dsp: p_firsym: Implement p_firsym

[lfochamon]

This implementation refers to issue #158. Implementation-wise, it follows the same ideas as p_fir (see PR #152 for details). Notes:

• As it is, p_firsym implements symmetric FIR filters of odd length (a.k.a., type I). In other words, filters such as 1, 2, 3, 4, 5, 4, 3, 2, 1. In this case, one would pass h = {1, 2, 3, 4, 5} and nh = 5 to the function.
• There is a longer discussion on the filter type issue on dsp: p_firsym: API issue and implementation #158, but it all boils down to this: filters that have even length (type 2, e.g., 1 2 3 4 4 3 2 1) cannot be high pass or band stop. Type I filters have no restrictions.
• p_firsym is generally faster than p_fir. On ARM cores, it is always faster. On Intel i7, it is only slightly slower for shorter filters (small nh). The result of a larger benchmark can be found in dsp: p_firsym: API issue and implementation #158.
• Valuing speed over accuracy, this function does not use fused add-multiply. The tests I ran showed that the filter becomes considerably slower and that the accuracy increase is negligible.

I'll push the tests for this function and p_fir ASAP.

--Chamon (luizchamon@gmail.com)

[aolofsson]

Looks good. Please sign off on all pull requests
(can do in comment below for now)

[lfochamon]

Sorry, do you mean name and email on PR? I edited my first comment, please let me know if that's OK...

[aolofsson]

Like Linux and many others we are using this format. Pleas ee guidelines.
https://github.com/parallella/pal/blob/master/CONTRIBUTING.md

Signed-off-by: Joe Smith joe.smith@email.com

Good enough for now, but keep in mind for future. Thanks!