tree 1f3086fddb78d03f9ccd8b989eb14c9ac44cc707
parent 60acd13dcd30893a86942e756ef490b0761fe895
author David Turner <david.turner@argondesign.com> 1539946193 +0100
committer Debargha Mukherjee <debargha@google.com> 1540493946 +0000

Optimise get_horver_correlation_full, SSE4 & AVX2

I have optimised the pure C get_horver_correlation_full in rdopt.c and
also developed SSE4.1 and AVX2 implementations for further speed-up.  I
have also added functional equivalence unit tests for these.

Optimised C speed-up over un-optimised C at the unit test level:
  4x4  : 2.7x
 32x32 : 6.2x
128x128: 7.4x

SSE4.1 speed-up over un-optimised C at the unit test level:
  4x4  : 2.7x
 32x32 : 13.7x
128x128: 18.6x

AVX2 speed-up over un-optimised C at the unit test level:
  4x4  : 2.2x
 32x32 : 15.6x
128x128: 26.4x

Speed-ups for other block sizes, including rectangular blocks, are
similar.

Change-Id: Idc75006d5682e5cbf336d0427038e855da607097