tree 8c16444348fde28b077a0e7b28c0018d43e7ee6b
parent 4ca0087fa0abcb6e2b3db7c9270f6e5e9dc9fb4d
author David Turner <david.turner@argondesign.com> 1539184868 +0100
committer David Turner <david.turner@argondesign.com> 1539276294 +0000

Fix and cleanup lowbd SIMD pixel_proj_error

In low bit-depth pixel_proj_error SSE4.1 and AVX2 implementations there
are two very similar codepaths depending whether only the first or only
the second SGR filter are enabled.  Combining these paths reduces
repetition with minimal extra complexity.

Also, in the SSE4.1 implementation was an apparent bug where a for loop

for (j = 0; j < width - 8; j += 8)

should have been

for (j = 0; j <= width - 8; j += 8)

This meant 8 (or 16 in another case) pixels were processed outside the
optimised SIMD loop, reducing speed-up.  Fixing this makes the SSE4.1
implementation of lowbd pixel_proj_error 20-70% faster depending on
filter configuration (and RUI size).

Change-Id: Iccdc7ea9da8131e2a4d6a0ee14b0a9924903e59b