[Numpy-discussion] Tensor Contraction (HPTT) and Tensor Transposition (TCL)
Charles R Harris
charlesr.harris at gmail.com
Thu Aug 17 12:58:33 EDT 2017
On Thu, Aug 17, 2017 at 10:15 AM, Chris Barker <chris.barker at noaa.gov>
> On Thu, Aug 17, 2017 at 12:55 AM, Sebastian Berg <
> sebastian at sipsolutions.net> wrote:
>> > How would the process look like if NumPY is distributed as a
>> > precompiled binary?
>> Well, numpy is BSD, and the official binaries will be BSD, someone else
>> could do less free binaries of course.
> Indeed, if you want it to be distributed as a binary with numpy, then the
> license needs to be compatible -- do you have a substantial objection to
> BSD? The BSD family is pretty much the standard for Python -- Python (and
> numpy) are very broadly used in proprietary software.
> I doubt we can have a hard
>> dependency unless it is part of the numpy source
> and no reason to -- if it is a hard dependency, it HAS to be compatible
> licensed, and it's a lot easier to keep the source together.
> However, it _could_ be a soft dependency, like LAPACK/BLAS -- I've
> honestly lost track, but numpy used come with a lapack-lite (or some such),
> so that it could be compiled and work with no external LAPACK
> implementation -- you wouldn't get the best performance, but it would work.
> I doubt including the source
>> itself is going to happen quickly since we would first have to decide
>> to actually use a modern C++ compiler (I have no idea if that is
>> problematic or not).
> could it be there as a conditional compilation? There is a lot of push to
> support C++11 elsewhere, so a compiled-with-a-modern-compiler numpy is
> not SO far off..
> (for py3 anyway...)
It would take a fair amount of grunge work to get there. Variables would
need renaming, for instance `new`, and other such things. Nothing mind
bending, but not completely trivial either.
> * Use TCL if you need faster einsum(like) operations
> That is, of course, the other option -- distribute it on its own or maybe
> in scipy, and then users can use it as an optimization for those few core
> functions where speed matters to them -- honestly, it's a pretty small
> fraction of numpy code.
> But it sure would be nice if it could be built in, and then folks would
> get better performance without even thinkning about it.
>> Just a few thoughts, did not think about details really. But yes, it is
>> sounds reasonable to me to re-add support for optional dependencies
>> such as fftw or your TCL. But packagers have to make use of that or I
>> fear it is actually less available than a standalone python module.
> true -- though I expect Anaconda / conda forge at least would be likely to
> pick it up if it works well.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the NumPy-Discussion