Hey,
2013/9/24 Toby St Clere Smithe <m...@tsmithe.net>
> Karl Rupp <r...@iue.tuwien.ac.at> writes:
> > Very nice! :-)
>
> Thanks! :)
>
> > Are you sure about the lack of precision in NumPy? Do you get relative
> > errors in the 10^{-6} range? Please let me know if there are any issues
> > with slice and range (or any other operations) in whatever scenarios.
>
> Well, for instance, calculating the inner product with ViennaCL for two
> vectors of length 11 with random float values produces a value of
> 662.9756438... whilst NumPy produces a value of 666. I don't really know
> what's happening, but I've not yet investigated the code path in depth.
>
>
It seems like Satan is here! Have you been using a few lines of Java or
something? :D
Actually, just trying some things at the interpreter prompt suggests
> I've got my numbers the wrong way round, and PyViennaCL always produces
> a value of 666. That's almost certainly a bug in _viennacl.cpp, which
> should be fairly easy to fix. I'll investigate it later today.
>
> > I think we should use a similar model as e.g. for the KDE development
> > libraries, i.e. ViennaCL and pyViennaCL each have their own version
> > numbers. For consistency, I opt for starting with PyViennaCL 1.0.0,
> > increment the last digit whenever minor bug fixes (or similar) without
> > changing the API are introduced, and increment the digit in the center
> > whenever features are added.
>
> Yeah, that's probably fine. I can define pyviennacl.__version__ as
> 1.0.0, for instance, and pyviennacl._viennacl.__version__ as 1.5.0, with
> an alias for that at pyviennacl.__viennacl_version__. That way, people
> have access to the information they might be interested in.
>
> > Great, that looks like a pretty short list already. I'm slowly
> > recovering from the relocation pain and find some time for
> > science/coding again, so I can help you with platform tests and the like.
>
> That would be great. Debian and Ubuntu are quite straightforward,
> because they contain software for automatically building a Python
> extension for all the available and supported Python versions, which can
> then be packaged up fairly neatly. I just need to learn how to use
> `pybuild`!
>
> It's platforms like Windows and Mac OS X that are more difficult,
> because of the non-standard / non-Linux tool set.. I'll probably just
> write a script to get the version of the default Python interpreter; but
> that's not ideal if you want, as Debian does, to provide an extension
> for all Python versions that might exist on a system.
>
if this can reassure you, I think that a vast majority of the linux users
interested in PyViennaCL (mostly researchers in computational science) run
ubuntu-based distro (Ubuntu or Mint). I think that you are more likely to
meet users on Ubuntu 10.04 looking for a ready debian package than users on
gentoo who want to compile it by hand... therefore I would tend to think
that the build process is not that important, as long as you are able to
provide easily installable binaries (.deb, .exe) that work on a variety of
platforms (Ubuntu old versions, Windows 8, Windows 7, maybe Windows XP).
I would be pretty surprised if you find a Windows user who would want to
compile PyViennaCL by himself!
> > I think the most important bits and pieces for the first release are the
> > installation and the documentation. There is no existing user base yet,
> > so everybody *has* to go through this step. We don't want to lose any
> > users during these first steps, do we?
>
> So, right now, installation is fairly simple, if you don't mind doing
> the basic configuration that I described to Philippe. But even that much
> is probably too much -- hence the auto-determination. As for docs, they
> are done, and pretty much ready to host on-line. I want to wait for the
> final first release before doing that though, and with that will come a
> final proof-read etc.
>
> > char/short and outer products will be fixed within the next few days.
>
> Great.
>
> > You should get instructions for the upload via email directly from
> > Google during this week. So far this has always been a fairly painless
> > process, so no need to worry about it :-)
>
> Yep, looks like there's a form on Melange right now, but no obvious
> upload location. I'll wait a while and then check it out tomorrow / the
> day after. I'm currently at a friend's house on Cape Cod, with a fairly
> bad internet connection, so there's not much I can do immediately!
>
> > Hmm, this is indeed a lot. Do you consistently use vector_base and
> > matrix_base as types for functions in order to avoid an explosion of
> > instantiation for {vector, vector_range, vector_slice} and {matrix,
> > matrix_range, matrix_slice}?
>
> Yeah, I do. I think it comes from the expansion of quite a few macros,
> but there's not much I can do about that whilst keeping the code
> maintainable. So I'll just have to split _viennacl.cpp into some more
> files, and link the objects all together. It should be fine, but might
> be a bit much re-organisation for the first release, if we want to get
> the first release out fairly soon (ie, next week or so).
>
>
Yes, I think it is impossible to use templates in such context... Have you
tried to compile it using another compiler such as clang? I've heard it
often had a lighter memory footprint.
I'm starting a maths course in the next fortnight, so I'd rather get the
> formal work of the release done before then.
>
Good luck!
Philippe
>
> Best,
>
>
> Toby
>
>
>
> ------------------------------------------------------------------------------
> October Webinars: Code for Performance
> Free Intel webinars can help you accelerate application performance.
> Explore tips for MPI, OpenMP, advanced profiling, and more. Get the most
> from
> the latest Intel processors and coprocessors. See abstracts and register >
> http://pubads.g.doubleclick.net/gampad/clk?id=60133471&iu=/4140/ostg.clktrk
> _______________________________________________
> ViennaCL-devel mailing list
> ViennaCL-devel@lists.sourceforge.net
> https://lists.sourceforge.net/lists/listinfo/viennacl-devel
>
------------------------------------------------------------------------------
October Webinars: Code for Performance
Free Intel webinars can help you accelerate application performance.
Explore tips for MPI, OpenMP, advanced profiling, and more. Get the most from
the latest Intel processors and coprocessors. See abstracts and register >
http://pubads.g.doubleclick.net/gampad/clk?id=60133471&iu=/4140/ostg.clktrk
_______________________________________________
ViennaCL-devel mailing list
ViennaCL-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/viennacl-devel
