Re: [Numpy-discussion] Building numpy with ATLAS

[Shitikanth Kashyap]

As it turns out, I just had to run "python setup.py build --force" after 
changing my site.cfg file in order to recompile numpy/core/multiarray.so.
___
NumPy-Discussion mailing list
NumPy-Discussion@scipy.org
https://mail.scipy.org/mailman/listinfo/numpy-discussion

Re: [Numpy-discussion] isnan() equivalent for np.NaT?

[Stephan Hoyer]

Agreed -- this would be really nice to have. For now, the best you can do
is something like the following:


def is_na(x):
x = np.asarray(x)
if np.issubdtype(x.dtype, (np.datetime64, np.timedelta64)):  # ugh
int_min = np.iinfo(np.int64).min
return x.view('int64') == int_min
else:
return np.isnan(x)


On Mon, Jul 18, 2016 at 3:39 PM, Gerrit Holl  wrote:

> On 18 July 2016 at 22:20, Scott Sanderson 
> wrote:
> > I'm working on upgrading Zipline (github.com/quantopian/zipline) to the
> > latest numpy, and I'm getting a FutureWarnings about the upcoming change
> in
> > the behavior of comparisons on np.NaT.  I'd like to be able to do checks
> for
> > NaT in a way that's forwards-compatible, but I couldn't find a function
> > analogous to `np.isnan` for NaT.  Am I missing something that already
> > exists?  If not, is there interest in such a function? I'd like to be
> able
> > to pre-emptively fix the warnings in Zipline so that we're ready when the
> > change actually happens, but it doesn't seem like the necessary tools are
> > available right now.
>
> Hi Scott,
>
> see https://github.com/numpy/numpy/issues/5610
>
> Gerrit.
> ___
> NumPy-Discussion mailing list
> NumPy-Discussion@scipy.org
> https://mail.scipy.org/mailman/listinfo/numpy-discussion
>
___
NumPy-Discussion mailing list
NumPy-Discussion@scipy.org
https://mail.scipy.org/mailman/listinfo/numpy-discussion

Re: [Numpy-discussion] isnan() equivalent for np.NaT?

[Gerrit Holl]

On 18 July 2016 at 22:20, Scott Sanderson  wrote:
> I'm working on upgrading Zipline (github.com/quantopian/zipline) to the
> latest numpy, and I'm getting a FutureWarnings about the upcoming change in
> the behavior of comparisons on np.NaT.  I'd like to be able to do checks for
> NaT in a way that's forwards-compatible, but I couldn't find a function
> analogous to `np.isnan` for NaT.  Am I missing something that already
> exists?  If not, is there interest in such a function? I'd like to be able
> to pre-emptively fix the warnings in Zipline so that we're ready when the
> change actually happens, but it doesn't seem like the necessary tools are
> available right now.

Hi Scott,

see https://github.com/numpy/numpy/issues/5610

Gerrit.
___
NumPy-Discussion mailing list
NumPy-Discussion@scipy.org
https://mail.scipy.org/mailman/listinfo/numpy-discussion

[Numpy-discussion] isnan() equivalent for np.NaT?

[Scott Sanderson]

Hi All,

I'm working on upgrading Zipline (github.com/quantopian/zipline) to the
latest numpy, and I'm getting a FutureWarnings about the upcoming change in
the behavior of comparisons on np.NaT.  I'd like to be able to do checks
for NaT in a way that's forwards-compatible, but I couldn't find a function
analogous to `np.isnan` for NaT.  Am I missing something that already
exists?  If not, is there interest in such a function? I'd like to be able
to pre-emptively fix the warnings in Zipline so that we're ready when the
change actually happens, but it doesn't seem like the necessary tools are
available right now.

-Scott
___
NumPy-Discussion mailing list
NumPy-Discussion@scipy.org
https://mail.scipy.org/mailman/listinfo/numpy-discussion

Re: [Numpy-discussion] deterministic, reproducible matmul / __matmult_

[Jason Newton]

On Mon, Jul 11, 2016 at 3:27 PM, Pauli Virtanen  wrote:
> Mon, 11 Jul 2016 13:01:49 -0400, Jason Newton kirjoitti:
>> Does the ML have any ideas on how one could get a matmul that will not
>> allow any funny business on the evaluation of the products?  Funny
>> business here is something like changing the evaluation order additions
>> of terms. I want strict IEEE 754 compliance - no 80 bit registers, and
>> perhaps control of the rounding mode, no unsafe math optimizations.
>
> If you link Numpy with a BLAS and LAPACK libraries that have been
> compiled for this purpose, and turn on the compiler flags that enforce
> strict IEEE (and disable SSE) when compiling Numpy, you probably will get
> reproducible builds. Numpy itself just offloads the dot computations to
> BLAS, so if your BLAS is reproducible, things should mainly be OK.

The matrix multiplication of the reference blas hard to follow, so
that's good.  Generalized Inverse is a little more difficult.  I've
actually had problems building numpy w/ ref blas under windows...,
anyone know where and how to take a ready made cblas dll and get an
existing numpy installation (e.g. anaconda) to use it?

>
> You may also need to turn off the SSE optimizations in Numpy, because
> these can make results depend on memory alignment --- not in dot
> products, but in other computations.
>
> Out of curiosity, what is the application where this is necessary?
> Maybe there is a numerically stable formulation?

This can come up in recursive processes or anywhere where parallelism
(vectorization or other kinds) and the need for reproducable code
exists (there are many reasons you'd want this, see slides below).  As
I mentioned, I desire to have reproducable calculations when involving
alternative implementations with things like c modules, MPI, FPGAs,
GPUs coming into the picture.  You  can usually figure out a strategy
to do this when you write everything yourself, but I'd love to write
things in numpy and have it just choose simple / straight forward
implementations that are usually what everything boils down to onto
these other devices, at least in meaningful peaces.  There may be
other times where it gets more complicated than what i mention here,
but it is still very useful as a building block for those cases (which
are often multiple accumulators/partitioning, tree like reduction
ordering).

I looked into reproblas further and also I asked the authors of
repoblas to add a cblas wrapper in the hopes I might sometime have
numpy working on top of it.  I read alot of their research recently
and it's pretty cool - they get better accuracy than most
implementations and the performance is pretty good.  Check out slides
here http://people.eecs.berkeley.edu/~hdnguyen/public/talks/ARITH21_Fast_Sum.pdf
(skip to slide 24 for accuracy) - the takeaway here is you actually do
quite a bit better in precision/accuracy with their sum method, than
the naive and alternative implementations.  The cpu performance is
worth the trade and really not bad at all for most operations and
purposes - especially since they hand scalability very well.

They also just posted a monster of a technical report here
https://www.eecs.berkeley.edu/Pubs/TechRpts/2016/EECS-2016-121.html -
this details recent performance using avx and sse, if anyone is
interested.

I'd love to have a library like this that I could just tell numpy to
switch out to at runtime - at the start of a script.

-Jason
___
NumPy-Discussion mailing list
NumPy-Discussion@scipy.org
https://mail.scipy.org/mailman/listinfo/numpy-discussion

[Numpy-discussion] Building numpy with ATLAS

[Shitikanth Kashyap]

Hey Everyone,

I am trying to build numpy-1.11.1 with ATLAS libraries. The build completes 
successfully, and 
I am able to use numpy, but there seems to be some problem with my config 
because 
numpy only ever uses one core, even when multiplying 5000x5000 matrices with 
numpy.dot.

My site.cfg is:

[default]
library_dirs = 
/u3/s4kashya/.local/atlas3.11/lib:/u3/s4kashya/.local/lib:/usr/local/lib:/usr/lib
include_dirs = 
/u3/s4kashya/.local/atlas3.11/include:/u3/s4kashya/.local/include:/usr/local/include:/usr/include

[atlas]
libraries = ptlapack, ptf77blas, ptcblas, atlas
library_dirs = /u3/s4kashya/.local/atlas3.11/lib
include_dirs = /u3/s4kashya/.local/atlas3.11/include


Output from numpy.__config__.show():

atlas_3_10_blas_threads_info:
   libraries = ['tatlas', 'ptlapack', 'ptf77blas', 'ptcblas', 'atlas']
   library_dirs = ['/u3/s4kashya/.local/atlas3.11/lib']
   define_macros = [('HAVE_CBLAS', None), ('NO_ATLAS_INFO', -1)]
   language = c
   include_dirs = ['/u3/s4kashya/.local/atlas3.11/include']
lapack_opt_info:
   libraries = ['tatlas', 'tatlas', 'tatlas', 'ptlapack', 'ptf77blas', 
'ptcblas', 'atlas']
   library_dirs = ['/u3/s4kashya/.local/atlas3.11/lib']
   define_macros = [('NO_ATLAS_INFO', -1)]
   language = f77
   include_dirs = ['/u3/s4kashya/.local/atlas3.11/include']
blas_opt_info:
   libraries = ['tatlas', 'ptlapack', 'ptf77blas', 'ptcblas', 'atlas']
   library_dirs = ['/u3/s4kashya/.local/atlas3.11/lib']
   define_macros = [('HAVE_CBLAS', None), ('NO_ATLAS_INFO', -1)]
   language = c
   include_dirs = ['/u3/s4kashya/.local/atlas3.11/include']
openblas_info:
 NOT AVAILABLE
openblas_lapack_info:
 NOT AVAILABLE
atlas_3_10_threads_info:
   libraries = ['tatlas', 'tatlas', 'tatlas', 'ptlapack', 'ptf77blas', 
'ptcblas', 'atlas']
   library_dirs = ['/u3/s4kashya/.local/atlas3.11/lib']
   define_macros = [('NO_ATLAS_INFO', -1)]
   language = f77
   include_dirs = ['/u3/s4kashya/.local/atlas3.11/include']
lapack_mkl_info:
 NOT AVAILABLE
blas_mkl_info:
 NOT AVAILABLE
mkl_info:
 NOT AVAILABLE

Any ideas what could be going wrong?

Thanks,
Shitikanth
___
NumPy-Discussion mailing list
NumPy-Discussion@scipy.org
https://mail.scipy.org/mailman/listinfo/numpy-discussion