An experimental Rust LD_PRELOAD hook for video in Discord's native Linux voice module. It uses NVIDIA NVDEC for received VP8, AV1, H.264, and H.265 frames and changes the module's NVIDIA encoder preference so its existing CUDA NVENC candidate can be selected for transmission. It does not add an encoder, change codec negotiation, or modify Discord's installed files.
This is an unofficial reverse-engineering project. The hook depends on parts of Discord's private C++ ABI. A Discord update can break it even when the required ELF symbols are still present. Review and test each updated voice module before relying on the patch.
The implementation was tested locally with Discord Linux host 1.0.166, discord_voice.node SHA-256 1a3fa65e21242ca84a27b6b654ba342c7ebf6c964fa3c1e1ee41682b9717fa24, NVIDIA driver 615.78.08, and a GeForce RTX 5080. Its WebRTC log listed VP8, AV1, H.264, and H.265 as supported decoders. Rust tests decoded synthetic samples of all four codecs, including H.265 Main10. A separate harness loaded the actual voice module and received decoded callbacks for all four codecs. Earlier HEVC harness runs also decoded one 4096×2304 Main frame. The harness confirmed that the CUDA NVENC preference patch matched the module's instructions.
Live observations: An AV1X broadcast on the earlier build selected nvdec but returned cuvidParseVideoData failed: 999, decoded zero frames, and repeatedly requested keyframes. A frame-by-frame AV1 test reproduced the same error when the final frame OBU omitted its size field, as in the received packets. The current build restores that size field before passing AV1 data to NVDEC and marks each WebRTC encoded frame as an end-of-picture packet. In a live AV1X broadcast after the fix, Discord reported AV1X (nvdec), 601 decoded frames at 60 fps, zero dropped frames, and 1410×860 resolution. An outgoing AV1 broadcast was observed as AV1 (nvidia: cuda) with encoded frames after its send stream became active.
Not yet validated: packet loss, latency, stream renegotiation, and behavior after a Discord update.
| Path | Behavior |
|---|---|
| Video receive | Intercepts VideoDecoderFactory::Create for VP8, AV1, H.264, and H.265. The GPU's NVDEC capability is checked for each codec before installing the custom decoder. Decoded surfaces are copied to host memory, converted from NV12 to I420, and delivered through the module's WebRTC callback. |
| Video transmit | Skips the module's NVIDIA Vulkan encoder candidate insertion so the existing NVIDIA CUDA candidate remains available. The module chooses the final encoder. |
| Unsupported codecs or GPUs | Delegates decoder creation to the original factory. |
The decoder currently accepts progressive 4:2:0 video within the GPU's reported limits. Higher-bit-depth output is reduced to 8-bit I420. Copying frames to host memory and converting them adds CPU work; this is not a zero-copy decoder. Codec support varies by GPU and driver; the hook checks NVDEC capabilities at runtime.
The hook reads the voice module's local ELF symbol table at load time and checks the decoder factory prologue and the encoder insertion call. It does not use fixed function offsets or a fixed Discord version path. If a required symbol or checked instruction sequence is missing, the hook skips the module and prints a reason to standard error. Stripped voice modules are unsupported. A matching symbol and instruction sequence does not verify the EncodedImage layout, decoder vtable, callback ABI, or other private structures.
Rust owns the decoder state and frame buffers. Calls into CUDA and Discord, vtable access, and instruction patching use unsafe. The private ABI prevents a general memory-safety guarantee across Discord releases.
- Linux x86-64. The hook reads
/proc/self/maps, parses ELF64, and writes x86-64 jump instructions. - An NVIDIA GPU and driver providing
libcuda.so.1andlibnvcuvid.so.1at runtime. Building also needs linkablelibcuda.soandlibnvcuvid.sodevelopment symlinks. - Rust 1.99.1 and Cargo were used for validation. Earlier toolchains are unverified. The repository has no Cargo dependencies.
- Discord's user-installed executable at
${XDG_CONFIG_HOME:-$HOME/.config}/discord/Discord. The voice module must retain the ELF symbols used by the hook. - Bash and standard Linux utilities used by the build and installer scripts. Desktop installation also expects the system entry at
/usr/share/applications/discord.desktop.
The repository includes generated FFI bindings and the nv-codec-headers declarations they came from. Building does not require proprietary CUDA Toolkit headers or bindgen. Bindgen CLI 0.72.1 is needed only to regenerate src/ffi.rs.
From the repository root:
./build.shThis runs cargo build --release --offline --locked and copies the result to ./libdiscord_nvenc_nvdec.so. The library is a local build artifact and is ignored by Git. If linking fails with cannot find -lcuda or cannot find -lnvcuvid, check that the development symlinks are installed and visible to the linker. The runtime driver libraries alone do not satisfy those link names.
To build without copying the result to the repository root:
cargo build --release --offline --lockedFor one launch from a terminal, after building:
./launch.shThe launcher sets LD_PRELOAD for the user-installed Discord executable and passes through any arguments. Close an already running Discord instance first. Opening a second instance can hand off to the existing process, which would not have the hook loaded.
To use the patch from the desktop menu:
./install.sh installThe installer copies the library and launcher into ${XDG_DATA_HOME:-$HOME/.local/share}/discord-nvenc-nvdec-patch and writes a user-specific ${XDG_DATA_HOME:-$HOME/.local/share}/applications/discord.desktop. It backs up an existing user-specific Discord desktop entry before replacing it. It does not edit /usr/share/discord, /usr/bin/discord, or the voice module. Installation does not change a running Discord process. Launch Discord from the desktop entry after closing the old process.
A direct /usr/bin/discord launch bypasses the user desktop entry. Use ./launch.sh for a terminal launch with the hook.
To remove the desktop integration and installed hook files:
./install.sh uninstallThe uninstall command restores the previous user-specific desktop entry when the installer backed one up. It refuses to replace a desktop entry whose installer marker has since been removed or changed. Running Discord processes are not stopped by either command.
Run ./launch.sh in a terminal to read messages prefixed discord-nvdec:. CUDA NVENC preference patch installed means the instruction pattern matched and was patched. <codec> selected for NVDEC decoding means the custom decoder was created for that codec. Neither message alone confirms successful live video receive or outgoing hardware encoding. Confirm those paths with actual broadcasts and inspect the resulting video behavior.
For AV1 parser diagnosis, set DISCORD_NVDEC_TRACE=1 before launching. The hook logs the first 16 encoded bytes of up to four AV1 packets per decoder and rejected sequence metadata. The byte prefixes are derived from received video and should be removed from logs before sharing them publicly.
When Discord updates, record the new host version and voice-module SHA-256:
discord_host=$(readlink -f "${XDG_CONFIG_HOME:-$HOME/.config}/discord/Discord")
sha256sum "${discord_host%/*}"/modules/discord_voice-*/discord_voice/discord_voice.nodeThen verify that the hook still finds its symbols and patterns. Test incoming VP8, AV1, H.264, and H.265 video as well as outgoing video before treating the new module as compatible. If the module is stripped or a pattern changes, the hook leaves it unpatched. If its private object layouts change while the patterns remain, a crash or incorrect frame handling is possible. Remove the desktop integration with ./install.sh uninstall if the updated module is not validated.
cargo fmt --check
cargo clippy --release --offline --locked -- -D warnings
cargo test --release --offline --locked -- --test-threads=1The NVDEC tests need a working NVIDIA GPU and driver. tests/fixtures contains locally generated H.265, H.264, AV1, and VP8 test patterns. AV1 is tested as separate frame packets, a WebRTC-style packet without the final OBU size field, and a batched stream. To regenerate the committed bindings from the bundled headers:
./scripts/generate-bindings.shThat development command requires bindgen CLI 0.72.1 and rustfmt. It preserves the provenance pointer at the top of src/ffi.rs.
The project's original code is under the MIT License. The three bundled files under include/ffnvcodec/ retain their own copyright and permission notices. src/ffi.rs was generated from those headers. Their exact source commit, file-level copyright lines, common permission and warranty text, and the relationship to the generated bindings are recorded in Third-party notices. The project's MIT license entry in Cargo.toml does not replace those header notices.
For a source distribution, keep LICENSE, THIRD_PARTY_NOTICES.md, and the notices embedded in the bundled headers. For a binary distribution that omits the headers, provide LICENSE and THIRD_PARTY_NOTICES.md alongside the library. build.sh produces a library, not a release archive; it does not assemble those files for you.
Discord binaries, NVIDIA driver libraries, the proprietary CUDA Toolkit headers, FFmpeg, libx265, and bindgen are not included in this repository. Users supply the runtime libraries installed on their own systems. This project is not affiliated with Discord or NVIDIA.