nx_vmvres.nx
buildroot/runtime/nx_vmvres.nx
about
nx_vmvres.nx -- efficient MV-RESIDUAL coding. After MV prediction (nx_vmvpred) most residual components are exactly 0
(neighbouring blocks share motion on a pan/head-pan), so paying the varint's fixed 5-bit length prefix on every
component is wasteful. Code instead: [1 bit nonzero-flag]; only if nonzero, the signed varint. A 0 residual costs 1
bit instead of 5 -> MVP's magnitude win finally shows up in the total bitrate. Composes nx_ventropy. license_tier: ORIGINAL
dependencies 1 imports · 1 importers
imports: nx_ventropy.nx
imported by: nx_vmvres_gate.nx
structs
| none |
consts
| none |
functions
| 8 | func mvr_put(buf: *u8, bp: i64, v: i64) -> i64 |
| 14 | func mvr_get(buf: *u8, bp: i64) -> i64 |
| 19 | func mvr_len(buf: *u8, bp: i64) -> i64 |