Re: [PD-dev] Pd-cvs Digest, Vol 92, Issue 2
... so if -ffast-math is already turned on in pd extended I think it's a fortiori safe to turn it on in vanilla :) Miller On Wed, Oct 03, 2012 at 08:28:38AM -0700, Miller Puckette wrote: > I have to think about thi one a little... nobody could ever test -fastmath > for all architectures. The danger I see is that some externals might > break. Maybe I should just leve it on during the 0.44 test phase and hope > I hear back if it's breaking things :) > > The reason for putting it in is that I stil get situations where Pd grinds > to a halt handling underflow interrupts - it's now a problem on the Pi. > > cheers > Miller > > On Wed, Oct 03, 2012 at 03:16:22PM +0100, Claude Heiland-Allen wrote: > > > > On 03/10/12 11:00, pd-cvs-requ...@iem.at wrote: > > > add -ffast-math flag to CC lines for linus and Mac > > > > Have you checked that this is safe on all architectures? > > > > IIRC, it optimizes with the assumption that everything is finite and > > not NaN, among other things. > > > > I know when I wrote 'tilde' (compiler from Pd dsp to C++), which > > incidentally used 'double' all the way through, I couldn't always > > use -ffast-math because it broke some patches very audibly. I > > didn't have time to debug the issue, so I just removed the flag > > globally. > > > > https://gitorious.org/maximus/tilde > > (currently unmaintained / dormant, but might still work) > > > > > > Claude > > -- > > http://mathr.co.uk > > ___ > Pd-dev mailing list > Pd-dev@iem.at > http://lists.puredata.info/listinfo/pd-dev ___ Pd-dev mailing list Pd-dev@iem.at http://lists.puredata.info/listinfo/pd-dev
Re: [PD-dev] Pd-cvs Digest, Vol 92, Issue 2
For what its worth, the past couple Pd-extended releases on Mac OS X have included the -fast flag, which does enable -ffast-math. I've never heard any complaints, but I've also never tested it closely. https://developer.apple.com/library/mac/#documentation/Darwin/Reference/ManPages/man1/gcc.1.html This would be a great use case for Katja's test framework in externals/testtools. Just write the tests using that framework, and I can set them to run on GNU/Linux i386 and amd64, and Mac OS X i386/x86_64/powerpc. Also, I have access to the Debian developer machines, so I could test them on any of those. .hc On 10/03/2012 11:28 AM, Miller Puckette wrote: > I have to think about thi one a little... nobody could ever test -fastmath > for all architectures. The danger I see is that some externals might > break. Maybe I should just leve it on during the 0.44 test phase and hope > I hear back if it's breaking things :) > > The reason for putting it in is that I stil get situations where Pd grinds > to a halt handling underflow interrupts - it's now a problem on the Pi. > > cheers > Millerhttps://developer.apple.com/library/mac/#documentation/Darwin/Reference/ManPages/man1/gcc.1.htm > > On Wed, Oct 03, 2012 at 03:16:22PM +0100, Claude Heiland-Allen wrote: >> On 03/10/12 11:00, pd-cvs-requ...@iem.at wrote: >>> add -ffast-math flag to CC lines for linus and Mac >> Have you checked that this is safe on all architectures? >> >> IIRC, it optimizes with the assumption that everything is finite and >> not NaN, among other things. >> >> I know when I wrote 'tilde' (compiler from Pd dsp to C++), which >> incidentally used 'double' all the way through, I couldn't always >> use -ffast-math because it broke some patches very audibly. I >> didn't have time to debug the issue, so I just removed the flag >> globally. >> >> https://gitorious.org/maximus/tilde >> (currently unmaintained / dormant, but might still work) >> >> >> Claude >> -- >> http://mathr.co.uk > ___ > Pd-dev mailing list > Pd-dev@iem.at > http://lists.puredata.info/listinfo/pd-dev ___ Pd-dev mailing list Pd-dev@iem.at http://lists.puredata.info/listinfo/pd-dev
Re: [PD-dev] Pd-cvs Digest, Vol 92, Issue 2
I have to think about thi one a little... nobody could ever test -fastmath for all architectures. The danger I see is that some externals might break. Maybe I should just leve it on during the 0.44 test phase and hope I hear back if it's breaking things :) The reason for putting it in is that I stil get situations where Pd grinds to a halt handling underflow interrupts - it's now a problem on the Pi. cheers Miller On Wed, Oct 03, 2012 at 03:16:22PM +0100, Claude Heiland-Allen wrote: > > On 03/10/12 11:00, pd-cvs-requ...@iem.at wrote: > > add -ffast-math flag to CC lines for linus and Mac > > Have you checked that this is safe on all architectures? > > IIRC, it optimizes with the assumption that everything is finite and > not NaN, among other things. > > I know when I wrote 'tilde' (compiler from Pd dsp to C++), which > incidentally used 'double' all the way through, I couldn't always > use -ffast-math because it broke some patches very audibly. I > didn't have time to debug the issue, so I just removed the flag > globally. > > https://gitorious.org/maximus/tilde > (currently unmaintained / dormant, but might still work) > > > Claude > -- > http://mathr.co.uk ___ Pd-dev mailing list Pd-dev@iem.at http://lists.puredata.info/listinfo/pd-dev
Re: [PD-dev] Pd-cvs Digest, Vol 92, Issue 2
On 03/10/12 11:00, pd-cvs-requ...@iem.at wrote: add -ffast-math flag to CC lines for linus and Mac Have you checked that this is safe on all architectures? IIRC, it optimizes with the assumption that everything is finite and not NaN, among other things. I know when I wrote 'tilde' (compiler from Pd dsp to C++), which incidentally used 'double' all the way through, I couldn't always use -ffast-math because it broke some patches very audibly. I didn't have time to debug the issue, so I just removed the flag globally. https://gitorious.org/maximus/tilde (currently unmaintained / dormant, but might still work) Claude -- http://mathr.co.uk ___ Pd-dev mailing list Pd-dev@iem.at http://lists.puredata.info/listinfo/pd-dev