tree 8db3145c5effb994e068eb1fbbded64fcb1417dd
parent c26cf7ff3c110749757305f548279a6236b02901
author George Steed <george.steed@arm.com> 1681914937 +0100
committer James Zern <jzern@google.com> 1682624126 +0000

Add Neon implementations for remaining h predictors

We already have Neon implementations of the h predictors for all square
block sizes except 64x64, so add the remaining cases and update
tests/speed to match. We also clean up and refactor the existing cases
to expose some helper functions we can reuse in the new implementations.

On Neoverse N1 these new implementations are worth about a 33% geomean
performance improvement over the C versions on both Clang 15 and GCC 12.
On Neoverse V1 there is a smaller improvement, with a geomean
improvement of around 3% on Clang 15 and 15% on GCC 12.

Change-Id: Ib9dd1e2c924c41d620635b8e3acb813242642cb3