refactor(host/W6.2): extract the video encode backends into the pf-encode crate
encode.rs + encode/* (NVENC, VAAPI, native AMF, AMF/QSV ffmpeg, direct-SDK NVENC/CUDA, raw Vulkan-Video, PyroWave, openh264) move into crates/pf-encode behind one Encoder trait + open_video selector (plan §W6). The crate speaks the shared frame vocabulary (pf-frame: CapturedFrame/PixelFormat + the DXGI identity D3d11Frame/make_device) and pf-zerocopy (CUDA context/buffers), and NEVER pf-capture — the capture→encode edge is one-way (ZeroCopyPolicy, prior commit). Dep moves: the heavy encoder deps (ffmpeg-next, the NVENC SDK, openh264, pyrowave-sys) move from the host to pf-encode; the host's nvenc/amf-qsv/vulkan-encode/pyrowave features now FORWARD to pf-encode/*. The host keeps a mod-encode shim (pub use pf_encode) so every crate::encode::* path (negotiator + GameStream/native/mgmt planes) is unchanged. resolve_render_adapter_luid moves from the host's windows/win_adapter.rs into pf-gpu (both pf-encode and pf-capture need it as a peer of GPU selection); its 5 call sites (encode amf/nvenc, capture idd_push/synthetic_nv12, vdisplay manager) rewire to pf_gpu::resolve_render_adapter_luid and win_adapter.rs is deleted. pf-frame's make_device gains a # Safety section (public-unsafe-fn lint, latent since the pf-frame carve — a full-workspace -D warnings clippy catches it). Verified: Linux clippy -D warnings (pf-encode + host nvenc,vulkan-encode,pyrowave --all-targets) + 13/13 pf-encode + 299/299 host tests; Windows clippy -D warnings (pf-encode nvenc,amf-qsv --all-targets + host nvenc,amf-qsv --all-targets) Finished exit 0. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
@@ -0,0 +1,857 @@
|
||||
//! NVENC encoder via `ffmpeg-next` (binds the system FFmpeg — `ffmpeg-sys-next` auto-detects the
|
||||
//! installed version, so this builds against FFmpeg 7.x/libavcodec 61 *or* 8.x/libavcodec 62;
|
||||
//! validated live on Ubuntu 26.04 (FFmpeg 8) and Bazzite F43 (FFmpeg 7.1)).
|
||||
//!
|
||||
//! Input is a packed RGB/BGR CPU frame; `*_nvenc` accepts `rgb0`/`bgr0`/`rgba`/`bgra`
|
||||
//! directly and does the RGB→YUV conversion on the GPU, so the host stays off the
|
||||
//! colour-conversion path. The portal commonly negotiates packed 24-bit `RGB`, which NVENC
|
||||
//! does *not* accept — we expand it to `rgb0` (one padding byte/pixel, no colour math).
|
||||
//! The encoder is opened *without* a global header so VPS/SPS/PPS are emitted in-band on
|
||||
//! every IDR — the output is both a playable raw Annex-B stream and self-contained AUs.
|
||||
// Every `unsafe` block in this file carries a `// SAFETY:` proof; enforce it (unsafe-proof program).
|
||||
#![deny(clippy::undocumented_unsafe_blocks)]
|
||||
|
||||
use super::{ChromaFormat, Codec, EncodedFrame, Encoder};
|
||||
use anyhow::{anyhow, bail, Context, Result};
|
||||
use ffmpeg::format::Pixel;
|
||||
use ffmpeg::util::frame::Video as VideoFrame;
|
||||
use ffmpeg::{codec, encoder, Dictionary};
|
||||
use ffmpeg_next as ffmpeg;
|
||||
use pf_frame::{CapturedFrame, FramePayload, PixelFormat};
|
||||
use std::os::raw::c_int;
|
||||
use std::ptr;
|
||||
|
||||
use super::libav::{
|
||||
apply_low_latency_rc, pixel_to_av, poll_encoder, PollOutcome, SWS_CS_ITU709, SWS_POINT,
|
||||
};
|
||||
use ffmpeg::ffi; // = ffmpeg_sys_next
|
||||
|
||||
/// The swscale *source* pixel format for a captured packed RGB/BGR layout (the real byte order, not
|
||||
/// the NVENC-padded `*0` form). Used by the 4:4:4 RGB→YUV444P conversion path. Mirrors the VAAPI
|
||||
/// CPU-input mapping; YUV/10-bit inputs can't feed this path (the 4:4:4 session forces packed RGB).
|
||||
fn sws_src_pixel(format: PixelFormat) -> Result<Pixel> {
|
||||
Ok(match format {
|
||||
PixelFormat::Bgrx => Pixel::BGRZ, // bgr0
|
||||
PixelFormat::Rgbx => Pixel::RGBZ, // rgb0
|
||||
PixelFormat::Bgra => Pixel::BGRA,
|
||||
PixelFormat::Rgba => Pixel::RGBA,
|
||||
PixelFormat::Rgb => Pixel::RGB24,
|
||||
PixelFormat::Bgr => Pixel::BGR24,
|
||||
PixelFormat::Nv12 | PixelFormat::P010 | PixelFormat::Rgb10a2 | PixelFormat::Yuv444 => {
|
||||
bail!("NVENC 4:4:4 CPU-input path supports packed RGB/BGR only; got {format:?}")
|
||||
}
|
||||
})
|
||||
}
|
||||
|
||||
/// `AVCUDADeviceContext` (libavutil/hwcontext_cuda.h) — not in the ffmpeg-sys bindings (the
|
||||
/// crate doesn't allowlist that header), so mirror its stable 3-pointer layout. We set the
|
||||
/// first field to *our* `CUcontext` so NVENC shares the context the EGL importer maps into.
|
||||
#[repr(C)]
|
||||
struct AVCUDADeviceContext {
|
||||
cuda_ctx: *mut std::ffi::c_void, // CUcontext
|
||||
stream: *mut std::ffi::c_void, // CUstream (null = default)
|
||||
internal: *mut std::ffi::c_void, // filled by ctx_init
|
||||
}
|
||||
|
||||
/// CUDA hardware-frame contexts that wrap our shared `CUcontext`, so `hevc_nvenc` reads the
|
||||
/// imported device buffer directly. Owns two `AVBufferRef`s, unref'd on drop.
|
||||
struct CudaHw {
|
||||
device_ref: *mut ffi::AVBufferRef,
|
||||
frames_ref: *mut ffi::AVBufferRef,
|
||||
}
|
||||
|
||||
impl CudaHw {
|
||||
/// Build a CUDA hwdevice wrapping `cu_ctx` and a frames pool (`sw_format` = `pixel`).
|
||||
unsafe fn new(cu_ctx: *mut std::ffi::c_void, sw_format: Pixel, w: u32, h: u32) -> Result<Self> {
|
||||
let mut device_ref = ffi::av_hwdevice_ctx_alloc(ffi::AVHWDeviceType::AV_HWDEVICE_TYPE_CUDA);
|
||||
if device_ref.is_null() {
|
||||
bail!("av_hwdevice_ctx_alloc(CUDA) failed");
|
||||
}
|
||||
let dev_ctx = (*device_ref).data as *mut ffi::AVHWDeviceContext;
|
||||
let cu = (*dev_ctx).hwctx as *mut AVCUDADeviceContext;
|
||||
(*cu).cuda_ctx = cu_ctx; // share the importer's context
|
||||
let r = ffi::av_hwdevice_ctx_init(device_ref);
|
||||
if r < 0 {
|
||||
ffi::av_buffer_unref(&mut device_ref);
|
||||
bail!("av_hwdevice_ctx_init failed ({r})");
|
||||
}
|
||||
|
||||
let mut frames_ref = ffi::av_hwframe_ctx_alloc(device_ref);
|
||||
if frames_ref.is_null() {
|
||||
ffi::av_buffer_unref(&mut device_ref);
|
||||
bail!("av_hwframe_ctx_alloc failed");
|
||||
}
|
||||
let fc = (*frames_ref).data as *mut ffi::AVHWFramesContext;
|
||||
(*fc).format = ffi::AVPixelFormat::AV_PIX_FMT_CUDA;
|
||||
(*fc).sw_format = pixel_to_av(sw_format);
|
||||
(*fc).width = w as c_int;
|
||||
(*fc).height = h as c_int;
|
||||
(*fc).initial_pool_size = 0; // we supply the device pointers
|
||||
let r = ffi::av_hwframe_ctx_init(frames_ref);
|
||||
if r < 0 {
|
||||
ffi::av_buffer_unref(&mut frames_ref);
|
||||
ffi::av_buffer_unref(&mut device_ref);
|
||||
bail!("av_hwframe_ctx_init failed ({r})");
|
||||
}
|
||||
Ok(CudaHw {
|
||||
device_ref,
|
||||
frames_ref,
|
||||
})
|
||||
}
|
||||
}
|
||||
|
||||
impl Drop for CudaHw {
|
||||
fn drop(&mut self) {
|
||||
// SAFETY: `frames_ref`/`device_ref` are the two non-null `AVBufferRef`s `CudaHw::new` created
|
||||
// (it bails before returning `Self` if either alloc fails, so a live `CudaHw` always holds
|
||||
// both). `av_buffer_unref` drops one reference and nulls the pointer through the `&mut`. This
|
||||
// `Drop` runs exactly once and `CudaHw` owns these refs exclusively → no double-free /
|
||||
// use-after-free. Frames are unref'd before the device (the frames ctx internally refs the
|
||||
// device; refcounted, so the order is sound regardless).
|
||||
unsafe {
|
||||
ffi::av_buffer_unref(&mut self.frames_ref);
|
||||
ffi::av_buffer_unref(&mut self.device_ref);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
/// Map a captured layout to the NVENC input pixel format, and whether a 3→4 byte expand is
|
||||
/// needed (packed RGB/BGR have no padding byte; the NVENC `*0` formats do).
|
||||
fn nvenc_input(format: PixelFormat) -> (Pixel, bool) {
|
||||
match format {
|
||||
PixelFormat::Bgrx => (Pixel::BGRZ, false), // bgr0
|
||||
PixelFormat::Rgbx => (Pixel::RGBZ, false), // rgb0
|
||||
PixelFormat::Bgra => (Pixel::BGRA, false),
|
||||
PixelFormat::Rgba => (Pixel::RGBA, false),
|
||||
PixelFormat::Rgb => (Pixel::RGBZ, true), // RGB -> rgb0
|
||||
PixelFormat::Bgr => (Pixel::BGRZ, true), // BGR -> bgr0
|
||||
// NV12 is native YUV: NVENC encodes it with NO internal RGB→YUV CSC (the Tier 2A win). On
|
||||
// Linux it's produced by the GPU convert on the zero-copy tiled path (`PUNKTFUNK_NV12`); on
|
||||
// Windows by the D3D11 video processor.
|
||||
PixelFormat::Nv12 => (Pixel::NV12, false),
|
||||
// Planar YUV444 from the zero-copy worker's GPU convert (a 4:4:4 session) — native
|
||||
// full-chroma YUV in, `hevc_nvenc` emits Range-Extensions 4:4:4.
|
||||
PixelFormat::Yuv444 => (Pixel::YUV444P, false),
|
||||
// Rgb10a2 (HDR) and P010 (the Windows 10-bit video-processor output) are produced only by
|
||||
// the Windows paths; the Linux capturer never emits them. Map to BGRA so the match is
|
||||
// exhaustive — unreachable here.
|
||||
PixelFormat::Rgb10a2 | PixelFormat::P010 => (Pixel::BGRA, false),
|
||||
}
|
||||
}
|
||||
|
||||
/// The [`NvencEncoder::open`] arguments, kept on the encoder so [`Encoder::reset`] can rebuild it
|
||||
/// in place with the session's negotiated parameters — the encode-stall watchdog's recovery lever
|
||||
/// (drop the wedged libavcodec encoder, reopen fresh, forfeit the owed AUs, restart at an IDR).
|
||||
#[derive(Clone, Copy)]
|
||||
struct OpenArgs {
|
||||
codec: Codec,
|
||||
format: PixelFormat,
|
||||
width: u32,
|
||||
height: u32,
|
||||
fps: u32,
|
||||
bitrate_bps: u64,
|
||||
cuda: bool,
|
||||
bit_depth: u8,
|
||||
chroma: ChromaFormat,
|
||||
}
|
||||
|
||||
pub struct NvencEncoder {
|
||||
enc: encoder::video::Encoder,
|
||||
/// Reusable 4-bpp CPU input frame (CPU path only; `None` for the zero-copy/CUDA path).
|
||||
/// Mutating it in place across frames is sound only because the encoder is opened with
|
||||
/// `delay=0`/`bf=0`/`max_b_frames=0` and the caller drains `poll()` after each `submit`,
|
||||
/// so libavcodec holds no reference to the previous frame's buffer when we overwrite it.
|
||||
frame: Option<VideoFrame>,
|
||||
/// Zero-copy path: CUDA hwdevice/hwframes contexts (the encoder takes `AV_PIX_FMT_CUDA`).
|
||||
cuda: Option<CudaHw>,
|
||||
/// 4:4:4 CPU path only: swscale context converting the captured packed RGB/BGR → planar
|
||||
/// YUV444P into [`Self::frame`], because `hevc_nvenc` only emits 4:4:4 from a YUV444 *input*
|
||||
/// (RGB-in is always 4:2:0). `None` on the 4:2:0 paths AND on the zero-copy 4:4:4 path (the
|
||||
/// worker's GPU convert delivers YUV444 CUDA frames). Freed in `Drop`.
|
||||
sws_444: Option<*mut ffi::SwsContext>,
|
||||
/// This session opened as full-chroma 4:4:4 (FREXT) — via either input path.
|
||||
want_444: bool,
|
||||
src_format: PixelFormat,
|
||||
expand: bool,
|
||||
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,
|
||||
/// Opened in intra-refresh mode (surfaced via [`caps`](Encoder::caps) so the session glue
|
||||
/// rate-limits forced IDRs — the wave heals loss without them).
|
||||
intra_refresh: bool,
|
||||
/// Resolved wave length in frames when [`intra_refresh`](Self::intra_refresh), else 0. Cached at
|
||||
/// open so the pump's per-AU `caps()` doesn't re-read `PUNKTFUNK_IR_PERIOD_FRAMES`; the pump marks
|
||||
/// every Nth AU with `USER_FLAG_RECOVERY_POINT` for the client's clean re-anchor.
|
||||
intra_refresh_period: u32,
|
||||
/// The open arguments, for the in-place [`reset`](Encoder::reset) rebuild.
|
||||
args: OpenArgs,
|
||||
}
|
||||
|
||||
// `CudaHw` holds raw `AVBufferRef`s and `sws_444` a raw `SwsContext`; the encoder lives on a single
|
||||
// thread. The CPU encoder is already `Send` via ffmpeg-next; assert it for the raw fields too.
|
||||
// SAFETY: `NvencEncoder` owns an ffmpeg-next `Encoder`/`VideoFrame` (already `Send`) plus a `CudaHw`
|
||||
// holding raw `AVBufferRef`s and an optional raw `SwsContext`, none of which are `Send` by default.
|
||||
// The `SwsContext` is a self-contained swscale state object with no thread affinity, touched only
|
||||
// through `&mut self` on the one encode thread. The encoder is owned and driven by
|
||||
// exactly ONE thread — the per-session encode thread it is moved to — and is only touched through
|
||||
// `&mut self` methods, so it is never aliased or accessed concurrently. The wrapped libav contexts
|
||||
// (and the shared `CUcontext` the `CudaHw` references) have no thread affinity, so transferring
|
||||
// ownership across threads is sound. This asserts `Send` (transfer) only, extending ffmpeg-next's
|
||||
// existing `Send` to the raw CUDA fields; `Sync` (shared `&`) is deliberately NOT implemented.
|
||||
unsafe impl Send for NvencEncoder {}
|
||||
|
||||
/// Latched true once an intra-refresh open failed with the device-capability error (ENOSYS from
|
||||
/// `NV_ENC_CAPS_SUPPORT_INTRA_REFRESH`), so later sessions skip the doomed attempt. Never set by
|
||||
/// other open failures (a bitrate EINVAL must not permanently disable the feature).
|
||||
static IR_UNSUPPORTED: std::sync::atomic::AtomicBool = std::sync::atomic::AtomicBool::new(false);
|
||||
|
||||
/// Whether this open should run the NVENC **intra-refresh** loss-recovery mode
|
||||
/// (`PUNKTFUNK_INTRA_REFRESH` truthy, opt-in until on-glass validated): a moving intra band +
|
||||
/// recovery-point SEI refreshes the whole picture every [`intra_refresh_period`] frames, so
|
||||
/// FEC-unrecoverable loss heals without the 20-40× full-IDR spike (which under loss causes more
|
||||
/// loss — the cascade). The session glue then rate-limits client keyframe requests
|
||||
/// ([`EncoderCaps::intra_refresh`](super::EncoderCaps)).
|
||||
fn intra_refresh_requested() -> bool {
|
||||
std::env::var("PUNKTFUNK_INTRA_REFRESH")
|
||||
.map(|v| matches!(v.trim(), "1" | "true" | "yes" | "on"))
|
||||
.unwrap_or(false)
|
||||
&& !IR_UNSUPPORTED.load(std::sync::atomic::Ordering::Relaxed)
|
||||
}
|
||||
|
||||
/// The intra-refresh wave length in frames — ffmpeg derives `intraRefreshPeriod`/`Cnt` from
|
||||
/// `gop_size` before forcing the real GOP infinite, so this is what `gop_size` is set to in IR
|
||||
/// mode. Default = half a second of frames (heals fast, spreads the intra cost to ~2-3% per
|
||||
/// frame); `PUNKTFUNK_IR_PERIOD_FRAMES` overrides.
|
||||
fn intra_refresh_period(fps: u32) -> i32 {
|
||||
std::env::var("PUNKTFUNK_IR_PERIOD_FRAMES")
|
||||
.ok()
|
||||
.and_then(|s| s.parse::<i32>().ok())
|
||||
.filter(|v| *v >= 2)
|
||||
.unwrap_or_else(|| (fps.max(16) / 2) as i32)
|
||||
}
|
||||
|
||||
impl NvencEncoder {
|
||||
#[allow(clippy::too_many_arguments)]
|
||||
pub fn open(
|
||||
codec: Codec,
|
||||
format: PixelFormat,
|
||||
width: u32,
|
||||
height: u32,
|
||||
fps: u32,
|
||||
bitrate_bps: u64,
|
||||
cuda: bool,
|
||||
bit_depth: u8,
|
||||
chroma: ChromaFormat,
|
||||
) -> Result<Self> {
|
||||
// TODO(hdr): Linux 10-bit parity. Unlike the Windows raw-SDK path (which upconverts 8-bit
|
||||
// ARGB → Main10 via pixelBitDepthMinus8), libavcodec hevc_nvenc needs a 10-bit input pixel
|
||||
// format (p010) for Main10, so it's a bigger change; deferred until a Linux GPU box is
|
||||
// available to validate. The Linux host stays 8-bit for now.
|
||||
if bit_depth != 8 {
|
||||
tracing::warn!(
|
||||
bit_depth,
|
||||
"Linux NVENC 10-bit not yet wired — encoding 8-bit"
|
||||
);
|
||||
}
|
||||
// Full-chroma 4:4:4 (HEVC Range Extensions). `hevc_nvenc` only emits 4:4:4 from a YUV444
|
||||
// *input* frame — feeding RGB always subsamples to 4:2:0 regardless of profile (verified on
|
||||
// the RTX 5070 Ti). Two ways to produce that input: the zero-copy worker's GPU convert
|
||||
// (planar-YUV444 CUDA frames — `cuda` true), or the CPU path's swscale RGB→YUV444P. Both
|
||||
// feed `profile=rext`; the range follows `PUNKTFUNK_444_FULLRANGE` in both.
|
||||
let want_444 = chroma.is_444() && codec == Codec::H265;
|
||||
ffmpeg::init().context("ffmpeg init")?;
|
||||
if std::env::var_os("PUNKTFUNK_FFMPEG_DEBUG").is_some() {
|
||||
// SAFETY: `av_log_set_level` sets libav's global integer log level; `48` (= AV_LOG_DEBUG)
|
||||
// is a valid level with no pointer args, and libav was just initialized by `ffmpeg::init()`
|
||||
// above — always sound.
|
||||
unsafe { ffi::av_log_set_level(48) }; // AV_LOG_DEBUG — surface NVENC hw-frame rejects
|
||||
}
|
||||
let name = codec.nvenc_name();
|
||||
let av_codec = encoder::find_by_name(name)
|
||||
.ok_or_else(|| anyhow!("{name} not built into libavcodec"))?;
|
||||
let (rgb_pixel, rgb_expand) = nvenc_input(format);
|
||||
// 4:4:4 feeds NVENC a planar YUV444P frame we produce by swscale; the ordinary path feeds the
|
||||
// captured RGB straight in and lets NVENC's internal CSC subsample to 4:2:0.
|
||||
let (nvenc_pixel, expand) = if want_444 {
|
||||
(Pixel::YUV444P, false)
|
||||
} else {
|
||||
(rgb_pixel, rgb_expand)
|
||||
};
|
||||
|
||||
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(nvenc_pixel); // NVENC converts RGB→YUV internally
|
||||
// Fixed rate, CBR, no B-frames, ~1-frame VBV — the shared low-latency RC contract.
|
||||
apply_low_latency_rc(&mut video, fps, bitrate_bps);
|
||||
// Infinite GOP — NO periodic IDR. A keyframe at 5120x1440 is ~20-40x a P-frame, so a
|
||||
// periodic IDR is a recurring multi-millisecond encode+packetize+send spike — the ~2s
|
||||
// "freeze". NVENC emits one IDR at stream start, then P-frames only; `forced-idr` (below)
|
||||
// turns a client recovery request (RFI, via `request_keyframe`) into an IDR on demand.
|
||||
// This is the Moonlight/Sunshine low-latency model.
|
||||
// In intra-refresh mode the GOP is still infinite — ffmpeg reads `gop_size` as the refresh
|
||||
// WAVE length (`intraRefreshPeriod`/`Cnt`) and then forces `gopLength` infinite itself, so
|
||||
// a positive `gop_size` here does NOT reintroduce periodic IDRs.
|
||||
let intra_refresh = intra_refresh_requested();
|
||||
// SAFETY: same `video` builder as above — a non-null, properly-aligned, sole-owned, not-yet-
|
||||
// opened `AVCodecContext`. We write the plain `gop_size` int field (-1 = infinite GOP, or the
|
||||
// intra-refresh wave length) before `open_with`, which ffmpeg-next has no setter for. No
|
||||
// aliasing; synchronous scalar write.
|
||||
unsafe {
|
||||
(*video.as_mut_ptr()).gop_size = if intra_refresh {
|
||||
intra_refresh_period(fps)
|
||||
} else {
|
||||
-1
|
||||
};
|
||||
}
|
||||
|
||||
// NV12 / 4:4:4 paths: we do the RGB→YUV conversion ourselves as BT.709 (swscale), so
|
||||
// signal that in the bitstream VUI (colorspace/range/primaries/transfer) — otherwise the
|
||||
// client decoder assumes a default and the picture comes out washed-out / wrong-contrast.
|
||||
// The RGB-input 4:2:0 path leaves these unset (NVENC's internal CSC writes its own VUI).
|
||||
// Matches the Windows NV12 path's BT.709 limited-range signalling.
|
||||
//
|
||||
// PUNKTFUNK_444_FULLRANGE=1 (experimental, 4:4:4-only): convert AND signal FULL range —
|
||||
// recovers the ~12% of code space limited-range quantization gives up, for the exact
|
||||
// text/UI chroma 4:4:4 exists for. Every punktfunk client honors the signaled range
|
||||
// (csc_rows / the Apple rows port); ship as default only if the on-glass A/B shows a
|
||||
// visible win. Linux-only: the Windows path's NVENC-internal CSC range is unmeasured.
|
||||
let full_range_444 =
|
||||
want_444 && std::env::var("PUNKTFUNK_444_FULLRANGE").is_ok_and(|v| v.trim() == "1");
|
||||
if matches!(format, PixelFormat::Nv12) || want_444 {
|
||||
// SAFETY: same `video` builder — `raw = video.as_mut_ptr()` is the non-null, properly-
|
||||
// aligned, sole-owned, not-yet-opened `AVCodecContext`. We set its four VUI colour enum
|
||||
// fields to valid `AVColorSpace`/`AVColorRange`/`AVColorPrimaries`/`AVColorTransfer-
|
||||
// Characteristic` variants before `open_with`. Sole owner → no aliasing; synchronous writes.
|
||||
unsafe {
|
||||
let raw = video.as_mut_ptr();
|
||||
(*raw).colorspace = ffi::AVColorSpace::AVCOL_SPC_BT709;
|
||||
(*raw).color_range = if full_range_444 {
|
||||
ffi::AVColorRange::AVCOL_RANGE_JPEG // full
|
||||
} else {
|
||||
ffi::AVColorRange::AVCOL_RANGE_MPEG // limited/studio
|
||||
};
|
||||
(*raw).color_primaries = ffi::AVColorPrimaries::AVCOL_PRI_BT709;
|
||||
(*raw).color_trc = ffi::AVColorTransferCharacteristic::AVCOL_TRC_BT709;
|
||||
}
|
||||
}
|
||||
|
||||
// For the zero-copy path, take CUDA surfaces: wrap the shared CUcontext in CUDA
|
||||
// hwdevice/hwframes contexts and set `pix_fmt = CUDA` on the raw encoder context
|
||||
// *before* open (NVENC derives the device from `hw_frames_ctx`).
|
||||
let cuda_hw = if cuda {
|
||||
let cu_ctx = pf_zerocopy::cuda::context().context("shared CUDA context")?;
|
||||
// SAFETY: `CudaHw::new` (an `unsafe fn`) requires libav initialized (the `ffmpeg::init()`
|
||||
// above ran) and a valid `CUcontext`; `cu_ctx` is the shared importer context from
|
||||
// `zerocopy::cuda::context()?`, non-null on the `Ok` path. `nvenc_pixel` is a valid `Pixel`
|
||||
// and `width`/`height` are the validated positive dims. It returns a RAII `CudaHw` wrapping
|
||||
// (not owning) `cu_ctx` and owning two `AVBufferRef`s freed on drop.
|
||||
let hw = unsafe { CudaHw::new(cu_ctx, nvenc_pixel, width, height)? };
|
||||
// SAFETY: `raw = video.as_mut_ptr()` is the non-null, sole-owned, not-yet-opened
|
||||
// `AVCodecContext`. We set `pix_fmt = CUDA` and attach NEW refs (`av_buffer_ref`) of
|
||||
// `hw.device_ref`/`hw.frames_ref` — both non-null (`CudaHw::new` guarantees) and from the
|
||||
// live `hw`, which is moved into `NvencEncoder.cuda` next to `enc` and so outlives the
|
||||
// encoder. The context owns its own refs (freed when the context closes). No aliasing.
|
||||
unsafe {
|
||||
let raw = video.as_mut_ptr();
|
||||
(*raw).pix_fmt = ffi::AVPixelFormat::AV_PIX_FMT_CUDA;
|
||||
(*raw).hw_device_ctx = ffi::av_buffer_ref(hw.device_ref);
|
||||
(*raw).hw_frames_ctx = ffi::av_buffer_ref(hw.frames_ref);
|
||||
}
|
||||
Some(hw)
|
||||
} else {
|
||||
None
|
||||
};
|
||||
|
||||
// 4:4:4 CPU path: build the RGB→YUV444P swscale (BT.709, range per the flag; no rescale).
|
||||
// Mirrors the VAAPI CPU path's RGB→NV12 scaler, but the dst is full-chroma planar 4:4:4.
|
||||
// Skipped on the zero-copy path (`cuda`): the worker's GPU convert already delivers
|
||||
// planar YUV444 CUDA frames — no CPU pixels exist to scale.
|
||||
let sws_444 = if want_444 && !cuda {
|
||||
let src_av = pixel_to_av(sws_src_pixel(format)?);
|
||||
// SAFETY: `sws_getContext` allocates a swscale context for the given src/dst dims + pixel
|
||||
// formats. Both dims are the encoder's positive `width`/`height` as `c_int`; `src_av` is a
|
||||
// valid `AVPixelFormat` (from the `sws_src_pixel`-validated, packed-RGB-only source), the
|
||||
// dst is YUV444P. The trailing filter/param pointers are null = "use defaults" (documented
|
||||
// as accepted). No Rust memory is borrowed; the returned pointer is null-checked below.
|
||||
let sws = unsafe {
|
||||
ffi::sws_getContext(
|
||||
width as c_int,
|
||||
height as c_int,
|
||||
src_av,
|
||||
width as c_int,
|
||||
height as c_int,
|
||||
ffi::AVPixelFormat::AV_PIX_FMT_YUV444P,
|
||||
SWS_POINT,
|
||||
ptr::null_mut(),
|
||||
ptr::null_mut(),
|
||||
ptr::null(),
|
||||
)
|
||||
};
|
||||
if sws.is_null() {
|
||||
bail!("sws_getContext(RGB→YUV444P) failed");
|
||||
}
|
||||
// SAFETY: `sws` is the non-null context from the call above (null-checked). The ITU-709
|
||||
// coefficient table from `sws_getCoefficients` is a process-lifetime libswscale static,
|
||||
// reused for src+dst matrices; `sws_setColorspaceDetails` only reads it and writes scalar
|
||||
// CSC settings into `sws` (dstRange matches the VUI: 0 = limited, 1 = the
|
||||
// PUNKTFUNK_444_FULLRANGE experiment). No Rust memory is passed.
|
||||
unsafe {
|
||||
let cs709 = ffi::sws_getCoefficients(SWS_CS_ITU709);
|
||||
let dst_range = i32::from(full_range_444);
|
||||
ffi::sws_setColorspaceDetails(sws, cs709, 1, cs709, dst_range, 0, 1 << 16, 1 << 16);
|
||||
}
|
||||
Some(sws)
|
||||
} else {
|
||||
None
|
||||
};
|
||||
|
||||
// Low-latency NVENC tuning (plan §7 / linux-setup doc).
|
||||
let mut opts = Dictionary::new();
|
||||
opts.set("preset", "p1"); // fastest
|
||||
opts.set("tune", "ull"); // ultra-low-latency
|
||||
opts.set("rc", "cbr");
|
||||
opts.set("bf", "0");
|
||||
opts.set("delay", "0");
|
||||
opts.set("forced-idr", "1"); // RFI/request_keyframe → real IDR under the infinite GOP
|
||||
if intra_refresh {
|
||||
// Moving intra band + recovery-point SEI (period set via gop_size above). Loss now
|
||||
// self-heals within the wave; forced IDRs remain available (rate-limited by the glue).
|
||||
opts.set("intra-refresh", "1");
|
||||
}
|
||||
if want_444 {
|
||||
// HEVC Range Extensions — the profile that carries chroma_format_idc=3. With a YUV444P
|
||||
// input `hevc_nvenc` auto-selects it, but pin it explicitly so the chroma is never silently
|
||||
// dropped on a future libavcodec.
|
||||
opts.set("profile", "rext");
|
||||
}
|
||||
|
||||
// Split-frame encode across both NVENC engines (GB203 has 2) when the pixel rate exceeds
|
||||
// a single engine's HEVC capacity (~1 Gpix/s); e.g. 5120x1440@240 = 1.77 Gpix/s needs it,
|
||||
// @120 = 0.88 Gpix/s does not. HEVC/AV1 only (not H.264). AUTO won't engage below ~2112px
|
||||
// height, so we force `2`; below the threshold we leave it AUTO (split costs ~2% BD-rate).
|
||||
// Output is standard HEVC — transparent to the client. Override with PUNKTFUNK_SPLIT_ENCODE.
|
||||
let pix_rate = width as u64 * height as u64 * fps as u64;
|
||||
let split = std::env::var("PUNKTFUNK_SPLIT_ENCODE").ok();
|
||||
match split.as_deref() {
|
||||
Some(mode) => opts.set("split_encode_mode", mode),
|
||||
None if matches!(codec, Codec::H265 | Codec::Av1) && pix_rate > 1_000_000_000 => {
|
||||
opts.set("split_encode_mode", "2");
|
||||
tracing::info!(
|
||||
pix_rate,
|
||||
"NVENC: forcing 2-way split encode (high pixel rate)"
|
||||
);
|
||||
}
|
||||
None => {}
|
||||
}
|
||||
|
||||
let enc = match video.open_with(opts) {
|
||||
Ok(enc) => enc,
|
||||
// The GPU lacks NV_ENC_CAPS_SUPPORT_INTRA_REFRESH — ffmpeg fails the open with
|
||||
// ENOSYS ("Function not implemented"). Latch it (skip the doomed attempt on later
|
||||
// sessions) and reopen this session without intra-refresh; any other failure — and
|
||||
// any failure when IR wasn't requested — propagates untouched (the bitrate probe
|
||||
// keys on EINVAL, which must not trip the latch).
|
||||
Err(e) if intra_refresh && format!("{e:#}").contains("Function not implemented") => {
|
||||
tracing::warn!(
|
||||
encoder = name,
|
||||
"NVENC intra-refresh not supported by this GPU — falling back to IDR-only \
|
||||
recovery"
|
||||
);
|
||||
IR_UNSUPPORTED.store(true, std::sync::atomic::Ordering::Relaxed);
|
||||
return Self::open(
|
||||
codec,
|
||||
format,
|
||||
width,
|
||||
height,
|
||||
fps,
|
||||
bitrate_bps,
|
||||
cuda,
|
||||
bit_depth,
|
||||
chroma,
|
||||
);
|
||||
}
|
||||
Err(e) => {
|
||||
return Err(e).with_context(|| {
|
||||
format!("open {name} ({width}x{height}@{fps}, {bitrate_bps} bps)")
|
||||
})
|
||||
}
|
||||
};
|
||||
if intra_refresh {
|
||||
tracing::info!(
|
||||
encoder = name,
|
||||
period_frames = intra_refresh_period(fps),
|
||||
"NVENC intra-refresh recovery active (no periodic IDR; wave heals loss)"
|
||||
);
|
||||
}
|
||||
|
||||
let frame = if cuda {
|
||||
None
|
||||
} else {
|
||||
Some(VideoFrame::new(nvenc_pixel, width, height))
|
||||
};
|
||||
Ok(NvencEncoder {
|
||||
enc,
|
||||
frame,
|
||||
cuda: cuda_hw,
|
||||
sws_444,
|
||||
want_444,
|
||||
src_format: format,
|
||||
expand,
|
||||
width,
|
||||
height,
|
||||
fps,
|
||||
frame_idx: 0,
|
||||
force_kf: false,
|
||||
intra_refresh,
|
||||
intra_refresh_period: if intra_refresh {
|
||||
intra_refresh_period(fps).max(1) as u32
|
||||
} else {
|
||||
0
|
||||
},
|
||||
args: OpenArgs {
|
||||
codec,
|
||||
format,
|
||||
width,
|
||||
height,
|
||||
fps,
|
||||
bitrate_bps,
|
||||
cuda,
|
||||
bit_depth,
|
||||
chroma,
|
||||
},
|
||||
})
|
||||
}
|
||||
}
|
||||
|
||||
impl Encoder for NvencEncoder {
|
||||
fn caps(&self) -> super::EncoderCaps {
|
||||
super::EncoderCaps {
|
||||
// 4:4:4 iff this session opened FREXT — the CPU swscale path or the zero-copy GPU
|
||||
// convert. RFI/HDR-SEI stay unsupported on libavcodec NVENC (the trait defaults).
|
||||
chroma_444: self.want_444,
|
||||
intra_refresh: self.intra_refresh,
|
||||
// NVENC intra-refresh is purpose-built GDR loss recovery (moving band + recovery-point
|
||||
// SEI): the wave heals a lost picture within one period, so mark the boundary AUs and let
|
||||
// the client re-anchor on them instead of forcing a full IDR. Tied to `intra_refresh`
|
||||
// (already the `PUNKTFUNK_INTRA_REFRESH` opt-in), unlike AMF/QSV which stay unvalidated.
|
||||
intra_refresh_recovery: self.intra_refresh,
|
||||
intra_refresh_period: self.intra_refresh_period,
|
||||
..super::EncoderCaps::default()
|
||||
}
|
||||
}
|
||||
|
||||
fn submit(&mut self, captured: &CapturedFrame) -> Result<()> {
|
||||
anyhow::ensure!(
|
||||
captured.width == self.width && captured.height == self.height,
|
||||
"captured frame {}x{} != encoder {}x{}",
|
||||
captured.width,
|
||||
captured.height,
|
||||
self.width,
|
||||
self.height
|
||||
);
|
||||
let pts = self.frame_idx;
|
||||
self.frame_idx += 1;
|
||||
// Force an IDR when requested (client RFI); otherwise let NVENC pick (GOP/P-frame).
|
||||
let idr = self.force_kf;
|
||||
self.force_kf = false;
|
||||
match &captured.payload {
|
||||
FramePayload::Cuda(buf) => self.submit_cuda(buf, pts, idr),
|
||||
FramePayload::Cpu(bytes) => self.submit_cpu(bytes, captured.format, pts, idr),
|
||||
FramePayload::Dmabuf(_) => {
|
||||
bail!("NVENC got a VAAPI dmabuf frame — capture/encoder backend mismatch")
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
fn request_keyframe(&mut self) {
|
||||
self.force_kf = true;
|
||||
}
|
||||
|
||||
/// Encode-stall recovery: drop the wedged libavcodec encoder and reopen it fresh with the
|
||||
/// session's negotiated parameters (the stored [`OpenArgs`]) — the drop-and-reopen lever the
|
||||
/// QSV/VAAPI paths use, so the encode-stall watchdog can heal a wedged NVENC/driver instead of
|
||||
/// ending the session. Owed AUs are forfeited; the fresh encoder opens on an IDR.
|
||||
fn reset(&mut self) -> bool {
|
||||
let a = self.args;
|
||||
match Self::open(
|
||||
a.codec,
|
||||
a.format,
|
||||
a.width,
|
||||
a.height,
|
||||
a.fps,
|
||||
a.bitrate_bps,
|
||||
a.cuda,
|
||||
a.bit_depth,
|
||||
a.chroma,
|
||||
) {
|
||||
Ok(mut fresh) => {
|
||||
fresh.force_kf = true;
|
||||
*self = fresh; // drops the wedged encoder (frees its contexts) in the same step
|
||||
true
|
||||
}
|
||||
Err(e) => {
|
||||
tracing::error!(error = %format!("{e:#}"), "NVENC in-place reopen failed");
|
||||
false
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
fn poll(&mut self) -> Result<Option<EncodedFrame>> {
|
||||
// Non-blocking single drain: a packet ships, EAGAIN (need another input frame) and EOF
|
||||
// (drained after flush) both mean "nothing this tick".
|
||||
match poll_encoder(&mut self.enc, self.fps)? {
|
||||
PollOutcome::Packet(au) => Ok(Some(au)),
|
||||
PollOutcome::Again | PollOutcome::Eof => Ok(None),
|
||||
}
|
||||
}
|
||||
|
||||
fn flush(&mut self) -> Result<()> {
|
||||
self.enc.send_eof().context("send_eof")?;
|
||||
Ok(())
|
||||
}
|
||||
}
|
||||
|
||||
impl NvencEncoder {
|
||||
/// CPU path: expand/copy the packed RGB/BGR bytes into the reusable 4-bpp frame, then send.
|
||||
fn submit_cpu(&mut self, bytes: &[u8], format: PixelFormat, pts: i64, idr: bool) -> Result<()> {
|
||||
anyhow::ensure!(
|
||||
format == self.src_format,
|
||||
"captured format {:?} != encoder source {:?}",
|
||||
format,
|
||||
self.src_format
|
||||
);
|
||||
let w = self.width as usize;
|
||||
let h = self.height as usize;
|
||||
let src_bpp = self.src_format.bytes_per_pixel();
|
||||
let src_row = w * src_bpp;
|
||||
anyhow::ensure!(
|
||||
bytes.len() >= src_row * h,
|
||||
"captured buffer {} bytes < required {}",
|
||||
bytes.len(),
|
||||
src_row * h
|
||||
);
|
||||
// 4:4:4: swscale the packed RGB straight into the planar YUV444P input frame (BT.709 limited),
|
||||
// then send it — no byte-expand. The 4:2:0 RGB path (below) feeds NVENC packed RGB directly.
|
||||
if let Some(sws) = self.sws_444 {
|
||||
let frame = self
|
||||
.frame
|
||||
.as_mut()
|
||||
.context("CPU frame missing (encoder opened in CUDA mode)")?;
|
||||
// SAFETY: `format == self.src_format` and `bytes.len() >= src_row * h` (the `ensure!`s
|
||||
// above), so `sws_scale` reads `h` rows of `src_row` bytes from `src_data[0] = bytes`
|
||||
// (packed RGB is single-plane; the other src planes are null/0) — all in bounds. `sws` is
|
||||
// the non-null context built in `open`. The dst is `frame`'s underlying `AVFrame`: its
|
||||
// `data`/`linesize` in-struct arrays were sized for YUV444P by `VideoFrame::new`, and the
|
||||
// 3 planes are each `width`×`height`. All pointers are live locals for this synchronous
|
||||
// call; the encoder runs only on this thread (`unsafe impl Send`), so no aliasing/race.
|
||||
unsafe {
|
||||
let dst_av = frame.as_mut_ptr();
|
||||
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(
|
||||
sws,
|
||||
src_data.as_ptr(),
|
||||
src_stride.as_ptr(),
|
||||
0,
|
||||
h as c_int,
|
||||
(*dst_av).data.as_ptr(),
|
||||
(*dst_av).linesize.as_ptr(),
|
||||
);
|
||||
if r < 0 {
|
||||
bail!("sws_scale(RGB→YUV444P) failed ({r})");
|
||||
}
|
||||
}
|
||||
frame.set_pts(Some(pts));
|
||||
frame.set_kind(if idr {
|
||||
ffmpeg::picture::Type::I
|
||||
} else {
|
||||
ffmpeg::picture::Type::None
|
||||
});
|
||||
self.enc.send_frame(frame).context("send_frame(444)")?;
|
||||
return Ok(());
|
||||
}
|
||||
let frame = self
|
||||
.frame
|
||||
.as_mut()
|
||||
.context("CPU frame missing (encoder opened in CUDA mode)")?;
|
||||
let stride = frame.stride(0); // dst is 4-bpp, aligned
|
||||
let dst = frame.data_mut(0);
|
||||
if self.expand {
|
||||
// packed 3-bpp RGB/BGR → 4-bpp *0 (copy 3 bytes, zero the pad byte)
|
||||
for y in 0..h {
|
||||
let s = &bytes[y * src_row..y * src_row + src_row];
|
||||
let drow = &mut dst[y * stride..y * stride + w * 4];
|
||||
for x in 0..w {
|
||||
drow[x * 4..x * 4 + 3].copy_from_slice(&s[x * 3..x * 3 + 3]);
|
||||
drow[x * 4 + 3] = 0;
|
||||
}
|
||||
}
|
||||
} else {
|
||||
// 4-bpp → 4-bpp, honoring the (possibly larger) dst stride
|
||||
for y in 0..h {
|
||||
dst[y * stride..y * stride + src_row]
|
||||
.copy_from_slice(&bytes[y * src_row..y * src_row + src_row]);
|
||||
}
|
||||
}
|
||||
frame.set_pts(Some(pts));
|
||||
frame.set_kind(if idr {
|
||||
ffmpeg::picture::Type::I
|
||||
} else {
|
||||
ffmpeg::picture::Type::None
|
||||
});
|
||||
self.enc.send_frame(frame).context("send_frame")?;
|
||||
Ok(())
|
||||
}
|
||||
|
||||
/// Zero-copy path: hand the imported CUDA device buffer to NVENC with no CPU touch.
|
||||
///
|
||||
/// We take a *pooled* surface from the CUDA hwframes context (`av_hwframe_get_buffer`) and
|
||||
/// device→device-copy our imported buffer into it, rather than wrapping our own pointer in a
|
||||
/// bare frame. Two reasons: (1) NVENC's `nvenc_send_frame` ignores frames whose `buf[0]` is
|
||||
/// null and the generic encode path's `av_frame_ref` needs a refcounted buffer — a bare
|
||||
/// frame is rejected with `EINVAL`; (2) NVENC caches CUDA-resource *registrations* keyed by
|
||||
/// device pointer with a bounded table, so a fresh pointer every frame would thrash/overflow
|
||||
/// it — the pool recycles a small set of pointers. The extra copy is device-local (~8 MB at
|
||||
/// 1080p, sub-millisecond on the GPU) and keeps the host fully off the pixel path.
|
||||
fn submit_cuda(&mut self, buf: &pf_zerocopy::DeviceBuffer, pts: i64, idr: bool) -> Result<()> {
|
||||
let frames_ref = self
|
||||
.cuda
|
||||
.as_ref()
|
||||
.context("CUDA hw context missing (encoder opened in CPU mode)")?
|
||||
.frames_ref;
|
||||
// The device→device copy below uses our shared context directly; make it current on the
|
||||
// encode thread (ffmpeg pushes its own around the pool alloc, so order is fine).
|
||||
pf_zerocopy::cuda::make_current().context("CUDA context current (encode thread)")?;
|
||||
// SAFETY: `frames_ref` is the non-null CUDA frames ctx from `self.cuda` (unwrapped via
|
||||
// `.context(..)?` above), and the shared CUDA context was just made current on THIS thread
|
||||
// (`make_current()?`), the precondition for the device-pointer copies below.
|
||||
// * `av_frame_alloc` → `f` (null-checked). `av_hwframe_get_buffer(frames_ref, f, 0)` fills `f`
|
||||
// with a pooled CUDA surface (sets `data[]`/`linesize[]`/`buf[0]`/`hw_frames_ctx`); on
|
||||
// failure we free `f` and bail.
|
||||
// * For NV12 we read `(*f).data[0..2]` / `linesize[0..2]` (Y + interleaved UV), else
|
||||
// `data[0]`/`linesize[0]` — in-struct fields of the non-null `f`, valid for the surface dims
|
||||
// ffmpeg allocated — and pass them to the cuda copy helpers, which device→device copy `buf`
|
||||
// (the imported `DeviceBuffer`, owned by the caller and live for this call) into the surface.
|
||||
// * On copy error we free `f` and return. Otherwise we write `pts`/`pict_type` through `f` and
|
||||
// `avcodec_send_frame` it into the live owned `self.enc` context (which takes its own ref of
|
||||
// the pooled surface), then free our `f` ref exactly once. Single-threaded encoder → no race.
|
||||
unsafe {
|
||||
let mut f = ffi::av_frame_alloc();
|
||||
if f.is_null() {
|
||||
bail!("av_frame_alloc failed");
|
||||
}
|
||||
// Pooled CUDA surface: sets format, width/height, data[0]/linesize[0], buf[0] and
|
||||
// hw_frames_ctx. Reused across frames (the pool recycles), keeping NVENC's
|
||||
// registration cache warm.
|
||||
let r = ffi::av_hwframe_get_buffer(frames_ref, f, 0);
|
||||
if r < 0 {
|
||||
ffi::av_frame_free(&mut f);
|
||||
bail!("av_hwframe_get_buffer(CUDA) failed ({r})");
|
||||
}
|
||||
// NV12 surfaces are two-plane (Y in data[0], interleaved UV in data[1]); YUV444
|
||||
// surfaces are three-plane (`yuv444p` frames ctx — data[0..3]); the RGB surfaces are
|
||||
// single-plane. Copy the matching layout into NVENC's pooled surface. A 4:4:4 session
|
||||
// whose buffer ISN'T YUV444 (a LINEAR/gamescope capture the worker can't convert)
|
||||
// fails loudly here rather than letting `hevc_nvenc` silently subsample RGB to 4:2:0.
|
||||
let copy_res = if buf.yuv444 {
|
||||
let dsts = core::array::from_fn(|i| {
|
||||
(
|
||||
(*f).data[i] as pf_zerocopy::cuda::CUdeviceptr,
|
||||
(*f).linesize[i] as usize,
|
||||
)
|
||||
});
|
||||
pf_zerocopy::cuda::copy_yuv444_to_device(buf, dsts)
|
||||
} else if self.want_444 {
|
||||
ffi::av_frame_free(&mut f);
|
||||
bail!(
|
||||
"4:4:4 session but the zero-copy frame is not YUV444 (LINEAR/gamescope \
|
||||
capture has no GPU 4:4:4 convert) — unset PUNKTFUNK_ZEROCOPY to use the \
|
||||
CPU 4:4:4 path on this compositor"
|
||||
);
|
||||
} else if buf.is_nv12() {
|
||||
let y_ptr = (*f).data[0] as pf_zerocopy::cuda::CUdeviceptr;
|
||||
let y_pitch = (*f).linesize[0] as usize;
|
||||
let uv_ptr = (*f).data[1] as pf_zerocopy::cuda::CUdeviceptr;
|
||||
let uv_pitch = (*f).linesize[1] as usize;
|
||||
pf_zerocopy::cuda::copy_nv12_to_device(buf, y_ptr, y_pitch, uv_ptr, uv_pitch)
|
||||
} else {
|
||||
let dst_ptr = (*f).data[0] as pf_zerocopy::cuda::CUdeviceptr;
|
||||
let dst_pitch = (*f).linesize[0] as usize;
|
||||
pf_zerocopy::cuda::copy_device_to_device(buf, dst_ptr, dst_pitch)
|
||||
};
|
||||
if let Err(e) = copy_res {
|
||||
ffi::av_frame_free(&mut f);
|
||||
return Err(e).context("copy imported buffer into NVENC surface");
|
||||
}
|
||||
(*f).pts = pts;
|
||||
(*f).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(), f);
|
||||
ffi::av_frame_free(&mut f);
|
||||
if r < 0 {
|
||||
bail!("avcodec_send_frame(CUDA) failed ({r})");
|
||||
}
|
||||
}
|
||||
Ok(())
|
||||
}
|
||||
}
|
||||
|
||||
impl Drop for NvencEncoder {
|
||||
fn drop(&mut self) {
|
||||
if let Some(sws) = self.sws_444.take() {
|
||||
// SAFETY: `sws` is the non-null `SwsContext` allocated by `sws_getContext` in `open` and
|
||||
// owned exclusively by this encoder (taken out of the field so it can't be freed twice).
|
||||
// `sws_freeContext` frees it; nothing else references it after this single-threaded drop.
|
||||
unsafe { ffi::sws_freeContext(sws) };
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
/// Probe whether this NVIDIA GPU + driver + libavcodec can actually encode HEVC **4:4:4** (Range
|
||||
/// Extensions). Opens a tiny real `hevc_nvenc` 4:4:4 session — the exact path [`NvencEncoder::open`]
|
||||
/// takes for a live 4:4:4 stream — and reports whether it succeeded. HEVC-only; the result is cached
|
||||
/// by the caller ([`crate::can_encode_444`]). A GPU/driver/ffmpeg without RExt 4:4:4 fails
|
||||
/// the open here, so the host resolves the session to 4:2:0 before the Welcome (honest downgrade).
|
||||
pub fn probe_can_encode_444(codec: Codec) -> bool {
|
||||
if codec != Codec::H265 {
|
||||
return false;
|
||||
}
|
||||
if ffmpeg::init().is_err() {
|
||||
return false;
|
||||
}
|
||||
// Quiet ffmpeg's open error on a GPU that lacks 4:4:4 — the probe failing is an expected outcome.
|
||||
// SAFETY: libav initialized above; `av_log_{get,set}_level` only read/write the global int level
|
||||
// (no pointer args) and are always sound post-init.
|
||||
let prev = unsafe {
|
||||
let p = ffi::av_log_get_level();
|
||||
ffi::av_log_set_level(ffi::AV_LOG_FATAL);
|
||||
p
|
||||
};
|
||||
let ok = NvencEncoder::open(
|
||||
codec,
|
||||
PixelFormat::Bgra,
|
||||
640,
|
||||
480,
|
||||
30,
|
||||
2_000_000,
|
||||
false, // CPU input (the 4:4:4 path never uses CUDA)
|
||||
8,
|
||||
ChromaFormat::Yuv444,
|
||||
)
|
||||
.is_ok();
|
||||
// SAFETY: restore the saved global log level (scalar arg, no pointers).
|
||||
unsafe { ffi::av_log_set_level(prev) };
|
||||
ok
|
||||
}
|
||||
Reference in New Issue
Block a user