Performance of asarray() when input is already an array

[anntzer]

I have noticed that np.asarray(x) and np.asarray(x, float) have pretty big overheads when the input, x, is already a float array. One could perhaps check whether that is already the case, immediately returning x if possible. The additional check would have a negligible cost in the case where a conversion is needed (because the conversion e.g. from a x that's a list is much slower than anything else).

Reproducing code example:

In [1]: x = np.random.rand(10, 10)                                                               

In [2]: %timeit np.asarray(x)                                                                    
345 ns ± 0.53 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)

In [3]: %timeit np.asarray(x, float)                                                             
613 ns ± 0.274 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)

# Check for fast path.
In [4]: %timeit x if type(x) is np.ndarray and x.dtype == float else np.asarray(x, float)        
214 ns ± 0.894 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)

In [5]: x = np.random.rand(10, 10).tolist()                                                      

In [6]: %timeit np.asarray(x)                                                                    
6.77 µs ± 6.7 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)

In [7]: %timeit np.asarray(x, float)                                                             
6.78 µs ± 14.8 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)

# The additional check is a bit slower but not so much, relatively speaking.
In [8]: %timeit x if type(x) is np.ndarray and x.dtype == float else np.asarray(x, float)        
6.93 µs ± 10.1 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)

Numpy/Python version information:

1.16.3 3.7.3 (default, Mar 26 2019, 21:43:19) 
[GCC 8.2.1 20181127]

[hmaarrfk]

I think the discussion #11804 is relevant.

[rgommers]

@seberg and I just discussed this: the expected speedup from moving asarray to C is almost 2x, and asarray can be a performance bottleneck on real-world code bases because it's called so often. It may also give a nice speedup on test suites.

[seberg]

This has been fixed by my new argparsing code and moving these array "aliases" to C.

Using asarray, etc. is now preferable to using array (if you pass fewer keywords). But generally using keyword arguments comes with only a very small overhead penalty now.

[rgommers]

That's great news, thanks @seberg!