feat(host/encode): VAAPI zero-copy dmabuf import (AMD/Intel GPU CSC)

Phase 2 of AMD/Intel support: the VAAPI encoder now takes the capture dmabuf directly and does the RGB->NV12 colour conversion on the GPU's video engine, eliminating the host-side de-pad + swscale CSC + upload the CPU path pays. - capture: a vendor-neutral FramePayload::Dmabuf (dup'd fd + fourcc/modifier/ layout). When zero-copy is on, the EGL->CUDA importer is unavailable (any non-NVIDIA host), and the backend is VAAPI, the capturer advertises LINEAR dmabuf and hands the raw buffer to the encoder instead of CPU-copying it. - encode/vaapi: the encoder self-configures from the first frame's payload (no open_video signature change). The dmabuf arm wraps the buffer as an AV_PIX_FMT_DRM_PRIME frame and pushes it through a filter graph buffer(drm_prime) -> hwmap(vaapi) -> scale_vaapi=nv12 -> buffersink; the encoder takes NV12 surfaces straight from the sink. The Phase 1 CPU-upload path is kept as the other arm (used when capture produces CPU frames). Live-validated on a Radeon 780M (real Sway/xdpw desktop capture): correct, pixel-perfect HEVC, and ~10x less host CPU at 1440p (4.2s -> 0.4s of CPU for 300 frames) -- the de-pad/CSC/upload moves to the GPU. NVIDIA unchanged (zero-copy still imports to CUDA; the passthrough path only engages on non-NVIDIA hosts). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-06-20 09:57:00 +00:00
parent 5e27f65f2e
commit 708c62788d
5 changed files with 696 additions and 237 deletions
@@ -4,27 +4,30 @@
 //! sibling of [`super::linux`] (NVENC/CUDA) behind the shared [`Encoder`] trait — selected in
 //! [`super::open_video`] (NVIDIA → NVENC, AMD/Intel → here).
 //!
-//! Two input paths:
-//! * **CPU (this file today).** The portal negotiates packed RGB/BGR; we swscale it to BT.709
-//!   limited-range NV12, upload that into a pooled VA surface (`av_hwframe_transfer_data`), and
-//!   encode in place. Robust on any VAAPI GPU with no capture-side changes — the capturer already
-//!   falls back to CPU frames on a non-NVIDIA box (its EGL→CUDA importer needs `libcuda`).
-//! * **Zero-copy dmabuf (deferred to Phase 2).** Import the capture dmabuf straight into a VA
-//!   surface (`av_hwframe_map` of an `AV_PIX_FMT_DRM_PRIME` frame) — no EGL/Vulkan/CUDA detour,
-//!   no host CSC. This is the inverse of the Linux client's VAAPI *decode* path.
+//! Two input paths, chosen lazily from the FIRST frame's payload (so `open_video`'s signature
+//! is unchanged and the encoder self-configures for whatever the capturer produces):
+//! * **CPU upload** ([`CpuInner`]): the portal hands packed RGB/BGR CPU frames; we swscale to
+//!   BT.709-limited NV12 and `av_hwframe_transfer_data` it into a pooled VA surface. Works on any
+//!   VAAPI GPU with no capture changes (the capturer falls back to CPU frames on non-NVIDIA).
+//! * **Zero-copy dmabuf** ([`DmabufInner`], `PUNKTFUNK_ZEROCOPY=1`): the capturer hands a packed-RGB
+//!   dmabuf. We wrap it as an `AV_PIX_FMT_DRM_PRIME` frame and push it through a tiny filter graph
+//!   `buffer(drm_prime) → hwmap=derive_device=vaapi → scale_vaapi=format=nv12 → buffersink`, so
+//!   the import AND the RGB→NV12 colour conversion run on the GPU's video engine — no host CSC, no
+//!   upload. The encoder takes the NV12 surfaces straight from the filter sink.
 //!
-//! Raw FFI: `ffmpeg-next` has no hwcontext wrappers, so the hwdevice/hwframes/transfer calls go
-//! through `ffmpeg::ffi` (= `ffmpeg_sys_next`), exactly as the CUDA encode path and the clients'
-//! decode paths already do. The encoder is opened *without* a global header, so VPS/SPS/PPS are
-//! in-band on every IDR.
+//! Raw FFI: `ffmpeg-next` has no hwcontext/filter wrappers for what we need, so the
+//! hwdevice/hwframes/buffersrc/buffersink calls go through `ffmpeg::ffi` (= `ffmpeg_sys_next`),
+//! as the CUDA encode path and the clients' decode paths already do. The encoder is opened
+//! *without* a global header, so VPS/SPS/PPS are in-band on every IDR.

 use super::{Codec, EncodedFrame, Encoder};
-use crate::capture::{CapturedFrame, FramePayload, PixelFormat};
+use crate::capture::{CapturedFrame, DmabufFrame, FramePayload, PixelFormat};
 use anyhow::{anyhow, bail, Context, Result};
 use ffmpeg::format::Pixel;
 use ffmpeg::{codec, encoder, Dictionary, Packet, Rational};
 use ffmpeg_next as ffmpeg;
 use std::ffi::{CStr, CString};
+use std::os::fd::AsRawFd;
 use std::os::raw::c_int;
 use std::ptr;

@@ -40,10 +43,19 @@ fn pixel_to_av(p: Pixel) -> ffi::AVPixelFormat {
    ffi::AVPixelFormat::from(p)
 }

-/// The swscale *source* pixel format for a captured CPU layout. The portal fixates packed
-/// 24/32-bit RGB/BGR; swscale converts any of these → NV12 directly (it even takes 3-bpp RGB24
-/// with no host-side 3→4 expand, unlike NVENC). NV12/P010/HDR only arrive on Windows or the
-/// deferred 10-bit path, so reject them here with a clear message.
+/// `fourcc(a,b,c,d)` — DRM FourCC packing (`a | b<<8 | c<<16 | d<<24`).
+const fn fourcc(a: u8, b: u8, c: u8, d: u8) -> u32 {
+    (a as u32) | ((b as u32) << 8) | ((c as u32) << 16) | ((d as u32) << 24)
+}
+
+/// The render node a VAAPI/DRM device should open. `PUNKTFUNK_RENDER_NODE` pins it on a multi-GPU
+/// box; the default is correct on a single-GPU host.
+fn render_node() -> CString {
+    let p = std::env::var("PUNKTFUNK_RENDER_NODE").unwrap_or_else(|_| "/dev/dri/renderD128".into());
+    CString::new(p).unwrap_or_else(|_| CString::new("/dev/dri/renderD128").unwrap())
+}
+
+/// The swscale *source* pixel format for a captured CPU layout (packed RGB/BGR only).
 fn vaapi_sws_src(format: PixelFormat) -> Result<Pixel> {
    Ok(match format {
        PixelFormat::Bgrx => Pixel::BGRZ, // bgr0
@@ -52,48 +64,115 @@ fn vaapi_sws_src(format: PixelFormat) -> Result<Pixel> {
        PixelFormat::Rgba => Pixel::RGBA,
        PixelFormat::Rgb => Pixel::RGB24,
        PixelFormat::Bgr => Pixel::BGR24,
-        PixelFormat::Nv12 | PixelFormat::P010 | PixelFormat::Rgb10a2 => bail!(
-            "VAAPI CPU-input path supports packed RGB/BGR only; got {format:?} \
-             (NV12/P010/HDR arrive only on the Windows or deferred 10-bit paths)"
-        ),
+        PixelFormat::Nv12 | PixelFormat::P010 | PixelFormat::Rgb10a2 => {
+            bail!("VAAPI CPU-input path supports packed RGB/BGR only; got {format:?}")
+        }
    })
 }

-/// VAAPI hardware contexts: a device created on a DRM render node and a frames pool the encoder
-/// draws input surfaces from. Owns two `AVBufferRef`s, unref'd on drop (refcounted, so the copies
-/// we hand the encoder outlive this).
+/// Build the FFmpeg encoder context (shared by both inner paths): name, mode, low-latency RC,
+/// infinite GOP, BT.709-limited VUI, `pix_fmt=VAAPI`, and the given hw device + frames contexts.
+/// Returns the opened encoder. `device_ref`/`frames_ref` are borrowed (ref'd into the context).
+unsafe fn open_vaapi_encoder(
+    codec: Codec,
+    width: u32,
+    height: u32,
+    fps: u32,
+    bitrate_bps: u64,
+    device_ref: *mut ffi::AVBufferRef,
+    frames_ref: *mut ffi::AVBufferRef,
+) -> Result<encoder::video::Encoder> {
+    let name = codec.vaapi_name();
+    let av_codec = encoder::find_by_name(name).ok_or_else(|| {
+        anyhow!("{name} not built into libavcodec (no VAAPI encoder for {codec:?})")
+    })?;
+    let mut video = codec::context::Context::new_with_codec(av_codec)
+        .encoder()
+        .video()
+        .context("alloc video encoder")?;
+    video.set_width(width);
+    video.set_height(height);
+    video.set_format(Pixel::NV12); // sw view; pix_fmt overridden to VAAPI below
+    video.set_time_base(Rational(1, fps as i32));
+    video.set_frame_rate(Some(Rational(fps as i32, 1)));
+    video.set_bit_rate(bitrate_bps as usize);
+    video.set_max_bit_rate(bitrate_bps as usize); // == target → vaapi_encode picks CBR when supported
+    let vbv_frames = std::env::var("PUNKTFUNK_VBV_FRAMES")
+        .ok()
+        .and_then(|s| s.parse::<f32>().ok())
+        .filter(|v| v.is_finite() && *v > 0.0)
+        .unwrap_or(1.0);
+    let vbv_bits =
+        ((bitrate_bps as f64 / fps.max(1) as f64) * vbv_frames as f64).clamp(1.0, i32::MAX as f64);
+    video.set_max_b_frames(0);
+    let raw = video.as_mut_ptr();
+    (*raw).rc_buffer_size = vbv_bits as i32;
+    (*raw).gop_size = i32::MAX; // no periodic IDR (forced-IDR via pict_type=I on RFI)
+                                // We hand the encoder BT.709 *limited* NV12 (swscale CSC, or scale_vaapi which preserves the
+                                // input range we tag), so signal that VUI — else the client decoder washes the picture out.
+    (*raw).colorspace = ffi::AVColorSpace::AVCOL_SPC_BT709;
+    (*raw).color_range = ffi::AVColorRange::AVCOL_RANGE_MPEG;
+    (*raw).color_primaries = ffi::AVColorPrimaries::AVCOL_PRI_BT709;
+    (*raw).color_trc = ffi::AVColorTransferCharacteristic::AVCOL_TRC_BT709;
+    (*raw).pix_fmt = ffi::AVPixelFormat::AV_PIX_FMT_VAAPI;
+    (*raw).hw_device_ctx = ffi::av_buffer_ref(device_ref);
+    (*raw).hw_frames_ctx = ffi::av_buffer_ref(frames_ref);
+
+    let mut opts = Dictionary::new();
+    opts.set("async_depth", "1"); // one-in/one-out — minimal encode-pipeline latency
+    video
+        .open_with(opts)
+        .with_context(|| format!("open {name} ({width}x{height}@{fps}, {bitrate_bps} bps)"))
+}
+
+/// Drain the encoder for one packet (shared poll logic).
+fn poll_encoder(enc: &mut encoder::video::Encoder, fps: u32) -> Result<Option<EncodedFrame>> {
+    let mut pkt = Packet::empty();
+    match enc.receive_packet(&mut pkt) {
+        Ok(()) => {
+            let data = pkt.data().map(|d| d.to_vec()).unwrap_or_default();
+            let pts = pkt.pts().unwrap_or(0).max(0) as u64;
+            Ok(Some(EncodedFrame {
+                data,
+                pts_ns: pts * 1_000_000_000 / fps as u64,
+                keyframe: pkt.is_key(),
+            }))
+        }
+        Err(ffmpeg::Error::Other { errno })
+            if errno == ffmpeg::util::error::EAGAIN
+                || errno == ffmpeg::util::error::EWOULDBLOCK =>
+        {
+            Ok(None)
+        }
+        Err(ffmpeg::Error::Eof) => Ok(None),
+        Err(e) => Err(e).context("receive_packet"),
+    }
+}
+
+// ---------------------------------------------------------------------------------------------
+// CPU upload path (Phase 1): swscale RGB→NV12 → upload into a pooled VA surface → encode.
+// ---------------------------------------------------------------------------------------------
+
+/// VAAPI device + NV12 frames pool (the encoder's input surfaces for the CPU path).
 struct VaapiHw {
    device_ref: *mut ffi::AVBufferRef,
    frames_ref: *mut ffi::AVBufferRef,
 }

 impl VaapiHw {
-    /// Create a VAAPI device (`node` = e.g. `/dev/dri/renderD128`, or `None` for libva's default
-    /// — correct on a single-GPU box) and an `AV_PIX_FMT_VAAPI` frames pool with `sw_format`.
-    unsafe fn new(
-        node: Option<&CStr>,
-        sw_format: ffi::AVPixelFormat,
-        w: u32,
-        h: u32,
-        pool: c_int,
-    ) -> Result<Self> {
+    unsafe fn new(sw_format: ffi::AVPixelFormat, w: u32, h: u32, pool: c_int) -> Result<Self> {
        let mut device_ref: *mut ffi::AVBufferRef = ptr::null_mut();
-        let node_ptr = node.map_or(ptr::null(), |c| c.as_ptr());
+        let node = render_node();
        let r = ffi::av_hwdevice_ctx_create(
            &mut device_ref,
            ffi::AVHWDeviceType::AV_HWDEVICE_TYPE_VAAPI,
-            node_ptr,
+            node.as_ptr(),
            ptr::null_mut(),
            0,
        );
        if r < 0 {
-            let where_ = node
-                .and_then(|c| c.to_str().ok())
-                .map(|s| format!(" ({s})"))
-                .unwrap_or_default();
-            bail!("no VAAPI device{where_}: {}", ffmpeg::Error::from(r));
+            bail!("no VAAPI device ({:?}): {}", node, ffmpeg::Error::from(r));
        }
-
        let mut frames_ref = ffi::av_hwframe_ctx_alloc(device_ref);
        if frames_ref.is_null() {
            ffi::av_buffer_unref(&mut device_ref);
@@ -127,125 +206,40 @@ impl Drop for VaapiHw {
    }
 }

-pub struct VaapiEncoder {
+struct CpuInner {
    enc: encoder::video::Encoder,
    hw: VaapiHw,
-    /// swscale context: packed RGB/BGR → NV12 (BT.709 limited). CPU-input path only.
    sws: *mut ffi::SwsContext,
-    /// Reusable software NV12 staging frame (swscale dst → `av_hwframe_transfer_data` src).
-    /// Overwriting it across frames is sound: the upload copies into a fresh pooled VA surface and
-    /// the caller drains `poll()` after each `submit`, so nothing holds a reference to it.
-    nv12: *mut ffi::AVFrame,
+    nv12: *mut ffi::AVFrame, // reusable software NV12 staging frame (swscale dst → upload src)
    src_format: PixelFormat,
    width: u32,
    height: u32,
-    fps: u32,
-    /// Monotonic presentation index, in `1/fps` time-base units.
-    frame_idx: i64,
-    /// Force the next submitted frame to be an IDR (set by [`request_keyframe`]).
-    force_kf: bool,
 }

-// Raw FFI pointers; the encoder lives on a single thread (same contract as `NvencEncoder`).
-unsafe impl Send for VaapiEncoder {}
-
-impl VaapiEncoder {
-    pub fn open(
+impl CpuInner {
+    fn open(
        codec: Codec,
        format: PixelFormat,
        width: u32,
        height: u32,
        fps: u32,
        bitrate_bps: u64,
-        bit_depth: u8,
    ) -> Result<Self> {
-        // 10-bit/HDR (P010 sw_format) is a follow-up — VAAPI supports it cleanly via Main10, but
-        // it needs the capture/negotiation 10-bit plumbing that the Linux host doesn't have yet.
-        if bit_depth != 8 {
-            tracing::warn!(bit_depth, "VAAPI 10-bit not yet wired — encoding 8-bit");
-        }
-        ffmpeg::init().context("ffmpeg init")?;
-        if std::env::var_os("PUNKTFUNK_FFMPEG_DEBUG").is_some() {
-            unsafe { ffi::av_log_set_level(48) }; // AV_LOG_DEBUG — surface VAAPI open/upload rejects
-        }
-        let name = codec.vaapi_name();
-        let av_codec = encoder::find_by_name(name).ok_or_else(|| {
-            anyhow!("{name} not built into libavcodec (no VAAPI encoder for {codec:?})")
-        })?;
        let src_pixel = vaapi_sws_src(format)?;
-
-        // VAAPI device + NV12 frames pool. `PUNKTFUNK_RENDER_NODE` pins the GPU on a multi-GPU box;
-        // unset = libva's default render node (right on a single-GPU host).
-        let node = std::env::var("PUNKTFUNK_RENDER_NODE").ok();
-        let node_c = node
-            .as_deref()
-            .map(CString::new)
-            .transpose()
-            .context("PUNKTFUNK_RENDER_NODE contained a NUL")?;
        const POOL: c_int = 16;
-        let hw = unsafe {
-            VaapiHw::new(
-                node_c.as_deref(),
-                ffi::AVPixelFormat::AV_PIX_FMT_NV12,
+        let hw = unsafe { VaapiHw::new(ffi::AVPixelFormat::AV_PIX_FMT_NV12, width, height, POOL)? };
+        let enc = unsafe {
+            open_vaapi_encoder(
+                codec,
                width,
                height,
-                POOL,
+                fps,
+                bitrate_bps,
+                hw.device_ref,
+                hw.frames_ref,
            )?
        };
-
-        let mut video = codec::context::Context::new_with_codec(av_codec)
-            .encoder()
-            .video()
-            .context("alloc video encoder")?;
-        video.set_width(width);
-        video.set_height(height);
-        video.set_format(Pixel::NV12); // sw_format; pix_fmt is overridden to VAAPI below
-        video.set_time_base(Rational(1, fps as i32));
-        video.set_frame_rate(Some(Rational(fps as i32, 1)));
-        video.set_bit_rate(bitrate_bps as usize);
-        // max == target so vaapi_encode selects CBR when the driver's RC entrypoint supports it
-        // (modern AMD/Intel), and gracefully degrades to VBR otherwise — without failing to open.
-        video.set_max_bit_rate(bitrate_bps as usize);
-        // VBV/HRD ~1 frame of bits — same rationale as NVENC: keep per-frame size roughly constant
-        // so a high-motion P-frame can't balloon past the bounded send queue. PUNKTFUNK_VBV_FRAMES
-        // tunes it (shared knob with NVENC).
-        let vbv_frames = std::env::var("PUNKTFUNK_VBV_FRAMES")
-            .ok()
-            .and_then(|s| s.parse::<f32>().ok())
-            .filter(|v| v.is_finite() && *v > 0.0)
-            .unwrap_or(1.0);
-        let vbv_bits = ((bitrate_bps as f64 / fps.max(1) as f64) * vbv_frames as f64)
-            .clamp(1.0, i32::MAX as f64);
-        video.set_max_b_frames(0);
-        unsafe {
-            let raw = video.as_mut_ptr();
-            (*raw).rc_buffer_size = vbv_bits as i32;
-            // Infinite GOP — no periodic IDR (the "freeze" fix). VAAPI has no NVENC `gop_size=-1`,
-            // so use a huge GOP and drive keyframes on demand via forced IDR (pict_type=I), the
-            // same Moonlight/Sunshine low-latency model.
-            (*raw).gop_size = i32::MAX;
-            // We CSC RGB→NV12 as BT.709 *limited* range in swscale (below), so signal that VUI —
-            // otherwise the client decoder assumes a default and the picture is washed-out / wrong
-            // contrast. Matches the NVENC NV12 path's signalling.
-            (*raw).colorspace = ffi::AVColorSpace::AVCOL_SPC_BT709;
-            (*raw).color_range = ffi::AVColorRange::AVCOL_RANGE_MPEG; // limited/studio
-            (*raw).color_primaries = ffi::AVColorPrimaries::AVCOL_PRI_BT709;
-            (*raw).color_trc = ffi::AVColorTransferCharacteristic::AVCOL_TRC_BT709;
-            // Take VAAPI hw surfaces: derive the device from the frames pool, set both before open.
-            (*raw).pix_fmt = ffi::AVPixelFormat::AV_PIX_FMT_VAAPI;
-            (*raw).hw_device_ctx = ffi::av_buffer_ref(hw.device_ref);
-            (*raw).hw_frames_ctx = ffi::av_buffer_ref(hw.frames_ref);
-        }
-
-        let mut opts = Dictionary::new();
-        opts.set("async_depth", "1"); // one-in/one-out — minimal encode-pipeline latency
-
-        let enc = video
-            .open_with(opts)
-            .with_context(|| format!("open {name} ({width}x{height}@{fps}, {bitrate_bps} bps)"))?;
-
-        // swscale: packed RGB/BGR → NV12, no rescale (POINT). Force BT.709 limited so the bytes
-        // match the VUI we signalled.
+        // swscale RGB→NV12, BT.709 limited (matches the VUI), no rescale.
        let src_av = pixel_to_av(src_pixel);
        let sws = unsafe {
            ffi::sws_getContext(
@@ -265,12 +259,9 @@ impl VaapiEncoder {
            bail!("sws_getContext(RGB→NV12) failed");
        }
        unsafe {
-            // src RGB = full range (1), dst YUV = limited/studio (0); BT.709 coefficients both sides.
            let cs709 = ffi::sws_getCoefficients(SWS_CS_ITU709);
            ffi::sws_setColorspaceDetails(sws, cs709, 1, cs709, 0, 0, 1 << 16, 1 << 16);
        }
-
-        // Reusable software NV12 staging frame.
        let nv12 = unsafe {
            let f = ffi::av_frame_alloc();
            if f.is_null() {
@@ -280,22 +271,19 @@ impl VaapiEncoder {
            (*f).format = ffi::AVPixelFormat::AV_PIX_FMT_NV12 as c_int;
            (*f).width = width as c_int;
            (*f).height = height as c_int;
-            let r = ffi::av_frame_get_buffer(f, 0);
-            if r < 0 {
+            if ffi::av_frame_get_buffer(f, 0) < 0 {
                let mut f = f;
                ffi::av_frame_free(&mut f);
                ffi::sws_freeContext(sws);
-                bail!("av_frame_get_buffer(NV12) failed ({r})");
+                bail!("av_frame_get_buffer(NV12) failed");
            }
            f
        };
-
        tracing::info!(
-            encoder = name,
-            render_node = node.as_deref().unwrap_or("default"),
+            encoder = codec.vaapi_name(),
            "VAAPI encode active ({width}x{height}@{fps}, CPU→NV12 upload path)"
        );
-        Ok(VaapiEncoder {
+        Ok(CpuInner {
            enc,
            hw,
            sws,
@@ -303,34 +291,23 @@ impl VaapiEncoder {
            src_format: format,
            width,
            height,
-            fps,
-            frame_idx: 0,
-            force_kf: false,
        })
    }

-    /// CPU path: swscale the packed RGB/BGR bytes into the reusable NV12 frame, upload that into a
-    /// pooled VA surface, and encode in place.
-    fn submit_cpu(&mut self, bytes: &[u8], format: PixelFormat, pts: i64, idr: bool) -> Result<()> {
+    fn submit(&mut self, bytes: &[u8], format: PixelFormat, pts: i64, idr: bool) -> Result<()> {
        anyhow::ensure!(
            format == self.src_format,
-            "captured format {:?} != encoder source {:?}",
-            format,
+            "captured format {format:?} != encoder source {:?}",
            self.src_format
        );
        let w = self.width as usize;
        let h = self.height as usize;
        let src_row = w * self.src_format.bytes_per_pixel();
-        anyhow::ensure!(
-            bytes.len() >= src_row * h,
-            "captured buffer {} bytes < required {}",
-            bytes.len(),
-            src_row * h
-        );
+        anyhow::ensure!(bytes.len() >= src_row * h, "captured buffer too small");
        unsafe {
            let src_data: [*const u8; 4] = [bytes.as_ptr(), ptr::null(), ptr::null(), ptr::null()];
            let src_stride: [c_int; 4] = [src_row as c_int, 0, 0, 0];
-            let r = ffi::sws_scale(
+            if ffi::sws_scale(
                self.sws,
                src_data.as_ptr(),
                src_stride.as_ptr(),
@@ -338,26 +315,21 @@ impl VaapiEncoder {
                h as c_int,
                (*self.nv12).data.as_ptr(),
                (*self.nv12).linesize.as_ptr(),
-            );
-            if r < 0 {
-                bail!("sws_scale RGB→NV12 failed ({r})");
+            ) < 0
+            {
+                bail!("sws_scale RGB→NV12 failed");
            }
-
-            // Pooled VA surface ← NV12 upload, then encode in place. Free the frame after send;
-            // avcodec_send_frame takes its own ref to the surface.
            let mut hwf = ffi::av_frame_alloc();
            if hwf.is_null() {
                bail!("av_frame_alloc(hw) failed");
            }
-            let r = ffi::av_hwframe_get_buffer(self.hw.frames_ref, hwf, 0);
-            if r < 0 {
+            if ffi::av_hwframe_get_buffer(self.hw.frames_ref, hwf, 0) < 0 {
                ffi::av_frame_free(&mut hwf);
-                bail!("av_hwframe_get_buffer(VAAPI) failed ({r})");
+                bail!("av_hwframe_get_buffer(VAAPI) failed");
            }
-            let r = ffi::av_hwframe_transfer_data(hwf, self.nv12, 0);
-            if r < 0 {
+            if ffi::av_hwframe_transfer_data(hwf, self.nv12, 0) < 0 {
                ffi::av_frame_free(&mut hwf);
-                bail!("av_hwframe_transfer_data(→VAAPI) failed ({r})");
+                bail!("av_hwframe_transfer_data(→VAAPI) failed");
            }
            (*hwf).pts = pts;
            (*hwf).pict_type = if idr {
@@ -375,6 +347,398 @@ impl VaapiEncoder {
    }
 }

+impl Drop for CpuInner {
+    fn drop(&mut self) {
+        unsafe {
+            if !self.nv12.is_null() {
+                ffi::av_frame_free(&mut self.nv12);
+            }
+            if !self.sws.is_null() {
+                ffi::sws_freeContext(self.sws);
+            }
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------------------------
+// Zero-copy dmabuf path: DRM-PRIME → hwmap(vaapi) → scale_vaapi(nv12) filter graph → encode.
+// ---------------------------------------------------------------------------------------------
+
+struct DmabufInner {
+    enc: encoder::video::Encoder,
+    /// DRM device the source dmabuf frames reference (the buffersrc's `hw_frames_ctx` device).
+    drm_device: *mut ffi::AVBufferRef,
+    /// VAAPI device driving `hwmap`/`scale_vaapi`/the encoder.
+    vaapi_device: *mut ffi::AVBufferRef,
+    /// DRM-PRIME frames context for the imported dmabufs (buffersrc input).
+    drm_frames: *mut ffi::AVBufferRef,
+    graph: *mut ffi::AVFilterGraph,
+    src: *mut ffi::AVFilterContext,
+    sink: *mut ffi::AVFilterContext,
+    width: u32,
+    height: u32,
+    fourcc: u32,
+}
+
+impl DmabufInner {
+    fn open(
+        codec: Codec,
+        format: PixelFormat,
+        width: u32,
+        height: u32,
+        fps: u32,
+        bitrate_bps: u64,
+    ) -> Result<Self> {
+        let drm_fourcc = crate::zerocopy::drm_fourcc(format)
+            .ok_or_else(|| anyhow!("no DRM fourcc for {format:?} (VAAPI zero-copy)"))?;
+        let node = render_node();
+        unsafe {
+            // DRM device (source dmabuf frames) + a VAAPI device derived from it (same GPU) for
+            // hwmap/scale_vaapi/the encoder.
+            let mut drm_device: *mut ffi::AVBufferRef = ptr::null_mut();
+            let r = ffi::av_hwdevice_ctx_create(
+                &mut drm_device,
+                ffi::AVHWDeviceType::AV_HWDEVICE_TYPE_DRM,
+                node.as_ptr(),
+                ptr::null_mut(),
+                0,
+            );
+            if r < 0 {
+                bail!(
+                    "av_hwdevice_ctx_create(DRM {:?}): {}",
+                    node,
+                    ffmpeg::Error::from(r)
+                );
+            }
+            let mut vaapi_device: *mut ffi::AVBufferRef = ptr::null_mut();
+            let r = ffi::av_hwdevice_ctx_create_derived(
+                &mut vaapi_device,
+                ffi::AVHWDeviceType::AV_HWDEVICE_TYPE_VAAPI,
+                drm_device,
+                0,
+            );
+            if r < 0 {
+                ffi::av_buffer_unref(&mut drm_device);
+                bail!("derive VAAPI from DRM: {}", ffmpeg::Error::from(r));
+            }
+
+            // DRM-PRIME frames context for the imported dmabufs.
+            let mut drm_frames = ffi::av_hwframe_ctx_alloc(drm_device);
+            if drm_frames.is_null() {
+                ffi::av_buffer_unref(&mut vaapi_device);
+                ffi::av_buffer_unref(&mut drm_device);
+                bail!("av_hwframe_ctx_alloc(DRM) failed");
+            }
+            let fc = (*drm_frames).data as *mut ffi::AVHWFramesContext;
+            (*fc).format = ffi::AVPixelFormat::AV_PIX_FMT_DRM_PRIME;
+            (*fc).sw_format = ffi::AVPixelFormat::AV_PIX_FMT_BGR0; // packed XR24 RGB plane
+            (*fc).width = width as c_int;
+            (*fc).height = height as c_int;
+            if ffi::av_hwframe_ctx_init(drm_frames) < 0 {
+                ffi::av_buffer_unref(&mut drm_frames);
+                ffi::av_buffer_unref(&mut vaapi_device);
+                ffi::av_buffer_unref(&mut drm_device);
+                bail!("av_hwframe_ctx_init(DRM) failed");
+            }
+
+            // Filter graph: buffer(drm_prime) → hwmap=derive_device=vaapi:mode=read →
+            // scale_vaapi=format=nv12 → buffersink.
+            let mut graph = ffi::avfilter_graph_alloc();
+            if graph.is_null() {
+                ffi::av_buffer_unref(&mut drm_frames);
+                ffi::av_buffer_unref(&mut vaapi_device);
+                ffi::av_buffer_unref(&mut drm_device);
+                bail!("avfilter_graph_alloc failed");
+            }
+
+            let mk = |name: &CStr, inst: &CStr| -> *mut ffi::AVFilterContext {
+                let f = ffi::avfilter_get_by_name(name.as_ptr());
+                if f.is_null() {
+                    return ptr::null_mut();
+                }
+                ffi::avfilter_graph_alloc_filter(graph, f, inst.as_ptr())
+            };
+            let src = mk(c"buffer", c"in");
+            let hwmap = mk(c"hwmap", c"map");
+            let scale = mk(c"scale_vaapi", c"csc");
+            let sink = mk(c"buffersink", c"out");
+            if src.is_null() || hwmap.is_null() || scale.is_null() || sink.is_null() {
+                ffi::avfilter_graph_free(&mut graph);
+                ffi::av_buffer_unref(&mut drm_frames);
+                ffi::av_buffer_unref(&mut vaapi_device);
+                ffi::av_buffer_unref(&mut drm_device);
+                bail!("a VAAPI filter (buffer/hwmap/scale_vaapi/buffersink) is missing");
+            }
+            // hwmap maps the DRM-PRIME input onto THIS vaapi device; scale_vaapi runs the CSC on
+            // it. Giving both our device (rather than `hwmap=derive_device`) keeps every surface —
+            // and the sink's output frames ctx the encoder adopts — on one VADisplay.
+            (*hwmap).hw_device_ctx = ffi::av_buffer_ref(vaapi_device);
+            (*scale).hw_device_ctx = ffi::av_buffer_ref(vaapi_device);
+
+            // buffersrc params: DRM-PRIME frames, the drm_frames ctx.
+            let par = ffi::av_buffersrc_parameters_alloc();
+            (*par).format = ffi::AVPixelFormat::AV_PIX_FMT_DRM_PRIME as c_int;
+            (*par).width = width as c_int;
+            (*par).height = height as c_int;
+            (*par).time_base = ffi::AVRational {
+                num: 1,
+                den: fps as c_int,
+            };
+            (*par).hw_frames_ctx = ffi::av_buffer_ref(drm_frames);
+            let r = ffi::av_buffersrc_parameters_set(src, par);
+            ffi::av_free(par as *mut _);
+            if r < 0 {
+                ffi::avfilter_graph_free(&mut graph);
+                ffi::av_buffer_unref(&mut drm_frames);
+                ffi::av_buffer_unref(&mut vaapi_device);
+                ffi::av_buffer_unref(&mut drm_device);
+                bail!("av_buffersrc_parameters_set failed ({r})");
+            }
+            macro_rules! init {
+                ($ctx:expr, $args:expr, $what:literal) => {{
+                    let r = ffi::avfilter_init_str($ctx, $args);
+                    if r < 0 {
+                        ffi::avfilter_graph_free(&mut graph);
+                        ffi::av_buffer_unref(&mut drm_frames);
+                        ffi::av_buffer_unref(&mut vaapi_device);
+                        ffi::av_buffer_unref(&mut drm_device);
+                        bail!(concat!("init ", $what, " failed ({})"), r);
+                    }
+                }};
+            }
+            init!(src, ptr::null(), "buffer");
+            init!(hwmap, c"mode=read".as_ptr(), "hwmap");
+            init!(scale, c"format=nv12".as_ptr(), "scale_vaapi");
+            init!(sink, ptr::null(), "buffersink");
+
+            let link = |a: *mut ffi::AVFilterContext, b: *mut ffi::AVFilterContext| -> c_int {
+                ffi::avfilter_link(a, 0, b, 0)
+            };
+            if link(src, hwmap) < 0 || link(hwmap, scale) < 0 || link(scale, sink) < 0 {
+                ffi::avfilter_graph_free(&mut graph);
+                ffi::av_buffer_unref(&mut drm_frames);
+                ffi::av_buffer_unref(&mut vaapi_device);
+                ffi::av_buffer_unref(&mut drm_device);
+                bail!("avfilter_link failed");
+            }
+            let r = ffi::avfilter_graph_config(graph, ptr::null_mut());
+            if r < 0 {
+                ffi::avfilter_graph_free(&mut graph);
+                ffi::av_buffer_unref(&mut drm_frames);
+                ffi::av_buffer_unref(&mut vaapi_device);
+                ffi::av_buffer_unref(&mut drm_device);
+                bail!("avfilter_graph_config failed ({r})");
+            }
+
+            // The encoder takes NV12 surfaces from the sink's output frames context.
+            let nv12_ctx = ffi::av_buffersink_get_hw_frames_ctx(sink);
+            if nv12_ctx.is_null() {
+                ffi::avfilter_graph_free(&mut graph);
+                ffi::av_buffer_unref(&mut drm_frames);
+                ffi::av_buffer_unref(&mut vaapi_device);
+                ffi::av_buffer_unref(&mut drm_device);
+                bail!("filter sink has no VAAPI frames context");
+            }
+            let enc = open_vaapi_encoder(
+                codec,
+                width,
+                height,
+                fps,
+                bitrate_bps,
+                vaapi_device,
+                nv12_ctx,
+            )?;
+
+            tracing::info!(
+                encoder = codec.vaapi_name(),
+                "VAAPI encode active ({width}x{height}@{fps}, zero-copy dmabuf → GPU NV12)"
+            );
+            Ok(DmabufInner {
+                enc,
+                drm_device,
+                vaapi_device,
+                drm_frames,
+                graph,
+                src,
+                sink,
+                width,
+                height,
+                fourcc: drm_fourcc,
+            })
+        }
+    }
+
+    fn submit(&mut self, dmabuf: &DmabufFrame, pts: i64, idr: bool) -> Result<()> {
+        anyhow::ensure!(
+            dmabuf.fourcc == self.fourcc,
+            "dmabuf fourcc {:#x} != encoder {:#x}",
+            dmabuf.fourcc,
+            self.fourcc
+        );
+        unsafe {
+            // Build a DRM-PRIME AVFrame describing the dmabuf (one object/fd, one layer/plane).
+            let mut desc: Box<ffi::AVDRMFrameDescriptor> = Box::new(std::mem::zeroed());
+            desc.nb_objects = 1;
+            desc.objects[0].fd = dmabuf.fd.as_raw_fd();
+            desc.objects[0].size = 0;
+            desc.objects[0].format_modifier = dmabuf.modifier;
+            desc.nb_layers = 1;
+            desc.layers[0].format = self.fourcc;
+            desc.layers[0].nb_planes = 1;
+            desc.layers[0].planes[0].object_index = 0;
+            desc.layers[0].planes[0].offset = dmabuf.offset as isize;
+            desc.layers[0].planes[0].pitch = dmabuf.stride as isize;
+
+            let mut drm = ffi::av_frame_alloc();
+            if drm.is_null() {
+                bail!("av_frame_alloc(drm) failed");
+            }
+            (*drm).format = ffi::AVPixelFormat::AV_PIX_FMT_DRM_PRIME as c_int;
+            (*drm).width = self.width as c_int;
+            (*drm).height = self.height as c_int;
+            (*drm).hw_frames_ctx = ffi::av_buffer_ref(self.drm_frames);
+            (*drm).data[0] = Box::into_raw(desc) as *mut u8;
+            // Own the descriptor so it frees with the frame (the fd is owned by the DmabufFrame,
+            // which outlives this call — the graph reads the surface before submit returns).
+            extern "C" fn free_desc(_opaque: *mut std::ffi::c_void, data: *mut u8) {
+                unsafe { drop(Box::from_raw(data as *mut ffi::AVDRMFrameDescriptor)) };
+            }
+            (*drm).buf[0] = ffi::av_buffer_create(
+                (*drm).data[0],
+                std::mem::size_of::<ffi::AVDRMFrameDescriptor>(),
+                Some(free_desc),
+                ptr::null_mut(),
+                0,
+            );
+
+            // Push through hwmap → scale_vaapi; pull the NV12 surface back out.
+            let r = ffi::av_buffersrc_add_frame_flags(
+                self.src,
+                drm,
+                ffi::AV_BUFFERSRC_FLAG_KEEP_REF as c_int,
+            );
+            ffi::av_frame_free(&mut drm);
+            if r < 0 {
+                bail!("av_buffersrc_add_frame failed ({r})");
+            }
+            let mut nv12 = ffi::av_frame_alloc();
+            if nv12.is_null() {
+                bail!("av_frame_alloc(nv12) failed");
+            }
+            let r = ffi::av_buffersink_get_frame(self.sink, nv12);
+            if r < 0 {
+                ffi::av_frame_free(&mut nv12);
+                bail!("av_buffersink_get_frame failed ({r})");
+            }
+            (*nv12).pts = pts;
+            (*nv12).pict_type = if idr {
+                ffi::AVPictureType::AV_PICTURE_TYPE_I
+            } else {
+                ffi::AVPictureType::AV_PICTURE_TYPE_NONE
+            };
+            let r = ffi::avcodec_send_frame(self.enc.as_mut_ptr(), nv12);
+            ffi::av_frame_free(&mut nv12);
+            if r < 0 {
+                bail!("avcodec_send_frame(VAAPI) failed ({r})");
+            }
+        }
+        Ok(())
+    }
+}
+
+impl Drop for DmabufInner {
+    fn drop(&mut self) {
+        unsafe {
+            ffi::avfilter_graph_free(&mut self.graph);
+            ffi::av_buffer_unref(&mut self.drm_frames);
+            ffi::av_buffer_unref(&mut self.vaapi_device);
+            ffi::av_buffer_unref(&mut self.drm_device);
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------------------------
+
+enum Inner {
+    Cpu(CpuInner),
+    Dmabuf(DmabufInner),
+}
+
+pub struct VaapiEncoder {
+    codec: Codec,
+    format: PixelFormat,
+    width: u32,
+    height: u32,
+    fps: u32,
+    bitrate_bps: u64,
+    /// Built lazily from the first frame's payload (CPU upload vs zero-copy dmabuf).
+    inner: Option<Inner>,
+    frame_idx: i64,
+    force_kf: bool,
+}
+
+// Raw FFI pointers; the encoder lives on a single thread (same contract as `NvencEncoder`).
+unsafe impl Send for VaapiEncoder {}
+
+impl VaapiEncoder {
+    pub fn open(
+        codec: Codec,
+        format: PixelFormat,
+        width: u32,
+        height: u32,
+        fps: u32,
+        bitrate_bps: u64,
+        bit_depth: u8,
+    ) -> Result<Self> {
+        if bit_depth != 8 {
+            tracing::warn!(bit_depth, "VAAPI 10-bit not yet wired — encoding 8-bit");
+        }
+        ffmpeg::init().context("ffmpeg init")?;
+        if std::env::var_os("PUNKTFUNK_FFMPEG_DEBUG").is_some() {
+            unsafe { ffi::av_log_set_level(48) };
+        }
+        // Validate the codec/format up front so a bad request fails at open, not on the first frame.
+        let _ = vaapi_sws_src(format)?;
+        Ok(VaapiEncoder {
+            codec,
+            format,
+            width,
+            height,
+            fps,
+            bitrate_bps,
+            inner: None,
+            frame_idx: 0,
+            force_kf: false,
+        })
+    }
+
+    fn ensure_inner(&mut self, want_dmabuf: bool) -> Result<&mut Inner> {
+        if self.inner.is_none() {
+            let inner = if want_dmabuf {
+                Inner::Dmabuf(DmabufInner::open(
+                    self.codec,
+                    self.format,
+                    self.width,
+                    self.height,
+                    self.fps,
+                    self.bitrate_bps,
+                )?)
+            } else {
+                Inner::Cpu(CpuInner::open(
+                    self.codec,
+                    self.format,
+                    self.width,
+                    self.height,
+                    self.fps,
+                    self.bitrate_bps,
+                )?)
+            };
+            self.inner = Some(inner);
+        }
+        Ok(self.inner.as_mut().unwrap())
+    }
+}
+
 impl Encoder for VaapiEncoder {
    fn submit(&mut self, captured: &CapturedFrame) -> Result<()> {
        anyhow::ensure!(
@@ -390,10 +754,14 @@ impl Encoder for VaapiEncoder {
        let idr = self.force_kf;
        self.force_kf = false;
        match &captured.payload {
-            FramePayload::Cpu(bytes) => self.submit_cpu(bytes, captured.format, pts, idr),
-            // CUDA frames are produced only by the NVIDIA zero-copy importer, which never runs on a
-            // VAAPI host. Reaching here means a misconfiguration (e.g. forced PUNKTFUNK_ENCODER=vaapi
-            // on an NVIDIA box with zero-copy on).
+            FramePayload::Cpu(bytes) => match self.ensure_inner(false)? {
+                Inner::Cpu(c) => c.submit(bytes, captured.format, pts, idr),
+                Inner::Dmabuf(_) => bail!("VAAPI encoder built for dmabuf got a CPU frame"),
+            },
+            FramePayload::Dmabuf(d) => match self.ensure_inner(true)? {
+                Inner::Dmabuf(dm) => dm.submit(d, pts, idr),
+                Inner::Cpu(_) => bail!("VAAPI encoder built for CPU got a dmabuf frame"),
+            },
            FramePayload::Cuda(_) => bail!(
                "VAAPI encoder received a CUDA frame — that payload is NVENC-only; \
                 unset PUNKTFUNK_ZEROCOPY or don't force PUNKTFUNK_ENCODER=vaapi on an NVIDIA host"
@@ -406,46 +774,19 @@ impl Encoder for VaapiEncoder {
    }

    fn poll(&mut self) -> Result<Option<EncodedFrame>> {
-        let mut pkt = Packet::empty();
-        match self.enc.receive_packet(&mut pkt) {
-            Ok(()) => {
-                let data = pkt.data().map(|d| d.to_vec()).unwrap_or_default();
-                let pts = pkt.pts().unwrap_or(0).max(0) as u64;
-                let pts_ns = pts * 1_000_000_000 / self.fps as u64;
-                Ok(Some(EncodedFrame {
-                    data,
-                    pts_ns,
-                    keyframe: pkt.is_key(),
-                }))
-            }
-            Err(ffmpeg::Error::Other { errno })
-                if errno == ffmpeg::util::error::EAGAIN
-                    || errno == ffmpeg::util::error::EWOULDBLOCK =>
-            {
-                Ok(None)
-            }
-            Err(ffmpeg::Error::Eof) => Ok(None),
-            Err(e) => Err(e).context("receive_packet"),
+        match &mut self.inner {
+            Some(Inner::Cpu(c)) => poll_encoder(&mut c.enc, self.fps),
+            Some(Inner::Dmabuf(d)) => poll_encoder(&mut d.enc, self.fps),
+            None => Ok(None),
        }
    }

    fn flush(&mut self) -> Result<()> {
-        self.enc.send_eof().context("send_eof")?;
+        match &mut self.inner {
+            Some(Inner::Cpu(c)) => c.enc.send_eof().context("send_eof")?,
+            Some(Inner::Dmabuf(d)) => d.enc.send_eof().context("send_eof")?,
+            None => {}
+        }
        Ok(())
    }
 }
-
-impl Drop for VaapiEncoder {
-    fn drop(&mut self) {
-        unsafe {
-            if !self.nv12.is_null() {
-                ffi::av_frame_free(&mut self.nv12);
-            }
-            if !self.sws.is_null() {
-                ffi::sws_freeContext(self.sws);
-            }
-        }
-        // `enc` (frees the codec ctx, unref'ing its hw-context copies) and `hw` (unref'ing the
-        // originals) drop via their own impls — refcounting makes the order irrelevant.
-    }
-}