array_scalar * numpy_scalar is slow.

[charris]

Some timings:

In [8]: a = np.float64(123)

In [9]: b = np.array(4.)

In [10]: timeit a*b
10000000 loops, best of 3: 110 ns per loop

In [11]: timeit b*a
1000000 loops, best of 3: 985 ns per loop

In [12]: timeit a*a
10000000 loops, best of 3: 66.6 ns per loop

In [13]: timeit b*b
1000000 loops, best of 3: 349 ns per loop

In [14]: timeit a + b
10000000 loops, best of 3: 109 ns per loop

In [15]: timeit b + a
1000000 loops, best of 3: 1.01 µs per loop

Maybe a fastpath it needed.

[juliantaylor]

[charris]

Cool! Or maybe, hot stuff ;) I love those graphs but I'm still coming to grips with how to understand them.

[juliantaylor]

it shows the sampled callstack with the width being the cumulative runtime. The main bottlenecks are ufunc_hasattr and GetPriority which have an array fastpath but no scalar one. Next largest fraction are a bunch of subtype check from PyArray_DescrFromScalar

[jorenham]

for the most part this no longer seems to be an issue

In [1]: a = np.float64(123)

In [2]: b = np.array(4.)

In [3]: %timeit a*b
318 ns ± 0.424 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

In [4]: %timeit b*a
307 ns ± 0.418 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

In [5]: %timeit a*a
23.7 ns ± 0.0384 ns per loop (mean ± std. dev. of 7 runs, 10,000,000 loops each)

In [6]: %timeit b*b
208 ns ± 0.0481 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

In [7]: %timeit a+b
323 ns ± 0.611 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

In [8]: %timeit b+a
308 ns ± 0.562 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

[mhvk]

Yes, this was fixed in #29819. I should have noticed there was an issue for it...