Switch on SLP vectorize.

[stuartarchibald]

As title.

[stuartarchibald]

Seeing what impact this has on testing run times.

[stuartarchibald]

Command: time ./runtests.py --random=0.1

Mainline does this:

Ran 995 tests in 636.697s

OK (skipped=19, expected failures=1)

real    10m42.166s
user    11m33.653s
sys     0m7.492s

This branch does this:

Ran 995 tests in 632.633s

OK (skipped=19, expected failures=1)

real    10m37.993s
user    11m38.904s
sys     0m7.392s

seems like the impact on the test suite run time is potentially small/zero.

[stuartarchibald]

Mainline:

In [1]: %cpaste                                                                                                                 
Pasting code; enter '--' alone on the line to stop or use Ctrl-D.
:from numba import njit
import numpy as np

@njit
def foo(a1, a2, b1, b2, A):
    for x in range(0, len(A), 4):
        Z = A[x:x + 4]
        Z[0] = a1*(a1 + b1)
        Z[1] = a2*(a2 + b2)
        Z[2] = a1*(a1 + b1)
        Z[3] = a2*(a2 + b2)

foo(1.2, 3.4, 5.6, 7.8, np.empty(4000))::::::::::::
:<EOF>

In [2]: G = np.empty(40000)                                                                                                     

In [3]: %timeit foo(1.2, 3.4, 5.6, 7.8, G)                                                                                      
12.9 µs ± 2.75 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)

This branch:

In [1]: %cpaste                                                                                                                 
Pasting code; enter '--' alone on the line to stop or use Ctrl-D.
:from numba import njit
import numpy as np

@njit
def foo(a1, a2, b1, b2, A):
    for x in range(0, len(A), 4):
        Z = A[x:x + 4]
        Z[0] = a1*(a1 + b1)
        Z[1] = a2*(a2 + b2)
        Z[2] = a1*(a1 + b1)
        Z[3] = a2*(a2 + b2)

foo(1.2, 3.4, 5.6, 7.8, np.empty(4000))::::::::::::
:<EOF>

In [2]: G = np.empty(40000)                                                                                                     

In [3]: %timeit foo(1.2, 3.4, 5.6, 7.8, G)                                                                                      
10.3 µs ± 22.8 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)

[sklam]

Shouldn't this also get a config flag? e.g. config.SLP_VECTORIZE

[stuartarchibald]

Can do, but I really want to move away from specific config flags for each feature in the optimiser, I think it should be in the hands of the user but also it shouldn't be spelled one at a time like this. For 0.52 shall we do a flag so it's easy to debug/turn off but this is the last one like it!?!

[sklam]

Ok, make it the last flag.

[sklam]

/azp run

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).