feat(moq-video): reintroduce VAAPI via discord/cros-codecs + NV12 surface upload (#1720)

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: kixelated

Cargo.lock

@@ -109,3 +109,9 @@ strip = true
 # doesn't have that feature yet, so consume a fork branch until it's merged.
 [patch.crates-io]
 nvidia-video-codec-sdk = { git = "https://github.com/kixelated/nvidia-video-codec-sdk.git", branch = "dynamic-loading" }
+# moq-video's `vaapi` feature uses discord/cros-codecs (`vaapi_dlopen`), which
+# pulls cros-libva 0.0.13. The `dlopen` feature it needs (load libva at runtime
+# so a vaapi-enabled binary still links without libva-dev and still loads on a
+# libva-less machine, falling back to software) lives only on discord's fork, so
+# point the published 0.0.13 at it. Drop once `dlopen` is upstreamed.
+cros-libva = { git = "https://github.com/discord/cros-libva.git", branch = "discord-0.0.13" }

Cargo.toml

@@ -70,4 +70,12 @@ allow-registry = ["https://github.com/rust-lang/crates.io-index"]
 # moq-video patches nvidia-video-codec-sdk to a fork that adds dlopen
 # (`dynamic-loading`); see the root Cargo.toml [patch.crates-io]. Drop this once
 # the feature is upstreamed and the patch is removed.
-allow-git = ["https://github.com/kixelated/nvidia-video-codec-sdk"]
+#
+# The `vaapi` feature pulls discord/cros-codecs (the maintained fork: cros-libva
+# 0.0.13 + a hardened H.264 VAAPI encoder) and patches cros-libva to discord's
+# fork for the `dlopen` feature. Drop these once `dlopen` is upstreamed.
+allow-git = [
+    "https://github.com/kixelated/nvidia-video-codec-sdk",
+    "https://github.com/discord/cros-codecs",
+    "https://github.com/discord/cros-libva",
+]

deny.toml

@@ -90,9 +90,9 @@ moq-cli publish --url https://relay.example.com --broadcast cam.hang capture --n
 
 Video capture uses a native per-platform backend (AVFoundation on macOS, V4L2 on
 Linux, Media Foundation on Windows) and picks a hardware H.264 encoder
-(VideoToolbox on macOS, NVENC on Linux NVIDIA) when one is present, falling back
-to the built-in software encoder (openh264); force either with `--hardware` /
-`--software`. `--camera` takes a bare integer as a device index, otherwise a
+(VideoToolbox on macOS, NVENC on Linux NVIDIA, VAAPI on Linux Intel/AMD) when one
+is present, falling back to the built-in software encoder (openh264); force either
+with `--hardware` / `--software`. `--camera` takes a bare integer as a device index, otherwise a
 device path (Linux) or name (a friendly-name substring on Windows, the
 AVFoundation `uniqueID` on macOS). Audio capture uses cpal (CoreAudio / WASAPI /
 ALSA) and encodes Opus.

doc/bin/cli.md

@@ -104,6 +104,10 @@
             # cpal's `alsa-sys` (moq-audio `capture` feature) links libasound on
             # Linux via pkg-config; macOS uses CoreAudio, so no dep there.
             pkgs.alsa-lib
+            # moq-video `vaapi` feature: cros-libva runs bindgen against the libva
+            # headers at build time (found via pkg-config), even though libva is
+            # dlopen'd at runtime (`vaapi_dlopen`). macOS has no VAAPI.
+            pkgs.libva
           ];
 
         # JavaScript dependencies

flake.nix

@@ -24,10 +24,24 @@ doctest = false
 # driver, falling back to software (see backend::open). `cuda-12020` pins the CUDA
 # API version so the build needs no CUDA toolkit.
 nvenc = ["dep:nvidia-video-codec-sdk", "nvidia-video-codec-sdk/dynamic-loading", "cudarc/fallback-dynamic-loading", "cudarc/cuda-12020"]
+# Intel/AMD VAAPI hardware encoder (Linux). Off by default; enable for a build
+# that should use VAAPI where present. Everything is dlopen'd at runtime:
+# discord/cros-codecs's `vaapi_dlopen` resolves libva through the cros-libva
+# `dlopen` feature (a fork feature, see the root [patch.crates-io]). So a
+# vaapi-enabled binary still links on a libva-less builder and still loads on
+# machines without libva, falling back to software (see backend::open). The
+# build still needs libva *headers* for bindgen (provided by the nix devShell).
+vaapi = ["dep:cros-codecs"]
 
 [dependencies]
 anyhow = "1"
 bytes = "1"
+# VAAPI H.264 encoder + zero-copy dmabuf import, gated behind `vaapi` + a linux
+# cfg. discord's fork is the maintained one: cros-libva 0.0.13 so it builds
+# against libva >= 2.23 (published cros-codecs pins the broken 0.0.12), plus a
+# hardened H.264 encoder (packed SPS/PPS + slice headers, frame_num, rate
+# control). `vaapi_dlopen` loads libva at runtime.
+cros-codecs = { git = "https://github.com/discord/cros-codecs.git", branch = "discord-0.0.5", optional = true, default-features = false, features = ["vaapi_dlopen"] }
 cudarc = { version = "0.19", optional = true, default-features = false, features = ["driver"] }
 moq-mux = { workspace = true }
 moq-net = { workspace = true }