tree c4d86e5ac1b6965c5f072f1121cd127cf12a793e
parent 416c1a9ff3198f9f4833ba6809ca65eba649273f
author Salome Thirot <salome.thirot@arm.com> 1709914437 +0000
committer James Zern <jzern@google.com> 1710794557 +0000

Specialize 4-tap HBD dist_wtd_convolve_x_sve2 on bitdepth

The rounding value at the end of the convolution depends on the bitdepth
(8/10, or 12). Add 2 specialized versions of the function, so that we
know the rounding value at compile time and therefore use only one
instruction instead of two to perform the final rounding and narrowing
step. This gives up to 20% uplift over the non-specialized version.

Change-Id: Id83a813ddefbb61704adf4b59d93f6910160f22d