On Thu, 5 Nov 2026, zhujing via ffmpeg-devel wrote:
Replace the store-to-stack + ld4 transpose in ff_hscale8to15_4_neon with a register ins + tbl transpose carried out entirely in caller-saved registers (no stack). Uses function/endfunc and portable const/movrel idioms for the tbl index tables, keeps 8-pixel vector processing down to a <=7-pixel scalar tail, and is bit-exact.
Thanks for your patch submission! For review convenience and for consistent testing of the patch, can you submit it through our Forgejo instance at code.ffmpeg.org?
1. Environment - CPU: Kunpeng 950 (HiSilicon, 192 cores, SVE 256-bit)
I haven't heard about this CPU before - does it have SVE2, or is this only SVE1?
// Martin _______________________________________________ ffmpeg-devel mailing list -- [email protected] To unsubscribe send an email to [email protected]
