flac dev - Apr 2015 - "keep qlp coeff precision such that only 32-bit math is required"

Martijn van Beurden wrote:
> Yes, but that MAX value is used to loop over the
> qlp_coeff_precision values between MIN and MAX. So, if the
> qlp_coeff_precision value is limited in the loop but MAX is not
> limited, the loop does the exact same thing multiple times: a
> waste of time. Therefore, only the MAX should be limited.
>
> I don't think the logic is needed directly after
>> min_qlp_coeff_precision = max_qlp_coeff_precision =
encoder->protected_->qlp_coeff_precision;
> because that value is either chosen by the encoder (it is a
> lookup value, and is always small enough not to trigger this,
> set in stream_encoder.c around line 700) or it is set through
> the encoder, in which case it should probably not be overriden
> without notice.
Currently evaluate_lpc_subframe_() automatically adjusts qlp_coeff_precision
when a user sets non-standard order value (via -l option). I don't think
that it's incorrect behavior. Why not?

(And my initial point was that the current limit of qlp_coeff_precision is not
enough,
and 32-bit datapath is *not* ensured for 16-bit input audio).

Op 22-04-15 om 22:45 schreef lvqcl:
> Currently evaluate_lpc_subframe_() automatically adjusts 
> qlp_coeff_precision when a user sets non-standard order value 
> (via -l option). I don't think that it's incorrect behavior. 
> Why not?
Because what is the point in setting a non-standard value if it 
not used anyway?
> (And my initial point was that the current limit of 
> qlp_coeff_precision is not enough, and 32-bit datapath is 
> *not* ensured for 16-bit input audio).
Yes, the code should be used for 17-bit audio as well. That's no 
reason to duplicate the code though.

Martijn van Beurden wrote:
> Op 22-04-15 om 22:45 schreef lvqcl:
>> Currently evaluate_lpc_subframe_() automatically adjusts
>> qlp_coeff_precision when a user sets non-standard order value
>> (via -l option). I don't think that it's incorrect behavior.
>> Why not?
>
> Because what is the point in setting a non-standard value if it
> not used anyway?

I think that the comment in the code ("try to keep qlp coeff precision
such that only 32-bit math is required for decode of <=16bps streams")
explains it:
flac overrides the value of qlp_coeff_precision in order to increase decoding
speed.
64-bit additions, multiplications and shifts are relatively slow on 32-bit CPUs,
and 32-bit operations are preferable.