enricobuehler
75c113f331
The gpu-contention plan's §5.B lever: today submit and the blocking lock_bitstream share one thread, so under a GPU-saturating game the pipeline serializes on the WDDM scheduling wait (1000/17ms ≈ 59fps — the depth-1 collapse; the old 'deeper pipeline just stacks latency' result was a same-thread implementation, not a disproof). Async mode opens the session enableEncodeAsync=1, registers an auto-reset completion event per pool bitstream, and moves the wait+lock+copy+ unlock onto an internal retrieve thread feeding poll() through a channel — the exact split the NVENC guide mandates. Register/map/unmap stay on the encode thread; teardown drops the job channel, joins the thread, THEN destroys the session. In-flight depth is bounded by PUNKTFUNK_NVENC_ASYNC_DEPTH (default 4, hard cap POOL-1) — both for output-buffer reuse and because NVENC encodes the capture ring's textures in place. Idle latency cost ≈ 0 (same-tick pickup); under contention completed frames queue instead of stalling capture. CI-compile validated only — on-glass A/B under game load on the RTX box still pending (box offline). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
1348 lines
70 KiB
Rust
1348 lines
70 KiB
Rust
//! NVENC hardware encoder (Windows, D3D11 input) — zero-copy capture→encode on the GPU.
|
|
//!
|
|
//! Drives the raw NVENC API via `nvidia_video_codec_sdk::{sys, ENCODE_API}` (the safe `Encoder`
|
|
//! wrapper is CUDA-only). Opens an encode session bound to the **same** `ID3D11Device` as the DXGI
|
|
//! capturer (the device is carried on `FramePayload::D3d11`), and **encodes the capturer's texture in
|
|
//! place** — it registers each input texture with NVENC once (cached by pointer) and `encode_picture`s
|
|
//! it directly, with NO per-frame `CopyResource`. (That's safe because the host encode loop is
|
|
//! synchronous — capture → submit → poll, where `poll`/`lock_bitstream` blocks until the encode
|
|
//! finishes — so the capturer never overwrites the texture mid-encode; if that loop ever becomes
|
|
//! pipelined, the capturer must hand a ring of textures.) Mirrors the Linux NVENC config: CBR +
|
|
//! ultra-low-latency, infinite GOP, P-frames only, forced-IDR for RFI, in-band SPS/PPS each keyframe.
|
|
//!
|
|
//! Needs a real NVIDIA GPU at runtime (session creation fails otherwise) — compiles GPU-less, but
|
|
//! `open`/`submit` only succeed on a GPU box. The software encoder (`super::sw`) is the fallback.
|
|
//!
|
|
//! **Two-thread async retrieve** (`PUNKTFUNK_NVENC_ASYNC=1`, opt-in until on-glass validated —
|
|
//! gpu-contention plan §5.B): the NVENC guide mandates that the main thread only *submit*
|
|
//! (`nvEncEncodePicture`) while a **secondary thread** waits on per-buffer completion events and
|
|
//! does `nvEncLockBitstream`. Today's sync mode does both on one thread, so under a GPU-saturating
|
|
//! game the whole pipeline serializes on the WDDM scheduling wait (`1000/17ms ≈ 59 fps` — the
|
|
//! depth-1 collapse). In async mode the session is opened `enableEncodeAsync=1`, each output
|
|
//! bitstream gets a registered auto-reset event, `submit` returns immediately, and an internal
|
|
//! retrieve thread waits + locks + copies + unlocks, handing finished AUs back through a channel
|
|
//! that `poll` drains without blocking. All input-resource calls (register/map/unmap) stay on the
|
|
//! encode thread; the retrieve thread touches ONLY the event + lock/unlock — the exact split the
|
|
//! guide blesses. Backpressure: `submit` blocks on the oldest completion when `POOL - 1` encodes
|
|
//! are in flight, so an output buffer is never reused mid-encode. Latency cost when idle ≈ 0 (the
|
|
//! AU completes within the same tick and `poll` picks it up); under contention completed frames
|
|
//! queue instead of stalling capture — throughput recovers up to the scheduler-granted share.
|
|
|
|
// Every `unsafe` block / impl in this file carries a `// SAFETY:` proof; enforce it.
|
|
#![deny(clippy::undocumented_unsafe_blocks)]
|
|
|
|
use super::{ChromaFormat, Codec, EncodedFrame, Encoder, EncoderCaps};
|
|
use crate::capture::{CapturedFrame, FramePayload, PixelFormat};
|
|
use anyhow::{anyhow, bail, Context, Result};
|
|
use std::collections::{HashMap, VecDeque};
|
|
use std::ffi::c_void;
|
|
use std::ptr;
|
|
use std::sync::mpsc;
|
|
use windows::core::{Interface, PCWSTR};
|
|
use windows::Win32::Foundation::{CloseHandle, HANDLE, WAIT_OBJECT_0};
|
|
use windows::Win32::Graphics::Direct3D11::{ID3D11Device, ID3D11Texture2D};
|
|
use windows::Win32::System::Threading::{CreateEventW, WaitForSingleObject};
|
|
|
|
use nvidia_video_codec_sdk::sys::nvEncodeAPI as nv;
|
|
use nvidia_video_codec_sdk::ENCODE_API as API;
|
|
|
|
// Output bitstream buffers = max in-flight encodes. The helper deep-pipelines (submits several frames
|
|
// before locking the oldest) so per-frame GPU-scheduling waits OVERLAP instead of serializing under a
|
|
// GPU-saturating game; this must be ≥ the helper's `PUNKTFUNK_ENCODE_DEPTH` (default 4, clamped ≤ 6).
|
|
const POOL: usize = 8;
|
|
|
|
/// Reference-frame DPB depth when RFI is supported (Apollo uses 5 for H.264/HEVC). A deeper DPB
|
|
/// lets an invalidated reference fall back to an older still-valid frame instead of a full IDR;
|
|
/// `numRefL0 = 1` keeps each P-frame single-reference for low latency.
|
|
const RFI_DPB: u32 = 5;
|
|
|
|
fn codec_guid(codec: Codec) -> nv::GUID {
|
|
match codec {
|
|
Codec::H264 => nv::NV_ENC_CODEC_H264_GUID,
|
|
Codec::H265 => nv::NV_ENC_CODEC_HEVC_GUID,
|
|
Codec::Av1 => nv::NV_ENC_CODEC_AV1_GUID,
|
|
}
|
|
}
|
|
|
|
/// Whether the operator asked for the two-thread async retrieve (`PUNKTFUNK_NVENC_ASYNC` truthy).
|
|
/// Combined with the GPU's `NV_ENC_CAPS_ASYNC_ENCODE_SUPPORT` in `init_session`. Opt-in until
|
|
/// on-glass validated; note an async-rejecting config surfaces as a failed session open — unset
|
|
/// the env in that case.
|
|
fn async_retrieve_requested() -> bool {
|
|
std::env::var("PUNKTFUNK_NVENC_ASYNC")
|
|
.map(|v| matches!(v.trim(), "1" | "true" | "yes" | "on"))
|
|
.unwrap_or(false)
|
|
}
|
|
|
|
/// Max encodes in flight in async mode (`PUNKTFUNK_NVENC_ASYNC_DEPTH`, default 4, clamped
|
|
/// `2..=POOL-1`). Two independent ceilings meet here: the output-bitstream pool (hard, `POOL-1` —
|
|
/// a buffer must never be reused mid-encode) and the capturer's texture ring (soft — NVENC encodes
|
|
/// the ring textures in place, so in-flight depth beyond the ring lets the capturer overwrite a
|
|
/// frame mid-encode: visual corruption, not UB). IDD-push rings are sized around
|
|
/// `PUNKTFUNK_IDD_DEPTH`; raise both together if deeper pipelining is needed.
|
|
fn async_inflight_cap() -> usize {
|
|
std::env::var("PUNKTFUNK_NVENC_ASYNC_DEPTH")
|
|
.ok()
|
|
.and_then(|s| s.parse::<usize>().ok())
|
|
.unwrap_or(4)
|
|
.clamp(2, POOL - 1)
|
|
}
|
|
|
|
/// One in-flight encode handed to the retrieve thread: the output bitstream to lock once its
|
|
/// completion `event` signals. Raw pointers travel as `usize` (the addresses are process-global
|
|
/// driver handles; the thread is joined before the session they belong to is destroyed).
|
|
struct RetrieveJob {
|
|
bs: usize,
|
|
event: usize,
|
|
}
|
|
|
|
/// A finished retrieve: the locked-and-copied AU (or the retrieve-side error) for the oldest
|
|
/// in-flight bitstream. `bs` lets the encode thread cross-check FIFO pairing with `pending`.
|
|
struct RetrieveDone {
|
|
bs: usize,
|
|
result: std::result::Result<(Vec<u8>, bool), String>,
|
|
}
|
|
|
|
/// The async-retrieve runtime: the job channel feeding the retrieve thread, the completion channel
|
|
/// back, the thread handle (joined in `teardown` BEFORE the session is destroyed), and AUs already
|
|
/// absorbed by backpressure that `poll` hands out first.
|
|
struct AsyncRetrieve {
|
|
work_tx: Option<mpsc::SyncSender<RetrieveJob>>,
|
|
done_rx: mpsc::Receiver<RetrieveDone>,
|
|
join: Option<std::thread::JoinHandle<()>>,
|
|
ready: VecDeque<EncodedFrame>,
|
|
}
|
|
|
|
/// The retrieve-thread body (gpu-contention plan §5.B): for each submitted frame, wait on its
|
|
/// completion event, lock the bitstream, copy the AU out, unlock, and send it back. Exits when the
|
|
/// job channel closes (teardown drops the sender and joins BEFORE destroying the session, so
|
|
/// `enc`/`bs`/`event` outlive every use here). Touches ONLY the event wait + lock/unlock — the
|
|
/// NVENC threading model's sanctioned secondary-thread surface.
|
|
fn retrieve_loop(
|
|
enc: usize,
|
|
work_rx: mpsc::Receiver<RetrieveJob>,
|
|
done_tx: mpsc::Sender<RetrieveDone>,
|
|
) {
|
|
crate::punktfunk1::boost_thread_priority(false);
|
|
while let Ok(job) = work_rx.recv() {
|
|
// SAFETY: `job.event` is one of the auto-reset events `init_session` created and
|
|
// registered for exactly this session, and `job.bs` one of its pool bitstreams; both stay
|
|
// valid until `teardown`, which joins this thread first. `WaitForSingleObject` takes the
|
|
// handle by value. On WAIT_OBJECT_0 the driver has completed the encode into `job.bs`, so
|
|
// `lock_bitstream` (version set, struct a live stack local for the synchronous call)
|
|
// yields a CPU-readable `bitstreamBufferPtr`/`bitstreamSizeInBytes` valid until
|
|
// `unlock_bitstream`; the slice is copied (`to_vec`) before the unlock on the same buffer.
|
|
// Lock/unlock from a secondary thread while the encode thread submits is the NVENC
|
|
// guide's documented threading model.
|
|
let result = unsafe {
|
|
if WaitForSingleObject(HANDLE(job.event as *mut c_void), 5000) != WAIT_OBJECT_0 {
|
|
Err("NVENC completion event timeout (5s) — encoder wedged?".to_string())
|
|
} else {
|
|
let mut lock = nv::NV_ENC_LOCK_BITSTREAM {
|
|
version: nv::NV_ENC_LOCK_BITSTREAM_VER,
|
|
outputBitstream: job.bs as *mut c_void,
|
|
..Default::default()
|
|
};
|
|
match (API.lock_bitstream)(enc as *mut c_void, &mut lock).result_without_string() {
|
|
Ok(()) => {
|
|
let data = std::slice::from_raw_parts(
|
|
lock.bitstreamBufferPtr as *const u8,
|
|
lock.bitstreamSizeInBytes as usize,
|
|
)
|
|
.to_vec();
|
|
let keyframe = matches!(
|
|
lock.pictureType,
|
|
nv::NV_ENC_PIC_TYPE::NV_ENC_PIC_TYPE_IDR
|
|
| nv::NV_ENC_PIC_TYPE::NV_ENC_PIC_TYPE_I
|
|
);
|
|
let _ = (API.unlock_bitstream)(enc as *mut c_void, job.bs as *mut c_void);
|
|
Ok((data, keyframe))
|
|
}
|
|
Err(e) => Err(format!("lock_bitstream (async): {e:?}")),
|
|
}
|
|
}
|
|
};
|
|
if done_tx.send(RetrieveDone { bs: job.bs, result }).is_err() {
|
|
break; // encoder side gone (teardown drains us via join)
|
|
}
|
|
}
|
|
}
|
|
|
|
pub struct NvencD3d11Encoder {
|
|
encoder: *mut c_void,
|
|
codec: Codec,
|
|
codec_guid: nv::GUID,
|
|
width: u32,
|
|
height: u32,
|
|
fps: u32,
|
|
bitrate_bps: u64,
|
|
buffer_fmt: nv::NV_ENC_BUFFER_FORMAT,
|
|
/// Encoded bit depth (8 or 10). 10 → HEVC Main10 (NVENC upconverts the 8-bit ARGB input).
|
|
bit_depth: u8,
|
|
/// Full-chroma 4:4:4 (HEVC Range Extensions, `chroma_format_idc = 3`) requested for this session.
|
|
/// NVENC ingests the RGB (ARGB/ABGR10) input and CSCs it to YUV444 internally; the `FREXT` profile
|
|
/// and `chromaFormatIDC = 3` in the encode config carry the chroma. Gated on the GPU's
|
|
/// `NV_ENC_CAPS_SUPPORT_YUV444_ENCODE` (cleared in `query_caps` on a card that lacks it) and on an
|
|
/// RGB input format (NV12/P010 capture can't reconstruct 4:4:4). HEVC-only.
|
|
chroma_444: bool,
|
|
/// `NV_ENC_CAPS_SUPPORT_YUV444_ENCODE` from the caps probe — whether this GPU can 4:4:4 encode at
|
|
/// all. `chroma_444` is forced off when this is false (graceful downgrade to 4:2:0).
|
|
yuv444_supported: bool,
|
|
/// HDR: the capturer is delivering BT.2020 PQ 10-bit (`PixelFormat::Rgb10a2`) frames. Sets the
|
|
/// `ABGR10` input format + the BT.2020/PQ colour VUI. Derived per-frame from the capture format
|
|
/// (HDR can toggle mid-session); a change re-inits the session.
|
|
hdr: bool,
|
|
/// The source's static HDR mastering metadata (from the capturer's `GetDesc1`), emitted as
|
|
/// in-band SEI (`mastering_display_colour_volume` + `content_light_level_info`) on each keyframe
|
|
/// when `hdr`. `None` = unknown → no SEI (the VUI still signals BT.2020 PQ). Set per-frame via
|
|
/// [`Encoder::set_hdr_meta`], so a mid-session regrade is picked up on the next keyframe.
|
|
hdr_meta: Option<punktfunk_core::quic::HdrMeta>,
|
|
/// Registrations of the capturer's input textures, cached by texture raw pointer — NVENC encodes
|
|
/// them in place (no per-frame copy). The cloned `ID3D11Texture2D` keeps each alive until we
|
|
/// unregister it (the capturer may drop its copy on a device recreate before our teardown runs).
|
|
regs: HashMap<isize, (nv::NV_ENC_REGISTERED_PTR, ID3D11Texture2D)>,
|
|
next: usize,
|
|
bitstreams: Vec<nv::NV_ENC_OUTPUT_PTR>,
|
|
/// Async mode: the registered completion event per pool bitstream (raw `HANDLE` as `usize`,
|
|
/// parallel to `bitstreams`); empty in sync mode. Unregistered + closed in `teardown`.
|
|
events: Vec<usize>,
|
|
/// Async mode: the retrieve thread + its channels (`None` = classic same-thread sync retrieve).
|
|
async_rt: Option<AsyncRetrieve>,
|
|
/// `NV_ENC_CAPS_ASYNC_ENCODE_SUPPORT` from the caps probe — gates the async retrieve mode.
|
|
async_supported: bool,
|
|
/// (bitstream, mapped input resource to unmap after retrieval, pts_ns) per in-flight encode.
|
|
pending: VecDeque<(nv::NV_ENC_OUTPUT_PTR, nv::NV_ENC_INPUT_PTR, u64)>,
|
|
frame_idx: i64,
|
|
force_kf: bool,
|
|
inited: bool,
|
|
/// GPU capabilities probed once via `nvEncGetEncodeCaps` before configuring (Apollo's
|
|
/// `get_encoder_cap`): gates 10-bit/custom-VBV/RFI on what this card actually supports instead
|
|
/// of failing later as an opaque `InvalidParam`. Set by [`query_caps`](Self::query_caps).
|
|
rfi_supported: bool,
|
|
custom_vbv: bool,
|
|
/// The last reference-frame range we invalidated — dedupes repeated RFI requests for the same
|
|
/// loss event (the client resends until it sees recovery).
|
|
last_rfi_range: Option<(i64, i64)>,
|
|
/// Raw ptr of the D3D11 device this session was initialized with. The capturer recreates the
|
|
/// device on a desktop switch (normal ↔ Winlogon secure); when a frame carries a new device we
|
|
/// tear down and re-init NVENC against it.
|
|
init_device: *mut c_void,
|
|
}
|
|
|
|
// SAFETY: the `!Send` fields are the raw NVENC session/device handles (`encoder`, `init_device`),
|
|
// the raw NVENC bitstream/registered/mapped pointers carried in `bitstreams`/`regs`/`pending`, and
|
|
// the `ID3D11Texture2D` COM refs — none of which may be touched concurrently from two threads
|
|
// EXCEPT along the NVENC guide's sanctioned split. The encoder object is owned by exactly one
|
|
// thread: it is moved onto the host encode thread once at construction, and every method
|
|
// (`submit`/`poll`/`invalidate_ref_frames`/`Drop`) runs there. In async mode the internal retrieve
|
|
// thread additionally calls `WaitForSingleObject`/`lock_bitstream`/`unlock_bitstream` on the same
|
|
// session — the exact two-thread model the NVENC API documents as thread-safe (submit-side vs
|
|
// output-side); it never touches registrations, mappings, or D3D11. `teardown` joins that thread
|
|
// BEFORE destroying the session, so no retrieve call can outlive the handles. Moving the encoder
|
|
// across its single ownership-transfer boundary is sound because no NVENC/D3D11 call is in flight
|
|
// during the move — so `Send` introduces no data race on the non-`Send` fields.
|
|
unsafe impl Send for NvencD3d11Encoder {}
|
|
|
|
impl NvencD3d11Encoder {
|
|
#[allow(clippy::too_many_arguments)]
|
|
pub fn open(
|
|
codec: Codec,
|
|
_format: PixelFormat,
|
|
width: u32,
|
|
height: u32,
|
|
fps: u32,
|
|
bitrate_bps: u64,
|
|
bit_depth: u8,
|
|
chroma: ChromaFormat,
|
|
) -> Result<Self> {
|
|
Ok(Self {
|
|
encoder: ptr::null_mut(),
|
|
codec,
|
|
codec_guid: codec_guid(codec),
|
|
width,
|
|
height,
|
|
fps,
|
|
bitrate_bps,
|
|
buffer_fmt: nv::NV_ENC_BUFFER_FORMAT::NV_ENC_BUFFER_FORMAT_ARGB,
|
|
bit_depth,
|
|
// 4:4:4 is HEVC-only; the GPU-support gate is applied in `query_caps`.
|
|
chroma_444: chroma.is_444() && codec == Codec::H265,
|
|
yuv444_supported: false,
|
|
hdr: false,
|
|
hdr_meta: None,
|
|
regs: HashMap::new(),
|
|
next: 0,
|
|
bitstreams: Vec::new(),
|
|
events: Vec::new(),
|
|
async_rt: None,
|
|
async_supported: false,
|
|
pending: VecDeque::new(),
|
|
frame_idx: 0,
|
|
force_kf: false,
|
|
inited: false,
|
|
rfi_supported: false,
|
|
custom_vbv: false,
|
|
last_rfi_range: None,
|
|
init_device: ptr::null_mut(),
|
|
})
|
|
}
|
|
|
|
/// Tear down the encode session + pooled resources. Reused on a capture-device change (desktop
|
|
/// switch) and at Drop.
|
|
unsafe fn teardown(&mut self) {
|
|
if self.encoder.is_null() {
|
|
return;
|
|
}
|
|
// Async mode: retire the retrieve thread FIRST — drop the job sender so it finishes every
|
|
// queued job (each references the still-live session) and exits, then join. Only after the
|
|
// join is it sound to unmap/destroy anything the thread might have been touching.
|
|
if let Some(mut rt) = self.async_rt.take() {
|
|
drop(rt.work_tx.take());
|
|
if let Some(j) = rt.join.take() {
|
|
let _ = j.join();
|
|
}
|
|
// Completions the thread produced that poll() never absorbed — their AUs are dropped
|
|
// (the session is going away), but the FIFO pairing stands, so nothing extra to do
|
|
// beyond the pending unmap below.
|
|
while rt.done_rx.try_recv().is_ok() {}
|
|
}
|
|
// Unmap any in-flight inputs, then unregister every cached texture and destroy the bitstreams.
|
|
for (_, map, _) in &self.pending {
|
|
if !map.is_null() {
|
|
let _ = (API.unmap_input_resource)(self.encoder, *map);
|
|
}
|
|
}
|
|
for (reg, _tex) in self.regs.values() {
|
|
let _ = (API.unregister_resource)(self.encoder, *reg);
|
|
}
|
|
// Async events: unregister from the session, then close the Win32 handles.
|
|
for &ev in &self.events {
|
|
let mut ep = nv::NV_ENC_EVENT_PARAMS {
|
|
version: nv::NV_ENC_EVENT_PARAMS_VER,
|
|
completionEvent: ev as *mut c_void,
|
|
..Default::default()
|
|
};
|
|
let _ = (API.unregister_async_event)(self.encoder, &mut ep);
|
|
let _ = CloseHandle(HANDLE(ev as *mut c_void));
|
|
}
|
|
self.events.clear();
|
|
for &bs in &self.bitstreams {
|
|
let _ = (API.destroy_bitstream_buffer)(self.encoder, bs);
|
|
}
|
|
let _ = (API.destroy_encoder)(self.encoder);
|
|
self.regs.clear(); // drops the texture clones, releasing our refs
|
|
self.bitstreams.clear();
|
|
self.pending.clear();
|
|
self.encoder = ptr::null_mut();
|
|
self.inited = false;
|
|
self.next = 0;
|
|
// The new session starts with an empty DPB (its first frame is an IDR), so any prior
|
|
// invalidation range is meaningless against it.
|
|
self.last_rfi_range = None;
|
|
}
|
|
|
|
/// Query one `NV_ENC_CAPS` value for this codec on an open session; 0 on any error (treat an
|
|
/// unqueryable cap as "unsupported", the conservative choice).
|
|
unsafe fn get_cap(&self, enc: *mut c_void, which: nv::NV_ENC_CAPS) -> i32 {
|
|
let mut param = nv::NV_ENC_CAPS_PARAM {
|
|
version: nv::NV_ENC_CAPS_PARAM_VER,
|
|
capsToQuery: which,
|
|
reserved: [0; 62],
|
|
};
|
|
let mut val: i32 = 0;
|
|
match (API.get_encode_caps)(enc, self.codec_guid, &mut param, &mut val)
|
|
.result_without_string()
|
|
{
|
|
Ok(()) => val,
|
|
Err(_) => 0,
|
|
}
|
|
}
|
|
|
|
/// Probe this GPU's real capabilities once (Apollo's `get_encoder_cap`) before the bitrate-probe
|
|
/// loop configures the session: opens a throwaway session, queries the codec's max dimensions +
|
|
/// 10-bit / custom-VBV / ref-pic-invalidation support, destroys it. Rejects an out-of-range mode
|
|
/// up front with a clear error, downgrades 10-bit→8-bit when unsupported, and records the
|
|
/// RFI/custom-VBV flags the config + [`invalidate_ref_frames`](Encoder::invalidate_ref_frames)
|
|
/// gate on. Without this, an unsupported config surfaces only as an opaque `InvalidParam` that
|
|
/// the bitrate-clamp search misreads as "bitrate too high" and binary-searches into the floor.
|
|
unsafe fn query_caps(&mut self, device: &ID3D11Device) -> Result<()> {
|
|
let mut params = nv::NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS {
|
|
version: nv::NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS_VER,
|
|
deviceType: nv::NV_ENC_DEVICE_TYPE::NV_ENC_DEVICE_TYPE_DIRECTX,
|
|
device: device.as_raw(),
|
|
apiVersion: nv::NVENCAPI_VERSION,
|
|
..Default::default()
|
|
};
|
|
let mut enc: *mut c_void = ptr::null_mut();
|
|
(API.open_encode_session_ex)(&mut params, &mut enc)
|
|
.result_without_string()
|
|
.map_err(|e| {
|
|
anyhow!("NVENC open_encode_session_ex (caps probe): {e:?} (no NVIDIA GPU?)")
|
|
})?;
|
|
let wmax = self.get_cap(enc, nv::NV_ENC_CAPS::NV_ENC_CAPS_WIDTH_MAX);
|
|
let hmax = self.get_cap(enc, nv::NV_ENC_CAPS::NV_ENC_CAPS_HEIGHT_MAX);
|
|
let ten_bit = self.get_cap(enc, nv::NV_ENC_CAPS::NV_ENC_CAPS_SUPPORT_10BIT_ENCODE);
|
|
let yuv444 = self.get_cap(enc, nv::NV_ENC_CAPS::NV_ENC_CAPS_SUPPORT_YUV444_ENCODE);
|
|
let rfi = self.get_cap(
|
|
enc,
|
|
nv::NV_ENC_CAPS::NV_ENC_CAPS_SUPPORT_REF_PIC_INVALIDATION,
|
|
);
|
|
let custom_vbv = self.get_cap(
|
|
enc,
|
|
nv::NV_ENC_CAPS::NV_ENC_CAPS_SUPPORT_CUSTOM_VBV_BUF_SIZE,
|
|
);
|
|
let async_enc = self.get_cap(enc, nv::NV_ENC_CAPS::NV_ENC_CAPS_ASYNC_ENCODE_SUPPORT);
|
|
let _ = (API.destroy_encoder)(enc);
|
|
|
|
// Reject an over-range mode with a clear message instead of an opaque InvalidParam.
|
|
if wmax > 0 && hmax > 0 && (self.width as i32 > wmax || self.height as i32 > hmax) {
|
|
bail!(
|
|
"this GPU's NVENC max encode size for {:?} is {wmax}x{hmax}; client requested \
|
|
{}x{} (lower the client resolution or use a codec/GPU that supports it)",
|
|
self.codec,
|
|
self.width,
|
|
self.height
|
|
);
|
|
}
|
|
// Degrade gracefully rather than fail: no 10-bit encode on this card → 8-bit SDR.
|
|
if self.bit_depth >= 10 && ten_bit == 0 {
|
|
tracing::warn!("NVENC: this GPU can't 10-bit encode — falling back to 8-bit SDR");
|
|
self.bit_depth = 8;
|
|
self.hdr = false;
|
|
}
|
|
// Same for 4:4:4: a card without YUV444 encode falls back to 4:2:0. (The host already probed
|
|
// this via `probe_can_encode_444` before the Welcome, so this is a belt-and-braces guard.)
|
|
self.yuv444_supported = yuv444 != 0;
|
|
if self.chroma_444 && !self.yuv444_supported {
|
|
tracing::warn!("NVENC: this GPU can't 4:4:4 encode — falling back to 4:2:0");
|
|
self.chroma_444 = false;
|
|
}
|
|
self.rfi_supported = rfi != 0;
|
|
self.custom_vbv = custom_vbv != 0;
|
|
self.async_supported = async_enc != 0;
|
|
tracing::info!(
|
|
rfi = self.rfi_supported,
|
|
custom_vbv = self.custom_vbv,
|
|
async_encode = self.async_supported,
|
|
max = %format!("{wmax}x{hmax}"),
|
|
ten_bit = ten_bit != 0,
|
|
"NVENC capabilities probed"
|
|
);
|
|
Ok(())
|
|
}
|
|
|
|
/// Open + configure + initialize ONE NVENC session at `bitrate` (bps) and `split_mode`. Returns
|
|
/// the session handle, or destroys it and returns the error. NVENC has no re-init after a failed
|
|
/// `initialize_encoder`, so the bitrate-clamp search in `init_session` calls this once per probe.
|
|
unsafe fn try_open_session(
|
|
&self,
|
|
device: &ID3D11Device,
|
|
bitrate: u64,
|
|
split_mode: u32,
|
|
enable_async: bool,
|
|
) -> Result<*mut c_void> {
|
|
let mut params = nv::NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS {
|
|
version: nv::NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS_VER,
|
|
deviceType: nv::NV_ENC_DEVICE_TYPE::NV_ENC_DEVICE_TYPE_DIRECTX,
|
|
device: device.as_raw(),
|
|
apiVersion: nv::NVENCAPI_VERSION,
|
|
..Default::default()
|
|
};
|
|
let mut enc: *mut c_void = ptr::null_mut();
|
|
(API.open_encode_session_ex)(&mut params, &mut enc)
|
|
.result_without_string()
|
|
.map_err(|e| anyhow!("NVENC open_encode_session_ex: {e:?} (no NVIDIA GPU?)"))?;
|
|
|
|
// Seed the P1 + ultra-low-latency preset config.
|
|
let mut preset = nv::NV_ENC_PRESET_CONFIG {
|
|
version: nv::NV_ENC_PRESET_CONFIG_VER,
|
|
presetCfg: nv::NV_ENC_CONFIG {
|
|
version: nv::NV_ENC_CONFIG_VER,
|
|
..Default::default()
|
|
},
|
|
..Default::default()
|
|
};
|
|
if let Err(e) = (API.get_encode_preset_config_ex)(
|
|
enc,
|
|
self.codec_guid,
|
|
nv::NV_ENC_PRESET_P1_GUID,
|
|
nv::NV_ENC_TUNING_INFO::NV_ENC_TUNING_INFO_ULTRA_LOW_LATENCY,
|
|
&mut preset,
|
|
)
|
|
.result_without_string()
|
|
{
|
|
let _ = (API.destroy_encoder)(enc);
|
|
return Err(anyhow!("get_encode_preset_config_ex: {e:?}"));
|
|
}
|
|
let mut cfg = preset.presetCfg;
|
|
|
|
// Mirror the Linux RC config: CBR, infinite GOP, P-only, ~1-frame VBV.
|
|
cfg.gopLength = nv::NVENC_INFINITE_GOPLENGTH;
|
|
cfg.frameIntervalP = 1;
|
|
cfg.rcParams.rateControlMode = nv::NV_ENC_PARAMS_RC_MODE::NV_ENC_PARAMS_RC_CBR;
|
|
let bps = bitrate.min(u32::MAX as u64) as u32;
|
|
cfg.rcParams.averageBitRate = bps;
|
|
cfg.rcParams.maxBitRate = bps;
|
|
// Shrink the VBV with the bitrate — NVENC validates it against the same level ceiling. Only
|
|
// when the GPU advertises custom-VBV support (else leave the preset default, per the caps probe).
|
|
if self.custom_vbv {
|
|
let vbv = (bitrate as f64 / self.fps.max(1) as f64) as u32;
|
|
cfg.rcParams.vbvBufferSize = vbv;
|
|
cfg.rcParams.vbvInitialDelay = vbv;
|
|
}
|
|
|
|
// HIGH tier + autoselect level. The codec's PER-LEVEL bitrate ceiling is otherwise the
|
|
// MAIN-tier cap — for HEVC at 5K that's Level 6.2 Main ≈ 240 Mbps. HIGH tier lifts the HEVC
|
|
// ceiling to ≈800 Mbps (AV1 higher still); autoselect lets NVENC pick the level for the
|
|
// tier+bitrate. `tier`/`level` are u32 (HIGH=1, AUTOSELECT=0); HEVC/AV1 share the union offset.
|
|
cfg.encodeCodecConfig.hevcConfig.tier = 1;
|
|
cfg.encodeCodecConfig.hevcConfig.level = 0;
|
|
|
|
// Chroma + bit depth. Full-chroma 4:4:4 (HEVC Range Extensions) takes precedence and composes
|
|
// with 10-bit (Main 4:4:4 10): NVENC ingests the RGB input (ARGB / ABGR10) and CSCs it to
|
|
// YUV444 internally when `chromaFormatIDC = 3` under the FREXT profile. Only valid on an RGB
|
|
// input — a subsampled NV12/P010 source can't reconstruct full chroma (so the capturer is
|
|
// forced to RGB for a 4:4:4 session, and we guard on the input format here too).
|
|
//
|
|
// ON-GLASS TODO (RTX box): confirm ARGB + chromaFormatIDC=3 + FREXT yields a *true* 4:4:4
|
|
// stream. NVENC's RGB→YUV CSC is documented to honor chromaFormatIDC (unlike libavcodec's
|
|
// wrapper, which always subsamples RGB to 4:2:0 — hence the Linux path feeds planar YUV444
|
|
// instead). If on-glass shows 4:2:0, the follow-up is a BGRA→AYUV shader feeding the native
|
|
// `NV_ENC_BUFFER_FORMAT_AYUV` 4:4:4 input format.
|
|
let rgb_input = matches!(
|
|
self.buffer_fmt,
|
|
nv::NV_ENC_BUFFER_FORMAT::NV_ENC_BUFFER_FORMAT_ARGB
|
|
| nv::NV_ENC_BUFFER_FORMAT::NV_ENC_BUFFER_FORMAT_ABGR10
|
|
);
|
|
if self.chroma_444 && rgb_input {
|
|
cfg.profileGUID = nv::NV_ENC_HEVC_PROFILE_FREXT_GUID;
|
|
cfg.encodeCodecConfig.hevcConfig.set_chromaFormatIDC(3);
|
|
if self.bit_depth == 10 {
|
|
cfg.encodeCodecConfig.hevcConfig.set_pixelBitDepthMinus8(2); // Main 4:4:4 10
|
|
}
|
|
} else if self.bit_depth == 10 {
|
|
// 10-bit HEVC Main10 (HDR foundation): NVENC upconverts the 8-bit input; 8-bit leaves the
|
|
// preset default (Main) untouched.
|
|
cfg.profileGUID = nv::NV_ENC_HEVC_PROFILE_MAIN10_GUID;
|
|
cfg.encodeCodecConfig.hevcConfig.set_pixelBitDepthMinus8(2); // 10 - 8
|
|
}
|
|
|
|
// HDR colour signaling: BT.2020 primaries + SMPTE ST.2084 (PQ) transfer + BT.2020-NCL
|
|
// matrix, limited (studio) range — NVENC's RGB→YUV default. HEVC/H.264 carry it in the VUI;
|
|
// AV1 has NO VUI, so the SAME CICP code points go in the sequence-header colour config
|
|
// (`colorPrimaries`/`transferCharacteristics`/`matrixCoefficients`/`colorRange`). Without
|
|
// this a non-HEVC decoder assumes BT.709 SDR → washed-out / colour-shifted HDR.
|
|
//
|
|
// This is the per-stream colour *description* only. The static mastering-display (ST.2086)
|
|
// and content-light (MaxCLL/MaxFALL) metadata — HEVC SEI / AV1 METADATA OBUs — is a
|
|
// separate follow-up, as is wiring AV1/H.264 to a true 10-bit (Main10) encode (only HEVC
|
|
// sets Main10 above today).
|
|
if self.hdr {
|
|
let prim = nv::NV_ENC_VUI_COLOR_PRIMARIES::NV_ENC_VUI_COLOR_PRIMARIES_BT2020;
|
|
let trc =
|
|
nv::NV_ENC_VUI_TRANSFER_CHARACTERISTIC::NV_ENC_VUI_TRANSFER_CHARACTERISTIC_SMPTE2084;
|
|
let mat = nv::NV_ENC_VUI_MATRIX_COEFFS::NV_ENC_VUI_MATRIX_COEFFS_BT2020_NCL;
|
|
match self.codec {
|
|
Codec::H265 => {
|
|
let vui = &mut cfg.encodeCodecConfig.hevcConfig.hevcVUIParameters;
|
|
vui.videoSignalTypePresentFlag = 1;
|
|
vui.videoFullRangeFlag = 0;
|
|
vui.colourDescriptionPresentFlag = 1;
|
|
vui.colourPrimaries = prim;
|
|
vui.transferCharacteristics = trc;
|
|
vui.colourMatrix = mat;
|
|
}
|
|
Codec::H264 => {
|
|
let vui = &mut cfg.encodeCodecConfig.h264Config.h264VUIParameters;
|
|
vui.videoSignalTypePresentFlag = 1;
|
|
vui.videoFullRangeFlag = 0;
|
|
vui.colourDescriptionPresentFlag = 1;
|
|
vui.colourPrimaries = prim;
|
|
vui.transferCharacteristics = trc;
|
|
vui.colourMatrix = mat;
|
|
}
|
|
Codec::Av1 => {
|
|
let av1 = &mut cfg.encodeCodecConfig.av1Config;
|
|
av1.colorPrimaries = prim;
|
|
av1.transferCharacteristics = trc;
|
|
av1.matrixCoefficients = mat;
|
|
av1.colorRange = 0; // studio/limited swing
|
|
}
|
|
}
|
|
}
|
|
|
|
// Reference-frame invalidation: keep a deeper DPB so an invalidated reference can fall back
|
|
// to an older still-valid frame instead of a full IDR, while `numRefL0 = 1` keeps each
|
|
// P-frame single-reference for low latency. Only when this GPU supports RFI (else leave the
|
|
// preset default — `invalidate_ref_frames` then returns false and the caller forces an IDR).
|
|
if self.rfi_supported {
|
|
let one = nv::NV_ENC_NUM_REF_FRAMES::NV_ENC_NUM_REF_FRAMES_1;
|
|
match self.codec {
|
|
Codec::H264 => {
|
|
cfg.encodeCodecConfig.h264Config.maxNumRefFrames = RFI_DPB;
|
|
cfg.encodeCodecConfig.h264Config.numRefL0 = one;
|
|
}
|
|
Codec::H265 => {
|
|
cfg.encodeCodecConfig.hevcConfig.maxNumRefFramesInDPB = RFI_DPB;
|
|
cfg.encodeCodecConfig.hevcConfig.numRefL0 = one;
|
|
}
|
|
Codec::Av1 => {
|
|
cfg.encodeCodecConfig.av1Config.maxNumRefFramesInDPB = RFI_DPB;
|
|
}
|
|
}
|
|
}
|
|
|
|
let mut init = nv::NV_ENC_INITIALIZE_PARAMS {
|
|
version: nv::NV_ENC_INITIALIZE_PARAMS_VER,
|
|
encodeGUID: self.codec_guid,
|
|
presetGUID: nv::NV_ENC_PRESET_P1_GUID,
|
|
tuningInfo: nv::NV_ENC_TUNING_INFO::NV_ENC_TUNING_INFO_ULTRA_LOW_LATENCY,
|
|
encodeWidth: self.width,
|
|
encodeHeight: self.height,
|
|
darWidth: self.width,
|
|
darHeight: self.height,
|
|
frameRateNum: self.fps,
|
|
frameRateDen: 1,
|
|
enablePTD: 1,
|
|
// Two-thread async retrieve (§5.B): completion events signal the retrieve thread
|
|
// instead of `lock_bitstream` blocking the submit thread.
|
|
enableEncodeAsync: enable_async as u32,
|
|
encodeConfig: &mut cfg,
|
|
..Default::default()
|
|
};
|
|
// splitEncodeMode is a C bitfield — set via the generated accessor, not a struct field.
|
|
init.set_splitEncodeMode(split_mode);
|
|
|
|
match (API.initialize_encoder)(enc, &mut init).result_without_string() {
|
|
Ok(()) => Ok(enc),
|
|
Err(e) => {
|
|
let _ = (API.destroy_encoder)(enc);
|
|
Err(anyhow!("initialize_encoder: {e:?}"))
|
|
}
|
|
}
|
|
}
|
|
|
|
/// Lazily create the session on the first frame's D3D11 device (so capture + encode share it).
|
|
fn init_session(&mut self, device: &ID3D11Device) -> Result<()> {
|
|
// SAFETY: every call below goes through a function pointer resolved once from the loaded
|
|
// `nvidia_video_codec_sdk::ENCODE_API` (`nvEncodeAPI`) table, or through this type's own
|
|
// `unsafe fn`s whose contract is met here. `query_caps`/`try_open_session` receive `device`,
|
|
// the live `ID3D11Device` the caller pulled off the first frame; each returns either a valid
|
|
// open NVENC session handle or an `Err`. `destroy_encoder` is only ever called on a handle a
|
|
// `try_open_session` just returned (and `best` only when `!best.is_null()`), so it never frees
|
|
// a dangling or null session. `create_bitstream_buffer` is passed `enc` — the one chosen live
|
|
// session — and `&mut cb`, a `#[repr(C)] NV_ENC_CREATE_BITSTREAM_BUFFER` whose `version` is set
|
|
// to `NV_ENC_CREATE_BITSTREAM_BUFFER_VER`; `cb` lives across the synchronous call and its
|
|
// returned `bitstreamBuffer` is copied into `self.bitstreams` before `cb` drops. No handle
|
|
// escapes the encode thread.
|
|
unsafe {
|
|
// Probe real GPU caps first (max dims / 10-bit / custom-VBV / RFI) so the config below is
|
|
// gated on what this card supports and an out-of-range mode fails with a clear error
|
|
// rather than being misread as a too-high bitrate by the clamp search.
|
|
self.query_caps(device)?;
|
|
// Bitrate clamp (see the search below): NVENC rejects `initialize_encoder` when the bitrate
|
|
// exceeds the GPU's max codec level. We try the requested rate, then binary-search down to
|
|
// the MAX the level accepts and clamp to it — so an over-asking client (e.g. 1 Gbps on HEVC)
|
|
// gets the highest the GPU can actually do, not a coarse fraction of it.
|
|
const FLOOR_BPS: u64 = 10_000_000;
|
|
let requested_bps = self.bitrate_bps;
|
|
// 2-way NVENC split-frame encoding (Ada dual-NVENC) — the high-pixel-rate throughput lever
|
|
// the Linux host enables via libavcodec `split_encode_mode`. A single Ada NVENC session tops
|
|
// out ~0.8 Gpix/s, so at high motion a 5K@240 (1.77 Gpix/s) frame takes ~8 ms to encode and
|
|
// the rate caps ~125 fps; splitting across both engines roughly halves that. Force 2-way
|
|
// above ~1 Gpix/s (matching encode/linux.rs), AUTO below (the ~2% BD-rate cost isn't worth
|
|
// it at low pixel rates). Env override PUNKTFUNK_SPLIT_ENCODE = 0/disable | 1/auto | 2 | 3.
|
|
// HEVC/AV1 only; the init-failure fallback below disables it if a codec/config rejects it.
|
|
let pixel_rate = self.width as u64 * self.height as u64 * self.fps.max(1) as u64;
|
|
let mut split_mode: u32 = match std::env::var("PUNKTFUNK_SPLIT_ENCODE").ok().as_deref()
|
|
{
|
|
Some("0") | Some("disable") => {
|
|
nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_DISABLE_MODE as u32
|
|
}
|
|
Some("1") | Some("auto") => {
|
|
nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_AUTO_FORCED_MODE as u32
|
|
}
|
|
Some("3") => nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_THREE_FORCED_MODE as u32,
|
|
Some("2") => nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_TWO_FORCED_MODE as u32,
|
|
// Main10 (10-bit / HDR): 2-way split is measurably SLOWER on Ada — at 5120x1440@240
|
|
// Main10, forced-2 took 7.6 ms/frame (~131 fps) vs 2.8 ms (~357 fps) single-engine
|
|
// (the split/merge overhead dominates for 10-bit). A single Ada NVENC engine already
|
|
// handles 5K@240 Main10 well under the 4.17 ms budget, so DON'T split — splitting was
|
|
// the "broken animations in HDR" (the stream capped at ~131 fps). Env still overrides.
|
|
_ if self.bit_depth >= 10 => {
|
|
nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_DISABLE_MODE as u32
|
|
}
|
|
_ if pixel_rate > 1_000_000_000 => {
|
|
nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_TWO_FORCED_MODE as u32
|
|
}
|
|
_ => nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_AUTO_MODE as u32,
|
|
};
|
|
tracing::info!(
|
|
split_mode,
|
|
bit_depth = self.bit_depth,
|
|
pixel_rate,
|
|
"NVENC split-encode mode (0=disable 1=auto-forced 2=two 3=three 4=auto)"
|
|
);
|
|
// Find the highest bitrate the GPU's codec LEVEL accepts and CLAMP to it. NVENC rejects
|
|
// `initialize_encoder` (InvalidParam) when the bitrate exceeds the level ceiling (e.g. a
|
|
// 1 Gbps request on HEVC). Strategy: try the requested rate; if the only problem is a forced
|
|
// split-encode mode the codec doesn't support, disable split and retry; if the bitrate
|
|
// itself is too high, binary-search [FLOOR, requested] for the MAX accepted rate and clamp
|
|
// to THAT (don't undershoot — the old ×¾ step-down landed well below the real ceiling).
|
|
const CLAMP_TOL_BPS: u64 = 20_000_000; // stop bisecting within ~20 Mbps of the ceiling
|
|
|
|
// Two-thread async retrieve: operator opt-in AND the GPU reports async-encode support
|
|
// (query_caps above). Threaded into every session-open probe so the chosen session is
|
|
// built in the right mode from the start.
|
|
let use_async = self.async_supported && async_retrieve_requested();
|
|
|
|
let mut probe = self.try_open_session(device, requested_bps, split_mode, use_async);
|
|
// Disambiguate a forced-split rejection from a bitrate-cap rejection: retry once at the
|
|
// requested rate with split disabled — if THAT succeeds, split was the problem, not bitrate.
|
|
let split_forced = split_mode
|
|
!= nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_AUTO_MODE as u32
|
|
&& split_mode != nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_DISABLE_MODE as u32;
|
|
if probe.is_err() && split_forced {
|
|
let no_split = nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_DISABLE_MODE as u32;
|
|
if let Ok(e) = self.try_open_session(device, requested_bps, no_split, use_async) {
|
|
tracing::warn!("NVENC: split-encode rejected by codec/config — disabled");
|
|
split_mode = no_split;
|
|
probe = Ok(e);
|
|
}
|
|
}
|
|
|
|
let enc = match probe {
|
|
Ok(enc) => {
|
|
self.bitrate_bps = requested_bps;
|
|
enc
|
|
}
|
|
Err(_) => {
|
|
// Requested bitrate exceeds the codec-level ceiling — binary-search the max accepted.
|
|
// `lo` is the highest known-good rate (FLOOR is assumed to fit), `hi` the lowest
|
|
// rejected; `best` holds the live session at `lo` so we end up with the clamped one.
|
|
let mut lo = FLOOR_BPS;
|
|
let mut hi = requested_bps;
|
|
let mut best: *mut c_void = ptr::null_mut();
|
|
let mut best_bps = 0u64;
|
|
while hi > lo + CLAMP_TOL_BPS {
|
|
let mid = lo + (hi - lo) / 2;
|
|
match self.try_open_session(device, mid, split_mode, use_async) {
|
|
Ok(e) => {
|
|
if !best.is_null() {
|
|
let _ = (API.destroy_encoder)(best);
|
|
}
|
|
best = e;
|
|
best_bps = mid;
|
|
lo = mid;
|
|
}
|
|
Err(_) => hi = mid,
|
|
}
|
|
}
|
|
if best.is_null() {
|
|
// Nothing in (FLOOR, requested] accepted — fall back to the floor itself, also
|
|
// trying split-disabled in case a forced split (not the bitrate) is the blocker.
|
|
let no_split =
|
|
nv::NV_ENC_SPLIT_ENCODE_MODE::NV_ENC_SPLIT_DISABLE_MODE as u32;
|
|
best = self
|
|
.try_open_session(device, FLOOR_BPS, split_mode, use_async)
|
|
.or_else(|_| {
|
|
self.try_open_session(device, FLOOR_BPS, no_split, use_async)
|
|
})
|
|
.context(
|
|
"NVENC initialize_encoder rejected even at the floor bitrate",
|
|
)?;
|
|
best_bps = FLOOR_BPS;
|
|
}
|
|
tracing::warn!(
|
|
requested_mbps = requested_bps / 1_000_000,
|
|
clamped_mbps = best_bps / 1_000_000,
|
|
"NVENC: requested bitrate above the GPU codec-level ceiling — clamped to the max accepted"
|
|
);
|
|
self.bitrate_bps = best_bps;
|
|
best
|
|
}
|
|
};
|
|
self.encoder = enc;
|
|
if self.bitrate_bps < requested_bps {
|
|
tracing::info!(
|
|
requested_mbps = requested_bps / 1_000_000,
|
|
applied_mbps = self.bitrate_bps / 1_000_000,
|
|
"NVENC bitrate capped to this GPU's max for the codec"
|
|
);
|
|
}
|
|
|
|
// 5. one output bitstream per in-flight slot. There is NO encoder-owned input pool: the
|
|
// capturer's textures are registered on demand in `submit` and encoded in place.
|
|
for _ in 0..POOL {
|
|
let mut cb = nv::NV_ENC_CREATE_BITSTREAM_BUFFER {
|
|
version: nv::NV_ENC_CREATE_BITSTREAM_BUFFER_VER,
|
|
..Default::default()
|
|
};
|
|
(API.create_bitstream_buffer)(enc, &mut cb)
|
|
.result_without_string()
|
|
.map_err(|e| anyhow!("create_bitstream_buffer: {e:?}"))?;
|
|
self.bitstreams.push(cb.bitstreamBuffer);
|
|
}
|
|
// Async retrieve: one auto-reset completion event per pool bitstream, registered with
|
|
// the session, plus the retrieve thread the events signal. The thread only ever sees
|
|
// raw addresses; `teardown` joins it before any of them die.
|
|
if use_async {
|
|
for _ in 0..POOL {
|
|
let ev = CreateEventW(None, false, false, PCWSTR::null())
|
|
.context("CreateEvent (NVENC completion)")?;
|
|
let mut ep = nv::NV_ENC_EVENT_PARAMS {
|
|
version: nv::NV_ENC_EVENT_PARAMS_VER,
|
|
completionEvent: ev.0,
|
|
..Default::default()
|
|
};
|
|
(API.register_async_event)(enc, &mut ep)
|
|
.result_without_string()
|
|
.map_err(|e| anyhow!("register_async_event: {e:?}"))?;
|
|
self.events.push(ev.0 as usize);
|
|
}
|
|
let (work_tx, work_rx) = mpsc::sync_channel::<RetrieveJob>(POOL);
|
|
let (done_tx, done_rx) = mpsc::channel::<RetrieveDone>();
|
|
let enc_addr = enc as usize;
|
|
let join = std::thread::Builder::new()
|
|
.name("punktfunk-nvenc-out".into())
|
|
.spawn(move || retrieve_loop(enc_addr, work_rx, done_tx))
|
|
.context("spawn NVENC retrieve thread")?;
|
|
self.async_rt = Some(AsyncRetrieve {
|
|
work_tx: Some(work_tx),
|
|
done_rx,
|
|
join: Some(join),
|
|
ready: VecDeque::new(),
|
|
});
|
|
tracing::info!(
|
|
pool = POOL,
|
|
"NVENC async retrieve active (two-thread encode: submit here, \
|
|
lock_bitstream on the retrieve thread)"
|
|
);
|
|
}
|
|
self.inited = true;
|
|
tracing::info!(
|
|
"NVENC D3D11 session: {}x{}@{} {}-bit{} {} Mbps {:?}",
|
|
self.width,
|
|
self.height,
|
|
self.fps,
|
|
self.bit_depth,
|
|
if self.hdr { " HDR(BT.2020 PQ)" } else { "" },
|
|
self.bitrate_bps / 1_000_000,
|
|
self.codec_guid
|
|
);
|
|
Ok(())
|
|
}
|
|
}
|
|
|
|
/// Fold one retrieve-thread completion back into encoder state ON THE ENCODE THREAD: pop the
|
|
/// oldest `pending` entry (completions are FIFO — one retrieve thread, in-order jobs), verify
|
|
/// the bitstream pairing, unmap the input resource, and queue the AU for `poll`. A retrieve
|
|
/// error surfaces AFTER the unmap (the resource is retired either way) so the session glue's
|
|
/// rebuild path starts from clean state.
|
|
fn absorb_done(&mut self, done: RetrieveDone) -> Result<()> {
|
|
let Some((bs, map, pts_ns)) = self.pending.pop_front() else {
|
|
bail!("NVENC async: completion with no in-flight frame (pairing bug)");
|
|
};
|
|
if bs as usize != done.bs {
|
|
bail!("NVENC async: completion out of order (pairing bug)");
|
|
}
|
|
// SAFETY: `map` is the mapped input `submit` recorded for exactly this now-completed
|
|
// encode; the session is live (`async_rt` exists only between `init_session` and
|
|
// `teardown`) and this runs on the encode thread — the single unmap here mirrors the sync
|
|
// path's poll-side unmap, exactly once per mapping.
|
|
unsafe {
|
|
if !map.is_null() {
|
|
let _ = (API.unmap_input_resource)(self.encoder, map);
|
|
}
|
|
}
|
|
let (data, keyframe) = done.result.map_err(|e| anyhow!("{e}"))?;
|
|
self.async_rt
|
|
.as_mut()
|
|
.expect("absorb_done is only reachable in async mode")
|
|
.ready
|
|
.push_back(EncodedFrame {
|
|
data,
|
|
pts_ns,
|
|
keyframe,
|
|
});
|
|
Ok(())
|
|
}
|
|
}
|
|
|
|
impl Encoder for NvencD3d11Encoder {
|
|
fn submit(&mut self, captured: &CapturedFrame) -> Result<()> {
|
|
let frame = match &captured.payload {
|
|
FramePayload::D3d11(f) => f,
|
|
FramePayload::Cpu(_) => {
|
|
bail!("NVENC D3D11 encoder needs a GPU texture frame (use the software encoder for CPU frames)")
|
|
}
|
|
};
|
|
// The capturer recreates its D3D11 device on a desktop switch (secure/Winlogon) and may come
|
|
// back at a different resolution (user session applies its own mode on login). Re-init when the
|
|
// frame arrives on a different device OR at a different size than our session was built on.
|
|
// HDR (BT.2020 PQ 10-bit) when the capturer hands us a 10-bit R10G10B10A2 frame. This can flip
|
|
// mid-session when the user toggles HDR (which arrives as a capture device recreate anyway).
|
|
// HDR (BT.2020 PQ) when the capturer hands a 10-bit frame — either R10G10B10A2 (the legacy
|
|
// shader path) or P010 (the video-processor path). 8-bit NV12/ARGB → SDR.
|
|
let hdr = matches!(captured.format, PixelFormat::Rgb10a2 | PixelFormat::P010);
|
|
let dev_raw = frame.device.as_raw();
|
|
let size_changed =
|
|
self.inited && (self.width != captured.width || self.height != captured.height);
|
|
let hdr_changed = self.inited && self.hdr != hdr;
|
|
if self.inited && (self.init_device != dev_raw || size_changed || hdr_changed) {
|
|
tracing::info!(
|
|
device_changed = self.init_device != dev_raw,
|
|
size_changed,
|
|
hdr_changed,
|
|
hdr,
|
|
new = format!("{}x{}", captured.width, captured.height),
|
|
"NVENC: capture device/size/HDR changed — re-initializing session"
|
|
);
|
|
// SAFETY: `teardown` (an `unsafe fn`) requires the encode thread with no NVENC call in
|
|
// flight and a session whose cached regs/bitstreams/pending all belong to `self.encoder`.
|
|
// All hold: this is the synchronous encode thread, `self.inited` so `self.encoder` is the
|
|
// live session every cached resource was created against, and the previous frame's encode
|
|
// has already been polled (synchronous submit→poll), so nothing is mid-encode.
|
|
unsafe { self.teardown() };
|
|
}
|
|
if !self.inited {
|
|
// Adopt the current frame size + colour so the encoder always matches the capturer output.
|
|
self.width = captured.width;
|
|
self.height = captured.height;
|
|
self.hdr = hdr;
|
|
// Pick the NVENC input format from the captured pixel format. YUV (NV12/P010) is the
|
|
// video-processor path — NVENC encodes it natively (no internal RGB→YUV, which is a hidden
|
|
// 3D/compute step that would fight a GPU-saturating game). RGB (ARGB/ABGR10) is the legacy
|
|
// shader path. 10-bit (P010/ABGR10) forces HEVC Main10 + the BT.2020 PQ VUI.
|
|
self.buffer_fmt = match captured.format {
|
|
PixelFormat::P010 => {
|
|
self.bit_depth = 10;
|
|
nv::NV_ENC_BUFFER_FORMAT::NV_ENC_BUFFER_FORMAT_YUV420_10BIT
|
|
}
|
|
PixelFormat::Rgb10a2 => {
|
|
self.bit_depth = 10;
|
|
nv::NV_ENC_BUFFER_FORMAT::NV_ENC_BUFFER_FORMAT_ABGR10
|
|
}
|
|
PixelFormat::Nv12 => {
|
|
// NV12 is 8-bit 4:2:0. Force 8-bit so a transition from a prior P010 (10-bit) session
|
|
// — or a 10-bit-negotiated client on an SDR display — re-inits at the matching depth.
|
|
// Unlike ARGB (which NVENC upconverts to Main10), NV12 cannot feed a 10-bit session:
|
|
// `register_resource` rejects it as InvalidParam (the HDR→SDR-toggle stream drop).
|
|
self.bit_depth = 8;
|
|
nv::NV_ENC_BUFFER_FORMAT::NV_ENC_BUFFER_FORMAT_NV12
|
|
}
|
|
_ => nv::NV_ENC_BUFFER_FORMAT::NV_ENC_BUFFER_FORMAT_ARGB,
|
|
};
|
|
let device = frame.device.clone();
|
|
self.init_session(&device)?;
|
|
self.init_device = dev_raw;
|
|
}
|
|
// Async backpressure: never hand NVENC an output bitstream that is still in flight, and
|
|
// keep in-flight depth within the capturer's texture ring (see `async_inflight_cap`). At
|
|
// the cap, block on the OLDEST completion (the retrieve thread is already waiting on its
|
|
// event) before submitting more — bounding depth exactly like the sync path's per-tick
|
|
// blocking poll, just `cap` deep instead of 1.
|
|
while self.async_rt.is_some() && self.pending.len() >= async_inflight_cap() {
|
|
let done = {
|
|
let rt = self.async_rt.as_mut().expect("checked in loop condition");
|
|
rt.done_rx
|
|
.recv_timeout(std::time::Duration::from_secs(5))
|
|
.map_err(|_| anyhow!("NVENC async retrieve stalled (5s) — encoder wedged?"))?
|
|
};
|
|
self.absorb_done(done)?;
|
|
}
|
|
let slot = self.next % POOL;
|
|
self.next += 1;
|
|
// SAFETY: every NVENC call goes through a function pointer from the loaded `ENCODE_API` table
|
|
// and takes `self.encoder`, the live session `init_session` just established (non-null on the
|
|
// path that reaches here). `NV_ENC_REGISTER_RESOURCE rr` has `version =
|
|
// NV_ENC_REGISTER_RESOURCE_VER` and registers `frame.texture` — a D3D11 texture from
|
|
// `frame.device`, which is the SAME device the session was opened against (any device change
|
|
// tears down and re-inits above, so `init_device == frame.device.as_raw()` here); the cloned
|
|
// `ID3D11Texture2D` is kept alive in `regs` so NVENC's registration never outlives the texture.
|
|
// `mp` (`NV_ENC_MAP_INPUT_RESOURCE`, version set) maps that registration and the map is recorded
|
|
// in `pending` to be unmapped exactly once in `poll`/`teardown`. `pic` (`NV_ENC_PIC_PARAMS`,
|
|
// version set) points `inputBuffer` at `mp.mappedResource` and `outputBitstream` at the live
|
|
// pool bitstream `bitstreams[slot]`; the optional SEI scratch (`mastering_sei`/`cll_sei` and the
|
|
// `sei` Vec whose `as_mut_ptr()` is written into the codec union) are stack locals that outlive
|
|
// the synchronous `encode_picture`. Every `#[repr(C)]` param is a live local borrowed `&mut`
|
|
// for the duration of its one synchronous call. (In-place encode without `CopyResource` is
|
|
// sound because the encode loop is synchronous, as the module docs state.)
|
|
unsafe {
|
|
// Register the capturer's texture with NVENC once (cached by raw pointer), then encode it
|
|
// IN PLACE — no `CopyResource` into an encoder-owned pool. This is the zero-copy win: the
|
|
// capturer already produced a stable GPU texture; we just register + map + encode it.
|
|
let key = frame.texture.as_raw() as isize;
|
|
if !self.regs.contains_key(&key) {
|
|
let mut rr = nv::NV_ENC_REGISTER_RESOURCE {
|
|
version: nv::NV_ENC_REGISTER_RESOURCE_VER,
|
|
resourceType:
|
|
nv::NV_ENC_INPUT_RESOURCE_TYPE::NV_ENC_INPUT_RESOURCE_TYPE_DIRECTX,
|
|
width: self.width,
|
|
height: self.height,
|
|
pitch: 0,
|
|
resourceToRegister: frame.texture.as_raw(),
|
|
bufferFormat: self.buffer_fmt,
|
|
bufferUsage: nv::NV_ENC_BUFFER_USAGE::NV_ENC_INPUT_IMAGE,
|
|
..Default::default()
|
|
};
|
|
(API.register_resource)(self.encoder, &mut rr)
|
|
.result_without_string()
|
|
.map_err(|e| anyhow!("register_resource: {e:?}"))?;
|
|
self.regs
|
|
.insert(key, (rr.registeredResource, frame.texture.clone()));
|
|
}
|
|
let reg = self.regs[&key].0;
|
|
|
|
let mut mp = nv::NV_ENC_MAP_INPUT_RESOURCE {
|
|
version: nv::NV_ENC_MAP_INPUT_RESOURCE_VER,
|
|
registeredResource: reg,
|
|
..Default::default()
|
|
};
|
|
(API.map_input_resource)(self.encoder, &mut mp)
|
|
.result_without_string()
|
|
.map_err(|e| anyhow!("map_input_resource: {e:?}"))?;
|
|
|
|
let pts = self.frame_idx as u64;
|
|
self.frame_idx += 1;
|
|
let flags = if std::mem::take(&mut self.force_kf) {
|
|
nv::NV_ENC_PIC_FLAGS::NV_ENC_PIC_FLAG_FORCEIDR as u32
|
|
| nv::NV_ENC_PIC_FLAGS::NV_ENC_PIC_FLAG_OUTPUT_SPSPPS as u32
|
|
} else {
|
|
0
|
|
};
|
|
let mut pic = nv::NV_ENC_PIC_PARAMS {
|
|
version: nv::NV_ENC_PIC_PARAMS_VER,
|
|
inputWidth: self.width,
|
|
inputHeight: self.height,
|
|
inputPitch: 0,
|
|
inputBuffer: mp.mappedResource,
|
|
bufferFmt: mp.mappedBufferFmt,
|
|
outputBitstream: self.bitstreams[slot],
|
|
pictureStruct: nv::NV_ENC_PIC_STRUCT::NV_ENC_PIC_STRUCT_FRAME,
|
|
inputTimeStamp: pts,
|
|
encodePicFlags: flags as u32,
|
|
// Async mode: the event the driver signals when this encode completes (the
|
|
// retrieve thread waits on it). Null in sync mode (`events` is empty).
|
|
completionEvent: self
|
|
.events
|
|
.get(slot)
|
|
.map(|&e| e as *mut c_void)
|
|
.unwrap_or(ptr::null_mut()),
|
|
..Default::default()
|
|
};
|
|
|
|
// In-band HDR10 SEI on every IDR (a forced keyframe, or the first frame NVENC opens with):
|
|
// `mastering_display_colour_volume` (ST.2086) + `content_light_level_info` (CEA-861.3),
|
|
// built from the source display's metadata. Any decoder — incl. stock Moonlight — then
|
|
// tone-maps from the real grade. HEVC/H.264 carry SEI; AV1 uses metadata OBUs (follow-up).
|
|
// The scratch buffers must outlive `encode_picture`, so they live in this scope.
|
|
let is_idr = flags != 0 || pts == 0;
|
|
let mastering_sei = self
|
|
.hdr_meta
|
|
.map(|m| crate::hdr::hevc_mastering_display_sei(&m));
|
|
let cll_sei = self
|
|
.hdr_meta
|
|
.map(|m| crate::hdr::hevc_content_light_level_sei(&m));
|
|
let mut sei: Vec<nv::NV_ENC_SEI_PAYLOAD> = Vec::new();
|
|
if is_idr && self.hdr {
|
|
if let Some(p) = mastering_sei.as_ref() {
|
|
sei.push(nv::NV_ENC_SEI_PAYLOAD {
|
|
payloadSize: p.len() as u32,
|
|
payloadType: crate::hdr::SEI_TYPE_MASTERING_DISPLAY_COLOUR_VOLUME,
|
|
payload: p.as_ptr() as *mut u8,
|
|
});
|
|
}
|
|
if let Some(p) = cll_sei.as_ref() {
|
|
sei.push(nv::NV_ENC_SEI_PAYLOAD {
|
|
payloadSize: p.len() as u32,
|
|
payloadType: crate::hdr::SEI_TYPE_CONTENT_LIGHT_LEVEL_INFO,
|
|
payload: p.as_ptr() as *mut u8,
|
|
});
|
|
}
|
|
}
|
|
if !sei.is_empty() {
|
|
// Writing a union field is safe; the pointers/len are read during encode_picture.
|
|
match self.codec {
|
|
Codec::H265 => {
|
|
pic.codecPicParams.hevcPicParams.seiPayloadArray = sei.as_mut_ptr();
|
|
pic.codecPicParams.hevcPicParams.seiPayloadArrayCnt = sei.len() as u32;
|
|
}
|
|
Codec::H264 => {
|
|
pic.codecPicParams.h264PicParams.seiPayloadArray = sei.as_mut_ptr();
|
|
pic.codecPicParams.h264PicParams.seiPayloadArrayCnt = sei.len() as u32;
|
|
}
|
|
// AV1 mastering/CLL ride METADATA OBUs, not SEI — separate follow-up.
|
|
Codec::Av1 => {}
|
|
}
|
|
}
|
|
(API.encode_picture)(self.encoder, &mut pic)
|
|
.result_without_string()
|
|
.map_err(|e| anyhow!("encode_picture: {e:?}"))?;
|
|
self.pending
|
|
.push_back((self.bitstreams[slot], mp.mappedResource, captured.pts_ns));
|
|
// Async: hand the in-flight encode to the retrieve thread (channel capacity = POOL ≥
|
|
// in-flight, so this send never blocks). The pending entry above pairs with its
|
|
// completion FIFO in `absorb_done`.
|
|
if let Some(rt) = &self.async_rt {
|
|
let job = RetrieveJob {
|
|
bs: self.bitstreams[slot] as usize,
|
|
event: self.events[slot],
|
|
};
|
|
if rt.work_tx.as_ref().is_none_or(|tx| tx.send(job).is_err()) {
|
|
bail!("NVENC retrieve thread gone — rebuilding the session");
|
|
}
|
|
}
|
|
}
|
|
Ok(())
|
|
}
|
|
|
|
fn request_keyframe(&mut self) {
|
|
self.force_kf = true;
|
|
}
|
|
|
|
fn caps(&self) -> EncoderCaps {
|
|
// RFI is probed once at open (`rfi_supported`); HDR SEI rides keyframes whenever the
|
|
// session is in HDR mode. Both are the real capabilities the session glue routes on.
|
|
EncoderCaps {
|
|
supports_rfi: self.rfi_supported,
|
|
supports_hdr_metadata: self.hdr,
|
|
// Reflects what the session actually configured (cleared in `query_caps` if the GPU lacks
|
|
// YUV444 encode), so the glue can confirm 4:4:4 vs the negotiated request.
|
|
chroma_444: self.chroma_444,
|
|
// The direct-NVENC path recovers via real RFI (or a forced IDR), not the Linux
|
|
// libavcodec intra-refresh mode.
|
|
intra_refresh: false,
|
|
}
|
|
}
|
|
|
|
fn set_hdr_meta(&mut self, meta: Option<punktfunk_core::quic::HdrMeta>) {
|
|
// Stored and emitted as in-band SEI on the next keyframe (see `submit`). Cheap to call every
|
|
// frame; only changes when the source is regraded or HDR toggles.
|
|
self.hdr_meta = meta;
|
|
}
|
|
|
|
fn invalidate_ref_frames(&mut self, first: i64, last: i64) -> bool {
|
|
// No live session, the GPU can't invalidate, or a nonsense range → caller forces a full IDR.
|
|
// (NVENC handles are single-threaded; this runs on the encode thread, like submit/poll.)
|
|
if self.encoder.is_null() || !self.rfi_supported || first < 0 || first > last {
|
|
return false;
|
|
}
|
|
// Already invalidated a covering range for this loss event — nothing more to do, no IDR.
|
|
if let Some((pf, pl)) = self.last_rfi_range {
|
|
if first >= pf && last <= pl {
|
|
return true;
|
|
}
|
|
}
|
|
// `frame_idx` is the NEXT timestamp to assign, so the last encoded frame is `frame_idx - 1`
|
|
// and the DPB holds `[frame_idx - RFI_DPB, frame_idx - 1]`. A lost frame older than that
|
|
// can't be invalidated, so the only correct recovery is an IDR.
|
|
let oldest_in_dpb = self.frame_idx - RFI_DPB as i64;
|
|
if first < oldest_in_dpb {
|
|
return false;
|
|
}
|
|
// Clamp to frames we've actually encoded (don't invalidate a timestamp we never assigned).
|
|
let last = last.min(self.frame_idx - 1);
|
|
if first > last {
|
|
return false;
|
|
}
|
|
// We tag each input with `inputTimeStamp = frame_idx` (0,1,2,…), which is also the client's
|
|
// frame number (the packetizer numbers frames in submit order), so the client's lost-frame
|
|
// range maps 1:1 onto the timestamps NVENC invalidates here.
|
|
// SAFETY: `invalidate_ref_frames` is a function pointer from the loaded `ENCODE_API` table.
|
|
// `self.encoder` was checked non-null at the top of this fn and is the live session; this runs
|
|
// on the encode thread (like submit/poll), so there is no concurrent NVENC use. Each `ts` was
|
|
// clamped to `[oldest_in_dpb, frame_idx - 1]` above, so it names a frame still in the session's
|
|
// DPB; the call passes only that `u64` timestamp (no struct), so there is no struct-size or
|
|
// lifetime concern.
|
|
unsafe {
|
|
for ts in first..=last {
|
|
if (API.invalidate_ref_frames)(self.encoder, ts as u64)
|
|
.result_without_string()
|
|
.is_err()
|
|
{
|
|
return false; // any failure → fall back to IDR
|
|
}
|
|
}
|
|
}
|
|
self.last_rfi_range = Some((first, last));
|
|
true
|
|
}
|
|
|
|
fn poll(&mut self) -> Result<Option<EncodedFrame>> {
|
|
// Async mode: drain whatever the retrieve thread has finished (non-blocking) and hand out
|
|
// the oldest ready AU. `None` = nothing completed yet — the session loop keeps the frame
|
|
// in flight and re-polls next tick, capture never blocks on the WDDM scheduling wait.
|
|
if self.async_rt.is_some() {
|
|
loop {
|
|
let done = match self
|
|
.async_rt
|
|
.as_mut()
|
|
.expect("checked just above")
|
|
.done_rx
|
|
.try_recv()
|
|
{
|
|
Ok(d) => d,
|
|
Err(_) => break,
|
|
};
|
|
self.absorb_done(done)?;
|
|
}
|
|
return Ok(self
|
|
.async_rt
|
|
.as_mut()
|
|
.expect("checked just above")
|
|
.ready
|
|
.pop_front());
|
|
}
|
|
let Some((bs, map, pts_ns)) = self.pending.pop_front() else {
|
|
return Ok(None);
|
|
};
|
|
// SAFETY: a non-empty `pending` implies `submit` ran, so `self.encoder` is the live session
|
|
// (`teardown` clears `pending` whenever it nulls the handle); all calls below use function
|
|
// pointers from the loaded `ENCODE_API` table on the encode thread. `NV_ENC_LOCK_BITSTREAM lock`
|
|
// (version = `NV_ENC_LOCK_BITSTREAM_VER`) locks `bs`, a pool bitstream a prior `encode_picture`
|
|
// targeted; `lock_bitstream` blocks until that encode finishes, so on success
|
|
// `lock.bitstreamBufferPtr` is non-null and points at `lock.bitstreamSizeInBytes` bytes of
|
|
// NVENC-owned, CPU-readable output valid until `unlock_bitstream`. The `from_raw_parts` slice is
|
|
// only read (copied via `to_vec()`) BEFORE `unlock_bitstream(bs)` — lock and unlock pair on the
|
|
// same buffer — so it never outlives the lock. `map` (the input resource paired with `bs` in
|
|
// `pending`) is unmapped here, after the encode completed, exactly once.
|
|
unsafe {
|
|
let mut lock = nv::NV_ENC_LOCK_BITSTREAM {
|
|
version: nv::NV_ENC_LOCK_BITSTREAM_VER,
|
|
outputBitstream: bs,
|
|
..Default::default()
|
|
};
|
|
(API.lock_bitstream)(self.encoder, &mut lock)
|
|
.result_without_string()
|
|
.map_err(|e| anyhow!("lock_bitstream: {e:?}"))?;
|
|
let data = std::slice::from_raw_parts(
|
|
lock.bitstreamBufferPtr as *const u8,
|
|
lock.bitstreamSizeInBytes as usize,
|
|
)
|
|
.to_vec();
|
|
let keyframe = matches!(
|
|
lock.pictureType,
|
|
nv::NV_ENC_PIC_TYPE::NV_ENC_PIC_TYPE_IDR | nv::NV_ENC_PIC_TYPE::NV_ENC_PIC_TYPE_I
|
|
);
|
|
(API.unlock_bitstream)(self.encoder, bs)
|
|
.result_without_string()
|
|
.map_err(|e| anyhow!("unlock_bitstream: {e:?}"))?;
|
|
if !map.is_null() {
|
|
let _ = (API.unmap_input_resource)(self.encoder, map);
|
|
}
|
|
Ok(Some(EncodedFrame {
|
|
data,
|
|
pts_ns,
|
|
keyframe,
|
|
}))
|
|
}
|
|
}
|
|
|
|
fn flush(&mut self) -> Result<()> {
|
|
Ok(()) // P1/ULL + frameIntervalP=1: each submit yields its AU; no internal queue to drain.
|
|
}
|
|
}
|
|
|
|
impl Drop for NvencD3d11Encoder {
|
|
fn drop(&mut self) {
|
|
// SAFETY: `teardown` (an `unsafe fn`) needs the owning thread with no NVENC call in flight and
|
|
// a session whose cached resources all belong to `self.encoder`. At Drop this encoder is owned
|
|
// exclusively (no other reference can exist), runs on the encode thread it was confined to, and
|
|
// `teardown` early-returns when `self.encoder` is null; otherwise every cached reg/bitstream/
|
|
// pending was created against that live session. It runs exactly once (here).
|
|
unsafe { self.teardown() };
|
|
}
|
|
}
|
|
|
|
/// Probe whether the active NVIDIA GPU can encode HEVC **4:4:4** (`NV_ENC_CAPS_SUPPORT_YUV444_ENCODE`).
|
|
/// Creates a throwaway hardware D3D11 device + NVENC session, queries the cap, and tears down. HEVC-only;
|
|
/// the result is cached by the caller ([`crate::encode::can_encode_444`]) and read *before* the Welcome
|
|
/// so the host advertises the chroma it can really encode (honest downgrade to 4:2:0 on a card without it).
|
|
pub fn probe_can_encode_444(codec: Codec) -> bool {
|
|
use windows::Win32::Foundation::HMODULE;
|
|
use windows::Win32::Graphics::Direct3D::{
|
|
D3D_DRIVER_TYPE_HARDWARE, D3D_DRIVER_TYPE_UNKNOWN, D3D_FEATURE_LEVEL_11_0,
|
|
};
|
|
use windows::Win32::Graphics::Direct3D11::{
|
|
D3D11CreateDevice, D3D11_CREATE_DEVICE_BGRA_SUPPORT, D3D11_SDK_VERSION,
|
|
};
|
|
use windows::Win32::Graphics::Dxgi::{CreateDXGIFactory1, IDXGIAdapter1, IDXGIFactory4};
|
|
if codec != Codec::H265 {
|
|
return false;
|
|
}
|
|
// SAFETY: a self-contained probe owning every handle it creates. `CreateDXGIFactory1`/
|
|
// `EnumAdapterByLuid` return owned COM objects or err (→ default-adapter fallback).
|
|
// `D3D11CreateDevice` (explicit adapter + UNKNOWN driver type, or NULL adapter + HARDWARE)
|
|
// fills `device` or returns Err (→ false). `open_encode_session_ex` opens an NVENC session
|
|
// against that device's raw pointer (valid while `device` is held) or errors (→ false, tearing
|
|
// nothing down). `get_encode_caps` reads one scalar cap into `val` via the loaded API table.
|
|
// `destroy_encoder` frees the session exactly once; `device`/its context drop with the COM
|
|
// wrappers. No handle escapes this call and nothing runs concurrently.
|
|
unsafe {
|
|
// Probe on the SELECTED render adapter — the GPU the session will actually encode on
|
|
// (web-console preference / PUNKTFUNK_RENDER_ADAPTER / max VRAM). The OS default adapter
|
|
// (NULL) can be the *other* GPU on a hybrid box, answering for hardware we won't use.
|
|
let adapter: Option<IDXGIAdapter1> = crate::win_adapter::resolve_render_adapter_luid()
|
|
.and_then(|luid| {
|
|
let factory: IDXGIFactory4 = CreateDXGIFactory1().ok()?;
|
|
factory.EnumAdapterByLuid(luid).ok()
|
|
});
|
|
let mut device: Option<ID3D11Device> = None;
|
|
let created = match &adapter {
|
|
Some(a) => D3D11CreateDevice(
|
|
a,
|
|
D3D_DRIVER_TYPE_UNKNOWN,
|
|
HMODULE::default(),
|
|
D3D11_CREATE_DEVICE_BGRA_SUPPORT,
|
|
Some(&[D3D_FEATURE_LEVEL_11_0]),
|
|
D3D11_SDK_VERSION,
|
|
Some(&mut device),
|
|
None,
|
|
None,
|
|
),
|
|
None => D3D11CreateDevice(
|
|
None,
|
|
D3D_DRIVER_TYPE_HARDWARE,
|
|
HMODULE::default(),
|
|
D3D11_CREATE_DEVICE_BGRA_SUPPORT,
|
|
Some(&[D3D_FEATURE_LEVEL_11_0]),
|
|
D3D11_SDK_VERSION,
|
|
Some(&mut device),
|
|
None,
|
|
None,
|
|
),
|
|
};
|
|
if created.is_err() {
|
|
return false;
|
|
}
|
|
let Some(device) = device else { return false };
|
|
let mut params = nv::NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS {
|
|
version: nv::NV_ENC_OPEN_ENCODE_SESSION_EX_PARAMS_VER,
|
|
deviceType: nv::NV_ENC_DEVICE_TYPE::NV_ENC_DEVICE_TYPE_DIRECTX,
|
|
device: device.as_raw(),
|
|
apiVersion: nv::NVENCAPI_VERSION,
|
|
..Default::default()
|
|
};
|
|
let mut enc: *mut c_void = ptr::null_mut();
|
|
if (API.open_encode_session_ex)(&mut params, &mut enc)
|
|
.result_without_string()
|
|
.is_err()
|
|
{
|
|
return false;
|
|
}
|
|
let mut param = nv::NV_ENC_CAPS_PARAM {
|
|
version: nv::NV_ENC_CAPS_PARAM_VER,
|
|
capsToQuery: nv::NV_ENC_CAPS::NV_ENC_CAPS_SUPPORT_YUV444_ENCODE,
|
|
reserved: [0; 62],
|
|
};
|
|
let mut val: i32 = 0;
|
|
let ok = (API.get_encode_caps)(enc, nv::NV_ENC_CODEC_HEVC_GUID, &mut param, &mut val)
|
|
.result_without_string()
|
|
.is_ok()
|
|
&& val != 0;
|
|
let _ = (API.destroy_encoder)(enc);
|
|
ok
|
|
}
|
|
}
|