ENH: ARM Neon implementation with intrinsic for np.argmax.

[Qiyu8]

I measured the execution time of np.argmax by using asv benchmark suite. The Performance has increased by 70%. Here is the detail result in Linux OpenEuler OS+Kunpeng 920 aarch64 machine:

============================
        dtype
 ---------------------------
        numpy.float32    217±20μs
        bool       47.7±0.5μs
============================
        dtype
 --------------------------
        numpy.float32    220±20μs
        bool       161±0.3μs
============================
       before           after         ratio
     [3f11db40]       [00b21d1b]
     <master>         <neon-argmax>
-       161±0.3μs       47.7±0.5μs     0.30  bench_reduce.ArgMax.time_argmax(<class 'bool'>)

SOME BENCHMARKS HAVE CHANGED SIGNIFICANTLY.
PERFORMANCE INCREASED.

Reference:

1. SSE2Neon : https://github.com/jratcliff63367/sse2neon
2. http://stackoverflow.com/questions/11870910/sse-mm-movemask-epi8-equivalent-method-for-arm-neon

[mattip]

Nice. This will need refactoring once the Universal Intrinsics code goes in. It might provide a nice benchmark for the conversion (to prove performance on non-x86 hardware is not degraded), so I think we should put it in. Do we have any other non-X86 intrinsic use?

[Qiyu8]

@mattip you are right, this will be a good benchmark case to demonstrate the power of Universal Intrinsics, one piece of code leads to performance improvements across all platforms(X86/ARM/Power), more non-X86 intrinsic usage will bring later.

[mattip]

LGTM. It is strange that argmax is not a ufunc. Just for completeness, how much does this change the size of _multiarray_umath.so?

[Qiyu8]

@mattip The size of _multiarray_umath.so at each branch is:
master: 16423KB
neon-argmax: 16423KB
The size remains the same. :)

[mattip]

Thanks @Qiyu8. I am a little concerned that this does no runtime detection of the neon instructions, I assume this is valid across all processors of a certain architecture (like we require SSE2 for X86).