Speex dev - Dec 2005 - More floating point errors

Hi!

It seems that using a different compiler (I assume the
development is done with MSC) uncovers various issues
with the floating point.

Here is the second one I'm seeing - DOMAIN error from
sqrt().

In the open_loop_nbest_pitch(), this line:

          g = DIV32(corr[i-start],
10+SHR32(MULT16_16(spx_sqrt(e0),spx_sqrt(energy[i-start])
),6));

The values are:

e0 is 7.2
i is 142,
start is 17
energy [i-start] is -2.1004911E-015


Hope this information is enough to go on.

Tia,
.a
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
http://lists.xiph.org/pipermail/speex-dev/attachments/20051212/931e9f96/attachment.html

> It seems that using a different compiler (I assume the development is
> done with MSC) uncovers various issues with the floating point.
No, all development is done on gcc/Linux.
> Here is the second one I'm seeing - DOMAIN error from sqrt().
> 
> In the open_loop_nbest_pitch(), this line:
> 
>           g = DIV32(corr[i-start], 10
> +SHR32(MULT16_16(spx_sqrt(e0),spx_sqrt(energy[i-start])),6));
change that to:

g = DIV32(corr[i-start], 10+SHR32(MULT16_16(spx_sqrt(e0),spx_sqrt(1
+energy[i-start])),6));

(note the "1+" in the spx_sqrt)

As for this appearing only on some version/compiler configurations, the
only explanation is sheer luck (order in which the float operations are
done and when it goes to registers).

	Jean-Marc

I had this issue a (couple of?) weeks ago...  the solution then (from jm) was-

g = DIV32(corr[i-start], SHR32(MULT16_16(spx_sqrt(10 + e0), 
spx_sqrt(10+energy[i-start])),6));

This effects more than fixed point- basically, once this goes bad it will 
eventually
cause the codec to only produce fuzz in floating point, and it will result in
ol_pitch_coef getting set to NAN, which is then preserved for all of eternity
(or until the speex enc gets destroyed) in vbr->soft_pitch- i.e. you
can't
just
reset the state to fix it.

Tom

At 10:31 PM 12/12/2005, Alex Bakaev wrote:
>Hi!
>
>It seems that using a different compiler (I assume the development is done 
>with MSC) uncovers various issues with the floating point.
>
>Here is the second one I'm seeing - DOMAIN error from sqrt().
>
>In the open_loop_nbest_pitch(), this line:
>
>           g = DIV32(corr[i-start], 
> 10+SHR32(MULT16_16(spx_sqrt(e0),spx_sqrt(energy[i-start])),6));
>
>The values are:
>
>e0 is 7.2
>i is 142,
>start is 17
>energy [i-start] is -2.1004911E-015
>
>Hope this information is enough to go on.
>
>Tia,
>.a
>_______________________________________________
>Speex-dev mailing list
>Speex-dev@xiph.org
>http://lists.xiph.org/mailman/listinfo/speex-dev
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
http://lists.xiph.org/pipermail/speex-dev/attachments/20051213/abb7bb72/attachment.html