[CR] Add VSX Hadamard

[rafaeldelucena]

Description

This PR add Hadamard Transform using the VSX instructions set, aiming to optimize this operation for Power8 in order to resolve the issue #6

Checklist:

• [x] Add code to Build System
• [x] Implement the Transpose 8x8 for signed shorts
• [x] Add Unit tests
• [x] Implement Hadamard 8x8
• [x] Implement Hadamard 16x16
• [x] Test for endianness and unaligned addresses issues
• [x] Make all tests pass

[rafaeldelucena]

@lu-zero I've successful test the implementation on both little and big endian architectures, also I'm using the vec_vsx_ld and vec_vsx_st to load and store data, so I'll mark the Test for endianness and unaligned addresses issues as checked for now, I think it's ready for review. :)

[rafaeldelucena]

This is a first version, I don't apply any further optimizations, but if you find it necessary I can make the vpx_transpose_8x8_s16 and vpx_hadamard_8x8_s16_one_pass inline functions also I can use an aligned address buffer and replace to simple load and store functions for char arrays, aiming to reduce the permutation overhead of the used instructions.

[lu-zero]

Using the test harness I'm not seeing any speedup, I guess some more work might be needed.

[lu-zero]

Seems that the test harness does not double as benchmark harness easily, do you have the alternative benchmark code ready to be in the tree?

[rafaeldelucena]

I've some test code here, I'll push until the end of the day.

[rafaeldelucena]

The benchmark code is very simple, runs the hadamard implementation with multiple strides from 10^1 to 10^6 times with the same input matrix.

[rafaeldelucena]

Already merged to upstream