[linux-dvb] Kernelspace floating pointer - was Re: Port of em28xx and xc3028 to new hybrid tuner framework.

Markus Rechberger mrechberger at gmail.com
Sat Sep 8 01:24:17 CEST 2007


On 9/8/07, Mauro Carvalho Chehab <mchehab at infradead.org> wrote:
> Markus wrote:
>
> > the em28xx uses the userland implementation since the current xceive
> > reference drivers use floating point algorithms.
>
> The usage of FP on DVB drivers is something that have been discussed for
> quite a long time at the community. I've asked Linus about the
> possibility of having a kernelspace driver using floating point. This
> may eventually happen (and, in fact, it already happens on a few
> drivers). There are, however, two points to consider:
>
> a) care should be taken to preserve FP state by the driver, since kernel
> won't do this by itself. This probably means that this may not be so
> fast as userspace calculation (although I suspect that it will be faster
> enough for the current needs);
>
> b) FP is processor dependent. So, old 80386 processors without FP
> instructions shouldn't work (IMO, this is not an issue ). Also, maybe we
> may have some troubles on a few non-x86 architectures (ARM ?).
>

As you know the userspace tuner framework is capable of solving this
and even more problems.
http://mcentral.de/wiki/index.php/Userspace_tuner

Although I'm waiting for an ack from several companies to reuse their
tuner sample drivers without doing much conversion (so far there's an
Ack for Quantek tuners)

In case of the userspace framework better see the em28xx driver as a
field test I won't submit an Ack for something else since I'm doing
ongoing support for newer silicons (which are not only limited to
Xceive based devices)

Markus



More information about the linux-dvb mailing list