which may ain't been used/changed
during a session.

And it happen in a non-reproducible way. Very strange, and indeed only
when -ffinite-math-only

is turned on.

I've tried for weeks to find out what really happen, but could only nail
it to -ffinite-math-only.

-ffast-math, on the other hand changes the resulting sound, it's
minimal, but, notable.

The only flag from the -ffast-math macro we still use is -fno-math-errno.

regards

hermann
_______________________________________________
Linux-audio-dev mailing list
Linux-audio-dev@lists.linuxaudio.org
https://lists.linuxaudio.org/listinfo/linux-audio-dev
Received on Fri Nov 23 08:15:01 2018