[Cython] Hash-based vtables

[Dag Sverre Seljebotn]

On 06/12/2012 08:12 PM, Robert Bradshaw wrote:
> On Tue, Jun 12, 2012 at 10:21 AM, Dag Sverre Seljebotn
> <d.s.seljebotn at astro.uio.no>  wrote:
>> On 06/12/2012 01:01 PM, Dag Sverre Seljebotn wrote:
>>>
>>> On 06/10/2012 11:53 AM, Robert Bradshaw wrote:
>>>>
>>>> On Sun, Jun 10, 2012 at 1:43 AM, Dag Sverre Seljebotn
>>>>>
>>>>> About signatures, a problem I see with following the C typing is that
>>>>> the
>>>>> signature "ill" wouldn't hash the same as "iii" on 32-bit Windows and
>>>>> "iqq"
>>>>> on 32-bit Linux, and so on. I think that would be really bad.
>>>>
>>>>
>>>> This is why I suggested promotion for scalars (divide ints into
>>>> <=sizeof(long) and sizeof(long)<  x<= sizeof(long long)), checked at
>>>> C compile time, though I guess you consider that evil. I don't
>>>> consider not matching really bad, just kind of bad.
>>>
>>>
>>> Right. At least a convention for promotion of scalars would be good
>>> anyway.
>>>
>>> Even MSVC supports stdint.h these days; basing ourselves on the random
>>> behaviour of "long" seems a bit outdated to me. "ssize_t" would be
>>> better motivated I feel.
>>>
>>> Many linear algebra libraries use 32-bit matrix indices by default, but
>>> can be swapped to 64-bit indices (this holds for many LAPACK
>>> implementations and most sparse linear algebra). So often there will at
>>> least be one typedef that is either 32 bits or 64 bits without the
>>> Cython compiler knowing.
>>>
>>> Promoting to a single type "[u]int64" is the only one that removes
>>> possible combinatorial explosion if you have multiple external typedefs
>>> that you don't know the size of (although I guess that's rather
>>> theoretical).
>>>
>>> Anyway, runtime table generation is quite fast, see below.
>>>
>>>>
>>>>> "l" must be banished -- but then one might as well do "i4i8i8".
>>>>>
>>>>> Designing a signature hash where you select between these at
>>>>> compile-time is
>>>>> perhaps doable but does generate a lot of code and makes everything
>>>>> complicated.
>>>>
>>>>
>>>> It especially gets messy when you're trying to pre-compute tables.
>>>>
>>>>> I think we should just start off with hashing at module load
>>>>> time when sizes are known, and then work with heuristics and/or build
>>>>> system
>>>>> integration to improve on that afterwards.
>>>>
>>>>
>>>> Finding 10,000 optimal tables at runtime better be really cheap than
>>>> for Sage's sake :).
>>>
>>>
>>> The code is highly unpolished as I write this, but it works so here's
>>> some preliminary benchmarks.
>>>
>>> Assuming the 64-bit pre-hashes are already computed, hashing a 64-slot
>>> table varies between 5 and 10 us (microseconds) depending on the set of
>>> hashes.
>>>
>>> Computing md5's with C code from ulib (not hashlib/OpenSSL) takes ~400ns
>>> per hash, so 26 us for the 64-slot table =>  it dominates!
>>>
>>> The crapwow64 hash takes ~10-20 ns, for ~1 us per 64-slot table.
>>> Admittedly, that's with hand-written non-portable assembly for the
>>> crapwow64.
>>>
>>> Assuming 10 000 64-slot tables we're looking at something like 0.3-0.4
>>> seconds for loading Sage using md5, or 0.1 seconds using crapwow64.
>>>
>>>
>>> https://github.com/dagss/pyextensibletype/blob/master/include/perfecthash.h
>>>
>>> http://www.team5150.com/~andrew/noncryptohashzoo/CrapWow64.html
>>
>>
>> Look: A big advantage of the hash-vtables is that subclasses stay
>> ABI-compatible with superclasses, and don't need recompilation when
>> superclasses adds or removes methods.
>>
>> =>  Finding the hash table must happen at run-time in a lot of cases anyway,
>> so I feel Robert's chase for a compile-time table building is moot.
>>
>> I feel this would also need to trigger automatically heap-allocated tables
>> if the statically allocated. Which is good to have in the very few cases
>> where a perfect table can't be found too.
>
> Finding the hash table at runtime should be supported, but the *vast*
> majority of methods sets is known at compile time. 0.4 seconds is a
> huge overhead to just add to Sage (yes, it's an exception, but an
> important one), and though crapwow64 helps I'd rather rely on a known,
> good standard hash. I need to actually look at Sage to see what the
> impact would be. Also, most tables would probably have 2 entries in
> them (e.g. a typed one and an all-object one).

Hopefully 0.4 was a severe overestimate once one actually looks at this.

What's loaded at startup -- is it the pyx files in sage/devel/sage? My 
count (just cloned from github.com/sagemath/sage):

$ find -name '*.pyx' -exec grep 'cdef class' {} \; | wc -l
641

And I doubt that *all* of that is loaded at Sage startup, you need to do 
some manual importing for at least some of those classes? So it's 
probably closer to 0.01-0.02 seconds than 0.4 even with md5?

About the *vast* majority of method sets being known: That may be the 
case for old code, but keep in mind that that situation might 
deteriorate. Once we have hash-based vtables, declaring methods of cdef 
classes in pxd files could become optional (and only be there to help 
callers, incl. subclasses, determine the signature). So any method 
that's only used in the superclass and is therefore not declared in the 
pxd file would consistently trigger a run-time build of the table of 
subclasses; the compile-time generated table would be useless then.

(OTOH, as duck-typing becomes the norm, more cdef classes will be 
without superclasses...)

> long int will continue to be an important type as long as it's the
> default for int literals and Python's "fast" ints (whether in type or
> implementation), so we can't just move to stdint. I also don't like
> that the form of the table (and whether certain signatures match)
> being platform-dependent: the less variance we have from one platform
> to the next is better.

Perhaps in Sage there's a lot of use of "long" and therefore this would 
make Sage code vary less between platforms.

But for much NumPy-using code you'd typically use int32 or int64, and 
since long is 32 bits on 32-bit Windows and 64 bits on Linux/Mac, 
choosing long sort of maximises inter-platform variation of signatures...

> On an orthogonal note, sizeof(long)-sensitive tables need not be
> entirely at odds with compile-time table compilation, as most
> functions will probably have 0 or 1 parameters that are of unknown
> size, so we could spit out 1 or 2 statically compiled tables and do
> generate the rest on the fly. I still would rather have fixed
> Cython-compile time tables though.

Well, I'd "rather have" that as well if it worked every time.

But there's no use designing a feature which works great unless you use 
the fftw_complex type (can be 64 or 128 bits). Or works great unless you 
use 64-bit LAPACK. Or works great unless you have a superclass with a 
partially defined pxd file.

Since one implementation of a concept is simpler than two, then as long 
as run-time generation code must always be there (or at least, be there 
in the common cases x, y, and z), the reasons should be very good for 
adding a compile-time implementation.

Sage taking 0.4 seconds extra would indeed be a very good reason, but I 
don't believe it. So when you can get around to it it'd be great to have 
the actual number of classes (and ideally an estimate for number of 
methods per class).

Dag