tree c7558bc711bfbd20332d0f3524912375348139bb
parent d7f8c2fbafa212880c987fc2be1f51d326144560
author Gerda Zsejke More <gerdazsejke.more@arm.com> 1762783761 +0100
committer Gerda Zsejke More <gerdazsejke.more@arm.com> 1763423129 -0800

Optimize vertical filter kernel in av1_warp_affine

Some filter kernels contain zero-padded coefficients in the
vertical_filter_4x1_f1 and vertical_filter_8x1_f1 functions. Skip
unnecessary multiplications by zero in the Neon (Armv8.0), Neon I8MM,
and SVE implementations to reduce instruction count, yielding up to
~5% performance improvement.

Move the vertical_filter_4x1_f1 and vertical_filter_8x1_f1 functions
into the header file, as their implementations are identical across
all extensions.

Change-Id: I23e8cd8e2516edd3323addd956549ce6e2a6d494