On 06/04/2012 09:44 PM, Dag Sverre Seljebotn wrote:
Me and Robert had a long discussion on the NumFOCUS list about this
already, but I figured it was better to continue it and provide more
in-depth benchmark results here.
It's basically a new idea of how to provide a vtable based on perfect
hashing, which should be a lot simpler to implement than what I first
imagined.
I'll write down some context first, if you're familiar with this
skip ahead a bit..
This means that you can do fast dispatches *without* the messy
business of binding vtable slots at compile time. To be concrete, this
might e.g. take the form
def f(obj):
obj.method(3.4) # try to find a vtable with "void method(double)" in it
or, a more typed approach,
# File A
cdef class MyImpl:
def double method(double x): return x * x
# File B
# Here we never know about MyImpl, hence "duck-typed"
@cython.interface
class MyIntf:
def double method(double x): pass
def f(MyIntf obj):
# obj *can* be MyImpl instance, or whatever else that supports
# that interface
obj.method(3.4)
Now, the idea to implement this is:
a) Both caller and callee pre-hash name/argument string
"mymethod:iidd" to 64 bits of hash data (probably lower 64 bits of
md5)
b) Callee (MyImpl) generates a vtable of its methods by *perfect*
hashing. What you do is define a final hash fh as a function
of the pre-hash ph, for instance
fh = ((ph >> vtable.r1) ^ (ph >> vtable.r2) ^ (ph >> vtable.r3)) & vtable.m
(Me and Robert are benchmarking different functions to use here.) By
playing with r1, r2, r3, you have 64**3 choices of hash function, and
will be able to pick a combination which gives *no* (or very few)
collisions.
c) Caller then combines the pre-hash generated at compile-time, with
r1, r2, r3, m stored in the vtable header, in order to find the
final location in the hash-table.
The exciting thing is that in benchmark, the performance penalty is
actually very slight over a C++-style v-table. (Of course you can
cache a proper vtable, but the fact that you get so close without
caring about caching means that this can be done much faster.)
Back to my and Robert's discussion on benchmarks:
I've uploaded benchmarks here:
https://github.com/dagss/hashvtable/tree/master/dispatchbench
I've changed the benchmark taking to give more robust numbers (at
least for me), you want to look at the 'min' column.
I changed the benchmark a bit so that it benchmarks a *callsite*.
So we don't pass 'h' on the stack, but either a) looks it up in a global
variable (default), or b) it's a compile-time constant (immediate in
assembly) (compile with -DIMHASH).
Similarly, the ID is either an "interned" global variable, or an
immediate (-DIMID).
The results are very different on my machine depending on this aspect.
My conclusions:
- Both three shifts with masking, two shifts with a "fallback slot"
(allowing for a single collision), three shifts, two shifts with
two masks allows for constructing good vtables. In the case of only
two shifts, one colliding method gets the twoshift+fback
performance and the rest gets the twoshift performance.
- Performance is really more affected by whether hashes are
immediates or global variables than the hash function. This is in
contrast to the interning vs. key benchmarks -- so I think that if
we looked up the vtable through PyTypeObject, rather than getting
the vtable directly, the loads of the global variables could
potentially be masked by that.
- My conclusion: Just use lower bits of md5 *both* for the hashing
and the ID-ing (don't bother with any interning), and compile the
thing as a 64-bit immediate. This can cause crashes/stack smashes
etc. if there's lower-64bit-of-md5 collisions, but a) the
probability is incredibly small, b) it would only matter in
situations that should cause an AttributeError anyway, c) if we
really care, we can always use an interning-like mechanism to
validate on module loading that its hashes doesn't collide with
other hashes (and raise an exception "Congratulations, you've
discovered a phenomenal md5 collision, get in touch with cython
devs and we'll work around it right away").
What I forgot to mention:
- I really want to avoid linear probing just because of the code bloat
in call sites. With two shifts, when there was a failure to find a
perfect hash it was always possible to find one with a single collision.
- Probing for the hash with two shifts is lightning fast, it can take
a while with three shifts (though you can always spend more memory on a
bigger table to make it fast again). However, it makes me uneasy to
penalize the performance of calling one of the random methods, so I'm
really in favour of three-shifts or double-mask (to be decided when
investigating the performance of probing for parameters in more detail).
- I tried using SSE to do shifts in parallel and failed (miserable
performance). The problem is quickly moving things between general
purpose registers and SSE registers, and the lack of SSE
immediates/constants in the instruction stream. At least, what my gcc
4.6 generates appeared to use the stack to communicate between SSE
registers and general purpose registers (but I can't have been doing the
right thing..).
The RTTI (i.e. the char*) is also put in there, but is not used for
comparison and is not interned.
At least, that's what I think we should do for duck-style vtables.
Do we then go to all the pain of defining key-encoding, interning
etc. just for SEP 201? Isn't it easier to just mandate a md5 dependency
and be done with it? (After all, md5 usually comes with Python in the
md5 and hashlib modules)
direct: Early-binding
index: Call slot 0 (C++-style vtable/function pointer)
noshift: h & m1
oneshift: (h >> r1) & m1
twoshift: ((h >> r1) ^ (h >> r2)) & m1
twoshift+fback: hash doesn't
I meant: Hash collision and then, after a branch miss, look up the one
fallback slot in the vtable header.
Dag
threeshift: ((h >> r1) ^ (h >> r2) ^ (h >> r3)) & m1
doublemask: ((h >> r1) & m1) ^ ((h >> r2) & m2)
doublemask2: ((h >> r1) & m1) ^ ((h & m2) >> r2)
Default distutils build (-O2):
------------------------------
Hash globalvar, ids globalvar
direct: min=5.38e-09 mean=5.45e-09 std=3.79e-11 val=1600000000.000000
index: min=5.38e-09 mean=5.44e-09 std=3.09e-11 val=1200000000.000000
noshift: min=5.99e-09 mean=6.14e-09 std=6.63e-11 val=1200000000.000000
oneshift: min=6.47e-09 mean=6.53e-09 std=3.21e-11 val=1200000000.000000
twoshift: min=7.00e-09 mean=7.08e-09 std=3.73e-11 val=1200000000.000000
twoshift+fback: min=7.54e-09 mean=7.61e-09 std=4.46e-11
val=1200000000.000000
threeshift: min=7.54e-09 mean=7.64e-09 std=3.79e-11 val=1200000000.000000
doublemask: min=7.56e-09 mean=7.64e-09 std=3.02e-11 val=1200000000.000000
doublemask2: min=7.55e-09 mean=7.62e-09 std=3.24e-11 val=1200000000.000000
hash immediate, ids globalvar
direct: min=5.38e-09 mean=5.45e-09 std=3.87e-11 val=1600000000.000000
index: min=5.40e-09 mean=5.45e-09 std=2.92e-11 val=1200000000.000000
noshift: min=5.38e-09 mean=5.44e-09 std=3.48e-11 val=1200000000.000000
oneshift: min=5.90e-09 mean=5.99e-09 std=4.05e-11 val=1200000000.000000
twoshift: min=6.09e-09 mean=6.17e-09 std=3.52e-11 val=1200000000.000000
twoshift+fback: min=7.00e-09 mean=7.08e-09 std=3.64e-11
val=1200000000.000000
threeshift: min=6.47e-09 mean=6.55e-09 std=6.04e-11 val=1200000000.000000
doublemask: min=6.46e-09 mean=6.50e-09 std=3.37e-11 val=1200000000.000000
doublemask2: min=6.46e-09 mean=6.51e-09 std=3.04e-11 val=1200000000.000000
all immediate:
direct: min=5.39e-09 mean=5.50e-09 std=5.22e-11 val=1600000000.000000
index: min=5.38e-09 mean=5.51e-09 std=6.25e-11 val=1200000000.000000
noshift: min=5.38e-09 mean=5.51e-09 std=6.90e-11 val=1200000000.000000
oneshift: min=5.40e-09 mean=5.51e-09 std=5.35e-11 val=1200000000.000000
twoshift: min=5.94e-09 mean=6.06e-09 std=5.91e-11 val=1200000000.000000
twoshift+fback: min=7.06e-09 mean=7.19e-09 std=5.39e-11
val=1200000000.000000
threeshift: min=5.96e-09 mean=6.07e-09 std=5.54e-11 val=1200000000.000000
doublemask: min=5.88e-09 mean=6.01e-09 std=6.06e-11 val=1200000000.000000
doublemask2: min=5.94e-09 mean=6.05e-09 std=6.16e-11 val=1200000000.000000
-O3 build
---------
all globalvars:
direct: min=1.61e-09 mean=1.63e-09 std=1.40e-11 val=1600000000.000000
index: min=5.38e-09 mean=5.43e-09 std=2.82e-11 val=1200000000.000000
noshift: min=6.04e-09 mean=6.13e-09 std=4.76e-11 val=1200000000.000000
oneshift: min=6.46e-09 mean=6.54e-09 std=3.82e-11 val=1200000000.000000
twoshift: min=7.01e-09 mean=7.06e-09 std=3.41e-11 val=1200000000.000000
twoshift+fback: min=7.57e-09 mean=7.64e-09 std=3.47e-11
val=1200000000.000000
threeshift: min=7.54e-09 mean=7.63e-09 std=4.17e-11 val=1200000000.000000
doublemask: min=7.54e-09 mean=7.61e-09 std=3.64e-11 val=1200000000.000000
doublemask2: min=7.55e-09 mean=7.63e-09 std=3.35e-11 val=1200000000.000000
hash immediate, ids globalvar:
direct: min=1.61e-09 mean=1.66e-09 std=3.30e-11 val=1600000000.000000
index: min=5.40e-09 mean=5.50e-09 std=4.94e-11 val=1200000000.000000
noshift: min=5.38e-09 mean=5.49e-09 std=6.02e-11 val=1200000000.000000
oneshift: min=5.95e-09 mean=6.06e-09 std=6.64e-11 val=1200000000.000000
twoshift: min=5.96e-09 mean=6.13e-09 std=7.22e-11 val=1200000000.000000
twoshift+fback: min=7.02e-09 mean=7.18e-09 std=7.04e-11
val=1200000000.000000
threeshift: min=6.52e-09 mean=6.65e-09 std=6.43e-11 val=1200000000.000000
doublemask: min=6.50e-09 mean=6.62e-09 std=5.28e-11 val=1200000000.000000
doublemask2: min=6.52e-09 mean=6.63e-09 std=5.23e-11 val=1200000000.000000
all immediate:
direct: min=1.61e-09 mean=1.62e-09 std=9.77e-12 val=1600000000.000000
index: min=5.38e-09 mean=5.39e-09 std=1.71e-11 val=1200000000.000000
noshift: min=5.38e-09 mean=5.40e-09 std=2.41e-11 val=1200000000.000000
oneshift: min=5.38e-09 mean=5.40e-09 std=1.81e-11 val=1200000000.000000
twoshift: min=5.92e-09 mean=5.93e-09 std=1.43e-11 val=1200000000.000000
twoshift+fback: min=7.00e-09 mean=7.01e-09 std=2.20e-11
val=1200000000.000000
threeshift: min=5.92e-09 mean=5.94e-09 std=1.99e-11 val=1200000000.000000
doublemask: min=5.79e-09 mean=5.82e-09 std=2.32e-11 val=1200000000.000000
doublemask2: min=5.92e-09 mean=5.94e-09 std=2.25e-11 val=1200000000.000000
_______________________________________________
cython-devel mailing list
cython-devel@python.org
http://mail.python.org/mailman/listinfo/cython-devel
_______________________________________________
cython-devel mailing list
cython-devel@python.org
http://mail.python.org/mailman/listinfo/cython-devel
