Tags: Charrine/rav1e
Tags
Ease auto vectorization of tx distortion computation The compiler was not autovectorizing this section of code. Switches the code to doing two seperate iteration on the input coefficients from using repeat and chain. The first iteration processes the stored reconstructed coefficients and the second processes the reconstructed coefficients that can be assumed to be zero. This isn't used by default, but Thomas is working on a project that uses it.
Optimize sgrproj_solve to avoid f64 arithmetic Results can always fit in i64 regardless of what bit depth is used. It doesn't seem like much vectorization is happening but it is faster. Some quick tests show that switching to a smaller type seems to enable auto-vectorization.
PreviousNext