ENH: Cython interface for special #6195

person142 · 2016-05-28T22:50:06Z

This defines typed, scalar versions of many of the functions in
scipy.special and makes them available to cimport in Cython. It uses
fused types so that the functions have the same names and signatures
as their ufunc counterparts.

It piggybacks off of the existing code generation in
generate_ufuncs.py. In particular:

The process for adding new ufuncs is unchanged
When a new ufunc is added a Cython version will automatically be
created (subject to some restrictions mentioned at the end), and the
documentation and tests will be automatically updated.

This does not attempt to add Cython versions of:

~~Ufuncs with multiple return arguments~~ Ufuncs with multiple return arguments are now supported
Ufuncs that have additional layers of python wrappers.

With more work these restrictions could be removed.

Closes gh-5964.

rgommers · 2016-06-05T10:01:01Z

This looks quite promising @person142!

There are 2 test failures for the latest commit, those are real.

rgommers · 2016-06-05T12:25:57Z

Benchmarks are quite consistent; numpy and cython same speed (except ~400ns less overhead for small arrays with cython), Python significantly slower:

[ 57.14%] ··· Running cython_special.CythonSpecial.time_gamma_d                                      ok
[ 57.14%] ···· 
               ========== ====== ========== ========== ==========
               --                              api               
               ----------------- --------------------------------
                argument    N      python     numpy      cython  
               ========== ====== ========== ========== ==========
                 (100,)     10    10.61μs     2.11μs     1.59μs  
                 (100,)    100    95.18μs    12.64μs    12.19μs  
                 (100,)    1000   890.96μs   122.49μs   117.66μs 
               ========== ====== ========== ========== ==========

And autogenerating benchmarks is a nice touch.

person142 · 2016-06-05T14:15:35Z

Yes, I need to track those failures down. I ran into an issue before that became gh-6208; might be more of the same.

I have noticed one circumstance where cython is slower than numpy, and that's functions with a lot of complex arguments and an underlying c function that's using numpy cdoubles (e.g. the airy benchmark). I guess the extra time is coming from all the casting back and forth between double complex and cdouble.

pv · 2016-06-05T14:19:44Z

It would be surprising if the time difference comes from casting, since that should not be very many cpu cycles, compared to the floating point math the function in the end does.

person142 · 2016-06-05T22:29:23Z

I looked at the generated cython wrappers for pro_rad2 and pro_rad2_cv and didn't see anything suspicious, and while gh-6211 doesn't specifically address the test failures it makes me suspect the underlying issue is in specfun.f. (Uninitialized memory getting accessed and making the outputs nondeterministic.)

I marked the radial spheroidal wave functions as known failures; if people are not happy with this I'll dig deeper into the Fortran code and try to pinpoint the error.

rgommers · 2016-06-12T07:14:05Z

@person142 it would be good to advertise this PR on the mailing list. I'm sure some people will be interested.

rgommers · 2016-06-12T07:15:04Z

I marked the radial spheroidal wave functions as known failures; if people are not happy with this I'll dig deeper into the Fortran code and try to pinpoint the error.

Addressing that separately is ok I think.

pv · 2016-06-12T20:12:36Z

scipy/special/generate_ufuncs.py

+Cython API for Special Functions
+================================
+
+Scalar, typed versions of many of the functions in ``scipy.special``


It could be simpler to just list all the available C signatures, rather than the internal representation used in generate.py --- this way it would be more self-explanatory.

person142 · 2016-06-14T00:32:11Z

@pv good point, changed it:

https://circle-artifacts.com/gh/scipy/scipy/174/artifacts/0/home/ubuntu/scipy/doc/build/html-scipyorg/special.cython_special.html

ev-br · 2016-06-17T21:48:55Z

While I won't be able to review the details, I played a bit with this from a user perspective and it looks great. Don't know if this makes this release or the next one, but this definitely should be a highlight of a release.

One longer-term question, not for this PR. Would it be possible to add some magic to dispatch to these functions automatically when a special function is called with scalar arguments?

pv · 2016-06-18T12:09:47Z

scipy/special/specfun_wrappers.c

@@ -77,7 +77,7 @@ npy_cdouble clngamma_wrap( npy_cdouble z) {

 npy_cdouble chyp2f1_wrap( double a, double b, double c, npy_cdouble z) {
  npy_cdouble outz;
-  int l1, l0;
+  int l1, l0, isfer;


Initialize isfer = 0 --- since it's not set by the fortran routine always, the result is undefined.

pv · 2016-06-18T13:15:36Z

Ok, a few comments:

I think autogeneration for tests and benchmarks is a bit too far --- I'd write these manually.
The test_cython_special api tests don't seem to pick the function with correct signatures (as noted above, the _float et al map to Python floats, so you always check a single specialization).

I have fixes for the above in my branch, cf person142/scipy@cython-interface...pv:pr-6195

pv · 2016-06-18T13:17:51Z

Apart from that, looks great to me.

person142 · 2016-06-18T15:20:57Z

Thanks for the fixes @pv. I moved the known failure check to earlier in the test so that those functions never get evaluated. I've had pbwa cause a segfault before.

pv · 2016-06-18T16:05:04Z

Other stuff: cython_special.pxd doesn't get installed, missing from setup.py

person142 · 2016-06-18T17:52:15Z

I think I fixed it.

This defines typed, scalar versions of many of the functions in scipy.special and makes them available to cimport in Cython. It uses fused types so that the functions have the same names and signatures as their ufunc counterparts. It piggybacks off of the existing code generation in generate_ufuncs.py. In particular: - The process for adding new ufuncs is unchanged - When a new ufunc is added a Cython version will automatically be created (subject to some restrictions mentioned at the end), and the documentation and tests will be automatically updated. This does not attempt to add Cython versions of: - Ufuncs with multiple return arguments - Ufuncs that have additional layers of python wrappers. With more work these restrictions could be removed. Closes scipygh-5964.

This mimics the behavior of iter_varaints in generate_ufuncs.py

Hopefully this will fix issues with conversions.

Modifies specfun.f to return sf_error codes in added flag variables instead of printing warnings. The error codes are caught by the c wrapper functions and the appropriate sf_error is raised. This silences a lot of noise in test_cython_special.py.

Also make generate_ufuncs.py automatically generate the benchmarks.

Functions with multiple return values are handled by passing in pointers as the final arguments. This is intended to be analogous to providing output arrays to a ufunc.

Functions are listed by their c signatures instead of typecodes

person142 · 2016-06-19T06:12:57Z

And rebased.

pv · 2016-07-21T19:22:36Z

Thanks, merged based on positive feedback.

person142 · 2016-07-21T20:18:41Z

Thanks for taking the time to review @pv.

rgommers added enhancement A new feature or improvement scipy.special labels May 29, 2016

person142 force-pushed the cython-interface branch from 9eb801c to 0a0d83b Compare June 1, 2016 23:44

pv reviewed Jun 12, 2016
View reviewed changes

pv reviewed Jun 18, 2016
View reviewed changes

pv mentioned this pull request Jun 18, 2016

Pr 6195 person142/scipy#1

Merged

pv added the needs-work Items that are pending response from the author label Jun 18, 2016

pv removed the needs-work Items that are pending response from the author label Jun 18, 2016

person142 added 6 commits June 19, 2016 00:12

STY: pep8 fixes.

33c920f

MAINT: convert all int arguments to long

7bcb7f1

This mimics the behavior of iter_varaints in generate_ufuncs.py

MAINT: use cython extension types in tests instead of numpy dtypes

399ddf1

Hopefully this will fix issues with conversions.

MAINT: make hyp2f1, hypU use sf_error

65ba70d

Modifies specfun.f to return sf_error codes in added flag variables instead of printing warnings. The error codes are caught by the c wrapper functions and the appropriate sf_error is raised. This silences a lot of noise in test_cython_special.py.

BENCH: add more benchmarks for cython_special

931b4b1

Also make generate_ufuncs.py automatically generate the benchmarks.

person142 and others added 13 commits June 19, 2016 00:16

ENH: allow functions with multiple return values in cython_special

c082f24

Functions with multiple return values are handled by passing in pointers as the final arguments. This is intended to be analogous to providing output arrays to a ufunc.

MAINT: mark some tests as known failures.

5a95f06

MAINT: make sf_error work for cython_special

df9b71b

MAINT: special: isolate errprint code

8582137

DOC: special: simplify cython_special documentation

51d57de

Functions are listed by their c signatures instead of typecodes

MAINT: special: remove code generation for benchmarks and tests

df42fae

MAINT: special: initialize isfer and emit also unexpected error codes

01e3995

MAINT: mark cython_special files as binary in .gitattributes

c06caf3

TST: special: fix cython_special routine type specialization tests

d93f858

GEN: special: regenerate files

d983662

TST: add missing from __future__ stanza

5206dd4

TST: make known failure check earlier to guard against segfaults

9c29179

MAINT: make cython_special.pxd get installed

be3e37f

person142 force-pushed the cython-interface branch from 3b3d064 to be3e37f Compare June 19, 2016 05:31

pv added this to the 0.19.0 milestone Jul 21, 2016

pv merged commit 0a03211 into scipy:master Jul 21, 2016

person142 deleted the cython-interface branch July 22, 2016 21:35

pv mentioned this pull request Jul 22, 2016

ENH: special: extend sici/shichi to complex arguments #6406

Merged

tylerjereddy mentioned this pull request Jan 25, 2019

Clarify that Cython bindings are recent scipy/scipy-articles#87

Merged

Uh oh!

ENH: Cython interface for special #6195

ENH: Cython interface for special #6195

Uh oh!

Conversation

person142 commented May 28, 2016 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

rgommers commented Jun 5, 2016

Uh oh!

rgommers commented Jun 5, 2016

Uh oh!

person142 commented Jun 5, 2016

Uh oh!

pv commented Jun 5, 2016 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

person142 commented Jun 5, 2016

Uh oh!

rgommers commented Jun 12, 2016

Uh oh!

rgommers commented Jun 12, 2016

Uh oh!

pv Jun 12, 2016

Choose a reason for hiding this comment

Uh oh!

person142 commented Jun 14, 2016

Uh oh!

ev-br commented Jun 17, 2016 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

pv Jun 18, 2016

Choose a reason for hiding this comment

Uh oh!

pv commented Jun 18, 2016

Uh oh!

pv commented Jun 18, 2016

Uh oh!

person142 commented Jun 18, 2016

Uh oh!

pv commented Jun 18, 2016

Uh oh!

person142 commented Jun 18, 2016

Uh oh!

person142 commented Jun 19, 2016

Uh oh!

pv commented Jul 21, 2016

Uh oh!

person142 commented Jul 21, 2016

Uh oh!

Uh oh!

person142 commented May 28, 2016 •

edited

Loading

pv commented Jun 5, 2016 •

edited

Loading

ev-br commented Jun 17, 2016 •

edited

Loading