On Wednesday, 27 de October de 2010 09:51:44 sakari.pou...@nokia.com wrote: > My take on this as a Nokia MeeGo architect is that we need hardfp. That is > where the market is going: Harmattan and linaro are already in that train. > IMO, we should start building softfp and hardfp for a 6 months grace > period. Once 1.2 is out we should drop softfp.
I've received a new report from our Technology Management people this morning.
The recommendation, in summary, is to use:
1) single-precision floating point (float, not double)
2) ARM's RunFast mode (i.e., not fully IEEE754 compliant)
3) hardfp argument passing
4) disable use of short vector mode
On the Cortex-A8, single precision FP with RunFast mode executes on the
pipelined Neon core (even if you don't use Neon instructions), which results
in single-cycle operations for multiply-accumulate. Anything else will execute
in the VFPlite core, which is not pipelined and will impose delays due to lack
of pipelining and branch misprediction.
Some applications need IEEE754 compliancy, so turning this back on should be
possible. But the OS should start apps in RunFast mode by default.
I'm not sure I understand what the "short vector mode" is. If the compiler is
generating such a code, it needs to stop. I've passed this recommendation to
the Linaro toolchain team already.
I'm not sure how things change if the processor doesn't have Neon (like the
Tegra II).
--
Thiago Macieira - thiago (AT) macieira.info - thiago (AT) kde.org
Senior Product Manager - Nokia, Qt Development Frameworks
PGP/GPG: 0x6EF45358; fingerprint:
E067 918B B660 DBD1 105C 966C 33F5 F005 6EF4 5358
signature.asc
Description: This is a digitally signed message part.
_______________________________________________ MeeGo-dev mailing list MeeGo-dev@meego.com http://lists.meego.com/listinfo/meego-dev
