The attached patch drastically speeds up PyLong_AsDouble for single digit longs:


-m timeit -s "x=2" "x*2.2 + 2 + x*2.5 + 1.0 - x / 2.0 + (x+0.1)/(x-0.1)*2 + (x+10)*(x-30)"

with patch: 0.414
without: 0.612

spectral_norm: 1.05x faster.    The results are even better when paired with patch from issue #21955.

New changeset 986184c355e8 by Yury Selivanov in branch 'default':
Issue #26288: Optimize PyLong_AsDouble.
https://hg.python.org/cpython/rev/986184c355e8

Thanks a lot for the review, Serhiy!

Nice enhancement.

/* Fast path; single digit will always fit decimal.
   This improves performance of FP/long operations by at
   least 20%. This is even visible on macro-benchmarks.
*/

I'm not sure that "spectral_norm" can be qualified as macro-benchmark. It's more a microbenchmark on numerics, no? I'm just nitpicking, your patch is fine ;-)

Actually, please fix the comment. We don't want someone wondering what those "macro-benchmarks" are.

Additionally, "single digit will always fit a double"?

> Actually, please fix the comment. We don't want someone wondering what those "macro-benchmarks" are.

If spectral-norm and nbody aren't good benchmarks then let's remove them from our benchmarks suite.

I'll remove that comment anyways, as it doesn't make a lot of sense :)

> Additionally, "single digit will always fit a double"?

What's wrong about that phrase?  Would this be better: "It's safe to cast a single-digit long (31 bits) to double"?

Sorry, I was a bit brief: The current comment says "decimal" instead of "double". It should be changed to "double".

Le 06/02/2016 18:07, Yury Selivanov a écrit :
> 
>> Actually, please fix the comment. We don't want someone wondering what those "macro-benchmarks" are.
> 
> If spectral-norm and nbody aren't good benchmarks then let's remove
> them from our benchmarks suite.

Probably, yes.

Actually, let me refine that. nbody and spectral-norm don't make sense
for people running CPython. Perhaps people running PyPy might care about
their performance... (though PyPy is supposed to support a subset of
Numpy too)

New changeset cfb77ccdc23a by Yury Selivanov in branch 'default':
Issue #26288: Fix comment
https://hg.python.org/cpython/rev/cfb77ccdc23a

> Sorry, I was a bit brief: The current comment says "decimal" instead of "double". It should be changed to "double".

Oh, got it now, sorry.  I rephrased the comment a bit, hopefully it's better now.  Please check.  Thanks!

The comment looks good to me -- I'll stay out of the benchmarking issue, I didn't check any of that. :)

Well I *did* run the decimal/float milli-benchmark now and it shows at least 15% improvement for floats consistently.

Given that the official benchmark suite does not seem to be very stable either (#21955), I actually prefer small and well-understood benchmarks.

I'm not sure why this issue is open... Closing it.