Add C99's log2() function to the math library

[rhettinger]

BPO	11888
Nosy	@rhettinger, @jcea, @mdickinson, @vstinner
Files	issue11888.patch issue11888-part2.patch

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = 'https://github.com/mdickinson'
closed_at = <Date 2011-05-11.20:08:22.426>
created_at = <Date 2011-04-20.15:58:48.806>
labels = ['type-feature']
title = "Add C99's log2() function to the math library"
updated_at = <Date 2011-05-12.07:29:45.446>
user = 'https://github.com/rhettinger'

bugs.python.org fields:

activity = <Date 2011-05-12.07:29:45.446>
actor = 'mark.dickinson'
assignee = 'mark.dickinson'
closed = True
closed_date = <Date 2011-05-11.20:08:22.426>
closer = 'vstinner'
components = []
creation = <Date 2011-04-20.15:58:48.806>
creator = 'rhettinger'
dependencies = []
files = ['21861', '21918']
hgrepos = []
issue_num = 11888
keywords = ['patch']
message_count = 35.0
messages = ['134159', '134197', '134198', '134200', '134202', '134203', '134427', '134428', '134946', '134998', '135001', '135201', '135352', '135426', '135427', '135428', '135549', '135569', '135572', '135575', '135576', '135577', '135578', '135579', '135583', '135585', '135586', '135733', '135734', '135735', '135737', '135742', '135748', '135800', '135822']
nosy_count = 6.0
nosy_names = ['rhettinger', 'jcea', 'mark.dickinson', 'vstinner', 'rpetrov', 'python-dev']
pr_nums = []
priority = 'normal'
resolution = 'fixed'
stage = None
status = 'closed'
superseder = None
type = 'enhancement'
url = 'https://bugs.python.org/issue11888'
versions = ['Python 3.3']

[rhettinger]

The three most popular logarithm bases are 10, e, and 2. The math library has direct function calls for the first two but not the latter which is important in informatics.

Since a direct call can use a custom algorithm or native hardware support (such as the FLDLN2 fpu instruction), it provides better speed and accuracy than our existing math.log(x, 2) option.

[mdickinson]

See also bpo-3724.

I'm -0 on this: between log(x, 2) and int.bit_length, there's not much need for log2. log(x, 2) should be plenty accurate enough for most numerical needs; the exception is when you're taking log base 2 of an integer and need a guarantee of exact results for powers of 2, and int.bit_length generally solves that problem.

The main issue is that we'd have to provide (and maintain) our own implementation of log2 for Windows (and other OSs that don't have all the C99 support. Solaris?) That implementation should, ideally:

• provide exact values for powers of 2, and
• be monotonic.

and that's not trivial. As Raymond points out, on x86 / x64 we might be able to use inline assembly directly; that would probably cover us for Windows.

[vstinner]

The main issue is that we'd have to provide (and maintain) our own
implementation of log2 for Windows (and other OSs that don't have all
the C99 support. Solaris?)

No, we don't have to. Python has already a lot of optional functions, see for example the os module. We can provide log2() only if the C library has this function.

[mdickinson]

We can provide log2() only if the C library has this function.

Big -1 from me: I'd hate to see working Python scripts written on Unix fail on Windows because of a missing log2.

[mdickinson]

Rather than reinventing the wheel, it may be worth looking at what numpy does here.

[mdickinson]

it may be worth looking at what numpy does here.

... or it may not. NumPy just uses (approximation to 1/log(2)) * log(x) when log2 doesn't already exist. And indeed, on Windows:

Python 2.7.1 |EPD 7.0-2 (64-bit)| (r271:86832, Dec  2 2010, 10:23:25) [MSC v.1500 64 bit (AMD64)] on win32
Type "help", "copyright", "credits" or "license" for more information.
>>> import numpy
>>> numpy.log2(8.0)
2.9999999999999996

I think we should be able to do better than this. :-)

[vstinner]

The main issue is that we'd have to provide (and maintain) our own
implementation of log2 for Windows (and other OSs that don't have all
the C99 support. Solaris?)

Can't we simply use (approximation to 1/log(2)) * log(x)? Is it worse than reimplementing it using log(x)/log(2) in Python?

That implementation should, ideally: ...

Because some platforms are less accurate, you prefer to not provide a more accurate function when log2() is available? Can't we start with something simple and improve it later?

It can be documented than the Python log2 function may be the same than log(x)/log(2) if the platform/CPU doesn't provide a C log2 function.

--

"... And indeed, on Windows ...

>>> numpy.log2(8.0)
2.9999999999999996"

Oh. Python is better on Linux:

Python 2.6.6 (r266:84292, Dec 26 2010, 22:31:48) 
[GCC 4.4.5] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import math
>>> math.log(8) / math.log(2)
3.0

[vstinner]

Can't we simply use (approximation to 1/log(2)) * log(x)?
Is it worse than reimplementing it using log(x)/log(2) in Python?

Hum. With a x86 and the right compiler optimization level, log(x)/log(2) in C can be more accurate than log(x)/log(2) in Python, because the FPU works with 80 bits float internally, and the result is only "truncated" to 64 bits float at the end. In Python, the result is truncated to 64 bits on each Python instruction.

I don't know if it should be called a feature or a bug. In PHP world, it would be called a bug :-D http://bugs.php.net/bug.php?id=53632

[vstinner]

Oh... math.log() has an optional second argument: base. math.log(x, 2). But it is equivalent as math.log(x) / math.log(2) in Python. math.log(x, 2) is implemented as:
num=math.log(x)
den=math.log(2)
return num / den
where num and den are Python floats (64 bits).

So we don't benefit from 80 bits float used internally in x87.

[mdickinson]

Here's a patch implementing log2. Still to do: use the system log2 where available.

[rhettinger]

Wow Mark, that is really nice work. Thanks.

[vstinner]

Updated patch to use the system log2() if it is available. The test pass with the system log2() on Linux (Debian Sid, eglibc 2.11.2).

[mdickinson]

Thanks, Victor. I suspect we're going to need to be a bit more careful, though: when the extra tests were added for math.log, it turned out that it had all sorts of strange special-case behaviour on various platforms.

So I suspect that even on platforms that have log2 natively, it'll be necessary to factor out special cases and deal with those first, only passing positive finite floats onto the system log2. Take a look at m_log and the comment directly above it to see how that works.

I'd also like to check in the non-system version first, just to give it a thorough test on the buildbots, before adding in the version that uses the system log2 when available.

[vstinner]

Thanks, Victor. I suspect we're going to need to be a bit more
careful, though: when the extra tests were added for math.log, it
turned out that it had all sorts of strange special-case behaviour on
various platforms.

So I suspect that even on platforms that have log2 natively, it'll be
necessary to factor out special cases and deal with those first, only
passing positive finite floats onto the system log2. Take a look at
m_log and the comment directly above it to see how that works.

Oh, I see:

/*
Various platforms (Solaris, OpenBSD) do nonstandard things for
log(0),
log(-ve), log(NaN). Here are wrappers for log and log10 that deal
with
special values directly, passing positive non-special values through
to
the system
log/log10.
*/

I'd also like to check in the non-system version first, just to give
it a thorough test on the buildbots, before adding in the version that
uses the system log2 when available.

Yes, we can use log2() only for the "x > 0.0" case. My secret plan was
to check system log2() using the buildbots. But if we know that system
log() only works correctly with strictly positive numbers, it's faster
to directly only use system log2() for x > 0.0.

Updated patch (version 3) implements that.

[vstinner]

(Oh, I hit the wrong keyboard shortcut and sent my email too fast)

You can commit bpo-11888.patch first if you would like to test it.

In this case, here is a patch to use system log2(), patch to apply *after* bpo-11888.patch. It only uses log2() if x > 0.0.

[vstinner]

By the way, bpo-11888.patch is just fine: you can commit it. I like your frexp "trick" to improve the accuracy.

[python-dev]

New changeset 6d1cbfcee45a by Victor Stinner in branch 'default':
Issue bpo-11888: Add log2 function to math module. Patch written by Mark
http://hg.python.org/cpython/rev/6d1cbfcee45a

[mdickinson]

Thanks, Victor. You caught me by surprise a bit: I had some more minor changes to that patch pending, so I've committed those separately.

Any news from the buildbots?

[vstinner]

Thanks, Victor. You caught me by surprise a bit

Oh, I thought that the patch was ready to be commited.

I had some more minor changes to that patch pending,
so I've committed those separately.

You should add "Issue bpo-11888: " prefix to your commit messages so a bot automatically add comments for the new commits to this issue.

"Fix cut-and-paste typo in comment: log10 -> log2."

Oh, I didn't notice it, but Terry Reedy

"Fix nonunique test ids in math_testcases.txt."

Oh, I didn't notice it: why do we need explicit identifiers? Can't we use the line number or something like that? What happens if two tests have the same identifier? Is the first test skipped?

Any news from the buildbots?

Yes, most of them are happy. Some of them are busy (don't have test log2() yet), others are unhappy but not because of log2(). Said differently: log2 tests pass on most buildbots, but we have to wait 12 hours or maybe one day to wait for all buildbots.

If the test pass on all buildbots, I will commit bpo-11888-part2.patch to use the system log2 function.

[vstinner]

we have to wait 12 hours or maybe one day to wait for all buildbots

Oh, it's faster than expected: test_math passed on FreeBSD 6.4 3.x buildbot. I was waiting for this one because it's an old OS and many tests fail on this buildbot (because it's old but also slow). So test_math passed on all buildbots. Let's try the system log2 :-)

[python-dev]

New changeset 565f43f6bed4 by Victor Stinner in branch 'default':
Issue bpo-11888: Use system log2() when available
http://hg.python.org/cpython/rev/565f43f6bed4

[vstinner]

Issue bpo-11888: Use system log2() when available
http://hg.python.org/cpython/rev/565f43f6bed4

"I expect the system libc to use more accurate functions than Python."

You know what? Mac OS X log2 is less accurate than Python log2! A log2 test failed on "x86 Tiger 3.x":

http://www.python.org/dev/buildbot/all/builders/x86%20Tiger%203.x/builds/2488
======================================================================
FAIL: testLog2 (test.test_math.MathTests)
----------------------------------------------------------------------

Traceback (most recent call last):
  File "/Users/db3l/buildarea/3.x.bolen-tiger/build/Lib/test/test_math.py", line 658, in testLog2
    self.assertEqual(actual, expected)
AssertionError: Lists differ: [-324.0, -323.0, -322.0, -321.... != [-324.0, -323.0, -322.0, -321....

First differing element 69:
-254.99999999999997
-255.0
----------------------------------------------------------------------

Should I revert my patch or should we test the system log2 in configure to check if it is as accurate or more accurate than Mark's algorithm?

[mdickinson]

You know what? Mac OS X log2 is less accurate than Python log2!

That doesn't surprise me much. Though it's probably still true that log2 from OS X is more accurate than our log2 for some other values. It's just that getting the answer wrong for a power of 2 is a 'high-visibility' error.

Testing log2 sounds long-winded and error-prone. I'd suggest just marking that test as an expected failure on Tiger. BTW, I don't see any such failure on Snow Leopard.

[mdickinson]

One other thought: we should check that it's not pow that's at fault here, rather than log2. The test uses math.log2(2.0**n). It would probably be better off using math.log2(ldexp(1.0, n)), or similar: the libm pow operation is also notorious for inaccuracies (due to poor implementations or otherwise) on various platforms.

[vstinner]

we should check that it's not pow that's at fault here

Some tests on Mac OS X Tiger:

>>> (2.0 ** -255).hex()
'0x1.0000000000000p-255'

=> pow is correct

>>> import ctypes; import ctypes.util, math
>>> libc = ctypes.cdll.LoadLibrary(ctypes.util.find_library('c'))
>>> clog2=libc.log2
>>> clog2.restype=ctypes.c_double
>>> clog2.argtypes=(ctypes.c_double,)
>>> clog2(2.0**-255)
-254.99999999999997
>>> math.log(2.0**-255) / math.log(2.0)
-255.0

>>> math.log(2.0**-255)
-176.75253104278605
>>> math.log(2.0**-255).hex()
'-0x1.61814bbfb3fb5p+7'
>>> math.log(2.0)
0.6931471805599453
>>> math.log(2.0).hex()
'0x1.62e42fefa39efp-1'

>>> clog2(2.0**-255).hex()
'-0x1.fdfffffffffffp+7'
>>> (math.log(2.0**-255) / math.log(2.0)).hex()
'-0x1.fe00000000000p+7'

clog2() is wrong for 2^-255.

[mdickinson]

Okay, thanks. We should still be using ldexp rather than 2.0**... in the tests, though; I've fixed this, and also fixed the incorrect (too small) range for those tests, so that all representable powers of 2 are now covered.

[mdickinson]

Grr. Got the issue number wrong in the commit message; see msg135584.

New changeset 1f23d63b578c by Mark Dickinson in branch 'default':
Issue bpo-11188: In log2 tests, create powers of 2 using ldexp(1, n) instead of the less reliable 2.0**n.
http://hg.python.org/cpython/rev/1f23d63b578c

[rpetrov]

Why configure script check two times for log2 function ?

[python-dev]

New changeset d3f9895e2e19 by Mark Dickinson in branch 'default':
Issue bpo-11888: remove duplicate check for log2 in configure.in.
http://hg.python.org/cpython/rev/d3f9895e2e19

[mdickinson]

Thanks, Roumen. Fixed.

[mdickinson]

Victor, what do you think about simply #undefining HAVE_LOG2 on Tiger (e.g. in pyport.h), so that the fallback log2 version is used there instead of the system version?

Does anyone know the appropriate preprocessor check for OS X <= 10.4? I can get as far as "#ifdef __APPLE__", but don't know how to check for specific versions of OS X.

[python-dev]

New changeset 34871c3072c9 by Victor Stinner in branch 'default':
Issue bpo-11888: skip some log2 tests on Mac OS X Tiger
http://hg.python.org/cpython/rev/34871c3072c9

[vstinner]

New changeset 34871c3072c9 by Victor Stinner in branch 'default':
Issue bpo-11888: skip some log2 tests on Mac OS X Tiger

Oh... I realized that the test doesn't fail on Mac OS X Tiger PPC, only on Mac OS X Tiger x86. But I am too lazy to patch the test. Or should I do it?

I wait for the following build to close this issue.
http://www.python.org/dev/buildbot/all/builders/x86%20Tiger%203.x/builds/2507

[vstinner]

I wait for the following build to close this issue.
http://www.python.org/dev/buildbot/all/builders/x86%20Tiger%203.x/builds/2507

Oh, it's the wrong build. The correct build is:
http://www.python.org/dev/buildbot/all/builders/x86%20Tiger%203.x/builds/2508

And it passed so I close this issue.

[mdickinson]

Thanks, Victor.