Repository navigation
Note
Microsoft Defender flags this release's libfxbridge.so as Trojan:Script/Wacatac.C!ml. This is a false positive, and it has been reported to Microsoft. It comes from Microsoft's machine-learning engine only, and no other engine on VirusTotal flags it. The binary rebuilds bit-identically from this tag, and the detection is most likely triggered by the new code that reads Resolve's functions to find its offsets. Details: v0.3.1.1.
If you do not trust it: use v0.3.1.1. It is the same bridge with fixed offsets for Resolve 21.1.2 only, and none of that code.
DaVinci Resolve 21.1.2 support. The bridge now reads the stock effect's member offsets out of Resolve's own instructions when it loads, instead of carrying them as constants.
Resolve 21.1.2 added eight bytes to AudioPlugin between +0x50 and +0x68, and every member of the stock effect past that point moved with it. The vtable, the map node size and the 21.1 plugin gate did not change, so tools/check-offsets.py on v0.3.0.1 reported 32 of 32 slots in place while every member read was eight bytes short.
| Member | 21.1.0 | 21.1.2 | Read out of |
|---|---|---|---|
| input / output channel counts | 0x150 / 0x158 | 0x158 / 0x160 | BMDAudioPluginImpl::UpdateChannelCount |
| bypass flag | 0x169 | 0x171 | BMDAudioPluginImpl::Bypass(bool) |
| dirty flag | 0x99 | 0xa1 | AudioPlugin::SetDirty |
| name record | 0x550 | 0x558 | BMDAudioPlugin<BMDStereoDelay>::GetEffectName |
| resource tree | 0x360 | 0x368 | BMDStereoDelay::GenerateUserInterface |
| panel width / height | 0x3a8 / 0x3ac | 0x3b0 / 0x3b4 | BMDStereoDelay::GenerateUserInterface |
| per-plugin lock / its flag | 0x218 / 0x250 | 0x220 / 0x258 | ResetHistory, PreProcess |
The fix
src/member_layout.cpp reads each member at load out of the exported function that owns it: an exact instruction form that must appear exactly once (twice for the panel size), sized from the function's own dynamic symbol entry, with adjacency and alignment cross-checks. If any member cannot be read with certainty, the bridge forwards Resolve's stock interface untouched — Resolve starts with its own effects and none of yours, instead of a bridge writing on a guess.
The bridge log (in ~/.local/share/DaVinciResolve/logs/ResolveDebug.txt) now names every offset it read:
[fxbridge] layout: channels +0x158/+0x160, bypass +0x171, dirty +0xa1, parent +0x40, name +0x558, panel tree +0x368 size +0x3b0/+0x3b4
This also works on 21.1.0 and earlier, which read their own numbers the same way.
Checking a new Resolve
tools/check-offsets.pydecodes the members by the same rules and prints what the bridge will read —ok, moved from …where a member moved,REFUSEDwhere the bridge would refuse.tools/member-layout-test.shruns the shipped decoder over a library file, so an installer can be checked before installing it.
Also
- Nix:
flake.nixwas missingsrc/chain_lock_fix.cppsince v0.3.0.1, which leaves a Nix-built bridge with an undefined symbol that Resolve'sdlopenrejects silently. Fixed; not built under Nix here.
Verified
On 21.1.0.0014 and 21.1.2.0010, from the library files and in memory; and live inside Resolve 21.1.2.0010 — plugins listed, instances named, channel counts 1 on mono and 2 on stereo plugins, panel, editors and project save all working. Not measured: the dirty-flag write (only with FXBRIDGE_STATE_STORE=1) and bypass by ear.
Full detail in CHANGELOG.md.
Updating
Close Resolve, then download the files into ~/.local/share/BMDAudioPlugins/ again — Step 1 of Method 2 in the readme fetches latest, so the same commands work. On 21.1 and later, refresh libfxbridge-gate.so and resolve-with-fxbridge too, and keep starting Resolve through resolve-with-fxbridge. config-fairlight.dat needs no change. SHA256SUMS lists every file.