Hi,
there is now a very experimental FP emulator available for Arm Linux. It is
approximately 4-8 times faster than NWFPE, but has also some disadvantages:
- only max. 32Bit precision for mantissa (ok for float, but insufficient for
double)
- no exceptions
- no overflow detecion and no support for unnormalized numbers
So it is probably not usable for scientific calculations. I have been testing
it for many hours now, and found no major problems (except the low precision).
But it caused a 4 times speed improvement for Gimp an xfractint.
I think I will increase its precision to 64bit in a next release, but before I
do that I would like to know if there are people interested in this, and if
there are still undetected bugs.
If you want to try it, you should first keep in mind that it could do harm to
your system and that you have been warned and that I will not take any responsibility
for loss of data or what know I. You should not test it on a system that contains
important data. With this in your mind you should get the kernel patch from:
http://rcswww.urz.tu-dresden.de/~teich-p/download/fastfpe.patch
You should apply the patch, then do a "make menuconfig", disable NWFPE and
choose FASTFPE. Then recompile the kernel and reboot the system with the new
kernel. I have tested it with kernels 2.2.14-rmk6 and 2.2.16-rmk2 and a
Debian 2.2 installation on an Acorn RiscPC with Strongarm processor.
I would like to get your bug reports and ideas for improvements.
Peter Teichmann
--
Email: [EMAIL PROTECTED] WWW: rcswww.urz.tu-dresden.de/~teich-p
_______________________________________________
http://lists.arm.linux.org.uk/mailman/listinfo/linux-arm
