Add GPU-accelerated AV1 encode/decode (VA-API, NVENC/NVDEC, Vulkan Video)

[staging-devin-ai-integration]

Summary

Add hardware-accelerated AV1 video encoding and decoding to StreamKit, leveraging GPU capabilities via VA-API, NVIDIA NVENC/NVDEC, and (future) Vulkan Video. This builds on the CPU AV1 foundation (#216) and the GPU compositing work (#194) to move toward a zero-copy GPU media pipeline.

Motivation

• CPU AV1 encoding (rav1e) at real-time speeds trades significant quality; HW encoders deliver better quality at the same latency
• GPU compositing (GPU-accelerated compositing via wgpu #194) already keeps frames on GPU — HW codecs reduce CPU↔GPU roundtrips
• AV1 has universal HW encode support: Intel (Arc+), AMD (RDNA 3+), NVIDIA (RTX 40xx+)
• The ultimate goal is a zero-copy pipeline: GPU decode → GPU composite → GPU encode with no CPU copies

Prerequisites

• Add CPU AV1 codec support (rav1e encoder + rav1d decoder) + codec negotiation refactor #216 — CPU AV1 support + codec negotiation refactor (provides the av1 feature, Av1EncoderNode/Av1DecoderNode interfaces, and MoQ codec negotiation)
• GPU-accelerated compositing via wgpu #194 — GPU compositing via wgpu (provides the GPU texture pipeline and wgpu device management)

HW Acceleration Options

Option A: VA-API via cros-codecs / cros-libva

Property	Detail
Crate	cros-codecs v0.0.6 (BSD-3-Clause)
AV1 decode	Supported (VA-API)
AV1 encode	Supported (VA-API)
VP9 decode/encode	Also supported
HW support	Intel, AMD, NVIDIA on Linux (via Mesa/intel-media-driver)
Maturity	Solid — developed for ChromeOS/crosvm, 546k downloads
wgpu interop	No native interop — requires DMA-BUF → Vulkan VkImage → wgpu texture bridge (custom work)

Option B: NVIDIA NVENC/NVDEC (native SDK bindings)

Property	Detail
SDK	NVIDIA Video Codec SDK
AV1 encode	Supported (RTX 40xx+, Ada Lovelace)
AV1 decode	Supported (RTX 30xx+, Ampere)
Rust bindings	Would need custom FFI bindings or use existing crates like nvidia-video-codec-sdk
HW support	NVIDIA GPUs only
Performance	Best-in-class for NVIDIA hardware, lower latency than VA-API on NVIDIA
wgpu interop	CUDA↔Vulkan interop is possible but complex

Option C: Vulkan Video via vk-video

Property	Detail
Crate	vk-video v0.2.1 (MIT)
AV1 support	Not yet — currently H.264 only
wgpu interop	Native! — built on wgpu, decoded frames are wgpu::Textures
Maturity	Very early (1.2k downloads), developed by Software Mansion for Smelter
Vulkan spec	VK_KHR_video_decode_av1 ratified Feb 2024, VK_KHR_video_encode_av1 ratified Nov 2024
Potential	Best long-term option for zero-copy pipeline, but needs AV1 support added

Proposed Phased Approach

Phase 2a: VA-API HW acceleration (most practical today)

// crates/nodes/src/video/vaapi.rs (new)

/// VA-API accelerated AV1 encoder using cros-codecs.
pub struct VaapiAv1EncoderNode { config: VaapiAv1EncoderConfig }
// Input: I420/NV12 raw frames
// Output: AV1 encoded packets
// Uses cros_codecs::encoder for VA-API AV1 encoding

/// VA-API accelerated AV1 decoder using cros-codecs.  
pub struct VaapiAv1DecoderNode { config: VaapiAv1DecoderConfig }
// Input: AV1 encoded packets
// Output: NV12 frames (VA surface → mapped to CPU buffer)

/// Runtime capability detection
pub fn probe_vaapi_capabilities() -> VaapiCapabilities { ... }
// Checks available VA-API profiles/entrypoints for AV1 encode/decode

• Feature: vaapi = ["dep:cros-codecs"]
• Fallback: auto-detect VA-API at runtime, fall back to CPU rav1e/rav1d if unavailable
• Works on Intel/AMD/NVIDIA on Linux
• Docker: requires --gpus flag + VA-API drivers

Phase 2b: NVIDIA NVENC/NVDEC (optional, for best NVIDIA performance)

• Feature: nvenc = ["dep:nvidia-video-codec-sdk"] (or custom FFI bindings)
• Only worth pursuing if VA-API performance on NVIDIA is insufficient
• NVENC typically has lower latency and better rate control than VA-API on NVIDIA

Phase 2c: Vulkan Video zero-copy pipeline (future, when ecosystem matures)

AV1 packets → Vulkan Video decode → wgpu::Texture (NV12)
  → GPU compositor (wgpu, already merged) → wgpu::Texture (output)
  → Vulkan Video encode → AV1 packets

• No CPU roundtrips — frames never leave GPU memory
• Requires vk-video crate to add AV1 support, OR building Vulkan Video AV1 directly using ash + wgpu Vulkan interop
• The wgpu Device and Queue are shared between compositor and codec
• This is the ultimate performance target

Tasks

• Evaluate VA-API AV1 encode/decode quality and latency with cros-codecs (PoC)
• Implement VaapiAv1EncoderNode and VaapiAv1DecoderNode
• Add runtime GPU capability detection and CPU fallback
• Evaluate NVIDIA NVENC/NVDEC bindings (if VA-API perf on NVIDIA is insufficient)
• Update Docker GPU images for VA-API/NVENC support
• Benchmark HW vs CPU encode/decode (latency, quality, throughput)
• Track vk-video AV1 progress for future zero-copy integration
• Update docs

Open Questions

1. VA-API vs NVENC priority — Start with VA-API (cross-vendor) or NVENC (best NVIDIA perf)?
2. DMA-BUF → wgpu bridge — Build custom interop for Phase 2a, or wait for Vulkan Video (Phase 2c)?
3. Auto-selection — Should the runtime automatically choose HW > CPU, or let the user configure?
4. Docker story — Separate Dockerfile.gpu-vaapi / Dockerfile.gpu-nvenc, or unified with runtime detection?