tree 17093ae411aad07411f68361148dc80af62d55d9
parent a88627a130a7b455cc14201d23bd11a75e038b5f
author Gerda Zsejke More <gerdazsejke.more@arm.com> 1715699066 +0200
committer Jonathan Wright <jonathan.wright@arm.com> 1717523757 +0000

Add Armv8.0 Neon horiz 2x1 scale spec. impl for convolve_2d_scale

AV1 has a limit on the scale ratio, specifically, the reference
resolution cannot be more than 2 times the source resolution in any
dimension. Given that the algorithm uses higher precision
(1/1024-pel) for the step size (chapter 7.11.3.4. [1]), the
horizontal scaling function can be easily optimised for this specific
case. The indices of the source pixel to be interpolated are
calculated using the (subpel_qn + x * step) >> 1024 equation, which
can be simplified if step is a multiple of 1024.

Add implementation that specialises on x_step_qn equals to 2048, that
gives an uplift of around 33% when a 2x1 scaling is applied.

[1]https://aomediacodec.github.io/av1-spec/av1-spec.pdf

Change-Id: I9127ca4e6b4188a4dabe4cfd416efe4d762b2e9f