tree 46c923512210378a67900ae5d5224824c34bdefa
parent 6f62e972fff1ee64c8c8fd3bd901d0b3b78d986b
author Jonathan Wright <jonathan.wright@arm.com> 1783691684 +0100
committer James Zern <jzern@google.com> 1784833498 -0700

Unroll Arm Neon DotProd variance kernels for w >= 32

Unroll the Arm Neon DotProd variance kernels and split the
accumulators for blocks having width >= 32. This change reduces the
time required to execute these kernels by up to 40% on a Neoverse V3
when compiled with Clang 22.

Change-Id: Ibc8541bde727cbba5df460704877dc5e61455d1f