[Numpy-discussion] Windows wheels, built, but should we deploy?

Discussion:

Matthew Brett

2016-03-04 04:42:42 UTC

Hi,

Summary:

I propose that we upload Windows wheels to pypi. The wheels are
likely to be stable and relatively easy to maintain, but will have
slower performance than other versions of numpy linked against faster
BLAS / LAPACK libraries.

Background:

There's a long discussion going on at issue github #5479 [1], where
the old problem of Windows wheels for numpy came up.

For those of you not following this issue, the current situation for
community-built numpy Windows binaries is dire:

* We have not so far provided windows wheels on pypi, so `pip install
numpy` on Windows will bring you a world of pain;
* Until recently we did provide .exe "superpack" installers on
sourceforge, but these became increasingly difficult to build and we
gave up building them as of the latest (1.10.4) release.

Despite this, popularity of Windows wheels on pypi is high. A few
weeks ago, Donald Stufft ran a query for the binary wheels most often
downloaded from pypi, for any platform [2] . The top five most
downloaded were (n_downloads, name):

6646, numpy-1.10.4-cp27-none-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl
5445, cryptography-1.2.1-cp27-none-win_amd64.whl
5243, matplotlib-1.4.0-cp34-none-win32.whl
5241, scikit_learn-0.15.1-cp34-none-win32.whl
4573, pandas-0.17.1-cp27-none-win_amd64.whl

So a) the OSX numpy wheel is very popular and b) despite the fact that
we don't provide a numpy wheel for Windows, matplotlib, sckit_learn
and pandas, that depend on numpy, are the 3rd, 4th and 5th most
downloaded wheels as of a few weeks ago.

So, there seems to be a large appetite for numpy wheels.

Current proposal:

I have now built numpy wheels, using the ATLAS blas / lapack library -
the build is automatic and reproducible [3].

I chose ATLAS to build against, rather than, say OpenBLAS, because
we've had some significant worries in the past about the reliability
of OpenBLAS, and I thought it better to err on the side of
correctness.

However, these builds are relatively slow for matrix multiply and
other linear algebra routines compared numpy built against OpenBLAS or
MKL (which we cannot use because of its license) [4]. In my very
crude array test of a dot product and matrix inversion, the ATLAS
wheels were 2-3 times slower than MKL. Other benchmarks on Julia
found about the same result for ATLAS vs OpenBLAS on 32-bit bit, but a
much bigger difference on 64-bit (for an earlier version of ATLAS than
we are currently using) [5].

So, our numpy wheels likely to be stable and give correct results, but
will be somewhat slow for linear algebra.

I propose that we upload these ATLAS wheels to pypi. The upside is
that this gives our Windows users a much better experience with pip,
and allows other developers to build Windows wheels that depend on
numpy. The downside is that these will not be optimized for
performance on modern processors. In order to signal that, I propose
adding the following text to the numpy pypi front page:

```
All numpy wheels distributed from pypi are BSD licensed.

Windows wheels are linked against the ATLAS BLAS / LAPACK library,
restricted to SSE2 instructions, so may not give optimal linear
algebra performance for your machine. See
http://docs.scipy.org/doc/numpy/user/install.html for alternatives.
```

In a way this is very similar to our previous situation, in that the
superpack installers also used ATLAS - in fact an older version of
ATLAS.

Once we are up and running with numpy wheels, we can consider whether
we should switch to other BLAS libraries, such as OpenBLAS or BLIS
(see [6]).

I'm posting here hoping for your feedback...

Cheers,

Matthew

[1] https://github.com/numpy/numpy/issues/5479
[2] https://gist.github.com/dstufft/1dda9a9f87ee7121e0ee
[3] https://ci.appveyor.com/project/matthew-brett/np-wheel-builder
[4] http://mingwpy.github.io/blas_lapack.html#intel-math-kernel-library
[5] https://github.com/numpy/numpy/issues/5479#issuecomment-185033668
[6] https://github.com/numpy/numpy/issues/7372

David Cournapeau

2016-03-04 08:29:27 UTC

Permalink

I would not worry too much about this: at worst, this gives us back the
situation where we were w/ so-called superpack, which have been successful
in the past to spread numpy use on windows.

My main worry is whether this locks us into ATLAS for a long time because
of package depending on numpy blas/lapack (scipy, scikit learn). I am not
sure how much this is the case.

David

Post by Matthew Brett
I propose that we upload these ATLAS wheels to pypi. The upside is
that this gives our Windows users a much better experience with pip,
and allows other developers to build Windows wheels that depend on
numpy. The downside is that these will not be optimized for
performance on modern processors. In order to signal that, I propose
```
All numpy wheels distributed from pypi are BSD licensed.
Windows wheels are linked against the ATLAS BLAS / LAPACK library,
restricted to SSE2 instructions, so may not give optimal linear
algebra performance for your machine. See
http://docs.scipy.org/doc/numpy/user/install.html for alternatives.
```
In a way this is very similar to our previous situation, in that the
superpack installers also used ATLAS - in fact an older version of
ATLAS.
Once we are up and running with numpy wheels, we can consider whether
we should switch to other BLAS libraries, such as OpenBLAS or BLIS
(see [6]).
I'm posting here hoping for your feedback...
Cheers,
Matthew
[1] https://github.com/numpy/numpy/issues/5479
[2] https://gist.github.com/dstufft/1dda9a9f87ee7121e0ee
[3] https://ci.appveyor.com/project/matthew-brett/np-wheel-builder
[4] http://mingwpy.github.io/blas_lapack.html#intel-math-kernel-library
[5] https://github.com/numpy/numpy/issues/5479#issuecomment-185033668
[6] https://github.com/numpy/numpy/issues/7372
_______________________________________________
NumPy-Discussion mailing list
https://mail.scipy.org/mailman/listinfo/numpy-discussion

Matthew Brett

2016-03-04 18:38:22 UTC

Permalink

Post by David Cournapeau

I would not worry too much about this: at worst, this gives us back the
situation where we were w/ so-called superpack, which have been successful
in the past to spread numpy use on windows.
My main worry is whether this locks us into ATLAS for a long time because
of package depending on numpy blas/lapack (scipy, scikit learn). I am not
sure how much this is the case.

You mean the situation where other packages try to find the BLAS /
LAPACK library and link against that? My impression was that neither
scipy or scikit-learn do that at the moment, but I'm happy to be
corrected.

You'd know better than me about this, but my understanding is that
BLAS / LAPACK has a standard interface that should allow code to run
the same way, regardless of which BLAS / LAPACK library it is linking
to. So, even if another package is trying to link against the numpy
BLAS, swapping the numpy BLAS library shouldn't cause a problem
(unless the package is trying to link to ATLAS-specific stuff, which
seems a bit unlikely).

Is that right?

Cheers,

Matthew

j***@gmail.com

2016-03-05 03:30:31 UTC

Permalink

Post by Matthew Brett

Post by David Cournapeau

numpy-1.10.4-cp27-none-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl

Post by David Cournapeau

Post by Matthew Brett
5445, cryptography-1.2.1-cp27-none-win_amd64.whl
5243, matplotlib-1.4.0-cp34-none-win32.whl
5241, scikit_learn-0.15.1-cp34-none-win32.whl
4573, pandas-0.17.1-cp27-none-win_amd64.whl
So a) the OSX numpy wheel is very popular and b) despite the fact that
we don't provide a numpy wheel for Windows, matplotlib, sckit_learn
and pandas, that depend on numpy, are the 3rd, 4th and 5th most
downloaded wheels as of a few weeks ago.
So, there seems to be a large appetite for numpy wheels.
I have now built numpy wheels, using the ATLAS blas / lapack library -
the build is automatic and reproducible [3].
I chose ATLAS to build against, rather than, say OpenBLAS, because
we've had some significant worries in the past about the reliability
of OpenBLAS, and I thought it better to err on the side of
correctness.
However, these builds are relatively slow for matrix multiply and
other linear algebra routines compared numpy built against OpenBLAS or
MKL (which we cannot use because of its license) [4]. In my very
crude array test of a dot product and matrix inversion, the ATLAS
wheels were 2-3 times slower than MKL. Other benchmarks on Julia
found about the same result for ATLAS vs OpenBLAS on 32-bit bit, but a
much bigger difference on 64-bit (for an earlier version of ATLAS than
we are currently using) [5].
So, our numpy wheels likely to be stable and give correct results, but
will be somewhat slow for linear algebra.

I would not worry too much about this: at worst, this gives us back the
situation where we were w/ so-called superpack, which have been

successful

Post by David Cournapeau
in the past to spread numpy use on windows.
My main worry is whether this locks us into ATLAS for a long time

because

Post by David Cournapeau
of package depending on numpy blas/lapack (scipy, scikit learn). I am not
sure how much this is the case.

You mean the situation where other packages try to find the BLAS /
LAPACK library and link against that? My impression was that neither
scipy or scikit-learn do that at the moment, but I'm happy to be
corrected.
You'd know better than me about this, but my understanding is that
BLAS / LAPACK has a standard interface that should allow code to run
the same way, regardless of which BLAS / LAPACK library it is linking
to. So, even if another package is trying to link against the numpy
BLAS, swapping the numpy BLAS library shouldn't cause a problem
(unless the package is trying to link to ATLAS-specific stuff, which
seems a bit unlikely).
Is that right?

AFAIK, numpy doesn't provide access to BLAS/LAPACK. scipy does. statsmodels
is linking to the installed BLAS/LAPACK in cython code through scipy. So
far we haven't seen problems with different versions. I think scipy
development works very well to isolate linalg library version specific
parts from the user interface.

AFAIU, The main problem will be linking to inconsistent Fortran libraries
in downstream packages that use Fortran.
Eg. AFAIU it won't work to pip install a ATLAS based numpy and then install
a MKL based scipy from Gohlke.

I don't know if there is a useful error message, or if this just results in
puzzled users.

Josef

Post by Matthew Brett
Cheers,
Matthew
_______________________________________________
NumPy-Discussion mailing list
https://mail.scipy.org/mailman/listinfo/numpy-discussion

Nathaniel Smith

2016-03-05 04:40:29 UTC

Permalink

On Fri, Mar 4, 2016 at 7:30 PM, <***@gmail.com> wrote:
[...]

Post by j***@gmail.com
AFAIK, numpy doesn't provide access to BLAS/LAPACK. scipy does. statsmodels
is linking to the installed BLAS/LAPACK in cython code through scipy. So far
we haven't seen problems with different versions. I think scipy development
works very well to isolate linalg library version specific parts from the
user interface.

Yeah, it should be invisible to users of both numpy and scipy which
BLAS/LAPACK is in use under the hood.

Post by j***@gmail.com
AFAIU, The main problem will be linking to inconsistent Fortran libraries in
downstream packages that use Fortran.
Eg. AFAIU it won't work to pip install a ATLAS based numpy and then install
a MKL based scipy from Gohlke.

The specific scenario you describe will be a problem, but not for the
reason you state -- the problem is that (IIUC) the Gohlke scipy build
has some specific hacks where it "knows" that it can find a copy of
MKL buried at a particular location inside the numpy package (and the
Gohlke numpy build has a specific hack to put a copy of MKL there). So
the Gohlke scipy requires the Gohlke numpy, but this is due to patches
that Christoph applies to his builds.

AFAIK, outside of downstream packages that poke around the inside of
numpy like this, there should be no way for downstream packages to
know or care which BLAS/LAPACK implementation numpy is using (except
for speed, bugs, etc.).

-n

--
Nathaniel J. Smith -- https://vorpus.org

Matthew Brett

2016-03-05 18:40:37 UTC

Permalink

Hi,

Post by Nathaniel Smith
[...]

Yeah, it should be invisible to users of both numpy and scipy which
BLAS/LAPACK is in use under the hood.

Carl Kleffner

2016-03-06 14:52:31 UTC

Permalink

+1 from me. I could prepare scipy builds based on these numpy builds.

Carl

Post by j***@gmail.com
Hi,

Post by Nathaniel Smith
[...]

Post by j***@gmail.com
AFAIK, numpy doesn't provide access to BLAS/LAPACK. scipy does.

statsmodels

Post by Nathaniel Smith

Post by j***@gmail.com
is linking to the installed BLAS/LAPACK in cython code through scipy.

So far

Post by Nathaniel Smith

Post by j***@gmail.com
we haven't seen problems with different versions. I think scipy

development

Post by Nathaniel Smith

Post by j***@gmail.com
works very well to isolate linalg library version specific parts from

the

Post by Nathaniel Smith

Post by j***@gmail.com
user interface.

Yeah, it should be invisible to users of both numpy and scipy which
BLAS/LAPACK is in use under the hood.

My impression is that the general mood here is positive, so I plan to
deploy these wheels to pypi on Monday, with the change to the pypi
text. Please do let me know if there are any strong objections.
Cheers,
Matthew
_______________________________________________
NumPy-Discussion mailing list
https://mail.scipy.org/mailman/listinfo/numpy-discussion

Matthew Brett

2016-03-07 20:29:45 UTC

Permalink

Post by Matthew Brett
Hi,

Post by Nathaniel Smith
[...]

Yeah, it should be invisible to users of both numpy and scipy which
BLAS/LAPACK is in use under the hood.

Done:

(py35) PS C:\tmp> pip install numpy
Collecting numpy
Downloading numpy-1.10.4-cp35-none-win32.whl (6.6MB)
100% |################################| 6.6MB 34kB/s
Installing collected packages: numpy

Cheers,

Matthew

Olivier Grisel

2016-03-08 21:11:33 UTC

Permalink

Thanks Matthew! I just installed it and ran the tests and it all works
(except for test_system_info.py that fails because I am missing a
vcvarsall.bat on that system but this is expected).

--
Olivier

Chris Barker

2016-03-04 20:29:52 UTC

Permalink

+1 -- thanks for doing all this work.

There is a HUGE amount you can do with numpy that doesn't give a whit about
how fast .dot() et all are. If you really do need that to be fast as
possible, you can pug in a faster build later.

This is great.

Just as one example -- I teach a general python class every year --I do
only one session on numpy/scipy. If I can expect my students to be able to
simply pip install the core scipy stack, this will be SO much easier.

-CHB

Post by Matthew Brett
Hi,
I propose that we upload Windows wheels to pypi. The wheels are
likely to be stable and relatively easy to maintain, but will have
slower performance than other versions of numpy linked against faster
BLAS / LAPACK libraries.
There's a long discussion going on at issue github #5479 [1], where
the old problem of Windows wheels for numpy came up.
For those of you not following this issue, the current situation for
* We have not so far provided windows wheels on pypi, so `pip install
numpy` on Windows will bring you a world of pain;
* Until recently we did provide .exe "superpack" installers on
sourceforge, but these became increasingly difficult to build and we
gave up building them as of the latest (1.10.4) release.
Despite this, popularity of Windows wheels on pypi is high. A few
weeks ago, Donald Stufft ran a query for the binary wheels most often
downloaded from pypi, for any platform [2] . The top five most
6646,
numpy-1.10.4-cp27-none-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl
5445, cryptography-1.2.1-cp27-none-win_amd64.whl
5243, matplotlib-1.4.0-cp34-none-win32.whl
5241, scikit_learn-0.15.1-cp34-none-win32.whl
4573, pandas-0.17.1-cp27-none-win_amd64.whl
So a) the OSX numpy wheel is very popular and b) despite the fact that
we don't provide a numpy wheel for Windows, matplotlib, sckit_learn
and pandas, that depend on numpy, are the 3rd, 4th and 5th most
downloaded wheels as of a few weeks ago.
So, there seems to be a large appetite for numpy wheels.
I have now built numpy wheels, using the ATLAS blas / lapack library -
the build is automatic and reproducible [3].
I chose ATLAS to build against, rather than, say OpenBLAS, because
we've had some significant worries in the past about the reliability
of OpenBLAS, and I thought it better to err on the side of
correctness.
However, these builds are relatively slow for matrix multiply and
other linear algebra routines compared numpy built against OpenBLAS or
MKL (which we cannot use because of its license) [4]. In my very
crude array test of a dot product and matrix inversion, the ATLAS
wheels were 2-3 times slower than MKL. Other benchmarks on Julia
found about the same result for ATLAS vs OpenBLAS on 32-bit bit, but a
much bigger difference on 64-bit (for an earlier version of ATLAS than
we are currently using) [5].
So, our numpy wheels likely to be stable and give correct results, but
will be somewhat slow for linear algebra.
I propose that we upload these ATLAS wheels to pypi. The upside is
that this gives our Windows users a much better experience with pip,
and allows other developers to build Windows wheels that depend on
numpy. The downside is that these will not be optimized for
performance on modern processors. In order to signal that, I propose
```
All numpy wheels distributed from pypi are BSD licensed.
Windows wheels are linked against the ATLAS BLAS / LAPACK library,
restricted to SSE2 instructions, so may not give optimal linear
algebra performance for your machine. See
http://docs.scipy.org/doc/numpy/user/install.html for alternatives.
```
In a way this is very similar to our previous situation, in that the
superpack installers also used ATLAS - in fact an older version of
ATLAS.
Once we are up and running with numpy wheels, we can consider whether
we should switch to other BLAS libraries, such as OpenBLAS or BLIS
(see [6]).
I'm posting here hoping for your feedback...
Cheers,
Matthew
[1] https://github.com/numpy/numpy/issues/5479
[2] https://gist.github.com/dstufft/1dda9a9f87ee7121e0ee
[3] https://ci.appveyor.com/project/matthew-brett/np-wheel-builder
[4] http://mingwpy.github.io/blas_lapack.html#intel-math-kernel-library
[5] https://github.com/numpy/numpy/issues/5479#issuecomment-185033668
[6] https://github.com/numpy/numpy/issues/7372
_______________________________________________
NumPy-Discussion mailing list
https://mail.scipy.org/mailman/listinfo/numpy-discussion

--
Christopher Barker, Ph.D.
Oceanographer

Emergency Response Division
NOAA/NOS/OR&R (206) 526-6959 voice
7600 Sand Point Way NE (206) 526-6329 fax
Seattle, WA 98115 (206) 526-6317 main reception

***@noaa.gov