The chroma is never taken to 4:4:4 for P and T; only the intensity uses an interpolated chroma value, so nothing has to be filtered back down.
For each pair of pixels, from the composed picture:
The 4:2:0 chroma goes to 4:2:2 vertically (the same 12-tap Lanczos step the YCbCr tunnel already does).
P and T are computed at the even pixel, which sits on the 4:2:2 chroma sample: its own Y and that chroma, through the RPU's matrix to R'G'B', then linear light, LMS, PQ and IPT. No horizontal filter touches them.
I is needed for both pixels. For the odd one, its chroma is interpolated horizontally with the same Lanczos, and only its I is kept.
hadesdevls Thanks for the details and reference.
Is that not identically equivalent to upsampling to 4:4:4, doing the conversion, and then dropping the odd chroma pixels rather doing standard chroma downsampling that first runs an anti-aliasing filter?
