nopython mode jit yields worse performance with unique function

[William Beard]

Here's a simplified version of `np.unique` that seems to be slowed by numba (in nopython mode)

    from numba import njit

    def unique_np(ar):
        std = np.sort(ar)
        flag = np.concatenate((np.array([True]), std[1:] != std[:-1]))
        return std[flag]

    unique_jit = njit(unique_np)


    # Check they both work
    inp = np.array([6, 1, 2, 2, 3])
    assert (unique_jit(inp) == unique_np(inp)).all()
    assert (np.unique(inp) == unique_np(inp)).all()

    

For following input the original has a better time on my machine

 

    nr.seed(0)
    ri = nr.randint(0, high=1000, size=10000)
    assert (unique_jit(ri)  == np.unique(ri)).all()  # checking again...
   
    %timeit unique_np(ri)   # => 1000 loops, best of 3: 427 µs per loop
    %timeit unique_jit(ri)  # => 1000 loops, best of 3: 594 µs per loop

    

Am I doing something un-numban here? I'm curious because this is the first example of performance degradation I've seen.

[Jason Sachs]

it's not your code, it's the way that numba seems to handle np.sort:

def just_sort(x):

    return np.sort(x)

jitsort = njit(just_sort)

jitsort(ri)   # prime the jit

%timeit just_sort(ri)  # 1000 loops, best of 3: 409 µs per loop

%timeit jitsort(ri)    # 1000 loops, best of 3: 634 µs per loop

--
You received this message because you are subscribed to the Google Groups "Numba Public Discussion - Public" group.
To unsubscribe from this group and stop receiving emails from it, send an email to numba-users+unsubscribe@continuum.io.
To post to this group, send email to numba...@continuum.io.
To view this discussion on the web visit https://groups.google.com/a/continuum.io/d/msgid/numba-users/4bf4ffc5-2b78-4e0a-a186-fbceaef76f2c%40continuum.io.
For more options, visit https://groups.google.com/a/continuum.io/d/optout.

[Jason Sachs]

see also: 

http://numba.pydata.org/numba-doc/0.26.0/reference/numpysupported.html

Warning

Sorting may be slightly slower than Numpy’s implementation.