alsa-lib and floating point support

alsa-lib and floating point support

[Ricard Wanderlof]

We recently came across an issue with the alsa-lib softvol plugin, which 
requires hardware floating-point (i.e. --with-softfloat not enabled during 
build) support in order to use anything but a set of default values (min 
gain -51 dB, max 0 dB, and 256 gain steps). With no hardware floating 
point support, the plugin just prints an error message when loaded.

static int softvol_load_control(snd_pcm_t *pcm, snd_pcm_softvol_t *svol,
 ...
#ifndef HAVE_SOFT_FLOAT
 ...
                        double v = (pow(10.0, db / 20.0) * (double)(1 << 
VOL_SCALE_SHIFT));
 ...
#else
                SNDERR("Cannot handle the given dB range and resolution");
                return -EINVAL;
#endif
        }
 ...

I'm trying to understand why it is critical to have hardware floating 
point support in this case. The pow() function is very resource hungry, 
true, but the function is only called when the plugin is loaded, and not 
for instance on every sample while the stream is running, so on the whole 
I would expect the impact to be minimal. Is there some other rationale 
that I'm missing?

/Ricard
-- 
Ricard Wolf Wanderlöf                           ricardw(at)axis.com
Axis Communications AB, Lund, Sweden            www.axis.com
Phone +46 46 272 2016                           Fax +46 46 13 61 30

Re: alsa-lib and floating point support

[Takashi Iwai]

On Wed, 20 Sep 2017 12:55:30 +0200,
Ricard Wanderlof wrote:
> 
> 
> We recently came across an issue with the alsa-lib softvol plugin, which 
> requires hardware floating-point (i.e. --with-softfloat not enabled during 
> build) support in order to use anything but a set of default values (min 
> gain -51 dB, max 0 dB, and 256 gain steps). With no hardware floating 
> point support, the plugin just prints an error message when loaded.
> 
> static int softvol_load_control(snd_pcm_t *pcm, snd_pcm_softvol_t *svol,
>  ...
> #ifndef HAVE_SOFT_FLOAT
>  ...
>                         double v = (pow(10.0, db / 20.0) * (double)(1 << 
> VOL_SCALE_SHIFT));
>  ...
> #else
>                 SNDERR("Cannot handle the given dB range and resolution");
>                 return -EINVAL;
> #endif
>         }
>  ...
> 
> I'm trying to understand why it is critical to have hardware floating 
> point support in this case. The pow() function is very resource hungry, 
> true, but the function is only called when the plugin is loaded, and not 
> for instance on every sample while the stream is running, so on the whole 
> I would expect the impact to be minimal. Is there some other rationale 
> that I'm missing?

The softfloat option is to avoid calculation in float as much as
possible, i.e. alsa-lib will be built without any usage of math
library.


Takashi

Re: alsa-lib and floating point support

[Ricard Wanderlof]

On Wed, 20 Sep 2017, Takashi Iwai wrote:

> > I'm trying to understand why it is critical to have hardware floating 
> > point support in this case. The pow() function is very resource hungry, 
> > true, but the function is only called when the plugin is loaded, and not 
> > for instance on every sample while the stream is running, so on the whole 
> > I would expect the impact to be minimal. Is there some other rationale 
> > that I'm missing?
> 
> The softfloat option is to avoid calculation in float as much as
> possible, i.e. alsa-lib will be built without any usage of math
> library.

"as much as possible" sounds like the softvol case would be a reasonable 
exception, on the other hand, if the goal is to completely avoid using the 
softfloat library I can see the point. Is the reason to avoid dependency 
issues, or to minimize the total resulting code size, especially on 
systems with limited memory?

/Ricard
-- 
Ricard Wolf Wanderlöf                           ricardw(at)axis.com
Axis Communications AB, Lund, Sweden            www.axis.com
Phone +46 46 272 2016                           Fax +46 46 13 61 30

Re: alsa-lib and floating point support

[Takashi Iwai]

On Wed, 20 Sep 2017 16:45:46 +0200,
Ricard Wanderlof wrote:
> 
> 
> On Wed, 20 Sep 2017, Takashi Iwai wrote:
> 
> > > I'm trying to understand why it is critical to have hardware floating 
> > > point support in this case. The pow() function is very resource hungry, 
> > > true, but the function is only called when the plugin is loaded, and not 
> > > for instance on every sample while the stream is running, so on the whole 
> > > I would expect the impact to be minimal. Is there some other rationale 
> > > that I'm missing?
> > 
> > The softfloat option is to avoid calculation in float as much as
> > possible, i.e. alsa-lib will be built without any usage of math
> > library.
> 
> "as much as possible" sounds like the softvol case would be a reasonable 
> exception, on the other hand, if the goal is to completely avoid using the 
> softfloat library I can see the point. Is the reason to avoid dependency 
> issues, or to minimize the total resulting code size, especially on 
> systems with limited memory?

Well, actually no, because we don't link with -lm when the option is
specified, so far.  That is, the softfloat option is for controlling
whether to allow the usage of math function or not.


Takashi