On Thu, Sep 4, 2014 at 9:49 AM, Tom Rondeau <t...@trondeau.com> wrote: > > On Wed, Sep 3, 2014 at 5:53 PM, Philip Balister <phi...@balister.org> wrote: >> >> >> >> On 09/03/2014 04:30 PM, Nowlan, Sean wrote: >> > I'm trying to boot a ZedBoard with the GNU Radio image here: >> > http://gnuradio.org/data/sdk/zedboard_armv7a-hf-vfp-neon/. I picked the >> > hard >> > float image even though the tutorial points to the soft float image >> > (http://gnuradio.org/redmine/projects/gnuradio/wiki/Zynq). I don't know why >> > I'd want the soft float image since using the FPU will be much faster. >> > >> >> Hard float isn't that much faster. It referes to the ARM abi used for >> function calls. Hard float lets you return floats in NEON registers, >> which makes functions that return floats faster. Pretty much everything >> else is the same. So, it is a win, but not huge, except for certain cases. >> >> Philip > > > Nathan and I have both done some tests. You can get a few percent > improvement in speed with just hard float. That by itself was not a > compelling enough reason to drop the ORC support. It was only because of the > NEON VOLK proto-kernels that made the change worth it. > > I had some data that I was going to publish at one point, but it was too > confused with different development stages so that the comparisons wouldn't > be clean or clear. > > Tom > >
That's correct, in my VOLK tests the only time hardfloat made any difference was when the compiler inserted a call to its runtime library for generic code. I expect this will come up in the VOLK and/or embedded working group sessions at grcon. _______________________________________________ Discuss-gnuradio mailing list Discuss-gnuradio@gnu.org https://lists.gnu.org/mailman/listinfo/discuss-gnuradio
